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_DENYWRITE)
152 allow_write_access(file);
153 if (vma->vm_flags & VM_SHARED)
154 mapping_unmap_writable(mapping);
156 flush_dcache_mmap_lock(mapping);
157 vma_interval_tree_remove(vma, &mapping->i_mmap);
158 flush_dcache_mmap_unlock(mapping);
162 * Unlink a file-based vm structure from its interval tree, to hide
163 * vma from rmap and vmtruncate before freeing its page tables.
165 void unlink_file_vma(struct vm_area_struct *vma)
167 struct file *file = vma->vm_file;
170 struct address_space *mapping = file->f_mapping;
171 i_mmap_lock_write(mapping);
172 __remove_shared_vm_struct(vma, file, mapping);
173 i_mmap_unlock_write(mapping);
178 * Close a vm structure and free it, returning the next.
180 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
182 struct vm_area_struct *next = vma->vm_next;
185 if (vma->vm_ops && vma->vm_ops->close)
186 vma->vm_ops->close(vma);
189 mpol_put(vma_policy(vma));
194 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags,
195 struct list_head *uf);
196 SYSCALL_DEFINE1(brk, unsigned long, brk)
198 unsigned long newbrk, oldbrk, origbrk;
199 struct mm_struct *mm = current->mm;
200 struct vm_area_struct *next;
201 unsigned long min_brk;
203 bool downgraded = false;
206 if (mmap_write_lock_killable(mm))
211 #ifdef CONFIG_COMPAT_BRK
213 * CONFIG_COMPAT_BRK can still be overridden by setting
214 * randomize_va_space to 2, which will still cause mm->start_brk
215 * to be arbitrarily shifted
217 if (current->brk_randomized)
218 min_brk = mm->start_brk;
220 min_brk = mm->end_data;
222 min_brk = mm->start_brk;
228 * Check against rlimit here. If this check is done later after the test
229 * of oldbrk with newbrk then it can escape the test and let the data
230 * segment grow beyond its set limit the in case where the limit is
231 * not page aligned -Ram Gupta
233 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
234 mm->end_data, mm->start_data))
237 newbrk = PAGE_ALIGN(brk);
238 oldbrk = PAGE_ALIGN(mm->brk);
239 if (oldbrk == newbrk) {
245 * Always allow shrinking brk.
246 * __do_munmap() may downgrade mmap_lock to read.
248 if (brk <= mm->brk) {
252 * mm->brk must to be protected by write mmap_lock so update it
253 * before downgrading mmap_lock. When __do_munmap() fails,
254 * mm->brk will be restored from origbrk.
257 ret = __do_munmap(mm, newbrk, oldbrk-newbrk, &uf, true);
261 } else if (ret == 1) {
267 /* Check against existing mmap mappings. */
268 next = find_vma(mm, oldbrk);
269 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
272 /* Ok, looks good - let it rip. */
273 if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
278 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
280 mmap_read_unlock(mm);
282 mmap_write_unlock(mm);
283 userfaultfd_unmap_complete(mm, &uf);
285 mm_populate(oldbrk, newbrk - oldbrk);
289 mmap_write_unlock(mm);
293 static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
295 unsigned long gap, prev_end;
298 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
299 * allow two stack_guard_gaps between them here, and when choosing
300 * an unmapped area; whereas when expanding we only require one.
301 * That's a little inconsistent, but keeps the code here simpler.
303 gap = vm_start_gap(vma);
305 prev_end = vm_end_gap(vma->vm_prev);
314 #ifdef CONFIG_DEBUG_VM_RB
315 static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma)
317 unsigned long max = vma_compute_gap(vma), subtree_gap;
318 if (vma->vm_rb.rb_left) {
319 subtree_gap = rb_entry(vma->vm_rb.rb_left,
320 struct vm_area_struct, vm_rb)->rb_subtree_gap;
321 if (subtree_gap > max)
324 if (vma->vm_rb.rb_right) {
325 subtree_gap = rb_entry(vma->vm_rb.rb_right,
326 struct vm_area_struct, vm_rb)->rb_subtree_gap;
327 if (subtree_gap > max)
333 static int browse_rb(struct mm_struct *mm)
335 struct rb_root *root = &mm->mm_rb;
336 int i = 0, j, bug = 0;
337 struct rb_node *nd, *pn = NULL;
338 unsigned long prev = 0, pend = 0;
340 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
341 struct vm_area_struct *vma;
342 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
343 if (vma->vm_start < prev) {
344 pr_emerg("vm_start %lx < prev %lx\n",
345 vma->vm_start, prev);
348 if (vma->vm_start < pend) {
349 pr_emerg("vm_start %lx < pend %lx\n",
350 vma->vm_start, pend);
353 if (vma->vm_start > vma->vm_end) {
354 pr_emerg("vm_start %lx > vm_end %lx\n",
355 vma->vm_start, vma->vm_end);
358 spin_lock(&mm->page_table_lock);
359 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
360 pr_emerg("free gap %lx, correct %lx\n",
362 vma_compute_subtree_gap(vma));
365 spin_unlock(&mm->page_table_lock);
368 prev = vma->vm_start;
372 for (nd = pn; nd; nd = rb_prev(nd))
375 pr_emerg("backwards %d, forwards %d\n", j, i);
381 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
385 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
386 struct vm_area_struct *vma;
387 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
388 VM_BUG_ON_VMA(vma != ignore &&
389 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
394 static void validate_mm(struct mm_struct *mm)
398 unsigned long highest_address = 0;
399 struct vm_area_struct *vma = mm->mmap;
402 struct anon_vma *anon_vma = vma->anon_vma;
403 struct anon_vma_chain *avc;
406 anon_vma_lock_read(anon_vma);
407 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
408 anon_vma_interval_tree_verify(avc);
409 anon_vma_unlock_read(anon_vma);
412 highest_address = vm_end_gap(vma);
416 if (i != mm->map_count) {
417 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
420 if (highest_address != mm->highest_vm_end) {
421 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
422 mm->highest_vm_end, highest_address);
426 if (i != mm->map_count) {
428 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
431 VM_BUG_ON_MM(bug, mm);
434 #define validate_mm_rb(root, ignore) do { } while (0)
435 #define validate_mm(mm) do { } while (0)
438 RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks,
439 struct vm_area_struct, vm_rb,
440 unsigned long, rb_subtree_gap, vma_compute_gap)
443 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
444 * vma->vm_prev->vm_end values changed, without modifying the vma's position
447 static void vma_gap_update(struct vm_area_struct *vma)
450 * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
451 * a callback function that does exactly what we want.
453 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
456 static inline void vma_rb_insert(struct vm_area_struct *vma,
457 struct rb_root *root)
459 /* All rb_subtree_gap values must be consistent prior to insertion */
460 validate_mm_rb(root, NULL);
462 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
465 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
468 * Note rb_erase_augmented is a fairly large inline function,
469 * so make sure we instantiate it only once with our desired
470 * augmented rbtree callbacks.
472 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
475 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
476 struct rb_root *root,
477 struct vm_area_struct *ignore)
480 * All rb_subtree_gap values must be consistent prior to erase,
481 * with the possible exception of
483 * a. the "next" vma being erased if next->vm_start was reduced in
484 * __vma_adjust() -> __vma_unlink()
485 * b. the vma being erased in detach_vmas_to_be_unmapped() ->
488 validate_mm_rb(root, ignore);
490 __vma_rb_erase(vma, root);
493 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
494 struct rb_root *root)
496 vma_rb_erase_ignore(vma, root, vma);
500 * vma has some anon_vma assigned, and is already inserted on that
501 * anon_vma's interval trees.
503 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
504 * vma must be removed from the anon_vma's interval trees using
505 * anon_vma_interval_tree_pre_update_vma().
507 * After the update, the vma will be reinserted using
508 * anon_vma_interval_tree_post_update_vma().
510 * The entire update must be protected by exclusive mmap_lock and by
511 * the root anon_vma's mutex.
514 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
516 struct anon_vma_chain *avc;
518 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
519 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
523 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
525 struct anon_vma_chain *avc;
527 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
528 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
531 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
532 unsigned long end, struct vm_area_struct **pprev,
533 struct rb_node ***rb_link, struct rb_node **rb_parent)
535 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
537 __rb_link = &mm->mm_rb.rb_node;
538 rb_prev = __rb_parent = NULL;
541 struct vm_area_struct *vma_tmp;
543 __rb_parent = *__rb_link;
544 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
546 if (vma_tmp->vm_end > addr) {
547 /* Fail if an existing vma overlaps the area */
548 if (vma_tmp->vm_start < end)
550 __rb_link = &__rb_parent->rb_left;
552 rb_prev = __rb_parent;
553 __rb_link = &__rb_parent->rb_right;
559 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
560 *rb_link = __rb_link;
561 *rb_parent = __rb_parent;
566 * vma_next() - Get the next VMA.
567 * @mm: The mm_struct.
568 * @vma: The current vma.
570 * If @vma is NULL, return the first vma in the mm.
572 * Returns: The next VMA after @vma.
574 static inline struct vm_area_struct *vma_next(struct mm_struct *mm,
575 struct vm_area_struct *vma)
584 * munmap_vma_range() - munmap VMAs that overlap a range.
586 * @start: The start of the range.
587 * @len: The length of the range.
588 * @pprev: pointer to the pointer that will be set to previous vm_area_struct
589 * @rb_link: the rb_node
590 * @rb_parent: the parent rb_node
592 * Find all the vm_area_struct that overlap from @start to
593 * @end and munmap them. Set @pprev to the previous vm_area_struct.
595 * Returns: -ENOMEM on munmap failure or 0 on success.
598 munmap_vma_range(struct mm_struct *mm, unsigned long start, unsigned long len,
599 struct vm_area_struct **pprev, struct rb_node ***link,
600 struct rb_node **parent, struct list_head *uf)
603 while (find_vma_links(mm, start, start + len, pprev, link, parent))
604 if (do_munmap(mm, start, len, uf))
609 static unsigned long count_vma_pages_range(struct mm_struct *mm,
610 unsigned long addr, unsigned long end)
612 unsigned long nr_pages = 0;
613 struct vm_area_struct *vma;
615 /* Find first overlapping mapping */
616 vma = find_vma_intersection(mm, addr, end);
620 nr_pages = (min(end, vma->vm_end) -
621 max(addr, vma->vm_start)) >> PAGE_SHIFT;
623 /* Iterate over the rest of the overlaps */
624 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
625 unsigned long overlap_len;
627 if (vma->vm_start > end)
630 overlap_len = min(end, vma->vm_end) - vma->vm_start;
631 nr_pages += overlap_len >> PAGE_SHIFT;
637 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
638 struct rb_node **rb_link, struct rb_node *rb_parent)
640 /* Update tracking information for the gap following the new vma. */
642 vma_gap_update(vma->vm_next);
644 mm->highest_vm_end = vm_end_gap(vma);
647 * vma->vm_prev wasn't known when we followed the rbtree to find the
648 * correct insertion point for that vma. As a result, we could not
649 * update the vma vm_rb parents rb_subtree_gap values on the way down.
650 * So, we first insert the vma with a zero rb_subtree_gap value
651 * (to be consistent with what we did on the way down), and then
652 * immediately update the gap to the correct value. Finally we
653 * rebalance the rbtree after all augmented values have been set.
655 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
656 vma->rb_subtree_gap = 0;
658 vma_rb_insert(vma, &mm->mm_rb);
661 static void __vma_link_file(struct vm_area_struct *vma)
667 struct address_space *mapping = file->f_mapping;
669 if (vma->vm_flags & VM_DENYWRITE)
670 put_write_access(file_inode(file));
671 if (vma->vm_flags & VM_SHARED)
672 mapping_allow_writable(mapping);
674 flush_dcache_mmap_lock(mapping);
675 vma_interval_tree_insert(vma, &mapping->i_mmap);
676 flush_dcache_mmap_unlock(mapping);
681 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
682 struct vm_area_struct *prev, struct rb_node **rb_link,
683 struct rb_node *rb_parent)
685 __vma_link_list(mm, vma, prev);
686 __vma_link_rb(mm, vma, rb_link, rb_parent);
689 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
690 struct vm_area_struct *prev, struct rb_node **rb_link,
691 struct rb_node *rb_parent)
693 struct address_space *mapping = NULL;
696 mapping = vma->vm_file->f_mapping;
697 i_mmap_lock_write(mapping);
700 __vma_link(mm, vma, prev, rb_link, rb_parent);
701 __vma_link_file(vma);
704 i_mmap_unlock_write(mapping);
711 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
712 * mm's list and rbtree. It has already been inserted into the interval tree.
714 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
716 struct vm_area_struct *prev;
717 struct rb_node **rb_link, *rb_parent;
719 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
720 &prev, &rb_link, &rb_parent))
722 __vma_link(mm, vma, prev, rb_link, rb_parent);
726 static __always_inline void __vma_unlink(struct mm_struct *mm,
727 struct vm_area_struct *vma,
728 struct vm_area_struct *ignore)
730 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
731 __vma_unlink_list(mm, vma);
733 vmacache_invalidate(mm);
737 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
738 * is already present in an i_mmap tree without adjusting the tree.
739 * The following helper function should be used when such adjustments
740 * are necessary. The "insert" vma (if any) is to be inserted
741 * before we drop the necessary locks.
743 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
744 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
745 struct vm_area_struct *expand)
747 struct mm_struct *mm = vma->vm_mm;
748 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
749 struct address_space *mapping = NULL;
750 struct rb_root_cached *root = NULL;
751 struct anon_vma *anon_vma = NULL;
752 struct file *file = vma->vm_file;
753 bool start_changed = false, end_changed = false;
754 long adjust_next = 0;
757 if (next && !insert) {
758 struct vm_area_struct *exporter = NULL, *importer = NULL;
760 if (end >= next->vm_end) {
762 * vma expands, overlapping all the next, and
763 * perhaps the one after too (mprotect case 6).
764 * The only other cases that gets here are
765 * case 1, case 7 and case 8.
767 if (next == expand) {
769 * The only case where we don't expand "vma"
770 * and we expand "next" instead is case 8.
772 VM_WARN_ON(end != next->vm_end);
774 * remove_next == 3 means we're
775 * removing "vma" and that to do so we
776 * swapped "vma" and "next".
779 VM_WARN_ON(file != next->vm_file);
782 VM_WARN_ON(expand != vma);
784 * case 1, 6, 7, remove_next == 2 is case 6,
785 * remove_next == 1 is case 1 or 7.
787 remove_next = 1 + (end > next->vm_end);
788 VM_WARN_ON(remove_next == 2 &&
789 end != next->vm_next->vm_end);
790 /* trim end to next, for case 6 first pass */
798 * If next doesn't have anon_vma, import from vma after
799 * next, if the vma overlaps with it.
801 if (remove_next == 2 && !next->anon_vma)
802 exporter = next->vm_next;
804 } else if (end > next->vm_start) {
806 * vma expands, overlapping part of the next:
807 * mprotect case 5 shifting the boundary up.
809 adjust_next = (end - next->vm_start);
812 VM_WARN_ON(expand != importer);
813 } else if (end < vma->vm_end) {
815 * vma shrinks, and !insert tells it's not
816 * split_vma inserting another: so it must be
817 * mprotect case 4 shifting the boundary down.
819 adjust_next = -(vma->vm_end - end);
822 VM_WARN_ON(expand != importer);
826 * Easily overlooked: when mprotect shifts the boundary,
827 * make sure the expanding vma has anon_vma set if the
828 * shrinking vma had, to cover any anon pages imported.
830 if (exporter && exporter->anon_vma && !importer->anon_vma) {
833 importer->anon_vma = exporter->anon_vma;
834 error = anon_vma_clone(importer, exporter);
840 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
843 mapping = file->f_mapping;
844 root = &mapping->i_mmap;
845 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
848 uprobe_munmap(next, next->vm_start, next->vm_end);
850 i_mmap_lock_write(mapping);
853 * Put into interval tree now, so instantiated pages
854 * are visible to arm/parisc __flush_dcache_page
855 * throughout; but we cannot insert into address
856 * space until vma start or end is updated.
858 __vma_link_file(insert);
862 anon_vma = vma->anon_vma;
863 if (!anon_vma && adjust_next)
864 anon_vma = next->anon_vma;
866 VM_WARN_ON(adjust_next && next->anon_vma &&
867 anon_vma != next->anon_vma);
868 anon_vma_lock_write(anon_vma);
869 anon_vma_interval_tree_pre_update_vma(vma);
871 anon_vma_interval_tree_pre_update_vma(next);
875 flush_dcache_mmap_lock(mapping);
876 vma_interval_tree_remove(vma, root);
878 vma_interval_tree_remove(next, root);
881 if (start != vma->vm_start) {
882 vma->vm_start = start;
883 start_changed = true;
885 if (end != vma->vm_end) {
889 vma->vm_pgoff = pgoff;
891 next->vm_start += adjust_next;
892 next->vm_pgoff += adjust_next >> PAGE_SHIFT;
897 vma_interval_tree_insert(next, root);
898 vma_interval_tree_insert(vma, root);
899 flush_dcache_mmap_unlock(mapping);
904 * vma_merge has merged next into vma, and needs
905 * us to remove next before dropping the locks.
907 if (remove_next != 3)
908 __vma_unlink(mm, next, next);
911 * vma is not before next if they've been
914 * pre-swap() next->vm_start was reduced so
915 * tell validate_mm_rb to ignore pre-swap()
916 * "next" (which is stored in post-swap()
919 __vma_unlink(mm, next, vma);
921 __remove_shared_vm_struct(next, file, mapping);
924 * split_vma has split insert from vma, and needs
925 * us to insert it before dropping the locks
926 * (it may either follow vma or precede it).
928 __insert_vm_struct(mm, insert);
934 mm->highest_vm_end = vm_end_gap(vma);
935 else if (!adjust_next)
936 vma_gap_update(next);
941 anon_vma_interval_tree_post_update_vma(vma);
943 anon_vma_interval_tree_post_update_vma(next);
944 anon_vma_unlock_write(anon_vma);
948 i_mmap_unlock_write(mapping);
957 uprobe_munmap(next, next->vm_start, next->vm_end);
961 anon_vma_merge(vma, next);
963 mpol_put(vma_policy(next));
966 * In mprotect's case 6 (see comments on vma_merge),
967 * we must remove another next too. It would clutter
968 * up the code too much to do both in one go.
970 if (remove_next != 3) {
972 * If "next" was removed and vma->vm_end was
973 * expanded (up) over it, in turn
974 * "next->vm_prev->vm_end" changed and the
975 * "vma->vm_next" gap must be updated.
980 * For the scope of the comment "next" and
981 * "vma" considered pre-swap(): if "vma" was
982 * removed, next->vm_start was expanded (down)
983 * over it and the "next" gap must be updated.
984 * Because of the swap() the post-swap() "vma"
985 * actually points to pre-swap() "next"
986 * (post-swap() "next" as opposed is now a
991 if (remove_next == 2) {
997 vma_gap_update(next);
1000 * If remove_next == 2 we obviously can't
1003 * If remove_next == 3 we can't reach this
1004 * path because pre-swap() next is always not
1005 * NULL. pre-swap() "next" is not being
1006 * removed and its next->vm_end is not altered
1007 * (and furthermore "end" already matches
1008 * next->vm_end in remove_next == 3).
1010 * We reach this only in the remove_next == 1
1011 * case if the "next" vma that was removed was
1012 * the highest vma of the mm. However in such
1013 * case next->vm_end == "end" and the extended
1014 * "vma" has vma->vm_end == next->vm_end so
1015 * mm->highest_vm_end doesn't need any update
1016 * in remove_next == 1 case.
1018 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
1022 uprobe_mmap(insert);
1030 * If the vma has a ->close operation then the driver probably needs to release
1031 * per-vma resources, so we don't attempt to merge those.
1033 static inline int is_mergeable_vma(struct vm_area_struct *vma,
1034 struct file *file, unsigned long vm_flags,
1035 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1038 * VM_SOFTDIRTY should not prevent from VMA merging, if we
1039 * match the flags but dirty bit -- the caller should mark
1040 * merged VMA as dirty. If dirty bit won't be excluded from
1041 * comparison, we increase pressure on the memory system forcing
1042 * the kernel to generate new VMAs when old one could be
1045 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
1047 if (vma->vm_file != file)
1049 if (vma->vm_ops && vma->vm_ops->close)
1051 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1056 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1057 struct anon_vma *anon_vma2,
1058 struct vm_area_struct *vma)
1061 * The list_is_singular() test is to avoid merging VMA cloned from
1062 * parents. This can improve scalability caused by anon_vma lock.
1064 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1065 list_is_singular(&vma->anon_vma_chain)))
1067 return anon_vma1 == anon_vma2;
1071 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1072 * in front of (at a lower virtual address and file offset than) the vma.
1074 * We cannot merge two vmas if they have differently assigned (non-NULL)
1075 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1077 * We don't check here for the merged mmap wrapping around the end of pagecache
1078 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
1079 * wrap, nor mmaps which cover the final page at index -1UL.
1082 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1083 struct anon_vma *anon_vma, struct file *file,
1085 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1087 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1088 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1089 if (vma->vm_pgoff == vm_pgoff)
1096 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1097 * beyond (at a higher virtual address and file offset than) the vma.
1099 * We cannot merge two vmas if they have differently assigned (non-NULL)
1100 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1103 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1104 struct anon_vma *anon_vma, struct file *file,
1106 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1108 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1109 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1111 vm_pglen = vma_pages(vma);
1112 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1119 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1120 * whether that can be merged with its predecessor or its successor.
1121 * Or both (it neatly fills a hole).
1123 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1124 * certain not to be mapped by the time vma_merge is called; but when
1125 * called for mprotect, it is certain to be already mapped (either at
1126 * an offset within prev, or at the start of next), and the flags of
1127 * this area are about to be changed to vm_flags - and the no-change
1128 * case has already been eliminated.
1130 * The following mprotect cases have to be considered, where AAAA is
1131 * the area passed down from mprotect_fixup, never extending beyond one
1132 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1135 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN
1136 * cannot merge might become might become
1137 * PPNNNNNNNNNN PPPPPPPPPPNN
1138 * mmap, brk or case 4 below case 5 below
1141 * PPPP NNNN PPPPNNNNXXXX
1142 * might become might become
1143 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
1144 * PPPPPPPPNNNN 2 or PPPPPPPPXXXX 7 or
1145 * PPPPNNNNNNNN 3 PPPPXXXXXXXX 8
1147 * It is important for case 8 that the vma NNNN overlapping the
1148 * region AAAA is never going to extended over XXXX. Instead XXXX must
1149 * be extended in region AAAA and NNNN must be removed. This way in
1150 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1151 * rmap_locks, the properties of the merged vma will be already
1152 * correct for the whole merged range. Some of those properties like
1153 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1154 * be correct for the whole merged range immediately after the
1155 * rmap_locks are released. Otherwise if XXXX would be removed and
1156 * NNNN would be extended over the XXXX range, remove_migration_ptes
1157 * or other rmap walkers (if working on addresses beyond the "end"
1158 * parameter) may establish ptes with the wrong permissions of NNNN
1159 * instead of the right permissions of XXXX.
1161 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1162 struct vm_area_struct *prev, unsigned long addr,
1163 unsigned long end, unsigned long vm_flags,
1164 struct anon_vma *anon_vma, struct file *file,
1165 pgoff_t pgoff, struct mempolicy *policy,
1166 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1168 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1169 struct vm_area_struct *area, *next;
1173 * We later require that vma->vm_flags == vm_flags,
1174 * so this tests vma->vm_flags & VM_SPECIAL, too.
1176 if (vm_flags & VM_SPECIAL)
1179 next = vma_next(mm, prev);
1181 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1182 next = next->vm_next;
1184 /* verify some invariant that must be enforced by the caller */
1185 VM_WARN_ON(prev && addr <= prev->vm_start);
1186 VM_WARN_ON(area && end > area->vm_end);
1187 VM_WARN_ON(addr >= end);
1190 * Can it merge with the predecessor?
1192 if (prev && prev->vm_end == addr &&
1193 mpol_equal(vma_policy(prev), policy) &&
1194 can_vma_merge_after(prev, vm_flags,
1195 anon_vma, file, pgoff,
1196 vm_userfaultfd_ctx)) {
1198 * OK, it can. Can we now merge in the successor as well?
1200 if (next && end == next->vm_start &&
1201 mpol_equal(policy, vma_policy(next)) &&
1202 can_vma_merge_before(next, vm_flags,
1205 vm_userfaultfd_ctx) &&
1206 is_mergeable_anon_vma(prev->anon_vma,
1207 next->anon_vma, NULL)) {
1209 err = __vma_adjust(prev, prev->vm_start,
1210 next->vm_end, prev->vm_pgoff, NULL,
1212 } else /* cases 2, 5, 7 */
1213 err = __vma_adjust(prev, prev->vm_start,
1214 end, prev->vm_pgoff, NULL, prev);
1217 khugepaged_enter_vma_merge(prev, vm_flags);
1222 * Can this new request be merged in front of next?
1224 if (next && end == next->vm_start &&
1225 mpol_equal(policy, vma_policy(next)) &&
1226 can_vma_merge_before(next, vm_flags,
1227 anon_vma, file, pgoff+pglen,
1228 vm_userfaultfd_ctx)) {
1229 if (prev && addr < prev->vm_end) /* case 4 */
1230 err = __vma_adjust(prev, prev->vm_start,
1231 addr, prev->vm_pgoff, NULL, next);
1232 else { /* cases 3, 8 */
1233 err = __vma_adjust(area, addr, next->vm_end,
1234 next->vm_pgoff - pglen, NULL, next);
1236 * In case 3 area is already equal to next and
1237 * this is a noop, but in case 8 "area" has
1238 * been removed and next was expanded over it.
1244 khugepaged_enter_vma_merge(area, vm_flags);
1252 * Rough compatibility check to quickly see if it's even worth looking
1253 * at sharing an anon_vma.
1255 * They need to have the same vm_file, and the flags can only differ
1256 * in things that mprotect may change.
1258 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1259 * we can merge the two vma's. For example, we refuse to merge a vma if
1260 * there is a vm_ops->close() function, because that indicates that the
1261 * driver is doing some kind of reference counting. But that doesn't
1262 * really matter for the anon_vma sharing case.
1264 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1266 return a->vm_end == b->vm_start &&
1267 mpol_equal(vma_policy(a), vma_policy(b)) &&
1268 a->vm_file == b->vm_file &&
1269 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1270 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1274 * Do some basic sanity checking to see if we can re-use the anon_vma
1275 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1276 * the same as 'old', the other will be the new one that is trying
1277 * to share the anon_vma.
1279 * NOTE! This runs with mm_sem held for reading, so it is possible that
1280 * the anon_vma of 'old' is concurrently in the process of being set up
1281 * by another page fault trying to merge _that_. But that's ok: if it
1282 * is being set up, that automatically means that it will be a singleton
1283 * acceptable for merging, so we can do all of this optimistically. But
1284 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1286 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1287 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1288 * is to return an anon_vma that is "complex" due to having gone through
1291 * We also make sure that the two vma's are compatible (adjacent,
1292 * and with the same memory policies). That's all stable, even with just
1293 * a read lock on the mm_sem.
1295 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1297 if (anon_vma_compatible(a, b)) {
1298 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1300 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1307 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1308 * neighbouring vmas for a suitable anon_vma, before it goes off
1309 * to allocate a new anon_vma. It checks because a repetitive
1310 * sequence of mprotects and faults may otherwise lead to distinct
1311 * anon_vmas being allocated, preventing vma merge in subsequent
1314 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1316 struct anon_vma *anon_vma = NULL;
1318 /* Try next first. */
1320 anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next);
1325 /* Try prev next. */
1327 anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma);
1330 * We might reach here with anon_vma == NULL if we can't find
1331 * any reusable anon_vma.
1332 * There's no absolute need to look only at touching neighbours:
1333 * we could search further afield for "compatible" anon_vmas.
1334 * But it would probably just be a waste of time searching,
1335 * or lead to too many vmas hanging off the same anon_vma.
1336 * We're trying to allow mprotect remerging later on,
1337 * not trying to minimize memory used for anon_vmas.
1343 * If a hint addr is less than mmap_min_addr change hint to be as
1344 * low as possible but still greater than mmap_min_addr
1346 static inline unsigned long round_hint_to_min(unsigned long hint)
1349 if (((void *)hint != NULL) &&
1350 (hint < mmap_min_addr))
1351 return PAGE_ALIGN(mmap_min_addr);
1355 static inline int mlock_future_check(struct mm_struct *mm,
1356 unsigned long flags,
1359 unsigned long locked, lock_limit;
1361 /* mlock MCL_FUTURE? */
1362 if (flags & VM_LOCKED) {
1363 locked = len >> PAGE_SHIFT;
1364 locked += mm->locked_vm;
1365 lock_limit = rlimit(RLIMIT_MEMLOCK);
1366 lock_limit >>= PAGE_SHIFT;
1367 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1373 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1375 if (S_ISREG(inode->i_mode))
1376 return MAX_LFS_FILESIZE;
1378 if (S_ISBLK(inode->i_mode))
1379 return MAX_LFS_FILESIZE;
1381 if (S_ISSOCK(inode->i_mode))
1382 return MAX_LFS_FILESIZE;
1384 /* Special "we do even unsigned file positions" case */
1385 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1388 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1392 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1393 unsigned long pgoff, unsigned long len)
1395 u64 maxsize = file_mmap_size_max(file, inode);
1397 if (maxsize && len > maxsize)
1400 if (pgoff > maxsize >> PAGE_SHIFT)
1406 * The caller must write-lock current->mm->mmap_lock.
1408 unsigned long do_mmap(struct file *file, unsigned long addr,
1409 unsigned long len, unsigned long prot,
1410 unsigned long flags, unsigned long pgoff,
1411 unsigned long *populate, struct list_head *uf)
1413 struct mm_struct *mm = current->mm;
1414 vm_flags_t vm_flags;
1423 * Does the application expect PROT_READ to imply PROT_EXEC?
1425 * (the exception is when the underlying filesystem is noexec
1426 * mounted, in which case we dont add PROT_EXEC.)
1428 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1429 if (!(file && path_noexec(&file->f_path)))
1432 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1433 if (flags & MAP_FIXED_NOREPLACE)
1436 if (!(flags & MAP_FIXED))
1437 addr = round_hint_to_min(addr);
1439 /* Careful about overflows.. */
1440 len = PAGE_ALIGN(len);
1444 /* offset overflow? */
1445 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1448 /* Too many mappings? */
1449 if (mm->map_count > sysctl_max_map_count)
1452 /* Obtain the address to map to. we verify (or select) it and ensure
1453 * that it represents a valid section of the address space.
1455 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1456 if (IS_ERR_VALUE(addr))
1459 if (flags & MAP_FIXED_NOREPLACE) {
1460 if (find_vma_intersection(mm, addr, addr + len))
1464 if (prot == PROT_EXEC) {
1465 pkey = execute_only_pkey(mm);
1470 /* Do simple checking here so the lower-level routines won't have
1471 * to. we assume access permissions have been handled by the open
1472 * of the memory object, so we don't do any here.
1474 vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1475 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1477 if (flags & MAP_LOCKED)
1478 if (!can_do_mlock())
1481 if (mlock_future_check(mm, vm_flags, len))
1485 struct inode *inode = file_inode(file);
1486 unsigned long flags_mask;
1488 if (!file_mmap_ok(file, inode, pgoff, len))
1491 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1493 switch (flags & MAP_TYPE) {
1496 * Force use of MAP_SHARED_VALIDATE with non-legacy
1497 * flags. E.g. MAP_SYNC is dangerous to use with
1498 * MAP_SHARED as you don't know which consistency model
1499 * you will get. We silently ignore unsupported flags
1500 * with MAP_SHARED to preserve backward compatibility.
1502 flags &= LEGACY_MAP_MASK;
1504 case MAP_SHARED_VALIDATE:
1505 if (flags & ~flags_mask)
1507 if (prot & PROT_WRITE) {
1508 if (!(file->f_mode & FMODE_WRITE))
1510 if (IS_SWAPFILE(file->f_mapping->host))
1515 * Make sure we don't allow writing to an append-only
1518 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1522 * Make sure there are no mandatory locks on the file.
1524 if (locks_verify_locked(file))
1527 vm_flags |= VM_SHARED | VM_MAYSHARE;
1528 if (!(file->f_mode & FMODE_WRITE))
1529 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1532 if (!(file->f_mode & FMODE_READ))
1534 if (path_noexec(&file->f_path)) {
1535 if (vm_flags & VM_EXEC)
1537 vm_flags &= ~VM_MAYEXEC;
1540 if (!file->f_op->mmap)
1542 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1550 switch (flags & MAP_TYPE) {
1552 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1558 vm_flags |= VM_SHARED | VM_MAYSHARE;
1562 * Set pgoff according to addr for anon_vma.
1564 pgoff = addr >> PAGE_SHIFT;
1572 * Set 'VM_NORESERVE' if we should not account for the
1573 * memory use of this mapping.
1575 if (flags & MAP_NORESERVE) {
1576 /* We honor MAP_NORESERVE if allowed to overcommit */
1577 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1578 vm_flags |= VM_NORESERVE;
1580 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1581 if (file && is_file_hugepages(file))
1582 vm_flags |= VM_NORESERVE;
1585 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1586 if (!IS_ERR_VALUE(addr) &&
1587 ((vm_flags & VM_LOCKED) ||
1588 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1593 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1594 unsigned long prot, unsigned long flags,
1595 unsigned long fd, unsigned long pgoff)
1597 struct file *file = NULL;
1598 unsigned long retval;
1600 if (!(flags & MAP_ANONYMOUS)) {
1601 audit_mmap_fd(fd, flags);
1605 if (is_file_hugepages(file)) {
1606 len = ALIGN(len, huge_page_size(hstate_file(file)));
1607 } else if (unlikely(flags & MAP_HUGETLB)) {
1611 } else if (flags & MAP_HUGETLB) {
1612 struct ucounts *ucounts = NULL;
1615 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1619 len = ALIGN(len, huge_page_size(hs));
1621 * VM_NORESERVE is used because the reservations will be
1622 * taken when vm_ops->mmap() is called
1623 * A dummy user value is used because we are not locking
1624 * memory so no accounting is necessary
1626 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1628 &ucounts, HUGETLB_ANONHUGE_INODE,
1629 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1631 return PTR_ERR(file);
1634 flags &= ~MAP_DENYWRITE;
1636 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1643 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1644 unsigned long, prot, unsigned long, flags,
1645 unsigned long, fd, unsigned long, pgoff)
1647 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1650 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1651 struct mmap_arg_struct {
1655 unsigned long flags;
1657 unsigned long offset;
1660 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1662 struct mmap_arg_struct a;
1664 if (copy_from_user(&a, arg, sizeof(a)))
1666 if (offset_in_page(a.offset))
1669 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1670 a.offset >> PAGE_SHIFT);
1672 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1675 * Some shared mappings will want the pages marked read-only
1676 * to track write events. If so, we'll downgrade vm_page_prot
1677 * to the private version (using protection_map[] without the
1680 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1682 vm_flags_t vm_flags = vma->vm_flags;
1683 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1685 /* If it was private or non-writable, the write bit is already clear */
1686 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1689 /* The backer wishes to know when pages are first written to? */
1690 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1693 /* The open routine did something to the protections that pgprot_modify
1694 * won't preserve? */
1695 if (pgprot_val(vm_page_prot) !=
1696 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1699 /* Do we need to track softdirty? */
1700 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1703 /* Specialty mapping? */
1704 if (vm_flags & VM_PFNMAP)
1707 /* Can the mapping track the dirty pages? */
1708 return vma->vm_file && vma->vm_file->f_mapping &&
1709 mapping_can_writeback(vma->vm_file->f_mapping);
1713 * We account for memory if it's a private writeable mapping,
1714 * not hugepages and VM_NORESERVE wasn't set.
1716 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1719 * hugetlb has its own accounting separate from the core VM
1720 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1722 if (file && is_file_hugepages(file))
1725 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1728 unsigned long mmap_region(struct file *file, unsigned long addr,
1729 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1730 struct list_head *uf)
1732 struct mm_struct *mm = current->mm;
1733 struct vm_area_struct *vma, *prev, *merge;
1735 struct rb_node **rb_link, *rb_parent;
1736 unsigned long charged = 0;
1738 /* Check against address space limit. */
1739 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1740 unsigned long nr_pages;
1743 * MAP_FIXED may remove pages of mappings that intersects with
1744 * requested mapping. Account for the pages it would unmap.
1746 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1748 if (!may_expand_vm(mm, vm_flags,
1749 (len >> PAGE_SHIFT) - nr_pages))
1753 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
1754 if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
1757 * Private writable mapping: check memory availability
1759 if (accountable_mapping(file, vm_flags)) {
1760 charged = len >> PAGE_SHIFT;
1761 if (security_vm_enough_memory_mm(mm, charged))
1763 vm_flags |= VM_ACCOUNT;
1767 * Can we just expand an old mapping?
1769 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1770 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1775 * Determine the object being mapped and call the appropriate
1776 * specific mapper. the address has already been validated, but
1777 * not unmapped, but the maps are removed from the list.
1779 vma = vm_area_alloc(mm);
1785 vma->vm_start = addr;
1786 vma->vm_end = addr + len;
1787 vma->vm_flags = vm_flags;
1788 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1789 vma->vm_pgoff = pgoff;
1792 if (vm_flags & VM_DENYWRITE) {
1793 error = deny_write_access(file);
1797 if (vm_flags & VM_SHARED) {
1798 error = mapping_map_writable(file->f_mapping);
1800 goto allow_write_and_free_vma;
1803 /* ->mmap() can change vma->vm_file, but must guarantee that
1804 * vma_link() below can deny write-access if VM_DENYWRITE is set
1805 * and map writably if VM_SHARED is set. This usually means the
1806 * new file must not have been exposed to user-space, yet.
1808 vma->vm_file = get_file(file);
1809 error = call_mmap(file, vma);
1811 goto unmap_and_free_vma;
1813 /* Can addr have changed??
1815 * Answer: Yes, several device drivers can do it in their
1816 * f_op->mmap method. -DaveM
1817 * Bug: If addr is changed, prev, rb_link, rb_parent should
1818 * be updated for vma_link()
1820 WARN_ON_ONCE(addr != vma->vm_start);
1822 addr = vma->vm_start;
1824 /* If vm_flags changed after call_mmap(), we should try merge vma again
1825 * as we may succeed this time.
1827 if (unlikely(vm_flags != vma->vm_flags && prev)) {
1828 merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
1829 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX);
1831 /* ->mmap() can change vma->vm_file and fput the original file. So
1832 * fput the vma->vm_file here or we would add an extra fput for file
1833 * and cause general protection fault ultimately.
1838 /* Update vm_flags to pick up the change. */
1839 vm_flags = vma->vm_flags;
1840 goto unmap_writable;
1844 vm_flags = vma->vm_flags;
1845 } else if (vm_flags & VM_SHARED) {
1846 error = shmem_zero_setup(vma);
1850 vma_set_anonymous(vma);
1853 /* Allow architectures to sanity-check the vm_flags */
1854 if (!arch_validate_flags(vma->vm_flags)) {
1857 goto unmap_and_free_vma;
1862 vma_link(mm, vma, prev, rb_link, rb_parent);
1863 /* Once vma denies write, undo our temporary denial count */
1866 if (vm_flags & VM_SHARED)
1867 mapping_unmap_writable(file->f_mapping);
1868 if (vm_flags & VM_DENYWRITE)
1869 allow_write_access(file);
1871 file = vma->vm_file;
1873 perf_event_mmap(vma);
1875 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1876 if (vm_flags & VM_LOCKED) {
1877 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1878 is_vm_hugetlb_page(vma) ||
1879 vma == get_gate_vma(current->mm))
1880 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1882 mm->locked_vm += (len >> PAGE_SHIFT);
1889 * New (or expanded) vma always get soft dirty status.
1890 * Otherwise user-space soft-dirty page tracker won't
1891 * be able to distinguish situation when vma area unmapped,
1892 * then new mapped in-place (which must be aimed as
1893 * a completely new data area).
1895 vma->vm_flags |= VM_SOFTDIRTY;
1897 vma_set_page_prot(vma);
1903 vma->vm_file = NULL;
1905 /* Undo any partial mapping done by a device driver. */
1906 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1908 if (vm_flags & VM_SHARED)
1909 mapping_unmap_writable(file->f_mapping);
1910 allow_write_and_free_vma:
1911 if (vm_flags & VM_DENYWRITE)
1912 allow_write_access(file);
1917 vm_unacct_memory(charged);
1921 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1924 * We implement the search by looking for an rbtree node that
1925 * immediately follows a suitable gap. That is,
1926 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1927 * - gap_end = vma->vm_start >= info->low_limit + length;
1928 * - gap_end - gap_start >= length
1931 struct mm_struct *mm = current->mm;
1932 struct vm_area_struct *vma;
1933 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1935 /* Adjust search length to account for worst case alignment overhead */
1936 length = info->length + info->align_mask;
1937 if (length < info->length)
1940 /* Adjust search limits by the desired length */
1941 if (info->high_limit < length)
1943 high_limit = info->high_limit - length;
1945 if (info->low_limit > high_limit)
1947 low_limit = info->low_limit + length;
1949 /* Check if rbtree root looks promising */
1950 if (RB_EMPTY_ROOT(&mm->mm_rb))
1952 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1953 if (vma->rb_subtree_gap < length)
1957 /* Visit left subtree if it looks promising */
1958 gap_end = vm_start_gap(vma);
1959 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1960 struct vm_area_struct *left =
1961 rb_entry(vma->vm_rb.rb_left,
1962 struct vm_area_struct, vm_rb);
1963 if (left->rb_subtree_gap >= length) {
1969 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1971 /* Check if current node has a suitable gap */
1972 if (gap_start > high_limit)
1974 if (gap_end >= low_limit &&
1975 gap_end > gap_start && gap_end - gap_start >= length)
1978 /* Visit right subtree if it looks promising */
1979 if (vma->vm_rb.rb_right) {
1980 struct vm_area_struct *right =
1981 rb_entry(vma->vm_rb.rb_right,
1982 struct vm_area_struct, vm_rb);
1983 if (right->rb_subtree_gap >= length) {
1989 /* Go back up the rbtree to find next candidate node */
1991 struct rb_node *prev = &vma->vm_rb;
1992 if (!rb_parent(prev))
1994 vma = rb_entry(rb_parent(prev),
1995 struct vm_area_struct, vm_rb);
1996 if (prev == vma->vm_rb.rb_left) {
1997 gap_start = vm_end_gap(vma->vm_prev);
1998 gap_end = vm_start_gap(vma);
2005 /* Check highest gap, which does not precede any rbtree node */
2006 gap_start = mm->highest_vm_end;
2007 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
2008 if (gap_start > high_limit)
2012 /* We found a suitable gap. Clip it with the original low_limit. */
2013 if (gap_start < info->low_limit)
2014 gap_start = info->low_limit;
2016 /* Adjust gap address to the desired alignment */
2017 gap_start += (info->align_offset - gap_start) & info->align_mask;
2019 VM_BUG_ON(gap_start + info->length > info->high_limit);
2020 VM_BUG_ON(gap_start + info->length > gap_end);
2024 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
2026 struct mm_struct *mm = current->mm;
2027 struct vm_area_struct *vma;
2028 unsigned long length, low_limit, high_limit, gap_start, gap_end;
2030 /* Adjust search length to account for worst case alignment overhead */
2031 length = info->length + info->align_mask;
2032 if (length < info->length)
2036 * Adjust search limits by the desired length.
2037 * See implementation comment at top of unmapped_area().
2039 gap_end = info->high_limit;
2040 if (gap_end < length)
2042 high_limit = gap_end - length;
2044 if (info->low_limit > high_limit)
2046 low_limit = info->low_limit + length;
2048 /* Check highest gap, which does not precede any rbtree node */
2049 gap_start = mm->highest_vm_end;
2050 if (gap_start <= high_limit)
2053 /* Check if rbtree root looks promising */
2054 if (RB_EMPTY_ROOT(&mm->mm_rb))
2056 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
2057 if (vma->rb_subtree_gap < length)
2061 /* Visit right subtree if it looks promising */
2062 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
2063 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
2064 struct vm_area_struct *right =
2065 rb_entry(vma->vm_rb.rb_right,
2066 struct vm_area_struct, vm_rb);
2067 if (right->rb_subtree_gap >= length) {
2074 /* Check if current node has a suitable gap */
2075 gap_end = vm_start_gap(vma);
2076 if (gap_end < low_limit)
2078 if (gap_start <= high_limit &&
2079 gap_end > gap_start && gap_end - gap_start >= length)
2082 /* Visit left subtree if it looks promising */
2083 if (vma->vm_rb.rb_left) {
2084 struct vm_area_struct *left =
2085 rb_entry(vma->vm_rb.rb_left,
2086 struct vm_area_struct, vm_rb);
2087 if (left->rb_subtree_gap >= length) {
2093 /* Go back up the rbtree to find next candidate node */
2095 struct rb_node *prev = &vma->vm_rb;
2096 if (!rb_parent(prev))
2098 vma = rb_entry(rb_parent(prev),
2099 struct vm_area_struct, vm_rb);
2100 if (prev == vma->vm_rb.rb_right) {
2101 gap_start = vma->vm_prev ?
2102 vm_end_gap(vma->vm_prev) : 0;
2109 /* We found a suitable gap. Clip it with the original high_limit. */
2110 if (gap_end > info->high_limit)
2111 gap_end = info->high_limit;
2114 /* Compute highest gap address at the desired alignment */
2115 gap_end -= info->length;
2116 gap_end -= (gap_end - info->align_offset) & info->align_mask;
2118 VM_BUG_ON(gap_end < info->low_limit);
2119 VM_BUG_ON(gap_end < gap_start);
2124 * Search for an unmapped address range.
2126 * We are looking for a range that:
2127 * - does not intersect with any VMA;
2128 * - is contained within the [low_limit, high_limit) interval;
2129 * - is at least the desired size.
2130 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
2132 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
2136 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
2137 addr = unmapped_area_topdown(info);
2139 addr = unmapped_area(info);
2141 trace_vm_unmapped_area(addr, info);
2145 #ifndef arch_get_mmap_end
2146 #define arch_get_mmap_end(addr) (TASK_SIZE)
2149 #ifndef arch_get_mmap_base
2150 #define arch_get_mmap_base(addr, base) (base)
2153 /* Get an address range which is currently unmapped.
2154 * For shmat() with addr=0.
2156 * Ugly calling convention alert:
2157 * Return value with the low bits set means error value,
2159 * if (ret & ~PAGE_MASK)
2162 * This function "knows" that -ENOMEM has the bits set.
2164 #ifndef HAVE_ARCH_UNMAPPED_AREA
2166 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2167 unsigned long len, unsigned long pgoff, unsigned long flags)
2169 struct mm_struct *mm = current->mm;
2170 struct vm_area_struct *vma, *prev;
2171 struct vm_unmapped_area_info info;
2172 const unsigned long mmap_end = arch_get_mmap_end(addr);
2174 if (len > mmap_end - mmap_min_addr)
2177 if (flags & MAP_FIXED)
2181 addr = PAGE_ALIGN(addr);
2182 vma = find_vma_prev(mm, addr, &prev);
2183 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2184 (!vma || addr + len <= vm_start_gap(vma)) &&
2185 (!prev || addr >= vm_end_gap(prev)))
2191 info.low_limit = mm->mmap_base;
2192 info.high_limit = mmap_end;
2193 info.align_mask = 0;
2194 info.align_offset = 0;
2195 return vm_unmapped_area(&info);
2200 * This mmap-allocator allocates new areas top-down from below the
2201 * stack's low limit (the base):
2203 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2205 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2206 unsigned long len, unsigned long pgoff,
2207 unsigned long flags)
2209 struct vm_area_struct *vma, *prev;
2210 struct mm_struct *mm = current->mm;
2211 struct vm_unmapped_area_info info;
2212 const unsigned long mmap_end = arch_get_mmap_end(addr);
2214 /* requested length too big for entire address space */
2215 if (len > mmap_end - mmap_min_addr)
2218 if (flags & MAP_FIXED)
2221 /* requesting a specific address */
2223 addr = PAGE_ALIGN(addr);
2224 vma = find_vma_prev(mm, addr, &prev);
2225 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2226 (!vma || addr + len <= vm_start_gap(vma)) &&
2227 (!prev || addr >= vm_end_gap(prev)))
2231 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2233 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2234 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2235 info.align_mask = 0;
2236 info.align_offset = 0;
2237 addr = vm_unmapped_area(&info);
2240 * A failed mmap() very likely causes application failure,
2241 * so fall back to the bottom-up function here. This scenario
2242 * can happen with large stack limits and large mmap()
2245 if (offset_in_page(addr)) {
2246 VM_BUG_ON(addr != -ENOMEM);
2248 info.low_limit = TASK_UNMAPPED_BASE;
2249 info.high_limit = mmap_end;
2250 addr = vm_unmapped_area(&info);
2258 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2259 unsigned long pgoff, unsigned long flags)
2261 unsigned long (*get_area)(struct file *, unsigned long,
2262 unsigned long, unsigned long, unsigned long);
2264 unsigned long error = arch_mmap_check(addr, len, flags);
2268 /* Careful about overflows.. */
2269 if (len > TASK_SIZE)
2272 get_area = current->mm->get_unmapped_area;
2274 if (file->f_op->get_unmapped_area)
2275 get_area = file->f_op->get_unmapped_area;
2276 } else if (flags & MAP_SHARED) {
2278 * mmap_region() will call shmem_zero_setup() to create a file,
2279 * so use shmem's get_unmapped_area in case it can be huge.
2280 * do_mmap() will clear pgoff, so match alignment.
2283 get_area = shmem_get_unmapped_area;
2286 addr = get_area(file, addr, len, pgoff, flags);
2287 if (IS_ERR_VALUE(addr))
2290 if (addr > TASK_SIZE - len)
2292 if (offset_in_page(addr))
2295 error = security_mmap_addr(addr);
2296 return error ? error : addr;
2299 EXPORT_SYMBOL(get_unmapped_area);
2301 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2302 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2304 struct rb_node *rb_node;
2305 struct vm_area_struct *vma;
2307 /* Check the cache first. */
2308 vma = vmacache_find(mm, addr);
2312 rb_node = mm->mm_rb.rb_node;
2315 struct vm_area_struct *tmp;
2317 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2319 if (tmp->vm_end > addr) {
2321 if (tmp->vm_start <= addr)
2323 rb_node = rb_node->rb_left;
2325 rb_node = rb_node->rb_right;
2329 vmacache_update(addr, vma);
2333 EXPORT_SYMBOL(find_vma);
2336 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2338 struct vm_area_struct *
2339 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2340 struct vm_area_struct **pprev)
2342 struct vm_area_struct *vma;
2344 vma = find_vma(mm, addr);
2346 *pprev = vma->vm_prev;
2348 struct rb_node *rb_node = rb_last(&mm->mm_rb);
2350 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
2356 * Verify that the stack growth is acceptable and
2357 * update accounting. This is shared with both the
2358 * grow-up and grow-down cases.
2360 static int acct_stack_growth(struct vm_area_struct *vma,
2361 unsigned long size, unsigned long grow)
2363 struct mm_struct *mm = vma->vm_mm;
2364 unsigned long new_start;
2366 /* address space limit tests */
2367 if (!may_expand_vm(mm, vma->vm_flags, grow))
2370 /* Stack limit test */
2371 if (size > rlimit(RLIMIT_STACK))
2374 /* mlock limit tests */
2375 if (vma->vm_flags & VM_LOCKED) {
2376 unsigned long locked;
2377 unsigned long limit;
2378 locked = mm->locked_vm + grow;
2379 limit = rlimit(RLIMIT_MEMLOCK);
2380 limit >>= PAGE_SHIFT;
2381 if (locked > limit && !capable(CAP_IPC_LOCK))
2385 /* Check to ensure the stack will not grow into a hugetlb-only region */
2386 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2388 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2392 * Overcommit.. This must be the final test, as it will
2393 * update security statistics.
2395 if (security_vm_enough_memory_mm(mm, grow))
2401 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2403 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2404 * vma is the last one with address > vma->vm_end. Have to extend vma.
2406 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2408 struct mm_struct *mm = vma->vm_mm;
2409 struct vm_area_struct *next;
2410 unsigned long gap_addr;
2413 if (!(vma->vm_flags & VM_GROWSUP))
2416 /* Guard against exceeding limits of the address space. */
2417 address &= PAGE_MASK;
2418 if (address >= (TASK_SIZE & PAGE_MASK))
2420 address += PAGE_SIZE;
2422 /* Enforce stack_guard_gap */
2423 gap_addr = address + stack_guard_gap;
2425 /* Guard against overflow */
2426 if (gap_addr < address || gap_addr > TASK_SIZE)
2427 gap_addr = TASK_SIZE;
2429 next = vma->vm_next;
2430 if (next && next->vm_start < gap_addr && vma_is_accessible(next)) {
2431 if (!(next->vm_flags & VM_GROWSUP))
2433 /* Check that both stack segments have the same anon_vma? */
2436 /* We must make sure the anon_vma is allocated. */
2437 if (unlikely(anon_vma_prepare(vma)))
2441 * vma->vm_start/vm_end cannot change under us because the caller
2442 * is required to hold the mmap_lock in read mode. We need the
2443 * anon_vma lock to serialize against concurrent expand_stacks.
2445 anon_vma_lock_write(vma->anon_vma);
2447 /* Somebody else might have raced and expanded it already */
2448 if (address > vma->vm_end) {
2449 unsigned long size, grow;
2451 size = address - vma->vm_start;
2452 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2455 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2456 error = acct_stack_growth(vma, size, grow);
2459 * vma_gap_update() doesn't support concurrent
2460 * updates, but we only hold a shared mmap_lock
2461 * lock here, so we need to protect against
2462 * concurrent vma expansions.
2463 * anon_vma_lock_write() doesn't help here, as
2464 * we don't guarantee that all growable vmas
2465 * in a mm share the same root anon vma.
2466 * So, we reuse mm->page_table_lock to guard
2467 * against concurrent vma expansions.
2469 spin_lock(&mm->page_table_lock);
2470 if (vma->vm_flags & VM_LOCKED)
2471 mm->locked_vm += grow;
2472 vm_stat_account(mm, vma->vm_flags, grow);
2473 anon_vma_interval_tree_pre_update_vma(vma);
2474 vma->vm_end = address;
2475 anon_vma_interval_tree_post_update_vma(vma);
2477 vma_gap_update(vma->vm_next);
2479 mm->highest_vm_end = vm_end_gap(vma);
2480 spin_unlock(&mm->page_table_lock);
2482 perf_event_mmap(vma);
2486 anon_vma_unlock_write(vma->anon_vma);
2487 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2491 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2494 * vma is the first one with address < vma->vm_start. Have to extend vma.
2496 int expand_downwards(struct vm_area_struct *vma,
2497 unsigned long address)
2499 struct mm_struct *mm = vma->vm_mm;
2500 struct vm_area_struct *prev;
2503 address &= PAGE_MASK;
2504 if (address < mmap_min_addr)
2507 /* Enforce stack_guard_gap */
2508 prev = vma->vm_prev;
2509 /* Check that both stack segments have the same anon_vma? */
2510 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2511 vma_is_accessible(prev)) {
2512 if (address - prev->vm_end < stack_guard_gap)
2516 /* We must make sure the anon_vma is allocated. */
2517 if (unlikely(anon_vma_prepare(vma)))
2521 * vma->vm_start/vm_end cannot change under us because the caller
2522 * is required to hold the mmap_lock in read mode. We need the
2523 * anon_vma lock to serialize against concurrent expand_stacks.
2525 anon_vma_lock_write(vma->anon_vma);
2527 /* Somebody else might have raced and expanded it already */
2528 if (address < vma->vm_start) {
2529 unsigned long size, grow;
2531 size = vma->vm_end - address;
2532 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2535 if (grow <= vma->vm_pgoff) {
2536 error = acct_stack_growth(vma, size, grow);
2539 * vma_gap_update() doesn't support concurrent
2540 * updates, but we only hold a shared mmap_lock
2541 * lock here, so we need to protect against
2542 * concurrent vma expansions.
2543 * anon_vma_lock_write() doesn't help here, as
2544 * we don't guarantee that all growable vmas
2545 * in a mm share the same root anon vma.
2546 * So, we reuse mm->page_table_lock to guard
2547 * against concurrent vma expansions.
2549 spin_lock(&mm->page_table_lock);
2550 if (vma->vm_flags & VM_LOCKED)
2551 mm->locked_vm += grow;
2552 vm_stat_account(mm, vma->vm_flags, grow);
2553 anon_vma_interval_tree_pre_update_vma(vma);
2554 vma->vm_start = address;
2555 vma->vm_pgoff -= grow;
2556 anon_vma_interval_tree_post_update_vma(vma);
2557 vma_gap_update(vma);
2558 spin_unlock(&mm->page_table_lock);
2560 perf_event_mmap(vma);
2564 anon_vma_unlock_write(vma->anon_vma);
2565 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2570 /* enforced gap between the expanding stack and other mappings. */
2571 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2573 static int __init cmdline_parse_stack_guard_gap(char *p)
2578 val = simple_strtoul(p, &endptr, 10);
2580 stack_guard_gap = val << PAGE_SHIFT;
2584 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2586 #ifdef CONFIG_STACK_GROWSUP
2587 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2589 return expand_upwards(vma, address);
2592 struct vm_area_struct *
2593 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2595 struct vm_area_struct *vma, *prev;
2598 vma = find_vma_prev(mm, addr, &prev);
2599 if (vma && (vma->vm_start <= addr))
2601 /* don't alter vm_end if the coredump is running */
2602 if (!prev || expand_stack(prev, addr))
2604 if (prev->vm_flags & VM_LOCKED)
2605 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2609 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2611 return expand_downwards(vma, address);
2614 struct vm_area_struct *
2615 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2617 struct vm_area_struct *vma;
2618 unsigned long start;
2621 vma = find_vma(mm, addr);
2624 if (vma->vm_start <= addr)
2626 if (!(vma->vm_flags & VM_GROWSDOWN))
2628 start = vma->vm_start;
2629 if (expand_stack(vma, addr))
2631 if (vma->vm_flags & VM_LOCKED)
2632 populate_vma_page_range(vma, addr, start, NULL);
2637 EXPORT_SYMBOL_GPL(find_extend_vma);
2640 * Ok - we have the memory areas we should free on the vma list,
2641 * so release them, and do the vma updates.
2643 * Called with the mm semaphore held.
2645 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2647 unsigned long nr_accounted = 0;
2649 /* Update high watermark before we lower total_vm */
2650 update_hiwater_vm(mm);
2652 long nrpages = vma_pages(vma);
2654 if (vma->vm_flags & VM_ACCOUNT)
2655 nr_accounted += nrpages;
2656 vm_stat_account(mm, vma->vm_flags, -nrpages);
2657 vma = remove_vma(vma);
2659 vm_unacct_memory(nr_accounted);
2664 * Get rid of page table information in the indicated region.
2666 * Called with the mm semaphore held.
2668 static void unmap_region(struct mm_struct *mm,
2669 struct vm_area_struct *vma, struct vm_area_struct *prev,
2670 unsigned long start, unsigned long end)
2672 struct vm_area_struct *next = vma_next(mm, prev);
2673 struct mmu_gather tlb;
2676 tlb_gather_mmu(&tlb, mm);
2677 update_hiwater_rss(mm);
2678 unmap_vmas(&tlb, vma, start, end);
2679 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2680 next ? next->vm_start : USER_PGTABLES_CEILING);
2681 tlb_finish_mmu(&tlb);
2685 * Create a list of vma's touched by the unmap, removing them from the mm's
2686 * vma list as we go..
2689 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2690 struct vm_area_struct *prev, unsigned long end)
2692 struct vm_area_struct **insertion_point;
2693 struct vm_area_struct *tail_vma = NULL;
2695 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2696 vma->vm_prev = NULL;
2698 vma_rb_erase(vma, &mm->mm_rb);
2702 } while (vma && vma->vm_start < end);
2703 *insertion_point = vma;
2705 vma->vm_prev = prev;
2706 vma_gap_update(vma);
2708 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2709 tail_vma->vm_next = NULL;
2711 /* Kill the cache */
2712 vmacache_invalidate(mm);
2715 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2716 * VM_GROWSUP VMA. Such VMAs can change their size under
2717 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2719 if (vma && (vma->vm_flags & VM_GROWSDOWN))
2721 if (prev && (prev->vm_flags & VM_GROWSUP))
2727 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2728 * has already been checked or doesn't make sense to fail.
2730 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2731 unsigned long addr, int new_below)
2733 struct vm_area_struct *new;
2736 if (vma->vm_ops && vma->vm_ops->may_split) {
2737 err = vma->vm_ops->may_split(vma, addr);
2742 new = vm_area_dup(vma);
2749 new->vm_start = addr;
2750 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2753 err = vma_dup_policy(vma, new);
2757 err = anon_vma_clone(new, vma);
2762 get_file(new->vm_file);
2764 if (new->vm_ops && new->vm_ops->open)
2765 new->vm_ops->open(new);
2768 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2769 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2771 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2777 /* Clean everything up if vma_adjust failed. */
2778 if (new->vm_ops && new->vm_ops->close)
2779 new->vm_ops->close(new);
2782 unlink_anon_vmas(new);
2784 mpol_put(vma_policy(new));
2791 * Split a vma into two pieces at address 'addr', a new vma is allocated
2792 * either for the first part or the tail.
2794 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2795 unsigned long addr, int new_below)
2797 if (mm->map_count >= sysctl_max_map_count)
2800 return __split_vma(mm, vma, addr, new_below);
2804 unlock_range(struct vm_area_struct *start, unsigned long limit)
2806 struct mm_struct *mm = start->vm_mm;
2807 struct vm_area_struct *tmp = start;
2809 while (tmp && tmp->vm_start < limit) {
2810 if (tmp->vm_flags & VM_LOCKED) {
2811 mm->locked_vm -= vma_pages(tmp);
2812 munlock_vma_pages_all(tmp);
2819 /* Munmap is split into 2 main parts -- this part which finds
2820 * what needs doing, and the areas themselves, which do the
2821 * work. This now handles partial unmappings.
2822 * Jeremy Fitzhardinge <jeremy@goop.org>
2824 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2825 struct list_head *uf, bool downgrade)
2828 struct vm_area_struct *vma, *prev, *last;
2830 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2833 len = PAGE_ALIGN(len);
2839 * arch_unmap() might do unmaps itself. It must be called
2840 * and finish any rbtree manipulation before this code
2841 * runs and also starts to manipulate the rbtree.
2843 arch_unmap(mm, start, end);
2845 /* Find the first overlapping VMA where start < vma->vm_end */
2846 vma = find_vma_intersection(mm, start, end);
2849 prev = vma->vm_prev;
2852 * If we need to split any vma, do it now to save pain later.
2854 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2855 * unmapped vm_area_struct will remain in use: so lower split_vma
2856 * places tmp vma above, and higher split_vma places tmp vma below.
2858 if (start > vma->vm_start) {
2862 * Make sure that map_count on return from munmap() will
2863 * not exceed its limit; but let map_count go just above
2864 * its limit temporarily, to help free resources as expected.
2866 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2869 error = __split_vma(mm, vma, start, 0);
2875 /* Does it split the last one? */
2876 last = find_vma(mm, end);
2877 if (last && end > last->vm_start) {
2878 int error = __split_vma(mm, last, end, 1);
2882 vma = vma_next(mm, prev);
2886 * If userfaultfd_unmap_prep returns an error the vmas
2887 * will remain split, but userland will get a
2888 * highly unexpected error anyway. This is no
2889 * different than the case where the first of the two
2890 * __split_vma fails, but we don't undo the first
2891 * split, despite we could. This is unlikely enough
2892 * failure that it's not worth optimizing it for.
2894 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2900 * unlock any mlock()ed ranges before detaching vmas
2903 unlock_range(vma, end);
2905 /* Detach vmas from rbtree */
2906 if (!detach_vmas_to_be_unmapped(mm, vma, prev, end))
2910 mmap_write_downgrade(mm);
2912 unmap_region(mm, vma, prev, start, end);
2914 /* Fix up all other VM information */
2915 remove_vma_list(mm, vma);
2917 return downgrade ? 1 : 0;
2920 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2921 struct list_head *uf)
2923 return __do_munmap(mm, start, len, uf, false);
2926 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2929 struct mm_struct *mm = current->mm;
2932 if (mmap_write_lock_killable(mm))
2935 ret = __do_munmap(mm, start, len, &uf, downgrade);
2937 * Returning 1 indicates mmap_lock is downgraded.
2938 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2939 * it to 0 before return.
2942 mmap_read_unlock(mm);
2945 mmap_write_unlock(mm);
2947 userfaultfd_unmap_complete(mm, &uf);
2951 int vm_munmap(unsigned long start, size_t len)
2953 return __vm_munmap(start, len, false);
2955 EXPORT_SYMBOL(vm_munmap);
2957 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2959 addr = untagged_addr(addr);
2960 profile_munmap(addr);
2961 return __vm_munmap(addr, len, true);
2966 * Emulation of deprecated remap_file_pages() syscall.
2968 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2969 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2972 struct mm_struct *mm = current->mm;
2973 struct vm_area_struct *vma;
2974 unsigned long populate = 0;
2975 unsigned long ret = -EINVAL;
2978 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2979 current->comm, current->pid);
2983 start = start & PAGE_MASK;
2984 size = size & PAGE_MASK;
2986 if (start + size <= start)
2989 /* Does pgoff wrap? */
2990 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2993 if (mmap_write_lock_killable(mm))
2996 vma = find_vma(mm, start);
2998 if (!vma || !(vma->vm_flags & VM_SHARED))
3001 if (start < vma->vm_start)
3004 if (start + size > vma->vm_end) {
3005 struct vm_area_struct *next;
3007 for (next = vma->vm_next; next; next = next->vm_next) {
3008 /* hole between vmas ? */
3009 if (next->vm_start != next->vm_prev->vm_end)
3012 if (next->vm_file != vma->vm_file)
3015 if (next->vm_flags != vma->vm_flags)
3018 if (start + size <= next->vm_end)
3026 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3027 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3028 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3030 flags &= MAP_NONBLOCK;
3031 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3032 if (vma->vm_flags & VM_LOCKED)
3033 flags |= MAP_LOCKED;
3035 file = get_file(vma->vm_file);
3036 ret = do_mmap(vma->vm_file, start, size,
3037 prot, flags, pgoff, &populate, NULL);
3040 mmap_write_unlock(mm);
3042 mm_populate(ret, populate);
3043 if (!IS_ERR_VALUE(ret))
3049 * this is really a simplified "do_mmap". it only handles
3050 * anonymous maps. eventually we may be able to do some
3051 * brk-specific accounting here.
3053 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
3055 struct mm_struct *mm = current->mm;
3056 struct vm_area_struct *vma, *prev;
3057 struct rb_node **rb_link, *rb_parent;
3058 pgoff_t pgoff = addr >> PAGE_SHIFT;
3060 unsigned long mapped_addr;
3062 /* Until we need other flags, refuse anything except VM_EXEC. */
3063 if ((flags & (~VM_EXEC)) != 0)
3065 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3067 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
3068 if (IS_ERR_VALUE(mapped_addr))
3071 error = mlock_future_check(mm, mm->def_flags, len);
3075 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
3076 if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
3079 /* Check against address space limits *after* clearing old maps... */
3080 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3083 if (mm->map_count > sysctl_max_map_count)
3086 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3089 /* Can we just expand an old private anonymous mapping? */
3090 vma = vma_merge(mm, prev, addr, addr + len, flags,
3091 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
3096 * create a vma struct for an anonymous mapping
3098 vma = vm_area_alloc(mm);
3100 vm_unacct_memory(len >> PAGE_SHIFT);
3104 vma_set_anonymous(vma);
3105 vma->vm_start = addr;
3106 vma->vm_end = addr + len;
3107 vma->vm_pgoff = pgoff;
3108 vma->vm_flags = flags;
3109 vma->vm_page_prot = vm_get_page_prot(flags);
3110 vma_link(mm, vma, prev, rb_link, rb_parent);
3112 perf_event_mmap(vma);
3113 mm->total_vm += len >> PAGE_SHIFT;
3114 mm->data_vm += len >> PAGE_SHIFT;
3115 if (flags & VM_LOCKED)
3116 mm->locked_vm += (len >> PAGE_SHIFT);
3117 vma->vm_flags |= VM_SOFTDIRTY;
3121 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3123 struct mm_struct *mm = current->mm;
3129 len = PAGE_ALIGN(request);
3135 if (mmap_write_lock_killable(mm))
3138 ret = do_brk_flags(addr, len, flags, &uf);
3139 populate = ((mm->def_flags & VM_LOCKED) != 0);
3140 mmap_write_unlock(mm);
3141 userfaultfd_unmap_complete(mm, &uf);
3142 if (populate && !ret)
3143 mm_populate(addr, len);
3146 EXPORT_SYMBOL(vm_brk_flags);
3148 int vm_brk(unsigned long addr, unsigned long len)
3150 return vm_brk_flags(addr, len, 0);
3152 EXPORT_SYMBOL(vm_brk);
3154 /* Release all mmaps. */
3155 void exit_mmap(struct mm_struct *mm)
3157 struct mmu_gather tlb;
3158 struct vm_area_struct *vma;
3159 unsigned long nr_accounted = 0;
3161 /* mm's last user has gone, and its about to be pulled down */
3162 mmu_notifier_release(mm);
3164 if (unlikely(mm_is_oom_victim(mm))) {
3166 * Manually reap the mm to free as much memory as possible.
3167 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3168 * this mm from further consideration. Taking mm->mmap_lock for
3169 * write after setting MMF_OOM_SKIP will guarantee that the oom
3170 * reaper will not run on this mm again after mmap_lock is
3173 * Nothing can be holding mm->mmap_lock here and the above call
3174 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3175 * __oom_reap_task_mm() will not block.
3177 * This needs to be done before calling munlock_vma_pages_all(),
3178 * which clears VM_LOCKED, otherwise the oom reaper cannot
3181 (void)__oom_reap_task_mm(mm);
3183 set_bit(MMF_OOM_SKIP, &mm->flags);
3184 mmap_write_lock(mm);
3185 mmap_write_unlock(mm);
3189 unlock_range(mm->mmap, ULONG_MAX);
3194 if (!vma) /* Can happen if dup_mmap() received an OOM */
3199 tlb_gather_mmu_fullmm(&tlb, mm);
3200 /* update_hiwater_rss(mm) here? but nobody should be looking */
3201 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3202 unmap_vmas(&tlb, vma, 0, -1);
3203 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3204 tlb_finish_mmu(&tlb);
3207 * Walk the list again, actually closing and freeing it,
3208 * with preemption enabled, without holding any MM locks.
3211 if (vma->vm_flags & VM_ACCOUNT)
3212 nr_accounted += vma_pages(vma);
3213 vma = remove_vma(vma);
3216 vm_unacct_memory(nr_accounted);
3219 /* Insert vm structure into process list sorted by address
3220 * and into the inode's i_mmap tree. If vm_file is non-NULL
3221 * then i_mmap_rwsem is taken here.
3223 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3225 struct vm_area_struct *prev;
3226 struct rb_node **rb_link, *rb_parent;
3228 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3229 &prev, &rb_link, &rb_parent))
3231 if ((vma->vm_flags & VM_ACCOUNT) &&
3232 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3236 * The vm_pgoff of a purely anonymous vma should be irrelevant
3237 * until its first write fault, when page's anon_vma and index
3238 * are set. But now set the vm_pgoff it will almost certainly
3239 * end up with (unless mremap moves it elsewhere before that
3240 * first wfault), so /proc/pid/maps tells a consistent story.
3242 * By setting it to reflect the virtual start address of the
3243 * vma, merges and splits can happen in a seamless way, just
3244 * using the existing file pgoff checks and manipulations.
3245 * Similarly in do_mmap and in do_brk_flags.
3247 if (vma_is_anonymous(vma)) {
3248 BUG_ON(vma->anon_vma);
3249 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3252 vma_link(mm, vma, prev, rb_link, rb_parent);
3257 * Copy the vma structure to a new location in the same mm,
3258 * prior to moving page table entries, to effect an mremap move.
3260 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3261 unsigned long addr, unsigned long len, pgoff_t pgoff,
3262 bool *need_rmap_locks)
3264 struct vm_area_struct *vma = *vmap;
3265 unsigned long vma_start = vma->vm_start;
3266 struct mm_struct *mm = vma->vm_mm;
3267 struct vm_area_struct *new_vma, *prev;
3268 struct rb_node **rb_link, *rb_parent;
3269 bool faulted_in_anon_vma = true;
3272 * If anonymous vma has not yet been faulted, update new pgoff
3273 * to match new location, to increase its chance of merging.
3275 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3276 pgoff = addr >> PAGE_SHIFT;
3277 faulted_in_anon_vma = false;
3280 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3281 return NULL; /* should never get here */
3282 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3283 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3284 vma->vm_userfaultfd_ctx);
3287 * Source vma may have been merged into new_vma
3289 if (unlikely(vma_start >= new_vma->vm_start &&
3290 vma_start < new_vma->vm_end)) {
3292 * The only way we can get a vma_merge with
3293 * self during an mremap is if the vma hasn't
3294 * been faulted in yet and we were allowed to
3295 * reset the dst vma->vm_pgoff to the
3296 * destination address of the mremap to allow
3297 * the merge to happen. mremap must change the
3298 * vm_pgoff linearity between src and dst vmas
3299 * (in turn preventing a vma_merge) to be
3300 * safe. It is only safe to keep the vm_pgoff
3301 * linear if there are no pages mapped yet.
3303 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3304 *vmap = vma = new_vma;
3306 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3308 new_vma = vm_area_dup(vma);
3311 new_vma->vm_start = addr;
3312 new_vma->vm_end = addr + len;
3313 new_vma->vm_pgoff = pgoff;
3314 if (vma_dup_policy(vma, new_vma))
3316 if (anon_vma_clone(new_vma, vma))
3317 goto out_free_mempol;
3318 if (new_vma->vm_file)
3319 get_file(new_vma->vm_file);
3320 if (new_vma->vm_ops && new_vma->vm_ops->open)
3321 new_vma->vm_ops->open(new_vma);
3322 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3323 *need_rmap_locks = false;
3328 mpol_put(vma_policy(new_vma));
3330 vm_area_free(new_vma);
3336 * Return true if the calling process may expand its vm space by the passed
3339 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3341 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3344 if (is_data_mapping(flags) &&
3345 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3346 /* Workaround for Valgrind */
3347 if (rlimit(RLIMIT_DATA) == 0 &&
3348 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3351 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3352 current->comm, current->pid,
3353 (mm->data_vm + npages) << PAGE_SHIFT,
3354 rlimit(RLIMIT_DATA),
3355 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3357 if (!ignore_rlimit_data)
3364 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3366 mm->total_vm += npages;
3368 if (is_exec_mapping(flags))
3369 mm->exec_vm += npages;
3370 else if (is_stack_mapping(flags))
3371 mm->stack_vm += npages;
3372 else if (is_data_mapping(flags))
3373 mm->data_vm += npages;
3376 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3379 * Having a close hook prevents vma merging regardless of flags.
3381 static void special_mapping_close(struct vm_area_struct *vma)
3385 static const char *special_mapping_name(struct vm_area_struct *vma)
3387 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3390 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3392 struct vm_special_mapping *sm = new_vma->vm_private_data;
3394 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3398 return sm->mremap(sm, new_vma);
3403 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3406 * Forbid splitting special mappings - kernel has expectations over
3407 * the number of pages in mapping. Together with VM_DONTEXPAND
3408 * the size of vma should stay the same over the special mapping's
3414 static const struct vm_operations_struct special_mapping_vmops = {
3415 .close = special_mapping_close,
3416 .fault = special_mapping_fault,
3417 .mremap = special_mapping_mremap,
3418 .name = special_mapping_name,
3419 /* vDSO code relies that VVAR can't be accessed remotely */
3421 .may_split = special_mapping_split,
3424 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3425 .close = special_mapping_close,
3426 .fault = special_mapping_fault,
3429 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3431 struct vm_area_struct *vma = vmf->vma;
3433 struct page **pages;
3435 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3436 pages = vma->vm_private_data;
3438 struct vm_special_mapping *sm = vma->vm_private_data;
3441 return sm->fault(sm, vmf->vma, vmf);
3446 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3450 struct page *page = *pages;
3456 return VM_FAULT_SIGBUS;
3459 static struct vm_area_struct *__install_special_mapping(
3460 struct mm_struct *mm,
3461 unsigned long addr, unsigned long len,
3462 unsigned long vm_flags, void *priv,
3463 const struct vm_operations_struct *ops)
3466 struct vm_area_struct *vma;
3468 vma = vm_area_alloc(mm);
3469 if (unlikely(vma == NULL))
3470 return ERR_PTR(-ENOMEM);
3472 vma->vm_start = addr;
3473 vma->vm_end = addr + len;
3475 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3476 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3479 vma->vm_private_data = priv;
3481 ret = insert_vm_struct(mm, vma);
3485 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3487 perf_event_mmap(vma);
3493 return ERR_PTR(ret);
3496 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3497 const struct vm_special_mapping *sm)
3499 return vma->vm_private_data == sm &&
3500 (vma->vm_ops == &special_mapping_vmops ||
3501 vma->vm_ops == &legacy_special_mapping_vmops);
3505 * Called with mm->mmap_lock held for writing.
3506 * Insert a new vma covering the given region, with the given flags.
3507 * Its pages are supplied by the given array of struct page *.
3508 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3509 * The region past the last page supplied will always produce SIGBUS.
3510 * The array pointer and the pages it points to are assumed to stay alive
3511 * for as long as this mapping might exist.
3513 struct vm_area_struct *_install_special_mapping(
3514 struct mm_struct *mm,
3515 unsigned long addr, unsigned long len,
3516 unsigned long vm_flags, const struct vm_special_mapping *spec)
3518 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3519 &special_mapping_vmops);
3522 int install_special_mapping(struct mm_struct *mm,
3523 unsigned long addr, unsigned long len,
3524 unsigned long vm_flags, struct page **pages)
3526 struct vm_area_struct *vma = __install_special_mapping(
3527 mm, addr, len, vm_flags, (void *)pages,
3528 &legacy_special_mapping_vmops);
3530 return PTR_ERR_OR_ZERO(vma);
3533 static DEFINE_MUTEX(mm_all_locks_mutex);
3535 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3537 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3539 * The LSB of head.next can't change from under us
3540 * because we hold the mm_all_locks_mutex.
3542 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3544 * We can safely modify head.next after taking the
3545 * anon_vma->root->rwsem. If some other vma in this mm shares
3546 * the same anon_vma we won't take it again.
3548 * No need of atomic instructions here, head.next
3549 * can't change from under us thanks to the
3550 * anon_vma->root->rwsem.
3552 if (__test_and_set_bit(0, (unsigned long *)
3553 &anon_vma->root->rb_root.rb_root.rb_node))
3558 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3560 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3562 * AS_MM_ALL_LOCKS can't change from under us because
3563 * we hold the mm_all_locks_mutex.
3565 * Operations on ->flags have to be atomic because
3566 * even if AS_MM_ALL_LOCKS is stable thanks to the
3567 * mm_all_locks_mutex, there may be other cpus
3568 * changing other bitflags in parallel to us.
3570 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3572 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3577 * This operation locks against the VM for all pte/vma/mm related
3578 * operations that could ever happen on a certain mm. This includes
3579 * vmtruncate, try_to_unmap, and all page faults.
3581 * The caller must take the mmap_lock in write mode before calling
3582 * mm_take_all_locks(). The caller isn't allowed to release the
3583 * mmap_lock until mm_drop_all_locks() returns.
3585 * mmap_lock in write mode is required in order to block all operations
3586 * that could modify pagetables and free pages without need of
3587 * altering the vma layout. It's also needed in write mode to avoid new
3588 * anon_vmas to be associated with existing vmas.
3590 * A single task can't take more than one mm_take_all_locks() in a row
3591 * or it would deadlock.
3593 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3594 * mapping->flags avoid to take the same lock twice, if more than one
3595 * vma in this mm is backed by the same anon_vma or address_space.
3597 * We take locks in following order, accordingly to comment at beginning
3599 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3601 * - all i_mmap_rwsem locks;
3602 * - all anon_vma->rwseml
3604 * We can take all locks within these types randomly because the VM code
3605 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3606 * mm_all_locks_mutex.
3608 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3609 * that may have to take thousand of locks.
3611 * mm_take_all_locks() can fail if it's interrupted by signals.
3613 int mm_take_all_locks(struct mm_struct *mm)
3615 struct vm_area_struct *vma;
3616 struct anon_vma_chain *avc;
3618 BUG_ON(mmap_read_trylock(mm));
3620 mutex_lock(&mm_all_locks_mutex);
3622 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3623 if (signal_pending(current))
3625 if (vma->vm_file && vma->vm_file->f_mapping &&
3626 is_vm_hugetlb_page(vma))
3627 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3630 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3631 if (signal_pending(current))
3633 if (vma->vm_file && vma->vm_file->f_mapping &&
3634 !is_vm_hugetlb_page(vma))
3635 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3638 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3639 if (signal_pending(current))
3642 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3643 vm_lock_anon_vma(mm, avc->anon_vma);
3649 mm_drop_all_locks(mm);
3653 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3655 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3657 * The LSB of head.next can't change to 0 from under
3658 * us because we hold the mm_all_locks_mutex.
3660 * We must however clear the bitflag before unlocking
3661 * the vma so the users using the anon_vma->rb_root will
3662 * never see our bitflag.
3664 * No need of atomic instructions here, head.next
3665 * can't change from under us until we release the
3666 * anon_vma->root->rwsem.
3668 if (!__test_and_clear_bit(0, (unsigned long *)
3669 &anon_vma->root->rb_root.rb_root.rb_node))
3671 anon_vma_unlock_write(anon_vma);
3675 static void vm_unlock_mapping(struct address_space *mapping)
3677 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3679 * AS_MM_ALL_LOCKS can't change to 0 from under us
3680 * because we hold the mm_all_locks_mutex.
3682 i_mmap_unlock_write(mapping);
3683 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3690 * The mmap_lock cannot be released by the caller until
3691 * mm_drop_all_locks() returns.
3693 void mm_drop_all_locks(struct mm_struct *mm)
3695 struct vm_area_struct *vma;
3696 struct anon_vma_chain *avc;
3698 BUG_ON(mmap_read_trylock(mm));
3699 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3701 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3703 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3704 vm_unlock_anon_vma(avc->anon_vma);
3705 if (vma->vm_file && vma->vm_file->f_mapping)
3706 vm_unlock_mapping(vma->vm_file->f_mapping);
3709 mutex_unlock(&mm_all_locks_mutex);
3713 * initialise the percpu counter for VM
3715 void __init mmap_init(void)
3719 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3724 * Initialise sysctl_user_reserve_kbytes.
3726 * This is intended to prevent a user from starting a single memory hogging
3727 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3730 * The default value is min(3% of free memory, 128MB)
3731 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3733 static int init_user_reserve(void)
3735 unsigned long free_kbytes;
3737 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3739 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3742 subsys_initcall(init_user_reserve);
3745 * Initialise sysctl_admin_reserve_kbytes.
3747 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3748 * to log in and kill a memory hogging process.
3750 * Systems with more than 256MB will reserve 8MB, enough to recover
3751 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3752 * only reserve 3% of free pages by default.
3754 static int init_admin_reserve(void)
3756 unsigned long free_kbytes;
3758 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3760 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3763 subsys_initcall(init_admin_reserve);
3766 * Reinititalise user and admin reserves if memory is added or removed.
3768 * The default user reserve max is 128MB, and the default max for the
3769 * admin reserve is 8MB. These are usually, but not always, enough to
3770 * enable recovery from a memory hogging process using login/sshd, a shell,
3771 * and tools like top. It may make sense to increase or even disable the
3772 * reserve depending on the existence of swap or variations in the recovery
3773 * tools. So, the admin may have changed them.
3775 * If memory is added and the reserves have been eliminated or increased above
3776 * the default max, then we'll trust the admin.
3778 * If memory is removed and there isn't enough free memory, then we
3779 * need to reset the reserves.
3781 * Otherwise keep the reserve set by the admin.
3783 static int reserve_mem_notifier(struct notifier_block *nb,
3784 unsigned long action, void *data)
3786 unsigned long tmp, free_kbytes;
3790 /* Default max is 128MB. Leave alone if modified by operator. */
3791 tmp = sysctl_user_reserve_kbytes;
3792 if (0 < tmp && tmp < (1UL << 17))
3793 init_user_reserve();
3795 /* Default max is 8MB. Leave alone if modified by operator. */
3796 tmp = sysctl_admin_reserve_kbytes;
3797 if (0 < tmp && tmp < (1UL << 13))
3798 init_admin_reserve();
3802 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3804 if (sysctl_user_reserve_kbytes > free_kbytes) {
3805 init_user_reserve();
3806 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3807 sysctl_user_reserve_kbytes);
3810 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3811 init_admin_reserve();
3812 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3813 sysctl_admin_reserve_kbytes);
3822 static struct notifier_block reserve_mem_nb = {
3823 .notifier_call = reserve_mem_notifier,
3826 static int __meminit init_reserve_notifier(void)
3828 if (register_hotmemory_notifier(&reserve_mem_nb))
3829 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3833 subsys_initcall(init_reserve_notifier);