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