mm: remove flags argument to mmap_region
[~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 #include <linux/sched/sysctl.h>
36
37 #include <asm/uaccess.h>
38 #include <asm/cacheflush.h>
39 #include <asm/tlb.h>
40 #include <asm/mmu_context.h>
41
42 #include "internal.h"
43
44 #ifndef arch_mmap_check
45 #define arch_mmap_check(addr, len, flags)       (0)
46 #endif
47
48 #ifndef arch_rebalance_pgtables
49 #define arch_rebalance_pgtables(addr, len)              (addr)
50 #endif
51
52 static void unmap_region(struct mm_struct *mm,
53                 struct vm_area_struct *vma, struct vm_area_struct *prev,
54                 unsigned long start, unsigned long end);
55
56 /* description of effects of mapping type and prot in current implementation.
57  * this is due to the limited x86 page protection hardware.  The expected
58  * behavior is in parens:
59  *
60  * map_type     prot
61  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
62  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
63  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
64  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
65  *              
66  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
67  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
68  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
69  *
70  */
71 pgprot_t protection_map[16] = {
72         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
73         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
74 };
75
76 pgprot_t vm_get_page_prot(unsigned long vm_flags)
77 {
78         return __pgprot(pgprot_val(protection_map[vm_flags &
79                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
80                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
81 }
82 EXPORT_SYMBOL(vm_get_page_prot);
83
84 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
85 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
86 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
87 /*
88  * Make sure vm_committed_as in one cacheline and not cacheline shared with
89  * other variables. It can be updated by several CPUs frequently.
90  */
91 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
92
93 /*
94  * The global memory commitment made in the system can be a metric
95  * that can be used to drive ballooning decisions when Linux is hosted
96  * as a guest. On Hyper-V, the host implements a policy engine for dynamically
97  * balancing memory across competing virtual machines that are hosted.
98  * Several metrics drive this policy engine including the guest reported
99  * memory commitment.
100  */
101 unsigned long vm_memory_committed(void)
102 {
103         return percpu_counter_read_positive(&vm_committed_as);
104 }
105 EXPORT_SYMBOL_GPL(vm_memory_committed);
106
107 /*
108  * Check that a process has enough memory to allocate a new virtual
109  * mapping. 0 means there is enough memory for the allocation to
110  * succeed and -ENOMEM implies there is not.
111  *
112  * We currently support three overcommit policies, which are set via the
113  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
114  *
115  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
116  * Additional code 2002 Jul 20 by Robert Love.
117  *
118  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
119  *
120  * Note this is a helper function intended to be used by LSMs which
121  * wish to use this logic.
122  */
123 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
124 {
125         unsigned long free, allowed;
126
127         vm_acct_memory(pages);
128
129         /*
130          * Sometimes we want to use more memory than we have
131          */
132         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
133                 return 0;
134
135         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
136                 free = global_page_state(NR_FREE_PAGES);
137                 free += global_page_state(NR_FILE_PAGES);
138
139                 /*
140                  * shmem pages shouldn't be counted as free in this
141                  * case, they can't be purged, only swapped out, and
142                  * that won't affect the overall amount of available
143                  * memory in the system.
144                  */
145                 free -= global_page_state(NR_SHMEM);
146
147                 free += nr_swap_pages;
148
149                 /*
150                  * Any slabs which are created with the
151                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
152                  * which are reclaimable, under pressure.  The dentry
153                  * cache and most inode caches should fall into this
154                  */
155                 free += global_page_state(NR_SLAB_RECLAIMABLE);
156
157                 /*
158                  * Leave reserved pages. The pages are not for anonymous pages.
159                  */
160                 if (free <= totalreserve_pages)
161                         goto error;
162                 else
163                         free -= totalreserve_pages;
164
165                 /*
166                  * Leave the last 3% for root
167                  */
168                 if (!cap_sys_admin)
169                         free -= free / 32;
170
171                 if (free > pages)
172                         return 0;
173
174                 goto error;
175         }
176
177         allowed = (totalram_pages - hugetlb_total_pages())
178                 * sysctl_overcommit_ratio / 100;
179         /*
180          * Leave the last 3% for root
181          */
182         if (!cap_sys_admin)
183                 allowed -= allowed / 32;
184         allowed += total_swap_pages;
185
186         /* Don't let a single process grow too big:
187            leave 3% of the size of this process for other processes */
188         if (mm)
189                 allowed -= mm->total_vm / 32;
190
191         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
192                 return 0;
193 error:
194         vm_unacct_memory(pages);
195
196         return -ENOMEM;
197 }
198
199 /*
200  * Requires inode->i_mapping->i_mmap_mutex
201  */
202 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
203                 struct file *file, struct address_space *mapping)
204 {
205         if (vma->vm_flags & VM_DENYWRITE)
206                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
207         if (vma->vm_flags & VM_SHARED)
208                 mapping->i_mmap_writable--;
209
210         flush_dcache_mmap_lock(mapping);
211         if (unlikely(vma->vm_flags & VM_NONLINEAR))
212                 list_del_init(&vma->shared.nonlinear);
213         else
214                 vma_interval_tree_remove(vma, &mapping->i_mmap);
215         flush_dcache_mmap_unlock(mapping);
216 }
217
218 /*
219  * Unlink a file-based vm structure from its interval tree, to hide
220  * vma from rmap and vmtruncate before freeing its page tables.
221  */
222 void unlink_file_vma(struct vm_area_struct *vma)
223 {
224         struct file *file = vma->vm_file;
225
226         if (file) {
227                 struct address_space *mapping = file->f_mapping;
228                 mutex_lock(&mapping->i_mmap_mutex);
229                 __remove_shared_vm_struct(vma, file, mapping);
230                 mutex_unlock(&mapping->i_mmap_mutex);
231         }
232 }
233
234 /*
235  * Close a vm structure and free it, returning the next.
236  */
237 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
238 {
239         struct vm_area_struct *next = vma->vm_next;
240
241         might_sleep();
242         if (vma->vm_ops && vma->vm_ops->close)
243                 vma->vm_ops->close(vma);
244         if (vma->vm_file)
245                 fput(vma->vm_file);
246         mpol_put(vma_policy(vma));
247         kmem_cache_free(vm_area_cachep, vma);
248         return next;
249 }
250
251 static unsigned long do_brk(unsigned long addr, unsigned long len);
252
253 SYSCALL_DEFINE1(brk, unsigned long, brk)
254 {
255         unsigned long rlim, retval;
256         unsigned long newbrk, oldbrk;
257         struct mm_struct *mm = current->mm;
258         unsigned long min_brk;
259         bool populate;
260
261         down_write(&mm->mmap_sem);
262
263 #ifdef CONFIG_COMPAT_BRK
264         /*
265          * CONFIG_COMPAT_BRK can still be overridden by setting
266          * randomize_va_space to 2, which will still cause mm->start_brk
267          * to be arbitrarily shifted
268          */
269         if (current->brk_randomized)
270                 min_brk = mm->start_brk;
271         else
272                 min_brk = mm->end_data;
273 #else
274         min_brk = mm->start_brk;
275 #endif
276         if (brk < min_brk)
277                 goto out;
278
279         /*
280          * Check against rlimit here. If this check is done later after the test
281          * of oldbrk with newbrk then it can escape the test and let the data
282          * segment grow beyond its set limit the in case where the limit is
283          * not page aligned -Ram Gupta
284          */
285         rlim = rlimit(RLIMIT_DATA);
286         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
287                         (mm->end_data - mm->start_data) > rlim)
288                 goto out;
289
290         newbrk = PAGE_ALIGN(brk);
291         oldbrk = PAGE_ALIGN(mm->brk);
292         if (oldbrk == newbrk)
293                 goto set_brk;
294
295         /* Always allow shrinking brk. */
296         if (brk <= mm->brk) {
297                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
298                         goto set_brk;
299                 goto out;
300         }
301
302         /* Check against existing mmap mappings. */
303         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
304                 goto out;
305
306         /* Ok, looks good - let it rip. */
307         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
308                 goto out;
309
310 set_brk:
311         mm->brk = brk;
312         populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
313         up_write(&mm->mmap_sem);
314         if (populate)
315                 mm_populate(oldbrk, newbrk - oldbrk);
316         return brk;
317
318 out:
319         retval = mm->brk;
320         up_write(&mm->mmap_sem);
321         return retval;
322 }
323
324 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
325 {
326         unsigned long max, subtree_gap;
327         max = vma->vm_start;
328         if (vma->vm_prev)
329                 max -= vma->vm_prev->vm_end;
330         if (vma->vm_rb.rb_left) {
331                 subtree_gap = rb_entry(vma->vm_rb.rb_left,
332                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
333                 if (subtree_gap > max)
334                         max = subtree_gap;
335         }
336         if (vma->vm_rb.rb_right) {
337                 subtree_gap = rb_entry(vma->vm_rb.rb_right,
338                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
339                 if (subtree_gap > max)
340                         max = subtree_gap;
341         }
342         return max;
343 }
344
345 #ifdef CONFIG_DEBUG_VM_RB
346 static int browse_rb(struct rb_root *root)
347 {
348         int i = 0, j, bug = 0;
349         struct rb_node *nd, *pn = NULL;
350         unsigned long prev = 0, pend = 0;
351
352         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
353                 struct vm_area_struct *vma;
354                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
355                 if (vma->vm_start < prev) {
356                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev);
357                         bug = 1;
358                 }
359                 if (vma->vm_start < pend) {
360                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
361                         bug = 1;
362                 }
363                 if (vma->vm_start > vma->vm_end) {
364                         printk("vm_end %lx < vm_start %lx\n",
365                                 vma->vm_end, vma->vm_start);
366                         bug = 1;
367                 }
368                 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
369                         printk("free gap %lx, correct %lx\n",
370                                vma->rb_subtree_gap,
371                                vma_compute_subtree_gap(vma));
372                         bug = 1;
373                 }
374                 i++;
375                 pn = nd;
376                 prev = vma->vm_start;
377                 pend = vma->vm_end;
378         }
379         j = 0;
380         for (nd = pn; nd; nd = rb_prev(nd))
381                 j++;
382         if (i != j) {
383                 printk("backwards %d, forwards %d\n", j, i);
384                 bug = 1;
385         }
386         return bug ? -1 : i;
387 }
388
389 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
390 {
391         struct rb_node *nd;
392
393         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
394                 struct vm_area_struct *vma;
395                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
396                 BUG_ON(vma != ignore &&
397                        vma->rb_subtree_gap != vma_compute_subtree_gap(vma));
398         }
399 }
400
401 void validate_mm(struct mm_struct *mm)
402 {
403         int bug = 0;
404         int i = 0;
405         unsigned long highest_address = 0;
406         struct vm_area_struct *vma = mm->mmap;
407         while (vma) {
408                 struct anon_vma_chain *avc;
409                 vma_lock_anon_vma(vma);
410                 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
411                         anon_vma_interval_tree_verify(avc);
412                 vma_unlock_anon_vma(vma);
413                 highest_address = vma->vm_end;
414                 vma = vma->vm_next;
415                 i++;
416         }
417         if (i != mm->map_count) {
418                 printk("map_count %d vm_next %d\n", mm->map_count, i);
419                 bug = 1;
420         }
421         if (highest_address != mm->highest_vm_end) {
422                 printk("mm->highest_vm_end %lx, found %lx\n",
423                        mm->highest_vm_end, highest_address);
424                 bug = 1;
425         }
426         i = browse_rb(&mm->mm_rb);
427         if (i != mm->map_count) {
428                 printk("map_count %d rb %d\n", mm->map_count, i);
429                 bug = 1;
430         }
431         BUG_ON(bug);
432 }
433 #else
434 #define validate_mm_rb(root, ignore) do { } while (0)
435 #define validate_mm(mm) do { } while (0)
436 #endif
437
438 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
439                      unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
440
441 /*
442  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
443  * vma->vm_prev->vm_end values changed, without modifying the vma's position
444  * in the rbtree.
445  */
446 static void vma_gap_update(struct vm_area_struct *vma)
447 {
448         /*
449          * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
450          * function that does exacltly what we want.
451          */
452         vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
453 }
454
455 static inline void vma_rb_insert(struct vm_area_struct *vma,
456                                  struct rb_root *root)
457 {
458         /* All rb_subtree_gap values must be consistent prior to insertion */
459         validate_mm_rb(root, NULL);
460
461         rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
462 }
463
464 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
465 {
466         /*
467          * All rb_subtree_gap values must be consistent prior to erase,
468          * with the possible exception of the vma being erased.
469          */
470         validate_mm_rb(root, vma);
471
472         /*
473          * Note rb_erase_augmented is a fairly large inline function,
474          * so make sure we instantiate it only once with our desired
475          * augmented rbtree callbacks.
476          */
477         rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
478 }
479
480 /*
481  * vma has some anon_vma assigned, and is already inserted on that
482  * anon_vma's interval trees.
483  *
484  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
485  * vma must be removed from the anon_vma's interval trees using
486  * anon_vma_interval_tree_pre_update_vma().
487  *
488  * After the update, the vma will be reinserted using
489  * anon_vma_interval_tree_post_update_vma().
490  *
491  * The entire update must be protected by exclusive mmap_sem and by
492  * the root anon_vma's mutex.
493  */
494 static inline void
495 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
496 {
497         struct anon_vma_chain *avc;
498
499         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
500                 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
501 }
502
503 static inline void
504 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
505 {
506         struct anon_vma_chain *avc;
507
508         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
509                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
510 }
511
512 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
513                 unsigned long end, struct vm_area_struct **pprev,
514                 struct rb_node ***rb_link, struct rb_node **rb_parent)
515 {
516         struct rb_node **__rb_link, *__rb_parent, *rb_prev;
517
518         __rb_link = &mm->mm_rb.rb_node;
519         rb_prev = __rb_parent = NULL;
520
521         while (*__rb_link) {
522                 struct vm_area_struct *vma_tmp;
523
524                 __rb_parent = *__rb_link;
525                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
526
527                 if (vma_tmp->vm_end > addr) {
528                         /* Fail if an existing vma overlaps the area */
529                         if (vma_tmp->vm_start < end)
530                                 return -ENOMEM;
531                         __rb_link = &__rb_parent->rb_left;
532                 } else {
533                         rb_prev = __rb_parent;
534                         __rb_link = &__rb_parent->rb_right;
535                 }
536         }
537
538         *pprev = NULL;
539         if (rb_prev)
540                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
541         *rb_link = __rb_link;
542         *rb_parent = __rb_parent;
543         return 0;
544 }
545
546 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
547                 struct rb_node **rb_link, struct rb_node *rb_parent)
548 {
549         /* Update tracking information for the gap following the new vma. */
550         if (vma->vm_next)
551                 vma_gap_update(vma->vm_next);
552         else
553                 mm->highest_vm_end = vma->vm_end;
554
555         /*
556          * vma->vm_prev wasn't known when we followed the rbtree to find the
557          * correct insertion point for that vma. As a result, we could not
558          * update the vma vm_rb parents rb_subtree_gap values on the way down.
559          * So, we first insert the vma with a zero rb_subtree_gap value
560          * (to be consistent with what we did on the way down), and then
561          * immediately update the gap to the correct value. Finally we
562          * rebalance the rbtree after all augmented values have been set.
563          */
564         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
565         vma->rb_subtree_gap = 0;
566         vma_gap_update(vma);
567         vma_rb_insert(vma, &mm->mm_rb);
568 }
569
570 static void __vma_link_file(struct vm_area_struct *vma)
571 {
572         struct file *file;
573
574         file = vma->vm_file;
575         if (file) {
576                 struct address_space *mapping = file->f_mapping;
577
578                 if (vma->vm_flags & VM_DENYWRITE)
579                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
580                 if (vma->vm_flags & VM_SHARED)
581                         mapping->i_mmap_writable++;
582
583                 flush_dcache_mmap_lock(mapping);
584                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
585                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
586                 else
587                         vma_interval_tree_insert(vma, &mapping->i_mmap);
588                 flush_dcache_mmap_unlock(mapping);
589         }
590 }
591
592 static void
593 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
594         struct vm_area_struct *prev, struct rb_node **rb_link,
595         struct rb_node *rb_parent)
596 {
597         __vma_link_list(mm, vma, prev, rb_parent);
598         __vma_link_rb(mm, vma, rb_link, rb_parent);
599 }
600
601 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
602                         struct vm_area_struct *prev, struct rb_node **rb_link,
603                         struct rb_node *rb_parent)
604 {
605         struct address_space *mapping = NULL;
606
607         if (vma->vm_file)
608                 mapping = vma->vm_file->f_mapping;
609
610         if (mapping)
611                 mutex_lock(&mapping->i_mmap_mutex);
612
613         __vma_link(mm, vma, prev, rb_link, rb_parent);
614         __vma_link_file(vma);
615
616         if (mapping)
617                 mutex_unlock(&mapping->i_mmap_mutex);
618
619         mm->map_count++;
620         validate_mm(mm);
621 }
622
623 /*
624  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
625  * mm's list and rbtree.  It has already been inserted into the interval tree.
626  */
627 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
628 {
629         struct vm_area_struct *prev;
630         struct rb_node **rb_link, *rb_parent;
631
632         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
633                            &prev, &rb_link, &rb_parent))
634                 BUG();
635         __vma_link(mm, vma, prev, rb_link, rb_parent);
636         mm->map_count++;
637 }
638
639 static inline void
640 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
641                 struct vm_area_struct *prev)
642 {
643         struct vm_area_struct *next;
644
645         vma_rb_erase(vma, &mm->mm_rb);
646         prev->vm_next = next = vma->vm_next;
647         if (next)
648                 next->vm_prev = prev;
649         if (mm->mmap_cache == vma)
650                 mm->mmap_cache = prev;
651 }
652
653 /*
654  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
655  * is already present in an i_mmap tree without adjusting the tree.
656  * The following helper function should be used when such adjustments
657  * are necessary.  The "insert" vma (if any) is to be inserted
658  * before we drop the necessary locks.
659  */
660 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
661         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
662 {
663         struct mm_struct *mm = vma->vm_mm;
664         struct vm_area_struct *next = vma->vm_next;
665         struct vm_area_struct *importer = NULL;
666         struct address_space *mapping = NULL;
667         struct rb_root *root = NULL;
668         struct anon_vma *anon_vma = NULL;
669         struct file *file = vma->vm_file;
670         bool start_changed = false, end_changed = false;
671         long adjust_next = 0;
672         int remove_next = 0;
673
674         if (next && !insert) {
675                 struct vm_area_struct *exporter = NULL;
676
677                 if (end >= next->vm_end) {
678                         /*
679                          * vma expands, overlapping all the next, and
680                          * perhaps the one after too (mprotect case 6).
681                          */
682 again:                  remove_next = 1 + (end > next->vm_end);
683                         end = next->vm_end;
684                         exporter = next;
685                         importer = vma;
686                 } else if (end > next->vm_start) {
687                         /*
688                          * vma expands, overlapping part of the next:
689                          * mprotect case 5 shifting the boundary up.
690                          */
691                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
692                         exporter = next;
693                         importer = vma;
694                 } else if (end < vma->vm_end) {
695                         /*
696                          * vma shrinks, and !insert tells it's not
697                          * split_vma inserting another: so it must be
698                          * mprotect case 4 shifting the boundary down.
699                          */
700                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
701                         exporter = vma;
702                         importer = next;
703                 }
704
705                 /*
706                  * Easily overlooked: when mprotect shifts the boundary,
707                  * make sure the expanding vma has anon_vma set if the
708                  * shrinking vma had, to cover any anon pages imported.
709                  */
710                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
711                         if (anon_vma_clone(importer, exporter))
712                                 return -ENOMEM;
713                         importer->anon_vma = exporter->anon_vma;
714                 }
715         }
716
717         if (file) {
718                 mapping = file->f_mapping;
719                 if (!(vma->vm_flags & VM_NONLINEAR)) {
720                         root = &mapping->i_mmap;
721                         uprobe_munmap(vma, vma->vm_start, vma->vm_end);
722
723                         if (adjust_next)
724                                 uprobe_munmap(next, next->vm_start,
725                                                         next->vm_end);
726                 }
727
728                 mutex_lock(&mapping->i_mmap_mutex);
729                 if (insert) {
730                         /*
731                          * Put into interval tree now, so instantiated pages
732                          * are visible to arm/parisc __flush_dcache_page
733                          * throughout; but we cannot insert into address
734                          * space until vma start or end is updated.
735                          */
736                         __vma_link_file(insert);
737                 }
738         }
739
740         vma_adjust_trans_huge(vma, start, end, adjust_next);
741
742         anon_vma = vma->anon_vma;
743         if (!anon_vma && adjust_next)
744                 anon_vma = next->anon_vma;
745         if (anon_vma) {
746                 VM_BUG_ON(adjust_next && next->anon_vma &&
747                           anon_vma != next->anon_vma);
748                 anon_vma_lock_write(anon_vma);
749                 anon_vma_interval_tree_pre_update_vma(vma);
750                 if (adjust_next)
751                         anon_vma_interval_tree_pre_update_vma(next);
752         }
753
754         if (root) {
755                 flush_dcache_mmap_lock(mapping);
756                 vma_interval_tree_remove(vma, root);
757                 if (adjust_next)
758                         vma_interval_tree_remove(next, root);
759         }
760
761         if (start != vma->vm_start) {
762                 vma->vm_start = start;
763                 start_changed = true;
764         }
765         if (end != vma->vm_end) {
766                 vma->vm_end = end;
767                 end_changed = true;
768         }
769         vma->vm_pgoff = pgoff;
770         if (adjust_next) {
771                 next->vm_start += adjust_next << PAGE_SHIFT;
772                 next->vm_pgoff += adjust_next;
773         }
774
775         if (root) {
776                 if (adjust_next)
777                         vma_interval_tree_insert(next, root);
778                 vma_interval_tree_insert(vma, root);
779                 flush_dcache_mmap_unlock(mapping);
780         }
781
782         if (remove_next) {
783                 /*
784                  * vma_merge has merged next into vma, and needs
785                  * us to remove next before dropping the locks.
786                  */
787                 __vma_unlink(mm, next, vma);
788                 if (file)
789                         __remove_shared_vm_struct(next, file, mapping);
790         } else if (insert) {
791                 /*
792                  * split_vma has split insert from vma, and needs
793                  * us to insert it before dropping the locks
794                  * (it may either follow vma or precede it).
795                  */
796                 __insert_vm_struct(mm, insert);
797         } else {
798                 if (start_changed)
799                         vma_gap_update(vma);
800                 if (end_changed) {
801                         if (!next)
802                                 mm->highest_vm_end = end;
803                         else if (!adjust_next)
804                                 vma_gap_update(next);
805                 }
806         }
807
808         if (anon_vma) {
809                 anon_vma_interval_tree_post_update_vma(vma);
810                 if (adjust_next)
811                         anon_vma_interval_tree_post_update_vma(next);
812                 anon_vma_unlock(anon_vma);
813         }
814         if (mapping)
815                 mutex_unlock(&mapping->i_mmap_mutex);
816
817         if (root) {
818                 uprobe_mmap(vma);
819
820                 if (adjust_next)
821                         uprobe_mmap(next);
822         }
823
824         if (remove_next) {
825                 if (file) {
826                         uprobe_munmap(next, next->vm_start, next->vm_end);
827                         fput(file);
828                 }
829                 if (next->anon_vma)
830                         anon_vma_merge(vma, next);
831                 mm->map_count--;
832                 mpol_put(vma_policy(next));
833                 kmem_cache_free(vm_area_cachep, next);
834                 /*
835                  * In mprotect's case 6 (see comments on vma_merge),
836                  * we must remove another next too. It would clutter
837                  * up the code too much to do both in one go.
838                  */
839                 next = vma->vm_next;
840                 if (remove_next == 2)
841                         goto again;
842                 else if (next)
843                         vma_gap_update(next);
844                 else
845                         mm->highest_vm_end = end;
846         }
847         if (insert && file)
848                 uprobe_mmap(insert);
849
850         validate_mm(mm);
851
852         return 0;
853 }
854
855 /*
856  * If the vma has a ->close operation then the driver probably needs to release
857  * per-vma resources, so we don't attempt to merge those.
858  */
859 static inline int is_mergeable_vma(struct vm_area_struct *vma,
860                         struct file *file, unsigned long vm_flags)
861 {
862         if (vma->vm_flags ^ vm_flags)
863                 return 0;
864         if (vma->vm_file != file)
865                 return 0;
866         if (vma->vm_ops && vma->vm_ops->close)
867                 return 0;
868         return 1;
869 }
870
871 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
872                                         struct anon_vma *anon_vma2,
873                                         struct vm_area_struct *vma)
874 {
875         /*
876          * The list_is_singular() test is to avoid merging VMA cloned from
877          * parents. This can improve scalability caused by anon_vma lock.
878          */
879         if ((!anon_vma1 || !anon_vma2) && (!vma ||
880                 list_is_singular(&vma->anon_vma_chain)))
881                 return 1;
882         return anon_vma1 == anon_vma2;
883 }
884
885 /*
886  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
887  * in front of (at a lower virtual address and file offset than) the vma.
888  *
889  * We cannot merge two vmas if they have differently assigned (non-NULL)
890  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
891  *
892  * We don't check here for the merged mmap wrapping around the end of pagecache
893  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
894  * wrap, nor mmaps which cover the final page at index -1UL.
895  */
896 static int
897 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
898         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
899 {
900         if (is_mergeable_vma(vma, file, vm_flags) &&
901             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
902                 if (vma->vm_pgoff == vm_pgoff)
903                         return 1;
904         }
905         return 0;
906 }
907
908 /*
909  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
910  * beyond (at a higher virtual address and file offset than) the vma.
911  *
912  * We cannot merge two vmas if they have differently assigned (non-NULL)
913  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
914  */
915 static int
916 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
917         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
918 {
919         if (is_mergeable_vma(vma, file, vm_flags) &&
920             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
921                 pgoff_t vm_pglen;
922                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
923                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
924                         return 1;
925         }
926         return 0;
927 }
928
929 /*
930  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
931  * whether that can be merged with its predecessor or its successor.
932  * Or both (it neatly fills a hole).
933  *
934  * In most cases - when called for mmap, brk or mremap - [addr,end) is
935  * certain not to be mapped by the time vma_merge is called; but when
936  * called for mprotect, it is certain to be already mapped (either at
937  * an offset within prev, or at the start of next), and the flags of
938  * this area are about to be changed to vm_flags - and the no-change
939  * case has already been eliminated.
940  *
941  * The following mprotect cases have to be considered, where AAAA is
942  * the area passed down from mprotect_fixup, never extending beyond one
943  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
944  *
945  *     AAAA             AAAA                AAAA          AAAA
946  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
947  *    cannot merge    might become    might become    might become
948  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
949  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
950  *    mremap move:                                    PPPPNNNNNNNN 8
951  *        AAAA
952  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
953  *    might become    case 1 below    case 2 below    case 3 below
954  *
955  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
956  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
957  */
958 struct vm_area_struct *vma_merge(struct mm_struct *mm,
959                         struct vm_area_struct *prev, unsigned long addr,
960                         unsigned long end, unsigned long vm_flags,
961                         struct anon_vma *anon_vma, struct file *file,
962                         pgoff_t pgoff, struct mempolicy *policy)
963 {
964         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
965         struct vm_area_struct *area, *next;
966         int err;
967
968         /*
969          * We later require that vma->vm_flags == vm_flags,
970          * so this tests vma->vm_flags & VM_SPECIAL, too.
971          */
972         if (vm_flags & VM_SPECIAL)
973                 return NULL;
974
975         if (prev)
976                 next = prev->vm_next;
977         else
978                 next = mm->mmap;
979         area = next;
980         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
981                 next = next->vm_next;
982
983         /*
984          * Can it merge with the predecessor?
985          */
986         if (prev && prev->vm_end == addr &&
987                         mpol_equal(vma_policy(prev), policy) &&
988                         can_vma_merge_after(prev, vm_flags,
989                                                 anon_vma, file, pgoff)) {
990                 /*
991                  * OK, it can.  Can we now merge in the successor as well?
992                  */
993                 if (next && end == next->vm_start &&
994                                 mpol_equal(policy, vma_policy(next)) &&
995                                 can_vma_merge_before(next, vm_flags,
996                                         anon_vma, file, pgoff+pglen) &&
997                                 is_mergeable_anon_vma(prev->anon_vma,
998                                                       next->anon_vma, NULL)) {
999                                                         /* cases 1, 6 */
1000                         err = vma_adjust(prev, prev->vm_start,
1001                                 next->vm_end, prev->vm_pgoff, NULL);
1002                 } else                                  /* cases 2, 5, 7 */
1003                         err = vma_adjust(prev, prev->vm_start,
1004                                 end, prev->vm_pgoff, NULL);
1005                 if (err)
1006                         return NULL;
1007                 khugepaged_enter_vma_merge(prev);
1008                 return prev;
1009         }
1010
1011         /*
1012          * Can this new request be merged in front of next?
1013          */
1014         if (next && end == next->vm_start &&
1015                         mpol_equal(policy, vma_policy(next)) &&
1016                         can_vma_merge_before(next, vm_flags,
1017                                         anon_vma, file, pgoff+pglen)) {
1018                 if (prev && addr < prev->vm_end)        /* case 4 */
1019                         err = vma_adjust(prev, prev->vm_start,
1020                                 addr, prev->vm_pgoff, NULL);
1021                 else                                    /* cases 3, 8 */
1022                         err = vma_adjust(area, addr, next->vm_end,
1023                                 next->vm_pgoff - pglen, NULL);
1024                 if (err)
1025                         return NULL;
1026                 khugepaged_enter_vma_merge(area);
1027                 return area;
1028         }
1029
1030         return NULL;
1031 }
1032
1033 /*
1034  * Rough compatbility check to quickly see if it's even worth looking
1035  * at sharing an anon_vma.
1036  *
1037  * They need to have the same vm_file, and the flags can only differ
1038  * in things that mprotect may change.
1039  *
1040  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1041  * we can merge the two vma's. For example, we refuse to merge a vma if
1042  * there is a vm_ops->close() function, because that indicates that the
1043  * driver is doing some kind of reference counting. But that doesn't
1044  * really matter for the anon_vma sharing case.
1045  */
1046 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1047 {
1048         return a->vm_end == b->vm_start &&
1049                 mpol_equal(vma_policy(a), vma_policy(b)) &&
1050                 a->vm_file == b->vm_file &&
1051                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
1052                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1053 }
1054
1055 /*
1056  * Do some basic sanity checking to see if we can re-use the anon_vma
1057  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1058  * the same as 'old', the other will be the new one that is trying
1059  * to share the anon_vma.
1060  *
1061  * NOTE! This runs with mm_sem held for reading, so it is possible that
1062  * the anon_vma of 'old' is concurrently in the process of being set up
1063  * by another page fault trying to merge _that_. But that's ok: if it
1064  * is being set up, that automatically means that it will be a singleton
1065  * acceptable for merging, so we can do all of this optimistically. But
1066  * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
1067  *
1068  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1069  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1070  * is to return an anon_vma that is "complex" due to having gone through
1071  * a fork).
1072  *
1073  * We also make sure that the two vma's are compatible (adjacent,
1074  * and with the same memory policies). That's all stable, even with just
1075  * a read lock on the mm_sem.
1076  */
1077 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1078 {
1079         if (anon_vma_compatible(a, b)) {
1080                 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
1081
1082                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1083                         return anon_vma;
1084         }
1085         return NULL;
1086 }
1087
1088 /*
1089  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1090  * neighbouring vmas for a suitable anon_vma, before it goes off
1091  * to allocate a new anon_vma.  It checks because a repetitive
1092  * sequence of mprotects and faults may otherwise lead to distinct
1093  * anon_vmas being allocated, preventing vma merge in subsequent
1094  * mprotect.
1095  */
1096 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1097 {
1098         struct anon_vma *anon_vma;
1099         struct vm_area_struct *near;
1100
1101         near = vma->vm_next;
1102         if (!near)
1103                 goto try_prev;
1104
1105         anon_vma = reusable_anon_vma(near, vma, near);
1106         if (anon_vma)
1107                 return anon_vma;
1108 try_prev:
1109         near = vma->vm_prev;
1110         if (!near)
1111                 goto none;
1112
1113         anon_vma = reusable_anon_vma(near, near, vma);
1114         if (anon_vma)
1115                 return anon_vma;
1116 none:
1117         /*
1118          * There's no absolute need to look only at touching neighbours:
1119          * we could search further afield for "compatible" anon_vmas.
1120          * But it would probably just be a waste of time searching,
1121          * or lead to too many vmas hanging off the same anon_vma.
1122          * We're trying to allow mprotect remerging later on,
1123          * not trying to minimize memory used for anon_vmas.
1124          */
1125         return NULL;
1126 }
1127
1128 #ifdef CONFIG_PROC_FS
1129 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
1130                                                 struct file *file, long pages)
1131 {
1132         const unsigned long stack_flags
1133                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
1134
1135         mm->total_vm += pages;
1136
1137         if (file) {
1138                 mm->shared_vm += pages;
1139                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
1140                         mm->exec_vm += pages;
1141         } else if (flags & stack_flags)
1142                 mm->stack_vm += pages;
1143 }
1144 #endif /* CONFIG_PROC_FS */
1145
1146 /*
1147  * If a hint addr is less than mmap_min_addr change hint to be as
1148  * low as possible but still greater than mmap_min_addr
1149  */
1150 static inline unsigned long round_hint_to_min(unsigned long hint)
1151 {
1152         hint &= PAGE_MASK;
1153         if (((void *)hint != NULL) &&
1154             (hint < mmap_min_addr))
1155                 return PAGE_ALIGN(mmap_min_addr);
1156         return hint;
1157 }
1158
1159 /*
1160  * The caller must hold down_write(&current->mm->mmap_sem).
1161  */
1162
1163 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
1164                         unsigned long len, unsigned long prot,
1165                         unsigned long flags, unsigned long pgoff,
1166                         bool *populate)
1167 {
1168         struct mm_struct * mm = current->mm;
1169         struct inode *inode;
1170         vm_flags_t vm_flags;
1171
1172         *populate = false;
1173
1174         /*
1175          * Does the application expect PROT_READ to imply PROT_EXEC?
1176          *
1177          * (the exception is when the underlying filesystem is noexec
1178          *  mounted, in which case we dont add PROT_EXEC.)
1179          */
1180         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1181                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
1182                         prot |= PROT_EXEC;
1183
1184         if (!len)
1185                 return -EINVAL;
1186
1187         if (!(flags & MAP_FIXED))
1188                 addr = round_hint_to_min(addr);
1189
1190         /* Careful about overflows.. */
1191         len = PAGE_ALIGN(len);
1192         if (!len)
1193                 return -ENOMEM;
1194
1195         /* offset overflow? */
1196         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1197                return -EOVERFLOW;
1198
1199         /* Too many mappings? */
1200         if (mm->map_count > sysctl_max_map_count)
1201                 return -ENOMEM;
1202
1203         /* Obtain the address to map to. we verify (or select) it and ensure
1204          * that it represents a valid section of the address space.
1205          */
1206         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1207         if (addr & ~PAGE_MASK)
1208                 return addr;
1209
1210         /* Do simple checking here so the lower-level routines won't have
1211          * to. we assume access permissions have been handled by the open
1212          * of the memory object, so we don't do any here.
1213          */
1214         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1215                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1216
1217         if (flags & MAP_LOCKED)
1218                 if (!can_do_mlock())
1219                         return -EPERM;
1220
1221         /* mlock MCL_FUTURE? */
1222         if (vm_flags & VM_LOCKED) {
1223                 unsigned long locked, lock_limit;
1224                 locked = len >> PAGE_SHIFT;
1225                 locked += mm->locked_vm;
1226                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1227                 lock_limit >>= PAGE_SHIFT;
1228                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1229                         return -EAGAIN;
1230         }
1231
1232         inode = file ? file->f_path.dentry->d_inode : NULL;
1233
1234         if (file) {
1235                 switch (flags & MAP_TYPE) {
1236                 case MAP_SHARED:
1237                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1238                                 return -EACCES;
1239
1240                         /*
1241                          * Make sure we don't allow writing to an append-only
1242                          * file..
1243                          */
1244                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1245                                 return -EACCES;
1246
1247                         /*
1248                          * Make sure there are no mandatory locks on the file.
1249                          */
1250                         if (locks_verify_locked(inode))
1251                                 return -EAGAIN;
1252
1253                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1254                         if (!(file->f_mode & FMODE_WRITE))
1255                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1256
1257                         /* fall through */
1258                 case MAP_PRIVATE:
1259                         if (!(file->f_mode & FMODE_READ))
1260                                 return -EACCES;
1261                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1262                                 if (vm_flags & VM_EXEC)
1263                                         return -EPERM;
1264                                 vm_flags &= ~VM_MAYEXEC;
1265                         }
1266
1267                         if (!file->f_op || !file->f_op->mmap)
1268                                 return -ENODEV;
1269                         break;
1270
1271                 default:
1272                         return -EINVAL;
1273                 }
1274         } else {
1275                 switch (flags & MAP_TYPE) {
1276                 case MAP_SHARED:
1277                         /*
1278                          * Ignore pgoff.
1279                          */
1280                         pgoff = 0;
1281                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1282                         break;
1283                 case MAP_PRIVATE:
1284                         /*
1285                          * Set pgoff according to addr for anon_vma.
1286                          */
1287                         pgoff = addr >> PAGE_SHIFT;
1288                         break;
1289                 default:
1290                         return -EINVAL;
1291                 }
1292         }
1293
1294         /*
1295          * Set 'VM_NORESERVE' if we should not account for the
1296          * memory use of this mapping.
1297          */
1298         if (flags & MAP_NORESERVE) {
1299                 /* We honor MAP_NORESERVE if allowed to overcommit */
1300                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1301                         vm_flags |= VM_NORESERVE;
1302
1303                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1304                 if (file && is_file_hugepages(file))
1305                         vm_flags |= VM_NORESERVE;
1306         }
1307
1308         addr = mmap_region(file, addr, len, vm_flags, pgoff);
1309         if (!IS_ERR_VALUE(addr) &&
1310             ((vm_flags & VM_LOCKED) ||
1311              (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1312                 *populate = true;
1313         return addr;
1314 }
1315
1316 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1317                 unsigned long, prot, unsigned long, flags,
1318                 unsigned long, fd, unsigned long, pgoff)
1319 {
1320         struct file *file = NULL;
1321         unsigned long retval = -EBADF;
1322
1323         if (!(flags & MAP_ANONYMOUS)) {
1324                 audit_mmap_fd(fd, flags);
1325                 if (unlikely(flags & MAP_HUGETLB))
1326                         return -EINVAL;
1327                 file = fget(fd);
1328                 if (!file)
1329                         goto out;
1330         } else if (flags & MAP_HUGETLB) {
1331                 struct user_struct *user = NULL;
1332                 /*
1333                  * VM_NORESERVE is used because the reservations will be
1334                  * taken when vm_ops->mmap() is called
1335                  * A dummy user value is used because we are not locking
1336                  * memory so no accounting is necessary
1337                  */
1338                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, addr, len,
1339                                 VM_NORESERVE,
1340                                 &user, HUGETLB_ANONHUGE_INODE,
1341                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1342                 if (IS_ERR(file))
1343                         return PTR_ERR(file);
1344         }
1345
1346         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1347
1348         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1349         if (file)
1350                 fput(file);
1351 out:
1352         return retval;
1353 }
1354
1355 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1356 struct mmap_arg_struct {
1357         unsigned long addr;
1358         unsigned long len;
1359         unsigned long prot;
1360         unsigned long flags;
1361         unsigned long fd;
1362         unsigned long offset;
1363 };
1364
1365 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1366 {
1367         struct mmap_arg_struct a;
1368
1369         if (copy_from_user(&a, arg, sizeof(a)))
1370                 return -EFAULT;
1371         if (a.offset & ~PAGE_MASK)
1372                 return -EINVAL;
1373
1374         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1375                               a.offset >> PAGE_SHIFT);
1376 }
1377 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1378
1379 /*
1380  * Some shared mappigns will want the pages marked read-only
1381  * to track write events. If so, we'll downgrade vm_page_prot
1382  * to the private version (using protection_map[] without the
1383  * VM_SHARED bit).
1384  */
1385 int vma_wants_writenotify(struct vm_area_struct *vma)
1386 {
1387         vm_flags_t vm_flags = vma->vm_flags;
1388
1389         /* If it was private or non-writable, the write bit is already clear */
1390         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1391                 return 0;
1392
1393         /* The backer wishes to know when pages are first written to? */
1394         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1395                 return 1;
1396
1397         /* The open routine did something to the protections already? */
1398         if (pgprot_val(vma->vm_page_prot) !=
1399             pgprot_val(vm_get_page_prot(vm_flags)))
1400                 return 0;
1401
1402         /* Specialty mapping? */
1403         if (vm_flags & VM_PFNMAP)
1404                 return 0;
1405
1406         /* Can the mapping track the dirty pages? */
1407         return vma->vm_file && vma->vm_file->f_mapping &&
1408                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1409 }
1410
1411 /*
1412  * We account for memory if it's a private writeable mapping,
1413  * not hugepages and VM_NORESERVE wasn't set.
1414  */
1415 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1416 {
1417         /*
1418          * hugetlb has its own accounting separate from the core VM
1419          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1420          */
1421         if (file && is_file_hugepages(file))
1422                 return 0;
1423
1424         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1425 }
1426
1427 unsigned long mmap_region(struct file *file, unsigned long addr,
1428                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1429 {
1430         struct mm_struct *mm = current->mm;
1431         struct vm_area_struct *vma, *prev;
1432         int correct_wcount = 0;
1433         int error;
1434         struct rb_node **rb_link, *rb_parent;
1435         unsigned long charged = 0;
1436         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1437
1438         /* Clear old maps */
1439         error = -ENOMEM;
1440 munmap_back:
1441         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
1442                 if (do_munmap(mm, addr, len))
1443                         return -ENOMEM;
1444                 goto munmap_back;
1445         }
1446
1447         /* Check against address space limit. */
1448         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1449                 return -ENOMEM;
1450
1451         /*
1452          * Private writable mapping: check memory availability
1453          */
1454         if (accountable_mapping(file, vm_flags)) {
1455                 charged = len >> PAGE_SHIFT;
1456                 if (security_vm_enough_memory_mm(mm, charged))
1457                         return -ENOMEM;
1458                 vm_flags |= VM_ACCOUNT;
1459         }
1460
1461         /*
1462          * Can we just expand an old mapping?
1463          */
1464         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1465         if (vma)
1466                 goto out;
1467
1468         /*
1469          * Determine the object being mapped and call the appropriate
1470          * specific mapper. the address has already been validated, but
1471          * not unmapped, but the maps are removed from the list.
1472          */
1473         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1474         if (!vma) {
1475                 error = -ENOMEM;
1476                 goto unacct_error;
1477         }
1478
1479         vma->vm_mm = mm;
1480         vma->vm_start = addr;
1481         vma->vm_end = addr + len;
1482         vma->vm_flags = vm_flags;
1483         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1484         vma->vm_pgoff = pgoff;
1485         INIT_LIST_HEAD(&vma->anon_vma_chain);
1486
1487         error = -EINVAL;        /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1488
1489         if (file) {
1490                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1491                         goto free_vma;
1492                 if (vm_flags & VM_DENYWRITE) {
1493                         error = deny_write_access(file);
1494                         if (error)
1495                                 goto free_vma;
1496                         correct_wcount = 1;
1497                 }
1498                 vma->vm_file = get_file(file);
1499                 error = file->f_op->mmap(file, vma);
1500                 if (error)
1501                         goto unmap_and_free_vma;
1502
1503                 /* Can addr have changed??
1504                  *
1505                  * Answer: Yes, several device drivers can do it in their
1506                  *         f_op->mmap method. -DaveM
1507                  * Bug: If addr is changed, prev, rb_link, rb_parent should
1508                  *      be updated for vma_link()
1509                  */
1510                 WARN_ON_ONCE(addr != vma->vm_start);
1511
1512                 addr = vma->vm_start;
1513                 pgoff = vma->vm_pgoff;
1514                 vm_flags = vma->vm_flags;
1515         } else if (vm_flags & VM_SHARED) {
1516                 if (unlikely(vm_flags & (VM_GROWSDOWN|VM_GROWSUP)))
1517                         goto free_vma;
1518                 error = shmem_zero_setup(vma);
1519                 if (error)
1520                         goto free_vma;
1521         }
1522
1523         if (vma_wants_writenotify(vma)) {
1524                 pgprot_t pprot = vma->vm_page_prot;
1525
1526                 /* Can vma->vm_page_prot have changed??
1527                  *
1528                  * Answer: Yes, drivers may have changed it in their
1529                  *         f_op->mmap method.
1530                  *
1531                  * Ensures that vmas marked as uncached stay that way.
1532                  */
1533                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1534                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1535                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1536         }
1537
1538         vma_link(mm, vma, prev, rb_link, rb_parent);
1539         file = vma->vm_file;
1540
1541         /* Once vma denies write, undo our temporary denial count */
1542         if (correct_wcount)
1543                 atomic_inc(&inode->i_writecount);
1544 out:
1545         perf_event_mmap(vma);
1546
1547         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1548         if (vm_flags & VM_LOCKED) {
1549                 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1550                                         vma == get_gate_vma(current->mm)))
1551                         mm->locked_vm += (len >> PAGE_SHIFT);
1552                 else
1553                         vma->vm_flags &= ~VM_LOCKED;
1554         }
1555
1556         if (file)
1557                 uprobe_mmap(vma);
1558
1559         return addr;
1560
1561 unmap_and_free_vma:
1562         if (correct_wcount)
1563                 atomic_inc(&inode->i_writecount);
1564         vma->vm_file = NULL;
1565         fput(file);
1566
1567         /* Undo any partial mapping done by a device driver. */
1568         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1569         charged = 0;
1570 free_vma:
1571         kmem_cache_free(vm_area_cachep, vma);
1572 unacct_error:
1573         if (charged)
1574                 vm_unacct_memory(charged);
1575         return error;
1576 }
1577
1578 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1579 {
1580         /*
1581          * We implement the search by looking for an rbtree node that
1582          * immediately follows a suitable gap. That is,
1583          * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1584          * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1585          * - gap_end - gap_start >= length
1586          */
1587
1588         struct mm_struct *mm = current->mm;
1589         struct vm_area_struct *vma;
1590         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1591
1592         /* Adjust search length to account for worst case alignment overhead */
1593         length = info->length + info->align_mask;
1594         if (length < info->length)
1595                 return -ENOMEM;
1596
1597         /* Adjust search limits by the desired length */
1598         if (info->high_limit < length)
1599                 return -ENOMEM;
1600         high_limit = info->high_limit - length;
1601
1602         if (info->low_limit > high_limit)
1603                 return -ENOMEM;
1604         low_limit = info->low_limit + length;
1605
1606         /* Check if rbtree root looks promising */
1607         if (RB_EMPTY_ROOT(&mm->mm_rb))
1608                 goto check_highest;
1609         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1610         if (vma->rb_subtree_gap < length)
1611                 goto check_highest;
1612
1613         while (true) {
1614                 /* Visit left subtree if it looks promising */
1615                 gap_end = vma->vm_start;
1616                 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1617                         struct vm_area_struct *left =
1618                                 rb_entry(vma->vm_rb.rb_left,
1619                                          struct vm_area_struct, vm_rb);
1620                         if (left->rb_subtree_gap >= length) {
1621                                 vma = left;
1622                                 continue;
1623                         }
1624                 }
1625
1626                 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1627 check_current:
1628                 /* Check if current node has a suitable gap */
1629                 if (gap_start > high_limit)
1630                         return -ENOMEM;
1631                 if (gap_end >= low_limit && gap_end - gap_start >= length)
1632                         goto found;
1633
1634                 /* Visit right subtree if it looks promising */
1635                 if (vma->vm_rb.rb_right) {
1636                         struct vm_area_struct *right =
1637                                 rb_entry(vma->vm_rb.rb_right,
1638                                          struct vm_area_struct, vm_rb);
1639                         if (right->rb_subtree_gap >= length) {
1640                                 vma = right;
1641                                 continue;
1642                         }
1643                 }
1644
1645                 /* Go back up the rbtree to find next candidate node */
1646                 while (true) {
1647                         struct rb_node *prev = &vma->vm_rb;
1648                         if (!rb_parent(prev))
1649                                 goto check_highest;
1650                         vma = rb_entry(rb_parent(prev),
1651                                        struct vm_area_struct, vm_rb);
1652                         if (prev == vma->vm_rb.rb_left) {
1653                                 gap_start = vma->vm_prev->vm_end;
1654                                 gap_end = vma->vm_start;
1655                                 goto check_current;
1656                         }
1657                 }
1658         }
1659
1660 check_highest:
1661         /* Check highest gap, which does not precede any rbtree node */
1662         gap_start = mm->highest_vm_end;
1663         gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1664         if (gap_start > high_limit)
1665                 return -ENOMEM;
1666
1667 found:
1668         /* We found a suitable gap. Clip it with the original low_limit. */
1669         if (gap_start < info->low_limit)
1670                 gap_start = info->low_limit;
1671
1672         /* Adjust gap address to the desired alignment */
1673         gap_start += (info->align_offset - gap_start) & info->align_mask;
1674
1675         VM_BUG_ON(gap_start + info->length > info->high_limit);
1676         VM_BUG_ON(gap_start + info->length > gap_end);
1677         return gap_start;
1678 }
1679
1680 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1681 {
1682         struct mm_struct *mm = current->mm;
1683         struct vm_area_struct *vma;
1684         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1685
1686         /* Adjust search length to account for worst case alignment overhead */
1687         length = info->length + info->align_mask;
1688         if (length < info->length)
1689                 return -ENOMEM;
1690
1691         /*
1692          * Adjust search limits by the desired length.
1693          * See implementation comment at top of unmapped_area().
1694          */
1695         gap_end = info->high_limit;
1696         if (gap_end < length)
1697                 return -ENOMEM;
1698         high_limit = gap_end - length;
1699
1700         if (info->low_limit > high_limit)
1701                 return -ENOMEM;
1702         low_limit = info->low_limit + length;
1703
1704         /* Check highest gap, which does not precede any rbtree node */
1705         gap_start = mm->highest_vm_end;
1706         if (gap_start <= high_limit)
1707                 goto found_highest;
1708
1709         /* Check if rbtree root looks promising */
1710         if (RB_EMPTY_ROOT(&mm->mm_rb))
1711                 return -ENOMEM;
1712         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1713         if (vma->rb_subtree_gap < length)
1714                 return -ENOMEM;
1715
1716         while (true) {
1717                 /* Visit right subtree if it looks promising */
1718                 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1719                 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1720                         struct vm_area_struct *right =
1721                                 rb_entry(vma->vm_rb.rb_right,
1722                                          struct vm_area_struct, vm_rb);
1723                         if (right->rb_subtree_gap >= length) {
1724                                 vma = right;
1725                                 continue;
1726                         }
1727                 }
1728
1729 check_current:
1730                 /* Check if current node has a suitable gap */
1731                 gap_end = vma->vm_start;
1732                 if (gap_end < low_limit)
1733                         return -ENOMEM;
1734                 if (gap_start <= high_limit && gap_end - gap_start >= length)
1735                         goto found;
1736
1737                 /* Visit left subtree if it looks promising */
1738                 if (vma->vm_rb.rb_left) {
1739                         struct vm_area_struct *left =
1740                                 rb_entry(vma->vm_rb.rb_left,
1741                                          struct vm_area_struct, vm_rb);
1742                         if (left->rb_subtree_gap >= length) {
1743                                 vma = left;
1744                                 continue;
1745                         }
1746                 }
1747
1748                 /* Go back up the rbtree to find next candidate node */
1749                 while (true) {
1750                         struct rb_node *prev = &vma->vm_rb;
1751                         if (!rb_parent(prev))
1752                                 return -ENOMEM;
1753                         vma = rb_entry(rb_parent(prev),
1754                                        struct vm_area_struct, vm_rb);
1755                         if (prev == vma->vm_rb.rb_right) {
1756                                 gap_start = vma->vm_prev ?
1757                                         vma->vm_prev->vm_end : 0;
1758                                 goto check_current;
1759                         }
1760                 }
1761         }
1762
1763 found:
1764         /* We found a suitable gap. Clip it with the original high_limit. */
1765         if (gap_end > info->high_limit)
1766                 gap_end = info->high_limit;
1767
1768 found_highest:
1769         /* Compute highest gap address at the desired alignment */
1770         gap_end -= info->length;
1771         gap_end -= (gap_end - info->align_offset) & info->align_mask;
1772
1773         VM_BUG_ON(gap_end < info->low_limit);
1774         VM_BUG_ON(gap_end < gap_start);
1775         return gap_end;
1776 }
1777
1778 /* Get an address range which is currently unmapped.
1779  * For shmat() with addr=0.
1780  *
1781  * Ugly calling convention alert:
1782  * Return value with the low bits set means error value,
1783  * ie
1784  *      if (ret & ~PAGE_MASK)
1785  *              error = ret;
1786  *
1787  * This function "knows" that -ENOMEM has the bits set.
1788  */
1789 #ifndef HAVE_ARCH_UNMAPPED_AREA
1790 unsigned long
1791 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1792                 unsigned long len, unsigned long pgoff, unsigned long flags)
1793 {
1794         struct mm_struct *mm = current->mm;
1795         struct vm_area_struct *vma;
1796         struct vm_unmapped_area_info info;
1797
1798         if (len > TASK_SIZE)
1799                 return -ENOMEM;
1800
1801         if (flags & MAP_FIXED)
1802                 return addr;
1803
1804         if (addr) {
1805                 addr = PAGE_ALIGN(addr);
1806                 vma = find_vma(mm, addr);
1807                 if (TASK_SIZE - len >= addr &&
1808                     (!vma || addr + len <= vma->vm_start))
1809                         return addr;
1810         }
1811
1812         info.flags = 0;
1813         info.length = len;
1814         info.low_limit = TASK_UNMAPPED_BASE;
1815         info.high_limit = TASK_SIZE;
1816         info.align_mask = 0;
1817         return vm_unmapped_area(&info);
1818 }
1819 #endif  
1820
1821 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1822 {
1823         /*
1824          * Is this a new hole at the lowest possible address?
1825          */
1826         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache)
1827                 mm->free_area_cache = addr;
1828 }
1829
1830 /*
1831  * This mmap-allocator allocates new areas top-down from below the
1832  * stack's low limit (the base):
1833  */
1834 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1835 unsigned long
1836 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1837                           const unsigned long len, const unsigned long pgoff,
1838                           const unsigned long flags)
1839 {
1840         struct vm_area_struct *vma;
1841         struct mm_struct *mm = current->mm;
1842         unsigned long addr = addr0;
1843         struct vm_unmapped_area_info info;
1844
1845         /* requested length too big for entire address space */
1846         if (len > TASK_SIZE)
1847                 return -ENOMEM;
1848
1849         if (flags & MAP_FIXED)
1850                 return addr;
1851
1852         /* requesting a specific address */
1853         if (addr) {
1854                 addr = PAGE_ALIGN(addr);
1855                 vma = find_vma(mm, addr);
1856                 if (TASK_SIZE - len >= addr &&
1857                                 (!vma || addr + len <= vma->vm_start))
1858                         return addr;
1859         }
1860
1861         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1862         info.length = len;
1863         info.low_limit = PAGE_SIZE;
1864         info.high_limit = mm->mmap_base;
1865         info.align_mask = 0;
1866         addr = vm_unmapped_area(&info);
1867
1868         /*
1869          * A failed mmap() very likely causes application failure,
1870          * so fall back to the bottom-up function here. This scenario
1871          * can happen with large stack limits and large mmap()
1872          * allocations.
1873          */
1874         if (addr & ~PAGE_MASK) {
1875                 VM_BUG_ON(addr != -ENOMEM);
1876                 info.flags = 0;
1877                 info.low_limit = TASK_UNMAPPED_BASE;
1878                 info.high_limit = TASK_SIZE;
1879                 addr = vm_unmapped_area(&info);
1880         }
1881
1882         return addr;
1883 }
1884 #endif
1885
1886 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1887 {
1888         /*
1889          * Is this a new hole at the highest possible address?
1890          */
1891         if (addr > mm->free_area_cache)
1892                 mm->free_area_cache = addr;
1893
1894         /* dont allow allocations above current base */
1895         if (mm->free_area_cache > mm->mmap_base)
1896                 mm->free_area_cache = mm->mmap_base;
1897 }
1898
1899 unsigned long
1900 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1901                 unsigned long pgoff, unsigned long flags)
1902 {
1903         unsigned long (*get_area)(struct file *, unsigned long,
1904                                   unsigned long, unsigned long, unsigned long);
1905
1906         unsigned long error = arch_mmap_check(addr, len, flags);
1907         if (error)
1908                 return error;
1909
1910         /* Careful about overflows.. */
1911         if (len > TASK_SIZE)
1912                 return -ENOMEM;
1913
1914         get_area = current->mm->get_unmapped_area;
1915         if (file && file->f_op && file->f_op->get_unmapped_area)
1916                 get_area = file->f_op->get_unmapped_area;
1917         addr = get_area(file, addr, len, pgoff, flags);
1918         if (IS_ERR_VALUE(addr))
1919                 return addr;
1920
1921         if (addr > TASK_SIZE - len)
1922                 return -ENOMEM;
1923         if (addr & ~PAGE_MASK)
1924                 return -EINVAL;
1925
1926         addr = arch_rebalance_pgtables(addr, len);
1927         error = security_mmap_addr(addr);
1928         return error ? error : addr;
1929 }
1930
1931 EXPORT_SYMBOL(get_unmapped_area);
1932
1933 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1934 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1935 {
1936         struct vm_area_struct *vma = NULL;
1937
1938         if (WARN_ON_ONCE(!mm))          /* Remove this in linux-3.6 */
1939                 return NULL;
1940
1941         /* Check the cache first. */
1942         /* (Cache hit rate is typically around 35%.) */
1943         vma = mm->mmap_cache;
1944         if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1945                 struct rb_node *rb_node;
1946
1947                 rb_node = mm->mm_rb.rb_node;
1948                 vma = NULL;
1949
1950                 while (rb_node) {
1951                         struct vm_area_struct *vma_tmp;
1952
1953                         vma_tmp = rb_entry(rb_node,
1954                                            struct vm_area_struct, vm_rb);
1955
1956                         if (vma_tmp->vm_end > addr) {
1957                                 vma = vma_tmp;
1958                                 if (vma_tmp->vm_start <= addr)
1959                                         break;
1960                                 rb_node = rb_node->rb_left;
1961                         } else
1962                                 rb_node = rb_node->rb_right;
1963                 }
1964                 if (vma)
1965                         mm->mmap_cache = vma;
1966         }
1967         return vma;
1968 }
1969
1970 EXPORT_SYMBOL(find_vma);
1971
1972 /*
1973  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1974  */
1975 struct vm_area_struct *
1976 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1977                         struct vm_area_struct **pprev)
1978 {
1979         struct vm_area_struct *vma;
1980
1981         vma = find_vma(mm, addr);
1982         if (vma) {
1983                 *pprev = vma->vm_prev;
1984         } else {
1985                 struct rb_node *rb_node = mm->mm_rb.rb_node;
1986                 *pprev = NULL;
1987                 while (rb_node) {
1988                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1989                         rb_node = rb_node->rb_right;
1990                 }
1991         }
1992         return vma;
1993 }
1994
1995 /*
1996  * Verify that the stack growth is acceptable and
1997  * update accounting. This is shared with both the
1998  * grow-up and grow-down cases.
1999  */
2000 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
2001 {
2002         struct mm_struct *mm = vma->vm_mm;
2003         struct rlimit *rlim = current->signal->rlim;
2004         unsigned long new_start;
2005
2006         /* address space limit tests */
2007         if (!may_expand_vm(mm, grow))
2008                 return -ENOMEM;
2009
2010         /* Stack limit test */
2011         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2012                 return -ENOMEM;
2013
2014         /* mlock limit tests */
2015         if (vma->vm_flags & VM_LOCKED) {
2016                 unsigned long locked;
2017                 unsigned long limit;
2018                 locked = mm->locked_vm + grow;
2019                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2020                 limit >>= PAGE_SHIFT;
2021                 if (locked > limit && !capable(CAP_IPC_LOCK))
2022                         return -ENOMEM;
2023         }
2024
2025         /* Check to ensure the stack will not grow into a hugetlb-only region */
2026         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2027                         vma->vm_end - size;
2028         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2029                 return -EFAULT;
2030
2031         /*
2032          * Overcommit..  This must be the final test, as it will
2033          * update security statistics.
2034          */
2035         if (security_vm_enough_memory_mm(mm, grow))
2036                 return -ENOMEM;
2037
2038         /* Ok, everything looks good - let it rip */
2039         if (vma->vm_flags & VM_LOCKED)
2040                 mm->locked_vm += grow;
2041         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
2042         return 0;
2043 }
2044
2045 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2046 /*
2047  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2048  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2049  */
2050 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2051 {
2052         int error;
2053
2054         if (!(vma->vm_flags & VM_GROWSUP))
2055                 return -EFAULT;
2056
2057         /*
2058          * We must make sure the anon_vma is allocated
2059          * so that the anon_vma locking is not a noop.
2060          */
2061         if (unlikely(anon_vma_prepare(vma)))
2062                 return -ENOMEM;
2063         vma_lock_anon_vma(vma);
2064
2065         /*
2066          * vma->vm_start/vm_end cannot change under us because the caller
2067          * is required to hold the mmap_sem in read mode.  We need the
2068          * anon_vma lock to serialize against concurrent expand_stacks.
2069          * Also guard against wrapping around to address 0.
2070          */
2071         if (address < PAGE_ALIGN(address+4))
2072                 address = PAGE_ALIGN(address+4);
2073         else {
2074                 vma_unlock_anon_vma(vma);
2075                 return -ENOMEM;
2076         }
2077         error = 0;
2078
2079         /* Somebody else might have raced and expanded it already */
2080         if (address > vma->vm_end) {
2081                 unsigned long size, grow;
2082
2083                 size = address - vma->vm_start;
2084                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2085
2086                 error = -ENOMEM;
2087                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2088                         error = acct_stack_growth(vma, size, grow);
2089                         if (!error) {
2090                                 /*
2091                                  * vma_gap_update() doesn't support concurrent
2092                                  * updates, but we only hold a shared mmap_sem
2093                                  * lock here, so we need to protect against
2094                                  * concurrent vma expansions.
2095                                  * vma_lock_anon_vma() doesn't help here, as
2096                                  * we don't guarantee that all growable vmas
2097                                  * in a mm share the same root anon vma.
2098                                  * So, we reuse mm->page_table_lock to guard
2099                                  * against concurrent vma expansions.
2100                                  */
2101                                 spin_lock(&vma->vm_mm->page_table_lock);
2102                                 anon_vma_interval_tree_pre_update_vma(vma);
2103                                 vma->vm_end = address;
2104                                 anon_vma_interval_tree_post_update_vma(vma);
2105                                 if (vma->vm_next)
2106                                         vma_gap_update(vma->vm_next);
2107                                 else
2108                                         vma->vm_mm->highest_vm_end = address;
2109                                 spin_unlock(&vma->vm_mm->page_table_lock);
2110
2111                                 perf_event_mmap(vma);
2112                         }
2113                 }
2114         }
2115         vma_unlock_anon_vma(vma);
2116         khugepaged_enter_vma_merge(vma);
2117         validate_mm(vma->vm_mm);
2118         return error;
2119 }
2120 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2121
2122 /*
2123  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2124  */
2125 int expand_downwards(struct vm_area_struct *vma,
2126                                    unsigned long address)
2127 {
2128         int error;
2129
2130         /*
2131          * We must make sure the anon_vma is allocated
2132          * so that the anon_vma locking is not a noop.
2133          */
2134         if (unlikely(anon_vma_prepare(vma)))
2135                 return -ENOMEM;
2136
2137         address &= PAGE_MASK;
2138         error = security_mmap_addr(address);
2139         if (error)
2140                 return error;
2141
2142         vma_lock_anon_vma(vma);
2143
2144         /*
2145          * vma->vm_start/vm_end cannot change under us because the caller
2146          * is required to hold the mmap_sem in read mode.  We need the
2147          * anon_vma lock to serialize against concurrent expand_stacks.
2148          */
2149
2150         /* Somebody else might have raced and expanded it already */
2151         if (address < vma->vm_start) {
2152                 unsigned long size, grow;
2153
2154                 size = vma->vm_end - address;
2155                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2156
2157                 error = -ENOMEM;
2158                 if (grow <= vma->vm_pgoff) {
2159                         error = acct_stack_growth(vma, size, grow);
2160                         if (!error) {
2161                                 /*
2162                                  * vma_gap_update() doesn't support concurrent
2163                                  * updates, but we only hold a shared mmap_sem
2164                                  * lock here, so we need to protect against
2165                                  * concurrent vma expansions.
2166                                  * vma_lock_anon_vma() doesn't help here, as
2167                                  * we don't guarantee that all growable vmas
2168                                  * in a mm share the same root anon vma.
2169                                  * So, we reuse mm->page_table_lock to guard
2170                                  * against concurrent vma expansions.
2171                                  */
2172                                 spin_lock(&vma->vm_mm->page_table_lock);
2173                                 anon_vma_interval_tree_pre_update_vma(vma);
2174                                 vma->vm_start = address;
2175                                 vma->vm_pgoff -= grow;
2176                                 anon_vma_interval_tree_post_update_vma(vma);
2177                                 vma_gap_update(vma);
2178                                 spin_unlock(&vma->vm_mm->page_table_lock);
2179
2180                                 perf_event_mmap(vma);
2181                         }
2182                 }
2183         }
2184         vma_unlock_anon_vma(vma);
2185         khugepaged_enter_vma_merge(vma);
2186         validate_mm(vma->vm_mm);
2187         return error;
2188 }
2189
2190 #ifdef CONFIG_STACK_GROWSUP
2191 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2192 {
2193         return expand_upwards(vma, address);
2194 }
2195
2196 struct vm_area_struct *
2197 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2198 {
2199         struct vm_area_struct *vma, *prev;
2200
2201         addr &= PAGE_MASK;
2202         vma = find_vma_prev(mm, addr, &prev);
2203         if (vma && (vma->vm_start <= addr))
2204                 return vma;
2205         if (!prev || expand_stack(prev, addr))
2206                 return NULL;
2207         if (prev->vm_flags & VM_LOCKED) {
2208                 mlock_vma_pages_range(prev, addr, prev->vm_end);
2209         }
2210         return prev;
2211 }
2212 #else
2213 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2214 {
2215         return expand_downwards(vma, address);
2216 }
2217
2218 struct vm_area_struct *
2219 find_extend_vma(struct mm_struct * mm, unsigned long addr)
2220 {
2221         struct vm_area_struct * vma;
2222         unsigned long start;
2223
2224         addr &= PAGE_MASK;
2225         vma = find_vma(mm,addr);
2226         if (!vma)
2227                 return NULL;
2228         if (vma->vm_start <= addr)
2229                 return vma;
2230         if (!(vma->vm_flags & VM_GROWSDOWN))
2231                 return NULL;
2232         start = vma->vm_start;
2233         if (expand_stack(vma, addr))
2234                 return NULL;
2235         if (vma->vm_flags & VM_LOCKED) {
2236                 mlock_vma_pages_range(vma, addr, start);
2237         }
2238         return vma;
2239 }
2240 #endif
2241
2242 /*
2243  * Ok - we have the memory areas we should free on the vma list,
2244  * so release them, and do the vma updates.
2245  *
2246  * Called with the mm semaphore held.
2247  */
2248 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2249 {
2250         unsigned long nr_accounted = 0;
2251
2252         /* Update high watermark before we lower total_vm */
2253         update_hiwater_vm(mm);
2254         do {
2255                 long nrpages = vma_pages(vma);
2256
2257                 if (vma->vm_flags & VM_ACCOUNT)
2258                         nr_accounted += nrpages;
2259                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
2260                 vma = remove_vma(vma);
2261         } while (vma);
2262         vm_unacct_memory(nr_accounted);
2263         validate_mm(mm);
2264 }
2265
2266 /*
2267  * Get rid of page table information in the indicated region.
2268  *
2269  * Called with the mm semaphore held.
2270  */
2271 static void unmap_region(struct mm_struct *mm,
2272                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2273                 unsigned long start, unsigned long end)
2274 {
2275         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
2276         struct mmu_gather tlb;
2277
2278         lru_add_drain();
2279         tlb_gather_mmu(&tlb, mm, 0);
2280         update_hiwater_rss(mm);
2281         unmap_vmas(&tlb, vma, start, end);
2282         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2283                                  next ? next->vm_start : 0);
2284         tlb_finish_mmu(&tlb, start, end);
2285 }
2286
2287 /*
2288  * Create a list of vma's touched by the unmap, removing them from the mm's
2289  * vma list as we go..
2290  */
2291 static void
2292 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2293         struct vm_area_struct *prev, unsigned long end)
2294 {
2295         struct vm_area_struct **insertion_point;
2296         struct vm_area_struct *tail_vma = NULL;
2297         unsigned long addr;
2298
2299         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2300         vma->vm_prev = NULL;
2301         do {
2302                 vma_rb_erase(vma, &mm->mm_rb);
2303                 mm->map_count--;
2304                 tail_vma = vma;
2305                 vma = vma->vm_next;
2306         } while (vma && vma->vm_start < end);
2307         *insertion_point = vma;
2308         if (vma) {
2309                 vma->vm_prev = prev;
2310                 vma_gap_update(vma);
2311         } else
2312                 mm->highest_vm_end = prev ? prev->vm_end : 0;
2313         tail_vma->vm_next = NULL;
2314         if (mm->unmap_area == arch_unmap_area)
2315                 addr = prev ? prev->vm_end : mm->mmap_base;
2316         else
2317                 addr = vma ?  vma->vm_start : mm->mmap_base;
2318         mm->unmap_area(mm, addr);
2319         mm->mmap_cache = NULL;          /* Kill the cache. */
2320 }
2321
2322 /*
2323  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
2324  * munmap path where it doesn't make sense to fail.
2325  */
2326 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
2327               unsigned long addr, int new_below)
2328 {
2329         struct mempolicy *pol;
2330         struct vm_area_struct *new;
2331         int err = -ENOMEM;
2332
2333         if (is_vm_hugetlb_page(vma) && (addr &
2334                                         ~(huge_page_mask(hstate_vma(vma)))))
2335                 return -EINVAL;
2336
2337         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2338         if (!new)
2339                 goto out_err;
2340
2341         /* most fields are the same, copy all, and then fixup */
2342         *new = *vma;
2343
2344         INIT_LIST_HEAD(&new->anon_vma_chain);
2345
2346         if (new_below)
2347                 new->vm_end = addr;
2348         else {
2349                 new->vm_start = addr;
2350                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2351         }
2352
2353         pol = mpol_dup(vma_policy(vma));
2354         if (IS_ERR(pol)) {
2355                 err = PTR_ERR(pol);
2356                 goto out_free_vma;
2357         }
2358         vma_set_policy(new, pol);
2359
2360         if (anon_vma_clone(new, vma))
2361                 goto out_free_mpol;
2362
2363         if (new->vm_file)
2364                 get_file(new->vm_file);
2365
2366         if (new->vm_ops && new->vm_ops->open)
2367                 new->vm_ops->open(new);
2368
2369         if (new_below)
2370                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2371                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2372         else
2373                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2374
2375         /* Success. */
2376         if (!err)
2377                 return 0;
2378
2379         /* Clean everything up if vma_adjust failed. */
2380         if (new->vm_ops && new->vm_ops->close)
2381                 new->vm_ops->close(new);
2382         if (new->vm_file)
2383                 fput(new->vm_file);
2384         unlink_anon_vmas(new);
2385  out_free_mpol:
2386         mpol_put(pol);
2387  out_free_vma:
2388         kmem_cache_free(vm_area_cachep, new);
2389  out_err:
2390         return err;
2391 }
2392
2393 /*
2394  * Split a vma into two pieces at address 'addr', a new vma is allocated
2395  * either for the first part or the tail.
2396  */
2397 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2398               unsigned long addr, int new_below)
2399 {
2400         if (mm->map_count >= sysctl_max_map_count)
2401                 return -ENOMEM;
2402
2403         return __split_vma(mm, vma, addr, new_below);
2404 }
2405
2406 /* Munmap is split into 2 main parts -- this part which finds
2407  * what needs doing, and the areas themselves, which do the
2408  * work.  This now handles partial unmappings.
2409  * Jeremy Fitzhardinge <jeremy@goop.org>
2410  */
2411 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2412 {
2413         unsigned long end;
2414         struct vm_area_struct *vma, *prev, *last;
2415
2416         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2417                 return -EINVAL;
2418
2419         if ((len = PAGE_ALIGN(len)) == 0)
2420                 return -EINVAL;
2421
2422         /* Find the first overlapping VMA */
2423         vma = find_vma(mm, start);
2424         if (!vma)
2425                 return 0;
2426         prev = vma->vm_prev;
2427         /* we have  start < vma->vm_end  */
2428
2429         /* if it doesn't overlap, we have nothing.. */
2430         end = start + len;
2431         if (vma->vm_start >= end)
2432                 return 0;
2433
2434         /*
2435          * If we need to split any vma, do it now to save pain later.
2436          *
2437          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2438          * unmapped vm_area_struct will remain in use: so lower split_vma
2439          * places tmp vma above, and higher split_vma places tmp vma below.
2440          */
2441         if (start > vma->vm_start) {
2442                 int error;
2443
2444                 /*
2445                  * Make sure that map_count on return from munmap() will
2446                  * not exceed its limit; but let map_count go just above
2447                  * its limit temporarily, to help free resources as expected.
2448                  */
2449                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2450                         return -ENOMEM;
2451
2452                 error = __split_vma(mm, vma, start, 0);
2453                 if (error)
2454                         return error;
2455                 prev = vma;
2456         }
2457
2458         /* Does it split the last one? */
2459         last = find_vma(mm, end);
2460         if (last && end > last->vm_start) {
2461                 int error = __split_vma(mm, last, end, 1);
2462                 if (error)
2463                         return error;
2464         }
2465         vma = prev? prev->vm_next: mm->mmap;
2466
2467         /*
2468          * unlock any mlock()ed ranges before detaching vmas
2469          */
2470         if (mm->locked_vm) {
2471                 struct vm_area_struct *tmp = vma;
2472                 while (tmp && tmp->vm_start < end) {
2473                         if (tmp->vm_flags & VM_LOCKED) {
2474                                 mm->locked_vm -= vma_pages(tmp);
2475                                 munlock_vma_pages_all(tmp);
2476                         }
2477                         tmp = tmp->vm_next;
2478                 }
2479         }
2480
2481         /*
2482          * Remove the vma's, and unmap the actual pages
2483          */
2484         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2485         unmap_region(mm, vma, prev, start, end);
2486
2487         /* Fix up all other VM information */
2488         remove_vma_list(mm, vma);
2489
2490         return 0;
2491 }
2492
2493 int vm_munmap(unsigned long start, size_t len)
2494 {
2495         int ret;
2496         struct mm_struct *mm = current->mm;
2497
2498         down_write(&mm->mmap_sem);
2499         ret = do_munmap(mm, start, len);
2500         up_write(&mm->mmap_sem);
2501         return ret;
2502 }
2503 EXPORT_SYMBOL(vm_munmap);
2504
2505 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2506 {
2507         profile_munmap(addr);
2508         return vm_munmap(addr, len);
2509 }
2510
2511 static inline void verify_mm_writelocked(struct mm_struct *mm)
2512 {
2513 #ifdef CONFIG_DEBUG_VM
2514         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2515                 WARN_ON(1);
2516                 up_read(&mm->mmap_sem);
2517         }
2518 #endif
2519 }
2520
2521 /*
2522  *  this is really a simplified "do_mmap".  it only handles
2523  *  anonymous maps.  eventually we may be able to do some
2524  *  brk-specific accounting here.
2525  */
2526 static unsigned long do_brk(unsigned long addr, unsigned long len)
2527 {
2528         struct mm_struct * mm = current->mm;
2529         struct vm_area_struct * vma, * prev;
2530         unsigned long flags;
2531         struct rb_node ** rb_link, * rb_parent;
2532         pgoff_t pgoff = addr >> PAGE_SHIFT;
2533         int error;
2534
2535         len = PAGE_ALIGN(len);
2536         if (!len)
2537                 return addr;
2538
2539         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2540
2541         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2542         if (error & ~PAGE_MASK)
2543                 return error;
2544
2545         /*
2546          * mlock MCL_FUTURE?
2547          */
2548         if (mm->def_flags & VM_LOCKED) {
2549                 unsigned long locked, lock_limit;
2550                 locked = len >> PAGE_SHIFT;
2551                 locked += mm->locked_vm;
2552                 lock_limit = rlimit(RLIMIT_MEMLOCK);
2553                 lock_limit >>= PAGE_SHIFT;
2554                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2555                         return -EAGAIN;
2556         }
2557
2558         /*
2559          * mm->mmap_sem is required to protect against another thread
2560          * changing the mappings in case we sleep.
2561          */
2562         verify_mm_writelocked(mm);
2563
2564         /*
2565          * Clear old maps.  this also does some error checking for us
2566          */
2567  munmap_back:
2568         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
2569                 if (do_munmap(mm, addr, len))
2570                         return -ENOMEM;
2571                 goto munmap_back;
2572         }
2573
2574         /* Check against address space limits *after* clearing old maps... */
2575         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2576                 return -ENOMEM;
2577
2578         if (mm->map_count > sysctl_max_map_count)
2579                 return -ENOMEM;
2580
2581         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2582                 return -ENOMEM;
2583
2584         /* Can we just expand an old private anonymous mapping? */
2585         vma = vma_merge(mm, prev, addr, addr + len, flags,
2586                                         NULL, NULL, pgoff, NULL);
2587         if (vma)
2588                 goto out;
2589
2590         /*
2591          * create a vma struct for an anonymous mapping
2592          */
2593         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2594         if (!vma) {
2595                 vm_unacct_memory(len >> PAGE_SHIFT);
2596                 return -ENOMEM;
2597         }
2598
2599         INIT_LIST_HEAD(&vma->anon_vma_chain);
2600         vma->vm_mm = mm;
2601         vma->vm_start = addr;
2602         vma->vm_end = addr + len;
2603         vma->vm_pgoff = pgoff;
2604         vma->vm_flags = flags;
2605         vma->vm_page_prot = vm_get_page_prot(flags);
2606         vma_link(mm, vma, prev, rb_link, rb_parent);
2607 out:
2608         perf_event_mmap(vma);
2609         mm->total_vm += len >> PAGE_SHIFT;
2610         if (flags & VM_LOCKED)
2611                 mm->locked_vm += (len >> PAGE_SHIFT);
2612         return addr;
2613 }
2614
2615 unsigned long vm_brk(unsigned long addr, unsigned long len)
2616 {
2617         struct mm_struct *mm = current->mm;
2618         unsigned long ret;
2619         bool populate;
2620
2621         down_write(&mm->mmap_sem);
2622         ret = do_brk(addr, len);
2623         populate = ((mm->def_flags & VM_LOCKED) != 0);
2624         up_write(&mm->mmap_sem);
2625         if (populate)
2626                 mm_populate(addr, len);
2627         return ret;
2628 }
2629 EXPORT_SYMBOL(vm_brk);
2630
2631 /* Release all mmaps. */
2632 void exit_mmap(struct mm_struct *mm)
2633 {
2634         struct mmu_gather tlb;
2635         struct vm_area_struct *vma;
2636         unsigned long nr_accounted = 0;
2637
2638         /* mm's last user has gone, and its about to be pulled down */
2639         mmu_notifier_release(mm);
2640
2641         if (mm->locked_vm) {
2642                 vma = mm->mmap;
2643                 while (vma) {
2644                         if (vma->vm_flags & VM_LOCKED)
2645                                 munlock_vma_pages_all(vma);
2646                         vma = vma->vm_next;
2647                 }
2648         }
2649
2650         arch_exit_mmap(mm);
2651
2652         vma = mm->mmap;
2653         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2654                 return;
2655
2656         lru_add_drain();
2657         flush_cache_mm(mm);
2658         tlb_gather_mmu(&tlb, mm, 1);
2659         /* update_hiwater_rss(mm) here? but nobody should be looking */
2660         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2661         unmap_vmas(&tlb, vma, 0, -1);
2662
2663         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2664         tlb_finish_mmu(&tlb, 0, -1);
2665
2666         /*
2667          * Walk the list again, actually closing and freeing it,
2668          * with preemption enabled, without holding any MM locks.
2669          */
2670         while (vma) {
2671                 if (vma->vm_flags & VM_ACCOUNT)
2672                         nr_accounted += vma_pages(vma);
2673                 vma = remove_vma(vma);
2674         }
2675         vm_unacct_memory(nr_accounted);
2676
2677         WARN_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2678 }
2679
2680 /* Insert vm structure into process list sorted by address
2681  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2682  * then i_mmap_mutex is taken here.
2683  */
2684 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2685 {
2686         struct vm_area_struct *prev;
2687         struct rb_node **rb_link, *rb_parent;
2688
2689         /*
2690          * The vm_pgoff of a purely anonymous vma should be irrelevant
2691          * until its first write fault, when page's anon_vma and index
2692          * are set.  But now set the vm_pgoff it will almost certainly
2693          * end up with (unless mremap moves it elsewhere before that
2694          * first wfault), so /proc/pid/maps tells a consistent story.
2695          *
2696          * By setting it to reflect the virtual start address of the
2697          * vma, merges and splits can happen in a seamless way, just
2698          * using the existing file pgoff checks and manipulations.
2699          * Similarly in do_mmap_pgoff and in do_brk.
2700          */
2701         if (!vma->vm_file) {
2702                 BUG_ON(vma->anon_vma);
2703                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2704         }
2705         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2706                            &prev, &rb_link, &rb_parent))
2707                 return -ENOMEM;
2708         if ((vma->vm_flags & VM_ACCOUNT) &&
2709              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2710                 return -ENOMEM;
2711
2712         vma_link(mm, vma, prev, rb_link, rb_parent);
2713         return 0;
2714 }
2715
2716 /*
2717  * Copy the vma structure to a new location in the same mm,
2718  * prior to moving page table entries, to effect an mremap move.
2719  */
2720 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2721         unsigned long addr, unsigned long len, pgoff_t pgoff,
2722         bool *need_rmap_locks)
2723 {
2724         struct vm_area_struct *vma = *vmap;
2725         unsigned long vma_start = vma->vm_start;
2726         struct mm_struct *mm = vma->vm_mm;
2727         struct vm_area_struct *new_vma, *prev;
2728         struct rb_node **rb_link, *rb_parent;
2729         struct mempolicy *pol;
2730         bool faulted_in_anon_vma = true;
2731
2732         /*
2733          * If anonymous vma has not yet been faulted, update new pgoff
2734          * to match new location, to increase its chance of merging.
2735          */
2736         if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2737                 pgoff = addr >> PAGE_SHIFT;
2738                 faulted_in_anon_vma = false;
2739         }
2740
2741         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2742                 return NULL;    /* should never get here */
2743         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2744                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2745         if (new_vma) {
2746                 /*
2747                  * Source vma may have been merged into new_vma
2748                  */
2749                 if (unlikely(vma_start >= new_vma->vm_start &&
2750                              vma_start < new_vma->vm_end)) {
2751                         /*
2752                          * The only way we can get a vma_merge with
2753                          * self during an mremap is if the vma hasn't
2754                          * been faulted in yet and we were allowed to
2755                          * reset the dst vma->vm_pgoff to the
2756                          * destination address of the mremap to allow
2757                          * the merge to happen. mremap must change the
2758                          * vm_pgoff linearity between src and dst vmas
2759                          * (in turn preventing a vma_merge) to be
2760                          * safe. It is only safe to keep the vm_pgoff
2761                          * linear if there are no pages mapped yet.
2762                          */
2763                         VM_BUG_ON(faulted_in_anon_vma);
2764                         *vmap = vma = new_vma;
2765                 }
2766                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2767         } else {
2768                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2769                 if (new_vma) {
2770                         *new_vma = *vma;
2771                         new_vma->vm_start = addr;
2772                         new_vma->vm_end = addr + len;
2773                         new_vma->vm_pgoff = pgoff;
2774                         pol = mpol_dup(vma_policy(vma));
2775                         if (IS_ERR(pol))
2776                                 goto out_free_vma;
2777                         vma_set_policy(new_vma, pol);
2778                         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2779                         if (anon_vma_clone(new_vma, vma))
2780                                 goto out_free_mempol;
2781                         if (new_vma->vm_file)
2782                                 get_file(new_vma->vm_file);
2783                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2784                                 new_vma->vm_ops->open(new_vma);
2785                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2786                         *need_rmap_locks = false;
2787                 }
2788         }
2789         return new_vma;
2790
2791  out_free_mempol:
2792         mpol_put(pol);
2793  out_free_vma:
2794         kmem_cache_free(vm_area_cachep, new_vma);
2795         return NULL;
2796 }
2797
2798 /*
2799  * Return true if the calling process may expand its vm space by the passed
2800  * number of pages
2801  */
2802 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2803 {
2804         unsigned long cur = mm->total_vm;       /* pages */
2805         unsigned long lim;
2806
2807         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2808
2809         if (cur + npages > lim)
2810                 return 0;
2811         return 1;
2812 }
2813
2814
2815 static int special_mapping_fault(struct vm_area_struct *vma,
2816                                 struct vm_fault *vmf)
2817 {
2818         pgoff_t pgoff;
2819         struct page **pages;
2820
2821         /*
2822          * special mappings have no vm_file, and in that case, the mm
2823          * uses vm_pgoff internally. So we have to subtract it from here.
2824          * We are allowed to do this because we are the mm; do not copy
2825          * this code into drivers!
2826          */
2827         pgoff = vmf->pgoff - vma->vm_pgoff;
2828
2829         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2830                 pgoff--;
2831
2832         if (*pages) {
2833                 struct page *page = *pages;
2834                 get_page(page);
2835                 vmf->page = page;
2836                 return 0;
2837         }
2838
2839         return VM_FAULT_SIGBUS;
2840 }
2841
2842 /*
2843  * Having a close hook prevents vma merging regardless of flags.
2844  */
2845 static void special_mapping_close(struct vm_area_struct *vma)
2846 {
2847 }
2848
2849 static const struct vm_operations_struct special_mapping_vmops = {
2850         .close = special_mapping_close,
2851         .fault = special_mapping_fault,
2852 };
2853
2854 /*
2855  * Called with mm->mmap_sem held for writing.
2856  * Insert a new vma covering the given region, with the given flags.
2857  * Its pages are supplied by the given array of struct page *.
2858  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2859  * The region past the last page supplied will always produce SIGBUS.
2860  * The array pointer and the pages it points to are assumed to stay alive
2861  * for as long as this mapping might exist.
2862  */
2863 int install_special_mapping(struct mm_struct *mm,
2864                             unsigned long addr, unsigned long len,
2865                             unsigned long vm_flags, struct page **pages)
2866 {
2867         int ret;
2868         struct vm_area_struct *vma;
2869
2870         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2871         if (unlikely(vma == NULL))
2872                 return -ENOMEM;
2873
2874         INIT_LIST_HEAD(&vma->anon_vma_chain);
2875         vma->vm_mm = mm;
2876         vma->vm_start = addr;
2877         vma->vm_end = addr + len;
2878
2879         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2880         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2881
2882         vma->vm_ops = &special_mapping_vmops;
2883         vma->vm_private_data = pages;
2884
2885         ret = insert_vm_struct(mm, vma);
2886         if (ret)
2887                 goto out;
2888
2889         mm->total_vm += len >> PAGE_SHIFT;
2890
2891         perf_event_mmap(vma);
2892
2893         return 0;
2894
2895 out:
2896         kmem_cache_free(vm_area_cachep, vma);
2897         return ret;
2898 }
2899
2900 static DEFINE_MUTEX(mm_all_locks_mutex);
2901
2902 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2903 {
2904         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
2905                 /*
2906                  * The LSB of head.next can't change from under us
2907                  * because we hold the mm_all_locks_mutex.
2908                  */
2909                 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
2910                 /*
2911                  * We can safely modify head.next after taking the
2912                  * anon_vma->root->rwsem. If some other vma in this mm shares
2913                  * the same anon_vma we won't take it again.
2914                  *
2915                  * No need of atomic instructions here, head.next
2916                  * can't change from under us thanks to the
2917                  * anon_vma->root->rwsem.
2918                  */
2919                 if (__test_and_set_bit(0, (unsigned long *)
2920                                        &anon_vma->root->rb_root.rb_node))
2921                         BUG();
2922         }
2923 }
2924
2925 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2926 {
2927         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2928                 /*
2929                  * AS_MM_ALL_LOCKS can't change from under us because
2930                  * we hold the mm_all_locks_mutex.
2931                  *
2932                  * Operations on ->flags have to be atomic because
2933                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2934                  * mm_all_locks_mutex, there may be other cpus
2935                  * changing other bitflags in parallel to us.
2936                  */
2937                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2938                         BUG();
2939                 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2940         }
2941 }
2942
2943 /*
2944  * This operation locks against the VM for all pte/vma/mm related
2945  * operations that could ever happen on a certain mm. This includes
2946  * vmtruncate, try_to_unmap, and all page faults.
2947  *
2948  * The caller must take the mmap_sem in write mode before calling
2949  * mm_take_all_locks(). The caller isn't allowed to release the
2950  * mmap_sem until mm_drop_all_locks() returns.
2951  *
2952  * mmap_sem in write mode is required in order to block all operations
2953  * that could modify pagetables and free pages without need of
2954  * altering the vma layout (for example populate_range() with
2955  * nonlinear vmas). It's also needed in write mode to avoid new
2956  * anon_vmas to be associated with existing vmas.
2957  *
2958  * A single task can't take more than one mm_take_all_locks() in a row
2959  * or it would deadlock.
2960  *
2961  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
2962  * mapping->flags avoid to take the same lock twice, if more than one
2963  * vma in this mm is backed by the same anon_vma or address_space.
2964  *
2965  * We can take all the locks in random order because the VM code
2966  * taking i_mmap_mutex or anon_vma->rwsem outside the mmap_sem never
2967  * takes more than one of them in a row. Secondly we're protected
2968  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2969  *
2970  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2971  * that may have to take thousand of locks.
2972  *
2973  * mm_take_all_locks() can fail if it's interrupted by signals.
2974  */
2975 int mm_take_all_locks(struct mm_struct *mm)
2976 {
2977         struct vm_area_struct *vma;
2978         struct anon_vma_chain *avc;
2979
2980         BUG_ON(down_read_trylock(&mm->mmap_sem));
2981
2982         mutex_lock(&mm_all_locks_mutex);
2983
2984         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2985                 if (signal_pending(current))
2986                         goto out_unlock;
2987                 if (vma->vm_file && vma->vm_file->f_mapping)
2988                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2989         }
2990
2991         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2992                 if (signal_pending(current))
2993                         goto out_unlock;
2994                 if (vma->anon_vma)
2995                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2996                                 vm_lock_anon_vma(mm, avc->anon_vma);
2997         }
2998
2999         return 0;
3000
3001 out_unlock:
3002         mm_drop_all_locks(mm);
3003         return -EINTR;
3004 }
3005
3006 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3007 {
3008         if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3009                 /*
3010                  * The LSB of head.next can't change to 0 from under
3011                  * us because we hold the mm_all_locks_mutex.
3012                  *
3013                  * We must however clear the bitflag before unlocking
3014                  * the vma so the users using the anon_vma->rb_root will
3015                  * never see our bitflag.
3016                  *
3017                  * No need of atomic instructions here, head.next
3018                  * can't change from under us until we release the
3019                  * anon_vma->root->rwsem.
3020                  */
3021                 if (!__test_and_clear_bit(0, (unsigned long *)
3022                                           &anon_vma->root->rb_root.rb_node))
3023                         BUG();
3024                 anon_vma_unlock(anon_vma);
3025         }
3026 }
3027
3028 static void vm_unlock_mapping(struct address_space *mapping)
3029 {
3030         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3031                 /*
3032                  * AS_MM_ALL_LOCKS can't change to 0 from under us
3033                  * because we hold the mm_all_locks_mutex.
3034                  */
3035                 mutex_unlock(&mapping->i_mmap_mutex);
3036                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3037                                         &mapping->flags))
3038                         BUG();
3039         }
3040 }
3041
3042 /*
3043  * The mmap_sem cannot be released by the caller until
3044  * mm_drop_all_locks() returns.
3045  */
3046 void mm_drop_all_locks(struct mm_struct *mm)
3047 {
3048         struct vm_area_struct *vma;
3049         struct anon_vma_chain *avc;
3050
3051         BUG_ON(down_read_trylock(&mm->mmap_sem));
3052         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3053
3054         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3055                 if (vma->anon_vma)
3056                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3057                                 vm_unlock_anon_vma(avc->anon_vma);
3058                 if (vma->vm_file && vma->vm_file->f_mapping)
3059                         vm_unlock_mapping(vma->vm_file->f_mapping);
3060         }
3061
3062         mutex_unlock(&mm_all_locks_mutex);
3063 }
3064
3065 /*
3066  * initialise the VMA slab
3067  */
3068 void __init mmap_init(void)
3069 {
3070         int ret;
3071
3072         ret = percpu_counter_init(&vm_committed_as, 0);
3073         VM_BUG_ON(ret);
3074 }