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>
47 #include <asm/uaccess.h>
48 #include <asm/cacheflush.h>
50 #include <asm/mmu_context.h>
54 #ifndef arch_mmap_check
55 #define arch_mmap_check(addr, len, flags) (0)
58 #ifndef arch_rebalance_pgtables
59 #define arch_rebalance_pgtables(addr, len) (addr)
62 static void unmap_region(struct mm_struct *mm,
63 struct vm_area_struct *vma, struct vm_area_struct *prev,
64 unsigned long start, unsigned long end);
66 /* description of effects of mapping type and prot in current implementation.
67 * this is due to the limited x86 page protection hardware. The expected
68 * behavior is in parens:
71 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
72 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
73 * w: (no) no w: (no) no w: (yes) yes w: (no) no
74 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
76 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
77 * w: (no) no w: (no) no w: (copy) copy w: (no) no
78 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
81 pgprot_t protection_map[16] = {
82 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
83 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
86 pgprot_t vm_get_page_prot(unsigned long vm_flags)
88 return __pgprot(pgprot_val(protection_map[vm_flags &
89 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
90 pgprot_val(arch_vm_get_page_prot(vm_flags)));
92 EXPORT_SYMBOL(vm_get_page_prot);
94 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
96 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
99 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
100 void vma_set_page_prot(struct vm_area_struct *vma)
102 unsigned long vm_flags = vma->vm_flags;
104 vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
105 if (vma_wants_writenotify(vma)) {
106 vm_flags &= ~VM_SHARED;
107 vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot,
113 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS; /* heuristic overcommit */
114 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
115 unsigned long sysctl_overcommit_kbytes __read_mostly;
116 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
117 unsigned long sysctl_user_reserve_kbytes __read_mostly = 1UL << 17; /* 128MB */
118 unsigned long sysctl_admin_reserve_kbytes __read_mostly = 1UL << 13; /* 8MB */
120 * Make sure vm_committed_as in one cacheline and not cacheline shared with
121 * other variables. It can be updated by several CPUs frequently.
123 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
126 * The global memory commitment made in the system can be a metric
127 * that can be used to drive ballooning decisions when Linux is hosted
128 * as a guest. On Hyper-V, the host implements a policy engine for dynamically
129 * balancing memory across competing virtual machines that are hosted.
130 * Several metrics drive this policy engine including the guest reported
133 unsigned long vm_memory_committed(void)
135 return percpu_counter_read_positive(&vm_committed_as);
137 EXPORT_SYMBOL_GPL(vm_memory_committed);
140 * Check that a process has enough memory to allocate a new virtual
141 * mapping. 0 means there is enough memory for the allocation to
142 * succeed and -ENOMEM implies there is not.
144 * We currently support three overcommit policies, which are set via the
145 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
147 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
148 * Additional code 2002 Jul 20 by Robert Love.
150 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
152 * Note this is a helper function intended to be used by LSMs which
153 * wish to use this logic.
155 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
157 long free, allowed, reserve;
159 VM_WARN_ONCE(percpu_counter_read(&vm_committed_as) <
160 -(s64)vm_committed_as_batch * num_online_cpus(),
161 "memory commitment underflow");
163 vm_acct_memory(pages);
166 * Sometimes we want to use more memory than we have
168 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
171 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
172 free = global_page_state(NR_FREE_PAGES);
173 free += global_page_state(NR_FILE_PAGES);
176 * shmem pages shouldn't be counted as free in this
177 * case, they can't be purged, only swapped out, and
178 * that won't affect the overall amount of available
179 * memory in the system.
181 free -= global_page_state(NR_SHMEM);
183 free += get_nr_swap_pages();
186 * Any slabs which are created with the
187 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
188 * which are reclaimable, under pressure. The dentry
189 * cache and most inode caches should fall into this
191 free += global_page_state(NR_SLAB_RECLAIMABLE);
194 * Leave reserved pages. The pages are not for anonymous pages.
196 if (free <= totalreserve_pages)
199 free -= totalreserve_pages;
202 * Reserve some for root
205 free -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
213 allowed = vm_commit_limit();
215 * Reserve some for root
218 allowed -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
221 * Don't let a single process grow so big a user can't recover
224 reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10);
225 allowed -= min_t(long, mm->total_vm / 32, reserve);
228 if (percpu_counter_read_positive(&vm_committed_as) < allowed)
231 vm_unacct_memory(pages);
237 * Requires inode->i_mapping->i_mmap_rwsem
239 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
240 struct file *file, struct address_space *mapping)
242 if (vma->vm_flags & VM_DENYWRITE)
243 atomic_inc(&file_inode(file)->i_writecount);
244 if (vma->vm_flags & VM_SHARED)
245 mapping_unmap_writable(mapping);
247 flush_dcache_mmap_lock(mapping);
248 vma_interval_tree_remove(vma, &mapping->i_mmap);
249 flush_dcache_mmap_unlock(mapping);
253 * Unlink a file-based vm structure from its interval tree, to hide
254 * vma from rmap and vmtruncate before freeing its page tables.
256 void unlink_file_vma(struct vm_area_struct *vma)
258 struct file *file = vma->vm_file;
261 struct address_space *mapping = file->f_mapping;
262 i_mmap_lock_write(mapping);
263 __remove_shared_vm_struct(vma, file, mapping);
264 i_mmap_unlock_write(mapping);
269 * Close a vm structure and free it, returning the next.
271 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
273 struct vm_area_struct *next = vma->vm_next;
276 if (vma->vm_ops && vma->vm_ops->close)
277 vma->vm_ops->close(vma);
280 mpol_put(vma_policy(vma));
281 kmem_cache_free(vm_area_cachep, vma);
285 static unsigned long do_brk(unsigned long addr, unsigned long len);
287 SYSCALL_DEFINE1(brk, unsigned long, brk)
289 unsigned long retval;
290 unsigned long newbrk, oldbrk;
291 struct mm_struct *mm = current->mm;
292 struct vm_area_struct *next;
293 unsigned long min_brk;
296 down_write(&mm->mmap_sem);
298 #ifdef CONFIG_COMPAT_BRK
300 * CONFIG_COMPAT_BRK can still be overridden by setting
301 * randomize_va_space to 2, which will still cause mm->start_brk
302 * to be arbitrarily shifted
304 if (current->brk_randomized)
305 min_brk = mm->start_brk;
307 min_brk = mm->end_data;
309 min_brk = mm->start_brk;
315 * Check against rlimit here. If this check is done later after the test
316 * of oldbrk with newbrk then it can escape the test and let the data
317 * segment grow beyond its set limit the in case where the limit is
318 * not page aligned -Ram Gupta
320 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
321 mm->end_data, mm->start_data))
324 newbrk = PAGE_ALIGN(brk);
325 oldbrk = PAGE_ALIGN(mm->brk);
326 if (oldbrk == newbrk)
329 /* Always allow shrinking brk. */
330 if (brk <= mm->brk) {
331 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
336 /* Check against existing mmap mappings. */
337 next = find_vma(mm, oldbrk);
338 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
341 /* Ok, looks good - let it rip. */
342 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
347 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
348 up_write(&mm->mmap_sem);
350 mm_populate(oldbrk, newbrk - oldbrk);
355 up_write(&mm->mmap_sem);
359 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
361 unsigned long max, prev_end, subtree_gap;
364 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
365 * allow two stack_guard_gaps between them here, and when choosing
366 * an unmapped area; whereas when expanding we only require one.
367 * That's a little inconsistent, but keeps the code here simpler.
369 max = vm_start_gap(vma);
371 prev_end = vm_end_gap(vma->vm_prev);
377 if (vma->vm_rb.rb_left) {
378 subtree_gap = rb_entry(vma->vm_rb.rb_left,
379 struct vm_area_struct, vm_rb)->rb_subtree_gap;
380 if (subtree_gap > max)
383 if (vma->vm_rb.rb_right) {
384 subtree_gap = rb_entry(vma->vm_rb.rb_right,
385 struct vm_area_struct, vm_rb)->rb_subtree_gap;
386 if (subtree_gap > max)
392 #ifdef CONFIG_DEBUG_VM_RB
393 static int browse_rb(struct rb_root *root)
395 int i = 0, j, bug = 0;
396 struct rb_node *nd, *pn = NULL;
397 unsigned long prev = 0, pend = 0;
399 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
400 struct vm_area_struct *vma;
401 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
402 if (vma->vm_start < prev) {
403 pr_emerg("vm_start %lx < prev %lx\n",
404 vma->vm_start, prev);
407 if (vma->vm_start < pend) {
408 pr_emerg("vm_start %lx < pend %lx\n",
409 vma->vm_start, pend);
412 if (vma->vm_start > vma->vm_end) {
413 pr_emerg("vm_start %lx > vm_end %lx\n",
414 vma->vm_start, vma->vm_end);
417 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
418 pr_emerg("free gap %lx, correct %lx\n",
420 vma_compute_subtree_gap(vma));
425 prev = vma->vm_start;
429 for (nd = pn; nd; nd = rb_prev(nd))
432 pr_emerg("backwards %d, forwards %d\n", j, i);
438 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
442 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
443 struct vm_area_struct *vma;
444 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
445 VM_BUG_ON_VMA(vma != ignore &&
446 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
451 static void validate_mm(struct mm_struct *mm)
455 unsigned long highest_address = 0;
456 struct vm_area_struct *vma = mm->mmap;
459 struct anon_vma *anon_vma = vma->anon_vma;
460 struct anon_vma_chain *avc;
463 anon_vma_lock_read(anon_vma);
464 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
465 anon_vma_interval_tree_verify(avc);
466 anon_vma_unlock_read(anon_vma);
469 highest_address = vm_end_gap(vma);
473 if (i != mm->map_count) {
474 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
477 if (highest_address != mm->highest_vm_end) {
478 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
479 mm->highest_vm_end, highest_address);
482 i = browse_rb(&mm->mm_rb);
483 if (i != mm->map_count) {
485 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
488 VM_BUG_ON_MM(bug, mm);
491 #define validate_mm_rb(root, ignore) do { } while (0)
492 #define validate_mm(mm) do { } while (0)
495 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
496 unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
499 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
500 * vma->vm_prev->vm_end values changed, without modifying the vma's position
503 static void vma_gap_update(struct vm_area_struct *vma)
506 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
507 * function that does exacltly what we want.
509 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
512 static inline void vma_rb_insert(struct vm_area_struct *vma,
513 struct rb_root *root)
515 /* All rb_subtree_gap values must be consistent prior to insertion */
516 validate_mm_rb(root, NULL);
518 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
521 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
524 * All rb_subtree_gap values must be consistent prior to erase,
525 * with the possible exception of the vma being erased.
527 validate_mm_rb(root, vma);
530 * Note rb_erase_augmented is a fairly large inline function,
531 * so make sure we instantiate it only once with our desired
532 * augmented rbtree callbacks.
534 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
538 * vma has some anon_vma assigned, and is already inserted on that
539 * anon_vma's interval trees.
541 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
542 * vma must be removed from the anon_vma's interval trees using
543 * anon_vma_interval_tree_pre_update_vma().
545 * After the update, the vma will be reinserted using
546 * anon_vma_interval_tree_post_update_vma().
548 * The entire update must be protected by exclusive mmap_sem and by
549 * the root anon_vma's mutex.
552 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
554 struct anon_vma_chain *avc;
556 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
557 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
561 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
563 struct anon_vma_chain *avc;
565 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
566 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
569 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
570 unsigned long end, struct vm_area_struct **pprev,
571 struct rb_node ***rb_link, struct rb_node **rb_parent)
573 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
575 __rb_link = &mm->mm_rb.rb_node;
576 rb_prev = __rb_parent = NULL;
579 struct vm_area_struct *vma_tmp;
581 __rb_parent = *__rb_link;
582 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
584 if (vma_tmp->vm_end > addr) {
585 /* Fail if an existing vma overlaps the area */
586 if (vma_tmp->vm_start < end)
588 __rb_link = &__rb_parent->rb_left;
590 rb_prev = __rb_parent;
591 __rb_link = &__rb_parent->rb_right;
597 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
598 *rb_link = __rb_link;
599 *rb_parent = __rb_parent;
603 static unsigned long count_vma_pages_range(struct mm_struct *mm,
604 unsigned long addr, unsigned long end)
606 unsigned long nr_pages = 0;
607 struct vm_area_struct *vma;
609 /* Find first overlaping mapping */
610 vma = find_vma_intersection(mm, addr, end);
614 nr_pages = (min(end, vma->vm_end) -
615 max(addr, vma->vm_start)) >> PAGE_SHIFT;
617 /* Iterate over the rest of the overlaps */
618 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
619 unsigned long overlap_len;
621 if (vma->vm_start > end)
624 overlap_len = min(end, vma->vm_end) - vma->vm_start;
625 nr_pages += overlap_len >> PAGE_SHIFT;
631 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
632 struct rb_node **rb_link, struct rb_node *rb_parent)
634 /* Update tracking information for the gap following the new vma. */
636 vma_gap_update(vma->vm_next);
638 mm->highest_vm_end = vm_end_gap(vma);
641 * vma->vm_prev wasn't known when we followed the rbtree to find the
642 * correct insertion point for that vma. As a result, we could not
643 * update the vma vm_rb parents rb_subtree_gap values on the way down.
644 * So, we first insert the vma with a zero rb_subtree_gap value
645 * (to be consistent with what we did on the way down), and then
646 * immediately update the gap to the correct value. Finally we
647 * rebalance the rbtree after all augmented values have been set.
649 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
650 vma->rb_subtree_gap = 0;
652 vma_rb_insert(vma, &mm->mm_rb);
655 static void __vma_link_file(struct vm_area_struct *vma)
661 struct address_space *mapping = file->f_mapping;
663 if (vma->vm_flags & VM_DENYWRITE)
664 atomic_dec(&file_inode(file)->i_writecount);
665 if (vma->vm_flags & VM_SHARED)
666 atomic_inc(&mapping->i_mmap_writable);
668 flush_dcache_mmap_lock(mapping);
669 vma_interval_tree_insert(vma, &mapping->i_mmap);
670 flush_dcache_mmap_unlock(mapping);
675 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
676 struct vm_area_struct *prev, struct rb_node **rb_link,
677 struct rb_node *rb_parent)
679 __vma_link_list(mm, vma, prev, rb_parent);
680 __vma_link_rb(mm, vma, rb_link, rb_parent);
683 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
684 struct vm_area_struct *prev, struct rb_node **rb_link,
685 struct rb_node *rb_parent)
687 struct address_space *mapping = NULL;
690 mapping = vma->vm_file->f_mapping;
691 i_mmap_lock_write(mapping);
694 __vma_link(mm, vma, prev, rb_link, rb_parent);
695 __vma_link_file(vma);
698 i_mmap_unlock_write(mapping);
705 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
706 * mm's list and rbtree. It has already been inserted into the interval tree.
708 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
710 struct vm_area_struct *prev;
711 struct rb_node **rb_link, *rb_parent;
713 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
714 &prev, &rb_link, &rb_parent))
716 __vma_link(mm, vma, prev, rb_link, rb_parent);
721 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
722 struct vm_area_struct *prev)
724 struct vm_area_struct *next;
726 vma_rb_erase(vma, &mm->mm_rb);
727 prev->vm_next = next = vma->vm_next;
729 next->vm_prev = prev;
732 vmacache_invalidate(mm);
736 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
737 * is already present in an i_mmap tree without adjusting the tree.
738 * The following helper function should be used when such adjustments
739 * are necessary. The "insert" vma (if any) is to be inserted
740 * before we drop the necessary locks.
742 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
743 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
745 struct mm_struct *mm = vma->vm_mm;
746 struct vm_area_struct *next = vma->vm_next;
747 struct vm_area_struct *importer = NULL;
748 struct address_space *mapping = NULL;
749 struct rb_root *root = NULL;
750 struct anon_vma *anon_vma = NULL;
751 struct file *file = vma->vm_file;
752 bool start_changed = false, end_changed = false;
753 long adjust_next = 0;
756 if (next && !insert) {
757 struct vm_area_struct *exporter = NULL;
759 if (end >= next->vm_end) {
761 * vma expands, overlapping all the next, and
762 * perhaps the one after too (mprotect case 6).
764 again: remove_next = 1 + (end > next->vm_end);
768 } else if (end > next->vm_start) {
770 * vma expands, overlapping part of the next:
771 * mprotect case 5 shifting the boundary up.
773 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
776 } else if (end < vma->vm_end) {
778 * vma shrinks, and !insert tells it's not
779 * split_vma inserting another: so it must be
780 * mprotect case 4 shifting the boundary down.
782 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
788 * Easily overlooked: when mprotect shifts the boundary,
789 * make sure the expanding vma has anon_vma set if the
790 * shrinking vma had, to cover any anon pages imported.
792 if (exporter && exporter->anon_vma && !importer->anon_vma) {
795 importer->anon_vma = exporter->anon_vma;
796 error = anon_vma_clone(importer, exporter);
803 mapping = file->f_mapping;
804 root = &mapping->i_mmap;
805 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
808 uprobe_munmap(next, next->vm_start, next->vm_end);
810 i_mmap_lock_write(mapping);
813 * Put into interval tree now, so instantiated pages
814 * are visible to arm/parisc __flush_dcache_page
815 * throughout; but we cannot insert into address
816 * space until vma start or end is updated.
818 __vma_link_file(insert);
822 vma_adjust_trans_huge(vma, start, end, adjust_next);
824 anon_vma = vma->anon_vma;
825 if (!anon_vma && adjust_next)
826 anon_vma = next->anon_vma;
828 VM_BUG_ON_VMA(adjust_next && next->anon_vma &&
829 anon_vma != next->anon_vma, next);
830 anon_vma_lock_write(anon_vma);
831 anon_vma_interval_tree_pre_update_vma(vma);
833 anon_vma_interval_tree_pre_update_vma(next);
837 flush_dcache_mmap_lock(mapping);
838 vma_interval_tree_remove(vma, root);
840 vma_interval_tree_remove(next, root);
843 if (start != vma->vm_start) {
844 vma->vm_start = start;
845 start_changed = true;
847 if (end != vma->vm_end) {
851 vma->vm_pgoff = pgoff;
853 next->vm_start += adjust_next << PAGE_SHIFT;
854 next->vm_pgoff += adjust_next;
859 vma_interval_tree_insert(next, root);
860 vma_interval_tree_insert(vma, root);
861 flush_dcache_mmap_unlock(mapping);
866 * vma_merge has merged next into vma, and needs
867 * us to remove next before dropping the locks.
869 __vma_unlink(mm, next, vma);
871 __remove_shared_vm_struct(next, file, mapping);
874 * split_vma has split insert from vma, and needs
875 * us to insert it before dropping the locks
876 * (it may either follow vma or precede it).
878 __insert_vm_struct(mm, insert);
884 mm->highest_vm_end = vm_end_gap(vma);
885 else if (!adjust_next)
886 vma_gap_update(next);
891 anon_vma_interval_tree_post_update_vma(vma);
893 anon_vma_interval_tree_post_update_vma(next);
894 anon_vma_unlock_write(anon_vma);
897 i_mmap_unlock_write(mapping);
908 uprobe_munmap(next, next->vm_start, next->vm_end);
912 anon_vma_merge(vma, next);
914 mpol_put(vma_policy(next));
915 kmem_cache_free(vm_area_cachep, next);
917 * In mprotect's case 6 (see comments on vma_merge),
918 * we must remove another next too. It would clutter
919 * up the code too much to do both in one go.
922 if (remove_next == 2)
925 vma_gap_update(next);
927 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
938 * If the vma has a ->close operation then the driver probably needs to release
939 * per-vma resources, so we don't attempt to merge those.
941 static inline int is_mergeable_vma(struct vm_area_struct *vma,
942 struct file *file, unsigned long vm_flags,
943 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
946 * VM_SOFTDIRTY should not prevent from VMA merging, if we
947 * match the flags but dirty bit -- the caller should mark
948 * merged VMA as dirty. If dirty bit won't be excluded from
949 * comparison, we increase pressue on the memory system forcing
950 * the kernel to generate new VMAs when old one could be
953 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
955 if (vma->vm_file != file)
957 if (vma->vm_ops && vma->vm_ops->close)
959 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
964 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
965 struct anon_vma *anon_vma2,
966 struct vm_area_struct *vma)
969 * The list_is_singular() test is to avoid merging VMA cloned from
970 * parents. This can improve scalability caused by anon_vma lock.
972 if ((!anon_vma1 || !anon_vma2) && (!vma ||
973 list_is_singular(&vma->anon_vma_chain)))
975 return anon_vma1 == anon_vma2;
979 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
980 * in front of (at a lower virtual address and file offset than) the vma.
982 * We cannot merge two vmas if they have differently assigned (non-NULL)
983 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
985 * We don't check here for the merged mmap wrapping around the end of pagecache
986 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
987 * wrap, nor mmaps which cover the final page at index -1UL.
990 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
991 struct anon_vma *anon_vma, struct file *file,
993 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
995 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
996 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
997 if (vma->vm_pgoff == vm_pgoff)
1004 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1005 * beyond (at a higher virtual address and file offset than) the vma.
1007 * We cannot merge two vmas if they have differently assigned (non-NULL)
1008 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1011 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1012 struct anon_vma *anon_vma, struct file *file,
1014 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1016 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1017 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1019 vm_pglen = vma_pages(vma);
1020 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1027 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1028 * whether that can be merged with its predecessor or its successor.
1029 * Or both (it neatly fills a hole).
1031 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1032 * certain not to be mapped by the time vma_merge is called; but when
1033 * called for mprotect, it is certain to be already mapped (either at
1034 * an offset within prev, or at the start of next), and the flags of
1035 * this area are about to be changed to vm_flags - and the no-change
1036 * case has already been eliminated.
1038 * The following mprotect cases have to be considered, where AAAA is
1039 * the area passed down from mprotect_fixup, never extending beyond one
1040 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1042 * AAAA AAAA AAAA AAAA
1043 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1044 * cannot merge might become might become might become
1045 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1046 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1047 * mremap move: PPPPNNNNNNNN 8
1049 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1050 * might become case 1 below case 2 below case 3 below
1052 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
1053 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
1055 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1056 struct vm_area_struct *prev, unsigned long addr,
1057 unsigned long end, unsigned long vm_flags,
1058 struct anon_vma *anon_vma, struct file *file,
1059 pgoff_t pgoff, struct mempolicy *policy,
1060 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1062 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1063 struct vm_area_struct *area, *next;
1067 * We later require that vma->vm_flags == vm_flags,
1068 * so this tests vma->vm_flags & VM_SPECIAL, too.
1070 if (vm_flags & VM_SPECIAL)
1074 next = prev->vm_next;
1078 if (next && next->vm_end == end) /* cases 6, 7, 8 */
1079 next = next->vm_next;
1082 * Can it merge with the predecessor?
1084 if (prev && prev->vm_end == addr &&
1085 mpol_equal(vma_policy(prev), policy) &&
1086 can_vma_merge_after(prev, vm_flags,
1087 anon_vma, file, pgoff,
1088 vm_userfaultfd_ctx)) {
1090 * OK, it can. Can we now merge in the successor as well?
1092 if (next && end == next->vm_start &&
1093 mpol_equal(policy, vma_policy(next)) &&
1094 can_vma_merge_before(next, vm_flags,
1097 vm_userfaultfd_ctx) &&
1098 is_mergeable_anon_vma(prev->anon_vma,
1099 next->anon_vma, NULL)) {
1101 err = vma_adjust(prev, prev->vm_start,
1102 next->vm_end, prev->vm_pgoff, NULL);
1103 } else /* cases 2, 5, 7 */
1104 err = vma_adjust(prev, prev->vm_start,
1105 end, prev->vm_pgoff, NULL);
1108 khugepaged_enter_vma_merge(prev, vm_flags);
1113 * Can this new request be merged in front of next?
1115 if (next && end == next->vm_start &&
1116 mpol_equal(policy, vma_policy(next)) &&
1117 can_vma_merge_before(next, vm_flags,
1118 anon_vma, file, pgoff+pglen,
1119 vm_userfaultfd_ctx)) {
1120 if (prev && addr < prev->vm_end) /* case 4 */
1121 err = vma_adjust(prev, prev->vm_start,
1122 addr, prev->vm_pgoff, NULL);
1123 else /* cases 3, 8 */
1124 err = vma_adjust(area, addr, next->vm_end,
1125 next->vm_pgoff - pglen, NULL);
1128 khugepaged_enter_vma_merge(area, vm_flags);
1136 * Rough compatbility check to quickly see if it's even worth looking
1137 * at sharing an anon_vma.
1139 * They need to have the same vm_file, and the flags can only differ
1140 * in things that mprotect may change.
1142 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1143 * we can merge the two vma's. For example, we refuse to merge a vma if
1144 * there is a vm_ops->close() function, because that indicates that the
1145 * driver is doing some kind of reference counting. But that doesn't
1146 * really matter for the anon_vma sharing case.
1148 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1150 return a->vm_end == b->vm_start &&
1151 mpol_equal(vma_policy(a), vma_policy(b)) &&
1152 a->vm_file == b->vm_file &&
1153 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1154 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1158 * Do some basic sanity checking to see if we can re-use the anon_vma
1159 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1160 * the same as 'old', the other will be the new one that is trying
1161 * to share the anon_vma.
1163 * NOTE! This runs with mm_sem held for reading, so it is possible that
1164 * the anon_vma of 'old' is concurrently in the process of being set up
1165 * by another page fault trying to merge _that_. But that's ok: if it
1166 * is being set up, that automatically means that it will be a singleton
1167 * acceptable for merging, so we can do all of this optimistically. But
1168 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1170 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1171 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1172 * is to return an anon_vma that is "complex" due to having gone through
1175 * We also make sure that the two vma's are compatible (adjacent,
1176 * and with the same memory policies). That's all stable, even with just
1177 * a read lock on the mm_sem.
1179 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1181 if (anon_vma_compatible(a, b)) {
1182 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1184 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1191 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1192 * neighbouring vmas for a suitable anon_vma, before it goes off
1193 * to allocate a new anon_vma. It checks because a repetitive
1194 * sequence of mprotects and faults may otherwise lead to distinct
1195 * anon_vmas being allocated, preventing vma merge in subsequent
1198 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1200 struct anon_vma *anon_vma;
1201 struct vm_area_struct *near;
1203 near = vma->vm_next;
1207 anon_vma = reusable_anon_vma(near, vma, near);
1211 near = vma->vm_prev;
1215 anon_vma = reusable_anon_vma(near, near, vma);
1220 * There's no absolute need to look only at touching neighbours:
1221 * we could search further afield for "compatible" anon_vmas.
1222 * But it would probably just be a waste of time searching,
1223 * or lead to too many vmas hanging off the same anon_vma.
1224 * We're trying to allow mprotect remerging later on,
1225 * not trying to minimize memory used for anon_vmas.
1230 #ifdef CONFIG_PROC_FS
1231 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
1232 struct file *file, long pages)
1234 const unsigned long stack_flags
1235 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
1237 mm->total_vm += pages;
1240 mm->shared_vm += pages;
1241 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
1242 mm->exec_vm += pages;
1243 } else if (flags & stack_flags)
1244 mm->stack_vm += pages;
1246 #endif /* CONFIG_PROC_FS */
1249 * If a hint addr is less than mmap_min_addr change hint to be as
1250 * low as possible but still greater than mmap_min_addr
1252 static inline unsigned long round_hint_to_min(unsigned long hint)
1255 if (((void *)hint != NULL) &&
1256 (hint < mmap_min_addr))
1257 return PAGE_ALIGN(mmap_min_addr);
1261 static inline int mlock_future_check(struct mm_struct *mm,
1262 unsigned long flags,
1265 unsigned long locked, lock_limit;
1267 /* mlock MCL_FUTURE? */
1268 if (flags & VM_LOCKED) {
1269 locked = len >> PAGE_SHIFT;
1270 locked += mm->locked_vm;
1271 lock_limit = rlimit(RLIMIT_MEMLOCK);
1272 lock_limit >>= PAGE_SHIFT;
1273 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1279 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1281 if (S_ISREG(inode->i_mode))
1282 return MAX_LFS_FILESIZE;
1284 if (S_ISBLK(inode->i_mode))
1285 return MAX_LFS_FILESIZE;
1287 /* Special "we do even unsigned file positions" case */
1288 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1291 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1295 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1296 unsigned long pgoff, unsigned long len)
1298 u64 maxsize = file_mmap_size_max(file, inode);
1300 if (maxsize && len > maxsize)
1303 if (pgoff > maxsize >> PAGE_SHIFT)
1309 * The caller must hold down_write(¤t->mm->mmap_sem).
1311 unsigned long do_mmap(struct file *file, unsigned long addr,
1312 unsigned long len, unsigned long prot,
1313 unsigned long flags, vm_flags_t vm_flags,
1314 unsigned long pgoff, unsigned long *populate)
1316 struct mm_struct *mm = current->mm;
1324 * Does the application expect PROT_READ to imply PROT_EXEC?
1326 * (the exception is when the underlying filesystem is noexec
1327 * mounted, in which case we dont add PROT_EXEC.)
1329 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1330 if (!(file && path_noexec(&file->f_path)))
1333 if (!(flags & MAP_FIXED))
1334 addr = round_hint_to_min(addr);
1336 /* Careful about overflows.. */
1337 len = PAGE_ALIGN(len);
1341 /* offset overflow? */
1342 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1345 /* Too many mappings? */
1346 if (mm->map_count > sysctl_max_map_count)
1349 /* Obtain the address to map to. we verify (or select) it and ensure
1350 * that it represents a valid section of the address space.
1352 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1353 if (offset_in_page(addr))
1356 /* Do simple checking here so the lower-level routines won't have
1357 * to. we assume access permissions have been handled by the open
1358 * of the memory object, so we don't do any here.
1360 vm_flags |= calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1361 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1363 if (flags & MAP_LOCKED)
1364 if (!can_do_mlock())
1367 if (mlock_future_check(mm, vm_flags, len))
1371 struct inode *inode = file_inode(file);
1373 if (!file_mmap_ok(file, inode, pgoff, len))
1376 switch (flags & MAP_TYPE) {
1378 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1382 * Make sure we don't allow writing to an append-only
1385 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1389 * Make sure there are no mandatory locks on the file.
1391 if (locks_verify_locked(file))
1394 vm_flags |= VM_SHARED | VM_MAYSHARE;
1395 if (!(file->f_mode & FMODE_WRITE))
1396 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1400 if (!(file->f_mode & FMODE_READ))
1402 if (path_noexec(&file->f_path)) {
1403 if (vm_flags & VM_EXEC)
1405 vm_flags &= ~VM_MAYEXEC;
1408 if (!file->f_op->mmap)
1410 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1418 switch (flags & MAP_TYPE) {
1420 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1426 vm_flags |= VM_SHARED | VM_MAYSHARE;
1430 * Set pgoff according to addr for anon_vma.
1432 pgoff = addr >> PAGE_SHIFT;
1440 * Set 'VM_NORESERVE' if we should not account for the
1441 * memory use of this mapping.
1443 if (flags & MAP_NORESERVE) {
1444 /* We honor MAP_NORESERVE if allowed to overcommit */
1445 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1446 vm_flags |= VM_NORESERVE;
1448 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1449 if (file && is_file_hugepages(file))
1450 vm_flags |= VM_NORESERVE;
1453 addr = mmap_region(file, addr, len, vm_flags, pgoff);
1454 if (!IS_ERR_VALUE(addr) &&
1455 ((vm_flags & VM_LOCKED) ||
1456 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1461 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1462 unsigned long, prot, unsigned long, flags,
1463 unsigned long, fd, unsigned long, pgoff)
1465 struct file *file = NULL;
1466 unsigned long retval;
1468 if (!(flags & MAP_ANONYMOUS)) {
1469 audit_mmap_fd(fd, flags);
1473 if (is_file_hugepages(file))
1474 len = ALIGN(len, huge_page_size(hstate_file(file)));
1476 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1478 } else if (flags & MAP_HUGETLB) {
1479 struct user_struct *user = NULL;
1482 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1486 len = ALIGN(len, huge_page_size(hs));
1488 * VM_NORESERVE is used because the reservations will be
1489 * taken when vm_ops->mmap() is called
1490 * A dummy user value is used because we are not locking
1491 * memory so no accounting is necessary
1493 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1495 &user, HUGETLB_ANONHUGE_INODE,
1496 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1498 return PTR_ERR(file);
1501 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1503 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1510 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1511 struct mmap_arg_struct {
1515 unsigned long flags;
1517 unsigned long offset;
1520 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1522 struct mmap_arg_struct a;
1524 if (copy_from_user(&a, arg, sizeof(a)))
1526 if (offset_in_page(a.offset))
1529 return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1530 a.offset >> PAGE_SHIFT);
1532 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1535 * Some shared mappigns will want the pages marked read-only
1536 * to track write events. If so, we'll downgrade vm_page_prot
1537 * to the private version (using protection_map[] without the
1540 int vma_wants_writenotify(struct vm_area_struct *vma)
1542 vm_flags_t vm_flags = vma->vm_flags;
1543 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1545 /* If it was private or non-writable, the write bit is already clear */
1546 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1549 /* The backer wishes to know when pages are first written to? */
1550 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1553 /* The open routine did something to the protections that pgprot_modify
1554 * won't preserve? */
1555 if (pgprot_val(vma->vm_page_prot) !=
1556 pgprot_val(vm_pgprot_modify(vma->vm_page_prot, vm_flags)))
1559 /* Do we need to track softdirty? */
1560 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1563 /* Specialty mapping? */
1564 if (vm_flags & VM_PFNMAP)
1567 /* Can the mapping track the dirty pages? */
1568 return vma->vm_file && vma->vm_file->f_mapping &&
1569 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1573 * We account for memory if it's a private writeable mapping,
1574 * not hugepages and VM_NORESERVE wasn't set.
1576 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1579 * hugetlb has its own accounting separate from the core VM
1580 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1582 if (file && is_file_hugepages(file))
1585 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1588 unsigned long mmap_region(struct file *file, unsigned long addr,
1589 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1591 struct mm_struct *mm = current->mm;
1592 struct vm_area_struct *vma, *prev;
1594 struct rb_node **rb_link, *rb_parent;
1595 unsigned long charged = 0;
1597 /* Check against address space limit. */
1598 if (!may_expand_vm(mm, len >> PAGE_SHIFT)) {
1599 unsigned long nr_pages;
1602 * MAP_FIXED may remove pages of mappings that intersects with
1603 * requested mapping. Account for the pages it would unmap.
1605 if (!(vm_flags & MAP_FIXED))
1608 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1610 if (!may_expand_vm(mm, (len >> PAGE_SHIFT) - nr_pages))
1614 /* Clear old maps */
1615 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1617 if (do_munmap(mm, addr, len))
1622 * Private writable mapping: check memory availability
1624 if (accountable_mapping(file, vm_flags)) {
1625 charged = len >> PAGE_SHIFT;
1626 if (security_vm_enough_memory_mm(mm, charged))
1628 vm_flags |= VM_ACCOUNT;
1632 * Can we just expand an old mapping?
1634 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1635 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1640 * Determine the object being mapped and call the appropriate
1641 * specific mapper. the address has already been validated, but
1642 * not unmapped, but the maps are removed from the list.
1644 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1651 vma->vm_start = addr;
1652 vma->vm_end = addr + len;
1653 vma->vm_flags = vm_flags;
1654 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1655 vma->vm_pgoff = pgoff;
1656 INIT_LIST_HEAD(&vma->anon_vma_chain);
1659 if (vm_flags & VM_DENYWRITE) {
1660 error = deny_write_access(file);
1664 if (vm_flags & VM_SHARED) {
1665 error = mapping_map_writable(file->f_mapping);
1667 goto allow_write_and_free_vma;
1670 /* ->mmap() can change vma->vm_file, but must guarantee that
1671 * vma_link() below can deny write-access if VM_DENYWRITE is set
1672 * and map writably if VM_SHARED is set. This usually means the
1673 * new file must not have been exposed to user-space, yet.
1675 vma->vm_file = get_file(file);
1676 error = file->f_op->mmap(file, vma);
1678 goto unmap_and_free_vma;
1680 /* Can addr have changed??
1682 * Answer: Yes, several device drivers can do it in their
1683 * f_op->mmap method. -DaveM
1684 * Bug: If addr is changed, prev, rb_link, rb_parent should
1685 * be updated for vma_link()
1687 WARN_ON_ONCE(addr != vma->vm_start);
1689 addr = vma->vm_start;
1690 vm_flags = vma->vm_flags;
1691 } else if (vm_flags & VM_SHARED) {
1692 error = shmem_zero_setup(vma);
1697 vma_link(mm, vma, prev, rb_link, rb_parent);
1698 /* Once vma denies write, undo our temporary denial count */
1700 if (vm_flags & VM_SHARED)
1701 mapping_unmap_writable(file->f_mapping);
1702 if (vm_flags & VM_DENYWRITE)
1703 allow_write_access(file);
1705 file = vma->vm_file;
1707 perf_event_mmap(vma);
1709 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1710 if (vm_flags & VM_LOCKED) {
1711 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1712 vma == get_gate_vma(current->mm)))
1713 mm->locked_vm += (len >> PAGE_SHIFT);
1715 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1722 * New (or expanded) vma always get soft dirty status.
1723 * Otherwise user-space soft-dirty page tracker won't
1724 * be able to distinguish situation when vma area unmapped,
1725 * then new mapped in-place (which must be aimed as
1726 * a completely new data area).
1728 vma->vm_flags |= VM_SOFTDIRTY;
1730 vma_set_page_prot(vma);
1735 vma->vm_file = NULL;
1738 /* Undo any partial mapping done by a device driver. */
1739 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1741 if (vm_flags & VM_SHARED)
1742 mapping_unmap_writable(file->f_mapping);
1743 allow_write_and_free_vma:
1744 if (vm_flags & VM_DENYWRITE)
1745 allow_write_access(file);
1747 kmem_cache_free(vm_area_cachep, vma);
1750 vm_unacct_memory(charged);
1754 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1757 * We implement the search by looking for an rbtree node that
1758 * immediately follows a suitable gap. That is,
1759 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1760 * - gap_end = vma->vm_start >= info->low_limit + length;
1761 * - gap_end - gap_start >= length
1764 struct mm_struct *mm = current->mm;
1765 struct vm_area_struct *vma;
1766 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1768 /* Adjust search length to account for worst case alignment overhead */
1769 length = info->length + info->align_mask;
1770 if (length < info->length)
1773 /* Adjust search limits by the desired length */
1774 if (info->high_limit < length)
1776 high_limit = info->high_limit - length;
1778 if (info->low_limit > high_limit)
1780 low_limit = info->low_limit + length;
1782 /* Check if rbtree root looks promising */
1783 if (RB_EMPTY_ROOT(&mm->mm_rb))
1785 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1786 if (vma->rb_subtree_gap < length)
1790 /* Visit left subtree if it looks promising */
1791 gap_end = vm_start_gap(vma);
1792 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1793 struct vm_area_struct *left =
1794 rb_entry(vma->vm_rb.rb_left,
1795 struct vm_area_struct, vm_rb);
1796 if (left->rb_subtree_gap >= length) {
1802 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1804 /* Check if current node has a suitable gap */
1805 if (gap_start > high_limit)
1807 if (gap_end >= low_limit &&
1808 gap_end > gap_start && gap_end - gap_start >= length)
1811 /* Visit right subtree if it looks promising */
1812 if (vma->vm_rb.rb_right) {
1813 struct vm_area_struct *right =
1814 rb_entry(vma->vm_rb.rb_right,
1815 struct vm_area_struct, vm_rb);
1816 if (right->rb_subtree_gap >= length) {
1822 /* Go back up the rbtree to find next candidate node */
1824 struct rb_node *prev = &vma->vm_rb;
1825 if (!rb_parent(prev))
1827 vma = rb_entry(rb_parent(prev),
1828 struct vm_area_struct, vm_rb);
1829 if (prev == vma->vm_rb.rb_left) {
1830 gap_start = vm_end_gap(vma->vm_prev);
1831 gap_end = vm_start_gap(vma);
1838 /* Check highest gap, which does not precede any rbtree node */
1839 gap_start = mm->highest_vm_end;
1840 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1841 if (gap_start > high_limit)
1845 /* We found a suitable gap. Clip it with the original low_limit. */
1846 if (gap_start < info->low_limit)
1847 gap_start = info->low_limit;
1849 /* Adjust gap address to the desired alignment */
1850 gap_start += (info->align_offset - gap_start) & info->align_mask;
1852 VM_BUG_ON(gap_start + info->length > info->high_limit);
1853 VM_BUG_ON(gap_start + info->length > gap_end);
1857 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1859 struct mm_struct *mm = current->mm;
1860 struct vm_area_struct *vma;
1861 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1863 /* Adjust search length to account for worst case alignment overhead */
1864 length = info->length + info->align_mask;
1865 if (length < info->length)
1869 * Adjust search limits by the desired length.
1870 * See implementation comment at top of unmapped_area().
1872 gap_end = info->high_limit;
1873 if (gap_end < length)
1875 high_limit = gap_end - length;
1877 if (info->low_limit > high_limit)
1879 low_limit = info->low_limit + length;
1881 /* Check highest gap, which does not precede any rbtree node */
1882 gap_start = mm->highest_vm_end;
1883 if (gap_start <= high_limit)
1886 /* Check if rbtree root looks promising */
1887 if (RB_EMPTY_ROOT(&mm->mm_rb))
1889 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1890 if (vma->rb_subtree_gap < length)
1894 /* Visit right subtree if it looks promising */
1895 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1896 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1897 struct vm_area_struct *right =
1898 rb_entry(vma->vm_rb.rb_right,
1899 struct vm_area_struct, vm_rb);
1900 if (right->rb_subtree_gap >= length) {
1907 /* Check if current node has a suitable gap */
1908 gap_end = vm_start_gap(vma);
1909 if (gap_end < low_limit)
1911 if (gap_start <= high_limit &&
1912 gap_end > gap_start && gap_end - gap_start >= length)
1915 /* Visit left subtree if it looks promising */
1916 if (vma->vm_rb.rb_left) {
1917 struct vm_area_struct *left =
1918 rb_entry(vma->vm_rb.rb_left,
1919 struct vm_area_struct, vm_rb);
1920 if (left->rb_subtree_gap >= length) {
1926 /* Go back up the rbtree to find next candidate node */
1928 struct rb_node *prev = &vma->vm_rb;
1929 if (!rb_parent(prev))
1931 vma = rb_entry(rb_parent(prev),
1932 struct vm_area_struct, vm_rb);
1933 if (prev == vma->vm_rb.rb_right) {
1934 gap_start = vma->vm_prev ?
1935 vm_end_gap(vma->vm_prev) : 0;
1942 /* We found a suitable gap. Clip it with the original high_limit. */
1943 if (gap_end > info->high_limit)
1944 gap_end = info->high_limit;
1947 /* Compute highest gap address at the desired alignment */
1948 gap_end -= info->length;
1949 gap_end -= (gap_end - info->align_offset) & info->align_mask;
1951 VM_BUG_ON(gap_end < info->low_limit);
1952 VM_BUG_ON(gap_end < gap_start);
1956 /* Get an address range which is currently unmapped.
1957 * For shmat() with addr=0.
1959 * Ugly calling convention alert:
1960 * Return value with the low bits set means error value,
1962 * if (ret & ~PAGE_MASK)
1965 * This function "knows" that -ENOMEM has the bits set.
1967 #ifndef HAVE_ARCH_UNMAPPED_AREA
1969 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1970 unsigned long len, unsigned long pgoff, unsigned long flags)
1972 struct mm_struct *mm = current->mm;
1973 struct vm_area_struct *vma, *prev;
1974 struct vm_unmapped_area_info info;
1976 if (len > TASK_SIZE - mmap_min_addr)
1979 if (flags & MAP_FIXED)
1983 addr = PAGE_ALIGN(addr);
1984 vma = find_vma_prev(mm, addr, &prev);
1985 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1986 (!vma || addr + len <= vm_start_gap(vma)) &&
1987 (!prev || addr >= vm_end_gap(prev)))
1993 info.low_limit = mm->mmap_base;
1994 info.high_limit = TASK_SIZE;
1995 info.align_mask = 0;
1996 return vm_unmapped_area(&info);
2001 * This mmap-allocator allocates new areas top-down from below the
2002 * stack's low limit (the base):
2004 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2006 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
2007 const unsigned long len, const unsigned long pgoff,
2008 const unsigned long flags)
2010 struct vm_area_struct *vma, *prev;
2011 struct mm_struct *mm = current->mm;
2012 unsigned long addr = addr0;
2013 struct vm_unmapped_area_info info;
2015 /* requested length too big for entire address space */
2016 if (len > TASK_SIZE - mmap_min_addr)
2019 if (flags & MAP_FIXED)
2022 /* requesting a specific address */
2024 addr = PAGE_ALIGN(addr);
2025 vma = find_vma_prev(mm, addr, &prev);
2026 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2027 (!vma || addr + len <= vm_start_gap(vma)) &&
2028 (!prev || addr >= vm_end_gap(prev)))
2032 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2034 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2035 info.high_limit = mm->mmap_base;
2036 info.align_mask = 0;
2037 addr = vm_unmapped_area(&info);
2040 * A failed mmap() very likely causes application failure,
2041 * so fall back to the bottom-up function here. This scenario
2042 * can happen with large stack limits and large mmap()
2045 if (offset_in_page(addr)) {
2046 VM_BUG_ON(addr != -ENOMEM);
2048 info.low_limit = TASK_UNMAPPED_BASE;
2049 info.high_limit = TASK_SIZE;
2050 addr = vm_unmapped_area(&info);
2058 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2059 unsigned long pgoff, unsigned long flags)
2061 unsigned long (*get_area)(struct file *, unsigned long,
2062 unsigned long, unsigned long, unsigned long);
2064 unsigned long error = arch_mmap_check(addr, len, flags);
2068 /* Careful about overflows.. */
2069 if (len > TASK_SIZE)
2072 get_area = current->mm->get_unmapped_area;
2073 if (file && file->f_op->get_unmapped_area)
2074 get_area = file->f_op->get_unmapped_area;
2075 addr = get_area(file, addr, len, pgoff, flags);
2076 if (IS_ERR_VALUE(addr))
2079 if (addr > TASK_SIZE - len)
2081 if (offset_in_page(addr))
2084 addr = arch_rebalance_pgtables(addr, len);
2085 error = security_mmap_addr(addr);
2086 return error ? error : addr;
2089 EXPORT_SYMBOL(get_unmapped_area);
2091 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2092 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2094 struct rb_node *rb_node;
2095 struct vm_area_struct *vma;
2097 /* Check the cache first. */
2098 vma = vmacache_find(mm, addr);
2102 rb_node = mm->mm_rb.rb_node;
2105 struct vm_area_struct *tmp;
2107 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2109 if (tmp->vm_end > addr) {
2111 if (tmp->vm_start <= addr)
2113 rb_node = rb_node->rb_left;
2115 rb_node = rb_node->rb_right;
2119 vmacache_update(addr, vma);
2123 EXPORT_SYMBOL(find_vma);
2126 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2128 struct vm_area_struct *
2129 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2130 struct vm_area_struct **pprev)
2132 struct vm_area_struct *vma;
2134 vma = find_vma(mm, addr);
2136 *pprev = vma->vm_prev;
2138 struct rb_node *rb_node = mm->mm_rb.rb_node;
2141 *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2142 rb_node = rb_node->rb_right;
2149 * Verify that the stack growth is acceptable and
2150 * update accounting. This is shared with both the
2151 * grow-up and grow-down cases.
2153 static int acct_stack_growth(struct vm_area_struct *vma,
2154 unsigned long size, unsigned long grow)
2156 struct mm_struct *mm = vma->vm_mm;
2157 struct rlimit *rlim = current->signal->rlim;
2158 unsigned long new_start;
2160 /* address space limit tests */
2161 if (!may_expand_vm(mm, grow))
2164 /* Stack limit test */
2165 if (size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2168 /* mlock limit tests */
2169 if (vma->vm_flags & VM_LOCKED) {
2170 unsigned long locked;
2171 unsigned long limit;
2172 locked = mm->locked_vm + grow;
2173 limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2174 limit >>= PAGE_SHIFT;
2175 if (locked > limit && !capable(CAP_IPC_LOCK))
2179 /* Check to ensure the stack will not grow into a hugetlb-only region */
2180 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2182 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2186 * Overcommit.. This must be the final test, as it will
2187 * update security statistics.
2189 if (security_vm_enough_memory_mm(mm, grow))
2195 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2197 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2198 * vma is the last one with address > vma->vm_end. Have to extend vma.
2200 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2202 struct mm_struct *mm = vma->vm_mm;
2203 struct vm_area_struct *next;
2204 unsigned long gap_addr;
2207 if (!(vma->vm_flags & VM_GROWSUP))
2210 /* Guard against exceeding limits of the address space. */
2211 address &= PAGE_MASK;
2212 if (address >= (TASK_SIZE & PAGE_MASK))
2214 address += PAGE_SIZE;
2216 /* Enforce stack_guard_gap */
2217 gap_addr = address + stack_guard_gap;
2219 /* Guard against overflow */
2220 if (gap_addr < address || gap_addr > TASK_SIZE)
2221 gap_addr = TASK_SIZE;
2223 next = vma->vm_next;
2224 if (next && next->vm_start < gap_addr &&
2225 (next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2226 if (!(next->vm_flags & VM_GROWSUP))
2228 /* Check that both stack segments have the same anon_vma? */
2231 /* We must make sure the anon_vma is allocated. */
2232 if (unlikely(anon_vma_prepare(vma)))
2236 * vma->vm_start/vm_end cannot change under us because the caller
2237 * is required to hold the mmap_sem in read mode. We need the
2238 * anon_vma lock to serialize against concurrent expand_stacks.
2240 anon_vma_lock_write(vma->anon_vma);
2242 /* Somebody else might have raced and expanded it already */
2243 if (address > vma->vm_end) {
2244 unsigned long size, grow;
2246 size = address - vma->vm_start;
2247 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2250 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2251 error = acct_stack_growth(vma, size, grow);
2254 * vma_gap_update() doesn't support concurrent
2255 * updates, but we only hold a shared mmap_sem
2256 * lock here, so we need to protect against
2257 * concurrent vma expansions.
2258 * anon_vma_lock_write() doesn't help here, as
2259 * we don't guarantee that all growable vmas
2260 * in a mm share the same root anon vma.
2261 * So, we reuse mm->page_table_lock to guard
2262 * against concurrent vma expansions.
2264 spin_lock(&mm->page_table_lock);
2265 if (vma->vm_flags & VM_LOCKED)
2266 mm->locked_vm += grow;
2267 vm_stat_account(mm, vma->vm_flags,
2268 vma->vm_file, grow);
2269 anon_vma_interval_tree_pre_update_vma(vma);
2270 vma->vm_end = address;
2271 anon_vma_interval_tree_post_update_vma(vma);
2273 vma_gap_update(vma->vm_next);
2275 mm->highest_vm_end = vm_end_gap(vma);
2276 spin_unlock(&mm->page_table_lock);
2278 perf_event_mmap(vma);
2282 anon_vma_unlock_write(vma->anon_vma);
2283 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2287 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2290 * vma is the first one with address < vma->vm_start. Have to extend vma.
2292 int expand_downwards(struct vm_area_struct *vma,
2293 unsigned long address)
2295 struct mm_struct *mm = vma->vm_mm;
2296 struct vm_area_struct *prev;
2297 unsigned long gap_addr;
2300 address &= PAGE_MASK;
2301 if (address < mmap_min_addr)
2304 /* Enforce stack_guard_gap */
2305 gap_addr = address - stack_guard_gap;
2306 if (gap_addr > address)
2308 prev = vma->vm_prev;
2309 if (prev && prev->vm_end > gap_addr &&
2310 (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2311 if (!(prev->vm_flags & VM_GROWSDOWN))
2313 /* Check that both stack segments have the same anon_vma? */
2316 /* We must make sure the anon_vma is allocated. */
2317 if (unlikely(anon_vma_prepare(vma)))
2321 * vma->vm_start/vm_end cannot change under us because the caller
2322 * is required to hold the mmap_sem in read mode. We need the
2323 * anon_vma lock to serialize against concurrent expand_stacks.
2325 anon_vma_lock_write(vma->anon_vma);
2327 /* Somebody else might have raced and expanded it already */
2328 if (address < vma->vm_start) {
2329 unsigned long size, grow;
2331 size = vma->vm_end - address;
2332 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2335 if (grow <= vma->vm_pgoff) {
2336 error = acct_stack_growth(vma, size, grow);
2339 * vma_gap_update() doesn't support concurrent
2340 * updates, but we only hold a shared mmap_sem
2341 * lock here, so we need to protect against
2342 * concurrent vma expansions.
2343 * anon_vma_lock_write() doesn't help here, as
2344 * we don't guarantee that all growable vmas
2345 * in a mm share the same root anon vma.
2346 * So, we reuse mm->page_table_lock to guard
2347 * against concurrent vma expansions.
2349 spin_lock(&mm->page_table_lock);
2350 if (vma->vm_flags & VM_LOCKED)
2351 mm->locked_vm += grow;
2352 vm_stat_account(mm, vma->vm_flags,
2353 vma->vm_file, grow);
2354 anon_vma_interval_tree_pre_update_vma(vma);
2355 vma->vm_start = address;
2356 vma->vm_pgoff -= grow;
2357 anon_vma_interval_tree_post_update_vma(vma);
2358 vma_gap_update(vma);
2359 spin_unlock(&mm->page_table_lock);
2361 perf_event_mmap(vma);
2365 anon_vma_unlock_write(vma->anon_vma);
2366 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2371 /* enforced gap between the expanding stack and other mappings. */
2372 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2374 static int __init cmdline_parse_stack_guard_gap(char *p)
2379 val = simple_strtoul(p, &endptr, 10);
2381 stack_guard_gap = val << PAGE_SHIFT;
2385 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2387 #ifdef CONFIG_STACK_GROWSUP
2388 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2390 return expand_upwards(vma, address);
2393 struct vm_area_struct *
2394 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2396 struct vm_area_struct *vma, *prev;
2399 vma = find_vma_prev(mm, addr, &prev);
2400 if (vma && (vma->vm_start <= addr))
2402 /* don't alter vm_end if the coredump is running */
2403 if (!prev || !mmget_still_valid(mm) || expand_stack(prev, addr))
2405 if (prev->vm_flags & VM_LOCKED)
2406 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2410 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2412 return expand_downwards(vma, address);
2415 struct vm_area_struct *
2416 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2418 struct vm_area_struct *vma;
2419 unsigned long start;
2422 vma = find_vma(mm, addr);
2425 if (vma->vm_start <= addr)
2427 if (!(vma->vm_flags & VM_GROWSDOWN))
2429 /* don't alter vm_start if the coredump is running */
2430 if (!mmget_still_valid(mm))
2432 start = vma->vm_start;
2433 if (expand_stack(vma, addr))
2435 if (vma->vm_flags & VM_LOCKED)
2436 populate_vma_page_range(vma, addr, start, NULL);
2441 EXPORT_SYMBOL_GPL(find_extend_vma);
2444 * Ok - we have the memory areas we should free on the vma list,
2445 * so release them, and do the vma updates.
2447 * Called with the mm semaphore held.
2449 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2451 unsigned long nr_accounted = 0;
2453 /* Update high watermark before we lower total_vm */
2454 update_hiwater_vm(mm);
2456 long nrpages = vma_pages(vma);
2458 if (vma->vm_flags & VM_ACCOUNT)
2459 nr_accounted += nrpages;
2460 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
2461 vma = remove_vma(vma);
2463 vm_unacct_memory(nr_accounted);
2468 * Get rid of page table information in the indicated region.
2470 * Called with the mm semaphore held.
2472 static void unmap_region(struct mm_struct *mm,
2473 struct vm_area_struct *vma, struct vm_area_struct *prev,
2474 unsigned long start, unsigned long end)
2476 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2477 struct mmu_gather tlb;
2480 tlb_gather_mmu(&tlb, mm, start, end);
2481 update_hiwater_rss(mm);
2482 unmap_vmas(&tlb, vma, start, end);
2483 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2484 next ? next->vm_start : USER_PGTABLES_CEILING);
2485 tlb_finish_mmu(&tlb, start, end);
2489 * Create a list of vma's touched by the unmap, removing them from the mm's
2490 * vma list as we go..
2493 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2494 struct vm_area_struct *prev, unsigned long end)
2496 struct vm_area_struct **insertion_point;
2497 struct vm_area_struct *tail_vma = NULL;
2499 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2500 vma->vm_prev = NULL;
2502 vma_rb_erase(vma, &mm->mm_rb);
2506 } while (vma && vma->vm_start < end);
2507 *insertion_point = vma;
2509 vma->vm_prev = prev;
2510 vma_gap_update(vma);
2512 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2513 tail_vma->vm_next = NULL;
2515 /* Kill the cache */
2516 vmacache_invalidate(mm);
2520 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2521 * munmap path where it doesn't make sense to fail.
2523 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2524 unsigned long addr, int new_below)
2526 struct vm_area_struct *new;
2529 if (is_vm_hugetlb_page(vma) && (addr &
2530 ~(huge_page_mask(hstate_vma(vma)))))
2533 new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2537 /* most fields are the same, copy all, and then fixup */
2540 INIT_LIST_HEAD(&new->anon_vma_chain);
2545 new->vm_start = addr;
2546 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2549 err = vma_dup_policy(vma, new);
2553 err = anon_vma_clone(new, vma);
2558 get_file(new->vm_file);
2560 if (new->vm_ops && new->vm_ops->open)
2561 new->vm_ops->open(new);
2564 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2565 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2567 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2573 /* Clean everything up if vma_adjust failed. */
2574 if (new->vm_ops && new->vm_ops->close)
2575 new->vm_ops->close(new);
2578 unlink_anon_vmas(new);
2580 mpol_put(vma_policy(new));
2582 kmem_cache_free(vm_area_cachep, new);
2587 * Split a vma into two pieces at address 'addr', a new vma is allocated
2588 * either for the first part or the tail.
2590 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2591 unsigned long addr, int new_below)
2593 if (mm->map_count >= sysctl_max_map_count)
2596 return __split_vma(mm, vma, addr, new_below);
2599 /* Munmap is split into 2 main parts -- this part which finds
2600 * what needs doing, and the areas themselves, which do the
2601 * work. This now handles partial unmappings.
2602 * Jeremy Fitzhardinge <jeremy@goop.org>
2604 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2607 struct vm_area_struct *vma, *prev, *last;
2609 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2612 len = PAGE_ALIGN(len);
2616 /* Find the first overlapping VMA */
2617 vma = find_vma(mm, start);
2620 prev = vma->vm_prev;
2621 /* we have start < vma->vm_end */
2623 /* if it doesn't overlap, we have nothing.. */
2625 if (vma->vm_start >= end)
2629 * If we need to split any vma, do it now to save pain later.
2631 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2632 * unmapped vm_area_struct will remain in use: so lower split_vma
2633 * places tmp vma above, and higher split_vma places tmp vma below.
2635 if (start > vma->vm_start) {
2639 * Make sure that map_count on return from munmap() will
2640 * not exceed its limit; but let map_count go just above
2641 * its limit temporarily, to help free resources as expected.
2643 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2646 error = __split_vma(mm, vma, start, 0);
2652 /* Does it split the last one? */
2653 last = find_vma(mm, end);
2654 if (last && end > last->vm_start) {
2655 int error = __split_vma(mm, last, end, 1);
2659 vma = prev ? prev->vm_next : mm->mmap;
2662 * unlock any mlock()ed ranges before detaching vmas
2664 if (mm->locked_vm) {
2665 struct vm_area_struct *tmp = vma;
2666 while (tmp && tmp->vm_start < end) {
2667 if (tmp->vm_flags & VM_LOCKED) {
2668 mm->locked_vm -= vma_pages(tmp);
2669 munlock_vma_pages_all(tmp);
2676 * Remove the vma's, and unmap the actual pages
2678 detach_vmas_to_be_unmapped(mm, vma, prev, end);
2679 unmap_region(mm, vma, prev, start, end);
2681 arch_unmap(mm, vma, start, end);
2683 /* Fix up all other VM information */
2684 remove_vma_list(mm, vma);
2689 int vm_munmap(unsigned long start, size_t len)
2692 struct mm_struct *mm = current->mm;
2694 down_write(&mm->mmap_sem);
2695 ret = do_munmap(mm, start, len);
2696 up_write(&mm->mmap_sem);
2699 EXPORT_SYMBOL(vm_munmap);
2701 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2703 profile_munmap(addr);
2704 return vm_munmap(addr, len);
2709 * Emulation of deprecated remap_file_pages() syscall.
2711 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2712 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2715 struct mm_struct *mm = current->mm;
2716 struct vm_area_struct *vma;
2717 unsigned long populate = 0;
2718 unsigned long ret = -EINVAL;
2721 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. "
2722 "See Documentation/vm/remap_file_pages.txt.\n",
2723 current->comm, current->pid);
2727 start = start & PAGE_MASK;
2728 size = size & PAGE_MASK;
2730 if (start + size <= start)
2733 /* Does pgoff wrap? */
2734 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2737 down_write(&mm->mmap_sem);
2738 vma = find_vma(mm, start);
2740 if (!vma || !(vma->vm_flags & VM_SHARED))
2743 if (start < vma->vm_start)
2746 if (start + size > vma->vm_end) {
2747 struct vm_area_struct *next;
2749 for (next = vma->vm_next; next; next = next->vm_next) {
2750 /* hole between vmas ? */
2751 if (next->vm_start != next->vm_prev->vm_end)
2754 if (next->vm_file != vma->vm_file)
2757 if (next->vm_flags != vma->vm_flags)
2760 if (start + size <= next->vm_end)
2768 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2769 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2770 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2772 flags &= MAP_NONBLOCK;
2773 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2774 if (vma->vm_flags & VM_LOCKED) {
2775 struct vm_area_struct *tmp;
2776 flags |= MAP_LOCKED;
2778 /* drop PG_Mlocked flag for over-mapped range */
2779 for (tmp = vma; tmp->vm_start >= start + size;
2780 tmp = tmp->vm_next) {
2781 munlock_vma_pages_range(tmp,
2782 max(tmp->vm_start, start),
2783 min(tmp->vm_end, start + size));
2787 file = get_file(vma->vm_file);
2788 ret = do_mmap_pgoff(vma->vm_file, start, size,
2789 prot, flags, pgoff, &populate);
2792 up_write(&mm->mmap_sem);
2794 mm_populate(ret, populate);
2795 if (!IS_ERR_VALUE(ret))
2800 static inline void verify_mm_writelocked(struct mm_struct *mm)
2802 #ifdef CONFIG_DEBUG_VM
2803 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2805 up_read(&mm->mmap_sem);
2811 * this is really a simplified "do_mmap". it only handles
2812 * anonymous maps. eventually we may be able to do some
2813 * brk-specific accounting here.
2815 static unsigned long do_brk(unsigned long addr, unsigned long len)
2817 struct mm_struct *mm = current->mm;
2818 struct vm_area_struct *vma, *prev;
2819 unsigned long flags;
2820 struct rb_node **rb_link, *rb_parent;
2821 pgoff_t pgoff = addr >> PAGE_SHIFT;
2824 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2826 error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2827 if (offset_in_page(error))
2830 error = mlock_future_check(mm, mm->def_flags, len);
2835 * mm->mmap_sem is required to protect against another thread
2836 * changing the mappings in case we sleep.
2838 verify_mm_writelocked(mm);
2841 * Clear old maps. this also does some error checking for us
2843 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2845 if (do_munmap(mm, addr, len))
2849 /* Check against address space limits *after* clearing old maps... */
2850 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2853 if (mm->map_count > sysctl_max_map_count)
2856 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2859 /* Can we just expand an old private anonymous mapping? */
2860 vma = vma_merge(mm, prev, addr, addr + len, flags,
2861 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2866 * create a vma struct for an anonymous mapping
2868 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2870 vm_unacct_memory(len >> PAGE_SHIFT);
2874 INIT_LIST_HEAD(&vma->anon_vma_chain);
2876 vma->vm_start = addr;
2877 vma->vm_end = addr + len;
2878 vma->vm_pgoff = pgoff;
2879 vma->vm_flags = flags;
2880 vma->vm_page_prot = vm_get_page_prot(flags);
2881 vma_link(mm, vma, prev, rb_link, rb_parent);
2883 perf_event_mmap(vma);
2884 mm->total_vm += len >> PAGE_SHIFT;
2885 if (flags & VM_LOCKED)
2886 mm->locked_vm += (len >> PAGE_SHIFT);
2887 vma->vm_flags |= VM_SOFTDIRTY;
2891 unsigned long vm_brk(unsigned long addr, unsigned long request)
2893 struct mm_struct *mm = current->mm;
2898 len = PAGE_ALIGN(request);
2904 down_write(&mm->mmap_sem);
2905 ret = do_brk(addr, len);
2906 populate = ((mm->def_flags & VM_LOCKED) != 0);
2907 up_write(&mm->mmap_sem);
2909 mm_populate(addr, len);
2912 EXPORT_SYMBOL(vm_brk);
2914 /* Release all mmaps. */
2915 void exit_mmap(struct mm_struct *mm)
2917 struct mmu_gather tlb;
2918 struct vm_area_struct *vma;
2919 unsigned long nr_accounted = 0;
2921 /* mm's last user has gone, and its about to be pulled down */
2922 mmu_notifier_release(mm);
2924 if (mm->locked_vm) {
2927 if (vma->vm_flags & VM_LOCKED)
2928 munlock_vma_pages_all(vma);
2936 if (!vma) /* Can happen if dup_mmap() received an OOM */
2941 tlb_gather_mmu(&tlb, mm, 0, -1);
2942 /* update_hiwater_rss(mm) here? but nobody should be looking */
2943 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2944 unmap_vmas(&tlb, vma, 0, -1);
2946 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
2947 tlb_finish_mmu(&tlb, 0, -1);
2950 * Walk the list again, actually closing and freeing it,
2951 * with preemption enabled, without holding any MM locks.
2954 if (vma->vm_flags & VM_ACCOUNT)
2955 nr_accounted += vma_pages(vma);
2956 vma = remove_vma(vma);
2958 vm_unacct_memory(nr_accounted);
2961 /* Insert vm structure into process list sorted by address
2962 * and into the inode's i_mmap tree. If vm_file is non-NULL
2963 * then i_mmap_rwsem is taken here.
2965 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2967 struct vm_area_struct *prev;
2968 struct rb_node **rb_link, *rb_parent;
2970 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2971 &prev, &rb_link, &rb_parent))
2973 if ((vma->vm_flags & VM_ACCOUNT) &&
2974 security_vm_enough_memory_mm(mm, vma_pages(vma)))
2978 * The vm_pgoff of a purely anonymous vma should be irrelevant
2979 * until its first write fault, when page's anon_vma and index
2980 * are set. But now set the vm_pgoff it will almost certainly
2981 * end up with (unless mremap moves it elsewhere before that
2982 * first wfault), so /proc/pid/maps tells a consistent story.
2984 * By setting it to reflect the virtual start address of the
2985 * vma, merges and splits can happen in a seamless way, just
2986 * using the existing file pgoff checks and manipulations.
2987 * Similarly in do_mmap_pgoff and in do_brk.
2989 if (vma_is_anonymous(vma)) {
2990 BUG_ON(vma->anon_vma);
2991 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2994 vma_link(mm, vma, prev, rb_link, rb_parent);
2999 * Copy the vma structure to a new location in the same mm,
3000 * prior to moving page table entries, to effect an mremap move.
3002 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3003 unsigned long addr, unsigned long len, pgoff_t pgoff,
3004 bool *need_rmap_locks)
3006 struct vm_area_struct *vma = *vmap;
3007 unsigned long vma_start = vma->vm_start;
3008 struct mm_struct *mm = vma->vm_mm;
3009 struct vm_area_struct *new_vma, *prev;
3010 struct rb_node **rb_link, *rb_parent;
3011 bool faulted_in_anon_vma = true;
3014 * If anonymous vma has not yet been faulted, update new pgoff
3015 * to match new location, to increase its chance of merging.
3017 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3018 pgoff = addr >> PAGE_SHIFT;
3019 faulted_in_anon_vma = false;
3022 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3023 return NULL; /* should never get here */
3024 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3025 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3026 vma->vm_userfaultfd_ctx);
3029 * Source vma may have been merged into new_vma
3031 if (unlikely(vma_start >= new_vma->vm_start &&
3032 vma_start < new_vma->vm_end)) {
3034 * The only way we can get a vma_merge with
3035 * self during an mremap is if the vma hasn't
3036 * been faulted in yet and we were allowed to
3037 * reset the dst vma->vm_pgoff to the
3038 * destination address of the mremap to allow
3039 * the merge to happen. mremap must change the
3040 * vm_pgoff linearity between src and dst vmas
3041 * (in turn preventing a vma_merge) to be
3042 * safe. It is only safe to keep the vm_pgoff
3043 * linear if there are no pages mapped yet.
3045 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3046 *vmap = vma = new_vma;
3048 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3050 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3054 new_vma->vm_start = addr;
3055 new_vma->vm_end = addr + len;
3056 new_vma->vm_pgoff = pgoff;
3057 if (vma_dup_policy(vma, new_vma))
3059 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3060 if (anon_vma_clone(new_vma, vma))
3061 goto out_free_mempol;
3062 if (new_vma->vm_file)
3063 get_file(new_vma->vm_file);
3064 if (new_vma->vm_ops && new_vma->vm_ops->open)
3065 new_vma->vm_ops->open(new_vma);
3066 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3067 *need_rmap_locks = false;
3072 mpol_put(vma_policy(new_vma));
3074 kmem_cache_free(vm_area_cachep, new_vma);
3080 * Return true if the calling process may expand its vm space by the passed
3083 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
3085 unsigned long cur = mm->total_vm; /* pages */
3088 lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
3090 if (cur + npages > lim)
3095 static int special_mapping_fault(struct vm_area_struct *vma,
3096 struct vm_fault *vmf);
3099 * Having a close hook prevents vma merging regardless of flags.
3101 static void special_mapping_close(struct vm_area_struct *vma)
3105 static const char *special_mapping_name(struct vm_area_struct *vma)
3107 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3110 static const struct vm_operations_struct special_mapping_vmops = {
3111 .close = special_mapping_close,
3112 .fault = special_mapping_fault,
3113 .name = special_mapping_name,
3116 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3117 .close = special_mapping_close,
3118 .fault = special_mapping_fault,
3121 static int special_mapping_fault(struct vm_area_struct *vma,
3122 struct vm_fault *vmf)
3125 struct page **pages;
3127 if (vma->vm_ops == &legacy_special_mapping_vmops)
3128 pages = vma->vm_private_data;
3130 pages = ((struct vm_special_mapping *)vma->vm_private_data)->
3133 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3137 struct page *page = *pages;
3143 return VM_FAULT_SIGBUS;
3146 static struct vm_area_struct *__install_special_mapping(
3147 struct mm_struct *mm,
3148 unsigned long addr, unsigned long len,
3149 unsigned long vm_flags, void *priv,
3150 const struct vm_operations_struct *ops)
3153 struct vm_area_struct *vma;
3155 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3156 if (unlikely(vma == NULL))
3157 return ERR_PTR(-ENOMEM);
3159 INIT_LIST_HEAD(&vma->anon_vma_chain);
3161 vma->vm_start = addr;
3162 vma->vm_end = addr + len;
3164 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3165 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3168 vma->vm_private_data = priv;
3170 ret = insert_vm_struct(mm, vma);
3174 mm->total_vm += len >> PAGE_SHIFT;
3176 perf_event_mmap(vma);
3181 kmem_cache_free(vm_area_cachep, vma);
3182 return ERR_PTR(ret);
3186 * Called with mm->mmap_sem held for writing.
3187 * Insert a new vma covering the given region, with the given flags.
3188 * Its pages are supplied by the given array of struct page *.
3189 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3190 * The region past the last page supplied will always produce SIGBUS.
3191 * The array pointer and the pages it points to are assumed to stay alive
3192 * for as long as this mapping might exist.
3194 struct vm_area_struct *_install_special_mapping(
3195 struct mm_struct *mm,
3196 unsigned long addr, unsigned long len,
3197 unsigned long vm_flags, const struct vm_special_mapping *spec)
3199 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3200 &special_mapping_vmops);
3203 int install_special_mapping(struct mm_struct *mm,
3204 unsigned long addr, unsigned long len,
3205 unsigned long vm_flags, struct page **pages)
3207 struct vm_area_struct *vma = __install_special_mapping(
3208 mm, addr, len, vm_flags, (void *)pages,
3209 &legacy_special_mapping_vmops);
3211 return PTR_ERR_OR_ZERO(vma);
3214 static DEFINE_MUTEX(mm_all_locks_mutex);
3216 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3218 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3220 * The LSB of head.next can't change from under us
3221 * because we hold the mm_all_locks_mutex.
3223 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3225 * We can safely modify head.next after taking the
3226 * anon_vma->root->rwsem. If some other vma in this mm shares
3227 * the same anon_vma we won't take it again.
3229 * No need of atomic instructions here, head.next
3230 * can't change from under us thanks to the
3231 * anon_vma->root->rwsem.
3233 if (__test_and_set_bit(0, (unsigned long *)
3234 &anon_vma->root->rb_root.rb_node))
3239 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3241 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3243 * AS_MM_ALL_LOCKS can't change from under us because
3244 * we hold the mm_all_locks_mutex.
3246 * Operations on ->flags have to be atomic because
3247 * even if AS_MM_ALL_LOCKS is stable thanks to the
3248 * mm_all_locks_mutex, there may be other cpus
3249 * changing other bitflags in parallel to us.
3251 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3253 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3258 * This operation locks against the VM for all pte/vma/mm related
3259 * operations that could ever happen on a certain mm. This includes
3260 * vmtruncate, try_to_unmap, and all page faults.
3262 * The caller must take the mmap_sem in write mode before calling
3263 * mm_take_all_locks(). The caller isn't allowed to release the
3264 * mmap_sem until mm_drop_all_locks() returns.
3266 * mmap_sem in write mode is required in order to block all operations
3267 * that could modify pagetables and free pages without need of
3268 * altering the vma layout. It's also needed in write mode to avoid new
3269 * anon_vmas to be associated with existing vmas.
3271 * A single task can't take more than one mm_take_all_locks() in a row
3272 * or it would deadlock.
3274 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3275 * mapping->flags avoid to take the same lock twice, if more than one
3276 * vma in this mm is backed by the same anon_vma or address_space.
3278 * We can take all the locks in random order because the VM code
3279 * taking i_mmap_rwsem or anon_vma->rwsem outside the mmap_sem never
3280 * takes more than one of them in a row. Secondly we're protected
3281 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3283 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3284 * that may have to take thousand of locks.
3286 * mm_take_all_locks() can fail if it's interrupted by signals.
3288 int mm_take_all_locks(struct mm_struct *mm)
3290 struct vm_area_struct *vma;
3291 struct anon_vma_chain *avc;
3293 BUG_ON(down_read_trylock(&mm->mmap_sem));
3295 mutex_lock(&mm_all_locks_mutex);
3297 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3298 if (signal_pending(current))
3300 if (vma->vm_file && vma->vm_file->f_mapping)
3301 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3304 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3305 if (signal_pending(current))
3308 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3309 vm_lock_anon_vma(mm, avc->anon_vma);
3315 mm_drop_all_locks(mm);
3319 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3321 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3323 * The LSB of head.next can't change to 0 from under
3324 * us because we hold the mm_all_locks_mutex.
3326 * We must however clear the bitflag before unlocking
3327 * the vma so the users using the anon_vma->rb_root will
3328 * never see our bitflag.
3330 * No need of atomic instructions here, head.next
3331 * can't change from under us until we release the
3332 * anon_vma->root->rwsem.
3334 if (!__test_and_clear_bit(0, (unsigned long *)
3335 &anon_vma->root->rb_root.rb_node))
3337 anon_vma_unlock_write(anon_vma);
3341 static void vm_unlock_mapping(struct address_space *mapping)
3343 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3345 * AS_MM_ALL_LOCKS can't change to 0 from under us
3346 * because we hold the mm_all_locks_mutex.
3348 i_mmap_unlock_write(mapping);
3349 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3356 * The mmap_sem cannot be released by the caller until
3357 * mm_drop_all_locks() returns.
3359 void mm_drop_all_locks(struct mm_struct *mm)
3361 struct vm_area_struct *vma;
3362 struct anon_vma_chain *avc;
3364 BUG_ON(down_read_trylock(&mm->mmap_sem));
3365 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3367 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3369 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3370 vm_unlock_anon_vma(avc->anon_vma);
3371 if (vma->vm_file && vma->vm_file->f_mapping)
3372 vm_unlock_mapping(vma->vm_file->f_mapping);
3375 mutex_unlock(&mm_all_locks_mutex);
3379 * initialise the VMA slab
3381 void __init mmap_init(void)
3385 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3390 * Initialise sysctl_user_reserve_kbytes.
3392 * This is intended to prevent a user from starting a single memory hogging
3393 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3396 * The default value is min(3% of free memory, 128MB)
3397 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3399 static int init_user_reserve(void)
3401 unsigned long free_kbytes;
3403 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3405 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3408 subsys_initcall(init_user_reserve);
3411 * Initialise sysctl_admin_reserve_kbytes.
3413 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3414 * to log in and kill a memory hogging process.
3416 * Systems with more than 256MB will reserve 8MB, enough to recover
3417 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3418 * only reserve 3% of free pages by default.
3420 static int init_admin_reserve(void)
3422 unsigned long free_kbytes;
3424 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3426 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3429 subsys_initcall(init_admin_reserve);
3432 * Reinititalise user and admin reserves if memory is added or removed.
3434 * The default user reserve max is 128MB, and the default max for the
3435 * admin reserve is 8MB. These are usually, but not always, enough to
3436 * enable recovery from a memory hogging process using login/sshd, a shell,
3437 * and tools like top. It may make sense to increase or even disable the
3438 * reserve depending on the existence of swap or variations in the recovery
3439 * tools. So, the admin may have changed them.
3441 * If memory is added and the reserves have been eliminated or increased above
3442 * the default max, then we'll trust the admin.
3444 * If memory is removed and there isn't enough free memory, then we
3445 * need to reset the reserves.
3447 * Otherwise keep the reserve set by the admin.
3449 static int reserve_mem_notifier(struct notifier_block *nb,
3450 unsigned long action, void *data)
3452 unsigned long tmp, free_kbytes;
3456 /* Default max is 128MB. Leave alone if modified by operator. */
3457 tmp = sysctl_user_reserve_kbytes;
3458 if (0 < tmp && tmp < (1UL << 17))
3459 init_user_reserve();
3461 /* Default max is 8MB. Leave alone if modified by operator. */
3462 tmp = sysctl_admin_reserve_kbytes;
3463 if (0 < tmp && tmp < (1UL << 13))
3464 init_admin_reserve();
3468 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3470 if (sysctl_user_reserve_kbytes > free_kbytes) {
3471 init_user_reserve();
3472 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3473 sysctl_user_reserve_kbytes);
3476 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3477 init_admin_reserve();
3478 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3479 sysctl_admin_reserve_kbytes);
3488 static struct notifier_block reserve_mem_nb = {
3489 .notifier_call = reserve_mem_notifier,
3492 static int __meminit init_reserve_notifier(void)
3494 if (register_hotmemory_notifier(&reserve_mem_nb))
3495 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3499 subsys_initcall(init_reserve_notifier);