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 allow_write_access(file);
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
479 * a. the "next" vma being erased if next->vm_start was reduced in
480 * __vma_adjust() -> __vma_unlink()
481 * b. the vma being erased in detach_vmas_to_be_unmapped() ->
484 validate_mm_rb(root, ignore);
486 __vma_rb_erase(vma, root);
489 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
490 struct rb_root *root)
492 vma_rb_erase_ignore(vma, root, vma);
496 * vma has some anon_vma assigned, and is already inserted on that
497 * anon_vma's interval trees.
499 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
500 * vma must be removed from the anon_vma's interval trees using
501 * anon_vma_interval_tree_pre_update_vma().
503 * After the update, the vma will be reinserted using
504 * anon_vma_interval_tree_post_update_vma().
506 * The entire update must be protected by exclusive mmap_lock and by
507 * the root anon_vma's mutex.
510 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
512 struct anon_vma_chain *avc;
514 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
515 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
519 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
521 struct anon_vma_chain *avc;
523 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
524 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
527 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
528 unsigned long end, struct vm_area_struct **pprev,
529 struct rb_node ***rb_link, struct rb_node **rb_parent)
531 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
533 __rb_link = &mm->mm_rb.rb_node;
534 rb_prev = __rb_parent = NULL;
537 struct vm_area_struct *vma_tmp;
539 __rb_parent = *__rb_link;
540 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
542 if (vma_tmp->vm_end > addr) {
543 /* Fail if an existing vma overlaps the area */
544 if (vma_tmp->vm_start < end)
546 __rb_link = &__rb_parent->rb_left;
548 rb_prev = __rb_parent;
549 __rb_link = &__rb_parent->rb_right;
555 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
556 *rb_link = __rb_link;
557 *rb_parent = __rb_parent;
562 * vma_next() - Get the next VMA.
563 * @mm: The mm_struct.
564 * @vma: The current vma.
566 * If @vma is NULL, return the first vma in the mm.
568 * Returns: The next VMA after @vma.
570 static inline struct vm_area_struct *vma_next(struct mm_struct *mm,
571 struct vm_area_struct *vma)
580 * munmap_vma_range() - munmap VMAs that overlap a range.
582 * @start: The start of the range.
583 * @len: The length of the range.
584 * @pprev: pointer to the pointer that will be set to previous vm_area_struct
585 * @rb_link: the rb_node
586 * @rb_parent: the parent rb_node
588 * Find all the vm_area_struct that overlap from @start to
589 * @end and munmap them. Set @pprev to the previous vm_area_struct.
591 * Returns: -ENOMEM on munmap failure or 0 on success.
594 munmap_vma_range(struct mm_struct *mm, unsigned long start, unsigned long len,
595 struct vm_area_struct **pprev, struct rb_node ***link,
596 struct rb_node **parent, struct list_head *uf)
599 while (find_vma_links(mm, start, start + len, pprev, link, parent))
600 if (do_munmap(mm, start, len, uf))
605 static unsigned long count_vma_pages_range(struct mm_struct *mm,
606 unsigned long addr, unsigned long end)
608 unsigned long nr_pages = 0;
609 struct vm_area_struct *vma;
611 /* Find first overlaping mapping */
612 vma = find_vma_intersection(mm, addr, end);
616 nr_pages = (min(end, vma->vm_end) -
617 max(addr, vma->vm_start)) >> PAGE_SHIFT;
619 /* Iterate over the rest of the overlaps */
620 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
621 unsigned long overlap_len;
623 if (vma->vm_start > end)
626 overlap_len = min(end, vma->vm_end) - vma->vm_start;
627 nr_pages += overlap_len >> PAGE_SHIFT;
633 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
634 struct rb_node **rb_link, struct rb_node *rb_parent)
636 /* Update tracking information for the gap following the new vma. */
638 vma_gap_update(vma->vm_next);
640 mm->highest_vm_end = vm_end_gap(vma);
643 * vma->vm_prev wasn't known when we followed the rbtree to find the
644 * correct insertion point for that vma. As a result, we could not
645 * update the vma vm_rb parents rb_subtree_gap values on the way down.
646 * So, we first insert the vma with a zero rb_subtree_gap value
647 * (to be consistent with what we did on the way down), and then
648 * immediately update the gap to the correct value. Finally we
649 * rebalance the rbtree after all augmented values have been set.
651 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
652 vma->rb_subtree_gap = 0;
654 vma_rb_insert(vma, &mm->mm_rb);
657 static void __vma_link_file(struct vm_area_struct *vma)
663 struct address_space *mapping = file->f_mapping;
665 if (vma->vm_flags & VM_DENYWRITE)
666 put_write_access(file_inode(file));
667 if (vma->vm_flags & VM_SHARED)
668 mapping_allow_writable(mapping);
670 flush_dcache_mmap_lock(mapping);
671 vma_interval_tree_insert(vma, &mapping->i_mmap);
672 flush_dcache_mmap_unlock(mapping);
677 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
678 struct vm_area_struct *prev, struct rb_node **rb_link,
679 struct rb_node *rb_parent)
681 __vma_link_list(mm, vma, prev);
682 __vma_link_rb(mm, vma, rb_link, rb_parent);
685 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
686 struct vm_area_struct *prev, struct rb_node **rb_link,
687 struct rb_node *rb_parent)
689 struct address_space *mapping = NULL;
692 mapping = vma->vm_file->f_mapping;
693 i_mmap_lock_write(mapping);
696 __vma_link(mm, vma, prev, rb_link, rb_parent);
697 __vma_link_file(vma);
700 i_mmap_unlock_write(mapping);
707 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
708 * mm's list and rbtree. It has already been inserted into the interval tree.
710 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
712 struct vm_area_struct *prev;
713 struct rb_node **rb_link, *rb_parent;
715 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
716 &prev, &rb_link, &rb_parent))
718 __vma_link(mm, vma, prev, rb_link, rb_parent);
722 static __always_inline void __vma_unlink(struct mm_struct *mm,
723 struct vm_area_struct *vma,
724 struct vm_area_struct *ignore)
726 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
727 __vma_unlink_list(mm, vma);
729 vmacache_invalidate(mm);
733 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
734 * is already present in an i_mmap tree without adjusting the tree.
735 * The following helper function should be used when such adjustments
736 * are necessary. The "insert" vma (if any) is to be inserted
737 * before we drop the necessary locks.
739 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
740 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
741 struct vm_area_struct *expand)
743 struct mm_struct *mm = vma->vm_mm;
744 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
745 struct address_space *mapping = NULL;
746 struct rb_root_cached *root = NULL;
747 struct anon_vma *anon_vma = NULL;
748 struct file *file = vma->vm_file;
749 bool start_changed = false, end_changed = false;
750 long adjust_next = 0;
753 if (next && !insert) {
754 struct vm_area_struct *exporter = NULL, *importer = NULL;
756 if (end >= next->vm_end) {
758 * vma expands, overlapping all the next, and
759 * perhaps the one after too (mprotect case 6).
760 * The only other cases that gets here are
761 * case 1, case 7 and case 8.
763 if (next == expand) {
765 * The only case where we don't expand "vma"
766 * and we expand "next" instead is case 8.
768 VM_WARN_ON(end != next->vm_end);
770 * remove_next == 3 means we're
771 * removing "vma" and that to do so we
772 * swapped "vma" and "next".
775 VM_WARN_ON(file != next->vm_file);
778 VM_WARN_ON(expand != vma);
780 * case 1, 6, 7, remove_next == 2 is case 6,
781 * remove_next == 1 is case 1 or 7.
783 remove_next = 1 + (end > next->vm_end);
784 VM_WARN_ON(remove_next == 2 &&
785 end != next->vm_next->vm_end);
786 /* trim end to next, for case 6 first pass */
794 * If next doesn't have anon_vma, import from vma after
795 * next, if the vma overlaps with it.
797 if (remove_next == 2 && !next->anon_vma)
798 exporter = next->vm_next;
800 } else if (end > next->vm_start) {
802 * vma expands, overlapping part of the next:
803 * mprotect case 5 shifting the boundary up.
805 adjust_next = (end - next->vm_start);
808 VM_WARN_ON(expand != importer);
809 } else if (end < vma->vm_end) {
811 * vma shrinks, and !insert tells it's not
812 * split_vma inserting another: so it must be
813 * mprotect case 4 shifting the boundary down.
815 adjust_next = -(vma->vm_end - end);
818 VM_WARN_ON(expand != importer);
822 * Easily overlooked: when mprotect shifts the boundary,
823 * make sure the expanding vma has anon_vma set if the
824 * shrinking vma had, to cover any anon pages imported.
826 if (exporter && exporter->anon_vma && !importer->anon_vma) {
829 importer->anon_vma = exporter->anon_vma;
830 error = anon_vma_clone(importer, exporter);
836 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
839 mapping = file->f_mapping;
840 root = &mapping->i_mmap;
841 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
844 uprobe_munmap(next, next->vm_start, next->vm_end);
846 i_mmap_lock_write(mapping);
849 * Put into interval tree now, so instantiated pages
850 * are visible to arm/parisc __flush_dcache_page
851 * throughout; but we cannot insert into address
852 * space until vma start or end is updated.
854 __vma_link_file(insert);
858 anon_vma = vma->anon_vma;
859 if (!anon_vma && adjust_next)
860 anon_vma = next->anon_vma;
862 VM_WARN_ON(adjust_next && next->anon_vma &&
863 anon_vma != next->anon_vma);
864 anon_vma_lock_write(anon_vma);
865 anon_vma_interval_tree_pre_update_vma(vma);
867 anon_vma_interval_tree_pre_update_vma(next);
871 flush_dcache_mmap_lock(mapping);
872 vma_interval_tree_remove(vma, root);
874 vma_interval_tree_remove(next, root);
877 if (start != vma->vm_start) {
878 vma->vm_start = start;
879 start_changed = true;
881 if (end != vma->vm_end) {
885 vma->vm_pgoff = pgoff;
887 next->vm_start += adjust_next;
888 next->vm_pgoff += adjust_next >> PAGE_SHIFT;
893 vma_interval_tree_insert(next, root);
894 vma_interval_tree_insert(vma, root);
895 flush_dcache_mmap_unlock(mapping);
900 * vma_merge has merged next into vma, and needs
901 * us to remove next before dropping the locks.
903 if (remove_next != 3)
904 __vma_unlink(mm, next, next);
907 * vma is not before next if they've been
910 * pre-swap() next->vm_start was reduced so
911 * tell validate_mm_rb to ignore pre-swap()
912 * "next" (which is stored in post-swap()
915 __vma_unlink(mm, next, vma);
917 __remove_shared_vm_struct(next, file, mapping);
920 * split_vma has split insert from vma, and needs
921 * us to insert it before dropping the locks
922 * (it may either follow vma or precede it).
924 __insert_vm_struct(mm, insert);
930 mm->highest_vm_end = vm_end_gap(vma);
931 else if (!adjust_next)
932 vma_gap_update(next);
937 anon_vma_interval_tree_post_update_vma(vma);
939 anon_vma_interval_tree_post_update_vma(next);
940 anon_vma_unlock_write(anon_vma);
944 i_mmap_unlock_write(mapping);
953 uprobe_munmap(next, next->vm_start, next->vm_end);
957 anon_vma_merge(vma, next);
959 mpol_put(vma_policy(next));
962 * In mprotect's case 6 (see comments on vma_merge),
963 * we must remove another next too. It would clutter
964 * up the code too much to do both in one go.
966 if (remove_next != 3) {
968 * If "next" was removed and vma->vm_end was
969 * expanded (up) over it, in turn
970 * "next->vm_prev->vm_end" changed and the
971 * "vma->vm_next" gap must be updated.
976 * For the scope of the comment "next" and
977 * "vma" considered pre-swap(): if "vma" was
978 * removed, next->vm_start was expanded (down)
979 * over it and the "next" gap must be updated.
980 * Because of the swap() the post-swap() "vma"
981 * actually points to pre-swap() "next"
982 * (post-swap() "next" as opposed is now a
987 if (remove_next == 2) {
993 vma_gap_update(next);
996 * If remove_next == 2 we obviously can't
999 * If remove_next == 3 we can't reach this
1000 * path because pre-swap() next is always not
1001 * NULL. pre-swap() "next" is not being
1002 * removed and its next->vm_end is not altered
1003 * (and furthermore "end" already matches
1004 * next->vm_end in remove_next == 3).
1006 * We reach this only in the remove_next == 1
1007 * case if the "next" vma that was removed was
1008 * the highest vma of the mm. However in such
1009 * case next->vm_end == "end" and the extended
1010 * "vma" has vma->vm_end == next->vm_end so
1011 * mm->highest_vm_end doesn't need any update
1012 * in remove_next == 1 case.
1014 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
1018 uprobe_mmap(insert);
1026 * If the vma has a ->close operation then the driver probably needs to release
1027 * per-vma resources, so we don't attempt to merge those.
1029 static inline int is_mergeable_vma(struct vm_area_struct *vma,
1030 struct file *file, unsigned long vm_flags,
1031 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1034 * VM_SOFTDIRTY should not prevent from VMA merging, if we
1035 * match the flags but dirty bit -- the caller should mark
1036 * merged VMA as dirty. If dirty bit won't be excluded from
1037 * comparison, we increase pressure on the memory system forcing
1038 * the kernel to generate new VMAs when old one could be
1041 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
1043 if (vma->vm_file != file)
1045 if (vma->vm_ops && vma->vm_ops->close)
1047 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1052 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1053 struct anon_vma *anon_vma2,
1054 struct vm_area_struct *vma)
1057 * The list_is_singular() test is to avoid merging VMA cloned from
1058 * parents. This can improve scalability caused by anon_vma lock.
1060 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1061 list_is_singular(&vma->anon_vma_chain)))
1063 return anon_vma1 == anon_vma2;
1067 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1068 * in front of (at a lower virtual address and file offset than) the vma.
1070 * We cannot merge two vmas if they have differently assigned (non-NULL)
1071 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1073 * We don't check here for the merged mmap wrapping around the end of pagecache
1074 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
1075 * wrap, nor mmaps which cover the final page at index -1UL.
1078 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1079 struct anon_vma *anon_vma, struct file *file,
1081 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1083 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1084 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1085 if (vma->vm_pgoff == vm_pgoff)
1092 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1093 * beyond (at a higher virtual address and file offset than) the vma.
1095 * We cannot merge two vmas if they have differently assigned (non-NULL)
1096 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1099 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1100 struct anon_vma *anon_vma, struct file *file,
1102 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1104 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1105 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1107 vm_pglen = vma_pages(vma);
1108 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1115 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1116 * whether that can be merged with its predecessor or its successor.
1117 * Or both (it neatly fills a hole).
1119 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1120 * certain not to be mapped by the time vma_merge is called; but when
1121 * called for mprotect, it is certain to be already mapped (either at
1122 * an offset within prev, or at the start of next), and the flags of
1123 * this area are about to be changed to vm_flags - and the no-change
1124 * case has already been eliminated.
1126 * The following mprotect cases have to be considered, where AAAA is
1127 * the area passed down from mprotect_fixup, never extending beyond one
1128 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1131 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN
1132 * cannot merge might become might become
1133 * PPNNNNNNNNNN PPPPPPPPPPNN
1134 * mmap, brk or case 4 below case 5 below
1137 * PPPP NNNN PPPPNNNNXXXX
1138 * might become might become
1139 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
1140 * PPPPPPPPNNNN 2 or PPPPPPPPXXXX 7 or
1141 * PPPPNNNNNNNN 3 PPPPXXXXXXXX 8
1143 * It is important for case 8 that the vma NNNN overlapping the
1144 * region AAAA is never going to extended over XXXX. Instead XXXX must
1145 * be extended in region AAAA and NNNN must be removed. This way in
1146 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1147 * rmap_locks, the properties of the merged vma will be already
1148 * correct for the whole merged range. Some of those properties like
1149 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1150 * be correct for the whole merged range immediately after the
1151 * rmap_locks are released. Otherwise if XXXX would be removed and
1152 * NNNN would be extended over the XXXX range, remove_migration_ptes
1153 * or other rmap walkers (if working on addresses beyond the "end"
1154 * parameter) may establish ptes with the wrong permissions of NNNN
1155 * instead of the right permissions of XXXX.
1157 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1158 struct vm_area_struct *prev, unsigned long addr,
1159 unsigned long end, unsigned long vm_flags,
1160 struct anon_vma *anon_vma, struct file *file,
1161 pgoff_t pgoff, struct mempolicy *policy,
1162 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1164 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1165 struct vm_area_struct *area, *next;
1169 * We later require that vma->vm_flags == vm_flags,
1170 * so this tests vma->vm_flags & VM_SPECIAL, too.
1172 if (vm_flags & VM_SPECIAL)
1175 next = vma_next(mm, prev);
1177 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1178 next = next->vm_next;
1180 /* verify some invariant that must be enforced by the caller */
1181 VM_WARN_ON(prev && addr <= prev->vm_start);
1182 VM_WARN_ON(area && end > area->vm_end);
1183 VM_WARN_ON(addr >= end);
1186 * Can it merge with the predecessor?
1188 if (prev && prev->vm_end == addr &&
1189 mpol_equal(vma_policy(prev), policy) &&
1190 can_vma_merge_after(prev, vm_flags,
1191 anon_vma, file, pgoff,
1192 vm_userfaultfd_ctx)) {
1194 * OK, it can. Can we now merge in the successor as well?
1196 if (next && end == next->vm_start &&
1197 mpol_equal(policy, vma_policy(next)) &&
1198 can_vma_merge_before(next, vm_flags,
1201 vm_userfaultfd_ctx) &&
1202 is_mergeable_anon_vma(prev->anon_vma,
1203 next->anon_vma, NULL)) {
1205 err = __vma_adjust(prev, prev->vm_start,
1206 next->vm_end, prev->vm_pgoff, NULL,
1208 } else /* cases 2, 5, 7 */
1209 err = __vma_adjust(prev, prev->vm_start,
1210 end, prev->vm_pgoff, NULL, prev);
1213 khugepaged_enter_vma_merge(prev, vm_flags);
1218 * Can this new request be merged in front of next?
1220 if (next && end == next->vm_start &&
1221 mpol_equal(policy, vma_policy(next)) &&
1222 can_vma_merge_before(next, vm_flags,
1223 anon_vma, file, pgoff+pglen,
1224 vm_userfaultfd_ctx)) {
1225 if (prev && addr < prev->vm_end) /* case 4 */
1226 err = __vma_adjust(prev, prev->vm_start,
1227 addr, prev->vm_pgoff, NULL, next);
1228 else { /* cases 3, 8 */
1229 err = __vma_adjust(area, addr, next->vm_end,
1230 next->vm_pgoff - pglen, NULL, next);
1232 * In case 3 area is already equal to next and
1233 * this is a noop, but in case 8 "area" has
1234 * been removed and next was expanded over it.
1240 khugepaged_enter_vma_merge(area, vm_flags);
1248 * Rough compatibility check to quickly see if it's even worth looking
1249 * at sharing an anon_vma.
1251 * They need to have the same vm_file, and the flags can only differ
1252 * in things that mprotect may change.
1254 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1255 * we can merge the two vma's. For example, we refuse to merge a vma if
1256 * there is a vm_ops->close() function, because that indicates that the
1257 * driver is doing some kind of reference counting. But that doesn't
1258 * really matter for the anon_vma sharing case.
1260 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1262 return a->vm_end == b->vm_start &&
1263 mpol_equal(vma_policy(a), vma_policy(b)) &&
1264 a->vm_file == b->vm_file &&
1265 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1266 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1270 * Do some basic sanity checking to see if we can re-use the anon_vma
1271 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1272 * the same as 'old', the other will be the new one that is trying
1273 * to share the anon_vma.
1275 * NOTE! This runs with mm_sem held for reading, so it is possible that
1276 * the anon_vma of 'old' is concurrently in the process of being set up
1277 * by another page fault trying to merge _that_. But that's ok: if it
1278 * is being set up, that automatically means that it will be a singleton
1279 * acceptable for merging, so we can do all of this optimistically. But
1280 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1282 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1283 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1284 * is to return an anon_vma that is "complex" due to having gone through
1287 * We also make sure that the two vma's are compatible (adjacent,
1288 * and with the same memory policies). That's all stable, even with just
1289 * a read lock on the mm_sem.
1291 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1293 if (anon_vma_compatible(a, b)) {
1294 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1296 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1303 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1304 * neighbouring vmas for a suitable anon_vma, before it goes off
1305 * to allocate a new anon_vma. It checks because a repetitive
1306 * sequence of mprotects and faults may otherwise lead to distinct
1307 * anon_vmas being allocated, preventing vma merge in subsequent
1310 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1312 struct anon_vma *anon_vma = NULL;
1314 /* Try next first. */
1316 anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next);
1321 /* Try prev next. */
1323 anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma);
1326 * We might reach here with anon_vma == NULL if we can't find
1327 * any reusable anon_vma.
1328 * There's no absolute need to look only at touching neighbours:
1329 * we could search further afield for "compatible" anon_vmas.
1330 * But it would probably just be a waste of time searching,
1331 * or lead to too many vmas hanging off the same anon_vma.
1332 * We're trying to allow mprotect remerging later on,
1333 * not trying to minimize memory used for anon_vmas.
1339 * If a hint addr is less than mmap_min_addr change hint to be as
1340 * low as possible but still greater than mmap_min_addr
1342 static inline unsigned long round_hint_to_min(unsigned long hint)
1345 if (((void *)hint != NULL) &&
1346 (hint < mmap_min_addr))
1347 return PAGE_ALIGN(mmap_min_addr);
1351 static inline int mlock_future_check(struct mm_struct *mm,
1352 unsigned long flags,
1355 unsigned long locked, lock_limit;
1357 /* mlock MCL_FUTURE? */
1358 if (flags & VM_LOCKED) {
1359 locked = len >> PAGE_SHIFT;
1360 locked += mm->locked_vm;
1361 lock_limit = rlimit(RLIMIT_MEMLOCK);
1362 lock_limit >>= PAGE_SHIFT;
1363 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1369 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1371 if (S_ISREG(inode->i_mode))
1372 return MAX_LFS_FILESIZE;
1374 if (S_ISBLK(inode->i_mode))
1375 return MAX_LFS_FILESIZE;
1377 if (S_ISSOCK(inode->i_mode))
1378 return MAX_LFS_FILESIZE;
1380 /* Special "we do even unsigned file positions" case */
1381 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1384 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1388 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1389 unsigned long pgoff, unsigned long len)
1391 u64 maxsize = file_mmap_size_max(file, inode);
1393 if (maxsize && len > maxsize)
1396 if (pgoff > maxsize >> PAGE_SHIFT)
1402 * The caller must write-lock current->mm->mmap_lock.
1404 unsigned long do_mmap(struct file *file, unsigned long addr,
1405 unsigned long len, unsigned long prot,
1406 unsigned long flags, unsigned long pgoff,
1407 unsigned long *populate, struct list_head *uf)
1409 struct mm_struct *mm = current->mm;
1410 vm_flags_t vm_flags;
1419 * Does the application expect PROT_READ to imply PROT_EXEC?
1421 * (the exception is when the underlying filesystem is noexec
1422 * mounted, in which case we dont add PROT_EXEC.)
1424 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1425 if (!(file && path_noexec(&file->f_path)))
1428 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1429 if (flags & MAP_FIXED_NOREPLACE)
1432 if (!(flags & MAP_FIXED))
1433 addr = round_hint_to_min(addr);
1435 /* Careful about overflows.. */
1436 len = PAGE_ALIGN(len);
1440 /* offset overflow? */
1441 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1444 /* Too many mappings? */
1445 if (mm->map_count > sysctl_max_map_count)
1448 /* Obtain the address to map to. we verify (or select) it and ensure
1449 * that it represents a valid section of the address space.
1451 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1452 if (IS_ERR_VALUE(addr))
1455 if (flags & MAP_FIXED_NOREPLACE) {
1456 struct vm_area_struct *vma = find_vma(mm, addr);
1458 if (vma && vma->vm_start < addr + len)
1462 if (prot == PROT_EXEC) {
1463 pkey = execute_only_pkey(mm);
1468 /* Do simple checking here so the lower-level routines won't have
1469 * to. we assume access permissions have been handled by the open
1470 * of the memory object, so we don't do any here.
1472 vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1473 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1475 if (flags & MAP_LOCKED)
1476 if (!can_do_mlock())
1479 if (mlock_future_check(mm, vm_flags, len))
1483 struct inode *inode = file_inode(file);
1484 unsigned long flags_mask;
1486 if (!file_mmap_ok(file, inode, pgoff, len))
1489 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1491 switch (flags & MAP_TYPE) {
1494 * Force use of MAP_SHARED_VALIDATE with non-legacy
1495 * flags. E.g. MAP_SYNC is dangerous to use with
1496 * MAP_SHARED as you don't know which consistency model
1497 * you will get. We silently ignore unsupported flags
1498 * with MAP_SHARED to preserve backward compatibility.
1500 flags &= LEGACY_MAP_MASK;
1502 case MAP_SHARED_VALIDATE:
1503 if (flags & ~flags_mask)
1505 if (prot & PROT_WRITE) {
1506 if (!(file->f_mode & FMODE_WRITE))
1508 if (IS_SWAPFILE(file->f_mapping->host))
1513 * Make sure we don't allow writing to an append-only
1516 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1520 * Make sure there are no mandatory locks on the file.
1522 if (locks_verify_locked(file))
1525 vm_flags |= VM_SHARED | VM_MAYSHARE;
1526 if (!(file->f_mode & FMODE_WRITE))
1527 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1530 if (!(file->f_mode & FMODE_READ))
1532 if (path_noexec(&file->f_path)) {
1533 if (vm_flags & VM_EXEC)
1535 vm_flags &= ~VM_MAYEXEC;
1538 if (!file->f_op->mmap)
1540 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1548 switch (flags & MAP_TYPE) {
1550 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1556 vm_flags |= VM_SHARED | VM_MAYSHARE;
1560 * Set pgoff according to addr for anon_vma.
1562 pgoff = addr >> PAGE_SHIFT;
1570 * Set 'VM_NORESERVE' if we should not account for the
1571 * memory use of this mapping.
1573 if (flags & MAP_NORESERVE) {
1574 /* We honor MAP_NORESERVE if allowed to overcommit */
1575 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1576 vm_flags |= VM_NORESERVE;
1578 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1579 if (file && is_file_hugepages(file))
1580 vm_flags |= VM_NORESERVE;
1583 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1584 if (!IS_ERR_VALUE(addr) &&
1585 ((vm_flags & VM_LOCKED) ||
1586 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1591 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1592 unsigned long prot, unsigned long flags,
1593 unsigned long fd, unsigned long pgoff)
1595 struct file *file = NULL;
1596 unsigned long retval;
1598 if (!(flags & MAP_ANONYMOUS)) {
1599 audit_mmap_fd(fd, flags);
1603 if (is_file_hugepages(file)) {
1604 len = ALIGN(len, huge_page_size(hstate_file(file)));
1605 } else if (unlikely(flags & MAP_HUGETLB)) {
1609 } else if (flags & MAP_HUGETLB) {
1610 struct user_struct *user = NULL;
1613 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1617 len = ALIGN(len, huge_page_size(hs));
1619 * VM_NORESERVE is used because the reservations will be
1620 * taken when vm_ops->mmap() is called
1621 * A dummy user value is used because we are not locking
1622 * memory so no accounting is necessary
1624 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1626 &user, HUGETLB_ANONHUGE_INODE,
1627 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1629 return PTR_ERR(file);
1632 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1634 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1641 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1642 unsigned long, prot, unsigned long, flags,
1643 unsigned long, fd, unsigned long, pgoff)
1645 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1648 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1649 struct mmap_arg_struct {
1653 unsigned long flags;
1655 unsigned long offset;
1658 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1660 struct mmap_arg_struct a;
1662 if (copy_from_user(&a, arg, sizeof(a)))
1664 if (offset_in_page(a.offset))
1667 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1668 a.offset >> PAGE_SHIFT);
1670 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1673 * Some shared mappings will want the pages marked read-only
1674 * to track write events. If so, we'll downgrade vm_page_prot
1675 * to the private version (using protection_map[] without the
1678 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1680 vm_flags_t vm_flags = vma->vm_flags;
1681 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1683 /* If it was private or non-writable, the write bit is already clear */
1684 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1687 /* The backer wishes to know when pages are first written to? */
1688 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1691 /* The open routine did something to the protections that pgprot_modify
1692 * won't preserve? */
1693 if (pgprot_val(vm_page_prot) !=
1694 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1697 /* Do we need to track softdirty? */
1698 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1701 /* Specialty mapping? */
1702 if (vm_flags & VM_PFNMAP)
1705 /* Can the mapping track the dirty pages? */
1706 return vma->vm_file && vma->vm_file->f_mapping &&
1707 mapping_can_writeback(vma->vm_file->f_mapping);
1711 * We account for memory if it's a private writeable mapping,
1712 * not hugepages and VM_NORESERVE wasn't set.
1714 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1717 * hugetlb has its own accounting separate from the core VM
1718 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1720 if (file && is_file_hugepages(file))
1723 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1726 unsigned long mmap_region(struct file *file, unsigned long addr,
1727 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1728 struct list_head *uf)
1730 struct mm_struct *mm = current->mm;
1731 struct vm_area_struct *vma, *prev, *merge;
1733 struct rb_node **rb_link, *rb_parent;
1734 unsigned long charged = 0;
1736 /* Check against address space limit. */
1737 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1738 unsigned long nr_pages;
1741 * MAP_FIXED may remove pages of mappings that intersects with
1742 * requested mapping. Account for the pages it would unmap.
1744 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1746 if (!may_expand_vm(mm, vm_flags,
1747 (len >> PAGE_SHIFT) - nr_pages))
1751 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
1752 if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
1755 * Private writable mapping: check memory availability
1757 if (accountable_mapping(file, vm_flags)) {
1758 charged = len >> PAGE_SHIFT;
1759 if (security_vm_enough_memory_mm(mm, charged))
1761 vm_flags |= VM_ACCOUNT;
1765 * Can we just expand an old mapping?
1767 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1768 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1773 * Determine the object being mapped and call the appropriate
1774 * specific mapper. the address has already been validated, but
1775 * not unmapped, but the maps are removed from the list.
1777 vma = vm_area_alloc(mm);
1783 vma->vm_start = addr;
1784 vma->vm_end = addr + len;
1785 vma->vm_flags = vm_flags;
1786 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1787 vma->vm_pgoff = pgoff;
1790 if (vm_flags & VM_DENYWRITE) {
1791 error = deny_write_access(file);
1795 if (vm_flags & VM_SHARED) {
1796 error = mapping_map_writable(file->f_mapping);
1798 goto allow_write_and_free_vma;
1801 /* ->mmap() can change vma->vm_file, but must guarantee that
1802 * vma_link() below can deny write-access if VM_DENYWRITE is set
1803 * and map writably if VM_SHARED is set. This usually means the
1804 * new file must not have been exposed to user-space, yet.
1806 vma->vm_file = get_file(file);
1807 error = call_mmap(file, vma);
1809 goto unmap_and_free_vma;
1811 /* Can addr have changed??
1813 * Answer: Yes, several device drivers can do it in their
1814 * f_op->mmap method. -DaveM
1815 * Bug: If addr is changed, prev, rb_link, rb_parent should
1816 * be updated for vma_link()
1818 WARN_ON_ONCE(addr != vma->vm_start);
1820 addr = vma->vm_start;
1822 /* If vm_flags changed after call_mmap(), we should try merge vma again
1823 * as we may succeed this time.
1825 if (unlikely(vm_flags != vma->vm_flags && prev)) {
1826 merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
1827 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX);
1829 /* ->mmap() can change vma->vm_file and fput the original file. So
1830 * fput the vma->vm_file here or we would add an extra fput for file
1831 * and cause general protection fault ultimately.
1836 /* Update vm_flags to pick up the change. */
1837 vm_flags = vma->vm_flags;
1838 goto unmap_writable;
1842 vm_flags = vma->vm_flags;
1843 } else if (vm_flags & VM_SHARED) {
1844 error = shmem_zero_setup(vma);
1848 vma_set_anonymous(vma);
1851 /* Allow architectures to sanity-check the vm_flags */
1852 if (!arch_validate_flags(vma->vm_flags)) {
1855 goto unmap_and_free_vma;
1860 vma_link(mm, vma, prev, rb_link, rb_parent);
1861 /* Once vma denies write, undo our temporary denial count */
1864 if (vm_flags & VM_SHARED)
1865 mapping_unmap_writable(file->f_mapping);
1866 if (vm_flags & VM_DENYWRITE)
1867 allow_write_access(file);
1869 file = vma->vm_file;
1871 perf_event_mmap(vma);
1873 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1874 if (vm_flags & VM_LOCKED) {
1875 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1876 is_vm_hugetlb_page(vma) ||
1877 vma == get_gate_vma(current->mm))
1878 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1880 mm->locked_vm += (len >> PAGE_SHIFT);
1887 * New (or expanded) vma always get soft dirty status.
1888 * Otherwise user-space soft-dirty page tracker won't
1889 * be able to distinguish situation when vma area unmapped,
1890 * then new mapped in-place (which must be aimed as
1891 * a completely new data area).
1893 vma->vm_flags |= VM_SOFTDIRTY;
1895 vma_set_page_prot(vma);
1901 vma->vm_file = NULL;
1903 /* Undo any partial mapping done by a device driver. */
1904 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1906 if (vm_flags & VM_SHARED)
1907 mapping_unmap_writable(file->f_mapping);
1908 allow_write_and_free_vma:
1909 if (vm_flags & VM_DENYWRITE)
1910 allow_write_access(file);
1915 vm_unacct_memory(charged);
1919 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1922 * We implement the search by looking for an rbtree node that
1923 * immediately follows a suitable gap. That is,
1924 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1925 * - gap_end = vma->vm_start >= info->low_limit + length;
1926 * - gap_end - gap_start >= length
1929 struct mm_struct *mm = current->mm;
1930 struct vm_area_struct *vma;
1931 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1933 /* Adjust search length to account for worst case alignment overhead */
1934 length = info->length + info->align_mask;
1935 if (length < info->length)
1938 /* Adjust search limits by the desired length */
1939 if (info->high_limit < length)
1941 high_limit = info->high_limit - length;
1943 if (info->low_limit > high_limit)
1945 low_limit = info->low_limit + length;
1947 /* Check if rbtree root looks promising */
1948 if (RB_EMPTY_ROOT(&mm->mm_rb))
1950 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1951 if (vma->rb_subtree_gap < length)
1955 /* Visit left subtree if it looks promising */
1956 gap_end = vm_start_gap(vma);
1957 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1958 struct vm_area_struct *left =
1959 rb_entry(vma->vm_rb.rb_left,
1960 struct vm_area_struct, vm_rb);
1961 if (left->rb_subtree_gap >= length) {
1967 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1969 /* Check if current node has a suitable gap */
1970 if (gap_start > high_limit)
1972 if (gap_end >= low_limit &&
1973 gap_end > gap_start && gap_end - gap_start >= length)
1976 /* Visit right subtree if it looks promising */
1977 if (vma->vm_rb.rb_right) {
1978 struct vm_area_struct *right =
1979 rb_entry(vma->vm_rb.rb_right,
1980 struct vm_area_struct, vm_rb);
1981 if (right->rb_subtree_gap >= length) {
1987 /* Go back up the rbtree to find next candidate node */
1989 struct rb_node *prev = &vma->vm_rb;
1990 if (!rb_parent(prev))
1992 vma = rb_entry(rb_parent(prev),
1993 struct vm_area_struct, vm_rb);
1994 if (prev == vma->vm_rb.rb_left) {
1995 gap_start = vm_end_gap(vma->vm_prev);
1996 gap_end = vm_start_gap(vma);
2003 /* Check highest gap, which does not precede any rbtree node */
2004 gap_start = mm->highest_vm_end;
2005 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
2006 if (gap_start > high_limit)
2010 /* We found a suitable gap. Clip it with the original low_limit. */
2011 if (gap_start < info->low_limit)
2012 gap_start = info->low_limit;
2014 /* Adjust gap address to the desired alignment */
2015 gap_start += (info->align_offset - gap_start) & info->align_mask;
2017 VM_BUG_ON(gap_start + info->length > info->high_limit);
2018 VM_BUG_ON(gap_start + info->length > gap_end);
2022 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
2024 struct mm_struct *mm = current->mm;
2025 struct vm_area_struct *vma;
2026 unsigned long length, low_limit, high_limit, gap_start, gap_end;
2028 /* Adjust search length to account for worst case alignment overhead */
2029 length = info->length + info->align_mask;
2030 if (length < info->length)
2034 * Adjust search limits by the desired length.
2035 * See implementation comment at top of unmapped_area().
2037 gap_end = info->high_limit;
2038 if (gap_end < length)
2040 high_limit = gap_end - length;
2042 if (info->low_limit > high_limit)
2044 low_limit = info->low_limit + length;
2046 /* Check highest gap, which does not precede any rbtree node */
2047 gap_start = mm->highest_vm_end;
2048 if (gap_start <= high_limit)
2051 /* Check if rbtree root looks promising */
2052 if (RB_EMPTY_ROOT(&mm->mm_rb))
2054 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
2055 if (vma->rb_subtree_gap < length)
2059 /* Visit right subtree if it looks promising */
2060 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
2061 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
2062 struct vm_area_struct *right =
2063 rb_entry(vma->vm_rb.rb_right,
2064 struct vm_area_struct, vm_rb);
2065 if (right->rb_subtree_gap >= length) {
2072 /* Check if current node has a suitable gap */
2073 gap_end = vm_start_gap(vma);
2074 if (gap_end < low_limit)
2076 if (gap_start <= high_limit &&
2077 gap_end > gap_start && gap_end - gap_start >= length)
2080 /* Visit left subtree if it looks promising */
2081 if (vma->vm_rb.rb_left) {
2082 struct vm_area_struct *left =
2083 rb_entry(vma->vm_rb.rb_left,
2084 struct vm_area_struct, vm_rb);
2085 if (left->rb_subtree_gap >= length) {
2091 /* Go back up the rbtree to find next candidate node */
2093 struct rb_node *prev = &vma->vm_rb;
2094 if (!rb_parent(prev))
2096 vma = rb_entry(rb_parent(prev),
2097 struct vm_area_struct, vm_rb);
2098 if (prev == vma->vm_rb.rb_right) {
2099 gap_start = vma->vm_prev ?
2100 vm_end_gap(vma->vm_prev) : 0;
2107 /* We found a suitable gap. Clip it with the original high_limit. */
2108 if (gap_end > info->high_limit)
2109 gap_end = info->high_limit;
2112 /* Compute highest gap address at the desired alignment */
2113 gap_end -= info->length;
2114 gap_end -= (gap_end - info->align_offset) & info->align_mask;
2116 VM_BUG_ON(gap_end < info->low_limit);
2117 VM_BUG_ON(gap_end < gap_start);
2122 * Search for an unmapped address range.
2124 * We are looking for a range that:
2125 * - does not intersect with any VMA;
2126 * - is contained within the [low_limit, high_limit) interval;
2127 * - is at least the desired size.
2128 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
2130 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
2134 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
2135 addr = unmapped_area_topdown(info);
2137 addr = unmapped_area(info);
2139 trace_vm_unmapped_area(addr, info);
2143 #ifndef arch_get_mmap_end
2144 #define arch_get_mmap_end(addr) (TASK_SIZE)
2147 #ifndef arch_get_mmap_base
2148 #define arch_get_mmap_base(addr, base) (base)
2151 /* Get an address range which is currently unmapped.
2152 * For shmat() with addr=0.
2154 * Ugly calling convention alert:
2155 * Return value with the low bits set means error value,
2157 * if (ret & ~PAGE_MASK)
2160 * This function "knows" that -ENOMEM has the bits set.
2162 #ifndef HAVE_ARCH_UNMAPPED_AREA
2164 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2165 unsigned long len, unsigned long pgoff, unsigned long flags)
2167 struct mm_struct *mm = current->mm;
2168 struct vm_area_struct *vma, *prev;
2169 struct vm_unmapped_area_info info;
2170 const unsigned long mmap_end = arch_get_mmap_end(addr);
2172 if (len > mmap_end - mmap_min_addr)
2175 if (flags & MAP_FIXED)
2179 addr = PAGE_ALIGN(addr);
2180 vma = find_vma_prev(mm, addr, &prev);
2181 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2182 (!vma || addr + len <= vm_start_gap(vma)) &&
2183 (!prev || addr >= vm_end_gap(prev)))
2189 info.low_limit = mm->mmap_base;
2190 info.high_limit = mmap_end;
2191 info.align_mask = 0;
2192 info.align_offset = 0;
2193 return vm_unmapped_area(&info);
2198 * This mmap-allocator allocates new areas top-down from below the
2199 * stack's low limit (the base):
2201 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2203 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2204 unsigned long len, unsigned long pgoff,
2205 unsigned long flags)
2207 struct vm_area_struct *vma, *prev;
2208 struct mm_struct *mm = current->mm;
2209 struct vm_unmapped_area_info info;
2210 const unsigned long mmap_end = arch_get_mmap_end(addr);
2212 /* requested length too big for entire address space */
2213 if (len > mmap_end - mmap_min_addr)
2216 if (flags & MAP_FIXED)
2219 /* requesting a specific address */
2221 addr = PAGE_ALIGN(addr);
2222 vma = find_vma_prev(mm, addr, &prev);
2223 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2224 (!vma || addr + len <= vm_start_gap(vma)) &&
2225 (!prev || addr >= vm_end_gap(prev)))
2229 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2231 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2232 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2233 info.align_mask = 0;
2234 info.align_offset = 0;
2235 addr = vm_unmapped_area(&info);
2238 * A failed mmap() very likely causes application failure,
2239 * so fall back to the bottom-up function here. This scenario
2240 * can happen with large stack limits and large mmap()
2243 if (offset_in_page(addr)) {
2244 VM_BUG_ON(addr != -ENOMEM);
2246 info.low_limit = TASK_UNMAPPED_BASE;
2247 info.high_limit = mmap_end;
2248 addr = vm_unmapped_area(&info);
2256 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2257 unsigned long pgoff, unsigned long flags)
2259 unsigned long (*get_area)(struct file *, unsigned long,
2260 unsigned long, unsigned long, unsigned long);
2262 unsigned long error = arch_mmap_check(addr, len, flags);
2266 /* Careful about overflows.. */
2267 if (len > TASK_SIZE)
2270 get_area = current->mm->get_unmapped_area;
2272 if (file->f_op->get_unmapped_area)
2273 get_area = file->f_op->get_unmapped_area;
2274 } else if (flags & MAP_SHARED) {
2276 * mmap_region() will call shmem_zero_setup() to create a file,
2277 * so use shmem's get_unmapped_area in case it can be huge.
2278 * do_mmap() will clear pgoff, so match alignment.
2281 get_area = shmem_get_unmapped_area;
2284 addr = get_area(file, addr, len, pgoff, flags);
2285 if (IS_ERR_VALUE(addr))
2288 if (addr > TASK_SIZE - len)
2290 if (offset_in_page(addr))
2293 error = security_mmap_addr(addr);
2294 return error ? error : addr;
2297 EXPORT_SYMBOL(get_unmapped_area);
2299 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2300 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2302 struct rb_node *rb_node;
2303 struct vm_area_struct *vma;
2305 /* Check the cache first. */
2306 vma = vmacache_find(mm, addr);
2310 rb_node = mm->mm_rb.rb_node;
2313 struct vm_area_struct *tmp;
2315 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2317 if (tmp->vm_end > addr) {
2319 if (tmp->vm_start <= addr)
2321 rb_node = rb_node->rb_left;
2323 rb_node = rb_node->rb_right;
2327 vmacache_update(addr, vma);
2331 EXPORT_SYMBOL(find_vma);
2334 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2336 struct vm_area_struct *
2337 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2338 struct vm_area_struct **pprev)
2340 struct vm_area_struct *vma;
2342 vma = find_vma(mm, addr);
2344 *pprev = vma->vm_prev;
2346 struct rb_node *rb_node = rb_last(&mm->mm_rb);
2348 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
2354 * Verify that the stack growth is acceptable and
2355 * update accounting. This is shared with both the
2356 * grow-up and grow-down cases.
2358 static int acct_stack_growth(struct vm_area_struct *vma,
2359 unsigned long size, unsigned long grow)
2361 struct mm_struct *mm = vma->vm_mm;
2362 unsigned long new_start;
2364 /* address space limit tests */
2365 if (!may_expand_vm(mm, vma->vm_flags, grow))
2368 /* Stack limit test */
2369 if (size > rlimit(RLIMIT_STACK))
2372 /* mlock limit tests */
2373 if (vma->vm_flags & VM_LOCKED) {
2374 unsigned long locked;
2375 unsigned long limit;
2376 locked = mm->locked_vm + grow;
2377 limit = rlimit(RLIMIT_MEMLOCK);
2378 limit >>= PAGE_SHIFT;
2379 if (locked > limit && !capable(CAP_IPC_LOCK))
2383 /* Check to ensure the stack will not grow into a hugetlb-only region */
2384 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2386 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2390 * Overcommit.. This must be the final test, as it will
2391 * update security statistics.
2393 if (security_vm_enough_memory_mm(mm, grow))
2399 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2401 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2402 * vma is the last one with address > vma->vm_end. Have to extend vma.
2404 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2406 struct mm_struct *mm = vma->vm_mm;
2407 struct vm_area_struct *next;
2408 unsigned long gap_addr;
2411 if (!(vma->vm_flags & VM_GROWSUP))
2414 /* Guard against exceeding limits of the address space. */
2415 address &= PAGE_MASK;
2416 if (address >= (TASK_SIZE & PAGE_MASK))
2418 address += PAGE_SIZE;
2420 /* Enforce stack_guard_gap */
2421 gap_addr = address + stack_guard_gap;
2423 /* Guard against overflow */
2424 if (gap_addr < address || gap_addr > TASK_SIZE)
2425 gap_addr = TASK_SIZE;
2427 next = vma->vm_next;
2428 if (next && next->vm_start < gap_addr && vma_is_accessible(next)) {
2429 if (!(next->vm_flags & VM_GROWSUP))
2431 /* Check that both stack segments have the same anon_vma? */
2434 /* We must make sure the anon_vma is allocated. */
2435 if (unlikely(anon_vma_prepare(vma)))
2439 * vma->vm_start/vm_end cannot change under us because the caller
2440 * is required to hold the mmap_lock in read mode. We need the
2441 * anon_vma lock to serialize against concurrent expand_stacks.
2443 anon_vma_lock_write(vma->anon_vma);
2445 /* Somebody else might have raced and expanded it already */
2446 if (address > vma->vm_end) {
2447 unsigned long size, grow;
2449 size = address - vma->vm_start;
2450 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2453 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2454 error = acct_stack_growth(vma, size, grow);
2457 * vma_gap_update() doesn't support concurrent
2458 * updates, but we only hold a shared mmap_lock
2459 * lock here, so we need to protect against
2460 * concurrent vma expansions.
2461 * anon_vma_lock_write() doesn't help here, as
2462 * we don't guarantee that all growable vmas
2463 * in a mm share the same root anon vma.
2464 * So, we reuse mm->page_table_lock to guard
2465 * against concurrent vma expansions.
2467 spin_lock(&mm->page_table_lock);
2468 if (vma->vm_flags & VM_LOCKED)
2469 mm->locked_vm += grow;
2470 vm_stat_account(mm, vma->vm_flags, grow);
2471 anon_vma_interval_tree_pre_update_vma(vma);
2472 vma->vm_end = address;
2473 anon_vma_interval_tree_post_update_vma(vma);
2475 vma_gap_update(vma->vm_next);
2477 mm->highest_vm_end = vm_end_gap(vma);
2478 spin_unlock(&mm->page_table_lock);
2480 perf_event_mmap(vma);
2484 anon_vma_unlock_write(vma->anon_vma);
2485 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2489 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2492 * vma is the first one with address < vma->vm_start. Have to extend vma.
2494 int expand_downwards(struct vm_area_struct *vma,
2495 unsigned long address)
2497 struct mm_struct *mm = vma->vm_mm;
2498 struct vm_area_struct *prev;
2501 address &= PAGE_MASK;
2502 if (address < mmap_min_addr)
2505 /* Enforce stack_guard_gap */
2506 prev = vma->vm_prev;
2507 /* Check that both stack segments have the same anon_vma? */
2508 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2509 vma_is_accessible(prev)) {
2510 if (address - prev->vm_end < stack_guard_gap)
2514 /* We must make sure the anon_vma is allocated. */
2515 if (unlikely(anon_vma_prepare(vma)))
2519 * vma->vm_start/vm_end cannot change under us because the caller
2520 * is required to hold the mmap_lock in read mode. We need the
2521 * anon_vma lock to serialize against concurrent expand_stacks.
2523 anon_vma_lock_write(vma->anon_vma);
2525 /* Somebody else might have raced and expanded it already */
2526 if (address < vma->vm_start) {
2527 unsigned long size, grow;
2529 size = vma->vm_end - address;
2530 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2533 if (grow <= vma->vm_pgoff) {
2534 error = acct_stack_growth(vma, size, grow);
2537 * vma_gap_update() doesn't support concurrent
2538 * updates, but we only hold a shared mmap_lock
2539 * lock here, so we need to protect against
2540 * concurrent vma expansions.
2541 * anon_vma_lock_write() doesn't help here, as
2542 * we don't guarantee that all growable vmas
2543 * in a mm share the same root anon vma.
2544 * So, we reuse mm->page_table_lock to guard
2545 * against concurrent vma expansions.
2547 spin_lock(&mm->page_table_lock);
2548 if (vma->vm_flags & VM_LOCKED)
2549 mm->locked_vm += grow;
2550 vm_stat_account(mm, vma->vm_flags, grow);
2551 anon_vma_interval_tree_pre_update_vma(vma);
2552 vma->vm_start = address;
2553 vma->vm_pgoff -= grow;
2554 anon_vma_interval_tree_post_update_vma(vma);
2555 vma_gap_update(vma);
2556 spin_unlock(&mm->page_table_lock);
2558 perf_event_mmap(vma);
2562 anon_vma_unlock_write(vma->anon_vma);
2563 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2568 /* enforced gap between the expanding stack and other mappings. */
2569 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2571 static int __init cmdline_parse_stack_guard_gap(char *p)
2576 val = simple_strtoul(p, &endptr, 10);
2578 stack_guard_gap = val << PAGE_SHIFT;
2582 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2584 #ifdef CONFIG_STACK_GROWSUP
2585 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2587 return expand_upwards(vma, address);
2590 struct vm_area_struct *
2591 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2593 struct vm_area_struct *vma, *prev;
2596 vma = find_vma_prev(mm, addr, &prev);
2597 if (vma && (vma->vm_start <= addr))
2599 /* don't alter vm_end if the coredump is running */
2600 if (!prev || expand_stack(prev, addr))
2602 if (prev->vm_flags & VM_LOCKED)
2603 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2607 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2609 return expand_downwards(vma, address);
2612 struct vm_area_struct *
2613 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2615 struct vm_area_struct *vma;
2616 unsigned long start;
2619 vma = find_vma(mm, addr);
2622 if (vma->vm_start <= addr)
2624 if (!(vma->vm_flags & VM_GROWSDOWN))
2626 start = vma->vm_start;
2627 if (expand_stack(vma, addr))
2629 if (vma->vm_flags & VM_LOCKED)
2630 populate_vma_page_range(vma, addr, start, NULL);
2635 EXPORT_SYMBOL_GPL(find_extend_vma);
2638 * Ok - we have the memory areas we should free on the vma list,
2639 * so release them, and do the vma updates.
2641 * Called with the mm semaphore held.
2643 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2645 unsigned long nr_accounted = 0;
2647 /* Update high watermark before we lower total_vm */
2648 update_hiwater_vm(mm);
2650 long nrpages = vma_pages(vma);
2652 if (vma->vm_flags & VM_ACCOUNT)
2653 nr_accounted += nrpages;
2654 vm_stat_account(mm, vma->vm_flags, -nrpages);
2655 vma = remove_vma(vma);
2657 vm_unacct_memory(nr_accounted);
2662 * Get rid of page table information in the indicated region.
2664 * Called with the mm semaphore held.
2666 static void unmap_region(struct mm_struct *mm,
2667 struct vm_area_struct *vma, struct vm_area_struct *prev,
2668 unsigned long start, unsigned long end)
2670 struct vm_area_struct *next = vma_next(mm, prev);
2671 struct mmu_gather tlb;
2674 tlb_gather_mmu(&tlb, mm, start, end);
2675 update_hiwater_rss(mm);
2676 unmap_vmas(&tlb, vma, start, end);
2677 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2678 next ? next->vm_start : USER_PGTABLES_CEILING);
2679 tlb_finish_mmu(&tlb, start, end);
2683 * Create a list of vma's touched by the unmap, removing them from the mm's
2684 * vma list as we go..
2687 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2688 struct vm_area_struct *prev, unsigned long end)
2690 struct vm_area_struct **insertion_point;
2691 struct vm_area_struct *tail_vma = NULL;
2693 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2694 vma->vm_prev = NULL;
2696 vma_rb_erase(vma, &mm->mm_rb);
2700 } while (vma && vma->vm_start < end);
2701 *insertion_point = vma;
2703 vma->vm_prev = prev;
2704 vma_gap_update(vma);
2706 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2707 tail_vma->vm_next = NULL;
2709 /* Kill the cache */
2710 vmacache_invalidate(mm);
2713 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2714 * VM_GROWSUP VMA. Such VMAs can change their size under
2715 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2717 if (vma && (vma->vm_flags & VM_GROWSDOWN))
2719 if (prev && (prev->vm_flags & VM_GROWSUP))
2725 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2726 * has already been checked or doesn't make sense to fail.
2728 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2729 unsigned long addr, int new_below)
2731 struct vm_area_struct *new;
2734 if (vma->vm_ops && vma->vm_ops->split) {
2735 err = vma->vm_ops->split(vma, addr);
2740 new = vm_area_dup(vma);
2747 new->vm_start = addr;
2748 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2751 err = vma_dup_policy(vma, new);
2755 err = anon_vma_clone(new, vma);
2760 get_file(new->vm_file);
2762 if (new->vm_ops && new->vm_ops->open)
2763 new->vm_ops->open(new);
2766 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2767 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2769 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2775 /* Clean everything up if vma_adjust failed. */
2776 if (new->vm_ops && new->vm_ops->close)
2777 new->vm_ops->close(new);
2780 unlink_anon_vmas(new);
2782 mpol_put(vma_policy(new));
2789 * Split a vma into two pieces at address 'addr', a new vma is allocated
2790 * either for the first part or the tail.
2792 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2793 unsigned long addr, int new_below)
2795 if (mm->map_count >= sysctl_max_map_count)
2798 return __split_vma(mm, vma, addr, new_below);
2801 /* Munmap is split into 2 main parts -- this part which finds
2802 * what needs doing, and the areas themselves, which do the
2803 * work. This now handles partial unmappings.
2804 * Jeremy Fitzhardinge <jeremy@goop.org>
2806 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2807 struct list_head *uf, bool downgrade)
2810 struct vm_area_struct *vma, *prev, *last;
2812 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2815 len = PAGE_ALIGN(len);
2821 * arch_unmap() might do unmaps itself. It must be called
2822 * and finish any rbtree manipulation before this code
2823 * runs and also starts to manipulate the rbtree.
2825 arch_unmap(mm, start, end);
2827 /* Find the first overlapping VMA */
2828 vma = find_vma(mm, start);
2831 prev = vma->vm_prev;
2832 /* we have start < vma->vm_end */
2834 /* if it doesn't overlap, we have nothing.. */
2835 if (vma->vm_start >= end)
2839 * If we need to split any vma, do it now to save pain later.
2841 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2842 * unmapped vm_area_struct will remain in use: so lower split_vma
2843 * places tmp vma above, and higher split_vma places tmp vma below.
2845 if (start > vma->vm_start) {
2849 * Make sure that map_count on return from munmap() will
2850 * not exceed its limit; but let map_count go just above
2851 * its limit temporarily, to help free resources as expected.
2853 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2856 error = __split_vma(mm, vma, start, 0);
2862 /* Does it split the last one? */
2863 last = find_vma(mm, end);
2864 if (last && end > last->vm_start) {
2865 int error = __split_vma(mm, last, end, 1);
2869 vma = vma_next(mm, prev);
2873 * If userfaultfd_unmap_prep returns an error the vmas
2874 * will remain splitted, but userland will get a
2875 * highly unexpected error anyway. This is no
2876 * different than the case where the first of the two
2877 * __split_vma fails, but we don't undo the first
2878 * split, despite we could. This is unlikely enough
2879 * failure that it's not worth optimizing it for.
2881 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2887 * unlock any mlock()ed ranges before detaching vmas
2889 if (mm->locked_vm) {
2890 struct vm_area_struct *tmp = vma;
2891 while (tmp && tmp->vm_start < end) {
2892 if (tmp->vm_flags & VM_LOCKED) {
2893 mm->locked_vm -= vma_pages(tmp);
2894 munlock_vma_pages_all(tmp);
2901 /* Detach vmas from rbtree */
2902 if (!detach_vmas_to_be_unmapped(mm, vma, prev, end))
2906 mmap_write_downgrade(mm);
2908 unmap_region(mm, vma, prev, start, end);
2910 /* Fix up all other VM information */
2911 remove_vma_list(mm, vma);
2913 return downgrade ? 1 : 0;
2916 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2917 struct list_head *uf)
2919 return __do_munmap(mm, start, len, uf, false);
2922 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2925 struct mm_struct *mm = current->mm;
2928 if (mmap_write_lock_killable(mm))
2931 ret = __do_munmap(mm, start, len, &uf, downgrade);
2933 * Returning 1 indicates mmap_lock is downgraded.
2934 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2935 * it to 0 before return.
2938 mmap_read_unlock(mm);
2941 mmap_write_unlock(mm);
2943 userfaultfd_unmap_complete(mm, &uf);
2947 int vm_munmap(unsigned long start, size_t len)
2949 return __vm_munmap(start, len, false);
2951 EXPORT_SYMBOL(vm_munmap);
2953 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2955 addr = untagged_addr(addr);
2956 profile_munmap(addr);
2957 return __vm_munmap(addr, len, true);
2962 * Emulation of deprecated remap_file_pages() syscall.
2964 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2965 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2968 struct mm_struct *mm = current->mm;
2969 struct vm_area_struct *vma;
2970 unsigned long populate = 0;
2971 unsigned long ret = -EINVAL;
2974 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2975 current->comm, current->pid);
2979 start = start & PAGE_MASK;
2980 size = size & PAGE_MASK;
2982 if (start + size <= start)
2985 /* Does pgoff wrap? */
2986 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2989 if (mmap_write_lock_killable(mm))
2992 vma = find_vma(mm, start);
2994 if (!vma || !(vma->vm_flags & VM_SHARED))
2997 if (start < vma->vm_start)
3000 if (start + size > vma->vm_end) {
3001 struct vm_area_struct *next;
3003 for (next = vma->vm_next; next; next = next->vm_next) {
3004 /* hole between vmas ? */
3005 if (next->vm_start != next->vm_prev->vm_end)
3008 if (next->vm_file != vma->vm_file)
3011 if (next->vm_flags != vma->vm_flags)
3014 if (start + size <= next->vm_end)
3022 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3023 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3024 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3026 flags &= MAP_NONBLOCK;
3027 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3028 if (vma->vm_flags & VM_LOCKED) {
3029 struct vm_area_struct *tmp;
3030 flags |= MAP_LOCKED;
3032 /* drop PG_Mlocked flag for over-mapped range */
3033 for (tmp = vma; tmp->vm_start >= start + size;
3034 tmp = tmp->vm_next) {
3036 * Split pmd and munlock page on the border
3039 vma_adjust_trans_huge(tmp, start, start + size, 0);
3041 munlock_vma_pages_range(tmp,
3042 max(tmp->vm_start, start),
3043 min(tmp->vm_end, start + size));
3047 file = get_file(vma->vm_file);
3048 ret = do_mmap(vma->vm_file, start, size,
3049 prot, flags, pgoff, &populate, NULL);
3052 mmap_write_unlock(mm);
3054 mm_populate(ret, populate);
3055 if (!IS_ERR_VALUE(ret))
3061 * this is really a simplified "do_mmap". it only handles
3062 * anonymous maps. eventually we may be able to do some
3063 * brk-specific accounting here.
3065 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
3067 struct mm_struct *mm = current->mm;
3068 struct vm_area_struct *vma, *prev;
3069 struct rb_node **rb_link, *rb_parent;
3070 pgoff_t pgoff = addr >> PAGE_SHIFT;
3072 unsigned long mapped_addr;
3074 /* Until we need other flags, refuse anything except VM_EXEC. */
3075 if ((flags & (~VM_EXEC)) != 0)
3077 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3079 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
3080 if (IS_ERR_VALUE(mapped_addr))
3083 error = mlock_future_check(mm, mm->def_flags, len);
3087 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
3088 if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
3091 /* Check against address space limits *after* clearing old maps... */
3092 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3095 if (mm->map_count > sysctl_max_map_count)
3098 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3101 /* Can we just expand an old private anonymous mapping? */
3102 vma = vma_merge(mm, prev, addr, addr + len, flags,
3103 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
3108 * create a vma struct for an anonymous mapping
3110 vma = vm_area_alloc(mm);
3112 vm_unacct_memory(len >> PAGE_SHIFT);
3116 vma_set_anonymous(vma);
3117 vma->vm_start = addr;
3118 vma->vm_end = addr + len;
3119 vma->vm_pgoff = pgoff;
3120 vma->vm_flags = flags;
3121 vma->vm_page_prot = vm_get_page_prot(flags);
3122 vma_link(mm, vma, prev, rb_link, rb_parent);
3124 perf_event_mmap(vma);
3125 mm->total_vm += len >> PAGE_SHIFT;
3126 mm->data_vm += len >> PAGE_SHIFT;
3127 if (flags & VM_LOCKED)
3128 mm->locked_vm += (len >> PAGE_SHIFT);
3129 vma->vm_flags |= VM_SOFTDIRTY;
3133 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3135 struct mm_struct *mm = current->mm;
3141 len = PAGE_ALIGN(request);
3147 if (mmap_write_lock_killable(mm))
3150 ret = do_brk_flags(addr, len, flags, &uf);
3151 populate = ((mm->def_flags & VM_LOCKED) != 0);
3152 mmap_write_unlock(mm);
3153 userfaultfd_unmap_complete(mm, &uf);
3154 if (populate && !ret)
3155 mm_populate(addr, len);
3158 EXPORT_SYMBOL(vm_brk_flags);
3160 int vm_brk(unsigned long addr, unsigned long len)
3162 return vm_brk_flags(addr, len, 0);
3164 EXPORT_SYMBOL(vm_brk);
3166 /* Release all mmaps. */
3167 void exit_mmap(struct mm_struct *mm)
3169 struct mmu_gather tlb;
3170 struct vm_area_struct *vma;
3171 unsigned long nr_accounted = 0;
3173 /* mm's last user has gone, and its about to be pulled down */
3174 mmu_notifier_release(mm);
3176 if (unlikely(mm_is_oom_victim(mm))) {
3178 * Manually reap the mm to free as much memory as possible.
3179 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3180 * this mm from further consideration. Taking mm->mmap_lock for
3181 * write after setting MMF_OOM_SKIP will guarantee that the oom
3182 * reaper will not run on this mm again after mmap_lock is
3185 * Nothing can be holding mm->mmap_lock here and the above call
3186 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3187 * __oom_reap_task_mm() will not block.
3189 * This needs to be done before calling munlock_vma_pages_all(),
3190 * which clears VM_LOCKED, otherwise the oom reaper cannot
3193 (void)__oom_reap_task_mm(mm);
3195 set_bit(MMF_OOM_SKIP, &mm->flags);
3196 mmap_write_lock(mm);
3197 mmap_write_unlock(mm);
3200 if (mm->locked_vm) {
3203 if (vma->vm_flags & VM_LOCKED)
3204 munlock_vma_pages_all(vma);
3212 if (!vma) /* Can happen if dup_mmap() received an OOM */
3217 tlb_gather_mmu(&tlb, mm, 0, -1);
3218 /* update_hiwater_rss(mm) here? but nobody should be looking */
3219 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3220 unmap_vmas(&tlb, vma, 0, -1);
3221 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3222 tlb_finish_mmu(&tlb, 0, -1);
3225 * Walk the list again, actually closing and freeing it,
3226 * with preemption enabled, without holding any MM locks.
3229 if (vma->vm_flags & VM_ACCOUNT)
3230 nr_accounted += vma_pages(vma);
3231 vma = remove_vma(vma);
3234 vm_unacct_memory(nr_accounted);
3237 /* Insert vm structure into process list sorted by address
3238 * and into the inode's i_mmap tree. If vm_file is non-NULL
3239 * then i_mmap_rwsem is taken here.
3241 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3243 struct vm_area_struct *prev;
3244 struct rb_node **rb_link, *rb_parent;
3246 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3247 &prev, &rb_link, &rb_parent))
3249 if ((vma->vm_flags & VM_ACCOUNT) &&
3250 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3254 * The vm_pgoff of a purely anonymous vma should be irrelevant
3255 * until its first write fault, when page's anon_vma and index
3256 * are set. But now set the vm_pgoff it will almost certainly
3257 * end up with (unless mremap moves it elsewhere before that
3258 * first wfault), so /proc/pid/maps tells a consistent story.
3260 * By setting it to reflect the virtual start address of the
3261 * vma, merges and splits can happen in a seamless way, just
3262 * using the existing file pgoff checks and manipulations.
3263 * Similarly in do_mmap and in do_brk_flags.
3265 if (vma_is_anonymous(vma)) {
3266 BUG_ON(vma->anon_vma);
3267 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3270 vma_link(mm, vma, prev, rb_link, rb_parent);
3275 * Copy the vma structure to a new location in the same mm,
3276 * prior to moving page table entries, to effect an mremap move.
3278 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3279 unsigned long addr, unsigned long len, pgoff_t pgoff,
3280 bool *need_rmap_locks)
3282 struct vm_area_struct *vma = *vmap;
3283 unsigned long vma_start = vma->vm_start;
3284 struct mm_struct *mm = vma->vm_mm;
3285 struct vm_area_struct *new_vma, *prev;
3286 struct rb_node **rb_link, *rb_parent;
3287 bool faulted_in_anon_vma = true;
3290 * If anonymous vma has not yet been faulted, update new pgoff
3291 * to match new location, to increase its chance of merging.
3293 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3294 pgoff = addr >> PAGE_SHIFT;
3295 faulted_in_anon_vma = false;
3298 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3299 return NULL; /* should never get here */
3300 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3301 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3302 vma->vm_userfaultfd_ctx);
3305 * Source vma may have been merged into new_vma
3307 if (unlikely(vma_start >= new_vma->vm_start &&
3308 vma_start < new_vma->vm_end)) {
3310 * The only way we can get a vma_merge with
3311 * self during an mremap is if the vma hasn't
3312 * been faulted in yet and we were allowed to
3313 * reset the dst vma->vm_pgoff to the
3314 * destination address of the mremap to allow
3315 * the merge to happen. mremap must change the
3316 * vm_pgoff linearity between src and dst vmas
3317 * (in turn preventing a vma_merge) to be
3318 * safe. It is only safe to keep the vm_pgoff
3319 * linear if there are no pages mapped yet.
3321 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3322 *vmap = vma = new_vma;
3324 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3326 new_vma = vm_area_dup(vma);
3329 new_vma->vm_start = addr;
3330 new_vma->vm_end = addr + len;
3331 new_vma->vm_pgoff = pgoff;
3332 if (vma_dup_policy(vma, new_vma))
3334 if (anon_vma_clone(new_vma, vma))
3335 goto out_free_mempol;
3336 if (new_vma->vm_file)
3337 get_file(new_vma->vm_file);
3338 if (new_vma->vm_ops && new_vma->vm_ops->open)
3339 new_vma->vm_ops->open(new_vma);
3340 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3341 *need_rmap_locks = false;
3346 mpol_put(vma_policy(new_vma));
3348 vm_area_free(new_vma);
3354 * Return true if the calling process may expand its vm space by the passed
3357 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3359 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3362 if (is_data_mapping(flags) &&
3363 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3364 /* Workaround for Valgrind */
3365 if (rlimit(RLIMIT_DATA) == 0 &&
3366 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3369 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3370 current->comm, current->pid,
3371 (mm->data_vm + npages) << PAGE_SHIFT,
3372 rlimit(RLIMIT_DATA),
3373 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3375 if (!ignore_rlimit_data)
3382 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3384 mm->total_vm += npages;
3386 if (is_exec_mapping(flags))
3387 mm->exec_vm += npages;
3388 else if (is_stack_mapping(flags))
3389 mm->stack_vm += npages;
3390 else if (is_data_mapping(flags))
3391 mm->data_vm += npages;
3394 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3397 * Having a close hook prevents vma merging regardless of flags.
3399 static void special_mapping_close(struct vm_area_struct *vma)
3403 static const char *special_mapping_name(struct vm_area_struct *vma)
3405 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3408 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3410 struct vm_special_mapping *sm = new_vma->vm_private_data;
3412 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3416 return sm->mremap(sm, new_vma);
3421 static const struct vm_operations_struct special_mapping_vmops = {
3422 .close = special_mapping_close,
3423 .fault = special_mapping_fault,
3424 .mremap = special_mapping_mremap,
3425 .name = special_mapping_name,
3426 /* vDSO code relies that VVAR can't be accessed remotely */
3430 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3431 .close = special_mapping_close,
3432 .fault = special_mapping_fault,
3435 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3437 struct vm_area_struct *vma = vmf->vma;
3439 struct page **pages;
3441 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3442 pages = vma->vm_private_data;
3444 struct vm_special_mapping *sm = vma->vm_private_data;
3447 return sm->fault(sm, vmf->vma, vmf);
3452 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3456 struct page *page = *pages;
3462 return VM_FAULT_SIGBUS;
3465 static struct vm_area_struct *__install_special_mapping(
3466 struct mm_struct *mm,
3467 unsigned long addr, unsigned long len,
3468 unsigned long vm_flags, void *priv,
3469 const struct vm_operations_struct *ops)
3472 struct vm_area_struct *vma;
3474 vma = vm_area_alloc(mm);
3475 if (unlikely(vma == NULL))
3476 return ERR_PTR(-ENOMEM);
3478 vma->vm_start = addr;
3479 vma->vm_end = addr + len;
3481 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3482 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3485 vma->vm_private_data = priv;
3487 ret = insert_vm_struct(mm, vma);
3491 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3493 perf_event_mmap(vma);
3499 return ERR_PTR(ret);
3502 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3503 const struct vm_special_mapping *sm)
3505 return vma->vm_private_data == sm &&
3506 (vma->vm_ops == &special_mapping_vmops ||
3507 vma->vm_ops == &legacy_special_mapping_vmops);
3511 * Called with mm->mmap_lock held for writing.
3512 * Insert a new vma covering the given region, with the given flags.
3513 * Its pages are supplied by the given array of struct page *.
3514 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3515 * The region past the last page supplied will always produce SIGBUS.
3516 * The array pointer and the pages it points to are assumed to stay alive
3517 * for as long as this mapping might exist.
3519 struct vm_area_struct *_install_special_mapping(
3520 struct mm_struct *mm,
3521 unsigned long addr, unsigned long len,
3522 unsigned long vm_flags, const struct vm_special_mapping *spec)
3524 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3525 &special_mapping_vmops);
3528 int install_special_mapping(struct mm_struct *mm,
3529 unsigned long addr, unsigned long len,
3530 unsigned long vm_flags, struct page **pages)
3532 struct vm_area_struct *vma = __install_special_mapping(
3533 mm, addr, len, vm_flags, (void *)pages,
3534 &legacy_special_mapping_vmops);
3536 return PTR_ERR_OR_ZERO(vma);
3539 static DEFINE_MUTEX(mm_all_locks_mutex);
3541 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3543 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3545 * The LSB of head.next can't change from under us
3546 * because we hold the mm_all_locks_mutex.
3548 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3550 * We can safely modify head.next after taking the
3551 * anon_vma->root->rwsem. If some other vma in this mm shares
3552 * the same anon_vma we won't take it again.
3554 * No need of atomic instructions here, head.next
3555 * can't change from under us thanks to the
3556 * anon_vma->root->rwsem.
3558 if (__test_and_set_bit(0, (unsigned long *)
3559 &anon_vma->root->rb_root.rb_root.rb_node))
3564 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3566 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3568 * AS_MM_ALL_LOCKS can't change from under us because
3569 * we hold the mm_all_locks_mutex.
3571 * Operations on ->flags have to be atomic because
3572 * even if AS_MM_ALL_LOCKS is stable thanks to the
3573 * mm_all_locks_mutex, there may be other cpus
3574 * changing other bitflags in parallel to us.
3576 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3578 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3583 * This operation locks against the VM for all pte/vma/mm related
3584 * operations that could ever happen on a certain mm. This includes
3585 * vmtruncate, try_to_unmap, and all page faults.
3587 * The caller must take the mmap_lock in write mode before calling
3588 * mm_take_all_locks(). The caller isn't allowed to release the
3589 * mmap_lock until mm_drop_all_locks() returns.
3591 * mmap_lock in write mode is required in order to block all operations
3592 * that could modify pagetables and free pages without need of
3593 * altering the vma layout. It's also needed in write mode to avoid new
3594 * anon_vmas to be associated with existing vmas.
3596 * A single task can't take more than one mm_take_all_locks() in a row
3597 * or it would deadlock.
3599 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3600 * mapping->flags avoid to take the same lock twice, if more than one
3601 * vma in this mm is backed by the same anon_vma or address_space.
3603 * We take locks in following order, accordingly to comment at beginning
3605 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3607 * - all i_mmap_rwsem locks;
3608 * - all anon_vma->rwseml
3610 * We can take all locks within these types randomly because the VM code
3611 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3612 * mm_all_locks_mutex.
3614 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3615 * that may have to take thousand of locks.
3617 * mm_take_all_locks() can fail if it's interrupted by signals.
3619 int mm_take_all_locks(struct mm_struct *mm)
3621 struct vm_area_struct *vma;
3622 struct anon_vma_chain *avc;
3624 BUG_ON(mmap_read_trylock(mm));
3626 mutex_lock(&mm_all_locks_mutex);
3628 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3629 if (signal_pending(current))
3631 if (vma->vm_file && vma->vm_file->f_mapping &&
3632 is_vm_hugetlb_page(vma))
3633 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3636 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3637 if (signal_pending(current))
3639 if (vma->vm_file && vma->vm_file->f_mapping &&
3640 !is_vm_hugetlb_page(vma))
3641 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3644 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3645 if (signal_pending(current))
3648 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3649 vm_lock_anon_vma(mm, avc->anon_vma);
3655 mm_drop_all_locks(mm);
3659 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3661 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3663 * The LSB of head.next can't change to 0 from under
3664 * us because we hold the mm_all_locks_mutex.
3666 * We must however clear the bitflag before unlocking
3667 * the vma so the users using the anon_vma->rb_root will
3668 * never see our bitflag.
3670 * No need of atomic instructions here, head.next
3671 * can't change from under us until we release the
3672 * anon_vma->root->rwsem.
3674 if (!__test_and_clear_bit(0, (unsigned long *)
3675 &anon_vma->root->rb_root.rb_root.rb_node))
3677 anon_vma_unlock_write(anon_vma);
3681 static void vm_unlock_mapping(struct address_space *mapping)
3683 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3685 * AS_MM_ALL_LOCKS can't change to 0 from under us
3686 * because we hold the mm_all_locks_mutex.
3688 i_mmap_unlock_write(mapping);
3689 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3696 * The mmap_lock cannot be released by the caller until
3697 * mm_drop_all_locks() returns.
3699 void mm_drop_all_locks(struct mm_struct *mm)
3701 struct vm_area_struct *vma;
3702 struct anon_vma_chain *avc;
3704 BUG_ON(mmap_read_trylock(mm));
3705 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3707 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3709 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3710 vm_unlock_anon_vma(avc->anon_vma);
3711 if (vma->vm_file && vma->vm_file->f_mapping)
3712 vm_unlock_mapping(vma->vm_file->f_mapping);
3715 mutex_unlock(&mm_all_locks_mutex);
3719 * initialise the percpu counter for VM
3721 void __init mmap_init(void)
3725 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3730 * Initialise sysctl_user_reserve_kbytes.
3732 * This is intended to prevent a user from starting a single memory hogging
3733 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3736 * The default value is min(3% of free memory, 128MB)
3737 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3739 static int init_user_reserve(void)
3741 unsigned long free_kbytes;
3743 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3745 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3748 subsys_initcall(init_user_reserve);
3751 * Initialise sysctl_admin_reserve_kbytes.
3753 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3754 * to log in and kill a memory hogging process.
3756 * Systems with more than 256MB will reserve 8MB, enough to recover
3757 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3758 * only reserve 3% of free pages by default.
3760 static int init_admin_reserve(void)
3762 unsigned long free_kbytes;
3764 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3766 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3769 subsys_initcall(init_admin_reserve);
3772 * Reinititalise user and admin reserves if memory is added or removed.
3774 * The default user reserve max is 128MB, and the default max for the
3775 * admin reserve is 8MB. These are usually, but not always, enough to
3776 * enable recovery from a memory hogging process using login/sshd, a shell,
3777 * and tools like top. It may make sense to increase or even disable the
3778 * reserve depending on the existence of swap or variations in the recovery
3779 * tools. So, the admin may have changed them.
3781 * If memory is added and the reserves have been eliminated or increased above
3782 * the default max, then we'll trust the admin.
3784 * If memory is removed and there isn't enough free memory, then we
3785 * need to reset the reserves.
3787 * Otherwise keep the reserve set by the admin.
3789 static int reserve_mem_notifier(struct notifier_block *nb,
3790 unsigned long action, void *data)
3792 unsigned long tmp, free_kbytes;
3796 /* Default max is 128MB. Leave alone if modified by operator. */
3797 tmp = sysctl_user_reserve_kbytes;
3798 if (0 < tmp && tmp < (1UL << 17))
3799 init_user_reserve();
3801 /* Default max is 8MB. Leave alone if modified by operator. */
3802 tmp = sysctl_admin_reserve_kbytes;
3803 if (0 < tmp && tmp < (1UL << 13))
3804 init_admin_reserve();
3808 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3810 if (sysctl_user_reserve_kbytes > free_kbytes) {
3811 init_user_reserve();
3812 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3813 sysctl_user_reserve_kbytes);
3816 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3817 init_admin_reserve();
3818 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3819 sysctl_admin_reserve_kbytes);
3828 static struct notifier_block reserve_mem_nb = {
3829 .notifier_call = reserve_mem_notifier,
3832 static int __meminit init_reserve_notifier(void)
3834 if (register_hotmemory_notifier(&reserve_mem_nb))
3835 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3839 subsys_initcall(init_reserve_notifier);