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:
455 * insert vm structure into list and rbtree and anon_vma,
456 * but it has already been inserted into prio_tree earlier.
458 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
460 struct vm_area_struct *__vma, *prev;
461 struct rb_node **rb_link, *rb_parent;
463 __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
464 BUG_ON(__vma && __vma->vm_start < vma->vm_end);
465 __vma_link(mm, vma, prev, rb_link, rb_parent);
470 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
471 struct vm_area_struct *prev)
473 struct vm_area_struct *next = vma->vm_next;
475 prev->vm_next = next;
477 next->vm_prev = prev;
478 rb_erase(&vma->vm_rb, &mm->mm_rb);
479 if (mm->mmap_cache == vma)
480 mm->mmap_cache = prev;
484 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
485 * is already present in an i_mmap tree without adjusting the tree.
486 * The following helper function should be used when such adjustments
487 * are necessary. The "insert" vma (if any) is to be inserted
488 * before we drop the necessary locks.
490 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
491 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
493 struct mm_struct *mm = vma->vm_mm;
494 struct vm_area_struct *next = vma->vm_next;
495 struct vm_area_struct *importer = NULL;
496 struct address_space *mapping = NULL;
497 struct prio_tree_root *root = NULL;
498 struct anon_vma *anon_vma = NULL;
499 struct file *file = vma->vm_file;
500 long adjust_next = 0;
503 if (next && !insert) {
504 struct vm_area_struct *exporter = NULL;
506 if (end >= next->vm_end) {
508 * vma expands, overlapping all the next, and
509 * perhaps the one after too (mprotect case 6).
511 again: remove_next = 1 + (end > next->vm_end);
515 } else if (end > next->vm_start) {
517 * vma expands, overlapping part of the next:
518 * mprotect case 5 shifting the boundary up.
520 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
523 } else if (end < vma->vm_end) {
525 * vma shrinks, and !insert tells it's not
526 * split_vma inserting another: so it must be
527 * mprotect case 4 shifting the boundary down.
529 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
535 * Easily overlooked: when mprotect shifts the boundary,
536 * make sure the expanding vma has anon_vma set if the
537 * shrinking vma had, to cover any anon pages imported.
539 if (exporter && exporter->anon_vma && !importer->anon_vma) {
540 if (anon_vma_clone(importer, exporter))
542 importer->anon_vma = exporter->anon_vma;
547 mapping = file->f_mapping;
548 if (!(vma->vm_flags & VM_NONLINEAR))
549 root = &mapping->i_mmap;
550 mutex_lock(&mapping->i_mmap_mutex);
553 * Put into prio_tree now, so instantiated pages
554 * are visible to arm/parisc __flush_dcache_page
555 * throughout; but we cannot insert into address
556 * space until vma start or end is updated.
558 __vma_link_file(insert);
562 vma_adjust_trans_huge(vma, start, end, adjust_next);
565 * When changing only vma->vm_end, we don't really need anon_vma
566 * lock. This is a fairly rare case by itself, but the anon_vma
567 * lock may be shared between many sibling processes. Skipping
568 * the lock for brk adjustments makes a difference sometimes.
570 if (vma->anon_vma && (importer || start != vma->vm_start)) {
571 anon_vma = vma->anon_vma;
572 anon_vma_lock(anon_vma);
576 flush_dcache_mmap_lock(mapping);
577 vma_prio_tree_remove(vma, root);
579 vma_prio_tree_remove(next, root);
582 vma->vm_start = start;
584 vma->vm_pgoff = pgoff;
586 next->vm_start += adjust_next << PAGE_SHIFT;
587 next->vm_pgoff += adjust_next;
592 vma_prio_tree_insert(next, root);
593 vma_prio_tree_insert(vma, root);
594 flush_dcache_mmap_unlock(mapping);
599 * vma_merge has merged next into vma, and needs
600 * us to remove next before dropping the locks.
602 __vma_unlink(mm, next, vma);
604 __remove_shared_vm_struct(next, file, mapping);
607 * split_vma has split insert from vma, and needs
608 * us to insert it before dropping the locks
609 * (it may either follow vma or precede it).
611 __insert_vm_struct(mm, insert);
615 anon_vma_unlock(anon_vma);
617 mutex_unlock(&mapping->i_mmap_mutex);
622 if (next->vm_flags & VM_EXECUTABLE)
623 removed_exe_file_vma(mm);
626 anon_vma_merge(vma, next);
628 mpol_put(vma_policy(next));
629 kmem_cache_free(vm_area_cachep, next);
631 * In mprotect's case 6 (see comments on vma_merge),
632 * we must remove another next too. It would clutter
633 * up the code too much to do both in one go.
635 if (remove_next == 2) {
647 * If the vma has a ->close operation then the driver probably needs to release
648 * per-vma resources, so we don't attempt to merge those.
650 static inline int is_mergeable_vma(struct vm_area_struct *vma,
651 struct file *file, unsigned long vm_flags)
653 /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
654 if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
656 if (vma->vm_file != file)
658 if (vma->vm_ops && vma->vm_ops->close)
663 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
664 struct anon_vma *anon_vma2,
665 struct vm_area_struct *vma)
668 * The list_is_singular() test is to avoid merging VMA cloned from
669 * parents. This can improve scalability caused by anon_vma lock.
671 if ((!anon_vma1 || !anon_vma2) && (!vma ||
672 list_is_singular(&vma->anon_vma_chain)))
674 return anon_vma1 == anon_vma2;
678 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
679 * in front of (at a lower virtual address and file offset than) the vma.
681 * We cannot merge two vmas if they have differently assigned (non-NULL)
682 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
684 * We don't check here for the merged mmap wrapping around the end of pagecache
685 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
686 * wrap, nor mmaps which cover the final page at index -1UL.
689 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
690 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
692 if (is_mergeable_vma(vma, file, vm_flags) &&
693 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
694 if (vma->vm_pgoff == vm_pgoff)
701 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
702 * beyond (at a higher virtual address and file offset than) the vma.
704 * We cannot merge two vmas if they have differently assigned (non-NULL)
705 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
708 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
709 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
711 if (is_mergeable_vma(vma, file, vm_flags) &&
712 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
714 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
715 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
722 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
723 * whether that can be merged with its predecessor or its successor.
724 * Or both (it neatly fills a hole).
726 * In most cases - when called for mmap, brk or mremap - [addr,end) is
727 * certain not to be mapped by the time vma_merge is called; but when
728 * called for mprotect, it is certain to be already mapped (either at
729 * an offset within prev, or at the start of next), and the flags of
730 * this area are about to be changed to vm_flags - and the no-change
731 * case has already been eliminated.
733 * The following mprotect cases have to be considered, where AAAA is
734 * the area passed down from mprotect_fixup, never extending beyond one
735 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
737 * AAAA AAAA AAAA AAAA
738 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
739 * cannot merge might become might become might become
740 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
741 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
742 * mremap move: PPPPNNNNNNNN 8
744 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
745 * might become case 1 below case 2 below case 3 below
747 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
748 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
750 struct vm_area_struct *vma_merge(struct mm_struct *mm,
751 struct vm_area_struct *prev, unsigned long addr,
752 unsigned long end, unsigned long vm_flags,
753 struct anon_vma *anon_vma, struct file *file,
754 pgoff_t pgoff, struct mempolicy *policy)
756 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
757 struct vm_area_struct *area, *next;
761 * We later require that vma->vm_flags == vm_flags,
762 * so this tests vma->vm_flags & VM_SPECIAL, too.
764 if (vm_flags & VM_SPECIAL)
768 next = prev->vm_next;
772 if (next && next->vm_end == end) /* cases 6, 7, 8 */
773 next = next->vm_next;
776 * Can it merge with the predecessor?
778 if (prev && prev->vm_end == addr &&
779 mpol_equal(vma_policy(prev), policy) &&
780 can_vma_merge_after(prev, vm_flags,
781 anon_vma, file, pgoff)) {
783 * OK, it can. Can we now merge in the successor as well?
785 if (next && end == next->vm_start &&
786 mpol_equal(policy, vma_policy(next)) &&
787 can_vma_merge_before(next, vm_flags,
788 anon_vma, file, pgoff+pglen) &&
789 is_mergeable_anon_vma(prev->anon_vma,
790 next->anon_vma, NULL)) {
792 err = vma_adjust(prev, prev->vm_start,
793 next->vm_end, prev->vm_pgoff, NULL);
794 } else /* cases 2, 5, 7 */
795 err = vma_adjust(prev, prev->vm_start,
796 end, prev->vm_pgoff, NULL);
799 khugepaged_enter_vma_merge(prev);
804 * Can this new request be merged in front of next?
806 if (next && end == next->vm_start &&
807 mpol_equal(policy, vma_policy(next)) &&
808 can_vma_merge_before(next, vm_flags,
809 anon_vma, file, pgoff+pglen)) {
810 if (prev && addr < prev->vm_end) /* case 4 */
811 err = vma_adjust(prev, prev->vm_start,
812 addr, prev->vm_pgoff, NULL);
813 else /* cases 3, 8 */
814 err = vma_adjust(area, addr, next->vm_end,
815 next->vm_pgoff - pglen, NULL);
818 khugepaged_enter_vma_merge(area);
826 * Rough compatbility check to quickly see if it's even worth looking
827 * at sharing an anon_vma.
829 * They need to have the same vm_file, and the flags can only differ
830 * in things that mprotect may change.
832 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
833 * we can merge the two vma's. For example, we refuse to merge a vma if
834 * there is a vm_ops->close() function, because that indicates that the
835 * driver is doing some kind of reference counting. But that doesn't
836 * really matter for the anon_vma sharing case.
838 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
840 return a->vm_end == b->vm_start &&
841 mpol_equal(vma_policy(a), vma_policy(b)) &&
842 a->vm_file == b->vm_file &&
843 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
844 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
848 * Do some basic sanity checking to see if we can re-use the anon_vma
849 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
850 * the same as 'old', the other will be the new one that is trying
851 * to share the anon_vma.
853 * NOTE! This runs with mm_sem held for reading, so it is possible that
854 * the anon_vma of 'old' is concurrently in the process of being set up
855 * by another page fault trying to merge _that_. But that's ok: if it
856 * is being set up, that automatically means that it will be a singleton
857 * acceptable for merging, so we can do all of this optimistically. But
858 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
860 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
861 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
862 * is to return an anon_vma that is "complex" due to having gone through
865 * We also make sure that the two vma's are compatible (adjacent,
866 * and with the same memory policies). That's all stable, even with just
867 * a read lock on the mm_sem.
869 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
871 if (anon_vma_compatible(a, b)) {
872 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
874 if (anon_vma && list_is_singular(&old->anon_vma_chain))
881 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
882 * neighbouring vmas for a suitable anon_vma, before it goes off
883 * to allocate a new anon_vma. It checks because a repetitive
884 * sequence of mprotects and faults may otherwise lead to distinct
885 * anon_vmas being allocated, preventing vma merge in subsequent
888 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
890 struct anon_vma *anon_vma;
891 struct vm_area_struct *near;
897 anon_vma = reusable_anon_vma(near, vma, near);
905 anon_vma = reusable_anon_vma(near, near, vma);
910 * There's no absolute need to look only at touching neighbours:
911 * we could search further afield for "compatible" anon_vmas.
912 * But it would probably just be a waste of time searching,
913 * or lead to too many vmas hanging off the same anon_vma.
914 * We're trying to allow mprotect remerging later on,
915 * not trying to minimize memory used for anon_vmas.
920 #ifdef CONFIG_PROC_FS
921 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
922 struct file *file, long pages)
924 const unsigned long stack_flags
925 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
928 mm->shared_vm += pages;
929 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
930 mm->exec_vm += pages;
931 } else if (flags & stack_flags)
932 mm->stack_vm += pages;
933 if (flags & (VM_RESERVED|VM_IO))
934 mm->reserved_vm += pages;
936 #endif /* CONFIG_PROC_FS */
939 * If a hint addr is less than mmap_min_addr change hint to be as
940 * low as possible but still greater than mmap_min_addr
942 static inline unsigned long round_hint_to_min(unsigned long hint)
945 if (((void *)hint != NULL) &&
946 (hint < mmap_min_addr))
947 return PAGE_ALIGN(mmap_min_addr);
952 * The caller must hold down_write(¤t->mm->mmap_sem).
955 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
956 unsigned long len, unsigned long prot,
957 unsigned long flags, unsigned long pgoff)
959 struct mm_struct * mm = current->mm;
963 unsigned long reqprot = prot;
966 * Does the application expect PROT_READ to imply PROT_EXEC?
968 * (the exception is when the underlying filesystem is noexec
969 * mounted, in which case we dont add PROT_EXEC.)
971 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
972 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
978 if (!(flags & MAP_FIXED))
979 addr = round_hint_to_min(addr);
981 /* Careful about overflows.. */
982 len = PAGE_ALIGN(len);
986 /* offset overflow? */
987 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
990 /* Too many mappings? */
991 if (mm->map_count > sysctl_max_map_count)
994 /* Obtain the address to map to. we verify (or select) it and ensure
995 * that it represents a valid section of the address space.
997 addr = get_unmapped_area(file, addr, len, pgoff, flags);
998 if (addr & ~PAGE_MASK)
1001 /* Do simple checking here so the lower-level routines won't have
1002 * to. we assume access permissions have been handled by the open
1003 * of the memory object, so we don't do any here.
1005 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1006 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1008 if (flags & MAP_LOCKED)
1009 if (!can_do_mlock())
1012 /* mlock MCL_FUTURE? */
1013 if (vm_flags & VM_LOCKED) {
1014 unsigned long locked, lock_limit;
1015 locked = len >> PAGE_SHIFT;
1016 locked += mm->locked_vm;
1017 lock_limit = rlimit(RLIMIT_MEMLOCK);
1018 lock_limit >>= PAGE_SHIFT;
1019 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1023 inode = file ? file->f_path.dentry->d_inode : NULL;
1026 switch (flags & MAP_TYPE) {
1028 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1032 * Make sure we don't allow writing to an append-only
1035 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1039 * Make sure there are no mandatory locks on the file.
1041 if (locks_verify_locked(inode))
1044 vm_flags |= VM_SHARED | VM_MAYSHARE;
1045 if (!(file->f_mode & FMODE_WRITE))
1046 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1050 if (!(file->f_mode & FMODE_READ))
1052 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1053 if (vm_flags & VM_EXEC)
1055 vm_flags &= ~VM_MAYEXEC;
1058 if (!file->f_op || !file->f_op->mmap)
1066 switch (flags & MAP_TYPE) {
1072 vm_flags |= VM_SHARED | VM_MAYSHARE;
1076 * Set pgoff according to addr for anon_vma.
1078 pgoff = addr >> PAGE_SHIFT;
1085 error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1089 return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1091 EXPORT_SYMBOL(do_mmap_pgoff);
1093 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1094 unsigned long, prot, unsigned long, flags,
1095 unsigned long, fd, unsigned long, pgoff)
1097 struct file *file = NULL;
1098 unsigned long retval = -EBADF;
1100 if (!(flags & MAP_ANONYMOUS)) {
1101 audit_mmap_fd(fd, flags);
1102 if (unlikely(flags & MAP_HUGETLB))
1107 } else if (flags & MAP_HUGETLB) {
1108 struct user_struct *user = NULL;
1110 * VM_NORESERVE is used because the reservations will be
1111 * taken when vm_ops->mmap() is called
1112 * A dummy user value is used because we are not locking
1113 * memory so no accounting is necessary
1115 len = ALIGN(len, huge_page_size(&default_hstate));
1116 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
1117 &user, HUGETLB_ANONHUGE_INODE);
1119 return PTR_ERR(file);
1122 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1124 down_write(¤t->mm->mmap_sem);
1125 retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1126 up_write(¤t->mm->mmap_sem);
1134 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1135 struct mmap_arg_struct {
1139 unsigned long flags;
1141 unsigned long offset;
1144 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1146 struct mmap_arg_struct a;
1148 if (copy_from_user(&a, arg, sizeof(a)))
1150 if (a.offset & ~PAGE_MASK)
1153 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1154 a.offset >> PAGE_SHIFT);
1156 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1159 * Some shared mappigns will want the pages marked read-only
1160 * to track write events. If so, we'll downgrade vm_page_prot
1161 * to the private version (using protection_map[] without the
1164 int vma_wants_writenotify(struct vm_area_struct *vma)
1166 vm_flags_t vm_flags = vma->vm_flags;
1168 /* If it was private or non-writable, the write bit is already clear */
1169 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1172 /* The backer wishes to know when pages are first written to? */
1173 if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1176 /* The open routine did something to the protections already? */
1177 if (pgprot_val(vma->vm_page_prot) !=
1178 pgprot_val(vm_get_page_prot(vm_flags)))
1181 /* Specialty mapping? */
1182 if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1185 /* Can the mapping track the dirty pages? */
1186 return vma->vm_file && vma->vm_file->f_mapping &&
1187 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1191 * We account for memory if it's a private writeable mapping,
1192 * not hugepages and VM_NORESERVE wasn't set.
1194 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1197 * hugetlb has its own accounting separate from the core VM
1198 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1200 if (file && is_file_hugepages(file))
1203 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1206 unsigned long mmap_region(struct file *file, unsigned long addr,
1207 unsigned long len, unsigned long flags,
1208 vm_flags_t vm_flags, unsigned long pgoff)
1210 struct mm_struct *mm = current->mm;
1211 struct vm_area_struct *vma, *prev;
1212 int correct_wcount = 0;
1214 struct rb_node **rb_link, *rb_parent;
1215 unsigned long charged = 0;
1216 struct inode *inode = file ? file->f_path.dentry->d_inode : NULL;
1218 /* Clear old maps */
1221 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1222 if (vma && vma->vm_start < addr + len) {
1223 if (do_munmap(mm, addr, len))
1228 /* Check against address space limit. */
1229 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1233 * Set 'VM_NORESERVE' if we should not account for the
1234 * memory use of this mapping.
1236 if ((flags & MAP_NORESERVE)) {
1237 /* We honor MAP_NORESERVE if allowed to overcommit */
1238 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1239 vm_flags |= VM_NORESERVE;
1241 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1242 if (file && is_file_hugepages(file))
1243 vm_flags |= VM_NORESERVE;
1247 * Private writable mapping: check memory availability
1249 if (accountable_mapping(file, vm_flags)) {
1250 charged = len >> PAGE_SHIFT;
1251 if (security_vm_enough_memory_mm(mm, charged))
1253 vm_flags |= VM_ACCOUNT;
1257 * Can we just expand an old mapping?
1259 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1264 * Determine the object being mapped and call the appropriate
1265 * specific mapper. the address has already been validated, but
1266 * not unmapped, but the maps are removed from the list.
1268 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1275 vma->vm_start = addr;
1276 vma->vm_end = addr + len;
1277 vma->vm_flags = vm_flags;
1278 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1279 vma->vm_pgoff = pgoff;
1280 INIT_LIST_HEAD(&vma->anon_vma_chain);
1284 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1286 if (vm_flags & VM_DENYWRITE) {
1287 error = deny_write_access(file);
1292 vma->vm_file = file;
1294 error = file->f_op->mmap(file, vma);
1296 goto unmap_and_free_vma;
1297 if (vm_flags & VM_EXECUTABLE)
1298 added_exe_file_vma(mm);
1300 /* Can addr have changed??
1302 * Answer: Yes, several device drivers can do it in their
1303 * f_op->mmap method. -DaveM
1305 addr = vma->vm_start;
1306 pgoff = vma->vm_pgoff;
1307 vm_flags = vma->vm_flags;
1308 } else if (vm_flags & VM_SHARED) {
1309 error = shmem_zero_setup(vma);
1314 if (vma_wants_writenotify(vma)) {
1315 pgprot_t pprot = vma->vm_page_prot;
1317 /* Can vma->vm_page_prot have changed??
1319 * Answer: Yes, drivers may have changed it in their
1320 * f_op->mmap method.
1322 * Ensures that vmas marked as uncached stay that way.
1324 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1325 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1326 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1329 vma_link(mm, vma, prev, rb_link, rb_parent);
1330 file = vma->vm_file;
1332 /* Once vma denies write, undo our temporary denial count */
1334 atomic_inc(&inode->i_writecount);
1336 perf_event_mmap(vma);
1338 mm->total_vm += len >> PAGE_SHIFT;
1339 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1340 if (vm_flags & VM_LOCKED) {
1341 if (!mlock_vma_pages_range(vma, addr, addr + len))
1342 mm->locked_vm += (len >> PAGE_SHIFT);
1343 } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1344 make_pages_present(addr, addr + len);
1349 atomic_inc(&inode->i_writecount);
1350 vma->vm_file = NULL;
1353 /* Undo any partial mapping done by a device driver. */
1354 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1357 kmem_cache_free(vm_area_cachep, vma);
1360 vm_unacct_memory(charged);
1364 /* Get an address range which is currently unmapped.
1365 * For shmat() with addr=0.
1367 * Ugly calling convention alert:
1368 * Return value with the low bits set means error value,
1370 * if (ret & ~PAGE_MASK)
1373 * This function "knows" that -ENOMEM has the bits set.
1375 #ifndef HAVE_ARCH_UNMAPPED_AREA
1377 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1378 unsigned long len, unsigned long pgoff, unsigned long flags)
1380 struct mm_struct *mm = current->mm;
1381 struct vm_area_struct *vma;
1382 unsigned long start_addr;
1384 if (len > TASK_SIZE)
1387 if (flags & MAP_FIXED)
1391 addr = PAGE_ALIGN(addr);
1392 vma = find_vma(mm, addr);
1393 if (TASK_SIZE - len >= addr &&
1394 (!vma || addr + len <= vma->vm_start))
1397 if (len > mm->cached_hole_size) {
1398 start_addr = addr = mm->free_area_cache;
1400 start_addr = addr = TASK_UNMAPPED_BASE;
1401 mm->cached_hole_size = 0;
1405 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1406 /* At this point: (!vma || addr < vma->vm_end). */
1407 if (TASK_SIZE - len < addr) {
1409 * Start a new search - just in case we missed
1412 if (start_addr != TASK_UNMAPPED_BASE) {
1413 addr = TASK_UNMAPPED_BASE;
1415 mm->cached_hole_size = 0;
1420 if (!vma || addr + len <= vma->vm_start) {
1422 * Remember the place where we stopped the search:
1424 mm->free_area_cache = addr + len;
1427 if (addr + mm->cached_hole_size < vma->vm_start)
1428 mm->cached_hole_size = vma->vm_start - addr;
1434 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1437 * Is this a new hole at the lowest possible address?
1439 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1440 mm->free_area_cache = addr;
1441 mm->cached_hole_size = ~0UL;
1446 * This mmap-allocator allocates new areas top-down from below the
1447 * stack's low limit (the base):
1449 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1451 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1452 const unsigned long len, const unsigned long pgoff,
1453 const unsigned long flags)
1455 struct vm_area_struct *vma;
1456 struct mm_struct *mm = current->mm;
1457 unsigned long addr = addr0;
1459 /* requested length too big for entire address space */
1460 if (len > TASK_SIZE)
1463 if (flags & MAP_FIXED)
1466 /* requesting a specific address */
1468 addr = PAGE_ALIGN(addr);
1469 vma = find_vma(mm, addr);
1470 if (TASK_SIZE - len >= addr &&
1471 (!vma || addr + len <= vma->vm_start))
1475 /* check if free_area_cache is useful for us */
1476 if (len <= mm->cached_hole_size) {
1477 mm->cached_hole_size = 0;
1478 mm->free_area_cache = mm->mmap_base;
1481 /* either no address requested or can't fit in requested address hole */
1482 addr = mm->free_area_cache;
1484 /* make sure it can fit in the remaining address space */
1486 vma = find_vma(mm, addr-len);
1487 if (!vma || addr <= vma->vm_start)
1488 /* remember the address as a hint for next time */
1489 return (mm->free_area_cache = addr-len);
1492 if (mm->mmap_base < len)
1495 addr = mm->mmap_base-len;
1499 * Lookup failure means no vma is above this address,
1500 * else if new region fits below vma->vm_start,
1501 * return with success:
1503 vma = find_vma(mm, addr);
1504 if (!vma || addr+len <= vma->vm_start)
1505 /* remember the address as a hint for next time */
1506 return (mm->free_area_cache = addr);
1508 /* remember the largest hole we saw so far */
1509 if (addr + mm->cached_hole_size < vma->vm_start)
1510 mm->cached_hole_size = vma->vm_start - addr;
1512 /* try just below the current vma->vm_start */
1513 addr = vma->vm_start-len;
1514 } while (len < vma->vm_start);
1518 * A failed mmap() very likely causes application failure,
1519 * so fall back to the bottom-up function here. This scenario
1520 * can happen with large stack limits and large mmap()
1523 mm->cached_hole_size = ~0UL;
1524 mm->free_area_cache = TASK_UNMAPPED_BASE;
1525 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1527 * Restore the topdown base:
1529 mm->free_area_cache = mm->mmap_base;
1530 mm->cached_hole_size = ~0UL;
1536 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1539 * Is this a new hole at the highest possible address?
1541 if (addr > mm->free_area_cache)
1542 mm->free_area_cache = addr;
1544 /* dont allow allocations above current base */
1545 if (mm->free_area_cache > mm->mmap_base)
1546 mm->free_area_cache = mm->mmap_base;
1550 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1551 unsigned long pgoff, unsigned long flags)
1553 unsigned long (*get_area)(struct file *, unsigned long,
1554 unsigned long, unsigned long, unsigned long);
1556 unsigned long error = arch_mmap_check(addr, len, flags);
1560 /* Careful about overflows.. */
1561 if (len > TASK_SIZE)
1564 get_area = current->mm->get_unmapped_area;
1565 if (file && file->f_op && file->f_op->get_unmapped_area)
1566 get_area = file->f_op->get_unmapped_area;
1567 addr = get_area(file, addr, len, pgoff, flags);
1568 if (IS_ERR_VALUE(addr))
1571 if (addr > TASK_SIZE - len)
1573 if (addr & ~PAGE_MASK)
1576 return arch_rebalance_pgtables(addr, len);
1579 EXPORT_SYMBOL(get_unmapped_area);
1581 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1582 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1584 struct vm_area_struct *vma = NULL;
1587 /* Check the cache first. */
1588 /* (Cache hit rate is typically around 35%.) */
1589 vma = mm->mmap_cache;
1590 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1591 struct rb_node * rb_node;
1593 rb_node = mm->mm_rb.rb_node;
1597 struct vm_area_struct * vma_tmp;
1599 vma_tmp = rb_entry(rb_node,
1600 struct vm_area_struct, vm_rb);
1602 if (vma_tmp->vm_end > addr) {
1604 if (vma_tmp->vm_start <= addr)
1606 rb_node = rb_node->rb_left;
1608 rb_node = rb_node->rb_right;
1611 mm->mmap_cache = vma;
1617 EXPORT_SYMBOL(find_vma);
1620 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1621 * Note: pprev is set to NULL when return value is NULL.
1623 struct vm_area_struct *
1624 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1625 struct vm_area_struct **pprev)
1627 struct vm_area_struct *vma;
1629 vma = find_vma(mm, addr);
1630 *pprev = vma ? vma->vm_prev : NULL;
1635 * Verify that the stack growth is acceptable and
1636 * update accounting. This is shared with both the
1637 * grow-up and grow-down cases.
1639 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1641 struct mm_struct *mm = vma->vm_mm;
1642 struct rlimit *rlim = current->signal->rlim;
1643 unsigned long new_start;
1645 /* address space limit tests */
1646 if (!may_expand_vm(mm, grow))
1649 /* Stack limit test */
1650 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1653 /* mlock limit tests */
1654 if (vma->vm_flags & VM_LOCKED) {
1655 unsigned long locked;
1656 unsigned long limit;
1657 locked = mm->locked_vm + grow;
1658 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1659 limit >>= PAGE_SHIFT;
1660 if (locked > limit && !capable(CAP_IPC_LOCK))
1664 /* Check to ensure the stack will not grow into a hugetlb-only region */
1665 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1667 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1671 * Overcommit.. This must be the final test, as it will
1672 * update security statistics.
1674 if (security_vm_enough_memory_mm(mm, grow))
1677 /* Ok, everything looks good - let it rip */
1678 mm->total_vm += grow;
1679 if (vma->vm_flags & VM_LOCKED)
1680 mm->locked_vm += grow;
1681 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1685 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1687 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1688 * vma is the last one with address > vma->vm_end. Have to extend vma.
1690 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1694 if (!(vma->vm_flags & VM_GROWSUP))
1698 * We must make sure the anon_vma is allocated
1699 * so that the anon_vma locking is not a noop.
1701 if (unlikely(anon_vma_prepare(vma)))
1703 vma_lock_anon_vma(vma);
1706 * vma->vm_start/vm_end cannot change under us because the caller
1707 * is required to hold the mmap_sem in read mode. We need the
1708 * anon_vma lock to serialize against concurrent expand_stacks.
1709 * Also guard against wrapping around to address 0.
1711 if (address < PAGE_ALIGN(address+4))
1712 address = PAGE_ALIGN(address+4);
1714 vma_unlock_anon_vma(vma);
1719 /* Somebody else might have raced and expanded it already */
1720 if (address > vma->vm_end) {
1721 unsigned long size, grow;
1723 size = address - vma->vm_start;
1724 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1727 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1728 error = acct_stack_growth(vma, size, grow);
1730 vma->vm_end = address;
1731 perf_event_mmap(vma);
1735 vma_unlock_anon_vma(vma);
1736 khugepaged_enter_vma_merge(vma);
1739 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1742 * vma is the first one with address < vma->vm_start. Have to extend vma.
1744 int expand_downwards(struct vm_area_struct *vma,
1745 unsigned long address)
1750 * We must make sure the anon_vma is allocated
1751 * so that the anon_vma locking is not a noop.
1753 if (unlikely(anon_vma_prepare(vma)))
1756 address &= PAGE_MASK;
1757 error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1761 vma_lock_anon_vma(vma);
1764 * vma->vm_start/vm_end cannot change under us because the caller
1765 * is required to hold the mmap_sem in read mode. We need the
1766 * anon_vma lock to serialize against concurrent expand_stacks.
1769 /* Somebody else might have raced and expanded it already */
1770 if (address < vma->vm_start) {
1771 unsigned long size, grow;
1773 size = vma->vm_end - address;
1774 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1777 if (grow <= vma->vm_pgoff) {
1778 error = acct_stack_growth(vma, size, grow);
1780 vma->vm_start = address;
1781 vma->vm_pgoff -= grow;
1782 perf_event_mmap(vma);
1786 vma_unlock_anon_vma(vma);
1787 khugepaged_enter_vma_merge(vma);
1791 #ifdef CONFIG_STACK_GROWSUP
1792 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1794 return expand_upwards(vma, address);
1797 struct vm_area_struct *
1798 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1800 struct vm_area_struct *vma, *prev;
1803 vma = find_vma_prev(mm, addr, &prev);
1804 if (vma && (vma->vm_start <= addr))
1806 if (!prev || expand_stack(prev, addr))
1808 if (prev->vm_flags & VM_LOCKED) {
1809 mlock_vma_pages_range(prev, addr, prev->vm_end);
1814 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1816 return expand_downwards(vma, address);
1819 struct vm_area_struct *
1820 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1822 struct vm_area_struct * vma;
1823 unsigned long start;
1826 vma = find_vma(mm,addr);
1829 if (vma->vm_start <= addr)
1831 if (!(vma->vm_flags & VM_GROWSDOWN))
1833 start = vma->vm_start;
1834 if (expand_stack(vma, addr))
1836 if (vma->vm_flags & VM_LOCKED) {
1837 mlock_vma_pages_range(vma, addr, start);
1844 * Ok - we have the memory areas we should free on the vma list,
1845 * so release them, and do the vma updates.
1847 * Called with the mm semaphore held.
1849 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1851 /* Update high watermark before we lower total_vm */
1852 update_hiwater_vm(mm);
1854 long nrpages = vma_pages(vma);
1856 mm->total_vm -= nrpages;
1857 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1858 vma = remove_vma(vma);
1864 * Get rid of page table information in the indicated region.
1866 * Called with the mm semaphore held.
1868 static void unmap_region(struct mm_struct *mm,
1869 struct vm_area_struct *vma, struct vm_area_struct *prev,
1870 unsigned long start, unsigned long end)
1872 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1873 struct mmu_gather tlb;
1874 unsigned long nr_accounted = 0;
1877 tlb_gather_mmu(&tlb, mm, 0);
1878 update_hiwater_rss(mm);
1879 unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1880 vm_unacct_memory(nr_accounted);
1881 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
1882 next ? next->vm_start : 0);
1883 tlb_finish_mmu(&tlb, start, end);
1887 * Create a list of vma's touched by the unmap, removing them from the mm's
1888 * vma list as we go..
1891 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1892 struct vm_area_struct *prev, unsigned long end)
1894 struct vm_area_struct **insertion_point;
1895 struct vm_area_struct *tail_vma = NULL;
1898 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1899 vma->vm_prev = NULL;
1901 rb_erase(&vma->vm_rb, &mm->mm_rb);
1905 } while (vma && vma->vm_start < end);
1906 *insertion_point = vma;
1908 vma->vm_prev = prev;
1909 tail_vma->vm_next = NULL;
1910 if (mm->unmap_area == arch_unmap_area)
1911 addr = prev ? prev->vm_end : mm->mmap_base;
1913 addr = vma ? vma->vm_start : mm->mmap_base;
1914 mm->unmap_area(mm, addr);
1915 mm->mmap_cache = NULL; /* Kill the cache. */
1919 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
1920 * munmap path where it doesn't make sense to fail.
1922 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1923 unsigned long addr, int new_below)
1925 struct mempolicy *pol;
1926 struct vm_area_struct *new;
1929 if (is_vm_hugetlb_page(vma) && (addr &
1930 ~(huge_page_mask(hstate_vma(vma)))))
1933 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1937 /* most fields are the same, copy all, and then fixup */
1940 INIT_LIST_HEAD(&new->anon_vma_chain);
1945 new->vm_start = addr;
1946 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1949 pol = mpol_dup(vma_policy(vma));
1954 vma_set_policy(new, pol);
1956 if (anon_vma_clone(new, vma))
1960 get_file(new->vm_file);
1961 if (vma->vm_flags & VM_EXECUTABLE)
1962 added_exe_file_vma(mm);
1965 if (new->vm_ops && new->vm_ops->open)
1966 new->vm_ops->open(new);
1969 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1970 ((addr - new->vm_start) >> PAGE_SHIFT), new);
1972 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1978 /* Clean everything up if vma_adjust failed. */
1979 if (new->vm_ops && new->vm_ops->close)
1980 new->vm_ops->close(new);
1982 if (vma->vm_flags & VM_EXECUTABLE)
1983 removed_exe_file_vma(mm);
1986 unlink_anon_vmas(new);
1990 kmem_cache_free(vm_area_cachep, new);
1996 * Split a vma into two pieces at address 'addr', a new vma is allocated
1997 * either for the first part or the tail.
1999 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2000 unsigned long addr, int new_below)
2002 if (mm->map_count >= sysctl_max_map_count)
2005 return __split_vma(mm, vma, addr, new_below);
2008 /* Munmap is split into 2 main parts -- this part which finds
2009 * what needs doing, and the areas themselves, which do the
2010 * work. This now handles partial unmappings.
2011 * Jeremy Fitzhardinge <jeremy@goop.org>
2013 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2016 struct vm_area_struct *vma, *prev, *last;
2018 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2021 if ((len = PAGE_ALIGN(len)) == 0)
2024 /* Find the first overlapping VMA */
2025 vma = find_vma(mm, start);
2028 prev = vma->vm_prev;
2029 /* we have start < vma->vm_end */
2031 /* if it doesn't overlap, we have nothing.. */
2033 if (vma->vm_start >= end)
2037 * If we need to split any vma, do it now to save pain later.
2039 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2040 * unmapped vm_area_struct will remain in use: so lower split_vma
2041 * places tmp vma above, and higher split_vma places tmp vma below.
2043 if (start > vma->vm_start) {
2047 * Make sure that map_count on return from munmap() will
2048 * not exceed its limit; but let map_count go just above
2049 * its limit temporarily, to help free resources as expected.
2051 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2054 error = __split_vma(mm, vma, start, 0);
2060 /* Does it split the last one? */
2061 last = find_vma(mm, end);
2062 if (last && end > last->vm_start) {
2063 int error = __split_vma(mm, last, end, 1);
2067 vma = prev? prev->vm_next: mm->mmap;
2070 * unlock any mlock()ed ranges before detaching vmas
2072 if (mm->locked_vm) {
2073 struct vm_area_struct *tmp = vma;
2074 while (tmp && tmp->vm_start < end) {
2075 if (tmp->vm_flags & VM_LOCKED) {
2076 mm->locked_vm -= vma_pages(tmp);
2077 munlock_vma_pages_all(tmp);
2084 * Remove the vma's, and unmap the actual pages
2086 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2087 unmap_region(mm, vma, prev, start, end);
2089 /* Fix up all other VM information */
2090 remove_vma_list(mm, vma);
2095 EXPORT_SYMBOL(do_munmap);
2097 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2100 struct mm_struct *mm = current->mm;
2102 profile_munmap(addr);
2104 down_write(&mm->mmap_sem);
2105 ret = do_munmap(mm, addr, len);
2106 up_write(&mm->mmap_sem);
2110 static inline void verify_mm_writelocked(struct mm_struct *mm)
2112 #ifdef CONFIG_DEBUG_VM
2113 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2115 up_read(&mm->mmap_sem);
2121 * this is really a simplified "do_mmap". it only handles
2122 * anonymous maps. eventually we may be able to do some
2123 * brk-specific accounting here.
2125 unsigned long do_brk(unsigned long addr, unsigned long len)
2127 struct mm_struct * mm = current->mm;
2128 struct vm_area_struct * vma, * prev;
2129 unsigned long flags;
2130 struct rb_node ** rb_link, * rb_parent;
2131 pgoff_t pgoff = addr >> PAGE_SHIFT;
2134 len = PAGE_ALIGN(len);
2138 error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2142 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2144 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2145 if (error & ~PAGE_MASK)
2151 if (mm->def_flags & VM_LOCKED) {
2152 unsigned long locked, lock_limit;
2153 locked = len >> PAGE_SHIFT;
2154 locked += mm->locked_vm;
2155 lock_limit = rlimit(RLIMIT_MEMLOCK);
2156 lock_limit >>= PAGE_SHIFT;
2157 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2162 * mm->mmap_sem is required to protect against another thread
2163 * changing the mappings in case we sleep.
2165 verify_mm_writelocked(mm);
2168 * Clear old maps. this also does some error checking for us
2171 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2172 if (vma && vma->vm_start < addr + len) {
2173 if (do_munmap(mm, addr, len))
2178 /* Check against address space limits *after* clearing old maps... */
2179 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2182 if (mm->map_count > sysctl_max_map_count)
2185 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2188 /* Can we just expand an old private anonymous mapping? */
2189 vma = vma_merge(mm, prev, addr, addr + len, flags,
2190 NULL, NULL, pgoff, NULL);
2195 * create a vma struct for an anonymous mapping
2197 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2199 vm_unacct_memory(len >> PAGE_SHIFT);
2203 INIT_LIST_HEAD(&vma->anon_vma_chain);
2205 vma->vm_start = addr;
2206 vma->vm_end = addr + len;
2207 vma->vm_pgoff = pgoff;
2208 vma->vm_flags = flags;
2209 vma->vm_page_prot = vm_get_page_prot(flags);
2210 vma_link(mm, vma, prev, rb_link, rb_parent);
2212 perf_event_mmap(vma);
2213 mm->total_vm += len >> PAGE_SHIFT;
2214 if (flags & VM_LOCKED) {
2215 if (!mlock_vma_pages_range(vma, addr, addr + len))
2216 mm->locked_vm += (len >> PAGE_SHIFT);
2221 EXPORT_SYMBOL(do_brk);
2223 /* Release all mmaps. */
2224 void exit_mmap(struct mm_struct *mm)
2226 struct mmu_gather tlb;
2227 struct vm_area_struct *vma;
2228 unsigned long nr_accounted = 0;
2231 /* mm's last user has gone, and its about to be pulled down */
2232 mmu_notifier_release(mm);
2234 if (mm->locked_vm) {
2237 if (vma->vm_flags & VM_LOCKED)
2238 munlock_vma_pages_all(vma);
2246 if (!vma) /* Can happen if dup_mmap() received an OOM */
2251 tlb_gather_mmu(&tlb, mm, 1);
2252 /* update_hiwater_rss(mm) here? but nobody should be looking */
2253 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2254 end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2255 vm_unacct_memory(nr_accounted);
2257 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2258 tlb_finish_mmu(&tlb, 0, end);
2261 * Walk the list again, actually closing and freeing it,
2262 * with preemption enabled, without holding any MM locks.
2265 vma = remove_vma(vma);
2267 BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2270 /* Insert vm structure into process list sorted by address
2271 * and into the inode's i_mmap tree. If vm_file is non-NULL
2272 * then i_mmap_mutex is taken here.
2274 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2276 struct vm_area_struct * __vma, * prev;
2277 struct rb_node ** rb_link, * rb_parent;
2280 * The vm_pgoff of a purely anonymous vma should be irrelevant
2281 * until its first write fault, when page's anon_vma and index
2282 * are set. But now set the vm_pgoff it will almost certainly
2283 * end up with (unless mremap moves it elsewhere before that
2284 * first wfault), so /proc/pid/maps tells a consistent story.
2286 * By setting it to reflect the virtual start address of the
2287 * vma, merges and splits can happen in a seamless way, just
2288 * using the existing file pgoff checks and manipulations.
2289 * Similarly in do_mmap_pgoff and in do_brk.
2291 if (!vma->vm_file) {
2292 BUG_ON(vma->anon_vma);
2293 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2295 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2296 if (__vma && __vma->vm_start < vma->vm_end)
2298 if ((vma->vm_flags & VM_ACCOUNT) &&
2299 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2301 vma_link(mm, vma, prev, rb_link, rb_parent);
2306 * Copy the vma structure to a new location in the same mm,
2307 * prior to moving page table entries, to effect an mremap move.
2309 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2310 unsigned long addr, unsigned long len, pgoff_t pgoff)
2312 struct vm_area_struct *vma = *vmap;
2313 unsigned long vma_start = vma->vm_start;
2314 struct mm_struct *mm = vma->vm_mm;
2315 struct vm_area_struct *new_vma, *prev;
2316 struct rb_node **rb_link, *rb_parent;
2317 struct mempolicy *pol;
2318 bool faulted_in_anon_vma = true;
2321 * If anonymous vma has not yet been faulted, update new pgoff
2322 * to match new location, to increase its chance of merging.
2324 if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2325 pgoff = addr >> PAGE_SHIFT;
2326 faulted_in_anon_vma = false;
2329 find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2330 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2331 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2334 * Source vma may have been merged into new_vma
2336 if (unlikely(vma_start >= new_vma->vm_start &&
2337 vma_start < new_vma->vm_end)) {
2339 * The only way we can get a vma_merge with
2340 * self during an mremap is if the vma hasn't
2341 * been faulted in yet and we were allowed to
2342 * reset the dst vma->vm_pgoff to the
2343 * destination address of the mremap to allow
2344 * the merge to happen. mremap must change the
2345 * vm_pgoff linearity between src and dst vmas
2346 * (in turn preventing a vma_merge) to be
2347 * safe. It is only safe to keep the vm_pgoff
2348 * linear if there are no pages mapped yet.
2350 VM_BUG_ON(faulted_in_anon_vma);
2353 anon_vma_moveto_tail(new_vma);
2355 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2358 pol = mpol_dup(vma_policy(vma));
2361 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2362 if (anon_vma_clone(new_vma, vma))
2363 goto out_free_mempol;
2364 vma_set_policy(new_vma, pol);
2365 new_vma->vm_start = addr;
2366 new_vma->vm_end = addr + len;
2367 new_vma->vm_pgoff = pgoff;
2368 if (new_vma->vm_file) {
2369 get_file(new_vma->vm_file);
2370 if (vma->vm_flags & VM_EXECUTABLE)
2371 added_exe_file_vma(mm);
2373 if (new_vma->vm_ops && new_vma->vm_ops->open)
2374 new_vma->vm_ops->open(new_vma);
2375 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2383 kmem_cache_free(vm_area_cachep, new_vma);
2388 * Return true if the calling process may expand its vm space by the passed
2391 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2393 unsigned long cur = mm->total_vm; /* pages */
2396 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2398 if (cur + npages > lim)
2404 static int special_mapping_fault(struct vm_area_struct *vma,
2405 struct vm_fault *vmf)
2408 struct page **pages;
2411 * special mappings have no vm_file, and in that case, the mm
2412 * uses vm_pgoff internally. So we have to subtract it from here.
2413 * We are allowed to do this because we are the mm; do not copy
2414 * this code into drivers!
2416 pgoff = vmf->pgoff - vma->vm_pgoff;
2418 for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2422 struct page *page = *pages;
2428 return VM_FAULT_SIGBUS;
2432 * Having a close hook prevents vma merging regardless of flags.
2434 static void special_mapping_close(struct vm_area_struct *vma)
2438 static const struct vm_operations_struct special_mapping_vmops = {
2439 .close = special_mapping_close,
2440 .fault = special_mapping_fault,
2444 * Called with mm->mmap_sem held for writing.
2445 * Insert a new vma covering the given region, with the given flags.
2446 * Its pages are supplied by the given array of struct page *.
2447 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2448 * The region past the last page supplied will always produce SIGBUS.
2449 * The array pointer and the pages it points to are assumed to stay alive
2450 * for as long as this mapping might exist.
2452 int install_special_mapping(struct mm_struct *mm,
2453 unsigned long addr, unsigned long len,
2454 unsigned long vm_flags, struct page **pages)
2457 struct vm_area_struct *vma;
2459 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2460 if (unlikely(vma == NULL))
2463 INIT_LIST_HEAD(&vma->anon_vma_chain);
2465 vma->vm_start = addr;
2466 vma->vm_end = addr + len;
2468 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2469 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2471 vma->vm_ops = &special_mapping_vmops;
2472 vma->vm_private_data = pages;
2474 ret = security_file_mmap(NULL, 0, 0, 0, vma->vm_start, 1);
2478 ret = insert_vm_struct(mm, vma);
2482 mm->total_vm += len >> PAGE_SHIFT;
2484 perf_event_mmap(vma);
2489 kmem_cache_free(vm_area_cachep, vma);
2493 static DEFINE_MUTEX(mm_all_locks_mutex);
2495 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2497 if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2499 * The LSB of head.next can't change from under us
2500 * because we hold the mm_all_locks_mutex.
2502 mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem);
2504 * We can safely modify head.next after taking the
2505 * anon_vma->root->mutex. If some other vma in this mm shares
2506 * the same anon_vma we won't take it again.
2508 * No need of atomic instructions here, head.next
2509 * can't change from under us thanks to the
2510 * anon_vma->root->mutex.
2512 if (__test_and_set_bit(0, (unsigned long *)
2513 &anon_vma->root->head.next))
2518 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2520 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2522 * AS_MM_ALL_LOCKS can't change from under us because
2523 * we hold the mm_all_locks_mutex.
2525 * Operations on ->flags have to be atomic because
2526 * even if AS_MM_ALL_LOCKS is stable thanks to the
2527 * mm_all_locks_mutex, there may be other cpus
2528 * changing other bitflags in parallel to us.
2530 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2532 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2537 * This operation locks against the VM for all pte/vma/mm related
2538 * operations that could ever happen on a certain mm. This includes
2539 * vmtruncate, try_to_unmap, and all page faults.
2541 * The caller must take the mmap_sem in write mode before calling
2542 * mm_take_all_locks(). The caller isn't allowed to release the
2543 * mmap_sem until mm_drop_all_locks() returns.
2545 * mmap_sem in write mode is required in order to block all operations
2546 * that could modify pagetables and free pages without need of
2547 * altering the vma layout (for example populate_range() with
2548 * nonlinear vmas). It's also needed in write mode to avoid new
2549 * anon_vmas to be associated with existing vmas.
2551 * A single task can't take more than one mm_take_all_locks() in a row
2552 * or it would deadlock.
2554 * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2555 * mapping->flags avoid to take the same lock twice, if more than one
2556 * vma in this mm is backed by the same anon_vma or address_space.
2558 * We can take all the locks in random order because the VM code
2559 * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2560 * takes more than one of them in a row. Secondly we're protected
2561 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2563 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2564 * that may have to take thousand of locks.
2566 * mm_take_all_locks() can fail if it's interrupted by signals.
2568 int mm_take_all_locks(struct mm_struct *mm)
2570 struct vm_area_struct *vma;
2571 struct anon_vma_chain *avc;
2573 BUG_ON(down_read_trylock(&mm->mmap_sem));
2575 mutex_lock(&mm_all_locks_mutex);
2577 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2578 if (signal_pending(current))
2580 if (vma->vm_file && vma->vm_file->f_mapping)
2581 vm_lock_mapping(mm, vma->vm_file->f_mapping);
2584 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2585 if (signal_pending(current))
2588 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2589 vm_lock_anon_vma(mm, avc->anon_vma);
2595 mm_drop_all_locks(mm);
2599 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2601 if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2603 * The LSB of head.next can't change to 0 from under
2604 * us because we hold the mm_all_locks_mutex.
2606 * We must however clear the bitflag before unlocking
2607 * the vma so the users using the anon_vma->head will
2608 * never see our bitflag.
2610 * No need of atomic instructions here, head.next
2611 * can't change from under us until we release the
2612 * anon_vma->root->mutex.
2614 if (!__test_and_clear_bit(0, (unsigned long *)
2615 &anon_vma->root->head.next))
2617 anon_vma_unlock(anon_vma);
2621 static void vm_unlock_mapping(struct address_space *mapping)
2623 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2625 * AS_MM_ALL_LOCKS can't change to 0 from under us
2626 * because we hold the mm_all_locks_mutex.
2628 mutex_unlock(&mapping->i_mmap_mutex);
2629 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2636 * The mmap_sem cannot be released by the caller until
2637 * mm_drop_all_locks() returns.
2639 void mm_drop_all_locks(struct mm_struct *mm)
2641 struct vm_area_struct *vma;
2642 struct anon_vma_chain *avc;
2644 BUG_ON(down_read_trylock(&mm->mmap_sem));
2645 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2647 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2649 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2650 vm_unlock_anon_vma(avc->anon_vma);
2651 if (vma->vm_file && vma->vm_file->f_mapping)
2652 vm_unlock_mapping(vma->vm_file->f_mapping);
2655 mutex_unlock(&mm_all_locks_mutex);
2659 * initialise the VMA slab
2661 void __init mmap_init(void)
2665 ret = percpu_counter_init(&vm_committed_as, 0);