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 newbrk, oldbrk, origbrk;
193 struct mm_struct *mm = current->mm;
194 struct vm_area_struct *next;
195 unsigned long min_brk;
197 bool downgraded = false;
200 if (mmap_write_lock_killable(mm))
205 #ifdef CONFIG_COMPAT_BRK
207 * CONFIG_COMPAT_BRK can still be overridden by setting
208 * randomize_va_space to 2, which will still cause mm->start_brk
209 * to be arbitrarily shifted
211 if (current->brk_randomized)
212 min_brk = mm->start_brk;
214 min_brk = mm->end_data;
216 min_brk = mm->start_brk;
222 * Check against rlimit here. If this check is done later after the test
223 * of oldbrk with newbrk then it can escape the test and let the data
224 * segment grow beyond its set limit the in case where the limit is
225 * not page aligned -Ram Gupta
227 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
228 mm->end_data, mm->start_data))
231 newbrk = PAGE_ALIGN(brk);
232 oldbrk = PAGE_ALIGN(mm->brk);
233 if (oldbrk == newbrk) {
239 * Always allow shrinking brk.
240 * __do_munmap() may downgrade mmap_lock to read.
242 if (brk <= mm->brk) {
246 * mm->brk must to be protected by write mmap_lock so update it
247 * before downgrading mmap_lock. When __do_munmap() fails,
248 * mm->brk will be restored from origbrk.
251 ret = __do_munmap(mm, newbrk, oldbrk-newbrk, &uf, true);
255 } else if (ret == 1) {
261 /* Check against existing mmap mappings. */
262 next = find_vma(mm, oldbrk);
263 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
266 /* Ok, looks good - let it rip. */
267 if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
272 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
274 mmap_read_unlock(mm);
276 mmap_write_unlock(mm);
277 userfaultfd_unmap_complete(mm, &uf);
279 mm_populate(oldbrk, newbrk - oldbrk);
283 mmap_write_unlock(mm);
287 static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
289 unsigned long gap, prev_end;
292 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
293 * allow two stack_guard_gaps between them here, and when choosing
294 * an unmapped area; whereas when expanding we only require one.
295 * That's a little inconsistent, but keeps the code here simpler.
297 gap = vm_start_gap(vma);
299 prev_end = vm_end_gap(vma->vm_prev);
308 #ifdef CONFIG_DEBUG_VM_RB
309 static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma)
311 unsigned long max = vma_compute_gap(vma), subtree_gap;
312 if (vma->vm_rb.rb_left) {
313 subtree_gap = rb_entry(vma->vm_rb.rb_left,
314 struct vm_area_struct, vm_rb)->rb_subtree_gap;
315 if (subtree_gap > max)
318 if (vma->vm_rb.rb_right) {
319 subtree_gap = rb_entry(vma->vm_rb.rb_right,
320 struct vm_area_struct, vm_rb)->rb_subtree_gap;
321 if (subtree_gap > max)
327 static int browse_rb(struct mm_struct *mm)
329 struct rb_root *root = &mm->mm_rb;
330 int i = 0, j, bug = 0;
331 struct rb_node *nd, *pn = NULL;
332 unsigned long prev = 0, pend = 0;
334 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
335 struct vm_area_struct *vma;
336 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
337 if (vma->vm_start < prev) {
338 pr_emerg("vm_start %lx < prev %lx\n",
339 vma->vm_start, prev);
342 if (vma->vm_start < pend) {
343 pr_emerg("vm_start %lx < pend %lx\n",
344 vma->vm_start, pend);
347 if (vma->vm_start > vma->vm_end) {
348 pr_emerg("vm_start %lx > vm_end %lx\n",
349 vma->vm_start, vma->vm_end);
352 spin_lock(&mm->page_table_lock);
353 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
354 pr_emerg("free gap %lx, correct %lx\n",
356 vma_compute_subtree_gap(vma));
359 spin_unlock(&mm->page_table_lock);
362 prev = vma->vm_start;
366 for (nd = pn; nd; nd = rb_prev(nd))
369 pr_emerg("backwards %d, forwards %d\n", j, i);
375 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
379 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
380 struct vm_area_struct *vma;
381 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
382 VM_BUG_ON_VMA(vma != ignore &&
383 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
388 static void validate_mm(struct mm_struct *mm)
392 unsigned long highest_address = 0;
393 struct vm_area_struct *vma = mm->mmap;
396 struct anon_vma *anon_vma = vma->anon_vma;
397 struct anon_vma_chain *avc;
400 anon_vma_lock_read(anon_vma);
401 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
402 anon_vma_interval_tree_verify(avc);
403 anon_vma_unlock_read(anon_vma);
406 highest_address = vm_end_gap(vma);
410 if (i != mm->map_count) {
411 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
414 if (highest_address != mm->highest_vm_end) {
415 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
416 mm->highest_vm_end, highest_address);
420 if (i != mm->map_count) {
422 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
425 VM_BUG_ON_MM(bug, mm);
428 #define validate_mm_rb(root, ignore) do { } while (0)
429 #define validate_mm(mm) do { } while (0)
432 RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks,
433 struct vm_area_struct, vm_rb,
434 unsigned long, rb_subtree_gap, vma_compute_gap)
437 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
438 * vma->vm_prev->vm_end values changed, without modifying the vma's position
441 static void vma_gap_update(struct vm_area_struct *vma)
444 * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
445 * a callback function that does exactly what we want.
447 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
450 static inline void vma_rb_insert(struct vm_area_struct *vma,
451 struct rb_root *root)
453 /* All rb_subtree_gap values must be consistent prior to insertion */
454 validate_mm_rb(root, NULL);
456 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
459 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
462 * Note rb_erase_augmented is a fairly large inline function,
463 * so make sure we instantiate it only once with our desired
464 * augmented rbtree callbacks.
466 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
469 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
470 struct rb_root *root,
471 struct vm_area_struct *ignore)
474 * All rb_subtree_gap values must be consistent prior to erase,
475 * with the possible exception of
477 * a. the "next" vma being erased if next->vm_start was reduced in
478 * __vma_adjust() -> __vma_unlink()
479 * b. the vma being erased in detach_vmas_to_be_unmapped() ->
482 validate_mm_rb(root, ignore);
484 __vma_rb_erase(vma, root);
487 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
488 struct rb_root *root)
490 vma_rb_erase_ignore(vma, root, vma);
494 * vma has some anon_vma assigned, and is already inserted on that
495 * anon_vma's interval trees.
497 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
498 * vma must be removed from the anon_vma's interval trees using
499 * anon_vma_interval_tree_pre_update_vma().
501 * After the update, the vma will be reinserted using
502 * anon_vma_interval_tree_post_update_vma().
504 * The entire update must be protected by exclusive mmap_lock and by
505 * the root anon_vma's mutex.
508 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
510 struct anon_vma_chain *avc;
512 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
513 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
517 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
519 struct anon_vma_chain *avc;
521 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
522 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
525 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
526 unsigned long end, struct vm_area_struct **pprev,
527 struct rb_node ***rb_link, struct rb_node **rb_parent)
529 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
531 __rb_link = &mm->mm_rb.rb_node;
532 rb_prev = __rb_parent = NULL;
535 struct vm_area_struct *vma_tmp;
537 __rb_parent = *__rb_link;
538 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
540 if (vma_tmp->vm_end > addr) {
541 /* Fail if an existing vma overlaps the area */
542 if (vma_tmp->vm_start < end)
544 __rb_link = &__rb_parent->rb_left;
546 rb_prev = __rb_parent;
547 __rb_link = &__rb_parent->rb_right;
553 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
554 *rb_link = __rb_link;
555 *rb_parent = __rb_parent;
560 * vma_next() - Get the next VMA.
561 * @mm: The mm_struct.
562 * @vma: The current vma.
564 * If @vma is NULL, return the first vma in the mm.
566 * Returns: The next VMA after @vma.
568 static inline struct vm_area_struct *vma_next(struct mm_struct *mm,
569 struct vm_area_struct *vma)
578 * munmap_vma_range() - munmap VMAs that overlap a range.
580 * @start: The start of the range.
581 * @len: The length of the range.
582 * @pprev: pointer to the pointer that will be set to previous vm_area_struct
583 * @rb_link: the rb_node
584 * @rb_parent: the parent rb_node
586 * Find all the vm_area_struct that overlap from @start to
587 * @end and munmap them. Set @pprev to the previous vm_area_struct.
589 * Returns: -ENOMEM on munmap failure or 0 on success.
592 munmap_vma_range(struct mm_struct *mm, unsigned long start, unsigned long len,
593 struct vm_area_struct **pprev, struct rb_node ***link,
594 struct rb_node **parent, struct list_head *uf)
597 while (find_vma_links(mm, start, start + len, pprev, link, parent))
598 if (do_munmap(mm, start, len, uf))
603 static unsigned long count_vma_pages_range(struct mm_struct *mm,
604 unsigned long addr, unsigned long end)
606 unsigned long nr_pages = 0;
607 struct vm_area_struct *vma;
609 /* Find first overlaping mapping */
610 vma = find_vma_intersection(mm, addr, end);
614 nr_pages = (min(end, vma->vm_end) -
615 max(addr, vma->vm_start)) >> PAGE_SHIFT;
617 /* Iterate over the rest of the overlaps */
618 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
619 unsigned long overlap_len;
621 if (vma->vm_start > end)
624 overlap_len = min(end, vma->vm_end) - vma->vm_start;
625 nr_pages += overlap_len >> PAGE_SHIFT;
631 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
632 struct rb_node **rb_link, struct rb_node *rb_parent)
634 /* Update tracking information for the gap following the new vma. */
636 vma_gap_update(vma->vm_next);
638 mm->highest_vm_end = vm_end_gap(vma);
641 * vma->vm_prev wasn't known when we followed the rbtree to find the
642 * correct insertion point for that vma. As a result, we could not
643 * update the vma vm_rb parents rb_subtree_gap values on the way down.
644 * So, we first insert the vma with a zero rb_subtree_gap value
645 * (to be consistent with what we did on the way down), and then
646 * immediately update the gap to the correct value. Finally we
647 * rebalance the rbtree after all augmented values have been set.
649 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
650 vma->rb_subtree_gap = 0;
652 vma_rb_insert(vma, &mm->mm_rb);
655 static void __vma_link_file(struct vm_area_struct *vma)
661 struct address_space *mapping = file->f_mapping;
663 if (vma->vm_flags & VM_DENYWRITE)
664 put_write_access(file_inode(file));
665 if (vma->vm_flags & VM_SHARED)
666 mapping_allow_writable(mapping);
668 flush_dcache_mmap_lock(mapping);
669 vma_interval_tree_insert(vma, &mapping->i_mmap);
670 flush_dcache_mmap_unlock(mapping);
675 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
676 struct vm_area_struct *prev, struct rb_node **rb_link,
677 struct rb_node *rb_parent)
679 __vma_link_list(mm, vma, prev);
680 __vma_link_rb(mm, vma, rb_link, rb_parent);
683 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
684 struct vm_area_struct *prev, struct rb_node **rb_link,
685 struct rb_node *rb_parent)
687 struct address_space *mapping = NULL;
690 mapping = vma->vm_file->f_mapping;
691 i_mmap_lock_write(mapping);
694 __vma_link(mm, vma, prev, rb_link, rb_parent);
695 __vma_link_file(vma);
698 i_mmap_unlock_write(mapping);
705 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
706 * mm's list and rbtree. It has already been inserted into the interval tree.
708 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
710 struct vm_area_struct *prev;
711 struct rb_node **rb_link, *rb_parent;
713 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
714 &prev, &rb_link, &rb_parent))
716 __vma_link(mm, vma, prev, rb_link, rb_parent);
720 static __always_inline void __vma_unlink(struct mm_struct *mm,
721 struct vm_area_struct *vma,
722 struct vm_area_struct *ignore)
724 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
725 __vma_unlink_list(mm, vma);
727 vmacache_invalidate(mm);
731 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
732 * is already present in an i_mmap tree without adjusting the tree.
733 * The following helper function should be used when such adjustments
734 * are necessary. The "insert" vma (if any) is to be inserted
735 * before we drop the necessary locks.
737 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
738 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
739 struct vm_area_struct *expand)
741 struct mm_struct *mm = vma->vm_mm;
742 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
743 struct address_space *mapping = NULL;
744 struct rb_root_cached *root = NULL;
745 struct anon_vma *anon_vma = NULL;
746 struct file *file = vma->vm_file;
747 bool start_changed = false, end_changed = false;
748 long adjust_next = 0;
751 if (next && !insert) {
752 struct vm_area_struct *exporter = NULL, *importer = NULL;
754 if (end >= next->vm_end) {
756 * vma expands, overlapping all the next, and
757 * perhaps the one after too (mprotect case 6).
758 * The only other cases that gets here are
759 * case 1, case 7 and case 8.
761 if (next == expand) {
763 * The only case where we don't expand "vma"
764 * and we expand "next" instead is case 8.
766 VM_WARN_ON(end != next->vm_end);
768 * remove_next == 3 means we're
769 * removing "vma" and that to do so we
770 * swapped "vma" and "next".
773 VM_WARN_ON(file != next->vm_file);
776 VM_WARN_ON(expand != vma);
778 * case 1, 6, 7, remove_next == 2 is case 6,
779 * remove_next == 1 is case 1 or 7.
781 remove_next = 1 + (end > next->vm_end);
782 VM_WARN_ON(remove_next == 2 &&
783 end != next->vm_next->vm_end);
784 /* trim end to next, for case 6 first pass */
792 * If next doesn't have anon_vma, import from vma after
793 * next, if the vma overlaps with it.
795 if (remove_next == 2 && !next->anon_vma)
796 exporter = next->vm_next;
798 } else if (end > next->vm_start) {
800 * vma expands, overlapping part of the next:
801 * mprotect case 5 shifting the boundary up.
803 adjust_next = (end - next->vm_start);
806 VM_WARN_ON(expand != importer);
807 } else if (end < vma->vm_end) {
809 * vma shrinks, and !insert tells it's not
810 * split_vma inserting another: so it must be
811 * mprotect case 4 shifting the boundary down.
813 adjust_next = -(vma->vm_end - end);
816 VM_WARN_ON(expand != importer);
820 * Easily overlooked: when mprotect shifts the boundary,
821 * make sure the expanding vma has anon_vma set if the
822 * shrinking vma had, to cover any anon pages imported.
824 if (exporter && exporter->anon_vma && !importer->anon_vma) {
827 importer->anon_vma = exporter->anon_vma;
828 error = anon_vma_clone(importer, exporter);
834 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
837 mapping = file->f_mapping;
838 root = &mapping->i_mmap;
839 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
842 uprobe_munmap(next, next->vm_start, next->vm_end);
844 i_mmap_lock_write(mapping);
847 * Put into interval tree now, so instantiated pages
848 * are visible to arm/parisc __flush_dcache_page
849 * throughout; but we cannot insert into address
850 * space until vma start or end is updated.
852 __vma_link_file(insert);
856 anon_vma = vma->anon_vma;
857 if (!anon_vma && adjust_next)
858 anon_vma = next->anon_vma;
860 VM_WARN_ON(adjust_next && next->anon_vma &&
861 anon_vma != next->anon_vma);
862 anon_vma_lock_write(anon_vma);
863 anon_vma_interval_tree_pre_update_vma(vma);
865 anon_vma_interval_tree_pre_update_vma(next);
869 flush_dcache_mmap_lock(mapping);
870 vma_interval_tree_remove(vma, root);
872 vma_interval_tree_remove(next, root);
875 if (start != vma->vm_start) {
876 vma->vm_start = start;
877 start_changed = true;
879 if (end != vma->vm_end) {
883 vma->vm_pgoff = pgoff;
885 next->vm_start += adjust_next;
886 next->vm_pgoff += adjust_next >> PAGE_SHIFT;
891 vma_interval_tree_insert(next, root);
892 vma_interval_tree_insert(vma, root);
893 flush_dcache_mmap_unlock(mapping);
898 * vma_merge has merged next into vma, and needs
899 * us to remove next before dropping the locks.
901 if (remove_next != 3)
902 __vma_unlink(mm, next, next);
905 * vma is not before next if they've been
908 * pre-swap() next->vm_start was reduced so
909 * tell validate_mm_rb to ignore pre-swap()
910 * "next" (which is stored in post-swap()
913 __vma_unlink(mm, next, vma);
915 __remove_shared_vm_struct(next, file, mapping);
918 * split_vma has split insert from vma, and needs
919 * us to insert it before dropping the locks
920 * (it may either follow vma or precede it).
922 __insert_vm_struct(mm, insert);
928 mm->highest_vm_end = vm_end_gap(vma);
929 else if (!adjust_next)
930 vma_gap_update(next);
935 anon_vma_interval_tree_post_update_vma(vma);
937 anon_vma_interval_tree_post_update_vma(next);
938 anon_vma_unlock_write(anon_vma);
942 i_mmap_unlock_write(mapping);
951 uprobe_munmap(next, next->vm_start, next->vm_end);
955 anon_vma_merge(vma, next);
957 mpol_put(vma_policy(next));
960 * In mprotect's case 6 (see comments on vma_merge),
961 * we must remove another next too. It would clutter
962 * up the code too much to do both in one go.
964 if (remove_next != 3) {
966 * If "next" was removed and vma->vm_end was
967 * expanded (up) over it, in turn
968 * "next->vm_prev->vm_end" changed and the
969 * "vma->vm_next" gap must be updated.
974 * For the scope of the comment "next" and
975 * "vma" considered pre-swap(): if "vma" was
976 * removed, next->vm_start was expanded (down)
977 * over it and the "next" gap must be updated.
978 * Because of the swap() the post-swap() "vma"
979 * actually points to pre-swap() "next"
980 * (post-swap() "next" as opposed is now a
985 if (remove_next == 2) {
991 vma_gap_update(next);
994 * If remove_next == 2 we obviously can't
997 * If remove_next == 3 we can't reach this
998 * path because pre-swap() next is always not
999 * NULL. pre-swap() "next" is not being
1000 * removed and its next->vm_end is not altered
1001 * (and furthermore "end" already matches
1002 * next->vm_end in remove_next == 3).
1004 * We reach this only in the remove_next == 1
1005 * case if the "next" vma that was removed was
1006 * the highest vma of the mm. However in such
1007 * case next->vm_end == "end" and the extended
1008 * "vma" has vma->vm_end == next->vm_end so
1009 * mm->highest_vm_end doesn't need any update
1010 * in remove_next == 1 case.
1012 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
1016 uprobe_mmap(insert);
1024 * If the vma has a ->close operation then the driver probably needs to release
1025 * per-vma resources, so we don't attempt to merge those.
1027 static inline int is_mergeable_vma(struct vm_area_struct *vma,
1028 struct file *file, unsigned long vm_flags,
1029 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1032 * VM_SOFTDIRTY should not prevent from VMA merging, if we
1033 * match the flags but dirty bit -- the caller should mark
1034 * merged VMA as dirty. If dirty bit won't be excluded from
1035 * comparison, we increase pressure on the memory system forcing
1036 * the kernel to generate new VMAs when old one could be
1039 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
1041 if (vma->vm_file != file)
1043 if (vma->vm_ops && vma->vm_ops->close)
1045 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1050 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1051 struct anon_vma *anon_vma2,
1052 struct vm_area_struct *vma)
1055 * The list_is_singular() test is to avoid merging VMA cloned from
1056 * parents. This can improve scalability caused by anon_vma lock.
1058 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1059 list_is_singular(&vma->anon_vma_chain)))
1061 return anon_vma1 == anon_vma2;
1065 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1066 * in front of (at a lower virtual address and file offset than) the vma.
1068 * We cannot merge two vmas if they have differently assigned (non-NULL)
1069 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1071 * We don't check here for the merged mmap wrapping around the end of pagecache
1072 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
1073 * wrap, nor mmaps which cover the final page at index -1UL.
1076 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1077 struct anon_vma *anon_vma, struct file *file,
1079 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1081 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1082 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1083 if (vma->vm_pgoff == vm_pgoff)
1090 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1091 * beyond (at a higher virtual address and file offset than) the vma.
1093 * We cannot merge two vmas if they have differently assigned (non-NULL)
1094 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1097 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1098 struct anon_vma *anon_vma, struct file *file,
1100 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1102 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1103 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1105 vm_pglen = vma_pages(vma);
1106 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1113 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1114 * whether that can be merged with its predecessor or its successor.
1115 * Or both (it neatly fills a hole).
1117 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1118 * certain not to be mapped by the time vma_merge is called; but when
1119 * called for mprotect, it is certain to be already mapped (either at
1120 * an offset within prev, or at the start of next), and the flags of
1121 * this area are about to be changed to vm_flags - and the no-change
1122 * case has already been eliminated.
1124 * The following mprotect cases have to be considered, where AAAA is
1125 * the area passed down from mprotect_fixup, never extending beyond one
1126 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1129 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN
1130 * cannot merge might become might become
1131 * PPNNNNNNNNNN PPPPPPPPPPNN
1132 * mmap, brk or case 4 below case 5 below
1135 * PPPP NNNN PPPPNNNNXXXX
1136 * might become might become
1137 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
1138 * PPPPPPPPNNNN 2 or PPPPPPPPXXXX 7 or
1139 * PPPPNNNNNNNN 3 PPPPXXXXXXXX 8
1141 * It is important for case 8 that the vma NNNN overlapping the
1142 * region AAAA is never going to extended over XXXX. Instead XXXX must
1143 * be extended in region AAAA and NNNN must be removed. This way in
1144 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1145 * rmap_locks, the properties of the merged vma will be already
1146 * correct for the whole merged range. Some of those properties like
1147 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1148 * be correct for the whole merged range immediately after the
1149 * rmap_locks are released. Otherwise if XXXX would be removed and
1150 * NNNN would be extended over the XXXX range, remove_migration_ptes
1151 * or other rmap walkers (if working on addresses beyond the "end"
1152 * parameter) may establish ptes with the wrong permissions of NNNN
1153 * instead of the right permissions of XXXX.
1155 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1156 struct vm_area_struct *prev, unsigned long addr,
1157 unsigned long end, unsigned long vm_flags,
1158 struct anon_vma *anon_vma, struct file *file,
1159 pgoff_t pgoff, struct mempolicy *policy,
1160 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1162 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1163 struct vm_area_struct *area, *next;
1167 * We later require that vma->vm_flags == vm_flags,
1168 * so this tests vma->vm_flags & VM_SPECIAL, too.
1170 if (vm_flags & VM_SPECIAL)
1173 next = vma_next(mm, prev);
1175 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1176 next = next->vm_next;
1178 /* verify some invariant that must be enforced by the caller */
1179 VM_WARN_ON(prev && addr <= prev->vm_start);
1180 VM_WARN_ON(area && end > area->vm_end);
1181 VM_WARN_ON(addr >= end);
1184 * Can it merge with the predecessor?
1186 if (prev && prev->vm_end == addr &&
1187 mpol_equal(vma_policy(prev), policy) &&
1188 can_vma_merge_after(prev, vm_flags,
1189 anon_vma, file, pgoff,
1190 vm_userfaultfd_ctx)) {
1192 * OK, it can. Can we now merge in the successor as well?
1194 if (next && end == next->vm_start &&
1195 mpol_equal(policy, vma_policy(next)) &&
1196 can_vma_merge_before(next, vm_flags,
1199 vm_userfaultfd_ctx) &&
1200 is_mergeable_anon_vma(prev->anon_vma,
1201 next->anon_vma, NULL)) {
1203 err = __vma_adjust(prev, prev->vm_start,
1204 next->vm_end, prev->vm_pgoff, NULL,
1206 } else /* cases 2, 5, 7 */
1207 err = __vma_adjust(prev, prev->vm_start,
1208 end, prev->vm_pgoff, NULL, prev);
1211 khugepaged_enter_vma_merge(prev, vm_flags);
1216 * Can this new request be merged in front of next?
1218 if (next && end == next->vm_start &&
1219 mpol_equal(policy, vma_policy(next)) &&
1220 can_vma_merge_before(next, vm_flags,
1221 anon_vma, file, pgoff+pglen,
1222 vm_userfaultfd_ctx)) {
1223 if (prev && addr < prev->vm_end) /* case 4 */
1224 err = __vma_adjust(prev, prev->vm_start,
1225 addr, prev->vm_pgoff, NULL, next);
1226 else { /* cases 3, 8 */
1227 err = __vma_adjust(area, addr, next->vm_end,
1228 next->vm_pgoff - pglen, NULL, next);
1230 * In case 3 area is already equal to next and
1231 * this is a noop, but in case 8 "area" has
1232 * been removed and next was expanded over it.
1238 khugepaged_enter_vma_merge(area, vm_flags);
1246 * Rough compatibility check to quickly see if it's even worth looking
1247 * at sharing an anon_vma.
1249 * They need to have the same vm_file, and the flags can only differ
1250 * in things that mprotect may change.
1252 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1253 * we can merge the two vma's. For example, we refuse to merge a vma if
1254 * there is a vm_ops->close() function, because that indicates that the
1255 * driver is doing some kind of reference counting. But that doesn't
1256 * really matter for the anon_vma sharing case.
1258 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1260 return a->vm_end == b->vm_start &&
1261 mpol_equal(vma_policy(a), vma_policy(b)) &&
1262 a->vm_file == b->vm_file &&
1263 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1264 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1268 * Do some basic sanity checking to see if we can re-use the anon_vma
1269 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1270 * the same as 'old', the other will be the new one that is trying
1271 * to share the anon_vma.
1273 * NOTE! This runs with mm_sem held for reading, so it is possible that
1274 * the anon_vma of 'old' is concurrently in the process of being set up
1275 * by another page fault trying to merge _that_. But that's ok: if it
1276 * is being set up, that automatically means that it will be a singleton
1277 * acceptable for merging, so we can do all of this optimistically. But
1278 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1280 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1281 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1282 * is to return an anon_vma that is "complex" due to having gone through
1285 * We also make sure that the two vma's are compatible (adjacent,
1286 * and with the same memory policies). That's all stable, even with just
1287 * a read lock on the mm_sem.
1289 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1291 if (anon_vma_compatible(a, b)) {
1292 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1294 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1301 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1302 * neighbouring vmas for a suitable anon_vma, before it goes off
1303 * to allocate a new anon_vma. It checks because a repetitive
1304 * sequence of mprotects and faults may otherwise lead to distinct
1305 * anon_vmas being allocated, preventing vma merge in subsequent
1308 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1310 struct anon_vma *anon_vma = NULL;
1312 /* Try next first. */
1314 anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next);
1319 /* Try prev next. */
1321 anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma);
1324 * We might reach here with anon_vma == NULL if we can't find
1325 * any reusable anon_vma.
1326 * There's no absolute need to look only at touching neighbours:
1327 * we could search further afield for "compatible" anon_vmas.
1328 * But it would probably just be a waste of time searching,
1329 * or lead to too many vmas hanging off the same anon_vma.
1330 * We're trying to allow mprotect remerging later on,
1331 * not trying to minimize memory used for anon_vmas.
1337 * If a hint addr is less than mmap_min_addr change hint to be as
1338 * low as possible but still greater than mmap_min_addr
1340 static inline unsigned long round_hint_to_min(unsigned long hint)
1343 if (((void *)hint != NULL) &&
1344 (hint < mmap_min_addr))
1345 return PAGE_ALIGN(mmap_min_addr);
1349 static inline int mlock_future_check(struct mm_struct *mm,
1350 unsigned long flags,
1353 unsigned long locked, lock_limit;
1355 /* mlock MCL_FUTURE? */
1356 if (flags & VM_LOCKED) {
1357 locked = len >> PAGE_SHIFT;
1358 locked += mm->locked_vm;
1359 lock_limit = rlimit(RLIMIT_MEMLOCK);
1360 lock_limit >>= PAGE_SHIFT;
1361 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1367 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1369 if (S_ISREG(inode->i_mode))
1370 return MAX_LFS_FILESIZE;
1372 if (S_ISBLK(inode->i_mode))
1373 return MAX_LFS_FILESIZE;
1375 if (S_ISSOCK(inode->i_mode))
1376 return MAX_LFS_FILESIZE;
1378 /* Special "we do even unsigned file positions" case */
1379 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1382 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1386 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1387 unsigned long pgoff, unsigned long len)
1389 u64 maxsize = file_mmap_size_max(file, inode);
1391 if (maxsize && len > maxsize)
1394 if (pgoff > maxsize >> PAGE_SHIFT)
1400 * The caller must write-lock current->mm->mmap_lock.
1402 unsigned long do_mmap(struct file *file, unsigned long addr,
1403 unsigned long len, unsigned long prot,
1404 unsigned long flags, unsigned long pgoff,
1405 unsigned long *populate, struct list_head *uf)
1407 struct mm_struct *mm = current->mm;
1408 vm_flags_t vm_flags;
1417 * Does the application expect PROT_READ to imply PROT_EXEC?
1419 * (the exception is when the underlying filesystem is noexec
1420 * mounted, in which case we dont add PROT_EXEC.)
1422 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1423 if (!(file && path_noexec(&file->f_path)))
1426 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1427 if (flags & MAP_FIXED_NOREPLACE)
1430 if (!(flags & MAP_FIXED))
1431 addr = round_hint_to_min(addr);
1433 /* Careful about overflows.. */
1434 len = PAGE_ALIGN(len);
1438 /* offset overflow? */
1439 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1442 /* Too many mappings? */
1443 if (mm->map_count > sysctl_max_map_count)
1446 /* Obtain the address to map to. we verify (or select) it and ensure
1447 * that it represents a valid section of the address space.
1449 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1450 if (IS_ERR_VALUE(addr))
1453 if (flags & MAP_FIXED_NOREPLACE) {
1454 struct vm_area_struct *vma = find_vma(mm, addr);
1456 if (vma && vma->vm_start < addr + len)
1460 if (prot == PROT_EXEC) {
1461 pkey = execute_only_pkey(mm);
1466 /* Do simple checking here so the lower-level routines won't have
1467 * to. we assume access permissions have been handled by the open
1468 * of the memory object, so we don't do any here.
1470 vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1471 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1473 if (flags & MAP_LOCKED)
1474 if (!can_do_mlock())
1477 if (mlock_future_check(mm, vm_flags, len))
1481 struct inode *inode = file_inode(file);
1482 unsigned long flags_mask;
1484 if (!file_mmap_ok(file, inode, pgoff, len))
1487 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1489 switch (flags & MAP_TYPE) {
1492 * Force use of MAP_SHARED_VALIDATE with non-legacy
1493 * flags. E.g. MAP_SYNC is dangerous to use with
1494 * MAP_SHARED as you don't know which consistency model
1495 * you will get. We silently ignore unsupported flags
1496 * with MAP_SHARED to preserve backward compatibility.
1498 flags &= LEGACY_MAP_MASK;
1500 case MAP_SHARED_VALIDATE:
1501 if (flags & ~flags_mask)
1503 if (prot & PROT_WRITE) {
1504 if (!(file->f_mode & FMODE_WRITE))
1506 if (IS_SWAPFILE(file->f_mapping->host))
1511 * Make sure we don't allow writing to an append-only
1514 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1518 * Make sure there are no mandatory locks on the file.
1520 if (locks_verify_locked(file))
1523 vm_flags |= VM_SHARED | VM_MAYSHARE;
1524 if (!(file->f_mode & FMODE_WRITE))
1525 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1528 if (!(file->f_mode & FMODE_READ))
1530 if (path_noexec(&file->f_path)) {
1531 if (vm_flags & VM_EXEC)
1533 vm_flags &= ~VM_MAYEXEC;
1536 if (!file->f_op->mmap)
1538 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1546 switch (flags & MAP_TYPE) {
1548 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1554 vm_flags |= VM_SHARED | VM_MAYSHARE;
1558 * Set pgoff according to addr for anon_vma.
1560 pgoff = addr >> PAGE_SHIFT;
1568 * Set 'VM_NORESERVE' if we should not account for the
1569 * memory use of this mapping.
1571 if (flags & MAP_NORESERVE) {
1572 /* We honor MAP_NORESERVE if allowed to overcommit */
1573 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1574 vm_flags |= VM_NORESERVE;
1576 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1577 if (file && is_file_hugepages(file))
1578 vm_flags |= VM_NORESERVE;
1581 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1582 if (!IS_ERR_VALUE(addr) &&
1583 ((vm_flags & VM_LOCKED) ||
1584 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1589 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1590 unsigned long prot, unsigned long flags,
1591 unsigned long fd, unsigned long pgoff)
1593 struct file *file = NULL;
1594 unsigned long retval;
1596 if (!(flags & MAP_ANONYMOUS)) {
1597 audit_mmap_fd(fd, flags);
1601 if (is_file_hugepages(file)) {
1602 len = ALIGN(len, huge_page_size(hstate_file(file)));
1603 } else if (unlikely(flags & MAP_HUGETLB)) {
1607 } else if (flags & MAP_HUGETLB) {
1608 struct user_struct *user = NULL;
1611 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1615 len = ALIGN(len, huge_page_size(hs));
1617 * VM_NORESERVE is used because the reservations will be
1618 * taken when vm_ops->mmap() is called
1619 * A dummy user value is used because we are not locking
1620 * memory so no accounting is necessary
1622 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1624 &user, HUGETLB_ANONHUGE_INODE,
1625 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1627 return PTR_ERR(file);
1630 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1632 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1639 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1640 unsigned long, prot, unsigned long, flags,
1641 unsigned long, fd, unsigned long, pgoff)
1643 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1646 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1647 struct mmap_arg_struct {
1651 unsigned long flags;
1653 unsigned long offset;
1656 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1658 struct mmap_arg_struct a;
1660 if (copy_from_user(&a, arg, sizeof(a)))
1662 if (offset_in_page(a.offset))
1665 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1666 a.offset >> PAGE_SHIFT);
1668 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1671 * Some shared mappings will want the pages marked read-only
1672 * to track write events. If so, we'll downgrade vm_page_prot
1673 * to the private version (using protection_map[] without the
1676 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1678 vm_flags_t vm_flags = vma->vm_flags;
1679 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1681 /* If it was private or non-writable, the write bit is already clear */
1682 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1685 /* The backer wishes to know when pages are first written to? */
1686 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1689 /* The open routine did something to the protections that pgprot_modify
1690 * won't preserve? */
1691 if (pgprot_val(vm_page_prot) !=
1692 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1695 /* Do we need to track softdirty? */
1696 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1699 /* Specialty mapping? */
1700 if (vm_flags & VM_PFNMAP)
1703 /* Can the mapping track the dirty pages? */
1704 return vma->vm_file && vma->vm_file->f_mapping &&
1705 mapping_can_writeback(vma->vm_file->f_mapping);
1709 * We account for memory if it's a private writeable mapping,
1710 * not hugepages and VM_NORESERVE wasn't set.
1712 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1715 * hugetlb has its own accounting separate from the core VM
1716 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1718 if (file && is_file_hugepages(file))
1721 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1724 unsigned long mmap_region(struct file *file, unsigned long addr,
1725 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1726 struct list_head *uf)
1728 struct mm_struct *mm = current->mm;
1729 struct vm_area_struct *vma, *prev, *merge;
1731 struct rb_node **rb_link, *rb_parent;
1732 unsigned long charged = 0;
1734 /* Check against address space limit. */
1735 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1736 unsigned long nr_pages;
1739 * MAP_FIXED may remove pages of mappings that intersects with
1740 * requested mapping. Account for the pages it would unmap.
1742 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1744 if (!may_expand_vm(mm, vm_flags,
1745 (len >> PAGE_SHIFT) - nr_pages))
1749 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
1750 if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
1753 * Private writable mapping: check memory availability
1755 if (accountable_mapping(file, vm_flags)) {
1756 charged = len >> PAGE_SHIFT;
1757 if (security_vm_enough_memory_mm(mm, charged))
1759 vm_flags |= VM_ACCOUNT;
1763 * Can we just expand an old mapping?
1765 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1766 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1771 * Determine the object being mapped and call the appropriate
1772 * specific mapper. the address has already been validated, but
1773 * not unmapped, but the maps are removed from the list.
1775 vma = vm_area_alloc(mm);
1781 vma->vm_start = addr;
1782 vma->vm_end = addr + len;
1783 vma->vm_flags = vm_flags;
1784 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1785 vma->vm_pgoff = pgoff;
1788 if (vm_flags & VM_DENYWRITE) {
1789 error = deny_write_access(file);
1793 if (vm_flags & VM_SHARED) {
1794 error = mapping_map_writable(file->f_mapping);
1796 goto allow_write_and_free_vma;
1799 /* ->mmap() can change vma->vm_file, but must guarantee that
1800 * vma_link() below can deny write-access if VM_DENYWRITE is set
1801 * and map writably if VM_SHARED is set. This usually means the
1802 * new file must not have been exposed to user-space, yet.
1804 vma->vm_file = get_file(file);
1805 error = call_mmap(file, vma);
1807 goto unmap_and_free_vma;
1809 /* Can addr have changed??
1811 * Answer: Yes, several device drivers can do it in their
1812 * f_op->mmap method. -DaveM
1813 * Bug: If addr is changed, prev, rb_link, rb_parent should
1814 * be updated for vma_link()
1816 WARN_ON_ONCE(addr != vma->vm_start);
1818 addr = vma->vm_start;
1820 /* If vm_flags changed after call_mmap(), we should try merge vma again
1821 * as we may succeed this time.
1823 if (unlikely(vm_flags != vma->vm_flags && prev)) {
1824 merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
1825 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX);
1827 /* ->mmap() can change vma->vm_file and fput the original file. So
1828 * fput the vma->vm_file here or we would add an extra fput for file
1829 * and cause general protection fault ultimately.
1834 /* Update vm_flags to pick up the change. */
1835 vm_flags = vma->vm_flags;
1836 goto unmap_writable;
1840 vm_flags = vma->vm_flags;
1841 } else if (vm_flags & VM_SHARED) {
1842 error = shmem_zero_setup(vma);
1846 vma_set_anonymous(vma);
1849 /* Allow architectures to sanity-check the vm_flags */
1850 if (!arch_validate_flags(vma->vm_flags)) {
1853 goto unmap_and_free_vma;
1858 vma_link(mm, vma, prev, rb_link, rb_parent);
1859 /* Once vma denies write, undo our temporary denial count */
1862 if (vm_flags & VM_SHARED)
1863 mapping_unmap_writable(file->f_mapping);
1864 if (vm_flags & VM_DENYWRITE)
1865 allow_write_access(file);
1867 file = vma->vm_file;
1869 perf_event_mmap(vma);
1871 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1872 if (vm_flags & VM_LOCKED) {
1873 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1874 is_vm_hugetlb_page(vma) ||
1875 vma == get_gate_vma(current->mm))
1876 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1878 mm->locked_vm += (len >> PAGE_SHIFT);
1885 * New (or expanded) vma always get soft dirty status.
1886 * Otherwise user-space soft-dirty page tracker won't
1887 * be able to distinguish situation when vma area unmapped,
1888 * then new mapped in-place (which must be aimed as
1889 * a completely new data area).
1891 vma->vm_flags |= VM_SOFTDIRTY;
1893 vma_set_page_prot(vma);
1899 vma->vm_file = NULL;
1901 /* Undo any partial mapping done by a device driver. */
1902 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1904 if (vm_flags & VM_SHARED)
1905 mapping_unmap_writable(file->f_mapping);
1906 allow_write_and_free_vma:
1907 if (vm_flags & VM_DENYWRITE)
1908 allow_write_access(file);
1913 vm_unacct_memory(charged);
1917 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1920 * We implement the search by looking for an rbtree node that
1921 * immediately follows a suitable gap. That is,
1922 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1923 * - gap_end = vma->vm_start >= info->low_limit + length;
1924 * - gap_end - gap_start >= length
1927 struct mm_struct *mm = current->mm;
1928 struct vm_area_struct *vma;
1929 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1931 /* Adjust search length to account for worst case alignment overhead */
1932 length = info->length + info->align_mask;
1933 if (length < info->length)
1936 /* Adjust search limits by the desired length */
1937 if (info->high_limit < length)
1939 high_limit = info->high_limit - length;
1941 if (info->low_limit > high_limit)
1943 low_limit = info->low_limit + length;
1945 /* Check if rbtree root looks promising */
1946 if (RB_EMPTY_ROOT(&mm->mm_rb))
1948 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1949 if (vma->rb_subtree_gap < length)
1953 /* Visit left subtree if it looks promising */
1954 gap_end = vm_start_gap(vma);
1955 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1956 struct vm_area_struct *left =
1957 rb_entry(vma->vm_rb.rb_left,
1958 struct vm_area_struct, vm_rb);
1959 if (left->rb_subtree_gap >= length) {
1965 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1967 /* Check if current node has a suitable gap */
1968 if (gap_start > high_limit)
1970 if (gap_end >= low_limit &&
1971 gap_end > gap_start && gap_end - gap_start >= length)
1974 /* Visit right subtree if it looks promising */
1975 if (vma->vm_rb.rb_right) {
1976 struct vm_area_struct *right =
1977 rb_entry(vma->vm_rb.rb_right,
1978 struct vm_area_struct, vm_rb);
1979 if (right->rb_subtree_gap >= length) {
1985 /* Go back up the rbtree to find next candidate node */
1987 struct rb_node *prev = &vma->vm_rb;
1988 if (!rb_parent(prev))
1990 vma = rb_entry(rb_parent(prev),
1991 struct vm_area_struct, vm_rb);
1992 if (prev == vma->vm_rb.rb_left) {
1993 gap_start = vm_end_gap(vma->vm_prev);
1994 gap_end = vm_start_gap(vma);
2001 /* Check highest gap, which does not precede any rbtree node */
2002 gap_start = mm->highest_vm_end;
2003 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
2004 if (gap_start > high_limit)
2008 /* We found a suitable gap. Clip it with the original low_limit. */
2009 if (gap_start < info->low_limit)
2010 gap_start = info->low_limit;
2012 /* Adjust gap address to the desired alignment */
2013 gap_start += (info->align_offset - gap_start) & info->align_mask;
2015 VM_BUG_ON(gap_start + info->length > info->high_limit);
2016 VM_BUG_ON(gap_start + info->length > gap_end);
2020 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
2022 struct mm_struct *mm = current->mm;
2023 struct vm_area_struct *vma;
2024 unsigned long length, low_limit, high_limit, gap_start, gap_end;
2026 /* Adjust search length to account for worst case alignment overhead */
2027 length = info->length + info->align_mask;
2028 if (length < info->length)
2032 * Adjust search limits by the desired length.
2033 * See implementation comment at top of unmapped_area().
2035 gap_end = info->high_limit;
2036 if (gap_end < length)
2038 high_limit = gap_end - length;
2040 if (info->low_limit > high_limit)
2042 low_limit = info->low_limit + length;
2044 /* Check highest gap, which does not precede any rbtree node */
2045 gap_start = mm->highest_vm_end;
2046 if (gap_start <= high_limit)
2049 /* Check if rbtree root looks promising */
2050 if (RB_EMPTY_ROOT(&mm->mm_rb))
2052 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
2053 if (vma->rb_subtree_gap < length)
2057 /* Visit right subtree if it looks promising */
2058 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
2059 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
2060 struct vm_area_struct *right =
2061 rb_entry(vma->vm_rb.rb_right,
2062 struct vm_area_struct, vm_rb);
2063 if (right->rb_subtree_gap >= length) {
2070 /* Check if current node has a suitable gap */
2071 gap_end = vm_start_gap(vma);
2072 if (gap_end < low_limit)
2074 if (gap_start <= high_limit &&
2075 gap_end > gap_start && gap_end - gap_start >= length)
2078 /* Visit left subtree if it looks promising */
2079 if (vma->vm_rb.rb_left) {
2080 struct vm_area_struct *left =
2081 rb_entry(vma->vm_rb.rb_left,
2082 struct vm_area_struct, vm_rb);
2083 if (left->rb_subtree_gap >= length) {
2089 /* Go back up the rbtree to find next candidate node */
2091 struct rb_node *prev = &vma->vm_rb;
2092 if (!rb_parent(prev))
2094 vma = rb_entry(rb_parent(prev),
2095 struct vm_area_struct, vm_rb);
2096 if (prev == vma->vm_rb.rb_right) {
2097 gap_start = vma->vm_prev ?
2098 vm_end_gap(vma->vm_prev) : 0;
2105 /* We found a suitable gap. Clip it with the original high_limit. */
2106 if (gap_end > info->high_limit)
2107 gap_end = info->high_limit;
2110 /* Compute highest gap address at the desired alignment */
2111 gap_end -= info->length;
2112 gap_end -= (gap_end - info->align_offset) & info->align_mask;
2114 VM_BUG_ON(gap_end < info->low_limit);
2115 VM_BUG_ON(gap_end < gap_start);
2120 * Search for an unmapped address range.
2122 * We are looking for a range that:
2123 * - does not intersect with any VMA;
2124 * - is contained within the [low_limit, high_limit) interval;
2125 * - is at least the desired size.
2126 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
2128 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
2132 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
2133 addr = unmapped_area_topdown(info);
2135 addr = unmapped_area(info);
2137 trace_vm_unmapped_area(addr, info);
2141 #ifndef arch_get_mmap_end
2142 #define arch_get_mmap_end(addr) (TASK_SIZE)
2145 #ifndef arch_get_mmap_base
2146 #define arch_get_mmap_base(addr, base) (base)
2149 /* Get an address range which is currently unmapped.
2150 * For shmat() with addr=0.
2152 * Ugly calling convention alert:
2153 * Return value with the low bits set means error value,
2155 * if (ret & ~PAGE_MASK)
2158 * This function "knows" that -ENOMEM has the bits set.
2160 #ifndef HAVE_ARCH_UNMAPPED_AREA
2162 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2163 unsigned long len, unsigned long pgoff, unsigned long flags)
2165 struct mm_struct *mm = current->mm;
2166 struct vm_area_struct *vma, *prev;
2167 struct vm_unmapped_area_info info;
2168 const unsigned long mmap_end = arch_get_mmap_end(addr);
2170 if (len > mmap_end - mmap_min_addr)
2173 if (flags & MAP_FIXED)
2177 addr = PAGE_ALIGN(addr);
2178 vma = find_vma_prev(mm, addr, &prev);
2179 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2180 (!vma || addr + len <= vm_start_gap(vma)) &&
2181 (!prev || addr >= vm_end_gap(prev)))
2187 info.low_limit = mm->mmap_base;
2188 info.high_limit = mmap_end;
2189 info.align_mask = 0;
2190 info.align_offset = 0;
2191 return vm_unmapped_area(&info);
2196 * This mmap-allocator allocates new areas top-down from below the
2197 * stack's low limit (the base):
2199 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2201 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2202 unsigned long len, unsigned long pgoff,
2203 unsigned long flags)
2205 struct vm_area_struct *vma, *prev;
2206 struct mm_struct *mm = current->mm;
2207 struct vm_unmapped_area_info info;
2208 const unsigned long mmap_end = arch_get_mmap_end(addr);
2210 /* requested length too big for entire address space */
2211 if (len > mmap_end - mmap_min_addr)
2214 if (flags & MAP_FIXED)
2217 /* requesting a specific address */
2219 addr = PAGE_ALIGN(addr);
2220 vma = find_vma_prev(mm, addr, &prev);
2221 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2222 (!vma || addr + len <= vm_start_gap(vma)) &&
2223 (!prev || addr >= vm_end_gap(prev)))
2227 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2229 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2230 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2231 info.align_mask = 0;
2232 info.align_offset = 0;
2233 addr = vm_unmapped_area(&info);
2236 * A failed mmap() very likely causes application failure,
2237 * so fall back to the bottom-up function here. This scenario
2238 * can happen with large stack limits and large mmap()
2241 if (offset_in_page(addr)) {
2242 VM_BUG_ON(addr != -ENOMEM);
2244 info.low_limit = TASK_UNMAPPED_BASE;
2245 info.high_limit = mmap_end;
2246 addr = vm_unmapped_area(&info);
2254 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2255 unsigned long pgoff, unsigned long flags)
2257 unsigned long (*get_area)(struct file *, unsigned long,
2258 unsigned long, unsigned long, unsigned long);
2260 unsigned long error = arch_mmap_check(addr, len, flags);
2264 /* Careful about overflows.. */
2265 if (len > TASK_SIZE)
2268 get_area = current->mm->get_unmapped_area;
2270 if (file->f_op->get_unmapped_area)
2271 get_area = file->f_op->get_unmapped_area;
2272 } else if (flags & MAP_SHARED) {
2274 * mmap_region() will call shmem_zero_setup() to create a file,
2275 * so use shmem's get_unmapped_area in case it can be huge.
2276 * do_mmap() will clear pgoff, so match alignment.
2279 get_area = shmem_get_unmapped_area;
2282 addr = get_area(file, addr, len, pgoff, flags);
2283 if (IS_ERR_VALUE(addr))
2286 if (addr > TASK_SIZE - len)
2288 if (offset_in_page(addr))
2291 error = security_mmap_addr(addr);
2292 return error ? error : addr;
2295 EXPORT_SYMBOL(get_unmapped_area);
2297 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2298 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2300 struct rb_node *rb_node;
2301 struct vm_area_struct *vma;
2303 /* Check the cache first. */
2304 vma = vmacache_find(mm, addr);
2308 rb_node = mm->mm_rb.rb_node;
2311 struct vm_area_struct *tmp;
2313 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2315 if (tmp->vm_end > addr) {
2317 if (tmp->vm_start <= addr)
2319 rb_node = rb_node->rb_left;
2321 rb_node = rb_node->rb_right;
2325 vmacache_update(addr, vma);
2329 EXPORT_SYMBOL(find_vma);
2332 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2334 struct vm_area_struct *
2335 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2336 struct vm_area_struct **pprev)
2338 struct vm_area_struct *vma;
2340 vma = find_vma(mm, addr);
2342 *pprev = vma->vm_prev;
2344 struct rb_node *rb_node = rb_last(&mm->mm_rb);
2346 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
2352 * Verify that the stack growth is acceptable and
2353 * update accounting. This is shared with both the
2354 * grow-up and grow-down cases.
2356 static int acct_stack_growth(struct vm_area_struct *vma,
2357 unsigned long size, unsigned long grow)
2359 struct mm_struct *mm = vma->vm_mm;
2360 unsigned long new_start;
2362 /* address space limit tests */
2363 if (!may_expand_vm(mm, vma->vm_flags, grow))
2366 /* Stack limit test */
2367 if (size > rlimit(RLIMIT_STACK))
2370 /* mlock limit tests */
2371 if (vma->vm_flags & VM_LOCKED) {
2372 unsigned long locked;
2373 unsigned long limit;
2374 locked = mm->locked_vm + grow;
2375 limit = rlimit(RLIMIT_MEMLOCK);
2376 limit >>= PAGE_SHIFT;
2377 if (locked > limit && !capable(CAP_IPC_LOCK))
2381 /* Check to ensure the stack will not grow into a hugetlb-only region */
2382 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2384 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2388 * Overcommit.. This must be the final test, as it will
2389 * update security statistics.
2391 if (security_vm_enough_memory_mm(mm, grow))
2397 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2399 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2400 * vma is the last one with address > vma->vm_end. Have to extend vma.
2402 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2404 struct mm_struct *mm = vma->vm_mm;
2405 struct vm_area_struct *next;
2406 unsigned long gap_addr;
2409 if (!(vma->vm_flags & VM_GROWSUP))
2412 /* Guard against exceeding limits of the address space. */
2413 address &= PAGE_MASK;
2414 if (address >= (TASK_SIZE & PAGE_MASK))
2416 address += PAGE_SIZE;
2418 /* Enforce stack_guard_gap */
2419 gap_addr = address + stack_guard_gap;
2421 /* Guard against overflow */
2422 if (gap_addr < address || gap_addr > TASK_SIZE)
2423 gap_addr = TASK_SIZE;
2425 next = vma->vm_next;
2426 if (next && next->vm_start < gap_addr && vma_is_accessible(next)) {
2427 if (!(next->vm_flags & VM_GROWSUP))
2429 /* Check that both stack segments have the same anon_vma? */
2432 /* We must make sure the anon_vma is allocated. */
2433 if (unlikely(anon_vma_prepare(vma)))
2437 * vma->vm_start/vm_end cannot change under us because the caller
2438 * is required to hold the mmap_lock in read mode. We need the
2439 * anon_vma lock to serialize against concurrent expand_stacks.
2441 anon_vma_lock_write(vma->anon_vma);
2443 /* Somebody else might have raced and expanded it already */
2444 if (address > vma->vm_end) {
2445 unsigned long size, grow;
2447 size = address - vma->vm_start;
2448 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2451 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2452 error = acct_stack_growth(vma, size, grow);
2455 * vma_gap_update() doesn't support concurrent
2456 * updates, but we only hold a shared mmap_lock
2457 * lock here, so we need to protect against
2458 * concurrent vma expansions.
2459 * anon_vma_lock_write() doesn't help here, as
2460 * we don't guarantee that all growable vmas
2461 * in a mm share the same root anon vma.
2462 * So, we reuse mm->page_table_lock to guard
2463 * against concurrent vma expansions.
2465 spin_lock(&mm->page_table_lock);
2466 if (vma->vm_flags & VM_LOCKED)
2467 mm->locked_vm += grow;
2468 vm_stat_account(mm, vma->vm_flags, grow);
2469 anon_vma_interval_tree_pre_update_vma(vma);
2470 vma->vm_end = address;
2471 anon_vma_interval_tree_post_update_vma(vma);
2473 vma_gap_update(vma->vm_next);
2475 mm->highest_vm_end = vm_end_gap(vma);
2476 spin_unlock(&mm->page_table_lock);
2478 perf_event_mmap(vma);
2482 anon_vma_unlock_write(vma->anon_vma);
2483 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2487 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2490 * vma is the first one with address < vma->vm_start. Have to extend vma.
2492 int expand_downwards(struct vm_area_struct *vma,
2493 unsigned long address)
2495 struct mm_struct *mm = vma->vm_mm;
2496 struct vm_area_struct *prev;
2499 address &= PAGE_MASK;
2500 if (address < mmap_min_addr)
2503 /* Enforce stack_guard_gap */
2504 prev = vma->vm_prev;
2505 /* Check that both stack segments have the same anon_vma? */
2506 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2507 vma_is_accessible(prev)) {
2508 if (address - prev->vm_end < stack_guard_gap)
2512 /* We must make sure the anon_vma is allocated. */
2513 if (unlikely(anon_vma_prepare(vma)))
2517 * vma->vm_start/vm_end cannot change under us because the caller
2518 * is required to hold the mmap_lock in read mode. We need the
2519 * anon_vma lock to serialize against concurrent expand_stacks.
2521 anon_vma_lock_write(vma->anon_vma);
2523 /* Somebody else might have raced and expanded it already */
2524 if (address < vma->vm_start) {
2525 unsigned long size, grow;
2527 size = vma->vm_end - address;
2528 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2531 if (grow <= vma->vm_pgoff) {
2532 error = acct_stack_growth(vma, size, grow);
2535 * vma_gap_update() doesn't support concurrent
2536 * updates, but we only hold a shared mmap_lock
2537 * lock here, so we need to protect against
2538 * concurrent vma expansions.
2539 * anon_vma_lock_write() doesn't help here, as
2540 * we don't guarantee that all growable vmas
2541 * in a mm share the same root anon vma.
2542 * So, we reuse mm->page_table_lock to guard
2543 * against concurrent vma expansions.
2545 spin_lock(&mm->page_table_lock);
2546 if (vma->vm_flags & VM_LOCKED)
2547 mm->locked_vm += grow;
2548 vm_stat_account(mm, vma->vm_flags, grow);
2549 anon_vma_interval_tree_pre_update_vma(vma);
2550 vma->vm_start = address;
2551 vma->vm_pgoff -= grow;
2552 anon_vma_interval_tree_post_update_vma(vma);
2553 vma_gap_update(vma);
2554 spin_unlock(&mm->page_table_lock);
2556 perf_event_mmap(vma);
2560 anon_vma_unlock_write(vma->anon_vma);
2561 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2566 /* enforced gap between the expanding stack and other mappings. */
2567 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2569 static int __init cmdline_parse_stack_guard_gap(char *p)
2574 val = simple_strtoul(p, &endptr, 10);
2576 stack_guard_gap = val << PAGE_SHIFT;
2580 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2582 #ifdef CONFIG_STACK_GROWSUP
2583 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2585 return expand_upwards(vma, address);
2588 struct vm_area_struct *
2589 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2591 struct vm_area_struct *vma, *prev;
2594 vma = find_vma_prev(mm, addr, &prev);
2595 if (vma && (vma->vm_start <= addr))
2597 /* don't alter vm_end if the coredump is running */
2598 if (!prev || expand_stack(prev, addr))
2600 if (prev->vm_flags & VM_LOCKED)
2601 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2605 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2607 return expand_downwards(vma, address);
2610 struct vm_area_struct *
2611 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2613 struct vm_area_struct *vma;
2614 unsigned long start;
2617 vma = find_vma(mm, addr);
2620 if (vma->vm_start <= addr)
2622 if (!(vma->vm_flags & VM_GROWSDOWN))
2624 start = vma->vm_start;
2625 if (expand_stack(vma, addr))
2627 if (vma->vm_flags & VM_LOCKED)
2628 populate_vma_page_range(vma, addr, start, NULL);
2633 EXPORT_SYMBOL_GPL(find_extend_vma);
2636 * Ok - we have the memory areas we should free on the vma list,
2637 * so release them, and do the vma updates.
2639 * Called with the mm semaphore held.
2641 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2643 unsigned long nr_accounted = 0;
2645 /* Update high watermark before we lower total_vm */
2646 update_hiwater_vm(mm);
2648 long nrpages = vma_pages(vma);
2650 if (vma->vm_flags & VM_ACCOUNT)
2651 nr_accounted += nrpages;
2652 vm_stat_account(mm, vma->vm_flags, -nrpages);
2653 vma = remove_vma(vma);
2655 vm_unacct_memory(nr_accounted);
2660 * Get rid of page table information in the indicated region.
2662 * Called with the mm semaphore held.
2664 static void unmap_region(struct mm_struct *mm,
2665 struct vm_area_struct *vma, struct vm_area_struct *prev,
2666 unsigned long start, unsigned long end)
2668 struct vm_area_struct *next = vma_next(mm, prev);
2669 struct mmu_gather tlb;
2672 tlb_gather_mmu(&tlb, mm);
2673 update_hiwater_rss(mm);
2674 unmap_vmas(&tlb, vma, start, end);
2675 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2676 next ? next->vm_start : USER_PGTABLES_CEILING);
2677 tlb_finish_mmu(&tlb);
2681 * Create a list of vma's touched by the unmap, removing them from the mm's
2682 * vma list as we go..
2685 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2686 struct vm_area_struct *prev, unsigned long end)
2688 struct vm_area_struct **insertion_point;
2689 struct vm_area_struct *tail_vma = NULL;
2691 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2692 vma->vm_prev = NULL;
2694 vma_rb_erase(vma, &mm->mm_rb);
2698 } while (vma && vma->vm_start < end);
2699 *insertion_point = vma;
2701 vma->vm_prev = prev;
2702 vma_gap_update(vma);
2704 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2705 tail_vma->vm_next = NULL;
2707 /* Kill the cache */
2708 vmacache_invalidate(mm);
2711 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2712 * VM_GROWSUP VMA. Such VMAs can change their size under
2713 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2715 if (vma && (vma->vm_flags & VM_GROWSDOWN))
2717 if (prev && (prev->vm_flags & VM_GROWSUP))
2723 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2724 * has already been checked or doesn't make sense to fail.
2726 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2727 unsigned long addr, int new_below)
2729 struct vm_area_struct *new;
2732 if (vma->vm_ops && vma->vm_ops->may_split) {
2733 err = vma->vm_ops->may_split(vma, addr);
2738 new = vm_area_dup(vma);
2745 new->vm_start = addr;
2746 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2749 err = vma_dup_policy(vma, new);
2753 err = anon_vma_clone(new, vma);
2758 get_file(new->vm_file);
2760 if (new->vm_ops && new->vm_ops->open)
2761 new->vm_ops->open(new);
2764 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2765 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2767 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2773 /* Clean everything up if vma_adjust failed. */
2774 if (new->vm_ops && new->vm_ops->close)
2775 new->vm_ops->close(new);
2778 unlink_anon_vmas(new);
2780 mpol_put(vma_policy(new));
2787 * Split a vma into two pieces at address 'addr', a new vma is allocated
2788 * either for the first part or the tail.
2790 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2791 unsigned long addr, int new_below)
2793 if (mm->map_count >= sysctl_max_map_count)
2796 return __split_vma(mm, vma, addr, new_below);
2799 /* Munmap is split into 2 main parts -- this part which finds
2800 * what needs doing, and the areas themselves, which do the
2801 * work. This now handles partial unmappings.
2802 * Jeremy Fitzhardinge <jeremy@goop.org>
2804 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2805 struct list_head *uf, bool downgrade)
2808 struct vm_area_struct *vma, *prev, *last;
2810 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2813 len = PAGE_ALIGN(len);
2819 * arch_unmap() might do unmaps itself. It must be called
2820 * and finish any rbtree manipulation before this code
2821 * runs and also starts to manipulate the rbtree.
2823 arch_unmap(mm, start, end);
2825 /* Find the first overlapping VMA */
2826 vma = find_vma(mm, start);
2829 prev = vma->vm_prev;
2830 /* we have start < vma->vm_end */
2832 /* if it doesn't overlap, we have nothing.. */
2833 if (vma->vm_start >= end)
2837 * If we need to split any vma, do it now to save pain later.
2839 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2840 * unmapped vm_area_struct will remain in use: so lower split_vma
2841 * places tmp vma above, and higher split_vma places tmp vma below.
2843 if (start > vma->vm_start) {
2847 * Make sure that map_count on return from munmap() will
2848 * not exceed its limit; but let map_count go just above
2849 * its limit temporarily, to help free resources as expected.
2851 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2854 error = __split_vma(mm, vma, start, 0);
2860 /* Does it split the last one? */
2861 last = find_vma(mm, end);
2862 if (last && end > last->vm_start) {
2863 int error = __split_vma(mm, last, end, 1);
2867 vma = vma_next(mm, prev);
2871 * If userfaultfd_unmap_prep returns an error the vmas
2872 * will remain splitted, but userland will get a
2873 * highly unexpected error anyway. This is no
2874 * different than the case where the first of the two
2875 * __split_vma fails, but we don't undo the first
2876 * split, despite we could. This is unlikely enough
2877 * failure that it's not worth optimizing it for.
2879 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2885 * unlock any mlock()ed ranges before detaching vmas
2887 if (mm->locked_vm) {
2888 struct vm_area_struct *tmp = vma;
2889 while (tmp && tmp->vm_start < end) {
2890 if (tmp->vm_flags & VM_LOCKED) {
2891 mm->locked_vm -= vma_pages(tmp);
2892 munlock_vma_pages_all(tmp);
2899 /* Detach vmas from rbtree */
2900 if (!detach_vmas_to_be_unmapped(mm, vma, prev, end))
2904 mmap_write_downgrade(mm);
2906 unmap_region(mm, vma, prev, start, end);
2908 /* Fix up all other VM information */
2909 remove_vma_list(mm, vma);
2911 return downgrade ? 1 : 0;
2914 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2915 struct list_head *uf)
2917 return __do_munmap(mm, start, len, uf, false);
2920 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2923 struct mm_struct *mm = current->mm;
2926 if (mmap_write_lock_killable(mm))
2929 ret = __do_munmap(mm, start, len, &uf, downgrade);
2931 * Returning 1 indicates mmap_lock is downgraded.
2932 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2933 * it to 0 before return.
2936 mmap_read_unlock(mm);
2939 mmap_write_unlock(mm);
2941 userfaultfd_unmap_complete(mm, &uf);
2945 int vm_munmap(unsigned long start, size_t len)
2947 return __vm_munmap(start, len, false);
2949 EXPORT_SYMBOL(vm_munmap);
2951 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2953 addr = untagged_addr(addr);
2954 profile_munmap(addr);
2955 return __vm_munmap(addr, len, true);
2960 * Emulation of deprecated remap_file_pages() syscall.
2962 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2963 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2966 struct mm_struct *mm = current->mm;
2967 struct vm_area_struct *vma;
2968 unsigned long populate = 0;
2969 unsigned long ret = -EINVAL;
2972 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2973 current->comm, current->pid);
2977 start = start & PAGE_MASK;
2978 size = size & PAGE_MASK;
2980 if (start + size <= start)
2983 /* Does pgoff wrap? */
2984 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2987 if (mmap_write_lock_killable(mm))
2990 vma = find_vma(mm, start);
2992 if (!vma || !(vma->vm_flags & VM_SHARED))
2995 if (start < vma->vm_start)
2998 if (start + size > vma->vm_end) {
2999 struct vm_area_struct *next;
3001 for (next = vma->vm_next; next; next = next->vm_next) {
3002 /* hole between vmas ? */
3003 if (next->vm_start != next->vm_prev->vm_end)
3006 if (next->vm_file != vma->vm_file)
3009 if (next->vm_flags != vma->vm_flags)
3012 if (start + size <= next->vm_end)
3020 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3021 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3022 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3024 flags &= MAP_NONBLOCK;
3025 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3026 if (vma->vm_flags & VM_LOCKED) {
3027 struct vm_area_struct *tmp;
3028 flags |= MAP_LOCKED;
3030 /* drop PG_Mlocked flag for over-mapped range */
3031 for (tmp = vma; tmp->vm_start >= start + size;
3032 tmp = tmp->vm_next) {
3034 * Split pmd and munlock page on the border
3037 vma_adjust_trans_huge(tmp, start, start + size, 0);
3039 munlock_vma_pages_range(tmp,
3040 max(tmp->vm_start, start),
3041 min(tmp->vm_end, start + size));
3045 file = get_file(vma->vm_file);
3046 ret = do_mmap(vma->vm_file, start, size,
3047 prot, flags, pgoff, &populate, NULL);
3050 mmap_write_unlock(mm);
3052 mm_populate(ret, populate);
3053 if (!IS_ERR_VALUE(ret))
3059 * this is really a simplified "do_mmap". it only handles
3060 * anonymous maps. eventually we may be able to do some
3061 * brk-specific accounting here.
3063 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
3065 struct mm_struct *mm = current->mm;
3066 struct vm_area_struct *vma, *prev;
3067 struct rb_node **rb_link, *rb_parent;
3068 pgoff_t pgoff = addr >> PAGE_SHIFT;
3070 unsigned long mapped_addr;
3072 /* Until we need other flags, refuse anything except VM_EXEC. */
3073 if ((flags & (~VM_EXEC)) != 0)
3075 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3077 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
3078 if (IS_ERR_VALUE(mapped_addr))
3081 error = mlock_future_check(mm, mm->def_flags, len);
3085 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
3086 if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
3089 /* Check against address space limits *after* clearing old maps... */
3090 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3093 if (mm->map_count > sysctl_max_map_count)
3096 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3099 /* Can we just expand an old private anonymous mapping? */
3100 vma = vma_merge(mm, prev, addr, addr + len, flags,
3101 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
3106 * create a vma struct for an anonymous mapping
3108 vma = vm_area_alloc(mm);
3110 vm_unacct_memory(len >> PAGE_SHIFT);
3114 vma_set_anonymous(vma);
3115 vma->vm_start = addr;
3116 vma->vm_end = addr + len;
3117 vma->vm_pgoff = pgoff;
3118 vma->vm_flags = flags;
3119 vma->vm_page_prot = vm_get_page_prot(flags);
3120 vma_link(mm, vma, prev, rb_link, rb_parent);
3122 perf_event_mmap(vma);
3123 mm->total_vm += len >> PAGE_SHIFT;
3124 mm->data_vm += len >> PAGE_SHIFT;
3125 if (flags & VM_LOCKED)
3126 mm->locked_vm += (len >> PAGE_SHIFT);
3127 vma->vm_flags |= VM_SOFTDIRTY;
3131 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3133 struct mm_struct *mm = current->mm;
3139 len = PAGE_ALIGN(request);
3145 if (mmap_write_lock_killable(mm))
3148 ret = do_brk_flags(addr, len, flags, &uf);
3149 populate = ((mm->def_flags & VM_LOCKED) != 0);
3150 mmap_write_unlock(mm);
3151 userfaultfd_unmap_complete(mm, &uf);
3152 if (populate && !ret)
3153 mm_populate(addr, len);
3156 EXPORT_SYMBOL(vm_brk_flags);
3158 int vm_brk(unsigned long addr, unsigned long len)
3160 return vm_brk_flags(addr, len, 0);
3162 EXPORT_SYMBOL(vm_brk);
3164 /* Release all mmaps. */
3165 void exit_mmap(struct mm_struct *mm)
3167 struct mmu_gather tlb;
3168 struct vm_area_struct *vma;
3169 unsigned long nr_accounted = 0;
3171 /* mm's last user has gone, and its about to be pulled down */
3172 mmu_notifier_release(mm);
3174 if (unlikely(mm_is_oom_victim(mm))) {
3176 * Manually reap the mm to free as much memory as possible.
3177 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3178 * this mm from further consideration. Taking mm->mmap_lock for
3179 * write after setting MMF_OOM_SKIP will guarantee that the oom
3180 * reaper will not run on this mm again after mmap_lock is
3183 * Nothing can be holding mm->mmap_lock here and the above call
3184 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3185 * __oom_reap_task_mm() will not block.
3187 * This needs to be done before calling munlock_vma_pages_all(),
3188 * which clears VM_LOCKED, otherwise the oom reaper cannot
3191 (void)__oom_reap_task_mm(mm);
3193 set_bit(MMF_OOM_SKIP, &mm->flags);
3194 mmap_write_lock(mm);
3195 mmap_write_unlock(mm);
3198 if (mm->locked_vm) {
3201 if (vma->vm_flags & VM_LOCKED)
3202 munlock_vma_pages_all(vma);
3210 if (!vma) /* Can happen if dup_mmap() received an OOM */
3215 tlb_gather_mmu_fullmm(&tlb, mm);
3216 /* update_hiwater_rss(mm) here? but nobody should be looking */
3217 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3218 unmap_vmas(&tlb, vma, 0, -1);
3219 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3220 tlb_finish_mmu(&tlb);
3223 * Walk the list again, actually closing and freeing it,
3224 * with preemption enabled, without holding any MM locks.
3227 if (vma->vm_flags & VM_ACCOUNT)
3228 nr_accounted += vma_pages(vma);
3229 vma = remove_vma(vma);
3232 vm_unacct_memory(nr_accounted);
3235 /* Insert vm structure into process list sorted by address
3236 * and into the inode's i_mmap tree. If vm_file is non-NULL
3237 * then i_mmap_rwsem is taken here.
3239 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3241 struct vm_area_struct *prev;
3242 struct rb_node **rb_link, *rb_parent;
3244 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3245 &prev, &rb_link, &rb_parent))
3247 if ((vma->vm_flags & VM_ACCOUNT) &&
3248 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3252 * The vm_pgoff of a purely anonymous vma should be irrelevant
3253 * until its first write fault, when page's anon_vma and index
3254 * are set. But now set the vm_pgoff it will almost certainly
3255 * end up with (unless mremap moves it elsewhere before that
3256 * first wfault), so /proc/pid/maps tells a consistent story.
3258 * By setting it to reflect the virtual start address of the
3259 * vma, merges and splits can happen in a seamless way, just
3260 * using the existing file pgoff checks and manipulations.
3261 * Similarly in do_mmap and in do_brk_flags.
3263 if (vma_is_anonymous(vma)) {
3264 BUG_ON(vma->anon_vma);
3265 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3268 vma_link(mm, vma, prev, rb_link, rb_parent);
3273 * Copy the vma structure to a new location in the same mm,
3274 * prior to moving page table entries, to effect an mremap move.
3276 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3277 unsigned long addr, unsigned long len, pgoff_t pgoff,
3278 bool *need_rmap_locks)
3280 struct vm_area_struct *vma = *vmap;
3281 unsigned long vma_start = vma->vm_start;
3282 struct mm_struct *mm = vma->vm_mm;
3283 struct vm_area_struct *new_vma, *prev;
3284 struct rb_node **rb_link, *rb_parent;
3285 bool faulted_in_anon_vma = true;
3288 * If anonymous vma has not yet been faulted, update new pgoff
3289 * to match new location, to increase its chance of merging.
3291 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3292 pgoff = addr >> PAGE_SHIFT;
3293 faulted_in_anon_vma = false;
3296 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3297 return NULL; /* should never get here */
3298 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3299 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3300 vma->vm_userfaultfd_ctx);
3303 * Source vma may have been merged into new_vma
3305 if (unlikely(vma_start >= new_vma->vm_start &&
3306 vma_start < new_vma->vm_end)) {
3308 * The only way we can get a vma_merge with
3309 * self during an mremap is if the vma hasn't
3310 * been faulted in yet and we were allowed to
3311 * reset the dst vma->vm_pgoff to the
3312 * destination address of the mremap to allow
3313 * the merge to happen. mremap must change the
3314 * vm_pgoff linearity between src and dst vmas
3315 * (in turn preventing a vma_merge) to be
3316 * safe. It is only safe to keep the vm_pgoff
3317 * linear if there are no pages mapped yet.
3319 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3320 *vmap = vma = new_vma;
3322 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3324 new_vma = vm_area_dup(vma);
3327 new_vma->vm_start = addr;
3328 new_vma->vm_end = addr + len;
3329 new_vma->vm_pgoff = pgoff;
3330 if (vma_dup_policy(vma, new_vma))
3332 if (anon_vma_clone(new_vma, vma))
3333 goto out_free_mempol;
3334 if (new_vma->vm_file)
3335 get_file(new_vma->vm_file);
3336 if (new_vma->vm_ops && new_vma->vm_ops->open)
3337 new_vma->vm_ops->open(new_vma);
3338 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3339 *need_rmap_locks = false;
3344 mpol_put(vma_policy(new_vma));
3346 vm_area_free(new_vma);
3352 * Return true if the calling process may expand its vm space by the passed
3355 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3357 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3360 if (is_data_mapping(flags) &&
3361 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3362 /* Workaround for Valgrind */
3363 if (rlimit(RLIMIT_DATA) == 0 &&
3364 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3367 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3368 current->comm, current->pid,
3369 (mm->data_vm + npages) << PAGE_SHIFT,
3370 rlimit(RLIMIT_DATA),
3371 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3373 if (!ignore_rlimit_data)
3380 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3382 mm->total_vm += npages;
3384 if (is_exec_mapping(flags))
3385 mm->exec_vm += npages;
3386 else if (is_stack_mapping(flags))
3387 mm->stack_vm += npages;
3388 else if (is_data_mapping(flags))
3389 mm->data_vm += npages;
3392 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3395 * Having a close hook prevents vma merging regardless of flags.
3397 static void special_mapping_close(struct vm_area_struct *vma)
3401 static const char *special_mapping_name(struct vm_area_struct *vma)
3403 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3406 static int special_mapping_mremap(struct vm_area_struct *new_vma,
3407 unsigned long flags)
3409 struct vm_special_mapping *sm = new_vma->vm_private_data;
3411 if (flags & MREMAP_DONTUNMAP)
3414 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3418 return sm->mremap(sm, new_vma);
3423 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3426 * Forbid splitting special mappings - kernel has expectations over
3427 * the number of pages in mapping. Together with VM_DONTEXPAND
3428 * the size of vma should stay the same over the special mapping's
3434 static const struct vm_operations_struct special_mapping_vmops = {
3435 .close = special_mapping_close,
3436 .fault = special_mapping_fault,
3437 .mremap = special_mapping_mremap,
3438 .name = special_mapping_name,
3439 /* vDSO code relies that VVAR can't be accessed remotely */
3441 .may_split = special_mapping_split,
3444 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3445 .close = special_mapping_close,
3446 .fault = special_mapping_fault,
3449 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3451 struct vm_area_struct *vma = vmf->vma;
3453 struct page **pages;
3455 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3456 pages = vma->vm_private_data;
3458 struct vm_special_mapping *sm = vma->vm_private_data;
3461 return sm->fault(sm, vmf->vma, vmf);
3466 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3470 struct page *page = *pages;
3476 return VM_FAULT_SIGBUS;
3479 static struct vm_area_struct *__install_special_mapping(
3480 struct mm_struct *mm,
3481 unsigned long addr, unsigned long len,
3482 unsigned long vm_flags, void *priv,
3483 const struct vm_operations_struct *ops)
3486 struct vm_area_struct *vma;
3488 vma = vm_area_alloc(mm);
3489 if (unlikely(vma == NULL))
3490 return ERR_PTR(-ENOMEM);
3492 vma->vm_start = addr;
3493 vma->vm_end = addr + len;
3495 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3496 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3499 vma->vm_private_data = priv;
3501 ret = insert_vm_struct(mm, vma);
3505 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3507 perf_event_mmap(vma);
3513 return ERR_PTR(ret);
3516 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3517 const struct vm_special_mapping *sm)
3519 return vma->vm_private_data == sm &&
3520 (vma->vm_ops == &special_mapping_vmops ||
3521 vma->vm_ops == &legacy_special_mapping_vmops);
3525 * Called with mm->mmap_lock held for writing.
3526 * Insert a new vma covering the given region, with the given flags.
3527 * Its pages are supplied by the given array of struct page *.
3528 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3529 * The region past the last page supplied will always produce SIGBUS.
3530 * The array pointer and the pages it points to are assumed to stay alive
3531 * for as long as this mapping might exist.
3533 struct vm_area_struct *_install_special_mapping(
3534 struct mm_struct *mm,
3535 unsigned long addr, unsigned long len,
3536 unsigned long vm_flags, const struct vm_special_mapping *spec)
3538 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3539 &special_mapping_vmops);
3542 int install_special_mapping(struct mm_struct *mm,
3543 unsigned long addr, unsigned long len,
3544 unsigned long vm_flags, struct page **pages)
3546 struct vm_area_struct *vma = __install_special_mapping(
3547 mm, addr, len, vm_flags, (void *)pages,
3548 &legacy_special_mapping_vmops);
3550 return PTR_ERR_OR_ZERO(vma);
3553 static DEFINE_MUTEX(mm_all_locks_mutex);
3555 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3557 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3559 * The LSB of head.next can't change from under us
3560 * because we hold the mm_all_locks_mutex.
3562 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3564 * We can safely modify head.next after taking the
3565 * anon_vma->root->rwsem. If some other vma in this mm shares
3566 * the same anon_vma we won't take it again.
3568 * No need of atomic instructions here, head.next
3569 * can't change from under us thanks to the
3570 * anon_vma->root->rwsem.
3572 if (__test_and_set_bit(0, (unsigned long *)
3573 &anon_vma->root->rb_root.rb_root.rb_node))
3578 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3580 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3582 * AS_MM_ALL_LOCKS can't change from under us because
3583 * we hold the mm_all_locks_mutex.
3585 * Operations on ->flags have to be atomic because
3586 * even if AS_MM_ALL_LOCKS is stable thanks to the
3587 * mm_all_locks_mutex, there may be other cpus
3588 * changing other bitflags in parallel to us.
3590 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3592 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3597 * This operation locks against the VM for all pte/vma/mm related
3598 * operations that could ever happen on a certain mm. This includes
3599 * vmtruncate, try_to_unmap, and all page faults.
3601 * The caller must take the mmap_lock in write mode before calling
3602 * mm_take_all_locks(). The caller isn't allowed to release the
3603 * mmap_lock until mm_drop_all_locks() returns.
3605 * mmap_lock in write mode is required in order to block all operations
3606 * that could modify pagetables and free pages without need of
3607 * altering the vma layout. It's also needed in write mode to avoid new
3608 * anon_vmas to be associated with existing vmas.
3610 * A single task can't take more than one mm_take_all_locks() in a row
3611 * or it would deadlock.
3613 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3614 * mapping->flags avoid to take the same lock twice, if more than one
3615 * vma in this mm is backed by the same anon_vma or address_space.
3617 * We take locks in following order, accordingly to comment at beginning
3619 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3621 * - all i_mmap_rwsem locks;
3622 * - all anon_vma->rwseml
3624 * We can take all locks within these types randomly because the VM code
3625 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3626 * mm_all_locks_mutex.
3628 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3629 * that may have to take thousand of locks.
3631 * mm_take_all_locks() can fail if it's interrupted by signals.
3633 int mm_take_all_locks(struct mm_struct *mm)
3635 struct vm_area_struct *vma;
3636 struct anon_vma_chain *avc;
3638 BUG_ON(mmap_read_trylock(mm));
3640 mutex_lock(&mm_all_locks_mutex);
3642 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3643 if (signal_pending(current))
3645 if (vma->vm_file && vma->vm_file->f_mapping &&
3646 is_vm_hugetlb_page(vma))
3647 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3650 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3651 if (signal_pending(current))
3653 if (vma->vm_file && vma->vm_file->f_mapping &&
3654 !is_vm_hugetlb_page(vma))
3655 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3658 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3659 if (signal_pending(current))
3662 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3663 vm_lock_anon_vma(mm, avc->anon_vma);
3669 mm_drop_all_locks(mm);
3673 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3675 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3677 * The LSB of head.next can't change to 0 from under
3678 * us because we hold the mm_all_locks_mutex.
3680 * We must however clear the bitflag before unlocking
3681 * the vma so the users using the anon_vma->rb_root will
3682 * never see our bitflag.
3684 * No need of atomic instructions here, head.next
3685 * can't change from under us until we release the
3686 * anon_vma->root->rwsem.
3688 if (!__test_and_clear_bit(0, (unsigned long *)
3689 &anon_vma->root->rb_root.rb_root.rb_node))
3691 anon_vma_unlock_write(anon_vma);
3695 static void vm_unlock_mapping(struct address_space *mapping)
3697 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3699 * AS_MM_ALL_LOCKS can't change to 0 from under us
3700 * because we hold the mm_all_locks_mutex.
3702 i_mmap_unlock_write(mapping);
3703 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3710 * The mmap_lock cannot be released by the caller until
3711 * mm_drop_all_locks() returns.
3713 void mm_drop_all_locks(struct mm_struct *mm)
3715 struct vm_area_struct *vma;
3716 struct anon_vma_chain *avc;
3718 BUG_ON(mmap_read_trylock(mm));
3719 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3721 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3723 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3724 vm_unlock_anon_vma(avc->anon_vma);
3725 if (vma->vm_file && vma->vm_file->f_mapping)
3726 vm_unlock_mapping(vma->vm_file->f_mapping);
3729 mutex_unlock(&mm_all_locks_mutex);
3733 * initialise the percpu counter for VM
3735 void __init mmap_init(void)
3739 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3744 * Initialise sysctl_user_reserve_kbytes.
3746 * This is intended to prevent a user from starting a single memory hogging
3747 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3750 * The default value is min(3% of free memory, 128MB)
3751 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3753 static int init_user_reserve(void)
3755 unsigned long free_kbytes;
3757 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3759 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3762 subsys_initcall(init_user_reserve);
3765 * Initialise sysctl_admin_reserve_kbytes.
3767 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3768 * to log in and kill a memory hogging process.
3770 * Systems with more than 256MB will reserve 8MB, enough to recover
3771 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3772 * only reserve 3% of free pages by default.
3774 static int init_admin_reserve(void)
3776 unsigned long free_kbytes;
3778 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3780 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3783 subsys_initcall(init_admin_reserve);
3786 * Reinititalise user and admin reserves if memory is added or removed.
3788 * The default user reserve max is 128MB, and the default max for the
3789 * admin reserve is 8MB. These are usually, but not always, enough to
3790 * enable recovery from a memory hogging process using login/sshd, a shell,
3791 * and tools like top. It may make sense to increase or even disable the
3792 * reserve depending on the existence of swap or variations in the recovery
3793 * tools. So, the admin may have changed them.
3795 * If memory is added and the reserves have been eliminated or increased above
3796 * the default max, then we'll trust the admin.
3798 * If memory is removed and there isn't enough free memory, then we
3799 * need to reset the reserves.
3801 * Otherwise keep the reserve set by the admin.
3803 static int reserve_mem_notifier(struct notifier_block *nb,
3804 unsigned long action, void *data)
3806 unsigned long tmp, free_kbytes;
3810 /* Default max is 128MB. Leave alone if modified by operator. */
3811 tmp = sysctl_user_reserve_kbytes;
3812 if (0 < tmp && tmp < (1UL << 17))
3813 init_user_reserve();
3815 /* Default max is 8MB. Leave alone if modified by operator. */
3816 tmp = sysctl_admin_reserve_kbytes;
3817 if (0 < tmp && tmp < (1UL << 13))
3818 init_admin_reserve();
3822 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3824 if (sysctl_user_reserve_kbytes > free_kbytes) {
3825 init_user_reserve();
3826 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3827 sysctl_user_reserve_kbytes);
3830 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3831 init_admin_reserve();
3832 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3833 sysctl_admin_reserve_kbytes);
3842 static struct notifier_block reserve_mem_nb = {
3843 .notifier_call = reserve_mem_notifier,
3846 static int __meminit init_reserve_notifier(void)
3848 if (register_hotmemory_notifier(&reserve_mem_nb))
3849 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3853 subsys_initcall(init_reserve_notifier);