ath10k: clean up assoc code
[linux-2.6-microblaze.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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10
11 #include <linux/kernel.h>
12 #include <linux/slab.h>
13 #include <linux/backing-dev.h>
14 #include <linux/mm.h>
15 #include <linux/vmacache.h>
16 #include <linux/shm.h>
17 #include <linux/mman.h>
18 #include <linux/pagemap.h>
19 #include <linux/swap.h>
20 #include <linux/syscalls.h>
21 #include <linux/capability.h>
22 #include <linux/init.h>
23 #include <linux/file.h>
24 #include <linux/fs.h>
25 #include <linux/personality.h>
26 #include <linux/security.h>
27 #include <linux/hugetlb.h>
28 #include <linux/profile.h>
29 #include <linux/export.h>
30 #include <linux/mount.h>
31 #include <linux/mempolicy.h>
32 #include <linux/rmap.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/mmdebug.h>
35 #include <linux/perf_event.h>
36 #include <linux/audit.h>
37 #include <linux/khugepaged.h>
38 #include <linux/uprobes.h>
39 #include <linux/rbtree_augmented.h>
40 #include <linux/sched/sysctl.h>
41 #include <linux/notifier.h>
42 #include <linux/memory.h>
43 #include <linux/printk.h>
44
45 #include <asm/uaccess.h>
46 #include <asm/cacheflush.h>
47 #include <asm/tlb.h>
48 #include <asm/mmu_context.h>
49
50 #include "internal.h"
51
52 #ifndef arch_mmap_check
53 #define arch_mmap_check(addr, len, flags)       (0)
54 #endif
55
56 #ifndef arch_rebalance_pgtables
57 #define arch_rebalance_pgtables(addr, len)              (addr)
58 #endif
59
60 static void unmap_region(struct mm_struct *mm,
61                 struct vm_area_struct *vma, struct vm_area_struct *prev,
62                 unsigned long start, unsigned long end);
63
64 /* description of effects of mapping type and prot in current implementation.
65  * this is due to the limited x86 page protection hardware.  The expected
66  * behavior is in parens:
67  *
68  * map_type     prot
69  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
70  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
71  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
72  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
73  *              
74  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
75  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
76  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
77  *
78  */
79 pgprot_t protection_map[16] = {
80         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
81         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
82 };
83
84 pgprot_t vm_get_page_prot(unsigned long vm_flags)
85 {
86         return __pgprot(pgprot_val(protection_map[vm_flags &
87                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
88                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
89 }
90 EXPORT_SYMBOL(vm_get_page_prot);
91
92 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
93 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
94 unsigned long sysctl_overcommit_kbytes __read_mostly;
95 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
96 unsigned long sysctl_user_reserve_kbytes __read_mostly = 1UL << 17; /* 128MB */
97 unsigned long sysctl_admin_reserve_kbytes __read_mostly = 1UL << 13; /* 8MB */
98 /*
99  * Make sure vm_committed_as in one cacheline and not cacheline shared with
100  * other variables. It can be updated by several CPUs frequently.
101  */
102 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
103
104 /*
105  * The global memory commitment made in the system can be a metric
106  * that can be used to drive ballooning decisions when Linux is hosted
107  * as a guest. On Hyper-V, the host implements a policy engine for dynamically
108  * balancing memory across competing virtual machines that are hosted.
109  * Several metrics drive this policy engine including the guest reported
110  * memory commitment.
111  */
112 unsigned long vm_memory_committed(void)
113 {
114         return percpu_counter_read_positive(&vm_committed_as);
115 }
116 EXPORT_SYMBOL_GPL(vm_memory_committed);
117
118 /*
119  * Check that a process has enough memory to allocate a new virtual
120  * mapping. 0 means there is enough memory for the allocation to
121  * succeed and -ENOMEM implies there is not.
122  *
123  * We currently support three overcommit policies, which are set via the
124  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
125  *
126  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
127  * Additional code 2002 Jul 20 by Robert Love.
128  *
129  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
130  *
131  * Note this is a helper function intended to be used by LSMs which
132  * wish to use this logic.
133  */
134 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
135 {
136         unsigned long free, allowed, reserve;
137
138         VM_WARN_ONCE(percpu_counter_read(&vm_committed_as) <
139                         -(s64)vm_committed_as_batch * num_online_cpus(),
140                         "memory commitment underflow");
141
142         vm_acct_memory(pages);
143
144         /*
145          * Sometimes we want to use more memory than we have
146          */
147         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
148                 return 0;
149
150         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
151                 free = global_page_state(NR_FREE_PAGES);
152                 free += global_page_state(NR_FILE_PAGES);
153
154                 /*
155                  * shmem pages shouldn't be counted as free in this
156                  * case, they can't be purged, only swapped out, and
157                  * that won't affect the overall amount of available
158                  * memory in the system.
159                  */
160                 free -= global_page_state(NR_SHMEM);
161
162                 free += get_nr_swap_pages();
163
164                 /*
165                  * Any slabs which are created with the
166                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
167                  * which are reclaimable, under pressure.  The dentry
168                  * cache and most inode caches should fall into this
169                  */
170                 free += global_page_state(NR_SLAB_RECLAIMABLE);
171
172                 /*
173                  * Leave reserved pages. The pages are not for anonymous pages.
174                  */
175                 if (free <= totalreserve_pages)
176                         goto error;
177                 else
178                         free -= totalreserve_pages;
179
180                 /*
181                  * Reserve some for root
182                  */
183                 if (!cap_sys_admin)
184                         free -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
185
186                 if (free > pages)
187                         return 0;
188
189                 goto error;
190         }
191
192         allowed = vm_commit_limit();
193         /*
194          * Reserve some for root
195          */
196         if (!cap_sys_admin)
197                 allowed -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
198
199         /*
200          * Don't let a single process grow so big a user can't recover
201          */
202         if (mm) {
203                 reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10);
204                 allowed -= min(mm->total_vm / 32, reserve);
205         }
206
207         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
208                 return 0;
209 error:
210         vm_unacct_memory(pages);
211
212         return -ENOMEM;
213 }
214
215 /*
216  * Requires inode->i_mapping->i_mmap_mutex
217  */
218 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
219                 struct file *file, struct address_space *mapping)
220 {
221         if (vma->vm_flags & VM_DENYWRITE)
222                 atomic_inc(&file_inode(file)->i_writecount);
223         if (vma->vm_flags & VM_SHARED)
224                 mapping_unmap_writable(mapping);
225
226         flush_dcache_mmap_lock(mapping);
227         if (unlikely(vma->vm_flags & VM_NONLINEAR))
228                 list_del_init(&vma->shared.nonlinear);
229         else
230                 vma_interval_tree_remove(vma, &mapping->i_mmap);
231         flush_dcache_mmap_unlock(mapping);
232 }
233
234 /*
235  * Unlink a file-based vm structure from its interval tree, to hide
236  * vma from rmap and vmtruncate before freeing its page tables.
237  */
238 void unlink_file_vma(struct vm_area_struct *vma)
239 {
240         struct file *file = vma->vm_file;
241
242         if (file) {
243                 struct address_space *mapping = file->f_mapping;
244                 mutex_lock(&mapping->i_mmap_mutex);
245                 __remove_shared_vm_struct(vma, file, mapping);
246                 mutex_unlock(&mapping->i_mmap_mutex);
247         }
248 }
249
250 /*
251  * Close a vm structure and free it, returning the next.
252  */
253 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
254 {
255         struct vm_area_struct *next = vma->vm_next;
256
257         might_sleep();
258         if (vma->vm_ops && vma->vm_ops->close)
259                 vma->vm_ops->close(vma);
260         if (vma->vm_file)
261                 fput(vma->vm_file);
262         mpol_put(vma_policy(vma));
263         kmem_cache_free(vm_area_cachep, vma);
264         return next;
265 }
266
267 static unsigned long do_brk(unsigned long addr, unsigned long len);
268
269 SYSCALL_DEFINE1(brk, unsigned long, brk)
270 {
271         unsigned long rlim, retval;
272         unsigned long newbrk, oldbrk;
273         struct mm_struct *mm = current->mm;
274         unsigned long min_brk;
275         bool populate;
276
277         down_write(&mm->mmap_sem);
278
279 #ifdef CONFIG_COMPAT_BRK
280         /*
281          * CONFIG_COMPAT_BRK can still be overridden by setting
282          * randomize_va_space to 2, which will still cause mm->start_brk
283          * to be arbitrarily shifted
284          */
285         if (current->brk_randomized)
286                 min_brk = mm->start_brk;
287         else
288                 min_brk = mm->end_data;
289 #else
290         min_brk = mm->start_brk;
291 #endif
292         if (brk < min_brk)
293                 goto out;
294
295         /*
296          * Check against rlimit here. If this check is done later after the test
297          * of oldbrk with newbrk then it can escape the test and let the data
298          * segment grow beyond its set limit the in case where the limit is
299          * not page aligned -Ram Gupta
300          */
301         rlim = rlimit(RLIMIT_DATA);
302         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
303                         (mm->end_data - mm->start_data) > rlim)
304                 goto out;
305
306         newbrk = PAGE_ALIGN(brk);
307         oldbrk = PAGE_ALIGN(mm->brk);
308         if (oldbrk == newbrk)
309                 goto set_brk;
310
311         /* Always allow shrinking brk. */
312         if (brk <= mm->brk) {
313                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
314                         goto set_brk;
315                 goto out;
316         }
317
318         /* Check against existing mmap mappings. */
319         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
320                 goto out;
321
322         /* Ok, looks good - let it rip. */
323         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
324                 goto out;
325
326 set_brk:
327         mm->brk = brk;
328         populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
329         up_write(&mm->mmap_sem);
330         if (populate)
331                 mm_populate(oldbrk, newbrk - oldbrk);
332         return brk;
333
334 out:
335         retval = mm->brk;
336         up_write(&mm->mmap_sem);
337         return retval;
338 }
339
340 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
341 {
342         unsigned long max, subtree_gap;
343         max = vma->vm_start;
344         if (vma->vm_prev)
345                 max -= vma->vm_prev->vm_end;
346         if (vma->vm_rb.rb_left) {
347                 subtree_gap = rb_entry(vma->vm_rb.rb_left,
348                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
349                 if (subtree_gap > max)
350                         max = subtree_gap;
351         }
352         if (vma->vm_rb.rb_right) {
353                 subtree_gap = rb_entry(vma->vm_rb.rb_right,
354                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
355                 if (subtree_gap > max)
356                         max = subtree_gap;
357         }
358         return max;
359 }
360
361 #ifdef CONFIG_DEBUG_VM_RB
362 static int browse_rb(struct rb_root *root)
363 {
364         int i = 0, j, bug = 0;
365         struct rb_node *nd, *pn = NULL;
366         unsigned long prev = 0, pend = 0;
367
368         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
369                 struct vm_area_struct *vma;
370                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
371                 if (vma->vm_start < prev) {
372                         pr_info("vm_start %lx prev %lx\n", vma->vm_start, prev);
373                         bug = 1;
374                 }
375                 if (vma->vm_start < pend) {
376                         pr_info("vm_start %lx pend %lx\n", vma->vm_start, pend);
377                         bug = 1;
378                 }
379                 if (vma->vm_start > vma->vm_end) {
380                         pr_info("vm_end %lx < vm_start %lx\n",
381                                 vma->vm_end, vma->vm_start);
382                         bug = 1;
383                 }
384                 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
385                         pr_info("free gap %lx, correct %lx\n",
386                                vma->rb_subtree_gap,
387                                vma_compute_subtree_gap(vma));
388                         bug = 1;
389                 }
390                 i++;
391                 pn = nd;
392                 prev = vma->vm_start;
393                 pend = vma->vm_end;
394         }
395         j = 0;
396         for (nd = pn; nd; nd = rb_prev(nd))
397                 j++;
398         if (i != j) {
399                 pr_info("backwards %d, forwards %d\n", j, i);
400                 bug = 1;
401         }
402         return bug ? -1 : i;
403 }
404
405 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
406 {
407         struct rb_node *nd;
408
409         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
410                 struct vm_area_struct *vma;
411                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
412                 BUG_ON(vma != ignore &&
413                        vma->rb_subtree_gap != vma_compute_subtree_gap(vma));
414         }
415 }
416
417 static void validate_mm(struct mm_struct *mm)
418 {
419         int bug = 0;
420         int i = 0;
421         unsigned long highest_address = 0;
422         struct vm_area_struct *vma = mm->mmap;
423         while (vma) {
424                 struct anon_vma_chain *avc;
425                 vma_lock_anon_vma(vma);
426                 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
427                         anon_vma_interval_tree_verify(avc);
428                 vma_unlock_anon_vma(vma);
429                 highest_address = vma->vm_end;
430                 vma = vma->vm_next;
431                 i++;
432         }
433         if (i != mm->map_count) {
434                 pr_info("map_count %d vm_next %d\n", mm->map_count, i);
435                 bug = 1;
436         }
437         if (highest_address != mm->highest_vm_end) {
438                 pr_info("mm->highest_vm_end %lx, found %lx\n",
439                        mm->highest_vm_end, highest_address);
440                 bug = 1;
441         }
442         i = browse_rb(&mm->mm_rb);
443         if (i != mm->map_count) {
444                 pr_info("map_count %d rb %d\n", mm->map_count, i);
445                 bug = 1;
446         }
447         BUG_ON(bug);
448 }
449 #else
450 #define validate_mm_rb(root, ignore) do { } while (0)
451 #define validate_mm(mm) do { } while (0)
452 #endif
453
454 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
455                      unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
456
457 /*
458  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
459  * vma->vm_prev->vm_end values changed, without modifying the vma's position
460  * in the rbtree.
461  */
462 static void vma_gap_update(struct vm_area_struct *vma)
463 {
464         /*
465          * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
466          * function that does exacltly what we want.
467          */
468         vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
469 }
470
471 static inline void vma_rb_insert(struct vm_area_struct *vma,
472                                  struct rb_root *root)
473 {
474         /* All rb_subtree_gap values must be consistent prior to insertion */
475         validate_mm_rb(root, NULL);
476
477         rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
478 }
479
480 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
481 {
482         /*
483          * All rb_subtree_gap values must be consistent prior to erase,
484          * with the possible exception of the vma being erased.
485          */
486         validate_mm_rb(root, vma);
487
488         /*
489          * Note rb_erase_augmented is a fairly large inline function,
490          * so make sure we instantiate it only once with our desired
491          * augmented rbtree callbacks.
492          */
493         rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
494 }
495
496 /*
497  * vma has some anon_vma assigned, and is already inserted on that
498  * anon_vma's interval trees.
499  *
500  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
501  * vma must be removed from the anon_vma's interval trees using
502  * anon_vma_interval_tree_pre_update_vma().
503  *
504  * After the update, the vma will be reinserted using
505  * anon_vma_interval_tree_post_update_vma().
506  *
507  * The entire update must be protected by exclusive mmap_sem and by
508  * the root anon_vma's mutex.
509  */
510 static inline void
511 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
512 {
513         struct anon_vma_chain *avc;
514
515         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
516                 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
517 }
518
519 static inline void
520 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
521 {
522         struct anon_vma_chain *avc;
523
524         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
525                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
526 }
527
528 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
529                 unsigned long end, struct vm_area_struct **pprev,
530                 struct rb_node ***rb_link, struct rb_node **rb_parent)
531 {
532         struct rb_node **__rb_link, *__rb_parent, *rb_prev;
533
534         __rb_link = &mm->mm_rb.rb_node;
535         rb_prev = __rb_parent = NULL;
536
537         while (*__rb_link) {
538                 struct vm_area_struct *vma_tmp;
539
540                 __rb_parent = *__rb_link;
541                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
542
543                 if (vma_tmp->vm_end > addr) {
544                         /* Fail if an existing vma overlaps the area */
545                         if (vma_tmp->vm_start < end)
546                                 return -ENOMEM;
547                         __rb_link = &__rb_parent->rb_left;
548                 } else {
549                         rb_prev = __rb_parent;
550                         __rb_link = &__rb_parent->rb_right;
551                 }
552         }
553
554         *pprev = NULL;
555         if (rb_prev)
556                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
557         *rb_link = __rb_link;
558         *rb_parent = __rb_parent;
559         return 0;
560 }
561
562 static unsigned long count_vma_pages_range(struct mm_struct *mm,
563                 unsigned long addr, unsigned long end)
564 {
565         unsigned long nr_pages = 0;
566         struct vm_area_struct *vma;
567
568         /* Find first overlaping mapping */
569         vma = find_vma_intersection(mm, addr, end);
570         if (!vma)
571                 return 0;
572
573         nr_pages = (min(end, vma->vm_end) -
574                 max(addr, vma->vm_start)) >> PAGE_SHIFT;
575
576         /* Iterate over the rest of the overlaps */
577         for (vma = vma->vm_next; vma; vma = vma->vm_next) {
578                 unsigned long overlap_len;
579
580                 if (vma->vm_start > end)
581                         break;
582
583                 overlap_len = min(end, vma->vm_end) - vma->vm_start;
584                 nr_pages += overlap_len >> PAGE_SHIFT;
585         }
586
587         return nr_pages;
588 }
589
590 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
591                 struct rb_node **rb_link, struct rb_node *rb_parent)
592 {
593         /* Update tracking information for the gap following the new vma. */
594         if (vma->vm_next)
595                 vma_gap_update(vma->vm_next);
596         else
597                 mm->highest_vm_end = vma->vm_end;
598
599         /*
600          * vma->vm_prev wasn't known when we followed the rbtree to find the
601          * correct insertion point for that vma. As a result, we could not
602          * update the vma vm_rb parents rb_subtree_gap values on the way down.
603          * So, we first insert the vma with a zero rb_subtree_gap value
604          * (to be consistent with what we did on the way down), and then
605          * immediately update the gap to the correct value. Finally we
606          * rebalance the rbtree after all augmented values have been set.
607          */
608         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
609         vma->rb_subtree_gap = 0;
610         vma_gap_update(vma);
611         vma_rb_insert(vma, &mm->mm_rb);
612 }
613
614 static void __vma_link_file(struct vm_area_struct *vma)
615 {
616         struct file *file;
617
618         file = vma->vm_file;
619         if (file) {
620                 struct address_space *mapping = file->f_mapping;
621
622                 if (vma->vm_flags & VM_DENYWRITE)
623                         atomic_dec(&file_inode(file)->i_writecount);
624                 if (vma->vm_flags & VM_SHARED)
625                         atomic_inc(&mapping->i_mmap_writable);
626
627                 flush_dcache_mmap_lock(mapping);
628                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
629                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
630                 else
631                         vma_interval_tree_insert(vma, &mapping->i_mmap);
632                 flush_dcache_mmap_unlock(mapping);
633         }
634 }
635
636 static void
637 __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         __vma_link_list(mm, vma, prev, rb_parent);
642         __vma_link_rb(mm, vma, rb_link, rb_parent);
643 }
644
645 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
646                         struct vm_area_struct *prev, struct rb_node **rb_link,
647                         struct rb_node *rb_parent)
648 {
649         struct address_space *mapping = NULL;
650
651         if (vma->vm_file) {
652                 mapping = vma->vm_file->f_mapping;
653                 mutex_lock(&mapping->i_mmap_mutex);
654         }
655
656         __vma_link(mm, vma, prev, rb_link, rb_parent);
657         __vma_link_file(vma);
658
659         if (mapping)
660                 mutex_unlock(&mapping->i_mmap_mutex);
661
662         mm->map_count++;
663         validate_mm(mm);
664 }
665
666 /*
667  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
668  * mm's list and rbtree.  It has already been inserted into the interval tree.
669  */
670 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
671 {
672         struct vm_area_struct *prev;
673         struct rb_node **rb_link, *rb_parent;
674
675         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
676                            &prev, &rb_link, &rb_parent))
677                 BUG();
678         __vma_link(mm, vma, prev, rb_link, rb_parent);
679         mm->map_count++;
680 }
681
682 static inline void
683 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
684                 struct vm_area_struct *prev)
685 {
686         struct vm_area_struct *next;
687
688         vma_rb_erase(vma, &mm->mm_rb);
689         prev->vm_next = next = vma->vm_next;
690         if (next)
691                 next->vm_prev = prev;
692
693         /* Kill the cache */
694         vmacache_invalidate(mm);
695 }
696
697 /*
698  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
699  * is already present in an i_mmap tree without adjusting the tree.
700  * The following helper function should be used when such adjustments
701  * are necessary.  The "insert" vma (if any) is to be inserted
702  * before we drop the necessary locks.
703  */
704 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
705         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
706 {
707         struct mm_struct *mm = vma->vm_mm;
708         struct vm_area_struct *next = vma->vm_next;
709         struct vm_area_struct *importer = NULL;
710         struct address_space *mapping = NULL;
711         struct rb_root *root = NULL;
712         struct anon_vma *anon_vma = NULL;
713         struct file *file = vma->vm_file;
714         bool start_changed = false, end_changed = false;
715         long adjust_next = 0;
716         int remove_next = 0;
717
718         if (next && !insert) {
719                 struct vm_area_struct *exporter = NULL;
720
721                 if (end >= next->vm_end) {
722                         /*
723                          * vma expands, overlapping all the next, and
724                          * perhaps the one after too (mprotect case 6).
725                          */
726 again:                  remove_next = 1 + (end > next->vm_end);
727                         end = next->vm_end;
728                         exporter = next;
729                         importer = vma;
730                 } else if (end > next->vm_start) {
731                         /*
732                          * vma expands, overlapping part of the next:
733                          * mprotect case 5 shifting the boundary up.
734                          */
735                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
736                         exporter = next;
737                         importer = vma;
738                 } else if (end < vma->vm_end) {
739                         /*
740                          * vma shrinks, and !insert tells it's not
741                          * split_vma inserting another: so it must be
742                          * mprotect case 4 shifting the boundary down.
743                          */
744                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
745                         exporter = vma;
746                         importer = next;
747                 }
748
749                 /*
750                  * Easily overlooked: when mprotect shifts the boundary,
751                  * make sure the expanding vma has anon_vma set if the
752                  * shrinking vma had, to cover any anon pages imported.
753                  */
754                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
755                         if (anon_vma_clone(importer, exporter))
756                                 return -ENOMEM;
757                         importer->anon_vma = exporter->anon_vma;
758                 }
759         }
760
761         if (file) {
762                 mapping = file->f_mapping;
763                 if (!(vma->vm_flags & VM_NONLINEAR)) {
764                         root = &mapping->i_mmap;
765                         uprobe_munmap(vma, vma->vm_start, vma->vm_end);
766
767                         if (adjust_next)
768                                 uprobe_munmap(next, next->vm_start,
769                                                         next->vm_end);
770                 }
771
772                 mutex_lock(&mapping->i_mmap_mutex);
773                 if (insert) {
774                         /*
775                          * Put into interval tree now, so instantiated pages
776                          * are visible to arm/parisc __flush_dcache_page
777                          * throughout; but we cannot insert into address
778                          * space until vma start or end is updated.
779                          */
780                         __vma_link_file(insert);
781                 }
782         }
783
784         vma_adjust_trans_huge(vma, start, end, adjust_next);
785
786         anon_vma = vma->anon_vma;
787         if (!anon_vma && adjust_next)
788                 anon_vma = next->anon_vma;
789         if (anon_vma) {
790                 VM_BUG_ON(adjust_next && next->anon_vma &&
791                           anon_vma != next->anon_vma);
792                 anon_vma_lock_write(anon_vma);
793                 anon_vma_interval_tree_pre_update_vma(vma);
794                 if (adjust_next)
795                         anon_vma_interval_tree_pre_update_vma(next);
796         }
797
798         if (root) {
799                 flush_dcache_mmap_lock(mapping);
800                 vma_interval_tree_remove(vma, root);
801                 if (adjust_next)
802                         vma_interval_tree_remove(next, root);
803         }
804
805         if (start != vma->vm_start) {
806                 vma->vm_start = start;
807                 start_changed = true;
808         }
809         if (end != vma->vm_end) {
810                 vma->vm_end = end;
811                 end_changed = true;
812         }
813         vma->vm_pgoff = pgoff;
814         if (adjust_next) {
815                 next->vm_start += adjust_next << PAGE_SHIFT;
816                 next->vm_pgoff += adjust_next;
817         }
818
819         if (root) {
820                 if (adjust_next)
821                         vma_interval_tree_insert(next, root);
822                 vma_interval_tree_insert(vma, root);
823                 flush_dcache_mmap_unlock(mapping);
824         }
825
826         if (remove_next) {
827                 /*
828                  * vma_merge has merged next into vma, and needs
829                  * us to remove next before dropping the locks.
830                  */
831                 __vma_unlink(mm, next, vma);
832                 if (file)
833                         __remove_shared_vm_struct(next, file, mapping);
834         } else if (insert) {
835                 /*
836                  * split_vma has split insert from vma, and needs
837                  * us to insert it before dropping the locks
838                  * (it may either follow vma or precede it).
839                  */
840                 __insert_vm_struct(mm, insert);
841         } else {
842                 if (start_changed)
843                         vma_gap_update(vma);
844                 if (end_changed) {
845                         if (!next)
846                                 mm->highest_vm_end = end;
847                         else if (!adjust_next)
848                                 vma_gap_update(next);
849                 }
850         }
851
852         if (anon_vma) {
853                 anon_vma_interval_tree_post_update_vma(vma);
854                 if (adjust_next)
855                         anon_vma_interval_tree_post_update_vma(next);
856                 anon_vma_unlock_write(anon_vma);
857         }
858         if (mapping)
859                 mutex_unlock(&mapping->i_mmap_mutex);
860
861         if (root) {
862                 uprobe_mmap(vma);
863
864                 if (adjust_next)
865                         uprobe_mmap(next);
866         }
867
868         if (remove_next) {
869                 if (file) {
870                         uprobe_munmap(next, next->vm_start, next->vm_end);
871                         fput(file);
872                 }
873                 if (next->anon_vma)
874                         anon_vma_merge(vma, next);
875                 mm->map_count--;
876                 mpol_put(vma_policy(next));
877                 kmem_cache_free(vm_area_cachep, next);
878                 /*
879                  * In mprotect's case 6 (see comments on vma_merge),
880                  * we must remove another next too. It would clutter
881                  * up the code too much to do both in one go.
882                  */
883                 next = vma->vm_next;
884                 if (remove_next == 2)
885                         goto again;
886                 else if (next)
887                         vma_gap_update(next);
888                 else
889                         mm->highest_vm_end = end;
890         }
891         if (insert && file)
892                 uprobe_mmap(insert);
893
894         validate_mm(mm);
895
896         return 0;
897 }
898
899 /*
900  * If the vma has a ->close operation then the driver probably needs to release
901  * per-vma resources, so we don't attempt to merge those.
902  */
903 static inline int is_mergeable_vma(struct vm_area_struct *vma,
904                         struct file *file, unsigned long vm_flags)
905 {
906         /*
907          * VM_SOFTDIRTY should not prevent from VMA merging, if we
908          * match the flags but dirty bit -- the caller should mark
909          * merged VMA as dirty. If dirty bit won't be excluded from
910          * comparison, we increase pressue on the memory system forcing
911          * the kernel to generate new VMAs when old one could be
912          * extended instead.
913          */
914         if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
915                 return 0;
916         if (vma->vm_file != file)
917                 return 0;
918         if (vma->vm_ops && vma->vm_ops->close)
919                 return 0;
920         return 1;
921 }
922
923 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
924                                         struct anon_vma *anon_vma2,
925                                         struct vm_area_struct *vma)
926 {
927         /*
928          * The list_is_singular() test is to avoid merging VMA cloned from
929          * parents. This can improve scalability caused by anon_vma lock.
930          */
931         if ((!anon_vma1 || !anon_vma2) && (!vma ||
932                 list_is_singular(&vma->anon_vma_chain)))
933                 return 1;
934         return anon_vma1 == anon_vma2;
935 }
936
937 /*
938  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
939  * in front of (at a lower virtual address and file offset than) the vma.
940  *
941  * We cannot merge two vmas if they have differently assigned (non-NULL)
942  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
943  *
944  * We don't check here for the merged mmap wrapping around the end of pagecache
945  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
946  * wrap, nor mmaps which cover the final page at index -1UL.
947  */
948 static int
949 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
950         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
951 {
952         if (is_mergeable_vma(vma, file, vm_flags) &&
953             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
954                 if (vma->vm_pgoff == vm_pgoff)
955                         return 1;
956         }
957         return 0;
958 }
959
960 /*
961  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
962  * beyond (at a higher virtual address and file offset than) the vma.
963  *
964  * We cannot merge two vmas if they have differently assigned (non-NULL)
965  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
966  */
967 static int
968 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
969         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
970 {
971         if (is_mergeable_vma(vma, file, vm_flags) &&
972             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
973                 pgoff_t vm_pglen;
974                 vm_pglen = vma_pages(vma);
975                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
976                         return 1;
977         }
978         return 0;
979 }
980
981 /*
982  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
983  * whether that can be merged with its predecessor or its successor.
984  * Or both (it neatly fills a hole).
985  *
986  * In most cases - when called for mmap, brk or mremap - [addr,end) is
987  * certain not to be mapped by the time vma_merge is called; but when
988  * called for mprotect, it is certain to be already mapped (either at
989  * an offset within prev, or at the start of next), and the flags of
990  * this area are about to be changed to vm_flags - and the no-change
991  * case has already been eliminated.
992  *
993  * The following mprotect cases have to be considered, where AAAA is
994  * the area passed down from mprotect_fixup, never extending beyond one
995  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
996  *
997  *     AAAA             AAAA                AAAA          AAAA
998  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
999  *    cannot merge    might become    might become    might become
1000  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
1001  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
1002  *    mremap move:                                    PPPPNNNNNNNN 8
1003  *        AAAA
1004  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
1005  *    might become    case 1 below    case 2 below    case 3 below
1006  *
1007  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
1008  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
1009  */
1010 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1011                         struct vm_area_struct *prev, unsigned long addr,
1012                         unsigned long end, unsigned long vm_flags,
1013                         struct anon_vma *anon_vma, struct file *file,
1014                         pgoff_t pgoff, struct mempolicy *policy)
1015 {
1016         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1017         struct vm_area_struct *area, *next;
1018         int err;
1019
1020         /*
1021          * We later require that vma->vm_flags == vm_flags,
1022          * so this tests vma->vm_flags & VM_SPECIAL, too.
1023          */
1024         if (vm_flags & VM_SPECIAL)
1025                 return NULL;
1026
1027         if (prev)
1028                 next = prev->vm_next;
1029         else
1030                 next = mm->mmap;
1031         area = next;
1032         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
1033                 next = next->vm_next;
1034
1035         /*
1036          * Can it merge with the predecessor?
1037          */
1038         if (prev && prev->vm_end == addr &&
1039                         mpol_equal(vma_policy(prev), policy) &&
1040                         can_vma_merge_after(prev, vm_flags,
1041                                                 anon_vma, file, pgoff)) {
1042                 /*
1043                  * OK, it can.  Can we now merge in the successor as well?
1044                  */
1045                 if (next && end == next->vm_start &&
1046                                 mpol_equal(policy, vma_policy(next)) &&
1047                                 can_vma_merge_before(next, vm_flags,
1048                                         anon_vma, file, pgoff+pglen) &&
1049                                 is_mergeable_anon_vma(prev->anon_vma,
1050                                                       next->anon_vma, NULL)) {
1051                                                         /* cases 1, 6 */
1052                         err = vma_adjust(prev, prev->vm_start,
1053                                 next->vm_end, prev->vm_pgoff, NULL);
1054                 } else                                  /* cases 2, 5, 7 */
1055                         err = vma_adjust(prev, prev->vm_start,
1056                                 end, prev->vm_pgoff, NULL);
1057                 if (err)
1058                         return NULL;
1059                 khugepaged_enter_vma_merge(prev);
1060                 return prev;
1061         }
1062
1063         /*
1064          * Can this new request be merged in front of next?
1065          */
1066         if (next && end == next->vm_start &&
1067                         mpol_equal(policy, vma_policy(next)) &&
1068                         can_vma_merge_before(next, vm_flags,
1069                                         anon_vma, file, pgoff+pglen)) {
1070                 if (prev && addr < prev->vm_end)        /* case 4 */
1071                         err = vma_adjust(prev, prev->vm_start,
1072                                 addr, prev->vm_pgoff, NULL);
1073                 else                                    /* cases 3, 8 */
1074                         err = vma_adjust(area, addr, next->vm_end,
1075                                 next->vm_pgoff - pglen, NULL);
1076                 if (err)
1077                         return NULL;
1078                 khugepaged_enter_vma_merge(area);
1079                 return area;
1080         }
1081
1082         return NULL;
1083 }
1084
1085 /*
1086  * Rough compatbility check to quickly see if it's even worth looking
1087  * at sharing an anon_vma.
1088  *
1089  * They need to have the same vm_file, and the flags can only differ
1090  * in things that mprotect may change.
1091  *
1092  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1093  * we can merge the two vma's. For example, we refuse to merge a vma if
1094  * there is a vm_ops->close() function, because that indicates that the
1095  * driver is doing some kind of reference counting. But that doesn't
1096  * really matter for the anon_vma sharing case.
1097  */
1098 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1099 {
1100         return a->vm_end == b->vm_start &&
1101                 mpol_equal(vma_policy(a), vma_policy(b)) &&
1102                 a->vm_file == b->vm_file &&
1103                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1104                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1105 }
1106
1107 /*
1108  * Do some basic sanity checking to see if we can re-use the anon_vma
1109  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1110  * the same as 'old', the other will be the new one that is trying
1111  * to share the anon_vma.
1112  *
1113  * NOTE! This runs with mm_sem held for reading, so it is possible that
1114  * the anon_vma of 'old' is concurrently in the process of being set up
1115  * by another page fault trying to merge _that_. But that's ok: if it
1116  * is being set up, that automatically means that it will be a singleton
1117  * acceptable for merging, so we can do all of this optimistically. But
1118  * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
1119  *
1120  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1121  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1122  * is to return an anon_vma that is "complex" due to having gone through
1123  * a fork).
1124  *
1125  * We also make sure that the two vma's are compatible (adjacent,
1126  * and with the same memory policies). That's all stable, even with just
1127  * a read lock on the mm_sem.
1128  */
1129 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1130 {
1131         if (anon_vma_compatible(a, b)) {
1132                 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
1133
1134                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1135                         return anon_vma;
1136         }
1137         return NULL;
1138 }
1139
1140 /*
1141  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1142  * neighbouring vmas for a suitable anon_vma, before it goes off
1143  * to allocate a new anon_vma.  It checks because a repetitive
1144  * sequence of mprotects and faults may otherwise lead to distinct
1145  * anon_vmas being allocated, preventing vma merge in subsequent
1146  * mprotect.
1147  */
1148 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1149 {
1150         struct anon_vma *anon_vma;
1151         struct vm_area_struct *near;
1152
1153         near = vma->vm_next;
1154         if (!near)
1155                 goto try_prev;
1156
1157         anon_vma = reusable_anon_vma(near, vma, near);
1158         if (anon_vma)
1159                 return anon_vma;
1160 try_prev:
1161         near = vma->vm_prev;
1162         if (!near)
1163                 goto none;
1164
1165         anon_vma = reusable_anon_vma(near, near, vma);
1166         if (anon_vma)
1167                 return anon_vma;
1168 none:
1169         /*
1170          * There's no absolute need to look only at touching neighbours:
1171          * we could search further afield for "compatible" anon_vmas.
1172          * But it would probably just be a waste of time searching,
1173          * or lead to too many vmas hanging off the same anon_vma.
1174          * We're trying to allow mprotect remerging later on,
1175          * not trying to minimize memory used for anon_vmas.
1176          */
1177         return NULL;
1178 }
1179
1180 #ifdef CONFIG_PROC_FS
1181 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
1182                                                 struct file *file, long pages)
1183 {
1184         const unsigned long stack_flags
1185                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
1186
1187         mm->total_vm += pages;
1188
1189         if (file) {
1190                 mm->shared_vm += pages;
1191                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
1192                         mm->exec_vm += pages;
1193         } else if (flags & stack_flags)
1194                 mm->stack_vm += pages;
1195 }
1196 #endif /* CONFIG_PROC_FS */
1197
1198 /*
1199  * If a hint addr is less than mmap_min_addr change hint to be as
1200  * low as possible but still greater than mmap_min_addr
1201  */
1202 static inline unsigned long round_hint_to_min(unsigned long hint)
1203 {
1204         hint &= PAGE_MASK;
1205         if (((void *)hint != NULL) &&
1206             (hint < mmap_min_addr))
1207                 return PAGE_ALIGN(mmap_min_addr);
1208         return hint;
1209 }
1210
1211 static inline int mlock_future_check(struct mm_struct *mm,
1212                                      unsigned long flags,
1213                                      unsigned long len)
1214 {
1215         unsigned long locked, lock_limit;
1216
1217         /*  mlock MCL_FUTURE? */
1218         if (flags & VM_LOCKED) {
1219                 locked = len >> PAGE_SHIFT;
1220                 locked += mm->locked_vm;
1221                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1222                 lock_limit >>= PAGE_SHIFT;
1223                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1224                         return -EAGAIN;
1225         }
1226         return 0;
1227 }
1228
1229 /*
1230  * The caller must hold down_write(&current->mm->mmap_sem).
1231  */
1232
1233 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
1234                         unsigned long len, unsigned long prot,
1235                         unsigned long flags, unsigned long pgoff,
1236                         unsigned long *populate)
1237 {
1238         struct mm_struct * mm = current->mm;
1239         vm_flags_t vm_flags;
1240
1241         *populate = 0;
1242
1243         /*
1244          * Does the application expect PROT_READ to imply PROT_EXEC?
1245          *
1246          * (the exception is when the underlying filesystem is noexec
1247          *  mounted, in which case we dont add PROT_EXEC.)
1248          */
1249         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1250                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
1251                         prot |= PROT_EXEC;
1252
1253         if (!len)
1254                 return -EINVAL;
1255
1256         if (!(flags & MAP_FIXED))
1257                 addr = round_hint_to_min(addr);
1258
1259         /* Careful about overflows.. */
1260         len = PAGE_ALIGN(len);
1261         if (!len)
1262                 return -ENOMEM;
1263
1264         /* offset overflow? */
1265         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1266                return -EOVERFLOW;
1267
1268         /* Too many mappings? */
1269         if (mm->map_count > sysctl_max_map_count)
1270                 return -ENOMEM;
1271
1272         /* Obtain the address to map to. we verify (or select) it and ensure
1273          * that it represents a valid section of the address space.
1274          */
1275         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1276         if (addr & ~PAGE_MASK)
1277                 return addr;
1278
1279         /* Do simple checking here so the lower-level routines won't have
1280          * to. we assume access permissions have been handled by the open
1281          * of the memory object, so we don't do any here.
1282          */
1283         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1284                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1285
1286         if (flags & MAP_LOCKED)
1287                 if (!can_do_mlock())
1288                         return -EPERM;
1289
1290         if (mlock_future_check(mm, vm_flags, len))
1291                 return -EAGAIN;
1292
1293         if (file) {
1294                 struct inode *inode = file_inode(file);
1295
1296                 switch (flags & MAP_TYPE) {
1297                 case MAP_SHARED:
1298                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1299                                 return -EACCES;
1300
1301                         /*
1302                          * Make sure we don't allow writing to an append-only
1303                          * file..
1304                          */
1305                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1306                                 return -EACCES;
1307
1308                         /*
1309                          * Make sure there are no mandatory locks on the file.
1310                          */
1311                         if (locks_verify_locked(file))
1312                                 return -EAGAIN;
1313
1314                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1315                         if (!(file->f_mode & FMODE_WRITE))
1316                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1317
1318                         /* fall through */
1319                 case MAP_PRIVATE:
1320                         if (!(file->f_mode & FMODE_READ))
1321                                 return -EACCES;
1322                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1323                                 if (vm_flags & VM_EXEC)
1324                                         return -EPERM;
1325                                 vm_flags &= ~VM_MAYEXEC;
1326                         }
1327
1328                         if (!file->f_op->mmap)
1329                                 return -ENODEV;
1330                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1331                                 return -EINVAL;
1332                         break;
1333
1334                 default:
1335                         return -EINVAL;
1336                 }
1337         } else {
1338                 switch (flags & MAP_TYPE) {
1339                 case MAP_SHARED:
1340                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1341                                 return -EINVAL;
1342                         /*
1343                          * Ignore pgoff.
1344                          */
1345                         pgoff = 0;
1346                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1347                         break;
1348                 case MAP_PRIVATE:
1349                         /*
1350                          * Set pgoff according to addr for anon_vma.
1351                          */
1352                         pgoff = addr >> PAGE_SHIFT;
1353                         break;
1354                 default:
1355                         return -EINVAL;
1356                 }
1357         }
1358
1359         /*
1360          * Set 'VM_NORESERVE' if we should not account for the
1361          * memory use of this mapping.
1362          */
1363         if (flags & MAP_NORESERVE) {
1364                 /* We honor MAP_NORESERVE if allowed to overcommit */
1365                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1366                         vm_flags |= VM_NORESERVE;
1367
1368                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1369                 if (file && is_file_hugepages(file))
1370                         vm_flags |= VM_NORESERVE;
1371         }
1372
1373         addr = mmap_region(file, addr, len, vm_flags, pgoff);
1374         if (!IS_ERR_VALUE(addr) &&
1375             ((vm_flags & VM_LOCKED) ||
1376              (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1377                 *populate = len;
1378         return addr;
1379 }
1380
1381 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1382                 unsigned long, prot, unsigned long, flags,
1383                 unsigned long, fd, unsigned long, pgoff)
1384 {
1385         struct file *file = NULL;
1386         unsigned long retval = -EBADF;
1387
1388         if (!(flags & MAP_ANONYMOUS)) {
1389                 audit_mmap_fd(fd, flags);
1390                 file = fget(fd);
1391                 if (!file)
1392                         goto out;
1393                 if (is_file_hugepages(file))
1394                         len = ALIGN(len, huge_page_size(hstate_file(file)));
1395                 retval = -EINVAL;
1396                 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1397                         goto out_fput;
1398         } else if (flags & MAP_HUGETLB) {
1399                 struct user_struct *user = NULL;
1400                 struct hstate *hs;
1401
1402                 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1403                 if (!hs)
1404                         return -EINVAL;
1405
1406                 len = ALIGN(len, huge_page_size(hs));
1407                 /*
1408                  * VM_NORESERVE is used because the reservations will be
1409                  * taken when vm_ops->mmap() is called
1410                  * A dummy user value is used because we are not locking
1411                  * memory so no accounting is necessary
1412                  */
1413                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1414                                 VM_NORESERVE,
1415                                 &user, HUGETLB_ANONHUGE_INODE,
1416                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1417                 if (IS_ERR(file))
1418                         return PTR_ERR(file);
1419         }
1420
1421         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1422
1423         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1424 out_fput:
1425         if (file)
1426                 fput(file);
1427 out:
1428         return retval;
1429 }
1430
1431 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1432 struct mmap_arg_struct {
1433         unsigned long addr;
1434         unsigned long len;
1435         unsigned long prot;
1436         unsigned long flags;
1437         unsigned long fd;
1438         unsigned long offset;
1439 };
1440
1441 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1442 {
1443         struct mmap_arg_struct a;
1444
1445         if (copy_from_user(&a, arg, sizeof(a)))
1446                 return -EFAULT;
1447         if (a.offset & ~PAGE_MASK)
1448                 return -EINVAL;
1449
1450         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1451                               a.offset >> PAGE_SHIFT);
1452 }
1453 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1454
1455 /*
1456  * Some shared mappigns will want the pages marked read-only
1457  * to track write events. If so, we'll downgrade vm_page_prot
1458  * to the private version (using protection_map[] without the
1459  * VM_SHARED bit).
1460  */
1461 int vma_wants_writenotify(struct vm_area_struct *vma)
1462 {
1463         vm_flags_t vm_flags = vma->vm_flags;
1464
1465         /* If it was private or non-writable, the write bit is already clear */
1466         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1467                 return 0;
1468
1469         /* The backer wishes to know when pages are first written to? */
1470         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1471                 return 1;
1472
1473         /* The open routine did something to the protections already? */
1474         if (pgprot_val(vma->vm_page_prot) !=
1475             pgprot_val(vm_get_page_prot(vm_flags)))
1476                 return 0;
1477
1478         /* Specialty mapping? */
1479         if (vm_flags & VM_PFNMAP)
1480                 return 0;
1481
1482         /* Can the mapping track the dirty pages? */
1483         return vma->vm_file && vma->vm_file->f_mapping &&
1484                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1485 }
1486
1487 /*
1488  * We account for memory if it's a private writeable mapping,
1489  * not hugepages and VM_NORESERVE wasn't set.
1490  */
1491 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1492 {
1493         /*
1494          * hugetlb has its own accounting separate from the core VM
1495          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1496          */
1497         if (file && is_file_hugepages(file))
1498                 return 0;
1499
1500         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1501 }
1502
1503 unsigned long mmap_region(struct file *file, unsigned long addr,
1504                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1505 {
1506         struct mm_struct *mm = current->mm;
1507         struct vm_area_struct *vma, *prev;
1508         int error;
1509         struct rb_node **rb_link, *rb_parent;
1510         unsigned long charged = 0;
1511
1512         /* Check against address space limit. */
1513         if (!may_expand_vm(mm, len >> PAGE_SHIFT)) {
1514                 unsigned long nr_pages;
1515
1516                 /*
1517                  * MAP_FIXED may remove pages of mappings that intersects with
1518                  * requested mapping. Account for the pages it would unmap.
1519                  */
1520                 if (!(vm_flags & MAP_FIXED))
1521                         return -ENOMEM;
1522
1523                 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1524
1525                 if (!may_expand_vm(mm, (len >> PAGE_SHIFT) - nr_pages))
1526                         return -ENOMEM;
1527         }
1528
1529         /* Clear old maps */
1530         error = -ENOMEM;
1531 munmap_back:
1532         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
1533                 if (do_munmap(mm, addr, len))
1534                         return -ENOMEM;
1535                 goto munmap_back;
1536         }
1537
1538         /*
1539          * Private writable mapping: check memory availability
1540          */
1541         if (accountable_mapping(file, vm_flags)) {
1542                 charged = len >> PAGE_SHIFT;
1543                 if (security_vm_enough_memory_mm(mm, charged))
1544                         return -ENOMEM;
1545                 vm_flags |= VM_ACCOUNT;
1546         }
1547
1548         /*
1549          * Can we just expand an old mapping?
1550          */
1551         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1552         if (vma)
1553                 goto out;
1554
1555         /*
1556          * Determine the object being mapped and call the appropriate
1557          * specific mapper. the address has already been validated, but
1558          * not unmapped, but the maps are removed from the list.
1559          */
1560         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1561         if (!vma) {
1562                 error = -ENOMEM;
1563                 goto unacct_error;
1564         }
1565
1566         vma->vm_mm = mm;
1567         vma->vm_start = addr;
1568         vma->vm_end = addr + len;
1569         vma->vm_flags = vm_flags;
1570         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1571         vma->vm_pgoff = pgoff;
1572         INIT_LIST_HEAD(&vma->anon_vma_chain);
1573
1574         if (file) {
1575                 if (vm_flags & VM_DENYWRITE) {
1576                         error = deny_write_access(file);
1577                         if (error)
1578                                 goto free_vma;
1579                 }
1580                 if (vm_flags & VM_SHARED) {
1581                         error = mapping_map_writable(file->f_mapping);
1582                         if (error)
1583                                 goto allow_write_and_free_vma;
1584                 }
1585
1586                 /* ->mmap() can change vma->vm_file, but must guarantee that
1587                  * vma_link() below can deny write-access if VM_DENYWRITE is set
1588                  * and map writably if VM_SHARED is set. This usually means the
1589                  * new file must not have been exposed to user-space, yet.
1590                  */
1591                 vma->vm_file = get_file(file);
1592                 error = file->f_op->mmap(file, vma);
1593                 if (error)
1594                         goto unmap_and_free_vma;
1595
1596                 /* Can addr have changed??
1597                  *
1598                  * Answer: Yes, several device drivers can do it in their
1599                  *         f_op->mmap method. -DaveM
1600                  * Bug: If addr is changed, prev, rb_link, rb_parent should
1601                  *      be updated for vma_link()
1602                  */
1603                 WARN_ON_ONCE(addr != vma->vm_start);
1604
1605                 addr = vma->vm_start;
1606                 vm_flags = vma->vm_flags;
1607         } else if (vm_flags & VM_SHARED) {
1608                 error = shmem_zero_setup(vma);
1609                 if (error)
1610                         goto free_vma;
1611         }
1612
1613         if (vma_wants_writenotify(vma)) {
1614                 pgprot_t pprot = vma->vm_page_prot;
1615
1616                 /* Can vma->vm_page_prot have changed??
1617                  *
1618                  * Answer: Yes, drivers may have changed it in their
1619                  *         f_op->mmap method.
1620                  *
1621                  * Ensures that vmas marked as uncached stay that way.
1622                  */
1623                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1624                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1625                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1626         }
1627
1628         vma_link(mm, vma, prev, rb_link, rb_parent);
1629         /* Once vma denies write, undo our temporary denial count */
1630         if (file) {
1631                 if (vm_flags & VM_SHARED)
1632                         mapping_unmap_writable(file->f_mapping);
1633                 if (vm_flags & VM_DENYWRITE)
1634                         allow_write_access(file);
1635         }
1636         file = vma->vm_file;
1637 out:
1638         perf_event_mmap(vma);
1639
1640         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1641         if (vm_flags & VM_LOCKED) {
1642                 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1643                                         vma == get_gate_vma(current->mm)))
1644                         mm->locked_vm += (len >> PAGE_SHIFT);
1645                 else
1646                         vma->vm_flags &= ~VM_LOCKED;
1647         }
1648
1649         if (file)
1650                 uprobe_mmap(vma);
1651
1652         /*
1653          * New (or expanded) vma always get soft dirty status.
1654          * Otherwise user-space soft-dirty page tracker won't
1655          * be able to distinguish situation when vma area unmapped,
1656          * then new mapped in-place (which must be aimed as
1657          * a completely new data area).
1658          */
1659         vma->vm_flags |= VM_SOFTDIRTY;
1660
1661         return addr;
1662
1663 unmap_and_free_vma:
1664         vma->vm_file = NULL;
1665         fput(file);
1666
1667         /* Undo any partial mapping done by a device driver. */
1668         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1669         charged = 0;
1670         if (vm_flags & VM_SHARED)
1671                 mapping_unmap_writable(file->f_mapping);
1672 allow_write_and_free_vma:
1673         if (vm_flags & VM_DENYWRITE)
1674                 allow_write_access(file);
1675 free_vma:
1676         kmem_cache_free(vm_area_cachep, vma);
1677 unacct_error:
1678         if (charged)
1679                 vm_unacct_memory(charged);
1680         return error;
1681 }
1682
1683 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1684 {
1685         /*
1686          * We implement the search by looking for an rbtree node that
1687          * immediately follows a suitable gap. That is,
1688          * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1689          * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1690          * - gap_end - gap_start >= length
1691          */
1692
1693         struct mm_struct *mm = current->mm;
1694         struct vm_area_struct *vma;
1695         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1696
1697         /* Adjust search length to account for worst case alignment overhead */
1698         length = info->length + info->align_mask;
1699         if (length < info->length)
1700                 return -ENOMEM;
1701
1702         /* Adjust search limits by the desired length */
1703         if (info->high_limit < length)
1704                 return -ENOMEM;
1705         high_limit = info->high_limit - length;
1706
1707         if (info->low_limit > high_limit)
1708                 return -ENOMEM;
1709         low_limit = info->low_limit + length;
1710
1711         /* Check if rbtree root looks promising */
1712         if (RB_EMPTY_ROOT(&mm->mm_rb))
1713                 goto check_highest;
1714         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1715         if (vma->rb_subtree_gap < length)
1716                 goto check_highest;
1717
1718         while (true) {
1719                 /* Visit left subtree if it looks promising */
1720                 gap_end = vma->vm_start;
1721                 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1722                         struct vm_area_struct *left =
1723                                 rb_entry(vma->vm_rb.rb_left,
1724                                          struct vm_area_struct, vm_rb);
1725                         if (left->rb_subtree_gap >= length) {
1726                                 vma = left;
1727                                 continue;
1728                         }
1729                 }
1730
1731                 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1732 check_current:
1733                 /* Check if current node has a suitable gap */
1734                 if (gap_start > high_limit)
1735                         return -ENOMEM;
1736                 if (gap_end >= low_limit && gap_end - gap_start >= length)
1737                         goto found;
1738
1739                 /* Visit right subtree if it looks promising */
1740                 if (vma->vm_rb.rb_right) {
1741                         struct vm_area_struct *right =
1742                                 rb_entry(vma->vm_rb.rb_right,
1743                                          struct vm_area_struct, vm_rb);
1744                         if (right->rb_subtree_gap >= length) {
1745                                 vma = right;
1746                                 continue;
1747                         }
1748                 }
1749
1750                 /* Go back up the rbtree to find next candidate node */
1751                 while (true) {
1752                         struct rb_node *prev = &vma->vm_rb;
1753                         if (!rb_parent(prev))
1754                                 goto check_highest;
1755                         vma = rb_entry(rb_parent(prev),
1756                                        struct vm_area_struct, vm_rb);
1757                         if (prev == vma->vm_rb.rb_left) {
1758                                 gap_start = vma->vm_prev->vm_end;
1759                                 gap_end = vma->vm_start;
1760                                 goto check_current;
1761                         }
1762                 }
1763         }
1764
1765 check_highest:
1766         /* Check highest gap, which does not precede any rbtree node */
1767         gap_start = mm->highest_vm_end;
1768         gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1769         if (gap_start > high_limit)
1770                 return -ENOMEM;
1771
1772 found:
1773         /* We found a suitable gap. Clip it with the original low_limit. */
1774         if (gap_start < info->low_limit)
1775                 gap_start = info->low_limit;
1776
1777         /* Adjust gap address to the desired alignment */
1778         gap_start += (info->align_offset - gap_start) & info->align_mask;
1779
1780         VM_BUG_ON(gap_start + info->length > info->high_limit);
1781         VM_BUG_ON(gap_start + info->length > gap_end);
1782         return gap_start;
1783 }
1784
1785 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1786 {
1787         struct mm_struct *mm = current->mm;
1788         struct vm_area_struct *vma;
1789         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1790
1791         /* Adjust search length to account for worst case alignment overhead */
1792         length = info->length + info->align_mask;
1793         if (length < info->length)
1794                 return -ENOMEM;
1795
1796         /*
1797          * Adjust search limits by the desired length.
1798          * See implementation comment at top of unmapped_area().
1799          */
1800         gap_end = info->high_limit;
1801         if (gap_end < length)
1802                 return -ENOMEM;
1803         high_limit = gap_end - length;
1804
1805         if (info->low_limit > high_limit)
1806                 return -ENOMEM;
1807         low_limit = info->low_limit + length;
1808
1809         /* Check highest gap, which does not precede any rbtree node */
1810         gap_start = mm->highest_vm_end;
1811         if (gap_start <= high_limit)
1812                 goto found_highest;
1813
1814         /* Check if rbtree root looks promising */
1815         if (RB_EMPTY_ROOT(&mm->mm_rb))
1816                 return -ENOMEM;
1817         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1818         if (vma->rb_subtree_gap < length)
1819                 return -ENOMEM;
1820
1821         while (true) {
1822                 /* Visit right subtree if it looks promising */
1823                 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1824                 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1825                         struct vm_area_struct *right =
1826                                 rb_entry(vma->vm_rb.rb_right,
1827                                          struct vm_area_struct, vm_rb);
1828                         if (right->rb_subtree_gap >= length) {
1829                                 vma = right;
1830                                 continue;
1831                         }
1832                 }
1833
1834 check_current:
1835                 /* Check if current node has a suitable gap */
1836                 gap_end = vma->vm_start;
1837                 if (gap_end < low_limit)
1838                         return -ENOMEM;
1839                 if (gap_start <= high_limit && gap_end - gap_start >= length)
1840                         goto found;
1841
1842                 /* Visit left subtree if it looks promising */
1843                 if (vma->vm_rb.rb_left) {
1844                         struct vm_area_struct *left =
1845                                 rb_entry(vma->vm_rb.rb_left,
1846                                          struct vm_area_struct, vm_rb);
1847                         if (left->rb_subtree_gap >= length) {
1848                                 vma = left;
1849                                 continue;
1850                         }
1851                 }
1852
1853                 /* Go back up the rbtree to find next candidate node */
1854                 while (true) {
1855                         struct rb_node *prev = &vma->vm_rb;
1856                         if (!rb_parent(prev))
1857                                 return -ENOMEM;
1858                         vma = rb_entry(rb_parent(prev),
1859                                        struct vm_area_struct, vm_rb);
1860                         if (prev == vma->vm_rb.rb_right) {
1861                                 gap_start = vma->vm_prev ?
1862                                         vma->vm_prev->vm_end : 0;
1863                                 goto check_current;
1864                         }
1865                 }
1866         }
1867
1868 found:
1869         /* We found a suitable gap. Clip it with the original high_limit. */
1870         if (gap_end > info->high_limit)
1871                 gap_end = info->high_limit;
1872
1873 found_highest:
1874         /* Compute highest gap address at the desired alignment */
1875         gap_end -= info->length;
1876         gap_end -= (gap_end - info->align_offset) & info->align_mask;
1877
1878         VM_BUG_ON(gap_end < info->low_limit);
1879         VM_BUG_ON(gap_end < gap_start);
1880         return gap_end;
1881 }
1882
1883 /* Get an address range which is currently unmapped.
1884  * For shmat() with addr=0.
1885  *
1886  * Ugly calling convention alert:
1887  * Return value with the low bits set means error value,
1888  * ie
1889  *      if (ret & ~PAGE_MASK)
1890  *              error = ret;
1891  *
1892  * This function "knows" that -ENOMEM has the bits set.
1893  */
1894 #ifndef HAVE_ARCH_UNMAPPED_AREA
1895 unsigned long
1896 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1897                 unsigned long len, unsigned long pgoff, unsigned long flags)
1898 {
1899         struct mm_struct *mm = current->mm;
1900         struct vm_area_struct *vma;
1901         struct vm_unmapped_area_info info;
1902
1903         if (len > TASK_SIZE - mmap_min_addr)
1904                 return -ENOMEM;
1905
1906         if (flags & MAP_FIXED)
1907                 return addr;
1908
1909         if (addr) {
1910                 addr = PAGE_ALIGN(addr);
1911                 vma = find_vma(mm, addr);
1912                 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1913                     (!vma || addr + len <= vma->vm_start))
1914                         return addr;
1915         }
1916
1917         info.flags = 0;
1918         info.length = len;
1919         info.low_limit = mm->mmap_base;
1920         info.high_limit = TASK_SIZE;
1921         info.align_mask = 0;
1922         return vm_unmapped_area(&info);
1923 }
1924 #endif  
1925
1926 /*
1927  * This mmap-allocator allocates new areas top-down from below the
1928  * stack's low limit (the base):
1929  */
1930 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1931 unsigned long
1932 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1933                           const unsigned long len, const unsigned long pgoff,
1934                           const unsigned long flags)
1935 {
1936         struct vm_area_struct *vma;
1937         struct mm_struct *mm = current->mm;
1938         unsigned long addr = addr0;
1939         struct vm_unmapped_area_info info;
1940
1941         /* requested length too big for entire address space */
1942         if (len > TASK_SIZE - mmap_min_addr)
1943                 return -ENOMEM;
1944
1945         if (flags & MAP_FIXED)
1946                 return addr;
1947
1948         /* requesting a specific address */
1949         if (addr) {
1950                 addr = PAGE_ALIGN(addr);
1951                 vma = find_vma(mm, addr);
1952                 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1953                                 (!vma || addr + len <= vma->vm_start))
1954                         return addr;
1955         }
1956
1957         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1958         info.length = len;
1959         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
1960         info.high_limit = mm->mmap_base;
1961         info.align_mask = 0;
1962         addr = vm_unmapped_area(&info);
1963
1964         /*
1965          * A failed mmap() very likely causes application failure,
1966          * so fall back to the bottom-up function here. This scenario
1967          * can happen with large stack limits and large mmap()
1968          * allocations.
1969          */
1970         if (addr & ~PAGE_MASK) {
1971                 VM_BUG_ON(addr != -ENOMEM);
1972                 info.flags = 0;
1973                 info.low_limit = TASK_UNMAPPED_BASE;
1974                 info.high_limit = TASK_SIZE;
1975                 addr = vm_unmapped_area(&info);
1976         }
1977
1978         return addr;
1979 }
1980 #endif
1981
1982 unsigned long
1983 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1984                 unsigned long pgoff, unsigned long flags)
1985 {
1986         unsigned long (*get_area)(struct file *, unsigned long,
1987                                   unsigned long, unsigned long, unsigned long);
1988
1989         unsigned long error = arch_mmap_check(addr, len, flags);
1990         if (error)
1991                 return error;
1992
1993         /* Careful about overflows.. */
1994         if (len > TASK_SIZE)
1995                 return -ENOMEM;
1996
1997         get_area = current->mm->get_unmapped_area;
1998         if (file && file->f_op->get_unmapped_area)
1999                 get_area = file->f_op->get_unmapped_area;
2000         addr = get_area(file, addr, len, pgoff, flags);
2001         if (IS_ERR_VALUE(addr))
2002                 return addr;
2003
2004         if (addr > TASK_SIZE - len)
2005                 return -ENOMEM;
2006         if (addr & ~PAGE_MASK)
2007                 return -EINVAL;
2008
2009         addr = arch_rebalance_pgtables(addr, len);
2010         error = security_mmap_addr(addr);
2011         return error ? error : addr;
2012 }
2013
2014 EXPORT_SYMBOL(get_unmapped_area);
2015
2016 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2017 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2018 {
2019         struct rb_node *rb_node;
2020         struct vm_area_struct *vma;
2021
2022         /* Check the cache first. */
2023         vma = vmacache_find(mm, addr);
2024         if (likely(vma))
2025                 return vma;
2026
2027         rb_node = mm->mm_rb.rb_node;
2028         vma = NULL;
2029
2030         while (rb_node) {
2031                 struct vm_area_struct *tmp;
2032
2033                 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2034
2035                 if (tmp->vm_end > addr) {
2036                         vma = tmp;
2037                         if (tmp->vm_start <= addr)
2038                                 break;
2039                         rb_node = rb_node->rb_left;
2040                 } else
2041                         rb_node = rb_node->rb_right;
2042         }
2043
2044         if (vma)
2045                 vmacache_update(addr, vma);
2046         return vma;
2047 }
2048
2049 EXPORT_SYMBOL(find_vma);
2050
2051 /*
2052  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2053  */
2054 struct vm_area_struct *
2055 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2056                         struct vm_area_struct **pprev)
2057 {
2058         struct vm_area_struct *vma;
2059
2060         vma = find_vma(mm, addr);
2061         if (vma) {
2062                 *pprev = vma->vm_prev;
2063         } else {
2064                 struct rb_node *rb_node = mm->mm_rb.rb_node;
2065                 *pprev = NULL;
2066                 while (rb_node) {
2067                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2068                         rb_node = rb_node->rb_right;
2069                 }
2070         }
2071         return vma;
2072 }
2073
2074 /*
2075  * Verify that the stack growth is acceptable and
2076  * update accounting. This is shared with both the
2077  * grow-up and grow-down cases.
2078  */
2079 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
2080 {
2081         struct mm_struct *mm = vma->vm_mm;
2082         struct rlimit *rlim = current->signal->rlim;
2083         unsigned long new_start;
2084
2085         /* address space limit tests */
2086         if (!may_expand_vm(mm, grow))
2087                 return -ENOMEM;
2088
2089         /* Stack limit test */
2090         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2091                 return -ENOMEM;
2092
2093         /* mlock limit tests */
2094         if (vma->vm_flags & VM_LOCKED) {
2095                 unsigned long locked;
2096                 unsigned long limit;
2097                 locked = mm->locked_vm + grow;
2098                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2099                 limit >>= PAGE_SHIFT;
2100                 if (locked > limit && !capable(CAP_IPC_LOCK))
2101                         return -ENOMEM;
2102         }
2103
2104         /* Check to ensure the stack will not grow into a hugetlb-only region */
2105         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2106                         vma->vm_end - size;
2107         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2108                 return -EFAULT;
2109
2110         /*
2111          * Overcommit..  This must be the final test, as it will
2112          * update security statistics.
2113          */
2114         if (security_vm_enough_memory_mm(mm, grow))
2115                 return -ENOMEM;
2116
2117         /* Ok, everything looks good - let it rip */
2118         if (vma->vm_flags & VM_LOCKED)
2119                 mm->locked_vm += grow;
2120         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
2121         return 0;
2122 }
2123
2124 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2125 /*
2126  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2127  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2128  */
2129 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2130 {
2131         int error;
2132
2133         if (!(vma->vm_flags & VM_GROWSUP))
2134                 return -EFAULT;
2135
2136         /*
2137          * We must make sure the anon_vma is allocated
2138          * so that the anon_vma locking is not a noop.
2139          */
2140         if (unlikely(anon_vma_prepare(vma)))
2141                 return -ENOMEM;
2142         vma_lock_anon_vma(vma);
2143
2144         /*
2145          * vma->vm_start/vm_end cannot change under us because the caller
2146          * is required to hold the mmap_sem in read mode.  We need the
2147          * anon_vma lock to serialize against concurrent expand_stacks.
2148          * Also guard against wrapping around to address 0.
2149          */
2150         if (address < PAGE_ALIGN(address+4))
2151                 address = PAGE_ALIGN(address+4);
2152         else {
2153                 vma_unlock_anon_vma(vma);
2154                 return -ENOMEM;
2155         }
2156         error = 0;
2157
2158         /* Somebody else might have raced and expanded it already */
2159         if (address > vma->vm_end) {
2160                 unsigned long size, grow;
2161
2162                 size = address - vma->vm_start;
2163                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2164
2165                 error = -ENOMEM;
2166                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2167                         error = acct_stack_growth(vma, size, grow);
2168                         if (!error) {
2169                                 /*
2170                                  * vma_gap_update() doesn't support concurrent
2171                                  * updates, but we only hold a shared mmap_sem
2172                                  * lock here, so we need to protect against
2173                                  * concurrent vma expansions.
2174                                  * vma_lock_anon_vma() doesn't help here, as
2175                                  * we don't guarantee that all growable vmas
2176                                  * in a mm share the same root anon vma.
2177                                  * So, we reuse mm->page_table_lock to guard
2178                                  * against concurrent vma expansions.
2179                                  */
2180                                 spin_lock(&vma->vm_mm->page_table_lock);
2181                                 anon_vma_interval_tree_pre_update_vma(vma);
2182                                 vma->vm_end = address;
2183                                 anon_vma_interval_tree_post_update_vma(vma);
2184                                 if (vma->vm_next)
2185                                         vma_gap_update(vma->vm_next);
2186                                 else
2187                                         vma->vm_mm->highest_vm_end = address;
2188                                 spin_unlock(&vma->vm_mm->page_table_lock);
2189
2190                                 perf_event_mmap(vma);
2191                         }
2192                 }
2193         }
2194         vma_unlock_anon_vma(vma);
2195         khugepaged_enter_vma_merge(vma);
2196         validate_mm(vma->vm_mm);
2197         return error;
2198 }
2199 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2200
2201 /*
2202  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2203  */
2204 int expand_downwards(struct vm_area_struct *vma,
2205                                    unsigned long address)
2206 {
2207         int error;
2208
2209         /*
2210          * We must make sure the anon_vma is allocated
2211          * so that the anon_vma locking is not a noop.
2212          */
2213         if (unlikely(anon_vma_prepare(vma)))
2214                 return -ENOMEM;
2215
2216         address &= PAGE_MASK;
2217         error = security_mmap_addr(address);
2218         if (error)
2219                 return error;
2220
2221         vma_lock_anon_vma(vma);
2222
2223         /*
2224          * vma->vm_start/vm_end cannot change under us because the caller
2225          * is required to hold the mmap_sem in read mode.  We need the
2226          * anon_vma lock to serialize against concurrent expand_stacks.
2227          */
2228
2229         /* Somebody else might have raced and expanded it already */
2230         if (address < vma->vm_start) {
2231                 unsigned long size, grow;
2232
2233                 size = vma->vm_end - address;
2234                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2235
2236                 error = -ENOMEM;
2237                 if (grow <= vma->vm_pgoff) {
2238                         error = acct_stack_growth(vma, size, grow);
2239                         if (!error) {
2240                                 /*
2241                                  * vma_gap_update() doesn't support concurrent
2242                                  * updates, but we only hold a shared mmap_sem
2243                                  * lock here, so we need to protect against
2244                                  * concurrent vma expansions.
2245                                  * vma_lock_anon_vma() doesn't help here, as
2246                                  * we don't guarantee that all growable vmas
2247                                  * in a mm share the same root anon vma.
2248                                  * So, we reuse mm->page_table_lock to guard
2249                                  * against concurrent vma expansions.
2250                                  */
2251                                 spin_lock(&vma->vm_mm->page_table_lock);
2252                                 anon_vma_interval_tree_pre_update_vma(vma);
2253                                 vma->vm_start = address;
2254                                 vma->vm_pgoff -= grow;
2255                                 anon_vma_interval_tree_post_update_vma(vma);
2256                                 vma_gap_update(vma);
2257                                 spin_unlock(&vma->vm_mm->page_table_lock);
2258
2259                                 perf_event_mmap(vma);
2260                         }
2261                 }
2262         }
2263         vma_unlock_anon_vma(vma);
2264         khugepaged_enter_vma_merge(vma);
2265         validate_mm(vma->vm_mm);
2266         return error;
2267 }
2268
2269 /*
2270  * Note how expand_stack() refuses to expand the stack all the way to
2271  * abut the next virtual mapping, *unless* that mapping itself is also
2272  * a stack mapping. We want to leave room for a guard page, after all
2273  * (the guard page itself is not added here, that is done by the
2274  * actual page faulting logic)
2275  *
2276  * This matches the behavior of the guard page logic (see mm/memory.c:
2277  * check_stack_guard_page()), which only allows the guard page to be
2278  * removed under these circumstances.
2279  */
2280 #ifdef CONFIG_STACK_GROWSUP
2281 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2282 {
2283         struct vm_area_struct *next;
2284
2285         address &= PAGE_MASK;
2286         next = vma->vm_next;
2287         if (next && next->vm_start == address + PAGE_SIZE) {
2288                 if (!(next->vm_flags & VM_GROWSUP))
2289                         return -ENOMEM;
2290         }
2291         return expand_upwards(vma, address);
2292 }
2293
2294 struct vm_area_struct *
2295 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2296 {
2297         struct vm_area_struct *vma, *prev;
2298
2299         addr &= PAGE_MASK;
2300         vma = find_vma_prev(mm, addr, &prev);
2301         if (vma && (vma->vm_start <= addr))
2302                 return vma;
2303         if (!prev || expand_stack(prev, addr))
2304                 return NULL;
2305         if (prev->vm_flags & VM_LOCKED)
2306                 __mlock_vma_pages_range(prev, addr, prev->vm_end, NULL);
2307         return prev;
2308 }
2309 #else
2310 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2311 {
2312         struct vm_area_struct *prev;
2313
2314         address &= PAGE_MASK;
2315         prev = vma->vm_prev;
2316         if (prev && prev->vm_end == address) {
2317                 if (!(prev->vm_flags & VM_GROWSDOWN))
2318                         return -ENOMEM;
2319         }
2320         return expand_downwards(vma, address);
2321 }
2322
2323 struct vm_area_struct *
2324 find_extend_vma(struct mm_struct * mm, unsigned long addr)
2325 {
2326         struct vm_area_struct * vma;
2327         unsigned long start;
2328
2329         addr &= PAGE_MASK;
2330         vma = find_vma(mm,addr);
2331         if (!vma)
2332                 return NULL;
2333         if (vma->vm_start <= addr)
2334                 return vma;
2335         if (!(vma->vm_flags & VM_GROWSDOWN))
2336                 return NULL;
2337         start = vma->vm_start;
2338         if (expand_stack(vma, addr))
2339                 return NULL;
2340         if (vma->vm_flags & VM_LOCKED)
2341                 __mlock_vma_pages_range(vma, addr, start, NULL);
2342         return vma;
2343 }
2344 #endif
2345
2346 /*
2347  * Ok - we have the memory areas we should free on the vma list,
2348  * so release them, and do the vma updates.
2349  *
2350  * Called with the mm semaphore held.
2351  */
2352 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2353 {
2354         unsigned long nr_accounted = 0;
2355
2356         /* Update high watermark before we lower total_vm */
2357         update_hiwater_vm(mm);
2358         do {
2359                 long nrpages = vma_pages(vma);
2360
2361                 if (vma->vm_flags & VM_ACCOUNT)
2362                         nr_accounted += nrpages;
2363                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
2364                 vma = remove_vma(vma);
2365         } while (vma);
2366         vm_unacct_memory(nr_accounted);
2367         validate_mm(mm);
2368 }
2369
2370 /*
2371  * Get rid of page table information in the indicated region.
2372  *
2373  * Called with the mm semaphore held.
2374  */
2375 static void unmap_region(struct mm_struct *mm,
2376                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2377                 unsigned long start, unsigned long end)
2378 {
2379         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
2380         struct mmu_gather tlb;
2381
2382         lru_add_drain();
2383         tlb_gather_mmu(&tlb, mm, start, end);
2384         update_hiwater_rss(mm);
2385         unmap_vmas(&tlb, vma, start, end);
2386         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2387                                  next ? next->vm_start : USER_PGTABLES_CEILING);
2388         tlb_finish_mmu(&tlb, start, end);
2389 }
2390
2391 /*
2392  * Create a list of vma's touched by the unmap, removing them from the mm's
2393  * vma list as we go..
2394  */
2395 static void
2396 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2397         struct vm_area_struct *prev, unsigned long end)
2398 {
2399         struct vm_area_struct **insertion_point;
2400         struct vm_area_struct *tail_vma = NULL;
2401
2402         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2403         vma->vm_prev = NULL;
2404         do {
2405                 vma_rb_erase(vma, &mm->mm_rb);
2406                 mm->map_count--;
2407                 tail_vma = vma;
2408                 vma = vma->vm_next;
2409         } while (vma && vma->vm_start < end);
2410         *insertion_point = vma;
2411         if (vma) {
2412                 vma->vm_prev = prev;
2413                 vma_gap_update(vma);
2414         } else
2415                 mm->highest_vm_end = prev ? prev->vm_end : 0;
2416         tail_vma->vm_next = NULL;
2417
2418         /* Kill the cache */
2419         vmacache_invalidate(mm);
2420 }
2421
2422 /*
2423  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
2424  * munmap path where it doesn't make sense to fail.
2425  */
2426 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
2427               unsigned long addr, int new_below)
2428 {
2429         struct vm_area_struct *new;
2430         int err = -ENOMEM;
2431
2432         if (is_vm_hugetlb_page(vma) && (addr &
2433                                         ~(huge_page_mask(hstate_vma(vma)))))
2434                 return -EINVAL;
2435
2436         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2437         if (!new)
2438                 goto out_err;
2439
2440         /* most fields are the same, copy all, and then fixup */
2441         *new = *vma;
2442
2443         INIT_LIST_HEAD(&new->anon_vma_chain);
2444
2445         if (new_below)
2446                 new->vm_end = addr;
2447         else {
2448                 new->vm_start = addr;
2449                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2450         }
2451
2452         err = vma_dup_policy(vma, new);
2453         if (err)
2454                 goto out_free_vma;
2455
2456         if (anon_vma_clone(new, vma))
2457                 goto out_free_mpol;
2458
2459         if (new->vm_file)
2460                 get_file(new->vm_file);
2461
2462         if (new->vm_ops && new->vm_ops->open)
2463                 new->vm_ops->open(new);
2464
2465         if (new_below)
2466                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2467                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2468         else
2469                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2470
2471         /* Success. */
2472         if (!err)
2473                 return 0;
2474
2475         /* Clean everything up if vma_adjust failed. */
2476         if (new->vm_ops && new->vm_ops->close)
2477                 new->vm_ops->close(new);
2478         if (new->vm_file)
2479                 fput(new->vm_file);
2480         unlink_anon_vmas(new);
2481  out_free_mpol:
2482         mpol_put(vma_policy(new));
2483  out_free_vma:
2484         kmem_cache_free(vm_area_cachep, new);
2485  out_err:
2486         return err;
2487 }
2488
2489 /*
2490  * Split a vma into two pieces at address 'addr', a new vma is allocated
2491  * either for the first part or the tail.
2492  */
2493 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2494               unsigned long addr, int new_below)
2495 {
2496         if (mm->map_count >= sysctl_max_map_count)
2497                 return -ENOMEM;
2498
2499         return __split_vma(mm, vma, addr, new_below);
2500 }
2501
2502 /* Munmap is split into 2 main parts -- this part which finds
2503  * what needs doing, and the areas themselves, which do the
2504  * work.  This now handles partial unmappings.
2505  * Jeremy Fitzhardinge <jeremy@goop.org>
2506  */
2507 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2508 {
2509         unsigned long end;
2510         struct vm_area_struct *vma, *prev, *last;
2511
2512         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2513                 return -EINVAL;
2514
2515         if ((len = PAGE_ALIGN(len)) == 0)
2516                 return -EINVAL;
2517
2518         /* Find the first overlapping VMA */
2519         vma = find_vma(mm, start);
2520         if (!vma)
2521                 return 0;
2522         prev = vma->vm_prev;
2523         /* we have  start < vma->vm_end  */
2524
2525         /* if it doesn't overlap, we have nothing.. */
2526         end = start + len;
2527         if (vma->vm_start >= end)
2528                 return 0;
2529
2530         /*
2531          * If we need to split any vma, do it now to save pain later.
2532          *
2533          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2534          * unmapped vm_area_struct will remain in use: so lower split_vma
2535          * places tmp vma above, and higher split_vma places tmp vma below.
2536          */
2537         if (start > vma->vm_start) {
2538                 int error;
2539
2540                 /*
2541                  * Make sure that map_count on return from munmap() will
2542                  * not exceed its limit; but let map_count go just above
2543                  * its limit temporarily, to help free resources as expected.
2544                  */
2545                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2546                         return -ENOMEM;
2547
2548                 error = __split_vma(mm, vma, start, 0);
2549                 if (error)
2550                         return error;
2551                 prev = vma;
2552         }
2553
2554         /* Does it split the last one? */
2555         last = find_vma(mm, end);
2556         if (last && end > last->vm_start) {
2557                 int error = __split_vma(mm, last, end, 1);
2558                 if (error)
2559                         return error;
2560         }
2561         vma = prev? prev->vm_next: mm->mmap;
2562
2563         /*
2564          * unlock any mlock()ed ranges before detaching vmas
2565          */
2566         if (mm->locked_vm) {
2567                 struct vm_area_struct *tmp = vma;
2568                 while (tmp && tmp->vm_start < end) {
2569                         if (tmp->vm_flags & VM_LOCKED) {
2570                                 mm->locked_vm -= vma_pages(tmp);
2571                                 munlock_vma_pages_all(tmp);
2572                         }
2573                         tmp = tmp->vm_next;
2574                 }
2575         }
2576
2577         /*
2578          * Remove the vma's, and unmap the actual pages
2579          */
2580         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2581         unmap_region(mm, vma, prev, start, end);
2582
2583         /* Fix up all other VM information */
2584         remove_vma_list(mm, vma);
2585
2586         return 0;
2587 }
2588
2589 int vm_munmap(unsigned long start, size_t len)
2590 {
2591         int ret;
2592         struct mm_struct *mm = current->mm;
2593
2594         down_write(&mm->mmap_sem);
2595         ret = do_munmap(mm, start, len);
2596         up_write(&mm->mmap_sem);
2597         return ret;
2598 }
2599 EXPORT_SYMBOL(vm_munmap);
2600
2601 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2602 {
2603         profile_munmap(addr);
2604         return vm_munmap(addr, len);
2605 }
2606
2607 static inline void verify_mm_writelocked(struct mm_struct *mm)
2608 {
2609 #ifdef CONFIG_DEBUG_VM
2610         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2611                 WARN_ON(1);
2612                 up_read(&mm->mmap_sem);
2613         }
2614 #endif
2615 }
2616
2617 /*
2618  *  this is really a simplified "do_mmap".  it only handles
2619  *  anonymous maps.  eventually we may be able to do some
2620  *  brk-specific accounting here.
2621  */
2622 static unsigned long do_brk(unsigned long addr, unsigned long len)
2623 {
2624         struct mm_struct * mm = current->mm;
2625         struct vm_area_struct * vma, * prev;
2626         unsigned long flags;
2627         struct rb_node ** rb_link, * rb_parent;
2628         pgoff_t pgoff = addr >> PAGE_SHIFT;
2629         int error;
2630
2631         len = PAGE_ALIGN(len);
2632         if (!len)
2633                 return addr;
2634
2635         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2636
2637         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2638         if (error & ~PAGE_MASK)
2639                 return error;
2640
2641         error = mlock_future_check(mm, mm->def_flags, len);
2642         if (error)
2643                 return error;
2644
2645         /*
2646          * mm->mmap_sem is required to protect against another thread
2647          * changing the mappings in case we sleep.
2648          */
2649         verify_mm_writelocked(mm);
2650
2651         /*
2652          * Clear old maps.  this also does some error checking for us
2653          */
2654  munmap_back:
2655         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
2656                 if (do_munmap(mm, addr, len))
2657                         return -ENOMEM;
2658                 goto munmap_back;
2659         }
2660
2661         /* Check against address space limits *after* clearing old maps... */
2662         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2663                 return -ENOMEM;
2664
2665         if (mm->map_count > sysctl_max_map_count)
2666                 return -ENOMEM;
2667
2668         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2669                 return -ENOMEM;
2670
2671         /* Can we just expand an old private anonymous mapping? */
2672         vma = vma_merge(mm, prev, addr, addr + len, flags,
2673                                         NULL, NULL, pgoff, NULL);
2674         if (vma)
2675                 goto out;
2676
2677         /*
2678          * create a vma struct for an anonymous mapping
2679          */
2680         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2681         if (!vma) {
2682                 vm_unacct_memory(len >> PAGE_SHIFT);
2683                 return -ENOMEM;
2684         }
2685
2686         INIT_LIST_HEAD(&vma->anon_vma_chain);
2687         vma->vm_mm = mm;
2688         vma->vm_start = addr;
2689         vma->vm_end = addr + len;
2690         vma->vm_pgoff = pgoff;
2691         vma->vm_flags = flags;
2692         vma->vm_page_prot = vm_get_page_prot(flags);
2693         vma_link(mm, vma, prev, rb_link, rb_parent);
2694 out:
2695         perf_event_mmap(vma);
2696         mm->total_vm += len >> PAGE_SHIFT;
2697         if (flags & VM_LOCKED)
2698                 mm->locked_vm += (len >> PAGE_SHIFT);
2699         vma->vm_flags |= VM_SOFTDIRTY;
2700         return addr;
2701 }
2702
2703 unsigned long vm_brk(unsigned long addr, unsigned long len)
2704 {
2705         struct mm_struct *mm = current->mm;
2706         unsigned long ret;
2707         bool populate;
2708
2709         down_write(&mm->mmap_sem);
2710         ret = do_brk(addr, len);
2711         populate = ((mm->def_flags & VM_LOCKED) != 0);
2712         up_write(&mm->mmap_sem);
2713         if (populate)
2714                 mm_populate(addr, len);
2715         return ret;
2716 }
2717 EXPORT_SYMBOL(vm_brk);
2718
2719 /* Release all mmaps. */
2720 void exit_mmap(struct mm_struct *mm)
2721 {
2722         struct mmu_gather tlb;
2723         struct vm_area_struct *vma;
2724         unsigned long nr_accounted = 0;
2725
2726         /* mm's last user has gone, and its about to be pulled down */
2727         mmu_notifier_release(mm);
2728
2729         if (mm->locked_vm) {
2730                 vma = mm->mmap;
2731                 while (vma) {
2732                         if (vma->vm_flags & VM_LOCKED)
2733                                 munlock_vma_pages_all(vma);
2734                         vma = vma->vm_next;
2735                 }
2736         }
2737
2738         arch_exit_mmap(mm);
2739
2740         vma = mm->mmap;
2741         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2742                 return;
2743
2744         lru_add_drain();
2745         flush_cache_mm(mm);
2746         tlb_gather_mmu(&tlb, mm, 0, -1);
2747         /* update_hiwater_rss(mm) here? but nobody should be looking */
2748         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2749         unmap_vmas(&tlb, vma, 0, -1);
2750
2751         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
2752         tlb_finish_mmu(&tlb, 0, -1);
2753
2754         /*
2755          * Walk the list again, actually closing and freeing it,
2756          * with preemption enabled, without holding any MM locks.
2757          */
2758         while (vma) {
2759                 if (vma->vm_flags & VM_ACCOUNT)
2760                         nr_accounted += vma_pages(vma);
2761                 vma = remove_vma(vma);
2762         }
2763         vm_unacct_memory(nr_accounted);
2764
2765         WARN_ON(atomic_long_read(&mm->nr_ptes) >
2766                         (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2767 }
2768
2769 /* Insert vm structure into process list sorted by address
2770  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2771  * then i_mmap_mutex is taken here.
2772  */
2773 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2774 {
2775         struct vm_area_struct *prev;
2776         struct rb_node **rb_link, *rb_parent;
2777
2778         /*
2779          * The vm_pgoff of a purely anonymous vma should be irrelevant
2780          * until its first write fault, when page's anon_vma and index
2781          * are set.  But now set the vm_pgoff it will almost certainly
2782          * end up with (unless mremap moves it elsewhere before that
2783          * first wfault), so /proc/pid/maps tells a consistent story.
2784          *
2785          * By setting it to reflect the virtual start address of the
2786          * vma, merges and splits can happen in a seamless way, just
2787          * using the existing file pgoff checks and manipulations.
2788          * Similarly in do_mmap_pgoff and in do_brk.
2789          */
2790         if (!vma->vm_file) {
2791                 BUG_ON(vma->anon_vma);
2792                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2793         }
2794         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2795                            &prev, &rb_link, &rb_parent))
2796                 return -ENOMEM;
2797         if ((vma->vm_flags & VM_ACCOUNT) &&
2798              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2799                 return -ENOMEM;
2800
2801         vma_link(mm, vma, prev, rb_link, rb_parent);
2802         return 0;
2803 }
2804
2805 /*
2806  * Copy the vma structure to a new location in the same mm,
2807  * prior to moving page table entries, to effect an mremap move.
2808  */
2809 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2810         unsigned long addr, unsigned long len, pgoff_t pgoff,
2811         bool *need_rmap_locks)
2812 {
2813         struct vm_area_struct *vma = *vmap;
2814         unsigned long vma_start = vma->vm_start;
2815         struct mm_struct *mm = vma->vm_mm;
2816         struct vm_area_struct *new_vma, *prev;
2817         struct rb_node **rb_link, *rb_parent;
2818         bool faulted_in_anon_vma = true;
2819
2820         /*
2821          * If anonymous vma has not yet been faulted, update new pgoff
2822          * to match new location, to increase its chance of merging.
2823          */
2824         if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2825                 pgoff = addr >> PAGE_SHIFT;
2826                 faulted_in_anon_vma = false;
2827         }
2828
2829         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2830                 return NULL;    /* should never get here */
2831         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2832                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2833         if (new_vma) {
2834                 /*
2835                  * Source vma may have been merged into new_vma
2836                  */
2837                 if (unlikely(vma_start >= new_vma->vm_start &&
2838                              vma_start < new_vma->vm_end)) {
2839                         /*
2840                          * The only way we can get a vma_merge with
2841                          * self during an mremap is if the vma hasn't
2842                          * been faulted in yet and we were allowed to
2843                          * reset the dst vma->vm_pgoff to the
2844                          * destination address of the mremap to allow
2845                          * the merge to happen. mremap must change the
2846                          * vm_pgoff linearity between src and dst vmas
2847                          * (in turn preventing a vma_merge) to be
2848                          * safe. It is only safe to keep the vm_pgoff
2849                          * linear if there are no pages mapped yet.
2850                          */
2851                         VM_BUG_ON(faulted_in_anon_vma);
2852                         *vmap = vma = new_vma;
2853                 }
2854                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2855         } else {
2856                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2857                 if (new_vma) {
2858                         *new_vma = *vma;
2859                         new_vma->vm_start = addr;
2860                         new_vma->vm_end = addr + len;
2861                         new_vma->vm_pgoff = pgoff;
2862                         if (vma_dup_policy(vma, new_vma))
2863                                 goto out_free_vma;
2864                         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2865                         if (anon_vma_clone(new_vma, vma))
2866                                 goto out_free_mempol;
2867                         if (new_vma->vm_file)
2868                                 get_file(new_vma->vm_file);
2869                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2870                                 new_vma->vm_ops->open(new_vma);
2871                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2872                         *need_rmap_locks = false;
2873                 }
2874         }
2875         return new_vma;
2876
2877  out_free_mempol:
2878         mpol_put(vma_policy(new_vma));
2879  out_free_vma:
2880         kmem_cache_free(vm_area_cachep, new_vma);
2881         return NULL;
2882 }
2883
2884 /*
2885  * Return true if the calling process may expand its vm space by the passed
2886  * number of pages
2887  */
2888 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2889 {
2890         unsigned long cur = mm->total_vm;       /* pages */
2891         unsigned long lim;
2892
2893         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2894
2895         if (cur + npages > lim)
2896                 return 0;
2897         return 1;
2898 }
2899
2900 static int special_mapping_fault(struct vm_area_struct *vma,
2901                                  struct vm_fault *vmf);
2902
2903 /*
2904  * Having a close hook prevents vma merging regardless of flags.
2905  */
2906 static void special_mapping_close(struct vm_area_struct *vma)
2907 {
2908 }
2909
2910 static const char *special_mapping_name(struct vm_area_struct *vma)
2911 {
2912         return ((struct vm_special_mapping *)vma->vm_private_data)->name;
2913 }
2914
2915 static const struct vm_operations_struct special_mapping_vmops = {
2916         .close = special_mapping_close,
2917         .fault = special_mapping_fault,
2918         .name = special_mapping_name,
2919 };
2920
2921 static const struct vm_operations_struct legacy_special_mapping_vmops = {
2922         .close = special_mapping_close,
2923         .fault = special_mapping_fault,
2924 };
2925
2926 static int special_mapping_fault(struct vm_area_struct *vma,
2927                                 struct vm_fault *vmf)
2928 {
2929         pgoff_t pgoff;
2930         struct page **pages;
2931
2932         /*
2933          * special mappings have no vm_file, and in that case, the mm
2934          * uses vm_pgoff internally. So we have to subtract it from here.
2935          * We are allowed to do this because we are the mm; do not copy
2936          * this code into drivers!
2937          */
2938         pgoff = vmf->pgoff - vma->vm_pgoff;
2939
2940         if (vma->vm_ops == &legacy_special_mapping_vmops)
2941                 pages = vma->vm_private_data;
2942         else
2943                 pages = ((struct vm_special_mapping *)vma->vm_private_data)->
2944                         pages;
2945
2946         for (; pgoff && *pages; ++pages)
2947                 pgoff--;
2948
2949         if (*pages) {
2950                 struct page *page = *pages;
2951                 get_page(page);
2952                 vmf->page = page;
2953                 return 0;
2954         }
2955
2956         return VM_FAULT_SIGBUS;
2957 }
2958
2959 static struct vm_area_struct *__install_special_mapping(
2960         struct mm_struct *mm,
2961         unsigned long addr, unsigned long len,
2962         unsigned long vm_flags, const struct vm_operations_struct *ops,
2963         void *priv)
2964 {
2965         int ret;
2966         struct vm_area_struct *vma;
2967
2968         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2969         if (unlikely(vma == NULL))
2970                 return ERR_PTR(-ENOMEM);
2971
2972         INIT_LIST_HEAD(&vma->anon_vma_chain);
2973         vma->vm_mm = mm;
2974         vma->vm_start = addr;
2975         vma->vm_end = addr + len;
2976
2977         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
2978         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2979
2980         vma->vm_ops = ops;
2981         vma->vm_private_data = priv;
2982
2983         ret = insert_vm_struct(mm, vma);
2984         if (ret)
2985                 goto out;
2986
2987         mm->total_vm += len >> PAGE_SHIFT;
2988
2989         perf_event_mmap(vma);
2990
2991         return vma;
2992
2993 out:
2994         kmem_cache_free(vm_area_cachep, vma);
2995         return ERR_PTR(ret);
2996 }
2997
2998 /*
2999  * Called with mm->mmap_sem held for writing.
3000  * Insert a new vma covering the given region, with the given flags.
3001  * Its pages are supplied by the given array of struct page *.
3002  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3003  * The region past the last page supplied will always produce SIGBUS.
3004  * The array pointer and the pages it points to are assumed to stay alive
3005  * for as long as this mapping might exist.
3006  */
3007 struct vm_area_struct *_install_special_mapping(
3008         struct mm_struct *mm,
3009         unsigned long addr, unsigned long len,
3010         unsigned long vm_flags, const struct vm_special_mapping *spec)
3011 {
3012         return __install_special_mapping(mm, addr, len, vm_flags,
3013                                          &special_mapping_vmops, (void *)spec);
3014 }
3015
3016 int install_special_mapping(struct mm_struct *mm,
3017                             unsigned long addr, unsigned long len,
3018                             unsigned long vm_flags, struct page **pages)
3019 {
3020         struct vm_area_struct *vma = __install_special_mapping(
3021                 mm, addr, len, vm_flags, &legacy_special_mapping_vmops,
3022                 (void *)pages);
3023
3024         return PTR_ERR_OR_ZERO(vma);
3025 }
3026
3027 static DEFINE_MUTEX(mm_all_locks_mutex);
3028
3029 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3030 {
3031         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3032                 /*
3033                  * The LSB of head.next can't change from under us
3034                  * because we hold the mm_all_locks_mutex.
3035                  */
3036                 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3037                 /*
3038                  * We can safely modify head.next after taking the
3039                  * anon_vma->root->rwsem. If some other vma in this mm shares
3040                  * the same anon_vma we won't take it again.
3041                  *
3042                  * No need of atomic instructions here, head.next
3043                  * can't change from under us thanks to the
3044                  * anon_vma->root->rwsem.
3045                  */
3046                 if (__test_and_set_bit(0, (unsigned long *)
3047                                        &anon_vma->root->rb_root.rb_node))
3048                         BUG();
3049         }
3050 }
3051
3052 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3053 {
3054         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3055                 /*
3056                  * AS_MM_ALL_LOCKS can't change from under us because
3057                  * we hold the mm_all_locks_mutex.
3058                  *
3059                  * Operations on ->flags have to be atomic because
3060                  * even if AS_MM_ALL_LOCKS is stable thanks to the
3061                  * mm_all_locks_mutex, there may be other cpus
3062                  * changing other bitflags in parallel to us.
3063                  */
3064                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3065                         BUG();
3066                 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
3067         }
3068 }
3069
3070 /*
3071  * This operation locks against the VM for all pte/vma/mm related
3072  * operations that could ever happen on a certain mm. This includes
3073  * vmtruncate, try_to_unmap, and all page faults.
3074  *
3075  * The caller must take the mmap_sem in write mode before calling
3076  * mm_take_all_locks(). The caller isn't allowed to release the
3077  * mmap_sem until mm_drop_all_locks() returns.
3078  *
3079  * mmap_sem in write mode is required in order to block all operations
3080  * that could modify pagetables and free pages without need of
3081  * altering the vma layout (for example populate_range() with
3082  * nonlinear vmas). It's also needed in write mode to avoid new
3083  * anon_vmas to be associated with existing vmas.
3084  *
3085  * A single task can't take more than one mm_take_all_locks() in a row
3086  * or it would deadlock.
3087  *
3088  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3089  * mapping->flags avoid to take the same lock twice, if more than one
3090  * vma in this mm is backed by the same anon_vma or address_space.
3091  *
3092  * We can take all the locks in random order because the VM code
3093  * taking i_mmap_mutex or anon_vma->rwsem outside the mmap_sem never
3094  * takes more than one of them in a row. Secondly we're protected
3095  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3096  *
3097  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3098  * that may have to take thousand of locks.
3099  *
3100  * mm_take_all_locks() can fail if it's interrupted by signals.
3101  */
3102 int mm_take_all_locks(struct mm_struct *mm)
3103 {
3104         struct vm_area_struct *vma;
3105         struct anon_vma_chain *avc;
3106
3107         BUG_ON(down_read_trylock(&mm->mmap_sem));
3108
3109         mutex_lock(&mm_all_locks_mutex);
3110
3111         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3112                 if (signal_pending(current))
3113                         goto out_unlock;
3114                 if (vma->vm_file && vma->vm_file->f_mapping)
3115                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3116         }
3117
3118         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3119                 if (signal_pending(current))
3120                         goto out_unlock;
3121                 if (vma->anon_vma)
3122                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3123                                 vm_lock_anon_vma(mm, avc->anon_vma);
3124         }
3125
3126         return 0;
3127
3128 out_unlock:
3129         mm_drop_all_locks(mm);
3130         return -EINTR;
3131 }
3132
3133 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3134 {
3135         if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3136                 /*
3137                  * The LSB of head.next can't change to 0 from under
3138                  * us because we hold the mm_all_locks_mutex.
3139                  *
3140                  * We must however clear the bitflag before unlocking
3141                  * the vma so the users using the anon_vma->rb_root will
3142                  * never see our bitflag.
3143                  *
3144                  * No need of atomic instructions here, head.next
3145                  * can't change from under us until we release the
3146                  * anon_vma->root->rwsem.
3147                  */
3148                 if (!__test_and_clear_bit(0, (unsigned long *)
3149                                           &anon_vma->root->rb_root.rb_node))
3150                         BUG();
3151                 anon_vma_unlock_write(anon_vma);
3152         }
3153 }
3154
3155 static void vm_unlock_mapping(struct address_space *mapping)
3156 {
3157         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3158                 /*
3159                  * AS_MM_ALL_LOCKS can't change to 0 from under us
3160                  * because we hold the mm_all_locks_mutex.
3161                  */
3162                 mutex_unlock(&mapping->i_mmap_mutex);
3163                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3164                                         &mapping->flags))
3165                         BUG();
3166         }
3167 }
3168
3169 /*
3170  * The mmap_sem cannot be released by the caller until
3171  * mm_drop_all_locks() returns.
3172  */
3173 void mm_drop_all_locks(struct mm_struct *mm)
3174 {
3175         struct vm_area_struct *vma;
3176         struct anon_vma_chain *avc;
3177
3178         BUG_ON(down_read_trylock(&mm->mmap_sem));
3179         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3180
3181         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3182                 if (vma->anon_vma)
3183                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3184                                 vm_unlock_anon_vma(avc->anon_vma);
3185                 if (vma->vm_file && vma->vm_file->f_mapping)
3186                         vm_unlock_mapping(vma->vm_file->f_mapping);
3187         }
3188
3189         mutex_unlock(&mm_all_locks_mutex);
3190 }
3191
3192 /*
3193  * initialise the VMA slab
3194  */
3195 void __init mmap_init(void)
3196 {
3197         int ret;
3198
3199         ret = percpu_counter_init(&vm_committed_as, 0);
3200         VM_BUG_ON(ret);
3201 }
3202
3203 /*
3204  * Initialise sysctl_user_reserve_kbytes.
3205  *
3206  * This is intended to prevent a user from starting a single memory hogging
3207  * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3208  * mode.
3209  *
3210  * The default value is min(3% of free memory, 128MB)
3211  * 128MB is enough to recover with sshd/login, bash, and top/kill.
3212  */
3213 static int init_user_reserve(void)
3214 {
3215         unsigned long free_kbytes;
3216
3217         free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3218
3219         sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3220         return 0;
3221 }
3222 subsys_initcall(init_user_reserve);
3223
3224 /*
3225  * Initialise sysctl_admin_reserve_kbytes.
3226  *
3227  * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3228  * to log in and kill a memory hogging process.
3229  *
3230  * Systems with more than 256MB will reserve 8MB, enough to recover
3231  * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3232  * only reserve 3% of free pages by default.
3233  */
3234 static int init_admin_reserve(void)
3235 {
3236         unsigned long free_kbytes;
3237
3238         free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3239
3240         sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3241         return 0;
3242 }
3243 subsys_initcall(init_admin_reserve);
3244
3245 /*
3246  * Reinititalise user and admin reserves if memory is added or removed.
3247  *
3248  * The default user reserve max is 128MB, and the default max for the
3249  * admin reserve is 8MB. These are usually, but not always, enough to
3250  * enable recovery from a memory hogging process using login/sshd, a shell,
3251  * and tools like top. It may make sense to increase or even disable the
3252  * reserve depending on the existence of swap or variations in the recovery
3253  * tools. So, the admin may have changed them.
3254  *
3255  * If memory is added and the reserves have been eliminated or increased above
3256  * the default max, then we'll trust the admin.
3257  *
3258  * If memory is removed and there isn't enough free memory, then we
3259  * need to reset the reserves.
3260  *
3261  * Otherwise keep the reserve set by the admin.
3262  */
3263 static int reserve_mem_notifier(struct notifier_block *nb,
3264                              unsigned long action, void *data)
3265 {
3266         unsigned long tmp, free_kbytes;
3267
3268         switch (action) {
3269         case MEM_ONLINE:
3270                 /* Default max is 128MB. Leave alone if modified by operator. */
3271                 tmp = sysctl_user_reserve_kbytes;
3272                 if (0 < tmp && tmp < (1UL << 17))
3273                         init_user_reserve();
3274
3275                 /* Default max is 8MB.  Leave alone if modified by operator. */
3276                 tmp = sysctl_admin_reserve_kbytes;
3277                 if (0 < tmp && tmp < (1UL << 13))
3278                         init_admin_reserve();
3279
3280                 break;
3281         case MEM_OFFLINE:
3282                 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3283
3284                 if (sysctl_user_reserve_kbytes > free_kbytes) {
3285                         init_user_reserve();
3286                         pr_info("vm.user_reserve_kbytes reset to %lu\n",
3287                                 sysctl_user_reserve_kbytes);
3288                 }
3289
3290                 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3291                         init_admin_reserve();
3292                         pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3293                                 sysctl_admin_reserve_kbytes);
3294                 }
3295                 break;
3296         default:
3297                 break;
3298         }
3299         return NOTIFY_OK;
3300 }
3301
3302 static struct notifier_block reserve_mem_nb = {
3303         .notifier_call = reserve_mem_notifier,
3304 };
3305
3306 static int __meminit init_reserve_notifier(void)
3307 {
3308         if (register_hotmemory_notifier(&reserve_mem_nb))
3309                 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3310
3311         return 0;
3312 }
3313 subsys_initcall(init_reserve_notifier);