Loading...
1/*
2 * mm/mmap.c
3 *
4 * Written by obz.
5 *
6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
7 */
8
9#include <linux/slab.h>
10#include <linux/backing-dev.h>
11#include <linux/mm.h>
12#include <linux/shm.h>
13#include <linux/mman.h>
14#include <linux/pagemap.h>
15#include <linux/swap.h>
16#include <linux/syscalls.h>
17#include <linux/capability.h>
18#include <linux/init.h>
19#include <linux/file.h>
20#include <linux/fs.h>
21#include <linux/personality.h>
22#include <linux/security.h>
23#include <linux/hugetlb.h>
24#include <linux/profile.h>
25#include <linux/module.h>
26#include <linux/mount.h>
27#include <linux/mempolicy.h>
28#include <linux/rmap.h>
29#include <linux/mmu_notifier.h>
30#include <linux/perf_event.h>
31#include <linux/audit.h>
32#include <linux/khugepaged.h>
33
34#include <asm/uaccess.h>
35#include <asm/cacheflush.h>
36#include <asm/tlb.h>
37#include <asm/mmu_context.h>
38
39#include "internal.h"
40
41#ifndef arch_mmap_check
42#define arch_mmap_check(addr, len, flags) (0)
43#endif
44
45#ifndef arch_rebalance_pgtables
46#define arch_rebalance_pgtables(addr, len) (addr)
47#endif
48
49static void unmap_region(struct mm_struct *mm,
50 struct vm_area_struct *vma, struct vm_area_struct *prev,
51 unsigned long start, unsigned long end);
52
53/*
54 * WARNING: the debugging will use recursive algorithms so never enable this
55 * unless you know what you are doing.
56 */
57#undef DEBUG_MM_RB
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 */
74pgprot_t protection_map[16] = {
75 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
76 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
77};
78
79pgprot_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}
85EXPORT_SYMBOL(vm_get_page_prot);
86
87int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS; /* heuristic overcommit */
88int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
89int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
90/*
91 * Make sure vm_committed_as in one cacheline and not cacheline shared with
92 * other variables. It can be updated by several CPUs frequently.
93 */
94struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
95
96/*
97 * Check that a process has enough memory to allocate a new virtual
98 * mapping. 0 means there is enough memory for the allocation to
99 * succeed and -ENOMEM implies there is not.
100 *
101 * We currently support three overcommit policies, which are set via the
102 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
103 *
104 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
105 * Additional code 2002 Jul 20 by Robert Love.
106 *
107 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
108 *
109 * Note this is a helper function intended to be used by LSMs which
110 * wish to use this logic.
111 */
112int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
113{
114 unsigned long free, allowed;
115
116 vm_acct_memory(pages);
117
118 /*
119 * Sometimes we want to use more memory than we have
120 */
121 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
122 return 0;
123
124 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
125 free = global_page_state(NR_FREE_PAGES);
126 free += global_page_state(NR_FILE_PAGES);
127
128 /*
129 * shmem pages shouldn't be counted as free in this
130 * case, they can't be purged, only swapped out, and
131 * that won't affect the overall amount of available
132 * memory in the system.
133 */
134 free -= global_page_state(NR_SHMEM);
135
136 free += nr_swap_pages;
137
138 /*
139 * Any slabs which are created with the
140 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
141 * which are reclaimable, under pressure. The dentry
142 * cache and most inode caches should fall into this
143 */
144 free += global_page_state(NR_SLAB_RECLAIMABLE);
145
146 /*
147 * Leave reserved pages. The pages are not for anonymous pages.
148 */
149 if (free <= totalreserve_pages)
150 goto error;
151 else
152 free -= totalreserve_pages;
153
154 /*
155 * Leave the last 3% for root
156 */
157 if (!cap_sys_admin)
158 free -= free / 32;
159
160 if (free > pages)
161 return 0;
162
163 goto error;
164 }
165
166 allowed = (totalram_pages - hugetlb_total_pages())
167 * sysctl_overcommit_ratio / 100;
168 /*
169 * Leave the last 3% for root
170 */
171 if (!cap_sys_admin)
172 allowed -= allowed / 32;
173 allowed += total_swap_pages;
174
175 /* Don't let a single process grow too big:
176 leave 3% of the size of this process for other processes */
177 if (mm)
178 allowed -= mm->total_vm / 32;
179
180 if (percpu_counter_read_positive(&vm_committed_as) < allowed)
181 return 0;
182error:
183 vm_unacct_memory(pages);
184
185 return -ENOMEM;
186}
187
188/*
189 * Requires inode->i_mapping->i_mmap_mutex
190 */
191static void __remove_shared_vm_struct(struct vm_area_struct *vma,
192 struct file *file, struct address_space *mapping)
193{
194 if (vma->vm_flags & VM_DENYWRITE)
195 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
196 if (vma->vm_flags & VM_SHARED)
197 mapping->i_mmap_writable--;
198
199 flush_dcache_mmap_lock(mapping);
200 if (unlikely(vma->vm_flags & VM_NONLINEAR))
201 list_del_init(&vma->shared.vm_set.list);
202 else
203 vma_prio_tree_remove(vma, &mapping->i_mmap);
204 flush_dcache_mmap_unlock(mapping);
205}
206
207/*
208 * Unlink a file-based vm structure from its prio_tree, to hide
209 * vma from rmap and vmtruncate before freeing its page tables.
210 */
211void unlink_file_vma(struct vm_area_struct *vma)
212{
213 struct file *file = vma->vm_file;
214
215 if (file) {
216 struct address_space *mapping = file->f_mapping;
217 mutex_lock(&mapping->i_mmap_mutex);
218 __remove_shared_vm_struct(vma, file, mapping);
219 mutex_unlock(&mapping->i_mmap_mutex);
220 }
221}
222
223/*
224 * Close a vm structure and free it, returning the next.
225 */
226static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
227{
228 struct vm_area_struct *next = vma->vm_next;
229
230 might_sleep();
231 if (vma->vm_ops && vma->vm_ops->close)
232 vma->vm_ops->close(vma);
233 if (vma->vm_file) {
234 fput(vma->vm_file);
235 if (vma->vm_flags & VM_EXECUTABLE)
236 removed_exe_file_vma(vma->vm_mm);
237 }
238 mpol_put(vma_policy(vma));
239 kmem_cache_free(vm_area_cachep, vma);
240 return next;
241}
242
243SYSCALL_DEFINE1(brk, unsigned long, brk)
244{
245 unsigned long rlim, retval;
246 unsigned long newbrk, oldbrk;
247 struct mm_struct *mm = current->mm;
248 unsigned long min_brk;
249
250 down_write(&mm->mmap_sem);
251
252#ifdef CONFIG_COMPAT_BRK
253 /*
254 * CONFIG_COMPAT_BRK can still be overridden by setting
255 * randomize_va_space to 2, which will still cause mm->start_brk
256 * to be arbitrarily shifted
257 */
258 if (current->brk_randomized)
259 min_brk = mm->start_brk;
260 else
261 min_brk = mm->end_data;
262#else
263 min_brk = mm->start_brk;
264#endif
265 if (brk < min_brk)
266 goto out;
267
268 /*
269 * Check against rlimit here. If this check is done later after the test
270 * of oldbrk with newbrk then it can escape the test and let the data
271 * segment grow beyond its set limit the in case where the limit is
272 * not page aligned -Ram Gupta
273 */
274 rlim = rlimit(RLIMIT_DATA);
275 if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
276 (mm->end_data - mm->start_data) > rlim)
277 goto out;
278
279 newbrk = PAGE_ALIGN(brk);
280 oldbrk = PAGE_ALIGN(mm->brk);
281 if (oldbrk == newbrk)
282 goto set_brk;
283
284 /* Always allow shrinking brk. */
285 if (brk <= mm->brk) {
286 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
287 goto set_brk;
288 goto out;
289 }
290
291 /* Check against existing mmap mappings. */
292 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
293 goto out;
294
295 /* Ok, looks good - let it rip. */
296 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
297 goto out;
298set_brk:
299 mm->brk = brk;
300out:
301 retval = mm->brk;
302 up_write(&mm->mmap_sem);
303 return retval;
304}
305
306#ifdef DEBUG_MM_RB
307static int browse_rb(struct rb_root *root)
308{
309 int i = 0, j;
310 struct rb_node *nd, *pn = NULL;
311 unsigned long prev = 0, pend = 0;
312
313 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
314 struct vm_area_struct *vma;
315 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
316 if (vma->vm_start < prev)
317 printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
318 if (vma->vm_start < pend)
319 printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
320 if (vma->vm_start > vma->vm_end)
321 printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
322 i++;
323 pn = nd;
324 prev = vma->vm_start;
325 pend = vma->vm_end;
326 }
327 j = 0;
328 for (nd = pn; nd; nd = rb_prev(nd)) {
329 j++;
330 }
331 if (i != j)
332 printk("backwards %d, forwards %d\n", j, i), i = 0;
333 return i;
334}
335
336void validate_mm(struct mm_struct *mm)
337{
338 int bug = 0;
339 int i = 0;
340 struct vm_area_struct *tmp = mm->mmap;
341 while (tmp) {
342 tmp = tmp->vm_next;
343 i++;
344 }
345 if (i != mm->map_count)
346 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
347 i = browse_rb(&mm->mm_rb);
348 if (i != mm->map_count)
349 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
350 BUG_ON(bug);
351}
352#else
353#define validate_mm(mm) do { } while (0)
354#endif
355
356static struct vm_area_struct *
357find_vma_prepare(struct mm_struct *mm, unsigned long addr,
358 struct vm_area_struct **pprev, struct rb_node ***rb_link,
359 struct rb_node ** rb_parent)
360{
361 struct vm_area_struct * vma;
362 struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
363
364 __rb_link = &mm->mm_rb.rb_node;
365 rb_prev = __rb_parent = NULL;
366 vma = NULL;
367
368 while (*__rb_link) {
369 struct vm_area_struct *vma_tmp;
370
371 __rb_parent = *__rb_link;
372 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
373
374 if (vma_tmp->vm_end > addr) {
375 vma = vma_tmp;
376 if (vma_tmp->vm_start <= addr)
377 break;
378 __rb_link = &__rb_parent->rb_left;
379 } else {
380 rb_prev = __rb_parent;
381 __rb_link = &__rb_parent->rb_right;
382 }
383 }
384
385 *pprev = NULL;
386 if (rb_prev)
387 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
388 *rb_link = __rb_link;
389 *rb_parent = __rb_parent;
390 return vma;
391}
392
393void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
394 struct rb_node **rb_link, struct rb_node *rb_parent)
395{
396 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
397 rb_insert_color(&vma->vm_rb, &mm->mm_rb);
398}
399
400static void __vma_link_file(struct vm_area_struct *vma)
401{
402 struct file *file;
403
404 file = vma->vm_file;
405 if (file) {
406 struct address_space *mapping = file->f_mapping;
407
408 if (vma->vm_flags & VM_DENYWRITE)
409 atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
410 if (vma->vm_flags & VM_SHARED)
411 mapping->i_mmap_writable++;
412
413 flush_dcache_mmap_lock(mapping);
414 if (unlikely(vma->vm_flags & VM_NONLINEAR))
415 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
416 else
417 vma_prio_tree_insert(vma, &mapping->i_mmap);
418 flush_dcache_mmap_unlock(mapping);
419 }
420}
421
422static void
423__vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
424 struct vm_area_struct *prev, struct rb_node **rb_link,
425 struct rb_node *rb_parent)
426{
427 __vma_link_list(mm, vma, prev, rb_parent);
428 __vma_link_rb(mm, vma, rb_link, rb_parent);
429}
430
431static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
432 struct vm_area_struct *prev, struct rb_node **rb_link,
433 struct rb_node *rb_parent)
434{
435 struct address_space *mapping = NULL;
436
437 if (vma->vm_file)
438 mapping = vma->vm_file->f_mapping;
439
440 if (mapping)
441 mutex_lock(&mapping->i_mmap_mutex);
442
443 __vma_link(mm, vma, prev, rb_link, rb_parent);
444 __vma_link_file(vma);
445
446 if (mapping)
447 mutex_unlock(&mapping->i_mmap_mutex);
448
449 mm->map_count++;
450 validate_mm(mm);
451}
452
453/*
454 * Helper for vma_adjust in the split_vma insert case:
455 * insert vm structure into list and rbtree and anon_vma,
456 * but it has already been inserted into prio_tree earlier.
457 */
458static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
459{
460 struct vm_area_struct *__vma, *prev;
461 struct rb_node **rb_link, *rb_parent;
462
463 __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
464 BUG_ON(__vma && __vma->vm_start < vma->vm_end);
465 __vma_link(mm, vma, prev, rb_link, rb_parent);
466 mm->map_count++;
467}
468
469static inline void
470__vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
471 struct vm_area_struct *prev)
472{
473 struct vm_area_struct *next = vma->vm_next;
474
475 prev->vm_next = next;
476 if (next)
477 next->vm_prev = prev;
478 rb_erase(&vma->vm_rb, &mm->mm_rb);
479 if (mm->mmap_cache == vma)
480 mm->mmap_cache = prev;
481}
482
483/*
484 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
485 * is already present in an i_mmap tree without adjusting the tree.
486 * The following helper function should be used when such adjustments
487 * are necessary. The "insert" vma (if any) is to be inserted
488 * before we drop the necessary locks.
489 */
490int vma_adjust(struct vm_area_struct *vma, unsigned long start,
491 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
492{
493 struct mm_struct *mm = vma->vm_mm;
494 struct vm_area_struct *next = vma->vm_next;
495 struct vm_area_struct *importer = NULL;
496 struct address_space *mapping = NULL;
497 struct prio_tree_root *root = NULL;
498 struct anon_vma *anon_vma = NULL;
499 struct file *file = vma->vm_file;
500 long adjust_next = 0;
501 int remove_next = 0;
502
503 if (next && !insert) {
504 struct vm_area_struct *exporter = NULL;
505
506 if (end >= next->vm_end) {
507 /*
508 * vma expands, overlapping all the next, and
509 * perhaps the one after too (mprotect case 6).
510 */
511again: remove_next = 1 + (end > next->vm_end);
512 end = next->vm_end;
513 exporter = next;
514 importer = vma;
515 } else if (end > next->vm_start) {
516 /*
517 * vma expands, overlapping part of the next:
518 * mprotect case 5 shifting the boundary up.
519 */
520 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
521 exporter = next;
522 importer = vma;
523 } else if (end < vma->vm_end) {
524 /*
525 * vma shrinks, and !insert tells it's not
526 * split_vma inserting another: so it must be
527 * mprotect case 4 shifting the boundary down.
528 */
529 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
530 exporter = vma;
531 importer = next;
532 }
533
534 /*
535 * Easily overlooked: when mprotect shifts the boundary,
536 * make sure the expanding vma has anon_vma set if the
537 * shrinking vma had, to cover any anon pages imported.
538 */
539 if (exporter && exporter->anon_vma && !importer->anon_vma) {
540 if (anon_vma_clone(importer, exporter))
541 return -ENOMEM;
542 importer->anon_vma = exporter->anon_vma;
543 }
544 }
545
546 if (file) {
547 mapping = file->f_mapping;
548 if (!(vma->vm_flags & VM_NONLINEAR))
549 root = &mapping->i_mmap;
550 mutex_lock(&mapping->i_mmap_mutex);
551 if (insert) {
552 /*
553 * Put into prio_tree now, so instantiated pages
554 * are visible to arm/parisc __flush_dcache_page
555 * throughout; but we cannot insert into address
556 * space until vma start or end is updated.
557 */
558 __vma_link_file(insert);
559 }
560 }
561
562 vma_adjust_trans_huge(vma, start, end, adjust_next);
563
564 /*
565 * When changing only vma->vm_end, we don't really need anon_vma
566 * lock. This is a fairly rare case by itself, but the anon_vma
567 * lock may be shared between many sibling processes. Skipping
568 * the lock for brk adjustments makes a difference sometimes.
569 */
570 if (vma->anon_vma && (importer || start != vma->vm_start)) {
571 anon_vma = vma->anon_vma;
572 anon_vma_lock(anon_vma);
573 }
574
575 if (root) {
576 flush_dcache_mmap_lock(mapping);
577 vma_prio_tree_remove(vma, root);
578 if (adjust_next)
579 vma_prio_tree_remove(next, root);
580 }
581
582 vma->vm_start = start;
583 vma->vm_end = end;
584 vma->vm_pgoff = pgoff;
585 if (adjust_next) {
586 next->vm_start += adjust_next << PAGE_SHIFT;
587 next->vm_pgoff += adjust_next;
588 }
589
590 if (root) {
591 if (adjust_next)
592 vma_prio_tree_insert(next, root);
593 vma_prio_tree_insert(vma, root);
594 flush_dcache_mmap_unlock(mapping);
595 }
596
597 if (remove_next) {
598 /*
599 * vma_merge has merged next into vma, and needs
600 * us to remove next before dropping the locks.
601 */
602 __vma_unlink(mm, next, vma);
603 if (file)
604 __remove_shared_vm_struct(next, file, mapping);
605 } else if (insert) {
606 /*
607 * split_vma has split insert from vma, and needs
608 * us to insert it before dropping the locks
609 * (it may either follow vma or precede it).
610 */
611 __insert_vm_struct(mm, insert);
612 }
613
614 if (anon_vma)
615 anon_vma_unlock(anon_vma);
616 if (mapping)
617 mutex_unlock(&mapping->i_mmap_mutex);
618
619 if (remove_next) {
620 if (file) {
621 fput(file);
622 if (next->vm_flags & VM_EXECUTABLE)
623 removed_exe_file_vma(mm);
624 }
625 if (next->anon_vma)
626 anon_vma_merge(vma, next);
627 mm->map_count--;
628 mpol_put(vma_policy(next));
629 kmem_cache_free(vm_area_cachep, next);
630 /*
631 * In mprotect's case 6 (see comments on vma_merge),
632 * we must remove another next too. It would clutter
633 * up the code too much to do both in one go.
634 */
635 if (remove_next == 2) {
636 next = vma->vm_next;
637 goto again;
638 }
639 }
640
641 validate_mm(mm);
642
643 return 0;
644}
645
646/*
647 * If the vma has a ->close operation then the driver probably needs to release
648 * per-vma resources, so we don't attempt to merge those.
649 */
650static inline int is_mergeable_vma(struct vm_area_struct *vma,
651 struct file *file, unsigned long vm_flags)
652{
653 /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
654 if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
655 return 0;
656 if (vma->vm_file != file)
657 return 0;
658 if (vma->vm_ops && vma->vm_ops->close)
659 return 0;
660 return 1;
661}
662
663static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
664 struct anon_vma *anon_vma2,
665 struct vm_area_struct *vma)
666{
667 /*
668 * The list_is_singular() test is to avoid merging VMA cloned from
669 * parents. This can improve scalability caused by anon_vma lock.
670 */
671 if ((!anon_vma1 || !anon_vma2) && (!vma ||
672 list_is_singular(&vma->anon_vma_chain)))
673 return 1;
674 return anon_vma1 == anon_vma2;
675}
676
677/*
678 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
679 * in front of (at a lower virtual address and file offset than) the vma.
680 *
681 * We cannot merge two vmas if they have differently assigned (non-NULL)
682 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
683 *
684 * We don't check here for the merged mmap wrapping around the end of pagecache
685 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
686 * wrap, nor mmaps which cover the final page at index -1UL.
687 */
688static int
689can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
690 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
691{
692 if (is_mergeable_vma(vma, file, vm_flags) &&
693 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
694 if (vma->vm_pgoff == vm_pgoff)
695 return 1;
696 }
697 return 0;
698}
699
700/*
701 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
702 * beyond (at a higher virtual address and file offset than) the vma.
703 *
704 * We cannot merge two vmas if they have differently assigned (non-NULL)
705 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
706 */
707static int
708can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
709 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
710{
711 if (is_mergeable_vma(vma, file, vm_flags) &&
712 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
713 pgoff_t vm_pglen;
714 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
715 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
716 return 1;
717 }
718 return 0;
719}
720
721/*
722 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
723 * whether that can be merged with its predecessor or its successor.
724 * Or both (it neatly fills a hole).
725 *
726 * In most cases - when called for mmap, brk or mremap - [addr,end) is
727 * certain not to be mapped by the time vma_merge is called; but when
728 * called for mprotect, it is certain to be already mapped (either at
729 * an offset within prev, or at the start of next), and the flags of
730 * this area are about to be changed to vm_flags - and the no-change
731 * case has already been eliminated.
732 *
733 * The following mprotect cases have to be considered, where AAAA is
734 * the area passed down from mprotect_fixup, never extending beyond one
735 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
736 *
737 * AAAA AAAA AAAA AAAA
738 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
739 * cannot merge might become might become might become
740 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
741 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
742 * mremap move: PPPPNNNNNNNN 8
743 * AAAA
744 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
745 * might become case 1 below case 2 below case 3 below
746 *
747 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
748 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
749 */
750struct vm_area_struct *vma_merge(struct mm_struct *mm,
751 struct vm_area_struct *prev, unsigned long addr,
752 unsigned long end, unsigned long vm_flags,
753 struct anon_vma *anon_vma, struct file *file,
754 pgoff_t pgoff, struct mempolicy *policy)
755{
756 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
757 struct vm_area_struct *area, *next;
758 int err;
759
760 /*
761 * We later require that vma->vm_flags == vm_flags,
762 * so this tests vma->vm_flags & VM_SPECIAL, too.
763 */
764 if (vm_flags & VM_SPECIAL)
765 return NULL;
766
767 if (prev)
768 next = prev->vm_next;
769 else
770 next = mm->mmap;
771 area = next;
772 if (next && next->vm_end == end) /* cases 6, 7, 8 */
773 next = next->vm_next;
774
775 /*
776 * Can it merge with the predecessor?
777 */
778 if (prev && prev->vm_end == addr &&
779 mpol_equal(vma_policy(prev), policy) &&
780 can_vma_merge_after(prev, vm_flags,
781 anon_vma, file, pgoff)) {
782 /*
783 * OK, it can. Can we now merge in the successor as well?
784 */
785 if (next && end == next->vm_start &&
786 mpol_equal(policy, vma_policy(next)) &&
787 can_vma_merge_before(next, vm_flags,
788 anon_vma, file, pgoff+pglen) &&
789 is_mergeable_anon_vma(prev->anon_vma,
790 next->anon_vma, NULL)) {
791 /* cases 1, 6 */
792 err = vma_adjust(prev, prev->vm_start,
793 next->vm_end, prev->vm_pgoff, NULL);
794 } else /* cases 2, 5, 7 */
795 err = vma_adjust(prev, prev->vm_start,
796 end, prev->vm_pgoff, NULL);
797 if (err)
798 return NULL;
799 khugepaged_enter_vma_merge(prev);
800 return prev;
801 }
802
803 /*
804 * Can this new request be merged in front of next?
805 */
806 if (next && end == next->vm_start &&
807 mpol_equal(policy, vma_policy(next)) &&
808 can_vma_merge_before(next, vm_flags,
809 anon_vma, file, pgoff+pglen)) {
810 if (prev && addr < prev->vm_end) /* case 4 */
811 err = vma_adjust(prev, prev->vm_start,
812 addr, prev->vm_pgoff, NULL);
813 else /* cases 3, 8 */
814 err = vma_adjust(area, addr, next->vm_end,
815 next->vm_pgoff - pglen, NULL);
816 if (err)
817 return NULL;
818 khugepaged_enter_vma_merge(area);
819 return area;
820 }
821
822 return NULL;
823}
824
825/*
826 * Rough compatbility check to quickly see if it's even worth looking
827 * at sharing an anon_vma.
828 *
829 * They need to have the same vm_file, and the flags can only differ
830 * in things that mprotect may change.
831 *
832 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
833 * we can merge the two vma's. For example, we refuse to merge a vma if
834 * there is a vm_ops->close() function, because that indicates that the
835 * driver is doing some kind of reference counting. But that doesn't
836 * really matter for the anon_vma sharing case.
837 */
838static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
839{
840 return a->vm_end == b->vm_start &&
841 mpol_equal(vma_policy(a), vma_policy(b)) &&
842 a->vm_file == b->vm_file &&
843 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
844 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
845}
846
847/*
848 * Do some basic sanity checking to see if we can re-use the anon_vma
849 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
850 * the same as 'old', the other will be the new one that is trying
851 * to share the anon_vma.
852 *
853 * NOTE! This runs with mm_sem held for reading, so it is possible that
854 * the anon_vma of 'old' is concurrently in the process of being set up
855 * by another page fault trying to merge _that_. But that's ok: if it
856 * is being set up, that automatically means that it will be a singleton
857 * acceptable for merging, so we can do all of this optimistically. But
858 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
859 *
860 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
861 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
862 * is to return an anon_vma that is "complex" due to having gone through
863 * a fork).
864 *
865 * We also make sure that the two vma's are compatible (adjacent,
866 * and with the same memory policies). That's all stable, even with just
867 * a read lock on the mm_sem.
868 */
869static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
870{
871 if (anon_vma_compatible(a, b)) {
872 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
873
874 if (anon_vma && list_is_singular(&old->anon_vma_chain))
875 return anon_vma;
876 }
877 return NULL;
878}
879
880/*
881 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
882 * neighbouring vmas for a suitable anon_vma, before it goes off
883 * to allocate a new anon_vma. It checks because a repetitive
884 * sequence of mprotects and faults may otherwise lead to distinct
885 * anon_vmas being allocated, preventing vma merge in subsequent
886 * mprotect.
887 */
888struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
889{
890 struct anon_vma *anon_vma;
891 struct vm_area_struct *near;
892
893 near = vma->vm_next;
894 if (!near)
895 goto try_prev;
896
897 anon_vma = reusable_anon_vma(near, vma, near);
898 if (anon_vma)
899 return anon_vma;
900try_prev:
901 near = vma->vm_prev;
902 if (!near)
903 goto none;
904
905 anon_vma = reusable_anon_vma(near, near, vma);
906 if (anon_vma)
907 return anon_vma;
908none:
909 /*
910 * There's no absolute need to look only at touching neighbours:
911 * we could search further afield for "compatible" anon_vmas.
912 * But it would probably just be a waste of time searching,
913 * or lead to too many vmas hanging off the same anon_vma.
914 * We're trying to allow mprotect remerging later on,
915 * not trying to minimize memory used for anon_vmas.
916 */
917 return NULL;
918}
919
920#ifdef CONFIG_PROC_FS
921void vm_stat_account(struct mm_struct *mm, unsigned long flags,
922 struct file *file, long pages)
923{
924 const unsigned long stack_flags
925 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
926
927 if (file) {
928 mm->shared_vm += pages;
929 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
930 mm->exec_vm += pages;
931 } else if (flags & stack_flags)
932 mm->stack_vm += pages;
933 if (flags & (VM_RESERVED|VM_IO))
934 mm->reserved_vm += pages;
935}
936#endif /* CONFIG_PROC_FS */
937
938/*
939 * The caller must hold down_write(¤t->mm->mmap_sem).
940 */
941
942unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
943 unsigned long len, unsigned long prot,
944 unsigned long flags, unsigned long pgoff)
945{
946 struct mm_struct * mm = current->mm;
947 struct inode *inode;
948 vm_flags_t vm_flags;
949 int error;
950 unsigned long reqprot = prot;
951
952 /*
953 * Does the application expect PROT_READ to imply PROT_EXEC?
954 *
955 * (the exception is when the underlying filesystem is noexec
956 * mounted, in which case we dont add PROT_EXEC.)
957 */
958 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
959 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
960 prot |= PROT_EXEC;
961
962 if (!len)
963 return -EINVAL;
964
965 if (!(flags & MAP_FIXED))
966 addr = round_hint_to_min(addr);
967
968 /* Careful about overflows.. */
969 len = PAGE_ALIGN(len);
970 if (!len)
971 return -ENOMEM;
972
973 /* offset overflow? */
974 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
975 return -EOVERFLOW;
976
977 /* Too many mappings? */
978 if (mm->map_count > sysctl_max_map_count)
979 return -ENOMEM;
980
981 /* Obtain the address to map to. we verify (or select) it and ensure
982 * that it represents a valid section of the address space.
983 */
984 addr = get_unmapped_area(file, addr, len, pgoff, flags);
985 if (addr & ~PAGE_MASK)
986 return addr;
987
988 /* Do simple checking here so the lower-level routines won't have
989 * to. we assume access permissions have been handled by the open
990 * of the memory object, so we don't do any here.
991 */
992 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
993 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
994
995 if (flags & MAP_LOCKED)
996 if (!can_do_mlock())
997 return -EPERM;
998
999 /* mlock MCL_FUTURE? */
1000 if (vm_flags & VM_LOCKED) {
1001 unsigned long locked, lock_limit;
1002 locked = len >> PAGE_SHIFT;
1003 locked += mm->locked_vm;
1004 lock_limit = rlimit(RLIMIT_MEMLOCK);
1005 lock_limit >>= PAGE_SHIFT;
1006 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1007 return -EAGAIN;
1008 }
1009
1010 inode = file ? file->f_path.dentry->d_inode : NULL;
1011
1012 if (file) {
1013 switch (flags & MAP_TYPE) {
1014 case MAP_SHARED:
1015 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1016 return -EACCES;
1017
1018 /*
1019 * Make sure we don't allow writing to an append-only
1020 * file..
1021 */
1022 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1023 return -EACCES;
1024
1025 /*
1026 * Make sure there are no mandatory locks on the file.
1027 */
1028 if (locks_verify_locked(inode))
1029 return -EAGAIN;
1030
1031 vm_flags |= VM_SHARED | VM_MAYSHARE;
1032 if (!(file->f_mode & FMODE_WRITE))
1033 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1034
1035 /* fall through */
1036 case MAP_PRIVATE:
1037 if (!(file->f_mode & FMODE_READ))
1038 return -EACCES;
1039 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1040 if (vm_flags & VM_EXEC)
1041 return -EPERM;
1042 vm_flags &= ~VM_MAYEXEC;
1043 }
1044
1045 if (!file->f_op || !file->f_op->mmap)
1046 return -ENODEV;
1047 break;
1048
1049 default:
1050 return -EINVAL;
1051 }
1052 } else {
1053 switch (flags & MAP_TYPE) {
1054 case MAP_SHARED:
1055 /*
1056 * Ignore pgoff.
1057 */
1058 pgoff = 0;
1059 vm_flags |= VM_SHARED | VM_MAYSHARE;
1060 break;
1061 case MAP_PRIVATE:
1062 /*
1063 * Set pgoff according to addr for anon_vma.
1064 */
1065 pgoff = addr >> PAGE_SHIFT;
1066 break;
1067 default:
1068 return -EINVAL;
1069 }
1070 }
1071
1072 error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1073 if (error)
1074 return error;
1075
1076 return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1077}
1078EXPORT_SYMBOL(do_mmap_pgoff);
1079
1080SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1081 unsigned long, prot, unsigned long, flags,
1082 unsigned long, fd, unsigned long, pgoff)
1083{
1084 struct file *file = NULL;
1085 unsigned long retval = -EBADF;
1086
1087 if (!(flags & MAP_ANONYMOUS)) {
1088 audit_mmap_fd(fd, flags);
1089 if (unlikely(flags & MAP_HUGETLB))
1090 return -EINVAL;
1091 file = fget(fd);
1092 if (!file)
1093 goto out;
1094 } else if (flags & MAP_HUGETLB) {
1095 struct user_struct *user = NULL;
1096 /*
1097 * VM_NORESERVE is used because the reservations will be
1098 * taken when vm_ops->mmap() is called
1099 * A dummy user value is used because we are not locking
1100 * memory so no accounting is necessary
1101 */
1102 len = ALIGN(len, huge_page_size(&default_hstate));
1103 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
1104 &user, HUGETLB_ANONHUGE_INODE);
1105 if (IS_ERR(file))
1106 return PTR_ERR(file);
1107 }
1108
1109 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1110
1111 down_write(¤t->mm->mmap_sem);
1112 retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1113 up_write(¤t->mm->mmap_sem);
1114
1115 if (file)
1116 fput(file);
1117out:
1118 return retval;
1119}
1120
1121#ifdef __ARCH_WANT_SYS_OLD_MMAP
1122struct mmap_arg_struct {
1123 unsigned long addr;
1124 unsigned long len;
1125 unsigned long prot;
1126 unsigned long flags;
1127 unsigned long fd;
1128 unsigned long offset;
1129};
1130
1131SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1132{
1133 struct mmap_arg_struct a;
1134
1135 if (copy_from_user(&a, arg, sizeof(a)))
1136 return -EFAULT;
1137 if (a.offset & ~PAGE_MASK)
1138 return -EINVAL;
1139
1140 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1141 a.offset >> PAGE_SHIFT);
1142}
1143#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1144
1145/*
1146 * Some shared mappigns will want the pages marked read-only
1147 * to track write events. If so, we'll downgrade vm_page_prot
1148 * to the private version (using protection_map[] without the
1149 * VM_SHARED bit).
1150 */
1151int vma_wants_writenotify(struct vm_area_struct *vma)
1152{
1153 vm_flags_t vm_flags = vma->vm_flags;
1154
1155 /* If it was private or non-writable, the write bit is already clear */
1156 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1157 return 0;
1158
1159 /* The backer wishes to know when pages are first written to? */
1160 if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1161 return 1;
1162
1163 /* The open routine did something to the protections already? */
1164 if (pgprot_val(vma->vm_page_prot) !=
1165 pgprot_val(vm_get_page_prot(vm_flags)))
1166 return 0;
1167
1168 /* Specialty mapping? */
1169 if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1170 return 0;
1171
1172 /* Can the mapping track the dirty pages? */
1173 return vma->vm_file && vma->vm_file->f_mapping &&
1174 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1175}
1176
1177/*
1178 * We account for memory if it's a private writeable mapping,
1179 * not hugepages and VM_NORESERVE wasn't set.
1180 */
1181static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1182{
1183 /*
1184 * hugetlb has its own accounting separate from the core VM
1185 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1186 */
1187 if (file && is_file_hugepages(file))
1188 return 0;
1189
1190 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1191}
1192
1193unsigned long mmap_region(struct file *file, unsigned long addr,
1194 unsigned long len, unsigned long flags,
1195 vm_flags_t vm_flags, unsigned long pgoff)
1196{
1197 struct mm_struct *mm = current->mm;
1198 struct vm_area_struct *vma, *prev;
1199 int correct_wcount = 0;
1200 int error;
1201 struct rb_node **rb_link, *rb_parent;
1202 unsigned long charged = 0;
1203 struct inode *inode = file ? file->f_path.dentry->d_inode : NULL;
1204
1205 /* Clear old maps */
1206 error = -ENOMEM;
1207munmap_back:
1208 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1209 if (vma && vma->vm_start < addr + len) {
1210 if (do_munmap(mm, addr, len))
1211 return -ENOMEM;
1212 goto munmap_back;
1213 }
1214
1215 /* Check against address space limit. */
1216 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1217 return -ENOMEM;
1218
1219 /*
1220 * Set 'VM_NORESERVE' if we should not account for the
1221 * memory use of this mapping.
1222 */
1223 if ((flags & MAP_NORESERVE)) {
1224 /* We honor MAP_NORESERVE if allowed to overcommit */
1225 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1226 vm_flags |= VM_NORESERVE;
1227
1228 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1229 if (file && is_file_hugepages(file))
1230 vm_flags |= VM_NORESERVE;
1231 }
1232
1233 /*
1234 * Private writable mapping: check memory availability
1235 */
1236 if (accountable_mapping(file, vm_flags)) {
1237 charged = len >> PAGE_SHIFT;
1238 if (security_vm_enough_memory(charged))
1239 return -ENOMEM;
1240 vm_flags |= VM_ACCOUNT;
1241 }
1242
1243 /*
1244 * Can we just expand an old mapping?
1245 */
1246 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1247 if (vma)
1248 goto out;
1249
1250 /*
1251 * Determine the object being mapped and call the appropriate
1252 * specific mapper. the address has already been validated, but
1253 * not unmapped, but the maps are removed from the list.
1254 */
1255 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1256 if (!vma) {
1257 error = -ENOMEM;
1258 goto unacct_error;
1259 }
1260
1261 vma->vm_mm = mm;
1262 vma->vm_start = addr;
1263 vma->vm_end = addr + len;
1264 vma->vm_flags = vm_flags;
1265 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1266 vma->vm_pgoff = pgoff;
1267 INIT_LIST_HEAD(&vma->anon_vma_chain);
1268
1269 if (file) {
1270 error = -EINVAL;
1271 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1272 goto free_vma;
1273 if (vm_flags & VM_DENYWRITE) {
1274 error = deny_write_access(file);
1275 if (error)
1276 goto free_vma;
1277 correct_wcount = 1;
1278 }
1279 vma->vm_file = file;
1280 get_file(file);
1281 error = file->f_op->mmap(file, vma);
1282 if (error)
1283 goto unmap_and_free_vma;
1284 if (vm_flags & VM_EXECUTABLE)
1285 added_exe_file_vma(mm);
1286
1287 /* Can addr have changed??
1288 *
1289 * Answer: Yes, several device drivers can do it in their
1290 * f_op->mmap method. -DaveM
1291 */
1292 addr = vma->vm_start;
1293 pgoff = vma->vm_pgoff;
1294 vm_flags = vma->vm_flags;
1295 } else if (vm_flags & VM_SHARED) {
1296 error = shmem_zero_setup(vma);
1297 if (error)
1298 goto free_vma;
1299 }
1300
1301 if (vma_wants_writenotify(vma)) {
1302 pgprot_t pprot = vma->vm_page_prot;
1303
1304 /* Can vma->vm_page_prot have changed??
1305 *
1306 * Answer: Yes, drivers may have changed it in their
1307 * f_op->mmap method.
1308 *
1309 * Ensures that vmas marked as uncached stay that way.
1310 */
1311 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1312 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1313 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1314 }
1315
1316 vma_link(mm, vma, prev, rb_link, rb_parent);
1317 file = vma->vm_file;
1318
1319 /* Once vma denies write, undo our temporary denial count */
1320 if (correct_wcount)
1321 atomic_inc(&inode->i_writecount);
1322out:
1323 perf_event_mmap(vma);
1324
1325 mm->total_vm += len >> PAGE_SHIFT;
1326 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1327 if (vm_flags & VM_LOCKED) {
1328 if (!mlock_vma_pages_range(vma, addr, addr + len))
1329 mm->locked_vm += (len >> PAGE_SHIFT);
1330 } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1331 make_pages_present(addr, addr + len);
1332 return addr;
1333
1334unmap_and_free_vma:
1335 if (correct_wcount)
1336 atomic_inc(&inode->i_writecount);
1337 vma->vm_file = NULL;
1338 fput(file);
1339
1340 /* Undo any partial mapping done by a device driver. */
1341 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1342 charged = 0;
1343free_vma:
1344 kmem_cache_free(vm_area_cachep, vma);
1345unacct_error:
1346 if (charged)
1347 vm_unacct_memory(charged);
1348 return error;
1349}
1350
1351/* Get an address range which is currently unmapped.
1352 * For shmat() with addr=0.
1353 *
1354 * Ugly calling convention alert:
1355 * Return value with the low bits set means error value,
1356 * ie
1357 * if (ret & ~PAGE_MASK)
1358 * error = ret;
1359 *
1360 * This function "knows" that -ENOMEM has the bits set.
1361 */
1362#ifndef HAVE_ARCH_UNMAPPED_AREA
1363unsigned long
1364arch_get_unmapped_area(struct file *filp, unsigned long addr,
1365 unsigned long len, unsigned long pgoff, unsigned long flags)
1366{
1367 struct mm_struct *mm = current->mm;
1368 struct vm_area_struct *vma;
1369 unsigned long start_addr;
1370
1371 if (len > TASK_SIZE)
1372 return -ENOMEM;
1373
1374 if (flags & MAP_FIXED)
1375 return addr;
1376
1377 if (addr) {
1378 addr = PAGE_ALIGN(addr);
1379 vma = find_vma(mm, addr);
1380 if (TASK_SIZE - len >= addr &&
1381 (!vma || addr + len <= vma->vm_start))
1382 return addr;
1383 }
1384 if (len > mm->cached_hole_size) {
1385 start_addr = addr = mm->free_area_cache;
1386 } else {
1387 start_addr = addr = TASK_UNMAPPED_BASE;
1388 mm->cached_hole_size = 0;
1389 }
1390
1391full_search:
1392 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1393 /* At this point: (!vma || addr < vma->vm_end). */
1394 if (TASK_SIZE - len < addr) {
1395 /*
1396 * Start a new search - just in case we missed
1397 * some holes.
1398 */
1399 if (start_addr != TASK_UNMAPPED_BASE) {
1400 addr = TASK_UNMAPPED_BASE;
1401 start_addr = addr;
1402 mm->cached_hole_size = 0;
1403 goto full_search;
1404 }
1405 return -ENOMEM;
1406 }
1407 if (!vma || addr + len <= vma->vm_start) {
1408 /*
1409 * Remember the place where we stopped the search:
1410 */
1411 mm->free_area_cache = addr + len;
1412 return addr;
1413 }
1414 if (addr + mm->cached_hole_size < vma->vm_start)
1415 mm->cached_hole_size = vma->vm_start - addr;
1416 addr = vma->vm_end;
1417 }
1418}
1419#endif
1420
1421void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1422{
1423 /*
1424 * Is this a new hole at the lowest possible address?
1425 */
1426 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1427 mm->free_area_cache = addr;
1428 mm->cached_hole_size = ~0UL;
1429 }
1430}
1431
1432/*
1433 * This mmap-allocator allocates new areas top-down from below the
1434 * stack's low limit (the base):
1435 */
1436#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1437unsigned long
1438arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1439 const unsigned long len, const unsigned long pgoff,
1440 const unsigned long flags)
1441{
1442 struct vm_area_struct *vma;
1443 struct mm_struct *mm = current->mm;
1444 unsigned long addr = addr0;
1445
1446 /* requested length too big for entire address space */
1447 if (len > TASK_SIZE)
1448 return -ENOMEM;
1449
1450 if (flags & MAP_FIXED)
1451 return addr;
1452
1453 /* requesting a specific address */
1454 if (addr) {
1455 addr = PAGE_ALIGN(addr);
1456 vma = find_vma(mm, addr);
1457 if (TASK_SIZE - len >= addr &&
1458 (!vma || addr + len <= vma->vm_start))
1459 return addr;
1460 }
1461
1462 /* check if free_area_cache is useful for us */
1463 if (len <= mm->cached_hole_size) {
1464 mm->cached_hole_size = 0;
1465 mm->free_area_cache = mm->mmap_base;
1466 }
1467
1468 /* either no address requested or can't fit in requested address hole */
1469 addr = mm->free_area_cache;
1470
1471 /* make sure it can fit in the remaining address space */
1472 if (addr > len) {
1473 vma = find_vma(mm, addr-len);
1474 if (!vma || addr <= vma->vm_start)
1475 /* remember the address as a hint for next time */
1476 return (mm->free_area_cache = addr-len);
1477 }
1478
1479 if (mm->mmap_base < len)
1480 goto bottomup;
1481
1482 addr = mm->mmap_base-len;
1483
1484 do {
1485 /*
1486 * Lookup failure means no vma is above this address,
1487 * else if new region fits below vma->vm_start,
1488 * return with success:
1489 */
1490 vma = find_vma(mm, addr);
1491 if (!vma || addr+len <= vma->vm_start)
1492 /* remember the address as a hint for next time */
1493 return (mm->free_area_cache = addr);
1494
1495 /* remember the largest hole we saw so far */
1496 if (addr + mm->cached_hole_size < vma->vm_start)
1497 mm->cached_hole_size = vma->vm_start - addr;
1498
1499 /* try just below the current vma->vm_start */
1500 addr = vma->vm_start-len;
1501 } while (len < vma->vm_start);
1502
1503bottomup:
1504 /*
1505 * A failed mmap() very likely causes application failure,
1506 * so fall back to the bottom-up function here. This scenario
1507 * can happen with large stack limits and large mmap()
1508 * allocations.
1509 */
1510 mm->cached_hole_size = ~0UL;
1511 mm->free_area_cache = TASK_UNMAPPED_BASE;
1512 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1513 /*
1514 * Restore the topdown base:
1515 */
1516 mm->free_area_cache = mm->mmap_base;
1517 mm->cached_hole_size = ~0UL;
1518
1519 return addr;
1520}
1521#endif
1522
1523void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1524{
1525 /*
1526 * Is this a new hole at the highest possible address?
1527 */
1528 if (addr > mm->free_area_cache)
1529 mm->free_area_cache = addr;
1530
1531 /* dont allow allocations above current base */
1532 if (mm->free_area_cache > mm->mmap_base)
1533 mm->free_area_cache = mm->mmap_base;
1534}
1535
1536unsigned long
1537get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1538 unsigned long pgoff, unsigned long flags)
1539{
1540 unsigned long (*get_area)(struct file *, unsigned long,
1541 unsigned long, unsigned long, unsigned long);
1542
1543 unsigned long error = arch_mmap_check(addr, len, flags);
1544 if (error)
1545 return error;
1546
1547 /* Careful about overflows.. */
1548 if (len > TASK_SIZE)
1549 return -ENOMEM;
1550
1551 get_area = current->mm->get_unmapped_area;
1552 if (file && file->f_op && file->f_op->get_unmapped_area)
1553 get_area = file->f_op->get_unmapped_area;
1554 addr = get_area(file, addr, len, pgoff, flags);
1555 if (IS_ERR_VALUE(addr))
1556 return addr;
1557
1558 if (addr > TASK_SIZE - len)
1559 return -ENOMEM;
1560 if (addr & ~PAGE_MASK)
1561 return -EINVAL;
1562
1563 return arch_rebalance_pgtables(addr, len);
1564}
1565
1566EXPORT_SYMBOL(get_unmapped_area);
1567
1568/* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1569struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1570{
1571 struct vm_area_struct *vma = NULL;
1572
1573 if (mm) {
1574 /* Check the cache first. */
1575 /* (Cache hit rate is typically around 35%.) */
1576 vma = mm->mmap_cache;
1577 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1578 struct rb_node * rb_node;
1579
1580 rb_node = mm->mm_rb.rb_node;
1581 vma = NULL;
1582
1583 while (rb_node) {
1584 struct vm_area_struct * vma_tmp;
1585
1586 vma_tmp = rb_entry(rb_node,
1587 struct vm_area_struct, vm_rb);
1588
1589 if (vma_tmp->vm_end > addr) {
1590 vma = vma_tmp;
1591 if (vma_tmp->vm_start <= addr)
1592 break;
1593 rb_node = rb_node->rb_left;
1594 } else
1595 rb_node = rb_node->rb_right;
1596 }
1597 if (vma)
1598 mm->mmap_cache = vma;
1599 }
1600 }
1601 return vma;
1602}
1603
1604EXPORT_SYMBOL(find_vma);
1605
1606/* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1607struct vm_area_struct *
1608find_vma_prev(struct mm_struct *mm, unsigned long addr,
1609 struct vm_area_struct **pprev)
1610{
1611 struct vm_area_struct *vma = NULL, *prev = NULL;
1612 struct rb_node *rb_node;
1613 if (!mm)
1614 goto out;
1615
1616 /* Guard against addr being lower than the first VMA */
1617 vma = mm->mmap;
1618
1619 /* Go through the RB tree quickly. */
1620 rb_node = mm->mm_rb.rb_node;
1621
1622 while (rb_node) {
1623 struct vm_area_struct *vma_tmp;
1624 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1625
1626 if (addr < vma_tmp->vm_end) {
1627 rb_node = rb_node->rb_left;
1628 } else {
1629 prev = vma_tmp;
1630 if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1631 break;
1632 rb_node = rb_node->rb_right;
1633 }
1634 }
1635
1636out:
1637 *pprev = prev;
1638 return prev ? prev->vm_next : vma;
1639}
1640
1641/*
1642 * Verify that the stack growth is acceptable and
1643 * update accounting. This is shared with both the
1644 * grow-up and grow-down cases.
1645 */
1646static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1647{
1648 struct mm_struct *mm = vma->vm_mm;
1649 struct rlimit *rlim = current->signal->rlim;
1650 unsigned long new_start;
1651
1652 /* address space limit tests */
1653 if (!may_expand_vm(mm, grow))
1654 return -ENOMEM;
1655
1656 /* Stack limit test */
1657 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1658 return -ENOMEM;
1659
1660 /* mlock limit tests */
1661 if (vma->vm_flags & VM_LOCKED) {
1662 unsigned long locked;
1663 unsigned long limit;
1664 locked = mm->locked_vm + grow;
1665 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1666 limit >>= PAGE_SHIFT;
1667 if (locked > limit && !capable(CAP_IPC_LOCK))
1668 return -ENOMEM;
1669 }
1670
1671 /* Check to ensure the stack will not grow into a hugetlb-only region */
1672 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1673 vma->vm_end - size;
1674 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1675 return -EFAULT;
1676
1677 /*
1678 * Overcommit.. This must be the final test, as it will
1679 * update security statistics.
1680 */
1681 if (security_vm_enough_memory_mm(mm, grow))
1682 return -ENOMEM;
1683
1684 /* Ok, everything looks good - let it rip */
1685 mm->total_vm += grow;
1686 if (vma->vm_flags & VM_LOCKED)
1687 mm->locked_vm += grow;
1688 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1689 return 0;
1690}
1691
1692#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1693/*
1694 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1695 * vma is the last one with address > vma->vm_end. Have to extend vma.
1696 */
1697int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1698{
1699 int error;
1700
1701 if (!(vma->vm_flags & VM_GROWSUP))
1702 return -EFAULT;
1703
1704 /*
1705 * We must make sure the anon_vma is allocated
1706 * so that the anon_vma locking is not a noop.
1707 */
1708 if (unlikely(anon_vma_prepare(vma)))
1709 return -ENOMEM;
1710 vma_lock_anon_vma(vma);
1711
1712 /*
1713 * vma->vm_start/vm_end cannot change under us because the caller
1714 * is required to hold the mmap_sem in read mode. We need the
1715 * anon_vma lock to serialize against concurrent expand_stacks.
1716 * Also guard against wrapping around to address 0.
1717 */
1718 if (address < PAGE_ALIGN(address+4))
1719 address = PAGE_ALIGN(address+4);
1720 else {
1721 vma_unlock_anon_vma(vma);
1722 return -ENOMEM;
1723 }
1724 error = 0;
1725
1726 /* Somebody else might have raced and expanded it already */
1727 if (address > vma->vm_end) {
1728 unsigned long size, grow;
1729
1730 size = address - vma->vm_start;
1731 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1732
1733 error = -ENOMEM;
1734 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1735 error = acct_stack_growth(vma, size, grow);
1736 if (!error) {
1737 vma->vm_end = address;
1738 perf_event_mmap(vma);
1739 }
1740 }
1741 }
1742 vma_unlock_anon_vma(vma);
1743 khugepaged_enter_vma_merge(vma);
1744 return error;
1745}
1746#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1747
1748/*
1749 * vma is the first one with address < vma->vm_start. Have to extend vma.
1750 */
1751int expand_downwards(struct vm_area_struct *vma,
1752 unsigned long address)
1753{
1754 int error;
1755
1756 /*
1757 * We must make sure the anon_vma is allocated
1758 * so that the anon_vma locking is not a noop.
1759 */
1760 if (unlikely(anon_vma_prepare(vma)))
1761 return -ENOMEM;
1762
1763 address &= PAGE_MASK;
1764 error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1765 if (error)
1766 return error;
1767
1768 vma_lock_anon_vma(vma);
1769
1770 /*
1771 * vma->vm_start/vm_end cannot change under us because the caller
1772 * is required to hold the mmap_sem in read mode. We need the
1773 * anon_vma lock to serialize against concurrent expand_stacks.
1774 */
1775
1776 /* Somebody else might have raced and expanded it already */
1777 if (address < vma->vm_start) {
1778 unsigned long size, grow;
1779
1780 size = vma->vm_end - address;
1781 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1782
1783 error = -ENOMEM;
1784 if (grow <= vma->vm_pgoff) {
1785 error = acct_stack_growth(vma, size, grow);
1786 if (!error) {
1787 vma->vm_start = address;
1788 vma->vm_pgoff -= grow;
1789 perf_event_mmap(vma);
1790 }
1791 }
1792 }
1793 vma_unlock_anon_vma(vma);
1794 khugepaged_enter_vma_merge(vma);
1795 return error;
1796}
1797
1798#ifdef CONFIG_STACK_GROWSUP
1799int expand_stack(struct vm_area_struct *vma, unsigned long address)
1800{
1801 return expand_upwards(vma, address);
1802}
1803
1804struct vm_area_struct *
1805find_extend_vma(struct mm_struct *mm, unsigned long addr)
1806{
1807 struct vm_area_struct *vma, *prev;
1808
1809 addr &= PAGE_MASK;
1810 vma = find_vma_prev(mm, addr, &prev);
1811 if (vma && (vma->vm_start <= addr))
1812 return vma;
1813 if (!prev || expand_stack(prev, addr))
1814 return NULL;
1815 if (prev->vm_flags & VM_LOCKED) {
1816 mlock_vma_pages_range(prev, addr, prev->vm_end);
1817 }
1818 return prev;
1819}
1820#else
1821int expand_stack(struct vm_area_struct *vma, unsigned long address)
1822{
1823 return expand_downwards(vma, address);
1824}
1825
1826struct vm_area_struct *
1827find_extend_vma(struct mm_struct * mm, unsigned long addr)
1828{
1829 struct vm_area_struct * vma;
1830 unsigned long start;
1831
1832 addr &= PAGE_MASK;
1833 vma = find_vma(mm,addr);
1834 if (!vma)
1835 return NULL;
1836 if (vma->vm_start <= addr)
1837 return vma;
1838 if (!(vma->vm_flags & VM_GROWSDOWN))
1839 return NULL;
1840 start = vma->vm_start;
1841 if (expand_stack(vma, addr))
1842 return NULL;
1843 if (vma->vm_flags & VM_LOCKED) {
1844 mlock_vma_pages_range(vma, addr, start);
1845 }
1846 return vma;
1847}
1848#endif
1849
1850/*
1851 * Ok - we have the memory areas we should free on the vma list,
1852 * so release them, and do the vma updates.
1853 *
1854 * Called with the mm semaphore held.
1855 */
1856static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1857{
1858 /* Update high watermark before we lower total_vm */
1859 update_hiwater_vm(mm);
1860 do {
1861 long nrpages = vma_pages(vma);
1862
1863 mm->total_vm -= nrpages;
1864 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1865 vma = remove_vma(vma);
1866 } while (vma);
1867 validate_mm(mm);
1868}
1869
1870/*
1871 * Get rid of page table information in the indicated region.
1872 *
1873 * Called with the mm semaphore held.
1874 */
1875static void unmap_region(struct mm_struct *mm,
1876 struct vm_area_struct *vma, struct vm_area_struct *prev,
1877 unsigned long start, unsigned long end)
1878{
1879 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1880 struct mmu_gather tlb;
1881 unsigned long nr_accounted = 0;
1882
1883 lru_add_drain();
1884 tlb_gather_mmu(&tlb, mm, 0);
1885 update_hiwater_rss(mm);
1886 unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1887 vm_unacct_memory(nr_accounted);
1888 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
1889 next ? next->vm_start : 0);
1890 tlb_finish_mmu(&tlb, start, end);
1891}
1892
1893/*
1894 * Create a list of vma's touched by the unmap, removing them from the mm's
1895 * vma list as we go..
1896 */
1897static void
1898detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1899 struct vm_area_struct *prev, unsigned long end)
1900{
1901 struct vm_area_struct **insertion_point;
1902 struct vm_area_struct *tail_vma = NULL;
1903 unsigned long addr;
1904
1905 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1906 vma->vm_prev = NULL;
1907 do {
1908 rb_erase(&vma->vm_rb, &mm->mm_rb);
1909 mm->map_count--;
1910 tail_vma = vma;
1911 vma = vma->vm_next;
1912 } while (vma && vma->vm_start < end);
1913 *insertion_point = vma;
1914 if (vma)
1915 vma->vm_prev = prev;
1916 tail_vma->vm_next = NULL;
1917 if (mm->unmap_area == arch_unmap_area)
1918 addr = prev ? prev->vm_end : mm->mmap_base;
1919 else
1920 addr = vma ? vma->vm_start : mm->mmap_base;
1921 mm->unmap_area(mm, addr);
1922 mm->mmap_cache = NULL; /* Kill the cache. */
1923}
1924
1925/*
1926 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
1927 * munmap path where it doesn't make sense to fail.
1928 */
1929static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1930 unsigned long addr, int new_below)
1931{
1932 struct mempolicy *pol;
1933 struct vm_area_struct *new;
1934 int err = -ENOMEM;
1935
1936 if (is_vm_hugetlb_page(vma) && (addr &
1937 ~(huge_page_mask(hstate_vma(vma)))))
1938 return -EINVAL;
1939
1940 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1941 if (!new)
1942 goto out_err;
1943
1944 /* most fields are the same, copy all, and then fixup */
1945 *new = *vma;
1946
1947 INIT_LIST_HEAD(&new->anon_vma_chain);
1948
1949 if (new_below)
1950 new->vm_end = addr;
1951 else {
1952 new->vm_start = addr;
1953 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1954 }
1955
1956 pol = mpol_dup(vma_policy(vma));
1957 if (IS_ERR(pol)) {
1958 err = PTR_ERR(pol);
1959 goto out_free_vma;
1960 }
1961 vma_set_policy(new, pol);
1962
1963 if (anon_vma_clone(new, vma))
1964 goto out_free_mpol;
1965
1966 if (new->vm_file) {
1967 get_file(new->vm_file);
1968 if (vma->vm_flags & VM_EXECUTABLE)
1969 added_exe_file_vma(mm);
1970 }
1971
1972 if (new->vm_ops && new->vm_ops->open)
1973 new->vm_ops->open(new);
1974
1975 if (new_below)
1976 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1977 ((addr - new->vm_start) >> PAGE_SHIFT), new);
1978 else
1979 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1980
1981 /* Success. */
1982 if (!err)
1983 return 0;
1984
1985 /* Clean everything up if vma_adjust failed. */
1986 if (new->vm_ops && new->vm_ops->close)
1987 new->vm_ops->close(new);
1988 if (new->vm_file) {
1989 if (vma->vm_flags & VM_EXECUTABLE)
1990 removed_exe_file_vma(mm);
1991 fput(new->vm_file);
1992 }
1993 unlink_anon_vmas(new);
1994 out_free_mpol:
1995 mpol_put(pol);
1996 out_free_vma:
1997 kmem_cache_free(vm_area_cachep, new);
1998 out_err:
1999 return err;
2000}
2001
2002/*
2003 * Split a vma into two pieces at address 'addr', a new vma is allocated
2004 * either for the first part or the tail.
2005 */
2006int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2007 unsigned long addr, int new_below)
2008{
2009 if (mm->map_count >= sysctl_max_map_count)
2010 return -ENOMEM;
2011
2012 return __split_vma(mm, vma, addr, new_below);
2013}
2014
2015/* Munmap is split into 2 main parts -- this part which finds
2016 * what needs doing, and the areas themselves, which do the
2017 * work. This now handles partial unmappings.
2018 * Jeremy Fitzhardinge <jeremy@goop.org>
2019 */
2020int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2021{
2022 unsigned long end;
2023 struct vm_area_struct *vma, *prev, *last;
2024
2025 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2026 return -EINVAL;
2027
2028 if ((len = PAGE_ALIGN(len)) == 0)
2029 return -EINVAL;
2030
2031 /* Find the first overlapping VMA */
2032 vma = find_vma(mm, start);
2033 if (!vma)
2034 return 0;
2035 prev = vma->vm_prev;
2036 /* we have start < vma->vm_end */
2037
2038 /* if it doesn't overlap, we have nothing.. */
2039 end = start + len;
2040 if (vma->vm_start >= end)
2041 return 0;
2042
2043 /*
2044 * If we need to split any vma, do it now to save pain later.
2045 *
2046 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2047 * unmapped vm_area_struct will remain in use: so lower split_vma
2048 * places tmp vma above, and higher split_vma places tmp vma below.
2049 */
2050 if (start > vma->vm_start) {
2051 int error;
2052
2053 /*
2054 * Make sure that map_count on return from munmap() will
2055 * not exceed its limit; but let map_count go just above
2056 * its limit temporarily, to help free resources as expected.
2057 */
2058 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2059 return -ENOMEM;
2060
2061 error = __split_vma(mm, vma, start, 0);
2062 if (error)
2063 return error;
2064 prev = vma;
2065 }
2066
2067 /* Does it split the last one? */
2068 last = find_vma(mm, end);
2069 if (last && end > last->vm_start) {
2070 int error = __split_vma(mm, last, end, 1);
2071 if (error)
2072 return error;
2073 }
2074 vma = prev? prev->vm_next: mm->mmap;
2075
2076 /*
2077 * unlock any mlock()ed ranges before detaching vmas
2078 */
2079 if (mm->locked_vm) {
2080 struct vm_area_struct *tmp = vma;
2081 while (tmp && tmp->vm_start < end) {
2082 if (tmp->vm_flags & VM_LOCKED) {
2083 mm->locked_vm -= vma_pages(tmp);
2084 munlock_vma_pages_all(tmp);
2085 }
2086 tmp = tmp->vm_next;
2087 }
2088 }
2089
2090 /*
2091 * Remove the vma's, and unmap the actual pages
2092 */
2093 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2094 unmap_region(mm, vma, prev, start, end);
2095
2096 /* Fix up all other VM information */
2097 remove_vma_list(mm, vma);
2098
2099 return 0;
2100}
2101
2102EXPORT_SYMBOL(do_munmap);
2103
2104SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2105{
2106 int ret;
2107 struct mm_struct *mm = current->mm;
2108
2109 profile_munmap(addr);
2110
2111 down_write(&mm->mmap_sem);
2112 ret = do_munmap(mm, addr, len);
2113 up_write(&mm->mmap_sem);
2114 return ret;
2115}
2116
2117static inline void verify_mm_writelocked(struct mm_struct *mm)
2118{
2119#ifdef CONFIG_DEBUG_VM
2120 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2121 WARN_ON(1);
2122 up_read(&mm->mmap_sem);
2123 }
2124#endif
2125}
2126
2127/*
2128 * this is really a simplified "do_mmap". it only handles
2129 * anonymous maps. eventually we may be able to do some
2130 * brk-specific accounting here.
2131 */
2132unsigned long do_brk(unsigned long addr, unsigned long len)
2133{
2134 struct mm_struct * mm = current->mm;
2135 struct vm_area_struct * vma, * prev;
2136 unsigned long flags;
2137 struct rb_node ** rb_link, * rb_parent;
2138 pgoff_t pgoff = addr >> PAGE_SHIFT;
2139 int error;
2140
2141 len = PAGE_ALIGN(len);
2142 if (!len)
2143 return addr;
2144
2145 error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2146 if (error)
2147 return error;
2148
2149 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2150
2151 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2152 if (error & ~PAGE_MASK)
2153 return error;
2154
2155 /*
2156 * mlock MCL_FUTURE?
2157 */
2158 if (mm->def_flags & VM_LOCKED) {
2159 unsigned long locked, lock_limit;
2160 locked = len >> PAGE_SHIFT;
2161 locked += mm->locked_vm;
2162 lock_limit = rlimit(RLIMIT_MEMLOCK);
2163 lock_limit >>= PAGE_SHIFT;
2164 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2165 return -EAGAIN;
2166 }
2167
2168 /*
2169 * mm->mmap_sem is required to protect against another thread
2170 * changing the mappings in case we sleep.
2171 */
2172 verify_mm_writelocked(mm);
2173
2174 /*
2175 * Clear old maps. this also does some error checking for us
2176 */
2177 munmap_back:
2178 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2179 if (vma && vma->vm_start < addr + len) {
2180 if (do_munmap(mm, addr, len))
2181 return -ENOMEM;
2182 goto munmap_back;
2183 }
2184
2185 /* Check against address space limits *after* clearing old maps... */
2186 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2187 return -ENOMEM;
2188
2189 if (mm->map_count > sysctl_max_map_count)
2190 return -ENOMEM;
2191
2192 if (security_vm_enough_memory(len >> PAGE_SHIFT))
2193 return -ENOMEM;
2194
2195 /* Can we just expand an old private anonymous mapping? */
2196 vma = vma_merge(mm, prev, addr, addr + len, flags,
2197 NULL, NULL, pgoff, NULL);
2198 if (vma)
2199 goto out;
2200
2201 /*
2202 * create a vma struct for an anonymous mapping
2203 */
2204 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2205 if (!vma) {
2206 vm_unacct_memory(len >> PAGE_SHIFT);
2207 return -ENOMEM;
2208 }
2209
2210 INIT_LIST_HEAD(&vma->anon_vma_chain);
2211 vma->vm_mm = mm;
2212 vma->vm_start = addr;
2213 vma->vm_end = addr + len;
2214 vma->vm_pgoff = pgoff;
2215 vma->vm_flags = flags;
2216 vma->vm_page_prot = vm_get_page_prot(flags);
2217 vma_link(mm, vma, prev, rb_link, rb_parent);
2218out:
2219 perf_event_mmap(vma);
2220 mm->total_vm += len >> PAGE_SHIFT;
2221 if (flags & VM_LOCKED) {
2222 if (!mlock_vma_pages_range(vma, addr, addr + len))
2223 mm->locked_vm += (len >> PAGE_SHIFT);
2224 }
2225 return addr;
2226}
2227
2228EXPORT_SYMBOL(do_brk);
2229
2230/* Release all mmaps. */
2231void exit_mmap(struct mm_struct *mm)
2232{
2233 struct mmu_gather tlb;
2234 struct vm_area_struct *vma;
2235 unsigned long nr_accounted = 0;
2236 unsigned long end;
2237
2238 /* mm's last user has gone, and its about to be pulled down */
2239 mmu_notifier_release(mm);
2240
2241 if (mm->locked_vm) {
2242 vma = mm->mmap;
2243 while (vma) {
2244 if (vma->vm_flags & VM_LOCKED)
2245 munlock_vma_pages_all(vma);
2246 vma = vma->vm_next;
2247 }
2248 }
2249
2250 arch_exit_mmap(mm);
2251
2252 vma = mm->mmap;
2253 if (!vma) /* Can happen if dup_mmap() received an OOM */
2254 return;
2255
2256 lru_add_drain();
2257 flush_cache_mm(mm);
2258 tlb_gather_mmu(&tlb, mm, 1);
2259 /* update_hiwater_rss(mm) here? but nobody should be looking */
2260 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2261 end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2262 vm_unacct_memory(nr_accounted);
2263
2264 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2265 tlb_finish_mmu(&tlb, 0, end);
2266
2267 /*
2268 * Walk the list again, actually closing and freeing it,
2269 * with preemption enabled, without holding any MM locks.
2270 */
2271 while (vma)
2272 vma = remove_vma(vma);
2273
2274 BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2275}
2276
2277/* Insert vm structure into process list sorted by address
2278 * and into the inode's i_mmap tree. If vm_file is non-NULL
2279 * then i_mmap_mutex is taken here.
2280 */
2281int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2282{
2283 struct vm_area_struct * __vma, * prev;
2284 struct rb_node ** rb_link, * rb_parent;
2285
2286 /*
2287 * The vm_pgoff of a purely anonymous vma should be irrelevant
2288 * until its first write fault, when page's anon_vma and index
2289 * are set. But now set the vm_pgoff it will almost certainly
2290 * end up with (unless mremap moves it elsewhere before that
2291 * first wfault), so /proc/pid/maps tells a consistent story.
2292 *
2293 * By setting it to reflect the virtual start address of the
2294 * vma, merges and splits can happen in a seamless way, just
2295 * using the existing file pgoff checks and manipulations.
2296 * Similarly in do_mmap_pgoff and in do_brk.
2297 */
2298 if (!vma->vm_file) {
2299 BUG_ON(vma->anon_vma);
2300 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2301 }
2302 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2303 if (__vma && __vma->vm_start < vma->vm_end)
2304 return -ENOMEM;
2305 if ((vma->vm_flags & VM_ACCOUNT) &&
2306 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2307 return -ENOMEM;
2308 vma_link(mm, vma, prev, rb_link, rb_parent);
2309 return 0;
2310}
2311
2312/*
2313 * Copy the vma structure to a new location in the same mm,
2314 * prior to moving page table entries, to effect an mremap move.
2315 */
2316struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2317 unsigned long addr, unsigned long len, pgoff_t pgoff)
2318{
2319 struct vm_area_struct *vma = *vmap;
2320 unsigned long vma_start = vma->vm_start;
2321 struct mm_struct *mm = vma->vm_mm;
2322 struct vm_area_struct *new_vma, *prev;
2323 struct rb_node **rb_link, *rb_parent;
2324 struct mempolicy *pol;
2325
2326 /*
2327 * If anonymous vma has not yet been faulted, update new pgoff
2328 * to match new location, to increase its chance of merging.
2329 */
2330 if (!vma->vm_file && !vma->anon_vma)
2331 pgoff = addr >> PAGE_SHIFT;
2332
2333 find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2334 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2335 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2336 if (new_vma) {
2337 /*
2338 * Source vma may have been merged into new_vma
2339 */
2340 if (vma_start >= new_vma->vm_start &&
2341 vma_start < new_vma->vm_end)
2342 *vmap = new_vma;
2343 } else {
2344 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2345 if (new_vma) {
2346 *new_vma = *vma;
2347 pol = mpol_dup(vma_policy(vma));
2348 if (IS_ERR(pol))
2349 goto out_free_vma;
2350 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2351 if (anon_vma_clone(new_vma, vma))
2352 goto out_free_mempol;
2353 vma_set_policy(new_vma, pol);
2354 new_vma->vm_start = addr;
2355 new_vma->vm_end = addr + len;
2356 new_vma->vm_pgoff = pgoff;
2357 if (new_vma->vm_file) {
2358 get_file(new_vma->vm_file);
2359 if (vma->vm_flags & VM_EXECUTABLE)
2360 added_exe_file_vma(mm);
2361 }
2362 if (new_vma->vm_ops && new_vma->vm_ops->open)
2363 new_vma->vm_ops->open(new_vma);
2364 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2365 }
2366 }
2367 return new_vma;
2368
2369 out_free_mempol:
2370 mpol_put(pol);
2371 out_free_vma:
2372 kmem_cache_free(vm_area_cachep, new_vma);
2373 return NULL;
2374}
2375
2376/*
2377 * Return true if the calling process may expand its vm space by the passed
2378 * number of pages
2379 */
2380int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2381{
2382 unsigned long cur = mm->total_vm; /* pages */
2383 unsigned long lim;
2384
2385 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2386
2387 if (cur + npages > lim)
2388 return 0;
2389 return 1;
2390}
2391
2392
2393static int special_mapping_fault(struct vm_area_struct *vma,
2394 struct vm_fault *vmf)
2395{
2396 pgoff_t pgoff;
2397 struct page **pages;
2398
2399 /*
2400 * special mappings have no vm_file, and in that case, the mm
2401 * uses vm_pgoff internally. So we have to subtract it from here.
2402 * We are allowed to do this because we are the mm; do not copy
2403 * this code into drivers!
2404 */
2405 pgoff = vmf->pgoff - vma->vm_pgoff;
2406
2407 for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2408 pgoff--;
2409
2410 if (*pages) {
2411 struct page *page = *pages;
2412 get_page(page);
2413 vmf->page = page;
2414 return 0;
2415 }
2416
2417 return VM_FAULT_SIGBUS;
2418}
2419
2420/*
2421 * Having a close hook prevents vma merging regardless of flags.
2422 */
2423static void special_mapping_close(struct vm_area_struct *vma)
2424{
2425}
2426
2427static const struct vm_operations_struct special_mapping_vmops = {
2428 .close = special_mapping_close,
2429 .fault = special_mapping_fault,
2430};
2431
2432/*
2433 * Called with mm->mmap_sem held for writing.
2434 * Insert a new vma covering the given region, with the given flags.
2435 * Its pages are supplied by the given array of struct page *.
2436 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2437 * The region past the last page supplied will always produce SIGBUS.
2438 * The array pointer and the pages it points to are assumed to stay alive
2439 * for as long as this mapping might exist.
2440 */
2441int install_special_mapping(struct mm_struct *mm,
2442 unsigned long addr, unsigned long len,
2443 unsigned long vm_flags, struct page **pages)
2444{
2445 int ret;
2446 struct vm_area_struct *vma;
2447
2448 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2449 if (unlikely(vma == NULL))
2450 return -ENOMEM;
2451
2452 INIT_LIST_HEAD(&vma->anon_vma_chain);
2453 vma->vm_mm = mm;
2454 vma->vm_start = addr;
2455 vma->vm_end = addr + len;
2456
2457 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2458 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2459
2460 vma->vm_ops = &special_mapping_vmops;
2461 vma->vm_private_data = pages;
2462
2463 ret = security_file_mmap(NULL, 0, 0, 0, vma->vm_start, 1);
2464 if (ret)
2465 goto out;
2466
2467 ret = insert_vm_struct(mm, vma);
2468 if (ret)
2469 goto out;
2470
2471 mm->total_vm += len >> PAGE_SHIFT;
2472
2473 perf_event_mmap(vma);
2474
2475 return 0;
2476
2477out:
2478 kmem_cache_free(vm_area_cachep, vma);
2479 return ret;
2480}
2481
2482static DEFINE_MUTEX(mm_all_locks_mutex);
2483
2484static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2485{
2486 if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2487 /*
2488 * The LSB of head.next can't change from under us
2489 * because we hold the mm_all_locks_mutex.
2490 */
2491 mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem);
2492 /*
2493 * We can safely modify head.next after taking the
2494 * anon_vma->root->mutex. If some other vma in this mm shares
2495 * the same anon_vma we won't take it again.
2496 *
2497 * No need of atomic instructions here, head.next
2498 * can't change from under us thanks to the
2499 * anon_vma->root->mutex.
2500 */
2501 if (__test_and_set_bit(0, (unsigned long *)
2502 &anon_vma->root->head.next))
2503 BUG();
2504 }
2505}
2506
2507static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2508{
2509 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2510 /*
2511 * AS_MM_ALL_LOCKS can't change from under us because
2512 * we hold the mm_all_locks_mutex.
2513 *
2514 * Operations on ->flags have to be atomic because
2515 * even if AS_MM_ALL_LOCKS is stable thanks to the
2516 * mm_all_locks_mutex, there may be other cpus
2517 * changing other bitflags in parallel to us.
2518 */
2519 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2520 BUG();
2521 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2522 }
2523}
2524
2525/*
2526 * This operation locks against the VM for all pte/vma/mm related
2527 * operations that could ever happen on a certain mm. This includes
2528 * vmtruncate, try_to_unmap, and all page faults.
2529 *
2530 * The caller must take the mmap_sem in write mode before calling
2531 * mm_take_all_locks(). The caller isn't allowed to release the
2532 * mmap_sem until mm_drop_all_locks() returns.
2533 *
2534 * mmap_sem in write mode is required in order to block all operations
2535 * that could modify pagetables and free pages without need of
2536 * altering the vma layout (for example populate_range() with
2537 * nonlinear vmas). It's also needed in write mode to avoid new
2538 * anon_vmas to be associated with existing vmas.
2539 *
2540 * A single task can't take more than one mm_take_all_locks() in a row
2541 * or it would deadlock.
2542 *
2543 * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2544 * mapping->flags avoid to take the same lock twice, if more than one
2545 * vma in this mm is backed by the same anon_vma or address_space.
2546 *
2547 * We can take all the locks in random order because the VM code
2548 * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2549 * takes more than one of them in a row. Secondly we're protected
2550 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2551 *
2552 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2553 * that may have to take thousand of locks.
2554 *
2555 * mm_take_all_locks() can fail if it's interrupted by signals.
2556 */
2557int mm_take_all_locks(struct mm_struct *mm)
2558{
2559 struct vm_area_struct *vma;
2560 struct anon_vma_chain *avc;
2561 int ret = -EINTR;
2562
2563 BUG_ON(down_read_trylock(&mm->mmap_sem));
2564
2565 mutex_lock(&mm_all_locks_mutex);
2566
2567 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2568 if (signal_pending(current))
2569 goto out_unlock;
2570 if (vma->vm_file && vma->vm_file->f_mapping)
2571 vm_lock_mapping(mm, vma->vm_file->f_mapping);
2572 }
2573
2574 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2575 if (signal_pending(current))
2576 goto out_unlock;
2577 if (vma->anon_vma)
2578 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2579 vm_lock_anon_vma(mm, avc->anon_vma);
2580 }
2581
2582 ret = 0;
2583
2584out_unlock:
2585 if (ret)
2586 mm_drop_all_locks(mm);
2587
2588 return ret;
2589}
2590
2591static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2592{
2593 if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2594 /*
2595 * The LSB of head.next can't change to 0 from under
2596 * us because we hold the mm_all_locks_mutex.
2597 *
2598 * We must however clear the bitflag before unlocking
2599 * the vma so the users using the anon_vma->head will
2600 * never see our bitflag.
2601 *
2602 * No need of atomic instructions here, head.next
2603 * can't change from under us until we release the
2604 * anon_vma->root->mutex.
2605 */
2606 if (!__test_and_clear_bit(0, (unsigned long *)
2607 &anon_vma->root->head.next))
2608 BUG();
2609 anon_vma_unlock(anon_vma);
2610 }
2611}
2612
2613static void vm_unlock_mapping(struct address_space *mapping)
2614{
2615 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2616 /*
2617 * AS_MM_ALL_LOCKS can't change to 0 from under us
2618 * because we hold the mm_all_locks_mutex.
2619 */
2620 mutex_unlock(&mapping->i_mmap_mutex);
2621 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2622 &mapping->flags))
2623 BUG();
2624 }
2625}
2626
2627/*
2628 * The mmap_sem cannot be released by the caller until
2629 * mm_drop_all_locks() returns.
2630 */
2631void mm_drop_all_locks(struct mm_struct *mm)
2632{
2633 struct vm_area_struct *vma;
2634 struct anon_vma_chain *avc;
2635
2636 BUG_ON(down_read_trylock(&mm->mmap_sem));
2637 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2638
2639 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2640 if (vma->anon_vma)
2641 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2642 vm_unlock_anon_vma(avc->anon_vma);
2643 if (vma->vm_file && vma->vm_file->f_mapping)
2644 vm_unlock_mapping(vma->vm_file->f_mapping);
2645 }
2646
2647 mutex_unlock(&mm_all_locks_mutex);
2648}
2649
2650/*
2651 * initialise the VMA slab
2652 */
2653void __init mmap_init(void)
2654{
2655 int ret;
2656
2657 ret = percpu_counter_init(&vm_committed_as, 0);
2658 VM_BUG_ON(ret);
2659}
1/*
2 * mm/mmap.c
3 *
4 * Written by obz.
5 *
6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
7 */
8
9#include <linux/slab.h>
10#include <linux/backing-dev.h>
11#include <linux/mm.h>
12#include <linux/shm.h>
13#include <linux/mman.h>
14#include <linux/pagemap.h>
15#include <linux/swap.h>
16#include <linux/syscalls.h>
17#include <linux/capability.h>
18#include <linux/init.h>
19#include <linux/file.h>
20#include <linux/fs.h>
21#include <linux/personality.h>
22#include <linux/security.h>
23#include <linux/hugetlb.h>
24#include <linux/profile.h>
25#include <linux/export.h>
26#include <linux/mount.h>
27#include <linux/mempolicy.h>
28#include <linux/rmap.h>
29#include <linux/mmu_notifier.h>
30#include <linux/perf_event.h>
31#include <linux/audit.h>
32#include <linux/khugepaged.h>
33#include <linux/uprobes.h>
34
35#include <asm/uaccess.h>
36#include <asm/cacheflush.h>
37#include <asm/tlb.h>
38#include <asm/mmu_context.h>
39
40#include "internal.h"
41
42#ifndef arch_mmap_check
43#define arch_mmap_check(addr, len, flags) (0)
44#endif
45
46#ifndef arch_rebalance_pgtables
47#define arch_rebalance_pgtables(addr, len) (addr)
48#endif
49
50static void unmap_region(struct mm_struct *mm,
51 struct vm_area_struct *vma, struct vm_area_struct *prev,
52 unsigned long start, unsigned long end);
53
54/*
55 * WARNING: the debugging will use recursive algorithms so never enable this
56 * unless you know what you are doing.
57 */
58#undef DEBUG_MM_RB
59
60/* description of effects of mapping type and prot in current implementation.
61 * this is due to the limited x86 page protection hardware. The expected
62 * behavior is in parens:
63 *
64 * map_type prot
65 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
66 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
67 * w: (no) no w: (no) no w: (yes) yes w: (no) no
68 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
69 *
70 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
71 * w: (no) no w: (no) no w: (copy) copy w: (no) no
72 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
73 *
74 */
75pgprot_t protection_map[16] = {
76 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
77 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
78};
79
80pgprot_t vm_get_page_prot(unsigned long vm_flags)
81{
82 return __pgprot(pgprot_val(protection_map[vm_flags &
83 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
84 pgprot_val(arch_vm_get_page_prot(vm_flags)));
85}
86EXPORT_SYMBOL(vm_get_page_prot);
87
88int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS; /* heuristic overcommit */
89int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
90int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
91/*
92 * Make sure vm_committed_as in one cacheline and not cacheline shared with
93 * other variables. It can be updated by several CPUs frequently.
94 */
95struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
96
97/*
98 * Check that a process has enough memory to allocate a new virtual
99 * mapping. 0 means there is enough memory for the allocation to
100 * succeed and -ENOMEM implies there is not.
101 *
102 * We currently support three overcommit policies, which are set via the
103 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
104 *
105 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
106 * Additional code 2002 Jul 20 by Robert Love.
107 *
108 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
109 *
110 * Note this is a helper function intended to be used by LSMs which
111 * wish to use this logic.
112 */
113int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
114{
115 unsigned long free, allowed;
116
117 vm_acct_memory(pages);
118
119 /*
120 * Sometimes we want to use more memory than we have
121 */
122 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
123 return 0;
124
125 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
126 free = global_page_state(NR_FREE_PAGES);
127 free += global_page_state(NR_FILE_PAGES);
128
129 /*
130 * shmem pages shouldn't be counted as free in this
131 * case, they can't be purged, only swapped out, and
132 * that won't affect the overall amount of available
133 * memory in the system.
134 */
135 free -= global_page_state(NR_SHMEM);
136
137 free += nr_swap_pages;
138
139 /*
140 * Any slabs which are created with the
141 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
142 * which are reclaimable, under pressure. The dentry
143 * cache and most inode caches should fall into this
144 */
145 free += global_page_state(NR_SLAB_RECLAIMABLE);
146
147 /*
148 * Leave reserved pages. The pages are not for anonymous pages.
149 */
150 if (free <= totalreserve_pages)
151 goto error;
152 else
153 free -= totalreserve_pages;
154
155 /*
156 * Leave the last 3% for root
157 */
158 if (!cap_sys_admin)
159 free -= free / 32;
160
161 if (free > pages)
162 return 0;
163
164 goto error;
165 }
166
167 allowed = (totalram_pages - hugetlb_total_pages())
168 * sysctl_overcommit_ratio / 100;
169 /*
170 * Leave the last 3% for root
171 */
172 if (!cap_sys_admin)
173 allowed -= allowed / 32;
174 allowed += total_swap_pages;
175
176 /* Don't let a single process grow too big:
177 leave 3% of the size of this process for other processes */
178 if (mm)
179 allowed -= mm->total_vm / 32;
180
181 if (percpu_counter_read_positive(&vm_committed_as) < allowed)
182 return 0;
183error:
184 vm_unacct_memory(pages);
185
186 return -ENOMEM;
187}
188
189/*
190 * Requires inode->i_mapping->i_mmap_mutex
191 */
192static void __remove_shared_vm_struct(struct vm_area_struct *vma,
193 struct file *file, struct address_space *mapping)
194{
195 if (vma->vm_flags & VM_DENYWRITE)
196 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
197 if (vma->vm_flags & VM_SHARED)
198 mapping->i_mmap_writable--;
199
200 flush_dcache_mmap_lock(mapping);
201 if (unlikely(vma->vm_flags & VM_NONLINEAR))
202 list_del_init(&vma->shared.vm_set.list);
203 else
204 vma_prio_tree_remove(vma, &mapping->i_mmap);
205 flush_dcache_mmap_unlock(mapping);
206}
207
208/*
209 * Unlink a file-based vm structure from its prio_tree, to hide
210 * vma from rmap and vmtruncate before freeing its page tables.
211 */
212void unlink_file_vma(struct vm_area_struct *vma)
213{
214 struct file *file = vma->vm_file;
215
216 if (file) {
217 struct address_space *mapping = file->f_mapping;
218 mutex_lock(&mapping->i_mmap_mutex);
219 __remove_shared_vm_struct(vma, file, mapping);
220 mutex_unlock(&mapping->i_mmap_mutex);
221 }
222}
223
224/*
225 * Close a vm structure and free it, returning the next.
226 */
227static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
228{
229 struct vm_area_struct *next = vma->vm_next;
230
231 might_sleep();
232 if (vma->vm_ops && vma->vm_ops->close)
233 vma->vm_ops->close(vma);
234 if (vma->vm_file) {
235 fput(vma->vm_file);
236 if (vma->vm_flags & VM_EXECUTABLE)
237 removed_exe_file_vma(vma->vm_mm);
238 }
239 mpol_put(vma_policy(vma));
240 kmem_cache_free(vm_area_cachep, vma);
241 return next;
242}
243
244static unsigned long do_brk(unsigned long addr, unsigned long len);
245
246SYSCALL_DEFINE1(brk, unsigned long, brk)
247{
248 unsigned long rlim, retval;
249 unsigned long newbrk, oldbrk;
250 struct mm_struct *mm = current->mm;
251 unsigned long min_brk;
252
253 down_write(&mm->mmap_sem);
254
255#ifdef CONFIG_COMPAT_BRK
256 /*
257 * CONFIG_COMPAT_BRK can still be overridden by setting
258 * randomize_va_space to 2, which will still cause mm->start_brk
259 * to be arbitrarily shifted
260 */
261 if (current->brk_randomized)
262 min_brk = mm->start_brk;
263 else
264 min_brk = mm->end_data;
265#else
266 min_brk = mm->start_brk;
267#endif
268 if (brk < min_brk)
269 goto out;
270
271 /*
272 * Check against rlimit here. If this check is done later after the test
273 * of oldbrk with newbrk then it can escape the test and let the data
274 * segment grow beyond its set limit the in case where the limit is
275 * not page aligned -Ram Gupta
276 */
277 rlim = rlimit(RLIMIT_DATA);
278 if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
279 (mm->end_data - mm->start_data) > rlim)
280 goto out;
281
282 newbrk = PAGE_ALIGN(brk);
283 oldbrk = PAGE_ALIGN(mm->brk);
284 if (oldbrk == newbrk)
285 goto set_brk;
286
287 /* Always allow shrinking brk. */
288 if (brk <= mm->brk) {
289 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
290 goto set_brk;
291 goto out;
292 }
293
294 /* Check against existing mmap mappings. */
295 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
296 goto out;
297
298 /* Ok, looks good - let it rip. */
299 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
300 goto out;
301set_brk:
302 mm->brk = brk;
303out:
304 retval = mm->brk;
305 up_write(&mm->mmap_sem);
306 return retval;
307}
308
309#ifdef DEBUG_MM_RB
310static int browse_rb(struct rb_root *root)
311{
312 int i = 0, j;
313 struct rb_node *nd, *pn = NULL;
314 unsigned long prev = 0, pend = 0;
315
316 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
317 struct vm_area_struct *vma;
318 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
319 if (vma->vm_start < prev)
320 printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
321 if (vma->vm_start < pend)
322 printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
323 if (vma->vm_start > vma->vm_end)
324 printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
325 i++;
326 pn = nd;
327 prev = vma->vm_start;
328 pend = vma->vm_end;
329 }
330 j = 0;
331 for (nd = pn; nd; nd = rb_prev(nd)) {
332 j++;
333 }
334 if (i != j)
335 printk("backwards %d, forwards %d\n", j, i), i = 0;
336 return i;
337}
338
339void validate_mm(struct mm_struct *mm)
340{
341 int bug = 0;
342 int i = 0;
343 struct vm_area_struct *tmp = mm->mmap;
344 while (tmp) {
345 tmp = tmp->vm_next;
346 i++;
347 }
348 if (i != mm->map_count)
349 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
350 i = browse_rb(&mm->mm_rb);
351 if (i != mm->map_count)
352 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
353 BUG_ON(bug);
354}
355#else
356#define validate_mm(mm) do { } while (0)
357#endif
358
359static struct vm_area_struct *
360find_vma_prepare(struct mm_struct *mm, unsigned long addr,
361 struct vm_area_struct **pprev, struct rb_node ***rb_link,
362 struct rb_node ** rb_parent)
363{
364 struct vm_area_struct * vma;
365 struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
366
367 __rb_link = &mm->mm_rb.rb_node;
368 rb_prev = __rb_parent = NULL;
369 vma = NULL;
370
371 while (*__rb_link) {
372 struct vm_area_struct *vma_tmp;
373
374 __rb_parent = *__rb_link;
375 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
376
377 if (vma_tmp->vm_end > addr) {
378 vma = vma_tmp;
379 if (vma_tmp->vm_start <= addr)
380 break;
381 __rb_link = &__rb_parent->rb_left;
382 } else {
383 rb_prev = __rb_parent;
384 __rb_link = &__rb_parent->rb_right;
385 }
386 }
387
388 *pprev = NULL;
389 if (rb_prev)
390 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
391 *rb_link = __rb_link;
392 *rb_parent = __rb_parent;
393 return vma;
394}
395
396void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
397 struct rb_node **rb_link, struct rb_node *rb_parent)
398{
399 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
400 rb_insert_color(&vma->vm_rb, &mm->mm_rb);
401}
402
403static void __vma_link_file(struct vm_area_struct *vma)
404{
405 struct file *file;
406
407 file = vma->vm_file;
408 if (file) {
409 struct address_space *mapping = file->f_mapping;
410
411 if (vma->vm_flags & VM_DENYWRITE)
412 atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
413 if (vma->vm_flags & VM_SHARED)
414 mapping->i_mmap_writable++;
415
416 flush_dcache_mmap_lock(mapping);
417 if (unlikely(vma->vm_flags & VM_NONLINEAR))
418 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
419 else
420 vma_prio_tree_insert(vma, &mapping->i_mmap);
421 flush_dcache_mmap_unlock(mapping);
422 }
423}
424
425static void
426__vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
427 struct vm_area_struct *prev, struct rb_node **rb_link,
428 struct rb_node *rb_parent)
429{
430 __vma_link_list(mm, vma, prev, rb_parent);
431 __vma_link_rb(mm, vma, rb_link, rb_parent);
432}
433
434static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
435 struct vm_area_struct *prev, struct rb_node **rb_link,
436 struct rb_node *rb_parent)
437{
438 struct address_space *mapping = NULL;
439
440 if (vma->vm_file)
441 mapping = vma->vm_file->f_mapping;
442
443 if (mapping)
444 mutex_lock(&mapping->i_mmap_mutex);
445
446 __vma_link(mm, vma, prev, rb_link, rb_parent);
447 __vma_link_file(vma);
448
449 if (mapping)
450 mutex_unlock(&mapping->i_mmap_mutex);
451
452 mm->map_count++;
453 validate_mm(mm);
454}
455
456/*
457 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
458 * mm's list and rbtree. It has already been inserted into the prio_tree.
459 */
460static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
461{
462 struct vm_area_struct *__vma, *prev;
463 struct rb_node **rb_link, *rb_parent;
464
465 __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
466 BUG_ON(__vma && __vma->vm_start < vma->vm_end);
467 __vma_link(mm, vma, prev, rb_link, rb_parent);
468 mm->map_count++;
469}
470
471static inline void
472__vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
473 struct vm_area_struct *prev)
474{
475 struct vm_area_struct *next = vma->vm_next;
476
477 prev->vm_next = next;
478 if (next)
479 next->vm_prev = prev;
480 rb_erase(&vma->vm_rb, &mm->mm_rb);
481 if (mm->mmap_cache == vma)
482 mm->mmap_cache = prev;
483}
484
485/*
486 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
487 * is already present in an i_mmap tree without adjusting the tree.
488 * The following helper function should be used when such adjustments
489 * are necessary. The "insert" vma (if any) is to be inserted
490 * before we drop the necessary locks.
491 */
492int vma_adjust(struct vm_area_struct *vma, unsigned long start,
493 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
494{
495 struct mm_struct *mm = vma->vm_mm;
496 struct vm_area_struct *next = vma->vm_next;
497 struct vm_area_struct *importer = NULL;
498 struct address_space *mapping = NULL;
499 struct prio_tree_root *root = NULL;
500 struct anon_vma *anon_vma = NULL;
501 struct file *file = vma->vm_file;
502 long adjust_next = 0;
503 int remove_next = 0;
504
505 if (next && !insert) {
506 struct vm_area_struct *exporter = NULL;
507
508 if (end >= next->vm_end) {
509 /*
510 * vma expands, overlapping all the next, and
511 * perhaps the one after too (mprotect case 6).
512 */
513again: remove_next = 1 + (end > next->vm_end);
514 end = next->vm_end;
515 exporter = next;
516 importer = vma;
517 } else if (end > next->vm_start) {
518 /*
519 * vma expands, overlapping part of the next:
520 * mprotect case 5 shifting the boundary up.
521 */
522 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
523 exporter = next;
524 importer = vma;
525 } else if (end < vma->vm_end) {
526 /*
527 * vma shrinks, and !insert tells it's not
528 * split_vma inserting another: so it must be
529 * mprotect case 4 shifting the boundary down.
530 */
531 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
532 exporter = vma;
533 importer = next;
534 }
535
536 /*
537 * Easily overlooked: when mprotect shifts the boundary,
538 * make sure the expanding vma has anon_vma set if the
539 * shrinking vma had, to cover any anon pages imported.
540 */
541 if (exporter && exporter->anon_vma && !importer->anon_vma) {
542 if (anon_vma_clone(importer, exporter))
543 return -ENOMEM;
544 importer->anon_vma = exporter->anon_vma;
545 }
546 }
547
548 if (file) {
549 mapping = file->f_mapping;
550 if (!(vma->vm_flags & VM_NONLINEAR)) {
551 root = &mapping->i_mmap;
552 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
553
554 if (adjust_next)
555 uprobe_munmap(next, next->vm_start,
556 next->vm_end);
557 }
558
559 mutex_lock(&mapping->i_mmap_mutex);
560 if (insert) {
561 /*
562 * Put into prio_tree now, so instantiated pages
563 * are visible to arm/parisc __flush_dcache_page
564 * throughout; but we cannot insert into address
565 * space until vma start or end is updated.
566 */
567 __vma_link_file(insert);
568 }
569 }
570
571 vma_adjust_trans_huge(vma, start, end, adjust_next);
572
573 /*
574 * When changing only vma->vm_end, we don't really need anon_vma
575 * lock. This is a fairly rare case by itself, but the anon_vma
576 * lock may be shared between many sibling processes. Skipping
577 * the lock for brk adjustments makes a difference sometimes.
578 */
579 if (vma->anon_vma && (importer || start != vma->vm_start)) {
580 anon_vma = vma->anon_vma;
581 anon_vma_lock(anon_vma);
582 }
583
584 if (root) {
585 flush_dcache_mmap_lock(mapping);
586 vma_prio_tree_remove(vma, root);
587 if (adjust_next)
588 vma_prio_tree_remove(next, root);
589 }
590
591 vma->vm_start = start;
592 vma->vm_end = end;
593 vma->vm_pgoff = pgoff;
594 if (adjust_next) {
595 next->vm_start += adjust_next << PAGE_SHIFT;
596 next->vm_pgoff += adjust_next;
597 }
598
599 if (root) {
600 if (adjust_next)
601 vma_prio_tree_insert(next, root);
602 vma_prio_tree_insert(vma, root);
603 flush_dcache_mmap_unlock(mapping);
604 }
605
606 if (remove_next) {
607 /*
608 * vma_merge has merged next into vma, and needs
609 * us to remove next before dropping the locks.
610 */
611 __vma_unlink(mm, next, vma);
612 if (file)
613 __remove_shared_vm_struct(next, file, mapping);
614 } else if (insert) {
615 /*
616 * split_vma has split insert from vma, and needs
617 * us to insert it before dropping the locks
618 * (it may either follow vma or precede it).
619 */
620 __insert_vm_struct(mm, insert);
621 }
622
623 if (anon_vma)
624 anon_vma_unlock(anon_vma);
625 if (mapping)
626 mutex_unlock(&mapping->i_mmap_mutex);
627
628 if (root) {
629 uprobe_mmap(vma);
630
631 if (adjust_next)
632 uprobe_mmap(next);
633 }
634
635 if (remove_next) {
636 if (file) {
637 uprobe_munmap(next, next->vm_start, next->vm_end);
638 fput(file);
639 if (next->vm_flags & VM_EXECUTABLE)
640 removed_exe_file_vma(mm);
641 }
642 if (next->anon_vma)
643 anon_vma_merge(vma, next);
644 mm->map_count--;
645 mpol_put(vma_policy(next));
646 kmem_cache_free(vm_area_cachep, next);
647 /*
648 * In mprotect's case 6 (see comments on vma_merge),
649 * we must remove another next too. It would clutter
650 * up the code too much to do both in one go.
651 */
652 if (remove_next == 2) {
653 next = vma->vm_next;
654 goto again;
655 }
656 }
657 if (insert && file)
658 uprobe_mmap(insert);
659
660 validate_mm(mm);
661
662 return 0;
663}
664
665/*
666 * If the vma has a ->close operation then the driver probably needs to release
667 * per-vma resources, so we don't attempt to merge those.
668 */
669static inline int is_mergeable_vma(struct vm_area_struct *vma,
670 struct file *file, unsigned long vm_flags)
671{
672 /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
673 if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
674 return 0;
675 if (vma->vm_file != file)
676 return 0;
677 if (vma->vm_ops && vma->vm_ops->close)
678 return 0;
679 return 1;
680}
681
682static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
683 struct anon_vma *anon_vma2,
684 struct vm_area_struct *vma)
685{
686 /*
687 * The list_is_singular() test is to avoid merging VMA cloned from
688 * parents. This can improve scalability caused by anon_vma lock.
689 */
690 if ((!anon_vma1 || !anon_vma2) && (!vma ||
691 list_is_singular(&vma->anon_vma_chain)))
692 return 1;
693 return anon_vma1 == anon_vma2;
694}
695
696/*
697 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
698 * in front of (at a lower virtual address and file offset than) the vma.
699 *
700 * We cannot merge two vmas if they have differently assigned (non-NULL)
701 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
702 *
703 * We don't check here for the merged mmap wrapping around the end of pagecache
704 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
705 * wrap, nor mmaps which cover the final page at index -1UL.
706 */
707static int
708can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
709 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
710{
711 if (is_mergeable_vma(vma, file, vm_flags) &&
712 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
713 if (vma->vm_pgoff == vm_pgoff)
714 return 1;
715 }
716 return 0;
717}
718
719/*
720 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
721 * beyond (at a higher virtual address and file offset than) the vma.
722 *
723 * We cannot merge two vmas if they have differently assigned (non-NULL)
724 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
725 */
726static int
727can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
728 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
729{
730 if (is_mergeable_vma(vma, file, vm_flags) &&
731 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
732 pgoff_t vm_pglen;
733 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
734 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
735 return 1;
736 }
737 return 0;
738}
739
740/*
741 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
742 * whether that can be merged with its predecessor or its successor.
743 * Or both (it neatly fills a hole).
744 *
745 * In most cases - when called for mmap, brk or mremap - [addr,end) is
746 * certain not to be mapped by the time vma_merge is called; but when
747 * called for mprotect, it is certain to be already mapped (either at
748 * an offset within prev, or at the start of next), and the flags of
749 * this area are about to be changed to vm_flags - and the no-change
750 * case has already been eliminated.
751 *
752 * The following mprotect cases have to be considered, where AAAA is
753 * the area passed down from mprotect_fixup, never extending beyond one
754 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
755 *
756 * AAAA AAAA AAAA AAAA
757 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
758 * cannot merge might become might become might become
759 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
760 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
761 * mremap move: PPPPNNNNNNNN 8
762 * AAAA
763 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
764 * might become case 1 below case 2 below case 3 below
765 *
766 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
767 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
768 */
769struct vm_area_struct *vma_merge(struct mm_struct *mm,
770 struct vm_area_struct *prev, unsigned long addr,
771 unsigned long end, unsigned long vm_flags,
772 struct anon_vma *anon_vma, struct file *file,
773 pgoff_t pgoff, struct mempolicy *policy)
774{
775 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
776 struct vm_area_struct *area, *next;
777 int err;
778
779 /*
780 * We later require that vma->vm_flags == vm_flags,
781 * so this tests vma->vm_flags & VM_SPECIAL, too.
782 */
783 if (vm_flags & VM_SPECIAL)
784 return NULL;
785
786 if (prev)
787 next = prev->vm_next;
788 else
789 next = mm->mmap;
790 area = next;
791 if (next && next->vm_end == end) /* cases 6, 7, 8 */
792 next = next->vm_next;
793
794 /*
795 * Can it merge with the predecessor?
796 */
797 if (prev && prev->vm_end == addr &&
798 mpol_equal(vma_policy(prev), policy) &&
799 can_vma_merge_after(prev, vm_flags,
800 anon_vma, file, pgoff)) {
801 /*
802 * OK, it can. Can we now merge in the successor as well?
803 */
804 if (next && end == next->vm_start &&
805 mpol_equal(policy, vma_policy(next)) &&
806 can_vma_merge_before(next, vm_flags,
807 anon_vma, file, pgoff+pglen) &&
808 is_mergeable_anon_vma(prev->anon_vma,
809 next->anon_vma, NULL)) {
810 /* cases 1, 6 */
811 err = vma_adjust(prev, prev->vm_start,
812 next->vm_end, prev->vm_pgoff, NULL);
813 } else /* cases 2, 5, 7 */
814 err = vma_adjust(prev, prev->vm_start,
815 end, prev->vm_pgoff, NULL);
816 if (err)
817 return NULL;
818 khugepaged_enter_vma_merge(prev);
819 return prev;
820 }
821
822 /*
823 * Can this new request be merged in front of next?
824 */
825 if (next && end == next->vm_start &&
826 mpol_equal(policy, vma_policy(next)) &&
827 can_vma_merge_before(next, vm_flags,
828 anon_vma, file, pgoff+pglen)) {
829 if (prev && addr < prev->vm_end) /* case 4 */
830 err = vma_adjust(prev, prev->vm_start,
831 addr, prev->vm_pgoff, NULL);
832 else /* cases 3, 8 */
833 err = vma_adjust(area, addr, next->vm_end,
834 next->vm_pgoff - pglen, NULL);
835 if (err)
836 return NULL;
837 khugepaged_enter_vma_merge(area);
838 return area;
839 }
840
841 return NULL;
842}
843
844/*
845 * Rough compatbility check to quickly see if it's even worth looking
846 * at sharing an anon_vma.
847 *
848 * They need to have the same vm_file, and the flags can only differ
849 * in things that mprotect may change.
850 *
851 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
852 * we can merge the two vma's. For example, we refuse to merge a vma if
853 * there is a vm_ops->close() function, because that indicates that the
854 * driver is doing some kind of reference counting. But that doesn't
855 * really matter for the anon_vma sharing case.
856 */
857static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
858{
859 return a->vm_end == b->vm_start &&
860 mpol_equal(vma_policy(a), vma_policy(b)) &&
861 a->vm_file == b->vm_file &&
862 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
863 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
864}
865
866/*
867 * Do some basic sanity checking to see if we can re-use the anon_vma
868 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
869 * the same as 'old', the other will be the new one that is trying
870 * to share the anon_vma.
871 *
872 * NOTE! This runs with mm_sem held for reading, so it is possible that
873 * the anon_vma of 'old' is concurrently in the process of being set up
874 * by another page fault trying to merge _that_. But that's ok: if it
875 * is being set up, that automatically means that it will be a singleton
876 * acceptable for merging, so we can do all of this optimistically. But
877 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
878 *
879 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
880 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
881 * is to return an anon_vma that is "complex" due to having gone through
882 * a fork).
883 *
884 * We also make sure that the two vma's are compatible (adjacent,
885 * and with the same memory policies). That's all stable, even with just
886 * a read lock on the mm_sem.
887 */
888static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
889{
890 if (anon_vma_compatible(a, b)) {
891 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
892
893 if (anon_vma && list_is_singular(&old->anon_vma_chain))
894 return anon_vma;
895 }
896 return NULL;
897}
898
899/*
900 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
901 * neighbouring vmas for a suitable anon_vma, before it goes off
902 * to allocate a new anon_vma. It checks because a repetitive
903 * sequence of mprotects and faults may otherwise lead to distinct
904 * anon_vmas being allocated, preventing vma merge in subsequent
905 * mprotect.
906 */
907struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
908{
909 struct anon_vma *anon_vma;
910 struct vm_area_struct *near;
911
912 near = vma->vm_next;
913 if (!near)
914 goto try_prev;
915
916 anon_vma = reusable_anon_vma(near, vma, near);
917 if (anon_vma)
918 return anon_vma;
919try_prev:
920 near = vma->vm_prev;
921 if (!near)
922 goto none;
923
924 anon_vma = reusable_anon_vma(near, near, vma);
925 if (anon_vma)
926 return anon_vma;
927none:
928 /*
929 * There's no absolute need to look only at touching neighbours:
930 * we could search further afield for "compatible" anon_vmas.
931 * But it would probably just be a waste of time searching,
932 * or lead to too many vmas hanging off the same anon_vma.
933 * We're trying to allow mprotect remerging later on,
934 * not trying to minimize memory used for anon_vmas.
935 */
936 return NULL;
937}
938
939#ifdef CONFIG_PROC_FS
940void vm_stat_account(struct mm_struct *mm, unsigned long flags,
941 struct file *file, long pages)
942{
943 const unsigned long stack_flags
944 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
945
946 if (file) {
947 mm->shared_vm += pages;
948 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
949 mm->exec_vm += pages;
950 } else if (flags & stack_flags)
951 mm->stack_vm += pages;
952 if (flags & (VM_RESERVED|VM_IO))
953 mm->reserved_vm += pages;
954}
955#endif /* CONFIG_PROC_FS */
956
957/*
958 * If a hint addr is less than mmap_min_addr change hint to be as
959 * low as possible but still greater than mmap_min_addr
960 */
961static inline unsigned long round_hint_to_min(unsigned long hint)
962{
963 hint &= PAGE_MASK;
964 if (((void *)hint != NULL) &&
965 (hint < mmap_min_addr))
966 return PAGE_ALIGN(mmap_min_addr);
967 return hint;
968}
969
970/*
971 * The caller must hold down_write(¤t->mm->mmap_sem).
972 */
973
974unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
975 unsigned long len, unsigned long prot,
976 unsigned long flags, unsigned long pgoff)
977{
978 struct mm_struct * mm = current->mm;
979 struct inode *inode;
980 vm_flags_t vm_flags;
981
982 /*
983 * Does the application expect PROT_READ to imply PROT_EXEC?
984 *
985 * (the exception is when the underlying filesystem is noexec
986 * mounted, in which case we dont add PROT_EXEC.)
987 */
988 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
989 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
990 prot |= PROT_EXEC;
991
992 if (!len)
993 return -EINVAL;
994
995 if (!(flags & MAP_FIXED))
996 addr = round_hint_to_min(addr);
997
998 /* Careful about overflows.. */
999 len = PAGE_ALIGN(len);
1000 if (!len)
1001 return -ENOMEM;
1002
1003 /* offset overflow? */
1004 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1005 return -EOVERFLOW;
1006
1007 /* Too many mappings? */
1008 if (mm->map_count > sysctl_max_map_count)
1009 return -ENOMEM;
1010
1011 /* Obtain the address to map to. we verify (or select) it and ensure
1012 * that it represents a valid section of the address space.
1013 */
1014 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1015 if (addr & ~PAGE_MASK)
1016 return addr;
1017
1018 /* Do simple checking here so the lower-level routines won't have
1019 * to. we assume access permissions have been handled by the open
1020 * of the memory object, so we don't do any here.
1021 */
1022 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1023 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1024
1025 if (flags & MAP_LOCKED)
1026 if (!can_do_mlock())
1027 return -EPERM;
1028
1029 /* mlock MCL_FUTURE? */
1030 if (vm_flags & VM_LOCKED) {
1031 unsigned long locked, lock_limit;
1032 locked = len >> PAGE_SHIFT;
1033 locked += mm->locked_vm;
1034 lock_limit = rlimit(RLIMIT_MEMLOCK);
1035 lock_limit >>= PAGE_SHIFT;
1036 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1037 return -EAGAIN;
1038 }
1039
1040 inode = file ? file->f_path.dentry->d_inode : NULL;
1041
1042 if (file) {
1043 switch (flags & MAP_TYPE) {
1044 case MAP_SHARED:
1045 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1046 return -EACCES;
1047
1048 /*
1049 * Make sure we don't allow writing to an append-only
1050 * file..
1051 */
1052 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1053 return -EACCES;
1054
1055 /*
1056 * Make sure there are no mandatory locks on the file.
1057 */
1058 if (locks_verify_locked(inode))
1059 return -EAGAIN;
1060
1061 vm_flags |= VM_SHARED | VM_MAYSHARE;
1062 if (!(file->f_mode & FMODE_WRITE))
1063 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1064
1065 /* fall through */
1066 case MAP_PRIVATE:
1067 if (!(file->f_mode & FMODE_READ))
1068 return -EACCES;
1069 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1070 if (vm_flags & VM_EXEC)
1071 return -EPERM;
1072 vm_flags &= ~VM_MAYEXEC;
1073 }
1074
1075 if (!file->f_op || !file->f_op->mmap)
1076 return -ENODEV;
1077 break;
1078
1079 default:
1080 return -EINVAL;
1081 }
1082 } else {
1083 switch (flags & MAP_TYPE) {
1084 case MAP_SHARED:
1085 /*
1086 * Ignore pgoff.
1087 */
1088 pgoff = 0;
1089 vm_flags |= VM_SHARED | VM_MAYSHARE;
1090 break;
1091 case MAP_PRIVATE:
1092 /*
1093 * Set pgoff according to addr for anon_vma.
1094 */
1095 pgoff = addr >> PAGE_SHIFT;
1096 break;
1097 default:
1098 return -EINVAL;
1099 }
1100 }
1101
1102 return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1103}
1104
1105SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1106 unsigned long, prot, unsigned long, flags,
1107 unsigned long, fd, unsigned long, pgoff)
1108{
1109 struct file *file = NULL;
1110 unsigned long retval = -EBADF;
1111
1112 if (!(flags & MAP_ANONYMOUS)) {
1113 audit_mmap_fd(fd, flags);
1114 if (unlikely(flags & MAP_HUGETLB))
1115 return -EINVAL;
1116 file = fget(fd);
1117 if (!file)
1118 goto out;
1119 } else if (flags & MAP_HUGETLB) {
1120 struct user_struct *user = NULL;
1121 /*
1122 * VM_NORESERVE is used because the reservations will be
1123 * taken when vm_ops->mmap() is called
1124 * A dummy user value is used because we are not locking
1125 * memory so no accounting is necessary
1126 */
1127 file = hugetlb_file_setup(HUGETLB_ANON_FILE, addr, len,
1128 VM_NORESERVE, &user,
1129 HUGETLB_ANONHUGE_INODE);
1130 if (IS_ERR(file))
1131 return PTR_ERR(file);
1132 }
1133
1134 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1135
1136 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1137 if (file)
1138 fput(file);
1139out:
1140 return retval;
1141}
1142
1143#ifdef __ARCH_WANT_SYS_OLD_MMAP
1144struct mmap_arg_struct {
1145 unsigned long addr;
1146 unsigned long len;
1147 unsigned long prot;
1148 unsigned long flags;
1149 unsigned long fd;
1150 unsigned long offset;
1151};
1152
1153SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1154{
1155 struct mmap_arg_struct a;
1156
1157 if (copy_from_user(&a, arg, sizeof(a)))
1158 return -EFAULT;
1159 if (a.offset & ~PAGE_MASK)
1160 return -EINVAL;
1161
1162 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1163 a.offset >> PAGE_SHIFT);
1164}
1165#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1166
1167/*
1168 * Some shared mappigns will want the pages marked read-only
1169 * to track write events. If so, we'll downgrade vm_page_prot
1170 * to the private version (using protection_map[] without the
1171 * VM_SHARED bit).
1172 */
1173int vma_wants_writenotify(struct vm_area_struct *vma)
1174{
1175 vm_flags_t vm_flags = vma->vm_flags;
1176
1177 /* If it was private or non-writable, the write bit is already clear */
1178 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1179 return 0;
1180
1181 /* The backer wishes to know when pages are first written to? */
1182 if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1183 return 1;
1184
1185 /* The open routine did something to the protections already? */
1186 if (pgprot_val(vma->vm_page_prot) !=
1187 pgprot_val(vm_get_page_prot(vm_flags)))
1188 return 0;
1189
1190 /* Specialty mapping? */
1191 if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1192 return 0;
1193
1194 /* Can the mapping track the dirty pages? */
1195 return vma->vm_file && vma->vm_file->f_mapping &&
1196 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1197}
1198
1199/*
1200 * We account for memory if it's a private writeable mapping,
1201 * not hugepages and VM_NORESERVE wasn't set.
1202 */
1203static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1204{
1205 /*
1206 * hugetlb has its own accounting separate from the core VM
1207 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1208 */
1209 if (file && is_file_hugepages(file))
1210 return 0;
1211
1212 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1213}
1214
1215unsigned long mmap_region(struct file *file, unsigned long addr,
1216 unsigned long len, unsigned long flags,
1217 vm_flags_t vm_flags, unsigned long pgoff)
1218{
1219 struct mm_struct *mm = current->mm;
1220 struct vm_area_struct *vma, *prev;
1221 int correct_wcount = 0;
1222 int error;
1223 struct rb_node **rb_link, *rb_parent;
1224 unsigned long charged = 0;
1225 struct inode *inode = file ? file->f_path.dentry->d_inode : NULL;
1226
1227 /* Clear old maps */
1228 error = -ENOMEM;
1229munmap_back:
1230 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1231 if (vma && vma->vm_start < addr + len) {
1232 if (do_munmap(mm, addr, len))
1233 return -ENOMEM;
1234 goto munmap_back;
1235 }
1236
1237 /* Check against address space limit. */
1238 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1239 return -ENOMEM;
1240
1241 /*
1242 * Set 'VM_NORESERVE' if we should not account for the
1243 * memory use of this mapping.
1244 */
1245 if ((flags & MAP_NORESERVE)) {
1246 /* We honor MAP_NORESERVE if allowed to overcommit */
1247 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1248 vm_flags |= VM_NORESERVE;
1249
1250 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1251 if (file && is_file_hugepages(file))
1252 vm_flags |= VM_NORESERVE;
1253 }
1254
1255 /*
1256 * Private writable mapping: check memory availability
1257 */
1258 if (accountable_mapping(file, vm_flags)) {
1259 charged = len >> PAGE_SHIFT;
1260 if (security_vm_enough_memory_mm(mm, charged))
1261 return -ENOMEM;
1262 vm_flags |= VM_ACCOUNT;
1263 }
1264
1265 /*
1266 * Can we just expand an old mapping?
1267 */
1268 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1269 if (vma)
1270 goto out;
1271
1272 /*
1273 * Determine the object being mapped and call the appropriate
1274 * specific mapper. the address has already been validated, but
1275 * not unmapped, but the maps are removed from the list.
1276 */
1277 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1278 if (!vma) {
1279 error = -ENOMEM;
1280 goto unacct_error;
1281 }
1282
1283 vma->vm_mm = mm;
1284 vma->vm_start = addr;
1285 vma->vm_end = addr + len;
1286 vma->vm_flags = vm_flags;
1287 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1288 vma->vm_pgoff = pgoff;
1289 INIT_LIST_HEAD(&vma->anon_vma_chain);
1290
1291 error = -EINVAL; /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1292
1293 if (file) {
1294 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1295 goto free_vma;
1296 if (vm_flags & VM_DENYWRITE) {
1297 error = deny_write_access(file);
1298 if (error)
1299 goto free_vma;
1300 correct_wcount = 1;
1301 }
1302 vma->vm_file = file;
1303 get_file(file);
1304 error = file->f_op->mmap(file, vma);
1305 if (error)
1306 goto unmap_and_free_vma;
1307 if (vm_flags & VM_EXECUTABLE)
1308 added_exe_file_vma(mm);
1309
1310 /* Can addr have changed??
1311 *
1312 * Answer: Yes, several device drivers can do it in their
1313 * f_op->mmap method. -DaveM
1314 */
1315 addr = vma->vm_start;
1316 pgoff = vma->vm_pgoff;
1317 vm_flags = vma->vm_flags;
1318 } else if (vm_flags & VM_SHARED) {
1319 if (unlikely(vm_flags & (VM_GROWSDOWN|VM_GROWSUP)))
1320 goto free_vma;
1321 error = shmem_zero_setup(vma);
1322 if (error)
1323 goto free_vma;
1324 }
1325
1326 if (vma_wants_writenotify(vma)) {
1327 pgprot_t pprot = vma->vm_page_prot;
1328
1329 /* Can vma->vm_page_prot have changed??
1330 *
1331 * Answer: Yes, drivers may have changed it in their
1332 * f_op->mmap method.
1333 *
1334 * Ensures that vmas marked as uncached stay that way.
1335 */
1336 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1337 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1338 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1339 }
1340
1341 vma_link(mm, vma, prev, rb_link, rb_parent);
1342 file = vma->vm_file;
1343
1344 /* Once vma denies write, undo our temporary denial count */
1345 if (correct_wcount)
1346 atomic_inc(&inode->i_writecount);
1347out:
1348 perf_event_mmap(vma);
1349
1350 mm->total_vm += len >> PAGE_SHIFT;
1351 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1352 if (vm_flags & VM_LOCKED) {
1353 if (!mlock_vma_pages_range(vma, addr, addr + len))
1354 mm->locked_vm += (len >> PAGE_SHIFT);
1355 } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1356 make_pages_present(addr, addr + len);
1357
1358 if (file)
1359 uprobe_mmap(vma);
1360
1361 return addr;
1362
1363unmap_and_free_vma:
1364 if (correct_wcount)
1365 atomic_inc(&inode->i_writecount);
1366 vma->vm_file = NULL;
1367 fput(file);
1368
1369 /* Undo any partial mapping done by a device driver. */
1370 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1371 charged = 0;
1372free_vma:
1373 kmem_cache_free(vm_area_cachep, vma);
1374unacct_error:
1375 if (charged)
1376 vm_unacct_memory(charged);
1377 return error;
1378}
1379
1380/* Get an address range which is currently unmapped.
1381 * For shmat() with addr=0.
1382 *
1383 * Ugly calling convention alert:
1384 * Return value with the low bits set means error value,
1385 * ie
1386 * if (ret & ~PAGE_MASK)
1387 * error = ret;
1388 *
1389 * This function "knows" that -ENOMEM has the bits set.
1390 */
1391#ifndef HAVE_ARCH_UNMAPPED_AREA
1392unsigned long
1393arch_get_unmapped_area(struct file *filp, unsigned long addr,
1394 unsigned long len, unsigned long pgoff, unsigned long flags)
1395{
1396 struct mm_struct *mm = current->mm;
1397 struct vm_area_struct *vma;
1398 unsigned long start_addr;
1399
1400 if (len > TASK_SIZE)
1401 return -ENOMEM;
1402
1403 if (flags & MAP_FIXED)
1404 return addr;
1405
1406 if (addr) {
1407 addr = PAGE_ALIGN(addr);
1408 vma = find_vma(mm, addr);
1409 if (TASK_SIZE - len >= addr &&
1410 (!vma || addr + len <= vma->vm_start))
1411 return addr;
1412 }
1413 if (len > mm->cached_hole_size) {
1414 start_addr = addr = mm->free_area_cache;
1415 } else {
1416 start_addr = addr = TASK_UNMAPPED_BASE;
1417 mm->cached_hole_size = 0;
1418 }
1419
1420full_search:
1421 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1422 /* At this point: (!vma || addr < vma->vm_end). */
1423 if (TASK_SIZE - len < addr) {
1424 /*
1425 * Start a new search - just in case we missed
1426 * some holes.
1427 */
1428 if (start_addr != TASK_UNMAPPED_BASE) {
1429 addr = TASK_UNMAPPED_BASE;
1430 start_addr = addr;
1431 mm->cached_hole_size = 0;
1432 goto full_search;
1433 }
1434 return -ENOMEM;
1435 }
1436 if (!vma || addr + len <= vma->vm_start) {
1437 /*
1438 * Remember the place where we stopped the search:
1439 */
1440 mm->free_area_cache = addr + len;
1441 return addr;
1442 }
1443 if (addr + mm->cached_hole_size < vma->vm_start)
1444 mm->cached_hole_size = vma->vm_start - addr;
1445 addr = vma->vm_end;
1446 }
1447}
1448#endif
1449
1450void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1451{
1452 /*
1453 * Is this a new hole at the lowest possible address?
1454 */
1455 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache)
1456 mm->free_area_cache = addr;
1457}
1458
1459/*
1460 * This mmap-allocator allocates new areas top-down from below the
1461 * stack's low limit (the base):
1462 */
1463#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1464unsigned long
1465arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1466 const unsigned long len, const unsigned long pgoff,
1467 const unsigned long flags)
1468{
1469 struct vm_area_struct *vma;
1470 struct mm_struct *mm = current->mm;
1471 unsigned long addr = addr0, start_addr;
1472
1473 /* requested length too big for entire address space */
1474 if (len > TASK_SIZE)
1475 return -ENOMEM;
1476
1477 if (flags & MAP_FIXED)
1478 return addr;
1479
1480 /* requesting a specific address */
1481 if (addr) {
1482 addr = PAGE_ALIGN(addr);
1483 vma = find_vma(mm, addr);
1484 if (TASK_SIZE - len >= addr &&
1485 (!vma || addr + len <= vma->vm_start))
1486 return addr;
1487 }
1488
1489 /* check if free_area_cache is useful for us */
1490 if (len <= mm->cached_hole_size) {
1491 mm->cached_hole_size = 0;
1492 mm->free_area_cache = mm->mmap_base;
1493 }
1494
1495try_again:
1496 /* either no address requested or can't fit in requested address hole */
1497 start_addr = addr = mm->free_area_cache;
1498
1499 if (addr < len)
1500 goto fail;
1501
1502 addr -= len;
1503 do {
1504 /*
1505 * Lookup failure means no vma is above this address,
1506 * else if new region fits below vma->vm_start,
1507 * return with success:
1508 */
1509 vma = find_vma(mm, addr);
1510 if (!vma || addr+len <= vma->vm_start)
1511 /* remember the address as a hint for next time */
1512 return (mm->free_area_cache = addr);
1513
1514 /* remember the largest hole we saw so far */
1515 if (addr + mm->cached_hole_size < vma->vm_start)
1516 mm->cached_hole_size = vma->vm_start - addr;
1517
1518 /* try just below the current vma->vm_start */
1519 addr = vma->vm_start-len;
1520 } while (len < vma->vm_start);
1521
1522fail:
1523 /*
1524 * if hint left us with no space for the requested
1525 * mapping then try again:
1526 *
1527 * Note: this is different with the case of bottomup
1528 * which does the fully line-search, but we use find_vma
1529 * here that causes some holes skipped.
1530 */
1531 if (start_addr != mm->mmap_base) {
1532 mm->free_area_cache = mm->mmap_base;
1533 mm->cached_hole_size = 0;
1534 goto try_again;
1535 }
1536
1537 /*
1538 * A failed mmap() very likely causes application failure,
1539 * so fall back to the bottom-up function here. This scenario
1540 * can happen with large stack limits and large mmap()
1541 * allocations.
1542 */
1543 mm->cached_hole_size = ~0UL;
1544 mm->free_area_cache = TASK_UNMAPPED_BASE;
1545 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1546 /*
1547 * Restore the topdown base:
1548 */
1549 mm->free_area_cache = mm->mmap_base;
1550 mm->cached_hole_size = ~0UL;
1551
1552 return addr;
1553}
1554#endif
1555
1556void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1557{
1558 /*
1559 * Is this a new hole at the highest possible address?
1560 */
1561 if (addr > mm->free_area_cache)
1562 mm->free_area_cache = addr;
1563
1564 /* dont allow allocations above current base */
1565 if (mm->free_area_cache > mm->mmap_base)
1566 mm->free_area_cache = mm->mmap_base;
1567}
1568
1569unsigned long
1570get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1571 unsigned long pgoff, unsigned long flags)
1572{
1573 unsigned long (*get_area)(struct file *, unsigned long,
1574 unsigned long, unsigned long, unsigned long);
1575
1576 unsigned long error = arch_mmap_check(addr, len, flags);
1577 if (error)
1578 return error;
1579
1580 /* Careful about overflows.. */
1581 if (len > TASK_SIZE)
1582 return -ENOMEM;
1583
1584 get_area = current->mm->get_unmapped_area;
1585 if (file && file->f_op && file->f_op->get_unmapped_area)
1586 get_area = file->f_op->get_unmapped_area;
1587 addr = get_area(file, addr, len, pgoff, flags);
1588 if (IS_ERR_VALUE(addr))
1589 return addr;
1590
1591 if (addr > TASK_SIZE - len)
1592 return -ENOMEM;
1593 if (addr & ~PAGE_MASK)
1594 return -EINVAL;
1595
1596 addr = arch_rebalance_pgtables(addr, len);
1597 error = security_mmap_addr(addr);
1598 return error ? error : addr;
1599}
1600
1601EXPORT_SYMBOL(get_unmapped_area);
1602
1603/* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1604struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1605{
1606 struct vm_area_struct *vma = NULL;
1607
1608 if (WARN_ON_ONCE(!mm)) /* Remove this in linux-3.6 */
1609 return NULL;
1610
1611 /* Check the cache first. */
1612 /* (Cache hit rate is typically around 35%.) */
1613 vma = mm->mmap_cache;
1614 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1615 struct rb_node *rb_node;
1616
1617 rb_node = mm->mm_rb.rb_node;
1618 vma = NULL;
1619
1620 while (rb_node) {
1621 struct vm_area_struct *vma_tmp;
1622
1623 vma_tmp = rb_entry(rb_node,
1624 struct vm_area_struct, vm_rb);
1625
1626 if (vma_tmp->vm_end > addr) {
1627 vma = vma_tmp;
1628 if (vma_tmp->vm_start <= addr)
1629 break;
1630 rb_node = rb_node->rb_left;
1631 } else
1632 rb_node = rb_node->rb_right;
1633 }
1634 if (vma)
1635 mm->mmap_cache = vma;
1636 }
1637 return vma;
1638}
1639
1640EXPORT_SYMBOL(find_vma);
1641
1642/*
1643 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1644 */
1645struct vm_area_struct *
1646find_vma_prev(struct mm_struct *mm, unsigned long addr,
1647 struct vm_area_struct **pprev)
1648{
1649 struct vm_area_struct *vma;
1650
1651 vma = find_vma(mm, addr);
1652 if (vma) {
1653 *pprev = vma->vm_prev;
1654 } else {
1655 struct rb_node *rb_node = mm->mm_rb.rb_node;
1656 *pprev = NULL;
1657 while (rb_node) {
1658 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1659 rb_node = rb_node->rb_right;
1660 }
1661 }
1662 return vma;
1663}
1664
1665/*
1666 * Verify that the stack growth is acceptable and
1667 * update accounting. This is shared with both the
1668 * grow-up and grow-down cases.
1669 */
1670static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1671{
1672 struct mm_struct *mm = vma->vm_mm;
1673 struct rlimit *rlim = current->signal->rlim;
1674 unsigned long new_start;
1675
1676 /* address space limit tests */
1677 if (!may_expand_vm(mm, grow))
1678 return -ENOMEM;
1679
1680 /* Stack limit test */
1681 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1682 return -ENOMEM;
1683
1684 /* mlock limit tests */
1685 if (vma->vm_flags & VM_LOCKED) {
1686 unsigned long locked;
1687 unsigned long limit;
1688 locked = mm->locked_vm + grow;
1689 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1690 limit >>= PAGE_SHIFT;
1691 if (locked > limit && !capable(CAP_IPC_LOCK))
1692 return -ENOMEM;
1693 }
1694
1695 /* Check to ensure the stack will not grow into a hugetlb-only region */
1696 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1697 vma->vm_end - size;
1698 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1699 return -EFAULT;
1700
1701 /*
1702 * Overcommit.. This must be the final test, as it will
1703 * update security statistics.
1704 */
1705 if (security_vm_enough_memory_mm(mm, grow))
1706 return -ENOMEM;
1707
1708 /* Ok, everything looks good - let it rip */
1709 mm->total_vm += grow;
1710 if (vma->vm_flags & VM_LOCKED)
1711 mm->locked_vm += grow;
1712 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1713 return 0;
1714}
1715
1716#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1717/*
1718 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1719 * vma is the last one with address > vma->vm_end. Have to extend vma.
1720 */
1721int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1722{
1723 int error;
1724
1725 if (!(vma->vm_flags & VM_GROWSUP))
1726 return -EFAULT;
1727
1728 /*
1729 * We must make sure the anon_vma is allocated
1730 * so that the anon_vma locking is not a noop.
1731 */
1732 if (unlikely(anon_vma_prepare(vma)))
1733 return -ENOMEM;
1734 vma_lock_anon_vma(vma);
1735
1736 /*
1737 * vma->vm_start/vm_end cannot change under us because the caller
1738 * is required to hold the mmap_sem in read mode. We need the
1739 * anon_vma lock to serialize against concurrent expand_stacks.
1740 * Also guard against wrapping around to address 0.
1741 */
1742 if (address < PAGE_ALIGN(address+4))
1743 address = PAGE_ALIGN(address+4);
1744 else {
1745 vma_unlock_anon_vma(vma);
1746 return -ENOMEM;
1747 }
1748 error = 0;
1749
1750 /* Somebody else might have raced and expanded it already */
1751 if (address > vma->vm_end) {
1752 unsigned long size, grow;
1753
1754 size = address - vma->vm_start;
1755 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1756
1757 error = -ENOMEM;
1758 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1759 error = acct_stack_growth(vma, size, grow);
1760 if (!error) {
1761 vma->vm_end = address;
1762 perf_event_mmap(vma);
1763 }
1764 }
1765 }
1766 vma_unlock_anon_vma(vma);
1767 khugepaged_enter_vma_merge(vma);
1768 return error;
1769}
1770#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1771
1772/*
1773 * vma is the first one with address < vma->vm_start. Have to extend vma.
1774 */
1775int expand_downwards(struct vm_area_struct *vma,
1776 unsigned long address)
1777{
1778 int error;
1779
1780 /*
1781 * We must make sure the anon_vma is allocated
1782 * so that the anon_vma locking is not a noop.
1783 */
1784 if (unlikely(anon_vma_prepare(vma)))
1785 return -ENOMEM;
1786
1787 address &= PAGE_MASK;
1788 error = security_mmap_addr(address);
1789 if (error)
1790 return error;
1791
1792 vma_lock_anon_vma(vma);
1793
1794 /*
1795 * vma->vm_start/vm_end cannot change under us because the caller
1796 * is required to hold the mmap_sem in read mode. We need the
1797 * anon_vma lock to serialize against concurrent expand_stacks.
1798 */
1799
1800 /* Somebody else might have raced and expanded it already */
1801 if (address < vma->vm_start) {
1802 unsigned long size, grow;
1803
1804 size = vma->vm_end - address;
1805 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1806
1807 error = -ENOMEM;
1808 if (grow <= vma->vm_pgoff) {
1809 error = acct_stack_growth(vma, size, grow);
1810 if (!error) {
1811 vma->vm_start = address;
1812 vma->vm_pgoff -= grow;
1813 perf_event_mmap(vma);
1814 }
1815 }
1816 }
1817 vma_unlock_anon_vma(vma);
1818 khugepaged_enter_vma_merge(vma);
1819 return error;
1820}
1821
1822#ifdef CONFIG_STACK_GROWSUP
1823int expand_stack(struct vm_area_struct *vma, unsigned long address)
1824{
1825 return expand_upwards(vma, address);
1826}
1827
1828struct vm_area_struct *
1829find_extend_vma(struct mm_struct *mm, unsigned long addr)
1830{
1831 struct vm_area_struct *vma, *prev;
1832
1833 addr &= PAGE_MASK;
1834 vma = find_vma_prev(mm, addr, &prev);
1835 if (vma && (vma->vm_start <= addr))
1836 return vma;
1837 if (!prev || expand_stack(prev, addr))
1838 return NULL;
1839 if (prev->vm_flags & VM_LOCKED) {
1840 mlock_vma_pages_range(prev, addr, prev->vm_end);
1841 }
1842 return prev;
1843}
1844#else
1845int expand_stack(struct vm_area_struct *vma, unsigned long address)
1846{
1847 return expand_downwards(vma, address);
1848}
1849
1850struct vm_area_struct *
1851find_extend_vma(struct mm_struct * mm, unsigned long addr)
1852{
1853 struct vm_area_struct * vma;
1854 unsigned long start;
1855
1856 addr &= PAGE_MASK;
1857 vma = find_vma(mm,addr);
1858 if (!vma)
1859 return NULL;
1860 if (vma->vm_start <= addr)
1861 return vma;
1862 if (!(vma->vm_flags & VM_GROWSDOWN))
1863 return NULL;
1864 start = vma->vm_start;
1865 if (expand_stack(vma, addr))
1866 return NULL;
1867 if (vma->vm_flags & VM_LOCKED) {
1868 mlock_vma_pages_range(vma, addr, start);
1869 }
1870 return vma;
1871}
1872#endif
1873
1874/*
1875 * Ok - we have the memory areas we should free on the vma list,
1876 * so release them, and do the vma updates.
1877 *
1878 * Called with the mm semaphore held.
1879 */
1880static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1881{
1882 unsigned long nr_accounted = 0;
1883
1884 /* Update high watermark before we lower total_vm */
1885 update_hiwater_vm(mm);
1886 do {
1887 long nrpages = vma_pages(vma);
1888
1889 if (vma->vm_flags & VM_ACCOUNT)
1890 nr_accounted += nrpages;
1891 mm->total_vm -= nrpages;
1892 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1893 vma = remove_vma(vma);
1894 } while (vma);
1895 vm_unacct_memory(nr_accounted);
1896 validate_mm(mm);
1897}
1898
1899/*
1900 * Get rid of page table information in the indicated region.
1901 *
1902 * Called with the mm semaphore held.
1903 */
1904static void unmap_region(struct mm_struct *mm,
1905 struct vm_area_struct *vma, struct vm_area_struct *prev,
1906 unsigned long start, unsigned long end)
1907{
1908 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1909 struct mmu_gather tlb;
1910
1911 lru_add_drain();
1912 tlb_gather_mmu(&tlb, mm, 0);
1913 update_hiwater_rss(mm);
1914 unmap_vmas(&tlb, vma, start, end);
1915 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
1916 next ? next->vm_start : 0);
1917 tlb_finish_mmu(&tlb, start, end);
1918}
1919
1920/*
1921 * Create a list of vma's touched by the unmap, removing them from the mm's
1922 * vma list as we go..
1923 */
1924static void
1925detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1926 struct vm_area_struct *prev, unsigned long end)
1927{
1928 struct vm_area_struct **insertion_point;
1929 struct vm_area_struct *tail_vma = NULL;
1930 unsigned long addr;
1931
1932 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1933 vma->vm_prev = NULL;
1934 do {
1935 rb_erase(&vma->vm_rb, &mm->mm_rb);
1936 mm->map_count--;
1937 tail_vma = vma;
1938 vma = vma->vm_next;
1939 } while (vma && vma->vm_start < end);
1940 *insertion_point = vma;
1941 if (vma)
1942 vma->vm_prev = prev;
1943 tail_vma->vm_next = NULL;
1944 if (mm->unmap_area == arch_unmap_area)
1945 addr = prev ? prev->vm_end : mm->mmap_base;
1946 else
1947 addr = vma ? vma->vm_start : mm->mmap_base;
1948 mm->unmap_area(mm, addr);
1949 mm->mmap_cache = NULL; /* Kill the cache. */
1950}
1951
1952/*
1953 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
1954 * munmap path where it doesn't make sense to fail.
1955 */
1956static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1957 unsigned long addr, int new_below)
1958{
1959 struct mempolicy *pol;
1960 struct vm_area_struct *new;
1961 int err = -ENOMEM;
1962
1963 if (is_vm_hugetlb_page(vma) && (addr &
1964 ~(huge_page_mask(hstate_vma(vma)))))
1965 return -EINVAL;
1966
1967 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1968 if (!new)
1969 goto out_err;
1970
1971 /* most fields are the same, copy all, and then fixup */
1972 *new = *vma;
1973
1974 INIT_LIST_HEAD(&new->anon_vma_chain);
1975
1976 if (new_below)
1977 new->vm_end = addr;
1978 else {
1979 new->vm_start = addr;
1980 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1981 }
1982
1983 pol = mpol_dup(vma_policy(vma));
1984 if (IS_ERR(pol)) {
1985 err = PTR_ERR(pol);
1986 goto out_free_vma;
1987 }
1988 vma_set_policy(new, pol);
1989
1990 if (anon_vma_clone(new, vma))
1991 goto out_free_mpol;
1992
1993 if (new->vm_file) {
1994 get_file(new->vm_file);
1995 if (vma->vm_flags & VM_EXECUTABLE)
1996 added_exe_file_vma(mm);
1997 }
1998
1999 if (new->vm_ops && new->vm_ops->open)
2000 new->vm_ops->open(new);
2001
2002 if (new_below)
2003 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2004 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2005 else
2006 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2007
2008 /* Success. */
2009 if (!err)
2010 return 0;
2011
2012 /* Clean everything up if vma_adjust failed. */
2013 if (new->vm_ops && new->vm_ops->close)
2014 new->vm_ops->close(new);
2015 if (new->vm_file) {
2016 if (vma->vm_flags & VM_EXECUTABLE)
2017 removed_exe_file_vma(mm);
2018 fput(new->vm_file);
2019 }
2020 unlink_anon_vmas(new);
2021 out_free_mpol:
2022 mpol_put(pol);
2023 out_free_vma:
2024 kmem_cache_free(vm_area_cachep, new);
2025 out_err:
2026 return err;
2027}
2028
2029/*
2030 * Split a vma into two pieces at address 'addr', a new vma is allocated
2031 * either for the first part or the tail.
2032 */
2033int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2034 unsigned long addr, int new_below)
2035{
2036 if (mm->map_count >= sysctl_max_map_count)
2037 return -ENOMEM;
2038
2039 return __split_vma(mm, vma, addr, new_below);
2040}
2041
2042/* Munmap is split into 2 main parts -- this part which finds
2043 * what needs doing, and the areas themselves, which do the
2044 * work. This now handles partial unmappings.
2045 * Jeremy Fitzhardinge <jeremy@goop.org>
2046 */
2047int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2048{
2049 unsigned long end;
2050 struct vm_area_struct *vma, *prev, *last;
2051
2052 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2053 return -EINVAL;
2054
2055 if ((len = PAGE_ALIGN(len)) == 0)
2056 return -EINVAL;
2057
2058 /* Find the first overlapping VMA */
2059 vma = find_vma(mm, start);
2060 if (!vma)
2061 return 0;
2062 prev = vma->vm_prev;
2063 /* we have start < vma->vm_end */
2064
2065 /* if it doesn't overlap, we have nothing.. */
2066 end = start + len;
2067 if (vma->vm_start >= end)
2068 return 0;
2069
2070 /*
2071 * If we need to split any vma, do it now to save pain later.
2072 *
2073 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2074 * unmapped vm_area_struct will remain in use: so lower split_vma
2075 * places tmp vma above, and higher split_vma places tmp vma below.
2076 */
2077 if (start > vma->vm_start) {
2078 int error;
2079
2080 /*
2081 * Make sure that map_count on return from munmap() will
2082 * not exceed its limit; but let map_count go just above
2083 * its limit temporarily, to help free resources as expected.
2084 */
2085 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2086 return -ENOMEM;
2087
2088 error = __split_vma(mm, vma, start, 0);
2089 if (error)
2090 return error;
2091 prev = vma;
2092 }
2093
2094 /* Does it split the last one? */
2095 last = find_vma(mm, end);
2096 if (last && end > last->vm_start) {
2097 int error = __split_vma(mm, last, end, 1);
2098 if (error)
2099 return error;
2100 }
2101 vma = prev? prev->vm_next: mm->mmap;
2102
2103 /*
2104 * unlock any mlock()ed ranges before detaching vmas
2105 */
2106 if (mm->locked_vm) {
2107 struct vm_area_struct *tmp = vma;
2108 while (tmp && tmp->vm_start < end) {
2109 if (tmp->vm_flags & VM_LOCKED) {
2110 mm->locked_vm -= vma_pages(tmp);
2111 munlock_vma_pages_all(tmp);
2112 }
2113 tmp = tmp->vm_next;
2114 }
2115 }
2116
2117 /*
2118 * Remove the vma's, and unmap the actual pages
2119 */
2120 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2121 unmap_region(mm, vma, prev, start, end);
2122
2123 /* Fix up all other VM information */
2124 remove_vma_list(mm, vma);
2125
2126 return 0;
2127}
2128
2129int vm_munmap(unsigned long start, size_t len)
2130{
2131 int ret;
2132 struct mm_struct *mm = current->mm;
2133
2134 down_write(&mm->mmap_sem);
2135 ret = do_munmap(mm, start, len);
2136 up_write(&mm->mmap_sem);
2137 return ret;
2138}
2139EXPORT_SYMBOL(vm_munmap);
2140
2141SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2142{
2143 profile_munmap(addr);
2144 return vm_munmap(addr, len);
2145}
2146
2147static inline void verify_mm_writelocked(struct mm_struct *mm)
2148{
2149#ifdef CONFIG_DEBUG_VM
2150 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2151 WARN_ON(1);
2152 up_read(&mm->mmap_sem);
2153 }
2154#endif
2155}
2156
2157/*
2158 * this is really a simplified "do_mmap". it only handles
2159 * anonymous maps. eventually we may be able to do some
2160 * brk-specific accounting here.
2161 */
2162static unsigned long do_brk(unsigned long addr, unsigned long len)
2163{
2164 struct mm_struct * mm = current->mm;
2165 struct vm_area_struct * vma, * prev;
2166 unsigned long flags;
2167 struct rb_node ** rb_link, * rb_parent;
2168 pgoff_t pgoff = addr >> PAGE_SHIFT;
2169 int error;
2170
2171 len = PAGE_ALIGN(len);
2172 if (!len)
2173 return addr;
2174
2175 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2176
2177 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2178 if (error & ~PAGE_MASK)
2179 return error;
2180
2181 /*
2182 * mlock MCL_FUTURE?
2183 */
2184 if (mm->def_flags & VM_LOCKED) {
2185 unsigned long locked, lock_limit;
2186 locked = len >> PAGE_SHIFT;
2187 locked += mm->locked_vm;
2188 lock_limit = rlimit(RLIMIT_MEMLOCK);
2189 lock_limit >>= PAGE_SHIFT;
2190 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2191 return -EAGAIN;
2192 }
2193
2194 /*
2195 * mm->mmap_sem is required to protect against another thread
2196 * changing the mappings in case we sleep.
2197 */
2198 verify_mm_writelocked(mm);
2199
2200 /*
2201 * Clear old maps. this also does some error checking for us
2202 */
2203 munmap_back:
2204 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2205 if (vma && vma->vm_start < addr + len) {
2206 if (do_munmap(mm, addr, len))
2207 return -ENOMEM;
2208 goto munmap_back;
2209 }
2210
2211 /* Check against address space limits *after* clearing old maps... */
2212 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2213 return -ENOMEM;
2214
2215 if (mm->map_count > sysctl_max_map_count)
2216 return -ENOMEM;
2217
2218 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2219 return -ENOMEM;
2220
2221 /* Can we just expand an old private anonymous mapping? */
2222 vma = vma_merge(mm, prev, addr, addr + len, flags,
2223 NULL, NULL, pgoff, NULL);
2224 if (vma)
2225 goto out;
2226
2227 /*
2228 * create a vma struct for an anonymous mapping
2229 */
2230 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2231 if (!vma) {
2232 vm_unacct_memory(len >> PAGE_SHIFT);
2233 return -ENOMEM;
2234 }
2235
2236 INIT_LIST_HEAD(&vma->anon_vma_chain);
2237 vma->vm_mm = mm;
2238 vma->vm_start = addr;
2239 vma->vm_end = addr + len;
2240 vma->vm_pgoff = pgoff;
2241 vma->vm_flags = flags;
2242 vma->vm_page_prot = vm_get_page_prot(flags);
2243 vma_link(mm, vma, prev, rb_link, rb_parent);
2244out:
2245 perf_event_mmap(vma);
2246 mm->total_vm += len >> PAGE_SHIFT;
2247 if (flags & VM_LOCKED) {
2248 if (!mlock_vma_pages_range(vma, addr, addr + len))
2249 mm->locked_vm += (len >> PAGE_SHIFT);
2250 }
2251 return addr;
2252}
2253
2254unsigned long vm_brk(unsigned long addr, unsigned long len)
2255{
2256 struct mm_struct *mm = current->mm;
2257 unsigned long ret;
2258
2259 down_write(&mm->mmap_sem);
2260 ret = do_brk(addr, len);
2261 up_write(&mm->mmap_sem);
2262 return ret;
2263}
2264EXPORT_SYMBOL(vm_brk);
2265
2266/* Release all mmaps. */
2267void exit_mmap(struct mm_struct *mm)
2268{
2269 struct mmu_gather tlb;
2270 struct vm_area_struct *vma;
2271 unsigned long nr_accounted = 0;
2272
2273 /* mm's last user has gone, and its about to be pulled down */
2274 mmu_notifier_release(mm);
2275
2276 if (mm->locked_vm) {
2277 vma = mm->mmap;
2278 while (vma) {
2279 if (vma->vm_flags & VM_LOCKED)
2280 munlock_vma_pages_all(vma);
2281 vma = vma->vm_next;
2282 }
2283 }
2284
2285 arch_exit_mmap(mm);
2286
2287 vma = mm->mmap;
2288 if (!vma) /* Can happen if dup_mmap() received an OOM */
2289 return;
2290
2291 lru_add_drain();
2292 flush_cache_mm(mm);
2293 tlb_gather_mmu(&tlb, mm, 1);
2294 /* update_hiwater_rss(mm) here? but nobody should be looking */
2295 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2296 unmap_vmas(&tlb, vma, 0, -1);
2297
2298 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2299 tlb_finish_mmu(&tlb, 0, -1);
2300
2301 /*
2302 * Walk the list again, actually closing and freeing it,
2303 * with preemption enabled, without holding any MM locks.
2304 */
2305 while (vma) {
2306 if (vma->vm_flags & VM_ACCOUNT)
2307 nr_accounted += vma_pages(vma);
2308 vma = remove_vma(vma);
2309 }
2310 vm_unacct_memory(nr_accounted);
2311
2312 BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2313}
2314
2315/* Insert vm structure into process list sorted by address
2316 * and into the inode's i_mmap tree. If vm_file is non-NULL
2317 * then i_mmap_mutex is taken here.
2318 */
2319int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2320{
2321 struct vm_area_struct * __vma, * prev;
2322 struct rb_node ** rb_link, * rb_parent;
2323
2324 /*
2325 * The vm_pgoff of a purely anonymous vma should be irrelevant
2326 * until its first write fault, when page's anon_vma and index
2327 * are set. But now set the vm_pgoff it will almost certainly
2328 * end up with (unless mremap moves it elsewhere before that
2329 * first wfault), so /proc/pid/maps tells a consistent story.
2330 *
2331 * By setting it to reflect the virtual start address of the
2332 * vma, merges and splits can happen in a seamless way, just
2333 * using the existing file pgoff checks and manipulations.
2334 * Similarly in do_mmap_pgoff and in do_brk.
2335 */
2336 if (!vma->vm_file) {
2337 BUG_ON(vma->anon_vma);
2338 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2339 }
2340 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2341 if (__vma && __vma->vm_start < vma->vm_end)
2342 return -ENOMEM;
2343 if ((vma->vm_flags & VM_ACCOUNT) &&
2344 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2345 return -ENOMEM;
2346
2347 if (vma->vm_file && uprobe_mmap(vma))
2348 return -EINVAL;
2349
2350 vma_link(mm, vma, prev, rb_link, rb_parent);
2351 return 0;
2352}
2353
2354/*
2355 * Copy the vma structure to a new location in the same mm,
2356 * prior to moving page table entries, to effect an mremap move.
2357 */
2358struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2359 unsigned long addr, unsigned long len, pgoff_t pgoff)
2360{
2361 struct vm_area_struct *vma = *vmap;
2362 unsigned long vma_start = vma->vm_start;
2363 struct mm_struct *mm = vma->vm_mm;
2364 struct vm_area_struct *new_vma, *prev;
2365 struct rb_node **rb_link, *rb_parent;
2366 struct mempolicy *pol;
2367 bool faulted_in_anon_vma = true;
2368
2369 /*
2370 * If anonymous vma has not yet been faulted, update new pgoff
2371 * to match new location, to increase its chance of merging.
2372 */
2373 if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2374 pgoff = addr >> PAGE_SHIFT;
2375 faulted_in_anon_vma = false;
2376 }
2377
2378 find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2379 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2380 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2381 if (new_vma) {
2382 /*
2383 * Source vma may have been merged into new_vma
2384 */
2385 if (unlikely(vma_start >= new_vma->vm_start &&
2386 vma_start < new_vma->vm_end)) {
2387 /*
2388 * The only way we can get a vma_merge with
2389 * self during an mremap is if the vma hasn't
2390 * been faulted in yet and we were allowed to
2391 * reset the dst vma->vm_pgoff to the
2392 * destination address of the mremap to allow
2393 * the merge to happen. mremap must change the
2394 * vm_pgoff linearity between src and dst vmas
2395 * (in turn preventing a vma_merge) to be
2396 * safe. It is only safe to keep the vm_pgoff
2397 * linear if there are no pages mapped yet.
2398 */
2399 VM_BUG_ON(faulted_in_anon_vma);
2400 *vmap = new_vma;
2401 } else
2402 anon_vma_moveto_tail(new_vma);
2403 } else {
2404 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2405 if (new_vma) {
2406 *new_vma = *vma;
2407 pol = mpol_dup(vma_policy(vma));
2408 if (IS_ERR(pol))
2409 goto out_free_vma;
2410 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2411 if (anon_vma_clone(new_vma, vma))
2412 goto out_free_mempol;
2413 vma_set_policy(new_vma, pol);
2414 new_vma->vm_start = addr;
2415 new_vma->vm_end = addr + len;
2416 new_vma->vm_pgoff = pgoff;
2417 if (new_vma->vm_file) {
2418 get_file(new_vma->vm_file);
2419
2420 if (uprobe_mmap(new_vma))
2421 goto out_free_mempol;
2422
2423 if (vma->vm_flags & VM_EXECUTABLE)
2424 added_exe_file_vma(mm);
2425 }
2426 if (new_vma->vm_ops && new_vma->vm_ops->open)
2427 new_vma->vm_ops->open(new_vma);
2428 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2429 }
2430 }
2431 return new_vma;
2432
2433 out_free_mempol:
2434 mpol_put(pol);
2435 out_free_vma:
2436 kmem_cache_free(vm_area_cachep, new_vma);
2437 return NULL;
2438}
2439
2440/*
2441 * Return true if the calling process may expand its vm space by the passed
2442 * number of pages
2443 */
2444int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2445{
2446 unsigned long cur = mm->total_vm; /* pages */
2447 unsigned long lim;
2448
2449 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2450
2451 if (cur + npages > lim)
2452 return 0;
2453 return 1;
2454}
2455
2456
2457static int special_mapping_fault(struct vm_area_struct *vma,
2458 struct vm_fault *vmf)
2459{
2460 pgoff_t pgoff;
2461 struct page **pages;
2462
2463 /*
2464 * special mappings have no vm_file, and in that case, the mm
2465 * uses vm_pgoff internally. So we have to subtract it from here.
2466 * We are allowed to do this because we are the mm; do not copy
2467 * this code into drivers!
2468 */
2469 pgoff = vmf->pgoff - vma->vm_pgoff;
2470
2471 for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2472 pgoff--;
2473
2474 if (*pages) {
2475 struct page *page = *pages;
2476 get_page(page);
2477 vmf->page = page;
2478 return 0;
2479 }
2480
2481 return VM_FAULT_SIGBUS;
2482}
2483
2484/*
2485 * Having a close hook prevents vma merging regardless of flags.
2486 */
2487static void special_mapping_close(struct vm_area_struct *vma)
2488{
2489}
2490
2491static const struct vm_operations_struct special_mapping_vmops = {
2492 .close = special_mapping_close,
2493 .fault = special_mapping_fault,
2494};
2495
2496/*
2497 * Called with mm->mmap_sem held for writing.
2498 * Insert a new vma covering the given region, with the given flags.
2499 * Its pages are supplied by the given array of struct page *.
2500 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2501 * The region past the last page supplied will always produce SIGBUS.
2502 * The array pointer and the pages it points to are assumed to stay alive
2503 * for as long as this mapping might exist.
2504 */
2505int install_special_mapping(struct mm_struct *mm,
2506 unsigned long addr, unsigned long len,
2507 unsigned long vm_flags, struct page **pages)
2508{
2509 int ret;
2510 struct vm_area_struct *vma;
2511
2512 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2513 if (unlikely(vma == NULL))
2514 return -ENOMEM;
2515
2516 INIT_LIST_HEAD(&vma->anon_vma_chain);
2517 vma->vm_mm = mm;
2518 vma->vm_start = addr;
2519 vma->vm_end = addr + len;
2520
2521 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2522 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2523
2524 vma->vm_ops = &special_mapping_vmops;
2525 vma->vm_private_data = pages;
2526
2527 ret = insert_vm_struct(mm, vma);
2528 if (ret)
2529 goto out;
2530
2531 mm->total_vm += len >> PAGE_SHIFT;
2532
2533 perf_event_mmap(vma);
2534
2535 return 0;
2536
2537out:
2538 kmem_cache_free(vm_area_cachep, vma);
2539 return ret;
2540}
2541
2542static DEFINE_MUTEX(mm_all_locks_mutex);
2543
2544static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2545{
2546 if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2547 /*
2548 * The LSB of head.next can't change from under us
2549 * because we hold the mm_all_locks_mutex.
2550 */
2551 mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem);
2552 /*
2553 * We can safely modify head.next after taking the
2554 * anon_vma->root->mutex. If some other vma in this mm shares
2555 * the same anon_vma we won't take it again.
2556 *
2557 * No need of atomic instructions here, head.next
2558 * can't change from under us thanks to the
2559 * anon_vma->root->mutex.
2560 */
2561 if (__test_and_set_bit(0, (unsigned long *)
2562 &anon_vma->root->head.next))
2563 BUG();
2564 }
2565}
2566
2567static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2568{
2569 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2570 /*
2571 * AS_MM_ALL_LOCKS can't change from under us because
2572 * we hold the mm_all_locks_mutex.
2573 *
2574 * Operations on ->flags have to be atomic because
2575 * even if AS_MM_ALL_LOCKS is stable thanks to the
2576 * mm_all_locks_mutex, there may be other cpus
2577 * changing other bitflags in parallel to us.
2578 */
2579 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2580 BUG();
2581 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2582 }
2583}
2584
2585/*
2586 * This operation locks against the VM for all pte/vma/mm related
2587 * operations that could ever happen on a certain mm. This includes
2588 * vmtruncate, try_to_unmap, and all page faults.
2589 *
2590 * The caller must take the mmap_sem in write mode before calling
2591 * mm_take_all_locks(). The caller isn't allowed to release the
2592 * mmap_sem until mm_drop_all_locks() returns.
2593 *
2594 * mmap_sem in write mode is required in order to block all operations
2595 * that could modify pagetables and free pages without need of
2596 * altering the vma layout (for example populate_range() with
2597 * nonlinear vmas). It's also needed in write mode to avoid new
2598 * anon_vmas to be associated with existing vmas.
2599 *
2600 * A single task can't take more than one mm_take_all_locks() in a row
2601 * or it would deadlock.
2602 *
2603 * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2604 * mapping->flags avoid to take the same lock twice, if more than one
2605 * vma in this mm is backed by the same anon_vma or address_space.
2606 *
2607 * We can take all the locks in random order because the VM code
2608 * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2609 * takes more than one of them in a row. Secondly we're protected
2610 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2611 *
2612 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2613 * that may have to take thousand of locks.
2614 *
2615 * mm_take_all_locks() can fail if it's interrupted by signals.
2616 */
2617int mm_take_all_locks(struct mm_struct *mm)
2618{
2619 struct vm_area_struct *vma;
2620 struct anon_vma_chain *avc;
2621
2622 BUG_ON(down_read_trylock(&mm->mmap_sem));
2623
2624 mutex_lock(&mm_all_locks_mutex);
2625
2626 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2627 if (signal_pending(current))
2628 goto out_unlock;
2629 if (vma->vm_file && vma->vm_file->f_mapping)
2630 vm_lock_mapping(mm, vma->vm_file->f_mapping);
2631 }
2632
2633 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2634 if (signal_pending(current))
2635 goto out_unlock;
2636 if (vma->anon_vma)
2637 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2638 vm_lock_anon_vma(mm, avc->anon_vma);
2639 }
2640
2641 return 0;
2642
2643out_unlock:
2644 mm_drop_all_locks(mm);
2645 return -EINTR;
2646}
2647
2648static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2649{
2650 if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2651 /*
2652 * The LSB of head.next can't change to 0 from under
2653 * us because we hold the mm_all_locks_mutex.
2654 *
2655 * We must however clear the bitflag before unlocking
2656 * the vma so the users using the anon_vma->head will
2657 * never see our bitflag.
2658 *
2659 * No need of atomic instructions here, head.next
2660 * can't change from under us until we release the
2661 * anon_vma->root->mutex.
2662 */
2663 if (!__test_and_clear_bit(0, (unsigned long *)
2664 &anon_vma->root->head.next))
2665 BUG();
2666 anon_vma_unlock(anon_vma);
2667 }
2668}
2669
2670static void vm_unlock_mapping(struct address_space *mapping)
2671{
2672 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2673 /*
2674 * AS_MM_ALL_LOCKS can't change to 0 from under us
2675 * because we hold the mm_all_locks_mutex.
2676 */
2677 mutex_unlock(&mapping->i_mmap_mutex);
2678 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2679 &mapping->flags))
2680 BUG();
2681 }
2682}
2683
2684/*
2685 * The mmap_sem cannot be released by the caller until
2686 * mm_drop_all_locks() returns.
2687 */
2688void mm_drop_all_locks(struct mm_struct *mm)
2689{
2690 struct vm_area_struct *vma;
2691 struct anon_vma_chain *avc;
2692
2693 BUG_ON(down_read_trylock(&mm->mmap_sem));
2694 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2695
2696 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2697 if (vma->anon_vma)
2698 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2699 vm_unlock_anon_vma(avc->anon_vma);
2700 if (vma->vm_file && vma->vm_file->f_mapping)
2701 vm_unlock_mapping(vma->vm_file->f_mapping);
2702 }
2703
2704 mutex_unlock(&mm_all_locks_mutex);
2705}
2706
2707/*
2708 * initialise the VMA slab
2709 */
2710void __init mmap_init(void)
2711{
2712 int ret;
2713
2714 ret = percpu_counter_init(&vm_committed_as, 0);
2715 VM_BUG_ON(ret);
2716}