6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
9 #include <linux/slab.h>
10 #include <linux/backing-dev.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>
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>
35 #include <asm/uaccess.h>
36 #include <asm/cacheflush.h>
38 #include <asm/mmu_context.h>
42 #ifndef arch_mmap_check
43 #define arch_mmap_check(addr, len, flags) (0)
46 #ifndef arch_rebalance_pgtables
47 #define arch_rebalance_pgtables(addr, len) (addr)
50 static 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);
55 * WARNING: the debugging will use recursive algorithms so never enable this
56 * unless you know what you are doing.
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:
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
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
75 pgprot_t protection_map[16] = {
76 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
77 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
80 pgprot_t vm_get_page_prot(unsigned long vm_flags)
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)));
86 EXPORT_SYMBOL(vm_get_page_prot);
88 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS; /* heuristic overcommit */
89 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
90 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
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.
95 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
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.
102 * We currently support three overcommit policies, which are set via the
103 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
105 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
106 * Additional code 2002 Jul 20 by Robert Love.
108 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
110 * Note this is a helper function intended to be used by LSMs which
111 * wish to use this logic.
113 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
115 unsigned long free, allowed;
117 vm_acct_memory(pages);
120 * Sometimes we want to use more memory than we have
122 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
125 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
126 free = global_page_state(NR_FREE_PAGES);
127 free += global_page_state(NR_FILE_PAGES);
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.
135 free -= global_page_state(NR_SHMEM);
137 free += nr_swap_pages;
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
145 free += global_page_state(NR_SLAB_RECLAIMABLE);
148 * Leave reserved pages. The pages are not for anonymous pages.
150 if (free <= totalreserve_pages)
153 free -= totalreserve_pages;
156 * Leave the last 3% for root
167 allowed = (totalram_pages - hugetlb_total_pages())
168 * sysctl_overcommit_ratio / 100;
170 * Leave the last 3% for root
173 allowed -= allowed / 32;
174 allowed += total_swap_pages;
176 /* Don't let a single process grow too big:
177 leave 3% of the size of this process for other processes */
179 allowed -= mm->total_vm / 32;
181 if (percpu_counter_read_positive(&vm_committed_as) < allowed)
184 vm_unacct_memory(pages);
190 * Requires inode->i_mapping->i_mmap_mutex
192 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
193 struct file *file, struct address_space *mapping)
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--;
200 flush_dcache_mmap_lock(mapping);
201 if (unlikely(vma->vm_flags & VM_NONLINEAR))
202 list_del_init(&vma->shared.vm_set.list);
204 vma_prio_tree_remove(vma, &mapping->i_mmap);
205 flush_dcache_mmap_unlock(mapping);
209 * Unlink a file-based vm structure from its prio_tree, to hide
210 * vma from rmap and vmtruncate before freeing its page tables.
212 void unlink_file_vma(struct vm_area_struct *vma)
214 struct file *file = vma->vm_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);
225 * Close a vm structure and free it, returning the next.
227 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
229 struct vm_area_struct *next = vma->vm_next;
232 if (vma->vm_ops && vma->vm_ops->close)
233 vma->vm_ops->close(vma);
236 mpol_put(vma_policy(vma));
237 kmem_cache_free(vm_area_cachep, vma);
241 static unsigned long do_brk(unsigned long addr, unsigned long len);
243 SYSCALL_DEFINE1(brk, unsigned long, brk)
245 unsigned long rlim, retval;
246 unsigned long newbrk, oldbrk;
247 struct mm_struct *mm = current->mm;
248 unsigned long min_brk;
250 down_write(&mm->mmap_sem);
252 #ifdef CONFIG_COMPAT_BRK
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
258 if (current->brk_randomized)
259 min_brk = mm->start_brk;
261 min_brk = mm->end_data;
263 min_brk = mm->start_brk;
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
274 rlim = rlimit(RLIMIT_DATA);
275 if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
276 (mm->end_data - mm->start_data) > rlim)
279 newbrk = PAGE_ALIGN(brk);
280 oldbrk = PAGE_ALIGN(mm->brk);
281 if (oldbrk == newbrk)
284 /* Always allow shrinking brk. */
285 if (brk <= mm->brk) {
286 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
291 /* Check against existing mmap mappings. */
292 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
295 /* Ok, looks good - let it rip. */
296 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
302 up_write(&mm->mmap_sem);
307 static int browse_rb(struct rb_root *root)
310 struct rb_node *nd, *pn = NULL;
311 unsigned long prev = 0, pend = 0;
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);
324 prev = vma->vm_start;
328 for (nd = pn; nd; nd = rb_prev(nd)) {
332 printk("backwards %d, forwards %d\n", j, i), i = 0;
336 void validate_mm(struct mm_struct *mm)
340 struct vm_area_struct *tmp = mm->mmap;
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;
353 #define validate_mm(mm) do { } while (0)
356 static struct vm_area_struct *
357 find_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)
361 struct vm_area_struct * vma;
362 struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
364 __rb_link = &mm->mm_rb.rb_node;
365 rb_prev = __rb_parent = NULL;
369 struct vm_area_struct *vma_tmp;
371 __rb_parent = *__rb_link;
372 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
374 if (vma_tmp->vm_end > addr) {
376 if (vma_tmp->vm_start <= addr)
378 __rb_link = &__rb_parent->rb_left;
380 rb_prev = __rb_parent;
381 __rb_link = &__rb_parent->rb_right;
387 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
388 *rb_link = __rb_link;
389 *rb_parent = __rb_parent;
393 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
394 struct rb_node **rb_link, struct rb_node *rb_parent)
396 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
397 rb_insert_color(&vma->vm_rb, &mm->mm_rb);
400 static void __vma_link_file(struct vm_area_struct *vma)
406 struct address_space *mapping = file->f_mapping;
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++;
413 flush_dcache_mmap_lock(mapping);
414 if (unlikely(vma->vm_flags & VM_NONLINEAR))
415 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
417 vma_prio_tree_insert(vma, &mapping->i_mmap);
418 flush_dcache_mmap_unlock(mapping);
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)
427 __vma_link_list(mm, vma, prev, rb_parent);
428 __vma_link_rb(mm, vma, rb_link, rb_parent);
431 static 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)
435 struct address_space *mapping = NULL;
438 mapping = vma->vm_file->f_mapping;
441 mutex_lock(&mapping->i_mmap_mutex);
443 __vma_link(mm, vma, prev, rb_link, rb_parent);
444 __vma_link_file(vma);
447 mutex_unlock(&mapping->i_mmap_mutex);
454 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
455 * mm's list and rbtree. It has already been inserted into the prio_tree.
457 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
459 struct vm_area_struct *__vma, *prev;
460 struct rb_node **rb_link, *rb_parent;
462 __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
463 BUG_ON(__vma && __vma->vm_start < vma->vm_end);
464 __vma_link(mm, vma, prev, rb_link, rb_parent);
469 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
470 struct vm_area_struct *prev)
472 struct vm_area_struct *next = vma->vm_next;
474 prev->vm_next = next;
476 next->vm_prev = prev;
477 rb_erase(&vma->vm_rb, &mm->mm_rb);
478 if (mm->mmap_cache == vma)
479 mm->mmap_cache = prev;
483 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
484 * is already present in an i_mmap tree without adjusting the tree.
485 * The following helper function should be used when such adjustments
486 * are necessary. The "insert" vma (if any) is to be inserted
487 * before we drop the necessary locks.
489 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
490 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
492 struct mm_struct *mm = vma->vm_mm;
493 struct vm_area_struct *next = vma->vm_next;
494 struct vm_area_struct *importer = NULL;
495 struct address_space *mapping = NULL;
496 struct prio_tree_root *root = NULL;
497 struct anon_vma *anon_vma = NULL;
498 struct file *file = vma->vm_file;
499 long adjust_next = 0;
502 if (next && !insert) {
503 struct vm_area_struct *exporter = NULL;
505 if (end >= next->vm_end) {
507 * vma expands, overlapping all the next, and
508 * perhaps the one after too (mprotect case 6).
510 again: remove_next = 1 + (end > next->vm_end);
514 } else if (end > next->vm_start) {
516 * vma expands, overlapping part of the next:
517 * mprotect case 5 shifting the boundary up.
519 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
522 } else if (end < vma->vm_end) {
524 * vma shrinks, and !insert tells it's not
525 * split_vma inserting another: so it must be
526 * mprotect case 4 shifting the boundary down.
528 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
534 * Easily overlooked: when mprotect shifts the boundary,
535 * make sure the expanding vma has anon_vma set if the
536 * shrinking vma had, to cover any anon pages imported.
538 if (exporter && exporter->anon_vma && !importer->anon_vma) {
539 if (anon_vma_clone(importer, exporter))
541 importer->anon_vma = exporter->anon_vma;
546 mapping = file->f_mapping;
547 if (!(vma->vm_flags & VM_NONLINEAR)) {
548 root = &mapping->i_mmap;
549 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
552 uprobe_munmap(next, next->vm_start,
556 mutex_lock(&mapping->i_mmap_mutex);
559 * Put into prio_tree now, so instantiated pages
560 * are visible to arm/parisc __flush_dcache_page
561 * throughout; but we cannot insert into address
562 * space until vma start or end is updated.
564 __vma_link_file(insert);
568 vma_adjust_trans_huge(vma, start, end, adjust_next);
571 * When changing only vma->vm_end, we don't really need anon_vma
572 * lock. This is a fairly rare case by itself, but the anon_vma
573 * lock may be shared between many sibling processes. Skipping
574 * the lock for brk adjustments makes a difference sometimes.
576 if (vma->anon_vma && (importer || start != vma->vm_start)) {
577 anon_vma = vma->anon_vma;
578 anon_vma_lock(anon_vma);
582 flush_dcache_mmap_lock(mapping);
583 vma_prio_tree_remove(vma, root);
585 vma_prio_tree_remove(next, root);
588 vma->vm_start = start;
590 vma->vm_pgoff = pgoff;
592 next->vm_start += adjust_next << PAGE_SHIFT;
593 next->vm_pgoff += adjust_next;
598 vma_prio_tree_insert(next, root);
599 vma_prio_tree_insert(vma, root);
600 flush_dcache_mmap_unlock(mapping);
605 * vma_merge has merged next into vma, and needs
606 * us to remove next before dropping the locks.
608 __vma_unlink(mm, next, vma);
610 __remove_shared_vm_struct(next, file, mapping);
613 * split_vma has split insert from vma, and needs
614 * us to insert it before dropping the locks
615 * (it may either follow vma or precede it).
617 __insert_vm_struct(mm, insert);
621 anon_vma_unlock(anon_vma);
623 mutex_unlock(&mapping->i_mmap_mutex);
634 uprobe_munmap(next, next->vm_start, next->vm_end);
638 anon_vma_merge(vma, next);
640 mpol_put(vma_policy(next));
641 kmem_cache_free(vm_area_cachep, next);
643 * In mprotect's case 6 (see comments on vma_merge),
644 * we must remove another next too. It would clutter
645 * up the code too much to do both in one go.
647 if (remove_next == 2) {
661 * If the vma has a ->close operation then the driver probably needs to release
662 * per-vma resources, so we don't attempt to merge those.
664 static inline int is_mergeable_vma(struct vm_area_struct *vma,
665 struct file *file, unsigned long vm_flags)
667 if (vma->vm_flags ^ vm_flags)
669 if (vma->vm_file != file)
671 if (vma->vm_ops && vma->vm_ops->close)
676 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
677 struct anon_vma *anon_vma2,
678 struct vm_area_struct *vma)
681 * The list_is_singular() test is to avoid merging VMA cloned from
682 * parents. This can improve scalability caused by anon_vma lock.
684 if ((!anon_vma1 || !anon_vma2) && (!vma ||
685 list_is_singular(&vma->anon_vma_chain)))
687 return anon_vma1 == anon_vma2;
691 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
692 * in front of (at a lower virtual address and file offset than) the vma.
694 * We cannot merge two vmas if they have differently assigned (non-NULL)
695 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
697 * We don't check here for the merged mmap wrapping around the end of pagecache
698 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
699 * wrap, nor mmaps which cover the final page at index -1UL.
702 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
703 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
705 if (is_mergeable_vma(vma, file, vm_flags) &&
706 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
707 if (vma->vm_pgoff == vm_pgoff)
714 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
715 * beyond (at a higher virtual address and file offset than) the vma.
717 * We cannot merge two vmas if they have differently assigned (non-NULL)
718 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
721 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
722 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
724 if (is_mergeable_vma(vma, file, vm_flags) &&
725 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
727 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
728 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
735 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
736 * whether that can be merged with its predecessor or its successor.
737 * Or both (it neatly fills a hole).
739 * In most cases - when called for mmap, brk or mremap - [addr,end) is
740 * certain not to be mapped by the time vma_merge is called; but when
741 * called for mprotect, it is certain to be already mapped (either at
742 * an offset within prev, or at the start of next), and the flags of
743 * this area are about to be changed to vm_flags - and the no-change
744 * case has already been eliminated.
746 * The following mprotect cases have to be considered, where AAAA is
747 * the area passed down from mprotect_fixup, never extending beyond one
748 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
750 * AAAA AAAA AAAA AAAA
751 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
752 * cannot merge might become might become might become
753 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
754 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
755 * mremap move: PPPPNNNNNNNN 8
757 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
758 * might become case 1 below case 2 below case 3 below
760 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
761 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
763 struct vm_area_struct *vma_merge(struct mm_struct *mm,
764 struct vm_area_struct *prev, unsigned long addr,
765 unsigned long end, unsigned long vm_flags,
766 struct anon_vma *anon_vma, struct file *file,
767 pgoff_t pgoff, struct mempolicy *policy)
769 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
770 struct vm_area_struct *area, *next;
774 * We later require that vma->vm_flags == vm_flags,
775 * so this tests vma->vm_flags & VM_SPECIAL, too.
777 if (vm_flags & VM_SPECIAL)
781 next = prev->vm_next;
785 if (next && next->vm_end == end) /* cases 6, 7, 8 */
786 next = next->vm_next;
789 * Can it merge with the predecessor?
791 if (prev && prev->vm_end == addr &&
792 mpol_equal(vma_policy(prev), policy) &&
793 can_vma_merge_after(prev, vm_flags,
794 anon_vma, file, pgoff)) {
796 * OK, it can. Can we now merge in the successor as well?
798 if (next && end == next->vm_start &&
799 mpol_equal(policy, vma_policy(next)) &&
800 can_vma_merge_before(next, vm_flags,
801 anon_vma, file, pgoff+pglen) &&
802 is_mergeable_anon_vma(prev->anon_vma,
803 next->anon_vma, NULL)) {
805 err = vma_adjust(prev, prev->vm_start,
806 next->vm_end, prev->vm_pgoff, NULL);
807 } else /* cases 2, 5, 7 */
808 err = vma_adjust(prev, prev->vm_start,
809 end, prev->vm_pgoff, NULL);
812 khugepaged_enter_vma_merge(prev);
817 * Can this new request be merged in front of next?
819 if (next && end == next->vm_start &&
820 mpol_equal(policy, vma_policy(next)) &&
821 can_vma_merge_before(next, vm_flags,
822 anon_vma, file, pgoff+pglen)) {
823 if (prev && addr < prev->vm_end) /* case 4 */
824 err = vma_adjust(prev, prev->vm_start,
825 addr, prev->vm_pgoff, NULL);
826 else /* cases 3, 8 */
827 err = vma_adjust(area, addr, next->vm_end,
828 next->vm_pgoff - pglen, NULL);
831 khugepaged_enter_vma_merge(area);
839 * Rough compatbility check to quickly see if it's even worth looking
840 * at sharing an anon_vma.
842 * They need to have the same vm_file, and the flags can only differ
843 * in things that mprotect may change.
845 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
846 * we can merge the two vma's. For example, we refuse to merge a vma if
847 * there is a vm_ops->close() function, because that indicates that the
848 * driver is doing some kind of reference counting. But that doesn't
849 * really matter for the anon_vma sharing case.
851 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
853 return a->vm_end == b->vm_start &&
854 mpol_equal(vma_policy(a), vma_policy(b)) &&
855 a->vm_file == b->vm_file &&
856 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
857 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
861 * Do some basic sanity checking to see if we can re-use the anon_vma
862 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
863 * the same as 'old', the other will be the new one that is trying
864 * to share the anon_vma.
866 * NOTE! This runs with mm_sem held for reading, so it is possible that
867 * the anon_vma of 'old' is concurrently in the process of being set up
868 * by another page fault trying to merge _that_. But that's ok: if it
869 * is being set up, that automatically means that it will be a singleton
870 * acceptable for merging, so we can do all of this optimistically. But
871 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
873 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
874 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
875 * is to return an anon_vma that is "complex" due to having gone through
878 * We also make sure that the two vma's are compatible (adjacent,
879 * and with the same memory policies). That's all stable, even with just
880 * a read lock on the mm_sem.
882 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
884 if (anon_vma_compatible(a, b)) {
885 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
887 if (anon_vma && list_is_singular(&old->anon_vma_chain))
894 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
895 * neighbouring vmas for a suitable anon_vma, before it goes off
896 * to allocate a new anon_vma. It checks because a repetitive
897 * sequence of mprotects and faults may otherwise lead to distinct
898 * anon_vmas being allocated, preventing vma merge in subsequent
901 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
903 struct anon_vma *anon_vma;
904 struct vm_area_struct *near;
910 anon_vma = reusable_anon_vma(near, vma, near);
918 anon_vma = reusable_anon_vma(near, near, vma);
923 * There's no absolute need to look only at touching neighbours:
924 * we could search further afield for "compatible" anon_vmas.
925 * But it would probably just be a waste of time searching,
926 * or lead to too many vmas hanging off the same anon_vma.
927 * We're trying to allow mprotect remerging later on,
928 * not trying to minimize memory used for anon_vmas.
933 #ifdef CONFIG_PROC_FS
934 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
935 struct file *file, long pages)
937 const unsigned long stack_flags
938 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
940 mm->total_vm += pages;
943 mm->shared_vm += pages;
944 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
945 mm->exec_vm += pages;
946 } else if (flags & stack_flags)
947 mm->stack_vm += pages;
948 if (flags & (VM_RESERVED|VM_IO))
949 mm->reserved_vm += pages;
951 #endif /* CONFIG_PROC_FS */
954 * If a hint addr is less than mmap_min_addr change hint to be as
955 * low as possible but still greater than mmap_min_addr
957 static inline unsigned long round_hint_to_min(unsigned long hint)
960 if (((void *)hint != NULL) &&
961 (hint < mmap_min_addr))
962 return PAGE_ALIGN(mmap_min_addr);
967 * The caller must hold down_write(¤t->mm->mmap_sem).
970 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
971 unsigned long len, unsigned long prot,
972 unsigned long flags, unsigned long pgoff)
974 struct mm_struct * mm = current->mm;
979 * Does the application expect PROT_READ to imply PROT_EXEC?
981 * (the exception is when the underlying filesystem is noexec
982 * mounted, in which case we dont add PROT_EXEC.)
984 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
985 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
991 if (!(flags & MAP_FIXED))
992 addr = round_hint_to_min(addr);
994 /* Careful about overflows.. */
995 len = PAGE_ALIGN(len);
999 /* offset overflow? */
1000 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1003 /* Too many mappings? */
1004 if (mm->map_count > sysctl_max_map_count)
1007 /* Obtain the address to map to. we verify (or select) it and ensure
1008 * that it represents a valid section of the address space.
1010 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1011 if (addr & ~PAGE_MASK)
1014 /* Do simple checking here so the lower-level routines won't have
1015 * to. we assume access permissions have been handled by the open
1016 * of the memory object, so we don't do any here.
1018 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1019 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1021 if (flags & MAP_LOCKED)
1022 if (!can_do_mlock())
1025 /* mlock MCL_FUTURE? */
1026 if (vm_flags & VM_LOCKED) {
1027 unsigned long locked, lock_limit;
1028 locked = len >> PAGE_SHIFT;
1029 locked += mm->locked_vm;
1030 lock_limit = rlimit(RLIMIT_MEMLOCK);
1031 lock_limit >>= PAGE_SHIFT;
1032 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1036 inode = file ? file->f_path.dentry->d_inode : NULL;
1039 switch (flags & MAP_TYPE) {
1041 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1045 * Make sure we don't allow writing to an append-only
1048 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1052 * Make sure there are no mandatory locks on the file.
1054 if (locks_verify_locked(inode))
1057 vm_flags |= VM_SHARED | VM_MAYSHARE;
1058 if (!(file->f_mode & FMODE_WRITE))
1059 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1063 if (!(file->f_mode & FMODE_READ))
1065 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1066 if (vm_flags & VM_EXEC)
1068 vm_flags &= ~VM_MAYEXEC;
1071 if (!file->f_op || !file->f_op->mmap)
1079 switch (flags & MAP_TYPE) {
1085 vm_flags |= VM_SHARED | VM_MAYSHARE;
1089 * Set pgoff according to addr for anon_vma.
1091 pgoff = addr >> PAGE_SHIFT;
1098 return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1101 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1102 unsigned long, prot, unsigned long, flags,
1103 unsigned long, fd, unsigned long, pgoff)
1105 struct file *file = NULL;
1106 unsigned long retval = -EBADF;
1108 if (!(flags & MAP_ANONYMOUS)) {
1109 audit_mmap_fd(fd, flags);
1110 if (unlikely(flags & MAP_HUGETLB))
1115 } else if (flags & MAP_HUGETLB) {
1116 struct user_struct *user = NULL;
1118 * VM_NORESERVE is used because the reservations will be
1119 * taken when vm_ops->mmap() is called
1120 * A dummy user value is used because we are not locking
1121 * memory so no accounting is necessary
1123 file = hugetlb_file_setup(HUGETLB_ANON_FILE, addr, len,
1124 VM_NORESERVE, &user,
1125 HUGETLB_ANONHUGE_INODE);
1127 return PTR_ERR(file);
1130 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1132 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1139 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1140 struct mmap_arg_struct {
1144 unsigned long flags;
1146 unsigned long offset;
1149 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1151 struct mmap_arg_struct a;
1153 if (copy_from_user(&a, arg, sizeof(a)))
1155 if (a.offset & ~PAGE_MASK)
1158 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1159 a.offset >> PAGE_SHIFT);
1161 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1164 * Some shared mappigns will want the pages marked read-only
1165 * to track write events. If so, we'll downgrade vm_page_prot
1166 * to the private version (using protection_map[] without the
1169 int vma_wants_writenotify(struct vm_area_struct *vma)
1171 vm_flags_t vm_flags = vma->vm_flags;
1173 /* If it was private or non-writable, the write bit is already clear */
1174 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1177 /* The backer wishes to know when pages are first written to? */
1178 if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1181 /* The open routine did something to the protections already? */
1182 if (pgprot_val(vma->vm_page_prot) !=
1183 pgprot_val(vm_get_page_prot(vm_flags)))
1186 /* Specialty mapping? */
1187 if (vm_flags & VM_PFNMAP)
1190 /* Can the mapping track the dirty pages? */
1191 return vma->vm_file && vma->vm_file->f_mapping &&
1192 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1196 * We account for memory if it's a private writeable mapping,
1197 * not hugepages and VM_NORESERVE wasn't set.
1199 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1202 * hugetlb has its own accounting separate from the core VM
1203 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1205 if (file && is_file_hugepages(file))
1208 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1211 unsigned long mmap_region(struct file *file, unsigned long addr,
1212 unsigned long len, unsigned long flags,
1213 vm_flags_t vm_flags, unsigned long pgoff)
1215 struct mm_struct *mm = current->mm;
1216 struct vm_area_struct *vma, *prev;
1217 int correct_wcount = 0;
1219 struct rb_node **rb_link, *rb_parent;
1220 unsigned long charged = 0;
1221 struct inode *inode = file ? file->f_path.dentry->d_inode : NULL;
1223 /* Clear old maps */
1226 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1227 if (vma && vma->vm_start < addr + len) {
1228 if (do_munmap(mm, addr, len))
1233 /* Check against address space limit. */
1234 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1238 * Set 'VM_NORESERVE' if we should not account for the
1239 * memory use of this mapping.
1241 if ((flags & MAP_NORESERVE)) {
1242 /* We honor MAP_NORESERVE if allowed to overcommit */
1243 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1244 vm_flags |= VM_NORESERVE;
1246 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1247 if (file && is_file_hugepages(file))
1248 vm_flags |= VM_NORESERVE;
1252 * Private writable mapping: check memory availability
1254 if (accountable_mapping(file, vm_flags)) {
1255 charged = len >> PAGE_SHIFT;
1256 if (security_vm_enough_memory_mm(mm, charged))
1258 vm_flags |= VM_ACCOUNT;
1262 * Can we just expand an old mapping?
1264 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1269 * Determine the object being mapped and call the appropriate
1270 * specific mapper. the address has already been validated, but
1271 * not unmapped, but the maps are removed from the list.
1273 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1280 vma->vm_start = addr;
1281 vma->vm_end = addr + len;
1282 vma->vm_flags = vm_flags;
1283 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1284 vma->vm_pgoff = pgoff;
1285 INIT_LIST_HEAD(&vma->anon_vma_chain);
1287 error = -EINVAL; /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1290 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1292 if (vm_flags & VM_DENYWRITE) {
1293 error = deny_write_access(file);
1298 vma->vm_file = file;
1300 error = file->f_op->mmap(file, vma);
1302 goto unmap_and_free_vma;
1304 /* Can addr have changed??
1306 * Answer: Yes, several device drivers can do it in their
1307 * f_op->mmap method. -DaveM
1309 addr = vma->vm_start;
1310 pgoff = vma->vm_pgoff;
1311 vm_flags = vma->vm_flags;
1312 } else if (vm_flags & VM_SHARED) {
1313 if (unlikely(vm_flags & (VM_GROWSDOWN|VM_GROWSUP)))
1315 error = shmem_zero_setup(vma);
1320 if (vma_wants_writenotify(vma)) {
1321 pgprot_t pprot = vma->vm_page_prot;
1323 /* Can vma->vm_page_prot have changed??
1325 * Answer: Yes, drivers may have changed it in their
1326 * f_op->mmap method.
1328 * Ensures that vmas marked as uncached stay that way.
1330 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1331 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1332 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1335 vma_link(mm, vma, prev, rb_link, rb_parent);
1336 file = vma->vm_file;
1338 /* Once vma denies write, undo our temporary denial count */
1340 atomic_inc(&inode->i_writecount);
1342 perf_event_mmap(vma);
1344 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1345 if (vm_flags & VM_LOCKED) {
1346 if (!mlock_vma_pages_range(vma, addr, addr + len))
1347 mm->locked_vm += (len >> PAGE_SHIFT);
1348 } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1349 make_pages_present(addr, addr + len);
1351 if (file && uprobe_mmap(vma))
1352 /* matching probes but cannot insert */
1353 goto unmap_and_free_vma;
1359 atomic_inc(&inode->i_writecount);
1360 vma->vm_file = NULL;
1363 /* Undo any partial mapping done by a device driver. */
1364 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1367 kmem_cache_free(vm_area_cachep, vma);
1370 vm_unacct_memory(charged);
1374 /* Get an address range which is currently unmapped.
1375 * For shmat() with addr=0.
1377 * Ugly calling convention alert:
1378 * Return value with the low bits set means error value,
1380 * if (ret & ~PAGE_MASK)
1383 * This function "knows" that -ENOMEM has the bits set.
1385 #ifndef HAVE_ARCH_UNMAPPED_AREA
1387 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1388 unsigned long len, unsigned long pgoff, unsigned long flags)
1390 struct mm_struct *mm = current->mm;
1391 struct vm_area_struct *vma;
1392 unsigned long start_addr;
1394 if (len > TASK_SIZE)
1397 if (flags & MAP_FIXED)
1401 addr = PAGE_ALIGN(addr);
1402 vma = find_vma(mm, addr);
1403 if (TASK_SIZE - len >= addr &&
1404 (!vma || addr + len <= vma->vm_start))
1407 if (len > mm->cached_hole_size) {
1408 start_addr = addr = mm->free_area_cache;
1410 start_addr = addr = TASK_UNMAPPED_BASE;
1411 mm->cached_hole_size = 0;
1415 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1416 /* At this point: (!vma || addr < vma->vm_end). */
1417 if (TASK_SIZE - len < addr) {
1419 * Start a new search - just in case we missed
1422 if (start_addr != TASK_UNMAPPED_BASE) {
1423 addr = TASK_UNMAPPED_BASE;
1425 mm->cached_hole_size = 0;
1430 if (!vma || addr + len <= vma->vm_start) {
1432 * Remember the place where we stopped the search:
1434 mm->free_area_cache = addr + len;
1437 if (addr + mm->cached_hole_size < vma->vm_start)
1438 mm->cached_hole_size = vma->vm_start - addr;
1444 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1447 * Is this a new hole at the lowest possible address?
1449 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache)
1450 mm->free_area_cache = addr;
1454 * This mmap-allocator allocates new areas top-down from below the
1455 * stack's low limit (the base):
1457 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1459 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1460 const unsigned long len, const unsigned long pgoff,
1461 const unsigned long flags)
1463 struct vm_area_struct *vma;
1464 struct mm_struct *mm = current->mm;
1465 unsigned long addr = addr0, start_addr;
1467 /* requested length too big for entire address space */
1468 if (len > TASK_SIZE)
1471 if (flags & MAP_FIXED)
1474 /* requesting a specific address */
1476 addr = PAGE_ALIGN(addr);
1477 vma = find_vma(mm, addr);
1478 if (TASK_SIZE - len >= addr &&
1479 (!vma || addr + len <= vma->vm_start))
1483 /* check if free_area_cache is useful for us */
1484 if (len <= mm->cached_hole_size) {
1485 mm->cached_hole_size = 0;
1486 mm->free_area_cache = mm->mmap_base;
1490 /* either no address requested or can't fit in requested address hole */
1491 start_addr = addr = mm->free_area_cache;
1499 * Lookup failure means no vma is above this address,
1500 * else if new region fits below vma->vm_start,
1501 * return with success:
1503 vma = find_vma(mm, addr);
1504 if (!vma || addr+len <= vma->vm_start)
1505 /* remember the address as a hint for next time */
1506 return (mm->free_area_cache = addr);
1508 /* remember the largest hole we saw so far */
1509 if (addr + mm->cached_hole_size < vma->vm_start)
1510 mm->cached_hole_size = vma->vm_start - addr;
1512 /* try just below the current vma->vm_start */
1513 addr = vma->vm_start-len;
1514 } while (len < vma->vm_start);
1518 * if hint left us with no space for the requested
1519 * mapping then try again:
1521 * Note: this is different with the case of bottomup
1522 * which does the fully line-search, but we use find_vma
1523 * here that causes some holes skipped.
1525 if (start_addr != mm->mmap_base) {
1526 mm->free_area_cache = mm->mmap_base;
1527 mm->cached_hole_size = 0;
1532 * A failed mmap() very likely causes application failure,
1533 * so fall back to the bottom-up function here. This scenario
1534 * can happen with large stack limits and large mmap()
1537 mm->cached_hole_size = ~0UL;
1538 mm->free_area_cache = TASK_UNMAPPED_BASE;
1539 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1541 * Restore the topdown base:
1543 mm->free_area_cache = mm->mmap_base;
1544 mm->cached_hole_size = ~0UL;
1550 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1553 * Is this a new hole at the highest possible address?
1555 if (addr > mm->free_area_cache)
1556 mm->free_area_cache = addr;
1558 /* dont allow allocations above current base */
1559 if (mm->free_area_cache > mm->mmap_base)
1560 mm->free_area_cache = mm->mmap_base;
1564 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1565 unsigned long pgoff, unsigned long flags)
1567 unsigned long (*get_area)(struct file *, unsigned long,
1568 unsigned long, unsigned long, unsigned long);
1570 unsigned long error = arch_mmap_check(addr, len, flags);
1574 /* Careful about overflows.. */
1575 if (len > TASK_SIZE)
1578 get_area = current->mm->get_unmapped_area;
1579 if (file && file->f_op && file->f_op->get_unmapped_area)
1580 get_area = file->f_op->get_unmapped_area;
1581 addr = get_area(file, addr, len, pgoff, flags);
1582 if (IS_ERR_VALUE(addr))
1585 if (addr > TASK_SIZE - len)
1587 if (addr & ~PAGE_MASK)
1590 addr = arch_rebalance_pgtables(addr, len);
1591 error = security_mmap_addr(addr);
1592 return error ? error : addr;
1595 EXPORT_SYMBOL(get_unmapped_area);
1597 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1598 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1600 struct vm_area_struct *vma = NULL;
1602 if (WARN_ON_ONCE(!mm)) /* Remove this in linux-3.6 */
1605 /* Check the cache first. */
1606 /* (Cache hit rate is typically around 35%.) */
1607 vma = mm->mmap_cache;
1608 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1609 struct rb_node *rb_node;
1611 rb_node = mm->mm_rb.rb_node;
1615 struct vm_area_struct *vma_tmp;
1617 vma_tmp = rb_entry(rb_node,
1618 struct vm_area_struct, vm_rb);
1620 if (vma_tmp->vm_end > addr) {
1622 if (vma_tmp->vm_start <= addr)
1624 rb_node = rb_node->rb_left;
1626 rb_node = rb_node->rb_right;
1629 mm->mmap_cache = vma;
1634 EXPORT_SYMBOL(find_vma);
1637 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1639 struct vm_area_struct *
1640 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1641 struct vm_area_struct **pprev)
1643 struct vm_area_struct *vma;
1645 vma = find_vma(mm, addr);
1647 *pprev = vma->vm_prev;
1649 struct rb_node *rb_node = mm->mm_rb.rb_node;
1652 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1653 rb_node = rb_node->rb_right;
1660 * Verify that the stack growth is acceptable and
1661 * update accounting. This is shared with both the
1662 * grow-up and grow-down cases.
1664 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1666 struct mm_struct *mm = vma->vm_mm;
1667 struct rlimit *rlim = current->signal->rlim;
1668 unsigned long new_start;
1670 /* address space limit tests */
1671 if (!may_expand_vm(mm, grow))
1674 /* Stack limit test */
1675 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1678 /* mlock limit tests */
1679 if (vma->vm_flags & VM_LOCKED) {
1680 unsigned long locked;
1681 unsigned long limit;
1682 locked = mm->locked_vm + grow;
1683 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1684 limit >>= PAGE_SHIFT;
1685 if (locked > limit && !capable(CAP_IPC_LOCK))
1689 /* Check to ensure the stack will not grow into a hugetlb-only region */
1690 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1692 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1696 * Overcommit.. This must be the final test, as it will
1697 * update security statistics.
1699 if (security_vm_enough_memory_mm(mm, grow))
1702 /* Ok, everything looks good - let it rip */
1703 if (vma->vm_flags & VM_LOCKED)
1704 mm->locked_vm += grow;
1705 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1709 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1711 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1712 * vma is the last one with address > vma->vm_end. Have to extend vma.
1714 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1718 if (!(vma->vm_flags & VM_GROWSUP))
1722 * We must make sure the anon_vma is allocated
1723 * so that the anon_vma locking is not a noop.
1725 if (unlikely(anon_vma_prepare(vma)))
1727 vma_lock_anon_vma(vma);
1730 * vma->vm_start/vm_end cannot change under us because the caller
1731 * is required to hold the mmap_sem in read mode. We need the
1732 * anon_vma lock to serialize against concurrent expand_stacks.
1733 * Also guard against wrapping around to address 0.
1735 if (address < PAGE_ALIGN(address+4))
1736 address = PAGE_ALIGN(address+4);
1738 vma_unlock_anon_vma(vma);
1743 /* Somebody else might have raced and expanded it already */
1744 if (address > vma->vm_end) {
1745 unsigned long size, grow;
1747 size = address - vma->vm_start;
1748 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1751 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1752 error = acct_stack_growth(vma, size, grow);
1754 vma->vm_end = address;
1755 perf_event_mmap(vma);
1759 vma_unlock_anon_vma(vma);
1760 khugepaged_enter_vma_merge(vma);
1763 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1766 * vma is the first one with address < vma->vm_start. Have to extend vma.
1768 int expand_downwards(struct vm_area_struct *vma,
1769 unsigned long address)
1774 * We must make sure the anon_vma is allocated
1775 * so that the anon_vma locking is not a noop.
1777 if (unlikely(anon_vma_prepare(vma)))
1780 address &= PAGE_MASK;
1781 error = security_mmap_addr(address);
1785 vma_lock_anon_vma(vma);
1788 * vma->vm_start/vm_end cannot change under us because the caller
1789 * is required to hold the mmap_sem in read mode. We need the
1790 * anon_vma lock to serialize against concurrent expand_stacks.
1793 /* Somebody else might have raced and expanded it already */
1794 if (address < vma->vm_start) {
1795 unsigned long size, grow;
1797 size = vma->vm_end - address;
1798 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1801 if (grow <= vma->vm_pgoff) {
1802 error = acct_stack_growth(vma, size, grow);
1804 vma->vm_start = address;
1805 vma->vm_pgoff -= grow;
1806 perf_event_mmap(vma);
1810 vma_unlock_anon_vma(vma);
1811 khugepaged_enter_vma_merge(vma);
1815 #ifdef CONFIG_STACK_GROWSUP
1816 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1818 return expand_upwards(vma, address);
1821 struct vm_area_struct *
1822 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1824 struct vm_area_struct *vma, *prev;
1827 vma = find_vma_prev(mm, addr, &prev);
1828 if (vma && (vma->vm_start <= addr))
1830 if (!prev || expand_stack(prev, addr))
1832 if (prev->vm_flags & VM_LOCKED) {
1833 mlock_vma_pages_range(prev, addr, prev->vm_end);
1838 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1840 return expand_downwards(vma, address);
1843 struct vm_area_struct *
1844 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1846 struct vm_area_struct * vma;
1847 unsigned long start;
1850 vma = find_vma(mm,addr);
1853 if (vma->vm_start <= addr)
1855 if (!(vma->vm_flags & VM_GROWSDOWN))
1857 start = vma->vm_start;
1858 if (expand_stack(vma, addr))
1860 if (vma->vm_flags & VM_LOCKED) {
1861 mlock_vma_pages_range(vma, addr, start);
1868 * Ok - we have the memory areas we should free on the vma list,
1869 * so release them, and do the vma updates.
1871 * Called with the mm semaphore held.
1873 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1875 unsigned long nr_accounted = 0;
1877 /* Update high watermark before we lower total_vm */
1878 update_hiwater_vm(mm);
1880 long nrpages = vma_pages(vma);
1882 if (vma->vm_flags & VM_ACCOUNT)
1883 nr_accounted += nrpages;
1884 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1885 vma = remove_vma(vma);
1887 vm_unacct_memory(nr_accounted);
1892 * Get rid of page table information in the indicated region.
1894 * Called with the mm semaphore held.
1896 static void unmap_region(struct mm_struct *mm,
1897 struct vm_area_struct *vma, struct vm_area_struct *prev,
1898 unsigned long start, unsigned long end)
1900 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1901 struct mmu_gather tlb;
1904 tlb_gather_mmu(&tlb, mm, 0);
1905 update_hiwater_rss(mm);
1906 unmap_vmas(&tlb, vma, start, end);
1907 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
1908 next ? next->vm_start : 0);
1909 tlb_finish_mmu(&tlb, start, end);
1913 * Create a list of vma's touched by the unmap, removing them from the mm's
1914 * vma list as we go..
1917 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1918 struct vm_area_struct *prev, unsigned long end)
1920 struct vm_area_struct **insertion_point;
1921 struct vm_area_struct *tail_vma = NULL;
1924 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1925 vma->vm_prev = NULL;
1927 rb_erase(&vma->vm_rb, &mm->mm_rb);
1931 } while (vma && vma->vm_start < end);
1932 *insertion_point = vma;
1934 vma->vm_prev = prev;
1935 tail_vma->vm_next = NULL;
1936 if (mm->unmap_area == arch_unmap_area)
1937 addr = prev ? prev->vm_end : mm->mmap_base;
1939 addr = vma ? vma->vm_start : mm->mmap_base;
1940 mm->unmap_area(mm, addr);
1941 mm->mmap_cache = NULL; /* Kill the cache. */
1945 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
1946 * munmap path where it doesn't make sense to fail.
1948 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1949 unsigned long addr, int new_below)
1951 struct mempolicy *pol;
1952 struct vm_area_struct *new;
1955 if (is_vm_hugetlb_page(vma) && (addr &
1956 ~(huge_page_mask(hstate_vma(vma)))))
1959 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1963 /* most fields are the same, copy all, and then fixup */
1966 INIT_LIST_HEAD(&new->anon_vma_chain);
1971 new->vm_start = addr;
1972 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1975 pol = mpol_dup(vma_policy(vma));
1980 vma_set_policy(new, pol);
1982 if (anon_vma_clone(new, vma))
1986 get_file(new->vm_file);
1988 if (new->vm_ops && new->vm_ops->open)
1989 new->vm_ops->open(new);
1992 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1993 ((addr - new->vm_start) >> PAGE_SHIFT), new);
1995 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2001 /* Clean everything up if vma_adjust failed. */
2002 if (new->vm_ops && new->vm_ops->close)
2003 new->vm_ops->close(new);
2006 unlink_anon_vmas(new);
2010 kmem_cache_free(vm_area_cachep, new);
2016 * Split a vma into two pieces at address 'addr', a new vma is allocated
2017 * either for the first part or the tail.
2019 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2020 unsigned long addr, int new_below)
2022 if (mm->map_count >= sysctl_max_map_count)
2025 return __split_vma(mm, vma, addr, new_below);
2028 /* Munmap is split into 2 main parts -- this part which finds
2029 * what needs doing, and the areas themselves, which do the
2030 * work. This now handles partial unmappings.
2031 * Jeremy Fitzhardinge <jeremy@goop.org>
2033 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2036 struct vm_area_struct *vma, *prev, *last;
2038 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2041 if ((len = PAGE_ALIGN(len)) == 0)
2044 /* Find the first overlapping VMA */
2045 vma = find_vma(mm, start);
2048 prev = vma->vm_prev;
2049 /* we have start < vma->vm_end */
2051 /* if it doesn't overlap, we have nothing.. */
2053 if (vma->vm_start >= end)
2057 * If we need to split any vma, do it now to save pain later.
2059 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2060 * unmapped vm_area_struct will remain in use: so lower split_vma
2061 * places tmp vma above, and higher split_vma places tmp vma below.
2063 if (start > vma->vm_start) {
2067 * Make sure that map_count on return from munmap() will
2068 * not exceed its limit; but let map_count go just above
2069 * its limit temporarily, to help free resources as expected.
2071 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2074 error = __split_vma(mm, vma, start, 0);
2080 /* Does it split the last one? */
2081 last = find_vma(mm, end);
2082 if (last && end > last->vm_start) {
2083 int error = __split_vma(mm, last, end, 1);
2087 vma = prev? prev->vm_next: mm->mmap;
2090 * unlock any mlock()ed ranges before detaching vmas
2092 if (mm->locked_vm) {
2093 struct vm_area_struct *tmp = vma;
2094 while (tmp && tmp->vm_start < end) {
2095 if (tmp->vm_flags & VM_LOCKED) {
2096 mm->locked_vm -= vma_pages(tmp);
2097 munlock_vma_pages_all(tmp);
2104 * Remove the vma's, and unmap the actual pages
2106 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2107 unmap_region(mm, vma, prev, start, end);
2109 /* Fix up all other VM information */
2110 remove_vma_list(mm, vma);
2115 int vm_munmap(unsigned long start, size_t len)
2118 struct mm_struct *mm = current->mm;
2120 down_write(&mm->mmap_sem);
2121 ret = do_munmap(mm, start, len);
2122 up_write(&mm->mmap_sem);
2125 EXPORT_SYMBOL(vm_munmap);
2127 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2129 profile_munmap(addr);
2130 return vm_munmap(addr, len);
2133 static inline void verify_mm_writelocked(struct mm_struct *mm)
2135 #ifdef CONFIG_DEBUG_VM
2136 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2138 up_read(&mm->mmap_sem);
2144 * this is really a simplified "do_mmap". it only handles
2145 * anonymous maps. eventually we may be able to do some
2146 * brk-specific accounting here.
2148 static unsigned long do_brk(unsigned long addr, unsigned long len)
2150 struct mm_struct * mm = current->mm;
2151 struct vm_area_struct * vma, * prev;
2152 unsigned long flags;
2153 struct rb_node ** rb_link, * rb_parent;
2154 pgoff_t pgoff = addr >> PAGE_SHIFT;
2157 len = PAGE_ALIGN(len);
2161 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2163 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2164 if (error & ~PAGE_MASK)
2170 if (mm->def_flags & VM_LOCKED) {
2171 unsigned long locked, lock_limit;
2172 locked = len >> PAGE_SHIFT;
2173 locked += mm->locked_vm;
2174 lock_limit = rlimit(RLIMIT_MEMLOCK);
2175 lock_limit >>= PAGE_SHIFT;
2176 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2181 * mm->mmap_sem is required to protect against another thread
2182 * changing the mappings in case we sleep.
2184 verify_mm_writelocked(mm);
2187 * Clear old maps. this also does some error checking for us
2190 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2191 if (vma && vma->vm_start < addr + len) {
2192 if (do_munmap(mm, addr, len))
2197 /* Check against address space limits *after* clearing old maps... */
2198 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2201 if (mm->map_count > sysctl_max_map_count)
2204 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2207 /* Can we just expand an old private anonymous mapping? */
2208 vma = vma_merge(mm, prev, addr, addr + len, flags,
2209 NULL, NULL, pgoff, NULL);
2214 * create a vma struct for an anonymous mapping
2216 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2218 vm_unacct_memory(len >> PAGE_SHIFT);
2222 INIT_LIST_HEAD(&vma->anon_vma_chain);
2224 vma->vm_start = addr;
2225 vma->vm_end = addr + len;
2226 vma->vm_pgoff = pgoff;
2227 vma->vm_flags = flags;
2228 vma->vm_page_prot = vm_get_page_prot(flags);
2229 vma_link(mm, vma, prev, rb_link, rb_parent);
2231 perf_event_mmap(vma);
2232 mm->total_vm += len >> PAGE_SHIFT;
2233 if (flags & VM_LOCKED) {
2234 if (!mlock_vma_pages_range(vma, addr, addr + len))
2235 mm->locked_vm += (len >> PAGE_SHIFT);
2240 unsigned long vm_brk(unsigned long addr, unsigned long len)
2242 struct mm_struct *mm = current->mm;
2245 down_write(&mm->mmap_sem);
2246 ret = do_brk(addr, len);
2247 up_write(&mm->mmap_sem);
2250 EXPORT_SYMBOL(vm_brk);
2252 /* Release all mmaps. */
2253 void exit_mmap(struct mm_struct *mm)
2255 struct mmu_gather tlb;
2256 struct vm_area_struct *vma;
2257 unsigned long nr_accounted = 0;
2259 /* mm's last user has gone, and its about to be pulled down */
2260 mmu_notifier_release(mm);
2262 if (mm->locked_vm) {
2265 if (vma->vm_flags & VM_LOCKED)
2266 munlock_vma_pages_all(vma);
2274 if (!vma) /* Can happen if dup_mmap() received an OOM */
2279 tlb_gather_mmu(&tlb, mm, 1);
2280 /* update_hiwater_rss(mm) here? but nobody should be looking */
2281 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2282 unmap_vmas(&tlb, vma, 0, -1);
2284 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2285 tlb_finish_mmu(&tlb, 0, -1);
2288 * Walk the list again, actually closing and freeing it,
2289 * with preemption enabled, without holding any MM locks.
2292 if (vma->vm_flags & VM_ACCOUNT)
2293 nr_accounted += vma_pages(vma);
2294 vma = remove_vma(vma);
2296 vm_unacct_memory(nr_accounted);
2298 WARN_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2301 /* Insert vm structure into process list sorted by address
2302 * and into the inode's i_mmap tree. If vm_file is non-NULL
2303 * then i_mmap_mutex is taken here.
2305 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2307 struct vm_area_struct * __vma, * prev;
2308 struct rb_node ** rb_link, * rb_parent;
2311 * The vm_pgoff of a purely anonymous vma should be irrelevant
2312 * until its first write fault, when page's anon_vma and index
2313 * are set. But now set the vm_pgoff it will almost certainly
2314 * end up with (unless mremap moves it elsewhere before that
2315 * first wfault), so /proc/pid/maps tells a consistent story.
2317 * By setting it to reflect the virtual start address of the
2318 * vma, merges and splits can happen in a seamless way, just
2319 * using the existing file pgoff checks and manipulations.
2320 * Similarly in do_mmap_pgoff and in do_brk.
2322 if (!vma->vm_file) {
2323 BUG_ON(vma->anon_vma);
2324 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2326 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2327 if (__vma && __vma->vm_start < vma->vm_end)
2329 if ((vma->vm_flags & VM_ACCOUNT) &&
2330 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2333 vma_link(mm, vma, prev, rb_link, rb_parent);
2338 * Copy the vma structure to a new location in the same mm,
2339 * prior to moving page table entries, to effect an mremap move.
2341 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2342 unsigned long addr, unsigned long len, pgoff_t pgoff)
2344 struct vm_area_struct *vma = *vmap;
2345 unsigned long vma_start = vma->vm_start;
2346 struct mm_struct *mm = vma->vm_mm;
2347 struct vm_area_struct *new_vma, *prev;
2348 struct rb_node **rb_link, *rb_parent;
2349 struct mempolicy *pol;
2350 bool faulted_in_anon_vma = true;
2353 * If anonymous vma has not yet been faulted, update new pgoff
2354 * to match new location, to increase its chance of merging.
2356 if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2357 pgoff = addr >> PAGE_SHIFT;
2358 faulted_in_anon_vma = false;
2361 find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2362 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2363 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2366 * Source vma may have been merged into new_vma
2368 if (unlikely(vma_start >= new_vma->vm_start &&
2369 vma_start < new_vma->vm_end)) {
2371 * The only way we can get a vma_merge with
2372 * self during an mremap is if the vma hasn't
2373 * been faulted in yet and we were allowed to
2374 * reset the dst vma->vm_pgoff to the
2375 * destination address of the mremap to allow
2376 * the merge to happen. mremap must change the
2377 * vm_pgoff linearity between src and dst vmas
2378 * (in turn preventing a vma_merge) to be
2379 * safe. It is only safe to keep the vm_pgoff
2380 * linear if there are no pages mapped yet.
2382 VM_BUG_ON(faulted_in_anon_vma);
2385 anon_vma_moveto_tail(new_vma);
2387 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2390 pol = mpol_dup(vma_policy(vma));
2393 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2394 if (anon_vma_clone(new_vma, vma))
2395 goto out_free_mempol;
2396 vma_set_policy(new_vma, pol);
2397 new_vma->vm_start = addr;
2398 new_vma->vm_end = addr + len;
2399 new_vma->vm_pgoff = pgoff;
2400 if (new_vma->vm_file)
2401 get_file(new_vma->vm_file);
2402 if (new_vma->vm_ops && new_vma->vm_ops->open)
2403 new_vma->vm_ops->open(new_vma);
2404 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2412 kmem_cache_free(vm_area_cachep, new_vma);
2417 * Return true if the calling process may expand its vm space by the passed
2420 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2422 unsigned long cur = mm->total_vm; /* pages */
2425 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2427 if (cur + npages > lim)
2433 static int special_mapping_fault(struct vm_area_struct *vma,
2434 struct vm_fault *vmf)
2437 struct page **pages;
2440 * special mappings have no vm_file, and in that case, the mm
2441 * uses vm_pgoff internally. So we have to subtract it from here.
2442 * We are allowed to do this because we are the mm; do not copy
2443 * this code into drivers!
2445 pgoff = vmf->pgoff - vma->vm_pgoff;
2447 for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2451 struct page *page = *pages;
2457 return VM_FAULT_SIGBUS;
2461 * Having a close hook prevents vma merging regardless of flags.
2463 static void special_mapping_close(struct vm_area_struct *vma)
2467 static const struct vm_operations_struct special_mapping_vmops = {
2468 .close = special_mapping_close,
2469 .fault = special_mapping_fault,
2473 * Called with mm->mmap_sem held for writing.
2474 * Insert a new vma covering the given region, with the given flags.
2475 * Its pages are supplied by the given array of struct page *.
2476 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2477 * The region past the last page supplied will always produce SIGBUS.
2478 * The array pointer and the pages it points to are assumed to stay alive
2479 * for as long as this mapping might exist.
2481 int install_special_mapping(struct mm_struct *mm,
2482 unsigned long addr, unsigned long len,
2483 unsigned long vm_flags, struct page **pages)
2486 struct vm_area_struct *vma;
2488 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2489 if (unlikely(vma == NULL))
2492 INIT_LIST_HEAD(&vma->anon_vma_chain);
2494 vma->vm_start = addr;
2495 vma->vm_end = addr + len;
2497 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2498 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2500 vma->vm_ops = &special_mapping_vmops;
2501 vma->vm_private_data = pages;
2503 ret = insert_vm_struct(mm, vma);
2507 mm->total_vm += len >> PAGE_SHIFT;
2509 perf_event_mmap(vma);
2514 kmem_cache_free(vm_area_cachep, vma);
2518 static DEFINE_MUTEX(mm_all_locks_mutex);
2520 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2522 if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2524 * The LSB of head.next can't change from under us
2525 * because we hold the mm_all_locks_mutex.
2527 mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem);
2529 * We can safely modify head.next after taking the
2530 * anon_vma->root->mutex. If some other vma in this mm shares
2531 * the same anon_vma we won't take it again.
2533 * No need of atomic instructions here, head.next
2534 * can't change from under us thanks to the
2535 * anon_vma->root->mutex.
2537 if (__test_and_set_bit(0, (unsigned long *)
2538 &anon_vma->root->head.next))
2543 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2545 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2547 * AS_MM_ALL_LOCKS can't change from under us because
2548 * we hold the mm_all_locks_mutex.
2550 * Operations on ->flags have to be atomic because
2551 * even if AS_MM_ALL_LOCKS is stable thanks to the
2552 * mm_all_locks_mutex, there may be other cpus
2553 * changing other bitflags in parallel to us.
2555 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2557 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2562 * This operation locks against the VM for all pte/vma/mm related
2563 * operations that could ever happen on a certain mm. This includes
2564 * vmtruncate, try_to_unmap, and all page faults.
2566 * The caller must take the mmap_sem in write mode before calling
2567 * mm_take_all_locks(). The caller isn't allowed to release the
2568 * mmap_sem until mm_drop_all_locks() returns.
2570 * mmap_sem in write mode is required in order to block all operations
2571 * that could modify pagetables and free pages without need of
2572 * altering the vma layout (for example populate_range() with
2573 * nonlinear vmas). It's also needed in write mode to avoid new
2574 * anon_vmas to be associated with existing vmas.
2576 * A single task can't take more than one mm_take_all_locks() in a row
2577 * or it would deadlock.
2579 * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2580 * mapping->flags avoid to take the same lock twice, if more than one
2581 * vma in this mm is backed by the same anon_vma or address_space.
2583 * We can take all the locks in random order because the VM code
2584 * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2585 * takes more than one of them in a row. Secondly we're protected
2586 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2588 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2589 * that may have to take thousand of locks.
2591 * mm_take_all_locks() can fail if it's interrupted by signals.
2593 int mm_take_all_locks(struct mm_struct *mm)
2595 struct vm_area_struct *vma;
2596 struct anon_vma_chain *avc;
2598 BUG_ON(down_read_trylock(&mm->mmap_sem));
2600 mutex_lock(&mm_all_locks_mutex);
2602 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2603 if (signal_pending(current))
2605 if (vma->vm_file && vma->vm_file->f_mapping)
2606 vm_lock_mapping(mm, vma->vm_file->f_mapping);
2609 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2610 if (signal_pending(current))
2613 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2614 vm_lock_anon_vma(mm, avc->anon_vma);
2620 mm_drop_all_locks(mm);
2624 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2626 if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2628 * The LSB of head.next can't change to 0 from under
2629 * us because we hold the mm_all_locks_mutex.
2631 * We must however clear the bitflag before unlocking
2632 * the vma so the users using the anon_vma->head will
2633 * never see our bitflag.
2635 * No need of atomic instructions here, head.next
2636 * can't change from under us until we release the
2637 * anon_vma->root->mutex.
2639 if (!__test_and_clear_bit(0, (unsigned long *)
2640 &anon_vma->root->head.next))
2642 anon_vma_unlock(anon_vma);
2646 static void vm_unlock_mapping(struct address_space *mapping)
2648 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2650 * AS_MM_ALL_LOCKS can't change to 0 from under us
2651 * because we hold the mm_all_locks_mutex.
2653 mutex_unlock(&mapping->i_mmap_mutex);
2654 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2661 * The mmap_sem cannot be released by the caller until
2662 * mm_drop_all_locks() returns.
2664 void mm_drop_all_locks(struct mm_struct *mm)
2666 struct vm_area_struct *vma;
2667 struct anon_vma_chain *avc;
2669 BUG_ON(down_read_trylock(&mm->mmap_sem));
2670 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2672 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2674 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2675 vm_unlock_anon_vma(avc->anon_vma);
2676 if (vma->vm_file && vma->vm_file->f_mapping)
2677 vm_unlock_mapping(vma->vm_file->f_mapping);
2680 mutex_unlock(&mm_all_locks_mutex);
2684 * initialise the VMA slab
2686 void __init mmap_init(void)
2690 ret = percpu_counter_init(&vm_committed_as, 0);