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 * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
98 * MAP_PRIVATE (with Enhanced PAN supported):
103 pgprot_t protection_map[16] __ro_after_init = {
104 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
105 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
108 #ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
109 static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
115 pgprot_t vm_get_page_prot(unsigned long vm_flags)
117 pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
118 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
119 pgprot_val(arch_vm_get_page_prot(vm_flags)));
121 return arch_filter_pgprot(ret);
123 EXPORT_SYMBOL(vm_get_page_prot);
125 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
127 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
130 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
131 void vma_set_page_prot(struct vm_area_struct *vma)
133 unsigned long vm_flags = vma->vm_flags;
134 pgprot_t vm_page_prot;
136 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
137 if (vma_wants_writenotify(vma, vm_page_prot)) {
138 vm_flags &= ~VM_SHARED;
139 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
141 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
142 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
146 * Requires inode->i_mapping->i_mmap_rwsem
148 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
149 struct file *file, struct address_space *mapping)
151 if (vma->vm_flags & VM_SHARED)
152 mapping_unmap_writable(mapping);
154 flush_dcache_mmap_lock(mapping);
155 vma_interval_tree_remove(vma, &mapping->i_mmap);
156 flush_dcache_mmap_unlock(mapping);
160 * Unlink a file-based vm structure from its interval tree, to hide
161 * vma from rmap and vmtruncate before freeing its page tables.
163 void unlink_file_vma(struct vm_area_struct *vma)
165 struct file *file = vma->vm_file;
168 struct address_space *mapping = file->f_mapping;
169 i_mmap_lock_write(mapping);
170 __remove_shared_vm_struct(vma, file, mapping);
171 i_mmap_unlock_write(mapping);
176 * Close a vm structure and free it, returning the next.
178 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
180 struct vm_area_struct *next = vma->vm_next;
183 if (vma->vm_ops && vma->vm_ops->close)
184 vma->vm_ops->close(vma);
187 mpol_put(vma_policy(vma));
192 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags,
193 struct list_head *uf);
194 SYSCALL_DEFINE1(brk, unsigned long, brk)
196 unsigned long newbrk, oldbrk, origbrk;
197 struct mm_struct *mm = current->mm;
198 struct vm_area_struct *next;
199 unsigned long min_brk;
201 bool downgraded = false;
204 if (mmap_write_lock_killable(mm))
209 #ifdef CONFIG_COMPAT_BRK
211 * CONFIG_COMPAT_BRK can still be overridden by setting
212 * randomize_va_space to 2, which will still cause mm->start_brk
213 * to be arbitrarily shifted
215 if (current->brk_randomized)
216 min_brk = mm->start_brk;
218 min_brk = mm->end_data;
220 min_brk = mm->start_brk;
226 * Check against rlimit here. If this check is done later after the test
227 * of oldbrk with newbrk then it can escape the test and let the data
228 * segment grow beyond its set limit the in case where the limit is
229 * not page aligned -Ram Gupta
231 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
232 mm->end_data, mm->start_data))
235 newbrk = PAGE_ALIGN(brk);
236 oldbrk = PAGE_ALIGN(mm->brk);
237 if (oldbrk == newbrk) {
243 * Always allow shrinking brk.
244 * __do_munmap() may downgrade mmap_lock to read.
246 if (brk <= mm->brk) {
250 * mm->brk must to be protected by write mmap_lock so update it
251 * before downgrading mmap_lock. When __do_munmap() fails,
252 * mm->brk will be restored from origbrk.
255 ret = __do_munmap(mm, newbrk, oldbrk-newbrk, &uf, true);
259 } else if (ret == 1) {
265 /* Check against existing mmap mappings. */
266 next = find_vma(mm, oldbrk);
267 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
270 /* Ok, looks good - let it rip. */
271 if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
276 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
278 mmap_read_unlock(mm);
280 mmap_write_unlock(mm);
281 userfaultfd_unmap_complete(mm, &uf);
283 mm_populate(oldbrk, newbrk - oldbrk);
287 mmap_write_unlock(mm);
291 static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
293 unsigned long gap, prev_end;
296 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
297 * allow two stack_guard_gaps between them here, and when choosing
298 * an unmapped area; whereas when expanding we only require one.
299 * That's a little inconsistent, but keeps the code here simpler.
301 gap = vm_start_gap(vma);
303 prev_end = vm_end_gap(vma->vm_prev);
312 #ifdef CONFIG_DEBUG_VM_RB
313 static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma)
315 unsigned long max = vma_compute_gap(vma), subtree_gap;
316 if (vma->vm_rb.rb_left) {
317 subtree_gap = rb_entry(vma->vm_rb.rb_left,
318 struct vm_area_struct, vm_rb)->rb_subtree_gap;
319 if (subtree_gap > max)
322 if (vma->vm_rb.rb_right) {
323 subtree_gap = rb_entry(vma->vm_rb.rb_right,
324 struct vm_area_struct, vm_rb)->rb_subtree_gap;
325 if (subtree_gap > max)
331 static int browse_rb(struct mm_struct *mm)
333 struct rb_root *root = &mm->mm_rb;
334 int i = 0, j, bug = 0;
335 struct rb_node *nd, *pn = NULL;
336 unsigned long prev = 0, pend = 0;
338 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
339 struct vm_area_struct *vma;
340 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
341 if (vma->vm_start < prev) {
342 pr_emerg("vm_start %lx < prev %lx\n",
343 vma->vm_start, prev);
346 if (vma->vm_start < pend) {
347 pr_emerg("vm_start %lx < pend %lx\n",
348 vma->vm_start, pend);
351 if (vma->vm_start > vma->vm_end) {
352 pr_emerg("vm_start %lx > vm_end %lx\n",
353 vma->vm_start, vma->vm_end);
356 spin_lock(&mm->page_table_lock);
357 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
358 pr_emerg("free gap %lx, correct %lx\n",
360 vma_compute_subtree_gap(vma));
363 spin_unlock(&mm->page_table_lock);
366 prev = vma->vm_start;
370 for (nd = pn; nd; nd = rb_prev(nd))
373 pr_emerg("backwards %d, forwards %d\n", j, i);
379 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
383 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
384 struct vm_area_struct *vma;
385 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
386 VM_BUG_ON_VMA(vma != ignore &&
387 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
392 static void validate_mm(struct mm_struct *mm)
396 unsigned long highest_address = 0;
397 struct vm_area_struct *vma = mm->mmap;
400 struct anon_vma *anon_vma = vma->anon_vma;
401 struct anon_vma_chain *avc;
404 anon_vma_lock_read(anon_vma);
405 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
406 anon_vma_interval_tree_verify(avc);
407 anon_vma_unlock_read(anon_vma);
410 highest_address = vm_end_gap(vma);
414 if (i != mm->map_count) {
415 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
418 if (highest_address != mm->highest_vm_end) {
419 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
420 mm->highest_vm_end, highest_address);
424 if (i != mm->map_count) {
426 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
429 VM_BUG_ON_MM(bug, mm);
432 #define validate_mm_rb(root, ignore) do { } while (0)
433 #define validate_mm(mm) do { } while (0)
436 RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks,
437 struct vm_area_struct, vm_rb,
438 unsigned long, rb_subtree_gap, vma_compute_gap)
441 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
442 * vma->vm_prev->vm_end values changed, without modifying the vma's position
445 static void vma_gap_update(struct vm_area_struct *vma)
448 * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
449 * a callback function that does exactly what we want.
451 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
454 static inline void vma_rb_insert(struct vm_area_struct *vma,
455 struct rb_root *root)
457 /* All rb_subtree_gap values must be consistent prior to insertion */
458 validate_mm_rb(root, NULL);
460 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
463 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
466 * Note rb_erase_augmented is a fairly large inline function,
467 * so make sure we instantiate it only once with our desired
468 * augmented rbtree callbacks.
470 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
473 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
474 struct rb_root *root,
475 struct vm_area_struct *ignore)
478 * All rb_subtree_gap values must be consistent prior to erase,
479 * with the possible exception of
481 * a. the "next" vma being erased if next->vm_start was reduced in
482 * __vma_adjust() -> __vma_unlink()
483 * b. the vma being erased in detach_vmas_to_be_unmapped() ->
486 validate_mm_rb(root, ignore);
488 __vma_rb_erase(vma, root);
491 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
492 struct rb_root *root)
494 vma_rb_erase_ignore(vma, root, vma);
498 * vma has some anon_vma assigned, and is already inserted on that
499 * anon_vma's interval trees.
501 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
502 * vma must be removed from the anon_vma's interval trees using
503 * anon_vma_interval_tree_pre_update_vma().
505 * After the update, the vma will be reinserted using
506 * anon_vma_interval_tree_post_update_vma().
508 * The entire update must be protected by exclusive mmap_lock and by
509 * the root anon_vma's mutex.
512 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
514 struct anon_vma_chain *avc;
516 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
517 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
521 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
523 struct anon_vma_chain *avc;
525 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
526 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
529 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
530 unsigned long end, struct vm_area_struct **pprev,
531 struct rb_node ***rb_link, struct rb_node **rb_parent)
533 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
535 mmap_assert_locked(mm);
536 __rb_link = &mm->mm_rb.rb_node;
537 rb_prev = __rb_parent = NULL;
540 struct vm_area_struct *vma_tmp;
542 __rb_parent = *__rb_link;
543 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
545 if (vma_tmp->vm_end > addr) {
546 /* Fail if an existing vma overlaps the area */
547 if (vma_tmp->vm_start < end)
549 __rb_link = &__rb_parent->rb_left;
551 rb_prev = __rb_parent;
552 __rb_link = &__rb_parent->rb_right;
558 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
559 *rb_link = __rb_link;
560 *rb_parent = __rb_parent;
565 * vma_next() - Get the next VMA.
566 * @mm: The mm_struct.
567 * @vma: The current vma.
569 * If @vma is NULL, return the first vma in the mm.
571 * Returns: The next VMA after @vma.
573 static inline struct vm_area_struct *vma_next(struct mm_struct *mm,
574 struct vm_area_struct *vma)
583 * munmap_vma_range() - munmap VMAs that overlap a range.
585 * @start: The start of the range.
586 * @len: The length of the range.
587 * @pprev: pointer to the pointer that will be set to previous vm_area_struct
588 * @rb_link: the rb_node
589 * @rb_parent: the parent rb_node
591 * Find all the vm_area_struct that overlap from @start to
592 * @end and munmap them. Set @pprev to the previous vm_area_struct.
594 * Returns: -ENOMEM on munmap failure or 0 on success.
597 munmap_vma_range(struct mm_struct *mm, unsigned long start, unsigned long len,
598 struct vm_area_struct **pprev, struct rb_node ***link,
599 struct rb_node **parent, struct list_head *uf)
602 while (find_vma_links(mm, start, start + len, pprev, link, parent))
603 if (do_munmap(mm, start, len, uf))
608 static unsigned long count_vma_pages_range(struct mm_struct *mm,
609 unsigned long addr, unsigned long end)
611 unsigned long nr_pages = 0;
612 struct vm_area_struct *vma;
614 /* Find first overlapping mapping */
615 vma = find_vma_intersection(mm, addr, end);
619 nr_pages = (min(end, vma->vm_end) -
620 max(addr, vma->vm_start)) >> PAGE_SHIFT;
622 /* Iterate over the rest of the overlaps */
623 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
624 unsigned long overlap_len;
626 if (vma->vm_start > end)
629 overlap_len = min(end, vma->vm_end) - vma->vm_start;
630 nr_pages += overlap_len >> PAGE_SHIFT;
636 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
637 struct rb_node **rb_link, struct rb_node *rb_parent)
639 /* Update tracking information for the gap following the new vma. */
641 vma_gap_update(vma->vm_next);
643 mm->highest_vm_end = vm_end_gap(vma);
646 * vma->vm_prev wasn't known when we followed the rbtree to find the
647 * correct insertion point for that vma. As a result, we could not
648 * update the vma vm_rb parents rb_subtree_gap values on the way down.
649 * So, we first insert the vma with a zero rb_subtree_gap value
650 * (to be consistent with what we did on the way down), and then
651 * immediately update the gap to the correct value. Finally we
652 * rebalance the rbtree after all augmented values have been set.
654 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
655 vma->rb_subtree_gap = 0;
657 vma_rb_insert(vma, &mm->mm_rb);
660 static void __vma_link_file(struct vm_area_struct *vma)
666 struct address_space *mapping = file->f_mapping;
668 if (vma->vm_flags & VM_SHARED)
669 mapping_allow_writable(mapping);
671 flush_dcache_mmap_lock(mapping);
672 vma_interval_tree_insert(vma, &mapping->i_mmap);
673 flush_dcache_mmap_unlock(mapping);
678 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
679 struct vm_area_struct *prev, struct rb_node **rb_link,
680 struct rb_node *rb_parent)
682 __vma_link_list(mm, vma, prev);
683 __vma_link_rb(mm, vma, rb_link, rb_parent);
686 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
687 struct vm_area_struct *prev, struct rb_node **rb_link,
688 struct rb_node *rb_parent)
690 struct address_space *mapping = NULL;
693 mapping = vma->vm_file->f_mapping;
694 i_mmap_lock_write(mapping);
697 __vma_link(mm, vma, prev, rb_link, rb_parent);
698 __vma_link_file(vma);
701 i_mmap_unlock_write(mapping);
708 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
709 * mm's list and rbtree. It has already been inserted into the interval tree.
711 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
713 struct vm_area_struct *prev;
714 struct rb_node **rb_link, *rb_parent;
716 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
717 &prev, &rb_link, &rb_parent))
719 __vma_link(mm, vma, prev, rb_link, rb_parent);
723 static __always_inline void __vma_unlink(struct mm_struct *mm,
724 struct vm_area_struct *vma,
725 struct vm_area_struct *ignore)
727 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
728 __vma_unlink_list(mm, vma);
730 vmacache_invalidate(mm);
734 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
735 * is already present in an i_mmap tree without adjusting the tree.
736 * The following helper function should be used when such adjustments
737 * are necessary. The "insert" vma (if any) is to be inserted
738 * before we drop the necessary locks.
740 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
741 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
742 struct vm_area_struct *expand)
744 struct mm_struct *mm = vma->vm_mm;
745 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
746 struct address_space *mapping = NULL;
747 struct rb_root_cached *root = NULL;
748 struct anon_vma *anon_vma = NULL;
749 struct file *file = vma->vm_file;
750 bool start_changed = false, end_changed = false;
751 long adjust_next = 0;
754 if (next && !insert) {
755 struct vm_area_struct *exporter = NULL, *importer = NULL;
757 if (end >= next->vm_end) {
759 * vma expands, overlapping all the next, and
760 * perhaps the one after too (mprotect case 6).
761 * The only other cases that gets here are
762 * case 1, case 7 and case 8.
764 if (next == expand) {
766 * The only case where we don't expand "vma"
767 * and we expand "next" instead is case 8.
769 VM_WARN_ON(end != next->vm_end);
771 * remove_next == 3 means we're
772 * removing "vma" and that to do so we
773 * swapped "vma" and "next".
776 VM_WARN_ON(file != next->vm_file);
779 VM_WARN_ON(expand != vma);
781 * case 1, 6, 7, remove_next == 2 is case 6,
782 * remove_next == 1 is case 1 or 7.
784 remove_next = 1 + (end > next->vm_end);
785 VM_WARN_ON(remove_next == 2 &&
786 end != next->vm_next->vm_end);
787 /* trim end to next, for case 6 first pass */
795 * If next doesn't have anon_vma, import from vma after
796 * next, if the vma overlaps with it.
798 if (remove_next == 2 && !next->anon_vma)
799 exporter = next->vm_next;
801 } else if (end > next->vm_start) {
803 * vma expands, overlapping part of the next:
804 * mprotect case 5 shifting the boundary up.
806 adjust_next = (end - next->vm_start);
809 VM_WARN_ON(expand != importer);
810 } else if (end < vma->vm_end) {
812 * vma shrinks, and !insert tells it's not
813 * split_vma inserting another: so it must be
814 * mprotect case 4 shifting the boundary down.
816 adjust_next = -(vma->vm_end - end);
819 VM_WARN_ON(expand != importer);
823 * Easily overlooked: when mprotect shifts the boundary,
824 * make sure the expanding vma has anon_vma set if the
825 * shrinking vma had, to cover any anon pages imported.
827 if (exporter && exporter->anon_vma && !importer->anon_vma) {
830 importer->anon_vma = exporter->anon_vma;
831 error = anon_vma_clone(importer, exporter);
837 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
840 mapping = file->f_mapping;
841 root = &mapping->i_mmap;
842 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
845 uprobe_munmap(next, next->vm_start, next->vm_end);
847 i_mmap_lock_write(mapping);
850 * Put into interval tree now, so instantiated pages
851 * are visible to arm/parisc __flush_dcache_page
852 * throughout; but we cannot insert into address
853 * space until vma start or end is updated.
855 __vma_link_file(insert);
859 anon_vma = vma->anon_vma;
860 if (!anon_vma && adjust_next)
861 anon_vma = next->anon_vma;
863 VM_WARN_ON(adjust_next && next->anon_vma &&
864 anon_vma != next->anon_vma);
865 anon_vma_lock_write(anon_vma);
866 anon_vma_interval_tree_pre_update_vma(vma);
868 anon_vma_interval_tree_pre_update_vma(next);
872 flush_dcache_mmap_lock(mapping);
873 vma_interval_tree_remove(vma, root);
875 vma_interval_tree_remove(next, root);
878 if (start != vma->vm_start) {
879 vma->vm_start = start;
880 start_changed = true;
882 if (end != vma->vm_end) {
886 vma->vm_pgoff = pgoff;
888 next->vm_start += adjust_next;
889 next->vm_pgoff += adjust_next >> PAGE_SHIFT;
894 vma_interval_tree_insert(next, root);
895 vma_interval_tree_insert(vma, root);
896 flush_dcache_mmap_unlock(mapping);
901 * vma_merge has merged next into vma, and needs
902 * us to remove next before dropping the locks.
904 if (remove_next != 3)
905 __vma_unlink(mm, next, next);
908 * vma is not before next if they've been
911 * pre-swap() next->vm_start was reduced so
912 * tell validate_mm_rb to ignore pre-swap()
913 * "next" (which is stored in post-swap()
916 __vma_unlink(mm, next, vma);
918 __remove_shared_vm_struct(next, file, mapping);
921 * split_vma has split insert from vma, and needs
922 * us to insert it before dropping the locks
923 * (it may either follow vma or precede it).
925 __insert_vm_struct(mm, insert);
931 mm->highest_vm_end = vm_end_gap(vma);
932 else if (!adjust_next)
933 vma_gap_update(next);
938 anon_vma_interval_tree_post_update_vma(vma);
940 anon_vma_interval_tree_post_update_vma(next);
941 anon_vma_unlock_write(anon_vma);
945 i_mmap_unlock_write(mapping);
954 uprobe_munmap(next, next->vm_start, next->vm_end);
958 anon_vma_merge(vma, next);
960 mpol_put(vma_policy(next));
963 * In mprotect's case 6 (see comments on vma_merge),
964 * we must remove another next too. It would clutter
965 * up the code too much to do both in one go.
967 if (remove_next != 3) {
969 * If "next" was removed and vma->vm_end was
970 * expanded (up) over it, in turn
971 * "next->vm_prev->vm_end" changed and the
972 * "vma->vm_next" gap must be updated.
977 * For the scope of the comment "next" and
978 * "vma" considered pre-swap(): if "vma" was
979 * removed, next->vm_start was expanded (down)
980 * over it and the "next" gap must be updated.
981 * Because of the swap() the post-swap() "vma"
982 * actually points to pre-swap() "next"
983 * (post-swap() "next" as opposed is now a
988 if (remove_next == 2) {
994 vma_gap_update(next);
997 * If remove_next == 2 we obviously can't
1000 * If remove_next == 3 we can't reach this
1001 * path because pre-swap() next is always not
1002 * NULL. pre-swap() "next" is not being
1003 * removed and its next->vm_end is not altered
1004 * (and furthermore "end" already matches
1005 * next->vm_end in remove_next == 3).
1007 * We reach this only in the remove_next == 1
1008 * case if the "next" vma that was removed was
1009 * the highest vma of the mm. However in such
1010 * case next->vm_end == "end" and the extended
1011 * "vma" has vma->vm_end == next->vm_end so
1012 * mm->highest_vm_end doesn't need any update
1013 * in remove_next == 1 case.
1015 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
1019 uprobe_mmap(insert);
1027 * If the vma has a ->close operation then the driver probably needs to release
1028 * per-vma resources, so we don't attempt to merge those.
1030 static inline int is_mergeable_vma(struct vm_area_struct *vma,
1031 struct file *file, unsigned long vm_flags,
1032 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1035 * VM_SOFTDIRTY should not prevent from VMA merging, if we
1036 * match the flags but dirty bit -- the caller should mark
1037 * merged VMA as dirty. If dirty bit won't be excluded from
1038 * comparison, we increase pressure on the memory system forcing
1039 * the kernel to generate new VMAs when old one could be
1042 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
1044 if (vma->vm_file != file)
1046 if (vma->vm_ops && vma->vm_ops->close)
1048 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1053 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1054 struct anon_vma *anon_vma2,
1055 struct vm_area_struct *vma)
1058 * The list_is_singular() test is to avoid merging VMA cloned from
1059 * parents. This can improve scalability caused by anon_vma lock.
1061 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1062 list_is_singular(&vma->anon_vma_chain)))
1064 return anon_vma1 == anon_vma2;
1068 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1069 * in front of (at a lower virtual address and file offset than) the vma.
1071 * We cannot merge two vmas if they have differently assigned (non-NULL)
1072 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1074 * We don't check here for the merged mmap wrapping around the end of pagecache
1075 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
1076 * wrap, nor mmaps which cover the final page at index -1UL.
1079 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1080 struct anon_vma *anon_vma, struct file *file,
1082 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1084 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1085 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1086 if (vma->vm_pgoff == vm_pgoff)
1093 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1094 * beyond (at a higher virtual address and file offset than) the vma.
1096 * We cannot merge two vmas if they have differently assigned (non-NULL)
1097 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1100 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1101 struct anon_vma *anon_vma, struct file *file,
1103 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1105 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1106 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1108 vm_pglen = vma_pages(vma);
1109 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1116 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1117 * whether that can be merged with its predecessor or its successor.
1118 * Or both (it neatly fills a hole).
1120 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1121 * certain not to be mapped by the time vma_merge is called; but when
1122 * called for mprotect, it is certain to be already mapped (either at
1123 * an offset within prev, or at the start of next), and the flags of
1124 * this area are about to be changed to vm_flags - and the no-change
1125 * case has already been eliminated.
1127 * The following mprotect cases have to be considered, where AAAA is
1128 * the area passed down from mprotect_fixup, never extending beyond one
1129 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1132 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN
1133 * cannot merge might become might become
1134 * PPNNNNNNNNNN PPPPPPPPPPNN
1135 * mmap, brk or case 4 below case 5 below
1138 * PPPP NNNN PPPPNNNNXXXX
1139 * might become might become
1140 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
1141 * PPPPPPPPNNNN 2 or PPPPPPPPXXXX 7 or
1142 * PPPPNNNNNNNN 3 PPPPXXXXXXXX 8
1144 * It is important for case 8 that the vma NNNN overlapping the
1145 * region AAAA is never going to extended over XXXX. Instead XXXX must
1146 * be extended in region AAAA and NNNN must be removed. This way in
1147 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1148 * rmap_locks, the properties of the merged vma will be already
1149 * correct for the whole merged range. Some of those properties like
1150 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1151 * be correct for the whole merged range immediately after the
1152 * rmap_locks are released. Otherwise if XXXX would be removed and
1153 * NNNN would be extended over the XXXX range, remove_migration_ptes
1154 * or other rmap walkers (if working on addresses beyond the "end"
1155 * parameter) may establish ptes with the wrong permissions of NNNN
1156 * instead of the right permissions of XXXX.
1158 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1159 struct vm_area_struct *prev, unsigned long addr,
1160 unsigned long end, unsigned long vm_flags,
1161 struct anon_vma *anon_vma, struct file *file,
1162 pgoff_t pgoff, struct mempolicy *policy,
1163 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1165 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1166 struct vm_area_struct *area, *next;
1170 * We later require that vma->vm_flags == vm_flags,
1171 * so this tests vma->vm_flags & VM_SPECIAL, too.
1173 if (vm_flags & VM_SPECIAL)
1176 next = vma_next(mm, prev);
1178 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1179 next = next->vm_next;
1181 /* verify some invariant that must be enforced by the caller */
1182 VM_WARN_ON(prev && addr <= prev->vm_start);
1183 VM_WARN_ON(area && end > area->vm_end);
1184 VM_WARN_ON(addr >= end);
1187 * Can it merge with the predecessor?
1189 if (prev && prev->vm_end == addr &&
1190 mpol_equal(vma_policy(prev), policy) &&
1191 can_vma_merge_after(prev, vm_flags,
1192 anon_vma, file, pgoff,
1193 vm_userfaultfd_ctx)) {
1195 * OK, it can. Can we now merge in the successor as well?
1197 if (next && end == next->vm_start &&
1198 mpol_equal(policy, vma_policy(next)) &&
1199 can_vma_merge_before(next, vm_flags,
1202 vm_userfaultfd_ctx) &&
1203 is_mergeable_anon_vma(prev->anon_vma,
1204 next->anon_vma, NULL)) {
1206 err = __vma_adjust(prev, prev->vm_start,
1207 next->vm_end, prev->vm_pgoff, NULL,
1209 } else /* cases 2, 5, 7 */
1210 err = __vma_adjust(prev, prev->vm_start,
1211 end, prev->vm_pgoff, NULL, prev);
1214 khugepaged_enter_vma_merge(prev, vm_flags);
1219 * Can this new request be merged in front of next?
1221 if (next && end == next->vm_start &&
1222 mpol_equal(policy, vma_policy(next)) &&
1223 can_vma_merge_before(next, vm_flags,
1224 anon_vma, file, pgoff+pglen,
1225 vm_userfaultfd_ctx)) {
1226 if (prev && addr < prev->vm_end) /* case 4 */
1227 err = __vma_adjust(prev, prev->vm_start,
1228 addr, prev->vm_pgoff, NULL, next);
1229 else { /* cases 3, 8 */
1230 err = __vma_adjust(area, addr, next->vm_end,
1231 next->vm_pgoff - pglen, NULL, next);
1233 * In case 3 area is already equal to next and
1234 * this is a noop, but in case 8 "area" has
1235 * been removed and next was expanded over it.
1241 khugepaged_enter_vma_merge(area, vm_flags);
1249 * Rough compatibility check to quickly see if it's even worth looking
1250 * at sharing an anon_vma.
1252 * They need to have the same vm_file, and the flags can only differ
1253 * in things that mprotect may change.
1255 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1256 * we can merge the two vma's. For example, we refuse to merge a vma if
1257 * there is a vm_ops->close() function, because that indicates that the
1258 * driver is doing some kind of reference counting. But that doesn't
1259 * really matter for the anon_vma sharing case.
1261 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1263 return a->vm_end == b->vm_start &&
1264 mpol_equal(vma_policy(a), vma_policy(b)) &&
1265 a->vm_file == b->vm_file &&
1266 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1267 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1271 * Do some basic sanity checking to see if we can re-use the anon_vma
1272 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1273 * the same as 'old', the other will be the new one that is trying
1274 * to share the anon_vma.
1276 * NOTE! This runs with mm_sem held for reading, so it is possible that
1277 * the anon_vma of 'old' is concurrently in the process of being set up
1278 * by another page fault trying to merge _that_. But that's ok: if it
1279 * is being set up, that automatically means that it will be a singleton
1280 * acceptable for merging, so we can do all of this optimistically. But
1281 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1283 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1284 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1285 * is to return an anon_vma that is "complex" due to having gone through
1288 * We also make sure that the two vma's are compatible (adjacent,
1289 * and with the same memory policies). That's all stable, even with just
1290 * a read lock on the mm_sem.
1292 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1294 if (anon_vma_compatible(a, b)) {
1295 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1297 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1304 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1305 * neighbouring vmas for a suitable anon_vma, before it goes off
1306 * to allocate a new anon_vma. It checks because a repetitive
1307 * sequence of mprotects and faults may otherwise lead to distinct
1308 * anon_vmas being allocated, preventing vma merge in subsequent
1311 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1313 struct anon_vma *anon_vma = NULL;
1315 /* Try next first. */
1317 anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next);
1322 /* Try prev next. */
1324 anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma);
1327 * We might reach here with anon_vma == NULL if we can't find
1328 * any reusable anon_vma.
1329 * There's no absolute need to look only at touching neighbours:
1330 * we could search further afield for "compatible" anon_vmas.
1331 * But it would probably just be a waste of time searching,
1332 * or lead to too many vmas hanging off the same anon_vma.
1333 * We're trying to allow mprotect remerging later on,
1334 * not trying to minimize memory used for anon_vmas.
1340 * If a hint addr is less than mmap_min_addr change hint to be as
1341 * low as possible but still greater than mmap_min_addr
1343 static inline unsigned long round_hint_to_min(unsigned long hint)
1346 if (((void *)hint != NULL) &&
1347 (hint < mmap_min_addr))
1348 return PAGE_ALIGN(mmap_min_addr);
1352 int mlock_future_check(struct mm_struct *mm, unsigned long flags,
1355 unsigned long locked, lock_limit;
1357 /* mlock MCL_FUTURE? */
1358 if (flags & VM_LOCKED) {
1359 locked = len >> PAGE_SHIFT;
1360 locked += mm->locked_vm;
1361 lock_limit = rlimit(RLIMIT_MEMLOCK);
1362 lock_limit >>= PAGE_SHIFT;
1363 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1369 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1371 if (S_ISREG(inode->i_mode))
1372 return MAX_LFS_FILESIZE;
1374 if (S_ISBLK(inode->i_mode))
1375 return MAX_LFS_FILESIZE;
1377 if (S_ISSOCK(inode->i_mode))
1378 return MAX_LFS_FILESIZE;
1380 /* Special "we do even unsigned file positions" case */
1381 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1384 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1388 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1389 unsigned long pgoff, unsigned long len)
1391 u64 maxsize = file_mmap_size_max(file, inode);
1393 if (maxsize && len > maxsize)
1396 if (pgoff > maxsize >> PAGE_SHIFT)
1402 * The caller must write-lock current->mm->mmap_lock.
1404 unsigned long do_mmap(struct file *file, unsigned long addr,
1405 unsigned long len, unsigned long prot,
1406 unsigned long flags, unsigned long pgoff,
1407 unsigned long *populate, struct list_head *uf)
1409 struct mm_struct *mm = current->mm;
1410 vm_flags_t vm_flags;
1419 * Does the application expect PROT_READ to imply PROT_EXEC?
1421 * (the exception is when the underlying filesystem is noexec
1422 * mounted, in which case we dont add PROT_EXEC.)
1424 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1425 if (!(file && path_noexec(&file->f_path)))
1428 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1429 if (flags & MAP_FIXED_NOREPLACE)
1432 if (!(flags & MAP_FIXED))
1433 addr = round_hint_to_min(addr);
1435 /* Careful about overflows.. */
1436 len = PAGE_ALIGN(len);
1440 /* offset overflow? */
1441 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1444 /* Too many mappings? */
1445 if (mm->map_count > sysctl_max_map_count)
1448 /* Obtain the address to map to. we verify (or select) it and ensure
1449 * that it represents a valid section of the address space.
1451 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1452 if (IS_ERR_VALUE(addr))
1455 if (flags & MAP_FIXED_NOREPLACE) {
1456 if (find_vma_intersection(mm, addr, 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))
1517 vm_flags |= VM_SHARED | VM_MAYSHARE;
1518 if (!(file->f_mode & FMODE_WRITE))
1519 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1522 if (!(file->f_mode & FMODE_READ))
1524 if (path_noexec(&file->f_path)) {
1525 if (vm_flags & VM_EXEC)
1527 vm_flags &= ~VM_MAYEXEC;
1530 if (!file->f_op->mmap)
1532 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1540 switch (flags & MAP_TYPE) {
1542 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1548 vm_flags |= VM_SHARED | VM_MAYSHARE;
1552 * Set pgoff according to addr for anon_vma.
1554 pgoff = addr >> PAGE_SHIFT;
1562 * Set 'VM_NORESERVE' if we should not account for the
1563 * memory use of this mapping.
1565 if (flags & MAP_NORESERVE) {
1566 /* We honor MAP_NORESERVE if allowed to overcommit */
1567 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1568 vm_flags |= VM_NORESERVE;
1570 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1571 if (file && is_file_hugepages(file))
1572 vm_flags |= VM_NORESERVE;
1575 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1576 if (!IS_ERR_VALUE(addr) &&
1577 ((vm_flags & VM_LOCKED) ||
1578 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1583 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1584 unsigned long prot, unsigned long flags,
1585 unsigned long fd, unsigned long pgoff)
1587 struct file *file = NULL;
1588 unsigned long retval;
1590 if (!(flags & MAP_ANONYMOUS)) {
1591 audit_mmap_fd(fd, flags);
1595 if (is_file_hugepages(file)) {
1596 len = ALIGN(len, huge_page_size(hstate_file(file)));
1597 } else if (unlikely(flags & MAP_HUGETLB)) {
1601 } else if (flags & MAP_HUGETLB) {
1602 struct ucounts *ucounts = NULL;
1605 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1609 len = ALIGN(len, huge_page_size(hs));
1611 * VM_NORESERVE is used because the reservations will be
1612 * taken when vm_ops->mmap() is called
1613 * A dummy user value is used because we are not locking
1614 * memory so no accounting is necessary
1616 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1618 &ucounts, HUGETLB_ANONHUGE_INODE,
1619 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1621 return PTR_ERR(file);
1624 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1631 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1632 unsigned long, prot, unsigned long, flags,
1633 unsigned long, fd, unsigned long, pgoff)
1635 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1638 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1639 struct mmap_arg_struct {
1643 unsigned long flags;
1645 unsigned long offset;
1648 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1650 struct mmap_arg_struct a;
1652 if (copy_from_user(&a, arg, sizeof(a)))
1654 if (offset_in_page(a.offset))
1657 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1658 a.offset >> PAGE_SHIFT);
1660 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1663 * Some shared mappings will want the pages marked read-only
1664 * to track write events. If so, we'll downgrade vm_page_prot
1665 * to the private version (using protection_map[] without the
1668 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1670 vm_flags_t vm_flags = vma->vm_flags;
1671 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1673 /* If it was private or non-writable, the write bit is already clear */
1674 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1677 /* The backer wishes to know when pages are first written to? */
1678 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1681 /* The open routine did something to the protections that pgprot_modify
1682 * won't preserve? */
1683 if (pgprot_val(vm_page_prot) !=
1684 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1688 * Do we need to track softdirty? hugetlb does not support softdirty
1691 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY) &&
1692 !is_vm_hugetlb_page(vma))
1695 /* Specialty mapping? */
1696 if (vm_flags & VM_PFNMAP)
1699 /* Can the mapping track the dirty pages? */
1700 return vma->vm_file && vma->vm_file->f_mapping &&
1701 mapping_can_writeback(vma->vm_file->f_mapping);
1705 * We account for memory if it's a private writeable mapping,
1706 * not hugepages and VM_NORESERVE wasn't set.
1708 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1711 * hugetlb has its own accounting separate from the core VM
1712 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1714 if (file && is_file_hugepages(file))
1717 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1720 unsigned long mmap_region(struct file *file, unsigned long addr,
1721 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1722 struct list_head *uf)
1724 struct mm_struct *mm = current->mm;
1725 struct vm_area_struct *vma, *prev, *merge;
1727 struct rb_node **rb_link, *rb_parent;
1728 unsigned long charged = 0;
1730 /* Check against address space limit. */
1731 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1732 unsigned long nr_pages;
1735 * MAP_FIXED may remove pages of mappings that intersects with
1736 * requested mapping. Account for the pages it would unmap.
1738 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1740 if (!may_expand_vm(mm, vm_flags,
1741 (len >> PAGE_SHIFT) - nr_pages))
1745 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
1746 if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
1749 * Private writable mapping: check memory availability
1751 if (accountable_mapping(file, vm_flags)) {
1752 charged = len >> PAGE_SHIFT;
1753 if (security_vm_enough_memory_mm(mm, charged))
1755 vm_flags |= VM_ACCOUNT;
1759 * Can we just expand an old mapping?
1761 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1762 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1767 * Determine the object being mapped and call the appropriate
1768 * specific mapper. the address has already been validated, but
1769 * not unmapped, but the maps are removed from the list.
1771 vma = vm_area_alloc(mm);
1777 vma->vm_start = addr;
1778 vma->vm_end = addr + len;
1779 vma->vm_flags = vm_flags;
1780 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1781 vma->vm_pgoff = pgoff;
1784 if (vm_flags & VM_SHARED) {
1785 error = mapping_map_writable(file->f_mapping);
1790 vma->vm_file = get_file(file);
1791 error = call_mmap(file, vma);
1793 goto unmap_and_free_vma;
1795 /* Can addr have changed??
1797 * Answer: Yes, several device drivers can do it in their
1798 * f_op->mmap method. -DaveM
1799 * Bug: If addr is changed, prev, rb_link, rb_parent should
1800 * be updated for vma_link()
1802 WARN_ON_ONCE(addr != vma->vm_start);
1804 addr = vma->vm_start;
1806 /* If vm_flags changed after call_mmap(), we should try merge vma again
1807 * as we may succeed this time.
1809 if (unlikely(vm_flags != vma->vm_flags && prev)) {
1810 merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
1811 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX);
1813 /* ->mmap() can change vma->vm_file and fput the original file. So
1814 * fput the vma->vm_file here or we would add an extra fput for file
1815 * and cause general protection fault ultimately.
1820 /* Update vm_flags to pick up the change. */
1821 vm_flags = vma->vm_flags;
1822 goto unmap_writable;
1826 vm_flags = vma->vm_flags;
1827 } else if (vm_flags & VM_SHARED) {
1828 error = shmem_zero_setup(vma);
1832 vma_set_anonymous(vma);
1835 /* Allow architectures to sanity-check the vm_flags */
1836 if (!arch_validate_flags(vma->vm_flags)) {
1839 goto close_and_free_vma;
1844 vma_link(mm, vma, prev, rb_link, rb_parent);
1845 /* Once vma denies write, undo our temporary denial count */
1847 if (file && vm_flags & VM_SHARED)
1848 mapping_unmap_writable(file->f_mapping);
1849 file = vma->vm_file;
1851 perf_event_mmap(vma);
1853 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1854 if (vm_flags & VM_LOCKED) {
1855 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1856 is_vm_hugetlb_page(vma) ||
1857 vma == get_gate_vma(current->mm))
1858 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1860 mm->locked_vm += (len >> PAGE_SHIFT);
1867 * New (or expanded) vma always get soft dirty status.
1868 * Otherwise user-space soft-dirty page tracker won't
1869 * be able to distinguish situation when vma area unmapped,
1870 * then new mapped in-place (which must be aimed as
1871 * a completely new data area).
1873 vma->vm_flags |= VM_SOFTDIRTY;
1875 vma_set_page_prot(vma);
1880 if (vma->vm_ops && vma->vm_ops->close)
1881 vma->vm_ops->close(vma);
1884 vma->vm_file = NULL;
1886 /* Undo any partial mapping done by a device driver. */
1887 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1888 if (vm_flags & VM_SHARED)
1889 mapping_unmap_writable(file->f_mapping);
1894 vm_unacct_memory(charged);
1898 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1901 * We implement the search by looking for an rbtree node that
1902 * immediately follows a suitable gap. That is,
1903 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1904 * - gap_end = vma->vm_start >= info->low_limit + length;
1905 * - gap_end - gap_start >= length
1908 struct mm_struct *mm = current->mm;
1909 struct vm_area_struct *vma;
1910 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1912 /* Adjust search length to account for worst case alignment overhead */
1913 length = info->length + info->align_mask;
1914 if (length < info->length)
1917 /* Adjust search limits by the desired length */
1918 if (info->high_limit < length)
1920 high_limit = info->high_limit - length;
1922 if (info->low_limit > high_limit)
1924 low_limit = info->low_limit + length;
1926 /* Check if rbtree root looks promising */
1927 if (RB_EMPTY_ROOT(&mm->mm_rb))
1929 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1930 if (vma->rb_subtree_gap < length)
1934 /* Visit left subtree if it looks promising */
1935 gap_end = vm_start_gap(vma);
1936 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1937 struct vm_area_struct *left =
1938 rb_entry(vma->vm_rb.rb_left,
1939 struct vm_area_struct, vm_rb);
1940 if (left->rb_subtree_gap >= length) {
1946 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1948 /* Check if current node has a suitable gap */
1949 if (gap_start > high_limit)
1951 if (gap_end >= low_limit &&
1952 gap_end > gap_start && gap_end - gap_start >= length)
1955 /* Visit right subtree if it looks promising */
1956 if (vma->vm_rb.rb_right) {
1957 struct vm_area_struct *right =
1958 rb_entry(vma->vm_rb.rb_right,
1959 struct vm_area_struct, vm_rb);
1960 if (right->rb_subtree_gap >= length) {
1966 /* Go back up the rbtree to find next candidate node */
1968 struct rb_node *prev = &vma->vm_rb;
1969 if (!rb_parent(prev))
1971 vma = rb_entry(rb_parent(prev),
1972 struct vm_area_struct, vm_rb);
1973 if (prev == vma->vm_rb.rb_left) {
1974 gap_start = vm_end_gap(vma->vm_prev);
1975 gap_end = vm_start_gap(vma);
1982 /* Check highest gap, which does not precede any rbtree node */
1983 gap_start = mm->highest_vm_end;
1984 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1985 if (gap_start > high_limit)
1989 /* We found a suitable gap. Clip it with the original low_limit. */
1990 if (gap_start < info->low_limit)
1991 gap_start = info->low_limit;
1993 /* Adjust gap address to the desired alignment */
1994 gap_start += (info->align_offset - gap_start) & info->align_mask;
1996 VM_BUG_ON(gap_start + info->length > info->high_limit);
1997 VM_BUG_ON(gap_start + info->length > gap_end);
2001 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
2003 struct mm_struct *mm = current->mm;
2004 struct vm_area_struct *vma;
2005 unsigned long length, low_limit, high_limit, gap_start, gap_end;
2007 /* Adjust search length to account for worst case alignment overhead */
2008 length = info->length + info->align_mask;
2009 if (length < info->length)
2013 * Adjust search limits by the desired length.
2014 * See implementation comment at top of unmapped_area().
2016 gap_end = info->high_limit;
2017 if (gap_end < length)
2019 high_limit = gap_end - length;
2021 if (info->low_limit > high_limit)
2023 low_limit = info->low_limit + length;
2025 /* Check highest gap, which does not precede any rbtree node */
2026 gap_start = mm->highest_vm_end;
2027 if (gap_start <= high_limit)
2030 /* Check if rbtree root looks promising */
2031 if (RB_EMPTY_ROOT(&mm->mm_rb))
2033 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
2034 if (vma->rb_subtree_gap < length)
2038 /* Visit right subtree if it looks promising */
2039 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
2040 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
2041 struct vm_area_struct *right =
2042 rb_entry(vma->vm_rb.rb_right,
2043 struct vm_area_struct, vm_rb);
2044 if (right->rb_subtree_gap >= length) {
2051 /* Check if current node has a suitable gap */
2052 gap_end = vm_start_gap(vma);
2053 if (gap_end < low_limit)
2055 if (gap_start <= high_limit &&
2056 gap_end > gap_start && gap_end - gap_start >= length)
2059 /* Visit left subtree if it looks promising */
2060 if (vma->vm_rb.rb_left) {
2061 struct vm_area_struct *left =
2062 rb_entry(vma->vm_rb.rb_left,
2063 struct vm_area_struct, vm_rb);
2064 if (left->rb_subtree_gap >= length) {
2070 /* Go back up the rbtree to find next candidate node */
2072 struct rb_node *prev = &vma->vm_rb;
2073 if (!rb_parent(prev))
2075 vma = rb_entry(rb_parent(prev),
2076 struct vm_area_struct, vm_rb);
2077 if (prev == vma->vm_rb.rb_right) {
2078 gap_start = vma->vm_prev ?
2079 vm_end_gap(vma->vm_prev) : 0;
2086 /* We found a suitable gap. Clip it with the original high_limit. */
2087 if (gap_end > info->high_limit)
2088 gap_end = info->high_limit;
2091 /* Compute highest gap address at the desired alignment */
2092 gap_end -= info->length;
2093 gap_end -= (gap_end - info->align_offset) & info->align_mask;
2095 VM_BUG_ON(gap_end < info->low_limit);
2096 VM_BUG_ON(gap_end < gap_start);
2101 * Search for an unmapped address range.
2103 * We are looking for a range that:
2104 * - does not intersect with any VMA;
2105 * - is contained within the [low_limit, high_limit) interval;
2106 * - is at least the desired size.
2107 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
2109 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
2113 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
2114 addr = unmapped_area_topdown(info);
2116 addr = unmapped_area(info);
2118 trace_vm_unmapped_area(addr, info);
2122 /* Get an address range which is currently unmapped.
2123 * For shmat() with addr=0.
2125 * Ugly calling convention alert:
2126 * Return value with the low bits set means error value,
2128 * if (ret & ~PAGE_MASK)
2131 * This function "knows" that -ENOMEM has the bits set.
2133 #ifndef HAVE_ARCH_UNMAPPED_AREA
2135 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2136 unsigned long len, unsigned long pgoff, unsigned long flags)
2138 struct mm_struct *mm = current->mm;
2139 struct vm_area_struct *vma, *prev;
2140 struct vm_unmapped_area_info info;
2141 const unsigned long mmap_end = arch_get_mmap_end(addr);
2143 if (len > mmap_end - mmap_min_addr)
2146 if (flags & MAP_FIXED)
2150 addr = PAGE_ALIGN(addr);
2151 vma = find_vma_prev(mm, addr, &prev);
2152 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2153 (!vma || addr + len <= vm_start_gap(vma)) &&
2154 (!prev || addr >= vm_end_gap(prev)))
2160 info.low_limit = mm->mmap_base;
2161 info.high_limit = mmap_end;
2162 info.align_mask = 0;
2163 info.align_offset = 0;
2164 return vm_unmapped_area(&info);
2169 * This mmap-allocator allocates new areas top-down from below the
2170 * stack's low limit (the base):
2172 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2174 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2175 unsigned long len, unsigned long pgoff,
2176 unsigned long flags)
2178 struct vm_area_struct *vma, *prev;
2179 struct mm_struct *mm = current->mm;
2180 struct vm_unmapped_area_info info;
2181 const unsigned long mmap_end = arch_get_mmap_end(addr);
2183 /* requested length too big for entire address space */
2184 if (len > mmap_end - mmap_min_addr)
2187 if (flags & MAP_FIXED)
2190 /* requesting a specific address */
2192 addr = PAGE_ALIGN(addr);
2193 vma = find_vma_prev(mm, addr, &prev);
2194 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2195 (!vma || addr + len <= vm_start_gap(vma)) &&
2196 (!prev || addr >= vm_end_gap(prev)))
2200 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2202 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2203 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2204 info.align_mask = 0;
2205 info.align_offset = 0;
2206 addr = vm_unmapped_area(&info);
2209 * A failed mmap() very likely causes application failure,
2210 * so fall back to the bottom-up function here. This scenario
2211 * can happen with large stack limits and large mmap()
2214 if (offset_in_page(addr)) {
2215 VM_BUG_ON(addr != -ENOMEM);
2217 info.low_limit = TASK_UNMAPPED_BASE;
2218 info.high_limit = mmap_end;
2219 addr = vm_unmapped_area(&info);
2227 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2228 unsigned long pgoff, unsigned long flags)
2230 unsigned long (*get_area)(struct file *, unsigned long,
2231 unsigned long, unsigned long, unsigned long);
2233 unsigned long error = arch_mmap_check(addr, len, flags);
2237 /* Careful about overflows.. */
2238 if (len > TASK_SIZE)
2241 get_area = current->mm->get_unmapped_area;
2243 if (file->f_op->get_unmapped_area)
2244 get_area = file->f_op->get_unmapped_area;
2245 } else if (flags & MAP_SHARED) {
2247 * mmap_region() will call shmem_zero_setup() to create a file,
2248 * so use shmem's get_unmapped_area in case it can be huge.
2249 * do_mmap() will clear pgoff, so match alignment.
2252 get_area = shmem_get_unmapped_area;
2255 addr = get_area(file, addr, len, pgoff, flags);
2256 if (IS_ERR_VALUE(addr))
2259 if (addr > TASK_SIZE - len)
2261 if (offset_in_page(addr))
2264 error = security_mmap_addr(addr);
2265 return error ? error : addr;
2268 EXPORT_SYMBOL(get_unmapped_area);
2270 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2271 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2273 struct rb_node *rb_node;
2274 struct vm_area_struct *vma;
2276 mmap_assert_locked(mm);
2277 /* Check the cache first. */
2278 vma = vmacache_find(mm, addr);
2282 rb_node = mm->mm_rb.rb_node;
2285 struct vm_area_struct *tmp;
2287 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2289 if (tmp->vm_end > addr) {
2291 if (tmp->vm_start <= addr)
2293 rb_node = rb_node->rb_left;
2295 rb_node = rb_node->rb_right;
2299 vmacache_update(addr, vma);
2303 EXPORT_SYMBOL(find_vma);
2306 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2308 struct vm_area_struct *
2309 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2310 struct vm_area_struct **pprev)
2312 struct vm_area_struct *vma;
2314 vma = find_vma(mm, addr);
2316 *pprev = vma->vm_prev;
2318 struct rb_node *rb_node = rb_last(&mm->mm_rb);
2320 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
2326 * Verify that the stack growth is acceptable and
2327 * update accounting. This is shared with both the
2328 * grow-up and grow-down cases.
2330 static int acct_stack_growth(struct vm_area_struct *vma,
2331 unsigned long size, unsigned long grow)
2333 struct mm_struct *mm = vma->vm_mm;
2334 unsigned long new_start;
2336 /* address space limit tests */
2337 if (!may_expand_vm(mm, vma->vm_flags, grow))
2340 /* Stack limit test */
2341 if (size > rlimit(RLIMIT_STACK))
2344 /* mlock limit tests */
2345 if (vma->vm_flags & VM_LOCKED) {
2346 unsigned long locked;
2347 unsigned long limit;
2348 locked = mm->locked_vm + grow;
2349 limit = rlimit(RLIMIT_MEMLOCK);
2350 limit >>= PAGE_SHIFT;
2351 if (locked > limit && !capable(CAP_IPC_LOCK))
2355 /* Check to ensure the stack will not grow into a hugetlb-only region */
2356 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2358 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2362 * Overcommit.. This must be the final test, as it will
2363 * update security statistics.
2365 if (security_vm_enough_memory_mm(mm, grow))
2371 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2373 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2374 * vma is the last one with address > vma->vm_end. Have to extend vma.
2376 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2378 struct mm_struct *mm = vma->vm_mm;
2379 struct vm_area_struct *next;
2380 unsigned long gap_addr;
2383 if (!(vma->vm_flags & VM_GROWSUP))
2386 /* Guard against exceeding limits of the address space. */
2387 address &= PAGE_MASK;
2388 if (address >= (TASK_SIZE & PAGE_MASK))
2390 address += PAGE_SIZE;
2392 /* Enforce stack_guard_gap */
2393 gap_addr = address + stack_guard_gap;
2395 /* Guard against overflow */
2396 if (gap_addr < address || gap_addr > TASK_SIZE)
2397 gap_addr = TASK_SIZE;
2399 next = vma->vm_next;
2400 if (next && next->vm_start < gap_addr && vma_is_accessible(next)) {
2401 if (!(next->vm_flags & VM_GROWSUP))
2403 /* Check that both stack segments have the same anon_vma? */
2406 /* We must make sure the anon_vma is allocated. */
2407 if (unlikely(anon_vma_prepare(vma)))
2411 * vma->vm_start/vm_end cannot change under us because the caller
2412 * is required to hold the mmap_lock in read mode. We need the
2413 * anon_vma lock to serialize against concurrent expand_stacks.
2415 anon_vma_lock_write(vma->anon_vma);
2417 /* Somebody else might have raced and expanded it already */
2418 if (address > vma->vm_end) {
2419 unsigned long size, grow;
2421 size = address - vma->vm_start;
2422 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2425 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2426 error = acct_stack_growth(vma, size, grow);
2429 * vma_gap_update() doesn't support concurrent
2430 * updates, but we only hold a shared mmap_lock
2431 * lock here, so we need to protect against
2432 * concurrent vma expansions.
2433 * anon_vma_lock_write() doesn't help here, as
2434 * we don't guarantee that all growable vmas
2435 * in a mm share the same root anon vma.
2436 * So, we reuse mm->page_table_lock to guard
2437 * against concurrent vma expansions.
2439 spin_lock(&mm->page_table_lock);
2440 if (vma->vm_flags & VM_LOCKED)
2441 mm->locked_vm += grow;
2442 vm_stat_account(mm, vma->vm_flags, grow);
2443 anon_vma_interval_tree_pre_update_vma(vma);
2444 vma->vm_end = address;
2445 anon_vma_interval_tree_post_update_vma(vma);
2447 vma_gap_update(vma->vm_next);
2449 mm->highest_vm_end = vm_end_gap(vma);
2450 spin_unlock(&mm->page_table_lock);
2452 perf_event_mmap(vma);
2456 anon_vma_unlock_write(vma->anon_vma);
2457 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2461 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2464 * vma is the first one with address < vma->vm_start. Have to extend vma.
2466 int expand_downwards(struct vm_area_struct *vma,
2467 unsigned long address)
2469 struct mm_struct *mm = vma->vm_mm;
2470 struct vm_area_struct *prev;
2473 address &= PAGE_MASK;
2474 if (address < mmap_min_addr)
2477 /* Enforce stack_guard_gap */
2478 prev = vma->vm_prev;
2479 /* Check that both stack segments have the same anon_vma? */
2480 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2481 vma_is_accessible(prev)) {
2482 if (address - prev->vm_end < stack_guard_gap)
2486 /* We must make sure the anon_vma is allocated. */
2487 if (unlikely(anon_vma_prepare(vma)))
2491 * vma->vm_start/vm_end cannot change under us because the caller
2492 * is required to hold the mmap_lock in read mode. We need the
2493 * anon_vma lock to serialize against concurrent expand_stacks.
2495 anon_vma_lock_write(vma->anon_vma);
2497 /* Somebody else might have raced and expanded it already */
2498 if (address < vma->vm_start) {
2499 unsigned long size, grow;
2501 size = vma->vm_end - address;
2502 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2505 if (grow <= vma->vm_pgoff) {
2506 error = acct_stack_growth(vma, size, grow);
2509 * vma_gap_update() doesn't support concurrent
2510 * updates, but we only hold a shared mmap_lock
2511 * lock here, so we need to protect against
2512 * concurrent vma expansions.
2513 * anon_vma_lock_write() doesn't help here, as
2514 * we don't guarantee that all growable vmas
2515 * in a mm share the same root anon vma.
2516 * So, we reuse mm->page_table_lock to guard
2517 * against concurrent vma expansions.
2519 spin_lock(&mm->page_table_lock);
2520 if (vma->vm_flags & VM_LOCKED)
2521 mm->locked_vm += grow;
2522 vm_stat_account(mm, vma->vm_flags, grow);
2523 anon_vma_interval_tree_pre_update_vma(vma);
2524 vma->vm_start = address;
2525 vma->vm_pgoff -= grow;
2526 anon_vma_interval_tree_post_update_vma(vma);
2527 vma_gap_update(vma);
2528 spin_unlock(&mm->page_table_lock);
2530 perf_event_mmap(vma);
2534 anon_vma_unlock_write(vma->anon_vma);
2535 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2540 /* enforced gap between the expanding stack and other mappings. */
2541 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2543 static int __init cmdline_parse_stack_guard_gap(char *p)
2548 val = simple_strtoul(p, &endptr, 10);
2550 stack_guard_gap = val << PAGE_SHIFT;
2554 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2556 #ifdef CONFIG_STACK_GROWSUP
2557 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2559 return expand_upwards(vma, address);
2562 struct vm_area_struct *
2563 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2565 struct vm_area_struct *vma, *prev;
2568 vma = find_vma_prev(mm, addr, &prev);
2569 if (vma && (vma->vm_start <= addr))
2571 /* don't alter vm_end if the coredump is running */
2572 if (!prev || expand_stack(prev, addr))
2574 if (prev->vm_flags & VM_LOCKED)
2575 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2579 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2581 return expand_downwards(vma, address);
2584 struct vm_area_struct *
2585 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2587 struct vm_area_struct *vma;
2588 unsigned long start;
2591 vma = find_vma(mm, addr);
2594 if (vma->vm_start <= addr)
2596 if (!(vma->vm_flags & VM_GROWSDOWN))
2598 start = vma->vm_start;
2599 if (expand_stack(vma, addr))
2601 if (vma->vm_flags & VM_LOCKED)
2602 populate_vma_page_range(vma, addr, start, NULL);
2607 EXPORT_SYMBOL_GPL(find_extend_vma);
2610 * Ok - we have the memory areas we should free on the vma list,
2611 * so release them, and do the vma updates.
2613 * Called with the mm semaphore held.
2615 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2617 unsigned long nr_accounted = 0;
2619 /* Update high watermark before we lower total_vm */
2620 update_hiwater_vm(mm);
2622 long nrpages = vma_pages(vma);
2624 if (vma->vm_flags & VM_ACCOUNT)
2625 nr_accounted += nrpages;
2626 vm_stat_account(mm, vma->vm_flags, -nrpages);
2627 vma = remove_vma(vma);
2629 vm_unacct_memory(nr_accounted);
2634 * Get rid of page table information in the indicated region.
2636 * Called with the mm semaphore held.
2638 static void unmap_region(struct mm_struct *mm,
2639 struct vm_area_struct *vma, struct vm_area_struct *prev,
2640 unsigned long start, unsigned long end)
2642 struct vm_area_struct *next = vma_next(mm, prev);
2643 struct mmu_gather tlb;
2644 struct vm_area_struct *cur_vma;
2647 tlb_gather_mmu(&tlb, mm);
2648 update_hiwater_rss(mm);
2649 unmap_vmas(&tlb, vma, start, end);
2652 * Ensure we have no stale TLB entries by the time this mapping is
2653 * removed from the rmap.
2654 * Note that we don't have to worry about nested flushes here because
2655 * we're holding the mm semaphore for removing the mapping - so any
2656 * concurrent flush in this region has to be coming through the rmap,
2657 * and we synchronize against that using the rmap lock.
2659 for (cur_vma = vma; cur_vma; cur_vma = cur_vma->vm_next) {
2660 if ((cur_vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) != 0) {
2661 tlb_flush_mmu(&tlb);
2666 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2667 next ? next->vm_start : USER_PGTABLES_CEILING);
2668 tlb_finish_mmu(&tlb);
2672 * Create a list of vma's touched by the unmap, removing them from the mm's
2673 * vma list as we go..
2676 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2677 struct vm_area_struct *prev, unsigned long end)
2679 struct vm_area_struct **insertion_point;
2680 struct vm_area_struct *tail_vma = NULL;
2682 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2683 vma->vm_prev = NULL;
2685 vma_rb_erase(vma, &mm->mm_rb);
2689 } while (vma && vma->vm_start < end);
2690 *insertion_point = vma;
2692 vma->vm_prev = prev;
2693 vma_gap_update(vma);
2695 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2696 tail_vma->vm_next = NULL;
2698 /* Kill the cache */
2699 vmacache_invalidate(mm);
2702 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2703 * VM_GROWSUP VMA. Such VMAs can change their size under
2704 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2706 if (vma && (vma->vm_flags & VM_GROWSDOWN))
2708 if (prev && (prev->vm_flags & VM_GROWSUP))
2714 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2715 * has already been checked or doesn't make sense to fail.
2717 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2718 unsigned long addr, int new_below)
2720 struct vm_area_struct *new;
2723 if (vma->vm_ops && vma->vm_ops->may_split) {
2724 err = vma->vm_ops->may_split(vma, addr);
2729 new = vm_area_dup(vma);
2736 new->vm_start = addr;
2737 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2740 err = vma_dup_policy(vma, new);
2744 err = anon_vma_clone(new, vma);
2749 get_file(new->vm_file);
2751 if (new->vm_ops && new->vm_ops->open)
2752 new->vm_ops->open(new);
2755 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2756 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2758 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2764 /* Clean everything up if vma_adjust failed. */
2765 if (new->vm_ops && new->vm_ops->close)
2766 new->vm_ops->close(new);
2769 unlink_anon_vmas(new);
2771 mpol_put(vma_policy(new));
2778 * Split a vma into two pieces at address 'addr', a new vma is allocated
2779 * either for the first part or the tail.
2781 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2782 unsigned long addr, int new_below)
2784 if (mm->map_count >= sysctl_max_map_count)
2787 return __split_vma(mm, vma, addr, new_below);
2791 unlock_range(struct vm_area_struct *start, unsigned long limit)
2793 struct mm_struct *mm = start->vm_mm;
2794 struct vm_area_struct *tmp = start;
2796 while (tmp && tmp->vm_start < limit) {
2797 if (tmp->vm_flags & VM_LOCKED) {
2798 mm->locked_vm -= vma_pages(tmp);
2799 munlock_vma_pages_all(tmp);
2806 /* Munmap is split into 2 main parts -- this part which finds
2807 * what needs doing, and the areas themselves, which do the
2808 * work. This now handles partial unmappings.
2809 * Jeremy Fitzhardinge <jeremy@goop.org>
2811 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2812 struct list_head *uf, bool downgrade)
2815 struct vm_area_struct *vma, *prev, *last;
2817 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2820 len = PAGE_ALIGN(len);
2826 * arch_unmap() might do unmaps itself. It must be called
2827 * and finish any rbtree manipulation before this code
2828 * runs and also starts to manipulate the rbtree.
2830 arch_unmap(mm, start, end);
2832 /* Find the first overlapping VMA where start < vma->vm_end */
2833 vma = find_vma_intersection(mm, start, end);
2836 prev = vma->vm_prev;
2839 * If we need to split any vma, do it now to save pain later.
2841 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2842 * unmapped vm_area_struct will remain in use: so lower split_vma
2843 * places tmp vma above, and higher split_vma places tmp vma below.
2845 if (start > vma->vm_start) {
2849 * Make sure that map_count on return from munmap() will
2850 * not exceed its limit; but let map_count go just above
2851 * its limit temporarily, to help free resources as expected.
2853 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2856 error = __split_vma(mm, vma, start, 0);
2862 /* Does it split the last one? */
2863 last = find_vma(mm, end);
2864 if (last && end > last->vm_start) {
2865 int error = __split_vma(mm, last, end, 1);
2869 vma = vma_next(mm, prev);
2873 * If userfaultfd_unmap_prep returns an error the vmas
2874 * will remain split, but userland will get a
2875 * highly unexpected error anyway. This is no
2876 * different than the case where the first of the two
2877 * __split_vma fails, but we don't undo the first
2878 * split, despite we could. This is unlikely enough
2879 * failure that it's not worth optimizing it for.
2881 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2887 * unlock any mlock()ed ranges before detaching vmas
2890 unlock_range(vma, end);
2892 /* Detach vmas from rbtree */
2893 if (!detach_vmas_to_be_unmapped(mm, vma, prev, end))
2897 mmap_write_downgrade(mm);
2899 unmap_region(mm, vma, prev, start, end);
2901 /* Fix up all other VM information */
2902 remove_vma_list(mm, vma);
2904 return downgrade ? 1 : 0;
2907 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2908 struct list_head *uf)
2910 return __do_munmap(mm, start, len, uf, false);
2913 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2916 struct mm_struct *mm = current->mm;
2919 if (mmap_write_lock_killable(mm))
2922 ret = __do_munmap(mm, start, len, &uf, downgrade);
2924 * Returning 1 indicates mmap_lock is downgraded.
2925 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2926 * it to 0 before return.
2929 mmap_read_unlock(mm);
2932 mmap_write_unlock(mm);
2934 userfaultfd_unmap_complete(mm, &uf);
2938 int vm_munmap(unsigned long start, size_t len)
2940 return __vm_munmap(start, len, false);
2942 EXPORT_SYMBOL(vm_munmap);
2944 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2946 addr = untagged_addr(addr);
2947 profile_munmap(addr);
2948 return __vm_munmap(addr, len, true);
2953 * Emulation of deprecated remap_file_pages() syscall.
2955 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2956 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2959 struct mm_struct *mm = current->mm;
2960 struct vm_area_struct *vma;
2961 unsigned long populate = 0;
2962 unsigned long ret = -EINVAL;
2965 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2966 current->comm, current->pid);
2970 start = start & PAGE_MASK;
2971 size = size & PAGE_MASK;
2973 if (start + size <= start)
2976 /* Does pgoff wrap? */
2977 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2980 if (mmap_write_lock_killable(mm))
2983 vma = vma_lookup(mm, start);
2985 if (!vma || !(vma->vm_flags & VM_SHARED))
2988 if (start + size > vma->vm_end) {
2989 struct vm_area_struct *next;
2991 for (next = vma->vm_next; next; next = next->vm_next) {
2992 /* hole between vmas ? */
2993 if (next->vm_start != next->vm_prev->vm_end)
2996 if (next->vm_file != vma->vm_file)
2999 if (next->vm_flags != vma->vm_flags)
3002 if (start + size <= next->vm_end)
3010 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3011 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3012 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3014 flags &= MAP_NONBLOCK;
3015 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3016 if (vma->vm_flags & VM_LOCKED)
3017 flags |= MAP_LOCKED;
3019 file = get_file(vma->vm_file);
3020 ret = do_mmap(vma->vm_file, start, size,
3021 prot, flags, pgoff, &populate, NULL);
3024 mmap_write_unlock(mm);
3026 mm_populate(ret, populate);
3027 if (!IS_ERR_VALUE(ret))
3033 * this is really a simplified "do_mmap". it only handles
3034 * anonymous maps. eventually we may be able to do some
3035 * brk-specific accounting here.
3037 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
3039 struct mm_struct *mm = current->mm;
3040 struct vm_area_struct *vma, *prev;
3041 struct rb_node **rb_link, *rb_parent;
3042 pgoff_t pgoff = addr >> PAGE_SHIFT;
3044 unsigned long mapped_addr;
3046 /* Until we need other flags, refuse anything except VM_EXEC. */
3047 if ((flags & (~VM_EXEC)) != 0)
3049 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3051 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
3052 if (IS_ERR_VALUE(mapped_addr))
3055 error = mlock_future_check(mm, mm->def_flags, len);
3059 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
3060 if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
3063 /* Check against address space limits *after* clearing old maps... */
3064 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3067 if (mm->map_count > sysctl_max_map_count)
3070 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3073 /* Can we just expand an old private anonymous mapping? */
3074 vma = vma_merge(mm, prev, addr, addr + len, flags,
3075 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
3080 * create a vma struct for an anonymous mapping
3082 vma = vm_area_alloc(mm);
3084 vm_unacct_memory(len >> PAGE_SHIFT);
3088 vma_set_anonymous(vma);
3089 vma->vm_start = addr;
3090 vma->vm_end = addr + len;
3091 vma->vm_pgoff = pgoff;
3092 vma->vm_flags = flags;
3093 vma->vm_page_prot = vm_get_page_prot(flags);
3094 vma_link(mm, vma, prev, rb_link, rb_parent);
3096 perf_event_mmap(vma);
3097 mm->total_vm += len >> PAGE_SHIFT;
3098 mm->data_vm += len >> PAGE_SHIFT;
3099 if (flags & VM_LOCKED)
3100 mm->locked_vm += (len >> PAGE_SHIFT);
3101 vma->vm_flags |= VM_SOFTDIRTY;
3105 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3107 struct mm_struct *mm = current->mm;
3113 len = PAGE_ALIGN(request);
3119 if (mmap_write_lock_killable(mm))
3122 ret = do_brk_flags(addr, len, flags, &uf);
3123 populate = ((mm->def_flags & VM_LOCKED) != 0);
3124 mmap_write_unlock(mm);
3125 userfaultfd_unmap_complete(mm, &uf);
3126 if (populate && !ret)
3127 mm_populate(addr, len);
3130 EXPORT_SYMBOL(vm_brk_flags);
3132 int vm_brk(unsigned long addr, unsigned long len)
3134 return vm_brk_flags(addr, len, 0);
3136 EXPORT_SYMBOL(vm_brk);
3138 /* Release all mmaps. */
3139 void exit_mmap(struct mm_struct *mm)
3141 struct mmu_gather tlb;
3142 struct vm_area_struct *vma;
3143 unsigned long nr_accounted = 0;
3145 /* mm's last user has gone, and its about to be pulled down */
3146 mmu_notifier_release(mm);
3148 if (unlikely(mm_is_oom_victim(mm))) {
3150 * Manually reap the mm to free as much memory as possible.
3151 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3152 * this mm from further consideration. Taking mm->mmap_lock for
3153 * write after setting MMF_OOM_SKIP will guarantee that the oom
3154 * reaper will not run on this mm again after mmap_lock is
3157 * Nothing can be holding mm->mmap_lock here and the above call
3158 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3159 * __oom_reap_task_mm() will not block.
3161 * This needs to be done before calling munlock_vma_pages_all(),
3162 * which clears VM_LOCKED, otherwise the oom reaper cannot
3165 (void)__oom_reap_task_mm(mm);
3167 set_bit(MMF_OOM_SKIP, &mm->flags);
3168 mmap_write_lock(mm);
3169 mmap_write_unlock(mm);
3173 unlock_range(mm->mmap, ULONG_MAX);
3178 if (!vma) /* Can happen if dup_mmap() received an OOM */
3183 tlb_gather_mmu_fullmm(&tlb, mm);
3184 /* update_hiwater_rss(mm) here? but nobody should be looking */
3185 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3186 unmap_vmas(&tlb, vma, 0, -1);
3187 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3188 tlb_finish_mmu(&tlb);
3191 * Walk the list again, actually closing and freeing it,
3192 * with preemption enabled, without holding any MM locks.
3195 if (vma->vm_flags & VM_ACCOUNT)
3196 nr_accounted += vma_pages(vma);
3197 vma = remove_vma(vma);
3200 vm_unacct_memory(nr_accounted);
3203 /* Insert vm structure into process list sorted by address
3204 * and into the inode's i_mmap tree. If vm_file is non-NULL
3205 * then i_mmap_rwsem is taken here.
3207 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3209 struct vm_area_struct *prev;
3210 struct rb_node **rb_link, *rb_parent;
3212 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3213 &prev, &rb_link, &rb_parent))
3215 if ((vma->vm_flags & VM_ACCOUNT) &&
3216 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3220 * The vm_pgoff of a purely anonymous vma should be irrelevant
3221 * until its first write fault, when page's anon_vma and index
3222 * are set. But now set the vm_pgoff it will almost certainly
3223 * end up with (unless mremap moves it elsewhere before that
3224 * first wfault), so /proc/pid/maps tells a consistent story.
3226 * By setting it to reflect the virtual start address of the
3227 * vma, merges and splits can happen in a seamless way, just
3228 * using the existing file pgoff checks and manipulations.
3229 * Similarly in do_mmap and in do_brk_flags.
3231 if (vma_is_anonymous(vma)) {
3232 BUG_ON(vma->anon_vma);
3233 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3236 vma_link(mm, vma, prev, rb_link, rb_parent);
3241 * Copy the vma structure to a new location in the same mm,
3242 * prior to moving page table entries, to effect an mremap move.
3244 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3245 unsigned long addr, unsigned long len, pgoff_t pgoff,
3246 bool *need_rmap_locks)
3248 struct vm_area_struct *vma = *vmap;
3249 unsigned long vma_start = vma->vm_start;
3250 struct mm_struct *mm = vma->vm_mm;
3251 struct vm_area_struct *new_vma, *prev;
3252 struct rb_node **rb_link, *rb_parent;
3253 bool faulted_in_anon_vma = true;
3256 * If anonymous vma has not yet been faulted, update new pgoff
3257 * to match new location, to increase its chance of merging.
3259 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3260 pgoff = addr >> PAGE_SHIFT;
3261 faulted_in_anon_vma = false;
3264 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3265 return NULL; /* should never get here */
3266 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3267 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3268 vma->vm_userfaultfd_ctx);
3271 * Source vma may have been merged into new_vma
3273 if (unlikely(vma_start >= new_vma->vm_start &&
3274 vma_start < new_vma->vm_end)) {
3276 * The only way we can get a vma_merge with
3277 * self during an mremap is if the vma hasn't
3278 * been faulted in yet and we were allowed to
3279 * reset the dst vma->vm_pgoff to the
3280 * destination address of the mremap to allow
3281 * the merge to happen. mremap must change the
3282 * vm_pgoff linearity between src and dst vmas
3283 * (in turn preventing a vma_merge) to be
3284 * safe. It is only safe to keep the vm_pgoff
3285 * linear if there are no pages mapped yet.
3287 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3288 *vmap = vma = new_vma;
3290 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3292 new_vma = vm_area_dup(vma);
3295 new_vma->vm_start = addr;
3296 new_vma->vm_end = addr + len;
3297 new_vma->vm_pgoff = pgoff;
3298 if (vma_dup_policy(vma, new_vma))
3300 if (anon_vma_clone(new_vma, vma))
3301 goto out_free_mempol;
3302 if (new_vma->vm_file)
3303 get_file(new_vma->vm_file);
3304 if (new_vma->vm_ops && new_vma->vm_ops->open)
3305 new_vma->vm_ops->open(new_vma);
3306 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3307 *need_rmap_locks = false;
3312 mpol_put(vma_policy(new_vma));
3314 vm_area_free(new_vma);
3320 * Return true if the calling process may expand its vm space by the passed
3323 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3325 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3328 if (is_data_mapping(flags) &&
3329 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3330 /* Workaround for Valgrind */
3331 if (rlimit(RLIMIT_DATA) == 0 &&
3332 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3335 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3336 current->comm, current->pid,
3337 (mm->data_vm + npages) << PAGE_SHIFT,
3338 rlimit(RLIMIT_DATA),
3339 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3341 if (!ignore_rlimit_data)
3348 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3350 mm->total_vm += npages;
3352 if (is_exec_mapping(flags))
3353 mm->exec_vm += npages;
3354 else if (is_stack_mapping(flags))
3355 mm->stack_vm += npages;
3356 else if (is_data_mapping(flags))
3357 mm->data_vm += npages;
3360 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3363 * Having a close hook prevents vma merging regardless of flags.
3365 static void special_mapping_close(struct vm_area_struct *vma)
3369 static const char *special_mapping_name(struct vm_area_struct *vma)
3371 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3374 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3376 struct vm_special_mapping *sm = new_vma->vm_private_data;
3378 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3382 return sm->mremap(sm, new_vma);
3387 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3390 * Forbid splitting special mappings - kernel has expectations over
3391 * the number of pages in mapping. Together with VM_DONTEXPAND
3392 * the size of vma should stay the same over the special mapping's
3398 static const struct vm_operations_struct special_mapping_vmops = {
3399 .close = special_mapping_close,
3400 .fault = special_mapping_fault,
3401 .mremap = special_mapping_mremap,
3402 .name = special_mapping_name,
3403 /* vDSO code relies that VVAR can't be accessed remotely */
3405 .may_split = special_mapping_split,
3408 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3409 .close = special_mapping_close,
3410 .fault = special_mapping_fault,
3413 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3415 struct vm_area_struct *vma = vmf->vma;
3417 struct page **pages;
3419 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3420 pages = vma->vm_private_data;
3422 struct vm_special_mapping *sm = vma->vm_private_data;
3425 return sm->fault(sm, vmf->vma, vmf);
3430 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3434 struct page *page = *pages;
3440 return VM_FAULT_SIGBUS;
3443 static struct vm_area_struct *__install_special_mapping(
3444 struct mm_struct *mm,
3445 unsigned long addr, unsigned long len,
3446 unsigned long vm_flags, void *priv,
3447 const struct vm_operations_struct *ops)
3450 struct vm_area_struct *vma;
3452 vma = vm_area_alloc(mm);
3453 if (unlikely(vma == NULL))
3454 return ERR_PTR(-ENOMEM);
3456 vma->vm_start = addr;
3457 vma->vm_end = addr + len;
3459 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3460 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3463 vma->vm_private_data = priv;
3465 ret = insert_vm_struct(mm, vma);
3469 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3471 perf_event_mmap(vma);
3477 return ERR_PTR(ret);
3480 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3481 const struct vm_special_mapping *sm)
3483 return vma->vm_private_data == sm &&
3484 (vma->vm_ops == &special_mapping_vmops ||
3485 vma->vm_ops == &legacy_special_mapping_vmops);
3489 * Called with mm->mmap_lock held for writing.
3490 * Insert a new vma covering the given region, with the given flags.
3491 * Its pages are supplied by the given array of struct page *.
3492 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3493 * The region past the last page supplied will always produce SIGBUS.
3494 * The array pointer and the pages it points to are assumed to stay alive
3495 * for as long as this mapping might exist.
3497 struct vm_area_struct *_install_special_mapping(
3498 struct mm_struct *mm,
3499 unsigned long addr, unsigned long len,
3500 unsigned long vm_flags, const struct vm_special_mapping *spec)
3502 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3503 &special_mapping_vmops);
3506 int install_special_mapping(struct mm_struct *mm,
3507 unsigned long addr, unsigned long len,
3508 unsigned long vm_flags, struct page **pages)
3510 struct vm_area_struct *vma = __install_special_mapping(
3511 mm, addr, len, vm_flags, (void *)pages,
3512 &legacy_special_mapping_vmops);
3514 return PTR_ERR_OR_ZERO(vma);
3517 static DEFINE_MUTEX(mm_all_locks_mutex);
3519 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3521 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3523 * The LSB of head.next can't change from under us
3524 * because we hold the mm_all_locks_mutex.
3526 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3528 * We can safely modify head.next after taking the
3529 * anon_vma->root->rwsem. If some other vma in this mm shares
3530 * the same anon_vma we won't take it again.
3532 * No need of atomic instructions here, head.next
3533 * can't change from under us thanks to the
3534 * anon_vma->root->rwsem.
3536 if (__test_and_set_bit(0, (unsigned long *)
3537 &anon_vma->root->rb_root.rb_root.rb_node))
3542 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3544 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3546 * AS_MM_ALL_LOCKS can't change from under us because
3547 * we hold the mm_all_locks_mutex.
3549 * Operations on ->flags have to be atomic because
3550 * even if AS_MM_ALL_LOCKS is stable thanks to the
3551 * mm_all_locks_mutex, there may be other cpus
3552 * changing other bitflags in parallel to us.
3554 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3556 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3561 * This operation locks against the VM for all pte/vma/mm related
3562 * operations that could ever happen on a certain mm. This includes
3563 * vmtruncate, try_to_unmap, and all page faults.
3565 * The caller must take the mmap_lock in write mode before calling
3566 * mm_take_all_locks(). The caller isn't allowed to release the
3567 * mmap_lock until mm_drop_all_locks() returns.
3569 * mmap_lock in write mode is required in order to block all operations
3570 * that could modify pagetables and free pages without need of
3571 * altering the vma layout. It's also needed in write mode to avoid new
3572 * anon_vmas to be associated with existing vmas.
3574 * A single task can't take more than one mm_take_all_locks() in a row
3575 * or it would deadlock.
3577 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3578 * mapping->flags avoid to take the same lock twice, if more than one
3579 * vma in this mm is backed by the same anon_vma or address_space.
3581 * We take locks in following order, accordingly to comment at beginning
3583 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3585 * - all i_mmap_rwsem locks;
3586 * - all anon_vma->rwseml
3588 * We can take all locks within these types randomly because the VM code
3589 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3590 * mm_all_locks_mutex.
3592 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3593 * that may have to take thousand of locks.
3595 * mm_take_all_locks() can fail if it's interrupted by signals.
3597 int mm_take_all_locks(struct mm_struct *mm)
3599 struct vm_area_struct *vma;
3600 struct anon_vma_chain *avc;
3602 BUG_ON(mmap_read_trylock(mm));
3604 mutex_lock(&mm_all_locks_mutex);
3606 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3607 if (signal_pending(current))
3609 if (vma->vm_file && vma->vm_file->f_mapping &&
3610 is_vm_hugetlb_page(vma))
3611 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3614 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3615 if (signal_pending(current))
3617 if (vma->vm_file && vma->vm_file->f_mapping &&
3618 !is_vm_hugetlb_page(vma))
3619 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3622 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3623 if (signal_pending(current))
3626 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3627 vm_lock_anon_vma(mm, avc->anon_vma);
3633 mm_drop_all_locks(mm);
3637 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3639 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3641 * The LSB of head.next can't change to 0 from under
3642 * us because we hold the mm_all_locks_mutex.
3644 * We must however clear the bitflag before unlocking
3645 * the vma so the users using the anon_vma->rb_root will
3646 * never see our bitflag.
3648 * No need of atomic instructions here, head.next
3649 * can't change from under us until we release the
3650 * anon_vma->root->rwsem.
3652 if (!__test_and_clear_bit(0, (unsigned long *)
3653 &anon_vma->root->rb_root.rb_root.rb_node))
3655 anon_vma_unlock_write(anon_vma);
3659 static void vm_unlock_mapping(struct address_space *mapping)
3661 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3663 * AS_MM_ALL_LOCKS can't change to 0 from under us
3664 * because we hold the mm_all_locks_mutex.
3666 i_mmap_unlock_write(mapping);
3667 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3674 * The mmap_lock cannot be released by the caller until
3675 * mm_drop_all_locks() returns.
3677 void mm_drop_all_locks(struct mm_struct *mm)
3679 struct vm_area_struct *vma;
3680 struct anon_vma_chain *avc;
3682 BUG_ON(mmap_read_trylock(mm));
3683 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3685 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3687 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3688 vm_unlock_anon_vma(avc->anon_vma);
3689 if (vma->vm_file && vma->vm_file->f_mapping)
3690 vm_unlock_mapping(vma->vm_file->f_mapping);
3693 mutex_unlock(&mm_all_locks_mutex);
3697 * initialise the percpu counter for VM
3699 void __init mmap_init(void)
3703 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3708 * Initialise sysctl_user_reserve_kbytes.
3710 * This is intended to prevent a user from starting a single memory hogging
3711 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3714 * The default value is min(3% of free memory, 128MB)
3715 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3717 static int init_user_reserve(void)
3719 unsigned long free_kbytes;
3721 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3723 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3726 subsys_initcall(init_user_reserve);
3729 * Initialise sysctl_admin_reserve_kbytes.
3731 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3732 * to log in and kill a memory hogging process.
3734 * Systems with more than 256MB will reserve 8MB, enough to recover
3735 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3736 * only reserve 3% of free pages by default.
3738 static int init_admin_reserve(void)
3740 unsigned long free_kbytes;
3742 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3744 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3747 subsys_initcall(init_admin_reserve);
3750 * Reinititalise user and admin reserves if memory is added or removed.
3752 * The default user reserve max is 128MB, and the default max for the
3753 * admin reserve is 8MB. These are usually, but not always, enough to
3754 * enable recovery from a memory hogging process using login/sshd, a shell,
3755 * and tools like top. It may make sense to increase or even disable the
3756 * reserve depending on the existence of swap or variations in the recovery
3757 * tools. So, the admin may have changed them.
3759 * If memory is added and the reserves have been eliminated or increased above
3760 * the default max, then we'll trust the admin.
3762 * If memory is removed and there isn't enough free memory, then we
3763 * need to reset the reserves.
3765 * Otherwise keep the reserve set by the admin.
3767 static int reserve_mem_notifier(struct notifier_block *nb,
3768 unsigned long action, void *data)
3770 unsigned long tmp, free_kbytes;
3774 /* Default max is 128MB. Leave alone if modified by operator. */
3775 tmp = sysctl_user_reserve_kbytes;
3776 if (0 < tmp && tmp < (1UL << 17))
3777 init_user_reserve();
3779 /* Default max is 8MB. Leave alone if modified by operator. */
3780 tmp = sysctl_admin_reserve_kbytes;
3781 if (0 < tmp && tmp < (1UL << 13))
3782 init_admin_reserve();
3786 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3788 if (sysctl_user_reserve_kbytes > free_kbytes) {
3789 init_user_reserve();
3790 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3791 sysctl_user_reserve_kbytes);
3794 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3795 init_admin_reserve();
3796 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3797 sysctl_admin_reserve_kbytes);
3806 static struct notifier_block reserve_mem_nb = {
3807 .notifier_call = reserve_mem_notifier,
3810 static int __meminit init_reserve_notifier(void)
3812 if (register_hotmemory_notifier(&reserve_mem_nb))
3813 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3817 subsys_initcall(init_reserve_notifier);