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