6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/export.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29 #include <linux/mmu_notifier.h>
30 #include <linux/perf_event.h>
31 #include <linux/audit.h>
32 #include <linux/khugepaged.h>
33 #include <linux/uprobes.h>
35 #include <asm/uaccess.h>
36 #include <asm/cacheflush.h>
38 #include <asm/mmu_context.h>
42 #ifndef arch_mmap_check
43 #define arch_mmap_check(addr, len, flags) (0)
46 #ifndef arch_rebalance_pgtables
47 #define arch_rebalance_pgtables(addr, len) (addr)
50 static void unmap_region(struct mm_struct *mm,
51 struct vm_area_struct *vma, struct vm_area_struct *prev,
52 unsigned long start, unsigned long end);
55 * WARNING: the debugging will use recursive algorithms so never enable this
56 * unless you know what you are doing.
60 /* description of effects of mapping type and prot in current implementation.
61 * this is due to the limited x86 page protection hardware. The expected
62 * behavior is in parens:
65 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
66 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
67 * w: (no) no w: (no) no w: (yes) yes w: (no) no
68 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
70 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
71 * w: (no) no w: (no) no w: (copy) copy w: (no) no
72 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
75 pgprot_t protection_map[16] = {
76 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
77 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
80 pgprot_t vm_get_page_prot(unsigned long vm_flags)
82 return __pgprot(pgprot_val(protection_map[vm_flags &
83 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
84 pgprot_val(arch_vm_get_page_prot(vm_flags)));
86 EXPORT_SYMBOL(vm_get_page_prot);
88 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS; /* heuristic overcommit */
89 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
90 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
92 * Make sure vm_committed_as in one cacheline and not cacheline shared with
93 * other variables. It can be updated by several CPUs frequently.
95 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
98 * Check that a process has enough memory to allocate a new virtual
99 * mapping. 0 means there is enough memory for the allocation to
100 * succeed and -ENOMEM implies there is not.
102 * We currently support three overcommit policies, which are set via the
103 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
105 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
106 * Additional code 2002 Jul 20 by Robert Love.
108 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
110 * Note this is a helper function intended to be used by LSMs which
111 * wish to use this logic.
113 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
115 unsigned long free, allowed;
117 vm_acct_memory(pages);
120 * Sometimes we want to use more memory than we have
122 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
125 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
126 free = global_page_state(NR_FREE_PAGES);
127 free += global_page_state(NR_FILE_PAGES);
130 * shmem pages shouldn't be counted as free in this
131 * case, they can't be purged, only swapped out, and
132 * that won't affect the overall amount of available
133 * memory in the system.
135 free -= global_page_state(NR_SHMEM);
137 free += nr_swap_pages;
140 * Any slabs which are created with the
141 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
142 * which are reclaimable, under pressure. The dentry
143 * cache and most inode caches should fall into this
145 free += global_page_state(NR_SLAB_RECLAIMABLE);
148 * Leave reserved pages. The pages are not for anonymous pages.
150 if (free <= totalreserve_pages)
153 free -= totalreserve_pages;
156 * Leave the last 3% for root
167 allowed = (totalram_pages - hugetlb_total_pages())
168 * sysctl_overcommit_ratio / 100;
170 * Leave the last 3% for root
173 allowed -= allowed / 32;
174 allowed += total_swap_pages;
176 /* Don't let a single process grow too big:
177 leave 3% of the size of this process for other processes */
179 allowed -= mm->total_vm / 32;
181 if (percpu_counter_read_positive(&vm_committed_as) < allowed)
184 vm_unacct_memory(pages);
190 * Requires inode->i_mapping->i_mmap_mutex
192 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
193 struct file *file, struct address_space *mapping)
195 if (vma->vm_flags & VM_DENYWRITE)
196 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
197 if (vma->vm_flags & VM_SHARED)
198 mapping->i_mmap_writable--;
200 flush_dcache_mmap_lock(mapping);
201 if (unlikely(vma->vm_flags & VM_NONLINEAR))
202 list_del_init(&vma->shared.vm_set.list);
204 vma_prio_tree_remove(vma, &mapping->i_mmap);
205 flush_dcache_mmap_unlock(mapping);
209 * Unlink a file-based vm structure from its prio_tree, to hide
210 * vma from rmap and vmtruncate before freeing its page tables.
212 void unlink_file_vma(struct vm_area_struct *vma)
214 struct file *file = vma->vm_file;
217 struct address_space *mapping = file->f_mapping;
218 mutex_lock(&mapping->i_mmap_mutex);
219 __remove_shared_vm_struct(vma, file, mapping);
220 mutex_unlock(&mapping->i_mmap_mutex);
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
256 * CONFIG_COMPAT_BRK can still be overridden by setting
257 * randomize_va_space to 2, which will still cause mm->start_brk
258 * to be arbitrarily shifted
260 if (current->brk_randomized)
261 min_brk = mm->start_brk;
263 min_brk = mm->end_data;
265 min_brk = mm->start_brk;
271 * Check against rlimit here. If this check is done later after the test
272 * of oldbrk with newbrk then it can escape the test and let the data
273 * segment grow beyond its set limit the in case where the limit is
274 * not page aligned -Ram Gupta
276 rlim = rlimit(RLIMIT_DATA);
277 if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
278 (mm->end_data - mm->start_data) > rlim)
281 newbrk = PAGE_ALIGN(brk);
282 oldbrk = PAGE_ALIGN(mm->brk);
283 if (oldbrk == newbrk)
286 /* Always allow shrinking brk. */
287 if (brk <= mm->brk) {
288 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
293 /* Check against existing mmap mappings. */
294 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
297 /* Ok, looks good - let it rip. */
298 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
304 up_write(&mm->mmap_sem);
309 static int browse_rb(struct rb_root *root)
312 struct rb_node *nd, *pn = NULL;
313 unsigned long prev = 0, pend = 0;
315 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
316 struct vm_area_struct *vma;
317 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
318 if (vma->vm_start < prev)
319 printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
320 if (vma->vm_start < pend)
321 printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
322 if (vma->vm_start > vma->vm_end)
323 printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
326 prev = vma->vm_start;
330 for (nd = pn; nd; nd = rb_prev(nd)) {
334 printk("backwards %d, forwards %d\n", j, i), i = 0;
338 void validate_mm(struct mm_struct *mm)
342 struct vm_area_struct *tmp = mm->mmap;
347 if (i != mm->map_count)
348 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
349 i = browse_rb(&mm->mm_rb);
350 if (i != mm->map_count)
351 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
355 #define validate_mm(mm) do { } while (0)
358 static struct vm_area_struct *
359 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
360 struct vm_area_struct **pprev, struct rb_node ***rb_link,
361 struct rb_node ** rb_parent)
363 struct vm_area_struct * vma;
364 struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
366 __rb_link = &mm->mm_rb.rb_node;
367 rb_prev = __rb_parent = NULL;
371 struct vm_area_struct *vma_tmp;
373 __rb_parent = *__rb_link;
374 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
376 if (vma_tmp->vm_end > addr) {
378 if (vma_tmp->vm_start <= addr)
380 __rb_link = &__rb_parent->rb_left;
382 rb_prev = __rb_parent;
383 __rb_link = &__rb_parent->rb_right;
389 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
390 *rb_link = __rb_link;
391 *rb_parent = __rb_parent;
395 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
396 struct rb_node **rb_link, struct rb_node *rb_parent)
398 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
399 rb_insert_color(&vma->vm_rb, &mm->mm_rb);
402 static void __vma_link_file(struct vm_area_struct *vma)
408 struct address_space *mapping = file->f_mapping;
410 if (vma->vm_flags & VM_DENYWRITE)
411 atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
412 if (vma->vm_flags & VM_SHARED)
413 mapping->i_mmap_writable++;
415 flush_dcache_mmap_lock(mapping);
416 if (unlikely(vma->vm_flags & VM_NONLINEAR))
417 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
419 vma_prio_tree_insert(vma, &mapping->i_mmap);
420 flush_dcache_mmap_unlock(mapping);
425 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
426 struct vm_area_struct *prev, struct rb_node **rb_link,
427 struct rb_node *rb_parent)
429 __vma_link_list(mm, vma, prev, rb_parent);
430 __vma_link_rb(mm, vma, rb_link, rb_parent);
433 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
434 struct vm_area_struct *prev, struct rb_node **rb_link,
435 struct rb_node *rb_parent)
437 struct address_space *mapping = NULL;
440 mapping = vma->vm_file->f_mapping;
443 mutex_lock(&mapping->i_mmap_mutex);
445 __vma_link(mm, vma, prev, rb_link, rb_parent);
446 __vma_link_file(vma);
449 mutex_unlock(&mapping->i_mmap_mutex);
456 * Helper for vma_adjust in the split_vma insert case:
457 * insert vm structure into list and rbtree and anon_vma,
458 * but it has already been inserted into prio_tree earlier.
460 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
462 struct vm_area_struct *__vma, *prev;
463 struct rb_node **rb_link, *rb_parent;
465 __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
466 BUG_ON(__vma && __vma->vm_start < vma->vm_end);
467 __vma_link(mm, vma, prev, rb_link, rb_parent);
472 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
473 struct vm_area_struct *prev)
475 struct vm_area_struct *next = vma->vm_next;
477 prev->vm_next = next;
479 next->vm_prev = prev;
480 rb_erase(&vma->vm_rb, &mm->mm_rb);
481 if (mm->mmap_cache == vma)
482 mm->mmap_cache = prev;
486 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
487 * is already present in an i_mmap tree without adjusting the tree.
488 * The following helper function should be used when such adjustments
489 * are necessary. The "insert" vma (if any) is to be inserted
490 * before we drop the necessary locks.
492 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
493 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
495 struct mm_struct *mm = vma->vm_mm;
496 struct vm_area_struct *next = vma->vm_next;
497 struct vm_area_struct *importer = NULL;
498 struct address_space *mapping = NULL;
499 struct prio_tree_root *root = NULL;
500 struct anon_vma *anon_vma = NULL;
501 struct file *file = vma->vm_file;
502 long adjust_next = 0;
505 if (next && !insert) {
506 struct vm_area_struct *exporter = NULL;
508 if (end >= next->vm_end) {
510 * vma expands, overlapping all the next, and
511 * perhaps the one after too (mprotect case 6).
513 again: remove_next = 1 + (end > next->vm_end);
517 } else if (end > next->vm_start) {
519 * vma expands, overlapping part of the next:
520 * mprotect case 5 shifting the boundary up.
522 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
525 } else if (end < vma->vm_end) {
527 * vma shrinks, and !insert tells it's not
528 * split_vma inserting another: so it must be
529 * mprotect case 4 shifting the boundary down.
531 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
537 * Easily overlooked: when mprotect shifts the boundary,
538 * make sure the expanding vma has anon_vma set if the
539 * shrinking vma had, to cover any anon pages imported.
541 if (exporter && exporter->anon_vma && !importer->anon_vma) {
542 if (anon_vma_clone(importer, exporter))
544 importer->anon_vma = exporter->anon_vma;
549 mapping = file->f_mapping;
550 if (!(vma->vm_flags & VM_NONLINEAR)) {
551 root = &mapping->i_mmap;
558 mutex_lock(&mapping->i_mmap_mutex);
561 * Put into prio_tree now, so instantiated pages
562 * are visible to arm/parisc __flush_dcache_page
563 * throughout; but we cannot insert into address
564 * space until vma start or end is updated.
566 __vma_link_file(insert);
570 vma_adjust_trans_huge(vma, start, end, adjust_next);
573 * When changing only vma->vm_end, we don't really need anon_vma
574 * lock. This is a fairly rare case by itself, but the anon_vma
575 * lock may be shared between many sibling processes. Skipping
576 * the lock for brk adjustments makes a difference sometimes.
578 if (vma->anon_vma && (importer || start != vma->vm_start)) {
579 anon_vma = vma->anon_vma;
580 anon_vma_lock(anon_vma);
584 flush_dcache_mmap_lock(mapping);
585 vma_prio_tree_remove(vma, root);
587 vma_prio_tree_remove(next, root);
590 vma->vm_start = start;
592 vma->vm_pgoff = pgoff;
594 next->vm_start += adjust_next << PAGE_SHIFT;
595 next->vm_pgoff += adjust_next;
600 vma_prio_tree_insert(next, root);
601 vma_prio_tree_insert(vma, root);
602 flush_dcache_mmap_unlock(mapping);
607 * vma_merge has merged next into vma, and needs
608 * us to remove next before dropping the locks.
610 __vma_unlink(mm, next, vma);
612 __remove_shared_vm_struct(next, file, mapping);
615 * split_vma has split insert from vma, and needs
616 * us to insert it before dropping the locks
617 * (it may either follow vma or precede it).
619 __insert_vm_struct(mm, insert);
623 anon_vma_unlock(anon_vma);
625 mutex_unlock(&mapping->i_mmap_mutex);
638 if (next->vm_flags & VM_EXECUTABLE)
639 removed_exe_file_vma(mm);
642 anon_vma_merge(vma, next);
644 mpol_put(vma_policy(next));
645 kmem_cache_free(vm_area_cachep, next);
647 * In mprotect's case 6 (see comments on vma_merge),
648 * we must remove another next too. It would clutter
649 * up the code too much to do both in one go.
651 if (remove_next == 2) {
665 * If the vma has a ->close operation then the driver probably needs to release
666 * per-vma resources, so we don't attempt to merge those.
668 static inline int is_mergeable_vma(struct vm_area_struct *vma,
669 struct file *file, unsigned long vm_flags)
671 /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
672 if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
674 if (vma->vm_file != file)
676 if (vma->vm_ops && vma->vm_ops->close)
681 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
682 struct anon_vma *anon_vma2,
683 struct vm_area_struct *vma)
686 * The list_is_singular() test is to avoid merging VMA cloned from
687 * parents. This can improve scalability caused by anon_vma lock.
689 if ((!anon_vma1 || !anon_vma2) && (!vma ||
690 list_is_singular(&vma->anon_vma_chain)))
692 return 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, 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, 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,
808 next->anon_vma, NULL)) {
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);
817 khugepaged_enter_vma_merge(prev);
822 * Can this new request be merged in front of next?
824 if (next && end == next->vm_start &&
825 mpol_equal(policy, vma_policy(next)) &&
826 can_vma_merge_before(next, vm_flags,
827 anon_vma, file, pgoff+pglen)) {
828 if (prev && addr < prev->vm_end) /* case 4 */
829 err = vma_adjust(prev, prev->vm_start,
830 addr, prev->vm_pgoff, NULL);
831 else /* cases 3, 8 */
832 err = vma_adjust(area, addr, next->vm_end,
833 next->vm_pgoff - pglen, NULL);
836 khugepaged_enter_vma_merge(area);
844 * Rough compatbility check to quickly see if it's even worth looking
845 * at sharing an anon_vma.
847 * They need to have the same vm_file, and the flags can only differ
848 * in things that mprotect may change.
850 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
851 * we can merge the two vma's. For example, we refuse to merge a vma if
852 * there is a vm_ops->close() function, because that indicates that the
853 * driver is doing some kind of reference counting. But that doesn't
854 * really matter for the anon_vma sharing case.
856 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
858 return a->vm_end == b->vm_start &&
859 mpol_equal(vma_policy(a), vma_policy(b)) &&
860 a->vm_file == b->vm_file &&
861 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
862 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
866 * Do some basic sanity checking to see if we can re-use the anon_vma
867 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
868 * the same as 'old', the other will be the new one that is trying
869 * to share the anon_vma.
871 * NOTE! This runs with mm_sem held for reading, so it is possible that
872 * the anon_vma of 'old' is concurrently in the process of being set up
873 * by another page fault trying to merge _that_. But that's ok: if it
874 * is being set up, that automatically means that it will be a singleton
875 * acceptable for merging, so we can do all of this optimistically. But
876 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
878 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
879 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
880 * is to return an anon_vma that is "complex" due to having gone through
883 * We also make sure that the two vma's are compatible (adjacent,
884 * and with the same memory policies). That's all stable, even with just
885 * a read lock on the mm_sem.
887 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
889 if (anon_vma_compatible(a, b)) {
890 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
892 if (anon_vma && list_is_singular(&old->anon_vma_chain))
899 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
900 * neighbouring vmas for a suitable anon_vma, before it goes off
901 * to allocate a new anon_vma. It checks because a repetitive
902 * sequence of mprotects and faults may otherwise lead to distinct
903 * anon_vmas being allocated, preventing vma merge in subsequent
906 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
908 struct anon_vma *anon_vma;
909 struct vm_area_struct *near;
915 anon_vma = reusable_anon_vma(near, vma, near);
923 anon_vma = reusable_anon_vma(near, near, vma);
928 * There's no absolute need to look only at touching neighbours:
929 * we could search further afield for "compatible" anon_vmas.
930 * But it would probably just be a waste of time searching,
931 * or lead to too many vmas hanging off the same anon_vma.
932 * We're trying to allow mprotect remerging later on,
933 * not trying to minimize memory used for anon_vmas.
938 #ifdef CONFIG_PROC_FS
939 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
940 struct file *file, long pages)
942 const unsigned long stack_flags
943 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
946 mm->shared_vm += pages;
947 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
948 mm->exec_vm += pages;
949 } else if (flags & stack_flags)
950 mm->stack_vm += pages;
951 if (flags & (VM_RESERVED|VM_IO))
952 mm->reserved_vm += pages;
954 #endif /* CONFIG_PROC_FS */
957 * The caller must hold down_write(¤t->mm->mmap_sem).
960 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
961 unsigned long len, unsigned long prot,
962 unsigned long flags, unsigned long pgoff)
964 struct mm_struct * mm = current->mm;
968 unsigned long reqprot = prot;
971 * Does the application expect PROT_READ to imply PROT_EXEC?
973 * (the exception is when the underlying filesystem is noexec
974 * mounted, in which case we dont add PROT_EXEC.)
976 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
977 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
983 if (!(flags & MAP_FIXED))
984 addr = round_hint_to_min(addr);
986 /* Careful about overflows.. */
987 len = PAGE_ALIGN(len);
991 /* offset overflow? */
992 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
995 /* Too many mappings? */
996 if (mm->map_count > sysctl_max_map_count)
999 /* Obtain the address to map to. we verify (or select) it and ensure
1000 * that it represents a valid section of the address space.
1002 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1003 if (addr & ~PAGE_MASK)
1006 /* Do simple checking here so the lower-level routines won't have
1007 * to. we assume access permissions have been handled by the open
1008 * of the memory object, so we don't do any here.
1010 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1011 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1013 if (flags & MAP_LOCKED)
1014 if (!can_do_mlock())
1017 /* mlock MCL_FUTURE? */
1018 if (vm_flags & VM_LOCKED) {
1019 unsigned long locked, lock_limit;
1020 locked = len >> PAGE_SHIFT;
1021 locked += mm->locked_vm;
1022 lock_limit = rlimit(RLIMIT_MEMLOCK);
1023 lock_limit >>= PAGE_SHIFT;
1024 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1028 inode = file ? file->f_path.dentry->d_inode : NULL;
1031 switch (flags & MAP_TYPE) {
1033 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1037 * Make sure we don't allow writing to an append-only
1040 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1044 * Make sure there are no mandatory locks on the file.
1046 if (locks_verify_locked(inode))
1049 vm_flags |= VM_SHARED | VM_MAYSHARE;
1050 if (!(file->f_mode & FMODE_WRITE))
1051 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1055 if (!(file->f_mode & FMODE_READ))
1057 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1058 if (vm_flags & VM_EXEC)
1060 vm_flags &= ~VM_MAYEXEC;
1063 if (!file->f_op || !file->f_op->mmap)
1071 switch (flags & MAP_TYPE) {
1077 vm_flags |= VM_SHARED | VM_MAYSHARE;
1081 * Set pgoff according to addr for anon_vma.
1083 pgoff = addr >> PAGE_SHIFT;
1090 error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1094 return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1096 EXPORT_SYMBOL(do_mmap_pgoff);
1098 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1099 unsigned long, prot, unsigned long, flags,
1100 unsigned long, fd, unsigned long, pgoff)
1102 struct file *file = NULL;
1103 unsigned long retval = -EBADF;
1105 if (!(flags & MAP_ANONYMOUS)) {
1106 audit_mmap_fd(fd, flags);
1107 if (unlikely(flags & MAP_HUGETLB))
1112 } else if (flags & MAP_HUGETLB) {
1113 struct user_struct *user = NULL;
1115 * VM_NORESERVE is used because the reservations will be
1116 * taken when vm_ops->mmap() is called
1117 * A dummy user value is used because we are not locking
1118 * memory so no accounting is necessary
1120 len = ALIGN(len, huge_page_size(&default_hstate));
1121 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
1122 &user, HUGETLB_ANONHUGE_INODE);
1124 return PTR_ERR(file);
1127 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1129 down_write(¤t->mm->mmap_sem);
1130 retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1131 up_write(¤t->mm->mmap_sem);
1139 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1140 struct mmap_arg_struct {
1144 unsigned long flags;
1146 unsigned long offset;
1149 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1151 struct mmap_arg_struct a;
1153 if (copy_from_user(&a, arg, sizeof(a)))
1155 if (a.offset & ~PAGE_MASK)
1158 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1159 a.offset >> PAGE_SHIFT);
1161 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1164 * Some shared mappigns will want the pages marked read-only
1165 * to track write events. If so, we'll downgrade vm_page_prot
1166 * to the private version (using protection_map[] without the
1169 int vma_wants_writenotify(struct vm_area_struct *vma)
1171 vm_flags_t vm_flags = vma->vm_flags;
1173 /* If it was private or non-writable, the write bit is already clear */
1174 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1177 /* The backer wishes to know when pages are first written to? */
1178 if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1181 /* The open routine did something to the protections already? */
1182 if (pgprot_val(vma->vm_page_prot) !=
1183 pgprot_val(vm_get_page_prot(vm_flags)))
1186 /* Specialty mapping? */
1187 if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1190 /* Can the mapping track the dirty pages? */
1191 return vma->vm_file && vma->vm_file->f_mapping &&
1192 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1196 * We account for memory if it's a private writeable mapping,
1197 * not hugepages and VM_NORESERVE wasn't set.
1199 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1202 * hugetlb has its own accounting separate from the core VM
1203 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1205 if (file && is_file_hugepages(file))
1208 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1211 unsigned long mmap_region(struct file *file, unsigned long addr,
1212 unsigned long len, unsigned long flags,
1213 vm_flags_t vm_flags, unsigned long pgoff)
1215 struct mm_struct *mm = current->mm;
1216 struct vm_area_struct *vma, *prev;
1217 int correct_wcount = 0;
1219 struct rb_node **rb_link, *rb_parent;
1220 unsigned long charged = 0;
1221 struct inode *inode = file ? file->f_path.dentry->d_inode : NULL;
1223 /* Clear old maps */
1226 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1227 if (vma && vma->vm_start < addr + len) {
1228 if (do_munmap(mm, addr, len))
1233 /* Check against address space limit. */
1234 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1238 * Set 'VM_NORESERVE' if we should not account for the
1239 * memory use of this mapping.
1241 if ((flags & MAP_NORESERVE)) {
1242 /* We honor MAP_NORESERVE if allowed to overcommit */
1243 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1244 vm_flags |= VM_NORESERVE;
1246 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1247 if (file && is_file_hugepages(file))
1248 vm_flags |= VM_NORESERVE;
1252 * Private writable mapping: check memory availability
1254 if (accountable_mapping(file, vm_flags)) {
1255 charged = len >> PAGE_SHIFT;
1256 if (security_vm_enough_memory(charged))
1258 vm_flags |= VM_ACCOUNT;
1262 * Can we just expand an old mapping?
1264 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1269 * Determine the object being mapped and call the appropriate
1270 * specific mapper. the address has already been validated, but
1271 * not unmapped, but the maps are removed from the list.
1273 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1280 vma->vm_start = addr;
1281 vma->vm_end = addr + len;
1282 vma->vm_flags = vm_flags;
1283 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1284 vma->vm_pgoff = pgoff;
1285 INIT_LIST_HEAD(&vma->anon_vma_chain);
1289 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1291 if (vm_flags & VM_DENYWRITE) {
1292 error = deny_write_access(file);
1297 vma->vm_file = file;
1299 error = file->f_op->mmap(file, vma);
1301 goto unmap_and_free_vma;
1302 if (vm_flags & VM_EXECUTABLE)
1303 added_exe_file_vma(mm);
1305 /* Can addr have changed??
1307 * Answer: Yes, several device drivers can do it in their
1308 * f_op->mmap method. -DaveM
1310 addr = vma->vm_start;
1311 pgoff = vma->vm_pgoff;
1312 vm_flags = vma->vm_flags;
1313 } else if (vm_flags & VM_SHARED) {
1314 error = shmem_zero_setup(vma);
1319 if (vma_wants_writenotify(vma)) {
1320 pgprot_t pprot = vma->vm_page_prot;
1322 /* Can vma->vm_page_prot have changed??
1324 * Answer: Yes, drivers may have changed it in their
1325 * f_op->mmap method.
1327 * Ensures that vmas marked as uncached stay that way.
1329 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1330 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1331 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1334 vma_link(mm, vma, prev, rb_link, rb_parent);
1335 file = vma->vm_file;
1337 /* Once vma denies write, undo our temporary denial count */
1339 atomic_inc(&inode->i_writecount);
1341 perf_event_mmap(vma);
1343 mm->total_vm += len >> PAGE_SHIFT;
1344 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1345 if (vm_flags & VM_LOCKED) {
1346 if (!mlock_vma_pages_range(vma, addr, addr + len))
1347 mm->locked_vm += (len >> PAGE_SHIFT);
1348 } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1349 make_pages_present(addr, addr + len);
1351 if (file && mmap_uprobe(vma))
1352 /* matching probes but cannot insert */
1353 goto unmap_and_free_vma;
1359 atomic_inc(&inode->i_writecount);
1360 vma->vm_file = NULL;
1363 /* Undo any partial mapping done by a device driver. */
1364 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1367 kmem_cache_free(vm_area_cachep, vma);
1370 vm_unacct_memory(charged);
1374 /* Get an address range which is currently unmapped.
1375 * For shmat() with addr=0.
1377 * Ugly calling convention alert:
1378 * Return value with the low bits set means error value,
1380 * if (ret & ~PAGE_MASK)
1383 * This function "knows" that -ENOMEM has the bits set.
1385 #ifndef HAVE_ARCH_UNMAPPED_AREA
1387 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1388 unsigned long len, unsigned long pgoff, unsigned long flags)
1390 struct mm_struct *mm = current->mm;
1391 struct vm_area_struct *vma;
1392 unsigned long start_addr;
1394 if (len > TASK_SIZE)
1397 if (flags & MAP_FIXED)
1401 addr = PAGE_ALIGN(addr);
1402 vma = find_vma(mm, addr);
1403 if (TASK_SIZE - len >= addr &&
1404 (!vma || addr + len <= vma->vm_start))
1407 if (len > mm->cached_hole_size) {
1408 start_addr = addr = mm->free_area_cache;
1410 start_addr = addr = TASK_UNMAPPED_BASE;
1411 mm->cached_hole_size = 0;
1415 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1416 /* At this point: (!vma || addr < vma->vm_end). */
1417 if (TASK_SIZE - len < addr) {
1419 * Start a new search - just in case we missed
1422 if (start_addr != TASK_UNMAPPED_BASE) {
1423 addr = TASK_UNMAPPED_BASE;
1425 mm->cached_hole_size = 0;
1430 if (!vma || addr + len <= vma->vm_start) {
1432 * Remember the place where we stopped the search:
1434 mm->free_area_cache = addr + len;
1437 if (addr + mm->cached_hole_size < vma->vm_start)
1438 mm->cached_hole_size = vma->vm_start - addr;
1444 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1447 * Is this a new hole at the lowest possible address?
1449 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1450 mm->free_area_cache = addr;
1451 mm->cached_hole_size = ~0UL;
1456 * This mmap-allocator allocates new areas top-down from below the
1457 * stack's low limit (the base):
1459 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1461 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1462 const unsigned long len, const unsigned long pgoff,
1463 const unsigned long flags)
1465 struct vm_area_struct *vma;
1466 struct mm_struct *mm = current->mm;
1467 unsigned long addr = addr0;
1469 /* requested length too big for entire address space */
1470 if (len > TASK_SIZE)
1473 if (flags & MAP_FIXED)
1476 /* requesting a specific address */
1478 addr = PAGE_ALIGN(addr);
1479 vma = find_vma(mm, addr);
1480 if (TASK_SIZE - len >= addr &&
1481 (!vma || addr + len <= vma->vm_start))
1485 /* check if free_area_cache is useful for us */
1486 if (len <= mm->cached_hole_size) {
1487 mm->cached_hole_size = 0;
1488 mm->free_area_cache = mm->mmap_base;
1491 /* either no address requested or can't fit in requested address hole */
1492 addr = mm->free_area_cache;
1494 /* make sure it can fit in the remaining address space */
1496 vma = find_vma(mm, addr-len);
1497 if (!vma || addr <= vma->vm_start)
1498 /* remember the address as a hint for next time */
1499 return (mm->free_area_cache = addr-len);
1502 if (mm->mmap_base < len)
1505 addr = mm->mmap_base-len;
1509 * Lookup failure means no vma is above this address,
1510 * else if new region fits below vma->vm_start,
1511 * return with success:
1513 vma = find_vma(mm, addr);
1514 if (!vma || addr+len <= vma->vm_start)
1515 /* remember the address as a hint for next time */
1516 return (mm->free_area_cache = addr);
1518 /* remember the largest hole we saw so far */
1519 if (addr + mm->cached_hole_size < vma->vm_start)
1520 mm->cached_hole_size = vma->vm_start - addr;
1522 /* try just below the current vma->vm_start */
1523 addr = vma->vm_start-len;
1524 } while (len < vma->vm_start);
1528 * A failed mmap() very likely causes application failure,
1529 * so fall back to the bottom-up function here. This scenario
1530 * can happen with large stack limits and large mmap()
1533 mm->cached_hole_size = ~0UL;
1534 mm->free_area_cache = TASK_UNMAPPED_BASE;
1535 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1537 * Restore the topdown base:
1539 mm->free_area_cache = mm->mmap_base;
1540 mm->cached_hole_size = ~0UL;
1546 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1549 * Is this a new hole at the highest possible address?
1551 if (addr > mm->free_area_cache)
1552 mm->free_area_cache = addr;
1554 /* dont allow allocations above current base */
1555 if (mm->free_area_cache > mm->mmap_base)
1556 mm->free_area_cache = mm->mmap_base;
1560 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1561 unsigned long pgoff, unsigned long flags)
1563 unsigned long (*get_area)(struct file *, unsigned long,
1564 unsigned long, unsigned long, unsigned long);
1566 unsigned long error = arch_mmap_check(addr, len, flags);
1570 /* Careful about overflows.. */
1571 if (len > TASK_SIZE)
1574 get_area = current->mm->get_unmapped_area;
1575 if (file && file->f_op && file->f_op->get_unmapped_area)
1576 get_area = file->f_op->get_unmapped_area;
1577 addr = get_area(file, addr, len, pgoff, flags);
1578 if (IS_ERR_VALUE(addr))
1581 if (addr > TASK_SIZE - len)
1583 if (addr & ~PAGE_MASK)
1586 return arch_rebalance_pgtables(addr, len);
1589 EXPORT_SYMBOL(get_unmapped_area);
1591 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1592 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1594 struct vm_area_struct *vma = NULL;
1597 /* Check the cache first. */
1598 /* (Cache hit rate is typically around 35%.) */
1599 vma = mm->mmap_cache;
1600 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1601 struct rb_node * rb_node;
1603 rb_node = mm->mm_rb.rb_node;
1607 struct vm_area_struct * vma_tmp;
1609 vma_tmp = rb_entry(rb_node,
1610 struct vm_area_struct, vm_rb);
1612 if (vma_tmp->vm_end > addr) {
1614 if (vma_tmp->vm_start <= addr)
1616 rb_node = rb_node->rb_left;
1618 rb_node = rb_node->rb_right;
1621 mm->mmap_cache = vma;
1627 EXPORT_SYMBOL(find_vma);
1629 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1630 struct vm_area_struct *
1631 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1632 struct vm_area_struct **pprev)
1634 struct vm_area_struct *vma = NULL, *prev = NULL;
1635 struct rb_node *rb_node;
1639 /* Guard against addr being lower than the first VMA */
1642 /* Go through the RB tree quickly. */
1643 rb_node = mm->mm_rb.rb_node;
1646 struct vm_area_struct *vma_tmp;
1647 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1649 if (addr < vma_tmp->vm_end) {
1650 rb_node = rb_node->rb_left;
1653 if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1655 rb_node = rb_node->rb_right;
1661 return prev ? prev->vm_next : vma;
1665 * Verify that the stack growth is acceptable and
1666 * update accounting. This is shared with both the
1667 * grow-up and grow-down cases.
1669 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1671 struct mm_struct *mm = vma->vm_mm;
1672 struct rlimit *rlim = current->signal->rlim;
1673 unsigned long new_start;
1675 /* address space limit tests */
1676 if (!may_expand_vm(mm, grow))
1679 /* Stack limit test */
1680 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1683 /* mlock limit tests */
1684 if (vma->vm_flags & VM_LOCKED) {
1685 unsigned long locked;
1686 unsigned long limit;
1687 locked = mm->locked_vm + grow;
1688 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1689 limit >>= PAGE_SHIFT;
1690 if (locked > limit && !capable(CAP_IPC_LOCK))
1694 /* Check to ensure the stack will not grow into a hugetlb-only region */
1695 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1697 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1701 * Overcommit.. This must be the final test, as it will
1702 * update security statistics.
1704 if (security_vm_enough_memory_mm(mm, grow))
1707 /* Ok, everything looks good - let it rip */
1708 mm->total_vm += grow;
1709 if (vma->vm_flags & VM_LOCKED)
1710 mm->locked_vm += grow;
1711 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1715 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1717 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1718 * vma is the last one with address > vma->vm_end. Have to extend vma.
1720 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1724 if (!(vma->vm_flags & VM_GROWSUP))
1728 * We must make sure the anon_vma is allocated
1729 * so that the anon_vma locking is not a noop.
1731 if (unlikely(anon_vma_prepare(vma)))
1733 vma_lock_anon_vma(vma);
1736 * vma->vm_start/vm_end cannot change under us because the caller
1737 * is required to hold the mmap_sem in read mode. We need the
1738 * anon_vma lock to serialize against concurrent expand_stacks.
1739 * Also guard against wrapping around to address 0.
1741 if (address < PAGE_ALIGN(address+4))
1742 address = PAGE_ALIGN(address+4);
1744 vma_unlock_anon_vma(vma);
1749 /* Somebody else might have raced and expanded it already */
1750 if (address > vma->vm_end) {
1751 unsigned long size, grow;
1753 size = address - vma->vm_start;
1754 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1757 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1758 error = acct_stack_growth(vma, size, grow);
1760 vma->vm_end = address;
1761 perf_event_mmap(vma);
1765 vma_unlock_anon_vma(vma);
1766 khugepaged_enter_vma_merge(vma);
1769 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1772 * vma is the first one with address < vma->vm_start. Have to extend vma.
1774 int expand_downwards(struct vm_area_struct *vma,
1775 unsigned long address)
1780 * We must make sure the anon_vma is allocated
1781 * so that the anon_vma locking is not a noop.
1783 if (unlikely(anon_vma_prepare(vma)))
1786 address &= PAGE_MASK;
1787 error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1791 vma_lock_anon_vma(vma);
1794 * vma->vm_start/vm_end cannot change under us because the caller
1795 * is required to hold the mmap_sem in read mode. We need the
1796 * anon_vma lock to serialize against concurrent expand_stacks.
1799 /* Somebody else might have raced and expanded it already */
1800 if (address < vma->vm_start) {
1801 unsigned long size, grow;
1803 size = vma->vm_end - address;
1804 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1807 if (grow <= vma->vm_pgoff) {
1808 error = acct_stack_growth(vma, size, grow);
1810 vma->vm_start = address;
1811 vma->vm_pgoff -= grow;
1812 perf_event_mmap(vma);
1816 vma_unlock_anon_vma(vma);
1817 khugepaged_enter_vma_merge(vma);
1821 #ifdef CONFIG_STACK_GROWSUP
1822 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1824 return expand_upwards(vma, address);
1827 struct vm_area_struct *
1828 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1830 struct vm_area_struct *vma, *prev;
1833 vma = find_vma_prev(mm, addr, &prev);
1834 if (vma && (vma->vm_start <= addr))
1836 if (!prev || expand_stack(prev, addr))
1838 if (prev->vm_flags & VM_LOCKED) {
1839 mlock_vma_pages_range(prev, addr, prev->vm_end);
1844 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1846 return expand_downwards(vma, address);
1849 struct vm_area_struct *
1850 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1852 struct vm_area_struct * vma;
1853 unsigned long start;
1856 vma = find_vma(mm,addr);
1859 if (vma->vm_start <= addr)
1861 if (!(vma->vm_flags & VM_GROWSDOWN))
1863 start = vma->vm_start;
1864 if (expand_stack(vma, addr))
1866 if (vma->vm_flags & VM_LOCKED) {
1867 mlock_vma_pages_range(vma, addr, start);
1874 * Ok - we have the memory areas we should free on the vma list,
1875 * so release them, and do the vma updates.
1877 * Called with the mm semaphore held.
1879 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1881 /* Update high watermark before we lower total_vm */
1882 update_hiwater_vm(mm);
1884 long nrpages = vma_pages(vma);
1886 mm->total_vm -= nrpages;
1887 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1888 vma = remove_vma(vma);
1894 * Get rid of page table information in the indicated region.
1896 * Called with the mm semaphore held.
1898 static void unmap_region(struct mm_struct *mm,
1899 struct vm_area_struct *vma, struct vm_area_struct *prev,
1900 unsigned long start, unsigned long end)
1902 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1903 struct mmu_gather tlb;
1904 unsigned long nr_accounted = 0;
1907 tlb_gather_mmu(&tlb, mm, 0);
1908 update_hiwater_rss(mm);
1909 unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1910 vm_unacct_memory(nr_accounted);
1911 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
1912 next ? next->vm_start : 0);
1913 tlb_finish_mmu(&tlb, start, end);
1917 * Create a list of vma's touched by the unmap, removing them from the mm's
1918 * vma list as we go..
1921 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1922 struct vm_area_struct *prev, unsigned long end)
1924 struct vm_area_struct **insertion_point;
1925 struct vm_area_struct *tail_vma = NULL;
1928 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1929 vma->vm_prev = NULL;
1931 rb_erase(&vma->vm_rb, &mm->mm_rb);
1935 } while (vma && vma->vm_start < end);
1936 *insertion_point = vma;
1938 vma->vm_prev = prev;
1939 tail_vma->vm_next = NULL;
1940 if (mm->unmap_area == arch_unmap_area)
1941 addr = prev ? prev->vm_end : mm->mmap_base;
1943 addr = vma ? vma->vm_start : mm->mmap_base;
1944 mm->unmap_area(mm, addr);
1945 mm->mmap_cache = NULL; /* Kill the cache. */
1949 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
1950 * munmap path where it doesn't make sense to fail.
1952 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1953 unsigned long addr, int new_below)
1955 struct mempolicy *pol;
1956 struct vm_area_struct *new;
1959 if (is_vm_hugetlb_page(vma) && (addr &
1960 ~(huge_page_mask(hstate_vma(vma)))))
1963 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1967 /* most fields are the same, copy all, and then fixup */
1970 INIT_LIST_HEAD(&new->anon_vma_chain);
1975 new->vm_start = addr;
1976 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1979 pol = mpol_dup(vma_policy(vma));
1984 vma_set_policy(new, pol);
1986 if (anon_vma_clone(new, vma))
1990 get_file(new->vm_file);
1991 if (vma->vm_flags & VM_EXECUTABLE)
1992 added_exe_file_vma(mm);
1995 if (new->vm_ops && new->vm_ops->open)
1996 new->vm_ops->open(new);
1999 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2000 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2002 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2008 /* Clean everything up if vma_adjust failed. */
2009 if (new->vm_ops && new->vm_ops->close)
2010 new->vm_ops->close(new);
2012 if (vma->vm_flags & VM_EXECUTABLE)
2013 removed_exe_file_vma(mm);
2016 unlink_anon_vmas(new);
2020 kmem_cache_free(vm_area_cachep, new);
2026 * Split a vma into two pieces at address 'addr', a new vma is allocated
2027 * either for the first part or the tail.
2029 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2030 unsigned long addr, int new_below)
2032 if (mm->map_count >= sysctl_max_map_count)
2035 return __split_vma(mm, vma, addr, new_below);
2038 /* Munmap is split into 2 main parts -- this part which finds
2039 * what needs doing, and the areas themselves, which do the
2040 * work. This now handles partial unmappings.
2041 * Jeremy Fitzhardinge <jeremy@goop.org>
2043 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2046 struct vm_area_struct *vma, *prev, *last;
2048 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2051 if ((len = PAGE_ALIGN(len)) == 0)
2054 /* Find the first overlapping VMA */
2055 vma = find_vma(mm, start);
2058 prev = vma->vm_prev;
2059 /* we have start < vma->vm_end */
2061 /* if it doesn't overlap, we have nothing.. */
2063 if (vma->vm_start >= end)
2067 * If we need to split any vma, do it now to save pain later.
2069 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2070 * unmapped vm_area_struct will remain in use: so lower split_vma
2071 * places tmp vma above, and higher split_vma places tmp vma below.
2073 if (start > vma->vm_start) {
2077 * Make sure that map_count on return from munmap() will
2078 * not exceed its limit; but let map_count go just above
2079 * its limit temporarily, to help free resources as expected.
2081 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2084 error = __split_vma(mm, vma, start, 0);
2090 /* Does it split the last one? */
2091 last = find_vma(mm, end);
2092 if (last && end > last->vm_start) {
2093 int error = __split_vma(mm, last, end, 1);
2097 vma = prev? prev->vm_next: mm->mmap;
2100 * unlock any mlock()ed ranges before detaching vmas
2102 if (mm->locked_vm) {
2103 struct vm_area_struct *tmp = vma;
2104 while (tmp && tmp->vm_start < end) {
2105 if (tmp->vm_flags & VM_LOCKED) {
2106 mm->locked_vm -= vma_pages(tmp);
2107 munlock_vma_pages_all(tmp);
2114 * Remove the vma's, and unmap the actual pages
2116 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2117 unmap_region(mm, vma, prev, start, end);
2119 /* Fix up all other VM information */
2120 remove_vma_list(mm, vma);
2125 EXPORT_SYMBOL(do_munmap);
2127 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2130 struct mm_struct *mm = current->mm;
2132 profile_munmap(addr);
2134 down_write(&mm->mmap_sem);
2135 ret = do_munmap(mm, addr, len);
2136 up_write(&mm->mmap_sem);
2140 static inline void verify_mm_writelocked(struct mm_struct *mm)
2142 #ifdef CONFIG_DEBUG_VM
2143 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2145 up_read(&mm->mmap_sem);
2151 * this is really a simplified "do_mmap". it only handles
2152 * anonymous maps. eventually we may be able to do some
2153 * brk-specific accounting here.
2155 unsigned long do_brk(unsigned long addr, unsigned long len)
2157 struct mm_struct * mm = current->mm;
2158 struct vm_area_struct * vma, * prev;
2159 unsigned long flags;
2160 struct rb_node ** rb_link, * rb_parent;
2161 pgoff_t pgoff = addr >> PAGE_SHIFT;
2164 len = PAGE_ALIGN(len);
2168 error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2172 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2174 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2175 if (error & ~PAGE_MASK)
2181 if (mm->def_flags & VM_LOCKED) {
2182 unsigned long locked, lock_limit;
2183 locked = len >> PAGE_SHIFT;
2184 locked += mm->locked_vm;
2185 lock_limit = rlimit(RLIMIT_MEMLOCK);
2186 lock_limit >>= PAGE_SHIFT;
2187 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2192 * mm->mmap_sem is required to protect against another thread
2193 * changing the mappings in case we sleep.
2195 verify_mm_writelocked(mm);
2198 * Clear old maps. this also does some error checking for us
2201 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2202 if (vma && vma->vm_start < addr + len) {
2203 if (do_munmap(mm, addr, len))
2208 /* Check against address space limits *after* clearing old maps... */
2209 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2212 if (mm->map_count > sysctl_max_map_count)
2215 if (security_vm_enough_memory(len >> PAGE_SHIFT))
2218 /* Can we just expand an old private anonymous mapping? */
2219 vma = vma_merge(mm, prev, addr, addr + len, flags,
2220 NULL, NULL, pgoff, NULL);
2225 * create a vma struct for an anonymous mapping
2227 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2229 vm_unacct_memory(len >> PAGE_SHIFT);
2233 INIT_LIST_HEAD(&vma->anon_vma_chain);
2235 vma->vm_start = addr;
2236 vma->vm_end = addr + len;
2237 vma->vm_pgoff = pgoff;
2238 vma->vm_flags = flags;
2239 vma->vm_page_prot = vm_get_page_prot(flags);
2240 vma_link(mm, vma, prev, rb_link, rb_parent);
2242 perf_event_mmap(vma);
2243 mm->total_vm += len >> PAGE_SHIFT;
2244 if (flags & VM_LOCKED) {
2245 if (!mlock_vma_pages_range(vma, addr, addr + len))
2246 mm->locked_vm += (len >> PAGE_SHIFT);
2251 EXPORT_SYMBOL(do_brk);
2253 /* Release all mmaps. */
2254 void exit_mmap(struct mm_struct *mm)
2256 struct mmu_gather tlb;
2257 struct vm_area_struct *vma;
2258 unsigned long nr_accounted = 0;
2261 /* mm's last user has gone, and its about to be pulled down */
2262 mmu_notifier_release(mm);
2264 if (mm->locked_vm) {
2267 if (vma->vm_flags & VM_LOCKED)
2268 munlock_vma_pages_all(vma);
2276 if (!vma) /* Can happen if dup_mmap() received an OOM */
2281 tlb_gather_mmu(&tlb, mm, 1);
2282 /* update_hiwater_rss(mm) here? but nobody should be looking */
2283 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2284 end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2285 vm_unacct_memory(nr_accounted);
2287 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2288 tlb_finish_mmu(&tlb, 0, end);
2291 * Walk the list again, actually closing and freeing it,
2292 * with preemption enabled, without holding any MM locks.
2295 vma = remove_vma(vma);
2297 BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2300 /* Insert vm structure into process list sorted by address
2301 * and into the inode's i_mmap tree. If vm_file is non-NULL
2302 * then i_mmap_mutex is taken here.
2304 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2306 struct vm_area_struct * __vma, * prev;
2307 struct rb_node ** rb_link, * rb_parent;
2310 * The vm_pgoff of a purely anonymous vma should be irrelevant
2311 * until its first write fault, when page's anon_vma and index
2312 * are set. But now set the vm_pgoff it will almost certainly
2313 * end up with (unless mremap moves it elsewhere before that
2314 * first wfault), so /proc/pid/maps tells a consistent story.
2316 * By setting it to reflect the virtual start address of the
2317 * vma, merges and splits can happen in a seamless way, just
2318 * using the existing file pgoff checks and manipulations.
2319 * Similarly in do_mmap_pgoff and in do_brk.
2321 if (!vma->vm_file) {
2322 BUG_ON(vma->anon_vma);
2323 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2325 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2326 if (__vma && __vma->vm_start < vma->vm_end)
2328 if ((vma->vm_flags & VM_ACCOUNT) &&
2329 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2332 if (vma->vm_file && mmap_uprobe(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);
2387 if (mmap_uprobe(new_vma))
2388 goto out_free_mempol;
2390 if (vma->vm_flags & VM_EXECUTABLE)
2391 added_exe_file_vma(mm);
2393 if (new_vma->vm_ops && new_vma->vm_ops->open)
2394 new_vma->vm_ops->open(new_vma);
2395 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2403 kmem_cache_free(vm_area_cachep, new_vma);
2408 * Return true if the calling process may expand its vm space by the passed
2411 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2413 unsigned long cur = mm->total_vm; /* pages */
2416 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2418 if (cur + npages > lim)
2424 static int special_mapping_fault(struct vm_area_struct *vma,
2425 struct vm_fault *vmf)
2428 struct page **pages;
2431 * special mappings have no vm_file, and in that case, the mm
2432 * uses vm_pgoff internally. So we have to subtract it from here.
2433 * We are allowed to do this because we are the mm; do not copy
2434 * this code into drivers!
2436 pgoff = vmf->pgoff - vma->vm_pgoff;
2438 for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2442 struct page *page = *pages;
2448 return VM_FAULT_SIGBUS;
2452 * Having a close hook prevents vma merging regardless of flags.
2454 static void special_mapping_close(struct vm_area_struct *vma)
2458 static const struct vm_operations_struct special_mapping_vmops = {
2459 .close = special_mapping_close,
2460 .fault = special_mapping_fault,
2464 * Called with mm->mmap_sem held for writing.
2465 * Insert a new vma covering the given region, with the given flags.
2466 * Its pages are supplied by the given array of struct page *.
2467 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2468 * The region past the last page supplied will always produce SIGBUS.
2469 * The array pointer and the pages it points to are assumed to stay alive
2470 * for as long as this mapping might exist.
2472 int install_special_mapping(struct mm_struct *mm,
2473 unsigned long addr, unsigned long len,
2474 unsigned long vm_flags, struct page **pages)
2477 struct vm_area_struct *vma;
2479 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2480 if (unlikely(vma == NULL))
2483 INIT_LIST_HEAD(&vma->anon_vma_chain);
2485 vma->vm_start = addr;
2486 vma->vm_end = addr + len;
2488 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2489 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2491 vma->vm_ops = &special_mapping_vmops;
2492 vma->vm_private_data = pages;
2494 ret = security_file_mmap(NULL, 0, 0, 0, vma->vm_start, 1);
2498 ret = insert_vm_struct(mm, vma);
2502 mm->total_vm += len >> PAGE_SHIFT;
2504 perf_event_mmap(vma);
2509 kmem_cache_free(vm_area_cachep, vma);
2513 static DEFINE_MUTEX(mm_all_locks_mutex);
2515 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2517 if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2519 * The LSB of head.next can't change from under us
2520 * because we hold the mm_all_locks_mutex.
2522 mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem);
2524 * We can safely modify head.next after taking the
2525 * anon_vma->root->mutex. If some other vma in this mm shares
2526 * the same anon_vma we won't take it again.
2528 * No need of atomic instructions here, head.next
2529 * can't change from under us thanks to the
2530 * anon_vma->root->mutex.
2532 if (__test_and_set_bit(0, (unsigned long *)
2533 &anon_vma->root->head.next))
2538 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2540 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2542 * AS_MM_ALL_LOCKS can't change from under us because
2543 * we hold the mm_all_locks_mutex.
2545 * Operations on ->flags have to be atomic because
2546 * even if AS_MM_ALL_LOCKS is stable thanks to the
2547 * mm_all_locks_mutex, there may be other cpus
2548 * changing other bitflags in parallel to us.
2550 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2552 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2557 * This operation locks against the VM for all pte/vma/mm related
2558 * operations that could ever happen on a certain mm. This includes
2559 * vmtruncate, try_to_unmap, and all page faults.
2561 * The caller must take the mmap_sem in write mode before calling
2562 * mm_take_all_locks(). The caller isn't allowed to release the
2563 * mmap_sem until mm_drop_all_locks() returns.
2565 * mmap_sem in write mode is required in order to block all operations
2566 * that could modify pagetables and free pages without need of
2567 * altering the vma layout (for example populate_range() with
2568 * nonlinear vmas). It's also needed in write mode to avoid new
2569 * anon_vmas to be associated with existing vmas.
2571 * A single task can't take more than one mm_take_all_locks() in a row
2572 * or it would deadlock.
2574 * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2575 * mapping->flags avoid to take the same lock twice, if more than one
2576 * vma in this mm is backed by the same anon_vma or address_space.
2578 * We can take all the locks in random order because the VM code
2579 * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2580 * takes more than one of them in a row. Secondly we're protected
2581 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2583 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2584 * that may have to take thousand of locks.
2586 * mm_take_all_locks() can fail if it's interrupted by signals.
2588 int mm_take_all_locks(struct mm_struct *mm)
2590 struct vm_area_struct *vma;
2591 struct anon_vma_chain *avc;
2593 BUG_ON(down_read_trylock(&mm->mmap_sem));
2595 mutex_lock(&mm_all_locks_mutex);
2597 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2598 if (signal_pending(current))
2600 if (vma->vm_file && vma->vm_file->f_mapping)
2601 vm_lock_mapping(mm, vma->vm_file->f_mapping);
2604 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2605 if (signal_pending(current))
2608 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2609 vm_lock_anon_vma(mm, avc->anon_vma);
2615 mm_drop_all_locks(mm);
2619 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2621 if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2623 * The LSB of head.next can't change to 0 from under
2624 * us because we hold the mm_all_locks_mutex.
2626 * We must however clear the bitflag before unlocking
2627 * the vma so the users using the anon_vma->head will
2628 * never see our bitflag.
2630 * No need of atomic instructions here, head.next
2631 * can't change from under us until we release the
2632 * anon_vma->root->mutex.
2634 if (!__test_and_clear_bit(0, (unsigned long *)
2635 &anon_vma->root->head.next))
2637 anon_vma_unlock(anon_vma);
2641 static void vm_unlock_mapping(struct address_space *mapping)
2643 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2645 * AS_MM_ALL_LOCKS can't change to 0 from under us
2646 * because we hold the mm_all_locks_mutex.
2648 mutex_unlock(&mapping->i_mmap_mutex);
2649 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2656 * The mmap_sem cannot be released by the caller until
2657 * mm_drop_all_locks() returns.
2659 void mm_drop_all_locks(struct mm_struct *mm)
2661 struct vm_area_struct *vma;
2662 struct anon_vma_chain *avc;
2664 BUG_ON(down_read_trylock(&mm->mmap_sem));
2665 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2667 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2669 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2670 vm_unlock_anon_vma(avc->anon_vma);
2671 if (vma->vm_file && vma->vm_file->f_mapping)
2672 vm_unlock_mapping(vma->vm_file->f_mapping);
2675 mutex_unlock(&mm_all_locks_mutex);
2679 * initialise the VMA slab
2681 void __init mmap_init(void)
2685 ret = percpu_counter_init(&vm_committed_as, 0);