1 // SPDX-License-Identifier: GPL-2.0-only
7 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12 #include <linux/kernel.h>
13 #include <linux/slab.h>
14 #include <linux/backing-dev.h>
16 #include <linux/vmacache.h>
17 #include <linux/shm.h>
18 #include <linux/mman.h>
19 #include <linux/pagemap.h>
20 #include <linux/swap.h>
21 #include <linux/syscalls.h>
22 #include <linux/capability.h>
23 #include <linux/init.h>
24 #include <linux/file.h>
26 #include <linux/personality.h>
27 #include <linux/security.h>
28 #include <linux/hugetlb.h>
29 #include <linux/shmem_fs.h>
30 #include <linux/profile.h>
31 #include <linux/export.h>
32 #include <linux/mount.h>
33 #include <linux/mempolicy.h>
34 #include <linux/rmap.h>
35 #include <linux/mmu_notifier.h>
36 #include <linux/mmdebug.h>
37 #include <linux/perf_event.h>
38 #include <linux/audit.h>
39 #include <linux/khugepaged.h>
40 #include <linux/uprobes.h>
41 #include <linux/rbtree_augmented.h>
42 #include <linux/notifier.h>
43 #include <linux/memory.h>
44 #include <linux/printk.h>
45 #include <linux/userfaultfd_k.h>
46 #include <linux/moduleparam.h>
47 #include <linux/pkeys.h>
48 #include <linux/oom.h>
49 #include <linux/sched/mm.h>
51 #include <linux/uaccess.h>
52 #include <asm/cacheflush.h>
54 #include <asm/mmu_context.h>
56 #define CREATE_TRACE_POINTS
57 #include <trace/events/mmap.h>
61 #ifndef arch_mmap_check
62 #define arch_mmap_check(addr, len, flags) (0)
65 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
66 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
67 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
68 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
70 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
71 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
72 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
73 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
76 static bool ignore_rlimit_data;
77 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
79 static void unmap_region(struct mm_struct *mm,
80 struct vm_area_struct *vma, struct vm_area_struct *prev,
81 unsigned long start, unsigned long end);
83 /* description of effects of mapping type and prot in current implementation.
84 * this is due to the limited x86 page protection hardware. The expected
85 * behavior is in parens:
88 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
89 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
90 * w: (no) no w: (no) no w: (yes) yes w: (no) no
91 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
93 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
94 * w: (no) no w: (no) no w: (copy) copy w: (no) no
95 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
97 pgprot_t protection_map[16] __ro_after_init = {
98 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
99 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
102 #ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
103 static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
109 pgprot_t vm_get_page_prot(unsigned long vm_flags)
111 pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
112 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
113 pgprot_val(arch_vm_get_page_prot(vm_flags)));
115 return arch_filter_pgprot(ret);
117 EXPORT_SYMBOL(vm_get_page_prot);
119 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
121 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
124 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
125 void vma_set_page_prot(struct vm_area_struct *vma)
127 unsigned long vm_flags = vma->vm_flags;
128 pgprot_t vm_page_prot;
130 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
131 if (vma_wants_writenotify(vma, vm_page_prot)) {
132 vm_flags &= ~VM_SHARED;
133 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
135 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
136 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
140 * Requires inode->i_mapping->i_mmap_rwsem
142 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
143 struct file *file, struct address_space *mapping)
145 if (vma->vm_flags & VM_DENYWRITE)
146 atomic_inc(&file_inode(file)->i_writecount);
147 if (vma->vm_flags & VM_SHARED)
148 mapping_unmap_writable(mapping);
150 flush_dcache_mmap_lock(mapping);
151 vma_interval_tree_remove(vma, &mapping->i_mmap);
152 flush_dcache_mmap_unlock(mapping);
156 * Unlink a file-based vm structure from its interval tree, to hide
157 * vma from rmap and vmtruncate before freeing its page tables.
159 void unlink_file_vma(struct vm_area_struct *vma)
161 struct file *file = vma->vm_file;
164 struct address_space *mapping = file->f_mapping;
165 i_mmap_lock_write(mapping);
166 __remove_shared_vm_struct(vma, file, mapping);
167 i_mmap_unlock_write(mapping);
172 * Close a vm structure and free it, returning the next.
174 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
176 struct vm_area_struct *next = vma->vm_next;
179 if (vma->vm_ops && vma->vm_ops->close)
180 vma->vm_ops->close(vma);
183 mpol_put(vma_policy(vma));
188 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags,
189 struct list_head *uf);
190 SYSCALL_DEFINE1(brk, unsigned long, brk)
192 unsigned long retval;
193 unsigned long newbrk, oldbrk, origbrk;
194 struct mm_struct *mm = current->mm;
195 struct vm_area_struct *next;
196 unsigned long min_brk;
198 bool downgraded = false;
201 if (mmap_write_lock_killable(mm))
206 #ifdef CONFIG_COMPAT_BRK
208 * CONFIG_COMPAT_BRK can still be overridden by setting
209 * randomize_va_space to 2, which will still cause mm->start_brk
210 * to be arbitrarily shifted
212 if (current->brk_randomized)
213 min_brk = mm->start_brk;
215 min_brk = mm->end_data;
217 min_brk = mm->start_brk;
223 * Check against rlimit here. If this check is done later after the test
224 * of oldbrk with newbrk then it can escape the test and let the data
225 * segment grow beyond its set limit the in case where the limit is
226 * not page aligned -Ram Gupta
228 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
229 mm->end_data, mm->start_data))
232 newbrk = PAGE_ALIGN(brk);
233 oldbrk = PAGE_ALIGN(mm->brk);
234 if (oldbrk == newbrk) {
240 * Always allow shrinking brk.
241 * __do_munmap() may downgrade mmap_lock to read.
243 if (brk <= mm->brk) {
247 * mm->brk must to be protected by write mmap_lock so update it
248 * before downgrading mmap_lock. When __do_munmap() fails,
249 * mm->brk will be restored from origbrk.
252 ret = __do_munmap(mm, newbrk, oldbrk-newbrk, &uf, true);
256 } else if (ret == 1) {
262 /* Check against existing mmap mappings. */
263 next = find_vma(mm, oldbrk);
264 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
267 /* Ok, looks good - let it rip. */
268 if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
273 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
275 mmap_read_unlock(mm);
277 mmap_write_unlock(mm);
278 userfaultfd_unmap_complete(mm, &uf);
280 mm_populate(oldbrk, newbrk - oldbrk);
285 mmap_write_unlock(mm);
289 static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
291 unsigned long gap, prev_end;
294 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
295 * allow two stack_guard_gaps between them here, and when choosing
296 * an unmapped area; whereas when expanding we only require one.
297 * That's a little inconsistent, but keeps the code here simpler.
299 gap = vm_start_gap(vma);
301 prev_end = vm_end_gap(vma->vm_prev);
310 #ifdef CONFIG_DEBUG_VM_RB
311 static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma)
313 unsigned long max = vma_compute_gap(vma), subtree_gap;
314 if (vma->vm_rb.rb_left) {
315 subtree_gap = rb_entry(vma->vm_rb.rb_left,
316 struct vm_area_struct, vm_rb)->rb_subtree_gap;
317 if (subtree_gap > max)
320 if (vma->vm_rb.rb_right) {
321 subtree_gap = rb_entry(vma->vm_rb.rb_right,
322 struct vm_area_struct, vm_rb)->rb_subtree_gap;
323 if (subtree_gap > max)
329 static int browse_rb(struct mm_struct *mm)
331 struct rb_root *root = &mm->mm_rb;
332 int i = 0, j, bug = 0;
333 struct rb_node *nd, *pn = NULL;
334 unsigned long prev = 0, pend = 0;
336 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
337 struct vm_area_struct *vma;
338 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
339 if (vma->vm_start < prev) {
340 pr_emerg("vm_start %lx < prev %lx\n",
341 vma->vm_start, prev);
344 if (vma->vm_start < pend) {
345 pr_emerg("vm_start %lx < pend %lx\n",
346 vma->vm_start, pend);
349 if (vma->vm_start > vma->vm_end) {
350 pr_emerg("vm_start %lx > vm_end %lx\n",
351 vma->vm_start, vma->vm_end);
354 spin_lock(&mm->page_table_lock);
355 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
356 pr_emerg("free gap %lx, correct %lx\n",
358 vma_compute_subtree_gap(vma));
361 spin_unlock(&mm->page_table_lock);
364 prev = vma->vm_start;
368 for (nd = pn; nd; nd = rb_prev(nd))
371 pr_emerg("backwards %d, forwards %d\n", j, i);
377 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
381 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
382 struct vm_area_struct *vma;
383 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
384 VM_BUG_ON_VMA(vma != ignore &&
385 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
390 static void validate_mm(struct mm_struct *mm)
394 unsigned long highest_address = 0;
395 struct vm_area_struct *vma = mm->mmap;
398 struct anon_vma *anon_vma = vma->anon_vma;
399 struct anon_vma_chain *avc;
402 anon_vma_lock_read(anon_vma);
403 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
404 anon_vma_interval_tree_verify(avc);
405 anon_vma_unlock_read(anon_vma);
408 highest_address = vm_end_gap(vma);
412 if (i != mm->map_count) {
413 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
416 if (highest_address != mm->highest_vm_end) {
417 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
418 mm->highest_vm_end, highest_address);
422 if (i != mm->map_count) {
424 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
427 VM_BUG_ON_MM(bug, mm);
430 #define validate_mm_rb(root, ignore) do { } while (0)
431 #define validate_mm(mm) do { } while (0)
434 RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks,
435 struct vm_area_struct, vm_rb,
436 unsigned long, rb_subtree_gap, vma_compute_gap)
439 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
440 * vma->vm_prev->vm_end values changed, without modifying the vma's position
443 static void vma_gap_update(struct vm_area_struct *vma)
446 * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
447 * a callback function that does exactly what we want.
449 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
452 static inline void vma_rb_insert(struct vm_area_struct *vma,
453 struct rb_root *root)
455 /* All rb_subtree_gap values must be consistent prior to insertion */
456 validate_mm_rb(root, NULL);
458 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
461 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
464 * Note rb_erase_augmented is a fairly large inline function,
465 * so make sure we instantiate it only once with our desired
466 * augmented rbtree callbacks.
468 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
471 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
472 struct rb_root *root,
473 struct vm_area_struct *ignore)
476 * All rb_subtree_gap values must be consistent prior to erase,
477 * with the possible exception of the "next" vma being erased if
478 * next->vm_start was reduced.
480 validate_mm_rb(root, ignore);
482 __vma_rb_erase(vma, root);
485 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
486 struct rb_root *root)
489 * All rb_subtree_gap values must be consistent prior to erase,
490 * with the possible exception of the vma being erased.
492 validate_mm_rb(root, vma);
494 __vma_rb_erase(vma, root);
498 * vma has some anon_vma assigned, and is already inserted on that
499 * anon_vma's interval trees.
501 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
502 * vma must be removed from the anon_vma's interval trees using
503 * anon_vma_interval_tree_pre_update_vma().
505 * After the update, the vma will be reinserted using
506 * anon_vma_interval_tree_post_update_vma().
508 * The entire update must be protected by exclusive mmap_lock and by
509 * the root anon_vma's mutex.
512 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
514 struct anon_vma_chain *avc;
516 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
517 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
521 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
523 struct anon_vma_chain *avc;
525 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
526 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
529 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
530 unsigned long end, struct vm_area_struct **pprev,
531 struct rb_node ***rb_link, struct rb_node **rb_parent)
533 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
535 __rb_link = &mm->mm_rb.rb_node;
536 rb_prev = __rb_parent = NULL;
539 struct vm_area_struct *vma_tmp;
541 __rb_parent = *__rb_link;
542 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
544 if (vma_tmp->vm_end > addr) {
545 /* Fail if an existing vma overlaps the area */
546 if (vma_tmp->vm_start < end)
548 __rb_link = &__rb_parent->rb_left;
550 rb_prev = __rb_parent;
551 __rb_link = &__rb_parent->rb_right;
557 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
558 *rb_link = __rb_link;
559 *rb_parent = __rb_parent;
563 static unsigned long count_vma_pages_range(struct mm_struct *mm,
564 unsigned long addr, unsigned long end)
566 unsigned long nr_pages = 0;
567 struct vm_area_struct *vma;
569 /* Find first overlaping mapping */
570 vma = find_vma_intersection(mm, addr, end);
574 nr_pages = (min(end, vma->vm_end) -
575 max(addr, vma->vm_start)) >> PAGE_SHIFT;
577 /* Iterate over the rest of the overlaps */
578 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
579 unsigned long overlap_len;
581 if (vma->vm_start > end)
584 overlap_len = min(end, vma->vm_end) - vma->vm_start;
585 nr_pages += overlap_len >> PAGE_SHIFT;
591 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
592 struct rb_node **rb_link, struct rb_node *rb_parent)
594 /* Update tracking information for the gap following the new vma. */
596 vma_gap_update(vma->vm_next);
598 mm->highest_vm_end = vm_end_gap(vma);
601 * vma->vm_prev wasn't known when we followed the rbtree to find the
602 * correct insertion point for that vma. As a result, we could not
603 * update the vma vm_rb parents rb_subtree_gap values on the way down.
604 * So, we first insert the vma with a zero rb_subtree_gap value
605 * (to be consistent with what we did on the way down), and then
606 * immediately update the gap to the correct value. Finally we
607 * rebalance the rbtree after all augmented values have been set.
609 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
610 vma->rb_subtree_gap = 0;
612 vma_rb_insert(vma, &mm->mm_rb);
615 static void __vma_link_file(struct vm_area_struct *vma)
621 struct address_space *mapping = file->f_mapping;
623 if (vma->vm_flags & VM_DENYWRITE)
624 atomic_dec(&file_inode(file)->i_writecount);
625 if (vma->vm_flags & VM_SHARED)
626 atomic_inc(&mapping->i_mmap_writable);
628 flush_dcache_mmap_lock(mapping);
629 vma_interval_tree_insert(vma, &mapping->i_mmap);
630 flush_dcache_mmap_unlock(mapping);
635 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
636 struct vm_area_struct *prev, struct rb_node **rb_link,
637 struct rb_node *rb_parent)
639 __vma_link_list(mm, vma, prev);
640 __vma_link_rb(mm, vma, rb_link, rb_parent);
643 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
644 struct vm_area_struct *prev, struct rb_node **rb_link,
645 struct rb_node *rb_parent)
647 struct address_space *mapping = NULL;
650 mapping = vma->vm_file->f_mapping;
651 i_mmap_lock_write(mapping);
654 __vma_link(mm, vma, prev, rb_link, rb_parent);
655 __vma_link_file(vma);
658 i_mmap_unlock_write(mapping);
665 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
666 * mm's list and rbtree. It has already been inserted into the interval tree.
668 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
670 struct vm_area_struct *prev;
671 struct rb_node **rb_link, *rb_parent;
673 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
674 &prev, &rb_link, &rb_parent))
676 __vma_link(mm, vma, prev, rb_link, rb_parent);
680 static __always_inline void __vma_unlink_common(struct mm_struct *mm,
681 struct vm_area_struct *vma,
682 struct vm_area_struct *ignore)
684 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
685 __vma_unlink_list(mm, vma);
687 vmacache_invalidate(mm);
691 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
692 * is already present in an i_mmap tree without adjusting the tree.
693 * The following helper function should be used when such adjustments
694 * are necessary. The "insert" vma (if any) is to be inserted
695 * before we drop the necessary locks.
697 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
698 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
699 struct vm_area_struct *expand)
701 struct mm_struct *mm = vma->vm_mm;
702 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
703 struct address_space *mapping = NULL;
704 struct rb_root_cached *root = NULL;
705 struct anon_vma *anon_vma = NULL;
706 struct file *file = vma->vm_file;
707 bool start_changed = false, end_changed = false;
708 long adjust_next = 0;
711 if (next && !insert) {
712 struct vm_area_struct *exporter = NULL, *importer = NULL;
714 if (end >= next->vm_end) {
716 * vma expands, overlapping all the next, and
717 * perhaps the one after too (mprotect case 6).
718 * The only other cases that gets here are
719 * case 1, case 7 and case 8.
721 if (next == expand) {
723 * The only case where we don't expand "vma"
724 * and we expand "next" instead is case 8.
726 VM_WARN_ON(end != next->vm_end);
728 * remove_next == 3 means we're
729 * removing "vma" and that to do so we
730 * swapped "vma" and "next".
733 VM_WARN_ON(file != next->vm_file);
736 VM_WARN_ON(expand != vma);
738 * case 1, 6, 7, remove_next == 2 is case 6,
739 * remove_next == 1 is case 1 or 7.
741 remove_next = 1 + (end > next->vm_end);
742 VM_WARN_ON(remove_next == 2 &&
743 end != next->vm_next->vm_end);
744 /* trim end to next, for case 6 first pass */
752 * If next doesn't have anon_vma, import from vma after
753 * next, if the vma overlaps with it.
755 if (remove_next == 2 && !next->anon_vma)
756 exporter = next->vm_next;
758 } else if (end > next->vm_start) {
760 * vma expands, overlapping part of the next:
761 * mprotect case 5 shifting the boundary up.
763 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
766 VM_WARN_ON(expand != importer);
767 } else if (end < vma->vm_end) {
769 * vma shrinks, and !insert tells it's not
770 * split_vma inserting another: so it must be
771 * mprotect case 4 shifting the boundary down.
773 adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
776 VM_WARN_ON(expand != importer);
780 * Easily overlooked: when mprotect shifts the boundary,
781 * make sure the expanding vma has anon_vma set if the
782 * shrinking vma had, to cover any anon pages imported.
784 if (exporter && exporter->anon_vma && !importer->anon_vma) {
787 importer->anon_vma = exporter->anon_vma;
788 error = anon_vma_clone(importer, exporter);
794 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
797 mapping = file->f_mapping;
798 root = &mapping->i_mmap;
799 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
802 uprobe_munmap(next, next->vm_start, next->vm_end);
804 i_mmap_lock_write(mapping);
807 * Put into interval tree now, so instantiated pages
808 * are visible to arm/parisc __flush_dcache_page
809 * throughout; but we cannot insert into address
810 * space until vma start or end is updated.
812 __vma_link_file(insert);
816 anon_vma = vma->anon_vma;
817 if (!anon_vma && adjust_next)
818 anon_vma = next->anon_vma;
820 VM_WARN_ON(adjust_next && next->anon_vma &&
821 anon_vma != next->anon_vma);
822 anon_vma_lock_write(anon_vma);
823 anon_vma_interval_tree_pre_update_vma(vma);
825 anon_vma_interval_tree_pre_update_vma(next);
829 flush_dcache_mmap_lock(mapping);
830 vma_interval_tree_remove(vma, root);
832 vma_interval_tree_remove(next, root);
835 if (start != vma->vm_start) {
836 vma->vm_start = start;
837 start_changed = true;
839 if (end != vma->vm_end) {
843 vma->vm_pgoff = pgoff;
845 next->vm_start += adjust_next << PAGE_SHIFT;
846 next->vm_pgoff += adjust_next;
851 vma_interval_tree_insert(next, root);
852 vma_interval_tree_insert(vma, root);
853 flush_dcache_mmap_unlock(mapping);
858 * vma_merge has merged next into vma, and needs
859 * us to remove next before dropping the locks.
861 if (remove_next != 3)
862 __vma_unlink_common(mm, next, next);
865 * vma is not before next if they've been
868 * pre-swap() next->vm_start was reduced so
869 * tell validate_mm_rb to ignore pre-swap()
870 * "next" (which is stored in post-swap()
873 __vma_unlink_common(mm, next, vma);
875 __remove_shared_vm_struct(next, file, mapping);
878 * split_vma has split insert from vma, and needs
879 * us to insert it before dropping the locks
880 * (it may either follow vma or precede it).
882 __insert_vm_struct(mm, insert);
888 mm->highest_vm_end = vm_end_gap(vma);
889 else if (!adjust_next)
890 vma_gap_update(next);
895 anon_vma_interval_tree_post_update_vma(vma);
897 anon_vma_interval_tree_post_update_vma(next);
898 anon_vma_unlock_write(anon_vma);
901 i_mmap_unlock_write(mapping);
912 uprobe_munmap(next, next->vm_start, next->vm_end);
916 anon_vma_merge(vma, next);
918 mpol_put(vma_policy(next));
921 * In mprotect's case 6 (see comments on vma_merge),
922 * we must remove another next too. It would clutter
923 * up the code too much to do both in one go.
925 if (remove_next != 3) {
927 * If "next" was removed and vma->vm_end was
928 * expanded (up) over it, in turn
929 * "next->vm_prev->vm_end" changed and the
930 * "vma->vm_next" gap must be updated.
935 * For the scope of the comment "next" and
936 * "vma" considered pre-swap(): if "vma" was
937 * removed, next->vm_start was expanded (down)
938 * over it and the "next" gap must be updated.
939 * Because of the swap() the post-swap() "vma"
940 * actually points to pre-swap() "next"
941 * (post-swap() "next" as opposed is now a
946 if (remove_next == 2) {
952 vma_gap_update(next);
955 * If remove_next == 2 we obviously can't
958 * If remove_next == 3 we can't reach this
959 * path because pre-swap() next is always not
960 * NULL. pre-swap() "next" is not being
961 * removed and its next->vm_end is not altered
962 * (and furthermore "end" already matches
963 * next->vm_end in remove_next == 3).
965 * We reach this only in the remove_next == 1
966 * case if the "next" vma that was removed was
967 * the highest vma of the mm. However in such
968 * case next->vm_end == "end" and the extended
969 * "vma" has vma->vm_end == next->vm_end so
970 * mm->highest_vm_end doesn't need any update
971 * in remove_next == 1 case.
973 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
985 * If the vma has a ->close operation then the driver probably needs to release
986 * per-vma resources, so we don't attempt to merge those.
988 static inline int is_mergeable_vma(struct vm_area_struct *vma,
989 struct file *file, unsigned long vm_flags,
990 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
993 * VM_SOFTDIRTY should not prevent from VMA merging, if we
994 * match the flags but dirty bit -- the caller should mark
995 * merged VMA as dirty. If dirty bit won't be excluded from
996 * comparison, we increase pressure on the memory system forcing
997 * the kernel to generate new VMAs when old one could be
1000 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
1002 if (vma->vm_file != file)
1004 if (vma->vm_ops && vma->vm_ops->close)
1006 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1011 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1012 struct anon_vma *anon_vma2,
1013 struct vm_area_struct *vma)
1016 * The list_is_singular() test is to avoid merging VMA cloned from
1017 * parents. This can improve scalability caused by anon_vma lock.
1019 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1020 list_is_singular(&vma->anon_vma_chain)))
1022 return anon_vma1 == anon_vma2;
1026 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1027 * in front of (at a lower virtual address and file offset than) the vma.
1029 * We cannot merge two vmas if they have differently assigned (non-NULL)
1030 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1032 * We don't check here for the merged mmap wrapping around the end of pagecache
1033 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
1034 * wrap, nor mmaps which cover the final page at index -1UL.
1037 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1038 struct anon_vma *anon_vma, struct file *file,
1040 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1042 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1043 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1044 if (vma->vm_pgoff == vm_pgoff)
1051 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1052 * beyond (at a higher virtual address and file offset than) the vma.
1054 * We cannot merge two vmas if they have differently assigned (non-NULL)
1055 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1058 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1059 struct anon_vma *anon_vma, struct file *file,
1061 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1063 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1064 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1066 vm_pglen = vma_pages(vma);
1067 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1074 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1075 * whether that can be merged with its predecessor or its successor.
1076 * Or both (it neatly fills a hole).
1078 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1079 * certain not to be mapped by the time vma_merge is called; but when
1080 * called for mprotect, it is certain to be already mapped (either at
1081 * an offset within prev, or at the start of next), and the flags of
1082 * this area are about to be changed to vm_flags - and the no-change
1083 * case has already been eliminated.
1085 * The following mprotect cases have to be considered, where AAAA is
1086 * the area passed down from mprotect_fixup, never extending beyond one
1087 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1090 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN
1091 * cannot merge might become might become
1092 * PPNNNNNNNNNN PPPPPPPPPPNN
1093 * mmap, brk or case 4 below case 5 below
1096 * PPPP NNNN PPPPNNNNXXXX
1097 * might become might become
1098 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
1099 * PPPPPPPPNNNN 2 or PPPPPPPPXXXX 7 or
1100 * PPPPNNNNNNNN 3 PPPPXXXXXXXX 8
1102 * It is important for case 8 that the vma NNNN overlapping the
1103 * region AAAA is never going to extended over XXXX. Instead XXXX must
1104 * be extended in region AAAA and NNNN must be removed. This way in
1105 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1106 * rmap_locks, the properties of the merged vma will be already
1107 * correct for the whole merged range. Some of those properties like
1108 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1109 * be correct for the whole merged range immediately after the
1110 * rmap_locks are released. Otherwise if XXXX would be removed and
1111 * NNNN would be extended over the XXXX range, remove_migration_ptes
1112 * or other rmap walkers (if working on addresses beyond the "end"
1113 * parameter) may establish ptes with the wrong permissions of NNNN
1114 * instead of the right permissions of XXXX.
1116 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1117 struct vm_area_struct *prev, unsigned long addr,
1118 unsigned long end, unsigned long vm_flags,
1119 struct anon_vma *anon_vma, struct file *file,
1120 pgoff_t pgoff, struct mempolicy *policy,
1121 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1123 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1124 struct vm_area_struct *area, *next;
1128 * We later require that vma->vm_flags == vm_flags,
1129 * so this tests vma->vm_flags & VM_SPECIAL, too.
1131 if (vm_flags & VM_SPECIAL)
1135 next = prev->vm_next;
1139 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1140 next = next->vm_next;
1142 /* verify some invariant that must be enforced by the caller */
1143 VM_WARN_ON(prev && addr <= prev->vm_start);
1144 VM_WARN_ON(area && end > area->vm_end);
1145 VM_WARN_ON(addr >= end);
1148 * Can it merge with the predecessor?
1150 if (prev && prev->vm_end == addr &&
1151 mpol_equal(vma_policy(prev), policy) &&
1152 can_vma_merge_after(prev, vm_flags,
1153 anon_vma, file, pgoff,
1154 vm_userfaultfd_ctx)) {
1156 * OK, it can. Can we now merge in the successor as well?
1158 if (next && end == next->vm_start &&
1159 mpol_equal(policy, vma_policy(next)) &&
1160 can_vma_merge_before(next, vm_flags,
1163 vm_userfaultfd_ctx) &&
1164 is_mergeable_anon_vma(prev->anon_vma,
1165 next->anon_vma, NULL)) {
1167 err = __vma_adjust(prev, prev->vm_start,
1168 next->vm_end, prev->vm_pgoff, NULL,
1170 } else /* cases 2, 5, 7 */
1171 err = __vma_adjust(prev, prev->vm_start,
1172 end, prev->vm_pgoff, NULL, prev);
1175 khugepaged_enter_vma_merge(prev, vm_flags);
1180 * Can this new request be merged in front of next?
1182 if (next && end == next->vm_start &&
1183 mpol_equal(policy, vma_policy(next)) &&
1184 can_vma_merge_before(next, vm_flags,
1185 anon_vma, file, pgoff+pglen,
1186 vm_userfaultfd_ctx)) {
1187 if (prev && addr < prev->vm_end) /* case 4 */
1188 err = __vma_adjust(prev, prev->vm_start,
1189 addr, prev->vm_pgoff, NULL, next);
1190 else { /* cases 3, 8 */
1191 err = __vma_adjust(area, addr, next->vm_end,
1192 next->vm_pgoff - pglen, NULL, next);
1194 * In case 3 area is already equal to next and
1195 * this is a noop, but in case 8 "area" has
1196 * been removed and next was expanded over it.
1202 khugepaged_enter_vma_merge(area, vm_flags);
1210 * Rough compatibility check to quickly see if it's even worth looking
1211 * at sharing an anon_vma.
1213 * They need to have the same vm_file, and the flags can only differ
1214 * in things that mprotect may change.
1216 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1217 * we can merge the two vma's. For example, we refuse to merge a vma if
1218 * there is a vm_ops->close() function, because that indicates that the
1219 * driver is doing some kind of reference counting. But that doesn't
1220 * really matter for the anon_vma sharing case.
1222 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1224 return a->vm_end == b->vm_start &&
1225 mpol_equal(vma_policy(a), vma_policy(b)) &&
1226 a->vm_file == b->vm_file &&
1227 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1228 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1232 * Do some basic sanity checking to see if we can re-use the anon_vma
1233 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1234 * the same as 'old', the other will be the new one that is trying
1235 * to share the anon_vma.
1237 * NOTE! This runs with mm_sem held for reading, so it is possible that
1238 * the anon_vma of 'old' is concurrently in the process of being set up
1239 * by another page fault trying to merge _that_. But that's ok: if it
1240 * is being set up, that automatically means that it will be a singleton
1241 * acceptable for merging, so we can do all of this optimistically. But
1242 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1244 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1245 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1246 * is to return an anon_vma that is "complex" due to having gone through
1249 * We also make sure that the two vma's are compatible (adjacent,
1250 * and with the same memory policies). That's all stable, even with just
1251 * a read lock on the mm_sem.
1253 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1255 if (anon_vma_compatible(a, b)) {
1256 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1258 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1265 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1266 * neighbouring vmas for a suitable anon_vma, before it goes off
1267 * to allocate a new anon_vma. It checks because a repetitive
1268 * sequence of mprotects and faults may otherwise lead to distinct
1269 * anon_vmas being allocated, preventing vma merge in subsequent
1272 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1274 struct anon_vma *anon_vma = NULL;
1276 /* Try next first. */
1278 anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next);
1283 /* Try prev next. */
1285 anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma);
1288 * We might reach here with anon_vma == NULL if we can't find
1289 * any reusable anon_vma.
1290 * There's no absolute need to look only at touching neighbours:
1291 * we could search further afield for "compatible" anon_vmas.
1292 * But it would probably just be a waste of time searching,
1293 * or lead to too many vmas hanging off the same anon_vma.
1294 * We're trying to allow mprotect remerging later on,
1295 * not trying to minimize memory used for anon_vmas.
1301 * If a hint addr is less than mmap_min_addr change hint to be as
1302 * low as possible but still greater than mmap_min_addr
1304 static inline unsigned long round_hint_to_min(unsigned long hint)
1307 if (((void *)hint != NULL) &&
1308 (hint < mmap_min_addr))
1309 return PAGE_ALIGN(mmap_min_addr);
1313 static inline int mlock_future_check(struct mm_struct *mm,
1314 unsigned long flags,
1317 unsigned long locked, lock_limit;
1319 /* mlock MCL_FUTURE? */
1320 if (flags & VM_LOCKED) {
1321 locked = len >> PAGE_SHIFT;
1322 locked += mm->locked_vm;
1323 lock_limit = rlimit(RLIMIT_MEMLOCK);
1324 lock_limit >>= PAGE_SHIFT;
1325 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1331 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1333 if (S_ISREG(inode->i_mode))
1334 return MAX_LFS_FILESIZE;
1336 if (S_ISBLK(inode->i_mode))
1337 return MAX_LFS_FILESIZE;
1339 if (S_ISSOCK(inode->i_mode))
1340 return MAX_LFS_FILESIZE;
1342 /* Special "we do even unsigned file positions" case */
1343 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1346 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1350 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1351 unsigned long pgoff, unsigned long len)
1353 u64 maxsize = file_mmap_size_max(file, inode);
1355 if (maxsize && len > maxsize)
1358 if (pgoff > maxsize >> PAGE_SHIFT)
1364 * The caller must write-lock current->mm->mmap_lock.
1366 unsigned long do_mmap(struct file *file, unsigned long addr,
1367 unsigned long len, unsigned long prot,
1368 unsigned long flags, unsigned long pgoff,
1369 unsigned long *populate, struct list_head *uf)
1371 struct mm_struct *mm = current->mm;
1372 vm_flags_t vm_flags;
1381 * Does the application expect PROT_READ to imply PROT_EXEC?
1383 * (the exception is when the underlying filesystem is noexec
1384 * mounted, in which case we dont add PROT_EXEC.)
1386 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1387 if (!(file && path_noexec(&file->f_path)))
1390 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1391 if (flags & MAP_FIXED_NOREPLACE)
1394 if (!(flags & MAP_FIXED))
1395 addr = round_hint_to_min(addr);
1397 /* Careful about overflows.. */
1398 len = PAGE_ALIGN(len);
1402 /* offset overflow? */
1403 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1406 /* Too many mappings? */
1407 if (mm->map_count > sysctl_max_map_count)
1410 /* Obtain the address to map to. we verify (or select) it and ensure
1411 * that it represents a valid section of the address space.
1413 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1414 if (IS_ERR_VALUE(addr))
1417 if (flags & MAP_FIXED_NOREPLACE) {
1418 struct vm_area_struct *vma = find_vma(mm, addr);
1420 if (vma && vma->vm_start < addr + len)
1424 if (prot == PROT_EXEC) {
1425 pkey = execute_only_pkey(mm);
1430 /* Do simple checking here so the lower-level routines won't have
1431 * to. we assume access permissions have been handled by the open
1432 * of the memory object, so we don't do any here.
1434 vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1435 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1437 if (flags & MAP_LOCKED)
1438 if (!can_do_mlock())
1441 if (mlock_future_check(mm, vm_flags, len))
1445 struct inode *inode = file_inode(file);
1446 unsigned long flags_mask;
1448 if (!file_mmap_ok(file, inode, pgoff, len))
1451 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1453 switch (flags & MAP_TYPE) {
1456 * Force use of MAP_SHARED_VALIDATE with non-legacy
1457 * flags. E.g. MAP_SYNC is dangerous to use with
1458 * MAP_SHARED as you don't know which consistency model
1459 * you will get. We silently ignore unsupported flags
1460 * with MAP_SHARED to preserve backward compatibility.
1462 flags &= LEGACY_MAP_MASK;
1464 case MAP_SHARED_VALIDATE:
1465 if (flags & ~flags_mask)
1467 if (prot & PROT_WRITE) {
1468 if (!(file->f_mode & FMODE_WRITE))
1470 if (IS_SWAPFILE(file->f_mapping->host))
1475 * Make sure we don't allow writing to an append-only
1478 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1482 * Make sure there are no mandatory locks on the file.
1484 if (locks_verify_locked(file))
1487 vm_flags |= VM_SHARED | VM_MAYSHARE;
1488 if (!(file->f_mode & FMODE_WRITE))
1489 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1492 if (!(file->f_mode & FMODE_READ))
1494 if (path_noexec(&file->f_path)) {
1495 if (vm_flags & VM_EXEC)
1497 vm_flags &= ~VM_MAYEXEC;
1500 if (!file->f_op->mmap)
1502 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1510 switch (flags & MAP_TYPE) {
1512 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1518 vm_flags |= VM_SHARED | VM_MAYSHARE;
1522 * Set pgoff according to addr for anon_vma.
1524 pgoff = addr >> PAGE_SHIFT;
1532 * Set 'VM_NORESERVE' if we should not account for the
1533 * memory use of this mapping.
1535 if (flags & MAP_NORESERVE) {
1536 /* We honor MAP_NORESERVE if allowed to overcommit */
1537 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1538 vm_flags |= VM_NORESERVE;
1540 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1541 if (file && is_file_hugepages(file))
1542 vm_flags |= VM_NORESERVE;
1545 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1546 if (!IS_ERR_VALUE(addr) &&
1547 ((vm_flags & VM_LOCKED) ||
1548 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1553 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1554 unsigned long prot, unsigned long flags,
1555 unsigned long fd, unsigned long pgoff)
1557 struct file *file = NULL;
1558 unsigned long retval;
1560 if (!(flags & MAP_ANONYMOUS)) {
1561 audit_mmap_fd(fd, flags);
1565 if (is_file_hugepages(file)) {
1566 len = ALIGN(len, huge_page_size(hstate_file(file)));
1567 } else if (unlikely(flags & MAP_HUGETLB)) {
1571 } else if (flags & MAP_HUGETLB) {
1572 struct user_struct *user = NULL;
1575 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1579 len = ALIGN(len, huge_page_size(hs));
1581 * VM_NORESERVE is used because the reservations will be
1582 * taken when vm_ops->mmap() is called
1583 * A dummy user value is used because we are not locking
1584 * memory so no accounting is necessary
1586 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1588 &user, HUGETLB_ANONHUGE_INODE,
1589 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1591 return PTR_ERR(file);
1594 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1596 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1603 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1604 unsigned long, prot, unsigned long, flags,
1605 unsigned long, fd, unsigned long, pgoff)
1607 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1610 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1611 struct mmap_arg_struct {
1615 unsigned long flags;
1617 unsigned long offset;
1620 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1622 struct mmap_arg_struct a;
1624 if (copy_from_user(&a, arg, sizeof(a)))
1626 if (offset_in_page(a.offset))
1629 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1630 a.offset >> PAGE_SHIFT);
1632 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1635 * Some shared mappings will want the pages marked read-only
1636 * to track write events. If so, we'll downgrade vm_page_prot
1637 * to the private version (using protection_map[] without the
1640 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1642 vm_flags_t vm_flags = vma->vm_flags;
1643 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1645 /* If it was private or non-writable, the write bit is already clear */
1646 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1649 /* The backer wishes to know when pages are first written to? */
1650 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1653 /* The open routine did something to the protections that pgprot_modify
1654 * won't preserve? */
1655 if (pgprot_val(vm_page_prot) !=
1656 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1659 /* Do we need to track softdirty? */
1660 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1663 /* Specialty mapping? */
1664 if (vm_flags & VM_PFNMAP)
1667 /* Can the mapping track the dirty pages? */
1668 return vma->vm_file && vma->vm_file->f_mapping &&
1669 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1673 * We account for memory if it's a private writeable mapping,
1674 * not hugepages and VM_NORESERVE wasn't set.
1676 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1679 * hugetlb has its own accounting separate from the core VM
1680 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1682 if (file && is_file_hugepages(file))
1685 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1688 unsigned long mmap_region(struct file *file, unsigned long addr,
1689 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1690 struct list_head *uf)
1692 struct mm_struct *mm = current->mm;
1693 struct vm_area_struct *vma, *prev, *merge;
1695 struct rb_node **rb_link, *rb_parent;
1696 unsigned long charged = 0;
1698 /* Check against address space limit. */
1699 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1700 unsigned long nr_pages;
1703 * MAP_FIXED may remove pages of mappings that intersects with
1704 * requested mapping. Account for the pages it would unmap.
1706 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1708 if (!may_expand_vm(mm, vm_flags,
1709 (len >> PAGE_SHIFT) - nr_pages))
1713 /* Clear old maps */
1714 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1716 if (do_munmap(mm, addr, len, uf))
1721 * Private writable mapping: check memory availability
1723 if (accountable_mapping(file, vm_flags)) {
1724 charged = len >> PAGE_SHIFT;
1725 if (security_vm_enough_memory_mm(mm, charged))
1727 vm_flags |= VM_ACCOUNT;
1731 * Can we just expand an old mapping?
1733 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1734 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1739 * Determine the object being mapped and call the appropriate
1740 * specific mapper. the address has already been validated, but
1741 * not unmapped, but the maps are removed from the list.
1743 vma = vm_area_alloc(mm);
1749 vma->vm_start = addr;
1750 vma->vm_end = addr + len;
1751 vma->vm_flags = vm_flags;
1752 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1753 vma->vm_pgoff = pgoff;
1756 if (vm_flags & VM_DENYWRITE) {
1757 error = deny_write_access(file);
1761 if (vm_flags & VM_SHARED) {
1762 error = mapping_map_writable(file->f_mapping);
1764 goto allow_write_and_free_vma;
1767 /* ->mmap() can change vma->vm_file, but must guarantee that
1768 * vma_link() below can deny write-access if VM_DENYWRITE is set
1769 * and map writably if VM_SHARED is set. This usually means the
1770 * new file must not have been exposed to user-space, yet.
1772 vma->vm_file = get_file(file);
1773 error = call_mmap(file, vma);
1775 goto unmap_and_free_vma;
1777 /* If vm_flags changed after call_mmap(), we should try merge vma again
1778 * as we may succeed this time.
1780 if (unlikely(vm_flags != vma->vm_flags && prev)) {
1781 merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
1782 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX);
1784 /* ->mmap() can change vma->vm_file and fput the original file. So
1785 * fput the vma->vm_file here or we would add an extra fput for file
1786 * and cause general protection fault ultimately.
1791 /* Update vm_flags and possible addr to pick up the change. We don't
1792 * warn here if addr changed as the vma is not linked by vma_link().
1794 addr = vma->vm_start;
1795 vm_flags = vma->vm_flags;
1796 goto unmap_writable;
1800 /* Can addr have changed??
1802 * Answer: Yes, several device drivers can do it in their
1803 * f_op->mmap method. -DaveM
1804 * Bug: If addr is changed, prev, rb_link, rb_parent should
1805 * be updated for vma_link()
1807 WARN_ON_ONCE(addr != vma->vm_start);
1809 addr = vma->vm_start;
1810 vm_flags = vma->vm_flags;
1811 } else if (vm_flags & VM_SHARED) {
1812 error = shmem_zero_setup(vma);
1816 vma_set_anonymous(vma);
1819 vma_link(mm, vma, prev, rb_link, rb_parent);
1820 /* Once vma denies write, undo our temporary denial count */
1823 if (vm_flags & VM_SHARED)
1824 mapping_unmap_writable(file->f_mapping);
1825 if (vm_flags & VM_DENYWRITE)
1826 allow_write_access(file);
1828 file = vma->vm_file;
1830 perf_event_mmap(vma);
1832 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1833 if (vm_flags & VM_LOCKED) {
1834 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1835 is_vm_hugetlb_page(vma) ||
1836 vma == get_gate_vma(current->mm))
1837 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1839 mm->locked_vm += (len >> PAGE_SHIFT);
1846 * New (or expanded) vma always get soft dirty status.
1847 * Otherwise user-space soft-dirty page tracker won't
1848 * be able to distinguish situation when vma area unmapped,
1849 * then new mapped in-place (which must be aimed as
1850 * a completely new data area).
1852 vma->vm_flags |= VM_SOFTDIRTY;
1854 vma_set_page_prot(vma);
1859 vma->vm_file = NULL;
1862 /* Undo any partial mapping done by a device driver. */
1863 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1865 if (vm_flags & VM_SHARED)
1866 mapping_unmap_writable(file->f_mapping);
1867 allow_write_and_free_vma:
1868 if (vm_flags & VM_DENYWRITE)
1869 allow_write_access(file);
1874 vm_unacct_memory(charged);
1878 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1881 * We implement the search by looking for an rbtree node that
1882 * immediately follows a suitable gap. That is,
1883 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1884 * - gap_end = vma->vm_start >= info->low_limit + length;
1885 * - gap_end - gap_start >= length
1888 struct mm_struct *mm = current->mm;
1889 struct vm_area_struct *vma;
1890 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1892 /* Adjust search length to account for worst case alignment overhead */
1893 length = info->length + info->align_mask;
1894 if (length < info->length)
1897 /* Adjust search limits by the desired length */
1898 if (info->high_limit < length)
1900 high_limit = info->high_limit - length;
1902 if (info->low_limit > high_limit)
1904 low_limit = info->low_limit + length;
1906 /* Check if rbtree root looks promising */
1907 if (RB_EMPTY_ROOT(&mm->mm_rb))
1909 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1910 if (vma->rb_subtree_gap < length)
1914 /* Visit left subtree if it looks promising */
1915 gap_end = vm_start_gap(vma);
1916 if (gap_end >= low_limit && 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 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1928 /* Check if current node has a suitable gap */
1929 if (gap_start > high_limit)
1931 if (gap_end >= low_limit &&
1932 gap_end > gap_start && gap_end - gap_start >= length)
1935 /* Visit right subtree if it looks promising */
1936 if (vma->vm_rb.rb_right) {
1937 struct vm_area_struct *right =
1938 rb_entry(vma->vm_rb.rb_right,
1939 struct vm_area_struct, vm_rb);
1940 if (right->rb_subtree_gap >= length) {
1946 /* Go back up the rbtree to find next candidate node */
1948 struct rb_node *prev = &vma->vm_rb;
1949 if (!rb_parent(prev))
1951 vma = rb_entry(rb_parent(prev),
1952 struct vm_area_struct, vm_rb);
1953 if (prev == vma->vm_rb.rb_left) {
1954 gap_start = vm_end_gap(vma->vm_prev);
1955 gap_end = vm_start_gap(vma);
1962 /* Check highest gap, which does not precede any rbtree node */
1963 gap_start = mm->highest_vm_end;
1964 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1965 if (gap_start > high_limit)
1969 /* We found a suitable gap. Clip it with the original low_limit. */
1970 if (gap_start < info->low_limit)
1971 gap_start = info->low_limit;
1973 /* Adjust gap address to the desired alignment */
1974 gap_start += (info->align_offset - gap_start) & info->align_mask;
1976 VM_BUG_ON(gap_start + info->length > info->high_limit);
1977 VM_BUG_ON(gap_start + info->length > gap_end);
1981 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1983 struct mm_struct *mm = current->mm;
1984 struct vm_area_struct *vma;
1985 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1987 /* Adjust search length to account for worst case alignment overhead */
1988 length = info->length + info->align_mask;
1989 if (length < info->length)
1993 * Adjust search limits by the desired length.
1994 * See implementation comment at top of unmapped_area().
1996 gap_end = info->high_limit;
1997 if (gap_end < length)
1999 high_limit = gap_end - length;
2001 if (info->low_limit > high_limit)
2003 low_limit = info->low_limit + length;
2005 /* Check highest gap, which does not precede any rbtree node */
2006 gap_start = mm->highest_vm_end;
2007 if (gap_start <= high_limit)
2010 /* Check if rbtree root looks promising */
2011 if (RB_EMPTY_ROOT(&mm->mm_rb))
2013 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
2014 if (vma->rb_subtree_gap < length)
2018 /* Visit right subtree if it looks promising */
2019 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
2020 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
2021 struct vm_area_struct *right =
2022 rb_entry(vma->vm_rb.rb_right,
2023 struct vm_area_struct, vm_rb);
2024 if (right->rb_subtree_gap >= length) {
2031 /* Check if current node has a suitable gap */
2032 gap_end = vm_start_gap(vma);
2033 if (gap_end < low_limit)
2035 if (gap_start <= high_limit &&
2036 gap_end > gap_start && gap_end - gap_start >= length)
2039 /* Visit left subtree if it looks promising */
2040 if (vma->vm_rb.rb_left) {
2041 struct vm_area_struct *left =
2042 rb_entry(vma->vm_rb.rb_left,
2043 struct vm_area_struct, vm_rb);
2044 if (left->rb_subtree_gap >= length) {
2050 /* Go back up the rbtree to find next candidate node */
2052 struct rb_node *prev = &vma->vm_rb;
2053 if (!rb_parent(prev))
2055 vma = rb_entry(rb_parent(prev),
2056 struct vm_area_struct, vm_rb);
2057 if (prev == vma->vm_rb.rb_right) {
2058 gap_start = vma->vm_prev ?
2059 vm_end_gap(vma->vm_prev) : 0;
2066 /* We found a suitable gap. Clip it with the original high_limit. */
2067 if (gap_end > info->high_limit)
2068 gap_end = info->high_limit;
2071 /* Compute highest gap address at the desired alignment */
2072 gap_end -= info->length;
2073 gap_end -= (gap_end - info->align_offset) & info->align_mask;
2075 VM_BUG_ON(gap_end < info->low_limit);
2076 VM_BUG_ON(gap_end < gap_start);
2081 * Search for an unmapped address range.
2083 * We are looking for a range that:
2084 * - does not intersect with any VMA;
2085 * - is contained within the [low_limit, high_limit) interval;
2086 * - is at least the desired size.
2087 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
2089 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
2093 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
2094 addr = unmapped_area_topdown(info);
2096 addr = unmapped_area(info);
2098 trace_vm_unmapped_area(addr, info);
2102 #ifndef arch_get_mmap_end
2103 #define arch_get_mmap_end(addr) (TASK_SIZE)
2106 #ifndef arch_get_mmap_base
2107 #define arch_get_mmap_base(addr, base) (base)
2110 /* Get an address range which is currently unmapped.
2111 * For shmat() with addr=0.
2113 * Ugly calling convention alert:
2114 * Return value with the low bits set means error value,
2116 * if (ret & ~PAGE_MASK)
2119 * This function "knows" that -ENOMEM has the bits set.
2121 #ifndef HAVE_ARCH_UNMAPPED_AREA
2123 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2124 unsigned long len, unsigned long pgoff, unsigned long flags)
2126 struct mm_struct *mm = current->mm;
2127 struct vm_area_struct *vma, *prev;
2128 struct vm_unmapped_area_info info;
2129 const unsigned long mmap_end = arch_get_mmap_end(addr);
2131 if (len > mmap_end - mmap_min_addr)
2134 if (flags & MAP_FIXED)
2138 addr = PAGE_ALIGN(addr);
2139 vma = find_vma_prev(mm, addr, &prev);
2140 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2141 (!vma || addr + len <= vm_start_gap(vma)) &&
2142 (!prev || addr >= vm_end_gap(prev)))
2148 info.low_limit = mm->mmap_base;
2149 info.high_limit = mmap_end;
2150 info.align_mask = 0;
2151 info.align_offset = 0;
2152 return vm_unmapped_area(&info);
2157 * This mmap-allocator allocates new areas top-down from below the
2158 * stack's low limit (the base):
2160 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2162 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2163 unsigned long len, unsigned long pgoff,
2164 unsigned long flags)
2166 struct vm_area_struct *vma, *prev;
2167 struct mm_struct *mm = current->mm;
2168 struct vm_unmapped_area_info info;
2169 const unsigned long mmap_end = arch_get_mmap_end(addr);
2171 /* requested length too big for entire address space */
2172 if (len > mmap_end - mmap_min_addr)
2175 if (flags & MAP_FIXED)
2178 /* requesting a specific address */
2180 addr = PAGE_ALIGN(addr);
2181 vma = find_vma_prev(mm, addr, &prev);
2182 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2183 (!vma || addr + len <= vm_start_gap(vma)) &&
2184 (!prev || addr >= vm_end_gap(prev)))
2188 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2190 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2191 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2192 info.align_mask = 0;
2193 info.align_offset = 0;
2194 addr = vm_unmapped_area(&info);
2197 * A failed mmap() very likely causes application failure,
2198 * so fall back to the bottom-up function here. This scenario
2199 * can happen with large stack limits and large mmap()
2202 if (offset_in_page(addr)) {
2203 VM_BUG_ON(addr != -ENOMEM);
2205 info.low_limit = TASK_UNMAPPED_BASE;
2206 info.high_limit = mmap_end;
2207 addr = vm_unmapped_area(&info);
2215 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2216 unsigned long pgoff, unsigned long flags)
2218 unsigned long (*get_area)(struct file *, unsigned long,
2219 unsigned long, unsigned long, unsigned long);
2221 unsigned long error = arch_mmap_check(addr, len, flags);
2225 /* Careful about overflows.. */
2226 if (len > TASK_SIZE)
2229 get_area = current->mm->get_unmapped_area;
2231 if (file->f_op->get_unmapped_area)
2232 get_area = file->f_op->get_unmapped_area;
2233 } else if (flags & MAP_SHARED) {
2235 * mmap_region() will call shmem_zero_setup() to create a file,
2236 * so use shmem's get_unmapped_area in case it can be huge.
2237 * do_mmap() will clear pgoff, so match alignment.
2240 get_area = shmem_get_unmapped_area;
2243 addr = get_area(file, addr, len, pgoff, flags);
2244 if (IS_ERR_VALUE(addr))
2247 if (addr > TASK_SIZE - len)
2249 if (offset_in_page(addr))
2252 error = security_mmap_addr(addr);
2253 return error ? error : addr;
2256 EXPORT_SYMBOL(get_unmapped_area);
2258 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2259 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2261 struct rb_node *rb_node;
2262 struct vm_area_struct *vma;
2264 /* Check the cache first. */
2265 vma = vmacache_find(mm, addr);
2269 rb_node = mm->mm_rb.rb_node;
2272 struct vm_area_struct *tmp;
2274 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2276 if (tmp->vm_end > addr) {
2278 if (tmp->vm_start <= addr)
2280 rb_node = rb_node->rb_left;
2282 rb_node = rb_node->rb_right;
2286 vmacache_update(addr, vma);
2290 EXPORT_SYMBOL(find_vma);
2293 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2295 struct vm_area_struct *
2296 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2297 struct vm_area_struct **pprev)
2299 struct vm_area_struct *vma;
2301 vma = find_vma(mm, addr);
2303 *pprev = vma->vm_prev;
2305 struct rb_node *rb_node = rb_last(&mm->mm_rb);
2307 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
2313 * Verify that the stack growth is acceptable and
2314 * update accounting. This is shared with both the
2315 * grow-up and grow-down cases.
2317 static int acct_stack_growth(struct vm_area_struct *vma,
2318 unsigned long size, unsigned long grow)
2320 struct mm_struct *mm = vma->vm_mm;
2321 unsigned long new_start;
2323 /* address space limit tests */
2324 if (!may_expand_vm(mm, vma->vm_flags, grow))
2327 /* Stack limit test */
2328 if (size > rlimit(RLIMIT_STACK))
2331 /* mlock limit tests */
2332 if (vma->vm_flags & VM_LOCKED) {
2333 unsigned long locked;
2334 unsigned long limit;
2335 locked = mm->locked_vm + grow;
2336 limit = rlimit(RLIMIT_MEMLOCK);
2337 limit >>= PAGE_SHIFT;
2338 if (locked > limit && !capable(CAP_IPC_LOCK))
2342 /* Check to ensure the stack will not grow into a hugetlb-only region */
2343 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2345 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2349 * Overcommit.. This must be the final test, as it will
2350 * update security statistics.
2352 if (security_vm_enough_memory_mm(mm, grow))
2358 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2360 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2361 * vma is the last one with address > vma->vm_end. Have to extend vma.
2363 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2365 struct mm_struct *mm = vma->vm_mm;
2366 struct vm_area_struct *next;
2367 unsigned long gap_addr;
2370 if (!(vma->vm_flags & VM_GROWSUP))
2373 /* Guard against exceeding limits of the address space. */
2374 address &= PAGE_MASK;
2375 if (address >= (TASK_SIZE & PAGE_MASK))
2377 address += PAGE_SIZE;
2379 /* Enforce stack_guard_gap */
2380 gap_addr = address + stack_guard_gap;
2382 /* Guard against overflow */
2383 if (gap_addr < address || gap_addr > TASK_SIZE)
2384 gap_addr = TASK_SIZE;
2386 next = vma->vm_next;
2387 if (next && next->vm_start < gap_addr && vma_is_accessible(next)) {
2388 if (!(next->vm_flags & VM_GROWSUP))
2390 /* Check that both stack segments have the same anon_vma? */
2393 /* We must make sure the anon_vma is allocated. */
2394 if (unlikely(anon_vma_prepare(vma)))
2398 * vma->vm_start/vm_end cannot change under us because the caller
2399 * is required to hold the mmap_lock in read mode. We need the
2400 * anon_vma lock to serialize against concurrent expand_stacks.
2402 anon_vma_lock_write(vma->anon_vma);
2404 /* Somebody else might have raced and expanded it already */
2405 if (address > vma->vm_end) {
2406 unsigned long size, grow;
2408 size = address - vma->vm_start;
2409 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2412 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2413 error = acct_stack_growth(vma, size, grow);
2416 * vma_gap_update() doesn't support concurrent
2417 * updates, but we only hold a shared mmap_lock
2418 * lock here, so we need to protect against
2419 * concurrent vma expansions.
2420 * anon_vma_lock_write() doesn't help here, as
2421 * we don't guarantee that all growable vmas
2422 * in a mm share the same root anon vma.
2423 * So, we reuse mm->page_table_lock to guard
2424 * against concurrent vma expansions.
2426 spin_lock(&mm->page_table_lock);
2427 if (vma->vm_flags & VM_LOCKED)
2428 mm->locked_vm += grow;
2429 vm_stat_account(mm, vma->vm_flags, grow);
2430 anon_vma_interval_tree_pre_update_vma(vma);
2431 vma->vm_end = address;
2432 anon_vma_interval_tree_post_update_vma(vma);
2434 vma_gap_update(vma->vm_next);
2436 mm->highest_vm_end = vm_end_gap(vma);
2437 spin_unlock(&mm->page_table_lock);
2439 perf_event_mmap(vma);
2443 anon_vma_unlock_write(vma->anon_vma);
2444 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2448 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2451 * vma is the first one with address < vma->vm_start. Have to extend vma.
2453 int expand_downwards(struct vm_area_struct *vma,
2454 unsigned long address)
2456 struct mm_struct *mm = vma->vm_mm;
2457 struct vm_area_struct *prev;
2460 address &= PAGE_MASK;
2461 if (address < mmap_min_addr)
2464 /* Enforce stack_guard_gap */
2465 prev = vma->vm_prev;
2466 /* Check that both stack segments have the same anon_vma? */
2467 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2468 vma_is_accessible(prev)) {
2469 if (address - prev->vm_end < stack_guard_gap)
2473 /* We must make sure the anon_vma is allocated. */
2474 if (unlikely(anon_vma_prepare(vma)))
2478 * vma->vm_start/vm_end cannot change under us because the caller
2479 * is required to hold the mmap_lock in read mode. We need the
2480 * anon_vma lock to serialize against concurrent expand_stacks.
2482 anon_vma_lock_write(vma->anon_vma);
2484 /* Somebody else might have raced and expanded it already */
2485 if (address < vma->vm_start) {
2486 unsigned long size, grow;
2488 size = vma->vm_end - address;
2489 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2492 if (grow <= vma->vm_pgoff) {
2493 error = acct_stack_growth(vma, size, grow);
2496 * vma_gap_update() doesn't support concurrent
2497 * updates, but we only hold a shared mmap_lock
2498 * lock here, so we need to protect against
2499 * concurrent vma expansions.
2500 * anon_vma_lock_write() doesn't help here, as
2501 * we don't guarantee that all growable vmas
2502 * in a mm share the same root anon vma.
2503 * So, we reuse mm->page_table_lock to guard
2504 * against concurrent vma expansions.
2506 spin_lock(&mm->page_table_lock);
2507 if (vma->vm_flags & VM_LOCKED)
2508 mm->locked_vm += grow;
2509 vm_stat_account(mm, vma->vm_flags, grow);
2510 anon_vma_interval_tree_pre_update_vma(vma);
2511 vma->vm_start = address;
2512 vma->vm_pgoff -= grow;
2513 anon_vma_interval_tree_post_update_vma(vma);
2514 vma_gap_update(vma);
2515 spin_unlock(&mm->page_table_lock);
2517 perf_event_mmap(vma);
2521 anon_vma_unlock_write(vma->anon_vma);
2522 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2527 /* enforced gap between the expanding stack and other mappings. */
2528 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2530 static int __init cmdline_parse_stack_guard_gap(char *p)
2535 val = simple_strtoul(p, &endptr, 10);
2537 stack_guard_gap = val << PAGE_SHIFT;
2541 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2543 #ifdef CONFIG_STACK_GROWSUP
2544 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2546 return expand_upwards(vma, address);
2549 struct vm_area_struct *
2550 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2552 struct vm_area_struct *vma, *prev;
2555 vma = find_vma_prev(mm, addr, &prev);
2556 if (vma && (vma->vm_start <= addr))
2558 /* don't alter vm_end if the coredump is running */
2559 if (!prev || !mmget_still_valid(mm) || expand_stack(prev, addr))
2561 if (prev->vm_flags & VM_LOCKED)
2562 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2566 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2568 return expand_downwards(vma, address);
2571 struct vm_area_struct *
2572 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2574 struct vm_area_struct *vma;
2575 unsigned long start;
2578 vma = find_vma(mm, addr);
2581 if (vma->vm_start <= addr)
2583 if (!(vma->vm_flags & VM_GROWSDOWN))
2585 /* don't alter vm_start if the coredump is running */
2586 if (!mmget_still_valid(mm))
2588 start = vma->vm_start;
2589 if (expand_stack(vma, addr))
2591 if (vma->vm_flags & VM_LOCKED)
2592 populate_vma_page_range(vma, addr, start, NULL);
2597 EXPORT_SYMBOL_GPL(find_extend_vma);
2600 * Ok - we have the memory areas we should free on the vma list,
2601 * so release them, and do the vma updates.
2603 * Called with the mm semaphore held.
2605 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2607 unsigned long nr_accounted = 0;
2609 /* Update high watermark before we lower total_vm */
2610 update_hiwater_vm(mm);
2612 long nrpages = vma_pages(vma);
2614 if (vma->vm_flags & VM_ACCOUNT)
2615 nr_accounted += nrpages;
2616 vm_stat_account(mm, vma->vm_flags, -nrpages);
2617 vma = remove_vma(vma);
2619 vm_unacct_memory(nr_accounted);
2624 * Get rid of page table information in the indicated region.
2626 * Called with the mm semaphore held.
2628 static void unmap_region(struct mm_struct *mm,
2629 struct vm_area_struct *vma, struct vm_area_struct *prev,
2630 unsigned long start, unsigned long end)
2632 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2633 struct mmu_gather tlb;
2636 tlb_gather_mmu(&tlb, mm, start, end);
2637 update_hiwater_rss(mm);
2638 unmap_vmas(&tlb, vma, start, end);
2639 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2640 next ? next->vm_start : USER_PGTABLES_CEILING);
2641 tlb_finish_mmu(&tlb, start, end);
2645 * Create a list of vma's touched by the unmap, removing them from the mm's
2646 * vma list as we go..
2649 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2650 struct vm_area_struct *prev, unsigned long end)
2652 struct vm_area_struct **insertion_point;
2653 struct vm_area_struct *tail_vma = NULL;
2655 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2656 vma->vm_prev = NULL;
2658 vma_rb_erase(vma, &mm->mm_rb);
2662 } while (vma && vma->vm_start < end);
2663 *insertion_point = vma;
2665 vma->vm_prev = prev;
2666 vma_gap_update(vma);
2668 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2669 tail_vma->vm_next = NULL;
2671 /* Kill the cache */
2672 vmacache_invalidate(mm);
2675 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2676 * VM_GROWSUP VMA. Such VMAs can change their size under
2677 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2679 if (vma && (vma->vm_flags & VM_GROWSDOWN))
2681 if (prev && (prev->vm_flags & VM_GROWSUP))
2687 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2688 * has already been checked or doesn't make sense to fail.
2690 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2691 unsigned long addr, int new_below)
2693 struct vm_area_struct *new;
2696 if (vma->vm_ops && vma->vm_ops->split) {
2697 err = vma->vm_ops->split(vma, addr);
2702 new = vm_area_dup(vma);
2709 new->vm_start = addr;
2710 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2713 err = vma_dup_policy(vma, new);
2717 err = anon_vma_clone(new, vma);
2722 get_file(new->vm_file);
2724 if (new->vm_ops && new->vm_ops->open)
2725 new->vm_ops->open(new);
2728 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2729 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2731 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2737 /* Clean everything up if vma_adjust failed. */
2738 if (new->vm_ops && new->vm_ops->close)
2739 new->vm_ops->close(new);
2742 unlink_anon_vmas(new);
2744 mpol_put(vma_policy(new));
2751 * Split a vma into two pieces at address 'addr', a new vma is allocated
2752 * either for the first part or the tail.
2754 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2755 unsigned long addr, int new_below)
2757 if (mm->map_count >= sysctl_max_map_count)
2760 return __split_vma(mm, vma, addr, new_below);
2763 /* Munmap is split into 2 main parts -- this part which finds
2764 * what needs doing, and the areas themselves, which do the
2765 * work. This now handles partial unmappings.
2766 * Jeremy Fitzhardinge <jeremy@goop.org>
2768 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2769 struct list_head *uf, bool downgrade)
2772 struct vm_area_struct *vma, *prev, *last;
2774 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2777 len = PAGE_ALIGN(len);
2783 * arch_unmap() might do unmaps itself. It must be called
2784 * and finish any rbtree manipulation before this code
2785 * runs and also starts to manipulate the rbtree.
2787 arch_unmap(mm, start, end);
2789 /* Find the first overlapping VMA */
2790 vma = find_vma(mm, start);
2793 prev = vma->vm_prev;
2794 /* we have start < vma->vm_end */
2796 /* if it doesn't overlap, we have nothing.. */
2797 if (vma->vm_start >= end)
2801 * If we need to split any vma, do it now to save pain later.
2803 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2804 * unmapped vm_area_struct will remain in use: so lower split_vma
2805 * places tmp vma above, and higher split_vma places tmp vma below.
2807 if (start > vma->vm_start) {
2811 * Make sure that map_count on return from munmap() will
2812 * not exceed its limit; but let map_count go just above
2813 * its limit temporarily, to help free resources as expected.
2815 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2818 error = __split_vma(mm, vma, start, 0);
2824 /* Does it split the last one? */
2825 last = find_vma(mm, end);
2826 if (last && end > last->vm_start) {
2827 int error = __split_vma(mm, last, end, 1);
2831 vma = prev ? prev->vm_next : mm->mmap;
2835 * If userfaultfd_unmap_prep returns an error the vmas
2836 * will remain splitted, but userland will get a
2837 * highly unexpected error anyway. This is no
2838 * different than the case where the first of the two
2839 * __split_vma fails, but we don't undo the first
2840 * split, despite we could. This is unlikely enough
2841 * failure that it's not worth optimizing it for.
2843 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2849 * unlock any mlock()ed ranges before detaching vmas
2851 if (mm->locked_vm) {
2852 struct vm_area_struct *tmp = vma;
2853 while (tmp && tmp->vm_start < end) {
2854 if (tmp->vm_flags & VM_LOCKED) {
2855 mm->locked_vm -= vma_pages(tmp);
2856 munlock_vma_pages_all(tmp);
2863 /* Detach vmas from rbtree */
2864 if (!detach_vmas_to_be_unmapped(mm, vma, prev, end))
2868 mmap_write_downgrade(mm);
2870 unmap_region(mm, vma, prev, start, end);
2872 /* Fix up all other VM information */
2873 remove_vma_list(mm, vma);
2875 return downgrade ? 1 : 0;
2878 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2879 struct list_head *uf)
2881 return __do_munmap(mm, start, len, uf, false);
2884 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2887 struct mm_struct *mm = current->mm;
2890 if (mmap_write_lock_killable(mm))
2893 ret = __do_munmap(mm, start, len, &uf, downgrade);
2895 * Returning 1 indicates mmap_lock is downgraded.
2896 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2897 * it to 0 before return.
2900 mmap_read_unlock(mm);
2903 mmap_write_unlock(mm);
2905 userfaultfd_unmap_complete(mm, &uf);
2909 int vm_munmap(unsigned long start, size_t len)
2911 return __vm_munmap(start, len, false);
2913 EXPORT_SYMBOL(vm_munmap);
2915 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2917 addr = untagged_addr(addr);
2918 profile_munmap(addr);
2919 return __vm_munmap(addr, len, true);
2924 * Emulation of deprecated remap_file_pages() syscall.
2926 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2927 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2930 struct mm_struct *mm = current->mm;
2931 struct vm_area_struct *vma;
2932 unsigned long populate = 0;
2933 unsigned long ret = -EINVAL;
2936 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2937 current->comm, current->pid);
2941 start = start & PAGE_MASK;
2942 size = size & PAGE_MASK;
2944 if (start + size <= start)
2947 /* Does pgoff wrap? */
2948 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2951 if (mmap_write_lock_killable(mm))
2954 vma = find_vma(mm, start);
2956 if (!vma || !(vma->vm_flags & VM_SHARED))
2959 if (start < vma->vm_start)
2962 if (start + size > vma->vm_end) {
2963 struct vm_area_struct *next;
2965 for (next = vma->vm_next; next; next = next->vm_next) {
2966 /* hole between vmas ? */
2967 if (next->vm_start != next->vm_prev->vm_end)
2970 if (next->vm_file != vma->vm_file)
2973 if (next->vm_flags != vma->vm_flags)
2976 if (start + size <= next->vm_end)
2984 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2985 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2986 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2988 flags &= MAP_NONBLOCK;
2989 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2990 if (vma->vm_flags & VM_LOCKED) {
2991 struct vm_area_struct *tmp;
2992 flags |= MAP_LOCKED;
2994 /* drop PG_Mlocked flag for over-mapped range */
2995 for (tmp = vma; tmp->vm_start >= start + size;
2996 tmp = tmp->vm_next) {
2998 * Split pmd and munlock page on the border
3001 vma_adjust_trans_huge(tmp, start, start + size, 0);
3003 munlock_vma_pages_range(tmp,
3004 max(tmp->vm_start, start),
3005 min(tmp->vm_end, start + size));
3009 file = get_file(vma->vm_file);
3010 ret = do_mmap(vma->vm_file, start, size,
3011 prot, flags, pgoff, &populate, NULL);
3014 mmap_write_unlock(mm);
3016 mm_populate(ret, populate);
3017 if (!IS_ERR_VALUE(ret))
3023 * this is really a simplified "do_mmap". it only handles
3024 * anonymous maps. eventually we may be able to do some
3025 * brk-specific accounting here.
3027 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
3029 struct mm_struct *mm = current->mm;
3030 struct vm_area_struct *vma, *prev;
3031 struct rb_node **rb_link, *rb_parent;
3032 pgoff_t pgoff = addr >> PAGE_SHIFT;
3034 unsigned long mapped_addr;
3036 /* Until we need other flags, refuse anything except VM_EXEC. */
3037 if ((flags & (~VM_EXEC)) != 0)
3039 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3041 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
3042 if (IS_ERR_VALUE(mapped_addr))
3045 error = mlock_future_check(mm, mm->def_flags, len);
3050 * Clear old maps. this also does some error checking for us
3052 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
3054 if (do_munmap(mm, addr, len, uf))
3058 /* Check against address space limits *after* clearing old maps... */
3059 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3062 if (mm->map_count > sysctl_max_map_count)
3065 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3068 /* Can we just expand an old private anonymous mapping? */
3069 vma = vma_merge(mm, prev, addr, addr + len, flags,
3070 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
3075 * create a vma struct for an anonymous mapping
3077 vma = vm_area_alloc(mm);
3079 vm_unacct_memory(len >> PAGE_SHIFT);
3083 vma_set_anonymous(vma);
3084 vma->vm_start = addr;
3085 vma->vm_end = addr + len;
3086 vma->vm_pgoff = pgoff;
3087 vma->vm_flags = flags;
3088 vma->vm_page_prot = vm_get_page_prot(flags);
3089 vma_link(mm, vma, prev, rb_link, rb_parent);
3091 perf_event_mmap(vma);
3092 mm->total_vm += len >> PAGE_SHIFT;
3093 mm->data_vm += len >> PAGE_SHIFT;
3094 if (flags & VM_LOCKED)
3095 mm->locked_vm += (len >> PAGE_SHIFT);
3096 vma->vm_flags |= VM_SOFTDIRTY;
3100 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3102 struct mm_struct *mm = current->mm;
3108 len = PAGE_ALIGN(request);
3114 if (mmap_write_lock_killable(mm))
3117 ret = do_brk_flags(addr, len, flags, &uf);
3118 populate = ((mm->def_flags & VM_LOCKED) != 0);
3119 mmap_write_unlock(mm);
3120 userfaultfd_unmap_complete(mm, &uf);
3121 if (populate && !ret)
3122 mm_populate(addr, len);
3125 EXPORT_SYMBOL(vm_brk_flags);
3127 int vm_brk(unsigned long addr, unsigned long len)
3129 return vm_brk_flags(addr, len, 0);
3131 EXPORT_SYMBOL(vm_brk);
3133 /* Release all mmaps. */
3134 void exit_mmap(struct mm_struct *mm)
3136 struct mmu_gather tlb;
3137 struct vm_area_struct *vma;
3138 unsigned long nr_accounted = 0;
3140 /* mm's last user has gone, and its about to be pulled down */
3141 mmu_notifier_release(mm);
3143 if (unlikely(mm_is_oom_victim(mm))) {
3145 * Manually reap the mm to free as much memory as possible.
3146 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3147 * this mm from further consideration. Taking mm->mmap_lock for
3148 * write after setting MMF_OOM_SKIP will guarantee that the oom
3149 * reaper will not run on this mm again after mmap_lock is
3152 * Nothing can be holding mm->mmap_lock here and the above call
3153 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3154 * __oom_reap_task_mm() will not block.
3156 * This needs to be done before calling munlock_vma_pages_all(),
3157 * which clears VM_LOCKED, otherwise the oom reaper cannot
3160 (void)__oom_reap_task_mm(mm);
3162 set_bit(MMF_OOM_SKIP, &mm->flags);
3163 mmap_write_lock(mm);
3164 mmap_write_unlock(mm);
3167 if (mm->locked_vm) {
3170 if (vma->vm_flags & VM_LOCKED)
3171 munlock_vma_pages_all(vma);
3179 if (!vma) /* Can happen if dup_mmap() received an OOM */
3184 tlb_gather_mmu(&tlb, mm, 0, -1);
3185 /* update_hiwater_rss(mm) here? but nobody should be looking */
3186 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3187 unmap_vmas(&tlb, vma, 0, -1);
3188 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3189 tlb_finish_mmu(&tlb, 0, -1);
3192 * Walk the list again, actually closing and freeing it,
3193 * with preemption enabled, without holding any MM locks.
3196 if (vma->vm_flags & VM_ACCOUNT)
3197 nr_accounted += vma_pages(vma);
3198 vma = remove_vma(vma);
3201 vm_unacct_memory(nr_accounted);
3204 /* Insert vm structure into process list sorted by address
3205 * and into the inode's i_mmap tree. If vm_file is non-NULL
3206 * then i_mmap_rwsem is taken here.
3208 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3210 struct vm_area_struct *prev;
3211 struct rb_node **rb_link, *rb_parent;
3213 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3214 &prev, &rb_link, &rb_parent))
3216 if ((vma->vm_flags & VM_ACCOUNT) &&
3217 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3221 * The vm_pgoff of a purely anonymous vma should be irrelevant
3222 * until its first write fault, when page's anon_vma and index
3223 * are set. But now set the vm_pgoff it will almost certainly
3224 * end up with (unless mremap moves it elsewhere before that
3225 * first wfault), so /proc/pid/maps tells a consistent story.
3227 * By setting it to reflect the virtual start address of the
3228 * vma, merges and splits can happen in a seamless way, just
3229 * using the existing file pgoff checks and manipulations.
3230 * Similarly in do_mmap and in do_brk.
3232 if (vma_is_anonymous(vma)) {
3233 BUG_ON(vma->anon_vma);
3234 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3237 vma_link(mm, vma, prev, rb_link, rb_parent);
3242 * Copy the vma structure to a new location in the same mm,
3243 * prior to moving page table entries, to effect an mremap move.
3245 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3246 unsigned long addr, unsigned long len, pgoff_t pgoff,
3247 bool *need_rmap_locks)
3249 struct vm_area_struct *vma = *vmap;
3250 unsigned long vma_start = vma->vm_start;
3251 struct mm_struct *mm = vma->vm_mm;
3252 struct vm_area_struct *new_vma, *prev;
3253 struct rb_node **rb_link, *rb_parent;
3254 bool faulted_in_anon_vma = true;
3257 * If anonymous vma has not yet been faulted, update new pgoff
3258 * to match new location, to increase its chance of merging.
3260 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3261 pgoff = addr >> PAGE_SHIFT;
3262 faulted_in_anon_vma = false;
3265 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3266 return NULL; /* should never get here */
3267 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3268 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3269 vma->vm_userfaultfd_ctx);
3272 * Source vma may have been merged into new_vma
3274 if (unlikely(vma_start >= new_vma->vm_start &&
3275 vma_start < new_vma->vm_end)) {
3277 * The only way we can get a vma_merge with
3278 * self during an mremap is if the vma hasn't
3279 * been faulted in yet and we were allowed to
3280 * reset the dst vma->vm_pgoff to the
3281 * destination address of the mremap to allow
3282 * the merge to happen. mremap must change the
3283 * vm_pgoff linearity between src and dst vmas
3284 * (in turn preventing a vma_merge) to be
3285 * safe. It is only safe to keep the vm_pgoff
3286 * linear if there are no pages mapped yet.
3288 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3289 *vmap = vma = new_vma;
3291 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3293 new_vma = vm_area_dup(vma);
3296 new_vma->vm_start = addr;
3297 new_vma->vm_end = addr + len;
3298 new_vma->vm_pgoff = pgoff;
3299 if (vma_dup_policy(vma, new_vma))
3301 if (anon_vma_clone(new_vma, vma))
3302 goto out_free_mempol;
3303 if (new_vma->vm_file)
3304 get_file(new_vma->vm_file);
3305 if (new_vma->vm_ops && new_vma->vm_ops->open)
3306 new_vma->vm_ops->open(new_vma);
3307 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3308 *need_rmap_locks = false;
3313 mpol_put(vma_policy(new_vma));
3315 vm_area_free(new_vma);
3321 * Return true if the calling process may expand its vm space by the passed
3324 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3326 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3329 if (is_data_mapping(flags) &&
3330 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3331 /* Workaround for Valgrind */
3332 if (rlimit(RLIMIT_DATA) == 0 &&
3333 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3336 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3337 current->comm, current->pid,
3338 (mm->data_vm + npages) << PAGE_SHIFT,
3339 rlimit(RLIMIT_DATA),
3340 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3342 if (!ignore_rlimit_data)
3349 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3351 mm->total_vm += npages;
3353 if (is_exec_mapping(flags))
3354 mm->exec_vm += npages;
3355 else if (is_stack_mapping(flags))
3356 mm->stack_vm += npages;
3357 else if (is_data_mapping(flags))
3358 mm->data_vm += npages;
3361 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3364 * Having a close hook prevents vma merging regardless of flags.
3366 static void special_mapping_close(struct vm_area_struct *vma)
3370 static const char *special_mapping_name(struct vm_area_struct *vma)
3372 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3375 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3377 struct vm_special_mapping *sm = new_vma->vm_private_data;
3379 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3383 return sm->mremap(sm, new_vma);
3388 static const struct vm_operations_struct special_mapping_vmops = {
3389 .close = special_mapping_close,
3390 .fault = special_mapping_fault,
3391 .mremap = special_mapping_mremap,
3392 .name = special_mapping_name,
3393 /* vDSO code relies that VVAR can't be accessed remotely */
3397 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3398 .close = special_mapping_close,
3399 .fault = special_mapping_fault,
3402 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3404 struct vm_area_struct *vma = vmf->vma;
3406 struct page **pages;
3408 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3409 pages = vma->vm_private_data;
3411 struct vm_special_mapping *sm = vma->vm_private_data;
3414 return sm->fault(sm, vmf->vma, vmf);
3419 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3423 struct page *page = *pages;
3429 return VM_FAULT_SIGBUS;
3432 static struct vm_area_struct *__install_special_mapping(
3433 struct mm_struct *mm,
3434 unsigned long addr, unsigned long len,
3435 unsigned long vm_flags, void *priv,
3436 const struct vm_operations_struct *ops)
3439 struct vm_area_struct *vma;
3441 vma = vm_area_alloc(mm);
3442 if (unlikely(vma == NULL))
3443 return ERR_PTR(-ENOMEM);
3445 vma->vm_start = addr;
3446 vma->vm_end = addr + len;
3448 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3449 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3452 vma->vm_private_data = priv;
3454 ret = insert_vm_struct(mm, vma);
3458 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3460 perf_event_mmap(vma);
3466 return ERR_PTR(ret);
3469 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3470 const struct vm_special_mapping *sm)
3472 return vma->vm_private_data == sm &&
3473 (vma->vm_ops == &special_mapping_vmops ||
3474 vma->vm_ops == &legacy_special_mapping_vmops);
3478 * Called with mm->mmap_lock held for writing.
3479 * Insert a new vma covering the given region, with the given flags.
3480 * Its pages are supplied by the given array of struct page *.
3481 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3482 * The region past the last page supplied will always produce SIGBUS.
3483 * The array pointer and the pages it points to are assumed to stay alive
3484 * for as long as this mapping might exist.
3486 struct vm_area_struct *_install_special_mapping(
3487 struct mm_struct *mm,
3488 unsigned long addr, unsigned long len,
3489 unsigned long vm_flags, const struct vm_special_mapping *spec)
3491 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3492 &special_mapping_vmops);
3495 int install_special_mapping(struct mm_struct *mm,
3496 unsigned long addr, unsigned long len,
3497 unsigned long vm_flags, struct page **pages)
3499 struct vm_area_struct *vma = __install_special_mapping(
3500 mm, addr, len, vm_flags, (void *)pages,
3501 &legacy_special_mapping_vmops);
3503 return PTR_ERR_OR_ZERO(vma);
3506 static DEFINE_MUTEX(mm_all_locks_mutex);
3508 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3510 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3512 * The LSB of head.next can't change from under us
3513 * because we hold the mm_all_locks_mutex.
3515 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3517 * We can safely modify head.next after taking the
3518 * anon_vma->root->rwsem. If some other vma in this mm shares
3519 * the same anon_vma we won't take it again.
3521 * No need of atomic instructions here, head.next
3522 * can't change from under us thanks to the
3523 * anon_vma->root->rwsem.
3525 if (__test_and_set_bit(0, (unsigned long *)
3526 &anon_vma->root->rb_root.rb_root.rb_node))
3531 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3533 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3535 * AS_MM_ALL_LOCKS can't change from under us because
3536 * we hold the mm_all_locks_mutex.
3538 * Operations on ->flags have to be atomic because
3539 * even if AS_MM_ALL_LOCKS is stable thanks to the
3540 * mm_all_locks_mutex, there may be other cpus
3541 * changing other bitflags in parallel to us.
3543 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3545 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3550 * This operation locks against the VM for all pte/vma/mm related
3551 * operations that could ever happen on a certain mm. This includes
3552 * vmtruncate, try_to_unmap, and all page faults.
3554 * The caller must take the mmap_lock in write mode before calling
3555 * mm_take_all_locks(). The caller isn't allowed to release the
3556 * mmap_lock until mm_drop_all_locks() returns.
3558 * mmap_lock in write mode is required in order to block all operations
3559 * that could modify pagetables and free pages without need of
3560 * altering the vma layout. It's also needed in write mode to avoid new
3561 * anon_vmas to be associated with existing vmas.
3563 * A single task can't take more than one mm_take_all_locks() in a row
3564 * or it would deadlock.
3566 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3567 * mapping->flags avoid to take the same lock twice, if more than one
3568 * vma in this mm is backed by the same anon_vma or address_space.
3570 * We take locks in following order, accordingly to comment at beginning
3572 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3574 * - all i_mmap_rwsem locks;
3575 * - all anon_vma->rwseml
3577 * We can take all locks within these types randomly because the VM code
3578 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3579 * mm_all_locks_mutex.
3581 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3582 * that may have to take thousand of locks.
3584 * mm_take_all_locks() can fail if it's interrupted by signals.
3586 int mm_take_all_locks(struct mm_struct *mm)
3588 struct vm_area_struct *vma;
3589 struct anon_vma_chain *avc;
3591 BUG_ON(mmap_read_trylock(mm));
3593 mutex_lock(&mm_all_locks_mutex);
3595 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3596 if (signal_pending(current))
3598 if (vma->vm_file && vma->vm_file->f_mapping &&
3599 is_vm_hugetlb_page(vma))
3600 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3603 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3604 if (signal_pending(current))
3606 if (vma->vm_file && vma->vm_file->f_mapping &&
3607 !is_vm_hugetlb_page(vma))
3608 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3611 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3612 if (signal_pending(current))
3615 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3616 vm_lock_anon_vma(mm, avc->anon_vma);
3622 mm_drop_all_locks(mm);
3626 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3628 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3630 * The LSB of head.next can't change to 0 from under
3631 * us because we hold the mm_all_locks_mutex.
3633 * We must however clear the bitflag before unlocking
3634 * the vma so the users using the anon_vma->rb_root will
3635 * never see our bitflag.
3637 * No need of atomic instructions here, head.next
3638 * can't change from under us until we release the
3639 * anon_vma->root->rwsem.
3641 if (!__test_and_clear_bit(0, (unsigned long *)
3642 &anon_vma->root->rb_root.rb_root.rb_node))
3644 anon_vma_unlock_write(anon_vma);
3648 static void vm_unlock_mapping(struct address_space *mapping)
3650 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3652 * AS_MM_ALL_LOCKS can't change to 0 from under us
3653 * because we hold the mm_all_locks_mutex.
3655 i_mmap_unlock_write(mapping);
3656 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3663 * The mmap_lock cannot be released by the caller until
3664 * mm_drop_all_locks() returns.
3666 void mm_drop_all_locks(struct mm_struct *mm)
3668 struct vm_area_struct *vma;
3669 struct anon_vma_chain *avc;
3671 BUG_ON(mmap_read_trylock(mm));
3672 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3674 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3676 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3677 vm_unlock_anon_vma(avc->anon_vma);
3678 if (vma->vm_file && vma->vm_file->f_mapping)
3679 vm_unlock_mapping(vma->vm_file->f_mapping);
3682 mutex_unlock(&mm_all_locks_mutex);
3686 * initialise the percpu counter for VM
3688 void __init mmap_init(void)
3692 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3697 * Initialise sysctl_user_reserve_kbytes.
3699 * This is intended to prevent a user from starting a single memory hogging
3700 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3703 * The default value is min(3% of free memory, 128MB)
3704 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3706 static int init_user_reserve(void)
3708 unsigned long free_kbytes;
3710 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3712 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3715 subsys_initcall(init_user_reserve);
3718 * Initialise sysctl_admin_reserve_kbytes.
3720 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3721 * to log in and kill a memory hogging process.
3723 * Systems with more than 256MB will reserve 8MB, enough to recover
3724 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3725 * only reserve 3% of free pages by default.
3727 static int init_admin_reserve(void)
3729 unsigned long free_kbytes;
3731 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3733 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3736 subsys_initcall(init_admin_reserve);
3739 * Reinititalise user and admin reserves if memory is added or removed.
3741 * The default user reserve max is 128MB, and the default max for the
3742 * admin reserve is 8MB. These are usually, but not always, enough to
3743 * enable recovery from a memory hogging process using login/sshd, a shell,
3744 * and tools like top. It may make sense to increase or even disable the
3745 * reserve depending on the existence of swap or variations in the recovery
3746 * tools. So, the admin may have changed them.
3748 * If memory is added and the reserves have been eliminated or increased above
3749 * the default max, then we'll trust the admin.
3751 * If memory is removed and there isn't enough free memory, then we
3752 * need to reset the reserves.
3754 * Otherwise keep the reserve set by the admin.
3756 static int reserve_mem_notifier(struct notifier_block *nb,
3757 unsigned long action, void *data)
3759 unsigned long tmp, free_kbytes;
3763 /* Default max is 128MB. Leave alone if modified by operator. */
3764 tmp = sysctl_user_reserve_kbytes;
3765 if (0 < tmp && tmp < (1UL << 17))
3766 init_user_reserve();
3768 /* Default max is 8MB. Leave alone if modified by operator. */
3769 tmp = sysctl_admin_reserve_kbytes;
3770 if (0 < tmp && tmp < (1UL << 13))
3771 init_admin_reserve();
3775 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3777 if (sysctl_user_reserve_kbytes > free_kbytes) {
3778 init_user_reserve();
3779 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3780 sysctl_user_reserve_kbytes);
3783 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3784 init_admin_reserve();
3785 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3786 sysctl_admin_reserve_kbytes);
3795 static struct notifier_block reserve_mem_nb = {
3796 .notifier_call = reserve_mem_notifier,
3799 static int __meminit init_reserve_notifier(void)
3801 if (register_hotmemory_notifier(&reserve_mem_nb))
3802 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3806 subsys_initcall(init_reserve_notifier);