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/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>
34 #include <asm/uaccess.h>
35 #include <asm/cacheflush.h>
37 #include <asm/mmu_context.h>
41 #ifndef arch_mmap_check
42 #define arch_mmap_check(addr, len, flags) (0)
45 #ifndef arch_rebalance_pgtables
46 #define arch_rebalance_pgtables(addr, len) (addr)
49 static 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);
54 * WARNING: the debugging will use recursive algorithms so never enable this
55 * unless you know what you are doing.
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:
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
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
74 pgprot_t protection_map[16] = {
75 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
76 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
79 pgprot_t vm_get_page_prot(unsigned long vm_flags)
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)));
85 EXPORT_SYMBOL(vm_get_page_prot);
87 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS; /* heuristic overcommit */
88 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
89 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
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.
94 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
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.
101 * We currently support three overcommit policies, which are set via the
102 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
104 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
105 * Additional code 2002 Jul 20 by Robert Love.
107 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
109 * Note this is a helper function intended to be used by LSMs which
110 * wish to use this logic.
112 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
114 unsigned long free, allowed;
116 vm_acct_memory(pages);
119 * Sometimes we want to use more memory than we have
121 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
124 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
127 free = global_page_state(NR_FILE_PAGES);
128 free += nr_swap_pages;
131 * Any slabs which are created with the
132 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
133 * which are reclaimable, under pressure. The dentry
134 * cache and most inode caches should fall into this
136 free += global_page_state(NR_SLAB_RECLAIMABLE);
139 * Leave the last 3% for root
148 * nr_free_pages() is very expensive on large systems,
149 * only call if we're about to fail.
154 * Leave reserved pages. The pages are not for anonymous pages.
156 if (n <= totalreserve_pages)
159 n -= totalreserve_pages;
162 * Leave the last 3% for root
174 allowed = (totalram_pages - hugetlb_total_pages())
175 * sysctl_overcommit_ratio / 100;
177 * Leave the last 3% for root
180 allowed -= allowed / 32;
181 allowed += total_swap_pages;
183 /* Don't let a single process grow too big:
184 leave 3% of the size of this process for other processes */
186 allowed -= mm->total_vm / 32;
188 if (percpu_counter_read_positive(&vm_committed_as) < allowed)
191 vm_unacct_memory(pages);
197 * Requires inode->i_mapping->i_mmap_lock
199 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
200 struct file *file, struct address_space *mapping)
202 if (vma->vm_flags & VM_DENYWRITE)
203 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
204 if (vma->vm_flags & VM_SHARED)
205 mapping->i_mmap_writable--;
207 flush_dcache_mmap_lock(mapping);
208 if (unlikely(vma->vm_flags & VM_NONLINEAR))
209 list_del_init(&vma->shared.vm_set.list);
211 vma_prio_tree_remove(vma, &mapping->i_mmap);
212 flush_dcache_mmap_unlock(mapping);
216 * Unlink a file-based vm structure from its prio_tree, to hide
217 * vma from rmap and vmtruncate before freeing its page tables.
219 void unlink_file_vma(struct vm_area_struct *vma)
221 struct file *file = vma->vm_file;
224 struct address_space *mapping = file->f_mapping;
225 spin_lock(&mapping->i_mmap_lock);
226 __remove_shared_vm_struct(vma, file, mapping);
227 spin_unlock(&mapping->i_mmap_lock);
232 * Close a vm structure and free it, returning the next.
234 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
236 struct vm_area_struct *next = vma->vm_next;
239 if (vma->vm_ops && vma->vm_ops->close)
240 vma->vm_ops->close(vma);
243 if (vma->vm_flags & VM_EXECUTABLE)
244 removed_exe_file_vma(vma->vm_mm);
246 mpol_put(vma_policy(vma));
247 kmem_cache_free(vm_area_cachep, vma);
251 SYSCALL_DEFINE1(brk, unsigned long, brk)
253 unsigned long rlim, retval;
254 unsigned long newbrk, oldbrk;
255 struct mm_struct *mm = current->mm;
256 unsigned long min_brk;
258 down_write(&mm->mmap_sem);
260 #ifdef CONFIG_COMPAT_BRK
262 * CONFIG_COMPAT_BRK can still be overridden by setting
263 * randomize_va_space to 2, which will still cause mm->start_brk
264 * to be arbitrarily shifted
266 if (current->brk_randomized)
267 min_brk = mm->start_brk;
269 min_brk = mm->end_data;
271 min_brk = mm->start_brk;
277 * Check against rlimit here. If this check is done later after the test
278 * of oldbrk with newbrk then it can escape the test and let the data
279 * segment grow beyond its set limit the in case where the limit is
280 * not page aligned -Ram Gupta
282 rlim = rlimit(RLIMIT_DATA);
283 if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
284 (mm->end_data - mm->start_data) > rlim)
287 newbrk = PAGE_ALIGN(brk);
288 oldbrk = PAGE_ALIGN(mm->brk);
289 if (oldbrk == newbrk)
292 /* Always allow shrinking brk. */
293 if (brk <= mm->brk) {
294 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
299 /* Check against existing mmap mappings. */
300 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
303 /* Ok, looks good - let it rip. */
304 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
310 up_write(&mm->mmap_sem);
315 static int browse_rb(struct rb_root *root)
318 struct rb_node *nd, *pn = NULL;
319 unsigned long prev = 0, pend = 0;
321 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
322 struct vm_area_struct *vma;
323 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
324 if (vma->vm_start < prev)
325 printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
326 if (vma->vm_start < pend)
327 printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
328 if (vma->vm_start > vma->vm_end)
329 printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
332 prev = vma->vm_start;
336 for (nd = pn; nd; nd = rb_prev(nd)) {
340 printk("backwards %d, forwards %d\n", j, i), i = 0;
344 void validate_mm(struct mm_struct *mm)
348 struct vm_area_struct *tmp = mm->mmap;
353 if (i != mm->map_count)
354 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
355 i = browse_rb(&mm->mm_rb);
356 if (i != mm->map_count)
357 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
361 #define validate_mm(mm) do { } while (0)
364 static struct vm_area_struct *
365 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
366 struct vm_area_struct **pprev, struct rb_node ***rb_link,
367 struct rb_node ** rb_parent)
369 struct vm_area_struct * vma;
370 struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
372 __rb_link = &mm->mm_rb.rb_node;
373 rb_prev = __rb_parent = NULL;
377 struct vm_area_struct *vma_tmp;
379 __rb_parent = *__rb_link;
380 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
382 if (vma_tmp->vm_end > addr) {
384 if (vma_tmp->vm_start <= addr)
386 __rb_link = &__rb_parent->rb_left;
388 rb_prev = __rb_parent;
389 __rb_link = &__rb_parent->rb_right;
395 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
396 *rb_link = __rb_link;
397 *rb_parent = __rb_parent;
402 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
403 struct vm_area_struct *prev, struct rb_node *rb_parent)
405 struct vm_area_struct *next;
409 next = prev->vm_next;
414 next = rb_entry(rb_parent,
415 struct vm_area_struct, vm_rb);
424 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
425 struct rb_node **rb_link, struct rb_node *rb_parent)
427 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
428 rb_insert_color(&vma->vm_rb, &mm->mm_rb);
431 static void __vma_link_file(struct vm_area_struct *vma)
437 struct address_space *mapping = file->f_mapping;
439 if (vma->vm_flags & VM_DENYWRITE)
440 atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
441 if (vma->vm_flags & VM_SHARED)
442 mapping->i_mmap_writable++;
444 flush_dcache_mmap_lock(mapping);
445 if (unlikely(vma->vm_flags & VM_NONLINEAR))
446 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
448 vma_prio_tree_insert(vma, &mapping->i_mmap);
449 flush_dcache_mmap_unlock(mapping);
454 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
455 struct vm_area_struct *prev, struct rb_node **rb_link,
456 struct rb_node *rb_parent)
458 __vma_link_list(mm, vma, prev, rb_parent);
459 __vma_link_rb(mm, vma, rb_link, rb_parent);
462 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
463 struct vm_area_struct *prev, struct rb_node **rb_link,
464 struct rb_node *rb_parent)
466 struct address_space *mapping = NULL;
469 mapping = vma->vm_file->f_mapping;
472 spin_lock(&mapping->i_mmap_lock);
473 vma->vm_truncate_count = mapping->truncate_count;
476 __vma_link(mm, vma, prev, rb_link, rb_parent);
477 __vma_link_file(vma);
480 spin_unlock(&mapping->i_mmap_lock);
487 * Helper for vma_adjust in the split_vma insert case:
488 * insert vm structure into list and rbtree and anon_vma,
489 * but it has already been inserted into prio_tree earlier.
491 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
493 struct vm_area_struct *__vma, *prev;
494 struct rb_node **rb_link, *rb_parent;
496 __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
497 BUG_ON(__vma && __vma->vm_start < vma->vm_end);
498 __vma_link(mm, vma, prev, rb_link, rb_parent);
503 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
504 struct vm_area_struct *prev)
506 struct vm_area_struct *next = vma->vm_next;
508 prev->vm_next = next;
510 next->vm_prev = prev;
511 rb_erase(&vma->vm_rb, &mm->mm_rb);
512 if (mm->mmap_cache == vma)
513 mm->mmap_cache = prev;
517 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
518 * is already present in an i_mmap tree without adjusting the tree.
519 * The following helper function should be used when such adjustments
520 * are necessary. The "insert" vma (if any) is to be inserted
521 * before we drop the necessary locks.
523 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
524 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
526 struct mm_struct *mm = vma->vm_mm;
527 struct vm_area_struct *next = vma->vm_next;
528 struct vm_area_struct *importer = NULL;
529 struct address_space *mapping = NULL;
530 struct prio_tree_root *root = NULL;
531 struct anon_vma *anon_vma = NULL;
532 struct file *file = vma->vm_file;
533 long adjust_next = 0;
536 if (next && !insert) {
537 struct vm_area_struct *exporter = NULL;
539 if (end >= next->vm_end) {
541 * vma expands, overlapping all the next, and
542 * perhaps the one after too (mprotect case 6).
544 again: remove_next = 1 + (end > next->vm_end);
548 } else if (end > next->vm_start) {
550 * vma expands, overlapping part of the next:
551 * mprotect case 5 shifting the boundary up.
553 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
556 } else if (end < vma->vm_end) {
558 * vma shrinks, and !insert tells it's not
559 * split_vma inserting another: so it must be
560 * mprotect case 4 shifting the boundary down.
562 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
568 * Easily overlooked: when mprotect shifts the boundary,
569 * make sure the expanding vma has anon_vma set if the
570 * shrinking vma had, to cover any anon pages imported.
572 if (exporter && exporter->anon_vma && !importer->anon_vma) {
573 if (anon_vma_clone(importer, exporter))
575 importer->anon_vma = exporter->anon_vma;
580 mapping = file->f_mapping;
581 if (!(vma->vm_flags & VM_NONLINEAR))
582 root = &mapping->i_mmap;
583 spin_lock(&mapping->i_mmap_lock);
585 vma->vm_truncate_count != next->vm_truncate_count) {
587 * unmap_mapping_range might be in progress:
588 * ensure that the expanding vma is rescanned.
590 importer->vm_truncate_count = 0;
593 insert->vm_truncate_count = vma->vm_truncate_count;
595 * Put into prio_tree now, so instantiated pages
596 * are visible to arm/parisc __flush_dcache_page
597 * throughout; but we cannot insert into address
598 * space until vma start or end is updated.
600 __vma_link_file(insert);
604 vma_adjust_trans_huge(vma, start, end, adjust_next);
607 * When changing only vma->vm_end, we don't really need anon_vma
608 * lock. This is a fairly rare case by itself, but the anon_vma
609 * lock may be shared between many sibling processes. Skipping
610 * the lock for brk adjustments makes a difference sometimes.
612 if (vma->anon_vma && (importer || start != vma->vm_start)) {
613 anon_vma = vma->anon_vma;
614 anon_vma_lock(anon_vma);
618 flush_dcache_mmap_lock(mapping);
619 vma_prio_tree_remove(vma, root);
621 vma_prio_tree_remove(next, root);
624 vma->vm_start = start;
626 vma->vm_pgoff = pgoff;
628 next->vm_start += adjust_next << PAGE_SHIFT;
629 next->vm_pgoff += adjust_next;
634 vma_prio_tree_insert(next, root);
635 vma_prio_tree_insert(vma, root);
636 flush_dcache_mmap_unlock(mapping);
641 * vma_merge has merged next into vma, and needs
642 * us to remove next before dropping the locks.
644 __vma_unlink(mm, next, vma);
646 __remove_shared_vm_struct(next, file, mapping);
649 * split_vma has split insert from vma, and needs
650 * us to insert it before dropping the locks
651 * (it may either follow vma or precede it).
653 __insert_vm_struct(mm, insert);
657 anon_vma_unlock(anon_vma);
659 spin_unlock(&mapping->i_mmap_lock);
664 if (next->vm_flags & VM_EXECUTABLE)
665 removed_exe_file_vma(mm);
668 anon_vma_merge(vma, next);
670 mpol_put(vma_policy(next));
671 kmem_cache_free(vm_area_cachep, next);
673 * In mprotect's case 6 (see comments on vma_merge),
674 * we must remove another next too. It would clutter
675 * up the code too much to do both in one go.
677 if (remove_next == 2) {
689 * If the vma has a ->close operation then the driver probably needs to release
690 * per-vma resources, so we don't attempt to merge those.
692 static inline int is_mergeable_vma(struct vm_area_struct *vma,
693 struct file *file, unsigned long vm_flags)
695 /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
696 if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
698 if (vma->vm_file != file)
700 if (vma->vm_ops && vma->vm_ops->close)
705 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
706 struct anon_vma *anon_vma2,
707 struct vm_area_struct *vma)
710 * The list_is_singular() test is to avoid merging VMA cloned from
711 * parents. This can improve scalability caused by anon_vma lock.
713 if ((!anon_vma1 || !anon_vma2) && (!vma ||
714 list_is_singular(&vma->anon_vma_chain)))
716 return anon_vma1 == anon_vma2;
720 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
721 * in front of (at a lower virtual address and file offset than) the vma.
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.
726 * We don't check here for the merged mmap wrapping around the end of pagecache
727 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
728 * wrap, nor mmaps which cover the final page at index -1UL.
731 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
732 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
734 if (is_mergeable_vma(vma, file, vm_flags) &&
735 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
736 if (vma->vm_pgoff == vm_pgoff)
743 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
744 * beyond (at a higher virtual address and file offset than) the vma.
746 * We cannot merge two vmas if they have differently assigned (non-NULL)
747 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
750 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
751 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
753 if (is_mergeable_vma(vma, file, vm_flags) &&
754 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
756 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
757 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
764 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
765 * whether that can be merged with its predecessor or its successor.
766 * Or both (it neatly fills a hole).
768 * In most cases - when called for mmap, brk or mremap - [addr,end) is
769 * certain not to be mapped by the time vma_merge is called; but when
770 * called for mprotect, it is certain to be already mapped (either at
771 * an offset within prev, or at the start of next), and the flags of
772 * this area are about to be changed to vm_flags - and the no-change
773 * case has already been eliminated.
775 * The following mprotect cases have to be considered, where AAAA is
776 * the area passed down from mprotect_fixup, never extending beyond one
777 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
779 * AAAA AAAA AAAA AAAA
780 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
781 * cannot merge might become might become might become
782 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
783 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
784 * mremap move: PPPPNNNNNNNN 8
786 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
787 * might become case 1 below case 2 below case 3 below
789 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
790 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
792 struct vm_area_struct *vma_merge(struct mm_struct *mm,
793 struct vm_area_struct *prev, unsigned long addr,
794 unsigned long end, unsigned long vm_flags,
795 struct anon_vma *anon_vma, struct file *file,
796 pgoff_t pgoff, struct mempolicy *policy)
798 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
799 struct vm_area_struct *area, *next;
803 * We later require that vma->vm_flags == vm_flags,
804 * so this tests vma->vm_flags & VM_SPECIAL, too.
806 if (vm_flags & VM_SPECIAL)
810 next = prev->vm_next;
814 if (next && next->vm_end == end) /* cases 6, 7, 8 */
815 next = next->vm_next;
818 * Can it merge with the predecessor?
820 if (prev && prev->vm_end == addr &&
821 mpol_equal(vma_policy(prev), policy) &&
822 can_vma_merge_after(prev, vm_flags,
823 anon_vma, file, pgoff)) {
825 * OK, it can. Can we now merge in the successor as well?
827 if (next && end == next->vm_start &&
828 mpol_equal(policy, vma_policy(next)) &&
829 can_vma_merge_before(next, vm_flags,
830 anon_vma, file, pgoff+pglen) &&
831 is_mergeable_anon_vma(prev->anon_vma,
832 next->anon_vma, NULL)) {
834 err = vma_adjust(prev, prev->vm_start,
835 next->vm_end, prev->vm_pgoff, NULL);
836 } else /* cases 2, 5, 7 */
837 err = vma_adjust(prev, prev->vm_start,
838 end, prev->vm_pgoff, NULL);
841 khugepaged_enter_vma_merge(prev);
846 * Can this new request be merged in front of next?
848 if (next && end == next->vm_start &&
849 mpol_equal(policy, vma_policy(next)) &&
850 can_vma_merge_before(next, vm_flags,
851 anon_vma, file, pgoff+pglen)) {
852 if (prev && addr < prev->vm_end) /* case 4 */
853 err = vma_adjust(prev, prev->vm_start,
854 addr, prev->vm_pgoff, NULL);
855 else /* cases 3, 8 */
856 err = vma_adjust(area, addr, next->vm_end,
857 next->vm_pgoff - pglen, NULL);
860 khugepaged_enter_vma_merge(area);
868 * Rough compatbility check to quickly see if it's even worth looking
869 * at sharing an anon_vma.
871 * They need to have the same vm_file, and the flags can only differ
872 * in things that mprotect may change.
874 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
875 * we can merge the two vma's. For example, we refuse to merge a vma if
876 * there is a vm_ops->close() function, because that indicates that the
877 * driver is doing some kind of reference counting. But that doesn't
878 * really matter for the anon_vma sharing case.
880 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
882 return a->vm_end == b->vm_start &&
883 mpol_equal(vma_policy(a), vma_policy(b)) &&
884 a->vm_file == b->vm_file &&
885 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
886 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
890 * Do some basic sanity checking to see if we can re-use the anon_vma
891 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
892 * the same as 'old', the other will be the new one that is trying
893 * to share the anon_vma.
895 * NOTE! This runs with mm_sem held for reading, so it is possible that
896 * the anon_vma of 'old' is concurrently in the process of being set up
897 * by another page fault trying to merge _that_. But that's ok: if it
898 * is being set up, that automatically means that it will be a singleton
899 * acceptable for merging, so we can do all of this optimistically. But
900 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
902 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
903 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
904 * is to return an anon_vma that is "complex" due to having gone through
907 * We also make sure that the two vma's are compatible (adjacent,
908 * and with the same memory policies). That's all stable, even with just
909 * a read lock on the mm_sem.
911 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
913 if (anon_vma_compatible(a, b)) {
914 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
916 if (anon_vma && list_is_singular(&old->anon_vma_chain))
923 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
924 * neighbouring vmas for a suitable anon_vma, before it goes off
925 * to allocate a new anon_vma. It checks because a repetitive
926 * sequence of mprotects and faults may otherwise lead to distinct
927 * anon_vmas being allocated, preventing vma merge in subsequent
930 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
932 struct anon_vma *anon_vma;
933 struct vm_area_struct *near;
939 anon_vma = reusable_anon_vma(near, vma, near);
944 * It is potentially slow to have to call find_vma_prev here.
945 * But it's only on the first write fault on the vma, not
946 * every time, and we could devise a way to avoid it later
947 * (e.g. stash info in next's anon_vma_node when assigning
948 * an anon_vma, or when trying vma_merge). Another time.
950 BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
954 anon_vma = reusable_anon_vma(near, near, vma);
959 * There's no absolute need to look only at touching neighbours:
960 * we could search further afield for "compatible" anon_vmas.
961 * But it would probably just be a waste of time searching,
962 * or lead to too many vmas hanging off the same anon_vma.
963 * We're trying to allow mprotect remerging later on,
964 * not trying to minimize memory used for anon_vmas.
969 #ifdef CONFIG_PROC_FS
970 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
971 struct file *file, long pages)
973 const unsigned long stack_flags
974 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
977 mm->shared_vm += pages;
978 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
979 mm->exec_vm += pages;
980 } else if (flags & stack_flags)
981 mm->stack_vm += pages;
982 if (flags & (VM_RESERVED|VM_IO))
983 mm->reserved_vm += pages;
985 #endif /* CONFIG_PROC_FS */
988 * The caller must hold down_write(¤t->mm->mmap_sem).
991 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
992 unsigned long len, unsigned long prot,
993 unsigned long flags, unsigned long pgoff)
995 struct mm_struct * mm = current->mm;
997 unsigned int vm_flags;
999 unsigned long reqprot = prot;
1002 * Does the application expect PROT_READ to imply PROT_EXEC?
1004 * (the exception is when the underlying filesystem is noexec
1005 * mounted, in which case we dont add PROT_EXEC.)
1007 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1008 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
1014 if (!(flags & MAP_FIXED))
1015 addr = round_hint_to_min(addr);
1017 /* Careful about overflows.. */
1018 len = PAGE_ALIGN(len);
1022 /* offset overflow? */
1023 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1026 /* Too many mappings? */
1027 if (mm->map_count > sysctl_max_map_count)
1030 /* Obtain the address to map to. we verify (or select) it and ensure
1031 * that it represents a valid section of the address space.
1033 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1034 if (addr & ~PAGE_MASK)
1037 /* Do simple checking here so the lower-level routines won't have
1038 * to. we assume access permissions have been handled by the open
1039 * of the memory object, so we don't do any here.
1041 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1042 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1044 if (flags & MAP_LOCKED)
1045 if (!can_do_mlock())
1048 /* mlock MCL_FUTURE? */
1049 if (vm_flags & VM_LOCKED) {
1050 unsigned long locked, lock_limit;
1051 locked = len >> PAGE_SHIFT;
1052 locked += mm->locked_vm;
1053 lock_limit = rlimit(RLIMIT_MEMLOCK);
1054 lock_limit >>= PAGE_SHIFT;
1055 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1059 inode = file ? file->f_path.dentry->d_inode : NULL;
1062 switch (flags & MAP_TYPE) {
1064 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1068 * Make sure we don't allow writing to an append-only
1071 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1075 * Make sure there are no mandatory locks on the file.
1077 if (locks_verify_locked(inode))
1080 vm_flags |= VM_SHARED | VM_MAYSHARE;
1081 if (!(file->f_mode & FMODE_WRITE))
1082 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1086 if (!(file->f_mode & FMODE_READ))
1088 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1089 if (vm_flags & VM_EXEC)
1091 vm_flags &= ~VM_MAYEXEC;
1094 if (!file->f_op || !file->f_op->mmap)
1102 switch (flags & MAP_TYPE) {
1108 vm_flags |= VM_SHARED | VM_MAYSHARE;
1112 * Set pgoff according to addr for anon_vma.
1114 pgoff = addr >> PAGE_SHIFT;
1121 error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1125 return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1127 EXPORT_SYMBOL(do_mmap_pgoff);
1129 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1130 unsigned long, prot, unsigned long, flags,
1131 unsigned long, fd, unsigned long, pgoff)
1133 struct file *file = NULL;
1134 unsigned long retval = -EBADF;
1136 if (!(flags & MAP_ANONYMOUS)) {
1137 audit_mmap_fd(fd, flags);
1138 if (unlikely(flags & MAP_HUGETLB))
1143 } else if (flags & MAP_HUGETLB) {
1144 struct user_struct *user = NULL;
1146 * VM_NORESERVE is used because the reservations will be
1147 * taken when vm_ops->mmap() is called
1148 * A dummy user value is used because we are not locking
1149 * memory so no accounting is necessary
1151 len = ALIGN(len, huge_page_size(&default_hstate));
1152 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
1153 &user, HUGETLB_ANONHUGE_INODE);
1155 return PTR_ERR(file);
1158 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1160 down_write(¤t->mm->mmap_sem);
1161 retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1162 up_write(¤t->mm->mmap_sem);
1170 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1171 struct mmap_arg_struct {
1175 unsigned long flags;
1177 unsigned long offset;
1180 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1182 struct mmap_arg_struct a;
1184 if (copy_from_user(&a, arg, sizeof(a)))
1186 if (a.offset & ~PAGE_MASK)
1189 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1190 a.offset >> PAGE_SHIFT);
1192 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1195 * Some shared mappigns will want the pages marked read-only
1196 * to track write events. If so, we'll downgrade vm_page_prot
1197 * to the private version (using protection_map[] without the
1200 int vma_wants_writenotify(struct vm_area_struct *vma)
1202 unsigned int vm_flags = vma->vm_flags;
1204 /* If it was private or non-writable, the write bit is already clear */
1205 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1208 /* The backer wishes to know when pages are first written to? */
1209 if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1212 /* The open routine did something to the protections already? */
1213 if (pgprot_val(vma->vm_page_prot) !=
1214 pgprot_val(vm_get_page_prot(vm_flags)))
1217 /* Specialty mapping? */
1218 if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1221 /* Can the mapping track the dirty pages? */
1222 return vma->vm_file && vma->vm_file->f_mapping &&
1223 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1227 * We account for memory if it's a private writeable mapping,
1228 * not hugepages and VM_NORESERVE wasn't set.
1230 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1233 * hugetlb has its own accounting separate from the core VM
1234 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1236 if (file && is_file_hugepages(file))
1239 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1242 unsigned long mmap_region(struct file *file, unsigned long addr,
1243 unsigned long len, unsigned long flags,
1244 unsigned int vm_flags, unsigned long pgoff)
1246 struct mm_struct *mm = current->mm;
1247 struct vm_area_struct *vma, *prev;
1248 int correct_wcount = 0;
1250 struct rb_node **rb_link, *rb_parent;
1251 unsigned long charged = 0;
1252 struct inode *inode = file ? file->f_path.dentry->d_inode : NULL;
1254 /* Clear old maps */
1257 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1258 if (vma && vma->vm_start < addr + len) {
1259 if (do_munmap(mm, addr, len))
1264 /* Check against address space limit. */
1265 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1269 * Set 'VM_NORESERVE' if we should not account for the
1270 * memory use of this mapping.
1272 if ((flags & MAP_NORESERVE)) {
1273 /* We honor MAP_NORESERVE if allowed to overcommit */
1274 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1275 vm_flags |= VM_NORESERVE;
1277 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1278 if (file && is_file_hugepages(file))
1279 vm_flags |= VM_NORESERVE;
1283 * Private writable mapping: check memory availability
1285 if (accountable_mapping(file, vm_flags)) {
1286 charged = len >> PAGE_SHIFT;
1287 if (security_vm_enough_memory(charged))
1289 vm_flags |= VM_ACCOUNT;
1293 * Can we just expand an old mapping?
1295 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1300 * Determine the object being mapped and call the appropriate
1301 * specific mapper. the address has already been validated, but
1302 * not unmapped, but the maps are removed from the list.
1304 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1311 vma->vm_start = addr;
1312 vma->vm_end = addr + len;
1313 vma->vm_flags = vm_flags;
1314 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1315 vma->vm_pgoff = pgoff;
1316 INIT_LIST_HEAD(&vma->anon_vma_chain);
1320 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1322 if (vm_flags & VM_DENYWRITE) {
1323 error = deny_write_access(file);
1328 vma->vm_file = file;
1330 error = file->f_op->mmap(file, vma);
1332 goto unmap_and_free_vma;
1333 if (vm_flags & VM_EXECUTABLE)
1334 added_exe_file_vma(mm);
1336 /* Can addr have changed??
1338 * Answer: Yes, several device drivers can do it in their
1339 * f_op->mmap method. -DaveM
1341 addr = vma->vm_start;
1342 pgoff = vma->vm_pgoff;
1343 vm_flags = vma->vm_flags;
1344 } else if (vm_flags & VM_SHARED) {
1345 error = shmem_zero_setup(vma);
1350 if (vma_wants_writenotify(vma)) {
1351 pgprot_t pprot = vma->vm_page_prot;
1353 /* Can vma->vm_page_prot have changed??
1355 * Answer: Yes, drivers may have changed it in their
1356 * f_op->mmap method.
1358 * Ensures that vmas marked as uncached stay that way.
1360 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1361 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1362 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1365 vma_link(mm, vma, prev, rb_link, rb_parent);
1366 file = vma->vm_file;
1368 /* Once vma denies write, undo our temporary denial count */
1370 atomic_inc(&inode->i_writecount);
1372 perf_event_mmap(vma);
1374 mm->total_vm += len >> PAGE_SHIFT;
1375 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1376 if (vm_flags & VM_LOCKED) {
1377 if (!mlock_vma_pages_range(vma, addr, addr + len))
1378 mm->locked_vm += (len >> PAGE_SHIFT);
1379 } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1380 make_pages_present(addr, addr + len);
1385 atomic_inc(&inode->i_writecount);
1386 vma->vm_file = NULL;
1389 /* Undo any partial mapping done by a device driver. */
1390 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1393 kmem_cache_free(vm_area_cachep, vma);
1396 vm_unacct_memory(charged);
1400 /* Get an address range which is currently unmapped.
1401 * For shmat() with addr=0.
1403 * Ugly calling convention alert:
1404 * Return value with the low bits set means error value,
1406 * if (ret & ~PAGE_MASK)
1409 * This function "knows" that -ENOMEM has the bits set.
1411 #ifndef HAVE_ARCH_UNMAPPED_AREA
1413 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1414 unsigned long len, unsigned long pgoff, unsigned long flags)
1416 struct mm_struct *mm = current->mm;
1417 struct vm_area_struct *vma;
1418 unsigned long start_addr;
1420 if (len > TASK_SIZE)
1423 if (flags & MAP_FIXED)
1427 addr = PAGE_ALIGN(addr);
1428 vma = find_vma(mm, addr);
1429 if (TASK_SIZE - len >= addr &&
1430 (!vma || addr + len <= vma->vm_start))
1433 if (len > mm->cached_hole_size) {
1434 start_addr = addr = mm->free_area_cache;
1436 start_addr = addr = TASK_UNMAPPED_BASE;
1437 mm->cached_hole_size = 0;
1441 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1442 /* At this point: (!vma || addr < vma->vm_end). */
1443 if (TASK_SIZE - len < addr) {
1445 * Start a new search - just in case we missed
1448 if (start_addr != TASK_UNMAPPED_BASE) {
1449 addr = TASK_UNMAPPED_BASE;
1451 mm->cached_hole_size = 0;
1456 if (!vma || addr + len <= vma->vm_start) {
1458 * Remember the place where we stopped the search:
1460 mm->free_area_cache = addr + len;
1463 if (addr + mm->cached_hole_size < vma->vm_start)
1464 mm->cached_hole_size = vma->vm_start - addr;
1470 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1473 * Is this a new hole at the lowest possible address?
1475 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1476 mm->free_area_cache = addr;
1477 mm->cached_hole_size = ~0UL;
1482 * This mmap-allocator allocates new areas top-down from below the
1483 * stack's low limit (the base):
1485 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1487 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1488 const unsigned long len, const unsigned long pgoff,
1489 const unsigned long flags)
1491 struct vm_area_struct *vma;
1492 struct mm_struct *mm = current->mm;
1493 unsigned long addr = addr0;
1495 /* requested length too big for entire address space */
1496 if (len > TASK_SIZE)
1499 if (flags & MAP_FIXED)
1502 /* requesting a specific address */
1504 addr = PAGE_ALIGN(addr);
1505 vma = find_vma(mm, addr);
1506 if (TASK_SIZE - len >= addr &&
1507 (!vma || addr + len <= vma->vm_start))
1511 /* check if free_area_cache is useful for us */
1512 if (len <= mm->cached_hole_size) {
1513 mm->cached_hole_size = 0;
1514 mm->free_area_cache = mm->mmap_base;
1517 /* either no address requested or can't fit in requested address hole */
1518 addr = mm->free_area_cache;
1520 /* make sure it can fit in the remaining address space */
1522 vma = find_vma(mm, addr-len);
1523 if (!vma || addr <= vma->vm_start)
1524 /* remember the address as a hint for next time */
1525 return (mm->free_area_cache = addr-len);
1528 if (mm->mmap_base < len)
1531 addr = mm->mmap_base-len;
1535 * Lookup failure means no vma is above this address,
1536 * else if new region fits below vma->vm_start,
1537 * return with success:
1539 vma = find_vma(mm, addr);
1540 if (!vma || addr+len <= vma->vm_start)
1541 /* remember the address as a hint for next time */
1542 return (mm->free_area_cache = addr);
1544 /* remember the largest hole we saw so far */
1545 if (addr + mm->cached_hole_size < vma->vm_start)
1546 mm->cached_hole_size = vma->vm_start - addr;
1548 /* try just below the current vma->vm_start */
1549 addr = vma->vm_start-len;
1550 } while (len < vma->vm_start);
1554 * A failed mmap() very likely causes application failure,
1555 * so fall back to the bottom-up function here. This scenario
1556 * can happen with large stack limits and large mmap()
1559 mm->cached_hole_size = ~0UL;
1560 mm->free_area_cache = TASK_UNMAPPED_BASE;
1561 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1563 * Restore the topdown base:
1565 mm->free_area_cache = mm->mmap_base;
1566 mm->cached_hole_size = ~0UL;
1572 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1575 * Is this a new hole at the highest possible address?
1577 if (addr > mm->free_area_cache)
1578 mm->free_area_cache = addr;
1580 /* dont allow allocations above current base */
1581 if (mm->free_area_cache > mm->mmap_base)
1582 mm->free_area_cache = mm->mmap_base;
1586 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1587 unsigned long pgoff, unsigned long flags)
1589 unsigned long (*get_area)(struct file *, unsigned long,
1590 unsigned long, unsigned long, unsigned long);
1592 unsigned long error = arch_mmap_check(addr, len, flags);
1596 /* Careful about overflows.. */
1597 if (len > TASK_SIZE)
1600 get_area = current->mm->get_unmapped_area;
1601 if (file && file->f_op && file->f_op->get_unmapped_area)
1602 get_area = file->f_op->get_unmapped_area;
1603 addr = get_area(file, addr, len, pgoff, flags);
1604 if (IS_ERR_VALUE(addr))
1607 if (addr > TASK_SIZE - len)
1609 if (addr & ~PAGE_MASK)
1612 return arch_rebalance_pgtables(addr, len);
1615 EXPORT_SYMBOL(get_unmapped_area);
1617 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1618 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1620 struct vm_area_struct *vma = NULL;
1623 /* Check the cache first. */
1624 /* (Cache hit rate is typically around 35%.) */
1625 vma = mm->mmap_cache;
1626 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1627 struct rb_node * rb_node;
1629 rb_node = mm->mm_rb.rb_node;
1633 struct vm_area_struct * vma_tmp;
1635 vma_tmp = rb_entry(rb_node,
1636 struct vm_area_struct, vm_rb);
1638 if (vma_tmp->vm_end > addr) {
1640 if (vma_tmp->vm_start <= addr)
1642 rb_node = rb_node->rb_left;
1644 rb_node = rb_node->rb_right;
1647 mm->mmap_cache = vma;
1653 EXPORT_SYMBOL(find_vma);
1655 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1656 struct vm_area_struct *
1657 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1658 struct vm_area_struct **pprev)
1660 struct vm_area_struct *vma = NULL, *prev = NULL;
1661 struct rb_node *rb_node;
1665 /* Guard against addr being lower than the first VMA */
1668 /* Go through the RB tree quickly. */
1669 rb_node = mm->mm_rb.rb_node;
1672 struct vm_area_struct *vma_tmp;
1673 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1675 if (addr < vma_tmp->vm_end) {
1676 rb_node = rb_node->rb_left;
1679 if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1681 rb_node = rb_node->rb_right;
1687 return prev ? prev->vm_next : vma;
1691 * Verify that the stack growth is acceptable and
1692 * update accounting. This is shared with both the
1693 * grow-up and grow-down cases.
1695 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1697 struct mm_struct *mm = vma->vm_mm;
1698 struct rlimit *rlim = current->signal->rlim;
1699 unsigned long new_start;
1701 /* address space limit tests */
1702 if (!may_expand_vm(mm, grow))
1705 /* Stack limit test */
1706 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1709 /* mlock limit tests */
1710 if (vma->vm_flags & VM_LOCKED) {
1711 unsigned long locked;
1712 unsigned long limit;
1713 locked = mm->locked_vm + grow;
1714 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1715 limit >>= PAGE_SHIFT;
1716 if (locked > limit && !capable(CAP_IPC_LOCK))
1720 /* Check to ensure the stack will not grow into a hugetlb-only region */
1721 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1723 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1727 * Overcommit.. This must be the final test, as it will
1728 * update security statistics.
1730 if (security_vm_enough_memory_mm(mm, grow))
1733 /* Ok, everything looks good - let it rip */
1734 mm->total_vm += grow;
1735 if (vma->vm_flags & VM_LOCKED)
1736 mm->locked_vm += grow;
1737 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1741 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1743 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1744 * vma is the last one with address > vma->vm_end. Have to extend vma.
1746 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1750 if (!(vma->vm_flags & VM_GROWSUP))
1754 * We must make sure the anon_vma is allocated
1755 * so that the anon_vma locking is not a noop.
1757 if (unlikely(anon_vma_prepare(vma)))
1759 vma_lock_anon_vma(vma);
1762 * vma->vm_start/vm_end cannot change under us because the caller
1763 * is required to hold the mmap_sem in read mode. We need the
1764 * anon_vma lock to serialize against concurrent expand_stacks.
1765 * Also guard against wrapping around to address 0.
1767 if (address < PAGE_ALIGN(address+4))
1768 address = PAGE_ALIGN(address+4);
1770 vma_unlock_anon_vma(vma);
1775 /* Somebody else might have raced and expanded it already */
1776 if (address > vma->vm_end) {
1777 unsigned long size, grow;
1779 size = address - vma->vm_start;
1780 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1783 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1784 error = acct_stack_growth(vma, size, grow);
1786 vma->vm_end = address;
1787 perf_event_mmap(vma);
1791 vma_unlock_anon_vma(vma);
1792 khugepaged_enter_vma_merge(vma);
1795 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1798 * vma is the first one with address < vma->vm_start. Have to extend vma.
1800 static int expand_downwards(struct vm_area_struct *vma,
1801 unsigned long address)
1806 * We must make sure the anon_vma is allocated
1807 * so that the anon_vma locking is not a noop.
1809 if (unlikely(anon_vma_prepare(vma)))
1812 address &= PAGE_MASK;
1813 error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1817 vma_lock_anon_vma(vma);
1820 * vma->vm_start/vm_end cannot change under us because the caller
1821 * is required to hold the mmap_sem in read mode. We need the
1822 * anon_vma lock to serialize against concurrent expand_stacks.
1825 /* Somebody else might have raced and expanded it already */
1826 if (address < vma->vm_start) {
1827 unsigned long size, grow;
1829 size = vma->vm_end - address;
1830 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1833 if (grow <= vma->vm_pgoff) {
1834 error = acct_stack_growth(vma, size, grow);
1836 vma->vm_start = address;
1837 vma->vm_pgoff -= grow;
1838 perf_event_mmap(vma);
1842 vma_unlock_anon_vma(vma);
1843 khugepaged_enter_vma_merge(vma);
1847 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1849 return expand_downwards(vma, address);
1852 #ifdef CONFIG_STACK_GROWSUP
1853 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1855 return expand_upwards(vma, address);
1858 struct vm_area_struct *
1859 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1861 struct vm_area_struct *vma, *prev;
1864 vma = find_vma_prev(mm, addr, &prev);
1865 if (vma && (vma->vm_start <= addr))
1867 if (!prev || expand_stack(prev, addr))
1869 if (prev->vm_flags & VM_LOCKED) {
1870 mlock_vma_pages_range(prev, addr, prev->vm_end);
1875 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1877 return expand_downwards(vma, address);
1880 struct vm_area_struct *
1881 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1883 struct vm_area_struct * vma;
1884 unsigned long start;
1887 vma = find_vma(mm,addr);
1890 if (vma->vm_start <= addr)
1892 if (!(vma->vm_flags & VM_GROWSDOWN))
1894 start = vma->vm_start;
1895 if (expand_stack(vma, addr))
1897 if (vma->vm_flags & VM_LOCKED) {
1898 mlock_vma_pages_range(vma, addr, start);
1905 * Ok - we have the memory areas we should free on the vma list,
1906 * so release them, and do the vma updates.
1908 * Called with the mm semaphore held.
1910 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1912 /* Update high watermark before we lower total_vm */
1913 update_hiwater_vm(mm);
1915 long nrpages = vma_pages(vma);
1917 mm->total_vm -= nrpages;
1918 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1919 vma = remove_vma(vma);
1925 * Get rid of page table information in the indicated region.
1927 * Called with the mm semaphore held.
1929 static void unmap_region(struct mm_struct *mm,
1930 struct vm_area_struct *vma, struct vm_area_struct *prev,
1931 unsigned long start, unsigned long end)
1933 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1934 struct mmu_gather *tlb;
1935 unsigned long nr_accounted = 0;
1938 tlb = tlb_gather_mmu(mm, 0);
1939 update_hiwater_rss(mm);
1940 unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1941 vm_unacct_memory(nr_accounted);
1942 free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1943 next? next->vm_start: 0);
1944 tlb_finish_mmu(tlb, start, end);
1948 * Create a list of vma's touched by the unmap, removing them from the mm's
1949 * vma list as we go..
1952 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1953 struct vm_area_struct *prev, unsigned long end)
1955 struct vm_area_struct **insertion_point;
1956 struct vm_area_struct *tail_vma = NULL;
1959 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1960 vma->vm_prev = NULL;
1962 rb_erase(&vma->vm_rb, &mm->mm_rb);
1966 } while (vma && vma->vm_start < end);
1967 *insertion_point = vma;
1969 vma->vm_prev = prev;
1970 tail_vma->vm_next = NULL;
1971 if (mm->unmap_area == arch_unmap_area)
1972 addr = prev ? prev->vm_end : mm->mmap_base;
1974 addr = vma ? vma->vm_start : mm->mmap_base;
1975 mm->unmap_area(mm, addr);
1976 mm->mmap_cache = NULL; /* Kill the cache. */
1980 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
1981 * munmap path where it doesn't make sense to fail.
1983 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1984 unsigned long addr, int new_below)
1986 struct mempolicy *pol;
1987 struct vm_area_struct *new;
1990 if (is_vm_hugetlb_page(vma) && (addr &
1991 ~(huge_page_mask(hstate_vma(vma)))))
1994 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1998 /* most fields are the same, copy all, and then fixup */
2001 INIT_LIST_HEAD(&new->anon_vma_chain);
2006 new->vm_start = addr;
2007 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2010 pol = mpol_dup(vma_policy(vma));
2015 vma_set_policy(new, pol);
2017 if (anon_vma_clone(new, vma))
2021 get_file(new->vm_file);
2022 if (vma->vm_flags & VM_EXECUTABLE)
2023 added_exe_file_vma(mm);
2026 if (new->vm_ops && new->vm_ops->open)
2027 new->vm_ops->open(new);
2030 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2031 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2033 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2039 /* Clean everything up if vma_adjust failed. */
2040 if (new->vm_ops && new->vm_ops->close)
2041 new->vm_ops->close(new);
2043 if (vma->vm_flags & VM_EXECUTABLE)
2044 removed_exe_file_vma(mm);
2047 unlink_anon_vmas(new);
2051 kmem_cache_free(vm_area_cachep, new);
2057 * Split a vma into two pieces at address 'addr', a new vma is allocated
2058 * either for the first part or the tail.
2060 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2061 unsigned long addr, int new_below)
2063 if (mm->map_count >= sysctl_max_map_count)
2066 return __split_vma(mm, vma, addr, new_below);
2069 /* Munmap is split into 2 main parts -- this part which finds
2070 * what needs doing, and the areas themselves, which do the
2071 * work. This now handles partial unmappings.
2072 * Jeremy Fitzhardinge <jeremy@goop.org>
2074 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2077 struct vm_area_struct *vma, *prev, *last;
2079 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2082 if ((len = PAGE_ALIGN(len)) == 0)
2085 /* Find the first overlapping VMA */
2086 vma = find_vma_prev(mm, start, &prev);
2089 /* we have start < vma->vm_end */
2091 /* if it doesn't overlap, we have nothing.. */
2093 if (vma->vm_start >= end)
2097 * If we need to split any vma, do it now to save pain later.
2099 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2100 * unmapped vm_area_struct will remain in use: so lower split_vma
2101 * places tmp vma above, and higher split_vma places tmp vma below.
2103 if (start > vma->vm_start) {
2107 * Make sure that map_count on return from munmap() will
2108 * not exceed its limit; but let map_count go just above
2109 * its limit temporarily, to help free resources as expected.
2111 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2114 error = __split_vma(mm, vma, start, 0);
2120 /* Does it split the last one? */
2121 last = find_vma(mm, end);
2122 if (last && end > last->vm_start) {
2123 int error = __split_vma(mm, last, end, 1);
2127 vma = prev? prev->vm_next: mm->mmap;
2130 * unlock any mlock()ed ranges before detaching vmas
2132 if (mm->locked_vm) {
2133 struct vm_area_struct *tmp = vma;
2134 while (tmp && tmp->vm_start < end) {
2135 if (tmp->vm_flags & VM_LOCKED) {
2136 mm->locked_vm -= vma_pages(tmp);
2137 munlock_vma_pages_all(tmp);
2144 * Remove the vma's, and unmap the actual pages
2146 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2147 unmap_region(mm, vma, prev, start, end);
2149 /* Fix up all other VM information */
2150 remove_vma_list(mm, vma);
2155 EXPORT_SYMBOL(do_munmap);
2157 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2160 struct mm_struct *mm = current->mm;
2162 profile_munmap(addr);
2164 down_write(&mm->mmap_sem);
2165 ret = do_munmap(mm, addr, len);
2166 up_write(&mm->mmap_sem);
2170 static inline void verify_mm_writelocked(struct mm_struct *mm)
2172 #ifdef CONFIG_DEBUG_VM
2173 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2175 up_read(&mm->mmap_sem);
2181 * this is really a simplified "do_mmap". it only handles
2182 * anonymous maps. eventually we may be able to do some
2183 * brk-specific accounting here.
2185 unsigned long do_brk(unsigned long addr, unsigned long len)
2187 struct mm_struct * mm = current->mm;
2188 struct vm_area_struct * vma, * prev;
2189 unsigned long flags;
2190 struct rb_node ** rb_link, * rb_parent;
2191 pgoff_t pgoff = addr >> PAGE_SHIFT;
2194 len = PAGE_ALIGN(len);
2198 error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2202 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2204 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2205 if (error & ~PAGE_MASK)
2211 if (mm->def_flags & VM_LOCKED) {
2212 unsigned long locked, lock_limit;
2213 locked = len >> PAGE_SHIFT;
2214 locked += mm->locked_vm;
2215 lock_limit = rlimit(RLIMIT_MEMLOCK);
2216 lock_limit >>= PAGE_SHIFT;
2217 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2222 * mm->mmap_sem is required to protect against another thread
2223 * changing the mappings in case we sleep.
2225 verify_mm_writelocked(mm);
2228 * Clear old maps. this also does some error checking for us
2231 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2232 if (vma && vma->vm_start < addr + len) {
2233 if (do_munmap(mm, addr, len))
2238 /* Check against address space limits *after* clearing old maps... */
2239 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2242 if (mm->map_count > sysctl_max_map_count)
2245 if (security_vm_enough_memory(len >> PAGE_SHIFT))
2248 /* Can we just expand an old private anonymous mapping? */
2249 vma = vma_merge(mm, prev, addr, addr + len, flags,
2250 NULL, NULL, pgoff, NULL);
2255 * create a vma struct for an anonymous mapping
2257 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2259 vm_unacct_memory(len >> PAGE_SHIFT);
2263 INIT_LIST_HEAD(&vma->anon_vma_chain);
2265 vma->vm_start = addr;
2266 vma->vm_end = addr + len;
2267 vma->vm_pgoff = pgoff;
2268 vma->vm_flags = flags;
2269 vma->vm_page_prot = vm_get_page_prot(flags);
2270 vma_link(mm, vma, prev, rb_link, rb_parent);
2272 perf_event_mmap(vma);
2273 mm->total_vm += len >> PAGE_SHIFT;
2274 if (flags & VM_LOCKED) {
2275 if (!mlock_vma_pages_range(vma, addr, addr + len))
2276 mm->locked_vm += (len >> PAGE_SHIFT);
2281 EXPORT_SYMBOL(do_brk);
2283 /* Release all mmaps. */
2284 void exit_mmap(struct mm_struct *mm)
2286 struct mmu_gather *tlb;
2287 struct vm_area_struct *vma;
2288 unsigned long nr_accounted = 0;
2291 /* mm's last user has gone, and its about to be pulled down */
2292 mmu_notifier_release(mm);
2294 if (mm->locked_vm) {
2297 if (vma->vm_flags & VM_LOCKED)
2298 munlock_vma_pages_all(vma);
2306 if (!vma) /* Can happen if dup_mmap() received an OOM */
2311 tlb = tlb_gather_mmu(mm, 1);
2312 /* update_hiwater_rss(mm) here? but nobody should be looking */
2313 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2314 end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2315 vm_unacct_memory(nr_accounted);
2317 free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2318 tlb_finish_mmu(tlb, 0, end);
2321 * Walk the list again, actually closing and freeing it,
2322 * with preemption enabled, without holding any MM locks.
2325 vma = remove_vma(vma);
2327 BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2330 /* Insert vm structure into process list sorted by address
2331 * and into the inode's i_mmap tree. If vm_file is non-NULL
2332 * then i_mmap_lock is taken here.
2334 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2336 struct vm_area_struct * __vma, * prev;
2337 struct rb_node ** rb_link, * rb_parent;
2340 * The vm_pgoff of a purely anonymous vma should be irrelevant
2341 * until its first write fault, when page's anon_vma and index
2342 * are set. But now set the vm_pgoff it will almost certainly
2343 * end up with (unless mremap moves it elsewhere before that
2344 * first wfault), so /proc/pid/maps tells a consistent story.
2346 * By setting it to reflect the virtual start address of the
2347 * vma, merges and splits can happen in a seamless way, just
2348 * using the existing file pgoff checks and manipulations.
2349 * Similarly in do_mmap_pgoff and in do_brk.
2351 if (!vma->vm_file) {
2352 BUG_ON(vma->anon_vma);
2353 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2355 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2356 if (__vma && __vma->vm_start < vma->vm_end)
2358 if ((vma->vm_flags & VM_ACCOUNT) &&
2359 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2361 vma_link(mm, vma, prev, rb_link, rb_parent);
2366 * Copy the vma structure to a new location in the same mm,
2367 * prior to moving page table entries, to effect an mremap move.
2369 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2370 unsigned long addr, unsigned long len, pgoff_t pgoff)
2372 struct vm_area_struct *vma = *vmap;
2373 unsigned long vma_start = vma->vm_start;
2374 struct mm_struct *mm = vma->vm_mm;
2375 struct vm_area_struct *new_vma, *prev;
2376 struct rb_node **rb_link, *rb_parent;
2377 struct mempolicy *pol;
2380 * If anonymous vma has not yet been faulted, update new pgoff
2381 * to match new location, to increase its chance of merging.
2383 if (!vma->vm_file && !vma->anon_vma)
2384 pgoff = addr >> PAGE_SHIFT;
2386 find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2387 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2388 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2391 * Source vma may have been merged into new_vma
2393 if (vma_start >= new_vma->vm_start &&
2394 vma_start < new_vma->vm_end)
2397 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2400 pol = mpol_dup(vma_policy(vma));
2403 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2404 if (anon_vma_clone(new_vma, vma))
2405 goto out_free_mempol;
2406 vma_set_policy(new_vma, pol);
2407 new_vma->vm_start = addr;
2408 new_vma->vm_end = addr + len;
2409 new_vma->vm_pgoff = pgoff;
2410 if (new_vma->vm_file) {
2411 get_file(new_vma->vm_file);
2412 if (vma->vm_flags & VM_EXECUTABLE)
2413 added_exe_file_vma(mm);
2415 if (new_vma->vm_ops && new_vma->vm_ops->open)
2416 new_vma->vm_ops->open(new_vma);
2417 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2425 kmem_cache_free(vm_area_cachep, new_vma);
2430 * Return true if the calling process may expand its vm space by the passed
2433 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2435 unsigned long cur = mm->total_vm; /* pages */
2438 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2440 if (cur + npages > lim)
2446 static int special_mapping_fault(struct vm_area_struct *vma,
2447 struct vm_fault *vmf)
2450 struct page **pages;
2453 * special mappings have no vm_file, and in that case, the mm
2454 * uses vm_pgoff internally. So we have to subtract it from here.
2455 * We are allowed to do this because we are the mm; do not copy
2456 * this code into drivers!
2458 pgoff = vmf->pgoff - vma->vm_pgoff;
2460 for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2464 struct page *page = *pages;
2470 return VM_FAULT_SIGBUS;
2474 * Having a close hook prevents vma merging regardless of flags.
2476 static void special_mapping_close(struct vm_area_struct *vma)
2480 static const struct vm_operations_struct special_mapping_vmops = {
2481 .close = special_mapping_close,
2482 .fault = special_mapping_fault,
2486 * Called with mm->mmap_sem held for writing.
2487 * Insert a new vma covering the given region, with the given flags.
2488 * Its pages are supplied by the given array of struct page *.
2489 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2490 * The region past the last page supplied will always produce SIGBUS.
2491 * The array pointer and the pages it points to are assumed to stay alive
2492 * for as long as this mapping might exist.
2494 int install_special_mapping(struct mm_struct *mm,
2495 unsigned long addr, unsigned long len,
2496 unsigned long vm_flags, struct page **pages)
2499 struct vm_area_struct *vma;
2501 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2502 if (unlikely(vma == NULL))
2505 INIT_LIST_HEAD(&vma->anon_vma_chain);
2507 vma->vm_start = addr;
2508 vma->vm_end = addr + len;
2510 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2511 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2513 vma->vm_ops = &special_mapping_vmops;
2514 vma->vm_private_data = pages;
2516 ret = security_file_mmap(NULL, 0, 0, 0, vma->vm_start, 1);
2520 ret = insert_vm_struct(mm, vma);
2524 mm->total_vm += len >> PAGE_SHIFT;
2526 perf_event_mmap(vma);
2531 kmem_cache_free(vm_area_cachep, vma);
2535 static DEFINE_MUTEX(mm_all_locks_mutex);
2537 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2539 if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2541 * The LSB of head.next can't change from under us
2542 * because we hold the mm_all_locks_mutex.
2544 spin_lock_nest_lock(&anon_vma->root->lock, &mm->mmap_sem);
2546 * We can safely modify head.next after taking the
2547 * anon_vma->root->lock. If some other vma in this mm shares
2548 * the same anon_vma we won't take it again.
2550 * No need of atomic instructions here, head.next
2551 * can't change from under us thanks to the
2552 * anon_vma->root->lock.
2554 if (__test_and_set_bit(0, (unsigned long *)
2555 &anon_vma->root->head.next))
2560 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2562 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2564 * AS_MM_ALL_LOCKS can't change from under us because
2565 * we hold the mm_all_locks_mutex.
2567 * Operations on ->flags have to be atomic because
2568 * even if AS_MM_ALL_LOCKS is stable thanks to the
2569 * mm_all_locks_mutex, there may be other cpus
2570 * changing other bitflags in parallel to us.
2572 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2574 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2579 * This operation locks against the VM for all pte/vma/mm related
2580 * operations that could ever happen on a certain mm. This includes
2581 * vmtruncate, try_to_unmap, and all page faults.
2583 * The caller must take the mmap_sem in write mode before calling
2584 * mm_take_all_locks(). The caller isn't allowed to release the
2585 * mmap_sem until mm_drop_all_locks() returns.
2587 * mmap_sem in write mode is required in order to block all operations
2588 * that could modify pagetables and free pages without need of
2589 * altering the vma layout (for example populate_range() with
2590 * nonlinear vmas). It's also needed in write mode to avoid new
2591 * anon_vmas to be associated with existing vmas.
2593 * A single task can't take more than one mm_take_all_locks() in a row
2594 * or it would deadlock.
2596 * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2597 * mapping->flags avoid to take the same lock twice, if more than one
2598 * vma in this mm is backed by the same anon_vma or address_space.
2600 * We can take all the locks in random order because the VM code
2601 * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2602 * takes more than one of them in a row. Secondly we're protected
2603 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2605 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2606 * that may have to take thousand of locks.
2608 * mm_take_all_locks() can fail if it's interrupted by signals.
2610 int mm_take_all_locks(struct mm_struct *mm)
2612 struct vm_area_struct *vma;
2613 struct anon_vma_chain *avc;
2616 BUG_ON(down_read_trylock(&mm->mmap_sem));
2618 mutex_lock(&mm_all_locks_mutex);
2620 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2621 if (signal_pending(current))
2623 if (vma->vm_file && vma->vm_file->f_mapping)
2624 vm_lock_mapping(mm, vma->vm_file->f_mapping);
2627 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2628 if (signal_pending(current))
2631 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2632 vm_lock_anon_vma(mm, avc->anon_vma);
2639 mm_drop_all_locks(mm);
2644 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2646 if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2648 * The LSB of head.next can't change to 0 from under
2649 * us because we hold the mm_all_locks_mutex.
2651 * We must however clear the bitflag before unlocking
2652 * the vma so the users using the anon_vma->head will
2653 * never see our bitflag.
2655 * No need of atomic instructions here, head.next
2656 * can't change from under us until we release the
2657 * anon_vma->root->lock.
2659 if (!__test_and_clear_bit(0, (unsigned long *)
2660 &anon_vma->root->head.next))
2662 anon_vma_unlock(anon_vma);
2666 static void vm_unlock_mapping(struct address_space *mapping)
2668 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2670 * AS_MM_ALL_LOCKS can't change to 0 from under us
2671 * because we hold the mm_all_locks_mutex.
2673 spin_unlock(&mapping->i_mmap_lock);
2674 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2681 * The mmap_sem cannot be released by the caller until
2682 * mm_drop_all_locks() returns.
2684 void mm_drop_all_locks(struct mm_struct *mm)
2686 struct vm_area_struct *vma;
2687 struct anon_vma_chain *avc;
2689 BUG_ON(down_read_trylock(&mm->mmap_sem));
2690 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2692 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2694 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2695 vm_unlock_anon_vma(avc->anon_vma);
2696 if (vma->vm_file && vma->vm_file->f_mapping)
2697 vm_unlock_mapping(vma->vm_file->f_mapping);
2700 mutex_unlock(&mm_all_locks_mutex);
2704 * initialise the VMA slab
2706 void __init mmap_init(void)
2710 ret = percpu_counter_init(&vm_committed_as, 0);