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 = OVERCOMMIT_GUESS; /* heuristic overcommit */
88 int sysctl_overcommit_ratio = 50; /* default is 50% */
89 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
90 struct percpu_counter vm_committed_as;
93 * Check that a process has enough memory to allocate a new virtual
94 * mapping. 0 means there is enough memory for the allocation to
95 * succeed and -ENOMEM implies there is not.
97 * We currently support three overcommit policies, which are set via the
98 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
100 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
101 * Additional code 2002 Jul 20 by Robert Love.
103 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
105 * Note this is a helper function intended to be used by LSMs which
106 * wish to use this logic.
108 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
110 unsigned long free, allowed;
112 vm_acct_memory(pages);
115 * Sometimes we want to use more memory than we have
117 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
120 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
123 free = global_page_state(NR_FILE_PAGES);
124 free += nr_swap_pages;
127 * Any slabs which are created with the
128 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
129 * which are reclaimable, under pressure. The dentry
130 * cache and most inode caches should fall into this
132 free += global_page_state(NR_SLAB_RECLAIMABLE);
135 * Leave the last 3% for root
144 * nr_free_pages() is very expensive on large systems,
145 * only call if we're about to fail.
150 * Leave reserved pages. The pages are not for anonymous pages.
152 if (n <= totalreserve_pages)
155 n -= totalreserve_pages;
158 * Leave the last 3% for root
170 allowed = (totalram_pages - hugetlb_total_pages())
171 * sysctl_overcommit_ratio / 100;
173 * Leave the last 3% for root
176 allowed -= allowed / 32;
177 allowed += total_swap_pages;
179 /* Don't let a single process grow too big:
180 leave 3% of the size of this process for other processes */
182 allowed -= mm->total_vm / 32;
184 if (percpu_counter_read_positive(&vm_committed_as) < allowed)
187 vm_unacct_memory(pages);
193 * Requires inode->i_mapping->i_mmap_lock
195 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
196 struct file *file, struct address_space *mapping)
198 if (vma->vm_flags & VM_DENYWRITE)
199 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
200 if (vma->vm_flags & VM_SHARED)
201 mapping->i_mmap_writable--;
203 flush_dcache_mmap_lock(mapping);
204 if (unlikely(vma->vm_flags & VM_NONLINEAR))
205 list_del_init(&vma->shared.vm_set.list);
207 vma_prio_tree_remove(vma, &mapping->i_mmap);
208 flush_dcache_mmap_unlock(mapping);
212 * Unlink a file-based vm structure from its prio_tree, to hide
213 * vma from rmap and vmtruncate before freeing its page tables.
215 void unlink_file_vma(struct vm_area_struct *vma)
217 struct file *file = vma->vm_file;
220 struct address_space *mapping = file->f_mapping;
221 spin_lock(&mapping->i_mmap_lock);
222 __remove_shared_vm_struct(vma, file, mapping);
223 spin_unlock(&mapping->i_mmap_lock);
228 * Close a vm structure and free it, returning the next.
230 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
232 struct vm_area_struct *next = vma->vm_next;
235 if (vma->vm_ops && vma->vm_ops->close)
236 vma->vm_ops->close(vma);
239 if (vma->vm_flags & VM_EXECUTABLE)
240 removed_exe_file_vma(vma->vm_mm);
242 mpol_put(vma_policy(vma));
243 kmem_cache_free(vm_area_cachep, vma);
247 SYSCALL_DEFINE1(brk, unsigned long, brk)
249 unsigned long rlim, retval;
250 unsigned long newbrk, oldbrk;
251 struct mm_struct *mm = current->mm;
252 unsigned long min_brk;
254 down_write(&mm->mmap_sem);
256 #ifdef CONFIG_COMPAT_BRK
257 min_brk = mm->end_code;
259 min_brk = mm->start_brk;
265 * Check against rlimit here. If this check is done later after the test
266 * of oldbrk with newbrk then it can escape the test and let the data
267 * segment grow beyond its set limit the in case where the limit is
268 * not page aligned -Ram Gupta
270 rlim = rlimit(RLIMIT_DATA);
271 if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
272 (mm->end_data - mm->start_data) > rlim)
275 newbrk = PAGE_ALIGN(brk);
276 oldbrk = PAGE_ALIGN(mm->brk);
277 if (oldbrk == newbrk)
280 /* Always allow shrinking brk. */
281 if (brk <= mm->brk) {
282 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
287 /* Check against existing mmap mappings. */
288 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
291 /* Ok, looks good - let it rip. */
292 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
298 up_write(&mm->mmap_sem);
303 static int browse_rb(struct rb_root *root)
306 struct rb_node *nd, *pn = NULL;
307 unsigned long prev = 0, pend = 0;
309 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
310 struct vm_area_struct *vma;
311 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
312 if (vma->vm_start < prev)
313 printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
314 if (vma->vm_start < pend)
315 printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
316 if (vma->vm_start > vma->vm_end)
317 printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
320 prev = vma->vm_start;
324 for (nd = pn; nd; nd = rb_prev(nd)) {
328 printk("backwards %d, forwards %d\n", j, i), i = 0;
332 void validate_mm(struct mm_struct *mm)
336 struct vm_area_struct *tmp = mm->mmap;
341 if (i != mm->map_count)
342 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
343 i = browse_rb(&mm->mm_rb);
344 if (i != mm->map_count)
345 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
349 #define validate_mm(mm) do { } while (0)
352 static struct vm_area_struct *
353 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
354 struct vm_area_struct **pprev, struct rb_node ***rb_link,
355 struct rb_node ** rb_parent)
357 struct vm_area_struct * vma;
358 struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
360 __rb_link = &mm->mm_rb.rb_node;
361 rb_prev = __rb_parent = NULL;
365 struct vm_area_struct *vma_tmp;
367 __rb_parent = *__rb_link;
368 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
370 if (vma_tmp->vm_end > addr) {
372 if (vma_tmp->vm_start <= addr)
374 __rb_link = &__rb_parent->rb_left;
376 rb_prev = __rb_parent;
377 __rb_link = &__rb_parent->rb_right;
383 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
384 *rb_link = __rb_link;
385 *rb_parent = __rb_parent;
390 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
391 struct vm_area_struct *prev, struct rb_node *rb_parent)
393 struct vm_area_struct *next;
397 next = prev->vm_next;
402 next = rb_entry(rb_parent,
403 struct vm_area_struct, vm_rb);
412 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
413 struct rb_node **rb_link, struct rb_node *rb_parent)
415 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
416 rb_insert_color(&vma->vm_rb, &mm->mm_rb);
419 static void __vma_link_file(struct vm_area_struct *vma)
425 struct address_space *mapping = file->f_mapping;
427 if (vma->vm_flags & VM_DENYWRITE)
428 atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
429 if (vma->vm_flags & VM_SHARED)
430 mapping->i_mmap_writable++;
432 flush_dcache_mmap_lock(mapping);
433 if (unlikely(vma->vm_flags & VM_NONLINEAR))
434 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
436 vma_prio_tree_insert(vma, &mapping->i_mmap);
437 flush_dcache_mmap_unlock(mapping);
442 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
443 struct vm_area_struct *prev, struct rb_node **rb_link,
444 struct rb_node *rb_parent)
446 __vma_link_list(mm, vma, prev, rb_parent);
447 __vma_link_rb(mm, vma, rb_link, rb_parent);
450 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
451 struct vm_area_struct *prev, struct rb_node **rb_link,
452 struct rb_node *rb_parent)
454 struct address_space *mapping = NULL;
457 mapping = vma->vm_file->f_mapping;
460 spin_lock(&mapping->i_mmap_lock);
461 vma->vm_truncate_count = mapping->truncate_count;
464 __vma_link(mm, vma, prev, rb_link, rb_parent);
465 __vma_link_file(vma);
468 spin_unlock(&mapping->i_mmap_lock);
475 * Helper for vma_adjust in the split_vma insert case:
476 * insert vm structure into list and rbtree and anon_vma,
477 * but it has already been inserted into prio_tree earlier.
479 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
481 struct vm_area_struct *__vma, *prev;
482 struct rb_node **rb_link, *rb_parent;
484 __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
485 BUG_ON(__vma && __vma->vm_start < vma->vm_end);
486 __vma_link(mm, vma, prev, rb_link, rb_parent);
491 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
492 struct vm_area_struct *prev)
494 struct vm_area_struct *next = vma->vm_next;
496 prev->vm_next = next;
498 next->vm_prev = prev;
499 rb_erase(&vma->vm_rb, &mm->mm_rb);
500 if (mm->mmap_cache == vma)
501 mm->mmap_cache = prev;
505 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
506 * is already present in an i_mmap tree without adjusting the tree.
507 * The following helper function should be used when such adjustments
508 * are necessary. The "insert" vma (if any) is to be inserted
509 * before we drop the necessary locks.
511 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
512 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
514 struct mm_struct *mm = vma->vm_mm;
515 struct vm_area_struct *next = vma->vm_next;
516 struct vm_area_struct *importer = NULL;
517 struct address_space *mapping = NULL;
518 struct prio_tree_root *root = NULL;
519 struct anon_vma *anon_vma = NULL;
520 struct file *file = vma->vm_file;
521 long adjust_next = 0;
524 if (next && !insert) {
525 struct vm_area_struct *exporter = NULL;
527 if (end >= next->vm_end) {
529 * vma expands, overlapping all the next, and
530 * perhaps the one after too (mprotect case 6).
532 again: remove_next = 1 + (end > next->vm_end);
536 } else if (end > next->vm_start) {
538 * vma expands, overlapping part of the next:
539 * mprotect case 5 shifting the boundary up.
541 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
544 } else if (end < vma->vm_end) {
546 * vma shrinks, and !insert tells it's not
547 * split_vma inserting another: so it must be
548 * mprotect case 4 shifting the boundary down.
550 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
556 * Easily overlooked: when mprotect shifts the boundary,
557 * make sure the expanding vma has anon_vma set if the
558 * shrinking vma had, to cover any anon pages imported.
560 if (exporter && exporter->anon_vma && !importer->anon_vma) {
561 if (anon_vma_clone(importer, exporter))
563 importer->anon_vma = exporter->anon_vma;
568 mapping = file->f_mapping;
569 if (!(vma->vm_flags & VM_NONLINEAR))
570 root = &mapping->i_mmap;
571 spin_lock(&mapping->i_mmap_lock);
573 vma->vm_truncate_count != next->vm_truncate_count) {
575 * unmap_mapping_range might be in progress:
576 * ensure that the expanding vma is rescanned.
578 importer->vm_truncate_count = 0;
581 insert->vm_truncate_count = vma->vm_truncate_count;
583 * Put into prio_tree now, so instantiated pages
584 * are visible to arm/parisc __flush_dcache_page
585 * throughout; but we cannot insert into address
586 * space until vma start or end is updated.
588 __vma_link_file(insert);
592 vma_adjust_trans_huge(vma, start, end, adjust_next);
595 * When changing only vma->vm_end, we don't really need anon_vma
596 * lock. This is a fairly rare case by itself, but the anon_vma
597 * lock may be shared between many sibling processes. Skipping
598 * the lock for brk adjustments makes a difference sometimes.
600 if (vma->anon_vma && (insert || importer || start != vma->vm_start)) {
601 anon_vma = vma->anon_vma;
602 anon_vma_lock(anon_vma);
606 flush_dcache_mmap_lock(mapping);
607 vma_prio_tree_remove(vma, root);
609 vma_prio_tree_remove(next, root);
612 vma->vm_start = start;
614 vma->vm_pgoff = pgoff;
616 next->vm_start += adjust_next << PAGE_SHIFT;
617 next->vm_pgoff += adjust_next;
622 vma_prio_tree_insert(next, root);
623 vma_prio_tree_insert(vma, root);
624 flush_dcache_mmap_unlock(mapping);
629 * vma_merge has merged next into vma, and needs
630 * us to remove next before dropping the locks.
632 __vma_unlink(mm, next, vma);
634 __remove_shared_vm_struct(next, file, mapping);
637 * split_vma has split insert from vma, and needs
638 * us to insert it before dropping the locks
639 * (it may either follow vma or precede it).
641 __insert_vm_struct(mm, insert);
645 anon_vma_unlock(anon_vma);
647 spin_unlock(&mapping->i_mmap_lock);
652 if (next->vm_flags & VM_EXECUTABLE)
653 removed_exe_file_vma(mm);
656 anon_vma_merge(vma, next);
658 mpol_put(vma_policy(next));
659 kmem_cache_free(vm_area_cachep, next);
661 * In mprotect's case 6 (see comments on vma_merge),
662 * we must remove another next too. It would clutter
663 * up the code too much to do both in one go.
665 if (remove_next == 2) {
677 * If the vma has a ->close operation then the driver probably needs to release
678 * per-vma resources, so we don't attempt to merge those.
680 static inline int is_mergeable_vma(struct vm_area_struct *vma,
681 struct file *file, unsigned long vm_flags)
683 /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
684 if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
686 if (vma->vm_file != file)
688 if (vma->vm_ops && vma->vm_ops->close)
693 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
694 struct anon_vma *anon_vma2)
696 return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
700 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
701 * in front of (at a lower virtual address and file offset than) the vma.
703 * We cannot merge two vmas if they have differently assigned (non-NULL)
704 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
706 * We don't check here for the merged mmap wrapping around the end of pagecache
707 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
708 * wrap, nor mmaps which cover the final page at index -1UL.
711 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
712 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
714 if (is_mergeable_vma(vma, file, vm_flags) &&
715 is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
716 if (vma->vm_pgoff == vm_pgoff)
723 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
724 * beyond (at a higher virtual address and file offset than) the vma.
726 * We cannot merge two vmas if they have differently assigned (non-NULL)
727 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
730 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
731 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
733 if (is_mergeable_vma(vma, file, vm_flags) &&
734 is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
736 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
737 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
744 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
745 * whether that can be merged with its predecessor or its successor.
746 * Or both (it neatly fills a hole).
748 * In most cases - when called for mmap, brk or mremap - [addr,end) is
749 * certain not to be mapped by the time vma_merge is called; but when
750 * called for mprotect, it is certain to be already mapped (either at
751 * an offset within prev, or at the start of next), and the flags of
752 * this area are about to be changed to vm_flags - and the no-change
753 * case has already been eliminated.
755 * The following mprotect cases have to be considered, where AAAA is
756 * the area passed down from mprotect_fixup, never extending beyond one
757 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
759 * AAAA AAAA AAAA AAAA
760 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
761 * cannot merge might become might become might become
762 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
763 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
764 * mremap move: PPPPNNNNNNNN 8
766 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
767 * might become case 1 below case 2 below case 3 below
769 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
770 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
772 struct vm_area_struct *vma_merge(struct mm_struct *mm,
773 struct vm_area_struct *prev, unsigned long addr,
774 unsigned long end, unsigned long vm_flags,
775 struct anon_vma *anon_vma, struct file *file,
776 pgoff_t pgoff, struct mempolicy *policy)
778 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
779 struct vm_area_struct *area, *next;
783 * We later require that vma->vm_flags == vm_flags,
784 * so this tests vma->vm_flags & VM_SPECIAL, too.
786 if (vm_flags & VM_SPECIAL)
790 next = prev->vm_next;
794 if (next && next->vm_end == end) /* cases 6, 7, 8 */
795 next = next->vm_next;
798 * Can it merge with the predecessor?
800 if (prev && prev->vm_end == addr &&
801 mpol_equal(vma_policy(prev), policy) &&
802 can_vma_merge_after(prev, vm_flags,
803 anon_vma, file, pgoff)) {
805 * OK, it can. Can we now merge in the successor as well?
807 if (next && end == next->vm_start &&
808 mpol_equal(policy, vma_policy(next)) &&
809 can_vma_merge_before(next, vm_flags,
810 anon_vma, file, pgoff+pglen) &&
811 is_mergeable_anon_vma(prev->anon_vma,
814 err = vma_adjust(prev, prev->vm_start,
815 next->vm_end, prev->vm_pgoff, NULL);
816 } else /* cases 2, 5, 7 */
817 err = vma_adjust(prev, prev->vm_start,
818 end, prev->vm_pgoff, NULL);
821 khugepaged_enter_vma_merge(prev);
826 * Can this new request be merged in front of next?
828 if (next && end == next->vm_start &&
829 mpol_equal(policy, vma_policy(next)) &&
830 can_vma_merge_before(next, vm_flags,
831 anon_vma, file, pgoff+pglen)) {
832 if (prev && addr < prev->vm_end) /* case 4 */
833 err = vma_adjust(prev, prev->vm_start,
834 addr, prev->vm_pgoff, NULL);
835 else /* cases 3, 8 */
836 err = vma_adjust(area, addr, next->vm_end,
837 next->vm_pgoff - pglen, NULL);
840 khugepaged_enter_vma_merge(area);
848 * Rough compatbility check to quickly see if it's even worth looking
849 * at sharing an anon_vma.
851 * They need to have the same vm_file, and the flags can only differ
852 * in things that mprotect may change.
854 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
855 * we can merge the two vma's. For example, we refuse to merge a vma if
856 * there is a vm_ops->close() function, because that indicates that the
857 * driver is doing some kind of reference counting. But that doesn't
858 * really matter for the anon_vma sharing case.
860 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
862 return a->vm_end == b->vm_start &&
863 mpol_equal(vma_policy(a), vma_policy(b)) &&
864 a->vm_file == b->vm_file &&
865 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
866 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
870 * Do some basic sanity checking to see if we can re-use the anon_vma
871 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
872 * the same as 'old', the other will be the new one that is trying
873 * to share the anon_vma.
875 * NOTE! This runs with mm_sem held for reading, so it is possible that
876 * the anon_vma of 'old' is concurrently in the process of being set up
877 * by another page fault trying to merge _that_. But that's ok: if it
878 * is being set up, that automatically means that it will be a singleton
879 * acceptable for merging, so we can do all of this optimistically. But
880 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
882 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
883 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
884 * is to return an anon_vma that is "complex" due to having gone through
887 * We also make sure that the two vma's are compatible (adjacent,
888 * and with the same memory policies). That's all stable, even with just
889 * a read lock on the mm_sem.
891 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
893 if (anon_vma_compatible(a, b)) {
894 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
896 if (anon_vma && list_is_singular(&old->anon_vma_chain))
903 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
904 * neighbouring vmas for a suitable anon_vma, before it goes off
905 * to allocate a new anon_vma. It checks because a repetitive
906 * sequence of mprotects and faults may otherwise lead to distinct
907 * anon_vmas being allocated, preventing vma merge in subsequent
910 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
912 struct anon_vma *anon_vma;
913 struct vm_area_struct *near;
919 anon_vma = reusable_anon_vma(near, vma, near);
924 * It is potentially slow to have to call find_vma_prev here.
925 * But it's only on the first write fault on the vma, not
926 * every time, and we could devise a way to avoid it later
927 * (e.g. stash info in next's anon_vma_node when assigning
928 * an anon_vma, or when trying vma_merge). Another time.
930 BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
934 anon_vma = reusable_anon_vma(near, near, vma);
939 * There's no absolute need to look only at touching neighbours:
940 * we could search further afield for "compatible" anon_vmas.
941 * But it would probably just be a waste of time searching,
942 * or lead to too many vmas hanging off the same anon_vma.
943 * We're trying to allow mprotect remerging later on,
944 * not trying to minimize memory used for anon_vmas.
949 #ifdef CONFIG_PROC_FS
950 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
951 struct file *file, long pages)
953 const unsigned long stack_flags
954 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
957 mm->shared_vm += pages;
958 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
959 mm->exec_vm += pages;
960 } else if (flags & stack_flags)
961 mm->stack_vm += pages;
962 if (flags & (VM_RESERVED|VM_IO))
963 mm->reserved_vm += pages;
965 #endif /* CONFIG_PROC_FS */
968 * The caller must hold down_write(¤t->mm->mmap_sem).
971 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
972 unsigned long len, unsigned long prot,
973 unsigned long flags, unsigned long pgoff)
975 struct mm_struct * mm = current->mm;
977 unsigned int vm_flags;
979 unsigned long reqprot = prot;
982 * Does the application expect PROT_READ to imply PROT_EXEC?
984 * (the exception is when the underlying filesystem is noexec
985 * mounted, in which case we dont add PROT_EXEC.)
987 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
988 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
994 if (!(flags & MAP_FIXED))
995 addr = round_hint_to_min(addr);
997 /* Careful about overflows.. */
998 len = PAGE_ALIGN(len);
1002 /* offset overflow? */
1003 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1006 /* Too many mappings? */
1007 if (mm->map_count > sysctl_max_map_count)
1010 /* Obtain the address to map to. we verify (or select) it and ensure
1011 * that it represents a valid section of the address space.
1013 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1014 if (addr & ~PAGE_MASK)
1017 /* Do simple checking here so the lower-level routines won't have
1018 * to. we assume access permissions have been handled by the open
1019 * of the memory object, so we don't do any here.
1021 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1022 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1024 if (flags & MAP_LOCKED)
1025 if (!can_do_mlock())
1028 /* mlock MCL_FUTURE? */
1029 if (vm_flags & VM_LOCKED) {
1030 unsigned long locked, lock_limit;
1031 locked = len >> PAGE_SHIFT;
1032 locked += mm->locked_vm;
1033 lock_limit = rlimit(RLIMIT_MEMLOCK);
1034 lock_limit >>= PAGE_SHIFT;
1035 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1039 inode = file ? file->f_path.dentry->d_inode : NULL;
1042 switch (flags & MAP_TYPE) {
1044 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1048 * Make sure we don't allow writing to an append-only
1051 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1055 * Make sure there are no mandatory locks on the file.
1057 if (locks_verify_locked(inode))
1060 vm_flags |= VM_SHARED | VM_MAYSHARE;
1061 if (!(file->f_mode & FMODE_WRITE))
1062 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1066 if (!(file->f_mode & FMODE_READ))
1068 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1069 if (vm_flags & VM_EXEC)
1071 vm_flags &= ~VM_MAYEXEC;
1074 if (!file->f_op || !file->f_op->mmap)
1082 switch (flags & MAP_TYPE) {
1088 vm_flags |= VM_SHARED | VM_MAYSHARE;
1092 * Set pgoff according to addr for anon_vma.
1094 pgoff = addr >> PAGE_SHIFT;
1101 error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1105 return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1107 EXPORT_SYMBOL(do_mmap_pgoff);
1109 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1110 unsigned long, prot, unsigned long, flags,
1111 unsigned long, fd, unsigned long, pgoff)
1113 struct file *file = NULL;
1114 unsigned long retval = -EBADF;
1116 if (!(flags & MAP_ANONYMOUS)) {
1117 audit_mmap_fd(fd, flags);
1118 if (unlikely(flags & MAP_HUGETLB))
1123 } else if (flags & MAP_HUGETLB) {
1124 struct user_struct *user = NULL;
1126 * VM_NORESERVE is used because the reservations will be
1127 * taken when vm_ops->mmap() is called
1128 * A dummy user value is used because we are not locking
1129 * memory so no accounting is necessary
1131 len = ALIGN(len, huge_page_size(&default_hstate));
1132 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
1133 &user, HUGETLB_ANONHUGE_INODE);
1135 return PTR_ERR(file);
1138 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1140 down_write(¤t->mm->mmap_sem);
1141 retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1142 up_write(¤t->mm->mmap_sem);
1150 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1151 struct mmap_arg_struct {
1155 unsigned long flags;
1157 unsigned long offset;
1160 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1162 struct mmap_arg_struct a;
1164 if (copy_from_user(&a, arg, sizeof(a)))
1166 if (a.offset & ~PAGE_MASK)
1169 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1170 a.offset >> PAGE_SHIFT);
1172 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1175 * Some shared mappigns will want the pages marked read-only
1176 * to track write events. If so, we'll downgrade vm_page_prot
1177 * to the private version (using protection_map[] without the
1180 int vma_wants_writenotify(struct vm_area_struct *vma)
1182 unsigned int vm_flags = vma->vm_flags;
1184 /* If it was private or non-writable, the write bit is already clear */
1185 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1188 /* The backer wishes to know when pages are first written to? */
1189 if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1192 /* The open routine did something to the protections already? */
1193 if (pgprot_val(vma->vm_page_prot) !=
1194 pgprot_val(vm_get_page_prot(vm_flags)))
1197 /* Specialty mapping? */
1198 if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1201 /* Can the mapping track the dirty pages? */
1202 return vma->vm_file && vma->vm_file->f_mapping &&
1203 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1207 * We account for memory if it's a private writeable mapping,
1208 * not hugepages and VM_NORESERVE wasn't set.
1210 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1213 * hugetlb has its own accounting separate from the core VM
1214 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1216 if (file && is_file_hugepages(file))
1219 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1222 unsigned long mmap_region(struct file *file, unsigned long addr,
1223 unsigned long len, unsigned long flags,
1224 unsigned int vm_flags, unsigned long pgoff)
1226 struct mm_struct *mm = current->mm;
1227 struct vm_area_struct *vma, *prev;
1228 int correct_wcount = 0;
1230 struct rb_node **rb_link, *rb_parent;
1231 unsigned long charged = 0;
1232 struct inode *inode = file ? file->f_path.dentry->d_inode : NULL;
1234 /* Clear old maps */
1237 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1238 if (vma && vma->vm_start < addr + len) {
1239 if (do_munmap(mm, addr, len))
1244 /* Check against address space limit. */
1245 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1249 * Set 'VM_NORESERVE' if we should not account for the
1250 * memory use of this mapping.
1252 if ((flags & MAP_NORESERVE)) {
1253 /* We honor MAP_NORESERVE if allowed to overcommit */
1254 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1255 vm_flags |= VM_NORESERVE;
1257 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1258 if (file && is_file_hugepages(file))
1259 vm_flags |= VM_NORESERVE;
1263 * Private writable mapping: check memory availability
1265 if (accountable_mapping(file, vm_flags)) {
1266 charged = len >> PAGE_SHIFT;
1267 if (security_vm_enough_memory(charged))
1269 vm_flags |= VM_ACCOUNT;
1273 * Can we just expand an old mapping?
1275 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1280 * Determine the object being mapped and call the appropriate
1281 * specific mapper. the address has already been validated, but
1282 * not unmapped, but the maps are removed from the list.
1284 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1291 vma->vm_start = addr;
1292 vma->vm_end = addr + len;
1293 vma->vm_flags = vm_flags;
1294 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1295 vma->vm_pgoff = pgoff;
1296 INIT_LIST_HEAD(&vma->anon_vma_chain);
1300 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1302 if (vm_flags & VM_DENYWRITE) {
1303 error = deny_write_access(file);
1308 vma->vm_file = file;
1310 error = file->f_op->mmap(file, vma);
1312 goto unmap_and_free_vma;
1313 if (vm_flags & VM_EXECUTABLE)
1314 added_exe_file_vma(mm);
1316 /* Can addr have changed??
1318 * Answer: Yes, several device drivers can do it in their
1319 * f_op->mmap method. -DaveM
1321 addr = vma->vm_start;
1322 pgoff = vma->vm_pgoff;
1323 vm_flags = vma->vm_flags;
1324 } else if (vm_flags & VM_SHARED) {
1325 error = shmem_zero_setup(vma);
1330 if (vma_wants_writenotify(vma)) {
1331 pgprot_t pprot = vma->vm_page_prot;
1333 /* Can vma->vm_page_prot have changed??
1335 * Answer: Yes, drivers may have changed it in their
1336 * f_op->mmap method.
1338 * Ensures that vmas marked as uncached stay that way.
1340 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1341 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1342 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1345 vma_link(mm, vma, prev, rb_link, rb_parent);
1346 file = vma->vm_file;
1348 /* Once vma denies write, undo our temporary denial count */
1350 atomic_inc(&inode->i_writecount);
1352 perf_event_mmap(vma);
1354 mm->total_vm += len >> PAGE_SHIFT;
1355 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1356 if (vm_flags & VM_LOCKED) {
1357 if (!mlock_vma_pages_range(vma, addr, addr + len))
1358 mm->locked_vm += (len >> PAGE_SHIFT);
1359 } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1360 make_pages_present(addr, addr + len);
1365 atomic_inc(&inode->i_writecount);
1366 vma->vm_file = NULL;
1369 /* Undo any partial mapping done by a device driver. */
1370 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1373 kmem_cache_free(vm_area_cachep, vma);
1376 vm_unacct_memory(charged);
1380 /* Get an address range which is currently unmapped.
1381 * For shmat() with addr=0.
1383 * Ugly calling convention alert:
1384 * Return value with the low bits set means error value,
1386 * if (ret & ~PAGE_MASK)
1389 * This function "knows" that -ENOMEM has the bits set.
1391 #ifndef HAVE_ARCH_UNMAPPED_AREA
1393 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1394 unsigned long len, unsigned long pgoff, unsigned long flags)
1396 struct mm_struct *mm = current->mm;
1397 struct vm_area_struct *vma;
1398 unsigned long start_addr;
1400 if (len > TASK_SIZE)
1403 if (flags & MAP_FIXED)
1407 addr = PAGE_ALIGN(addr);
1408 vma = find_vma(mm, addr);
1409 if (TASK_SIZE - len >= addr &&
1410 (!vma || addr + len <= vma->vm_start))
1413 if (len > mm->cached_hole_size) {
1414 start_addr = addr = mm->free_area_cache;
1416 start_addr = addr = TASK_UNMAPPED_BASE;
1417 mm->cached_hole_size = 0;
1421 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1422 /* At this point: (!vma || addr < vma->vm_end). */
1423 if (TASK_SIZE - len < addr) {
1425 * Start a new search - just in case we missed
1428 if (start_addr != TASK_UNMAPPED_BASE) {
1429 addr = TASK_UNMAPPED_BASE;
1431 mm->cached_hole_size = 0;
1436 if (!vma || addr + len <= vma->vm_start) {
1438 * Remember the place where we stopped the search:
1440 mm->free_area_cache = addr + len;
1443 if (addr + mm->cached_hole_size < vma->vm_start)
1444 mm->cached_hole_size = vma->vm_start - addr;
1450 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1453 * Is this a new hole at the lowest possible address?
1455 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1456 mm->free_area_cache = addr;
1457 mm->cached_hole_size = ~0UL;
1462 * This mmap-allocator allocates new areas top-down from below the
1463 * stack's low limit (the base):
1465 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1467 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1468 const unsigned long len, const unsigned long pgoff,
1469 const unsigned long flags)
1471 struct vm_area_struct *vma;
1472 struct mm_struct *mm = current->mm;
1473 unsigned long addr = addr0;
1475 /* requested length too big for entire address space */
1476 if (len > TASK_SIZE)
1479 if (flags & MAP_FIXED)
1482 /* requesting a specific address */
1484 addr = PAGE_ALIGN(addr);
1485 vma = find_vma(mm, addr);
1486 if (TASK_SIZE - len >= addr &&
1487 (!vma || addr + len <= vma->vm_start))
1491 /* check if free_area_cache is useful for us */
1492 if (len <= mm->cached_hole_size) {
1493 mm->cached_hole_size = 0;
1494 mm->free_area_cache = mm->mmap_base;
1497 /* either no address requested or can't fit in requested address hole */
1498 addr = mm->free_area_cache;
1500 /* make sure it can fit in the remaining address space */
1502 vma = find_vma(mm, addr-len);
1503 if (!vma || addr <= vma->vm_start)
1504 /* remember the address as a hint for next time */
1505 return (mm->free_area_cache = addr-len);
1508 if (mm->mmap_base < len)
1511 addr = mm->mmap_base-len;
1515 * Lookup failure means no vma is above this address,
1516 * else if new region fits below vma->vm_start,
1517 * return with success:
1519 vma = find_vma(mm, addr);
1520 if (!vma || addr+len <= vma->vm_start)
1521 /* remember the address as a hint for next time */
1522 return (mm->free_area_cache = addr);
1524 /* remember the largest hole we saw so far */
1525 if (addr + mm->cached_hole_size < vma->vm_start)
1526 mm->cached_hole_size = vma->vm_start - addr;
1528 /* try just below the current vma->vm_start */
1529 addr = vma->vm_start-len;
1530 } while (len < vma->vm_start);
1534 * A failed mmap() very likely causes application failure,
1535 * so fall back to the bottom-up function here. This scenario
1536 * can happen with large stack limits and large mmap()
1539 mm->cached_hole_size = ~0UL;
1540 mm->free_area_cache = TASK_UNMAPPED_BASE;
1541 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1543 * Restore the topdown base:
1545 mm->free_area_cache = mm->mmap_base;
1546 mm->cached_hole_size = ~0UL;
1552 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1555 * Is this a new hole at the highest possible address?
1557 if (addr > mm->free_area_cache)
1558 mm->free_area_cache = addr;
1560 /* dont allow allocations above current base */
1561 if (mm->free_area_cache > mm->mmap_base)
1562 mm->free_area_cache = mm->mmap_base;
1566 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1567 unsigned long pgoff, unsigned long flags)
1569 unsigned long (*get_area)(struct file *, unsigned long,
1570 unsigned long, unsigned long, unsigned long);
1572 unsigned long error = arch_mmap_check(addr, len, flags);
1576 /* Careful about overflows.. */
1577 if (len > TASK_SIZE)
1580 get_area = current->mm->get_unmapped_area;
1581 if (file && file->f_op && file->f_op->get_unmapped_area)
1582 get_area = file->f_op->get_unmapped_area;
1583 addr = get_area(file, addr, len, pgoff, flags);
1584 if (IS_ERR_VALUE(addr))
1587 if (addr > TASK_SIZE - len)
1589 if (addr & ~PAGE_MASK)
1592 return arch_rebalance_pgtables(addr, len);
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;
1603 /* Check the cache first. */
1604 /* (Cache hit rate is typically around 35%.) */
1605 vma = mm->mmap_cache;
1606 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1607 struct rb_node * rb_node;
1609 rb_node = mm->mm_rb.rb_node;
1613 struct vm_area_struct * vma_tmp;
1615 vma_tmp = rb_entry(rb_node,
1616 struct vm_area_struct, vm_rb);
1618 if (vma_tmp->vm_end > addr) {
1620 if (vma_tmp->vm_start <= addr)
1622 rb_node = rb_node->rb_left;
1624 rb_node = rb_node->rb_right;
1627 mm->mmap_cache = vma;
1633 EXPORT_SYMBOL(find_vma);
1635 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1636 struct vm_area_struct *
1637 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1638 struct vm_area_struct **pprev)
1640 struct vm_area_struct *vma = NULL, *prev = NULL;
1641 struct rb_node *rb_node;
1645 /* Guard against addr being lower than the first VMA */
1648 /* Go through the RB tree quickly. */
1649 rb_node = mm->mm_rb.rb_node;
1652 struct vm_area_struct *vma_tmp;
1653 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1655 if (addr < vma_tmp->vm_end) {
1656 rb_node = rb_node->rb_left;
1659 if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1661 rb_node = rb_node->rb_right;
1667 return prev ? prev->vm_next : vma;
1671 * Verify that the stack growth is acceptable and
1672 * update accounting. This is shared with both the
1673 * grow-up and grow-down cases.
1675 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1677 struct mm_struct *mm = vma->vm_mm;
1678 struct rlimit *rlim = current->signal->rlim;
1679 unsigned long new_start;
1681 /* address space limit tests */
1682 if (!may_expand_vm(mm, grow))
1685 /* Stack limit test */
1686 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1689 /* mlock limit tests */
1690 if (vma->vm_flags & VM_LOCKED) {
1691 unsigned long locked;
1692 unsigned long limit;
1693 locked = mm->locked_vm + grow;
1694 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1695 limit >>= PAGE_SHIFT;
1696 if (locked > limit && !capable(CAP_IPC_LOCK))
1700 /* Check to ensure the stack will not grow into a hugetlb-only region */
1701 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1703 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1707 * Overcommit.. This must be the final test, as it will
1708 * update security statistics.
1710 if (security_vm_enough_memory_mm(mm, grow))
1713 /* Ok, everything looks good - let it rip */
1714 mm->total_vm += grow;
1715 if (vma->vm_flags & VM_LOCKED)
1716 mm->locked_vm += grow;
1717 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1721 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1723 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1724 * vma is the last one with address > vma->vm_end. Have to extend vma.
1726 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1730 if (!(vma->vm_flags & VM_GROWSUP))
1734 * We must make sure the anon_vma is allocated
1735 * so that the anon_vma locking is not a noop.
1737 if (unlikely(anon_vma_prepare(vma)))
1739 vma_lock_anon_vma(vma);
1742 * vma->vm_start/vm_end cannot change under us because the caller
1743 * is required to hold the mmap_sem in read mode. We need the
1744 * anon_vma lock to serialize against concurrent expand_stacks.
1745 * Also guard against wrapping around to address 0.
1747 if (address < PAGE_ALIGN(address+4))
1748 address = PAGE_ALIGN(address+4);
1750 vma_unlock_anon_vma(vma);
1755 /* Somebody else might have raced and expanded it already */
1756 if (address > vma->vm_end) {
1757 unsigned long size, grow;
1759 size = address - vma->vm_start;
1760 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1762 error = acct_stack_growth(vma, size, grow);
1764 vma->vm_end = address;
1765 perf_event_mmap(vma);
1768 vma_unlock_anon_vma(vma);
1769 khugepaged_enter_vma_merge(vma);
1772 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1775 * vma is the first one with address < vma->vm_start. Have to extend vma.
1777 static int expand_downwards(struct vm_area_struct *vma,
1778 unsigned long address)
1783 * We must make sure the anon_vma is allocated
1784 * so that the anon_vma locking is not a noop.
1786 if (unlikely(anon_vma_prepare(vma)))
1789 address &= PAGE_MASK;
1790 error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1794 vma_lock_anon_vma(vma);
1797 * vma->vm_start/vm_end cannot change under us because the caller
1798 * is required to hold the mmap_sem in read mode. We need the
1799 * anon_vma lock to serialize against concurrent expand_stacks.
1802 /* Somebody else might have raced and expanded it already */
1803 if (address < vma->vm_start) {
1804 unsigned long size, grow;
1806 size = vma->vm_end - address;
1807 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1809 error = acct_stack_growth(vma, size, grow);
1811 vma->vm_start = address;
1812 vma->vm_pgoff -= grow;
1813 perf_event_mmap(vma);
1816 vma_unlock_anon_vma(vma);
1817 khugepaged_enter_vma_merge(vma);
1821 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1823 return expand_downwards(vma, address);
1826 #ifdef CONFIG_STACK_GROWSUP
1827 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1829 return expand_upwards(vma, address);
1832 struct vm_area_struct *
1833 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1835 struct vm_area_struct *vma, *prev;
1838 vma = find_vma_prev(mm, addr, &prev);
1839 if (vma && (vma->vm_start <= addr))
1841 if (!prev || expand_stack(prev, addr))
1843 if (prev->vm_flags & VM_LOCKED) {
1844 mlock_vma_pages_range(prev, addr, prev->vm_end);
1849 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1851 return expand_downwards(vma, address);
1854 struct vm_area_struct *
1855 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1857 struct vm_area_struct * vma;
1858 unsigned long start;
1861 vma = find_vma(mm,addr);
1864 if (vma->vm_start <= addr)
1866 if (!(vma->vm_flags & VM_GROWSDOWN))
1868 start = vma->vm_start;
1869 if (expand_stack(vma, addr))
1871 if (vma->vm_flags & VM_LOCKED) {
1872 mlock_vma_pages_range(vma, addr, start);
1879 * Ok - we have the memory areas we should free on the vma list,
1880 * so release them, and do the vma updates.
1882 * Called with the mm semaphore held.
1884 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1886 /* Update high watermark before we lower total_vm */
1887 update_hiwater_vm(mm);
1889 long nrpages = vma_pages(vma);
1891 mm->total_vm -= nrpages;
1892 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1893 vma = remove_vma(vma);
1899 * Get rid of page table information in the indicated region.
1901 * Called with the mm semaphore held.
1903 static void unmap_region(struct mm_struct *mm,
1904 struct vm_area_struct *vma, struct vm_area_struct *prev,
1905 unsigned long start, unsigned long end)
1907 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1908 struct mmu_gather *tlb;
1909 unsigned long nr_accounted = 0;
1912 tlb = tlb_gather_mmu(mm, 0);
1913 update_hiwater_rss(mm);
1914 unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1915 vm_unacct_memory(nr_accounted);
1916 free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1917 next? next->vm_start: 0);
1918 tlb_finish_mmu(tlb, start, end);
1922 * Create a list of vma's touched by the unmap, removing them from the mm's
1923 * vma list as we go..
1926 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1927 struct vm_area_struct *prev, unsigned long end)
1929 struct vm_area_struct **insertion_point;
1930 struct vm_area_struct *tail_vma = NULL;
1933 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1934 vma->vm_prev = NULL;
1936 rb_erase(&vma->vm_rb, &mm->mm_rb);
1940 } while (vma && vma->vm_start < end);
1941 *insertion_point = vma;
1943 vma->vm_prev = prev;
1944 tail_vma->vm_next = NULL;
1945 if (mm->unmap_area == arch_unmap_area)
1946 addr = prev ? prev->vm_end : mm->mmap_base;
1948 addr = vma ? vma->vm_start : mm->mmap_base;
1949 mm->unmap_area(mm, addr);
1950 mm->mmap_cache = NULL; /* Kill the cache. */
1954 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
1955 * munmap path where it doesn't make sense to fail.
1957 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1958 unsigned long addr, int new_below)
1960 struct mempolicy *pol;
1961 struct vm_area_struct *new;
1964 if (is_vm_hugetlb_page(vma) && (addr &
1965 ~(huge_page_mask(hstate_vma(vma)))))
1968 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1972 /* most fields are the same, copy all, and then fixup */
1975 INIT_LIST_HEAD(&new->anon_vma_chain);
1980 new->vm_start = addr;
1981 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1984 pol = mpol_dup(vma_policy(vma));
1989 vma_set_policy(new, pol);
1991 if (anon_vma_clone(new, vma))
1995 get_file(new->vm_file);
1996 if (vma->vm_flags & VM_EXECUTABLE)
1997 added_exe_file_vma(mm);
2000 if (new->vm_ops && new->vm_ops->open)
2001 new->vm_ops->open(new);
2004 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2005 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2007 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2013 /* Clean everything up if vma_adjust failed. */
2014 if (new->vm_ops && new->vm_ops->close)
2015 new->vm_ops->close(new);
2017 if (vma->vm_flags & VM_EXECUTABLE)
2018 removed_exe_file_vma(mm);
2021 unlink_anon_vmas(new);
2025 kmem_cache_free(vm_area_cachep, new);
2031 * Split a vma into two pieces at address 'addr', a new vma is allocated
2032 * either for the first part or the tail.
2034 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2035 unsigned long addr, int new_below)
2037 if (mm->map_count >= sysctl_max_map_count)
2040 return __split_vma(mm, vma, addr, new_below);
2043 /* Munmap is split into 2 main parts -- this part which finds
2044 * what needs doing, and the areas themselves, which do the
2045 * work. This now handles partial unmappings.
2046 * Jeremy Fitzhardinge <jeremy@goop.org>
2048 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2051 struct vm_area_struct *vma, *prev, *last;
2053 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2056 if ((len = PAGE_ALIGN(len)) == 0)
2059 /* Find the first overlapping VMA */
2060 vma = find_vma_prev(mm, start, &prev);
2063 /* we have start < vma->vm_end */
2065 /* if it doesn't overlap, we have nothing.. */
2067 if (vma->vm_start >= end)
2071 * If we need to split any vma, do it now to save pain later.
2073 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2074 * unmapped vm_area_struct will remain in use: so lower split_vma
2075 * places tmp vma above, and higher split_vma places tmp vma below.
2077 if (start > vma->vm_start) {
2081 * Make sure that map_count on return from munmap() will
2082 * not exceed its limit; but let map_count go just above
2083 * its limit temporarily, to help free resources as expected.
2085 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2088 error = __split_vma(mm, vma, start, 0);
2094 /* Does it split the last one? */
2095 last = find_vma(mm, end);
2096 if (last && end > last->vm_start) {
2097 int error = __split_vma(mm, last, end, 1);
2101 vma = prev? prev->vm_next: mm->mmap;
2104 * unlock any mlock()ed ranges before detaching vmas
2106 if (mm->locked_vm) {
2107 struct vm_area_struct *tmp = vma;
2108 while (tmp && tmp->vm_start < end) {
2109 if (tmp->vm_flags & VM_LOCKED) {
2110 mm->locked_vm -= vma_pages(tmp);
2111 munlock_vma_pages_all(tmp);
2118 * Remove the vma's, and unmap the actual pages
2120 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2121 unmap_region(mm, vma, prev, start, end);
2123 /* Fix up all other VM information */
2124 remove_vma_list(mm, vma);
2129 EXPORT_SYMBOL(do_munmap);
2131 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2134 struct mm_struct *mm = current->mm;
2136 profile_munmap(addr);
2138 down_write(&mm->mmap_sem);
2139 ret = do_munmap(mm, addr, len);
2140 up_write(&mm->mmap_sem);
2144 static inline void verify_mm_writelocked(struct mm_struct *mm)
2146 #ifdef CONFIG_DEBUG_VM
2147 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2149 up_read(&mm->mmap_sem);
2155 * this is really a simplified "do_mmap". it only handles
2156 * anonymous maps. eventually we may be able to do some
2157 * brk-specific accounting here.
2159 unsigned long do_brk(unsigned long addr, unsigned long len)
2161 struct mm_struct * mm = current->mm;
2162 struct vm_area_struct * vma, * prev;
2163 unsigned long flags;
2164 struct rb_node ** rb_link, * rb_parent;
2165 pgoff_t pgoff = addr >> PAGE_SHIFT;
2168 len = PAGE_ALIGN(len);
2172 error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2176 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2178 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2179 if (error & ~PAGE_MASK)
2185 if (mm->def_flags & VM_LOCKED) {
2186 unsigned long locked, lock_limit;
2187 locked = len >> PAGE_SHIFT;
2188 locked += mm->locked_vm;
2189 lock_limit = rlimit(RLIMIT_MEMLOCK);
2190 lock_limit >>= PAGE_SHIFT;
2191 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2196 * mm->mmap_sem is required to protect against another thread
2197 * changing the mappings in case we sleep.
2199 verify_mm_writelocked(mm);
2202 * Clear old maps. this also does some error checking for us
2205 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2206 if (vma && vma->vm_start < addr + len) {
2207 if (do_munmap(mm, addr, len))
2212 /* Check against address space limits *after* clearing old maps... */
2213 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2216 if (mm->map_count > sysctl_max_map_count)
2219 if (security_vm_enough_memory(len >> PAGE_SHIFT))
2222 /* Can we just expand an old private anonymous mapping? */
2223 vma = vma_merge(mm, prev, addr, addr + len, flags,
2224 NULL, NULL, pgoff, NULL);
2229 * create a vma struct for an anonymous mapping
2231 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2233 vm_unacct_memory(len >> PAGE_SHIFT);
2237 INIT_LIST_HEAD(&vma->anon_vma_chain);
2239 vma->vm_start = addr;
2240 vma->vm_end = addr + len;
2241 vma->vm_pgoff = pgoff;
2242 vma->vm_flags = flags;
2243 vma->vm_page_prot = vm_get_page_prot(flags);
2244 vma_link(mm, vma, prev, rb_link, rb_parent);
2246 perf_event_mmap(vma);
2247 mm->total_vm += len >> PAGE_SHIFT;
2248 if (flags & VM_LOCKED) {
2249 if (!mlock_vma_pages_range(vma, addr, addr + len))
2250 mm->locked_vm += (len >> PAGE_SHIFT);
2255 EXPORT_SYMBOL(do_brk);
2257 /* Release all mmaps. */
2258 void exit_mmap(struct mm_struct *mm)
2260 struct mmu_gather *tlb;
2261 struct vm_area_struct *vma;
2262 unsigned long nr_accounted = 0;
2265 /* mm's last user has gone, and its about to be pulled down */
2266 mmu_notifier_release(mm);
2268 if (mm->locked_vm) {
2271 if (vma->vm_flags & VM_LOCKED)
2272 munlock_vma_pages_all(vma);
2280 if (!vma) /* Can happen if dup_mmap() received an OOM */
2285 tlb = tlb_gather_mmu(mm, 1);
2286 /* update_hiwater_rss(mm) here? but nobody should be looking */
2287 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2288 end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2289 vm_unacct_memory(nr_accounted);
2291 free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2292 tlb_finish_mmu(tlb, 0, end);
2295 * Walk the list again, actually closing and freeing it,
2296 * with preemption enabled, without holding any MM locks.
2299 vma = remove_vma(vma);
2301 BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2304 /* Insert vm structure into process list sorted by address
2305 * and into the inode's i_mmap tree. If vm_file is non-NULL
2306 * then i_mmap_lock is taken here.
2308 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2310 struct vm_area_struct * __vma, * prev;
2311 struct rb_node ** rb_link, * rb_parent;
2314 * The vm_pgoff of a purely anonymous vma should be irrelevant
2315 * until its first write fault, when page's anon_vma and index
2316 * are set. But now set the vm_pgoff it will almost certainly
2317 * end up with (unless mremap moves it elsewhere before that
2318 * first wfault), so /proc/pid/maps tells a consistent story.
2320 * By setting it to reflect the virtual start address of the
2321 * vma, merges and splits can happen in a seamless way, just
2322 * using the existing file pgoff checks and manipulations.
2323 * Similarly in do_mmap_pgoff and in do_brk.
2325 if (!vma->vm_file) {
2326 BUG_ON(vma->anon_vma);
2327 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2329 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2330 if (__vma && __vma->vm_start < vma->vm_end)
2332 if ((vma->vm_flags & VM_ACCOUNT) &&
2333 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2335 vma_link(mm, vma, prev, rb_link, rb_parent);
2340 * Copy the vma structure to a new location in the same mm,
2341 * prior to moving page table entries, to effect an mremap move.
2343 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2344 unsigned long addr, unsigned long len, pgoff_t pgoff)
2346 struct vm_area_struct *vma = *vmap;
2347 unsigned long vma_start = vma->vm_start;
2348 struct mm_struct *mm = vma->vm_mm;
2349 struct vm_area_struct *new_vma, *prev;
2350 struct rb_node **rb_link, *rb_parent;
2351 struct mempolicy *pol;
2354 * If anonymous vma has not yet been faulted, update new pgoff
2355 * to match new location, to increase its chance of merging.
2357 if (!vma->vm_file && !vma->anon_vma)
2358 pgoff = addr >> PAGE_SHIFT;
2360 find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2361 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2362 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2365 * Source vma may have been merged into new_vma
2367 if (vma_start >= new_vma->vm_start &&
2368 vma_start < new_vma->vm_end)
2371 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2374 pol = mpol_dup(vma_policy(vma));
2377 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2378 if (anon_vma_clone(new_vma, vma))
2379 goto out_free_mempol;
2380 vma_set_policy(new_vma, pol);
2381 new_vma->vm_start = addr;
2382 new_vma->vm_end = addr + len;
2383 new_vma->vm_pgoff = pgoff;
2384 if (new_vma->vm_file) {
2385 get_file(new_vma->vm_file);
2386 if (vma->vm_flags & VM_EXECUTABLE)
2387 added_exe_file_vma(mm);
2389 if (new_vma->vm_ops && new_vma->vm_ops->open)
2390 new_vma->vm_ops->open(new_vma);
2391 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2399 kmem_cache_free(vm_area_cachep, new_vma);
2404 * Return true if the calling process may expand its vm space by the passed
2407 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2409 unsigned long cur = mm->total_vm; /* pages */
2412 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2414 if (cur + npages > lim)
2420 static int special_mapping_fault(struct vm_area_struct *vma,
2421 struct vm_fault *vmf)
2424 struct page **pages;
2427 * special mappings have no vm_file, and in that case, the mm
2428 * uses vm_pgoff internally. So we have to subtract it from here.
2429 * We are allowed to do this because we are the mm; do not copy
2430 * this code into drivers!
2432 pgoff = vmf->pgoff - vma->vm_pgoff;
2434 for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2438 struct page *page = *pages;
2444 return VM_FAULT_SIGBUS;
2448 * Having a close hook prevents vma merging regardless of flags.
2450 static void special_mapping_close(struct vm_area_struct *vma)
2454 static const struct vm_operations_struct special_mapping_vmops = {
2455 .close = special_mapping_close,
2456 .fault = special_mapping_fault,
2460 * Called with mm->mmap_sem held for writing.
2461 * Insert a new vma covering the given region, with the given flags.
2462 * Its pages are supplied by the given array of struct page *.
2463 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2464 * The region past the last page supplied will always produce SIGBUS.
2465 * The array pointer and the pages it points to are assumed to stay alive
2466 * for as long as this mapping might exist.
2468 int install_special_mapping(struct mm_struct *mm,
2469 unsigned long addr, unsigned long len,
2470 unsigned long vm_flags, struct page **pages)
2473 struct vm_area_struct *vma;
2475 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2476 if (unlikely(vma == NULL))
2479 INIT_LIST_HEAD(&vma->anon_vma_chain);
2481 vma->vm_start = addr;
2482 vma->vm_end = addr + len;
2484 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2485 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2487 vma->vm_ops = &special_mapping_vmops;
2488 vma->vm_private_data = pages;
2490 ret = security_file_mmap(NULL, 0, 0, 0, vma->vm_start, 1);
2494 ret = insert_vm_struct(mm, vma);
2498 mm->total_vm += len >> PAGE_SHIFT;
2500 perf_event_mmap(vma);
2505 kmem_cache_free(vm_area_cachep, vma);
2509 static DEFINE_MUTEX(mm_all_locks_mutex);
2511 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2513 if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2515 * The LSB of head.next can't change from under us
2516 * because we hold the mm_all_locks_mutex.
2518 spin_lock_nest_lock(&anon_vma->root->lock, &mm->mmap_sem);
2520 * We can safely modify head.next after taking the
2521 * anon_vma->root->lock. If some other vma in this mm shares
2522 * the same anon_vma we won't take it again.
2524 * No need of atomic instructions here, head.next
2525 * can't change from under us thanks to the
2526 * anon_vma->root->lock.
2528 if (__test_and_set_bit(0, (unsigned long *)
2529 &anon_vma->root->head.next))
2534 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2536 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2538 * AS_MM_ALL_LOCKS can't change from under us because
2539 * we hold the mm_all_locks_mutex.
2541 * Operations on ->flags have to be atomic because
2542 * even if AS_MM_ALL_LOCKS is stable thanks to the
2543 * mm_all_locks_mutex, there may be other cpus
2544 * changing other bitflags in parallel to us.
2546 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2548 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2553 * This operation locks against the VM for all pte/vma/mm related
2554 * operations that could ever happen on a certain mm. This includes
2555 * vmtruncate, try_to_unmap, and all page faults.
2557 * The caller must take the mmap_sem in write mode before calling
2558 * mm_take_all_locks(). The caller isn't allowed to release the
2559 * mmap_sem until mm_drop_all_locks() returns.
2561 * mmap_sem in write mode is required in order to block all operations
2562 * that could modify pagetables and free pages without need of
2563 * altering the vma layout (for example populate_range() with
2564 * nonlinear vmas). It's also needed in write mode to avoid new
2565 * anon_vmas to be associated with existing vmas.
2567 * A single task can't take more than one mm_take_all_locks() in a row
2568 * or it would deadlock.
2570 * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2571 * mapping->flags avoid to take the same lock twice, if more than one
2572 * vma in this mm is backed by the same anon_vma or address_space.
2574 * We can take all the locks in random order because the VM code
2575 * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2576 * takes more than one of them in a row. Secondly we're protected
2577 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2579 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2580 * that may have to take thousand of locks.
2582 * mm_take_all_locks() can fail if it's interrupted by signals.
2584 int mm_take_all_locks(struct mm_struct *mm)
2586 struct vm_area_struct *vma;
2587 struct anon_vma_chain *avc;
2590 BUG_ON(down_read_trylock(&mm->mmap_sem));
2592 mutex_lock(&mm_all_locks_mutex);
2594 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2595 if (signal_pending(current))
2597 if (vma->vm_file && vma->vm_file->f_mapping)
2598 vm_lock_mapping(mm, vma->vm_file->f_mapping);
2601 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2602 if (signal_pending(current))
2605 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2606 vm_lock_anon_vma(mm, avc->anon_vma);
2613 mm_drop_all_locks(mm);
2618 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2620 if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2622 * The LSB of head.next can't change to 0 from under
2623 * us because we hold the mm_all_locks_mutex.
2625 * We must however clear the bitflag before unlocking
2626 * the vma so the users using the anon_vma->head will
2627 * never see our bitflag.
2629 * No need of atomic instructions here, head.next
2630 * can't change from under us until we release the
2631 * anon_vma->root->lock.
2633 if (!__test_and_clear_bit(0, (unsigned long *)
2634 &anon_vma->root->head.next))
2636 anon_vma_unlock(anon_vma);
2640 static void vm_unlock_mapping(struct address_space *mapping)
2642 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2644 * AS_MM_ALL_LOCKS can't change to 0 from under us
2645 * because we hold the mm_all_locks_mutex.
2647 spin_unlock(&mapping->i_mmap_lock);
2648 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2655 * The mmap_sem cannot be released by the caller until
2656 * mm_drop_all_locks() returns.
2658 void mm_drop_all_locks(struct mm_struct *mm)
2660 struct vm_area_struct *vma;
2661 struct anon_vma_chain *avc;
2663 BUG_ON(down_read_trylock(&mm->mmap_sem));
2664 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2666 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2668 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2669 vm_unlock_anon_vma(avc->anon_vma);
2670 if (vma->vm_file && vma->vm_file->f_mapping)
2671 vm_unlock_mapping(vma->vm_file->f_mapping);
2674 mutex_unlock(&mm_all_locks_mutex);
2678 * initialise the VMA slab
2680 void __init mmap_init(void)
2684 ret = percpu_counter_init(&vm_committed_as, 0);