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>
32 #include <asm/uaccess.h>
33 #include <asm/cacheflush.h>
35 #include <asm/mmu_context.h>
39 #ifndef arch_mmap_check
40 #define arch_mmap_check(addr, len, flags) (0)
43 #ifndef arch_rebalance_pgtables
44 #define arch_rebalance_pgtables(addr, len) (addr)
47 static void unmap_region(struct mm_struct *mm,
48 struct vm_area_struct *vma, struct vm_area_struct *prev,
49 unsigned long start, unsigned long end);
52 * WARNING: the debugging will use recursive algorithms so never enable this
53 * unless you know what you are doing.
57 /* description of effects of mapping type and prot in current implementation.
58 * this is due to the limited x86 page protection hardware. The expected
59 * behavior is in parens:
62 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
63 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
64 * w: (no) no w: (no) no w: (yes) yes w: (no) no
65 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
67 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
68 * w: (no) no w: (no) no w: (copy) copy w: (no) no
69 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
72 pgprot_t protection_map[16] = {
73 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
74 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
77 pgprot_t vm_get_page_prot(unsigned long vm_flags)
79 return __pgprot(pgprot_val(protection_map[vm_flags &
80 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
81 pgprot_val(arch_vm_get_page_prot(vm_flags)));
83 EXPORT_SYMBOL(vm_get_page_prot);
85 int sysctl_overcommit_memory = OVERCOMMIT_GUESS; /* heuristic overcommit */
86 int sysctl_overcommit_ratio = 50; /* default is 50% */
87 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
88 struct percpu_counter vm_committed_as;
91 * Check that a process has enough memory to allocate a new virtual
92 * mapping. 0 means there is enough memory for the allocation to
93 * succeed and -ENOMEM implies there is not.
95 * We currently support three overcommit policies, which are set via the
96 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
98 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
99 * Additional code 2002 Jul 20 by Robert Love.
101 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
103 * Note this is a helper function intended to be used by LSMs which
104 * wish to use this logic.
106 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
108 unsigned long free, allowed;
110 vm_acct_memory(pages);
113 * Sometimes we want to use more memory than we have
115 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
118 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
121 free = global_page_state(NR_FILE_PAGES);
122 free += nr_swap_pages;
125 * Any slabs which are created with the
126 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
127 * which are reclaimable, under pressure. The dentry
128 * cache and most inode caches should fall into this
130 free += global_page_state(NR_SLAB_RECLAIMABLE);
133 * Leave the last 3% for root
142 * nr_free_pages() is very expensive on large systems,
143 * only call if we're about to fail.
148 * Leave reserved pages. The pages are not for anonymous pages.
150 if (n <= totalreserve_pages)
153 n -= totalreserve_pages;
156 * Leave the last 3% for root
168 allowed = (totalram_pages - hugetlb_total_pages())
169 * sysctl_overcommit_ratio / 100;
171 * Leave the last 3% for root
174 allowed -= allowed / 32;
175 allowed += total_swap_pages;
177 /* Don't let a single process grow too big:
178 leave 3% of the size of this process for other processes */
180 allowed -= mm->total_vm / 32;
182 if (percpu_counter_read_positive(&vm_committed_as) < allowed)
185 vm_unacct_memory(pages);
191 * Requires inode->i_mapping->i_mmap_lock
193 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
194 struct file *file, struct address_space *mapping)
196 if (vma->vm_flags & VM_DENYWRITE)
197 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
198 if (vma->vm_flags & VM_SHARED)
199 mapping->i_mmap_writable--;
201 flush_dcache_mmap_lock(mapping);
202 if (unlikely(vma->vm_flags & VM_NONLINEAR))
203 list_del_init(&vma->shared.vm_set.list);
205 vma_prio_tree_remove(vma, &mapping->i_mmap);
206 flush_dcache_mmap_unlock(mapping);
210 * Unlink a file-based vm structure from its prio_tree, to hide
211 * vma from rmap and vmtruncate before freeing its page tables.
213 void unlink_file_vma(struct vm_area_struct *vma)
215 struct file *file = vma->vm_file;
218 struct address_space *mapping = file->f_mapping;
219 spin_lock(&mapping->i_mmap_lock);
220 __remove_shared_vm_struct(vma, file, mapping);
221 spin_unlock(&mapping->i_mmap_lock);
226 * Close a vm structure and free it, returning the next.
228 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
230 struct vm_area_struct *next = vma->vm_next;
233 if (vma->vm_ops && vma->vm_ops->close)
234 vma->vm_ops->close(vma);
237 if (vma->vm_flags & VM_EXECUTABLE)
238 removed_exe_file_vma(vma->vm_mm);
240 mpol_put(vma_policy(vma));
241 kmem_cache_free(vm_area_cachep, vma);
245 SYSCALL_DEFINE1(brk, unsigned long, brk)
247 unsigned long rlim, retval;
248 unsigned long newbrk, oldbrk;
249 struct mm_struct *mm = current->mm;
250 unsigned long min_brk;
252 down_write(&mm->mmap_sem);
254 #ifdef CONFIG_COMPAT_BRK
255 min_brk = mm->end_code;
257 min_brk = mm->start_brk;
263 * Check against rlimit here. If this check is done later after the test
264 * of oldbrk with newbrk then it can escape the test and let the data
265 * segment grow beyond its set limit the in case where the limit is
266 * not page aligned -Ram Gupta
268 rlim = rlimit(RLIMIT_DATA);
269 if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
270 (mm->end_data - mm->start_data) > rlim)
273 newbrk = PAGE_ALIGN(brk);
274 oldbrk = PAGE_ALIGN(mm->brk);
275 if (oldbrk == newbrk)
278 /* Always allow shrinking brk. */
279 if (brk <= mm->brk) {
280 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
285 /* Check against existing mmap mappings. */
286 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
289 /* Ok, looks good - let it rip. */
290 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
296 up_write(&mm->mmap_sem);
301 static int browse_rb(struct rb_root *root)
304 struct rb_node *nd, *pn = NULL;
305 unsigned long prev = 0, pend = 0;
307 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
308 struct vm_area_struct *vma;
309 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
310 if (vma->vm_start < prev)
311 printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
312 if (vma->vm_start < pend)
313 printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
314 if (vma->vm_start > vma->vm_end)
315 printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
318 prev = vma->vm_start;
322 for (nd = pn; nd; nd = rb_prev(nd)) {
326 printk("backwards %d, forwards %d\n", j, i), i = 0;
330 void validate_mm(struct mm_struct *mm)
334 struct vm_area_struct *tmp = mm->mmap;
339 if (i != mm->map_count)
340 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
341 i = browse_rb(&mm->mm_rb);
342 if (i != mm->map_count)
343 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
347 #define validate_mm(mm) do { } while (0)
350 static struct vm_area_struct *
351 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
352 struct vm_area_struct **pprev, struct rb_node ***rb_link,
353 struct rb_node ** rb_parent)
355 struct vm_area_struct * vma;
356 struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
358 __rb_link = &mm->mm_rb.rb_node;
359 rb_prev = __rb_parent = NULL;
363 struct vm_area_struct *vma_tmp;
365 __rb_parent = *__rb_link;
366 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
368 if (vma_tmp->vm_end > addr) {
370 if (vma_tmp->vm_start <= addr)
372 __rb_link = &__rb_parent->rb_left;
374 rb_prev = __rb_parent;
375 __rb_link = &__rb_parent->rb_right;
381 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
382 *rb_link = __rb_link;
383 *rb_parent = __rb_parent;
388 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
389 struct vm_area_struct *prev, struct rb_node *rb_parent)
391 struct vm_area_struct *next;
395 next = prev->vm_next;
400 next = rb_entry(rb_parent,
401 struct vm_area_struct, vm_rb);
410 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
411 struct rb_node **rb_link, struct rb_node *rb_parent)
413 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
414 rb_insert_color(&vma->vm_rb, &mm->mm_rb);
417 static void __vma_link_file(struct vm_area_struct *vma)
423 struct address_space *mapping = file->f_mapping;
425 if (vma->vm_flags & VM_DENYWRITE)
426 atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
427 if (vma->vm_flags & VM_SHARED)
428 mapping->i_mmap_writable++;
430 flush_dcache_mmap_lock(mapping);
431 if (unlikely(vma->vm_flags & VM_NONLINEAR))
432 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
434 vma_prio_tree_insert(vma, &mapping->i_mmap);
435 flush_dcache_mmap_unlock(mapping);
440 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
441 struct vm_area_struct *prev, struct rb_node **rb_link,
442 struct rb_node *rb_parent)
444 __vma_link_list(mm, vma, prev, rb_parent);
445 __vma_link_rb(mm, vma, rb_link, rb_parent);
448 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
449 struct vm_area_struct *prev, struct rb_node **rb_link,
450 struct rb_node *rb_parent)
452 struct address_space *mapping = NULL;
455 mapping = vma->vm_file->f_mapping;
458 spin_lock(&mapping->i_mmap_lock);
459 vma->vm_truncate_count = mapping->truncate_count;
462 __vma_link(mm, vma, prev, rb_link, rb_parent);
463 __vma_link_file(vma);
466 spin_unlock(&mapping->i_mmap_lock);
473 * Helper for vma_adjust in the split_vma insert case:
474 * insert vm structure into list and rbtree and anon_vma,
475 * but it has already been inserted into prio_tree earlier.
477 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
479 struct vm_area_struct *__vma, *prev;
480 struct rb_node **rb_link, *rb_parent;
482 __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
483 BUG_ON(__vma && __vma->vm_start < vma->vm_end);
484 __vma_link(mm, vma, prev, rb_link, rb_parent);
489 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
490 struct vm_area_struct *prev)
492 struct vm_area_struct *next = vma->vm_next;
494 prev->vm_next = next;
496 next->vm_prev = prev;
497 rb_erase(&vma->vm_rb, &mm->mm_rb);
498 if (mm->mmap_cache == vma)
499 mm->mmap_cache = prev;
503 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
504 * is already present in an i_mmap tree without adjusting the tree.
505 * The following helper function should be used when such adjustments
506 * are necessary. The "insert" vma (if any) is to be inserted
507 * before we drop the necessary locks.
509 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
510 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
512 struct mm_struct *mm = vma->vm_mm;
513 struct vm_area_struct *next = vma->vm_next;
514 struct vm_area_struct *importer = NULL;
515 struct address_space *mapping = NULL;
516 struct prio_tree_root *root = NULL;
517 struct anon_vma *anon_vma = NULL;
518 struct file *file = vma->vm_file;
519 long adjust_next = 0;
522 if (next && !insert) {
523 struct vm_area_struct *exporter = NULL;
525 if (end >= next->vm_end) {
527 * vma expands, overlapping all the next, and
528 * perhaps the one after too (mprotect case 6).
530 again: remove_next = 1 + (end > next->vm_end);
534 } else if (end > next->vm_start) {
536 * vma expands, overlapping part of the next:
537 * mprotect case 5 shifting the boundary up.
539 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
542 } else if (end < vma->vm_end) {
544 * vma shrinks, and !insert tells it's not
545 * split_vma inserting another: so it must be
546 * mprotect case 4 shifting the boundary down.
548 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
554 * Easily overlooked: when mprotect shifts the boundary,
555 * make sure the expanding vma has anon_vma set if the
556 * shrinking vma had, to cover any anon pages imported.
558 if (exporter && exporter->anon_vma && !importer->anon_vma) {
559 if (anon_vma_clone(importer, exporter))
561 importer->anon_vma = exporter->anon_vma;
566 mapping = file->f_mapping;
567 if (!(vma->vm_flags & VM_NONLINEAR))
568 root = &mapping->i_mmap;
569 spin_lock(&mapping->i_mmap_lock);
571 vma->vm_truncate_count != next->vm_truncate_count) {
573 * unmap_mapping_range might be in progress:
574 * ensure that the expanding vma is rescanned.
576 importer->vm_truncate_count = 0;
579 insert->vm_truncate_count = vma->vm_truncate_count;
581 * Put into prio_tree now, so instantiated pages
582 * are visible to arm/parisc __flush_dcache_page
583 * throughout; but we cannot insert into address
584 * space until vma start or end is updated.
586 __vma_link_file(insert);
591 * When changing only vma->vm_end, we don't really need anon_vma
592 * lock. This is a fairly rare case by itself, but the anon_vma
593 * lock may be shared between many sibling processes. Skipping
594 * the lock for brk adjustments makes a difference sometimes.
596 if (vma->anon_vma && (insert || importer || start != vma->vm_start)) {
597 anon_vma = vma->anon_vma;
598 anon_vma_lock(anon_vma);
602 flush_dcache_mmap_lock(mapping);
603 vma_prio_tree_remove(vma, root);
605 vma_prio_tree_remove(next, root);
608 vma->vm_start = start;
610 vma->vm_pgoff = pgoff;
612 next->vm_start += adjust_next << PAGE_SHIFT;
613 next->vm_pgoff += adjust_next;
618 vma_prio_tree_insert(next, root);
619 vma_prio_tree_insert(vma, root);
620 flush_dcache_mmap_unlock(mapping);
625 * vma_merge has merged next into vma, and needs
626 * us to remove next before dropping the locks.
628 __vma_unlink(mm, next, vma);
630 __remove_shared_vm_struct(next, file, mapping);
633 * split_vma has split insert from vma, and needs
634 * us to insert it before dropping the locks
635 * (it may either follow vma or precede it).
637 __insert_vm_struct(mm, insert);
641 anon_vma_unlock(anon_vma);
643 spin_unlock(&mapping->i_mmap_lock);
648 if (next->vm_flags & VM_EXECUTABLE)
649 removed_exe_file_vma(mm);
652 anon_vma_merge(vma, next);
654 mpol_put(vma_policy(next));
655 kmem_cache_free(vm_area_cachep, next);
657 * In mprotect's case 6 (see comments on vma_merge),
658 * we must remove another next too. It would clutter
659 * up the code too much to do both in one go.
661 if (remove_next == 2) {
673 * If the vma has a ->close operation then the driver probably needs to release
674 * per-vma resources, so we don't attempt to merge those.
676 static inline int is_mergeable_vma(struct vm_area_struct *vma,
677 struct file *file, unsigned long vm_flags)
679 /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
680 if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
682 if (vma->vm_file != file)
684 if (vma->vm_ops && vma->vm_ops->close)
689 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
690 struct anon_vma *anon_vma2)
692 return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
696 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
697 * in front of (at a lower virtual address and file offset than) the vma.
699 * We cannot merge two vmas if they have differently assigned (non-NULL)
700 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
702 * We don't check here for the merged mmap wrapping around the end of pagecache
703 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
704 * wrap, nor mmaps which cover the final page at index -1UL.
707 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
708 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
710 if (is_mergeable_vma(vma, file, vm_flags) &&
711 is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
712 if (vma->vm_pgoff == vm_pgoff)
719 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
720 * beyond (at a higher virtual address and file offset than) the vma.
722 * We cannot merge two vmas if they have differently assigned (non-NULL)
723 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
726 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
727 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
729 if (is_mergeable_vma(vma, file, vm_flags) &&
730 is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
732 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
733 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
740 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
741 * whether that can be merged with its predecessor or its successor.
742 * Or both (it neatly fills a hole).
744 * In most cases - when called for mmap, brk or mremap - [addr,end) is
745 * certain not to be mapped by the time vma_merge is called; but when
746 * called for mprotect, it is certain to be already mapped (either at
747 * an offset within prev, or at the start of next), and the flags of
748 * this area are about to be changed to vm_flags - and the no-change
749 * case has already been eliminated.
751 * The following mprotect cases have to be considered, where AAAA is
752 * the area passed down from mprotect_fixup, never extending beyond one
753 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
755 * AAAA AAAA AAAA AAAA
756 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
757 * cannot merge might become might become might become
758 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
759 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
760 * mremap move: PPPPNNNNNNNN 8
762 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
763 * might become case 1 below case 2 below case 3 below
765 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
766 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
768 struct vm_area_struct *vma_merge(struct mm_struct *mm,
769 struct vm_area_struct *prev, unsigned long addr,
770 unsigned long end, unsigned long vm_flags,
771 struct anon_vma *anon_vma, struct file *file,
772 pgoff_t pgoff, struct mempolicy *policy)
774 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
775 struct vm_area_struct *area, *next;
779 * We later require that vma->vm_flags == vm_flags,
780 * so this tests vma->vm_flags & VM_SPECIAL, too.
782 if (vm_flags & VM_SPECIAL)
786 next = prev->vm_next;
790 if (next && next->vm_end == end) /* cases 6, 7, 8 */
791 next = next->vm_next;
794 * Can it merge with the predecessor?
796 if (prev && prev->vm_end == addr &&
797 mpol_equal(vma_policy(prev), policy) &&
798 can_vma_merge_after(prev, vm_flags,
799 anon_vma, file, pgoff)) {
801 * OK, it can. Can we now merge in the successor as well?
803 if (next && end == next->vm_start &&
804 mpol_equal(policy, vma_policy(next)) &&
805 can_vma_merge_before(next, vm_flags,
806 anon_vma, file, pgoff+pglen) &&
807 is_mergeable_anon_vma(prev->anon_vma,
810 err = vma_adjust(prev, prev->vm_start,
811 next->vm_end, prev->vm_pgoff, NULL);
812 } else /* cases 2, 5, 7 */
813 err = vma_adjust(prev, prev->vm_start,
814 end, prev->vm_pgoff, NULL);
821 * Can this new request be merged in front of next?
823 if (next && end == next->vm_start &&
824 mpol_equal(policy, vma_policy(next)) &&
825 can_vma_merge_before(next, vm_flags,
826 anon_vma, file, pgoff+pglen)) {
827 if (prev && addr < prev->vm_end) /* case 4 */
828 err = vma_adjust(prev, prev->vm_start,
829 addr, prev->vm_pgoff, NULL);
830 else /* cases 3, 8 */
831 err = vma_adjust(area, addr, next->vm_end,
832 next->vm_pgoff - pglen, NULL);
842 * Rough compatbility check to quickly see if it's even worth looking
843 * at sharing an anon_vma.
845 * They need to have the same vm_file, and the flags can only differ
846 * in things that mprotect may change.
848 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
849 * we can merge the two vma's. For example, we refuse to merge a vma if
850 * there is a vm_ops->close() function, because that indicates that the
851 * driver is doing some kind of reference counting. But that doesn't
852 * really matter for the anon_vma sharing case.
854 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
856 return a->vm_end == b->vm_start &&
857 mpol_equal(vma_policy(a), vma_policy(b)) &&
858 a->vm_file == b->vm_file &&
859 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
860 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
864 * Do some basic sanity checking to see if we can re-use the anon_vma
865 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
866 * the same as 'old', the other will be the new one that is trying
867 * to share the anon_vma.
869 * NOTE! This runs with mm_sem held for reading, so it is possible that
870 * the anon_vma of 'old' is concurrently in the process of being set up
871 * by another page fault trying to merge _that_. But that's ok: if it
872 * is being set up, that automatically means that it will be a singleton
873 * acceptable for merging, so we can do all of this optimistically. But
874 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
876 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
877 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
878 * is to return an anon_vma that is "complex" due to having gone through
881 * We also make sure that the two vma's are compatible (adjacent,
882 * and with the same memory policies). That's all stable, even with just
883 * a read lock on the mm_sem.
885 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
887 if (anon_vma_compatible(a, b)) {
888 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
890 if (anon_vma && list_is_singular(&old->anon_vma_chain))
897 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
898 * neighbouring vmas for a suitable anon_vma, before it goes off
899 * to allocate a new anon_vma. It checks because a repetitive
900 * sequence of mprotects and faults may otherwise lead to distinct
901 * anon_vmas being allocated, preventing vma merge in subsequent
904 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
906 struct anon_vma *anon_vma;
907 struct vm_area_struct *near;
913 anon_vma = reusable_anon_vma(near, vma, near);
918 * It is potentially slow to have to call find_vma_prev here.
919 * But it's only on the first write fault on the vma, not
920 * every time, and we could devise a way to avoid it later
921 * (e.g. stash info in next's anon_vma_node when assigning
922 * an anon_vma, or when trying vma_merge). Another time.
924 BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
928 anon_vma = reusable_anon_vma(near, near, vma);
933 * There's no absolute need to look only at touching neighbours:
934 * we could search further afield for "compatible" anon_vmas.
935 * But it would probably just be a waste of time searching,
936 * or lead to too many vmas hanging off the same anon_vma.
937 * We're trying to allow mprotect remerging later on,
938 * not trying to minimize memory used for anon_vmas.
943 #ifdef CONFIG_PROC_FS
944 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
945 struct file *file, long pages)
947 const unsigned long stack_flags
948 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
951 mm->shared_vm += pages;
952 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
953 mm->exec_vm += pages;
954 } else if (flags & stack_flags)
955 mm->stack_vm += pages;
956 if (flags & (VM_RESERVED|VM_IO))
957 mm->reserved_vm += pages;
959 #endif /* CONFIG_PROC_FS */
962 * The caller must hold down_write(¤t->mm->mmap_sem).
965 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
966 unsigned long len, unsigned long prot,
967 unsigned long flags, unsigned long pgoff)
969 struct mm_struct * mm = current->mm;
971 unsigned int vm_flags;
973 unsigned long reqprot = prot;
976 * Does the application expect PROT_READ to imply PROT_EXEC?
978 * (the exception is when the underlying filesystem is noexec
979 * mounted, in which case we dont add PROT_EXEC.)
981 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
982 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
988 if (!(flags & MAP_FIXED))
989 addr = round_hint_to_min(addr);
991 /* Careful about overflows.. */
992 len = PAGE_ALIGN(len);
996 /* offset overflow? */
997 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1000 /* Too many mappings? */
1001 if (mm->map_count > sysctl_max_map_count)
1004 /* Obtain the address to map to. we verify (or select) it and ensure
1005 * that it represents a valid section of the address space.
1007 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1008 if (addr & ~PAGE_MASK)
1011 /* Do simple checking here so the lower-level routines won't have
1012 * to. we assume access permissions have been handled by the open
1013 * of the memory object, so we don't do any here.
1015 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1016 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1018 if (flags & MAP_LOCKED)
1019 if (!can_do_mlock())
1022 /* mlock MCL_FUTURE? */
1023 if (vm_flags & VM_LOCKED) {
1024 unsigned long locked, lock_limit;
1025 locked = len >> PAGE_SHIFT;
1026 locked += mm->locked_vm;
1027 lock_limit = rlimit(RLIMIT_MEMLOCK);
1028 lock_limit >>= PAGE_SHIFT;
1029 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1033 inode = file ? file->f_path.dentry->d_inode : NULL;
1036 switch (flags & MAP_TYPE) {
1038 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1042 * Make sure we don't allow writing to an append-only
1045 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1049 * Make sure there are no mandatory locks on the file.
1051 if (locks_verify_locked(inode))
1054 vm_flags |= VM_SHARED | VM_MAYSHARE;
1055 if (!(file->f_mode & FMODE_WRITE))
1056 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1060 if (!(file->f_mode & FMODE_READ))
1062 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1063 if (vm_flags & VM_EXEC)
1065 vm_flags &= ~VM_MAYEXEC;
1068 if (!file->f_op || !file->f_op->mmap)
1076 switch (flags & MAP_TYPE) {
1082 vm_flags |= VM_SHARED | VM_MAYSHARE;
1086 * Set pgoff according to addr for anon_vma.
1088 pgoff = addr >> PAGE_SHIFT;
1095 error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1099 return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1101 EXPORT_SYMBOL(do_mmap_pgoff);
1103 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1104 unsigned long, prot, unsigned long, flags,
1105 unsigned long, fd, unsigned long, pgoff)
1107 struct file *file = NULL;
1108 unsigned long retval = -EBADF;
1110 if (!(flags & MAP_ANONYMOUS)) {
1111 if (unlikely(flags & MAP_HUGETLB))
1116 } else if (flags & MAP_HUGETLB) {
1117 struct user_struct *user = NULL;
1119 * VM_NORESERVE is used because the reservations will be
1120 * taken when vm_ops->mmap() is called
1121 * A dummy user value is used because we are not locking
1122 * memory so no accounting is necessary
1124 len = ALIGN(len, huge_page_size(&default_hstate));
1125 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
1126 &user, HUGETLB_ANONHUGE_INODE);
1128 return PTR_ERR(file);
1131 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1133 down_write(¤t->mm->mmap_sem);
1134 retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1135 up_write(¤t->mm->mmap_sem);
1143 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1144 struct mmap_arg_struct {
1148 unsigned long flags;
1150 unsigned long offset;
1153 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1155 struct mmap_arg_struct a;
1157 if (copy_from_user(&a, arg, sizeof(a)))
1159 if (a.offset & ~PAGE_MASK)
1162 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1163 a.offset >> PAGE_SHIFT);
1165 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1168 * Some shared mappigns will want the pages marked read-only
1169 * to track write events. If so, we'll downgrade vm_page_prot
1170 * to the private version (using protection_map[] without the
1173 int vma_wants_writenotify(struct vm_area_struct *vma)
1175 unsigned int vm_flags = vma->vm_flags;
1177 /* If it was private or non-writable, the write bit is already clear */
1178 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1181 /* The backer wishes to know when pages are first written to? */
1182 if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1185 /* The open routine did something to the protections already? */
1186 if (pgprot_val(vma->vm_page_prot) !=
1187 pgprot_val(vm_get_page_prot(vm_flags)))
1190 /* Specialty mapping? */
1191 if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1194 /* Can the mapping track the dirty pages? */
1195 return vma->vm_file && vma->vm_file->f_mapping &&
1196 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1200 * We account for memory if it's a private writeable mapping,
1201 * not hugepages and VM_NORESERVE wasn't set.
1203 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1206 * hugetlb has its own accounting separate from the core VM
1207 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1209 if (file && is_file_hugepages(file))
1212 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1215 unsigned long mmap_region(struct file *file, unsigned long addr,
1216 unsigned long len, unsigned long flags,
1217 unsigned int vm_flags, unsigned long pgoff)
1219 struct mm_struct *mm = current->mm;
1220 struct vm_area_struct *vma, *prev;
1221 int correct_wcount = 0;
1223 struct rb_node **rb_link, *rb_parent;
1224 unsigned long charged = 0;
1225 struct inode *inode = file ? file->f_path.dentry->d_inode : NULL;
1227 /* Clear old maps */
1230 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1231 if (vma && vma->vm_start < addr + len) {
1232 if (do_munmap(mm, addr, len))
1237 /* Check against address space limit. */
1238 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1242 * Set 'VM_NORESERVE' if we should not account for the
1243 * memory use of this mapping.
1245 if ((flags & MAP_NORESERVE)) {
1246 /* We honor MAP_NORESERVE if allowed to overcommit */
1247 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1248 vm_flags |= VM_NORESERVE;
1250 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1251 if (file && is_file_hugepages(file))
1252 vm_flags |= VM_NORESERVE;
1256 * Private writable mapping: check memory availability
1258 if (accountable_mapping(file, vm_flags)) {
1259 charged = len >> PAGE_SHIFT;
1260 if (security_vm_enough_memory(charged))
1262 vm_flags |= VM_ACCOUNT;
1266 * Can we just expand an old mapping?
1268 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1273 * Determine the object being mapped and call the appropriate
1274 * specific mapper. the address has already been validated, but
1275 * not unmapped, but the maps are removed from the list.
1277 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1284 vma->vm_start = addr;
1285 vma->vm_end = addr + len;
1286 vma->vm_flags = vm_flags;
1287 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1288 vma->vm_pgoff = pgoff;
1289 INIT_LIST_HEAD(&vma->anon_vma_chain);
1293 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1295 if (vm_flags & VM_DENYWRITE) {
1296 error = deny_write_access(file);
1301 vma->vm_file = file;
1303 error = file->f_op->mmap(file, vma);
1305 goto unmap_and_free_vma;
1306 if (vm_flags & VM_EXECUTABLE)
1307 added_exe_file_vma(mm);
1309 /* Can addr have changed??
1311 * Answer: Yes, several device drivers can do it in their
1312 * f_op->mmap method. -DaveM
1314 addr = vma->vm_start;
1315 pgoff = vma->vm_pgoff;
1316 vm_flags = vma->vm_flags;
1317 } else if (vm_flags & VM_SHARED) {
1318 error = shmem_zero_setup(vma);
1323 if (vma_wants_writenotify(vma)) {
1324 pgprot_t pprot = vma->vm_page_prot;
1326 /* Can vma->vm_page_prot have changed??
1328 * Answer: Yes, drivers may have changed it in their
1329 * f_op->mmap method.
1331 * Ensures that vmas marked as uncached stay that way.
1333 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1334 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1335 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1338 vma_link(mm, vma, prev, rb_link, rb_parent);
1339 file = vma->vm_file;
1341 /* Once vma denies write, undo our temporary denial count */
1343 atomic_inc(&inode->i_writecount);
1345 perf_event_mmap(vma);
1347 mm->total_vm += len >> PAGE_SHIFT;
1348 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1349 if (vm_flags & VM_LOCKED) {
1350 if (!mlock_vma_pages_range(vma, addr, addr + len))
1351 mm->locked_vm += (len >> PAGE_SHIFT);
1352 } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1353 make_pages_present(addr, addr + len);
1358 atomic_inc(&inode->i_writecount);
1359 vma->vm_file = NULL;
1362 /* Undo any partial mapping done by a device driver. */
1363 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1366 kmem_cache_free(vm_area_cachep, vma);
1369 vm_unacct_memory(charged);
1373 /* Get an address range which is currently unmapped.
1374 * For shmat() with addr=0.
1376 * Ugly calling convention alert:
1377 * Return value with the low bits set means error value,
1379 * if (ret & ~PAGE_MASK)
1382 * This function "knows" that -ENOMEM has the bits set.
1384 #ifndef HAVE_ARCH_UNMAPPED_AREA
1386 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1387 unsigned long len, unsigned long pgoff, unsigned long flags)
1389 struct mm_struct *mm = current->mm;
1390 struct vm_area_struct *vma;
1391 unsigned long start_addr;
1393 if (len > TASK_SIZE)
1396 if (flags & MAP_FIXED)
1400 addr = PAGE_ALIGN(addr);
1401 vma = find_vma(mm, addr);
1402 if (TASK_SIZE - len >= addr &&
1403 (!vma || addr + len <= vma->vm_start))
1406 if (len > mm->cached_hole_size) {
1407 start_addr = addr = mm->free_area_cache;
1409 start_addr = addr = TASK_UNMAPPED_BASE;
1410 mm->cached_hole_size = 0;
1414 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1415 /* At this point: (!vma || addr < vma->vm_end). */
1416 if (TASK_SIZE - len < addr) {
1418 * Start a new search - just in case we missed
1421 if (start_addr != TASK_UNMAPPED_BASE) {
1422 addr = TASK_UNMAPPED_BASE;
1424 mm->cached_hole_size = 0;
1429 if (!vma || addr + len <= vma->vm_start) {
1431 * Remember the place where we stopped the search:
1433 mm->free_area_cache = addr + len;
1436 if (addr + mm->cached_hole_size < vma->vm_start)
1437 mm->cached_hole_size = vma->vm_start - addr;
1443 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1446 * Is this a new hole at the lowest possible address?
1448 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1449 mm->free_area_cache = addr;
1450 mm->cached_hole_size = ~0UL;
1455 * This mmap-allocator allocates new areas top-down from below the
1456 * stack's low limit (the base):
1458 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1460 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1461 const unsigned long len, const unsigned long pgoff,
1462 const unsigned long flags)
1464 struct vm_area_struct *vma;
1465 struct mm_struct *mm = current->mm;
1466 unsigned long addr = addr0;
1468 /* requested length too big for entire address space */
1469 if (len > TASK_SIZE)
1472 if (flags & MAP_FIXED)
1475 /* requesting a specific address */
1477 addr = PAGE_ALIGN(addr);
1478 vma = find_vma(mm, addr);
1479 if (TASK_SIZE - len >= addr &&
1480 (!vma || addr + len <= vma->vm_start))
1484 /* check if free_area_cache is useful for us */
1485 if (len <= mm->cached_hole_size) {
1486 mm->cached_hole_size = 0;
1487 mm->free_area_cache = mm->mmap_base;
1490 /* either no address requested or can't fit in requested address hole */
1491 addr = mm->free_area_cache;
1493 /* make sure it can fit in the remaining address space */
1495 vma = find_vma(mm, addr-len);
1496 if (!vma || addr <= vma->vm_start)
1497 /* remember the address as a hint for next time */
1498 return (mm->free_area_cache = addr-len);
1501 if (mm->mmap_base < len)
1504 addr = mm->mmap_base-len;
1508 * Lookup failure means no vma is above this address,
1509 * else if new region fits below vma->vm_start,
1510 * return with success:
1512 vma = find_vma(mm, addr);
1513 if (!vma || addr+len <= vma->vm_start)
1514 /* remember the address as a hint for next time */
1515 return (mm->free_area_cache = addr);
1517 /* remember the largest hole we saw so far */
1518 if (addr + mm->cached_hole_size < vma->vm_start)
1519 mm->cached_hole_size = vma->vm_start - addr;
1521 /* try just below the current vma->vm_start */
1522 addr = vma->vm_start-len;
1523 } while (len < vma->vm_start);
1527 * A failed mmap() very likely causes application failure,
1528 * so fall back to the bottom-up function here. This scenario
1529 * can happen with large stack limits and large mmap()
1532 mm->cached_hole_size = ~0UL;
1533 mm->free_area_cache = TASK_UNMAPPED_BASE;
1534 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1536 * Restore the topdown base:
1538 mm->free_area_cache = mm->mmap_base;
1539 mm->cached_hole_size = ~0UL;
1545 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1548 * Is this a new hole at the highest possible address?
1550 if (addr > mm->free_area_cache)
1551 mm->free_area_cache = addr;
1553 /* dont allow allocations above current base */
1554 if (mm->free_area_cache > mm->mmap_base)
1555 mm->free_area_cache = mm->mmap_base;
1559 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1560 unsigned long pgoff, unsigned long flags)
1562 unsigned long (*get_area)(struct file *, unsigned long,
1563 unsigned long, unsigned long, unsigned long);
1565 unsigned long error = arch_mmap_check(addr, len, flags);
1569 /* Careful about overflows.. */
1570 if (len > TASK_SIZE)
1573 get_area = current->mm->get_unmapped_area;
1574 if (file && file->f_op && file->f_op->get_unmapped_area)
1575 get_area = file->f_op->get_unmapped_area;
1576 addr = get_area(file, addr, len, pgoff, flags);
1577 if (IS_ERR_VALUE(addr))
1580 if (addr > TASK_SIZE - len)
1582 if (addr & ~PAGE_MASK)
1585 return arch_rebalance_pgtables(addr, len);
1588 EXPORT_SYMBOL(get_unmapped_area);
1590 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1591 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1593 struct vm_area_struct *vma = NULL;
1596 /* Check the cache first. */
1597 /* (Cache hit rate is typically around 35%.) */
1598 vma = mm->mmap_cache;
1599 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1600 struct rb_node * rb_node;
1602 rb_node = mm->mm_rb.rb_node;
1606 struct vm_area_struct * vma_tmp;
1608 vma_tmp = rb_entry(rb_node,
1609 struct vm_area_struct, vm_rb);
1611 if (vma_tmp->vm_end > addr) {
1613 if (vma_tmp->vm_start <= addr)
1615 rb_node = rb_node->rb_left;
1617 rb_node = rb_node->rb_right;
1620 mm->mmap_cache = vma;
1626 EXPORT_SYMBOL(find_vma);
1628 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1629 struct vm_area_struct *
1630 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1631 struct vm_area_struct **pprev)
1633 struct vm_area_struct *vma = NULL, *prev = NULL;
1634 struct rb_node *rb_node;
1638 /* Guard against addr being lower than the first VMA */
1641 /* Go through the RB tree quickly. */
1642 rb_node = mm->mm_rb.rb_node;
1645 struct vm_area_struct *vma_tmp;
1646 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1648 if (addr < vma_tmp->vm_end) {
1649 rb_node = rb_node->rb_left;
1652 if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1654 rb_node = rb_node->rb_right;
1660 return prev ? prev->vm_next : vma;
1664 * Verify that the stack growth is acceptable and
1665 * update accounting. This is shared with both the
1666 * grow-up and grow-down cases.
1668 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1670 struct mm_struct *mm = vma->vm_mm;
1671 struct rlimit *rlim = current->signal->rlim;
1672 unsigned long new_start;
1674 /* address space limit tests */
1675 if (!may_expand_vm(mm, grow))
1678 /* Stack limit test */
1679 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1682 /* mlock limit tests */
1683 if (vma->vm_flags & VM_LOCKED) {
1684 unsigned long locked;
1685 unsigned long limit;
1686 locked = mm->locked_vm + grow;
1687 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1688 limit >>= PAGE_SHIFT;
1689 if (locked > limit && !capable(CAP_IPC_LOCK))
1693 /* Check to ensure the stack will not grow into a hugetlb-only region */
1694 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1696 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1700 * Overcommit.. This must be the final test, as it will
1701 * update security statistics.
1703 if (security_vm_enough_memory_mm(mm, grow))
1706 /* Ok, everything looks good - let it rip */
1707 mm->total_vm += grow;
1708 if (vma->vm_flags & VM_LOCKED)
1709 mm->locked_vm += grow;
1710 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1714 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1716 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1717 * vma is the last one with address > vma->vm_end. Have to extend vma.
1719 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1723 if (!(vma->vm_flags & VM_GROWSUP))
1727 * We must make sure the anon_vma is allocated
1728 * so that the anon_vma locking is not a noop.
1730 if (unlikely(anon_vma_prepare(vma)))
1732 vma_lock_anon_vma(vma);
1735 * vma->vm_start/vm_end cannot change under us because the caller
1736 * is required to hold the mmap_sem in read mode. We need the
1737 * anon_vma lock to serialize against concurrent expand_stacks.
1738 * Also guard against wrapping around to address 0.
1740 if (address < PAGE_ALIGN(address+4))
1741 address = PAGE_ALIGN(address+4);
1743 vma_unlock_anon_vma(vma);
1748 /* Somebody else might have raced and expanded it already */
1749 if (address > vma->vm_end) {
1750 unsigned long size, grow;
1752 size = address - vma->vm_start;
1753 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1755 error = acct_stack_growth(vma, size, grow);
1757 vma->vm_end = address;
1758 perf_event_mmap(vma);
1761 vma_unlock_anon_vma(vma);
1764 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1767 * vma is the first one with address < vma->vm_start. Have to extend vma.
1769 static int expand_downwards(struct vm_area_struct *vma,
1770 unsigned long address)
1775 * We must make sure the anon_vma is allocated
1776 * so that the anon_vma locking is not a noop.
1778 if (unlikely(anon_vma_prepare(vma)))
1781 address &= PAGE_MASK;
1782 error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1786 vma_lock_anon_vma(vma);
1789 * vma->vm_start/vm_end cannot change under us because the caller
1790 * is required to hold the mmap_sem in read mode. We need the
1791 * anon_vma lock to serialize against concurrent expand_stacks.
1794 /* Somebody else might have raced and expanded it already */
1795 if (address < vma->vm_start) {
1796 unsigned long size, grow;
1798 size = vma->vm_end - address;
1799 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1801 error = acct_stack_growth(vma, size, grow);
1803 vma->vm_start = address;
1804 vma->vm_pgoff -= grow;
1805 perf_event_mmap(vma);
1808 vma_unlock_anon_vma(vma);
1812 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1814 return expand_downwards(vma, address);
1817 #ifdef CONFIG_STACK_GROWSUP
1818 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1820 return expand_upwards(vma, address);
1823 struct vm_area_struct *
1824 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1826 struct vm_area_struct *vma, *prev;
1829 vma = find_vma_prev(mm, addr, &prev);
1830 if (vma && (vma->vm_start <= addr))
1832 if (!prev || expand_stack(prev, addr))
1834 if (prev->vm_flags & VM_LOCKED) {
1835 mlock_vma_pages_range(prev, addr, prev->vm_end);
1840 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1842 return expand_downwards(vma, address);
1845 struct vm_area_struct *
1846 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1848 struct vm_area_struct * vma;
1849 unsigned long start;
1852 vma = find_vma(mm,addr);
1855 if (vma->vm_start <= addr)
1857 if (!(vma->vm_flags & VM_GROWSDOWN))
1859 start = vma->vm_start;
1860 if (expand_stack(vma, addr))
1862 if (vma->vm_flags & VM_LOCKED) {
1863 mlock_vma_pages_range(vma, addr, start);
1870 * Ok - we have the memory areas we should free on the vma list,
1871 * so release them, and do the vma updates.
1873 * Called with the mm semaphore held.
1875 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1877 /* Update high watermark before we lower total_vm */
1878 update_hiwater_vm(mm);
1880 long nrpages = vma_pages(vma);
1882 mm->total_vm -= nrpages;
1883 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1884 vma = remove_vma(vma);
1890 * Get rid of page table information in the indicated region.
1892 * Called with the mm semaphore held.
1894 static void unmap_region(struct mm_struct *mm,
1895 struct vm_area_struct *vma, struct vm_area_struct *prev,
1896 unsigned long start, unsigned long end)
1898 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1899 struct mmu_gather *tlb;
1900 unsigned long nr_accounted = 0;
1903 tlb = tlb_gather_mmu(mm, 0);
1904 update_hiwater_rss(mm);
1905 unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1906 vm_unacct_memory(nr_accounted);
1907 free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1908 next? next->vm_start: 0);
1909 tlb_finish_mmu(tlb, start, end);
1913 * Create a list of vma's touched by the unmap, removing them from the mm's
1914 * vma list as we go..
1917 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1918 struct vm_area_struct *prev, unsigned long end)
1920 struct vm_area_struct **insertion_point;
1921 struct vm_area_struct *tail_vma = NULL;
1924 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1925 vma->vm_prev = NULL;
1927 rb_erase(&vma->vm_rb, &mm->mm_rb);
1931 } while (vma && vma->vm_start < end);
1932 *insertion_point = vma;
1934 vma->vm_prev = prev;
1935 tail_vma->vm_next = NULL;
1936 if (mm->unmap_area == arch_unmap_area)
1937 addr = prev ? prev->vm_end : mm->mmap_base;
1939 addr = vma ? vma->vm_start : mm->mmap_base;
1940 mm->unmap_area(mm, addr);
1941 mm->mmap_cache = NULL; /* Kill the cache. */
1945 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
1946 * munmap path where it doesn't make sense to fail.
1948 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1949 unsigned long addr, int new_below)
1951 struct mempolicy *pol;
1952 struct vm_area_struct *new;
1955 if (is_vm_hugetlb_page(vma) && (addr &
1956 ~(huge_page_mask(hstate_vma(vma)))))
1959 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1963 /* most fields are the same, copy all, and then fixup */
1966 INIT_LIST_HEAD(&new->anon_vma_chain);
1971 new->vm_start = addr;
1972 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1975 pol = mpol_dup(vma_policy(vma));
1980 vma_set_policy(new, pol);
1982 if (anon_vma_clone(new, vma))
1986 get_file(new->vm_file);
1987 if (vma->vm_flags & VM_EXECUTABLE)
1988 added_exe_file_vma(mm);
1991 if (new->vm_ops && new->vm_ops->open)
1992 new->vm_ops->open(new);
1995 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1996 ((addr - new->vm_start) >> PAGE_SHIFT), new);
1998 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2004 /* Clean everything up if vma_adjust failed. */
2005 if (new->vm_ops && new->vm_ops->close)
2006 new->vm_ops->close(new);
2008 if (vma->vm_flags & VM_EXECUTABLE)
2009 removed_exe_file_vma(mm);
2012 unlink_anon_vmas(new);
2016 kmem_cache_free(vm_area_cachep, new);
2022 * Split a vma into two pieces at address 'addr', a new vma is allocated
2023 * either for the first part or the tail.
2025 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2026 unsigned long addr, int new_below)
2028 if (mm->map_count >= sysctl_max_map_count)
2031 return __split_vma(mm, vma, addr, new_below);
2034 /* Munmap is split into 2 main parts -- this part which finds
2035 * what needs doing, and the areas themselves, which do the
2036 * work. This now handles partial unmappings.
2037 * Jeremy Fitzhardinge <jeremy@goop.org>
2039 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2042 struct vm_area_struct *vma, *prev, *last;
2044 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2047 if ((len = PAGE_ALIGN(len)) == 0)
2050 /* Find the first overlapping VMA */
2051 vma = find_vma_prev(mm, start, &prev);
2054 /* we have start < vma->vm_end */
2056 /* if it doesn't overlap, we have nothing.. */
2058 if (vma->vm_start >= end)
2062 * If we need to split any vma, do it now to save pain later.
2064 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2065 * unmapped vm_area_struct will remain in use: so lower split_vma
2066 * places tmp vma above, and higher split_vma places tmp vma below.
2068 if (start > vma->vm_start) {
2072 * Make sure that map_count on return from munmap() will
2073 * not exceed its limit; but let map_count go just above
2074 * its limit temporarily, to help free resources as expected.
2076 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2079 error = __split_vma(mm, vma, start, 0);
2085 /* Does it split the last one? */
2086 last = find_vma(mm, end);
2087 if (last && end > last->vm_start) {
2088 int error = __split_vma(mm, last, end, 1);
2092 vma = prev? prev->vm_next: mm->mmap;
2095 * unlock any mlock()ed ranges before detaching vmas
2097 if (mm->locked_vm) {
2098 struct vm_area_struct *tmp = vma;
2099 while (tmp && tmp->vm_start < end) {
2100 if (tmp->vm_flags & VM_LOCKED) {
2101 mm->locked_vm -= vma_pages(tmp);
2102 munlock_vma_pages_all(tmp);
2109 * Remove the vma's, and unmap the actual pages
2111 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2112 unmap_region(mm, vma, prev, start, end);
2114 /* Fix up all other VM information */
2115 remove_vma_list(mm, vma);
2120 EXPORT_SYMBOL(do_munmap);
2122 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2125 struct mm_struct *mm = current->mm;
2127 profile_munmap(addr);
2129 down_write(&mm->mmap_sem);
2130 ret = do_munmap(mm, addr, len);
2131 up_write(&mm->mmap_sem);
2135 static inline void verify_mm_writelocked(struct mm_struct *mm)
2137 #ifdef CONFIG_DEBUG_VM
2138 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2140 up_read(&mm->mmap_sem);
2146 * this is really a simplified "do_mmap". it only handles
2147 * anonymous maps. eventually we may be able to do some
2148 * brk-specific accounting here.
2150 unsigned long do_brk(unsigned long addr, unsigned long len)
2152 struct mm_struct * mm = current->mm;
2153 struct vm_area_struct * vma, * prev;
2154 unsigned long flags;
2155 struct rb_node ** rb_link, * rb_parent;
2156 pgoff_t pgoff = addr >> PAGE_SHIFT;
2159 len = PAGE_ALIGN(len);
2163 error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2167 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2169 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2170 if (error & ~PAGE_MASK)
2176 if (mm->def_flags & VM_LOCKED) {
2177 unsigned long locked, lock_limit;
2178 locked = len >> PAGE_SHIFT;
2179 locked += mm->locked_vm;
2180 lock_limit = rlimit(RLIMIT_MEMLOCK);
2181 lock_limit >>= PAGE_SHIFT;
2182 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2187 * mm->mmap_sem is required to protect against another thread
2188 * changing the mappings in case we sleep.
2190 verify_mm_writelocked(mm);
2193 * Clear old maps. this also does some error checking for us
2196 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2197 if (vma && vma->vm_start < addr + len) {
2198 if (do_munmap(mm, addr, len))
2203 /* Check against address space limits *after* clearing old maps... */
2204 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2207 if (mm->map_count > sysctl_max_map_count)
2210 if (security_vm_enough_memory(len >> PAGE_SHIFT))
2213 /* Can we just expand an old private anonymous mapping? */
2214 vma = vma_merge(mm, prev, addr, addr + len, flags,
2215 NULL, NULL, pgoff, NULL);
2220 * create a vma struct for an anonymous mapping
2222 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2224 vm_unacct_memory(len >> PAGE_SHIFT);
2228 INIT_LIST_HEAD(&vma->anon_vma_chain);
2230 vma->vm_start = addr;
2231 vma->vm_end = addr + len;
2232 vma->vm_pgoff = pgoff;
2233 vma->vm_flags = flags;
2234 vma->vm_page_prot = vm_get_page_prot(flags);
2235 vma_link(mm, vma, prev, rb_link, rb_parent);
2237 perf_event_mmap(vma);
2238 mm->total_vm += len >> PAGE_SHIFT;
2239 if (flags & VM_LOCKED) {
2240 if (!mlock_vma_pages_range(vma, addr, addr + len))
2241 mm->locked_vm += (len >> PAGE_SHIFT);
2246 EXPORT_SYMBOL(do_brk);
2248 /* Release all mmaps. */
2249 void exit_mmap(struct mm_struct *mm)
2251 struct mmu_gather *tlb;
2252 struct vm_area_struct *vma;
2253 unsigned long nr_accounted = 0;
2256 /* mm's last user has gone, and its about to be pulled down */
2257 mmu_notifier_release(mm);
2259 if (mm->locked_vm) {
2262 if (vma->vm_flags & VM_LOCKED)
2263 munlock_vma_pages_all(vma);
2271 if (!vma) /* Can happen if dup_mmap() received an OOM */
2276 tlb = tlb_gather_mmu(mm, 1);
2277 /* update_hiwater_rss(mm) here? but nobody should be looking */
2278 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2279 end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2280 vm_unacct_memory(nr_accounted);
2282 free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2283 tlb_finish_mmu(tlb, 0, end);
2286 * Walk the list again, actually closing and freeing it,
2287 * with preemption enabled, without holding any MM locks.
2290 vma = remove_vma(vma);
2292 BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2295 /* Insert vm structure into process list sorted by address
2296 * and into the inode's i_mmap tree. If vm_file is non-NULL
2297 * then i_mmap_lock is taken here.
2299 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2301 struct vm_area_struct * __vma, * prev;
2302 struct rb_node ** rb_link, * rb_parent;
2305 * The vm_pgoff of a purely anonymous vma should be irrelevant
2306 * until its first write fault, when page's anon_vma and index
2307 * are set. But now set the vm_pgoff it will almost certainly
2308 * end up with (unless mremap moves it elsewhere before that
2309 * first wfault), so /proc/pid/maps tells a consistent story.
2311 * By setting it to reflect the virtual start address of the
2312 * vma, merges and splits can happen in a seamless way, just
2313 * using the existing file pgoff checks and manipulations.
2314 * Similarly in do_mmap_pgoff and in do_brk.
2316 if (!vma->vm_file) {
2317 BUG_ON(vma->anon_vma);
2318 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2320 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2321 if (__vma && __vma->vm_start < vma->vm_end)
2323 if ((vma->vm_flags & VM_ACCOUNT) &&
2324 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2326 vma_link(mm, vma, prev, rb_link, rb_parent);
2331 * Copy the vma structure to a new location in the same mm,
2332 * prior to moving page table entries, to effect an mremap move.
2334 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2335 unsigned long addr, unsigned long len, pgoff_t pgoff)
2337 struct vm_area_struct *vma = *vmap;
2338 unsigned long vma_start = vma->vm_start;
2339 struct mm_struct *mm = vma->vm_mm;
2340 struct vm_area_struct *new_vma, *prev;
2341 struct rb_node **rb_link, *rb_parent;
2342 struct mempolicy *pol;
2345 * If anonymous vma has not yet been faulted, update new pgoff
2346 * to match new location, to increase its chance of merging.
2348 if (!vma->vm_file && !vma->anon_vma)
2349 pgoff = addr >> PAGE_SHIFT;
2351 find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2352 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2353 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2356 * Source vma may have been merged into new_vma
2358 if (vma_start >= new_vma->vm_start &&
2359 vma_start < new_vma->vm_end)
2362 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2365 pol = mpol_dup(vma_policy(vma));
2368 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2369 if (anon_vma_clone(new_vma, vma))
2370 goto out_free_mempol;
2371 vma_set_policy(new_vma, pol);
2372 new_vma->vm_start = addr;
2373 new_vma->vm_end = addr + len;
2374 new_vma->vm_pgoff = pgoff;
2375 if (new_vma->vm_file) {
2376 get_file(new_vma->vm_file);
2377 if (vma->vm_flags & VM_EXECUTABLE)
2378 added_exe_file_vma(mm);
2380 if (new_vma->vm_ops && new_vma->vm_ops->open)
2381 new_vma->vm_ops->open(new_vma);
2382 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2390 kmem_cache_free(vm_area_cachep, new_vma);
2395 * Return true if the calling process may expand its vm space by the passed
2398 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2400 unsigned long cur = mm->total_vm; /* pages */
2403 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2405 if (cur + npages > lim)
2411 static int special_mapping_fault(struct vm_area_struct *vma,
2412 struct vm_fault *vmf)
2415 struct page **pages;
2418 * special mappings have no vm_file, and in that case, the mm
2419 * uses vm_pgoff internally. So we have to subtract it from here.
2420 * We are allowed to do this because we are the mm; do not copy
2421 * this code into drivers!
2423 pgoff = vmf->pgoff - vma->vm_pgoff;
2425 for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2429 struct page *page = *pages;
2435 return VM_FAULT_SIGBUS;
2439 * Having a close hook prevents vma merging regardless of flags.
2441 static void special_mapping_close(struct vm_area_struct *vma)
2445 static const struct vm_operations_struct special_mapping_vmops = {
2446 .close = special_mapping_close,
2447 .fault = special_mapping_fault,
2451 * Called with mm->mmap_sem held for writing.
2452 * Insert a new vma covering the given region, with the given flags.
2453 * Its pages are supplied by the given array of struct page *.
2454 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2455 * The region past the last page supplied will always produce SIGBUS.
2456 * The array pointer and the pages it points to are assumed to stay alive
2457 * for as long as this mapping might exist.
2459 int install_special_mapping(struct mm_struct *mm,
2460 unsigned long addr, unsigned long len,
2461 unsigned long vm_flags, struct page **pages)
2463 struct vm_area_struct *vma;
2465 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2466 if (unlikely(vma == NULL))
2469 INIT_LIST_HEAD(&vma->anon_vma_chain);
2471 vma->vm_start = addr;
2472 vma->vm_end = addr + len;
2474 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2475 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2477 vma->vm_ops = &special_mapping_vmops;
2478 vma->vm_private_data = pages;
2480 if (unlikely(insert_vm_struct(mm, vma))) {
2481 kmem_cache_free(vm_area_cachep, vma);
2485 mm->total_vm += len >> PAGE_SHIFT;
2487 perf_event_mmap(vma);
2492 static DEFINE_MUTEX(mm_all_locks_mutex);
2494 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2496 if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2498 * The LSB of head.next can't change from under us
2499 * because we hold the mm_all_locks_mutex.
2501 spin_lock_nest_lock(&anon_vma->root->lock, &mm->mmap_sem);
2503 * We can safely modify head.next after taking the
2504 * anon_vma->root->lock. If some other vma in this mm shares
2505 * the same anon_vma we won't take it again.
2507 * No need of atomic instructions here, head.next
2508 * can't change from under us thanks to the
2509 * anon_vma->root->lock.
2511 if (__test_and_set_bit(0, (unsigned long *)
2512 &anon_vma->root->head.next))
2517 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2519 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2521 * AS_MM_ALL_LOCKS can't change from under us because
2522 * we hold the mm_all_locks_mutex.
2524 * Operations on ->flags have to be atomic because
2525 * even if AS_MM_ALL_LOCKS is stable thanks to the
2526 * mm_all_locks_mutex, there may be other cpus
2527 * changing other bitflags in parallel to us.
2529 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2531 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2536 * This operation locks against the VM for all pte/vma/mm related
2537 * operations that could ever happen on a certain mm. This includes
2538 * vmtruncate, try_to_unmap, and all page faults.
2540 * The caller must take the mmap_sem in write mode before calling
2541 * mm_take_all_locks(). The caller isn't allowed to release the
2542 * mmap_sem until mm_drop_all_locks() returns.
2544 * mmap_sem in write mode is required in order to block all operations
2545 * that could modify pagetables and free pages without need of
2546 * altering the vma layout (for example populate_range() with
2547 * nonlinear vmas). It's also needed in write mode to avoid new
2548 * anon_vmas to be associated with existing vmas.
2550 * A single task can't take more than one mm_take_all_locks() in a row
2551 * or it would deadlock.
2553 * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2554 * mapping->flags avoid to take the same lock twice, if more than one
2555 * vma in this mm is backed by the same anon_vma or address_space.
2557 * We can take all the locks in random order because the VM code
2558 * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2559 * takes more than one of them in a row. Secondly we're protected
2560 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2562 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2563 * that may have to take thousand of locks.
2565 * mm_take_all_locks() can fail if it's interrupted by signals.
2567 int mm_take_all_locks(struct mm_struct *mm)
2569 struct vm_area_struct *vma;
2570 struct anon_vma_chain *avc;
2573 BUG_ON(down_read_trylock(&mm->mmap_sem));
2575 mutex_lock(&mm_all_locks_mutex);
2577 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2578 if (signal_pending(current))
2580 if (vma->vm_file && vma->vm_file->f_mapping)
2581 vm_lock_mapping(mm, vma->vm_file->f_mapping);
2584 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2585 if (signal_pending(current))
2588 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2589 vm_lock_anon_vma(mm, avc->anon_vma);
2596 mm_drop_all_locks(mm);
2601 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2603 if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2605 * The LSB of head.next can't change to 0 from under
2606 * us because we hold the mm_all_locks_mutex.
2608 * We must however clear the bitflag before unlocking
2609 * the vma so the users using the anon_vma->head will
2610 * never see our bitflag.
2612 * No need of atomic instructions here, head.next
2613 * can't change from under us until we release the
2614 * anon_vma->root->lock.
2616 if (!__test_and_clear_bit(0, (unsigned long *)
2617 &anon_vma->root->head.next))
2619 anon_vma_unlock(anon_vma);
2623 static void vm_unlock_mapping(struct address_space *mapping)
2625 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2627 * AS_MM_ALL_LOCKS can't change to 0 from under us
2628 * because we hold the mm_all_locks_mutex.
2630 spin_unlock(&mapping->i_mmap_lock);
2631 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2638 * The mmap_sem cannot be released by the caller until
2639 * mm_drop_all_locks() returns.
2641 void mm_drop_all_locks(struct mm_struct *mm)
2643 struct vm_area_struct *vma;
2644 struct anon_vma_chain *avc;
2646 BUG_ON(down_read_trylock(&mm->mmap_sem));
2647 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2649 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2651 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2652 vm_unlock_anon_vma(avc->anon_vma);
2653 if (vma->vm_file && vma->vm_file->f_mapping)
2654 vm_unlock_mapping(vma->vm_file->f_mapping);
2657 mutex_unlock(&mm_all_locks_mutex);
2661 * initialise the VMA slab
2663 void __init mmap_init(void)
2667 ret = percpu_counter_init(&vm_committed_as, 0);