mm: migration: add migrate_entry_wait_huge()
[linux-3.10.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
7  */
8
9 #include <linux/kernel.h>
10 #include <linux/slab.h>
11 #include <linux/backing-dev.h>
12 #include <linux/mm.h>
13 #include <linux/shm.h>
14 #include <linux/mman.h>
15 #include <linux/pagemap.h>
16 #include <linux/swap.h>
17 #include <linux/syscalls.h>
18 #include <linux/capability.h>
19 #include <linux/init.h>
20 #include <linux/file.h>
21 #include <linux/fs.h>
22 #include <linux/personality.h>
23 #include <linux/security.h>
24 #include <linux/hugetlb.h>
25 #include <linux/profile.h>
26 #include <linux/export.h>
27 #include <linux/mount.h>
28 #include <linux/mempolicy.h>
29 #include <linux/rmap.h>
30 #include <linux/mmu_notifier.h>
31 #include <linux/perf_event.h>
32 #include <linux/audit.h>
33 #include <linux/khugepaged.h>
34 #include <linux/uprobes.h>
35 #include <linux/rbtree_augmented.h>
36 #include <linux/sched/sysctl.h>
37 #include <linux/notifier.h>
38 #include <linux/memory.h>
39
40 #include <asm/uaccess.h>
41 #include <asm/cacheflush.h>
42 #include <asm/tlb.h>
43 #include <asm/mmu_context.h>
44
45 #include "internal.h"
46
47 #ifndef arch_mmap_check
48 #define arch_mmap_check(addr, len, flags)       (0)
49 #endif
50
51 #ifndef arch_rebalance_pgtables
52 #define arch_rebalance_pgtables(addr, len)              (addr)
53 #endif
54
55 static void unmap_region(struct mm_struct *mm,
56                 struct vm_area_struct *vma, struct vm_area_struct *prev,
57                 unsigned long start, unsigned long end);
58
59 /* description of effects of mapping type and prot in current implementation.
60  * this is due to the limited x86 page protection hardware.  The expected
61  * behavior is in parens:
62  *
63  * map_type     prot
64  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
65  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
66  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
67  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
68  *              
69  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
70  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
71  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
72  *
73  */
74 pgprot_t protection_map[16] = {
75         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
76         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
77 };
78
79 pgprot_t vm_get_page_prot(unsigned long vm_flags)
80 {
81         return __pgprot(pgprot_val(protection_map[vm_flags &
82                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
83                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
84 }
85 EXPORT_SYMBOL(vm_get_page_prot);
86
87 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
88 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
89 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
90 unsigned long sysctl_user_reserve_kbytes __read_mostly = 1UL << 17; /* 128MB */
91 unsigned long sysctl_admin_reserve_kbytes __read_mostly = 1UL << 13; /* 8MB */
92 /*
93  * Make sure vm_committed_as in one cacheline and not cacheline shared with
94  * other variables. It can be updated by several CPUs frequently.
95  */
96 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
97
98 /*
99  * The global memory commitment made in the system can be a metric
100  * that can be used to drive ballooning decisions when Linux is hosted
101  * as a guest. On Hyper-V, the host implements a policy engine for dynamically
102  * balancing memory across competing virtual machines that are hosted.
103  * Several metrics drive this policy engine including the guest reported
104  * memory commitment.
105  */
106 unsigned long vm_memory_committed(void)
107 {
108         return percpu_counter_read_positive(&vm_committed_as);
109 }
110 EXPORT_SYMBOL_GPL(vm_memory_committed);
111
112 /*
113  * Check that a process has enough memory to allocate a new virtual
114  * mapping. 0 means there is enough memory for the allocation to
115  * succeed and -ENOMEM implies there is not.
116  *
117  * We currently support three overcommit policies, which are set via the
118  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
119  *
120  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
121  * Additional code 2002 Jul 20 by Robert Love.
122  *
123  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
124  *
125  * Note this is a helper function intended to be used by LSMs which
126  * wish to use this logic.
127  */
128 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
129 {
130         unsigned long free, allowed, reserve;
131
132         vm_acct_memory(pages);
133
134         /*
135          * Sometimes we want to use more memory than we have
136          */
137         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
138                 return 0;
139
140         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
141                 free = global_page_state(NR_FREE_PAGES);
142                 free += global_page_state(NR_FILE_PAGES);
143
144                 /*
145                  * shmem pages shouldn't be counted as free in this
146                  * case, they can't be purged, only swapped out, and
147                  * that won't affect the overall amount of available
148                  * memory in the system.
149                  */
150                 free -= global_page_state(NR_SHMEM);
151
152                 free += get_nr_swap_pages();
153
154                 /*
155                  * Any slabs which are created with the
156                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
157                  * which are reclaimable, under pressure.  The dentry
158                  * cache and most inode caches should fall into this
159                  */
160                 free += global_page_state(NR_SLAB_RECLAIMABLE);
161
162                 /*
163                  * Leave reserved pages. The pages are not for anonymous pages.
164                  */
165                 if (free <= totalreserve_pages)
166                         goto error;
167                 else
168                         free -= totalreserve_pages;
169
170                 /*
171                  * Reserve some for root
172                  */
173                 if (!cap_sys_admin)
174                         free -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
175
176                 if (free > pages)
177                         return 0;
178
179                 goto error;
180         }
181
182         allowed = (totalram_pages - hugetlb_total_pages())
183                 * sysctl_overcommit_ratio / 100;
184         /*
185          * Reserve some for root
186          */
187         if (!cap_sys_admin)
188                 allowed -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
189         allowed += total_swap_pages;
190
191         /*
192          * Don't let a single process grow so big a user can't recover
193          */
194         if (mm) {
195                 reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10);
196                 allowed -= min(mm->total_vm / 32, reserve);
197         }
198
199         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
200                 return 0;
201 error:
202         vm_unacct_memory(pages);
203
204         return -ENOMEM;
205 }
206
207 /*
208  * Requires inode->i_mapping->i_mmap_mutex
209  */
210 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
211                 struct file *file, struct address_space *mapping)
212 {
213         if (vma->vm_flags & VM_DENYWRITE)
214                 atomic_inc(&file_inode(file)->i_writecount);
215         if (vma->vm_flags & VM_SHARED)
216                 mapping->i_mmap_writable--;
217
218         flush_dcache_mmap_lock(mapping);
219         if (unlikely(vma->vm_flags & VM_NONLINEAR))
220                 list_del_init(&vma->shared.nonlinear);
221         else
222                 vma_interval_tree_remove(vma, &mapping->i_mmap);
223         flush_dcache_mmap_unlock(mapping);
224 }
225
226 /*
227  * Unlink a file-based vm structure from its interval tree, to hide
228  * vma from rmap and vmtruncate before freeing its page tables.
229  */
230 void unlink_file_vma(struct vm_area_struct *vma)
231 {
232         struct file *file = vma->vm_file;
233
234         if (file) {
235                 struct address_space *mapping = file->f_mapping;
236                 mutex_lock(&mapping->i_mmap_mutex);
237                 __remove_shared_vm_struct(vma, file, mapping);
238                 mutex_unlock(&mapping->i_mmap_mutex);
239         }
240 }
241
242 /*
243  * Close a vm structure and free it, returning the next.
244  */
245 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
246 {
247         struct vm_area_struct *next = vma->vm_next;
248
249         might_sleep();
250         if (vma->vm_ops && vma->vm_ops->close)
251                 vma->vm_ops->close(vma);
252         if (vma->vm_file)
253                 fput(vma->vm_file);
254         mpol_put(vma_policy(vma));
255         kmem_cache_free(vm_area_cachep, vma);
256         return next;
257 }
258
259 static unsigned long do_brk(unsigned long addr, unsigned long len);
260
261 SYSCALL_DEFINE1(brk, unsigned long, brk)
262 {
263         unsigned long rlim, retval;
264         unsigned long newbrk, oldbrk;
265         struct mm_struct *mm = current->mm;
266         unsigned long min_brk;
267         bool populate;
268
269         down_write(&mm->mmap_sem);
270
271 #ifdef CONFIG_COMPAT_BRK
272         /*
273          * CONFIG_COMPAT_BRK can still be overridden by setting
274          * randomize_va_space to 2, which will still cause mm->start_brk
275          * to be arbitrarily shifted
276          */
277         if (current->brk_randomized)
278                 min_brk = mm->start_brk;
279         else
280                 min_brk = mm->end_data;
281 #else
282         min_brk = mm->start_brk;
283 #endif
284         if (brk < min_brk)
285                 goto out;
286
287         /*
288          * Check against rlimit here. If this check is done later after the test
289          * of oldbrk with newbrk then it can escape the test and let the data
290          * segment grow beyond its set limit the in case where the limit is
291          * not page aligned -Ram Gupta
292          */
293         rlim = rlimit(RLIMIT_DATA);
294         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
295                         (mm->end_data - mm->start_data) > rlim)
296                 goto out;
297
298         newbrk = PAGE_ALIGN(brk);
299         oldbrk = PAGE_ALIGN(mm->brk);
300         if (oldbrk == newbrk)
301                 goto set_brk;
302
303         /* Always allow shrinking brk. */
304         if (brk <= mm->brk) {
305                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
306                         goto set_brk;
307                 goto out;
308         }
309
310         /* Check against existing mmap mappings. */
311         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
312                 goto out;
313
314         /* Ok, looks good - let it rip. */
315         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
316                 goto out;
317
318 set_brk:
319         mm->brk = brk;
320         populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
321         up_write(&mm->mmap_sem);
322         if (populate)
323                 mm_populate(oldbrk, newbrk - oldbrk);
324         return brk;
325
326 out:
327         retval = mm->brk;
328         up_write(&mm->mmap_sem);
329         return retval;
330 }
331
332 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
333 {
334         unsigned long max, subtree_gap;
335         max = vma->vm_start;
336         if (vma->vm_prev)
337                 max -= vma->vm_prev->vm_end;
338         if (vma->vm_rb.rb_left) {
339                 subtree_gap = rb_entry(vma->vm_rb.rb_left,
340                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
341                 if (subtree_gap > max)
342                         max = subtree_gap;
343         }
344         if (vma->vm_rb.rb_right) {
345                 subtree_gap = rb_entry(vma->vm_rb.rb_right,
346                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
347                 if (subtree_gap > max)
348                         max = subtree_gap;
349         }
350         return max;
351 }
352
353 #ifdef CONFIG_DEBUG_VM_RB
354 static int browse_rb(struct rb_root *root)
355 {
356         int i = 0, j, bug = 0;
357         struct rb_node *nd, *pn = NULL;
358         unsigned long prev = 0, pend = 0;
359
360         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
361                 struct vm_area_struct *vma;
362                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
363                 if (vma->vm_start < prev) {
364                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev);
365                         bug = 1;
366                 }
367                 if (vma->vm_start < pend) {
368                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
369                         bug = 1;
370                 }
371                 if (vma->vm_start > vma->vm_end) {
372                         printk("vm_end %lx < vm_start %lx\n",
373                                 vma->vm_end, vma->vm_start);
374                         bug = 1;
375                 }
376                 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
377                         printk("free gap %lx, correct %lx\n",
378                                vma->rb_subtree_gap,
379                                vma_compute_subtree_gap(vma));
380                         bug = 1;
381                 }
382                 i++;
383                 pn = nd;
384                 prev = vma->vm_start;
385                 pend = vma->vm_end;
386         }
387         j = 0;
388         for (nd = pn; nd; nd = rb_prev(nd))
389                 j++;
390         if (i != j) {
391                 printk("backwards %d, forwards %d\n", j, i);
392                 bug = 1;
393         }
394         return bug ? -1 : i;
395 }
396
397 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
398 {
399         struct rb_node *nd;
400
401         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
402                 struct vm_area_struct *vma;
403                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
404                 BUG_ON(vma != ignore &&
405                        vma->rb_subtree_gap != vma_compute_subtree_gap(vma));
406         }
407 }
408
409 void validate_mm(struct mm_struct *mm)
410 {
411         int bug = 0;
412         int i = 0;
413         unsigned long highest_address = 0;
414         struct vm_area_struct *vma = mm->mmap;
415         while (vma) {
416                 struct anon_vma_chain *avc;
417                 vma_lock_anon_vma(vma);
418                 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
419                         anon_vma_interval_tree_verify(avc);
420                 vma_unlock_anon_vma(vma);
421                 highest_address = vma->vm_end;
422                 vma = vma->vm_next;
423                 i++;
424         }
425         if (i != mm->map_count) {
426                 printk("map_count %d vm_next %d\n", mm->map_count, i);
427                 bug = 1;
428         }
429         if (highest_address != mm->highest_vm_end) {
430                 printk("mm->highest_vm_end %lx, found %lx\n",
431                        mm->highest_vm_end, highest_address);
432                 bug = 1;
433         }
434         i = browse_rb(&mm->mm_rb);
435         if (i != mm->map_count) {
436                 printk("map_count %d rb %d\n", mm->map_count, i);
437                 bug = 1;
438         }
439         BUG_ON(bug);
440 }
441 #else
442 #define validate_mm_rb(root, ignore) do { } while (0)
443 #define validate_mm(mm) do { } while (0)
444 #endif
445
446 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
447                      unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
448
449 /*
450  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
451  * vma->vm_prev->vm_end values changed, without modifying the vma's position
452  * in the rbtree.
453  */
454 static void vma_gap_update(struct vm_area_struct *vma)
455 {
456         /*
457          * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
458          * function that does exacltly what we want.
459          */
460         vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
461 }
462
463 static inline void vma_rb_insert(struct vm_area_struct *vma,
464                                  struct rb_root *root)
465 {
466         /* All rb_subtree_gap values must be consistent prior to insertion */
467         validate_mm_rb(root, NULL);
468
469         rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
470 }
471
472 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
473 {
474         /*
475          * All rb_subtree_gap values must be consistent prior to erase,
476          * with the possible exception of the vma being erased.
477          */
478         validate_mm_rb(root, vma);
479
480         /*
481          * Note rb_erase_augmented is a fairly large inline function,
482          * so make sure we instantiate it only once with our desired
483          * augmented rbtree callbacks.
484          */
485         rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
486 }
487
488 /*
489  * vma has some anon_vma assigned, and is already inserted on that
490  * anon_vma's interval trees.
491  *
492  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
493  * vma must be removed from the anon_vma's interval trees using
494  * anon_vma_interval_tree_pre_update_vma().
495  *
496  * After the update, the vma will be reinserted using
497  * anon_vma_interval_tree_post_update_vma().
498  *
499  * The entire update must be protected by exclusive mmap_sem and by
500  * the root anon_vma's mutex.
501  */
502 static inline void
503 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
504 {
505         struct anon_vma_chain *avc;
506
507         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
508                 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
509 }
510
511 static inline void
512 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
513 {
514         struct anon_vma_chain *avc;
515
516         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
517                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
518 }
519
520 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
521                 unsigned long end, struct vm_area_struct **pprev,
522                 struct rb_node ***rb_link, struct rb_node **rb_parent)
523 {
524         struct rb_node **__rb_link, *__rb_parent, *rb_prev;
525
526         __rb_link = &mm->mm_rb.rb_node;
527         rb_prev = __rb_parent = NULL;
528
529         while (*__rb_link) {
530                 struct vm_area_struct *vma_tmp;
531
532                 __rb_parent = *__rb_link;
533                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
534
535                 if (vma_tmp->vm_end > addr) {
536                         /* Fail if an existing vma overlaps the area */
537                         if (vma_tmp->vm_start < end)
538                                 return -ENOMEM;
539                         __rb_link = &__rb_parent->rb_left;
540                 } else {
541                         rb_prev = __rb_parent;
542                         __rb_link = &__rb_parent->rb_right;
543                 }
544         }
545
546         *pprev = NULL;
547         if (rb_prev)
548                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
549         *rb_link = __rb_link;
550         *rb_parent = __rb_parent;
551         return 0;
552 }
553
554 static unsigned long count_vma_pages_range(struct mm_struct *mm,
555                 unsigned long addr, unsigned long end)
556 {
557         unsigned long nr_pages = 0;
558         struct vm_area_struct *vma;
559
560         /* Find first overlaping mapping */
561         vma = find_vma_intersection(mm, addr, end);
562         if (!vma)
563                 return 0;
564
565         nr_pages = (min(end, vma->vm_end) -
566                 max(addr, vma->vm_start)) >> PAGE_SHIFT;
567
568         /* Iterate over the rest of the overlaps */
569         for (vma = vma->vm_next; vma; vma = vma->vm_next) {
570                 unsigned long overlap_len;
571
572                 if (vma->vm_start > end)
573                         break;
574
575                 overlap_len = min(end, vma->vm_end) - vma->vm_start;
576                 nr_pages += overlap_len >> PAGE_SHIFT;
577         }
578
579         return nr_pages;
580 }
581
582 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
583                 struct rb_node **rb_link, struct rb_node *rb_parent)
584 {
585         /* Update tracking information for the gap following the new vma. */
586         if (vma->vm_next)
587                 vma_gap_update(vma->vm_next);
588         else
589                 mm->highest_vm_end = vma->vm_end;
590
591         /*
592          * vma->vm_prev wasn't known when we followed the rbtree to find the
593          * correct insertion point for that vma. As a result, we could not
594          * update the vma vm_rb parents rb_subtree_gap values on the way down.
595          * So, we first insert the vma with a zero rb_subtree_gap value
596          * (to be consistent with what we did on the way down), and then
597          * immediately update the gap to the correct value. Finally we
598          * rebalance the rbtree after all augmented values have been set.
599          */
600         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
601         vma->rb_subtree_gap = 0;
602         vma_gap_update(vma);
603         vma_rb_insert(vma, &mm->mm_rb);
604 }
605
606 static void __vma_link_file(struct vm_area_struct *vma)
607 {
608         struct file *file;
609
610         file = vma->vm_file;
611         if (file) {
612                 struct address_space *mapping = file->f_mapping;
613
614                 if (vma->vm_flags & VM_DENYWRITE)
615                         atomic_dec(&file_inode(file)->i_writecount);
616                 if (vma->vm_flags & VM_SHARED)
617                         mapping->i_mmap_writable++;
618
619                 flush_dcache_mmap_lock(mapping);
620                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
621                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
622                 else
623                         vma_interval_tree_insert(vma, &mapping->i_mmap);
624                 flush_dcache_mmap_unlock(mapping);
625         }
626 }
627
628 static void
629 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
630         struct vm_area_struct *prev, struct rb_node **rb_link,
631         struct rb_node *rb_parent)
632 {
633         __vma_link_list(mm, vma, prev, rb_parent);
634         __vma_link_rb(mm, vma, rb_link, rb_parent);
635 }
636
637 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
638                         struct vm_area_struct *prev, struct rb_node **rb_link,
639                         struct rb_node *rb_parent)
640 {
641         struct address_space *mapping = NULL;
642
643         if (vma->vm_file)
644                 mapping = vma->vm_file->f_mapping;
645
646         if (mapping)
647                 mutex_lock(&mapping->i_mmap_mutex);
648
649         __vma_link(mm, vma, prev, rb_link, rb_parent);
650         __vma_link_file(vma);
651
652         if (mapping)
653                 mutex_unlock(&mapping->i_mmap_mutex);
654
655         mm->map_count++;
656         validate_mm(mm);
657 }
658
659 /*
660  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
661  * mm's list and rbtree.  It has already been inserted into the interval tree.
662  */
663 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
664 {
665         struct vm_area_struct *prev;
666         struct rb_node **rb_link, *rb_parent;
667
668         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
669                            &prev, &rb_link, &rb_parent))
670                 BUG();
671         __vma_link(mm, vma, prev, rb_link, rb_parent);
672         mm->map_count++;
673 }
674
675 static inline void
676 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
677                 struct vm_area_struct *prev)
678 {
679         struct vm_area_struct *next;
680
681         vma_rb_erase(vma, &mm->mm_rb);
682         prev->vm_next = next = vma->vm_next;
683         if (next)
684                 next->vm_prev = prev;
685         if (mm->mmap_cache == vma)
686                 mm->mmap_cache = prev;
687 }
688
689 /*
690  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
691  * is already present in an i_mmap tree without adjusting the tree.
692  * The following helper function should be used when such adjustments
693  * are necessary.  The "insert" vma (if any) is to be inserted
694  * before we drop the necessary locks.
695  */
696 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
697         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
698 {
699         struct mm_struct *mm = vma->vm_mm;
700         struct vm_area_struct *next = vma->vm_next;
701         struct vm_area_struct *importer = NULL;
702         struct address_space *mapping = NULL;
703         struct rb_root *root = NULL;
704         struct anon_vma *anon_vma = NULL;
705         struct file *file = vma->vm_file;
706         bool start_changed = false, end_changed = false;
707         long adjust_next = 0;
708         int remove_next = 0;
709
710         if (next && !insert) {
711                 struct vm_area_struct *exporter = NULL;
712
713                 if (end >= next->vm_end) {
714                         /*
715                          * vma expands, overlapping all the next, and
716                          * perhaps the one after too (mprotect case 6).
717                          */
718 again:                  remove_next = 1 + (end > next->vm_end);
719                         end = next->vm_end;
720                         exporter = next;
721                         importer = vma;
722                 } else if (end > next->vm_start) {
723                         /*
724                          * vma expands, overlapping part of the next:
725                          * mprotect case 5 shifting the boundary up.
726                          */
727                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
728                         exporter = next;
729                         importer = vma;
730                 } else if (end < vma->vm_end) {
731                         /*
732                          * vma shrinks, and !insert tells it's not
733                          * split_vma inserting another: so it must be
734                          * mprotect case 4 shifting the boundary down.
735                          */
736                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
737                         exporter = vma;
738                         importer = next;
739                 }
740
741                 /*
742                  * Easily overlooked: when mprotect shifts the boundary,
743                  * make sure the expanding vma has anon_vma set if the
744                  * shrinking vma had, to cover any anon pages imported.
745                  */
746                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
747                         if (anon_vma_clone(importer, exporter))
748                                 return -ENOMEM;
749                         importer->anon_vma = exporter->anon_vma;
750                 }
751         }
752
753         if (file) {
754                 mapping = file->f_mapping;
755                 if (!(vma->vm_flags & VM_NONLINEAR)) {
756                         root = &mapping->i_mmap;
757                         uprobe_munmap(vma, vma->vm_start, vma->vm_end);
758
759                         if (adjust_next)
760                                 uprobe_munmap(next, next->vm_start,
761                                                         next->vm_end);
762                 }
763
764                 mutex_lock(&mapping->i_mmap_mutex);
765                 if (insert) {
766                         /*
767                          * Put into interval tree now, so instantiated pages
768                          * are visible to arm/parisc __flush_dcache_page
769                          * throughout; but we cannot insert into address
770                          * space until vma start or end is updated.
771                          */
772                         __vma_link_file(insert);
773                 }
774         }
775
776         vma_adjust_trans_huge(vma, start, end, adjust_next);
777
778         anon_vma = vma->anon_vma;
779         if (!anon_vma && adjust_next)
780                 anon_vma = next->anon_vma;
781         if (anon_vma) {
782                 VM_BUG_ON(adjust_next && next->anon_vma &&
783                           anon_vma != next->anon_vma);
784                 anon_vma_lock_write(anon_vma);
785                 anon_vma_interval_tree_pre_update_vma(vma);
786                 if (adjust_next)
787                         anon_vma_interval_tree_pre_update_vma(next);
788         }
789
790         if (root) {
791                 flush_dcache_mmap_lock(mapping);
792                 vma_interval_tree_remove(vma, root);
793                 if (adjust_next)
794                         vma_interval_tree_remove(next, root);
795         }
796
797         if (start != vma->vm_start) {
798                 vma->vm_start = start;
799                 start_changed = true;
800         }
801         if (end != vma->vm_end) {
802                 vma->vm_end = end;
803                 end_changed = true;
804         }
805         vma->vm_pgoff = pgoff;
806         if (adjust_next) {
807                 next->vm_start += adjust_next << PAGE_SHIFT;
808                 next->vm_pgoff += adjust_next;
809         }
810
811         if (root) {
812                 if (adjust_next)
813                         vma_interval_tree_insert(next, root);
814                 vma_interval_tree_insert(vma, root);
815                 flush_dcache_mmap_unlock(mapping);
816         }
817
818         if (remove_next) {
819                 /*
820                  * vma_merge has merged next into vma, and needs
821                  * us to remove next before dropping the locks.
822                  */
823                 __vma_unlink(mm, next, vma);
824                 if (file)
825                         __remove_shared_vm_struct(next, file, mapping);
826         } else if (insert) {
827                 /*
828                  * split_vma has split insert from vma, and needs
829                  * us to insert it before dropping the locks
830                  * (it may either follow vma or precede it).
831                  */
832                 __insert_vm_struct(mm, insert);
833         } else {
834                 if (start_changed)
835                         vma_gap_update(vma);
836                 if (end_changed) {
837                         if (!next)
838                                 mm->highest_vm_end = end;
839                         else if (!adjust_next)
840                                 vma_gap_update(next);
841                 }
842         }
843
844         if (anon_vma) {
845                 anon_vma_interval_tree_post_update_vma(vma);
846                 if (adjust_next)
847                         anon_vma_interval_tree_post_update_vma(next);
848                 anon_vma_unlock_write(anon_vma);
849         }
850         if (mapping)
851                 mutex_unlock(&mapping->i_mmap_mutex);
852
853         if (root) {
854                 uprobe_mmap(vma);
855
856                 if (adjust_next)
857                         uprobe_mmap(next);
858         }
859
860         if (remove_next) {
861                 if (file) {
862                         uprobe_munmap(next, next->vm_start, next->vm_end);
863                         fput(file);
864                 }
865                 if (next->anon_vma)
866                         anon_vma_merge(vma, next);
867                 mm->map_count--;
868                 vma_set_policy(vma, vma_policy(next));
869                 kmem_cache_free(vm_area_cachep, next);
870                 /*
871                  * In mprotect's case 6 (see comments on vma_merge),
872                  * we must remove another next too. It would clutter
873                  * up the code too much to do both in one go.
874                  */
875                 next = vma->vm_next;
876                 if (remove_next == 2)
877                         goto again;
878                 else if (next)
879                         vma_gap_update(next);
880                 else
881                         mm->highest_vm_end = end;
882         }
883         if (insert && file)
884                 uprobe_mmap(insert);
885
886         validate_mm(mm);
887
888         return 0;
889 }
890
891 /*
892  * If the vma has a ->close operation then the driver probably needs to release
893  * per-vma resources, so we don't attempt to merge those.
894  */
895 static inline int is_mergeable_vma(struct vm_area_struct *vma,
896                         struct file *file, unsigned long vm_flags)
897 {
898         if (vma->vm_flags ^ vm_flags)
899                 return 0;
900         if (vma->vm_file != file)
901                 return 0;
902         if (vma->vm_ops && vma->vm_ops->close)
903                 return 0;
904         return 1;
905 }
906
907 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
908                                         struct anon_vma *anon_vma2,
909                                         struct vm_area_struct *vma)
910 {
911         /*
912          * The list_is_singular() test is to avoid merging VMA cloned from
913          * parents. This can improve scalability caused by anon_vma lock.
914          */
915         if ((!anon_vma1 || !anon_vma2) && (!vma ||
916                 list_is_singular(&vma->anon_vma_chain)))
917                 return 1;
918         return anon_vma1 == anon_vma2;
919 }
920
921 /*
922  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
923  * in front of (at a lower virtual address and file offset than) the vma.
924  *
925  * We cannot merge two vmas if they have differently assigned (non-NULL)
926  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
927  *
928  * We don't check here for the merged mmap wrapping around the end of pagecache
929  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
930  * wrap, nor mmaps which cover the final page at index -1UL.
931  */
932 static int
933 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
934         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
935 {
936         if (is_mergeable_vma(vma, file, vm_flags) &&
937             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
938                 if (vma->vm_pgoff == vm_pgoff)
939                         return 1;
940         }
941         return 0;
942 }
943
944 /*
945  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
946  * beyond (at a higher virtual address and file offset than) the vma.
947  *
948  * We cannot merge two vmas if they have differently assigned (non-NULL)
949  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
950  */
951 static int
952 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
953         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
954 {
955         if (is_mergeable_vma(vma, file, vm_flags) &&
956             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
957                 pgoff_t vm_pglen;
958                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
959                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
960                         return 1;
961         }
962         return 0;
963 }
964
965 /*
966  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
967  * whether that can be merged with its predecessor or its successor.
968  * Or both (it neatly fills a hole).
969  *
970  * In most cases - when called for mmap, brk or mremap - [addr,end) is
971  * certain not to be mapped by the time vma_merge is called; but when
972  * called for mprotect, it is certain to be already mapped (either at
973  * an offset within prev, or at the start of next), and the flags of
974  * this area are about to be changed to vm_flags - and the no-change
975  * case has already been eliminated.
976  *
977  * The following mprotect cases have to be considered, where AAAA is
978  * the area passed down from mprotect_fixup, never extending beyond one
979  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
980  *
981  *     AAAA             AAAA                AAAA          AAAA
982  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
983  *    cannot merge    might become    might become    might become
984  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
985  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
986  *    mremap move:                                    PPPPNNNNNNNN 8
987  *        AAAA
988  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
989  *    might become    case 1 below    case 2 below    case 3 below
990  *
991  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
992  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
993  */
994 struct vm_area_struct *vma_merge(struct mm_struct *mm,
995                         struct vm_area_struct *prev, unsigned long addr,
996                         unsigned long end, unsigned long vm_flags,
997                         struct anon_vma *anon_vma, struct file *file,
998                         pgoff_t pgoff, struct mempolicy *policy)
999 {
1000         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1001         struct vm_area_struct *area, *next;
1002         int err;
1003
1004         /*
1005          * We later require that vma->vm_flags == vm_flags,
1006          * so this tests vma->vm_flags & VM_SPECIAL, too.
1007          */
1008         if (vm_flags & VM_SPECIAL)
1009                 return NULL;
1010
1011         if (prev)
1012                 next = prev->vm_next;
1013         else
1014                 next = mm->mmap;
1015         area = next;
1016         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
1017                 next = next->vm_next;
1018
1019         /*
1020          * Can it merge with the predecessor?
1021          */
1022         if (prev && prev->vm_end == addr &&
1023                         mpol_equal(vma_policy(prev), policy) &&
1024                         can_vma_merge_after(prev, vm_flags,
1025                                                 anon_vma, file, pgoff)) {
1026                 /*
1027                  * OK, it can.  Can we now merge in the successor as well?
1028                  */
1029                 if (next && end == next->vm_start &&
1030                                 mpol_equal(policy, vma_policy(next)) &&
1031                                 can_vma_merge_before(next, vm_flags,
1032                                         anon_vma, file, pgoff+pglen) &&
1033                                 is_mergeable_anon_vma(prev->anon_vma,
1034                                                       next->anon_vma, NULL)) {
1035                                                         /* cases 1, 6 */
1036                         err = vma_adjust(prev, prev->vm_start,
1037                                 next->vm_end, prev->vm_pgoff, NULL);
1038                 } else                                  /* cases 2, 5, 7 */
1039                         err = vma_adjust(prev, prev->vm_start,
1040                                 end, prev->vm_pgoff, NULL);
1041                 if (err)
1042                         return NULL;
1043                 khugepaged_enter_vma_merge(prev);
1044                 return prev;
1045         }
1046
1047         /*
1048          * Can this new request be merged in front of next?
1049          */
1050         if (next && end == next->vm_start &&
1051                         mpol_equal(policy, vma_policy(next)) &&
1052                         can_vma_merge_before(next, vm_flags,
1053                                         anon_vma, file, pgoff+pglen)) {
1054                 if (prev && addr < prev->vm_end)        /* case 4 */
1055                         err = vma_adjust(prev, prev->vm_start,
1056                                 addr, prev->vm_pgoff, NULL);
1057                 else                                    /* cases 3, 8 */
1058                         err = vma_adjust(area, addr, next->vm_end,
1059                                 next->vm_pgoff - pglen, NULL);
1060                 if (err)
1061                         return NULL;
1062                 khugepaged_enter_vma_merge(area);
1063                 return area;
1064         }
1065
1066         return NULL;
1067 }
1068
1069 /*
1070  * Rough compatbility check to quickly see if it's even worth looking
1071  * at sharing an anon_vma.
1072  *
1073  * They need to have the same vm_file, and the flags can only differ
1074  * in things that mprotect may change.
1075  *
1076  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1077  * we can merge the two vma's. For example, we refuse to merge a vma if
1078  * there is a vm_ops->close() function, because that indicates that the
1079  * driver is doing some kind of reference counting. But that doesn't
1080  * really matter for the anon_vma sharing case.
1081  */
1082 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1083 {
1084         return a->vm_end == b->vm_start &&
1085                 mpol_equal(vma_policy(a), vma_policy(b)) &&
1086                 a->vm_file == b->vm_file &&
1087                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
1088                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1089 }
1090
1091 /*
1092  * Do some basic sanity checking to see if we can re-use the anon_vma
1093  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1094  * the same as 'old', the other will be the new one that is trying
1095  * to share the anon_vma.
1096  *
1097  * NOTE! This runs with mm_sem held for reading, so it is possible that
1098  * the anon_vma of 'old' is concurrently in the process of being set up
1099  * by another page fault trying to merge _that_. But that's ok: if it
1100  * is being set up, that automatically means that it will be a singleton
1101  * acceptable for merging, so we can do all of this optimistically. But
1102  * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
1103  *
1104  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1105  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1106  * is to return an anon_vma that is "complex" due to having gone through
1107  * a fork).
1108  *
1109  * We also make sure that the two vma's are compatible (adjacent,
1110  * and with the same memory policies). That's all stable, even with just
1111  * a read lock on the mm_sem.
1112  */
1113 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1114 {
1115         if (anon_vma_compatible(a, b)) {
1116                 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
1117
1118                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1119                         return anon_vma;
1120         }
1121         return NULL;
1122 }
1123
1124 /*
1125  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1126  * neighbouring vmas for a suitable anon_vma, before it goes off
1127  * to allocate a new anon_vma.  It checks because a repetitive
1128  * sequence of mprotects and faults may otherwise lead to distinct
1129  * anon_vmas being allocated, preventing vma merge in subsequent
1130  * mprotect.
1131  */
1132 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1133 {
1134         struct anon_vma *anon_vma;
1135         struct vm_area_struct *near;
1136
1137         near = vma->vm_next;
1138         if (!near)
1139                 goto try_prev;
1140
1141         anon_vma = reusable_anon_vma(near, vma, near);
1142         if (anon_vma)
1143                 return anon_vma;
1144 try_prev:
1145         near = vma->vm_prev;
1146         if (!near)
1147                 goto none;
1148
1149         anon_vma = reusable_anon_vma(near, near, vma);
1150         if (anon_vma)
1151                 return anon_vma;
1152 none:
1153         /*
1154          * There's no absolute need to look only at touching neighbours:
1155          * we could search further afield for "compatible" anon_vmas.
1156          * But it would probably just be a waste of time searching,
1157          * or lead to too many vmas hanging off the same anon_vma.
1158          * We're trying to allow mprotect remerging later on,
1159          * not trying to minimize memory used for anon_vmas.
1160          */
1161         return NULL;
1162 }
1163
1164 #ifdef CONFIG_PROC_FS
1165 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
1166                                                 struct file *file, long pages)
1167 {
1168         const unsigned long stack_flags
1169                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
1170
1171         mm->total_vm += pages;
1172
1173         if (file) {
1174                 mm->shared_vm += pages;
1175                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
1176                         mm->exec_vm += pages;
1177         } else if (flags & stack_flags)
1178                 mm->stack_vm += pages;
1179 }
1180 #endif /* CONFIG_PROC_FS */
1181
1182 /*
1183  * If a hint addr is less than mmap_min_addr change hint to be as
1184  * low as possible but still greater than mmap_min_addr
1185  */
1186 static inline unsigned long round_hint_to_min(unsigned long hint)
1187 {
1188         hint &= PAGE_MASK;
1189         if (((void *)hint != NULL) &&
1190             (hint < mmap_min_addr))
1191                 return PAGE_ALIGN(mmap_min_addr);
1192         return hint;
1193 }
1194
1195 /*
1196  * The caller must hold down_write(&current->mm->mmap_sem).
1197  */
1198
1199 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
1200                         unsigned long len, unsigned long prot,
1201                         unsigned long flags, unsigned long pgoff,
1202                         unsigned long *populate)
1203 {
1204         struct mm_struct * mm = current->mm;
1205         struct inode *inode;
1206         vm_flags_t vm_flags;
1207
1208         *populate = 0;
1209
1210         /*
1211          * Does the application expect PROT_READ to imply PROT_EXEC?
1212          *
1213          * (the exception is when the underlying filesystem is noexec
1214          *  mounted, in which case we dont add PROT_EXEC.)
1215          */
1216         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1217                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
1218                         prot |= PROT_EXEC;
1219
1220         if (!len)
1221                 return -EINVAL;
1222
1223         if (!(flags & MAP_FIXED))
1224                 addr = round_hint_to_min(addr);
1225
1226         /* Careful about overflows.. */
1227         len = PAGE_ALIGN(len);
1228         if (!len)
1229                 return -ENOMEM;
1230
1231         /* offset overflow? */
1232         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1233                return -EOVERFLOW;
1234
1235         /* Too many mappings? */
1236         if (mm->map_count > sysctl_max_map_count)
1237                 return -ENOMEM;
1238
1239         /* Obtain the address to map to. we verify (or select) it and ensure
1240          * that it represents a valid section of the address space.
1241          */
1242         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1243         if (addr & ~PAGE_MASK)
1244                 return addr;
1245
1246         /* Do simple checking here so the lower-level routines won't have
1247          * to. we assume access permissions have been handled by the open
1248          * of the memory object, so we don't do any here.
1249          */
1250         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1251                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1252
1253         if (flags & MAP_LOCKED)
1254                 if (!can_do_mlock())
1255                         return -EPERM;
1256
1257         /* mlock MCL_FUTURE? */
1258         if (vm_flags & VM_LOCKED) {
1259                 unsigned long locked, lock_limit;
1260                 locked = len >> PAGE_SHIFT;
1261                 locked += mm->locked_vm;
1262                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1263                 lock_limit >>= PAGE_SHIFT;
1264                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1265                         return -EAGAIN;
1266         }
1267
1268         inode = file ? file_inode(file) : NULL;
1269
1270         if (file) {
1271                 switch (flags & MAP_TYPE) {
1272                 case MAP_SHARED:
1273                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1274                                 return -EACCES;
1275
1276                         /*
1277                          * Make sure we don't allow writing to an append-only
1278                          * file..
1279                          */
1280                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1281                                 return -EACCES;
1282
1283                         /*
1284                          * Make sure there are no mandatory locks on the file.
1285                          */
1286                         if (locks_verify_locked(inode))
1287                                 return -EAGAIN;
1288
1289                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1290                         if (!(file->f_mode & FMODE_WRITE))
1291                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1292
1293                         /* fall through */
1294                 case MAP_PRIVATE:
1295                         if (!(file->f_mode & FMODE_READ))
1296                                 return -EACCES;
1297                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1298                                 if (vm_flags & VM_EXEC)
1299                                         return -EPERM;
1300                                 vm_flags &= ~VM_MAYEXEC;
1301                         }
1302
1303                         if (!file->f_op || !file->f_op->mmap)
1304                                 return -ENODEV;
1305                         break;
1306
1307                 default:
1308                         return -EINVAL;
1309                 }
1310         } else {
1311                 switch (flags & MAP_TYPE) {
1312                 case MAP_SHARED:
1313                         /*
1314                          * Ignore pgoff.
1315                          */
1316                         pgoff = 0;
1317                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1318                         break;
1319                 case MAP_PRIVATE:
1320                         /*
1321                          * Set pgoff according to addr for anon_vma.
1322                          */
1323                         pgoff = addr >> PAGE_SHIFT;
1324                         break;
1325                 default:
1326                         return -EINVAL;
1327                 }
1328         }
1329
1330         /*
1331          * Set 'VM_NORESERVE' if we should not account for the
1332          * memory use of this mapping.
1333          */
1334         if (flags & MAP_NORESERVE) {
1335                 /* We honor MAP_NORESERVE if allowed to overcommit */
1336                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1337                         vm_flags |= VM_NORESERVE;
1338
1339                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1340                 if (file && is_file_hugepages(file))
1341                         vm_flags |= VM_NORESERVE;
1342         }
1343
1344         addr = mmap_region(file, addr, len, vm_flags, pgoff);
1345         if (!IS_ERR_VALUE(addr) &&
1346             ((vm_flags & VM_LOCKED) ||
1347              (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1348                 *populate = len;
1349         return addr;
1350 }
1351
1352 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1353                 unsigned long, prot, unsigned long, flags,
1354                 unsigned long, fd, unsigned long, pgoff)
1355 {
1356         struct file *file = NULL;
1357         unsigned long retval = -EBADF;
1358
1359         if (!(flags & MAP_ANONYMOUS)) {
1360                 audit_mmap_fd(fd, flags);
1361                 if (unlikely(flags & MAP_HUGETLB))
1362                         return -EINVAL;
1363                 file = fget(fd);
1364                 if (!file)
1365                         goto out;
1366                 if (is_file_hugepages(file))
1367                         len = ALIGN(len, huge_page_size(hstate_file(file)));
1368         } else if (flags & MAP_HUGETLB) {
1369                 struct user_struct *user = NULL;
1370                 struct hstate *hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) &
1371                                                    SHM_HUGE_MASK);
1372
1373                 if (!hs)
1374                         return -EINVAL;
1375
1376                 len = ALIGN(len, huge_page_size(hs));
1377                 /*
1378                  * VM_NORESERVE is used because the reservations will be
1379                  * taken when vm_ops->mmap() is called
1380                  * A dummy user value is used because we are not locking
1381                  * memory so no accounting is necessary
1382                  */
1383                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1384                                 VM_NORESERVE,
1385                                 &user, HUGETLB_ANONHUGE_INODE,
1386                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1387                 if (IS_ERR(file))
1388                         return PTR_ERR(file);
1389         }
1390
1391         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1392
1393         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1394         if (file)
1395                 fput(file);
1396 out:
1397         return retval;
1398 }
1399
1400 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1401 struct mmap_arg_struct {
1402         unsigned long addr;
1403         unsigned long len;
1404         unsigned long prot;
1405         unsigned long flags;
1406         unsigned long fd;
1407         unsigned long offset;
1408 };
1409
1410 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1411 {
1412         struct mmap_arg_struct a;
1413
1414         if (copy_from_user(&a, arg, sizeof(a)))
1415                 return -EFAULT;
1416         if (a.offset & ~PAGE_MASK)
1417                 return -EINVAL;
1418
1419         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1420                               a.offset >> PAGE_SHIFT);
1421 }
1422 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1423
1424 /*
1425  * Some shared mappigns will want the pages marked read-only
1426  * to track write events. If so, we'll downgrade vm_page_prot
1427  * to the private version (using protection_map[] without the
1428  * VM_SHARED bit).
1429  */
1430 int vma_wants_writenotify(struct vm_area_struct *vma)
1431 {
1432         vm_flags_t vm_flags = vma->vm_flags;
1433
1434         /* If it was private or non-writable, the write bit is already clear */
1435         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1436                 return 0;
1437
1438         /* The backer wishes to know when pages are first written to? */
1439         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1440                 return 1;
1441
1442         /* The open routine did something to the protections already? */
1443         if (pgprot_val(vma->vm_page_prot) !=
1444             pgprot_val(vm_get_page_prot(vm_flags)))
1445                 return 0;
1446
1447         /* Specialty mapping? */
1448         if (vm_flags & VM_PFNMAP)
1449                 return 0;
1450
1451         /* Can the mapping track the dirty pages? */
1452         return vma->vm_file && vma->vm_file->f_mapping &&
1453                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1454 }
1455
1456 /*
1457  * We account for memory if it's a private writeable mapping,
1458  * not hugepages and VM_NORESERVE wasn't set.
1459  */
1460 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1461 {
1462         /*
1463          * hugetlb has its own accounting separate from the core VM
1464          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1465          */
1466         if (file && is_file_hugepages(file))
1467                 return 0;
1468
1469         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1470 }
1471
1472 unsigned long mmap_region(struct file *file, unsigned long addr,
1473                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1474 {
1475         struct mm_struct *mm = current->mm;
1476         struct vm_area_struct *vma, *prev;
1477         int correct_wcount = 0;
1478         int error;
1479         struct rb_node **rb_link, *rb_parent;
1480         unsigned long charged = 0;
1481         struct inode *inode =  file ? file_inode(file) : NULL;
1482
1483         /* Check against address space limit. */
1484         if (!may_expand_vm(mm, len >> PAGE_SHIFT)) {
1485                 unsigned long nr_pages;
1486
1487                 /*
1488                  * MAP_FIXED may remove pages of mappings that intersects with
1489                  * requested mapping. Account for the pages it would unmap.
1490                  */
1491                 if (!(vm_flags & MAP_FIXED))
1492                         return -ENOMEM;
1493
1494                 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1495
1496                 if (!may_expand_vm(mm, (len >> PAGE_SHIFT) - nr_pages))
1497                         return -ENOMEM;
1498         }
1499
1500         /* Clear old maps */
1501         error = -ENOMEM;
1502 munmap_back:
1503         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
1504                 if (do_munmap(mm, addr, len))
1505                         return -ENOMEM;
1506                 goto munmap_back;
1507         }
1508
1509         /*
1510          * Private writable mapping: check memory availability
1511          */
1512         if (accountable_mapping(file, vm_flags)) {
1513                 charged = len >> PAGE_SHIFT;
1514                 if (security_vm_enough_memory_mm(mm, charged))
1515                         return -ENOMEM;
1516                 vm_flags |= VM_ACCOUNT;
1517         }
1518
1519         /*
1520          * Can we just expand an old mapping?
1521          */
1522         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1523         if (vma)
1524                 goto out;
1525
1526         /*
1527          * Determine the object being mapped and call the appropriate
1528          * specific mapper. the address has already been validated, but
1529          * not unmapped, but the maps are removed from the list.
1530          */
1531         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1532         if (!vma) {
1533                 error = -ENOMEM;
1534                 goto unacct_error;
1535         }
1536
1537         vma->vm_mm = mm;
1538         vma->vm_start = addr;
1539         vma->vm_end = addr + len;
1540         vma->vm_flags = vm_flags;
1541         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1542         vma->vm_pgoff = pgoff;
1543         INIT_LIST_HEAD(&vma->anon_vma_chain);
1544
1545         error = -EINVAL;        /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1546
1547         if (file) {
1548                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1549                         goto free_vma;
1550                 if (vm_flags & VM_DENYWRITE) {
1551                         error = deny_write_access(file);
1552                         if (error)
1553                                 goto free_vma;
1554                         correct_wcount = 1;
1555                 }
1556                 vma->vm_file = get_file(file);
1557                 error = file->f_op->mmap(file, vma);
1558                 if (error)
1559                         goto unmap_and_free_vma;
1560
1561                 /* Can addr have changed??
1562                  *
1563                  * Answer: Yes, several device drivers can do it in their
1564                  *         f_op->mmap method. -DaveM
1565                  * Bug: If addr is changed, prev, rb_link, rb_parent should
1566                  *      be updated for vma_link()
1567                  */
1568                 WARN_ON_ONCE(addr != vma->vm_start);
1569
1570                 addr = vma->vm_start;
1571                 pgoff = vma->vm_pgoff;
1572                 vm_flags = vma->vm_flags;
1573         } else if (vm_flags & VM_SHARED) {
1574                 if (unlikely(vm_flags & (VM_GROWSDOWN|VM_GROWSUP)))
1575                         goto free_vma;
1576                 error = shmem_zero_setup(vma);
1577                 if (error)
1578                         goto free_vma;
1579         }
1580
1581         if (vma_wants_writenotify(vma)) {
1582                 pgprot_t pprot = vma->vm_page_prot;
1583
1584                 /* Can vma->vm_page_prot have changed??
1585                  *
1586                  * Answer: Yes, drivers may have changed it in their
1587                  *         f_op->mmap method.
1588                  *
1589                  * Ensures that vmas marked as uncached stay that way.
1590                  */
1591                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1592                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1593                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1594         }
1595
1596         vma_link(mm, vma, prev, rb_link, rb_parent);
1597         file = vma->vm_file;
1598
1599         /* Once vma denies write, undo our temporary denial count */
1600         if (correct_wcount)
1601                 atomic_inc(&inode->i_writecount);
1602 out:
1603         perf_event_mmap(vma);
1604
1605         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1606         if (vm_flags & VM_LOCKED) {
1607                 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1608                                         vma == get_gate_vma(current->mm)))
1609                         mm->locked_vm += (len >> PAGE_SHIFT);
1610                 else
1611                         vma->vm_flags &= ~VM_LOCKED;
1612         }
1613
1614         if (file)
1615                 uprobe_mmap(vma);
1616
1617         return addr;
1618
1619 unmap_and_free_vma:
1620         if (correct_wcount)
1621                 atomic_inc(&inode->i_writecount);
1622         vma->vm_file = NULL;
1623         fput(file);
1624
1625         /* Undo any partial mapping done by a device driver. */
1626         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1627         charged = 0;
1628 free_vma:
1629         kmem_cache_free(vm_area_cachep, vma);
1630 unacct_error:
1631         if (charged)
1632                 vm_unacct_memory(charged);
1633         return error;
1634 }
1635
1636 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1637 {
1638         /*
1639          * We implement the search by looking for an rbtree node that
1640          * immediately follows a suitable gap. That is,
1641          * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1642          * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1643          * - gap_end - gap_start >= length
1644          */
1645
1646         struct mm_struct *mm = current->mm;
1647         struct vm_area_struct *vma;
1648         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1649
1650         /* Adjust search length to account for worst case alignment overhead */
1651         length = info->length + info->align_mask;
1652         if (length < info->length)
1653                 return -ENOMEM;
1654
1655         /* Adjust search limits by the desired length */
1656         if (info->high_limit < length)
1657                 return -ENOMEM;
1658         high_limit = info->high_limit - length;
1659
1660         if (info->low_limit > high_limit)
1661                 return -ENOMEM;
1662         low_limit = info->low_limit + length;
1663
1664         /* Check if rbtree root looks promising */
1665         if (RB_EMPTY_ROOT(&mm->mm_rb))
1666                 goto check_highest;
1667         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1668         if (vma->rb_subtree_gap < length)
1669                 goto check_highest;
1670
1671         while (true) {
1672                 /* Visit left subtree if it looks promising */
1673                 gap_end = vma->vm_start;
1674                 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1675                         struct vm_area_struct *left =
1676                                 rb_entry(vma->vm_rb.rb_left,
1677                                          struct vm_area_struct, vm_rb);
1678                         if (left->rb_subtree_gap >= length) {
1679                                 vma = left;
1680                                 continue;
1681                         }
1682                 }
1683
1684                 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1685 check_current:
1686                 /* Check if current node has a suitable gap */
1687                 if (gap_start > high_limit)
1688                         return -ENOMEM;
1689                 if (gap_end >= low_limit && gap_end - gap_start >= length)
1690                         goto found;
1691
1692                 /* Visit right subtree if it looks promising */
1693                 if (vma->vm_rb.rb_right) {
1694                         struct vm_area_struct *right =
1695                                 rb_entry(vma->vm_rb.rb_right,
1696                                          struct vm_area_struct, vm_rb);
1697                         if (right->rb_subtree_gap >= length) {
1698                                 vma = right;
1699                                 continue;
1700                         }
1701                 }
1702
1703                 /* Go back up the rbtree to find next candidate node */
1704                 while (true) {
1705                         struct rb_node *prev = &vma->vm_rb;
1706                         if (!rb_parent(prev))
1707                                 goto check_highest;
1708                         vma = rb_entry(rb_parent(prev),
1709                                        struct vm_area_struct, vm_rb);
1710                         if (prev == vma->vm_rb.rb_left) {
1711                                 gap_start = vma->vm_prev->vm_end;
1712                                 gap_end = vma->vm_start;
1713                                 goto check_current;
1714                         }
1715                 }
1716         }
1717
1718 check_highest:
1719         /* Check highest gap, which does not precede any rbtree node */
1720         gap_start = mm->highest_vm_end;
1721         gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1722         if (gap_start > high_limit)
1723                 return -ENOMEM;
1724
1725 found:
1726         /* We found a suitable gap. Clip it with the original low_limit. */
1727         if (gap_start < info->low_limit)
1728                 gap_start = info->low_limit;
1729
1730         /* Adjust gap address to the desired alignment */
1731         gap_start += (info->align_offset - gap_start) & info->align_mask;
1732
1733         VM_BUG_ON(gap_start + info->length > info->high_limit);
1734         VM_BUG_ON(gap_start + info->length > gap_end);
1735         return gap_start;
1736 }
1737
1738 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1739 {
1740         struct mm_struct *mm = current->mm;
1741         struct vm_area_struct *vma;
1742         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1743
1744         /* Adjust search length to account for worst case alignment overhead */
1745         length = info->length + info->align_mask;
1746         if (length < info->length)
1747                 return -ENOMEM;
1748
1749         /*
1750          * Adjust search limits by the desired length.
1751          * See implementation comment at top of unmapped_area().
1752          */
1753         gap_end = info->high_limit;
1754         if (gap_end < length)
1755                 return -ENOMEM;
1756         high_limit = gap_end - length;
1757
1758         if (info->low_limit > high_limit)
1759                 return -ENOMEM;
1760         low_limit = info->low_limit + length;
1761
1762         /* Check highest gap, which does not precede any rbtree node */
1763         gap_start = mm->highest_vm_end;
1764         if (gap_start <= high_limit)
1765                 goto found_highest;
1766
1767         /* Check if rbtree root looks promising */
1768         if (RB_EMPTY_ROOT(&mm->mm_rb))
1769                 return -ENOMEM;
1770         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1771         if (vma->rb_subtree_gap < length)
1772                 return -ENOMEM;
1773
1774         while (true) {
1775                 /* Visit right subtree if it looks promising */
1776                 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1777                 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1778                         struct vm_area_struct *right =
1779                                 rb_entry(vma->vm_rb.rb_right,
1780                                          struct vm_area_struct, vm_rb);
1781                         if (right->rb_subtree_gap >= length) {
1782                                 vma = right;
1783                                 continue;
1784                         }
1785                 }
1786
1787 check_current:
1788                 /* Check if current node has a suitable gap */
1789                 gap_end = vma->vm_start;
1790                 if (gap_end < low_limit)
1791                         return -ENOMEM;
1792                 if (gap_start <= high_limit && gap_end - gap_start >= length)
1793                         goto found;
1794
1795                 /* Visit left subtree if it looks promising */
1796                 if (vma->vm_rb.rb_left) {
1797                         struct vm_area_struct *left =
1798                                 rb_entry(vma->vm_rb.rb_left,
1799                                          struct vm_area_struct, vm_rb);
1800                         if (left->rb_subtree_gap >= length) {
1801                                 vma = left;
1802                                 continue;
1803                         }
1804                 }
1805
1806                 /* Go back up the rbtree to find next candidate node */
1807                 while (true) {
1808                         struct rb_node *prev = &vma->vm_rb;
1809                         if (!rb_parent(prev))
1810                                 return -ENOMEM;
1811                         vma = rb_entry(rb_parent(prev),
1812                                        struct vm_area_struct, vm_rb);
1813                         if (prev == vma->vm_rb.rb_right) {
1814                                 gap_start = vma->vm_prev ?
1815                                         vma->vm_prev->vm_end : 0;
1816                                 goto check_current;
1817                         }
1818                 }
1819         }
1820
1821 found:
1822         /* We found a suitable gap. Clip it with the original high_limit. */
1823         if (gap_end > info->high_limit)
1824                 gap_end = info->high_limit;
1825
1826 found_highest:
1827         /* Compute highest gap address at the desired alignment */
1828         gap_end -= info->length;
1829         gap_end -= (gap_end - info->align_offset) & info->align_mask;
1830
1831         VM_BUG_ON(gap_end < info->low_limit);
1832         VM_BUG_ON(gap_end < gap_start);
1833         return gap_end;
1834 }
1835
1836 /* Get an address range which is currently unmapped.
1837  * For shmat() with addr=0.
1838  *
1839  * Ugly calling convention alert:
1840  * Return value with the low bits set means error value,
1841  * ie
1842  *      if (ret & ~PAGE_MASK)
1843  *              error = ret;
1844  *
1845  * This function "knows" that -ENOMEM has the bits set.
1846  */
1847 #ifndef HAVE_ARCH_UNMAPPED_AREA
1848 unsigned long
1849 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1850                 unsigned long len, unsigned long pgoff, unsigned long flags)
1851 {
1852         struct mm_struct *mm = current->mm;
1853         struct vm_area_struct *vma;
1854         struct vm_unmapped_area_info info;
1855
1856         if (len > TASK_SIZE)
1857                 return -ENOMEM;
1858
1859         if (flags & MAP_FIXED)
1860                 return addr;
1861
1862         if (addr) {
1863                 addr = PAGE_ALIGN(addr);
1864                 vma = find_vma(mm, addr);
1865                 if (TASK_SIZE - len >= addr &&
1866                     (!vma || addr + len <= vma->vm_start))
1867                         return addr;
1868         }
1869
1870         info.flags = 0;
1871         info.length = len;
1872         info.low_limit = TASK_UNMAPPED_BASE;
1873         info.high_limit = TASK_SIZE;
1874         info.align_mask = 0;
1875         return vm_unmapped_area(&info);
1876 }
1877 #endif  
1878
1879 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1880 {
1881         /*
1882          * Is this a new hole at the lowest possible address?
1883          */
1884         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache)
1885                 mm->free_area_cache = addr;
1886 }
1887
1888 /*
1889  * This mmap-allocator allocates new areas top-down from below the
1890  * stack's low limit (the base):
1891  */
1892 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1893 unsigned long
1894 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1895                           const unsigned long len, const unsigned long pgoff,
1896                           const unsigned long flags)
1897 {
1898         struct vm_area_struct *vma;
1899         struct mm_struct *mm = current->mm;
1900         unsigned long addr = addr0;
1901         struct vm_unmapped_area_info info;
1902
1903         /* requested length too big for entire address space */
1904         if (len > TASK_SIZE)
1905                 return -ENOMEM;
1906
1907         if (flags & MAP_FIXED)
1908                 return addr;
1909
1910         /* requesting a specific address */
1911         if (addr) {
1912                 addr = PAGE_ALIGN(addr);
1913                 vma = find_vma(mm, addr);
1914                 if (TASK_SIZE - len >= addr &&
1915                                 (!vma || addr + len <= vma->vm_start))
1916                         return addr;
1917         }
1918
1919         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1920         info.length = len;
1921         info.low_limit = PAGE_SIZE;
1922         info.high_limit = mm->mmap_base;
1923         info.align_mask = 0;
1924         addr = vm_unmapped_area(&info);
1925
1926         /*
1927          * A failed mmap() very likely causes application failure,
1928          * so fall back to the bottom-up function here. This scenario
1929          * can happen with large stack limits and large mmap()
1930          * allocations.
1931          */
1932         if (addr & ~PAGE_MASK) {
1933                 VM_BUG_ON(addr != -ENOMEM);
1934                 info.flags = 0;
1935                 info.low_limit = TASK_UNMAPPED_BASE;
1936                 info.high_limit = TASK_SIZE;
1937                 addr = vm_unmapped_area(&info);
1938         }
1939
1940         return addr;
1941 }
1942 #endif
1943
1944 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1945 {
1946         /*
1947          * Is this a new hole at the highest possible address?
1948          */
1949         if (addr > mm->free_area_cache)
1950                 mm->free_area_cache = addr;
1951
1952         /* dont allow allocations above current base */
1953         if (mm->free_area_cache > mm->mmap_base)
1954                 mm->free_area_cache = mm->mmap_base;
1955 }
1956
1957 unsigned long
1958 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1959                 unsigned long pgoff, unsigned long flags)
1960 {
1961         unsigned long (*get_area)(struct file *, unsigned long,
1962                                   unsigned long, unsigned long, unsigned long);
1963
1964         unsigned long error = arch_mmap_check(addr, len, flags);
1965         if (error)
1966                 return error;
1967
1968         /* Careful about overflows.. */
1969         if (len > TASK_SIZE)
1970                 return -ENOMEM;
1971
1972         get_area = current->mm->get_unmapped_area;
1973         if (file && file->f_op && file->f_op->get_unmapped_area)
1974                 get_area = file->f_op->get_unmapped_area;
1975         addr = get_area(file, addr, len, pgoff, flags);
1976         if (IS_ERR_VALUE(addr))
1977                 return addr;
1978
1979         if (addr > TASK_SIZE - len)
1980                 return -ENOMEM;
1981         if (addr & ~PAGE_MASK)
1982                 return -EINVAL;
1983
1984         addr = arch_rebalance_pgtables(addr, len);
1985         error = security_mmap_addr(addr);
1986         return error ? error : addr;
1987 }
1988
1989 EXPORT_SYMBOL(get_unmapped_area);
1990
1991 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1992 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1993 {
1994         struct vm_area_struct *vma = NULL;
1995
1996         /* Check the cache first. */
1997         /* (Cache hit rate is typically around 35%.) */
1998         vma = ACCESS_ONCE(mm->mmap_cache);
1999         if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
2000                 struct rb_node *rb_node;
2001
2002                 rb_node = mm->mm_rb.rb_node;
2003                 vma = NULL;
2004
2005                 while (rb_node) {
2006                         struct vm_area_struct *vma_tmp;
2007
2008                         vma_tmp = rb_entry(rb_node,
2009                                            struct vm_area_struct, vm_rb);
2010
2011                         if (vma_tmp->vm_end > addr) {
2012                                 vma = vma_tmp;
2013                                 if (vma_tmp->vm_start <= addr)
2014                                         break;
2015                                 rb_node = rb_node->rb_left;
2016                         } else
2017                                 rb_node = rb_node->rb_right;
2018                 }
2019                 if (vma)
2020                         mm->mmap_cache = vma;
2021         }
2022         return vma;
2023 }
2024
2025 EXPORT_SYMBOL(find_vma);
2026
2027 /*
2028  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2029  */
2030 struct vm_area_struct *
2031 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2032                         struct vm_area_struct **pprev)
2033 {
2034         struct vm_area_struct *vma;
2035
2036         vma = find_vma(mm, addr);
2037         if (vma) {
2038                 *pprev = vma->vm_prev;
2039         } else {
2040                 struct rb_node *rb_node = mm->mm_rb.rb_node;
2041                 *pprev = NULL;
2042                 while (rb_node) {
2043                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2044                         rb_node = rb_node->rb_right;
2045                 }
2046         }
2047         return vma;
2048 }
2049
2050 /*
2051  * Verify that the stack growth is acceptable and
2052  * update accounting. This is shared with both the
2053  * grow-up and grow-down cases.
2054  */
2055 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
2056 {
2057         struct mm_struct *mm = vma->vm_mm;
2058         struct rlimit *rlim = current->signal->rlim;
2059         unsigned long new_start;
2060
2061         /* address space limit tests */
2062         if (!may_expand_vm(mm, grow))
2063                 return -ENOMEM;
2064
2065         /* Stack limit test */
2066         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2067                 return -ENOMEM;
2068
2069         /* mlock limit tests */
2070         if (vma->vm_flags & VM_LOCKED) {
2071                 unsigned long locked;
2072                 unsigned long limit;
2073                 locked = mm->locked_vm + grow;
2074                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2075                 limit >>= PAGE_SHIFT;
2076                 if (locked > limit && !capable(CAP_IPC_LOCK))
2077                         return -ENOMEM;
2078         }
2079
2080         /* Check to ensure the stack will not grow into a hugetlb-only region */
2081         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2082                         vma->vm_end - size;
2083         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2084                 return -EFAULT;
2085
2086         /*
2087          * Overcommit..  This must be the final test, as it will
2088          * update security statistics.
2089          */
2090         if (security_vm_enough_memory_mm(mm, grow))
2091                 return -ENOMEM;
2092
2093         /* Ok, everything looks good - let it rip */
2094         if (vma->vm_flags & VM_LOCKED)
2095                 mm->locked_vm += grow;
2096         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
2097         return 0;
2098 }
2099
2100 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2101 /*
2102  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2103  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2104  */
2105 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2106 {
2107         int error;
2108
2109         if (!(vma->vm_flags & VM_GROWSUP))
2110                 return -EFAULT;
2111
2112         /*
2113          * We must make sure the anon_vma is allocated
2114          * so that the anon_vma locking is not a noop.
2115          */
2116         if (unlikely(anon_vma_prepare(vma)))
2117                 return -ENOMEM;
2118         vma_lock_anon_vma(vma);
2119
2120         /*
2121          * vma->vm_start/vm_end cannot change under us because the caller
2122          * is required to hold the mmap_sem in read mode.  We need the
2123          * anon_vma lock to serialize against concurrent expand_stacks.
2124          * Also guard against wrapping around to address 0.
2125          */
2126         if (address < PAGE_ALIGN(address+4))
2127                 address = PAGE_ALIGN(address+4);
2128         else {
2129                 vma_unlock_anon_vma(vma);
2130                 return -ENOMEM;
2131         }
2132         error = 0;
2133
2134         /* Somebody else might have raced and expanded it already */
2135         if (address > vma->vm_end) {
2136                 unsigned long size, grow;
2137
2138                 size = address - vma->vm_start;
2139                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2140
2141                 error = -ENOMEM;
2142                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2143                         error = acct_stack_growth(vma, size, grow);
2144                         if (!error) {
2145                                 /*
2146                                  * vma_gap_update() doesn't support concurrent
2147                                  * updates, but we only hold a shared mmap_sem
2148                                  * lock here, so we need to protect against
2149                                  * concurrent vma expansions.
2150                                  * vma_lock_anon_vma() doesn't help here, as
2151                                  * we don't guarantee that all growable vmas
2152                                  * in a mm share the same root anon vma.
2153                                  * So, we reuse mm->page_table_lock to guard
2154                                  * against concurrent vma expansions.
2155                                  */
2156                                 spin_lock(&vma->vm_mm->page_table_lock);
2157                                 anon_vma_interval_tree_pre_update_vma(vma);
2158                                 vma->vm_end = address;
2159                                 anon_vma_interval_tree_post_update_vma(vma);
2160                                 if (vma->vm_next)
2161                                         vma_gap_update(vma->vm_next);
2162                                 else
2163                                         vma->vm_mm->highest_vm_end = address;
2164                                 spin_unlock(&vma->vm_mm->page_table_lock);
2165
2166                                 perf_event_mmap(vma);
2167                         }
2168                 }
2169         }
2170         vma_unlock_anon_vma(vma);
2171         khugepaged_enter_vma_merge(vma);
2172         validate_mm(vma->vm_mm);
2173         return error;
2174 }
2175 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2176
2177 /*
2178  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2179  */
2180 int expand_downwards(struct vm_area_struct *vma,
2181                                    unsigned long address)
2182 {
2183         int error;
2184
2185         /*
2186          * We must make sure the anon_vma is allocated
2187          * so that the anon_vma locking is not a noop.
2188          */
2189         if (unlikely(anon_vma_prepare(vma)))
2190                 return -ENOMEM;
2191
2192         address &= PAGE_MASK;
2193         error = security_mmap_addr(address);
2194         if (error)
2195                 return error;
2196
2197         vma_lock_anon_vma(vma);
2198
2199         /*
2200          * vma->vm_start/vm_end cannot change under us because the caller
2201          * is required to hold the mmap_sem in read mode.  We need the
2202          * anon_vma lock to serialize against concurrent expand_stacks.
2203          */
2204
2205         /* Somebody else might have raced and expanded it already */
2206         if (address < vma->vm_start) {
2207                 unsigned long size, grow;
2208
2209                 size = vma->vm_end - address;
2210                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2211
2212                 error = -ENOMEM;
2213                 if (grow <= vma->vm_pgoff) {
2214                         error = acct_stack_growth(vma, size, grow);
2215                         if (!error) {
2216                                 /*
2217                                  * vma_gap_update() doesn't support concurrent
2218                                  * updates, but we only hold a shared mmap_sem
2219                                  * lock here, so we need to protect against
2220                                  * concurrent vma expansions.
2221                                  * vma_lock_anon_vma() doesn't help here, as
2222                                  * we don't guarantee that all growable vmas
2223                                  * in a mm share the same root anon vma.
2224                                  * So, we reuse mm->page_table_lock to guard
2225                                  * against concurrent vma expansions.
2226                                  */
2227                                 spin_lock(&vma->vm_mm->page_table_lock);
2228                                 anon_vma_interval_tree_pre_update_vma(vma);
2229                                 vma->vm_start = address;
2230                                 vma->vm_pgoff -= grow;
2231                                 anon_vma_interval_tree_post_update_vma(vma);
2232                                 vma_gap_update(vma);
2233                                 spin_unlock(&vma->vm_mm->page_table_lock);
2234
2235                                 perf_event_mmap(vma);
2236                         }
2237                 }
2238         }
2239         vma_unlock_anon_vma(vma);
2240         khugepaged_enter_vma_merge(vma);
2241         validate_mm(vma->vm_mm);
2242         return error;
2243 }
2244
2245 /*
2246  * Note how expand_stack() refuses to expand the stack all the way to
2247  * abut the next virtual mapping, *unless* that mapping itself is also
2248  * a stack mapping. We want to leave room for a guard page, after all
2249  * (the guard page itself is not added here, that is done by the
2250  * actual page faulting logic)
2251  *
2252  * This matches the behavior of the guard page logic (see mm/memory.c:
2253  * check_stack_guard_page()), which only allows the guard page to be
2254  * removed under these circumstances.
2255  */
2256 #ifdef CONFIG_STACK_GROWSUP
2257 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2258 {
2259         struct vm_area_struct *next;
2260
2261         address &= PAGE_MASK;
2262         next = vma->vm_next;
2263         if (next && next->vm_start == address + PAGE_SIZE) {
2264                 if (!(next->vm_flags & VM_GROWSUP))
2265                         return -ENOMEM;
2266         }
2267         return expand_upwards(vma, address);
2268 }
2269
2270 struct vm_area_struct *
2271 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2272 {
2273         struct vm_area_struct *vma, *prev;
2274
2275         addr &= PAGE_MASK;
2276         vma = find_vma_prev(mm, addr, &prev);
2277         if (vma && (vma->vm_start <= addr))
2278                 return vma;
2279         if (!prev || expand_stack(prev, addr))
2280                 return NULL;
2281         if (prev->vm_flags & VM_LOCKED)
2282                 __mlock_vma_pages_range(prev, addr, prev->vm_end, NULL);
2283         return prev;
2284 }
2285 #else
2286 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2287 {
2288         struct vm_area_struct *prev;
2289
2290         address &= PAGE_MASK;
2291         prev = vma->vm_prev;
2292         if (prev && prev->vm_end == address) {
2293                 if (!(prev->vm_flags & VM_GROWSDOWN))
2294                         return -ENOMEM;
2295         }
2296         return expand_downwards(vma, address);
2297 }
2298
2299 struct vm_area_struct *
2300 find_extend_vma(struct mm_struct * mm, unsigned long addr)
2301 {
2302         struct vm_area_struct * vma;
2303         unsigned long start;
2304
2305         addr &= PAGE_MASK;
2306         vma = find_vma(mm,addr);
2307         if (!vma)
2308                 return NULL;
2309         if (vma->vm_start <= addr)
2310                 return vma;
2311         if (!(vma->vm_flags & VM_GROWSDOWN))
2312                 return NULL;
2313         start = vma->vm_start;
2314         if (expand_stack(vma, addr))
2315                 return NULL;
2316         if (vma->vm_flags & VM_LOCKED)
2317                 __mlock_vma_pages_range(vma, addr, start, NULL);
2318         return vma;
2319 }
2320 #endif
2321
2322 /*
2323  * Ok - we have the memory areas we should free on the vma list,
2324  * so release them, and do the vma updates.
2325  *
2326  * Called with the mm semaphore held.
2327  */
2328 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2329 {
2330         unsigned long nr_accounted = 0;
2331
2332         /* Update high watermark before we lower total_vm */
2333         update_hiwater_vm(mm);
2334         do {
2335                 long nrpages = vma_pages(vma);
2336
2337                 if (vma->vm_flags & VM_ACCOUNT)
2338                         nr_accounted += nrpages;
2339                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
2340                 vma = remove_vma(vma);
2341         } while (vma);
2342         vm_unacct_memory(nr_accounted);
2343         validate_mm(mm);
2344 }
2345
2346 /*
2347  * Get rid of page table information in the indicated region.
2348  *
2349  * Called with the mm semaphore held.
2350  */
2351 static void unmap_region(struct mm_struct *mm,
2352                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2353                 unsigned long start, unsigned long end)
2354 {
2355         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
2356         struct mmu_gather tlb;
2357
2358         lru_add_drain();
2359         tlb_gather_mmu(&tlb, mm, 0);
2360         update_hiwater_rss(mm);
2361         unmap_vmas(&tlb, vma, start, end);
2362         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2363                                  next ? next->vm_start : USER_PGTABLES_CEILING);
2364         tlb_finish_mmu(&tlb, start, end);
2365 }
2366
2367 /*
2368  * Create a list of vma's touched by the unmap, removing them from the mm's
2369  * vma list as we go..
2370  */
2371 static void
2372 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2373         struct vm_area_struct *prev, unsigned long end)
2374 {
2375         struct vm_area_struct **insertion_point;
2376         struct vm_area_struct *tail_vma = NULL;
2377         unsigned long addr;
2378
2379         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2380         vma->vm_prev = NULL;
2381         do {
2382                 vma_rb_erase(vma, &mm->mm_rb);
2383                 mm->map_count--;
2384                 tail_vma = vma;
2385                 vma = vma->vm_next;
2386         } while (vma && vma->vm_start < end);
2387         *insertion_point = vma;
2388         if (vma) {
2389                 vma->vm_prev = prev;
2390                 vma_gap_update(vma);
2391         } else
2392                 mm->highest_vm_end = prev ? prev->vm_end : 0;
2393         tail_vma->vm_next = NULL;
2394         if (mm->unmap_area == arch_unmap_area)
2395                 addr = prev ? prev->vm_end : mm->mmap_base;
2396         else
2397                 addr = vma ?  vma->vm_start : mm->mmap_base;
2398         mm->unmap_area(mm, addr);
2399         mm->mmap_cache = NULL;          /* Kill the cache. */
2400 }
2401
2402 /*
2403  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
2404  * munmap path where it doesn't make sense to fail.
2405  */
2406 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
2407               unsigned long addr, int new_below)
2408 {
2409         struct mempolicy *pol;
2410         struct vm_area_struct *new;
2411         int err = -ENOMEM;
2412
2413         if (is_vm_hugetlb_page(vma) && (addr &
2414                                         ~(huge_page_mask(hstate_vma(vma)))))
2415                 return -EINVAL;
2416
2417         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2418         if (!new)
2419                 goto out_err;
2420
2421         /* most fields are the same, copy all, and then fixup */
2422         *new = *vma;
2423
2424         INIT_LIST_HEAD(&new->anon_vma_chain);
2425
2426         if (new_below)
2427                 new->vm_end = addr;
2428         else {
2429                 new->vm_start = addr;
2430                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2431         }
2432
2433         pol = mpol_dup(vma_policy(vma));
2434         if (IS_ERR(pol)) {
2435                 err = PTR_ERR(pol);
2436                 goto out_free_vma;
2437         }
2438         vma_set_policy(new, pol);
2439
2440         if (anon_vma_clone(new, vma))
2441                 goto out_free_mpol;
2442
2443         if (new->vm_file)
2444                 get_file(new->vm_file);
2445
2446         if (new->vm_ops && new->vm_ops->open)
2447                 new->vm_ops->open(new);
2448
2449         if (new_below)
2450                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2451                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2452         else
2453                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2454
2455         /* Success. */
2456         if (!err)
2457                 return 0;
2458
2459         /* Clean everything up if vma_adjust failed. */
2460         if (new->vm_ops && new->vm_ops->close)
2461                 new->vm_ops->close(new);
2462         if (new->vm_file)
2463                 fput(new->vm_file);
2464         unlink_anon_vmas(new);
2465  out_free_mpol:
2466         mpol_put(pol);
2467  out_free_vma:
2468         kmem_cache_free(vm_area_cachep, new);
2469  out_err:
2470         return err;
2471 }
2472
2473 /*
2474  * Split a vma into two pieces at address 'addr', a new vma is allocated
2475  * either for the first part or the tail.
2476  */
2477 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2478               unsigned long addr, int new_below)
2479 {
2480         if (mm->map_count >= sysctl_max_map_count)
2481                 return -ENOMEM;
2482
2483         return __split_vma(mm, vma, addr, new_below);
2484 }
2485
2486 /* Munmap is split into 2 main parts -- this part which finds
2487  * what needs doing, and the areas themselves, which do the
2488  * work.  This now handles partial unmappings.
2489  * Jeremy Fitzhardinge <jeremy@goop.org>
2490  */
2491 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2492 {
2493         unsigned long end;
2494         struct vm_area_struct *vma, *prev, *last;
2495
2496         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2497                 return -EINVAL;
2498
2499         if ((len = PAGE_ALIGN(len)) == 0)
2500                 return -EINVAL;
2501
2502         /* Find the first overlapping VMA */
2503         vma = find_vma(mm, start);
2504         if (!vma)
2505                 return 0;
2506         prev = vma->vm_prev;
2507         /* we have  start < vma->vm_end  */
2508
2509         /* if it doesn't overlap, we have nothing.. */
2510         end = start + len;
2511         if (vma->vm_start >= end)
2512                 return 0;
2513
2514         /*
2515          * If we need to split any vma, do it now to save pain later.
2516          *
2517          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2518          * unmapped vm_area_struct will remain in use: so lower split_vma
2519          * places tmp vma above, and higher split_vma places tmp vma below.
2520          */
2521         if (start > vma->vm_start) {
2522                 int error;
2523
2524                 /*
2525                  * Make sure that map_count on return from munmap() will
2526                  * not exceed its limit; but let map_count go just above
2527                  * its limit temporarily, to help free resources as expected.
2528                  */
2529                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2530                         return -ENOMEM;
2531
2532                 error = __split_vma(mm, vma, start, 0);
2533                 if (error)
2534                         return error;
2535                 prev = vma;
2536         }
2537
2538         /* Does it split the last one? */
2539         last = find_vma(mm, end);
2540         if (last && end > last->vm_start) {
2541                 int error = __split_vma(mm, last, end, 1);
2542                 if (error)
2543                         return error;
2544         }
2545         vma = prev? prev->vm_next: mm->mmap;
2546
2547         /*
2548          * unlock any mlock()ed ranges before detaching vmas
2549          */
2550         if (mm->locked_vm) {
2551                 struct vm_area_struct *tmp = vma;
2552                 while (tmp && tmp->vm_start < end) {
2553                         if (tmp->vm_flags & VM_LOCKED) {
2554                                 mm->locked_vm -= vma_pages(tmp);
2555                                 munlock_vma_pages_all(tmp);
2556                         }
2557                         tmp = tmp->vm_next;
2558                 }
2559         }
2560
2561         /*
2562          * Remove the vma's, and unmap the actual pages
2563          */
2564         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2565         unmap_region(mm, vma, prev, start, end);
2566
2567         /* Fix up all other VM information */
2568         remove_vma_list(mm, vma);
2569
2570         return 0;
2571 }
2572
2573 int vm_munmap(unsigned long start, size_t len)
2574 {
2575         int ret;
2576         struct mm_struct *mm = current->mm;
2577
2578         down_write(&mm->mmap_sem);
2579         ret = do_munmap(mm, start, len);
2580         up_write(&mm->mmap_sem);
2581         return ret;
2582 }
2583 EXPORT_SYMBOL(vm_munmap);
2584
2585 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2586 {
2587         profile_munmap(addr);
2588         return vm_munmap(addr, len);
2589 }
2590
2591 static inline void verify_mm_writelocked(struct mm_struct *mm)
2592 {
2593 #ifdef CONFIG_DEBUG_VM
2594         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2595                 WARN_ON(1);
2596                 up_read(&mm->mmap_sem);
2597         }
2598 #endif
2599 }
2600
2601 /*
2602  *  this is really a simplified "do_mmap".  it only handles
2603  *  anonymous maps.  eventually we may be able to do some
2604  *  brk-specific accounting here.
2605  */
2606 static unsigned long do_brk(unsigned long addr, unsigned long len)
2607 {
2608         struct mm_struct * mm = current->mm;
2609         struct vm_area_struct * vma, * prev;
2610         unsigned long flags;
2611         struct rb_node ** rb_link, * rb_parent;
2612         pgoff_t pgoff = addr >> PAGE_SHIFT;
2613         int error;
2614
2615         len = PAGE_ALIGN(len);
2616         if (!len)
2617                 return addr;
2618
2619         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2620
2621         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2622         if (error & ~PAGE_MASK)
2623                 return error;
2624
2625         /*
2626          * mlock MCL_FUTURE?
2627          */
2628         if (mm->def_flags & VM_LOCKED) {
2629                 unsigned long locked, lock_limit;
2630                 locked = len >> PAGE_SHIFT;
2631                 locked += mm->locked_vm;
2632                 lock_limit = rlimit(RLIMIT_MEMLOCK);
2633                 lock_limit >>= PAGE_SHIFT;
2634                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2635                         return -EAGAIN;
2636         }
2637
2638         /*
2639          * mm->mmap_sem is required to protect against another thread
2640          * changing the mappings in case we sleep.
2641          */
2642         verify_mm_writelocked(mm);
2643
2644         /*
2645          * Clear old maps.  this also does some error checking for us
2646          */
2647  munmap_back:
2648         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
2649                 if (do_munmap(mm, addr, len))
2650                         return -ENOMEM;
2651                 goto munmap_back;
2652         }
2653
2654         /* Check against address space limits *after* clearing old maps... */
2655         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2656                 return -ENOMEM;
2657
2658         if (mm->map_count > sysctl_max_map_count)
2659                 return -ENOMEM;
2660
2661         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2662                 return -ENOMEM;
2663
2664         /* Can we just expand an old private anonymous mapping? */
2665         vma = vma_merge(mm, prev, addr, addr + len, flags,
2666                                         NULL, NULL, pgoff, NULL);
2667         if (vma)
2668                 goto out;
2669
2670         /*
2671          * create a vma struct for an anonymous mapping
2672          */
2673         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2674         if (!vma) {
2675                 vm_unacct_memory(len >> PAGE_SHIFT);
2676                 return -ENOMEM;
2677         }
2678
2679         INIT_LIST_HEAD(&vma->anon_vma_chain);
2680         vma->vm_mm = mm;
2681         vma->vm_start = addr;
2682         vma->vm_end = addr + len;
2683         vma->vm_pgoff = pgoff;
2684         vma->vm_flags = flags;
2685         vma->vm_page_prot = vm_get_page_prot(flags);
2686         vma_link(mm, vma, prev, rb_link, rb_parent);
2687 out:
2688         perf_event_mmap(vma);
2689         mm->total_vm += len >> PAGE_SHIFT;
2690         if (flags & VM_LOCKED)
2691                 mm->locked_vm += (len >> PAGE_SHIFT);
2692         return addr;
2693 }
2694
2695 unsigned long vm_brk(unsigned long addr, unsigned long len)
2696 {
2697         struct mm_struct *mm = current->mm;
2698         unsigned long ret;
2699         bool populate;
2700
2701         down_write(&mm->mmap_sem);
2702         ret = do_brk(addr, len);
2703         populate = ((mm->def_flags & VM_LOCKED) != 0);
2704         up_write(&mm->mmap_sem);
2705         if (populate)
2706                 mm_populate(addr, len);
2707         return ret;
2708 }
2709 EXPORT_SYMBOL(vm_brk);
2710
2711 /* Release all mmaps. */
2712 void exit_mmap(struct mm_struct *mm)
2713 {
2714         struct mmu_gather tlb;
2715         struct vm_area_struct *vma;
2716         unsigned long nr_accounted = 0;
2717
2718         /* mm's last user has gone, and its about to be pulled down */
2719         mmu_notifier_release(mm);
2720
2721         if (mm->locked_vm) {
2722                 vma = mm->mmap;
2723                 while (vma) {
2724                         if (vma->vm_flags & VM_LOCKED)
2725                                 munlock_vma_pages_all(vma);
2726                         vma = vma->vm_next;
2727                 }
2728         }
2729
2730         arch_exit_mmap(mm);
2731
2732         vma = mm->mmap;
2733         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2734                 return;
2735
2736         lru_add_drain();
2737         flush_cache_mm(mm);
2738         tlb_gather_mmu(&tlb, mm, 1);
2739         /* update_hiwater_rss(mm) here? but nobody should be looking */
2740         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2741         unmap_vmas(&tlb, vma, 0, -1);
2742
2743         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
2744         tlb_finish_mmu(&tlb, 0, -1);
2745
2746         /*
2747          * Walk the list again, actually closing and freeing it,
2748          * with preemption enabled, without holding any MM locks.
2749          */
2750         while (vma) {
2751                 if (vma->vm_flags & VM_ACCOUNT)
2752                         nr_accounted += vma_pages(vma);
2753                 vma = remove_vma(vma);
2754         }
2755         vm_unacct_memory(nr_accounted);
2756
2757         WARN_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2758 }
2759
2760 /* Insert vm structure into process list sorted by address
2761  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2762  * then i_mmap_mutex is taken here.
2763  */
2764 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2765 {
2766         struct vm_area_struct *prev;
2767         struct rb_node **rb_link, *rb_parent;
2768
2769         /*
2770          * The vm_pgoff of a purely anonymous vma should be irrelevant
2771          * until its first write fault, when page's anon_vma and index
2772          * are set.  But now set the vm_pgoff it will almost certainly
2773          * end up with (unless mremap moves it elsewhere before that
2774          * first wfault), so /proc/pid/maps tells a consistent story.
2775          *
2776          * By setting it to reflect the virtual start address of the
2777          * vma, merges and splits can happen in a seamless way, just
2778          * using the existing file pgoff checks and manipulations.
2779          * Similarly in do_mmap_pgoff and in do_brk.
2780          */
2781         if (!vma->vm_file) {
2782                 BUG_ON(vma->anon_vma);
2783                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2784         }
2785         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2786                            &prev, &rb_link, &rb_parent))
2787                 return -ENOMEM;
2788         if ((vma->vm_flags & VM_ACCOUNT) &&
2789              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2790                 return -ENOMEM;
2791
2792         vma_link(mm, vma, prev, rb_link, rb_parent);
2793         return 0;
2794 }
2795
2796 /*
2797  * Copy the vma structure to a new location in the same mm,
2798  * prior to moving page table entries, to effect an mremap move.
2799  */
2800 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2801         unsigned long addr, unsigned long len, pgoff_t pgoff,
2802         bool *need_rmap_locks)
2803 {
2804         struct vm_area_struct *vma = *vmap;
2805         unsigned long vma_start = vma->vm_start;
2806         struct mm_struct *mm = vma->vm_mm;
2807         struct vm_area_struct *new_vma, *prev;
2808         struct rb_node **rb_link, *rb_parent;
2809         struct mempolicy *pol;
2810         bool faulted_in_anon_vma = true;
2811
2812         /*
2813          * If anonymous vma has not yet been faulted, update new pgoff
2814          * to match new location, to increase its chance of merging.
2815          */
2816         if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2817                 pgoff = addr >> PAGE_SHIFT;
2818                 faulted_in_anon_vma = false;
2819         }
2820
2821         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2822                 return NULL;    /* should never get here */
2823         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2824                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2825         if (new_vma) {
2826                 /*
2827                  * Source vma may have been merged into new_vma
2828                  */
2829                 if (unlikely(vma_start >= new_vma->vm_start &&
2830                              vma_start < new_vma->vm_end)) {
2831                         /*
2832                          * The only way we can get a vma_merge with
2833                          * self during an mremap is if the vma hasn't
2834                          * been faulted in yet and we were allowed to
2835                          * reset the dst vma->vm_pgoff to the
2836                          * destination address of the mremap to allow
2837                          * the merge to happen. mremap must change the
2838                          * vm_pgoff linearity between src and dst vmas
2839                          * (in turn preventing a vma_merge) to be
2840                          * safe. It is only safe to keep the vm_pgoff
2841                          * linear if there are no pages mapped yet.
2842                          */
2843                         VM_BUG_ON(faulted_in_anon_vma);
2844                         *vmap = vma = new_vma;
2845                 }
2846                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2847         } else {
2848                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2849                 if (new_vma) {
2850                         *new_vma = *vma;
2851                         new_vma->vm_start = addr;
2852                         new_vma->vm_end = addr + len;
2853                         new_vma->vm_pgoff = pgoff;
2854                         pol = mpol_dup(vma_policy(vma));
2855                         if (IS_ERR(pol))
2856                                 goto out_free_vma;
2857                         vma_set_policy(new_vma, pol);
2858                         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2859                         if (anon_vma_clone(new_vma, vma))
2860                                 goto out_free_mempol;
2861                         if (new_vma->vm_file)
2862                                 get_file(new_vma->vm_file);
2863                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2864                                 new_vma->vm_ops->open(new_vma);
2865                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2866                         *need_rmap_locks = false;
2867                 }
2868         }
2869         return new_vma;
2870
2871  out_free_mempol:
2872         mpol_put(pol);
2873  out_free_vma:
2874         kmem_cache_free(vm_area_cachep, new_vma);
2875         return NULL;
2876 }
2877
2878 /*
2879  * Return true if the calling process may expand its vm space by the passed
2880  * number of pages
2881  */
2882 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2883 {
2884         unsigned long cur = mm->total_vm;       /* pages */
2885         unsigned long lim;
2886
2887         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2888
2889         if (cur + npages > lim)
2890                 return 0;
2891         return 1;
2892 }
2893
2894
2895 static int special_mapping_fault(struct vm_area_struct *vma,
2896                                 struct vm_fault *vmf)
2897 {
2898         pgoff_t pgoff;
2899         struct page **pages;
2900
2901         /*
2902          * special mappings have no vm_file, and in that case, the mm
2903          * uses vm_pgoff internally. So we have to subtract it from here.
2904          * We are allowed to do this because we are the mm; do not copy
2905          * this code into drivers!
2906          */
2907         pgoff = vmf->pgoff - vma->vm_pgoff;
2908
2909         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2910                 pgoff--;
2911
2912         if (*pages) {
2913                 struct page *page = *pages;
2914                 get_page(page);
2915                 vmf->page = page;
2916                 return 0;
2917         }
2918
2919         return VM_FAULT_SIGBUS;
2920 }
2921
2922 /*
2923  * Having a close hook prevents vma merging regardless of flags.
2924  */
2925 static void special_mapping_close(struct vm_area_struct *vma)
2926 {
2927 }
2928
2929 static const struct vm_operations_struct special_mapping_vmops = {
2930         .close = special_mapping_close,
2931         .fault = special_mapping_fault,
2932 };
2933
2934 /*
2935  * Called with mm->mmap_sem held for writing.
2936  * Insert a new vma covering the given region, with the given flags.
2937  * Its pages are supplied by the given array of struct page *.
2938  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2939  * The region past the last page supplied will always produce SIGBUS.
2940  * The array pointer and the pages it points to are assumed to stay alive
2941  * for as long as this mapping might exist.
2942  */
2943 int install_special_mapping(struct mm_struct *mm,
2944                             unsigned long addr, unsigned long len,
2945                             unsigned long vm_flags, struct page **pages)
2946 {
2947         int ret;
2948         struct vm_area_struct *vma;
2949
2950         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2951         if (unlikely(vma == NULL))
2952                 return -ENOMEM;
2953
2954         INIT_LIST_HEAD(&vma->anon_vma_chain);
2955         vma->vm_mm = mm;
2956         vma->vm_start = addr;
2957         vma->vm_end = addr + len;
2958
2959         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2960         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2961
2962         vma->vm_ops = &special_mapping_vmops;
2963         vma->vm_private_data = pages;
2964
2965         ret = insert_vm_struct(mm, vma);
2966         if (ret)
2967                 goto out;
2968
2969         mm->total_vm += len >> PAGE_SHIFT;
2970
2971         perf_event_mmap(vma);
2972
2973         return 0;
2974
2975 out:
2976         kmem_cache_free(vm_area_cachep, vma);
2977         return ret;
2978 }
2979
2980 static DEFINE_MUTEX(mm_all_locks_mutex);
2981
2982 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2983 {
2984         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
2985                 /*
2986                  * The LSB of head.next can't change from under us
2987                  * because we hold the mm_all_locks_mutex.
2988                  */
2989                 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
2990                 /*
2991                  * We can safely modify head.next after taking the
2992                  * anon_vma->root->rwsem. If some other vma in this mm shares
2993                  * the same anon_vma we won't take it again.
2994                  *
2995                  * No need of atomic instructions here, head.next
2996                  * can't change from under us thanks to the
2997                  * anon_vma->root->rwsem.
2998                  */
2999                 if (__test_and_set_bit(0, (unsigned long *)
3000                                        &anon_vma->root->rb_root.rb_node))
3001                         BUG();
3002         }
3003 }
3004
3005 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3006 {
3007         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3008                 /*
3009                  * AS_MM_ALL_LOCKS can't change from under us because
3010                  * we hold the mm_all_locks_mutex.
3011                  *
3012                  * Operations on ->flags have to be atomic because
3013                  * even if AS_MM_ALL_LOCKS is stable thanks to the
3014                  * mm_all_locks_mutex, there may be other cpus
3015                  * changing other bitflags in parallel to us.
3016                  */
3017                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3018                         BUG();
3019                 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
3020         }
3021 }
3022
3023 /*
3024  * This operation locks against the VM for all pte/vma/mm related
3025  * operations that could ever happen on a certain mm. This includes
3026  * vmtruncate, try_to_unmap, and all page faults.
3027  *
3028  * The caller must take the mmap_sem in write mode before calling
3029  * mm_take_all_locks(). The caller isn't allowed to release the
3030  * mmap_sem until mm_drop_all_locks() returns.
3031  *
3032  * mmap_sem in write mode is required in order to block all operations
3033  * that could modify pagetables and free pages without need of
3034  * altering the vma layout (for example populate_range() with
3035  * nonlinear vmas). It's also needed in write mode to avoid new
3036  * anon_vmas to be associated with existing vmas.
3037  *
3038  * A single task can't take more than one mm_take_all_locks() in a row
3039  * or it would deadlock.
3040  *
3041  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3042  * mapping->flags avoid to take the same lock twice, if more than one
3043  * vma in this mm is backed by the same anon_vma or address_space.
3044  *
3045  * We can take all the locks in random order because the VM code
3046  * taking i_mmap_mutex or anon_vma->rwsem outside the mmap_sem never
3047  * takes more than one of them in a row. Secondly we're protected
3048  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3049  *
3050  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3051  * that may have to take thousand of locks.
3052  *
3053  * mm_take_all_locks() can fail if it's interrupted by signals.
3054  */
3055 int mm_take_all_locks(struct mm_struct *mm)
3056 {
3057         struct vm_area_struct *vma;
3058         struct anon_vma_chain *avc;
3059
3060         BUG_ON(down_read_trylock(&mm->mmap_sem));
3061
3062         mutex_lock(&mm_all_locks_mutex);
3063
3064         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3065                 if (signal_pending(current))
3066                         goto out_unlock;
3067                 if (vma->vm_file && vma->vm_file->f_mapping)
3068                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3069         }
3070
3071         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3072                 if (signal_pending(current))
3073                         goto out_unlock;
3074                 if (vma->anon_vma)
3075                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3076                                 vm_lock_anon_vma(mm, avc->anon_vma);
3077         }
3078
3079         return 0;
3080
3081 out_unlock:
3082         mm_drop_all_locks(mm);
3083         return -EINTR;
3084 }
3085
3086 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3087 {
3088         if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3089                 /*
3090                  * The LSB of head.next can't change to 0 from under
3091                  * us because we hold the mm_all_locks_mutex.
3092                  *
3093                  * We must however clear the bitflag before unlocking
3094                  * the vma so the users using the anon_vma->rb_root will
3095                  * never see our bitflag.
3096                  *
3097                  * No need of atomic instructions here, head.next
3098                  * can't change from under us until we release the
3099                  * anon_vma->root->rwsem.
3100                  */
3101                 if (!__test_and_clear_bit(0, (unsigned long *)
3102                                           &anon_vma->root->rb_root.rb_node))
3103                         BUG();
3104                 anon_vma_unlock_write(anon_vma);
3105         }
3106 }
3107
3108 static void vm_unlock_mapping(struct address_space *mapping)
3109 {
3110         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3111                 /*
3112                  * AS_MM_ALL_LOCKS can't change to 0 from under us
3113                  * because we hold the mm_all_locks_mutex.
3114                  */
3115                 mutex_unlock(&mapping->i_mmap_mutex);
3116                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3117                                         &mapping->flags))
3118                         BUG();
3119         }
3120 }
3121
3122 /*
3123  * The mmap_sem cannot be released by the caller until
3124  * mm_drop_all_locks() returns.
3125  */
3126 void mm_drop_all_locks(struct mm_struct *mm)
3127 {
3128         struct vm_area_struct *vma;
3129         struct anon_vma_chain *avc;
3130
3131         BUG_ON(down_read_trylock(&mm->mmap_sem));
3132         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3133
3134         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3135                 if (vma->anon_vma)
3136                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3137                                 vm_unlock_anon_vma(avc->anon_vma);
3138                 if (vma->vm_file && vma->vm_file->f_mapping)
3139                         vm_unlock_mapping(vma->vm_file->f_mapping);
3140         }
3141
3142         mutex_unlock(&mm_all_locks_mutex);
3143 }
3144
3145 /*
3146  * initialise the VMA slab
3147  */
3148 void __init mmap_init(void)
3149 {
3150         int ret;
3151
3152         ret = percpu_counter_init(&vm_committed_as, 0);
3153         VM_BUG_ON(ret);
3154 }
3155
3156 /*
3157  * Initialise sysctl_user_reserve_kbytes.
3158  *
3159  * This is intended to prevent a user from starting a single memory hogging
3160  * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3161  * mode.
3162  *
3163  * The default value is min(3% of free memory, 128MB)
3164  * 128MB is enough to recover with sshd/login, bash, and top/kill.
3165  */
3166 static int init_user_reserve(void)
3167 {
3168         unsigned long free_kbytes;
3169
3170         free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3171
3172         sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3173         return 0;
3174 }
3175 module_init(init_user_reserve)
3176
3177 /*
3178  * Initialise sysctl_admin_reserve_kbytes.
3179  *
3180  * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3181  * to log in and kill a memory hogging process.
3182  *
3183  * Systems with more than 256MB will reserve 8MB, enough to recover
3184  * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3185  * only reserve 3% of free pages by default.
3186  */
3187 static int init_admin_reserve(void)
3188 {
3189         unsigned long free_kbytes;
3190
3191         free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3192
3193         sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3194         return 0;
3195 }
3196 module_init(init_admin_reserve)
3197
3198 /*
3199  * Reinititalise user and admin reserves if memory is added or removed.
3200  *
3201  * The default user reserve max is 128MB, and the default max for the
3202  * admin reserve is 8MB. These are usually, but not always, enough to
3203  * enable recovery from a memory hogging process using login/sshd, a shell,
3204  * and tools like top. It may make sense to increase or even disable the
3205  * reserve depending on the existence of swap or variations in the recovery
3206  * tools. So, the admin may have changed them.
3207  *
3208  * If memory is added and the reserves have been eliminated or increased above
3209  * the default max, then we'll trust the admin.
3210  *
3211  * If memory is removed and there isn't enough free memory, then we
3212  * need to reset the reserves.
3213  *
3214  * Otherwise keep the reserve set by the admin.
3215  */
3216 static int reserve_mem_notifier(struct notifier_block *nb,
3217                              unsigned long action, void *data)
3218 {
3219         unsigned long tmp, free_kbytes;
3220
3221         switch (action) {
3222         case MEM_ONLINE:
3223                 /* Default max is 128MB. Leave alone if modified by operator. */
3224                 tmp = sysctl_user_reserve_kbytes;
3225                 if (0 < tmp && tmp < (1UL << 17))
3226                         init_user_reserve();
3227
3228                 /* Default max is 8MB.  Leave alone if modified by operator. */
3229                 tmp = sysctl_admin_reserve_kbytes;
3230                 if (0 < tmp && tmp < (1UL << 13))
3231                         init_admin_reserve();
3232
3233                 break;
3234         case MEM_OFFLINE:
3235                 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3236
3237                 if (sysctl_user_reserve_kbytes > free_kbytes) {
3238                         init_user_reserve();
3239                         pr_info("vm.user_reserve_kbytes reset to %lu\n",
3240                                 sysctl_user_reserve_kbytes);
3241                 }
3242
3243                 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3244                         init_admin_reserve();
3245                         pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3246                                 sysctl_admin_reserve_kbytes);
3247                 }
3248                 break;
3249         default:
3250                 break;
3251         }
3252         return NOTIFY_OK;
3253 }
3254
3255 static struct notifier_block reserve_mem_nb = {
3256         .notifier_call = reserve_mem_notifier,
3257 };
3258
3259 static int __meminit init_reserve_notifier(void)
3260 {
3261         if (register_hotmemory_notifier(&reserve_mem_nb))
3262                 printk("Failed registering memory add/remove notifier for admin reserve");
3263
3264         return 0;
3265 }
3266 module_init(init_reserve_notifier)