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>
34 #include <asm/uaccess.h>
35 #include <asm/cacheflush.h>
37 #include <asm/mmu_context.h>
41 #ifndef arch_mmap_check
42 #define arch_mmap_check(addr, len, flags) (0)
45 #ifndef arch_rebalance_pgtables
46 #define arch_rebalance_pgtables(addr, len) (addr)
49 static void unmap_region(struct mm_struct *mm,
50 struct vm_area_struct *vma, struct vm_area_struct *prev,
51 unsigned long start, unsigned long end);
54 * WARNING: the debugging will use recursive algorithms so never enable this
55 * unless you know what you are doing.
59 /* description of effects of mapping type and prot in current implementation.
60 * this is due to the limited x86 page protection hardware. The expected
61 * behavior is in parens:
64 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
65 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
66 * w: (no) no w: (no) no w: (yes) yes w: (no) no
67 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
69 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
70 * w: (no) no w: (no) no w: (copy) copy w: (no) no
71 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
74 pgprot_t protection_map[16] = {
75 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
76 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
79 pgprot_t vm_get_page_prot(unsigned long vm_flags)
81 return __pgprot(pgprot_val(protection_map[vm_flags &
82 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
83 pgprot_val(arch_vm_get_page_prot(vm_flags)));
85 EXPORT_SYMBOL(vm_get_page_prot);
87 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS; /* heuristic overcommit */
88 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
89 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
91 * Make sure vm_committed_as in one cacheline and not cacheline shared with
92 * other variables. It can be updated by several CPUs frequently.
94 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
97 * Check that a process has enough memory to allocate a new virtual
98 * mapping. 0 means there is enough memory for the allocation to
99 * succeed and -ENOMEM implies there is not.
101 * We currently support three overcommit policies, which are set via the
102 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
104 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
105 * Additional code 2002 Jul 20 by Robert Love.
107 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
109 * Note this is a helper function intended to be used by LSMs which
110 * wish to use this logic.
112 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
114 unsigned long free, allowed;
116 vm_acct_memory(pages);
119 * Sometimes we want to use more memory than we have
121 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
124 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
125 free = global_page_state(NR_FREE_PAGES);
126 free += global_page_state(NR_FILE_PAGES);
129 * shmem pages shouldn't be counted as free in this
130 * case, they can't be purged, only swapped out, and
131 * that won't affect the overall amount of available
132 * memory in the system.
134 free -= global_page_state(NR_SHMEM);
136 free += nr_swap_pages;
139 * Any slabs which are created with the
140 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
141 * which are reclaimable, under pressure. The dentry
142 * cache and most inode caches should fall into this
144 free += global_page_state(NR_SLAB_RECLAIMABLE);
147 * Leave reserved pages. The pages are not for anonymous pages.
149 if (free <= totalreserve_pages)
152 free -= totalreserve_pages;
155 * Leave the last 3% for root
166 allowed = (totalram_pages - hugetlb_total_pages())
167 * sysctl_overcommit_ratio / 100;
169 * Leave the last 3% for root
172 allowed -= allowed / 32;
173 allowed += total_swap_pages;
175 /* Don't let a single process grow too big:
176 leave 3% of the size of this process for other processes */
178 allowed -= mm->total_vm / 32;
180 if (percpu_counter_read_positive(&vm_committed_as) < allowed)
183 vm_unacct_memory(pages);
189 * Requires inode->i_mapping->i_mmap_mutex
191 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
192 struct file *file, struct address_space *mapping)
194 if (vma->vm_flags & VM_DENYWRITE)
195 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
196 if (vma->vm_flags & VM_SHARED)
197 mapping->i_mmap_writable--;
199 flush_dcache_mmap_lock(mapping);
200 if (unlikely(vma->vm_flags & VM_NONLINEAR))
201 list_del_init(&vma->shared.vm_set.list);
203 vma_prio_tree_remove(vma, &mapping->i_mmap);
204 flush_dcache_mmap_unlock(mapping);
208 * Unlink a file-based vm structure from its prio_tree, to hide
209 * vma from rmap and vmtruncate before freeing its page tables.
211 void unlink_file_vma(struct vm_area_struct *vma)
213 struct file *file = vma->vm_file;
216 struct address_space *mapping = file->f_mapping;
217 mutex_lock(&mapping->i_mmap_mutex);
218 __remove_shared_vm_struct(vma, file, mapping);
219 mutex_unlock(&mapping->i_mmap_mutex);
224 * Close a vm structure and free it, returning the next.
226 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
228 struct vm_area_struct *next = vma->vm_next;
231 if (vma->vm_ops && vma->vm_ops->close)
232 vma->vm_ops->close(vma);
235 if (vma->vm_flags & VM_EXECUTABLE)
236 removed_exe_file_vma(vma->vm_mm);
238 mpol_put(vma_policy(vma));
239 kmem_cache_free(vm_area_cachep, vma);
243 SYSCALL_DEFINE1(brk, unsigned long, brk)
245 unsigned long rlim, retval;
246 unsigned long newbrk, oldbrk;
247 struct mm_struct *mm = current->mm;
248 unsigned long min_brk;
250 down_write(&mm->mmap_sem);
252 #ifdef CONFIG_COMPAT_BRK
254 * CONFIG_COMPAT_BRK can still be overridden by setting
255 * randomize_va_space to 2, which will still cause mm->start_brk
256 * to be arbitrarily shifted
258 if (current->brk_randomized)
259 min_brk = mm->start_brk;
261 min_brk = mm->end_data;
263 min_brk = mm->start_brk;
269 * Check against rlimit here. If this check is done later after the test
270 * of oldbrk with newbrk then it can escape the test and let the data
271 * segment grow beyond its set limit the in case where the limit is
272 * not page aligned -Ram Gupta
274 rlim = rlimit(RLIMIT_DATA);
275 if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
276 (mm->end_data - mm->start_data) > rlim)
279 newbrk = PAGE_ALIGN(brk);
280 oldbrk = PAGE_ALIGN(mm->brk);
281 if (oldbrk == newbrk)
284 /* Always allow shrinking brk. */
285 if (brk <= mm->brk) {
286 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
291 /* Check against existing mmap mappings. */
292 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
295 /* Ok, looks good - let it rip. */
296 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
302 up_write(&mm->mmap_sem);
307 static int browse_rb(struct rb_root *root)
310 struct rb_node *nd, *pn = NULL;
311 unsigned long prev = 0, pend = 0;
313 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
314 struct vm_area_struct *vma;
315 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
316 if (vma->vm_start < prev)
317 printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
318 if (vma->vm_start < pend)
319 printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
320 if (vma->vm_start > vma->vm_end)
321 printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
324 prev = vma->vm_start;
328 for (nd = pn; nd; nd = rb_prev(nd)) {
332 printk("backwards %d, forwards %d\n", j, i), i = 0;
336 void validate_mm(struct mm_struct *mm)
340 struct vm_area_struct *tmp = mm->mmap;
345 if (i != mm->map_count)
346 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
347 i = browse_rb(&mm->mm_rb);
348 if (i != mm->map_count)
349 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
353 #define validate_mm(mm) do { } while (0)
356 static struct vm_area_struct *
357 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
358 struct vm_area_struct **pprev, struct rb_node ***rb_link,
359 struct rb_node ** rb_parent)
361 struct vm_area_struct * vma;
362 struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
364 __rb_link = &mm->mm_rb.rb_node;
365 rb_prev = __rb_parent = NULL;
369 struct vm_area_struct *vma_tmp;
371 __rb_parent = *__rb_link;
372 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
374 if (vma_tmp->vm_end > addr) {
376 if (vma_tmp->vm_start <= addr)
378 __rb_link = &__rb_parent->rb_left;
380 rb_prev = __rb_parent;
381 __rb_link = &__rb_parent->rb_right;
387 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
388 *rb_link = __rb_link;
389 *rb_parent = __rb_parent;
393 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
394 struct rb_node **rb_link, struct rb_node *rb_parent)
396 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
397 rb_insert_color(&vma->vm_rb, &mm->mm_rb);
400 static void __vma_link_file(struct vm_area_struct *vma)
406 struct address_space *mapping = file->f_mapping;
408 if (vma->vm_flags & VM_DENYWRITE)
409 atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
410 if (vma->vm_flags & VM_SHARED)
411 mapping->i_mmap_writable++;
413 flush_dcache_mmap_lock(mapping);
414 if (unlikely(vma->vm_flags & VM_NONLINEAR))
415 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
417 vma_prio_tree_insert(vma, &mapping->i_mmap);
418 flush_dcache_mmap_unlock(mapping);
423 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
424 struct vm_area_struct *prev, struct rb_node **rb_link,
425 struct rb_node *rb_parent)
427 __vma_link_list(mm, vma, prev, rb_parent);
428 __vma_link_rb(mm, vma, rb_link, rb_parent);
431 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
432 struct vm_area_struct *prev, struct rb_node **rb_link,
433 struct rb_node *rb_parent)
435 struct address_space *mapping = NULL;
438 mapping = vma->vm_file->f_mapping;
441 mutex_lock(&mapping->i_mmap_mutex);
443 __vma_link(mm, vma, prev, rb_link, rb_parent);
444 __vma_link_file(vma);
447 mutex_unlock(&mapping->i_mmap_mutex);
454 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
455 * mm's list and rbtree. It has already been inserted into the prio_tree.
457 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
459 struct vm_area_struct *__vma, *prev;
460 struct rb_node **rb_link, *rb_parent;
462 __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
463 BUG_ON(__vma && __vma->vm_start < vma->vm_end);
464 __vma_link(mm, vma, prev, rb_link, rb_parent);
469 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
470 struct vm_area_struct *prev)
472 struct vm_area_struct *next = vma->vm_next;
474 prev->vm_next = next;
476 next->vm_prev = prev;
477 rb_erase(&vma->vm_rb, &mm->mm_rb);
478 if (mm->mmap_cache == vma)
479 mm->mmap_cache = prev;
483 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
484 * is already present in an i_mmap tree without adjusting the tree.
485 * The following helper function should be used when such adjustments
486 * are necessary. The "insert" vma (if any) is to be inserted
487 * before we drop the necessary locks.
489 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
490 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
492 struct mm_struct *mm = vma->vm_mm;
493 struct vm_area_struct *next = vma->vm_next;
494 struct vm_area_struct *importer = NULL;
495 struct address_space *mapping = NULL;
496 struct prio_tree_root *root = NULL;
497 struct anon_vma *anon_vma = NULL;
498 struct file *file = vma->vm_file;
499 long adjust_next = 0;
502 if (next && !insert) {
503 struct vm_area_struct *exporter = NULL;
505 if (end >= next->vm_end) {
507 * vma expands, overlapping all the next, and
508 * perhaps the one after too (mprotect case 6).
510 again: remove_next = 1 + (end > next->vm_end);
514 } else if (end > next->vm_start) {
516 * vma expands, overlapping part of the next:
517 * mprotect case 5 shifting the boundary up.
519 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
522 } else if (end < vma->vm_end) {
524 * vma shrinks, and !insert tells it's not
525 * split_vma inserting another: so it must be
526 * mprotect case 4 shifting the boundary down.
528 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
534 * Easily overlooked: when mprotect shifts the boundary,
535 * make sure the expanding vma has anon_vma set if the
536 * shrinking vma had, to cover any anon pages imported.
538 if (exporter && exporter->anon_vma && !importer->anon_vma) {
539 if (anon_vma_clone(importer, exporter))
541 importer->anon_vma = exporter->anon_vma;
546 mapping = file->f_mapping;
547 if (!(vma->vm_flags & VM_NONLINEAR))
548 root = &mapping->i_mmap;
549 mutex_lock(&mapping->i_mmap_mutex);
552 * Put into prio_tree now, so instantiated pages
553 * are visible to arm/parisc __flush_dcache_page
554 * throughout; but we cannot insert into address
555 * space until vma start or end is updated.
557 __vma_link_file(insert);
561 vma_adjust_trans_huge(vma, start, end, adjust_next);
564 * When changing only vma->vm_end, we don't really need anon_vma
565 * lock. This is a fairly rare case by itself, but the anon_vma
566 * lock may be shared between many sibling processes. Skipping
567 * the lock for brk adjustments makes a difference sometimes.
569 if (vma->anon_vma && (importer || start != vma->vm_start)) {
570 anon_vma = vma->anon_vma;
571 anon_vma_lock(anon_vma);
575 flush_dcache_mmap_lock(mapping);
576 vma_prio_tree_remove(vma, root);
578 vma_prio_tree_remove(next, root);
581 vma->vm_start = start;
583 vma->vm_pgoff = pgoff;
585 next->vm_start += adjust_next << PAGE_SHIFT;
586 next->vm_pgoff += adjust_next;
591 vma_prio_tree_insert(next, root);
592 vma_prio_tree_insert(vma, root);
593 flush_dcache_mmap_unlock(mapping);
598 * vma_merge has merged next into vma, and needs
599 * us to remove next before dropping the locks.
601 __vma_unlink(mm, next, vma);
603 __remove_shared_vm_struct(next, file, mapping);
606 * split_vma has split insert from vma, and needs
607 * us to insert it before dropping the locks
608 * (it may either follow vma or precede it).
610 __insert_vm_struct(mm, insert);
614 anon_vma_unlock(anon_vma);
616 mutex_unlock(&mapping->i_mmap_mutex);
621 if (next->vm_flags & VM_EXECUTABLE)
622 removed_exe_file_vma(mm);
625 anon_vma_merge(vma, next);
627 mpol_put(vma_policy(next));
628 kmem_cache_free(vm_area_cachep, next);
630 * In mprotect's case 6 (see comments on vma_merge),
631 * we must remove another next too. It would clutter
632 * up the code too much to do both in one go.
634 if (remove_next == 2) {
646 * If the vma has a ->close operation then the driver probably needs to release
647 * per-vma resources, so we don't attempt to merge those.
649 static inline int is_mergeable_vma(struct vm_area_struct *vma,
650 struct file *file, unsigned long vm_flags)
652 /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
653 if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
655 if (vma->vm_file != file)
657 if (vma->vm_ops && vma->vm_ops->close)
662 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
663 struct anon_vma *anon_vma2,
664 struct vm_area_struct *vma)
667 * The list_is_singular() test is to avoid merging VMA cloned from
668 * parents. This can improve scalability caused by anon_vma lock.
670 if ((!anon_vma1 || !anon_vma2) && (!vma ||
671 list_is_singular(&vma->anon_vma_chain)))
673 return anon_vma1 == anon_vma2;
677 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
678 * in front of (at a lower virtual address and file offset than) the vma.
680 * We cannot merge two vmas if they have differently assigned (non-NULL)
681 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
683 * We don't check here for the merged mmap wrapping around the end of pagecache
684 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
685 * wrap, nor mmaps which cover the final page at index -1UL.
688 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
689 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
691 if (is_mergeable_vma(vma, file, vm_flags) &&
692 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
693 if (vma->vm_pgoff == vm_pgoff)
700 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
701 * beyond (at a higher virtual address and file offset than) the vma.
703 * We cannot merge two vmas if they have differently assigned (non-NULL)
704 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
707 can_vma_merge_after(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)) {
713 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
714 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
721 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
722 * whether that can be merged with its predecessor or its successor.
723 * Or both (it neatly fills a hole).
725 * In most cases - when called for mmap, brk or mremap - [addr,end) is
726 * certain not to be mapped by the time vma_merge is called; but when
727 * called for mprotect, it is certain to be already mapped (either at
728 * an offset within prev, or at the start of next), and the flags of
729 * this area are about to be changed to vm_flags - and the no-change
730 * case has already been eliminated.
732 * The following mprotect cases have to be considered, where AAAA is
733 * the area passed down from mprotect_fixup, never extending beyond one
734 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
736 * AAAA AAAA AAAA AAAA
737 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
738 * cannot merge might become might become might become
739 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
740 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
741 * mremap move: PPPPNNNNNNNN 8
743 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
744 * might become case 1 below case 2 below case 3 below
746 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
747 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
749 struct vm_area_struct *vma_merge(struct mm_struct *mm,
750 struct vm_area_struct *prev, unsigned long addr,
751 unsigned long end, unsigned long vm_flags,
752 struct anon_vma *anon_vma, struct file *file,
753 pgoff_t pgoff, struct mempolicy *policy)
755 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
756 struct vm_area_struct *area, *next;
760 * We later require that vma->vm_flags == vm_flags,
761 * so this tests vma->vm_flags & VM_SPECIAL, too.
763 if (vm_flags & VM_SPECIAL)
767 next = prev->vm_next;
771 if (next && next->vm_end == end) /* cases 6, 7, 8 */
772 next = next->vm_next;
775 * Can it merge with the predecessor?
777 if (prev && prev->vm_end == addr &&
778 mpol_equal(vma_policy(prev), policy) &&
779 can_vma_merge_after(prev, vm_flags,
780 anon_vma, file, pgoff)) {
782 * OK, it can. Can we now merge in the successor as well?
784 if (next && end == next->vm_start &&
785 mpol_equal(policy, vma_policy(next)) &&
786 can_vma_merge_before(next, vm_flags,
787 anon_vma, file, pgoff+pglen) &&
788 is_mergeable_anon_vma(prev->anon_vma,
789 next->anon_vma, NULL)) {
791 err = vma_adjust(prev, prev->vm_start,
792 next->vm_end, prev->vm_pgoff, NULL);
793 } else /* cases 2, 5, 7 */
794 err = vma_adjust(prev, prev->vm_start,
795 end, prev->vm_pgoff, NULL);
798 khugepaged_enter_vma_merge(prev);
803 * Can this new request be merged in front of next?
805 if (next && end == next->vm_start &&
806 mpol_equal(policy, vma_policy(next)) &&
807 can_vma_merge_before(next, vm_flags,
808 anon_vma, file, pgoff+pglen)) {
809 if (prev && addr < prev->vm_end) /* case 4 */
810 err = vma_adjust(prev, prev->vm_start,
811 addr, prev->vm_pgoff, NULL);
812 else /* cases 3, 8 */
813 err = vma_adjust(area, addr, next->vm_end,
814 next->vm_pgoff - pglen, NULL);
817 khugepaged_enter_vma_merge(area);
825 * Rough compatbility check to quickly see if it's even worth looking
826 * at sharing an anon_vma.
828 * They need to have the same vm_file, and the flags can only differ
829 * in things that mprotect may change.
831 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
832 * we can merge the two vma's. For example, we refuse to merge a vma if
833 * there is a vm_ops->close() function, because that indicates that the
834 * driver is doing some kind of reference counting. But that doesn't
835 * really matter for the anon_vma sharing case.
837 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
839 return a->vm_end == b->vm_start &&
840 mpol_equal(vma_policy(a), vma_policy(b)) &&
841 a->vm_file == b->vm_file &&
842 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
843 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
847 * Do some basic sanity checking to see if we can re-use the anon_vma
848 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
849 * the same as 'old', the other will be the new one that is trying
850 * to share the anon_vma.
852 * NOTE! This runs with mm_sem held for reading, so it is possible that
853 * the anon_vma of 'old' is concurrently in the process of being set up
854 * by another page fault trying to merge _that_. But that's ok: if it
855 * is being set up, that automatically means that it will be a singleton
856 * acceptable for merging, so we can do all of this optimistically. But
857 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
859 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
860 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
861 * is to return an anon_vma that is "complex" due to having gone through
864 * We also make sure that the two vma's are compatible (adjacent,
865 * and with the same memory policies). That's all stable, even with just
866 * a read lock on the mm_sem.
868 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
870 if (anon_vma_compatible(a, b)) {
871 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
873 if (anon_vma && list_is_singular(&old->anon_vma_chain))
880 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
881 * neighbouring vmas for a suitable anon_vma, before it goes off
882 * to allocate a new anon_vma. It checks because a repetitive
883 * sequence of mprotects and faults may otherwise lead to distinct
884 * anon_vmas being allocated, preventing vma merge in subsequent
887 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
889 struct anon_vma *anon_vma;
890 struct vm_area_struct *near;
896 anon_vma = reusable_anon_vma(near, vma, near);
904 anon_vma = reusable_anon_vma(near, near, vma);
909 * There's no absolute need to look only at touching neighbours:
910 * we could search further afield for "compatible" anon_vmas.
911 * But it would probably just be a waste of time searching,
912 * or lead to too many vmas hanging off the same anon_vma.
913 * We're trying to allow mprotect remerging later on,
914 * not trying to minimize memory used for anon_vmas.
919 #ifdef CONFIG_PROC_FS
920 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
921 struct file *file, long pages)
923 const unsigned long stack_flags
924 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
927 mm->shared_vm += pages;
928 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
929 mm->exec_vm += pages;
930 } else if (flags & stack_flags)
931 mm->stack_vm += pages;
932 if (flags & (VM_RESERVED|VM_IO))
933 mm->reserved_vm += pages;
935 #endif /* CONFIG_PROC_FS */
938 * The caller must hold down_write(¤t->mm->mmap_sem).
941 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
942 unsigned long len, unsigned long prot,
943 unsigned long flags, unsigned long pgoff)
945 struct mm_struct * mm = current->mm;
949 unsigned long reqprot = prot;
952 * Does the application expect PROT_READ to imply PROT_EXEC?
954 * (the exception is when the underlying filesystem is noexec
955 * mounted, in which case we dont add PROT_EXEC.)
957 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
958 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
964 if (!(flags & MAP_FIXED))
965 addr = round_hint_to_min(addr);
967 /* Careful about overflows.. */
968 len = PAGE_ALIGN(len);
972 /* offset overflow? */
973 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
976 /* Too many mappings? */
977 if (mm->map_count > sysctl_max_map_count)
980 /* Obtain the address to map to. we verify (or select) it and ensure
981 * that it represents a valid section of the address space.
983 addr = get_unmapped_area(file, addr, len, pgoff, flags);
984 if (addr & ~PAGE_MASK)
987 /* Do simple checking here so the lower-level routines won't have
988 * to. we assume access permissions have been handled by the open
989 * of the memory object, so we don't do any here.
991 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
992 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
994 if (flags & MAP_LOCKED)
998 /* mlock MCL_FUTURE? */
999 if (vm_flags & VM_LOCKED) {
1000 unsigned long locked, lock_limit;
1001 locked = len >> PAGE_SHIFT;
1002 locked += mm->locked_vm;
1003 lock_limit = rlimit(RLIMIT_MEMLOCK);
1004 lock_limit >>= PAGE_SHIFT;
1005 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1009 inode = file ? file->f_path.dentry->d_inode : NULL;
1012 switch (flags & MAP_TYPE) {
1014 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1018 * Make sure we don't allow writing to an append-only
1021 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1025 * Make sure there are no mandatory locks on the file.
1027 if (locks_verify_locked(inode))
1030 vm_flags |= VM_SHARED | VM_MAYSHARE;
1031 if (!(file->f_mode & FMODE_WRITE))
1032 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1036 if (!(file->f_mode & FMODE_READ))
1038 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1039 if (vm_flags & VM_EXEC)
1041 vm_flags &= ~VM_MAYEXEC;
1044 if (!file->f_op || !file->f_op->mmap)
1052 switch (flags & MAP_TYPE) {
1058 vm_flags |= VM_SHARED | VM_MAYSHARE;
1062 * Set pgoff according to addr for anon_vma.
1064 pgoff = addr >> PAGE_SHIFT;
1071 error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1075 return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1077 EXPORT_SYMBOL(do_mmap_pgoff);
1079 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1080 unsigned long, prot, unsigned long, flags,
1081 unsigned long, fd, unsigned long, pgoff)
1083 struct file *file = NULL;
1084 unsigned long retval = -EBADF;
1086 if (!(flags & MAP_ANONYMOUS)) {
1087 audit_mmap_fd(fd, flags);
1088 if (unlikely(flags & MAP_HUGETLB))
1093 } else if (flags & MAP_HUGETLB) {
1094 struct user_struct *user = NULL;
1096 * VM_NORESERVE is used because the reservations will be
1097 * taken when vm_ops->mmap() is called
1098 * A dummy user value is used because we are not locking
1099 * memory so no accounting is necessary
1101 file = hugetlb_file_setup(HUGETLB_ANON_FILE, addr, len,
1102 VM_NORESERVE, &user,
1103 HUGETLB_ANONHUGE_INODE);
1105 return PTR_ERR(file);
1108 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1110 down_write(¤t->mm->mmap_sem);
1111 retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1112 up_write(¤t->mm->mmap_sem);
1120 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1121 struct mmap_arg_struct {
1125 unsigned long flags;
1127 unsigned long offset;
1130 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1132 struct mmap_arg_struct a;
1134 if (copy_from_user(&a, arg, sizeof(a)))
1136 if (a.offset & ~PAGE_MASK)
1139 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1140 a.offset >> PAGE_SHIFT);
1142 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1145 * Some shared mappigns will want the pages marked read-only
1146 * to track write events. If so, we'll downgrade vm_page_prot
1147 * to the private version (using protection_map[] without the
1150 int vma_wants_writenotify(struct vm_area_struct *vma)
1152 vm_flags_t vm_flags = vma->vm_flags;
1154 /* If it was private or non-writable, the write bit is already clear */
1155 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1158 /* The backer wishes to know when pages are first written to? */
1159 if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1162 /* The open routine did something to the protections already? */
1163 if (pgprot_val(vma->vm_page_prot) !=
1164 pgprot_val(vm_get_page_prot(vm_flags)))
1167 /* Specialty mapping? */
1168 if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1171 /* Can the mapping track the dirty pages? */
1172 return vma->vm_file && vma->vm_file->f_mapping &&
1173 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1177 * We account for memory if it's a private writeable mapping,
1178 * not hugepages and VM_NORESERVE wasn't set.
1180 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1183 * hugetlb has its own accounting separate from the core VM
1184 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1186 if (file && is_file_hugepages(file))
1189 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1192 unsigned long mmap_region(struct file *file, unsigned long addr,
1193 unsigned long len, unsigned long flags,
1194 vm_flags_t vm_flags, unsigned long pgoff)
1196 struct mm_struct *mm = current->mm;
1197 struct vm_area_struct *vma, *prev;
1198 int correct_wcount = 0;
1200 struct rb_node **rb_link, *rb_parent;
1201 unsigned long charged = 0;
1202 struct inode *inode = file ? file->f_path.dentry->d_inode : NULL;
1204 /* Clear old maps */
1207 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1208 if (vma && vma->vm_start < addr + len) {
1209 if (do_munmap(mm, addr, len))
1214 /* Check against address space limit. */
1215 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1219 * Set 'VM_NORESERVE' if we should not account for the
1220 * memory use of this mapping.
1222 if ((flags & MAP_NORESERVE)) {
1223 /* We honor MAP_NORESERVE if allowed to overcommit */
1224 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1225 vm_flags |= VM_NORESERVE;
1227 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1228 if (file && is_file_hugepages(file))
1229 vm_flags |= VM_NORESERVE;
1233 * Private writable mapping: check memory availability
1235 if (accountable_mapping(file, vm_flags)) {
1236 charged = len >> PAGE_SHIFT;
1237 if (security_vm_enough_memory(charged))
1239 vm_flags |= VM_ACCOUNT;
1243 * Can we just expand an old mapping?
1245 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1250 * Determine the object being mapped and call the appropriate
1251 * specific mapper. the address has already been validated, but
1252 * not unmapped, but the maps are removed from the list.
1254 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1261 vma->vm_start = addr;
1262 vma->vm_end = addr + len;
1263 vma->vm_flags = vm_flags;
1264 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1265 vma->vm_pgoff = pgoff;
1266 INIT_LIST_HEAD(&vma->anon_vma_chain);
1268 error = -EINVAL; /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1271 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1273 if (vm_flags & VM_DENYWRITE) {
1274 error = deny_write_access(file);
1279 vma->vm_file = file;
1281 error = file->f_op->mmap(file, vma);
1283 goto unmap_and_free_vma;
1284 if (vm_flags & VM_EXECUTABLE)
1285 added_exe_file_vma(mm);
1287 /* Can addr have changed??
1289 * Answer: Yes, several device drivers can do it in their
1290 * f_op->mmap method. -DaveM
1292 addr = vma->vm_start;
1293 pgoff = vma->vm_pgoff;
1294 vm_flags = vma->vm_flags;
1295 } else if (vm_flags & VM_SHARED) {
1296 if (unlikely(vm_flags & (VM_GROWSDOWN|VM_GROWSUP)))
1298 error = shmem_zero_setup(vma);
1303 if (vma_wants_writenotify(vma)) {
1304 pgprot_t pprot = vma->vm_page_prot;
1306 /* Can vma->vm_page_prot have changed??
1308 * Answer: Yes, drivers may have changed it in their
1309 * f_op->mmap method.
1311 * Ensures that vmas marked as uncached stay that way.
1313 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1314 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1315 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1318 vma_link(mm, vma, prev, rb_link, rb_parent);
1319 file = vma->vm_file;
1321 /* Once vma denies write, undo our temporary denial count */
1323 atomic_inc(&inode->i_writecount);
1325 perf_event_mmap(vma);
1327 mm->total_vm += len >> PAGE_SHIFT;
1328 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1329 if (vm_flags & VM_LOCKED) {
1330 if (!mlock_vma_pages_range(vma, addr, addr + len))
1331 mm->locked_vm += (len >> PAGE_SHIFT);
1332 } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1333 make_pages_present(addr, addr + len);
1338 atomic_inc(&inode->i_writecount);
1339 vma->vm_file = NULL;
1342 /* Undo any partial mapping done by a device driver. */
1343 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1346 kmem_cache_free(vm_area_cachep, vma);
1349 vm_unacct_memory(charged);
1353 /* Get an address range which is currently unmapped.
1354 * For shmat() with addr=0.
1356 * Ugly calling convention alert:
1357 * Return value with the low bits set means error value,
1359 * if (ret & ~PAGE_MASK)
1362 * This function "knows" that -ENOMEM has the bits set.
1364 #ifndef HAVE_ARCH_UNMAPPED_AREA
1366 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1367 unsigned long len, unsigned long pgoff, unsigned long flags)
1369 struct mm_struct *mm = current->mm;
1370 struct vm_area_struct *vma;
1371 unsigned long start_addr;
1373 if (len > TASK_SIZE)
1376 if (flags & MAP_FIXED)
1380 addr = PAGE_ALIGN(addr);
1381 vma = find_vma(mm, addr);
1382 if (TASK_SIZE - len >= addr &&
1383 (!vma || addr + len <= vma->vm_start))
1386 if (len > mm->cached_hole_size) {
1387 start_addr = addr = mm->free_area_cache;
1389 start_addr = addr = TASK_UNMAPPED_BASE;
1390 mm->cached_hole_size = 0;
1394 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1395 /* At this point: (!vma || addr < vma->vm_end). */
1396 if (TASK_SIZE - len < addr) {
1398 * Start a new search - just in case we missed
1401 if (start_addr != TASK_UNMAPPED_BASE) {
1402 addr = TASK_UNMAPPED_BASE;
1404 mm->cached_hole_size = 0;
1409 if (!vma || addr + len <= vma->vm_start) {
1411 * Remember the place where we stopped the search:
1413 mm->free_area_cache = addr + len;
1416 if (addr + mm->cached_hole_size < vma->vm_start)
1417 mm->cached_hole_size = vma->vm_start - addr;
1423 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1426 * Is this a new hole at the lowest possible address?
1428 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache)
1429 mm->free_area_cache = addr;
1433 * This mmap-allocator allocates new areas top-down from below the
1434 * stack's low limit (the base):
1436 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1438 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1439 const unsigned long len, const unsigned long pgoff,
1440 const unsigned long flags)
1442 struct vm_area_struct *vma;
1443 struct mm_struct *mm = current->mm;
1444 unsigned long addr = addr0, start_addr;
1446 /* requested length too big for entire address space */
1447 if (len > TASK_SIZE)
1450 if (flags & MAP_FIXED)
1453 /* requesting a specific address */
1455 addr = PAGE_ALIGN(addr);
1456 vma = find_vma(mm, addr);
1457 if (TASK_SIZE - len >= addr &&
1458 (!vma || addr + len <= vma->vm_start))
1462 /* check if free_area_cache is useful for us */
1463 if (len <= mm->cached_hole_size) {
1464 mm->cached_hole_size = 0;
1465 mm->free_area_cache = mm->mmap_base;
1469 /* either no address requested or can't fit in requested address hole */
1470 start_addr = addr = mm->free_area_cache;
1478 * Lookup failure means no vma is above this address,
1479 * else if new region fits below vma->vm_start,
1480 * return with success:
1482 vma = find_vma(mm, addr);
1483 if (!vma || addr+len <= vma->vm_start)
1484 /* remember the address as a hint for next time */
1485 return (mm->free_area_cache = addr);
1487 /* remember the largest hole we saw so far */
1488 if (addr + mm->cached_hole_size < vma->vm_start)
1489 mm->cached_hole_size = vma->vm_start - addr;
1491 /* try just below the current vma->vm_start */
1492 addr = vma->vm_start-len;
1493 } while (len < vma->vm_start);
1497 * if hint left us with no space for the requested
1498 * mapping then try again:
1500 * Note: this is different with the case of bottomup
1501 * which does the fully line-search, but we use find_vma
1502 * here that causes some holes skipped.
1504 if (start_addr != mm->mmap_base) {
1505 mm->free_area_cache = mm->mmap_base;
1506 mm->cached_hole_size = 0;
1511 * A failed mmap() very likely causes application failure,
1512 * so fall back to the bottom-up function here. This scenario
1513 * can happen with large stack limits and large mmap()
1516 mm->cached_hole_size = ~0UL;
1517 mm->free_area_cache = TASK_UNMAPPED_BASE;
1518 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1520 * Restore the topdown base:
1522 mm->free_area_cache = mm->mmap_base;
1523 mm->cached_hole_size = ~0UL;
1529 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1532 * Is this a new hole at the highest possible address?
1534 if (addr > mm->free_area_cache)
1535 mm->free_area_cache = addr;
1537 /* dont allow allocations above current base */
1538 if (mm->free_area_cache > mm->mmap_base)
1539 mm->free_area_cache = mm->mmap_base;
1543 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1544 unsigned long pgoff, unsigned long flags)
1546 unsigned long (*get_area)(struct file *, unsigned long,
1547 unsigned long, unsigned long, unsigned long);
1549 unsigned long error = arch_mmap_check(addr, len, flags);
1553 /* Careful about overflows.. */
1554 if (len > TASK_SIZE)
1557 get_area = current->mm->get_unmapped_area;
1558 if (file && file->f_op && file->f_op->get_unmapped_area)
1559 get_area = file->f_op->get_unmapped_area;
1560 addr = get_area(file, addr, len, pgoff, flags);
1561 if (IS_ERR_VALUE(addr))
1564 if (addr > TASK_SIZE - len)
1566 if (addr & ~PAGE_MASK)
1569 return arch_rebalance_pgtables(addr, len);
1572 EXPORT_SYMBOL(get_unmapped_area);
1574 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1575 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1577 struct vm_area_struct *vma = NULL;
1580 /* Check the cache first. */
1581 /* (Cache hit rate is typically around 35%.) */
1582 vma = mm->mmap_cache;
1583 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1584 struct rb_node * rb_node;
1586 rb_node = mm->mm_rb.rb_node;
1590 struct vm_area_struct * vma_tmp;
1592 vma_tmp = rb_entry(rb_node,
1593 struct vm_area_struct, vm_rb);
1595 if (vma_tmp->vm_end > addr) {
1597 if (vma_tmp->vm_start <= addr)
1599 rb_node = rb_node->rb_left;
1601 rb_node = rb_node->rb_right;
1604 mm->mmap_cache = vma;
1610 EXPORT_SYMBOL(find_vma);
1613 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1615 struct vm_area_struct *
1616 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1617 struct vm_area_struct **pprev)
1619 struct vm_area_struct *vma;
1621 vma = find_vma(mm, addr);
1623 *pprev = vma->vm_prev;
1625 struct rb_node *rb_node = mm->mm_rb.rb_node;
1628 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1629 rb_node = rb_node->rb_right;
1636 * Verify that the stack growth is acceptable and
1637 * update accounting. This is shared with both the
1638 * grow-up and grow-down cases.
1640 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1642 struct mm_struct *mm = vma->vm_mm;
1643 struct rlimit *rlim = current->signal->rlim;
1644 unsigned long new_start;
1646 /* address space limit tests */
1647 if (!may_expand_vm(mm, grow))
1650 /* Stack limit test */
1651 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1654 /* mlock limit tests */
1655 if (vma->vm_flags & VM_LOCKED) {
1656 unsigned long locked;
1657 unsigned long limit;
1658 locked = mm->locked_vm + grow;
1659 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1660 limit >>= PAGE_SHIFT;
1661 if (locked > limit && !capable(CAP_IPC_LOCK))
1665 /* Check to ensure the stack will not grow into a hugetlb-only region */
1666 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1668 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1672 * Overcommit.. This must be the final test, as it will
1673 * update security statistics.
1675 if (security_vm_enough_memory_mm(mm, grow))
1678 /* Ok, everything looks good - let it rip */
1679 mm->total_vm += grow;
1680 if (vma->vm_flags & VM_LOCKED)
1681 mm->locked_vm += grow;
1682 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1686 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1688 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1689 * vma is the last one with address > vma->vm_end. Have to extend vma.
1691 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1695 if (!(vma->vm_flags & VM_GROWSUP))
1699 * We must make sure the anon_vma is allocated
1700 * so that the anon_vma locking is not a noop.
1702 if (unlikely(anon_vma_prepare(vma)))
1704 vma_lock_anon_vma(vma);
1707 * vma->vm_start/vm_end cannot change under us because the caller
1708 * is required to hold the mmap_sem in read mode. We need the
1709 * anon_vma lock to serialize against concurrent expand_stacks.
1710 * Also guard against wrapping around to address 0.
1712 if (address < PAGE_ALIGN(address+4))
1713 address = PAGE_ALIGN(address+4);
1715 vma_unlock_anon_vma(vma);
1720 /* Somebody else might have raced and expanded it already */
1721 if (address > vma->vm_end) {
1722 unsigned long size, grow;
1724 size = address - vma->vm_start;
1725 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1728 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1729 error = acct_stack_growth(vma, size, grow);
1731 vma->vm_end = address;
1732 perf_event_mmap(vma);
1736 vma_unlock_anon_vma(vma);
1737 khugepaged_enter_vma_merge(vma);
1740 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1743 * vma is the first one with address < vma->vm_start. Have to extend vma.
1745 int expand_downwards(struct vm_area_struct *vma,
1746 unsigned long address)
1751 * We must make sure the anon_vma is allocated
1752 * so that the anon_vma locking is not a noop.
1754 if (unlikely(anon_vma_prepare(vma)))
1757 address &= PAGE_MASK;
1758 error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1762 vma_lock_anon_vma(vma);
1765 * vma->vm_start/vm_end cannot change under us because the caller
1766 * is required to hold the mmap_sem in read mode. We need the
1767 * anon_vma lock to serialize against concurrent expand_stacks.
1770 /* Somebody else might have raced and expanded it already */
1771 if (address < vma->vm_start) {
1772 unsigned long size, grow;
1774 size = vma->vm_end - address;
1775 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1778 if (grow <= vma->vm_pgoff) {
1779 error = acct_stack_growth(vma, size, grow);
1781 vma->vm_start = address;
1782 vma->vm_pgoff -= grow;
1783 perf_event_mmap(vma);
1787 vma_unlock_anon_vma(vma);
1788 khugepaged_enter_vma_merge(vma);
1792 #ifdef CONFIG_STACK_GROWSUP
1793 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1795 return expand_upwards(vma, address);
1798 struct vm_area_struct *
1799 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1801 struct vm_area_struct *vma, *prev;
1804 vma = find_vma_prev(mm, addr, &prev);
1805 if (vma && (vma->vm_start <= addr))
1807 if (!prev || expand_stack(prev, addr))
1809 if (prev->vm_flags & VM_LOCKED) {
1810 mlock_vma_pages_range(prev, addr, prev->vm_end);
1815 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1817 return expand_downwards(vma, address);
1820 struct vm_area_struct *
1821 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1823 struct vm_area_struct * vma;
1824 unsigned long start;
1827 vma = find_vma(mm,addr);
1830 if (vma->vm_start <= addr)
1832 if (!(vma->vm_flags & VM_GROWSDOWN))
1834 start = vma->vm_start;
1835 if (expand_stack(vma, addr))
1837 if (vma->vm_flags & VM_LOCKED) {
1838 mlock_vma_pages_range(vma, addr, start);
1845 * Ok - we have the memory areas we should free on the vma list,
1846 * so release them, and do the vma updates.
1848 * Called with the mm semaphore held.
1850 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1852 /* Update high watermark before we lower total_vm */
1853 update_hiwater_vm(mm);
1855 long nrpages = vma_pages(vma);
1857 mm->total_vm -= nrpages;
1858 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1859 vma = remove_vma(vma);
1865 * Get rid of page table information in the indicated region.
1867 * Called with the mm semaphore held.
1869 static void unmap_region(struct mm_struct *mm,
1870 struct vm_area_struct *vma, struct vm_area_struct *prev,
1871 unsigned long start, unsigned long end)
1873 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1874 struct mmu_gather tlb;
1875 unsigned long nr_accounted = 0;
1878 tlb_gather_mmu(&tlb, mm, 0);
1879 update_hiwater_rss(mm);
1880 unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1881 vm_unacct_memory(nr_accounted);
1882 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
1883 next ? next->vm_start : 0);
1884 tlb_finish_mmu(&tlb, start, end);
1888 * Create a list of vma's touched by the unmap, removing them from the mm's
1889 * vma list as we go..
1892 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1893 struct vm_area_struct *prev, unsigned long end)
1895 struct vm_area_struct **insertion_point;
1896 struct vm_area_struct *tail_vma = NULL;
1899 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1900 vma->vm_prev = NULL;
1902 rb_erase(&vma->vm_rb, &mm->mm_rb);
1906 } while (vma && vma->vm_start < end);
1907 *insertion_point = vma;
1909 vma->vm_prev = prev;
1910 tail_vma->vm_next = NULL;
1911 if (mm->unmap_area == arch_unmap_area)
1912 addr = prev ? prev->vm_end : mm->mmap_base;
1914 addr = vma ? vma->vm_start : mm->mmap_base;
1915 mm->unmap_area(mm, addr);
1916 mm->mmap_cache = NULL; /* Kill the cache. */
1920 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
1921 * munmap path where it doesn't make sense to fail.
1923 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1924 unsigned long addr, int new_below)
1926 struct mempolicy *pol;
1927 struct vm_area_struct *new;
1930 if (is_vm_hugetlb_page(vma) && (addr &
1931 ~(huge_page_mask(hstate_vma(vma)))))
1934 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1938 /* most fields are the same, copy all, and then fixup */
1941 INIT_LIST_HEAD(&new->anon_vma_chain);
1946 new->vm_start = addr;
1947 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1950 pol = mpol_dup(vma_policy(vma));
1955 vma_set_policy(new, pol);
1957 if (anon_vma_clone(new, vma))
1961 get_file(new->vm_file);
1962 if (vma->vm_flags & VM_EXECUTABLE)
1963 added_exe_file_vma(mm);
1966 if (new->vm_ops && new->vm_ops->open)
1967 new->vm_ops->open(new);
1970 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1971 ((addr - new->vm_start) >> PAGE_SHIFT), new);
1973 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1979 /* Clean everything up if vma_adjust failed. */
1980 if (new->vm_ops && new->vm_ops->close)
1981 new->vm_ops->close(new);
1983 if (vma->vm_flags & VM_EXECUTABLE)
1984 removed_exe_file_vma(mm);
1987 unlink_anon_vmas(new);
1991 kmem_cache_free(vm_area_cachep, new);
1997 * Split a vma into two pieces at address 'addr', a new vma is allocated
1998 * either for the first part or the tail.
2000 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2001 unsigned long addr, int new_below)
2003 if (mm->map_count >= sysctl_max_map_count)
2006 return __split_vma(mm, vma, addr, new_below);
2009 /* Munmap is split into 2 main parts -- this part which finds
2010 * what needs doing, and the areas themselves, which do the
2011 * work. This now handles partial unmappings.
2012 * Jeremy Fitzhardinge <jeremy@goop.org>
2014 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2017 struct vm_area_struct *vma, *prev, *last;
2019 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2022 if ((len = PAGE_ALIGN(len)) == 0)
2025 /* Find the first overlapping VMA */
2026 vma = find_vma(mm, start);
2029 prev = vma->vm_prev;
2030 /* we have start < vma->vm_end */
2032 /* if it doesn't overlap, we have nothing.. */
2034 if (vma->vm_start >= end)
2038 * If we need to split any vma, do it now to save pain later.
2040 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2041 * unmapped vm_area_struct will remain in use: so lower split_vma
2042 * places tmp vma above, and higher split_vma places tmp vma below.
2044 if (start > vma->vm_start) {
2048 * Make sure that map_count on return from munmap() will
2049 * not exceed its limit; but let map_count go just above
2050 * its limit temporarily, to help free resources as expected.
2052 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2055 error = __split_vma(mm, vma, start, 0);
2061 /* Does it split the last one? */
2062 last = find_vma(mm, end);
2063 if (last && end > last->vm_start) {
2064 int error = __split_vma(mm, last, end, 1);
2068 vma = prev? prev->vm_next: mm->mmap;
2071 * unlock any mlock()ed ranges before detaching vmas
2073 if (mm->locked_vm) {
2074 struct vm_area_struct *tmp = vma;
2075 while (tmp && tmp->vm_start < end) {
2076 if (tmp->vm_flags & VM_LOCKED) {
2077 mm->locked_vm -= vma_pages(tmp);
2078 munlock_vma_pages_all(tmp);
2085 * Remove the vma's, and unmap the actual pages
2087 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2088 unmap_region(mm, vma, prev, start, end);
2090 /* Fix up all other VM information */
2091 remove_vma_list(mm, vma);
2096 EXPORT_SYMBOL(do_munmap);
2098 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2101 struct mm_struct *mm = current->mm;
2103 profile_munmap(addr);
2105 down_write(&mm->mmap_sem);
2106 ret = do_munmap(mm, addr, len);
2107 up_write(&mm->mmap_sem);
2111 static inline void verify_mm_writelocked(struct mm_struct *mm)
2113 #ifdef CONFIG_DEBUG_VM
2114 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2116 up_read(&mm->mmap_sem);
2122 * this is really a simplified "do_mmap". it only handles
2123 * anonymous maps. eventually we may be able to do some
2124 * brk-specific accounting here.
2126 unsigned long do_brk(unsigned long addr, unsigned long len)
2128 struct mm_struct * mm = current->mm;
2129 struct vm_area_struct * vma, * prev;
2130 unsigned long flags;
2131 struct rb_node ** rb_link, * rb_parent;
2132 pgoff_t pgoff = addr >> PAGE_SHIFT;
2135 len = PAGE_ALIGN(len);
2139 error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2143 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2145 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2146 if (error & ~PAGE_MASK)
2152 if (mm->def_flags & VM_LOCKED) {
2153 unsigned long locked, lock_limit;
2154 locked = len >> PAGE_SHIFT;
2155 locked += mm->locked_vm;
2156 lock_limit = rlimit(RLIMIT_MEMLOCK);
2157 lock_limit >>= PAGE_SHIFT;
2158 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2163 * mm->mmap_sem is required to protect against another thread
2164 * changing the mappings in case we sleep.
2166 verify_mm_writelocked(mm);
2169 * Clear old maps. this also does some error checking for us
2172 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2173 if (vma && vma->vm_start < addr + len) {
2174 if (do_munmap(mm, addr, len))
2179 /* Check against address space limits *after* clearing old maps... */
2180 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2183 if (mm->map_count > sysctl_max_map_count)
2186 if (security_vm_enough_memory(len >> PAGE_SHIFT))
2189 /* Can we just expand an old private anonymous mapping? */
2190 vma = vma_merge(mm, prev, addr, addr + len, flags,
2191 NULL, NULL, pgoff, NULL);
2196 * create a vma struct for an anonymous mapping
2198 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2200 vm_unacct_memory(len >> PAGE_SHIFT);
2204 INIT_LIST_HEAD(&vma->anon_vma_chain);
2206 vma->vm_start = addr;
2207 vma->vm_end = addr + len;
2208 vma->vm_pgoff = pgoff;
2209 vma->vm_flags = flags;
2210 vma->vm_page_prot = vm_get_page_prot(flags);
2211 vma_link(mm, vma, prev, rb_link, rb_parent);
2213 perf_event_mmap(vma);
2214 mm->total_vm += len >> PAGE_SHIFT;
2215 if (flags & VM_LOCKED) {
2216 if (!mlock_vma_pages_range(vma, addr, addr + len))
2217 mm->locked_vm += (len >> PAGE_SHIFT);
2222 EXPORT_SYMBOL(do_brk);
2224 /* Release all mmaps. */
2225 void exit_mmap(struct mm_struct *mm)
2227 struct mmu_gather tlb;
2228 struct vm_area_struct *vma;
2229 unsigned long nr_accounted = 0;
2232 /* mm's last user has gone, and its about to be pulled down */
2233 mmu_notifier_release(mm);
2235 if (mm->locked_vm) {
2238 if (vma->vm_flags & VM_LOCKED)
2239 munlock_vma_pages_all(vma);
2247 if (!vma) /* Can happen if dup_mmap() received an OOM */
2252 tlb_gather_mmu(&tlb, mm, 1);
2253 /* update_hiwater_rss(mm) here? but nobody should be looking */
2254 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2255 end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2256 vm_unacct_memory(nr_accounted);
2258 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2259 tlb_finish_mmu(&tlb, 0, end);
2262 * Walk the list again, actually closing and freeing it,
2263 * with preemption enabled, without holding any MM locks.
2266 vma = remove_vma(vma);
2268 BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2271 /* Insert vm structure into process list sorted by address
2272 * and into the inode's i_mmap tree. If vm_file is non-NULL
2273 * then i_mmap_mutex is taken here.
2275 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2277 struct vm_area_struct * __vma, * prev;
2278 struct rb_node ** rb_link, * rb_parent;
2281 * The vm_pgoff of a purely anonymous vma should be irrelevant
2282 * until its first write fault, when page's anon_vma and index
2283 * are set. But now set the vm_pgoff it will almost certainly
2284 * end up with (unless mremap moves it elsewhere before that
2285 * first wfault), so /proc/pid/maps tells a consistent story.
2287 * By setting it to reflect the virtual start address of the
2288 * vma, merges and splits can happen in a seamless way, just
2289 * using the existing file pgoff checks and manipulations.
2290 * Similarly in do_mmap_pgoff and in do_brk.
2292 if (!vma->vm_file) {
2293 BUG_ON(vma->anon_vma);
2294 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2296 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2297 if (__vma && __vma->vm_start < vma->vm_end)
2299 if ((vma->vm_flags & VM_ACCOUNT) &&
2300 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2302 vma_link(mm, vma, prev, rb_link, rb_parent);
2307 * Copy the vma structure to a new location in the same mm,
2308 * prior to moving page table entries, to effect an mremap move.
2310 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2311 unsigned long addr, unsigned long len, pgoff_t pgoff)
2313 struct vm_area_struct *vma = *vmap;
2314 unsigned long vma_start = vma->vm_start;
2315 struct mm_struct *mm = vma->vm_mm;
2316 struct vm_area_struct *new_vma, *prev;
2317 struct rb_node **rb_link, *rb_parent;
2318 struct mempolicy *pol;
2319 bool faulted_in_anon_vma = true;
2322 * If anonymous vma has not yet been faulted, update new pgoff
2323 * to match new location, to increase its chance of merging.
2325 if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2326 pgoff = addr >> PAGE_SHIFT;
2327 faulted_in_anon_vma = false;
2330 find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2331 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2332 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2335 * Source vma may have been merged into new_vma
2337 if (unlikely(vma_start >= new_vma->vm_start &&
2338 vma_start < new_vma->vm_end)) {
2340 * The only way we can get a vma_merge with
2341 * self during an mremap is if the vma hasn't
2342 * been faulted in yet and we were allowed to
2343 * reset the dst vma->vm_pgoff to the
2344 * destination address of the mremap to allow
2345 * the merge to happen. mremap must change the
2346 * vm_pgoff linearity between src and dst vmas
2347 * (in turn preventing a vma_merge) to be
2348 * safe. It is only safe to keep the vm_pgoff
2349 * linear if there are no pages mapped yet.
2351 VM_BUG_ON(faulted_in_anon_vma);
2354 anon_vma_moveto_tail(new_vma);
2356 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2359 pol = mpol_dup(vma_policy(vma));
2362 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2363 if (anon_vma_clone(new_vma, vma))
2364 goto out_free_mempol;
2365 vma_set_policy(new_vma, pol);
2366 new_vma->vm_start = addr;
2367 new_vma->vm_end = addr + len;
2368 new_vma->vm_pgoff = pgoff;
2369 if (new_vma->vm_file) {
2370 get_file(new_vma->vm_file);
2371 if (vma->vm_flags & VM_EXECUTABLE)
2372 added_exe_file_vma(mm);
2374 if (new_vma->vm_ops && new_vma->vm_ops->open)
2375 new_vma->vm_ops->open(new_vma);
2376 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2384 kmem_cache_free(vm_area_cachep, new_vma);
2389 * Return true if the calling process may expand its vm space by the passed
2392 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2394 unsigned long cur = mm->total_vm; /* pages */
2397 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2399 if (cur + npages > lim)
2405 static int special_mapping_fault(struct vm_area_struct *vma,
2406 struct vm_fault *vmf)
2409 struct page **pages;
2412 * special mappings have no vm_file, and in that case, the mm
2413 * uses vm_pgoff internally. So we have to subtract it from here.
2414 * We are allowed to do this because we are the mm; do not copy
2415 * this code into drivers!
2417 pgoff = vmf->pgoff - vma->vm_pgoff;
2419 for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2423 struct page *page = *pages;
2429 return VM_FAULT_SIGBUS;
2433 * Having a close hook prevents vma merging regardless of flags.
2435 static void special_mapping_close(struct vm_area_struct *vma)
2439 static const struct vm_operations_struct special_mapping_vmops = {
2440 .close = special_mapping_close,
2441 .fault = special_mapping_fault,
2445 * Called with mm->mmap_sem held for writing.
2446 * Insert a new vma covering the given region, with the given flags.
2447 * Its pages are supplied by the given array of struct page *.
2448 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2449 * The region past the last page supplied will always produce SIGBUS.
2450 * The array pointer and the pages it points to are assumed to stay alive
2451 * for as long as this mapping might exist.
2453 int install_special_mapping(struct mm_struct *mm,
2454 unsigned long addr, unsigned long len,
2455 unsigned long vm_flags, struct page **pages)
2458 struct vm_area_struct *vma;
2460 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2461 if (unlikely(vma == NULL))
2464 INIT_LIST_HEAD(&vma->anon_vma_chain);
2466 vma->vm_start = addr;
2467 vma->vm_end = addr + len;
2469 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2470 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2472 vma->vm_ops = &special_mapping_vmops;
2473 vma->vm_private_data = pages;
2475 ret = security_file_mmap(NULL, 0, 0, 0, vma->vm_start, 1);
2479 ret = insert_vm_struct(mm, vma);
2483 mm->total_vm += len >> PAGE_SHIFT;
2485 perf_event_mmap(vma);
2490 kmem_cache_free(vm_area_cachep, vma);
2494 static DEFINE_MUTEX(mm_all_locks_mutex);
2496 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2498 if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2500 * The LSB of head.next can't change from under us
2501 * because we hold the mm_all_locks_mutex.
2503 mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem);
2505 * We can safely modify head.next after taking the
2506 * anon_vma->root->mutex. If some other vma in this mm shares
2507 * the same anon_vma we won't take it again.
2509 * No need of atomic instructions here, head.next
2510 * can't change from under us thanks to the
2511 * anon_vma->root->mutex.
2513 if (__test_and_set_bit(0, (unsigned long *)
2514 &anon_vma->root->head.next))
2519 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2521 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2523 * AS_MM_ALL_LOCKS can't change from under us because
2524 * we hold the mm_all_locks_mutex.
2526 * Operations on ->flags have to be atomic because
2527 * even if AS_MM_ALL_LOCKS is stable thanks to the
2528 * mm_all_locks_mutex, there may be other cpus
2529 * changing other bitflags in parallel to us.
2531 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2533 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2538 * This operation locks against the VM for all pte/vma/mm related
2539 * operations that could ever happen on a certain mm. This includes
2540 * vmtruncate, try_to_unmap, and all page faults.
2542 * The caller must take the mmap_sem in write mode before calling
2543 * mm_take_all_locks(). The caller isn't allowed to release the
2544 * mmap_sem until mm_drop_all_locks() returns.
2546 * mmap_sem in write mode is required in order to block all operations
2547 * that could modify pagetables and free pages without need of
2548 * altering the vma layout (for example populate_range() with
2549 * nonlinear vmas). It's also needed in write mode to avoid new
2550 * anon_vmas to be associated with existing vmas.
2552 * A single task can't take more than one mm_take_all_locks() in a row
2553 * or it would deadlock.
2555 * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2556 * mapping->flags avoid to take the same lock twice, if more than one
2557 * vma in this mm is backed by the same anon_vma or address_space.
2559 * We can take all the locks in random order because the VM code
2560 * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2561 * takes more than one of them in a row. Secondly we're protected
2562 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2564 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2565 * that may have to take thousand of locks.
2567 * mm_take_all_locks() can fail if it's interrupted by signals.
2569 int mm_take_all_locks(struct mm_struct *mm)
2571 struct vm_area_struct *vma;
2572 struct anon_vma_chain *avc;
2574 BUG_ON(down_read_trylock(&mm->mmap_sem));
2576 mutex_lock(&mm_all_locks_mutex);
2578 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2579 if (signal_pending(current))
2581 if (vma->vm_file && vma->vm_file->f_mapping)
2582 vm_lock_mapping(mm, vma->vm_file->f_mapping);
2585 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2586 if (signal_pending(current))
2589 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2590 vm_lock_anon_vma(mm, avc->anon_vma);
2596 mm_drop_all_locks(mm);
2600 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2602 if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2604 * The LSB of head.next can't change to 0 from under
2605 * us because we hold the mm_all_locks_mutex.
2607 * We must however clear the bitflag before unlocking
2608 * the vma so the users using the anon_vma->head will
2609 * never see our bitflag.
2611 * No need of atomic instructions here, head.next
2612 * can't change from under us until we release the
2613 * anon_vma->root->mutex.
2615 if (!__test_and_clear_bit(0, (unsigned long *)
2616 &anon_vma->root->head.next))
2618 anon_vma_unlock(anon_vma);
2622 static void vm_unlock_mapping(struct address_space *mapping)
2624 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2626 * AS_MM_ALL_LOCKS can't change to 0 from under us
2627 * because we hold the mm_all_locks_mutex.
2629 mutex_unlock(&mapping->i_mmap_mutex);
2630 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2637 * The mmap_sem cannot be released by the caller until
2638 * mm_drop_all_locks() returns.
2640 void mm_drop_all_locks(struct mm_struct *mm)
2642 struct vm_area_struct *vma;
2643 struct anon_vma_chain *avc;
2645 BUG_ON(down_read_trylock(&mm->mmap_sem));
2646 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2648 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2650 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2651 vm_unlock_anon_vma(avc->anon_vma);
2652 if (vma->vm_file && vma->vm_file->f_mapping)
2653 vm_unlock_mapping(vma->vm_file->f_mapping);
2656 mutex_unlock(&mm_all_locks_mutex);
2660 * initialise the VMA slab
2662 void __init mmap_init(void)
2666 ret = percpu_counter_init(&vm_committed_as, 0);