6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
9 #include <linux/kernel.h>
10 #include <linux/slab.h>
11 #include <linux/backing-dev.h>
13 #include <linux/shm.h>
14 #include <linux/mman.h>
15 #include <linux/pagemap.h>
16 #include <linux/swap.h>
17 #include <linux/syscalls.h>
18 #include <linux/capability.h>
19 #include <linux/init.h>
20 #include <linux/file.h>
22 #include <linux/personality.h>
23 #include <linux/security.h>
24 #include <linux/hugetlb.h>
25 #include <linux/profile.h>
26 #include <linux/export.h>
27 #include <linux/mount.h>
28 #include <linux/mempolicy.h>
29 #include <linux/rmap.h>
30 #include <linux/mmu_notifier.h>
31 #include <linux/perf_event.h>
32 #include <linux/audit.h>
33 #include <linux/khugepaged.h>
34 #include <linux/uprobes.h>
35 #include <linux/rbtree_augmented.h>
36 #include <linux/sched/sysctl.h>
37 #include <linux/notifier.h>
38 #include <linux/memory.h>
40 #include <asm/uaccess.h>
41 #include <asm/cacheflush.h>
43 #include <asm/mmu_context.h>
47 #ifndef arch_mmap_check
48 #define arch_mmap_check(addr, len, flags) (0)
51 #ifndef arch_rebalance_pgtables
52 #define arch_rebalance_pgtables(addr, len) (addr)
55 static void unmap_region(struct mm_struct *mm,
56 struct vm_area_struct *vma, struct vm_area_struct *prev,
57 unsigned long start, unsigned long end);
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;
90 unsigned long sysctl_user_reserve_kbytes __read_mostly = 1UL << 17; /* 128MB */
91 unsigned long sysctl_admin_reserve_kbytes __read_mostly = 1UL << 13; /* 8MB */
93 * Make sure vm_committed_as in one cacheline and not cacheline shared with
94 * other variables. It can be updated by several CPUs frequently.
96 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
99 * The global memory commitment made in the system can be a metric
100 * that can be used to drive ballooning decisions when Linux is hosted
101 * as a guest. On Hyper-V, the host implements a policy engine for dynamically
102 * balancing memory across competing virtual machines that are hosted.
103 * Several metrics drive this policy engine including the guest reported
106 unsigned long vm_memory_committed(void)
108 return percpu_counter_read_positive(&vm_committed_as);
110 EXPORT_SYMBOL_GPL(vm_memory_committed);
113 * Check that a process has enough memory to allocate a new virtual
114 * mapping. 0 means there is enough memory for the allocation to
115 * succeed and -ENOMEM implies there is not.
117 * We currently support three overcommit policies, which are set via the
118 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
120 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
121 * Additional code 2002 Jul 20 by Robert Love.
123 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
125 * Note this is a helper function intended to be used by LSMs which
126 * wish to use this logic.
128 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
130 unsigned long free, allowed, reserve;
132 vm_acct_memory(pages);
135 * Sometimes we want to use more memory than we have
137 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
140 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
141 free = global_page_state(NR_FREE_PAGES);
142 free += global_page_state(NR_FILE_PAGES);
145 * shmem pages shouldn't be counted as free in this
146 * case, they can't be purged, only swapped out, and
147 * that won't affect the overall amount of available
148 * memory in the system.
150 free -= global_page_state(NR_SHMEM);
152 free += get_nr_swap_pages();
155 * Any slabs which are created with the
156 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
157 * which are reclaimable, under pressure. The dentry
158 * cache and most inode caches should fall into this
160 free += global_page_state(NR_SLAB_RECLAIMABLE);
163 * Leave reserved pages. The pages are not for anonymous pages.
165 if (free <= totalreserve_pages)
168 free -= totalreserve_pages;
171 * Reserve some for root
174 free -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
182 allowed = (totalram_pages - hugetlb_total_pages())
183 * sysctl_overcommit_ratio / 100;
185 * Reserve some for root
188 allowed -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
189 allowed += total_swap_pages;
192 * Don't let a single process grow so big a user can't recover
195 reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10);
196 allowed -= min(mm->total_vm / 32, reserve);
199 if (percpu_counter_read_positive(&vm_committed_as) < allowed)
202 vm_unacct_memory(pages);
208 * Requires inode->i_mapping->i_mmap_mutex
210 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
211 struct file *file, struct address_space *mapping)
213 if (vma->vm_flags & VM_DENYWRITE)
214 atomic_inc(&file_inode(file)->i_writecount);
215 if (vma->vm_flags & VM_SHARED)
216 mapping->i_mmap_writable--;
218 flush_dcache_mmap_lock(mapping);
219 if (unlikely(vma->vm_flags & VM_NONLINEAR))
220 list_del_init(&vma->shared.nonlinear);
222 vma_interval_tree_remove(vma, &mapping->i_mmap);
223 flush_dcache_mmap_unlock(mapping);
227 * Unlink a file-based vm structure from its interval tree, to hide
228 * vma from rmap and vmtruncate before freeing its page tables.
230 void unlink_file_vma(struct vm_area_struct *vma)
232 struct file *file = vma->vm_file;
235 struct address_space *mapping = file->f_mapping;
236 mutex_lock(&mapping->i_mmap_mutex);
237 __remove_shared_vm_struct(vma, file, mapping);
238 mutex_unlock(&mapping->i_mmap_mutex);
243 * Close a vm structure and free it, returning the next.
245 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
247 struct vm_area_struct *next = vma->vm_next;
250 if (vma->vm_ops && vma->vm_ops->close)
251 vma->vm_ops->close(vma);
254 mpol_put(vma_policy(vma));
255 kmem_cache_free(vm_area_cachep, vma);
259 static unsigned long do_brk(unsigned long addr, unsigned long len);
261 SYSCALL_DEFINE1(brk, unsigned long, brk)
263 unsigned long rlim, retval;
264 unsigned long newbrk, oldbrk;
265 struct mm_struct *mm = current->mm;
266 unsigned long min_brk;
269 down_write(&mm->mmap_sem);
271 #ifdef CONFIG_COMPAT_BRK
273 * CONFIG_COMPAT_BRK can still be overridden by setting
274 * randomize_va_space to 2, which will still cause mm->start_brk
275 * to be arbitrarily shifted
277 if (current->brk_randomized)
278 min_brk = mm->start_brk;
280 min_brk = mm->end_data;
282 min_brk = mm->start_brk;
288 * Check against rlimit here. If this check is done later after the test
289 * of oldbrk with newbrk then it can escape the test and let the data
290 * segment grow beyond its set limit the in case where the limit is
291 * not page aligned -Ram Gupta
293 rlim = rlimit(RLIMIT_DATA);
294 if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
295 (mm->end_data - mm->start_data) > rlim)
298 newbrk = PAGE_ALIGN(brk);
299 oldbrk = PAGE_ALIGN(mm->brk);
300 if (oldbrk == newbrk)
303 /* Always allow shrinking brk. */
304 if (brk <= mm->brk) {
305 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
310 /* Check against existing mmap mappings. */
311 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
314 /* Ok, looks good - let it rip. */
315 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
320 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
321 up_write(&mm->mmap_sem);
323 mm_populate(oldbrk, newbrk - oldbrk);
328 up_write(&mm->mmap_sem);
332 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
334 unsigned long max, subtree_gap;
337 max -= vma->vm_prev->vm_end;
338 if (vma->vm_rb.rb_left) {
339 subtree_gap = rb_entry(vma->vm_rb.rb_left,
340 struct vm_area_struct, vm_rb)->rb_subtree_gap;
341 if (subtree_gap > max)
344 if (vma->vm_rb.rb_right) {
345 subtree_gap = rb_entry(vma->vm_rb.rb_right,
346 struct vm_area_struct, vm_rb)->rb_subtree_gap;
347 if (subtree_gap > max)
353 #ifdef CONFIG_DEBUG_VM_RB
354 static int browse_rb(struct rb_root *root)
356 int i = 0, j, bug = 0;
357 struct rb_node *nd, *pn = NULL;
358 unsigned long prev = 0, pend = 0;
360 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
361 struct vm_area_struct *vma;
362 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
363 if (vma->vm_start < prev) {
364 printk("vm_start %lx prev %lx\n", vma->vm_start, prev);
367 if (vma->vm_start < pend) {
368 printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
371 if (vma->vm_start > vma->vm_end) {
372 printk("vm_end %lx < vm_start %lx\n",
373 vma->vm_end, vma->vm_start);
376 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
377 printk("free gap %lx, correct %lx\n",
379 vma_compute_subtree_gap(vma));
384 prev = vma->vm_start;
388 for (nd = pn; nd; nd = rb_prev(nd))
391 printk("backwards %d, forwards %d\n", j, i);
397 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
401 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
402 struct vm_area_struct *vma;
403 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
404 BUG_ON(vma != ignore &&
405 vma->rb_subtree_gap != vma_compute_subtree_gap(vma));
409 void validate_mm(struct mm_struct *mm)
413 unsigned long highest_address = 0;
414 struct vm_area_struct *vma = mm->mmap;
416 struct anon_vma_chain *avc;
417 vma_lock_anon_vma(vma);
418 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
419 anon_vma_interval_tree_verify(avc);
420 vma_unlock_anon_vma(vma);
421 highest_address = vma->vm_end;
425 if (i != mm->map_count) {
426 printk("map_count %d vm_next %d\n", mm->map_count, i);
429 if (highest_address != mm->highest_vm_end) {
430 printk("mm->highest_vm_end %lx, found %lx\n",
431 mm->highest_vm_end, highest_address);
434 i = browse_rb(&mm->mm_rb);
435 if (i != mm->map_count) {
436 printk("map_count %d rb %d\n", mm->map_count, i);
442 #define validate_mm_rb(root, ignore) do { } while (0)
443 #define validate_mm(mm) do { } while (0)
446 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
447 unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
450 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
451 * vma->vm_prev->vm_end values changed, without modifying the vma's position
454 static void vma_gap_update(struct vm_area_struct *vma)
457 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
458 * function that does exacltly what we want.
460 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
463 static inline void vma_rb_insert(struct vm_area_struct *vma,
464 struct rb_root *root)
466 /* All rb_subtree_gap values must be consistent prior to insertion */
467 validate_mm_rb(root, NULL);
469 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
472 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
475 * All rb_subtree_gap values must be consistent prior to erase,
476 * with the possible exception of the vma being erased.
478 validate_mm_rb(root, vma);
481 * Note rb_erase_augmented is a fairly large inline function,
482 * so make sure we instantiate it only once with our desired
483 * augmented rbtree callbacks.
485 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
489 * vma has some anon_vma assigned, and is already inserted on that
490 * anon_vma's interval trees.
492 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
493 * vma must be removed from the anon_vma's interval trees using
494 * anon_vma_interval_tree_pre_update_vma().
496 * After the update, the vma will be reinserted using
497 * anon_vma_interval_tree_post_update_vma().
499 * The entire update must be protected by exclusive mmap_sem and by
500 * the root anon_vma's mutex.
503 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
505 struct anon_vma_chain *avc;
507 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
508 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
512 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
514 struct anon_vma_chain *avc;
516 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
517 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
520 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
521 unsigned long end, struct vm_area_struct **pprev,
522 struct rb_node ***rb_link, struct rb_node **rb_parent)
524 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
526 __rb_link = &mm->mm_rb.rb_node;
527 rb_prev = __rb_parent = NULL;
530 struct vm_area_struct *vma_tmp;
532 __rb_parent = *__rb_link;
533 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
535 if (vma_tmp->vm_end > addr) {
536 /* Fail if an existing vma overlaps the area */
537 if (vma_tmp->vm_start < end)
539 __rb_link = &__rb_parent->rb_left;
541 rb_prev = __rb_parent;
542 __rb_link = &__rb_parent->rb_right;
548 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
549 *rb_link = __rb_link;
550 *rb_parent = __rb_parent;
554 static unsigned long count_vma_pages_range(struct mm_struct *mm,
555 unsigned long addr, unsigned long end)
557 unsigned long nr_pages = 0;
558 struct vm_area_struct *vma;
560 /* Find first overlaping mapping */
561 vma = find_vma_intersection(mm, addr, end);
565 nr_pages = (min(end, vma->vm_end) -
566 max(addr, vma->vm_start)) >> PAGE_SHIFT;
568 /* Iterate over the rest of the overlaps */
569 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
570 unsigned long overlap_len;
572 if (vma->vm_start > end)
575 overlap_len = min(end, vma->vm_end) - vma->vm_start;
576 nr_pages += overlap_len >> PAGE_SHIFT;
582 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
583 struct rb_node **rb_link, struct rb_node *rb_parent)
585 /* Update tracking information for the gap following the new vma. */
587 vma_gap_update(vma->vm_next);
589 mm->highest_vm_end = vma->vm_end;
592 * vma->vm_prev wasn't known when we followed the rbtree to find the
593 * correct insertion point for that vma. As a result, we could not
594 * update the vma vm_rb parents rb_subtree_gap values on the way down.
595 * So, we first insert the vma with a zero rb_subtree_gap value
596 * (to be consistent with what we did on the way down), and then
597 * immediately update the gap to the correct value. Finally we
598 * rebalance the rbtree after all augmented values have been set.
600 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
601 vma->rb_subtree_gap = 0;
603 vma_rb_insert(vma, &mm->mm_rb);
606 static void __vma_link_file(struct vm_area_struct *vma)
612 struct address_space *mapping = file->f_mapping;
614 if (vma->vm_flags & VM_DENYWRITE)
615 atomic_dec(&file_inode(file)->i_writecount);
616 if (vma->vm_flags & VM_SHARED)
617 mapping->i_mmap_writable++;
619 flush_dcache_mmap_lock(mapping);
620 if (unlikely(vma->vm_flags & VM_NONLINEAR))
621 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
623 vma_interval_tree_insert(vma, &mapping->i_mmap);
624 flush_dcache_mmap_unlock(mapping);
629 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
630 struct vm_area_struct *prev, struct rb_node **rb_link,
631 struct rb_node *rb_parent)
633 __vma_link_list(mm, vma, prev, rb_parent);
634 __vma_link_rb(mm, vma, rb_link, rb_parent);
637 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
638 struct vm_area_struct *prev, struct rb_node **rb_link,
639 struct rb_node *rb_parent)
641 struct address_space *mapping = NULL;
644 mapping = vma->vm_file->f_mapping;
647 mutex_lock(&mapping->i_mmap_mutex);
649 __vma_link(mm, vma, prev, rb_link, rb_parent);
650 __vma_link_file(vma);
653 mutex_unlock(&mapping->i_mmap_mutex);
660 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
661 * mm's list and rbtree. It has already been inserted into the interval tree.
663 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
665 struct vm_area_struct *prev;
666 struct rb_node **rb_link, *rb_parent;
668 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
669 &prev, &rb_link, &rb_parent))
671 __vma_link(mm, vma, prev, rb_link, rb_parent);
676 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
677 struct vm_area_struct *prev)
679 struct vm_area_struct *next;
681 vma_rb_erase(vma, &mm->mm_rb);
682 prev->vm_next = next = vma->vm_next;
684 next->vm_prev = prev;
685 if (mm->mmap_cache == vma)
686 mm->mmap_cache = prev;
690 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
691 * is already present in an i_mmap tree without adjusting the tree.
692 * The following helper function should be used when such adjustments
693 * are necessary. The "insert" vma (if any) is to be inserted
694 * before we drop the necessary locks.
696 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
697 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
699 struct mm_struct *mm = vma->vm_mm;
700 struct vm_area_struct *next = vma->vm_next;
701 struct vm_area_struct *importer = NULL;
702 struct address_space *mapping = NULL;
703 struct rb_root *root = NULL;
704 struct anon_vma *anon_vma = NULL;
705 struct file *file = vma->vm_file;
706 bool start_changed = false, end_changed = false;
707 long adjust_next = 0;
710 if (next && !insert) {
711 struct vm_area_struct *exporter = NULL;
713 if (end >= next->vm_end) {
715 * vma expands, overlapping all the next, and
716 * perhaps the one after too (mprotect case 6).
718 again: remove_next = 1 + (end > next->vm_end);
722 } else if (end > next->vm_start) {
724 * vma expands, overlapping part of the next:
725 * mprotect case 5 shifting the boundary up.
727 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
730 } else if (end < vma->vm_end) {
732 * vma shrinks, and !insert tells it's not
733 * split_vma inserting another: so it must be
734 * mprotect case 4 shifting the boundary down.
736 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
742 * Easily overlooked: when mprotect shifts the boundary,
743 * make sure the expanding vma has anon_vma set if the
744 * shrinking vma had, to cover any anon pages imported.
746 if (exporter && exporter->anon_vma && !importer->anon_vma) {
747 if (anon_vma_clone(importer, exporter))
749 importer->anon_vma = exporter->anon_vma;
754 mapping = file->f_mapping;
755 if (!(vma->vm_flags & VM_NONLINEAR)) {
756 root = &mapping->i_mmap;
757 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
760 uprobe_munmap(next, next->vm_start,
764 mutex_lock(&mapping->i_mmap_mutex);
767 * Put into interval tree now, so instantiated pages
768 * are visible to arm/parisc __flush_dcache_page
769 * throughout; but we cannot insert into address
770 * space until vma start or end is updated.
772 __vma_link_file(insert);
776 vma_adjust_trans_huge(vma, start, end, adjust_next);
778 anon_vma = vma->anon_vma;
779 if (!anon_vma && adjust_next)
780 anon_vma = next->anon_vma;
782 VM_BUG_ON(adjust_next && next->anon_vma &&
783 anon_vma != next->anon_vma);
784 anon_vma_lock_write(anon_vma);
785 anon_vma_interval_tree_pre_update_vma(vma);
787 anon_vma_interval_tree_pre_update_vma(next);
791 flush_dcache_mmap_lock(mapping);
792 vma_interval_tree_remove(vma, root);
794 vma_interval_tree_remove(next, root);
797 if (start != vma->vm_start) {
798 vma->vm_start = start;
799 start_changed = true;
801 if (end != vma->vm_end) {
805 vma->vm_pgoff = pgoff;
807 next->vm_start += adjust_next << PAGE_SHIFT;
808 next->vm_pgoff += adjust_next;
813 vma_interval_tree_insert(next, root);
814 vma_interval_tree_insert(vma, root);
815 flush_dcache_mmap_unlock(mapping);
820 * vma_merge has merged next into vma, and needs
821 * us to remove next before dropping the locks.
823 __vma_unlink(mm, next, vma);
825 __remove_shared_vm_struct(next, file, mapping);
828 * split_vma has split insert from vma, and needs
829 * us to insert it before dropping the locks
830 * (it may either follow vma or precede it).
832 __insert_vm_struct(mm, insert);
838 mm->highest_vm_end = end;
839 else if (!adjust_next)
840 vma_gap_update(next);
845 anon_vma_interval_tree_post_update_vma(vma);
847 anon_vma_interval_tree_post_update_vma(next);
848 anon_vma_unlock_write(anon_vma);
851 mutex_unlock(&mapping->i_mmap_mutex);
862 uprobe_munmap(next, next->vm_start, next->vm_end);
866 anon_vma_merge(vma, next);
868 mpol_put(vma_policy(next));
869 kmem_cache_free(vm_area_cachep, next);
871 * In mprotect's case 6 (see comments on vma_merge),
872 * we must remove another next too. It would clutter
873 * up the code too much to do both in one go.
876 if (remove_next == 2)
879 vma_gap_update(next);
881 mm->highest_vm_end = end;
892 * If the vma has a ->close operation then the driver probably needs to release
893 * per-vma resources, so we don't attempt to merge those.
895 static inline int is_mergeable_vma(struct vm_area_struct *vma,
896 struct file *file, unsigned long vm_flags)
898 if (vma->vm_flags ^ vm_flags)
900 if (vma->vm_file != file)
902 if (vma->vm_ops && vma->vm_ops->close)
907 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
908 struct anon_vma *anon_vma2,
909 struct vm_area_struct *vma)
912 * The list_is_singular() test is to avoid merging VMA cloned from
913 * parents. This can improve scalability caused by anon_vma lock.
915 if ((!anon_vma1 || !anon_vma2) && (!vma ||
916 list_is_singular(&vma->anon_vma_chain)))
918 return anon_vma1 == anon_vma2;
922 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
923 * in front of (at a lower virtual address and file offset than) the vma.
925 * We cannot merge two vmas if they have differently assigned (non-NULL)
926 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
928 * We don't check here for the merged mmap wrapping around the end of pagecache
929 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
930 * wrap, nor mmaps which cover the final page at index -1UL.
933 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
934 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
936 if (is_mergeable_vma(vma, file, vm_flags) &&
937 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
938 if (vma->vm_pgoff == vm_pgoff)
945 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
946 * beyond (at a higher virtual address and file offset than) the vma.
948 * We cannot merge two vmas if they have differently assigned (non-NULL)
949 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
952 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
953 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
955 if (is_mergeable_vma(vma, file, vm_flags) &&
956 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
958 vm_pglen = vma_pages(vma);
959 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
966 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
967 * whether that can be merged with its predecessor or its successor.
968 * Or both (it neatly fills a hole).
970 * In most cases - when called for mmap, brk or mremap - [addr,end) is
971 * certain not to be mapped by the time vma_merge is called; but when
972 * called for mprotect, it is certain to be already mapped (either at
973 * an offset within prev, or at the start of next), and the flags of
974 * this area are about to be changed to vm_flags - and the no-change
975 * case has already been eliminated.
977 * The following mprotect cases have to be considered, where AAAA is
978 * the area passed down from mprotect_fixup, never extending beyond one
979 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
981 * AAAA AAAA AAAA AAAA
982 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
983 * cannot merge might become might become might become
984 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
985 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
986 * mremap move: PPPPNNNNNNNN 8
988 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
989 * might become case 1 below case 2 below case 3 below
991 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
992 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
994 struct vm_area_struct *vma_merge(struct mm_struct *mm,
995 struct vm_area_struct *prev, unsigned long addr,
996 unsigned long end, unsigned long vm_flags,
997 struct anon_vma *anon_vma, struct file *file,
998 pgoff_t pgoff, struct mempolicy *policy)
1000 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1001 struct vm_area_struct *area, *next;
1005 * We later require that vma->vm_flags == vm_flags,
1006 * so this tests vma->vm_flags & VM_SPECIAL, too.
1008 if (vm_flags & VM_SPECIAL)
1012 next = prev->vm_next;
1016 if (next && next->vm_end == end) /* cases 6, 7, 8 */
1017 next = next->vm_next;
1020 * Can it merge with the predecessor?
1022 if (prev && prev->vm_end == addr &&
1023 mpol_equal(vma_policy(prev), policy) &&
1024 can_vma_merge_after(prev, vm_flags,
1025 anon_vma, file, pgoff)) {
1027 * OK, it can. Can we now merge in the successor as well?
1029 if (next && end == next->vm_start &&
1030 mpol_equal(policy, vma_policy(next)) &&
1031 can_vma_merge_before(next, vm_flags,
1032 anon_vma, file, pgoff+pglen) &&
1033 is_mergeable_anon_vma(prev->anon_vma,
1034 next->anon_vma, NULL)) {
1036 err = vma_adjust(prev, prev->vm_start,
1037 next->vm_end, prev->vm_pgoff, NULL);
1038 } else /* cases 2, 5, 7 */
1039 err = vma_adjust(prev, prev->vm_start,
1040 end, prev->vm_pgoff, NULL);
1043 khugepaged_enter_vma_merge(prev);
1048 * Can this new request be merged in front of next?
1050 if (next && end == next->vm_start &&
1051 mpol_equal(policy, vma_policy(next)) &&
1052 can_vma_merge_before(next, vm_flags,
1053 anon_vma, file, pgoff+pglen)) {
1054 if (prev && addr < prev->vm_end) /* case 4 */
1055 err = vma_adjust(prev, prev->vm_start,
1056 addr, prev->vm_pgoff, NULL);
1057 else /* cases 3, 8 */
1058 err = vma_adjust(area, addr, next->vm_end,
1059 next->vm_pgoff - pglen, NULL);
1062 khugepaged_enter_vma_merge(area);
1070 * Rough compatbility check to quickly see if it's even worth looking
1071 * at sharing an anon_vma.
1073 * They need to have the same vm_file, and the flags can only differ
1074 * in things that mprotect may change.
1076 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1077 * we can merge the two vma's. For example, we refuse to merge a vma if
1078 * there is a vm_ops->close() function, because that indicates that the
1079 * driver is doing some kind of reference counting. But that doesn't
1080 * really matter for the anon_vma sharing case.
1082 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1084 return a->vm_end == b->vm_start &&
1085 mpol_equal(vma_policy(a), vma_policy(b)) &&
1086 a->vm_file == b->vm_file &&
1087 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
1088 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1092 * Do some basic sanity checking to see if we can re-use the anon_vma
1093 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1094 * the same as 'old', the other will be the new one that is trying
1095 * to share the anon_vma.
1097 * NOTE! This runs with mm_sem held for reading, so it is possible that
1098 * the anon_vma of 'old' is concurrently in the process of being set up
1099 * by another page fault trying to merge _that_. But that's ok: if it
1100 * is being set up, that automatically means that it will be a singleton
1101 * acceptable for merging, so we can do all of this optimistically. But
1102 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
1104 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1105 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1106 * is to return an anon_vma that is "complex" due to having gone through
1109 * We also make sure that the two vma's are compatible (adjacent,
1110 * and with the same memory policies). That's all stable, even with just
1111 * a read lock on the mm_sem.
1113 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1115 if (anon_vma_compatible(a, b)) {
1116 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
1118 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1125 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1126 * neighbouring vmas for a suitable anon_vma, before it goes off
1127 * to allocate a new anon_vma. It checks because a repetitive
1128 * sequence of mprotects and faults may otherwise lead to distinct
1129 * anon_vmas being allocated, preventing vma merge in subsequent
1132 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1134 struct anon_vma *anon_vma;
1135 struct vm_area_struct *near;
1137 near = vma->vm_next;
1141 anon_vma = reusable_anon_vma(near, vma, near);
1145 near = vma->vm_prev;
1149 anon_vma = reusable_anon_vma(near, near, vma);
1154 * There's no absolute need to look only at touching neighbours:
1155 * we could search further afield for "compatible" anon_vmas.
1156 * But it would probably just be a waste of time searching,
1157 * or lead to too many vmas hanging off the same anon_vma.
1158 * We're trying to allow mprotect remerging later on,
1159 * not trying to minimize memory used for anon_vmas.
1164 #ifdef CONFIG_PROC_FS
1165 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
1166 struct file *file, long pages)
1168 const unsigned long stack_flags
1169 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
1171 mm->total_vm += pages;
1174 mm->shared_vm += pages;
1175 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
1176 mm->exec_vm += pages;
1177 } else if (flags & stack_flags)
1178 mm->stack_vm += pages;
1180 #endif /* CONFIG_PROC_FS */
1183 * If a hint addr is less than mmap_min_addr change hint to be as
1184 * low as possible but still greater than mmap_min_addr
1186 static inline unsigned long round_hint_to_min(unsigned long hint)
1189 if (((void *)hint != NULL) &&
1190 (hint < mmap_min_addr))
1191 return PAGE_ALIGN(mmap_min_addr);
1196 * The caller must hold down_write(¤t->mm->mmap_sem).
1199 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
1200 unsigned long len, unsigned long prot,
1201 unsigned long flags, unsigned long pgoff,
1202 unsigned long *populate)
1204 struct mm_struct * mm = current->mm;
1205 struct inode *inode;
1206 vm_flags_t vm_flags;
1211 * Does the application expect PROT_READ to imply PROT_EXEC?
1213 * (the exception is when the underlying filesystem is noexec
1214 * mounted, in which case we dont add PROT_EXEC.)
1216 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1217 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
1223 if (!(flags & MAP_FIXED))
1224 addr = round_hint_to_min(addr);
1226 /* Careful about overflows.. */
1227 len = PAGE_ALIGN(len);
1231 /* offset overflow? */
1232 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1235 /* Too many mappings? */
1236 if (mm->map_count > sysctl_max_map_count)
1239 /* Obtain the address to map to. we verify (or select) it and ensure
1240 * that it represents a valid section of the address space.
1242 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1243 if (addr & ~PAGE_MASK)
1246 /* Do simple checking here so the lower-level routines won't have
1247 * to. we assume access permissions have been handled by the open
1248 * of the memory object, so we don't do any here.
1250 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1251 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1253 if (flags & MAP_LOCKED)
1254 if (!can_do_mlock())
1257 /* mlock MCL_FUTURE? */
1258 if (vm_flags & VM_LOCKED) {
1259 unsigned long locked, lock_limit;
1260 locked = len >> PAGE_SHIFT;
1261 locked += mm->locked_vm;
1262 lock_limit = rlimit(RLIMIT_MEMLOCK);
1263 lock_limit >>= PAGE_SHIFT;
1264 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1268 inode = file ? file_inode(file) : NULL;
1271 switch (flags & MAP_TYPE) {
1273 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1277 * Make sure we don't allow writing to an append-only
1280 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1284 * Make sure there are no mandatory locks on the file.
1286 if (locks_verify_locked(inode))
1289 vm_flags |= VM_SHARED | VM_MAYSHARE;
1290 if (!(file->f_mode & FMODE_WRITE))
1291 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1295 if (!(file->f_mode & FMODE_READ))
1297 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1298 if (vm_flags & VM_EXEC)
1300 vm_flags &= ~VM_MAYEXEC;
1303 if (!file->f_op || !file->f_op->mmap)
1305 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1313 switch (flags & MAP_TYPE) {
1315 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1321 vm_flags |= VM_SHARED | VM_MAYSHARE;
1325 * Set pgoff according to addr for anon_vma.
1327 pgoff = addr >> PAGE_SHIFT;
1335 * Set 'VM_NORESERVE' if we should not account for the
1336 * memory use of this mapping.
1338 if (flags & MAP_NORESERVE) {
1339 /* We honor MAP_NORESERVE if allowed to overcommit */
1340 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1341 vm_flags |= VM_NORESERVE;
1343 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1344 if (file && is_file_hugepages(file))
1345 vm_flags |= VM_NORESERVE;
1348 addr = mmap_region(file, addr, len, vm_flags, pgoff);
1349 if (!IS_ERR_VALUE(addr) &&
1350 ((vm_flags & VM_LOCKED) ||
1351 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1356 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1357 unsigned long, prot, unsigned long, flags,
1358 unsigned long, fd, unsigned long, pgoff)
1360 struct file *file = NULL;
1361 unsigned long retval = -EBADF;
1363 if (!(flags & MAP_ANONYMOUS)) {
1364 audit_mmap_fd(fd, flags);
1368 if (is_file_hugepages(file))
1369 len = ALIGN(len, huge_page_size(hstate_file(file)));
1371 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1373 } else if (flags & MAP_HUGETLB) {
1374 struct user_struct *user = NULL;
1377 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1381 len = ALIGN(len, huge_page_size(hs));
1383 * VM_NORESERVE is used because the reservations will be
1384 * taken when vm_ops->mmap() is called
1385 * A dummy user value is used because we are not locking
1386 * memory so no accounting is necessary
1388 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1390 &user, HUGETLB_ANONHUGE_INODE,
1391 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1393 return PTR_ERR(file);
1396 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1398 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1406 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1407 struct mmap_arg_struct {
1411 unsigned long flags;
1413 unsigned long offset;
1416 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1418 struct mmap_arg_struct a;
1420 if (copy_from_user(&a, arg, sizeof(a)))
1422 if (a.offset & ~PAGE_MASK)
1425 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1426 a.offset >> PAGE_SHIFT);
1428 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1431 * Some shared mappigns will want the pages marked read-only
1432 * to track write events. If so, we'll downgrade vm_page_prot
1433 * to the private version (using protection_map[] without the
1436 int vma_wants_writenotify(struct vm_area_struct *vma)
1438 vm_flags_t vm_flags = vma->vm_flags;
1440 /* If it was private or non-writable, the write bit is already clear */
1441 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1444 /* The backer wishes to know when pages are first written to? */
1445 if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1448 /* The open routine did something to the protections already? */
1449 if (pgprot_val(vma->vm_page_prot) !=
1450 pgprot_val(vm_get_page_prot(vm_flags)))
1453 /* Specialty mapping? */
1454 if (vm_flags & VM_PFNMAP)
1457 /* Can the mapping track the dirty pages? */
1458 return vma->vm_file && vma->vm_file->f_mapping &&
1459 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1463 * We account for memory if it's a private writeable mapping,
1464 * not hugepages and VM_NORESERVE wasn't set.
1466 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1469 * hugetlb has its own accounting separate from the core VM
1470 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1472 if (file && is_file_hugepages(file))
1475 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1478 unsigned long mmap_region(struct file *file, unsigned long addr,
1479 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1481 struct mm_struct *mm = current->mm;
1482 struct vm_area_struct *vma, *prev;
1483 int correct_wcount = 0;
1485 struct rb_node **rb_link, *rb_parent;
1486 unsigned long charged = 0;
1487 struct inode *inode = file ? file_inode(file) : NULL;
1489 /* Check against address space limit. */
1490 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) {
1491 unsigned long nr_pages;
1494 * MAP_FIXED may remove pages of mappings that intersects with
1495 * requested mapping. Account for the pages it would unmap.
1497 if (!(vm_flags & MAP_FIXED))
1500 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1502 if (!may_expand_vm(mm, (len >> PAGE_SHIFT) - nr_pages))
1506 /* Clear old maps */
1509 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
1510 if (do_munmap(mm, addr, len))
1516 * Private writable mapping: check memory availability
1518 if (accountable_mapping(file, vm_flags)) {
1519 charged = len >> PAGE_SHIFT;
1520 if (security_vm_enough_memory_mm(mm, charged))
1522 vm_flags |= VM_ACCOUNT;
1526 * Can we just expand an old mapping?
1528 vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1533 * Determine the object being mapped and call the appropriate
1534 * specific mapper. the address has already been validated, but
1535 * not unmapped, but the maps are removed from the list.
1537 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1544 vma->vm_start = addr;
1545 vma->vm_end = addr + len;
1546 vma->vm_flags = vm_flags;
1547 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1548 vma->vm_pgoff = pgoff;
1549 INIT_LIST_HEAD(&vma->anon_vma_chain);
1552 if (vm_flags & VM_DENYWRITE) {
1553 error = deny_write_access(file);
1558 vma->vm_file = get_file(file);
1559 error = file->f_op->mmap(file, vma);
1561 goto unmap_and_free_vma;
1563 /* Can addr have changed??
1565 * Answer: Yes, several device drivers can do it in their
1566 * f_op->mmap method. -DaveM
1567 * Bug: If addr is changed, prev, rb_link, rb_parent should
1568 * be updated for vma_link()
1570 WARN_ON_ONCE(addr != vma->vm_start);
1572 addr = vma->vm_start;
1573 pgoff = vma->vm_pgoff;
1574 vm_flags = vma->vm_flags;
1575 } else if (vm_flags & VM_SHARED) {
1576 error = shmem_zero_setup(vma);
1581 if (vma_wants_writenotify(vma)) {
1582 pgprot_t pprot = vma->vm_page_prot;
1584 /* Can vma->vm_page_prot have changed??
1586 * Answer: Yes, drivers may have changed it in their
1587 * f_op->mmap method.
1589 * Ensures that vmas marked as uncached stay that way.
1591 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1592 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1593 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1596 vma_link(mm, vma, prev, rb_link, rb_parent);
1597 file = vma->vm_file;
1599 /* Once vma denies write, undo our temporary denial count */
1601 atomic_inc(&inode->i_writecount);
1603 perf_event_mmap(vma);
1605 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1606 if (vm_flags & VM_LOCKED) {
1607 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1608 vma == get_gate_vma(current->mm)))
1609 mm->locked_vm += (len >> PAGE_SHIFT);
1611 vma->vm_flags &= ~VM_LOCKED;
1621 atomic_inc(&inode->i_writecount);
1622 vma->vm_file = NULL;
1625 /* Undo any partial mapping done by a device driver. */
1626 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1629 kmem_cache_free(vm_area_cachep, vma);
1632 vm_unacct_memory(charged);
1636 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1639 * We implement the search by looking for an rbtree node that
1640 * immediately follows a suitable gap. That is,
1641 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1642 * - gap_end = vma->vm_start >= info->low_limit + length;
1643 * - gap_end - gap_start >= length
1646 struct mm_struct *mm = current->mm;
1647 struct vm_area_struct *vma;
1648 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1650 /* Adjust search length to account for worst case alignment overhead */
1651 length = info->length + info->align_mask;
1652 if (length < info->length)
1655 /* Adjust search limits by the desired length */
1656 if (info->high_limit < length)
1658 high_limit = info->high_limit - length;
1660 if (info->low_limit > high_limit)
1662 low_limit = info->low_limit + length;
1664 /* Check if rbtree root looks promising */
1665 if (RB_EMPTY_ROOT(&mm->mm_rb))
1667 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1668 if (vma->rb_subtree_gap < length)
1672 /* Visit left subtree if it looks promising */
1673 gap_end = vma->vm_start;
1674 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1675 struct vm_area_struct *left =
1676 rb_entry(vma->vm_rb.rb_left,
1677 struct vm_area_struct, vm_rb);
1678 if (left->rb_subtree_gap >= length) {
1684 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1686 /* Check if current node has a suitable gap */
1687 if (gap_start > high_limit)
1689 if (gap_end >= low_limit && gap_end - gap_start >= length)
1692 /* Visit right subtree if it looks promising */
1693 if (vma->vm_rb.rb_right) {
1694 struct vm_area_struct *right =
1695 rb_entry(vma->vm_rb.rb_right,
1696 struct vm_area_struct, vm_rb);
1697 if (right->rb_subtree_gap >= length) {
1703 /* Go back up the rbtree to find next candidate node */
1705 struct rb_node *prev = &vma->vm_rb;
1706 if (!rb_parent(prev))
1708 vma = rb_entry(rb_parent(prev),
1709 struct vm_area_struct, vm_rb);
1710 if (prev == vma->vm_rb.rb_left) {
1711 gap_start = vma->vm_prev->vm_end;
1712 gap_end = vma->vm_start;
1719 /* Check highest gap, which does not precede any rbtree node */
1720 gap_start = mm->highest_vm_end;
1721 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1722 if (gap_start > high_limit)
1726 /* We found a suitable gap. Clip it with the original low_limit. */
1727 if (gap_start < info->low_limit)
1728 gap_start = info->low_limit;
1730 /* Adjust gap address to the desired alignment */
1731 gap_start += (info->align_offset - gap_start) & info->align_mask;
1733 VM_BUG_ON(gap_start + info->length > info->high_limit);
1734 VM_BUG_ON(gap_start + info->length > gap_end);
1738 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1740 struct mm_struct *mm = current->mm;
1741 struct vm_area_struct *vma;
1742 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1744 /* Adjust search length to account for worst case alignment overhead */
1745 length = info->length + info->align_mask;
1746 if (length < info->length)
1750 * Adjust search limits by the desired length.
1751 * See implementation comment at top of unmapped_area().
1753 gap_end = info->high_limit;
1754 if (gap_end < length)
1756 high_limit = gap_end - length;
1758 if (info->low_limit > high_limit)
1760 low_limit = info->low_limit + length;
1762 /* Check highest gap, which does not precede any rbtree node */
1763 gap_start = mm->highest_vm_end;
1764 if (gap_start <= high_limit)
1767 /* Check if rbtree root looks promising */
1768 if (RB_EMPTY_ROOT(&mm->mm_rb))
1770 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1771 if (vma->rb_subtree_gap < length)
1775 /* Visit right subtree if it looks promising */
1776 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1777 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1778 struct vm_area_struct *right =
1779 rb_entry(vma->vm_rb.rb_right,
1780 struct vm_area_struct, vm_rb);
1781 if (right->rb_subtree_gap >= length) {
1788 /* Check if current node has a suitable gap */
1789 gap_end = vma->vm_start;
1790 if (gap_end < low_limit)
1792 if (gap_start <= high_limit && gap_end - gap_start >= length)
1795 /* Visit left subtree if it looks promising */
1796 if (vma->vm_rb.rb_left) {
1797 struct vm_area_struct *left =
1798 rb_entry(vma->vm_rb.rb_left,
1799 struct vm_area_struct, vm_rb);
1800 if (left->rb_subtree_gap >= length) {
1806 /* Go back up the rbtree to find next candidate node */
1808 struct rb_node *prev = &vma->vm_rb;
1809 if (!rb_parent(prev))
1811 vma = rb_entry(rb_parent(prev),
1812 struct vm_area_struct, vm_rb);
1813 if (prev == vma->vm_rb.rb_right) {
1814 gap_start = vma->vm_prev ?
1815 vma->vm_prev->vm_end : 0;
1822 /* We found a suitable gap. Clip it with the original high_limit. */
1823 if (gap_end > info->high_limit)
1824 gap_end = info->high_limit;
1827 /* Compute highest gap address at the desired alignment */
1828 gap_end -= info->length;
1829 gap_end -= (gap_end - info->align_offset) & info->align_mask;
1831 VM_BUG_ON(gap_end < info->low_limit);
1832 VM_BUG_ON(gap_end < gap_start);
1836 /* Get an address range which is currently unmapped.
1837 * For shmat() with addr=0.
1839 * Ugly calling convention alert:
1840 * Return value with the low bits set means error value,
1842 * if (ret & ~PAGE_MASK)
1845 * This function "knows" that -ENOMEM has the bits set.
1847 #ifndef HAVE_ARCH_UNMAPPED_AREA
1849 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1850 unsigned long len, unsigned long pgoff, unsigned long flags)
1852 struct mm_struct *mm = current->mm;
1853 struct vm_area_struct *vma;
1854 struct vm_unmapped_area_info info;
1856 if (len > TASK_SIZE)
1859 if (flags & MAP_FIXED)
1863 addr = PAGE_ALIGN(addr);
1864 vma = find_vma(mm, addr);
1865 if (TASK_SIZE - len >= addr &&
1866 (!vma || addr + len <= vma->vm_start))
1872 info.low_limit = TASK_UNMAPPED_BASE;
1873 info.high_limit = TASK_SIZE;
1874 info.align_mask = 0;
1875 return vm_unmapped_area(&info);
1880 * This mmap-allocator allocates new areas top-down from below the
1881 * stack's low limit (the base):
1883 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1885 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1886 const unsigned long len, const unsigned long pgoff,
1887 const unsigned long flags)
1889 struct vm_area_struct *vma;
1890 struct mm_struct *mm = current->mm;
1891 unsigned long addr = addr0;
1892 struct vm_unmapped_area_info info;
1894 /* requested length too big for entire address space */
1895 if (len > TASK_SIZE)
1898 if (flags & MAP_FIXED)
1901 /* requesting a specific address */
1903 addr = PAGE_ALIGN(addr);
1904 vma = find_vma(mm, addr);
1905 if (TASK_SIZE - len >= addr &&
1906 (!vma || addr + len <= vma->vm_start))
1910 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1912 info.low_limit = PAGE_SIZE;
1913 info.high_limit = mm->mmap_base;
1914 info.align_mask = 0;
1915 addr = vm_unmapped_area(&info);
1918 * A failed mmap() very likely causes application failure,
1919 * so fall back to the bottom-up function here. This scenario
1920 * can happen with large stack limits and large mmap()
1923 if (addr & ~PAGE_MASK) {
1924 VM_BUG_ON(addr != -ENOMEM);
1926 info.low_limit = TASK_UNMAPPED_BASE;
1927 info.high_limit = TASK_SIZE;
1928 addr = vm_unmapped_area(&info);
1936 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1937 unsigned long pgoff, unsigned long flags)
1939 unsigned long (*get_area)(struct file *, unsigned long,
1940 unsigned long, unsigned long, unsigned long);
1942 unsigned long error = arch_mmap_check(addr, len, flags);
1946 /* Careful about overflows.. */
1947 if (len > TASK_SIZE)
1950 get_area = current->mm->get_unmapped_area;
1951 if (file && file->f_op && file->f_op->get_unmapped_area)
1952 get_area = file->f_op->get_unmapped_area;
1953 addr = get_area(file, addr, len, pgoff, flags);
1954 if (IS_ERR_VALUE(addr))
1957 if (addr > TASK_SIZE - len)
1959 if (addr & ~PAGE_MASK)
1962 addr = arch_rebalance_pgtables(addr, len);
1963 error = security_mmap_addr(addr);
1964 return error ? error : addr;
1967 EXPORT_SYMBOL(get_unmapped_area);
1969 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1970 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1972 struct vm_area_struct *vma = NULL;
1974 /* Check the cache first. */
1975 /* (Cache hit rate is typically around 35%.) */
1976 vma = ACCESS_ONCE(mm->mmap_cache);
1977 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1978 struct rb_node *rb_node;
1980 rb_node = mm->mm_rb.rb_node;
1984 struct vm_area_struct *vma_tmp;
1986 vma_tmp = rb_entry(rb_node,
1987 struct vm_area_struct, vm_rb);
1989 if (vma_tmp->vm_end > addr) {
1991 if (vma_tmp->vm_start <= addr)
1993 rb_node = rb_node->rb_left;
1995 rb_node = rb_node->rb_right;
1998 mm->mmap_cache = vma;
2003 EXPORT_SYMBOL(find_vma);
2006 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2008 struct vm_area_struct *
2009 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2010 struct vm_area_struct **pprev)
2012 struct vm_area_struct *vma;
2014 vma = find_vma(mm, addr);
2016 *pprev = vma->vm_prev;
2018 struct rb_node *rb_node = mm->mm_rb.rb_node;
2021 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2022 rb_node = rb_node->rb_right;
2029 * Verify that the stack growth is acceptable and
2030 * update accounting. This is shared with both the
2031 * grow-up and grow-down cases.
2033 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
2035 struct mm_struct *mm = vma->vm_mm;
2036 struct rlimit *rlim = current->signal->rlim;
2037 unsigned long new_start;
2039 /* address space limit tests */
2040 if (!may_expand_vm(mm, grow))
2043 /* Stack limit test */
2044 if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2047 /* mlock limit tests */
2048 if (vma->vm_flags & VM_LOCKED) {
2049 unsigned long locked;
2050 unsigned long limit;
2051 locked = mm->locked_vm + grow;
2052 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2053 limit >>= PAGE_SHIFT;
2054 if (locked > limit && !capable(CAP_IPC_LOCK))
2058 /* Check to ensure the stack will not grow into a hugetlb-only region */
2059 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2061 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2065 * Overcommit.. This must be the final test, as it will
2066 * update security statistics.
2068 if (security_vm_enough_memory_mm(mm, grow))
2071 /* Ok, everything looks good - let it rip */
2072 if (vma->vm_flags & VM_LOCKED)
2073 mm->locked_vm += grow;
2074 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
2078 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2080 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2081 * vma is the last one with address > vma->vm_end. Have to extend vma.
2083 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2087 if (!(vma->vm_flags & VM_GROWSUP))
2091 * We must make sure the anon_vma is allocated
2092 * so that the anon_vma locking is not a noop.
2094 if (unlikely(anon_vma_prepare(vma)))
2096 vma_lock_anon_vma(vma);
2099 * vma->vm_start/vm_end cannot change under us because the caller
2100 * is required to hold the mmap_sem in read mode. We need the
2101 * anon_vma lock to serialize against concurrent expand_stacks.
2102 * Also guard against wrapping around to address 0.
2104 if (address < PAGE_ALIGN(address+4))
2105 address = PAGE_ALIGN(address+4);
2107 vma_unlock_anon_vma(vma);
2112 /* Somebody else might have raced and expanded it already */
2113 if (address > vma->vm_end) {
2114 unsigned long size, grow;
2116 size = address - vma->vm_start;
2117 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2120 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2121 error = acct_stack_growth(vma, size, grow);
2124 * vma_gap_update() doesn't support concurrent
2125 * updates, but we only hold a shared mmap_sem
2126 * lock here, so we need to protect against
2127 * concurrent vma expansions.
2128 * vma_lock_anon_vma() doesn't help here, as
2129 * we don't guarantee that all growable vmas
2130 * in a mm share the same root anon vma.
2131 * So, we reuse mm->page_table_lock to guard
2132 * against concurrent vma expansions.
2134 spin_lock(&vma->vm_mm->page_table_lock);
2135 anon_vma_interval_tree_pre_update_vma(vma);
2136 vma->vm_end = address;
2137 anon_vma_interval_tree_post_update_vma(vma);
2139 vma_gap_update(vma->vm_next);
2141 vma->vm_mm->highest_vm_end = address;
2142 spin_unlock(&vma->vm_mm->page_table_lock);
2144 perf_event_mmap(vma);
2148 vma_unlock_anon_vma(vma);
2149 khugepaged_enter_vma_merge(vma);
2150 validate_mm(vma->vm_mm);
2153 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2156 * vma is the first one with address < vma->vm_start. Have to extend vma.
2158 int expand_downwards(struct vm_area_struct *vma,
2159 unsigned long address)
2164 * We must make sure the anon_vma is allocated
2165 * so that the anon_vma locking is not a noop.
2167 if (unlikely(anon_vma_prepare(vma)))
2170 address &= PAGE_MASK;
2171 error = security_mmap_addr(address);
2175 vma_lock_anon_vma(vma);
2178 * vma->vm_start/vm_end cannot change under us because the caller
2179 * is required to hold the mmap_sem in read mode. We need the
2180 * anon_vma lock to serialize against concurrent expand_stacks.
2183 /* Somebody else might have raced and expanded it already */
2184 if (address < vma->vm_start) {
2185 unsigned long size, grow;
2187 size = vma->vm_end - address;
2188 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2191 if (grow <= vma->vm_pgoff) {
2192 error = acct_stack_growth(vma, size, grow);
2195 * vma_gap_update() doesn't support concurrent
2196 * updates, but we only hold a shared mmap_sem
2197 * lock here, so we need to protect against
2198 * concurrent vma expansions.
2199 * vma_lock_anon_vma() doesn't help here, as
2200 * we don't guarantee that all growable vmas
2201 * in a mm share the same root anon vma.
2202 * So, we reuse mm->page_table_lock to guard
2203 * against concurrent vma expansions.
2205 spin_lock(&vma->vm_mm->page_table_lock);
2206 anon_vma_interval_tree_pre_update_vma(vma);
2207 vma->vm_start = address;
2208 vma->vm_pgoff -= grow;
2209 anon_vma_interval_tree_post_update_vma(vma);
2210 vma_gap_update(vma);
2211 spin_unlock(&vma->vm_mm->page_table_lock);
2213 perf_event_mmap(vma);
2217 vma_unlock_anon_vma(vma);
2218 khugepaged_enter_vma_merge(vma);
2219 validate_mm(vma->vm_mm);
2224 * Note how expand_stack() refuses to expand the stack all the way to
2225 * abut the next virtual mapping, *unless* that mapping itself is also
2226 * a stack mapping. We want to leave room for a guard page, after all
2227 * (the guard page itself is not added here, that is done by the
2228 * actual page faulting logic)
2230 * This matches the behavior of the guard page logic (see mm/memory.c:
2231 * check_stack_guard_page()), which only allows the guard page to be
2232 * removed under these circumstances.
2234 #ifdef CONFIG_STACK_GROWSUP
2235 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2237 struct vm_area_struct *next;
2239 address &= PAGE_MASK;
2240 next = vma->vm_next;
2241 if (next && next->vm_start == address + PAGE_SIZE) {
2242 if (!(next->vm_flags & VM_GROWSUP))
2245 return expand_upwards(vma, address);
2248 struct vm_area_struct *
2249 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2251 struct vm_area_struct *vma, *prev;
2254 vma = find_vma_prev(mm, addr, &prev);
2255 if (vma && (vma->vm_start <= addr))
2257 if (!prev || expand_stack(prev, addr))
2259 if (prev->vm_flags & VM_LOCKED)
2260 __mlock_vma_pages_range(prev, addr, prev->vm_end, NULL);
2264 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2266 struct vm_area_struct *prev;
2268 address &= PAGE_MASK;
2269 prev = vma->vm_prev;
2270 if (prev && prev->vm_end == address) {
2271 if (!(prev->vm_flags & VM_GROWSDOWN))
2274 return expand_downwards(vma, address);
2277 struct vm_area_struct *
2278 find_extend_vma(struct mm_struct * mm, unsigned long addr)
2280 struct vm_area_struct * vma;
2281 unsigned long start;
2284 vma = find_vma(mm,addr);
2287 if (vma->vm_start <= addr)
2289 if (!(vma->vm_flags & VM_GROWSDOWN))
2291 start = vma->vm_start;
2292 if (expand_stack(vma, addr))
2294 if (vma->vm_flags & VM_LOCKED)
2295 __mlock_vma_pages_range(vma, addr, start, NULL);
2301 * Ok - we have the memory areas we should free on the vma list,
2302 * so release them, and do the vma updates.
2304 * Called with the mm semaphore held.
2306 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2308 unsigned long nr_accounted = 0;
2310 /* Update high watermark before we lower total_vm */
2311 update_hiwater_vm(mm);
2313 long nrpages = vma_pages(vma);
2315 if (vma->vm_flags & VM_ACCOUNT)
2316 nr_accounted += nrpages;
2317 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
2318 vma = remove_vma(vma);
2320 vm_unacct_memory(nr_accounted);
2325 * Get rid of page table information in the indicated region.
2327 * Called with the mm semaphore held.
2329 static void unmap_region(struct mm_struct *mm,
2330 struct vm_area_struct *vma, struct vm_area_struct *prev,
2331 unsigned long start, unsigned long end)
2333 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
2334 struct mmu_gather tlb;
2337 tlb_gather_mmu(&tlb, mm, start, end);
2338 update_hiwater_rss(mm);
2339 unmap_vmas(&tlb, vma, start, end);
2340 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2341 next ? next->vm_start : USER_PGTABLES_CEILING);
2342 tlb_finish_mmu(&tlb, start, end);
2346 * Create a list of vma's touched by the unmap, removing them from the mm's
2347 * vma list as we go..
2350 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2351 struct vm_area_struct *prev, unsigned long end)
2353 struct vm_area_struct **insertion_point;
2354 struct vm_area_struct *tail_vma = NULL;
2356 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2357 vma->vm_prev = NULL;
2359 vma_rb_erase(vma, &mm->mm_rb);
2363 } while (vma && vma->vm_start < end);
2364 *insertion_point = vma;
2366 vma->vm_prev = prev;
2367 vma_gap_update(vma);
2369 mm->highest_vm_end = prev ? prev->vm_end : 0;
2370 tail_vma->vm_next = NULL;
2371 mm->mmap_cache = NULL; /* Kill the cache. */
2375 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2376 * munmap path where it doesn't make sense to fail.
2378 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
2379 unsigned long addr, int new_below)
2381 struct vm_area_struct *new;
2384 if (is_vm_hugetlb_page(vma) && (addr &
2385 ~(huge_page_mask(hstate_vma(vma)))))
2388 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2392 /* most fields are the same, copy all, and then fixup */
2395 INIT_LIST_HEAD(&new->anon_vma_chain);
2400 new->vm_start = addr;
2401 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2404 err = vma_dup_policy(vma, new);
2408 if (anon_vma_clone(new, vma))
2412 get_file(new->vm_file);
2414 if (new->vm_ops && new->vm_ops->open)
2415 new->vm_ops->open(new);
2418 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2419 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2421 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2427 /* Clean everything up if vma_adjust failed. */
2428 if (new->vm_ops && new->vm_ops->close)
2429 new->vm_ops->close(new);
2432 unlink_anon_vmas(new);
2434 mpol_put(vma_policy(new));
2436 kmem_cache_free(vm_area_cachep, new);
2442 * Split a vma into two pieces at address 'addr', a new vma is allocated
2443 * either for the first part or the tail.
2445 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2446 unsigned long addr, int new_below)
2448 if (mm->map_count >= sysctl_max_map_count)
2451 return __split_vma(mm, vma, addr, new_below);
2454 /* Munmap is split into 2 main parts -- this part which finds
2455 * what needs doing, and the areas themselves, which do the
2456 * work. This now handles partial unmappings.
2457 * Jeremy Fitzhardinge <jeremy@goop.org>
2459 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2462 struct vm_area_struct *vma, *prev, *last;
2464 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2467 if ((len = PAGE_ALIGN(len)) == 0)
2470 /* Find the first overlapping VMA */
2471 vma = find_vma(mm, start);
2474 prev = vma->vm_prev;
2475 /* we have start < vma->vm_end */
2477 /* if it doesn't overlap, we have nothing.. */
2479 if (vma->vm_start >= end)
2483 * If we need to split any vma, do it now to save pain later.
2485 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2486 * unmapped vm_area_struct will remain in use: so lower split_vma
2487 * places tmp vma above, and higher split_vma places tmp vma below.
2489 if (start > vma->vm_start) {
2493 * Make sure that map_count on return from munmap() will
2494 * not exceed its limit; but let map_count go just above
2495 * its limit temporarily, to help free resources as expected.
2497 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2500 error = __split_vma(mm, vma, start, 0);
2506 /* Does it split the last one? */
2507 last = find_vma(mm, end);
2508 if (last && end > last->vm_start) {
2509 int error = __split_vma(mm, last, end, 1);
2513 vma = prev? prev->vm_next: mm->mmap;
2516 * unlock any mlock()ed ranges before detaching vmas
2518 if (mm->locked_vm) {
2519 struct vm_area_struct *tmp = vma;
2520 while (tmp && tmp->vm_start < end) {
2521 if (tmp->vm_flags & VM_LOCKED) {
2522 mm->locked_vm -= vma_pages(tmp);
2523 munlock_vma_pages_all(tmp);
2530 * Remove the vma's, and unmap the actual pages
2532 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2533 unmap_region(mm, vma, prev, start, end);
2535 /* Fix up all other VM information */
2536 remove_vma_list(mm, vma);
2541 int vm_munmap(unsigned long start, size_t len)
2544 struct mm_struct *mm = current->mm;
2546 down_write(&mm->mmap_sem);
2547 ret = do_munmap(mm, start, len);
2548 up_write(&mm->mmap_sem);
2551 EXPORT_SYMBOL(vm_munmap);
2553 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2555 profile_munmap(addr);
2556 return vm_munmap(addr, len);
2559 static inline void verify_mm_writelocked(struct mm_struct *mm)
2561 #ifdef CONFIG_DEBUG_VM
2562 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2564 up_read(&mm->mmap_sem);
2570 * this is really a simplified "do_mmap". it only handles
2571 * anonymous maps. eventually we may be able to do some
2572 * brk-specific accounting here.
2574 static unsigned long do_brk(unsigned long addr, unsigned long len)
2576 struct mm_struct * mm = current->mm;
2577 struct vm_area_struct * vma, * prev;
2578 unsigned long flags;
2579 struct rb_node ** rb_link, * rb_parent;
2580 pgoff_t pgoff = addr >> PAGE_SHIFT;
2583 len = PAGE_ALIGN(len);
2587 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2589 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2590 if (error & ~PAGE_MASK)
2596 if (mm->def_flags & VM_LOCKED) {
2597 unsigned long locked, lock_limit;
2598 locked = len >> PAGE_SHIFT;
2599 locked += mm->locked_vm;
2600 lock_limit = rlimit(RLIMIT_MEMLOCK);
2601 lock_limit >>= PAGE_SHIFT;
2602 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2607 * mm->mmap_sem is required to protect against another thread
2608 * changing the mappings in case we sleep.
2610 verify_mm_writelocked(mm);
2613 * Clear old maps. this also does some error checking for us
2616 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
2617 if (do_munmap(mm, addr, len))
2622 /* Check against address space limits *after* clearing old maps... */
2623 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2626 if (mm->map_count > sysctl_max_map_count)
2629 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2632 /* Can we just expand an old private anonymous mapping? */
2633 vma = vma_merge(mm, prev, addr, addr + len, flags,
2634 NULL, NULL, pgoff, NULL);
2639 * create a vma struct for an anonymous mapping
2641 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2643 vm_unacct_memory(len >> PAGE_SHIFT);
2647 INIT_LIST_HEAD(&vma->anon_vma_chain);
2649 vma->vm_start = addr;
2650 vma->vm_end = addr + len;
2651 vma->vm_pgoff = pgoff;
2652 vma->vm_flags = flags;
2653 vma->vm_page_prot = vm_get_page_prot(flags);
2654 vma_link(mm, vma, prev, rb_link, rb_parent);
2656 perf_event_mmap(vma);
2657 mm->total_vm += len >> PAGE_SHIFT;
2658 if (flags & VM_LOCKED)
2659 mm->locked_vm += (len >> PAGE_SHIFT);
2663 unsigned long vm_brk(unsigned long addr, unsigned long len)
2665 struct mm_struct *mm = current->mm;
2669 down_write(&mm->mmap_sem);
2670 ret = do_brk(addr, len);
2671 populate = ((mm->def_flags & VM_LOCKED) != 0);
2672 up_write(&mm->mmap_sem);
2674 mm_populate(addr, len);
2677 EXPORT_SYMBOL(vm_brk);
2679 /* Release all mmaps. */
2680 void exit_mmap(struct mm_struct *mm)
2682 struct mmu_gather tlb;
2683 struct vm_area_struct *vma;
2684 unsigned long nr_accounted = 0;
2686 /* mm's last user has gone, and its about to be pulled down */
2687 mmu_notifier_release(mm);
2689 if (mm->locked_vm) {
2692 if (vma->vm_flags & VM_LOCKED)
2693 munlock_vma_pages_all(vma);
2701 if (!vma) /* Can happen if dup_mmap() received an OOM */
2706 tlb_gather_mmu(&tlb, mm, 0, -1);
2707 /* update_hiwater_rss(mm) here? but nobody should be looking */
2708 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2709 unmap_vmas(&tlb, vma, 0, -1);
2711 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
2712 tlb_finish_mmu(&tlb, 0, -1);
2715 * Walk the list again, actually closing and freeing it,
2716 * with preemption enabled, without holding any MM locks.
2719 if (vma->vm_flags & VM_ACCOUNT)
2720 nr_accounted += vma_pages(vma);
2721 vma = remove_vma(vma);
2723 vm_unacct_memory(nr_accounted);
2725 WARN_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2728 /* Insert vm structure into process list sorted by address
2729 * and into the inode's i_mmap tree. If vm_file is non-NULL
2730 * then i_mmap_mutex is taken here.
2732 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2734 struct vm_area_struct *prev;
2735 struct rb_node **rb_link, *rb_parent;
2738 * The vm_pgoff of a purely anonymous vma should be irrelevant
2739 * until its first write fault, when page's anon_vma and index
2740 * are set. But now set the vm_pgoff it will almost certainly
2741 * end up with (unless mremap moves it elsewhere before that
2742 * first wfault), so /proc/pid/maps tells a consistent story.
2744 * By setting it to reflect the virtual start address of the
2745 * vma, merges and splits can happen in a seamless way, just
2746 * using the existing file pgoff checks and manipulations.
2747 * Similarly in do_mmap_pgoff and in do_brk.
2749 if (!vma->vm_file) {
2750 BUG_ON(vma->anon_vma);
2751 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2753 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2754 &prev, &rb_link, &rb_parent))
2756 if ((vma->vm_flags & VM_ACCOUNT) &&
2757 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2760 vma_link(mm, vma, prev, rb_link, rb_parent);
2765 * Copy the vma structure to a new location in the same mm,
2766 * prior to moving page table entries, to effect an mremap move.
2768 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2769 unsigned long addr, unsigned long len, pgoff_t pgoff,
2770 bool *need_rmap_locks)
2772 struct vm_area_struct *vma = *vmap;
2773 unsigned long vma_start = vma->vm_start;
2774 struct mm_struct *mm = vma->vm_mm;
2775 struct vm_area_struct *new_vma, *prev;
2776 struct rb_node **rb_link, *rb_parent;
2777 bool faulted_in_anon_vma = true;
2780 * If anonymous vma has not yet been faulted, update new pgoff
2781 * to match new location, to increase its chance of merging.
2783 if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2784 pgoff = addr >> PAGE_SHIFT;
2785 faulted_in_anon_vma = false;
2788 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2789 return NULL; /* should never get here */
2790 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2791 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2794 * Source vma may have been merged into new_vma
2796 if (unlikely(vma_start >= new_vma->vm_start &&
2797 vma_start < new_vma->vm_end)) {
2799 * The only way we can get a vma_merge with
2800 * self during an mremap is if the vma hasn't
2801 * been faulted in yet and we were allowed to
2802 * reset the dst vma->vm_pgoff to the
2803 * destination address of the mremap to allow
2804 * the merge to happen. mremap must change the
2805 * vm_pgoff linearity between src and dst vmas
2806 * (in turn preventing a vma_merge) to be
2807 * safe. It is only safe to keep the vm_pgoff
2808 * linear if there are no pages mapped yet.
2810 VM_BUG_ON(faulted_in_anon_vma);
2811 *vmap = vma = new_vma;
2813 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2815 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2818 new_vma->vm_start = addr;
2819 new_vma->vm_end = addr + len;
2820 new_vma->vm_pgoff = pgoff;
2821 if (vma_dup_policy(vma, new_vma))
2823 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2824 if (anon_vma_clone(new_vma, vma))
2825 goto out_free_mempol;
2826 if (new_vma->vm_file)
2827 get_file(new_vma->vm_file);
2828 if (new_vma->vm_ops && new_vma->vm_ops->open)
2829 new_vma->vm_ops->open(new_vma);
2830 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2831 *need_rmap_locks = false;
2837 mpol_put(vma_policy(new_vma));
2839 kmem_cache_free(vm_area_cachep, new_vma);
2844 * Return true if the calling process may expand its vm space by the passed
2847 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2849 unsigned long cur = mm->total_vm; /* pages */
2852 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2854 if (cur + npages > lim)
2860 static int special_mapping_fault(struct vm_area_struct *vma,
2861 struct vm_fault *vmf)
2864 struct page **pages;
2867 * special mappings have no vm_file, and in that case, the mm
2868 * uses vm_pgoff internally. So we have to subtract it from here.
2869 * We are allowed to do this because we are the mm; do not copy
2870 * this code into drivers!
2872 pgoff = vmf->pgoff - vma->vm_pgoff;
2874 for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2878 struct page *page = *pages;
2884 return VM_FAULT_SIGBUS;
2888 * Having a close hook prevents vma merging regardless of flags.
2890 static void special_mapping_close(struct vm_area_struct *vma)
2894 static const struct vm_operations_struct special_mapping_vmops = {
2895 .close = special_mapping_close,
2896 .fault = special_mapping_fault,
2900 * Called with mm->mmap_sem held for writing.
2901 * Insert a new vma covering the given region, with the given flags.
2902 * Its pages are supplied by the given array of struct page *.
2903 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2904 * The region past the last page supplied will always produce SIGBUS.
2905 * The array pointer and the pages it points to are assumed to stay alive
2906 * for as long as this mapping might exist.
2908 int install_special_mapping(struct mm_struct *mm,
2909 unsigned long addr, unsigned long len,
2910 unsigned long vm_flags, struct page **pages)
2913 struct vm_area_struct *vma;
2915 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2916 if (unlikely(vma == NULL))
2919 INIT_LIST_HEAD(&vma->anon_vma_chain);
2921 vma->vm_start = addr;
2922 vma->vm_end = addr + len;
2924 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2925 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2927 vma->vm_ops = &special_mapping_vmops;
2928 vma->vm_private_data = pages;
2930 ret = insert_vm_struct(mm, vma);
2934 mm->total_vm += len >> PAGE_SHIFT;
2936 perf_event_mmap(vma);
2941 kmem_cache_free(vm_area_cachep, vma);
2945 static DEFINE_MUTEX(mm_all_locks_mutex);
2947 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2949 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
2951 * The LSB of head.next can't change from under us
2952 * because we hold the mm_all_locks_mutex.
2954 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
2956 * We can safely modify head.next after taking the
2957 * anon_vma->root->rwsem. If some other vma in this mm shares
2958 * the same anon_vma we won't take it again.
2960 * No need of atomic instructions here, head.next
2961 * can't change from under us thanks to the
2962 * anon_vma->root->rwsem.
2964 if (__test_and_set_bit(0, (unsigned long *)
2965 &anon_vma->root->rb_root.rb_node))
2970 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2972 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2974 * AS_MM_ALL_LOCKS can't change from under us because
2975 * we hold the mm_all_locks_mutex.
2977 * Operations on ->flags have to be atomic because
2978 * even if AS_MM_ALL_LOCKS is stable thanks to the
2979 * mm_all_locks_mutex, there may be other cpus
2980 * changing other bitflags in parallel to us.
2982 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2984 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2989 * This operation locks against the VM for all pte/vma/mm related
2990 * operations that could ever happen on a certain mm. This includes
2991 * vmtruncate, try_to_unmap, and all page faults.
2993 * The caller must take the mmap_sem in write mode before calling
2994 * mm_take_all_locks(). The caller isn't allowed to release the
2995 * mmap_sem until mm_drop_all_locks() returns.
2997 * mmap_sem in write mode is required in order to block all operations
2998 * that could modify pagetables and free pages without need of
2999 * altering the vma layout (for example populate_range() with
3000 * nonlinear vmas). It's also needed in write mode to avoid new
3001 * anon_vmas to be associated with existing vmas.
3003 * A single task can't take more than one mm_take_all_locks() in a row
3004 * or it would deadlock.
3006 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3007 * mapping->flags avoid to take the same lock twice, if more than one
3008 * vma in this mm is backed by the same anon_vma or address_space.
3010 * We can take all the locks in random order because the VM code
3011 * taking i_mmap_mutex or anon_vma->rwsem outside the mmap_sem never
3012 * takes more than one of them in a row. Secondly we're protected
3013 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3015 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3016 * that may have to take thousand of locks.
3018 * mm_take_all_locks() can fail if it's interrupted by signals.
3020 int mm_take_all_locks(struct mm_struct *mm)
3022 struct vm_area_struct *vma;
3023 struct anon_vma_chain *avc;
3025 BUG_ON(down_read_trylock(&mm->mmap_sem));
3027 mutex_lock(&mm_all_locks_mutex);
3029 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3030 if (signal_pending(current))
3032 if (vma->vm_file && vma->vm_file->f_mapping)
3033 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3036 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3037 if (signal_pending(current))
3040 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3041 vm_lock_anon_vma(mm, avc->anon_vma);
3047 mm_drop_all_locks(mm);
3051 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3053 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3055 * The LSB of head.next can't change to 0 from under
3056 * us because we hold the mm_all_locks_mutex.
3058 * We must however clear the bitflag before unlocking
3059 * the vma so the users using the anon_vma->rb_root will
3060 * never see our bitflag.
3062 * No need of atomic instructions here, head.next
3063 * can't change from under us until we release the
3064 * anon_vma->root->rwsem.
3066 if (!__test_and_clear_bit(0, (unsigned long *)
3067 &anon_vma->root->rb_root.rb_node))
3069 anon_vma_unlock_write(anon_vma);
3073 static void vm_unlock_mapping(struct address_space *mapping)
3075 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3077 * AS_MM_ALL_LOCKS can't change to 0 from under us
3078 * because we hold the mm_all_locks_mutex.
3080 mutex_unlock(&mapping->i_mmap_mutex);
3081 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3088 * The mmap_sem cannot be released by the caller until
3089 * mm_drop_all_locks() returns.
3091 void mm_drop_all_locks(struct mm_struct *mm)
3093 struct vm_area_struct *vma;
3094 struct anon_vma_chain *avc;
3096 BUG_ON(down_read_trylock(&mm->mmap_sem));
3097 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3099 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3101 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3102 vm_unlock_anon_vma(avc->anon_vma);
3103 if (vma->vm_file && vma->vm_file->f_mapping)
3104 vm_unlock_mapping(vma->vm_file->f_mapping);
3107 mutex_unlock(&mm_all_locks_mutex);
3111 * initialise the VMA slab
3113 void __init mmap_init(void)
3117 ret = percpu_counter_init(&vm_committed_as, 0);
3122 * Initialise sysctl_user_reserve_kbytes.
3124 * This is intended to prevent a user from starting a single memory hogging
3125 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3128 * The default value is min(3% of free memory, 128MB)
3129 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3131 static int init_user_reserve(void)
3133 unsigned long free_kbytes;
3135 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3137 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3140 module_init(init_user_reserve)
3143 * Initialise sysctl_admin_reserve_kbytes.
3145 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3146 * to log in and kill a memory hogging process.
3148 * Systems with more than 256MB will reserve 8MB, enough to recover
3149 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3150 * only reserve 3% of free pages by default.
3152 static int init_admin_reserve(void)
3154 unsigned long free_kbytes;
3156 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3158 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3161 module_init(init_admin_reserve)
3164 * Reinititalise user and admin reserves if memory is added or removed.
3166 * The default user reserve max is 128MB, and the default max for the
3167 * admin reserve is 8MB. These are usually, but not always, enough to
3168 * enable recovery from a memory hogging process using login/sshd, a shell,
3169 * and tools like top. It may make sense to increase or even disable the
3170 * reserve depending on the existence of swap or variations in the recovery
3171 * tools. So, the admin may have changed them.
3173 * If memory is added and the reserves have been eliminated or increased above
3174 * the default max, then we'll trust the admin.
3176 * If memory is removed and there isn't enough free memory, then we
3177 * need to reset the reserves.
3179 * Otherwise keep the reserve set by the admin.
3181 static int reserve_mem_notifier(struct notifier_block *nb,
3182 unsigned long action, void *data)
3184 unsigned long tmp, free_kbytes;
3188 /* Default max is 128MB. Leave alone if modified by operator. */
3189 tmp = sysctl_user_reserve_kbytes;
3190 if (0 < tmp && tmp < (1UL << 17))
3191 init_user_reserve();
3193 /* Default max is 8MB. Leave alone if modified by operator. */
3194 tmp = sysctl_admin_reserve_kbytes;
3195 if (0 < tmp && tmp < (1UL << 13))
3196 init_admin_reserve();
3200 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3202 if (sysctl_user_reserve_kbytes > free_kbytes) {
3203 init_user_reserve();
3204 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3205 sysctl_user_reserve_kbytes);
3208 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3209 init_admin_reserve();
3210 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3211 sysctl_admin_reserve_kbytes);
3220 static struct notifier_block reserve_mem_nb = {
3221 .notifier_call = reserve_mem_notifier,
3224 static int __meminit init_reserve_notifier(void)
3226 if (register_hotmemory_notifier(&reserve_mem_nb))
3227 printk("Failed registering memory add/remove notifier for admin reserve");
3231 module_init(init_reserve_notifier)