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)
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) {
542 importer->anon_vma = exporter->anon_vma;
543 error = anon_vma_clone(importer, exporter);
550 mapping = file->f_mapping;
551 if (!(vma->vm_flags & VM_NONLINEAR))
552 root = &mapping->i_mmap;
553 mutex_lock(&mapping->i_mmap_mutex);
556 * Put into prio_tree now, so instantiated pages
557 * are visible to arm/parisc __flush_dcache_page
558 * throughout; but we cannot insert into address
559 * space until vma start or end is updated.
561 __vma_link_file(insert);
565 vma_adjust_trans_huge(vma, start, end, adjust_next);
568 * When changing only vma->vm_end, we don't really need anon_vma
569 * lock. This is a fairly rare case by itself, but the anon_vma
570 * lock may be shared between many sibling processes. Skipping
571 * the lock for brk adjustments makes a difference sometimes.
573 if (vma->anon_vma && (importer || start != vma->vm_start)) {
574 anon_vma = vma->anon_vma;
575 anon_vma_lock(anon_vma);
579 flush_dcache_mmap_lock(mapping);
580 vma_prio_tree_remove(vma, root);
582 vma_prio_tree_remove(next, root);
585 vma->vm_start = start;
587 vma->vm_pgoff = pgoff;
589 next->vm_start += adjust_next << PAGE_SHIFT;
590 next->vm_pgoff += adjust_next;
595 vma_prio_tree_insert(next, root);
596 vma_prio_tree_insert(vma, root);
597 flush_dcache_mmap_unlock(mapping);
602 * vma_merge has merged next into vma, and needs
603 * us to remove next before dropping the locks.
605 __vma_unlink(mm, next, vma);
607 __remove_shared_vm_struct(next, file, mapping);
610 * split_vma has split insert from vma, and needs
611 * us to insert it before dropping the locks
612 * (it may either follow vma or precede it).
614 __insert_vm_struct(mm, insert);
618 anon_vma_unlock(anon_vma);
620 mutex_unlock(&mapping->i_mmap_mutex);
625 if (next->vm_flags & VM_EXECUTABLE)
626 removed_exe_file_vma(mm);
629 anon_vma_merge(vma, next);
631 mpol_put(vma_policy(next));
632 kmem_cache_free(vm_area_cachep, next);
634 * In mprotect's case 6 (see comments on vma_merge),
635 * we must remove another next too. It would clutter
636 * up the code too much to do both in one go.
638 if (remove_next == 2) {
650 * If the vma has a ->close operation then the driver probably needs to release
651 * per-vma resources, so we don't attempt to merge those.
653 static inline int is_mergeable_vma(struct vm_area_struct *vma,
654 struct file *file, unsigned long vm_flags)
656 /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
657 if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
659 if (vma->vm_file != file)
661 if (vma->vm_ops && vma->vm_ops->close)
666 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
667 struct anon_vma *anon_vma2,
668 struct vm_area_struct *vma)
671 * The list_is_singular() test is to avoid merging VMA cloned from
672 * parents. This can improve scalability caused by anon_vma lock.
674 if ((!anon_vma1 || !anon_vma2) && (!vma ||
675 list_is_singular(&vma->anon_vma_chain)))
677 return anon_vma1 == anon_vma2;
681 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
682 * in front of (at a lower virtual address and file offset than) the vma.
684 * We cannot merge two vmas if they have differently assigned (non-NULL)
685 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
687 * We don't check here for the merged mmap wrapping around the end of pagecache
688 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
689 * wrap, nor mmaps which cover the final page at index -1UL.
692 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
693 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
695 if (is_mergeable_vma(vma, file, vm_flags) &&
696 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
697 if (vma->vm_pgoff == vm_pgoff)
704 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
705 * beyond (at a higher virtual address and file offset than) the vma.
707 * We cannot merge two vmas if they have differently assigned (non-NULL)
708 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
711 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
712 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
714 if (is_mergeable_vma(vma, file, vm_flags) &&
715 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
717 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
718 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
725 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
726 * whether that can be merged with its predecessor or its successor.
727 * Or both (it neatly fills a hole).
729 * In most cases - when called for mmap, brk or mremap - [addr,end) is
730 * certain not to be mapped by the time vma_merge is called; but when
731 * called for mprotect, it is certain to be already mapped (either at
732 * an offset within prev, or at the start of next), and the flags of
733 * this area are about to be changed to vm_flags - and the no-change
734 * case has already been eliminated.
736 * The following mprotect cases have to be considered, where AAAA is
737 * the area passed down from mprotect_fixup, never extending beyond one
738 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
740 * AAAA AAAA AAAA AAAA
741 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
742 * cannot merge might become might become might become
743 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
744 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
745 * mremap move: PPPPNNNNNNNN 8
747 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
748 * might become case 1 below case 2 below case 3 below
750 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
751 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
753 struct vm_area_struct *vma_merge(struct mm_struct *mm,
754 struct vm_area_struct *prev, unsigned long addr,
755 unsigned long end, unsigned long vm_flags,
756 struct anon_vma *anon_vma, struct file *file,
757 pgoff_t pgoff, struct mempolicy *policy)
759 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
760 struct vm_area_struct *area, *next;
764 * We later require that vma->vm_flags == vm_flags,
765 * so this tests vma->vm_flags & VM_SPECIAL, too.
767 if (vm_flags & VM_SPECIAL)
771 next = prev->vm_next;
775 if (next && next->vm_end == end) /* cases 6, 7, 8 */
776 next = next->vm_next;
779 * Can it merge with the predecessor?
781 if (prev && prev->vm_end == addr &&
782 mpol_equal(vma_policy(prev), policy) &&
783 can_vma_merge_after(prev, vm_flags,
784 anon_vma, file, pgoff)) {
786 * OK, it can. Can we now merge in the successor as well?
788 if (next && end == next->vm_start &&
789 mpol_equal(policy, vma_policy(next)) &&
790 can_vma_merge_before(next, vm_flags,
791 anon_vma, file, pgoff+pglen) &&
792 is_mergeable_anon_vma(prev->anon_vma,
793 next->anon_vma, NULL)) {
795 err = vma_adjust(prev, prev->vm_start,
796 next->vm_end, prev->vm_pgoff, NULL);
797 } else /* cases 2, 5, 7 */
798 err = vma_adjust(prev, prev->vm_start,
799 end, prev->vm_pgoff, NULL);
802 khugepaged_enter_vma_merge(prev, vm_flags);
807 * Can this new request be merged in front of next?
809 if (next && end == next->vm_start &&
810 mpol_equal(policy, vma_policy(next)) &&
811 can_vma_merge_before(next, vm_flags,
812 anon_vma, file, pgoff+pglen)) {
813 if (prev && addr < prev->vm_end) /* case 4 */
814 err = vma_adjust(prev, prev->vm_start,
815 addr, prev->vm_pgoff, NULL);
816 else /* cases 3, 8 */
817 err = vma_adjust(area, addr, next->vm_end,
818 next->vm_pgoff - pglen, NULL);
821 khugepaged_enter_vma_merge(area, vm_flags);
829 * Rough compatbility check to quickly see if it's even worth looking
830 * at sharing an anon_vma.
832 * They need to have the same vm_file, and the flags can only differ
833 * in things that mprotect may change.
835 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
836 * we can merge the two vma's. For example, we refuse to merge a vma if
837 * there is a vm_ops->close() function, because that indicates that the
838 * driver is doing some kind of reference counting. But that doesn't
839 * really matter for the anon_vma sharing case.
841 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
843 return a->vm_end == b->vm_start &&
844 mpol_equal(vma_policy(a), vma_policy(b)) &&
845 a->vm_file == b->vm_file &&
846 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
847 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
851 * Do some basic sanity checking to see if we can re-use the anon_vma
852 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
853 * the same as 'old', the other will be the new one that is trying
854 * to share the anon_vma.
856 * NOTE! This runs with mm_sem held for reading, so it is possible that
857 * the anon_vma of 'old' is concurrently in the process of being set up
858 * by another page fault trying to merge _that_. But that's ok: if it
859 * is being set up, that automatically means that it will be a singleton
860 * acceptable for merging, so we can do all of this optimistically. But
861 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
863 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
864 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
865 * is to return an anon_vma that is "complex" due to having gone through
868 * We also make sure that the two vma's are compatible (adjacent,
869 * and with the same memory policies). That's all stable, even with just
870 * a read lock on the mm_sem.
872 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
874 if (anon_vma_compatible(a, b)) {
875 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
877 if (anon_vma && list_is_singular(&old->anon_vma_chain))
884 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
885 * neighbouring vmas for a suitable anon_vma, before it goes off
886 * to allocate a new anon_vma. It checks because a repetitive
887 * sequence of mprotects and faults may otherwise lead to distinct
888 * anon_vmas being allocated, preventing vma merge in subsequent
891 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
893 struct anon_vma *anon_vma;
894 struct vm_area_struct *near;
900 anon_vma = reusable_anon_vma(near, vma, near);
908 anon_vma = reusable_anon_vma(near, near, vma);
913 * There's no absolute need to look only at touching neighbours:
914 * we could search further afield for "compatible" anon_vmas.
915 * But it would probably just be a waste of time searching,
916 * or lead to too many vmas hanging off the same anon_vma.
917 * We're trying to allow mprotect remerging later on,
918 * not trying to minimize memory used for anon_vmas.
923 #ifdef CONFIG_PROC_FS
924 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
925 struct file *file, long pages)
927 const unsigned long stack_flags
928 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
931 mm->shared_vm += pages;
932 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
933 mm->exec_vm += pages;
934 } else if (flags & stack_flags)
935 mm->stack_vm += pages;
936 if (flags & (VM_RESERVED|VM_IO))
937 mm->reserved_vm += pages;
939 #endif /* CONFIG_PROC_FS */
942 * The caller must hold down_write(¤t->mm->mmap_sem).
945 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
946 unsigned long len, unsigned long prot,
947 unsigned long flags, unsigned long pgoff)
949 struct mm_struct * mm = current->mm;
953 unsigned long reqprot = prot;
956 * Does the application expect PROT_READ to imply PROT_EXEC?
958 * (the exception is when the underlying filesystem is noexec
959 * mounted, in which case we dont add PROT_EXEC.)
961 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
962 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
968 if (!(flags & MAP_FIXED))
969 addr = round_hint_to_min(addr);
971 /* Careful about overflows.. */
972 len = PAGE_ALIGN(len);
976 /* offset overflow? */
977 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
980 /* Too many mappings? */
981 if (mm->map_count > sysctl_max_map_count)
984 /* Obtain the address to map to. we verify (or select) it and ensure
985 * that it represents a valid section of the address space.
987 addr = get_unmapped_area(file, addr, len, pgoff, flags);
988 if (addr & ~PAGE_MASK)
991 /* Do simple checking here so the lower-level routines won't have
992 * to. we assume access permissions have been handled by the open
993 * of the memory object, so we don't do any here.
995 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
996 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
998 if (flags & MAP_LOCKED)
1002 /* mlock MCL_FUTURE? */
1003 if (vm_flags & VM_LOCKED) {
1004 unsigned long locked, lock_limit;
1005 locked = len >> PAGE_SHIFT;
1006 locked += mm->locked_vm;
1007 lock_limit = rlimit(RLIMIT_MEMLOCK);
1008 lock_limit >>= PAGE_SHIFT;
1009 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1013 inode = file ? file->f_path.dentry->d_inode : NULL;
1016 switch (flags & MAP_TYPE) {
1018 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1022 * Make sure we don't allow writing to an append-only
1025 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1029 * Make sure there are no mandatory locks on the file.
1031 if (locks_verify_locked(inode))
1034 vm_flags |= VM_SHARED | VM_MAYSHARE;
1035 if (!(file->f_mode & FMODE_WRITE))
1036 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1040 if (!(file->f_mode & FMODE_READ))
1042 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1043 if (vm_flags & VM_EXEC)
1045 vm_flags &= ~VM_MAYEXEC;
1048 if (!file->f_op || !file->f_op->mmap)
1056 switch (flags & MAP_TYPE) {
1062 vm_flags |= VM_SHARED | VM_MAYSHARE;
1066 * Set pgoff according to addr for anon_vma.
1068 pgoff = addr >> PAGE_SHIFT;
1075 error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1079 return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1081 EXPORT_SYMBOL(do_mmap_pgoff);
1083 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1084 unsigned long, prot, unsigned long, flags,
1085 unsigned long, fd, unsigned long, pgoff)
1087 struct file *file = NULL;
1088 unsigned long retval = -EBADF;
1090 if (!(flags & MAP_ANONYMOUS)) {
1091 audit_mmap_fd(fd, flags);
1092 if (unlikely(flags & MAP_HUGETLB))
1097 } else if (flags & MAP_HUGETLB) {
1098 struct user_struct *user = NULL;
1100 * VM_NORESERVE is used because the reservations will be
1101 * taken when vm_ops->mmap() is called
1102 * A dummy user value is used because we are not locking
1103 * memory so no accounting is necessary
1105 len = ALIGN(len, huge_page_size(&default_hstate));
1106 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
1107 &user, HUGETLB_ANONHUGE_INODE);
1109 return PTR_ERR(file);
1112 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1114 down_write(¤t->mm->mmap_sem);
1115 retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1116 up_write(¤t->mm->mmap_sem);
1124 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1125 struct mmap_arg_struct {
1129 unsigned long flags;
1131 unsigned long offset;
1134 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1136 struct mmap_arg_struct a;
1138 if (copy_from_user(&a, arg, sizeof(a)))
1140 if (a.offset & ~PAGE_MASK)
1143 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1144 a.offset >> PAGE_SHIFT);
1146 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1149 * Some shared mappigns will want the pages marked read-only
1150 * to track write events. If so, we'll downgrade vm_page_prot
1151 * to the private version (using protection_map[] without the
1154 int vma_wants_writenotify(struct vm_area_struct *vma)
1156 vm_flags_t vm_flags = vma->vm_flags;
1158 /* If it was private or non-writable, the write bit is already clear */
1159 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1162 /* The backer wishes to know when pages are first written to? */
1163 if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1166 /* The open routine did something to the protections already? */
1167 if (pgprot_val(vma->vm_page_prot) !=
1168 pgprot_val(vm_get_page_prot(vm_flags)))
1171 /* Specialty mapping? */
1172 if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1175 /* Can the mapping track the dirty pages? */
1176 return vma->vm_file && vma->vm_file->f_mapping &&
1177 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1181 * We account for memory if it's a private writeable mapping,
1182 * not hugepages and VM_NORESERVE wasn't set.
1184 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1187 * hugetlb has its own accounting separate from the core VM
1188 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1190 if (file && is_file_hugepages(file))
1193 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1196 unsigned long mmap_region(struct file *file, unsigned long addr,
1197 unsigned long len, unsigned long flags,
1198 vm_flags_t vm_flags, unsigned long pgoff)
1200 struct mm_struct *mm = current->mm;
1201 struct vm_area_struct *vma, *prev;
1202 int correct_wcount = 0;
1204 struct rb_node **rb_link, *rb_parent;
1205 unsigned long charged = 0;
1206 struct inode *inode = file ? file->f_path.dentry->d_inode : NULL;
1208 /* Clear old maps */
1211 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1212 if (vma && vma->vm_start < addr + len) {
1213 if (do_munmap(mm, addr, len))
1218 /* Check against address space limit. */
1219 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1223 * Set 'VM_NORESERVE' if we should not account for the
1224 * memory use of this mapping.
1226 if ((flags & MAP_NORESERVE)) {
1227 /* We honor MAP_NORESERVE if allowed to overcommit */
1228 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1229 vm_flags |= VM_NORESERVE;
1231 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1232 if (file && is_file_hugepages(file))
1233 vm_flags |= VM_NORESERVE;
1237 * Private writable mapping: check memory availability
1239 if (accountable_mapping(file, vm_flags)) {
1240 charged = len >> PAGE_SHIFT;
1241 if (security_vm_enough_memory(charged))
1243 vm_flags |= VM_ACCOUNT;
1247 * Can we just expand an old mapping?
1249 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1254 * Determine the object being mapped and call the appropriate
1255 * specific mapper. the address has already been validated, but
1256 * not unmapped, but the maps are removed from the list.
1258 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1265 vma->vm_start = addr;
1266 vma->vm_end = addr + len;
1267 vma->vm_flags = vm_flags;
1268 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1269 vma->vm_pgoff = pgoff;
1270 INIT_LIST_HEAD(&vma->anon_vma_chain);
1274 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1276 if (vm_flags & VM_DENYWRITE) {
1277 error = deny_write_access(file);
1282 vma->vm_file = file;
1284 error = file->f_op->mmap(file, vma);
1286 goto unmap_and_free_vma;
1287 if (vm_flags & VM_EXECUTABLE)
1288 added_exe_file_vma(mm);
1290 /* Can addr have changed??
1292 * Answer: Yes, several device drivers can do it in their
1293 * f_op->mmap method. -DaveM
1295 addr = vma->vm_start;
1296 pgoff = vma->vm_pgoff;
1297 vm_flags = vma->vm_flags;
1298 } else if (vm_flags & VM_SHARED) {
1299 error = shmem_zero_setup(vma);
1304 if (vma_wants_writenotify(vma)) {
1305 pgprot_t pprot = vma->vm_page_prot;
1307 /* Can vma->vm_page_prot have changed??
1309 * Answer: Yes, drivers may have changed it in their
1310 * f_op->mmap method.
1312 * Ensures that vmas marked as uncached stay that way.
1314 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1315 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1316 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1319 vma_link(mm, vma, prev, rb_link, rb_parent);
1320 file = vma->vm_file;
1322 /* Once vma denies write, undo our temporary denial count */
1324 atomic_inc(&inode->i_writecount);
1326 perf_event_mmap(vma);
1328 mm->total_vm += len >> PAGE_SHIFT;
1329 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1330 if (vm_flags & VM_LOCKED) {
1331 if (!mlock_vma_pages_range(vma, addr, addr + len))
1332 mm->locked_vm += (len >> PAGE_SHIFT);
1333 } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1334 make_pages_present(addr, addr + len);
1339 atomic_inc(&inode->i_writecount);
1340 vma->vm_file = NULL;
1343 /* Undo any partial mapping done by a device driver. */
1344 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1347 kmem_cache_free(vm_area_cachep, vma);
1350 vm_unacct_memory(charged);
1354 /* Get an address range which is currently unmapped.
1355 * For shmat() with addr=0.
1357 * Ugly calling convention alert:
1358 * Return value with the low bits set means error value,
1360 * if (ret & ~PAGE_MASK)
1363 * This function "knows" that -ENOMEM has the bits set.
1365 #ifndef HAVE_ARCH_UNMAPPED_AREA
1367 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1368 unsigned long len, unsigned long pgoff, unsigned long flags)
1370 struct mm_struct *mm = current->mm;
1371 struct vm_area_struct *vma;
1372 unsigned long start_addr;
1374 if (len > TASK_SIZE - mmap_min_addr)
1377 if (flags & MAP_FIXED)
1381 addr = PAGE_ALIGN(addr);
1382 vma = find_vma(mm, addr);
1383 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1384 (!vma || addr + len <= vma->vm_start))
1387 if (len > mm->cached_hole_size) {
1388 start_addr = addr = mm->free_area_cache;
1390 start_addr = addr = TASK_UNMAPPED_BASE;
1391 mm->cached_hole_size = 0;
1395 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1396 /* At this point: (!vma || addr < vma->vm_end). */
1397 if (TASK_SIZE - len < addr) {
1399 * Start a new search - just in case we missed
1402 if (start_addr != TASK_UNMAPPED_BASE) {
1403 addr = TASK_UNMAPPED_BASE;
1405 mm->cached_hole_size = 0;
1410 if (!vma || addr + len <= vma->vm_start) {
1412 * Remember the place where we stopped the search:
1414 mm->free_area_cache = addr + len;
1417 if (addr + mm->cached_hole_size < vma->vm_start)
1418 mm->cached_hole_size = vma->vm_start - addr;
1424 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1427 * Is this a new hole at the lowest possible address?
1429 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1430 mm->free_area_cache = addr;
1431 mm->cached_hole_size = ~0UL;
1436 * This mmap-allocator allocates new areas top-down from below the
1437 * stack's low limit (the base):
1439 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1441 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1442 const unsigned long len, const unsigned long pgoff,
1443 const unsigned long flags)
1445 struct vm_area_struct *vma;
1446 struct mm_struct *mm = current->mm;
1447 unsigned long addr = addr0;
1448 unsigned long low_limit = max(PAGE_SIZE, mmap_min_addr);
1450 /* requested length too big for entire address space */
1451 if (len > TASK_SIZE - mmap_min_addr)
1454 if (flags & MAP_FIXED)
1457 /* requesting a specific address */
1459 addr = PAGE_ALIGN(addr);
1460 vma = find_vma(mm, addr);
1461 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1462 (!vma || addr + len <= vma->vm_start))
1466 /* check if free_area_cache is useful for us */
1467 if (len <= mm->cached_hole_size) {
1468 mm->cached_hole_size = 0;
1469 mm->free_area_cache = mm->mmap_base;
1472 /* either no address requested or can't fit in requested address hole */
1473 addr = mm->free_area_cache;
1475 /* make sure it can fit in the remaining address space */
1476 if (addr >= low_limit + len) {
1477 vma = find_vma(mm, addr-len);
1478 if (!vma || addr <= vma->vm_start)
1479 /* remember the address as a hint for next time */
1480 return (mm->free_area_cache = addr-len);
1483 if (mm->mmap_base < low_limit + len)
1486 addr = mm->mmap_base-len;
1490 * Lookup failure means no vma is above this address,
1491 * else if new region fits below vma->vm_start,
1492 * return with success:
1494 vma = find_vma(mm, addr);
1495 if (!vma || addr+len <= vma->vm_start)
1496 /* remember the address as a hint for next time */
1497 return (mm->free_area_cache = addr);
1499 /* remember the largest hole we saw so far */
1500 if (addr + mm->cached_hole_size < vma->vm_start)
1501 mm->cached_hole_size = vma->vm_start - addr;
1503 /* try just below the current vma->vm_start */
1504 addr = vma->vm_start-len;
1505 } while (vma->vm_start >= low_limit + len);
1509 * A failed mmap() very likely causes application failure,
1510 * so fall back to the bottom-up function here. This scenario
1511 * can happen with large stack limits and large mmap()
1514 mm->cached_hole_size = ~0UL;
1515 mm->free_area_cache = TASK_UNMAPPED_BASE;
1516 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1518 * Restore the topdown base:
1520 mm->free_area_cache = mm->mmap_base;
1521 mm->cached_hole_size = ~0UL;
1527 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1530 * Is this a new hole at the highest possible address?
1532 if (addr > mm->free_area_cache)
1533 mm->free_area_cache = addr;
1535 /* dont allow allocations above current base */
1536 if (mm->free_area_cache > mm->mmap_base)
1537 mm->free_area_cache = mm->mmap_base;
1541 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1542 unsigned long pgoff, unsigned long flags)
1544 unsigned long (*get_area)(struct file *, unsigned long,
1545 unsigned long, unsigned long, unsigned long);
1547 unsigned long error = arch_mmap_check(addr, len, flags);
1551 /* Careful about overflows.. */
1552 if (len > TASK_SIZE)
1555 get_area = current->mm->get_unmapped_area;
1556 if (file && file->f_op && file->f_op->get_unmapped_area)
1557 get_area = file->f_op->get_unmapped_area;
1558 addr = get_area(file, addr, len, pgoff, flags);
1559 if (IS_ERR_VALUE(addr))
1562 if (addr > TASK_SIZE - len)
1564 if (addr & ~PAGE_MASK)
1567 return arch_rebalance_pgtables(addr, len);
1570 EXPORT_SYMBOL(get_unmapped_area);
1572 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1573 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1575 struct vm_area_struct *vma = NULL;
1578 /* Check the cache first. */
1579 /* (Cache hit rate is typically around 35%.) */
1580 vma = ACCESS_ONCE(mm->mmap_cache);
1581 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1582 struct rb_node * rb_node;
1584 rb_node = mm->mm_rb.rb_node;
1588 struct vm_area_struct * vma_tmp;
1590 vma_tmp = rb_entry(rb_node,
1591 struct vm_area_struct, vm_rb);
1593 if (vma_tmp->vm_end > addr) {
1595 if (vma_tmp->vm_start <= addr)
1597 rb_node = rb_node->rb_left;
1599 rb_node = rb_node->rb_right;
1602 mm->mmap_cache = vma;
1608 EXPORT_SYMBOL(find_vma);
1610 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1611 struct vm_area_struct *
1612 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1613 struct vm_area_struct **pprev)
1615 struct vm_area_struct *vma = NULL, *prev = NULL;
1616 struct rb_node *rb_node;
1620 /* Guard against addr being lower than the first VMA */
1623 /* Go through the RB tree quickly. */
1624 rb_node = mm->mm_rb.rb_node;
1627 struct vm_area_struct *vma_tmp;
1628 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1630 if (addr < vma_tmp->vm_end) {
1631 rb_node = rb_node->rb_left;
1634 if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1636 rb_node = rb_node->rb_right;
1642 return prev ? prev->vm_next : vma;
1646 * Verify that the stack growth is acceptable and
1647 * update accounting. This is shared with both the
1648 * grow-up and grow-down cases.
1650 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1652 struct mm_struct *mm = vma->vm_mm;
1653 struct rlimit *rlim = current->signal->rlim;
1654 unsigned long new_start, actual_size;
1656 /* address space limit tests */
1657 if (!may_expand_vm(mm, grow))
1660 /* Stack limit test */
1662 if (size && (vma->vm_flags & (VM_GROWSUP | VM_GROWSDOWN)))
1663 actual_size -= PAGE_SIZE;
1664 if (actual_size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1667 /* mlock limit tests */
1668 if (vma->vm_flags & VM_LOCKED) {
1669 unsigned long locked;
1670 unsigned long limit;
1671 locked = mm->locked_vm + grow;
1672 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1673 limit >>= PAGE_SHIFT;
1674 if (locked > limit && !capable(CAP_IPC_LOCK))
1678 /* Check to ensure the stack will not grow into a hugetlb-only region */
1679 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1681 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1685 * Overcommit.. This must be the final test, as it will
1686 * update security statistics.
1688 if (security_vm_enough_memory_mm(mm, grow))
1691 /* Ok, everything looks good - let it rip */
1692 mm->total_vm += grow;
1693 if (vma->vm_flags & VM_LOCKED)
1694 mm->locked_vm += grow;
1695 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1699 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1701 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1702 * vma is the last one with address > vma->vm_end. Have to extend vma.
1704 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1708 if (!(vma->vm_flags & VM_GROWSUP))
1712 * We must make sure the anon_vma is allocated
1713 * so that the anon_vma locking is not a noop.
1715 if (unlikely(anon_vma_prepare(vma)))
1717 vma_lock_anon_vma(vma);
1720 * vma->vm_start/vm_end cannot change under us because the caller
1721 * is required to hold the mmap_sem in read mode. We need the
1722 * anon_vma lock to serialize against concurrent expand_stacks.
1723 * Also guard against wrapping around to address 0.
1725 if (address < PAGE_ALIGN(address+4))
1726 address = PAGE_ALIGN(address+4);
1728 vma_unlock_anon_vma(vma);
1733 /* Somebody else might have raced and expanded it already */
1734 if (address > vma->vm_end) {
1735 unsigned long size, grow;
1737 size = address - vma->vm_start;
1738 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1741 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1742 error = acct_stack_growth(vma, size, grow);
1744 vma->vm_end = address;
1745 perf_event_mmap(vma);
1749 vma_unlock_anon_vma(vma);
1750 khugepaged_enter_vma_merge(vma, vma->vm_flags);
1753 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1756 * vma is the first one with address < vma->vm_start. Have to extend vma.
1758 int expand_downwards(struct vm_area_struct *vma,
1759 unsigned long address)
1764 * We must make sure the anon_vma is allocated
1765 * so that the anon_vma locking is not a noop.
1767 if (unlikely(anon_vma_prepare(vma)))
1770 address &= PAGE_MASK;
1771 error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1775 vma_lock_anon_vma(vma);
1778 * vma->vm_start/vm_end cannot change under us because the caller
1779 * is required to hold the mmap_sem in read mode. We need the
1780 * anon_vma lock to serialize against concurrent expand_stacks.
1783 /* Somebody else might have raced and expanded it already */
1784 if (address < vma->vm_start) {
1785 unsigned long size, grow;
1787 size = vma->vm_end - address;
1788 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1791 if (grow <= vma->vm_pgoff) {
1792 error = acct_stack_growth(vma, size, grow);
1794 vma->vm_start = address;
1795 vma->vm_pgoff -= grow;
1796 perf_event_mmap(vma);
1800 vma_unlock_anon_vma(vma);
1801 khugepaged_enter_vma_merge(vma, vma->vm_flags);
1806 * Note how expand_stack() refuses to expand the stack all the way to
1807 * abut the next virtual mapping, *unless* that mapping itself is also
1808 * a stack mapping. We want to leave room for a guard page, after all
1809 * (the guard page itself is not added here, that is done by the
1810 * actual page faulting logic)
1812 * This matches the behavior of the guard page logic (see mm/memory.c:
1813 * check_stack_guard_page()), which only allows the guard page to be
1814 * removed under these circumstances.
1816 #ifdef CONFIG_STACK_GROWSUP
1817 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1819 struct vm_area_struct *next;
1821 address &= PAGE_MASK;
1822 next = vma->vm_next;
1823 if (next && next->vm_start == address + PAGE_SIZE) {
1824 if (!(next->vm_flags & VM_GROWSUP))
1827 return expand_upwards(vma, address);
1830 struct vm_area_struct *
1831 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1833 struct vm_area_struct *vma, *prev;
1836 vma = find_vma_prev(mm, addr, &prev);
1837 if (vma && (vma->vm_start <= addr))
1839 if (!prev || expand_stack(prev, addr))
1841 if (prev->vm_flags & VM_LOCKED) {
1842 mlock_vma_pages_range(prev, addr, prev->vm_end);
1847 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1849 struct vm_area_struct *prev;
1851 address &= PAGE_MASK;
1852 prev = vma->vm_prev;
1853 if (prev && prev->vm_end == address) {
1854 if (!(prev->vm_flags & VM_GROWSDOWN))
1857 return expand_downwards(vma, address);
1860 struct vm_area_struct *
1861 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1863 struct vm_area_struct * vma;
1864 unsigned long start;
1867 vma = find_vma(mm,addr);
1870 if (vma->vm_start <= addr)
1872 if (!(vma->vm_flags & VM_GROWSDOWN))
1874 start = vma->vm_start;
1875 if (expand_stack(vma, addr))
1877 if (vma->vm_flags & VM_LOCKED) {
1878 mlock_vma_pages_range(vma, addr, start);
1885 * Ok - we have the memory areas we should free on the vma list,
1886 * so release them, and do the vma updates.
1888 * Called with the mm semaphore held.
1890 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1892 /* Update high watermark before we lower total_vm */
1893 update_hiwater_vm(mm);
1895 long nrpages = vma_pages(vma);
1897 mm->total_vm -= nrpages;
1898 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1899 vma = remove_vma(vma);
1905 * Get rid of page table information in the indicated region.
1907 * Called with the mm semaphore held.
1909 static void unmap_region(struct mm_struct *mm,
1910 struct vm_area_struct *vma, struct vm_area_struct *prev,
1911 unsigned long start, unsigned long end)
1913 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1914 struct mmu_gather tlb;
1915 unsigned long nr_accounted = 0;
1918 tlb_gather_mmu(&tlb, mm, 0);
1919 update_hiwater_rss(mm);
1920 unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1921 vm_unacct_memory(nr_accounted);
1922 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
1923 next ? next->vm_start : 0);
1924 tlb_finish_mmu(&tlb, start, end);
1928 * Create a list of vma's touched by the unmap, removing them from the mm's
1929 * vma list as we go..
1932 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1933 struct vm_area_struct *prev, unsigned long end)
1935 struct vm_area_struct **insertion_point;
1936 struct vm_area_struct *tail_vma = NULL;
1939 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1940 vma->vm_prev = NULL;
1942 rb_erase(&vma->vm_rb, &mm->mm_rb);
1946 } while (vma && vma->vm_start < end);
1947 *insertion_point = vma;
1949 vma->vm_prev = prev;
1950 tail_vma->vm_next = NULL;
1951 if (mm->unmap_area == arch_unmap_area)
1952 addr = prev ? prev->vm_end : mm->mmap_base;
1954 addr = vma ? vma->vm_start : mm->mmap_base;
1955 mm->unmap_area(mm, addr);
1956 mm->mmap_cache = NULL; /* Kill the cache. */
1960 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
1961 * munmap path where it doesn't make sense to fail.
1963 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1964 unsigned long addr, int new_below)
1966 struct mempolicy *pol;
1967 struct vm_area_struct *new;
1970 if (is_vm_hugetlb_page(vma) && (addr &
1971 ~(huge_page_mask(hstate_vma(vma)))))
1974 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1978 /* most fields are the same, copy all, and then fixup */
1981 INIT_LIST_HEAD(&new->anon_vma_chain);
1986 new->vm_start = addr;
1987 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1990 pol = mpol_dup(vma_policy(vma));
1995 vma_set_policy(new, pol);
1997 if (anon_vma_clone(new, vma))
2001 get_file(new->vm_file);
2002 if (vma->vm_flags & VM_EXECUTABLE)
2003 added_exe_file_vma(mm);
2006 if (new->vm_ops && new->vm_ops->open)
2007 new->vm_ops->open(new);
2010 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2011 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2013 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2019 /* Clean everything up if vma_adjust failed. */
2020 if (new->vm_ops && new->vm_ops->close)
2021 new->vm_ops->close(new);
2023 if (vma->vm_flags & VM_EXECUTABLE)
2024 removed_exe_file_vma(mm);
2027 unlink_anon_vmas(new);
2031 kmem_cache_free(vm_area_cachep, new);
2037 * Split a vma into two pieces at address 'addr', a new vma is allocated
2038 * either for the first part or the tail.
2040 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2041 unsigned long addr, int new_below)
2043 if (mm->map_count >= sysctl_max_map_count)
2046 return __split_vma(mm, vma, addr, new_below);
2049 /* Munmap is split into 2 main parts -- this part which finds
2050 * what needs doing, and the areas themselves, which do the
2051 * work. This now handles partial unmappings.
2052 * Jeremy Fitzhardinge <jeremy@goop.org>
2054 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2057 struct vm_area_struct *vma, *prev, *last;
2059 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2062 if ((len = PAGE_ALIGN(len)) == 0)
2065 /* Find the first overlapping VMA */
2066 vma = find_vma(mm, start);
2069 prev = vma->vm_prev;
2070 /* we have start < vma->vm_end */
2072 /* if it doesn't overlap, we have nothing.. */
2074 if (vma->vm_start >= end)
2078 * If we need to split any vma, do it now to save pain later.
2080 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2081 * unmapped vm_area_struct will remain in use: so lower split_vma
2082 * places tmp vma above, and higher split_vma places tmp vma below.
2084 if (start > vma->vm_start) {
2088 * Make sure that map_count on return from munmap() will
2089 * not exceed its limit; but let map_count go just above
2090 * its limit temporarily, to help free resources as expected.
2092 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2095 error = __split_vma(mm, vma, start, 0);
2101 /* Does it split the last one? */
2102 last = find_vma(mm, end);
2103 if (last && end > last->vm_start) {
2104 int error = __split_vma(mm, last, end, 1);
2108 vma = prev? prev->vm_next: mm->mmap;
2111 * unlock any mlock()ed ranges before detaching vmas
2113 if (mm->locked_vm) {
2114 struct vm_area_struct *tmp = vma;
2115 while (tmp && tmp->vm_start < end) {
2116 if (tmp->vm_flags & VM_LOCKED) {
2117 mm->locked_vm -= vma_pages(tmp);
2118 munlock_vma_pages_all(tmp);
2125 * Remove the vma's, and unmap the actual pages
2127 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2128 unmap_region(mm, vma, prev, start, end);
2130 /* Fix up all other VM information */
2131 remove_vma_list(mm, vma);
2136 EXPORT_SYMBOL(do_munmap);
2138 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2141 struct mm_struct *mm = current->mm;
2143 profile_munmap(addr);
2145 down_write(&mm->mmap_sem);
2146 ret = do_munmap(mm, addr, len);
2147 up_write(&mm->mmap_sem);
2151 static inline void verify_mm_writelocked(struct mm_struct *mm)
2153 #ifdef CONFIG_DEBUG_VM
2154 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2156 up_read(&mm->mmap_sem);
2162 * this is really a simplified "do_mmap". it only handles
2163 * anonymous maps. eventually we may be able to do some
2164 * brk-specific accounting here.
2166 unsigned long do_brk(unsigned long addr, unsigned long len)
2168 struct mm_struct * mm = current->mm;
2169 struct vm_area_struct * vma, * prev;
2170 unsigned long flags;
2171 struct rb_node ** rb_link, * rb_parent;
2172 pgoff_t pgoff = addr >> PAGE_SHIFT;
2175 len = PAGE_ALIGN(len);
2179 error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2183 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2185 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2186 if (error & ~PAGE_MASK)
2192 if (mm->def_flags & VM_LOCKED) {
2193 unsigned long locked, lock_limit;
2194 locked = len >> PAGE_SHIFT;
2195 locked += mm->locked_vm;
2196 lock_limit = rlimit(RLIMIT_MEMLOCK);
2197 lock_limit >>= PAGE_SHIFT;
2198 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2203 * mm->mmap_sem is required to protect against another thread
2204 * changing the mappings in case we sleep.
2206 verify_mm_writelocked(mm);
2209 * Clear old maps. this also does some error checking for us
2212 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2213 if (vma && vma->vm_start < addr + len) {
2214 if (do_munmap(mm, addr, len))
2219 /* Check against address space limits *after* clearing old maps... */
2220 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2223 if (mm->map_count > sysctl_max_map_count)
2226 if (security_vm_enough_memory(len >> PAGE_SHIFT))
2229 /* Can we just expand an old private anonymous mapping? */
2230 vma = vma_merge(mm, prev, addr, addr + len, flags,
2231 NULL, NULL, pgoff, NULL);
2236 * create a vma struct for an anonymous mapping
2238 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2240 vm_unacct_memory(len >> PAGE_SHIFT);
2244 INIT_LIST_HEAD(&vma->anon_vma_chain);
2246 vma->vm_start = addr;
2247 vma->vm_end = addr + len;
2248 vma->vm_pgoff = pgoff;
2249 vma->vm_flags = flags;
2250 vma->vm_page_prot = vm_get_page_prot(flags);
2251 vma_link(mm, vma, prev, rb_link, rb_parent);
2253 perf_event_mmap(vma);
2254 mm->total_vm += len >> PAGE_SHIFT;
2255 if (flags & VM_LOCKED) {
2256 if (!mlock_vma_pages_range(vma, addr, addr + len))
2257 mm->locked_vm += (len >> PAGE_SHIFT);
2262 EXPORT_SYMBOL(do_brk);
2264 /* Release all mmaps. */
2265 void exit_mmap(struct mm_struct *mm)
2267 struct mmu_gather tlb;
2268 struct vm_area_struct *vma;
2269 unsigned long nr_accounted = 0;
2272 /* mm's last user has gone, and its about to be pulled down */
2273 mmu_notifier_release(mm);
2275 if (mm->locked_vm) {
2278 if (vma->vm_flags & VM_LOCKED)
2279 munlock_vma_pages_all(vma);
2287 if (!vma) /* Can happen if dup_mmap() received an OOM */
2292 tlb_gather_mmu(&tlb, mm, 1);
2293 /* update_hiwater_rss(mm) here? but nobody should be looking */
2294 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2295 end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2296 vm_unacct_memory(nr_accounted);
2298 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2299 tlb_finish_mmu(&tlb, 0, end);
2302 * Walk the list again, actually closing and freeing it,
2303 * with preemption enabled, without holding any MM locks.
2306 vma = remove_vma(vma);
2308 BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2311 /* Insert vm structure into process list sorted by address
2312 * and into the inode's i_mmap tree. If vm_file is non-NULL
2313 * then i_mmap_mutex is taken here.
2315 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2317 struct vm_area_struct * __vma, * prev;
2318 struct rb_node ** rb_link, * rb_parent;
2321 * The vm_pgoff of a purely anonymous vma should be irrelevant
2322 * until its first write fault, when page's anon_vma and index
2323 * are set. But now set the vm_pgoff it will almost certainly
2324 * end up with (unless mremap moves it elsewhere before that
2325 * first wfault), so /proc/pid/maps tells a consistent story.
2327 * By setting it to reflect the virtual start address of the
2328 * vma, merges and splits can happen in a seamless way, just
2329 * using the existing file pgoff checks and manipulations.
2330 * Similarly in do_mmap_pgoff and in do_brk.
2332 if (!vma->vm_file) {
2333 BUG_ON(vma->anon_vma);
2334 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2336 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2337 if (__vma && __vma->vm_start < vma->vm_end)
2339 if ((vma->vm_flags & VM_ACCOUNT) &&
2340 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2342 vma_link(mm, vma, prev, rb_link, rb_parent);
2347 * Copy the vma structure to a new location in the same mm,
2348 * prior to moving page table entries, to effect an mremap move.
2350 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2351 unsigned long addr, unsigned long len, pgoff_t pgoff)
2353 struct vm_area_struct *vma = *vmap;
2354 unsigned long vma_start = vma->vm_start;
2355 struct mm_struct *mm = vma->vm_mm;
2356 struct vm_area_struct *new_vma, *prev;
2357 struct rb_node **rb_link, *rb_parent;
2358 struct mempolicy *pol;
2361 * If anonymous vma has not yet been faulted, update new pgoff
2362 * to match new location, to increase its chance of merging.
2364 if (!vma->vm_file && !vma->anon_vma)
2365 pgoff = addr >> PAGE_SHIFT;
2367 find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2368 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2369 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2372 * Source vma may have been merged into new_vma
2374 if (vma_start >= new_vma->vm_start &&
2375 vma_start < new_vma->vm_end)
2378 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2381 pol = mpol_dup(vma_policy(vma));
2384 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2385 if (anon_vma_clone(new_vma, vma))
2386 goto out_free_mempol;
2387 vma_set_policy(new_vma, pol);
2388 new_vma->vm_start = addr;
2389 new_vma->vm_end = addr + len;
2390 new_vma->vm_pgoff = pgoff;
2391 if (new_vma->vm_file) {
2392 get_file(new_vma->vm_file);
2393 if (vma->vm_flags & VM_EXECUTABLE)
2394 added_exe_file_vma(mm);
2396 if (new_vma->vm_ops && new_vma->vm_ops->open)
2397 new_vma->vm_ops->open(new_vma);
2398 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2406 kmem_cache_free(vm_area_cachep, new_vma);
2411 * Return true if the calling process may expand its vm space by the passed
2414 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2416 unsigned long cur = mm->total_vm; /* pages */
2419 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2421 if (cur + npages > lim)
2427 static int special_mapping_fault(struct vm_area_struct *vma,
2428 struct vm_fault *vmf)
2431 struct page **pages;
2434 * special mappings have no vm_file, and in that case, the mm
2435 * uses vm_pgoff internally. So we have to subtract it from here.
2436 * We are allowed to do this because we are the mm; do not copy
2437 * this code into drivers!
2439 pgoff = vmf->pgoff - vma->vm_pgoff;
2441 for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2445 struct page *page = *pages;
2451 return VM_FAULT_SIGBUS;
2455 * Having a close hook prevents vma merging regardless of flags.
2457 static void special_mapping_close(struct vm_area_struct *vma)
2461 static const struct vm_operations_struct special_mapping_vmops = {
2462 .close = special_mapping_close,
2463 .fault = special_mapping_fault,
2467 * Called with mm->mmap_sem held for writing.
2468 * Insert a new vma covering the given region, with the given flags.
2469 * Its pages are supplied by the given array of struct page *.
2470 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2471 * The region past the last page supplied will always produce SIGBUS.
2472 * The array pointer and the pages it points to are assumed to stay alive
2473 * for as long as this mapping might exist.
2475 int install_special_mapping(struct mm_struct *mm,
2476 unsigned long addr, unsigned long len,
2477 unsigned long vm_flags, struct page **pages)
2480 struct vm_area_struct *vma;
2482 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2483 if (unlikely(vma == NULL))
2486 INIT_LIST_HEAD(&vma->anon_vma_chain);
2488 vma->vm_start = addr;
2489 vma->vm_end = addr + len;
2491 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2492 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2494 vma->vm_ops = &special_mapping_vmops;
2495 vma->vm_private_data = pages;
2497 ret = security_file_mmap(NULL, 0, 0, 0, vma->vm_start, 1);
2501 ret = insert_vm_struct(mm, vma);
2505 mm->total_vm += len >> PAGE_SHIFT;
2507 perf_event_mmap(vma);
2512 kmem_cache_free(vm_area_cachep, vma);
2516 static DEFINE_MUTEX(mm_all_locks_mutex);
2518 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2520 if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2522 * The LSB of head.next can't change from under us
2523 * because we hold the mm_all_locks_mutex.
2525 mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem);
2527 * We can safely modify head.next after taking the
2528 * anon_vma->root->mutex. If some other vma in this mm shares
2529 * the same anon_vma we won't take it again.
2531 * No need of atomic instructions here, head.next
2532 * can't change from under us thanks to the
2533 * anon_vma->root->mutex.
2535 if (__test_and_set_bit(0, (unsigned long *)
2536 &anon_vma->root->head.next))
2541 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2543 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2545 * AS_MM_ALL_LOCKS can't change from under us because
2546 * we hold the mm_all_locks_mutex.
2548 * Operations on ->flags have to be atomic because
2549 * even if AS_MM_ALL_LOCKS is stable thanks to the
2550 * mm_all_locks_mutex, there may be other cpus
2551 * changing other bitflags in parallel to us.
2553 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2555 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2560 * This operation locks against the VM for all pte/vma/mm related
2561 * operations that could ever happen on a certain mm. This includes
2562 * vmtruncate, try_to_unmap, and all page faults.
2564 * The caller must take the mmap_sem in write mode before calling
2565 * mm_take_all_locks(). The caller isn't allowed to release the
2566 * mmap_sem until mm_drop_all_locks() returns.
2568 * mmap_sem in write mode is required in order to block all operations
2569 * that could modify pagetables and free pages without need of
2570 * altering the vma layout (for example populate_range() with
2571 * nonlinear vmas). It's also needed in write mode to avoid new
2572 * anon_vmas to be associated with existing vmas.
2574 * A single task can't take more than one mm_take_all_locks() in a row
2575 * or it would deadlock.
2577 * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2578 * mapping->flags avoid to take the same lock twice, if more than one
2579 * vma in this mm is backed by the same anon_vma or address_space.
2581 * We can take all the locks in random order because the VM code
2582 * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2583 * takes more than one of them in a row. Secondly we're protected
2584 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2586 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2587 * that may have to take thousand of locks.
2589 * mm_take_all_locks() can fail if it's interrupted by signals.
2591 int mm_take_all_locks(struct mm_struct *mm)
2593 struct vm_area_struct *vma;
2594 struct anon_vma_chain *avc;
2596 BUG_ON(down_read_trylock(&mm->mmap_sem));
2598 mutex_lock(&mm_all_locks_mutex);
2600 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2601 if (signal_pending(current))
2603 if (vma->vm_file && vma->vm_file->f_mapping)
2604 vm_lock_mapping(mm, vma->vm_file->f_mapping);
2607 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2608 if (signal_pending(current))
2611 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2612 vm_lock_anon_vma(mm, avc->anon_vma);
2618 mm_drop_all_locks(mm);
2622 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2624 if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2626 * The LSB of head.next can't change to 0 from under
2627 * us because we hold the mm_all_locks_mutex.
2629 * We must however clear the bitflag before unlocking
2630 * the vma so the users using the anon_vma->head will
2631 * never see our bitflag.
2633 * No need of atomic instructions here, head.next
2634 * can't change from under us until we release the
2635 * anon_vma->root->mutex.
2637 if (!__test_and_clear_bit(0, (unsigned long *)
2638 &anon_vma->root->head.next))
2640 anon_vma_unlock(anon_vma);
2644 static void vm_unlock_mapping(struct address_space *mapping)
2646 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2648 * AS_MM_ALL_LOCKS can't change to 0 from under us
2649 * because we hold the mm_all_locks_mutex.
2651 mutex_unlock(&mapping->i_mmap_mutex);
2652 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2659 * The mmap_sem cannot be released by the caller until
2660 * mm_drop_all_locks() returns.
2662 void mm_drop_all_locks(struct mm_struct *mm)
2664 struct vm_area_struct *vma;
2665 struct anon_vma_chain *avc;
2667 BUG_ON(down_read_trylock(&mm->mmap_sem));
2668 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2670 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2672 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2673 vm_unlock_anon_vma(avc->anon_vma);
2674 if (vma->vm_file && vma->vm_file->f_mapping)
2675 vm_unlock_mapping(vma->vm_file->f_mapping);
2678 mutex_unlock(&mm_all_locks_mutex);
2682 * initialise the VMA slab
2684 void __init mmap_init(void)
2688 ret = percpu_counter_init(&vm_committed_as, 0);