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;
561 static unsigned long count_vma_pages_range(struct mm_struct *mm,
562 unsigned long addr, unsigned long end)
564 unsigned long nr_pages = 0;
565 struct vm_area_struct *vma;
567 /* Find first overlaping mapping */
568 vma = find_vma_intersection(mm, addr, end);
572 nr_pages = (min(end, vma->vm_end) -
573 max(addr, vma->vm_start)) >> PAGE_SHIFT;
575 /* Iterate over the rest of the overlaps */
576 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
577 unsigned long overlap_len;
579 if (vma->vm_start > end)
582 overlap_len = min(end, vma->vm_end) - vma->vm_start;
583 nr_pages += overlap_len >> PAGE_SHIFT;
589 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
590 struct rb_node **rb_link, struct rb_node *rb_parent)
592 /* Update tracking information for the gap following the new vma. */
594 vma_gap_update(vma->vm_next);
596 mm->highest_vm_end = vm_end_gap(vma);
599 * vma->vm_prev wasn't known when we followed the rbtree to find the
600 * correct insertion point for that vma. As a result, we could not
601 * update the vma vm_rb parents rb_subtree_gap values on the way down.
602 * So, we first insert the vma with a zero rb_subtree_gap value
603 * (to be consistent with what we did on the way down), and then
604 * immediately update the gap to the correct value. Finally we
605 * rebalance the rbtree after all augmented values have been set.
607 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
608 vma->rb_subtree_gap = 0;
610 vma_rb_insert(vma, &mm->mm_rb);
613 static void __vma_link_file(struct vm_area_struct *vma)
619 struct address_space *mapping = file->f_mapping;
621 if (vma->vm_flags & VM_DENYWRITE)
622 put_write_access(file_inode(file));
623 if (vma->vm_flags & VM_SHARED)
624 mapping_allow_writable(mapping);
626 flush_dcache_mmap_lock(mapping);
627 vma_interval_tree_insert(vma, &mapping->i_mmap);
628 flush_dcache_mmap_unlock(mapping);
633 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
634 struct vm_area_struct *prev, struct rb_node **rb_link,
635 struct rb_node *rb_parent)
637 __vma_link_list(mm, vma, prev);
638 __vma_link_rb(mm, vma, rb_link, rb_parent);
641 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
642 struct vm_area_struct *prev, struct rb_node **rb_link,
643 struct rb_node *rb_parent)
645 struct address_space *mapping = NULL;
648 mapping = vma->vm_file->f_mapping;
649 i_mmap_lock_write(mapping);
652 __vma_link(mm, vma, prev, rb_link, rb_parent);
653 __vma_link_file(vma);
656 i_mmap_unlock_write(mapping);
663 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
664 * mm's list and rbtree. It has already been inserted into the interval tree.
666 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
668 struct vm_area_struct *prev;
669 struct rb_node **rb_link, *rb_parent;
671 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
672 &prev, &rb_link, &rb_parent))
674 __vma_link(mm, vma, prev, rb_link, rb_parent);
678 static __always_inline void __vma_unlink(struct mm_struct *mm,
679 struct vm_area_struct *vma,
680 struct vm_area_struct *ignore)
682 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
683 __vma_unlink_list(mm, vma);
685 vmacache_invalidate(mm);
689 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
690 * is already present in an i_mmap tree without adjusting the tree.
691 * The following helper function should be used when such adjustments
692 * are necessary. The "insert" vma (if any) is to be inserted
693 * before we drop the necessary locks.
695 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
696 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
697 struct vm_area_struct *expand)
699 struct mm_struct *mm = vma->vm_mm;
700 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
701 struct address_space *mapping = NULL;
702 struct rb_root_cached *root = NULL;
703 struct anon_vma *anon_vma = NULL;
704 struct file *file = vma->vm_file;
705 bool start_changed = false, end_changed = false;
706 long adjust_next = 0;
709 if (next && !insert) {
710 struct vm_area_struct *exporter = NULL, *importer = NULL;
712 if (end >= next->vm_end) {
714 * vma expands, overlapping all the next, and
715 * perhaps the one after too (mprotect case 6).
716 * The only other cases that gets here are
717 * case 1, case 7 and case 8.
719 if (next == expand) {
721 * The only case where we don't expand "vma"
722 * and we expand "next" instead is case 8.
724 VM_WARN_ON(end != next->vm_end);
726 * remove_next == 3 means we're
727 * removing "vma" and that to do so we
728 * swapped "vma" and "next".
731 VM_WARN_ON(file != next->vm_file);
734 VM_WARN_ON(expand != vma);
736 * case 1, 6, 7, remove_next == 2 is case 6,
737 * remove_next == 1 is case 1 or 7.
739 remove_next = 1 + (end > next->vm_end);
740 VM_WARN_ON(remove_next == 2 &&
741 end != next->vm_next->vm_end);
742 /* trim end to next, for case 6 first pass */
750 * If next doesn't have anon_vma, import from vma after
751 * next, if the vma overlaps with it.
753 if (remove_next == 2 && !next->anon_vma)
754 exporter = next->vm_next;
756 } else if (end > next->vm_start) {
758 * vma expands, overlapping part of the next:
759 * mprotect case 5 shifting the boundary up.
761 adjust_next = (end - next->vm_start);
764 VM_WARN_ON(expand != importer);
765 } else if (end < vma->vm_end) {
767 * vma shrinks, and !insert tells it's not
768 * split_vma inserting another: so it must be
769 * mprotect case 4 shifting the boundary down.
771 adjust_next = -(vma->vm_end - end);
774 VM_WARN_ON(expand != importer);
778 * Easily overlooked: when mprotect shifts the boundary,
779 * make sure the expanding vma has anon_vma set if the
780 * shrinking vma had, to cover any anon pages imported.
782 if (exporter && exporter->anon_vma && !importer->anon_vma) {
785 importer->anon_vma = exporter->anon_vma;
786 error = anon_vma_clone(importer, exporter);
792 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
795 mapping = file->f_mapping;
796 root = &mapping->i_mmap;
797 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
800 uprobe_munmap(next, next->vm_start, next->vm_end);
802 i_mmap_lock_write(mapping);
805 * Put into interval tree now, so instantiated pages
806 * are visible to arm/parisc __flush_dcache_page
807 * throughout; but we cannot insert into address
808 * space until vma start or end is updated.
810 __vma_link_file(insert);
814 anon_vma = vma->anon_vma;
815 if (!anon_vma && adjust_next)
816 anon_vma = next->anon_vma;
818 VM_WARN_ON(adjust_next && next->anon_vma &&
819 anon_vma != next->anon_vma);
820 anon_vma_lock_write(anon_vma);
821 anon_vma_interval_tree_pre_update_vma(vma);
823 anon_vma_interval_tree_pre_update_vma(next);
827 flush_dcache_mmap_lock(mapping);
828 vma_interval_tree_remove(vma, root);
830 vma_interval_tree_remove(next, root);
833 if (start != vma->vm_start) {
834 vma->vm_start = start;
835 start_changed = true;
837 if (end != vma->vm_end) {
841 vma->vm_pgoff = pgoff;
843 next->vm_start += adjust_next;
844 next->vm_pgoff += adjust_next >> PAGE_SHIFT;
849 vma_interval_tree_insert(next, root);
850 vma_interval_tree_insert(vma, root);
851 flush_dcache_mmap_unlock(mapping);
856 * vma_merge has merged next into vma, and needs
857 * us to remove next before dropping the locks.
859 if (remove_next != 3)
860 __vma_unlink(mm, next, next);
863 * vma is not before next if they've been
866 * pre-swap() next->vm_start was reduced so
867 * tell validate_mm_rb to ignore pre-swap()
868 * "next" (which is stored in post-swap()
871 __vma_unlink(mm, next, vma);
873 __remove_shared_vm_struct(next, file, mapping);
876 * split_vma has split insert from vma, and needs
877 * us to insert it before dropping the locks
878 * (it may either follow vma or precede it).
880 __insert_vm_struct(mm, insert);
886 mm->highest_vm_end = vm_end_gap(vma);
887 else if (!adjust_next)
888 vma_gap_update(next);
893 anon_vma_interval_tree_post_update_vma(vma);
895 anon_vma_interval_tree_post_update_vma(next);
896 anon_vma_unlock_write(anon_vma);
900 i_mmap_unlock_write(mapping);
909 uprobe_munmap(next, next->vm_start, next->vm_end);
913 anon_vma_merge(vma, next);
915 mpol_put(vma_policy(next));
918 * In mprotect's case 6 (see comments on vma_merge),
919 * we must remove another next too. It would clutter
920 * up the code too much to do both in one go.
922 if (remove_next != 3) {
924 * If "next" was removed and vma->vm_end was
925 * expanded (up) over it, in turn
926 * "next->vm_prev->vm_end" changed and the
927 * "vma->vm_next" gap must be updated.
932 * For the scope of the comment "next" and
933 * "vma" considered pre-swap(): if "vma" was
934 * removed, next->vm_start was expanded (down)
935 * over it and the "next" gap must be updated.
936 * Because of the swap() the post-swap() "vma"
937 * actually points to pre-swap() "next"
938 * (post-swap() "next" as opposed is now a
943 if (remove_next == 2) {
949 vma_gap_update(next);
952 * If remove_next == 2 we obviously can't
955 * If remove_next == 3 we can't reach this
956 * path because pre-swap() next is always not
957 * NULL. pre-swap() "next" is not being
958 * removed and its next->vm_end is not altered
959 * (and furthermore "end" already matches
960 * next->vm_end in remove_next == 3).
962 * We reach this only in the remove_next == 1
963 * case if the "next" vma that was removed was
964 * the highest vma of the mm. However in such
965 * case next->vm_end == "end" and the extended
966 * "vma" has vma->vm_end == next->vm_end so
967 * mm->highest_vm_end doesn't need any update
968 * in remove_next == 1 case.
970 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
982 * If the vma has a ->close operation then the driver probably needs to release
983 * per-vma resources, so we don't attempt to merge those.
985 static inline int is_mergeable_vma(struct vm_area_struct *vma,
986 struct file *file, unsigned long vm_flags,
987 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
990 * VM_SOFTDIRTY should not prevent from VMA merging, if we
991 * match the flags but dirty bit -- the caller should mark
992 * merged VMA as dirty. If dirty bit won't be excluded from
993 * comparison, we increase pressure on the memory system forcing
994 * the kernel to generate new VMAs when old one could be
997 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
999 if (vma->vm_file != file)
1001 if (vma->vm_ops && vma->vm_ops->close)
1003 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1008 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1009 struct anon_vma *anon_vma2,
1010 struct vm_area_struct *vma)
1013 * The list_is_singular() test is to avoid merging VMA cloned from
1014 * parents. This can improve scalability caused by anon_vma lock.
1016 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1017 list_is_singular(&vma->anon_vma_chain)))
1019 return anon_vma1 == anon_vma2;
1023 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1024 * in front of (at a lower virtual address and file offset than) the vma.
1026 * We cannot merge two vmas if they have differently assigned (non-NULL)
1027 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1029 * We don't check here for the merged mmap wrapping around the end of pagecache
1030 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
1031 * wrap, nor mmaps which cover the final page at index -1UL.
1034 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1035 struct anon_vma *anon_vma, struct file *file,
1037 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1039 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1040 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1041 if (vma->vm_pgoff == vm_pgoff)
1048 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1049 * beyond (at a higher virtual address and file offset than) the vma.
1051 * We cannot merge two vmas if they have differently assigned (non-NULL)
1052 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1055 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1056 struct anon_vma *anon_vma, struct file *file,
1058 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1060 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1061 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1063 vm_pglen = vma_pages(vma);
1064 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1071 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1072 * whether that can be merged with its predecessor or its successor.
1073 * Or both (it neatly fills a hole).
1075 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1076 * certain not to be mapped by the time vma_merge is called; but when
1077 * called for mprotect, it is certain to be already mapped (either at
1078 * an offset within prev, or at the start of next), and the flags of
1079 * this area are about to be changed to vm_flags - and the no-change
1080 * case has already been eliminated.
1082 * The following mprotect cases have to be considered, where AAAA is
1083 * the area passed down from mprotect_fixup, never extending beyond one
1084 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1087 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN
1088 * cannot merge might become might become
1089 * PPNNNNNNNNNN PPPPPPPPPPNN
1090 * mmap, brk or case 4 below case 5 below
1093 * PPPP NNNN PPPPNNNNXXXX
1094 * might become might become
1095 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
1096 * PPPPPPPPNNNN 2 or PPPPPPPPXXXX 7 or
1097 * PPPPNNNNNNNN 3 PPPPXXXXXXXX 8
1099 * It is important for case 8 that the vma NNNN overlapping the
1100 * region AAAA is never going to extended over XXXX. Instead XXXX must
1101 * be extended in region AAAA and NNNN must be removed. This way in
1102 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1103 * rmap_locks, the properties of the merged vma will be already
1104 * correct for the whole merged range. Some of those properties like
1105 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1106 * be correct for the whole merged range immediately after the
1107 * rmap_locks are released. Otherwise if XXXX would be removed and
1108 * NNNN would be extended over the XXXX range, remove_migration_ptes
1109 * or other rmap walkers (if working on addresses beyond the "end"
1110 * parameter) may establish ptes with the wrong permissions of NNNN
1111 * instead of the right permissions of XXXX.
1113 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1114 struct vm_area_struct *prev, unsigned long addr,
1115 unsigned long end, unsigned long vm_flags,
1116 struct anon_vma *anon_vma, struct file *file,
1117 pgoff_t pgoff, struct mempolicy *policy,
1118 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1120 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1121 struct vm_area_struct *area, *next;
1125 * We later require that vma->vm_flags == vm_flags,
1126 * so this tests vma->vm_flags & VM_SPECIAL, too.
1128 if (vm_flags & VM_SPECIAL)
1132 next = prev->vm_next;
1136 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1137 next = next->vm_next;
1139 /* verify some invariant that must be enforced by the caller */
1140 VM_WARN_ON(prev && addr <= prev->vm_start);
1141 VM_WARN_ON(area && end > area->vm_end);
1142 VM_WARN_ON(addr >= end);
1145 * Can it merge with the predecessor?
1147 if (prev && prev->vm_end == addr &&
1148 mpol_equal(vma_policy(prev), policy) &&
1149 can_vma_merge_after(prev, vm_flags,
1150 anon_vma, file, pgoff,
1151 vm_userfaultfd_ctx)) {
1153 * OK, it can. Can we now merge in the successor as well?
1155 if (next && end == next->vm_start &&
1156 mpol_equal(policy, vma_policy(next)) &&
1157 can_vma_merge_before(next, vm_flags,
1160 vm_userfaultfd_ctx) &&
1161 is_mergeable_anon_vma(prev->anon_vma,
1162 next->anon_vma, NULL)) {
1164 err = __vma_adjust(prev, prev->vm_start,
1165 next->vm_end, prev->vm_pgoff, NULL,
1167 } else /* cases 2, 5, 7 */
1168 err = __vma_adjust(prev, prev->vm_start,
1169 end, prev->vm_pgoff, NULL, prev);
1172 khugepaged_enter_vma_merge(prev, vm_flags);
1177 * Can this new request be merged in front of next?
1179 if (next && end == next->vm_start &&
1180 mpol_equal(policy, vma_policy(next)) &&
1181 can_vma_merge_before(next, vm_flags,
1182 anon_vma, file, pgoff+pglen,
1183 vm_userfaultfd_ctx)) {
1184 if (prev && addr < prev->vm_end) /* case 4 */
1185 err = __vma_adjust(prev, prev->vm_start,
1186 addr, prev->vm_pgoff, NULL, next);
1187 else { /* cases 3, 8 */
1188 err = __vma_adjust(area, addr, next->vm_end,
1189 next->vm_pgoff - pglen, NULL, next);
1191 * In case 3 area is already equal to next and
1192 * this is a noop, but in case 8 "area" has
1193 * been removed and next was expanded over it.
1199 khugepaged_enter_vma_merge(area, vm_flags);
1207 * Rough compatibility check to quickly see if it's even worth looking
1208 * at sharing an anon_vma.
1210 * They need to have the same vm_file, and the flags can only differ
1211 * in things that mprotect may change.
1213 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1214 * we can merge the two vma's. For example, we refuse to merge a vma if
1215 * there is a vm_ops->close() function, because that indicates that the
1216 * driver is doing some kind of reference counting. But that doesn't
1217 * really matter for the anon_vma sharing case.
1219 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1221 return a->vm_end == b->vm_start &&
1222 mpol_equal(vma_policy(a), vma_policy(b)) &&
1223 a->vm_file == b->vm_file &&
1224 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1225 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1229 * Do some basic sanity checking to see if we can re-use the anon_vma
1230 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1231 * the same as 'old', the other will be the new one that is trying
1232 * to share the anon_vma.
1234 * NOTE! This runs with mm_sem held for reading, so it is possible that
1235 * the anon_vma of 'old' is concurrently in the process of being set up
1236 * by another page fault trying to merge _that_. But that's ok: if it
1237 * is being set up, that automatically means that it will be a singleton
1238 * acceptable for merging, so we can do all of this optimistically. But
1239 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1241 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1242 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1243 * is to return an anon_vma that is "complex" due to having gone through
1246 * We also make sure that the two vma's are compatible (adjacent,
1247 * and with the same memory policies). That's all stable, even with just
1248 * a read lock on the mm_sem.
1250 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1252 if (anon_vma_compatible(a, b)) {
1253 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1255 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1262 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1263 * neighbouring vmas for a suitable anon_vma, before it goes off
1264 * to allocate a new anon_vma. It checks because a repetitive
1265 * sequence of mprotects and faults may otherwise lead to distinct
1266 * anon_vmas being allocated, preventing vma merge in subsequent
1269 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1271 struct anon_vma *anon_vma = NULL;
1273 /* Try next first. */
1275 anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next);
1280 /* Try prev next. */
1282 anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma);
1285 * We might reach here with anon_vma == NULL if we can't find
1286 * any reusable anon_vma.
1287 * There's no absolute need to look only at touching neighbours:
1288 * we could search further afield for "compatible" anon_vmas.
1289 * But it would probably just be a waste of time searching,
1290 * or lead to too many vmas hanging off the same anon_vma.
1291 * We're trying to allow mprotect remerging later on,
1292 * not trying to minimize memory used for anon_vmas.
1298 * If a hint addr is less than mmap_min_addr change hint to be as
1299 * low as possible but still greater than mmap_min_addr
1301 static inline unsigned long round_hint_to_min(unsigned long hint)
1304 if (((void *)hint != NULL) &&
1305 (hint < mmap_min_addr))
1306 return PAGE_ALIGN(mmap_min_addr);
1310 static inline int mlock_future_check(struct mm_struct *mm,
1311 unsigned long flags,
1314 unsigned long locked, lock_limit;
1316 /* mlock MCL_FUTURE? */
1317 if (flags & VM_LOCKED) {
1318 locked = len >> PAGE_SHIFT;
1319 locked += mm->locked_vm;
1320 lock_limit = rlimit(RLIMIT_MEMLOCK);
1321 lock_limit >>= PAGE_SHIFT;
1322 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1328 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1330 if (S_ISREG(inode->i_mode))
1331 return MAX_LFS_FILESIZE;
1333 if (S_ISBLK(inode->i_mode))
1334 return MAX_LFS_FILESIZE;
1336 if (S_ISSOCK(inode->i_mode))
1337 return MAX_LFS_FILESIZE;
1339 /* Special "we do even unsigned file positions" case */
1340 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1343 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1347 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1348 unsigned long pgoff, unsigned long len)
1350 u64 maxsize = file_mmap_size_max(file, inode);
1352 if (maxsize && len > maxsize)
1355 if (pgoff > maxsize >> PAGE_SHIFT)
1361 * The caller must write-lock current->mm->mmap_lock.
1363 unsigned long do_mmap(struct file *file, unsigned long addr,
1364 unsigned long len, unsigned long prot,
1365 unsigned long flags, unsigned long pgoff,
1366 unsigned long *populate, struct list_head *uf)
1368 struct mm_struct *mm = current->mm;
1369 vm_flags_t vm_flags;
1378 * Does the application expect PROT_READ to imply PROT_EXEC?
1380 * (the exception is when the underlying filesystem is noexec
1381 * mounted, in which case we dont add PROT_EXEC.)
1383 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1384 if (!(file && path_noexec(&file->f_path)))
1387 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1388 if (flags & MAP_FIXED_NOREPLACE)
1391 if (!(flags & MAP_FIXED))
1392 addr = round_hint_to_min(addr);
1394 /* Careful about overflows.. */
1395 len = PAGE_ALIGN(len);
1399 /* offset overflow? */
1400 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1403 /* Too many mappings? */
1404 if (mm->map_count > sysctl_max_map_count)
1407 /* Obtain the address to map to. we verify (or select) it and ensure
1408 * that it represents a valid section of the address space.
1410 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1411 if (IS_ERR_VALUE(addr))
1414 if (flags & MAP_FIXED_NOREPLACE) {
1415 struct vm_area_struct *vma = find_vma(mm, addr);
1417 if (vma && vma->vm_start < addr + len)
1421 if (prot == PROT_EXEC) {
1422 pkey = execute_only_pkey(mm);
1427 /* Do simple checking here so the lower-level routines won't have
1428 * to. we assume access permissions have been handled by the open
1429 * of the memory object, so we don't do any here.
1431 vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1432 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1434 if (flags & MAP_LOCKED)
1435 if (!can_do_mlock())
1438 if (mlock_future_check(mm, vm_flags, len))
1442 struct inode *inode = file_inode(file);
1443 unsigned long flags_mask;
1445 if (!file_mmap_ok(file, inode, pgoff, len))
1448 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1450 switch (flags & MAP_TYPE) {
1453 * Force use of MAP_SHARED_VALIDATE with non-legacy
1454 * flags. E.g. MAP_SYNC is dangerous to use with
1455 * MAP_SHARED as you don't know which consistency model
1456 * you will get. We silently ignore unsupported flags
1457 * with MAP_SHARED to preserve backward compatibility.
1459 flags &= LEGACY_MAP_MASK;
1461 case MAP_SHARED_VALIDATE:
1462 if (flags & ~flags_mask)
1464 if (prot & PROT_WRITE) {
1465 if (!(file->f_mode & FMODE_WRITE))
1467 if (IS_SWAPFILE(file->f_mapping->host))
1472 * Make sure we don't allow writing to an append-only
1475 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1479 * Make sure there are no mandatory locks on the file.
1481 if (locks_verify_locked(file))
1484 vm_flags |= VM_SHARED | VM_MAYSHARE;
1485 if (!(file->f_mode & FMODE_WRITE))
1486 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1489 if (!(file->f_mode & FMODE_READ))
1491 if (path_noexec(&file->f_path)) {
1492 if (vm_flags & VM_EXEC)
1494 vm_flags &= ~VM_MAYEXEC;
1497 if (!file->f_op->mmap)
1499 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1507 switch (flags & MAP_TYPE) {
1509 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1515 vm_flags |= VM_SHARED | VM_MAYSHARE;
1519 * Set pgoff according to addr for anon_vma.
1521 pgoff = addr >> PAGE_SHIFT;
1529 * Set 'VM_NORESERVE' if we should not account for the
1530 * memory use of this mapping.
1532 if (flags & MAP_NORESERVE) {
1533 /* We honor MAP_NORESERVE if allowed to overcommit */
1534 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1535 vm_flags |= VM_NORESERVE;
1537 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1538 if (file && is_file_hugepages(file))
1539 vm_flags |= VM_NORESERVE;
1542 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1543 if (!IS_ERR_VALUE(addr) &&
1544 ((vm_flags & VM_LOCKED) ||
1545 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1550 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1551 unsigned long prot, unsigned long flags,
1552 unsigned long fd, unsigned long pgoff)
1554 struct file *file = NULL;
1555 unsigned long retval;
1557 if (!(flags & MAP_ANONYMOUS)) {
1558 audit_mmap_fd(fd, flags);
1562 if (is_file_hugepages(file)) {
1563 len = ALIGN(len, huge_page_size(hstate_file(file)));
1564 } else if (unlikely(flags & MAP_HUGETLB)) {
1568 } else if (flags & MAP_HUGETLB) {
1569 struct user_struct *user = NULL;
1572 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1576 len = ALIGN(len, huge_page_size(hs));
1578 * VM_NORESERVE is used because the reservations will be
1579 * taken when vm_ops->mmap() is called
1580 * A dummy user value is used because we are not locking
1581 * memory so no accounting is necessary
1583 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1585 &user, HUGETLB_ANONHUGE_INODE,
1586 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1588 return PTR_ERR(file);
1591 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1593 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1600 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1601 unsigned long, prot, unsigned long, flags,
1602 unsigned long, fd, unsigned long, pgoff)
1604 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1607 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1608 struct mmap_arg_struct {
1612 unsigned long flags;
1614 unsigned long offset;
1617 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1619 struct mmap_arg_struct a;
1621 if (copy_from_user(&a, arg, sizeof(a)))
1623 if (offset_in_page(a.offset))
1626 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1627 a.offset >> PAGE_SHIFT);
1629 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1632 * Some shared mappings will want the pages marked read-only
1633 * to track write events. If so, we'll downgrade vm_page_prot
1634 * to the private version (using protection_map[] without the
1637 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1639 vm_flags_t vm_flags = vma->vm_flags;
1640 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1642 /* If it was private or non-writable, the write bit is already clear */
1643 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1646 /* The backer wishes to know when pages are first written to? */
1647 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1650 /* The open routine did something to the protections that pgprot_modify
1651 * won't preserve? */
1652 if (pgprot_val(vm_page_prot) !=
1653 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1656 /* Do we need to track softdirty? */
1657 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1660 /* Specialty mapping? */
1661 if (vm_flags & VM_PFNMAP)
1664 /* Can the mapping track the dirty pages? */
1665 return vma->vm_file && vma->vm_file->f_mapping &&
1666 mapping_can_writeback(vma->vm_file->f_mapping);
1670 * We account for memory if it's a private writeable mapping,
1671 * not hugepages and VM_NORESERVE wasn't set.
1673 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1676 * hugetlb has its own accounting separate from the core VM
1677 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1679 if (file && is_file_hugepages(file))
1682 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1685 unsigned long mmap_region(struct file *file, unsigned long addr,
1686 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1687 struct list_head *uf)
1689 struct mm_struct *mm = current->mm;
1690 struct vm_area_struct *vma, *prev, *merge;
1692 struct rb_node **rb_link, *rb_parent;
1693 unsigned long charged = 0;
1695 /* Check against address space limit. */
1696 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1697 unsigned long nr_pages;
1700 * MAP_FIXED may remove pages of mappings that intersects with
1701 * requested mapping. Account for the pages it would unmap.
1703 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1705 if (!may_expand_vm(mm, vm_flags,
1706 (len >> PAGE_SHIFT) - nr_pages))
1710 /* Clear old maps */
1711 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1713 if (do_munmap(mm, addr, len, uf))
1718 * Private writable mapping: check memory availability
1720 if (accountable_mapping(file, vm_flags)) {
1721 charged = len >> PAGE_SHIFT;
1722 if (security_vm_enough_memory_mm(mm, charged))
1724 vm_flags |= VM_ACCOUNT;
1728 * Can we just expand an old mapping?
1730 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1731 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1736 * Determine the object being mapped and call the appropriate
1737 * specific mapper. the address has already been validated, but
1738 * not unmapped, but the maps are removed from the list.
1740 vma = vm_area_alloc(mm);
1746 vma->vm_start = addr;
1747 vma->vm_end = addr + len;
1748 vma->vm_flags = vm_flags;
1749 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1750 vma->vm_pgoff = pgoff;
1753 if (vm_flags & VM_DENYWRITE) {
1754 error = deny_write_access(file);
1758 if (vm_flags & VM_SHARED) {
1759 error = mapping_map_writable(file->f_mapping);
1761 goto allow_write_and_free_vma;
1764 /* ->mmap() can change vma->vm_file, but must guarantee that
1765 * vma_link() below can deny write-access if VM_DENYWRITE is set
1766 * and map writably if VM_SHARED is set. This usually means the
1767 * new file must not have been exposed to user-space, yet.
1769 vma->vm_file = get_file(file);
1770 error = call_mmap(file, vma);
1772 goto unmap_and_free_vma;
1774 /* If vm_flags changed after call_mmap(), we should try merge vma again
1775 * as we may succeed this time.
1777 if (unlikely(vm_flags != vma->vm_flags && prev)) {
1778 merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
1779 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX);
1781 /* ->mmap() can change vma->vm_file and fput the original file. So
1782 * fput the vma->vm_file here or we would add an extra fput for file
1783 * and cause general protection fault ultimately.
1788 /* Update vm_flags and possible addr to pick up the change. We don't
1789 * warn here if addr changed as the vma is not linked by vma_link().
1791 addr = vma->vm_start;
1792 vm_flags = vma->vm_flags;
1793 goto unmap_writable;
1797 /* Can addr have changed??
1799 * Answer: Yes, several device drivers can do it in their
1800 * f_op->mmap method. -DaveM
1801 * Bug: If addr is changed, prev, rb_link, rb_parent should
1802 * be updated for vma_link()
1804 WARN_ON_ONCE(addr != vma->vm_start);
1806 addr = vma->vm_start;
1807 vm_flags = vma->vm_flags;
1808 } else if (vm_flags & VM_SHARED) {
1809 error = shmem_zero_setup(vma);
1813 vma_set_anonymous(vma);
1816 /* Allow architectures to sanity-check the vm_flags */
1817 if (!arch_validate_flags(vma->vm_flags)) {
1820 goto unmap_and_free_vma;
1825 vma_link(mm, vma, prev, rb_link, rb_parent);
1826 /* Once vma denies write, undo our temporary denial count */
1829 if (vm_flags & VM_SHARED)
1830 mapping_unmap_writable(file->f_mapping);
1831 if (vm_flags & VM_DENYWRITE)
1832 allow_write_access(file);
1834 file = vma->vm_file;
1836 perf_event_mmap(vma);
1838 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1839 if (vm_flags & VM_LOCKED) {
1840 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1841 is_vm_hugetlb_page(vma) ||
1842 vma == get_gate_vma(current->mm))
1843 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1845 mm->locked_vm += (len >> PAGE_SHIFT);
1852 * New (or expanded) vma always get soft dirty status.
1853 * Otherwise user-space soft-dirty page tracker won't
1854 * be able to distinguish situation when vma area unmapped,
1855 * then new mapped in-place (which must be aimed as
1856 * a completely new data area).
1858 vma->vm_flags |= VM_SOFTDIRTY;
1860 vma_set_page_prot(vma);
1865 vma->vm_file = NULL;
1868 /* Undo any partial mapping done by a device driver. */
1869 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1871 if (vm_flags & VM_SHARED)
1872 mapping_unmap_writable(file->f_mapping);
1873 allow_write_and_free_vma:
1874 if (vm_flags & VM_DENYWRITE)
1875 allow_write_access(file);
1880 vm_unacct_memory(charged);
1884 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1887 * We implement the search by looking for an rbtree node that
1888 * immediately follows a suitable gap. That is,
1889 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1890 * - gap_end = vma->vm_start >= info->low_limit + length;
1891 * - gap_end - gap_start >= length
1894 struct mm_struct *mm = current->mm;
1895 struct vm_area_struct *vma;
1896 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1898 /* Adjust search length to account for worst case alignment overhead */
1899 length = info->length + info->align_mask;
1900 if (length < info->length)
1903 /* Adjust search limits by the desired length */
1904 if (info->high_limit < length)
1906 high_limit = info->high_limit - length;
1908 if (info->low_limit > high_limit)
1910 low_limit = info->low_limit + length;
1912 /* Check if rbtree root looks promising */
1913 if (RB_EMPTY_ROOT(&mm->mm_rb))
1915 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1916 if (vma->rb_subtree_gap < length)
1920 /* Visit left subtree if it looks promising */
1921 gap_end = vm_start_gap(vma);
1922 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1923 struct vm_area_struct *left =
1924 rb_entry(vma->vm_rb.rb_left,
1925 struct vm_area_struct, vm_rb);
1926 if (left->rb_subtree_gap >= length) {
1932 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1934 /* Check if current node has a suitable gap */
1935 if (gap_start > high_limit)
1937 if (gap_end >= low_limit &&
1938 gap_end > gap_start && gap_end - gap_start >= length)
1941 /* Visit right subtree if it looks promising */
1942 if (vma->vm_rb.rb_right) {
1943 struct vm_area_struct *right =
1944 rb_entry(vma->vm_rb.rb_right,
1945 struct vm_area_struct, vm_rb);
1946 if (right->rb_subtree_gap >= length) {
1952 /* Go back up the rbtree to find next candidate node */
1954 struct rb_node *prev = &vma->vm_rb;
1955 if (!rb_parent(prev))
1957 vma = rb_entry(rb_parent(prev),
1958 struct vm_area_struct, vm_rb);
1959 if (prev == vma->vm_rb.rb_left) {
1960 gap_start = vm_end_gap(vma->vm_prev);
1961 gap_end = vm_start_gap(vma);
1968 /* Check highest gap, which does not precede any rbtree node */
1969 gap_start = mm->highest_vm_end;
1970 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1971 if (gap_start > high_limit)
1975 /* We found a suitable gap. Clip it with the original low_limit. */
1976 if (gap_start < info->low_limit)
1977 gap_start = info->low_limit;
1979 /* Adjust gap address to the desired alignment */
1980 gap_start += (info->align_offset - gap_start) & info->align_mask;
1982 VM_BUG_ON(gap_start + info->length > info->high_limit);
1983 VM_BUG_ON(gap_start + info->length > gap_end);
1987 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1989 struct mm_struct *mm = current->mm;
1990 struct vm_area_struct *vma;
1991 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1993 /* Adjust search length to account for worst case alignment overhead */
1994 length = info->length + info->align_mask;
1995 if (length < info->length)
1999 * Adjust search limits by the desired length.
2000 * See implementation comment at top of unmapped_area().
2002 gap_end = info->high_limit;
2003 if (gap_end < length)
2005 high_limit = gap_end - length;
2007 if (info->low_limit > high_limit)
2009 low_limit = info->low_limit + length;
2011 /* Check highest gap, which does not precede any rbtree node */
2012 gap_start = mm->highest_vm_end;
2013 if (gap_start <= high_limit)
2016 /* Check if rbtree root looks promising */
2017 if (RB_EMPTY_ROOT(&mm->mm_rb))
2019 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
2020 if (vma->rb_subtree_gap < length)
2024 /* Visit right subtree if it looks promising */
2025 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
2026 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
2027 struct vm_area_struct *right =
2028 rb_entry(vma->vm_rb.rb_right,
2029 struct vm_area_struct, vm_rb);
2030 if (right->rb_subtree_gap >= length) {
2037 /* Check if current node has a suitable gap */
2038 gap_end = vm_start_gap(vma);
2039 if (gap_end < low_limit)
2041 if (gap_start <= high_limit &&
2042 gap_end > gap_start && gap_end - gap_start >= length)
2045 /* Visit left subtree if it looks promising */
2046 if (vma->vm_rb.rb_left) {
2047 struct vm_area_struct *left =
2048 rb_entry(vma->vm_rb.rb_left,
2049 struct vm_area_struct, vm_rb);
2050 if (left->rb_subtree_gap >= length) {
2056 /* Go back up the rbtree to find next candidate node */
2058 struct rb_node *prev = &vma->vm_rb;
2059 if (!rb_parent(prev))
2061 vma = rb_entry(rb_parent(prev),
2062 struct vm_area_struct, vm_rb);
2063 if (prev == vma->vm_rb.rb_right) {
2064 gap_start = vma->vm_prev ?
2065 vm_end_gap(vma->vm_prev) : 0;
2072 /* We found a suitable gap. Clip it with the original high_limit. */
2073 if (gap_end > info->high_limit)
2074 gap_end = info->high_limit;
2077 /* Compute highest gap address at the desired alignment */
2078 gap_end -= info->length;
2079 gap_end -= (gap_end - info->align_offset) & info->align_mask;
2081 VM_BUG_ON(gap_end < info->low_limit);
2082 VM_BUG_ON(gap_end < gap_start);
2087 * Search for an unmapped address range.
2089 * We are looking for a range that:
2090 * - does not intersect with any VMA;
2091 * - is contained within the [low_limit, high_limit) interval;
2092 * - is at least the desired size.
2093 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
2095 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
2099 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
2100 addr = unmapped_area_topdown(info);
2102 addr = unmapped_area(info);
2104 trace_vm_unmapped_area(addr, info);
2108 #ifndef arch_get_mmap_end
2109 #define arch_get_mmap_end(addr) (TASK_SIZE)
2112 #ifndef arch_get_mmap_base
2113 #define arch_get_mmap_base(addr, base) (base)
2116 /* Get an address range which is currently unmapped.
2117 * For shmat() with addr=0.
2119 * Ugly calling convention alert:
2120 * Return value with the low bits set means error value,
2122 * if (ret & ~PAGE_MASK)
2125 * This function "knows" that -ENOMEM has the bits set.
2127 #ifndef HAVE_ARCH_UNMAPPED_AREA
2129 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2130 unsigned long len, unsigned long pgoff, unsigned long flags)
2132 struct mm_struct *mm = current->mm;
2133 struct vm_area_struct *vma, *prev;
2134 struct vm_unmapped_area_info info;
2135 const unsigned long mmap_end = arch_get_mmap_end(addr);
2137 if (len > mmap_end - mmap_min_addr)
2140 if (flags & MAP_FIXED)
2144 addr = PAGE_ALIGN(addr);
2145 vma = find_vma_prev(mm, addr, &prev);
2146 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2147 (!vma || addr + len <= vm_start_gap(vma)) &&
2148 (!prev || addr >= vm_end_gap(prev)))
2154 info.low_limit = mm->mmap_base;
2155 info.high_limit = mmap_end;
2156 info.align_mask = 0;
2157 info.align_offset = 0;
2158 return vm_unmapped_area(&info);
2163 * This mmap-allocator allocates new areas top-down from below the
2164 * stack's low limit (the base):
2166 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2168 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2169 unsigned long len, unsigned long pgoff,
2170 unsigned long flags)
2172 struct vm_area_struct *vma, *prev;
2173 struct mm_struct *mm = current->mm;
2174 struct vm_unmapped_area_info info;
2175 const unsigned long mmap_end = arch_get_mmap_end(addr);
2177 /* requested length too big for entire address space */
2178 if (len > mmap_end - mmap_min_addr)
2181 if (flags & MAP_FIXED)
2184 /* requesting a specific address */
2186 addr = PAGE_ALIGN(addr);
2187 vma = find_vma_prev(mm, addr, &prev);
2188 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2189 (!vma || addr + len <= vm_start_gap(vma)) &&
2190 (!prev || addr >= vm_end_gap(prev)))
2194 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2196 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2197 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2198 info.align_mask = 0;
2199 info.align_offset = 0;
2200 addr = vm_unmapped_area(&info);
2203 * A failed mmap() very likely causes application failure,
2204 * so fall back to the bottom-up function here. This scenario
2205 * can happen with large stack limits and large mmap()
2208 if (offset_in_page(addr)) {
2209 VM_BUG_ON(addr != -ENOMEM);
2211 info.low_limit = TASK_UNMAPPED_BASE;
2212 info.high_limit = mmap_end;
2213 addr = vm_unmapped_area(&info);
2221 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2222 unsigned long pgoff, unsigned long flags)
2224 unsigned long (*get_area)(struct file *, unsigned long,
2225 unsigned long, unsigned long, unsigned long);
2227 unsigned long error = arch_mmap_check(addr, len, flags);
2231 /* Careful about overflows.. */
2232 if (len > TASK_SIZE)
2235 get_area = current->mm->get_unmapped_area;
2237 if (file->f_op->get_unmapped_area)
2238 get_area = file->f_op->get_unmapped_area;
2239 } else if (flags & MAP_SHARED) {
2241 * mmap_region() will call shmem_zero_setup() to create a file,
2242 * so use shmem's get_unmapped_area in case it can be huge.
2243 * do_mmap() will clear pgoff, so match alignment.
2246 get_area = shmem_get_unmapped_area;
2249 addr = get_area(file, addr, len, pgoff, flags);
2250 if (IS_ERR_VALUE(addr))
2253 if (addr > TASK_SIZE - len)
2255 if (offset_in_page(addr))
2258 error = security_mmap_addr(addr);
2259 return error ? error : addr;
2262 EXPORT_SYMBOL(get_unmapped_area);
2264 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2265 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2267 struct rb_node *rb_node;
2268 struct vm_area_struct *vma;
2270 /* Check the cache first. */
2271 vma = vmacache_find(mm, addr);
2275 rb_node = mm->mm_rb.rb_node;
2278 struct vm_area_struct *tmp;
2280 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2282 if (tmp->vm_end > addr) {
2284 if (tmp->vm_start <= addr)
2286 rb_node = rb_node->rb_left;
2288 rb_node = rb_node->rb_right;
2292 vmacache_update(addr, vma);
2296 EXPORT_SYMBOL(find_vma);
2299 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2301 struct vm_area_struct *
2302 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2303 struct vm_area_struct **pprev)
2305 struct vm_area_struct *vma;
2307 vma = find_vma(mm, addr);
2309 *pprev = vma->vm_prev;
2311 struct rb_node *rb_node = rb_last(&mm->mm_rb);
2313 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
2319 * Verify that the stack growth is acceptable and
2320 * update accounting. This is shared with both the
2321 * grow-up and grow-down cases.
2323 static int acct_stack_growth(struct vm_area_struct *vma,
2324 unsigned long size, unsigned long grow)
2326 struct mm_struct *mm = vma->vm_mm;
2327 unsigned long new_start;
2329 /* address space limit tests */
2330 if (!may_expand_vm(mm, vma->vm_flags, grow))
2333 /* Stack limit test */
2334 if (size > rlimit(RLIMIT_STACK))
2337 /* mlock limit tests */
2338 if (vma->vm_flags & VM_LOCKED) {
2339 unsigned long locked;
2340 unsigned long limit;
2341 locked = mm->locked_vm + grow;
2342 limit = rlimit(RLIMIT_MEMLOCK);
2343 limit >>= PAGE_SHIFT;
2344 if (locked > limit && !capable(CAP_IPC_LOCK))
2348 /* Check to ensure the stack will not grow into a hugetlb-only region */
2349 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2351 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2355 * Overcommit.. This must be the final test, as it will
2356 * update security statistics.
2358 if (security_vm_enough_memory_mm(mm, grow))
2364 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2366 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2367 * vma is the last one with address > vma->vm_end. Have to extend vma.
2369 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2371 struct mm_struct *mm = vma->vm_mm;
2372 struct vm_area_struct *next;
2373 unsigned long gap_addr;
2376 if (!(vma->vm_flags & VM_GROWSUP))
2379 /* Guard against exceeding limits of the address space. */
2380 address &= PAGE_MASK;
2381 if (address >= (TASK_SIZE & PAGE_MASK))
2383 address += PAGE_SIZE;
2385 /* Enforce stack_guard_gap */
2386 gap_addr = address + stack_guard_gap;
2388 /* Guard against overflow */
2389 if (gap_addr < address || gap_addr > TASK_SIZE)
2390 gap_addr = TASK_SIZE;
2392 next = vma->vm_next;
2393 if (next && next->vm_start < gap_addr && vma_is_accessible(next)) {
2394 if (!(next->vm_flags & VM_GROWSUP))
2396 /* Check that both stack segments have the same anon_vma? */
2399 /* We must make sure the anon_vma is allocated. */
2400 if (unlikely(anon_vma_prepare(vma)))
2404 * vma->vm_start/vm_end cannot change under us because the caller
2405 * is required to hold the mmap_lock in read mode. We need the
2406 * anon_vma lock to serialize against concurrent expand_stacks.
2408 anon_vma_lock_write(vma->anon_vma);
2410 /* Somebody else might have raced and expanded it already */
2411 if (address > vma->vm_end) {
2412 unsigned long size, grow;
2414 size = address - vma->vm_start;
2415 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2418 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2419 error = acct_stack_growth(vma, size, grow);
2422 * vma_gap_update() doesn't support concurrent
2423 * updates, but we only hold a shared mmap_lock
2424 * lock here, so we need to protect against
2425 * concurrent vma expansions.
2426 * anon_vma_lock_write() doesn't help here, as
2427 * we don't guarantee that all growable vmas
2428 * in a mm share the same root anon vma.
2429 * So, we reuse mm->page_table_lock to guard
2430 * against concurrent vma expansions.
2432 spin_lock(&mm->page_table_lock);
2433 if (vma->vm_flags & VM_LOCKED)
2434 mm->locked_vm += grow;
2435 vm_stat_account(mm, vma->vm_flags, grow);
2436 anon_vma_interval_tree_pre_update_vma(vma);
2437 vma->vm_end = address;
2438 anon_vma_interval_tree_post_update_vma(vma);
2440 vma_gap_update(vma->vm_next);
2442 mm->highest_vm_end = vm_end_gap(vma);
2443 spin_unlock(&mm->page_table_lock);
2445 perf_event_mmap(vma);
2449 anon_vma_unlock_write(vma->anon_vma);
2450 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2454 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2457 * vma is the first one with address < vma->vm_start. Have to extend vma.
2459 int expand_downwards(struct vm_area_struct *vma,
2460 unsigned long address)
2462 struct mm_struct *mm = vma->vm_mm;
2463 struct vm_area_struct *prev;
2466 address &= PAGE_MASK;
2467 if (address < mmap_min_addr)
2470 /* Enforce stack_guard_gap */
2471 prev = vma->vm_prev;
2472 /* Check that both stack segments have the same anon_vma? */
2473 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2474 vma_is_accessible(prev)) {
2475 if (address - prev->vm_end < stack_guard_gap)
2479 /* We must make sure the anon_vma is allocated. */
2480 if (unlikely(anon_vma_prepare(vma)))
2484 * vma->vm_start/vm_end cannot change under us because the caller
2485 * is required to hold the mmap_lock in read mode. We need the
2486 * anon_vma lock to serialize against concurrent expand_stacks.
2488 anon_vma_lock_write(vma->anon_vma);
2490 /* Somebody else might have raced and expanded it already */
2491 if (address < vma->vm_start) {
2492 unsigned long size, grow;
2494 size = vma->vm_end - address;
2495 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2498 if (grow <= vma->vm_pgoff) {
2499 error = acct_stack_growth(vma, size, grow);
2502 * vma_gap_update() doesn't support concurrent
2503 * updates, but we only hold a shared mmap_lock
2504 * lock here, so we need to protect against
2505 * concurrent vma expansions.
2506 * anon_vma_lock_write() doesn't help here, as
2507 * we don't guarantee that all growable vmas
2508 * in a mm share the same root anon vma.
2509 * So, we reuse mm->page_table_lock to guard
2510 * against concurrent vma expansions.
2512 spin_lock(&mm->page_table_lock);
2513 if (vma->vm_flags & VM_LOCKED)
2514 mm->locked_vm += grow;
2515 vm_stat_account(mm, vma->vm_flags, grow);
2516 anon_vma_interval_tree_pre_update_vma(vma);
2517 vma->vm_start = address;
2518 vma->vm_pgoff -= grow;
2519 anon_vma_interval_tree_post_update_vma(vma);
2520 vma_gap_update(vma);
2521 spin_unlock(&mm->page_table_lock);
2523 perf_event_mmap(vma);
2527 anon_vma_unlock_write(vma->anon_vma);
2528 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2533 /* enforced gap between the expanding stack and other mappings. */
2534 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2536 static int __init cmdline_parse_stack_guard_gap(char *p)
2541 val = simple_strtoul(p, &endptr, 10);
2543 stack_guard_gap = val << PAGE_SHIFT;
2547 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2549 #ifdef CONFIG_STACK_GROWSUP
2550 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2552 return expand_upwards(vma, address);
2555 struct vm_area_struct *
2556 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2558 struct vm_area_struct *vma, *prev;
2561 vma = find_vma_prev(mm, addr, &prev);
2562 if (vma && (vma->vm_start <= addr))
2564 /* don't alter vm_end if the coredump is running */
2565 if (!prev || expand_stack(prev, addr))
2567 if (prev->vm_flags & VM_LOCKED)
2568 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2572 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2574 return expand_downwards(vma, address);
2577 struct vm_area_struct *
2578 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2580 struct vm_area_struct *vma;
2581 unsigned long start;
2584 vma = find_vma(mm, addr);
2587 if (vma->vm_start <= addr)
2589 if (!(vma->vm_flags & VM_GROWSDOWN))
2591 start = vma->vm_start;
2592 if (expand_stack(vma, addr))
2594 if (vma->vm_flags & VM_LOCKED)
2595 populate_vma_page_range(vma, addr, start, NULL);
2600 EXPORT_SYMBOL_GPL(find_extend_vma);
2603 * Ok - we have the memory areas we should free on the vma list,
2604 * so release them, and do the vma updates.
2606 * Called with the mm semaphore held.
2608 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2610 unsigned long nr_accounted = 0;
2612 /* Update high watermark before we lower total_vm */
2613 update_hiwater_vm(mm);
2615 long nrpages = vma_pages(vma);
2617 if (vma->vm_flags & VM_ACCOUNT)
2618 nr_accounted += nrpages;
2619 vm_stat_account(mm, vma->vm_flags, -nrpages);
2620 vma = remove_vma(vma);
2622 vm_unacct_memory(nr_accounted);
2627 * Get rid of page table information in the indicated region.
2629 * Called with the mm semaphore held.
2631 static void unmap_region(struct mm_struct *mm,
2632 struct vm_area_struct *vma, struct vm_area_struct *prev,
2633 unsigned long start, unsigned long end)
2635 struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2636 struct mmu_gather tlb;
2639 tlb_gather_mmu(&tlb, mm, start, end);
2640 update_hiwater_rss(mm);
2641 unmap_vmas(&tlb, vma, start, end);
2642 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2643 next ? next->vm_start : USER_PGTABLES_CEILING);
2644 tlb_finish_mmu(&tlb, start, end);
2648 * Create a list of vma's touched by the unmap, removing them from the mm's
2649 * vma list as we go..
2652 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2653 struct vm_area_struct *prev, unsigned long end)
2655 struct vm_area_struct **insertion_point;
2656 struct vm_area_struct *tail_vma = NULL;
2658 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2659 vma->vm_prev = NULL;
2661 vma_rb_erase(vma, &mm->mm_rb);
2665 } while (vma && vma->vm_start < end);
2666 *insertion_point = vma;
2668 vma->vm_prev = prev;
2669 vma_gap_update(vma);
2671 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2672 tail_vma->vm_next = NULL;
2674 /* Kill the cache */
2675 vmacache_invalidate(mm);
2678 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2679 * VM_GROWSUP VMA. Such VMAs can change their size under
2680 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2682 if (vma && (vma->vm_flags & VM_GROWSDOWN))
2684 if (prev && (prev->vm_flags & VM_GROWSUP))
2690 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2691 * has already been checked or doesn't make sense to fail.
2693 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2694 unsigned long addr, int new_below)
2696 struct vm_area_struct *new;
2699 if (vma->vm_ops && vma->vm_ops->split) {
2700 err = vma->vm_ops->split(vma, addr);
2705 new = vm_area_dup(vma);
2712 new->vm_start = addr;
2713 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2716 err = vma_dup_policy(vma, new);
2720 err = anon_vma_clone(new, vma);
2725 get_file(new->vm_file);
2727 if (new->vm_ops && new->vm_ops->open)
2728 new->vm_ops->open(new);
2731 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2732 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2734 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2740 /* Clean everything up if vma_adjust failed. */
2741 if (new->vm_ops && new->vm_ops->close)
2742 new->vm_ops->close(new);
2745 unlink_anon_vmas(new);
2747 mpol_put(vma_policy(new));
2754 * Split a vma into two pieces at address 'addr', a new vma is allocated
2755 * either for the first part or the tail.
2757 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2758 unsigned long addr, int new_below)
2760 if (mm->map_count >= sysctl_max_map_count)
2763 return __split_vma(mm, vma, addr, new_below);
2766 /* Munmap is split into 2 main parts -- this part which finds
2767 * what needs doing, and the areas themselves, which do the
2768 * work. This now handles partial unmappings.
2769 * Jeremy Fitzhardinge <jeremy@goop.org>
2771 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2772 struct list_head *uf, bool downgrade)
2775 struct vm_area_struct *vma, *prev, *last;
2777 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2780 len = PAGE_ALIGN(len);
2786 * arch_unmap() might do unmaps itself. It must be called
2787 * and finish any rbtree manipulation before this code
2788 * runs and also starts to manipulate the rbtree.
2790 arch_unmap(mm, start, end);
2792 /* Find the first overlapping VMA */
2793 vma = find_vma(mm, start);
2796 prev = vma->vm_prev;
2797 /* we have start < vma->vm_end */
2799 /* if it doesn't overlap, we have nothing.. */
2800 if (vma->vm_start >= end)
2804 * If we need to split any vma, do it now to save pain later.
2806 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2807 * unmapped vm_area_struct will remain in use: so lower split_vma
2808 * places tmp vma above, and higher split_vma places tmp vma below.
2810 if (start > vma->vm_start) {
2814 * Make sure that map_count on return from munmap() will
2815 * not exceed its limit; but let map_count go just above
2816 * its limit temporarily, to help free resources as expected.
2818 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2821 error = __split_vma(mm, vma, start, 0);
2827 /* Does it split the last one? */
2828 last = find_vma(mm, end);
2829 if (last && end > last->vm_start) {
2830 int error = __split_vma(mm, last, end, 1);
2834 vma = prev ? prev->vm_next : mm->mmap;
2838 * If userfaultfd_unmap_prep returns an error the vmas
2839 * will remain splitted, but userland will get a
2840 * highly unexpected error anyway. This is no
2841 * different than the case where the first of the two
2842 * __split_vma fails, but we don't undo the first
2843 * split, despite we could. This is unlikely enough
2844 * failure that it's not worth optimizing it for.
2846 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2852 * unlock any mlock()ed ranges before detaching vmas
2854 if (mm->locked_vm) {
2855 struct vm_area_struct *tmp = vma;
2856 while (tmp && tmp->vm_start < end) {
2857 if (tmp->vm_flags & VM_LOCKED) {
2858 mm->locked_vm -= vma_pages(tmp);
2859 munlock_vma_pages_all(tmp);
2866 /* Detach vmas from rbtree */
2867 if (!detach_vmas_to_be_unmapped(mm, vma, prev, end))
2871 mmap_write_downgrade(mm);
2873 unmap_region(mm, vma, prev, start, end);
2875 /* Fix up all other VM information */
2876 remove_vma_list(mm, vma);
2878 return downgrade ? 1 : 0;
2881 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2882 struct list_head *uf)
2884 return __do_munmap(mm, start, len, uf, false);
2887 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2890 struct mm_struct *mm = current->mm;
2893 if (mmap_write_lock_killable(mm))
2896 ret = __do_munmap(mm, start, len, &uf, downgrade);
2898 * Returning 1 indicates mmap_lock is downgraded.
2899 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2900 * it to 0 before return.
2903 mmap_read_unlock(mm);
2906 mmap_write_unlock(mm);
2908 userfaultfd_unmap_complete(mm, &uf);
2912 int vm_munmap(unsigned long start, size_t len)
2914 return __vm_munmap(start, len, false);
2916 EXPORT_SYMBOL(vm_munmap);
2918 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2920 addr = untagged_addr(addr);
2921 profile_munmap(addr);
2922 return __vm_munmap(addr, len, true);
2927 * Emulation of deprecated remap_file_pages() syscall.
2929 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2930 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2933 struct mm_struct *mm = current->mm;
2934 struct vm_area_struct *vma;
2935 unsigned long populate = 0;
2936 unsigned long ret = -EINVAL;
2939 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2940 current->comm, current->pid);
2944 start = start & PAGE_MASK;
2945 size = size & PAGE_MASK;
2947 if (start + size <= start)
2950 /* Does pgoff wrap? */
2951 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2954 if (mmap_write_lock_killable(mm))
2957 vma = find_vma(mm, start);
2959 if (!vma || !(vma->vm_flags & VM_SHARED))
2962 if (start < vma->vm_start)
2965 if (start + size > vma->vm_end) {
2966 struct vm_area_struct *next;
2968 for (next = vma->vm_next; next; next = next->vm_next) {
2969 /* hole between vmas ? */
2970 if (next->vm_start != next->vm_prev->vm_end)
2973 if (next->vm_file != vma->vm_file)
2976 if (next->vm_flags != vma->vm_flags)
2979 if (start + size <= next->vm_end)
2987 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2988 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2989 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2991 flags &= MAP_NONBLOCK;
2992 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2993 if (vma->vm_flags & VM_LOCKED) {
2994 struct vm_area_struct *tmp;
2995 flags |= MAP_LOCKED;
2997 /* drop PG_Mlocked flag for over-mapped range */
2998 for (tmp = vma; tmp->vm_start >= start + size;
2999 tmp = tmp->vm_next) {
3001 * Split pmd and munlock page on the border
3004 vma_adjust_trans_huge(tmp, start, start + size, 0);
3006 munlock_vma_pages_range(tmp,
3007 max(tmp->vm_start, start),
3008 min(tmp->vm_end, start + size));
3012 file = get_file(vma->vm_file);
3013 ret = do_mmap(vma->vm_file, start, size,
3014 prot, flags, pgoff, &populate, NULL);
3017 mmap_write_unlock(mm);
3019 mm_populate(ret, populate);
3020 if (!IS_ERR_VALUE(ret))
3026 * this is really a simplified "do_mmap". it only handles
3027 * anonymous maps. eventually we may be able to do some
3028 * brk-specific accounting here.
3030 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
3032 struct mm_struct *mm = current->mm;
3033 struct vm_area_struct *vma, *prev;
3034 struct rb_node **rb_link, *rb_parent;
3035 pgoff_t pgoff = addr >> PAGE_SHIFT;
3037 unsigned long mapped_addr;
3039 /* Until we need other flags, refuse anything except VM_EXEC. */
3040 if ((flags & (~VM_EXEC)) != 0)
3042 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3044 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
3045 if (IS_ERR_VALUE(mapped_addr))
3048 error = mlock_future_check(mm, mm->def_flags, len);
3053 * Clear old maps. this also does some error checking for us
3055 while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
3057 if (do_munmap(mm, addr, len, uf))
3061 /* Check against address space limits *after* clearing old maps... */
3062 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3065 if (mm->map_count > sysctl_max_map_count)
3068 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3071 /* Can we just expand an old private anonymous mapping? */
3072 vma = vma_merge(mm, prev, addr, addr + len, flags,
3073 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
3078 * create a vma struct for an anonymous mapping
3080 vma = vm_area_alloc(mm);
3082 vm_unacct_memory(len >> PAGE_SHIFT);
3086 vma_set_anonymous(vma);
3087 vma->vm_start = addr;
3088 vma->vm_end = addr + len;
3089 vma->vm_pgoff = pgoff;
3090 vma->vm_flags = flags;
3091 vma->vm_page_prot = vm_get_page_prot(flags);
3092 vma_link(mm, vma, prev, rb_link, rb_parent);
3094 perf_event_mmap(vma);
3095 mm->total_vm += len >> PAGE_SHIFT;
3096 mm->data_vm += len >> PAGE_SHIFT;
3097 if (flags & VM_LOCKED)
3098 mm->locked_vm += (len >> PAGE_SHIFT);
3099 vma->vm_flags |= VM_SOFTDIRTY;
3103 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3105 struct mm_struct *mm = current->mm;
3111 len = PAGE_ALIGN(request);
3117 if (mmap_write_lock_killable(mm))
3120 ret = do_brk_flags(addr, len, flags, &uf);
3121 populate = ((mm->def_flags & VM_LOCKED) != 0);
3122 mmap_write_unlock(mm);
3123 userfaultfd_unmap_complete(mm, &uf);
3124 if (populate && !ret)
3125 mm_populate(addr, len);
3128 EXPORT_SYMBOL(vm_brk_flags);
3130 int vm_brk(unsigned long addr, unsigned long len)
3132 return vm_brk_flags(addr, len, 0);
3134 EXPORT_SYMBOL(vm_brk);
3136 /* Release all mmaps. */
3137 void exit_mmap(struct mm_struct *mm)
3139 struct mmu_gather tlb;
3140 struct vm_area_struct *vma;
3141 unsigned long nr_accounted = 0;
3143 /* mm's last user has gone, and its about to be pulled down */
3144 mmu_notifier_release(mm);
3146 if (unlikely(mm_is_oom_victim(mm))) {
3148 * Manually reap the mm to free as much memory as possible.
3149 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3150 * this mm from further consideration. Taking mm->mmap_lock for
3151 * write after setting MMF_OOM_SKIP will guarantee that the oom
3152 * reaper will not run on this mm again after mmap_lock is
3155 * Nothing can be holding mm->mmap_lock here and the above call
3156 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3157 * __oom_reap_task_mm() will not block.
3159 * This needs to be done before calling munlock_vma_pages_all(),
3160 * which clears VM_LOCKED, otherwise the oom reaper cannot
3163 (void)__oom_reap_task_mm(mm);
3165 set_bit(MMF_OOM_SKIP, &mm->flags);
3166 mmap_write_lock(mm);
3167 mmap_write_unlock(mm);
3170 if (mm->locked_vm) {
3173 if (vma->vm_flags & VM_LOCKED)
3174 munlock_vma_pages_all(vma);
3182 if (!vma) /* Can happen if dup_mmap() received an OOM */
3187 tlb_gather_mmu(&tlb, mm, 0, -1);
3188 /* update_hiwater_rss(mm) here? but nobody should be looking */
3189 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3190 unmap_vmas(&tlb, vma, 0, -1);
3191 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3192 tlb_finish_mmu(&tlb, 0, -1);
3195 * Walk the list again, actually closing and freeing it,
3196 * with preemption enabled, without holding any MM locks.
3199 if (vma->vm_flags & VM_ACCOUNT)
3200 nr_accounted += vma_pages(vma);
3201 vma = remove_vma(vma);
3204 vm_unacct_memory(nr_accounted);
3207 /* Insert vm structure into process list sorted by address
3208 * and into the inode's i_mmap tree. If vm_file is non-NULL
3209 * then i_mmap_rwsem is taken here.
3211 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3213 struct vm_area_struct *prev;
3214 struct rb_node **rb_link, *rb_parent;
3216 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3217 &prev, &rb_link, &rb_parent))
3219 if ((vma->vm_flags & VM_ACCOUNT) &&
3220 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3224 * The vm_pgoff of a purely anonymous vma should be irrelevant
3225 * until its first write fault, when page's anon_vma and index
3226 * are set. But now set the vm_pgoff it will almost certainly
3227 * end up with (unless mremap moves it elsewhere before that
3228 * first wfault), so /proc/pid/maps tells a consistent story.
3230 * By setting it to reflect the virtual start address of the
3231 * vma, merges and splits can happen in a seamless way, just
3232 * using the existing file pgoff checks and manipulations.
3233 * Similarly in do_mmap and in do_brk_flags.
3235 if (vma_is_anonymous(vma)) {
3236 BUG_ON(vma->anon_vma);
3237 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3240 vma_link(mm, vma, prev, rb_link, rb_parent);
3245 * Copy the vma structure to a new location in the same mm,
3246 * prior to moving page table entries, to effect an mremap move.
3248 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3249 unsigned long addr, unsigned long len, pgoff_t pgoff,
3250 bool *need_rmap_locks)
3252 struct vm_area_struct *vma = *vmap;
3253 unsigned long vma_start = vma->vm_start;
3254 struct mm_struct *mm = vma->vm_mm;
3255 struct vm_area_struct *new_vma, *prev;
3256 struct rb_node **rb_link, *rb_parent;
3257 bool faulted_in_anon_vma = true;
3260 * If anonymous vma has not yet been faulted, update new pgoff
3261 * to match new location, to increase its chance of merging.
3263 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3264 pgoff = addr >> PAGE_SHIFT;
3265 faulted_in_anon_vma = false;
3268 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3269 return NULL; /* should never get here */
3270 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3271 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3272 vma->vm_userfaultfd_ctx);
3275 * Source vma may have been merged into new_vma
3277 if (unlikely(vma_start >= new_vma->vm_start &&
3278 vma_start < new_vma->vm_end)) {
3280 * The only way we can get a vma_merge with
3281 * self during an mremap is if the vma hasn't
3282 * been faulted in yet and we were allowed to
3283 * reset the dst vma->vm_pgoff to the
3284 * destination address of the mremap to allow
3285 * the merge to happen. mremap must change the
3286 * vm_pgoff linearity between src and dst vmas
3287 * (in turn preventing a vma_merge) to be
3288 * safe. It is only safe to keep the vm_pgoff
3289 * linear if there are no pages mapped yet.
3291 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3292 *vmap = vma = new_vma;
3294 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3296 new_vma = vm_area_dup(vma);
3299 new_vma->vm_start = addr;
3300 new_vma->vm_end = addr + len;
3301 new_vma->vm_pgoff = pgoff;
3302 if (vma_dup_policy(vma, new_vma))
3304 if (anon_vma_clone(new_vma, vma))
3305 goto out_free_mempol;
3306 if (new_vma->vm_file)
3307 get_file(new_vma->vm_file);
3308 if (new_vma->vm_ops && new_vma->vm_ops->open)
3309 new_vma->vm_ops->open(new_vma);
3310 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3311 *need_rmap_locks = false;
3316 mpol_put(vma_policy(new_vma));
3318 vm_area_free(new_vma);
3324 * Return true if the calling process may expand its vm space by the passed
3327 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3329 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3332 if (is_data_mapping(flags) &&
3333 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3334 /* Workaround for Valgrind */
3335 if (rlimit(RLIMIT_DATA) == 0 &&
3336 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3339 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3340 current->comm, current->pid,
3341 (mm->data_vm + npages) << PAGE_SHIFT,
3342 rlimit(RLIMIT_DATA),
3343 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3345 if (!ignore_rlimit_data)
3352 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3354 mm->total_vm += npages;
3356 if (is_exec_mapping(flags))
3357 mm->exec_vm += npages;
3358 else if (is_stack_mapping(flags))
3359 mm->stack_vm += npages;
3360 else if (is_data_mapping(flags))
3361 mm->data_vm += npages;
3364 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3367 * Having a close hook prevents vma merging regardless of flags.
3369 static void special_mapping_close(struct vm_area_struct *vma)
3373 static const char *special_mapping_name(struct vm_area_struct *vma)
3375 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3378 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3380 struct vm_special_mapping *sm = new_vma->vm_private_data;
3382 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3386 return sm->mremap(sm, new_vma);
3391 static const struct vm_operations_struct special_mapping_vmops = {
3392 .close = special_mapping_close,
3393 .fault = special_mapping_fault,
3394 .mremap = special_mapping_mremap,
3395 .name = special_mapping_name,
3396 /* vDSO code relies that VVAR can't be accessed remotely */
3400 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3401 .close = special_mapping_close,
3402 .fault = special_mapping_fault,
3405 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3407 struct vm_area_struct *vma = vmf->vma;
3409 struct page **pages;
3411 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3412 pages = vma->vm_private_data;
3414 struct vm_special_mapping *sm = vma->vm_private_data;
3417 return sm->fault(sm, vmf->vma, vmf);
3422 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3426 struct page *page = *pages;
3432 return VM_FAULT_SIGBUS;
3435 static struct vm_area_struct *__install_special_mapping(
3436 struct mm_struct *mm,
3437 unsigned long addr, unsigned long len,
3438 unsigned long vm_flags, void *priv,
3439 const struct vm_operations_struct *ops)
3442 struct vm_area_struct *vma;
3444 vma = vm_area_alloc(mm);
3445 if (unlikely(vma == NULL))
3446 return ERR_PTR(-ENOMEM);
3448 vma->vm_start = addr;
3449 vma->vm_end = addr + len;
3451 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3452 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3455 vma->vm_private_data = priv;
3457 ret = insert_vm_struct(mm, vma);
3461 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3463 perf_event_mmap(vma);
3469 return ERR_PTR(ret);
3472 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3473 const struct vm_special_mapping *sm)
3475 return vma->vm_private_data == sm &&
3476 (vma->vm_ops == &special_mapping_vmops ||
3477 vma->vm_ops == &legacy_special_mapping_vmops);
3481 * Called with mm->mmap_lock held for writing.
3482 * Insert a new vma covering the given region, with the given flags.
3483 * Its pages are supplied by the given array of struct page *.
3484 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3485 * The region past the last page supplied will always produce SIGBUS.
3486 * The array pointer and the pages it points to are assumed to stay alive
3487 * for as long as this mapping might exist.
3489 struct vm_area_struct *_install_special_mapping(
3490 struct mm_struct *mm,
3491 unsigned long addr, unsigned long len,
3492 unsigned long vm_flags, const struct vm_special_mapping *spec)
3494 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3495 &special_mapping_vmops);
3498 int install_special_mapping(struct mm_struct *mm,
3499 unsigned long addr, unsigned long len,
3500 unsigned long vm_flags, struct page **pages)
3502 struct vm_area_struct *vma = __install_special_mapping(
3503 mm, addr, len, vm_flags, (void *)pages,
3504 &legacy_special_mapping_vmops);
3506 return PTR_ERR_OR_ZERO(vma);
3509 static DEFINE_MUTEX(mm_all_locks_mutex);
3511 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3513 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3515 * The LSB of head.next can't change from under us
3516 * because we hold the mm_all_locks_mutex.
3518 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3520 * We can safely modify head.next after taking the
3521 * anon_vma->root->rwsem. If some other vma in this mm shares
3522 * the same anon_vma we won't take it again.
3524 * No need of atomic instructions here, head.next
3525 * can't change from under us thanks to the
3526 * anon_vma->root->rwsem.
3528 if (__test_and_set_bit(0, (unsigned long *)
3529 &anon_vma->root->rb_root.rb_root.rb_node))
3534 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3536 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3538 * AS_MM_ALL_LOCKS can't change from under us because
3539 * we hold the mm_all_locks_mutex.
3541 * Operations on ->flags have to be atomic because
3542 * even if AS_MM_ALL_LOCKS is stable thanks to the
3543 * mm_all_locks_mutex, there may be other cpus
3544 * changing other bitflags in parallel to us.
3546 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3548 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3553 * This operation locks against the VM for all pte/vma/mm related
3554 * operations that could ever happen on a certain mm. This includes
3555 * vmtruncate, try_to_unmap, and all page faults.
3557 * The caller must take the mmap_lock in write mode before calling
3558 * mm_take_all_locks(). The caller isn't allowed to release the
3559 * mmap_lock until mm_drop_all_locks() returns.
3561 * mmap_lock in write mode is required in order to block all operations
3562 * that could modify pagetables and free pages without need of
3563 * altering the vma layout. It's also needed in write mode to avoid new
3564 * anon_vmas to be associated with existing vmas.
3566 * A single task can't take more than one mm_take_all_locks() in a row
3567 * or it would deadlock.
3569 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3570 * mapping->flags avoid to take the same lock twice, if more than one
3571 * vma in this mm is backed by the same anon_vma or address_space.
3573 * We take locks in following order, accordingly to comment at beginning
3575 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3577 * - all i_mmap_rwsem locks;
3578 * - all anon_vma->rwseml
3580 * We can take all locks within these types randomly because the VM code
3581 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3582 * mm_all_locks_mutex.
3584 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3585 * that may have to take thousand of locks.
3587 * mm_take_all_locks() can fail if it's interrupted by signals.
3589 int mm_take_all_locks(struct mm_struct *mm)
3591 struct vm_area_struct *vma;
3592 struct anon_vma_chain *avc;
3594 BUG_ON(mmap_read_trylock(mm));
3596 mutex_lock(&mm_all_locks_mutex);
3598 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3599 if (signal_pending(current))
3601 if (vma->vm_file && vma->vm_file->f_mapping &&
3602 is_vm_hugetlb_page(vma))
3603 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3606 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3607 if (signal_pending(current))
3609 if (vma->vm_file && vma->vm_file->f_mapping &&
3610 !is_vm_hugetlb_page(vma))
3611 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3614 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3615 if (signal_pending(current))
3618 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3619 vm_lock_anon_vma(mm, avc->anon_vma);
3625 mm_drop_all_locks(mm);
3629 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3631 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3633 * The LSB of head.next can't change to 0 from under
3634 * us because we hold the mm_all_locks_mutex.
3636 * We must however clear the bitflag before unlocking
3637 * the vma so the users using the anon_vma->rb_root will
3638 * never see our bitflag.
3640 * No need of atomic instructions here, head.next
3641 * can't change from under us until we release the
3642 * anon_vma->root->rwsem.
3644 if (!__test_and_clear_bit(0, (unsigned long *)
3645 &anon_vma->root->rb_root.rb_root.rb_node))
3647 anon_vma_unlock_write(anon_vma);
3651 static void vm_unlock_mapping(struct address_space *mapping)
3653 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3655 * AS_MM_ALL_LOCKS can't change to 0 from under us
3656 * because we hold the mm_all_locks_mutex.
3658 i_mmap_unlock_write(mapping);
3659 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3666 * The mmap_lock cannot be released by the caller until
3667 * mm_drop_all_locks() returns.
3669 void mm_drop_all_locks(struct mm_struct *mm)
3671 struct vm_area_struct *vma;
3672 struct anon_vma_chain *avc;
3674 BUG_ON(mmap_read_trylock(mm));
3675 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3677 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3679 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3680 vm_unlock_anon_vma(avc->anon_vma);
3681 if (vma->vm_file && vma->vm_file->f_mapping)
3682 vm_unlock_mapping(vma->vm_file->f_mapping);
3685 mutex_unlock(&mm_all_locks_mutex);
3689 * initialise the percpu counter for VM
3691 void __init mmap_init(void)
3695 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3700 * Initialise sysctl_user_reserve_kbytes.
3702 * This is intended to prevent a user from starting a single memory hogging
3703 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3706 * The default value is min(3% of free memory, 128MB)
3707 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3709 static int init_user_reserve(void)
3711 unsigned long free_kbytes;
3713 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3715 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3718 subsys_initcall(init_user_reserve);
3721 * Initialise sysctl_admin_reserve_kbytes.
3723 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3724 * to log in and kill a memory hogging process.
3726 * Systems with more than 256MB will reserve 8MB, enough to recover
3727 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3728 * only reserve 3% of free pages by default.
3730 static int init_admin_reserve(void)
3732 unsigned long free_kbytes;
3734 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3736 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3739 subsys_initcall(init_admin_reserve);
3742 * Reinititalise user and admin reserves if memory is added or removed.
3744 * The default user reserve max is 128MB, and the default max for the
3745 * admin reserve is 8MB. These are usually, but not always, enough to
3746 * enable recovery from a memory hogging process using login/sshd, a shell,
3747 * and tools like top. It may make sense to increase or even disable the
3748 * reserve depending on the existence of swap or variations in the recovery
3749 * tools. So, the admin may have changed them.
3751 * If memory is added and the reserves have been eliminated or increased above
3752 * the default max, then we'll trust the admin.
3754 * If memory is removed and there isn't enough free memory, then we
3755 * need to reset the reserves.
3757 * Otherwise keep the reserve set by the admin.
3759 static int reserve_mem_notifier(struct notifier_block *nb,
3760 unsigned long action, void *data)
3762 unsigned long tmp, free_kbytes;
3766 /* Default max is 128MB. Leave alone if modified by operator. */
3767 tmp = sysctl_user_reserve_kbytes;
3768 if (0 < tmp && tmp < (1UL << 17))
3769 init_user_reserve();
3771 /* Default max is 8MB. Leave alone if modified by operator. */
3772 tmp = sysctl_admin_reserve_kbytes;
3773 if (0 < tmp && tmp < (1UL << 13))
3774 init_admin_reserve();
3778 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3780 if (sysctl_user_reserve_kbytes > free_kbytes) {
3781 init_user_reserve();
3782 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3783 sysctl_user_reserve_kbytes);
3786 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3787 init_admin_reserve();
3788 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3789 sysctl_admin_reserve_kbytes);
3798 static struct notifier_block reserve_mem_nb = {
3799 .notifier_call = reserve_mem_notifier,
3802 static int __meminit init_reserve_notifier(void)
3804 if (register_hotmemory_notifier(&reserve_mem_nb))
3805 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3809 subsys_initcall(init_reserve_notifier);
projects / linux-2.6-microblaze.git / blob