6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/kernel.h>
12 #include <linux/slab.h>
13 #include <linux/backing-dev.h>
15 #include <linux/vmacache.h>
16 #include <linux/shm.h>
17 #include <linux/mman.h>
18 #include <linux/pagemap.h>
19 #include <linux/swap.h>
20 #include <linux/syscalls.h>
21 #include <linux/capability.h>
22 #include <linux/init.h>
23 #include <linux/file.h>
25 #include <linux/personality.h>
26 #include <linux/security.h>
27 #include <linux/hugetlb.h>
28 #include <linux/profile.h>
29 #include <linux/export.h>
30 #include <linux/mount.h>
31 #include <linux/mempolicy.h>
32 #include <linux/rmap.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/mmdebug.h>
35 #include <linux/perf_event.h>
36 #include <linux/audit.h>
37 #include <linux/khugepaged.h>
38 #include <linux/uprobes.h>
39 #include <linux/rbtree_augmented.h>
40 #include <linux/sched/sysctl.h>
41 #include <linux/notifier.h>
42 #include <linux/memory.h>
43 #include <linux/printk.h>
44 #include <linux/userfaultfd_k.h>
46 #include <asm/uaccess.h>
47 #include <asm/cacheflush.h>
49 #include <asm/mmu_context.h>
53 #ifndef arch_mmap_check
54 #define arch_mmap_check(addr, len, flags) (0)
57 #ifndef arch_rebalance_pgtables
58 #define arch_rebalance_pgtables(addr, len) (addr)
61 static void unmap_region(struct mm_struct *mm,
62 struct vm_area_struct *vma, struct vm_area_struct *prev,
63 unsigned long start, unsigned long end);
65 /* description of effects of mapping type and prot in current implementation.
66 * this is due to the limited x86 page protection hardware. The expected
67 * behavior is in parens:
70 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
71 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
72 * w: (no) no w: (no) no w: (yes) yes w: (no) no
73 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
75 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
76 * w: (no) no w: (no) no w: (copy) copy w: (no) no
77 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
80 pgprot_t protection_map[16] = {
81 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
82 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
85 pgprot_t vm_get_page_prot(unsigned long vm_flags)
87 return __pgprot(pgprot_val(protection_map[vm_flags &
88 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
89 pgprot_val(arch_vm_get_page_prot(vm_flags)));
91 EXPORT_SYMBOL(vm_get_page_prot);
93 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
95 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
98 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
99 void vma_set_page_prot(struct vm_area_struct *vma)
101 unsigned long vm_flags = vma->vm_flags;
103 vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
104 if (vma_wants_writenotify(vma)) {
105 vm_flags &= ~VM_SHARED;
106 vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot,
112 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS; /* heuristic overcommit */
113 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
114 unsigned long sysctl_overcommit_kbytes __read_mostly;
115 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
116 unsigned long sysctl_user_reserve_kbytes __read_mostly = 1UL << 17; /* 128MB */
117 unsigned long sysctl_admin_reserve_kbytes __read_mostly = 1UL << 13; /* 8MB */
119 * Make sure vm_committed_as in one cacheline and not cacheline shared with
120 * other variables. It can be updated by several CPUs frequently.
122 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
125 * The global memory commitment made in the system can be a metric
126 * that can be used to drive ballooning decisions when Linux is hosted
127 * as a guest. On Hyper-V, the host implements a policy engine for dynamically
128 * balancing memory across competing virtual machines that are hosted.
129 * Several metrics drive this policy engine including the guest reported
132 unsigned long vm_memory_committed(void)
134 return percpu_counter_read_positive(&vm_committed_as);
136 EXPORT_SYMBOL_GPL(vm_memory_committed);
139 * Check that a process has enough memory to allocate a new virtual
140 * mapping. 0 means there is enough memory for the allocation to
141 * succeed and -ENOMEM implies there is not.
143 * We currently support three overcommit policies, which are set via the
144 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
146 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
147 * Additional code 2002 Jul 20 by Robert Love.
149 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
151 * Note this is a helper function intended to be used by LSMs which
152 * wish to use this logic.
154 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
156 long free, allowed, reserve;
158 VM_WARN_ONCE(percpu_counter_read(&vm_committed_as) <
159 -(s64)vm_committed_as_batch * num_online_cpus(),
160 "memory commitment underflow");
162 vm_acct_memory(pages);
165 * Sometimes we want to use more memory than we have
167 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
170 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
171 free = global_page_state(NR_FREE_PAGES);
172 free += global_page_state(NR_FILE_PAGES);
175 * shmem pages shouldn't be counted as free in this
176 * case, they can't be purged, only swapped out, and
177 * that won't affect the overall amount of available
178 * memory in the system.
180 free -= global_page_state(NR_SHMEM);
182 free += get_nr_swap_pages();
185 * Any slabs which are created with the
186 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
187 * which are reclaimable, under pressure. The dentry
188 * cache and most inode caches should fall into this
190 free += global_page_state(NR_SLAB_RECLAIMABLE);
193 * Leave reserved pages. The pages are not for anonymous pages.
195 if (free <= totalreserve_pages)
198 free -= totalreserve_pages;
201 * Reserve some for root
204 free -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
212 allowed = vm_commit_limit();
214 * Reserve some for root
217 allowed -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
220 * Don't let a single process grow so big a user can't recover
223 reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10);
224 allowed -= min_t(long, mm->total_vm / 32, reserve);
227 if (percpu_counter_read_positive(&vm_committed_as) < allowed)
230 vm_unacct_memory(pages);
236 * Requires inode->i_mapping->i_mmap_rwsem
238 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
239 struct file *file, struct address_space *mapping)
241 if (vma->vm_flags & VM_DENYWRITE)
242 atomic_inc(&file_inode(file)->i_writecount);
243 if (vma->vm_flags & VM_SHARED)
244 mapping_unmap_writable(mapping);
246 flush_dcache_mmap_lock(mapping);
247 vma_interval_tree_remove(vma, &mapping->i_mmap);
248 flush_dcache_mmap_unlock(mapping);
252 * Unlink a file-based vm structure from its interval tree, to hide
253 * vma from rmap and vmtruncate before freeing its page tables.
255 void unlink_file_vma(struct vm_area_struct *vma)
257 struct file *file = vma->vm_file;
260 struct address_space *mapping = file->f_mapping;
261 i_mmap_lock_write(mapping);
262 __remove_shared_vm_struct(vma, file, mapping);
263 i_mmap_unlock_write(mapping);
268 * Close a vm structure and free it, returning the next.
270 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
272 struct vm_area_struct *next = vma->vm_next;
275 if (vma->vm_ops && vma->vm_ops->close)
276 vma->vm_ops->close(vma);
279 mpol_put(vma_policy(vma));
280 kmem_cache_free(vm_area_cachep, vma);
284 static unsigned long do_brk(unsigned long addr, unsigned long len);
286 SYSCALL_DEFINE1(brk, unsigned long, brk)
288 unsigned long retval;
289 unsigned long newbrk, oldbrk;
290 struct mm_struct *mm = current->mm;
291 struct vm_area_struct *next;
292 unsigned long min_brk;
295 down_write(&mm->mmap_sem);
297 #ifdef CONFIG_COMPAT_BRK
299 * CONFIG_COMPAT_BRK can still be overridden by setting
300 * randomize_va_space to 2, which will still cause mm->start_brk
301 * to be arbitrarily shifted
303 if (current->brk_randomized)
304 min_brk = mm->start_brk;
306 min_brk = mm->end_data;
308 min_brk = mm->start_brk;
314 * Check against rlimit here. If this check is done later after the test
315 * of oldbrk with newbrk then it can escape the test and let the data
316 * segment grow beyond its set limit the in case where the limit is
317 * not page aligned -Ram Gupta
319 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
320 mm->end_data, mm->start_data))
323 newbrk = PAGE_ALIGN(brk);
324 oldbrk = PAGE_ALIGN(mm->brk);
325 if (oldbrk == newbrk)
328 /* Always allow shrinking brk. */
329 if (brk <= mm->brk) {
330 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
335 /* Check against existing mmap mappings. */
336 next = find_vma(mm, oldbrk);
337 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
340 /* Ok, looks good - let it rip. */
341 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
346 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
347 up_write(&mm->mmap_sem);
349 mm_populate(oldbrk, newbrk - oldbrk);
354 up_write(&mm->mmap_sem);
358 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
360 unsigned long max, prev_end, subtree_gap;
363 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
364 * allow two stack_guard_gaps between them here, and when choosing
365 * an unmapped area; whereas when expanding we only require one.
366 * That's a little inconsistent, but keeps the code here simpler.
368 max = vm_start_gap(vma);
370 prev_end = vm_end_gap(vma->vm_prev);
376 if (vma->vm_rb.rb_left) {
377 subtree_gap = rb_entry(vma->vm_rb.rb_left,
378 struct vm_area_struct, vm_rb)->rb_subtree_gap;
379 if (subtree_gap > max)
382 if (vma->vm_rb.rb_right) {
383 subtree_gap = rb_entry(vma->vm_rb.rb_right,
384 struct vm_area_struct, vm_rb)->rb_subtree_gap;
385 if (subtree_gap > max)
391 #ifdef CONFIG_DEBUG_VM_RB
392 static int browse_rb(struct rb_root *root)
394 int i = 0, j, bug = 0;
395 struct rb_node *nd, *pn = NULL;
396 unsigned long prev = 0, pend = 0;
398 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
399 struct vm_area_struct *vma;
400 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
401 if (vma->vm_start < prev) {
402 pr_emerg("vm_start %lx < prev %lx\n",
403 vma->vm_start, prev);
406 if (vma->vm_start < pend) {
407 pr_emerg("vm_start %lx < pend %lx\n",
408 vma->vm_start, pend);
411 if (vma->vm_start > vma->vm_end) {
412 pr_emerg("vm_start %lx > vm_end %lx\n",
413 vma->vm_start, vma->vm_end);
416 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
417 pr_emerg("free gap %lx, correct %lx\n",
419 vma_compute_subtree_gap(vma));
424 prev = vma->vm_start;
428 for (nd = pn; nd; nd = rb_prev(nd))
431 pr_emerg("backwards %d, forwards %d\n", j, i);
437 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
441 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
442 struct vm_area_struct *vma;
443 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
444 VM_BUG_ON_VMA(vma != ignore &&
445 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
450 static void validate_mm(struct mm_struct *mm)
454 unsigned long highest_address = 0;
455 struct vm_area_struct *vma = mm->mmap;
458 struct anon_vma *anon_vma = vma->anon_vma;
459 struct anon_vma_chain *avc;
462 anon_vma_lock_read(anon_vma);
463 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
464 anon_vma_interval_tree_verify(avc);
465 anon_vma_unlock_read(anon_vma);
468 highest_address = vm_end_gap(vma);
472 if (i != mm->map_count) {
473 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
476 if (highest_address != mm->highest_vm_end) {
477 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
478 mm->highest_vm_end, highest_address);
481 i = browse_rb(&mm->mm_rb);
482 if (i != mm->map_count) {
484 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
487 VM_BUG_ON_MM(bug, mm);
490 #define validate_mm_rb(root, ignore) do { } while (0)
491 #define validate_mm(mm) do { } while (0)
494 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
495 unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
498 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
499 * vma->vm_prev->vm_end values changed, without modifying the vma's position
502 static void vma_gap_update(struct vm_area_struct *vma)
505 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
506 * function that does exacltly what we want.
508 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
511 static inline void vma_rb_insert(struct vm_area_struct *vma,
512 struct rb_root *root)
514 /* All rb_subtree_gap values must be consistent prior to insertion */
515 validate_mm_rb(root, NULL);
517 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
520 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
523 * All rb_subtree_gap values must be consistent prior to erase,
524 * with the possible exception of the vma being erased.
526 validate_mm_rb(root, vma);
529 * Note rb_erase_augmented is a fairly large inline function,
530 * so make sure we instantiate it only once with our desired
531 * augmented rbtree callbacks.
533 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
537 * vma has some anon_vma assigned, and is already inserted on that
538 * anon_vma's interval trees.
540 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
541 * vma must be removed from the anon_vma's interval trees using
542 * anon_vma_interval_tree_pre_update_vma().
544 * After the update, the vma will be reinserted using
545 * anon_vma_interval_tree_post_update_vma().
547 * The entire update must be protected by exclusive mmap_sem and by
548 * the root anon_vma's mutex.
551 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
553 struct anon_vma_chain *avc;
555 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
556 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
560 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
562 struct anon_vma_chain *avc;
564 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
565 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
568 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
569 unsigned long end, struct vm_area_struct **pprev,
570 struct rb_node ***rb_link, struct rb_node **rb_parent)
572 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
574 __rb_link = &mm->mm_rb.rb_node;
575 rb_prev = __rb_parent = NULL;
578 struct vm_area_struct *vma_tmp;
580 __rb_parent = *__rb_link;
581 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
583 if (vma_tmp->vm_end > addr) {
584 /* Fail if an existing vma overlaps the area */
585 if (vma_tmp->vm_start < end)
587 __rb_link = &__rb_parent->rb_left;
589 rb_prev = __rb_parent;
590 __rb_link = &__rb_parent->rb_right;
596 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
597 *rb_link = __rb_link;
598 *rb_parent = __rb_parent;
602 static unsigned long count_vma_pages_range(struct mm_struct *mm,
603 unsigned long addr, unsigned long end)
605 unsigned long nr_pages = 0;
606 struct vm_area_struct *vma;
608 /* Find first overlaping mapping */
609 vma = find_vma_intersection(mm, addr, end);
613 nr_pages = (min(end, vma->vm_end) -
614 max(addr, vma->vm_start)) >> PAGE_SHIFT;
616 /* Iterate over the rest of the overlaps */
617 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
618 unsigned long overlap_len;
620 if (vma->vm_start > end)
623 overlap_len = min(end, vma->vm_end) - vma->vm_start;
624 nr_pages += overlap_len >> PAGE_SHIFT;
630 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
631 struct rb_node **rb_link, struct rb_node *rb_parent)
633 /* Update tracking information for the gap following the new vma. */
635 vma_gap_update(vma->vm_next);
637 mm->highest_vm_end = vm_end_gap(vma);
640 * vma->vm_prev wasn't known when we followed the rbtree to find the
641 * correct insertion point for that vma. As a result, we could not
642 * update the vma vm_rb parents rb_subtree_gap values on the way down.
643 * So, we first insert the vma with a zero rb_subtree_gap value
644 * (to be consistent with what we did on the way down), and then
645 * immediately update the gap to the correct value. Finally we
646 * rebalance the rbtree after all augmented values have been set.
648 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
649 vma->rb_subtree_gap = 0;
651 vma_rb_insert(vma, &mm->mm_rb);
654 static void __vma_link_file(struct vm_area_struct *vma)
660 struct address_space *mapping = file->f_mapping;
662 if (vma->vm_flags & VM_DENYWRITE)
663 atomic_dec(&file_inode(file)->i_writecount);
664 if (vma->vm_flags & VM_SHARED)
665 atomic_inc(&mapping->i_mmap_writable);
667 flush_dcache_mmap_lock(mapping);
668 vma_interval_tree_insert(vma, &mapping->i_mmap);
669 flush_dcache_mmap_unlock(mapping);
674 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
675 struct vm_area_struct *prev, struct rb_node **rb_link,
676 struct rb_node *rb_parent)
678 __vma_link_list(mm, vma, prev, rb_parent);
679 __vma_link_rb(mm, vma, rb_link, rb_parent);
682 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
683 struct vm_area_struct *prev, struct rb_node **rb_link,
684 struct rb_node *rb_parent)
686 struct address_space *mapping = NULL;
689 mapping = vma->vm_file->f_mapping;
690 i_mmap_lock_write(mapping);
693 __vma_link(mm, vma, prev, rb_link, rb_parent);
694 __vma_link_file(vma);
697 i_mmap_unlock_write(mapping);
704 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
705 * mm's list and rbtree. It has already been inserted into the interval tree.
707 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
709 struct vm_area_struct *prev;
710 struct rb_node **rb_link, *rb_parent;
712 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
713 &prev, &rb_link, &rb_parent))
715 __vma_link(mm, vma, prev, rb_link, rb_parent);
720 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
721 struct vm_area_struct *prev)
723 struct vm_area_struct *next;
725 vma_rb_erase(vma, &mm->mm_rb);
726 prev->vm_next = next = vma->vm_next;
728 next->vm_prev = prev;
731 vmacache_invalidate(mm);
735 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
736 * is already present in an i_mmap tree without adjusting the tree.
737 * The following helper function should be used when such adjustments
738 * are necessary. The "insert" vma (if any) is to be inserted
739 * before we drop the necessary locks.
741 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
742 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
744 struct mm_struct *mm = vma->vm_mm;
745 struct vm_area_struct *next = vma->vm_next;
746 struct vm_area_struct *importer = NULL;
747 struct address_space *mapping = NULL;
748 struct rb_root *root = NULL;
749 struct anon_vma *anon_vma = NULL;
750 struct file *file = vma->vm_file;
751 bool start_changed = false, end_changed = false;
752 long adjust_next = 0;
755 if (next && !insert) {
756 struct vm_area_struct *exporter = NULL;
758 if (end >= next->vm_end) {
760 * vma expands, overlapping all the next, and
761 * perhaps the one after too (mprotect case 6).
763 again: remove_next = 1 + (end > next->vm_end);
767 } else if (end > next->vm_start) {
769 * vma expands, overlapping part of the next:
770 * mprotect case 5 shifting the boundary up.
772 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
775 } else if (end < vma->vm_end) {
777 * vma shrinks, and !insert tells it's not
778 * split_vma inserting another: so it must be
779 * mprotect case 4 shifting the boundary down.
781 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
787 * Easily overlooked: when mprotect shifts the boundary,
788 * make sure the expanding vma has anon_vma set if the
789 * shrinking vma had, to cover any anon pages imported.
791 if (exporter && exporter->anon_vma && !importer->anon_vma) {
794 importer->anon_vma = exporter->anon_vma;
795 error = anon_vma_clone(importer, exporter);
802 mapping = file->f_mapping;
803 root = &mapping->i_mmap;
804 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
807 uprobe_munmap(next, next->vm_start, next->vm_end);
809 i_mmap_lock_write(mapping);
812 * Put into interval tree now, so instantiated pages
813 * are visible to arm/parisc __flush_dcache_page
814 * throughout; but we cannot insert into address
815 * space until vma start or end is updated.
817 __vma_link_file(insert);
821 vma_adjust_trans_huge(vma, start, end, adjust_next);
823 anon_vma = vma->anon_vma;
824 if (!anon_vma && adjust_next)
825 anon_vma = next->anon_vma;
827 VM_BUG_ON_VMA(adjust_next && next->anon_vma &&
828 anon_vma != next->anon_vma, next);
829 anon_vma_lock_write(anon_vma);
830 anon_vma_interval_tree_pre_update_vma(vma);
832 anon_vma_interval_tree_pre_update_vma(next);
836 flush_dcache_mmap_lock(mapping);
837 vma_interval_tree_remove(vma, root);
839 vma_interval_tree_remove(next, root);
842 if (start != vma->vm_start) {
843 vma->vm_start = start;
844 start_changed = true;
846 if (end != vma->vm_end) {
850 vma->vm_pgoff = pgoff;
852 next->vm_start += adjust_next << PAGE_SHIFT;
853 next->vm_pgoff += adjust_next;
858 vma_interval_tree_insert(next, root);
859 vma_interval_tree_insert(vma, root);
860 flush_dcache_mmap_unlock(mapping);
865 * vma_merge has merged next into vma, and needs
866 * us to remove next before dropping the locks.
868 __vma_unlink(mm, next, vma);
870 __remove_shared_vm_struct(next, file, mapping);
873 * split_vma has split insert from vma, and needs
874 * us to insert it before dropping the locks
875 * (it may either follow vma or precede it).
877 __insert_vm_struct(mm, insert);
883 mm->highest_vm_end = vm_end_gap(vma);
884 else if (!adjust_next)
885 vma_gap_update(next);
890 anon_vma_interval_tree_post_update_vma(vma);
892 anon_vma_interval_tree_post_update_vma(next);
893 anon_vma_unlock_write(anon_vma);
896 i_mmap_unlock_write(mapping);
907 uprobe_munmap(next, next->vm_start, next->vm_end);
911 anon_vma_merge(vma, next);
913 mpol_put(vma_policy(next));
914 kmem_cache_free(vm_area_cachep, next);
916 * In mprotect's case 6 (see comments on vma_merge),
917 * we must remove another next too. It would clutter
918 * up the code too much to do both in one go.
921 if (remove_next == 2)
924 vma_gap_update(next);
926 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
937 * If the vma has a ->close operation then the driver probably needs to release
938 * per-vma resources, so we don't attempt to merge those.
940 static inline int is_mergeable_vma(struct vm_area_struct *vma,
941 struct file *file, unsigned long vm_flags,
942 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
945 * VM_SOFTDIRTY should not prevent from VMA merging, if we
946 * match the flags but dirty bit -- the caller should mark
947 * merged VMA as dirty. If dirty bit won't be excluded from
948 * comparison, we increase pressue on the memory system forcing
949 * the kernel to generate new VMAs when old one could be
952 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
954 if (vma->vm_file != file)
956 if (vma->vm_ops && vma->vm_ops->close)
958 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
963 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
964 struct anon_vma *anon_vma2,
965 struct vm_area_struct *vma)
968 * The list_is_singular() test is to avoid merging VMA cloned from
969 * parents. This can improve scalability caused by anon_vma lock.
971 if ((!anon_vma1 || !anon_vma2) && (!vma ||
972 list_is_singular(&vma->anon_vma_chain)))
974 return anon_vma1 == anon_vma2;
978 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
979 * in front of (at a lower virtual address and file offset than) the vma.
981 * We cannot merge two vmas if they have differently assigned (non-NULL)
982 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
984 * We don't check here for the merged mmap wrapping around the end of pagecache
985 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
986 * wrap, nor mmaps which cover the final page at index -1UL.
989 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
990 struct anon_vma *anon_vma, struct file *file,
992 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
994 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
995 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
996 if (vma->vm_pgoff == vm_pgoff)
1003 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1004 * beyond (at a higher virtual address and file offset than) the vma.
1006 * We cannot merge two vmas if they have differently assigned (non-NULL)
1007 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1010 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1011 struct anon_vma *anon_vma, struct file *file,
1013 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1015 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1016 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1018 vm_pglen = vma_pages(vma);
1019 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1026 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1027 * whether that can be merged with its predecessor or its successor.
1028 * Or both (it neatly fills a hole).
1030 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1031 * certain not to be mapped by the time vma_merge is called; but when
1032 * called for mprotect, it is certain to be already mapped (either at
1033 * an offset within prev, or at the start of next), and the flags of
1034 * this area are about to be changed to vm_flags - and the no-change
1035 * case has already been eliminated.
1037 * The following mprotect cases have to be considered, where AAAA is
1038 * the area passed down from mprotect_fixup, never extending beyond one
1039 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1041 * AAAA AAAA AAAA AAAA
1042 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1043 * cannot merge might become might become might become
1044 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1045 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1046 * mremap move: PPPPNNNNNNNN 8
1048 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1049 * might become case 1 below case 2 below case 3 below
1051 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
1052 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
1054 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1055 struct vm_area_struct *prev, unsigned long addr,
1056 unsigned long end, unsigned long vm_flags,
1057 struct anon_vma *anon_vma, struct file *file,
1058 pgoff_t pgoff, struct mempolicy *policy,
1059 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1061 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1062 struct vm_area_struct *area, *next;
1066 * We later require that vma->vm_flags == vm_flags,
1067 * so this tests vma->vm_flags & VM_SPECIAL, too.
1069 if (vm_flags & VM_SPECIAL)
1073 next = prev->vm_next;
1077 if (next && next->vm_end == end) /* cases 6, 7, 8 */
1078 next = next->vm_next;
1081 * Can it merge with the predecessor?
1083 if (prev && prev->vm_end == addr &&
1084 mpol_equal(vma_policy(prev), policy) &&
1085 can_vma_merge_after(prev, vm_flags,
1086 anon_vma, file, pgoff,
1087 vm_userfaultfd_ctx)) {
1089 * OK, it can. Can we now merge in the successor as well?
1091 if (next && end == next->vm_start &&
1092 mpol_equal(policy, vma_policy(next)) &&
1093 can_vma_merge_before(next, vm_flags,
1096 vm_userfaultfd_ctx) &&
1097 is_mergeable_anon_vma(prev->anon_vma,
1098 next->anon_vma, NULL)) {
1100 err = vma_adjust(prev, prev->vm_start,
1101 next->vm_end, prev->vm_pgoff, NULL);
1102 } else /* cases 2, 5, 7 */
1103 err = vma_adjust(prev, prev->vm_start,
1104 end, prev->vm_pgoff, NULL);
1107 khugepaged_enter_vma_merge(prev, vm_flags);
1112 * Can this new request be merged in front of next?
1114 if (next && end == next->vm_start &&
1115 mpol_equal(policy, vma_policy(next)) &&
1116 can_vma_merge_before(next, vm_flags,
1117 anon_vma, file, pgoff+pglen,
1118 vm_userfaultfd_ctx)) {
1119 if (prev && addr < prev->vm_end) /* case 4 */
1120 err = vma_adjust(prev, prev->vm_start,
1121 addr, prev->vm_pgoff, NULL);
1122 else /* cases 3, 8 */
1123 err = vma_adjust(area, addr, next->vm_end,
1124 next->vm_pgoff - pglen, NULL);
1127 khugepaged_enter_vma_merge(area, vm_flags);
1135 * Rough compatbility check to quickly see if it's even worth looking
1136 * at sharing an anon_vma.
1138 * They need to have the same vm_file, and the flags can only differ
1139 * in things that mprotect may change.
1141 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1142 * we can merge the two vma's. For example, we refuse to merge a vma if
1143 * there is a vm_ops->close() function, because that indicates that the
1144 * driver is doing some kind of reference counting. But that doesn't
1145 * really matter for the anon_vma sharing case.
1147 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1149 return a->vm_end == b->vm_start &&
1150 mpol_equal(vma_policy(a), vma_policy(b)) &&
1151 a->vm_file == b->vm_file &&
1152 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1153 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1157 * Do some basic sanity checking to see if we can re-use the anon_vma
1158 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1159 * the same as 'old', the other will be the new one that is trying
1160 * to share the anon_vma.
1162 * NOTE! This runs with mm_sem held for reading, so it is possible that
1163 * the anon_vma of 'old' is concurrently in the process of being set up
1164 * by another page fault trying to merge _that_. But that's ok: if it
1165 * is being set up, that automatically means that it will be a singleton
1166 * acceptable for merging, so we can do all of this optimistically. But
1167 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1169 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1170 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1171 * is to return an anon_vma that is "complex" due to having gone through
1174 * We also make sure that the two vma's are compatible (adjacent,
1175 * and with the same memory policies). That's all stable, even with just
1176 * a read lock on the mm_sem.
1178 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1180 if (anon_vma_compatible(a, b)) {
1181 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1183 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1190 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1191 * neighbouring vmas for a suitable anon_vma, before it goes off
1192 * to allocate a new anon_vma. It checks because a repetitive
1193 * sequence of mprotects and faults may otherwise lead to distinct
1194 * anon_vmas being allocated, preventing vma merge in subsequent
1197 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1199 struct anon_vma *anon_vma;
1200 struct vm_area_struct *near;
1202 near = vma->vm_next;
1206 anon_vma = reusable_anon_vma(near, vma, near);
1210 near = vma->vm_prev;
1214 anon_vma = reusable_anon_vma(near, near, vma);
1219 * There's no absolute need to look only at touching neighbours:
1220 * we could search further afield for "compatible" anon_vmas.
1221 * But it would probably just be a waste of time searching,
1222 * or lead to too many vmas hanging off the same anon_vma.
1223 * We're trying to allow mprotect remerging later on,
1224 * not trying to minimize memory used for anon_vmas.
1229 #ifdef CONFIG_PROC_FS
1230 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
1231 struct file *file, long pages)
1233 const unsigned long stack_flags
1234 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
1236 mm->total_vm += pages;
1239 mm->shared_vm += pages;
1240 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
1241 mm->exec_vm += pages;
1242 } else if (flags & stack_flags)
1243 mm->stack_vm += pages;
1245 #endif /* CONFIG_PROC_FS */
1248 * If a hint addr is less than mmap_min_addr change hint to be as
1249 * low as possible but still greater than mmap_min_addr
1251 static inline unsigned long round_hint_to_min(unsigned long hint)
1254 if (((void *)hint != NULL) &&
1255 (hint < mmap_min_addr))
1256 return PAGE_ALIGN(mmap_min_addr);
1260 static inline int mlock_future_check(struct mm_struct *mm,
1261 unsigned long flags,
1264 unsigned long locked, lock_limit;
1266 /* mlock MCL_FUTURE? */
1267 if (flags & VM_LOCKED) {
1268 locked = len >> PAGE_SHIFT;
1269 locked += mm->locked_vm;
1270 lock_limit = rlimit(RLIMIT_MEMLOCK);
1271 lock_limit >>= PAGE_SHIFT;
1272 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1279 * The caller must hold down_write(¤t->mm->mmap_sem).
1281 unsigned long do_mmap(struct file *file, unsigned long addr,
1282 unsigned long len, unsigned long prot,
1283 unsigned long flags, vm_flags_t vm_flags,
1284 unsigned long pgoff, unsigned long *populate)
1286 struct mm_struct *mm = current->mm;
1294 * Does the application expect PROT_READ to imply PROT_EXEC?
1296 * (the exception is when the underlying filesystem is noexec
1297 * mounted, in which case we dont add PROT_EXEC.)
1299 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1300 if (!(file && path_noexec(&file->f_path)))
1303 if (!(flags & MAP_FIXED))
1304 addr = round_hint_to_min(addr);
1306 /* Careful about overflows.. */
1307 len = PAGE_ALIGN(len);
1311 /* offset overflow? */
1312 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1315 /* Too many mappings? */
1316 if (mm->map_count > sysctl_max_map_count)
1319 /* Obtain the address to map to. we verify (or select) it and ensure
1320 * that it represents a valid section of the address space.
1322 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1323 if (offset_in_page(addr))
1326 /* Do simple checking here so the lower-level routines won't have
1327 * to. we assume access permissions have been handled by the open
1328 * of the memory object, so we don't do any here.
1330 vm_flags |= calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1331 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1333 if (flags & MAP_LOCKED)
1334 if (!can_do_mlock())
1337 if (mlock_future_check(mm, vm_flags, len))
1341 struct inode *inode = file_inode(file);
1343 switch (flags & MAP_TYPE) {
1345 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1349 * Make sure we don't allow writing to an append-only
1352 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1356 * Make sure there are no mandatory locks on the file.
1358 if (locks_verify_locked(file))
1361 vm_flags |= VM_SHARED | VM_MAYSHARE;
1362 if (!(file->f_mode & FMODE_WRITE))
1363 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1367 if (!(file->f_mode & FMODE_READ))
1369 if (path_noexec(&file->f_path)) {
1370 if (vm_flags & VM_EXEC)
1372 vm_flags &= ~VM_MAYEXEC;
1375 if (!file->f_op->mmap)
1377 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1385 switch (flags & MAP_TYPE) {
1387 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1393 vm_flags |= VM_SHARED | VM_MAYSHARE;
1397 * Set pgoff according to addr for anon_vma.
1399 pgoff = addr >> PAGE_SHIFT;
1407 * Set 'VM_NORESERVE' if we should not account for the
1408 * memory use of this mapping.
1410 if (flags & MAP_NORESERVE) {
1411 /* We honor MAP_NORESERVE if allowed to overcommit */
1412 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1413 vm_flags |= VM_NORESERVE;
1415 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1416 if (file && is_file_hugepages(file))
1417 vm_flags |= VM_NORESERVE;
1420 addr = mmap_region(file, addr, len, vm_flags, pgoff);
1421 if (!IS_ERR_VALUE(addr) &&
1422 ((vm_flags & VM_LOCKED) ||
1423 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1428 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1429 unsigned long, prot, unsigned long, flags,
1430 unsigned long, fd, unsigned long, pgoff)
1432 struct file *file = NULL;
1433 unsigned long retval;
1435 if (!(flags & MAP_ANONYMOUS)) {
1436 audit_mmap_fd(fd, flags);
1440 if (is_file_hugepages(file))
1441 len = ALIGN(len, huge_page_size(hstate_file(file)));
1443 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1445 } else if (flags & MAP_HUGETLB) {
1446 struct user_struct *user = NULL;
1449 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1453 len = ALIGN(len, huge_page_size(hs));
1455 * VM_NORESERVE is used because the reservations will be
1456 * taken when vm_ops->mmap() is called
1457 * A dummy user value is used because we are not locking
1458 * memory so no accounting is necessary
1460 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1462 &user, HUGETLB_ANONHUGE_INODE,
1463 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1465 return PTR_ERR(file);
1468 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1470 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1477 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1478 struct mmap_arg_struct {
1482 unsigned long flags;
1484 unsigned long offset;
1487 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1489 struct mmap_arg_struct a;
1491 if (copy_from_user(&a, arg, sizeof(a)))
1493 if (offset_in_page(a.offset))
1496 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1497 a.offset >> PAGE_SHIFT);
1499 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1502 * Some shared mappigns will want the pages marked read-only
1503 * to track write events. If so, we'll downgrade vm_page_prot
1504 * to the private version (using protection_map[] without the
1507 int vma_wants_writenotify(struct vm_area_struct *vma)
1509 vm_flags_t vm_flags = vma->vm_flags;
1510 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1512 /* If it was private or non-writable, the write bit is already clear */
1513 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1516 /* The backer wishes to know when pages are first written to? */
1517 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1520 /* The open routine did something to the protections that pgprot_modify
1521 * won't preserve? */
1522 if (pgprot_val(vma->vm_page_prot) !=
1523 pgprot_val(vm_pgprot_modify(vma->vm_page_prot, vm_flags)))
1526 /* Do we need to track softdirty? */
1527 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1530 /* Specialty mapping? */
1531 if (vm_flags & VM_PFNMAP)
1534 /* Can the mapping track the dirty pages? */
1535 return vma->vm_file && vma->vm_file->f_mapping &&
1536 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1540 * We account for memory if it's a private writeable mapping,
1541 * not hugepages and VM_NORESERVE wasn't set.
1543 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1546 * hugetlb has its own accounting separate from the core VM
1547 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1549 if (file && is_file_hugepages(file))
1552 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1555 unsigned long mmap_region(struct file *file, unsigned long addr,
1556 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1558 struct mm_struct *mm = current->mm;
1559 struct vm_area_struct *vma, *prev;
1561 struct rb_node **rb_link, *rb_parent;
1562 unsigned long charged = 0;
1564 /* Check against address space limit. */
1565 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) {
1566 unsigned long nr_pages;
1569 * MAP_FIXED may remove pages of mappings that intersects with
1570 * requested mapping. Account for the pages it would unmap.
1572 if (!(vm_flags & MAP_FIXED))
1575 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1577 if (!may_expand_vm(mm, (len >> PAGE_SHIFT) - nr_pages))
1581 /* Clear old maps */
1582 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1584 if (do_munmap(mm, addr, len))
1589 * Private writable mapping: check memory availability
1591 if (accountable_mapping(file, vm_flags)) {
1592 charged = len >> PAGE_SHIFT;
1593 if (security_vm_enough_memory_mm(mm, charged))
1595 vm_flags |= VM_ACCOUNT;
1599 * Can we just expand an old mapping?
1601 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1602 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1607 * Determine the object being mapped and call the appropriate
1608 * specific mapper. the address has already been validated, but
1609 * not unmapped, but the maps are removed from the list.
1611 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1618 vma->vm_start = addr;
1619 vma->vm_end = addr + len;
1620 vma->vm_flags = vm_flags;
1621 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1622 vma->vm_pgoff = pgoff;
1623 INIT_LIST_HEAD(&vma->anon_vma_chain);
1626 if (vm_flags & VM_DENYWRITE) {
1627 error = deny_write_access(file);
1631 if (vm_flags & VM_SHARED) {
1632 error = mapping_map_writable(file->f_mapping);
1634 goto allow_write_and_free_vma;
1637 /* ->mmap() can change vma->vm_file, but must guarantee that
1638 * vma_link() below can deny write-access if VM_DENYWRITE is set
1639 * and map writably if VM_SHARED is set. This usually means the
1640 * new file must not have been exposed to user-space, yet.
1642 vma->vm_file = get_file(file);
1643 error = file->f_op->mmap(file, vma);
1645 goto unmap_and_free_vma;
1647 /* Can addr have changed??
1649 * Answer: Yes, several device drivers can do it in their
1650 * f_op->mmap method. -DaveM
1651 * Bug: If addr is changed, prev, rb_link, rb_parent should
1652 * be updated for vma_link()
1654 WARN_ON_ONCE(addr != vma->vm_start);
1656 addr = vma->vm_start;
1657 vm_flags = vma->vm_flags;
1658 } else if (vm_flags & VM_SHARED) {
1659 error = shmem_zero_setup(vma);
1664 vma_link(mm, vma, prev, rb_link, rb_parent);
1665 /* Once vma denies write, undo our temporary denial count */
1667 if (vm_flags & VM_SHARED)
1668 mapping_unmap_writable(file->f_mapping);
1669 if (vm_flags & VM_DENYWRITE)
1670 allow_write_access(file);
1672 file = vma->vm_file;
1674 perf_event_mmap(vma);
1676 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1677 if (vm_flags & VM_LOCKED) {
1678 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1679 vma == get_gate_vma(current->mm)))
1680 mm->locked_vm += (len >> PAGE_SHIFT);
1682 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1689 * New (or expanded) vma always get soft dirty status.
1690 * Otherwise user-space soft-dirty page tracker won't
1691 * be able to distinguish situation when vma area unmapped,
1692 * then new mapped in-place (which must be aimed as
1693 * a completely new data area).
1695 vma->vm_flags |= VM_SOFTDIRTY;
1697 vma_set_page_prot(vma);
1702 vma->vm_file = NULL;
1705 /* Undo any partial mapping done by a device driver. */
1706 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1708 if (vm_flags & VM_SHARED)
1709 mapping_unmap_writable(file->f_mapping);
1710 allow_write_and_free_vma:
1711 if (vm_flags & VM_DENYWRITE)
1712 allow_write_access(file);
1714 kmem_cache_free(vm_area_cachep, vma);
1717 vm_unacct_memory(charged);
1721 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1724 * We implement the search by looking for an rbtree node that
1725 * immediately follows a suitable gap. That is,
1726 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1727 * - gap_end = vma->vm_start >= info->low_limit + length;
1728 * - gap_end - gap_start >= length
1731 struct mm_struct *mm = current->mm;
1732 struct vm_area_struct *vma;
1733 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1735 /* Adjust search length to account for worst case alignment overhead */
1736 length = info->length + info->align_mask;
1737 if (length < info->length)
1740 /* Adjust search limits by the desired length */
1741 if (info->high_limit < length)
1743 high_limit = info->high_limit - length;
1745 if (info->low_limit > high_limit)
1747 low_limit = info->low_limit + length;
1749 /* Check if rbtree root looks promising */
1750 if (RB_EMPTY_ROOT(&mm->mm_rb))
1752 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1753 if (vma->rb_subtree_gap < length)
1757 /* Visit left subtree if it looks promising */
1758 gap_end = vm_start_gap(vma);
1759 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1760 struct vm_area_struct *left =
1761 rb_entry(vma->vm_rb.rb_left,
1762 struct vm_area_struct, vm_rb);
1763 if (left->rb_subtree_gap >= length) {
1769 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1771 /* Check if current node has a suitable gap */
1772 if (gap_start > high_limit)
1774 if (gap_end >= low_limit &&
1775 gap_end > gap_start && gap_end - gap_start >= length)
1778 /* Visit right subtree if it looks promising */
1779 if (vma->vm_rb.rb_right) {
1780 struct vm_area_struct *right =
1781 rb_entry(vma->vm_rb.rb_right,
1782 struct vm_area_struct, vm_rb);
1783 if (right->rb_subtree_gap >= length) {
1789 /* Go back up the rbtree to find next candidate node */
1791 struct rb_node *prev = &vma->vm_rb;
1792 if (!rb_parent(prev))
1794 vma = rb_entry(rb_parent(prev),
1795 struct vm_area_struct, vm_rb);
1796 if (prev == vma->vm_rb.rb_left) {
1797 gap_start = vm_end_gap(vma->vm_prev);
1798 gap_end = vm_start_gap(vma);
1805 /* Check highest gap, which does not precede any rbtree node */
1806 gap_start = mm->highest_vm_end;
1807 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1808 if (gap_start > high_limit)
1812 /* We found a suitable gap. Clip it with the original low_limit. */
1813 if (gap_start < info->low_limit)
1814 gap_start = info->low_limit;
1816 /* Adjust gap address to the desired alignment */
1817 gap_start += (info->align_offset - gap_start) & info->align_mask;
1819 VM_BUG_ON(gap_start + info->length > info->high_limit);
1820 VM_BUG_ON(gap_start + info->length > gap_end);
1824 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1826 struct mm_struct *mm = current->mm;
1827 struct vm_area_struct *vma;
1828 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1830 /* Adjust search length to account for worst case alignment overhead */
1831 length = info->length + info->align_mask;
1832 if (length < info->length)
1836 * Adjust search limits by the desired length.
1837 * See implementation comment at top of unmapped_area().
1839 gap_end = info->high_limit;
1840 if (gap_end < length)
1842 high_limit = gap_end - length;
1844 if (info->low_limit > high_limit)
1846 low_limit = info->low_limit + length;
1848 /* Check highest gap, which does not precede any rbtree node */
1849 gap_start = mm->highest_vm_end;
1850 if (gap_start <= high_limit)
1853 /* Check if rbtree root looks promising */
1854 if (RB_EMPTY_ROOT(&mm->mm_rb))
1856 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1857 if (vma->rb_subtree_gap < length)
1861 /* Visit right subtree if it looks promising */
1862 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1863 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1864 struct vm_area_struct *right =
1865 rb_entry(vma->vm_rb.rb_right,
1866 struct vm_area_struct, vm_rb);
1867 if (right->rb_subtree_gap >= length) {
1874 /* Check if current node has a suitable gap */
1875 gap_end = vm_start_gap(vma);
1876 if (gap_end < low_limit)
1878 if (gap_start <= high_limit &&
1879 gap_end > gap_start && gap_end - gap_start >= length)
1882 /* Visit left subtree if it looks promising */
1883 if (vma->vm_rb.rb_left) {
1884 struct vm_area_struct *left =
1885 rb_entry(vma->vm_rb.rb_left,
1886 struct vm_area_struct, vm_rb);
1887 if (left->rb_subtree_gap >= length) {
1893 /* Go back up the rbtree to find next candidate node */
1895 struct rb_node *prev = &vma->vm_rb;
1896 if (!rb_parent(prev))
1898 vma = rb_entry(rb_parent(prev),
1899 struct vm_area_struct, vm_rb);
1900 if (prev == vma->vm_rb.rb_right) {
1901 gap_start = vma->vm_prev ?
1902 vm_end_gap(vma->vm_prev) : 0;
1909 /* We found a suitable gap. Clip it with the original high_limit. */
1910 if (gap_end > info->high_limit)
1911 gap_end = info->high_limit;
1914 /* Compute highest gap address at the desired alignment */
1915 gap_end -= info->length;
1916 gap_end -= (gap_end - info->align_offset) & info->align_mask;
1918 VM_BUG_ON(gap_end < info->low_limit);
1919 VM_BUG_ON(gap_end < gap_start);
1923 /* Get an address range which is currently unmapped.
1924 * For shmat() with addr=0.
1926 * Ugly calling convention alert:
1927 * Return value with the low bits set means error value,
1929 * if (ret & ~PAGE_MASK)
1932 * This function "knows" that -ENOMEM has the bits set.
1934 #ifndef HAVE_ARCH_UNMAPPED_AREA
1936 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1937 unsigned long len, unsigned long pgoff, unsigned long flags)
1939 struct mm_struct *mm = current->mm;
1940 struct vm_area_struct *vma, *prev;
1941 struct vm_unmapped_area_info info;
1943 if (len > TASK_SIZE - mmap_min_addr)
1946 if (flags & MAP_FIXED)
1950 addr = PAGE_ALIGN(addr);
1951 vma = find_vma_prev(mm, addr, &prev);
1952 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1953 (!vma || addr + len <= vm_start_gap(vma)) &&
1954 (!prev || addr >= vm_end_gap(prev)))
1960 info.low_limit = mm->mmap_base;
1961 info.high_limit = TASK_SIZE;
1962 info.align_mask = 0;
1963 return vm_unmapped_area(&info);
1968 * This mmap-allocator allocates new areas top-down from below the
1969 * stack's low limit (the base):
1971 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1973 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1974 const unsigned long len, const unsigned long pgoff,
1975 const unsigned long flags)
1977 struct vm_area_struct *vma, *prev;
1978 struct mm_struct *mm = current->mm;
1979 unsigned long addr = addr0;
1980 struct vm_unmapped_area_info info;
1982 /* requested length too big for entire address space */
1983 if (len > TASK_SIZE - mmap_min_addr)
1986 if (flags & MAP_FIXED)
1989 /* requesting a specific address */
1991 addr = PAGE_ALIGN(addr);
1992 vma = find_vma_prev(mm, addr, &prev);
1993 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1994 (!vma || addr + len <= vm_start_gap(vma)) &&
1995 (!prev || addr >= vm_end_gap(prev)))
1999 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2001 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2002 info.high_limit = mm->mmap_base;
2003 info.align_mask = 0;
2004 addr = vm_unmapped_area(&info);
2007 * A failed mmap() very likely causes application failure,
2008 * so fall back to the bottom-up function here. This scenario
2009 * can happen with large stack limits and large mmap()
2012 if (offset_in_page(addr)) {
2013 VM_BUG_ON(addr != -ENOMEM);
2015 info.low_limit = TASK_UNMAPPED_BASE;
2016 info.high_limit = TASK_SIZE;
2017 addr = vm_unmapped_area(&info);
2025 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2026 unsigned long pgoff, unsigned long flags)
2028 unsigned long (*get_area)(struct file *, unsigned long,
2029 unsigned long, unsigned long, unsigned long);
2031 unsigned long error = arch_mmap_check(addr, len, flags);
2035 /* Careful about overflows.. */
2036 if (len > TASK_SIZE)
2039 get_area = current->mm->get_unmapped_area;
2040 if (file && file->f_op->get_unmapped_area)
2041 get_area = file->f_op->get_unmapped_area;
2042 addr = get_area(file, addr, len, pgoff, flags);
2043 if (IS_ERR_VALUE(addr))
2046 if (addr > TASK_SIZE - len)
2048 if (offset_in_page(addr))
2051 addr = arch_rebalance_pgtables(addr, len);
2052 error = security_mmap_addr(addr);
2053 return error ? error : addr;
2056 EXPORT_SYMBOL(get_unmapped_area);
2058 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2059 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2061 struct rb_node *rb_node;
2062 struct vm_area_struct *vma;
2064 /* Check the cache first. */
2065 vma = vmacache_find(mm, addr);
2069 rb_node = mm->mm_rb.rb_node;
2072 struct vm_area_struct *tmp;
2074 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2076 if (tmp->vm_end > addr) {
2078 if (tmp->vm_start <= addr)
2080 rb_node = rb_node->rb_left;
2082 rb_node = rb_node->rb_right;
2086 vmacache_update(addr, vma);
2090 EXPORT_SYMBOL(find_vma);
2093 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2095 struct vm_area_struct *
2096 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2097 struct vm_area_struct **pprev)
2099 struct vm_area_struct *vma;
2101 vma = find_vma(mm, addr);
2103 *pprev = vma->vm_prev;
2105 struct rb_node *rb_node = mm->mm_rb.rb_node;
2108 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2109 rb_node = rb_node->rb_right;
2116 * Verify that the stack growth is acceptable and
2117 * update accounting. This is shared with both the
2118 * grow-up and grow-down cases.
2120 static int acct_stack_growth(struct vm_area_struct *vma,
2121 unsigned long size, unsigned long grow)
2123 struct mm_struct *mm = vma->vm_mm;
2124 struct rlimit *rlim = current->signal->rlim;
2125 unsigned long new_start;
2127 /* address space limit tests */
2128 if (!may_expand_vm(mm, grow))
2131 /* Stack limit test */
2132 if (size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2135 /* mlock limit tests */
2136 if (vma->vm_flags & VM_LOCKED) {
2137 unsigned long locked;
2138 unsigned long limit;
2139 locked = mm->locked_vm + grow;
2140 limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2141 limit >>= PAGE_SHIFT;
2142 if (locked > limit && !capable(CAP_IPC_LOCK))
2146 /* Check to ensure the stack will not grow into a hugetlb-only region */
2147 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2149 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2153 * Overcommit.. This must be the final test, as it will
2154 * update security statistics.
2156 if (security_vm_enough_memory_mm(mm, grow))
2162 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2164 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2165 * vma is the last one with address > vma->vm_end. Have to extend vma.
2167 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2169 struct mm_struct *mm = vma->vm_mm;
2170 struct vm_area_struct *next;
2171 unsigned long gap_addr;
2174 if (!(vma->vm_flags & VM_GROWSUP))
2177 /* Guard against exceeding limits of the address space. */
2178 address &= PAGE_MASK;
2179 if (address >= (TASK_SIZE & PAGE_MASK))
2181 address += PAGE_SIZE;
2183 /* Enforce stack_guard_gap */
2184 gap_addr = address + stack_guard_gap;
2186 /* Guard against overflow */
2187 if (gap_addr < address || gap_addr > TASK_SIZE)
2188 gap_addr = TASK_SIZE;
2190 next = vma->vm_next;
2191 if (next && next->vm_start < gap_addr) {
2192 if (!(next->vm_flags & VM_GROWSUP))
2194 /* Check that both stack segments have the same anon_vma? */
2197 /* We must make sure the anon_vma is allocated. */
2198 if (unlikely(anon_vma_prepare(vma)))
2202 * vma->vm_start/vm_end cannot change under us because the caller
2203 * is required to hold the mmap_sem in read mode. We need the
2204 * anon_vma lock to serialize against concurrent expand_stacks.
2206 anon_vma_lock_write(vma->anon_vma);
2208 /* Somebody else might have raced and expanded it already */
2209 if (address > vma->vm_end) {
2210 unsigned long size, grow;
2212 size = address - vma->vm_start;
2213 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2216 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2217 error = acct_stack_growth(vma, size, grow);
2220 * vma_gap_update() doesn't support concurrent
2221 * updates, but we only hold a shared mmap_sem
2222 * lock here, so we need to protect against
2223 * concurrent vma expansions.
2224 * anon_vma_lock_write() doesn't help here, as
2225 * we don't guarantee that all growable vmas
2226 * in a mm share the same root anon vma.
2227 * So, we reuse mm->page_table_lock to guard
2228 * against concurrent vma expansions.
2230 spin_lock(&mm->page_table_lock);
2231 if (vma->vm_flags & VM_LOCKED)
2232 mm->locked_vm += grow;
2233 vm_stat_account(mm, vma->vm_flags,
2234 vma->vm_file, grow);
2235 anon_vma_interval_tree_pre_update_vma(vma);
2236 vma->vm_end = address;
2237 anon_vma_interval_tree_post_update_vma(vma);
2239 vma_gap_update(vma->vm_next);
2241 mm->highest_vm_end = vm_end_gap(vma);
2242 spin_unlock(&mm->page_table_lock);
2244 perf_event_mmap(vma);
2248 anon_vma_unlock_write(vma->anon_vma);
2249 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2253 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2256 * vma is the first one with address < vma->vm_start. Have to extend vma.
2258 int expand_downwards(struct vm_area_struct *vma,
2259 unsigned long address)
2261 struct mm_struct *mm = vma->vm_mm;
2262 struct vm_area_struct *prev;
2263 unsigned long gap_addr;
2266 address &= PAGE_MASK;
2267 error = security_mmap_addr(address);
2271 /* Enforce stack_guard_gap */
2272 gap_addr = address - stack_guard_gap;
2273 if (gap_addr > address)
2275 prev = vma->vm_prev;
2276 if (prev && prev->vm_end > gap_addr) {
2277 if (!(prev->vm_flags & VM_GROWSDOWN))
2279 /* Check that both stack segments have the same anon_vma? */
2282 /* We must make sure the anon_vma is allocated. */
2283 if (unlikely(anon_vma_prepare(vma)))
2287 * vma->vm_start/vm_end cannot change under us because the caller
2288 * is required to hold the mmap_sem in read mode. We need the
2289 * anon_vma lock to serialize against concurrent expand_stacks.
2291 anon_vma_lock_write(vma->anon_vma);
2293 /* Somebody else might have raced and expanded it already */
2294 if (address < vma->vm_start) {
2295 unsigned long size, grow;
2297 size = vma->vm_end - address;
2298 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2301 if (grow <= vma->vm_pgoff) {
2302 error = acct_stack_growth(vma, size, grow);
2305 * vma_gap_update() doesn't support concurrent
2306 * updates, but we only hold a shared mmap_sem
2307 * lock here, so we need to protect against
2308 * concurrent vma expansions.
2309 * anon_vma_lock_write() doesn't help here, as
2310 * we don't guarantee that all growable vmas
2311 * in a mm share the same root anon vma.
2312 * So, we reuse mm->page_table_lock to guard
2313 * against concurrent vma expansions.
2315 spin_lock(&mm->page_table_lock);
2316 if (vma->vm_flags & VM_LOCKED)
2317 mm->locked_vm += grow;
2318 vm_stat_account(mm, vma->vm_flags,
2319 vma->vm_file, grow);
2320 anon_vma_interval_tree_pre_update_vma(vma);
2321 vma->vm_start = address;
2322 vma->vm_pgoff -= grow;
2323 anon_vma_interval_tree_post_update_vma(vma);
2324 vma_gap_update(vma);
2325 spin_unlock(&mm->page_table_lock);
2327 perf_event_mmap(vma);
2331 anon_vma_unlock_write(vma->anon_vma);
2332 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2337 /* enforced gap between the expanding stack and other mappings. */
2338 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2340 static int __init cmdline_parse_stack_guard_gap(char *p)
2345 val = simple_strtoul(p, &endptr, 10);
2347 stack_guard_gap = val << PAGE_SHIFT;
2351 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2353 #ifdef CONFIG_STACK_GROWSUP
2354 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2356 return expand_upwards(vma, address);
2359 struct vm_area_struct *
2360 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2362 struct vm_area_struct *vma, *prev;
2365 vma = find_vma_prev(mm, addr, &prev);
2366 if (vma && (vma->vm_start <= addr))
2368 if (!prev || expand_stack(prev, addr))
2370 if (prev->vm_flags & VM_LOCKED)
2371 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2375 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2377 return expand_downwards(vma, address);
2380 struct vm_area_struct *
2381 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2383 struct vm_area_struct *vma;
2384 unsigned long start;
2387 vma = find_vma(mm, addr);
2390 if (vma->vm_start <= addr)
2392 if (!(vma->vm_flags & VM_GROWSDOWN))
2394 start = vma->vm_start;
2395 if (expand_stack(vma, addr))
2397 if (vma->vm_flags & VM_LOCKED)
2398 populate_vma_page_range(vma, addr, start, NULL);
2403 EXPORT_SYMBOL_GPL(find_extend_vma);
2406 * Ok - we have the memory areas we should free on the vma list,
2407 * so release them, and do the vma updates.
2409 * Called with the mm semaphore held.
2411 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2413 unsigned long nr_accounted = 0;
2415 /* Update high watermark before we lower total_vm */
2416 update_hiwater_vm(mm);
2418 long nrpages = vma_pages(vma);
2420 if (vma->vm_flags & VM_ACCOUNT)
2421 nr_accounted += nrpages;
2422 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
2423 vma = remove_vma(vma);
2425 vm_unacct_memory(nr_accounted);
2430 * Get rid of page table information in the indicated region.
2432 * Called with the mm semaphore held.
2434 static void unmap_region(struct mm_struct *mm,
2435 struct vm_area_struct *vma, struct vm_area_struct *prev,
2436 unsigned long start, unsigned long end)
2438 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2439 struct mmu_gather tlb;
2442 tlb_gather_mmu(&tlb, mm, start, end);
2443 update_hiwater_rss(mm);
2444 unmap_vmas(&tlb, vma, start, end);
2445 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2446 next ? next->vm_start : USER_PGTABLES_CEILING);
2447 tlb_finish_mmu(&tlb, start, end);
2451 * Create a list of vma's touched by the unmap, removing them from the mm's
2452 * vma list as we go..
2455 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2456 struct vm_area_struct *prev, unsigned long end)
2458 struct vm_area_struct **insertion_point;
2459 struct vm_area_struct *tail_vma = NULL;
2461 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2462 vma->vm_prev = NULL;
2464 vma_rb_erase(vma, &mm->mm_rb);
2468 } while (vma && vma->vm_start < end);
2469 *insertion_point = vma;
2471 vma->vm_prev = prev;
2472 vma_gap_update(vma);
2474 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2475 tail_vma->vm_next = NULL;
2477 /* Kill the cache */
2478 vmacache_invalidate(mm);
2482 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2483 * munmap path where it doesn't make sense to fail.
2485 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2486 unsigned long addr, int new_below)
2488 struct vm_area_struct *new;
2491 if (is_vm_hugetlb_page(vma) && (addr &
2492 ~(huge_page_mask(hstate_vma(vma)))))
2495 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2499 /* most fields are the same, copy all, and then fixup */
2502 INIT_LIST_HEAD(&new->anon_vma_chain);
2507 new->vm_start = addr;
2508 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2511 err = vma_dup_policy(vma, new);
2515 err = anon_vma_clone(new, vma);
2520 get_file(new->vm_file);
2522 if (new->vm_ops && new->vm_ops->open)
2523 new->vm_ops->open(new);
2526 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2527 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2529 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2535 /* Clean everything up if vma_adjust failed. */
2536 if (new->vm_ops && new->vm_ops->close)
2537 new->vm_ops->close(new);
2540 unlink_anon_vmas(new);
2542 mpol_put(vma_policy(new));
2544 kmem_cache_free(vm_area_cachep, new);
2549 * Split a vma into two pieces at address 'addr', a new vma is allocated
2550 * either for the first part or the tail.
2552 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2553 unsigned long addr, int new_below)
2555 if (mm->map_count >= sysctl_max_map_count)
2558 return __split_vma(mm, vma, addr, new_below);
2561 /* Munmap is split into 2 main parts -- this part which finds
2562 * what needs doing, and the areas themselves, which do the
2563 * work. This now handles partial unmappings.
2564 * Jeremy Fitzhardinge <jeremy@goop.org>
2566 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2569 struct vm_area_struct *vma, *prev, *last;
2571 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2574 len = PAGE_ALIGN(len);
2578 /* Find the first overlapping VMA */
2579 vma = find_vma(mm, start);
2582 prev = vma->vm_prev;
2583 /* we have start < vma->vm_end */
2585 /* if it doesn't overlap, we have nothing.. */
2587 if (vma->vm_start >= end)
2591 * If we need to split any vma, do it now to save pain later.
2593 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2594 * unmapped vm_area_struct will remain in use: so lower split_vma
2595 * places tmp vma above, and higher split_vma places tmp vma below.
2597 if (start > vma->vm_start) {
2601 * Make sure that map_count on return from munmap() will
2602 * not exceed its limit; but let map_count go just above
2603 * its limit temporarily, to help free resources as expected.
2605 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2608 error = __split_vma(mm, vma, start, 0);
2614 /* Does it split the last one? */
2615 last = find_vma(mm, end);
2616 if (last && end > last->vm_start) {
2617 int error = __split_vma(mm, last, end, 1);
2621 vma = prev ? prev->vm_next : mm->mmap;
2624 * unlock any mlock()ed ranges before detaching vmas
2626 if (mm->locked_vm) {
2627 struct vm_area_struct *tmp = vma;
2628 while (tmp && tmp->vm_start < end) {
2629 if (tmp->vm_flags & VM_LOCKED) {
2630 mm->locked_vm -= vma_pages(tmp);
2631 munlock_vma_pages_all(tmp);
2638 * Remove the vma's, and unmap the actual pages
2640 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2641 unmap_region(mm, vma, prev, start, end);
2643 arch_unmap(mm, vma, start, end);
2645 /* Fix up all other VM information */
2646 remove_vma_list(mm, vma);
2651 int vm_munmap(unsigned long start, size_t len)
2654 struct mm_struct *mm = current->mm;
2656 down_write(&mm->mmap_sem);
2657 ret = do_munmap(mm, start, len);
2658 up_write(&mm->mmap_sem);
2661 EXPORT_SYMBOL(vm_munmap);
2663 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2665 profile_munmap(addr);
2666 return vm_munmap(addr, len);
2671 * Emulation of deprecated remap_file_pages() syscall.
2673 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2674 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2677 struct mm_struct *mm = current->mm;
2678 struct vm_area_struct *vma;
2679 unsigned long populate = 0;
2680 unsigned long ret = -EINVAL;
2683 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. "
2684 "See Documentation/vm/remap_file_pages.txt.\n",
2685 current->comm, current->pid);
2689 start = start & PAGE_MASK;
2690 size = size & PAGE_MASK;
2692 if (start + size <= start)
2695 /* Does pgoff wrap? */
2696 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2699 down_write(&mm->mmap_sem);
2700 vma = find_vma(mm, start);
2702 if (!vma || !(vma->vm_flags & VM_SHARED))
2705 if (start < vma->vm_start)
2708 if (start + size > vma->vm_end) {
2709 struct vm_area_struct *next;
2711 for (next = vma->vm_next; next; next = next->vm_next) {
2712 /* hole between vmas ? */
2713 if (next->vm_start != next->vm_prev->vm_end)
2716 if (next->vm_file != vma->vm_file)
2719 if (next->vm_flags != vma->vm_flags)
2722 if (start + size <= next->vm_end)
2730 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2731 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2732 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2734 flags &= MAP_NONBLOCK;
2735 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2736 if (vma->vm_flags & VM_LOCKED) {
2737 struct vm_area_struct *tmp;
2738 flags |= MAP_LOCKED;
2740 /* drop PG_Mlocked flag for over-mapped range */
2741 for (tmp = vma; tmp->vm_start >= start + size;
2742 tmp = tmp->vm_next) {
2743 munlock_vma_pages_range(tmp,
2744 max(tmp->vm_start, start),
2745 min(tmp->vm_end, start + size));
2749 file = get_file(vma->vm_file);
2750 ret = do_mmap_pgoff(vma->vm_file, start, size,
2751 prot, flags, pgoff, &populate);
2754 up_write(&mm->mmap_sem);
2756 mm_populate(ret, populate);
2757 if (!IS_ERR_VALUE(ret))
2762 static inline void verify_mm_writelocked(struct mm_struct *mm)
2764 #ifdef CONFIG_DEBUG_VM
2765 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2767 up_read(&mm->mmap_sem);
2773 * this is really a simplified "do_mmap". it only handles
2774 * anonymous maps. eventually we may be able to do some
2775 * brk-specific accounting here.
2777 static unsigned long do_brk(unsigned long addr, unsigned long len)
2779 struct mm_struct *mm = current->mm;
2780 struct vm_area_struct *vma, *prev;
2781 unsigned long flags;
2782 struct rb_node **rb_link, *rb_parent;
2783 pgoff_t pgoff = addr >> PAGE_SHIFT;
2786 len = PAGE_ALIGN(len);
2790 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2792 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2793 if (offset_in_page(error))
2796 error = mlock_future_check(mm, mm->def_flags, len);
2801 * mm->mmap_sem is required to protect against another thread
2802 * changing the mappings in case we sleep.
2804 verify_mm_writelocked(mm);
2807 * Clear old maps. this also does some error checking for us
2809 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2811 if (do_munmap(mm, addr, len))
2815 /* Check against address space limits *after* clearing old maps... */
2816 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2819 if (mm->map_count > sysctl_max_map_count)
2822 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2825 /* Can we just expand an old private anonymous mapping? */
2826 vma = vma_merge(mm, prev, addr, addr + len, flags,
2827 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2832 * create a vma struct for an anonymous mapping
2834 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2836 vm_unacct_memory(len >> PAGE_SHIFT);
2840 INIT_LIST_HEAD(&vma->anon_vma_chain);
2842 vma->vm_start = addr;
2843 vma->vm_end = addr + len;
2844 vma->vm_pgoff = pgoff;
2845 vma->vm_flags = flags;
2846 vma->vm_page_prot = vm_get_page_prot(flags);
2847 vma_link(mm, vma, prev, rb_link, rb_parent);
2849 perf_event_mmap(vma);
2850 mm->total_vm += len >> PAGE_SHIFT;
2851 if (flags & VM_LOCKED)
2852 mm->locked_vm += (len >> PAGE_SHIFT);
2853 vma->vm_flags |= VM_SOFTDIRTY;
2857 unsigned long vm_brk(unsigned long addr, unsigned long len)
2859 struct mm_struct *mm = current->mm;
2863 down_write(&mm->mmap_sem);
2864 ret = do_brk(addr, len);
2865 populate = ((mm->def_flags & VM_LOCKED) != 0);
2866 up_write(&mm->mmap_sem);
2868 mm_populate(addr, len);
2871 EXPORT_SYMBOL(vm_brk);
2873 /* Release all mmaps. */
2874 void exit_mmap(struct mm_struct *mm)
2876 struct mmu_gather tlb;
2877 struct vm_area_struct *vma;
2878 unsigned long nr_accounted = 0;
2880 /* mm's last user has gone, and its about to be pulled down */
2881 mmu_notifier_release(mm);
2883 if (mm->locked_vm) {
2886 if (vma->vm_flags & VM_LOCKED)
2887 munlock_vma_pages_all(vma);
2895 if (!vma) /* Can happen if dup_mmap() received an OOM */
2900 tlb_gather_mmu(&tlb, mm, 0, -1);
2901 /* update_hiwater_rss(mm) here? but nobody should be looking */
2902 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2903 unmap_vmas(&tlb, vma, 0, -1);
2905 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
2906 tlb_finish_mmu(&tlb, 0, -1);
2909 * Walk the list again, actually closing and freeing it,
2910 * with preemption enabled, without holding any MM locks.
2913 if (vma->vm_flags & VM_ACCOUNT)
2914 nr_accounted += vma_pages(vma);
2915 vma = remove_vma(vma);
2917 vm_unacct_memory(nr_accounted);
2920 /* Insert vm structure into process list sorted by address
2921 * and into the inode's i_mmap tree. If vm_file is non-NULL
2922 * then i_mmap_rwsem is taken here.
2924 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2926 struct vm_area_struct *prev;
2927 struct rb_node **rb_link, *rb_parent;
2929 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2930 &prev, &rb_link, &rb_parent))
2932 if ((vma->vm_flags & VM_ACCOUNT) &&
2933 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2937 * The vm_pgoff of a purely anonymous vma should be irrelevant
2938 * until its first write fault, when page's anon_vma and index
2939 * are set. But now set the vm_pgoff it will almost certainly
2940 * end up with (unless mremap moves it elsewhere before that
2941 * first wfault), so /proc/pid/maps tells a consistent story.
2943 * By setting it to reflect the virtual start address of the
2944 * vma, merges and splits can happen in a seamless way, just
2945 * using the existing file pgoff checks and manipulations.
2946 * Similarly in do_mmap_pgoff and in do_brk.
2948 if (vma_is_anonymous(vma)) {
2949 BUG_ON(vma->anon_vma);
2950 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2953 vma_link(mm, vma, prev, rb_link, rb_parent);
2958 * Copy the vma structure to a new location in the same mm,
2959 * prior to moving page table entries, to effect an mremap move.
2961 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2962 unsigned long addr, unsigned long len, pgoff_t pgoff,
2963 bool *need_rmap_locks)
2965 struct vm_area_struct *vma = *vmap;
2966 unsigned long vma_start = vma->vm_start;
2967 struct mm_struct *mm = vma->vm_mm;
2968 struct vm_area_struct *new_vma, *prev;
2969 struct rb_node **rb_link, *rb_parent;
2970 bool faulted_in_anon_vma = true;
2973 * If anonymous vma has not yet been faulted, update new pgoff
2974 * to match new location, to increase its chance of merging.
2976 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
2977 pgoff = addr >> PAGE_SHIFT;
2978 faulted_in_anon_vma = false;
2981 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2982 return NULL; /* should never get here */
2983 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2984 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
2985 vma->vm_userfaultfd_ctx);
2988 * Source vma may have been merged into new_vma
2990 if (unlikely(vma_start >= new_vma->vm_start &&
2991 vma_start < new_vma->vm_end)) {
2993 * The only way we can get a vma_merge with
2994 * self during an mremap is if the vma hasn't
2995 * been faulted in yet and we were allowed to
2996 * reset the dst vma->vm_pgoff to the
2997 * destination address of the mremap to allow
2998 * the merge to happen. mremap must change the
2999 * vm_pgoff linearity between src and dst vmas
3000 * (in turn preventing a vma_merge) to be
3001 * safe. It is only safe to keep the vm_pgoff
3002 * linear if there are no pages mapped yet.
3004 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3005 *vmap = vma = new_vma;
3007 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3009 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3013 new_vma->vm_start = addr;
3014 new_vma->vm_end = addr + len;
3015 new_vma->vm_pgoff = pgoff;
3016 if (vma_dup_policy(vma, new_vma))
3018 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3019 if (anon_vma_clone(new_vma, vma))
3020 goto out_free_mempol;
3021 if (new_vma->vm_file)
3022 get_file(new_vma->vm_file);
3023 if (new_vma->vm_ops && new_vma->vm_ops->open)
3024 new_vma->vm_ops->open(new_vma);
3025 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3026 *need_rmap_locks = false;
3031 mpol_put(vma_policy(new_vma));
3033 kmem_cache_free(vm_area_cachep, new_vma);
3039 * Return true if the calling process may expand its vm space by the passed
3042 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
3044 unsigned long cur = mm->total_vm; /* pages */
3047 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
3049 if (cur + npages > lim)
3054 static int special_mapping_fault(struct vm_area_struct *vma,
3055 struct vm_fault *vmf);
3058 * Having a close hook prevents vma merging regardless of flags.
3060 static void special_mapping_close(struct vm_area_struct *vma)
3064 static const char *special_mapping_name(struct vm_area_struct *vma)
3066 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3069 static const struct vm_operations_struct special_mapping_vmops = {
3070 .close = special_mapping_close,
3071 .fault = special_mapping_fault,
3072 .name = special_mapping_name,
3075 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3076 .close = special_mapping_close,
3077 .fault = special_mapping_fault,
3080 static int special_mapping_fault(struct vm_area_struct *vma,
3081 struct vm_fault *vmf)
3084 struct page **pages;
3086 if (vma->vm_ops == &legacy_special_mapping_vmops)
3087 pages = vma->vm_private_data;
3089 pages = ((struct vm_special_mapping *)vma->vm_private_data)->
3092 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3096 struct page *page = *pages;
3102 return VM_FAULT_SIGBUS;
3105 static struct vm_area_struct *__install_special_mapping(
3106 struct mm_struct *mm,
3107 unsigned long addr, unsigned long len,
3108 unsigned long vm_flags, void *priv,
3109 const struct vm_operations_struct *ops)
3112 struct vm_area_struct *vma;
3114 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3115 if (unlikely(vma == NULL))
3116 return ERR_PTR(-ENOMEM);
3118 INIT_LIST_HEAD(&vma->anon_vma_chain);
3120 vma->vm_start = addr;
3121 vma->vm_end = addr + len;
3123 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3124 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3127 vma->vm_private_data = priv;
3129 ret = insert_vm_struct(mm, vma);
3133 mm->total_vm += len >> PAGE_SHIFT;
3135 perf_event_mmap(vma);
3140 kmem_cache_free(vm_area_cachep, vma);
3141 return ERR_PTR(ret);
3145 * Called with mm->mmap_sem held for writing.
3146 * Insert a new vma covering the given region, with the given flags.
3147 * Its pages are supplied by the given array of struct page *.
3148 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3149 * The region past the last page supplied will always produce SIGBUS.
3150 * The array pointer and the pages it points to are assumed to stay alive
3151 * for as long as this mapping might exist.
3153 struct vm_area_struct *_install_special_mapping(
3154 struct mm_struct *mm,
3155 unsigned long addr, unsigned long len,
3156 unsigned long vm_flags, const struct vm_special_mapping *spec)
3158 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3159 &special_mapping_vmops);
3162 int install_special_mapping(struct mm_struct *mm,
3163 unsigned long addr, unsigned long len,
3164 unsigned long vm_flags, struct page **pages)
3166 struct vm_area_struct *vma = __install_special_mapping(
3167 mm, addr, len, vm_flags, (void *)pages,
3168 &legacy_special_mapping_vmops);
3170 return PTR_ERR_OR_ZERO(vma);
3173 static DEFINE_MUTEX(mm_all_locks_mutex);
3175 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3177 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3179 * The LSB of head.next can't change from under us
3180 * because we hold the mm_all_locks_mutex.
3182 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3184 * We can safely modify head.next after taking the
3185 * anon_vma->root->rwsem. If some other vma in this mm shares
3186 * the same anon_vma we won't take it again.
3188 * No need of atomic instructions here, head.next
3189 * can't change from under us thanks to the
3190 * anon_vma->root->rwsem.
3192 if (__test_and_set_bit(0, (unsigned long *)
3193 &anon_vma->root->rb_root.rb_node))
3198 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3200 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3202 * AS_MM_ALL_LOCKS can't change from under us because
3203 * we hold the mm_all_locks_mutex.
3205 * Operations on ->flags have to be atomic because
3206 * even if AS_MM_ALL_LOCKS is stable thanks to the
3207 * mm_all_locks_mutex, there may be other cpus
3208 * changing other bitflags in parallel to us.
3210 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3212 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3217 * This operation locks against the VM for all pte/vma/mm related
3218 * operations that could ever happen on a certain mm. This includes
3219 * vmtruncate, try_to_unmap, and all page faults.
3221 * The caller must take the mmap_sem in write mode before calling
3222 * mm_take_all_locks(). The caller isn't allowed to release the
3223 * mmap_sem until mm_drop_all_locks() returns.
3225 * mmap_sem in write mode is required in order to block all operations
3226 * that could modify pagetables and free pages without need of
3227 * altering the vma layout. It's also needed in write mode to avoid new
3228 * anon_vmas to be associated with existing vmas.
3230 * A single task can't take more than one mm_take_all_locks() in a row
3231 * or it would deadlock.
3233 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3234 * mapping->flags avoid to take the same lock twice, if more than one
3235 * vma in this mm is backed by the same anon_vma or address_space.
3237 * We can take all the locks in random order because the VM code
3238 * taking i_mmap_rwsem or anon_vma->rwsem outside the mmap_sem never
3239 * takes more than one of them in a row. Secondly we're protected
3240 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3242 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3243 * that may have to take thousand of locks.
3245 * mm_take_all_locks() can fail if it's interrupted by signals.
3247 int mm_take_all_locks(struct mm_struct *mm)
3249 struct vm_area_struct *vma;
3250 struct anon_vma_chain *avc;
3252 BUG_ON(down_read_trylock(&mm->mmap_sem));
3254 mutex_lock(&mm_all_locks_mutex);
3256 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3257 if (signal_pending(current))
3259 if (vma->vm_file && vma->vm_file->f_mapping)
3260 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3263 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3264 if (signal_pending(current))
3267 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3268 vm_lock_anon_vma(mm, avc->anon_vma);
3274 mm_drop_all_locks(mm);
3278 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3280 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3282 * The LSB of head.next can't change to 0 from under
3283 * us because we hold the mm_all_locks_mutex.
3285 * We must however clear the bitflag before unlocking
3286 * the vma so the users using the anon_vma->rb_root will
3287 * never see our bitflag.
3289 * No need of atomic instructions here, head.next
3290 * can't change from under us until we release the
3291 * anon_vma->root->rwsem.
3293 if (!__test_and_clear_bit(0, (unsigned long *)
3294 &anon_vma->root->rb_root.rb_node))
3296 anon_vma_unlock_write(anon_vma);
3300 static void vm_unlock_mapping(struct address_space *mapping)
3302 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3304 * AS_MM_ALL_LOCKS can't change to 0 from under us
3305 * because we hold the mm_all_locks_mutex.
3307 i_mmap_unlock_write(mapping);
3308 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3315 * The mmap_sem cannot be released by the caller until
3316 * mm_drop_all_locks() returns.
3318 void mm_drop_all_locks(struct mm_struct *mm)
3320 struct vm_area_struct *vma;
3321 struct anon_vma_chain *avc;
3323 BUG_ON(down_read_trylock(&mm->mmap_sem));
3324 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3326 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3328 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3329 vm_unlock_anon_vma(avc->anon_vma);
3330 if (vma->vm_file && vma->vm_file->f_mapping)
3331 vm_unlock_mapping(vma->vm_file->f_mapping);
3334 mutex_unlock(&mm_all_locks_mutex);
3338 * initialise the VMA slab
3340 void __init mmap_init(void)
3344 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3349 * Initialise sysctl_user_reserve_kbytes.
3351 * This is intended to prevent a user from starting a single memory hogging
3352 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3355 * The default value is min(3% of free memory, 128MB)
3356 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3358 static int init_user_reserve(void)
3360 unsigned long free_kbytes;
3362 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3364 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3367 subsys_initcall(init_user_reserve);
3370 * Initialise sysctl_admin_reserve_kbytes.
3372 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3373 * to log in and kill a memory hogging process.
3375 * Systems with more than 256MB will reserve 8MB, enough to recover
3376 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3377 * only reserve 3% of free pages by default.
3379 static int init_admin_reserve(void)
3381 unsigned long free_kbytes;
3383 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3385 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3388 subsys_initcall(init_admin_reserve);
3391 * Reinititalise user and admin reserves if memory is added or removed.
3393 * The default user reserve max is 128MB, and the default max for the
3394 * admin reserve is 8MB. These are usually, but not always, enough to
3395 * enable recovery from a memory hogging process using login/sshd, a shell,
3396 * and tools like top. It may make sense to increase or even disable the
3397 * reserve depending on the existence of swap or variations in the recovery
3398 * tools. So, the admin may have changed them.
3400 * If memory is added and the reserves have been eliminated or increased above
3401 * the default max, then we'll trust the admin.
3403 * If memory is removed and there isn't enough free memory, then we
3404 * need to reset the reserves.
3406 * Otherwise keep the reserve set by the admin.
3408 static int reserve_mem_notifier(struct notifier_block *nb,
3409 unsigned long action, void *data)
3411 unsigned long tmp, free_kbytes;
3415 /* Default max is 128MB. Leave alone if modified by operator. */
3416 tmp = sysctl_user_reserve_kbytes;
3417 if (0 < tmp && tmp < (1UL << 17))
3418 init_user_reserve();
3420 /* Default max is 8MB. Leave alone if modified by operator. */
3421 tmp = sysctl_admin_reserve_kbytes;
3422 if (0 < tmp && tmp < (1UL << 13))
3423 init_admin_reserve();
3427 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3429 if (sysctl_user_reserve_kbytes > free_kbytes) {
3430 init_user_reserve();
3431 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3432 sysctl_user_reserve_kbytes);
3435 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3436 init_admin_reserve();
3437 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3438 sysctl_admin_reserve_kbytes);
3447 static struct notifier_block reserve_mem_nb = {
3448 .notifier_call = reserve_mem_notifier,
3451 static int __meminit init_reserve_notifier(void)
3453 if (register_hotmemory_notifier(&reserve_mem_nb))
3454 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3458 subsys_initcall(init_reserve_notifier);