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/mm_inline.h>
17 #include <linux/vmacache.h>
18 #include <linux/shm.h>
19 #include <linux/mman.h>
20 #include <linux/pagemap.h>
21 #include <linux/swap.h>
22 #include <linux/syscalls.h>
23 #include <linux/capability.h>
24 #include <linux/init.h>
25 #include <linux/file.h>
27 #include <linux/personality.h>
28 #include <linux/security.h>
29 #include <linux/hugetlb.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/profile.h>
32 #include <linux/export.h>
33 #include <linux/mount.h>
34 #include <linux/mempolicy.h>
35 #include <linux/rmap.h>
36 #include <linux/mmu_notifier.h>
37 #include <linux/mmdebug.h>
38 #include <linux/perf_event.h>
39 #include <linux/audit.h>
40 #include <linux/khugepaged.h>
41 #include <linux/uprobes.h>
42 #include <linux/rbtree_augmented.h>
43 #include <linux/notifier.h>
44 #include <linux/memory.h>
45 #include <linux/printk.h>
46 #include <linux/userfaultfd_k.h>
47 #include <linux/moduleparam.h>
48 #include <linux/pkeys.h>
49 #include <linux/oom.h>
50 #include <linux/sched/mm.h>
52 #include <linux/uaccess.h>
53 #include <asm/cacheflush.h>
55 #include <asm/mmu_context.h>
57 #define CREATE_TRACE_POINTS
58 #include <trace/events/mmap.h>
62 #ifndef arch_mmap_check
63 #define arch_mmap_check(addr, len, flags) (0)
66 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
67 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
68 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
69 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
71 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
72 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
73 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
74 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
77 static bool ignore_rlimit_data;
78 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
80 static void unmap_region(struct mm_struct *mm,
81 struct vm_area_struct *vma, struct vm_area_struct *prev,
82 unsigned long start, unsigned long end);
84 /* description of effects of mapping type and prot in current implementation.
85 * this is due to the limited x86 page protection hardware. The expected
86 * behavior is in parens:
89 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
90 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
91 * w: (no) no w: (no) no w: (yes) yes w: (no) no
92 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
94 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
95 * w: (no) no w: (no) no w: (copy) copy w: (no) no
96 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
98 * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
99 * MAP_PRIVATE (with Enhanced PAN supported):
104 pgprot_t protection_map[16] __ro_after_init = {
105 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
106 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
109 #ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
110 static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
116 pgprot_t vm_get_page_prot(unsigned long vm_flags)
118 pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
119 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
120 pgprot_val(arch_vm_get_page_prot(vm_flags)));
122 return arch_filter_pgprot(ret);
124 EXPORT_SYMBOL(vm_get_page_prot);
126 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
128 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
131 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
132 void vma_set_page_prot(struct vm_area_struct *vma)
134 unsigned long vm_flags = vma->vm_flags;
135 pgprot_t vm_page_prot;
137 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
138 if (vma_wants_writenotify(vma, vm_page_prot)) {
139 vm_flags &= ~VM_SHARED;
140 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
142 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
143 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
147 * Requires inode->i_mapping->i_mmap_rwsem
149 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
150 struct file *file, struct address_space *mapping)
152 if (vma->vm_flags & VM_SHARED)
153 mapping_unmap_writable(mapping);
155 flush_dcache_mmap_lock(mapping);
156 vma_interval_tree_remove(vma, &mapping->i_mmap);
157 flush_dcache_mmap_unlock(mapping);
161 * Unlink a file-based vm structure from its interval tree, to hide
162 * vma from rmap and vmtruncate before freeing its page tables.
164 void unlink_file_vma(struct vm_area_struct *vma)
166 struct file *file = vma->vm_file;
169 struct address_space *mapping = file->f_mapping;
170 i_mmap_lock_write(mapping);
171 __remove_shared_vm_struct(vma, file, mapping);
172 i_mmap_unlock_write(mapping);
177 * Close a vm structure and free it, returning the next.
179 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
181 struct vm_area_struct *next = vma->vm_next;
184 if (vma->vm_ops && vma->vm_ops->close)
185 vma->vm_ops->close(vma);
188 mpol_put(vma_policy(vma));
193 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags,
194 struct list_head *uf);
195 SYSCALL_DEFINE1(brk, unsigned long, brk)
197 unsigned long newbrk, oldbrk, origbrk;
198 struct mm_struct *mm = current->mm;
199 struct vm_area_struct *next;
200 unsigned long min_brk;
202 bool downgraded = false;
205 if (mmap_write_lock_killable(mm))
210 #ifdef CONFIG_COMPAT_BRK
212 * CONFIG_COMPAT_BRK can still be overridden by setting
213 * randomize_va_space to 2, which will still cause mm->start_brk
214 * to be arbitrarily shifted
216 if (current->brk_randomized)
217 min_brk = mm->start_brk;
219 min_brk = mm->end_data;
221 min_brk = mm->start_brk;
227 * Check against rlimit here. If this check is done later after the test
228 * of oldbrk with newbrk then it can escape the test and let the data
229 * segment grow beyond its set limit the in case where the limit is
230 * not page aligned -Ram Gupta
232 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
233 mm->end_data, mm->start_data))
236 newbrk = PAGE_ALIGN(brk);
237 oldbrk = PAGE_ALIGN(mm->brk);
238 if (oldbrk == newbrk) {
244 * Always allow shrinking brk.
245 * __do_munmap() may downgrade mmap_lock to read.
247 if (brk <= mm->brk) {
251 * mm->brk must to be protected by write mmap_lock so update it
252 * before downgrading mmap_lock. When __do_munmap() fails,
253 * mm->brk will be restored from origbrk.
256 ret = __do_munmap(mm, newbrk, oldbrk-newbrk, &uf, true);
260 } else if (ret == 1) {
266 /* Check against existing mmap mappings. */
267 next = find_vma(mm, oldbrk);
268 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
271 /* Ok, looks good - let it rip. */
272 if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
277 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
279 mmap_read_unlock(mm);
281 mmap_write_unlock(mm);
282 userfaultfd_unmap_complete(mm, &uf);
284 mm_populate(oldbrk, newbrk - oldbrk);
288 mmap_write_unlock(mm);
292 static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
294 unsigned long gap, prev_end;
297 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
298 * allow two stack_guard_gaps between them here, and when choosing
299 * an unmapped area; whereas when expanding we only require one.
300 * That's a little inconsistent, but keeps the code here simpler.
302 gap = vm_start_gap(vma);
304 prev_end = vm_end_gap(vma->vm_prev);
313 #ifdef CONFIG_DEBUG_VM_RB
314 static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma)
316 unsigned long max = vma_compute_gap(vma), subtree_gap;
317 if (vma->vm_rb.rb_left) {
318 subtree_gap = rb_entry(vma->vm_rb.rb_left,
319 struct vm_area_struct, vm_rb)->rb_subtree_gap;
320 if (subtree_gap > max)
323 if (vma->vm_rb.rb_right) {
324 subtree_gap = rb_entry(vma->vm_rb.rb_right,
325 struct vm_area_struct, vm_rb)->rb_subtree_gap;
326 if (subtree_gap > max)
332 static int browse_rb(struct mm_struct *mm)
334 struct rb_root *root = &mm->mm_rb;
335 int i = 0, j, bug = 0;
336 struct rb_node *nd, *pn = NULL;
337 unsigned long prev = 0, pend = 0;
339 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
340 struct vm_area_struct *vma;
341 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
342 if (vma->vm_start < prev) {
343 pr_emerg("vm_start %lx < prev %lx\n",
344 vma->vm_start, prev);
347 if (vma->vm_start < pend) {
348 pr_emerg("vm_start %lx < pend %lx\n",
349 vma->vm_start, pend);
352 if (vma->vm_start > vma->vm_end) {
353 pr_emerg("vm_start %lx > vm_end %lx\n",
354 vma->vm_start, vma->vm_end);
357 spin_lock(&mm->page_table_lock);
358 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
359 pr_emerg("free gap %lx, correct %lx\n",
361 vma_compute_subtree_gap(vma));
364 spin_unlock(&mm->page_table_lock);
367 prev = vma->vm_start;
371 for (nd = pn; nd; nd = rb_prev(nd))
374 pr_emerg("backwards %d, forwards %d\n", j, i);
380 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
384 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
385 struct vm_area_struct *vma;
386 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
387 VM_BUG_ON_VMA(vma != ignore &&
388 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
393 static void validate_mm(struct mm_struct *mm)
397 unsigned long highest_address = 0;
398 struct vm_area_struct *vma = mm->mmap;
401 struct anon_vma *anon_vma = vma->anon_vma;
402 struct anon_vma_chain *avc;
405 anon_vma_lock_read(anon_vma);
406 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
407 anon_vma_interval_tree_verify(avc);
408 anon_vma_unlock_read(anon_vma);
411 highest_address = vm_end_gap(vma);
415 if (i != mm->map_count) {
416 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
419 if (highest_address != mm->highest_vm_end) {
420 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
421 mm->highest_vm_end, highest_address);
425 if (i != mm->map_count) {
427 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
430 VM_BUG_ON_MM(bug, mm);
433 #define validate_mm_rb(root, ignore) do { } while (0)
434 #define validate_mm(mm) do { } while (0)
437 RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks,
438 struct vm_area_struct, vm_rb,
439 unsigned long, rb_subtree_gap, vma_compute_gap)
442 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
443 * vma->vm_prev->vm_end values changed, without modifying the vma's position
446 static void vma_gap_update(struct vm_area_struct *vma)
449 * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
450 * a callback function that does exactly what we want.
452 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
455 static inline void vma_rb_insert(struct vm_area_struct *vma,
456 struct rb_root *root)
458 /* All rb_subtree_gap values must be consistent prior to insertion */
459 validate_mm_rb(root, NULL);
461 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
464 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
467 * Note rb_erase_augmented is a fairly large inline function,
468 * so make sure we instantiate it only once with our desired
469 * augmented rbtree callbacks.
471 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
474 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
475 struct rb_root *root,
476 struct vm_area_struct *ignore)
479 * All rb_subtree_gap values must be consistent prior to erase,
480 * with the possible exception of
482 * a. the "next" vma being erased if next->vm_start was reduced in
483 * __vma_adjust() -> __vma_unlink()
484 * b. the vma being erased in detach_vmas_to_be_unmapped() ->
487 validate_mm_rb(root, ignore);
489 __vma_rb_erase(vma, root);
492 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
493 struct rb_root *root)
495 vma_rb_erase_ignore(vma, root, vma);
499 * vma has some anon_vma assigned, and is already inserted on that
500 * anon_vma's interval trees.
502 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
503 * vma must be removed from the anon_vma's interval trees using
504 * anon_vma_interval_tree_pre_update_vma().
506 * After the update, the vma will be reinserted using
507 * anon_vma_interval_tree_post_update_vma().
509 * The entire update must be protected by exclusive mmap_lock and by
510 * the root anon_vma's mutex.
513 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
515 struct anon_vma_chain *avc;
517 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
518 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
522 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
524 struct anon_vma_chain *avc;
526 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
527 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
530 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
531 unsigned long end, struct vm_area_struct **pprev,
532 struct rb_node ***rb_link, struct rb_node **rb_parent)
534 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
536 mmap_assert_locked(mm);
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_SHARED)
670 mapping_allow_writable(mapping);
672 flush_dcache_mmap_lock(mapping);
673 vma_interval_tree_insert(vma, &mapping->i_mmap);
674 flush_dcache_mmap_unlock(mapping);
679 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
680 struct vm_area_struct *prev, struct rb_node **rb_link,
681 struct rb_node *rb_parent)
683 __vma_link_list(mm, vma, prev);
684 __vma_link_rb(mm, vma, rb_link, rb_parent);
687 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
688 struct vm_area_struct *prev, struct rb_node **rb_link,
689 struct rb_node *rb_parent)
691 struct address_space *mapping = NULL;
694 mapping = vma->vm_file->f_mapping;
695 i_mmap_lock_write(mapping);
698 __vma_link(mm, vma, prev, rb_link, rb_parent);
699 __vma_link_file(vma);
702 i_mmap_unlock_write(mapping);
709 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
710 * mm's list and rbtree. It has already been inserted into the interval tree.
712 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
714 struct vm_area_struct *prev;
715 struct rb_node **rb_link, *rb_parent;
717 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
718 &prev, &rb_link, &rb_parent))
720 __vma_link(mm, vma, prev, rb_link, rb_parent);
724 static __always_inline void __vma_unlink(struct mm_struct *mm,
725 struct vm_area_struct *vma,
726 struct vm_area_struct *ignore)
728 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
729 __vma_unlink_list(mm, vma);
731 vmacache_invalidate(mm);
735 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
736 * is already present in an i_mmap tree without adjusting the tree.
737 * The following helper function should be used when such adjustments
738 * are necessary. The "insert" vma (if any) is to be inserted
739 * before we drop the necessary locks.
741 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
742 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
743 struct vm_area_struct *expand)
745 struct mm_struct *mm = vma->vm_mm;
746 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
747 struct address_space *mapping = NULL;
748 struct rb_root_cached *root = NULL;
749 struct anon_vma *anon_vma = NULL;
750 struct file *file = vma->vm_file;
751 bool start_changed = false, end_changed = false;
752 long adjust_next = 0;
755 if (next && !insert) {
756 struct vm_area_struct *exporter = NULL, *importer = NULL;
758 if (end >= next->vm_end) {
760 * vma expands, overlapping all the next, and
761 * perhaps the one after too (mprotect case 6).
762 * The only other cases that gets here are
763 * case 1, case 7 and case 8.
765 if (next == expand) {
767 * The only case where we don't expand "vma"
768 * and we expand "next" instead is case 8.
770 VM_WARN_ON(end != next->vm_end);
772 * remove_next == 3 means we're
773 * removing "vma" and that to do so we
774 * swapped "vma" and "next".
777 VM_WARN_ON(file != next->vm_file);
780 VM_WARN_ON(expand != vma);
782 * case 1, 6, 7, remove_next == 2 is case 6,
783 * remove_next == 1 is case 1 or 7.
785 remove_next = 1 + (end > next->vm_end);
786 VM_WARN_ON(remove_next == 2 &&
787 end != next->vm_next->vm_end);
788 /* trim end to next, for case 6 first pass */
796 * If next doesn't have anon_vma, import from vma after
797 * next, if the vma overlaps with it.
799 if (remove_next == 2 && !next->anon_vma)
800 exporter = next->vm_next;
802 } else if (end > next->vm_start) {
804 * vma expands, overlapping part of the next:
805 * mprotect case 5 shifting the boundary up.
807 adjust_next = (end - next->vm_start);
810 VM_WARN_ON(expand != importer);
811 } else if (end < vma->vm_end) {
813 * vma shrinks, and !insert tells it's not
814 * split_vma inserting another: so it must be
815 * mprotect case 4 shifting the boundary down.
817 adjust_next = -(vma->vm_end - end);
820 VM_WARN_ON(expand != importer);
824 * Easily overlooked: when mprotect shifts the boundary,
825 * make sure the expanding vma has anon_vma set if the
826 * shrinking vma had, to cover any anon pages imported.
828 if (exporter && exporter->anon_vma && !importer->anon_vma) {
831 importer->anon_vma = exporter->anon_vma;
832 error = anon_vma_clone(importer, exporter);
838 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
841 mapping = file->f_mapping;
842 root = &mapping->i_mmap;
843 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
846 uprobe_munmap(next, next->vm_start, next->vm_end);
848 i_mmap_lock_write(mapping);
851 * Put into interval tree now, so instantiated pages
852 * are visible to arm/parisc __flush_dcache_page
853 * throughout; but we cannot insert into address
854 * space until vma start or end is updated.
856 __vma_link_file(insert);
860 anon_vma = vma->anon_vma;
861 if (!anon_vma && adjust_next)
862 anon_vma = next->anon_vma;
864 VM_WARN_ON(adjust_next && next->anon_vma &&
865 anon_vma != next->anon_vma);
866 anon_vma_lock_write(anon_vma);
867 anon_vma_interval_tree_pre_update_vma(vma);
869 anon_vma_interval_tree_pre_update_vma(next);
873 flush_dcache_mmap_lock(mapping);
874 vma_interval_tree_remove(vma, root);
876 vma_interval_tree_remove(next, root);
879 if (start != vma->vm_start) {
880 vma->vm_start = start;
881 start_changed = true;
883 if (end != vma->vm_end) {
887 vma->vm_pgoff = pgoff;
889 next->vm_start += adjust_next;
890 next->vm_pgoff += adjust_next >> PAGE_SHIFT;
895 vma_interval_tree_insert(next, root);
896 vma_interval_tree_insert(vma, root);
897 flush_dcache_mmap_unlock(mapping);
902 * vma_merge has merged next into vma, and needs
903 * us to remove next before dropping the locks.
905 if (remove_next != 3)
906 __vma_unlink(mm, next, next);
909 * vma is not before next if they've been
912 * pre-swap() next->vm_start was reduced so
913 * tell validate_mm_rb to ignore pre-swap()
914 * "next" (which is stored in post-swap()
917 __vma_unlink(mm, next, vma);
919 __remove_shared_vm_struct(next, file, mapping);
922 * split_vma has split insert from vma, and needs
923 * us to insert it before dropping the locks
924 * (it may either follow vma or precede it).
926 __insert_vm_struct(mm, insert);
932 mm->highest_vm_end = vm_end_gap(vma);
933 else if (!adjust_next)
934 vma_gap_update(next);
939 anon_vma_interval_tree_post_update_vma(vma);
941 anon_vma_interval_tree_post_update_vma(next);
942 anon_vma_unlock_write(anon_vma);
946 i_mmap_unlock_write(mapping);
955 uprobe_munmap(next, next->vm_start, next->vm_end);
959 anon_vma_merge(vma, next);
961 mpol_put(vma_policy(next));
964 * In mprotect's case 6 (see comments on vma_merge),
965 * we must remove another next too. It would clutter
966 * up the code too much to do both in one go.
968 if (remove_next != 3) {
970 * If "next" was removed and vma->vm_end was
971 * expanded (up) over it, in turn
972 * "next->vm_prev->vm_end" changed and the
973 * "vma->vm_next" gap must be updated.
978 * For the scope of the comment "next" and
979 * "vma" considered pre-swap(): if "vma" was
980 * removed, next->vm_start was expanded (down)
981 * over it and the "next" gap must be updated.
982 * Because of the swap() the post-swap() "vma"
983 * actually points to pre-swap() "next"
984 * (post-swap() "next" as opposed is now a
989 if (remove_next == 2) {
995 vma_gap_update(next);
998 * If remove_next == 2 we obviously can't
1001 * If remove_next == 3 we can't reach this
1002 * path because pre-swap() next is always not
1003 * NULL. pre-swap() "next" is not being
1004 * removed and its next->vm_end is not altered
1005 * (and furthermore "end" already matches
1006 * next->vm_end in remove_next == 3).
1008 * We reach this only in the remove_next == 1
1009 * case if the "next" vma that was removed was
1010 * the highest vma of the mm. However in such
1011 * case next->vm_end == "end" and the extended
1012 * "vma" has vma->vm_end == next->vm_end so
1013 * mm->highest_vm_end doesn't need any update
1014 * in remove_next == 1 case.
1016 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
1020 uprobe_mmap(insert);
1028 * If the vma has a ->close operation then the driver probably needs to release
1029 * per-vma resources, so we don't attempt to merge those.
1031 static inline int is_mergeable_vma(struct vm_area_struct *vma,
1032 struct file *file, unsigned long vm_flags,
1033 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1034 const char *anon_name)
1037 * VM_SOFTDIRTY should not prevent from VMA merging, if we
1038 * match the flags but dirty bit -- the caller should mark
1039 * merged VMA as dirty. If dirty bit won't be excluded from
1040 * comparison, we increase pressure on the memory system forcing
1041 * the kernel to generate new VMAs when old one could be
1044 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
1046 if (vma->vm_file != file)
1048 if (vma->vm_ops && vma->vm_ops->close)
1050 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1052 if (!is_same_vma_anon_name(vma, anon_name))
1057 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1058 struct anon_vma *anon_vma2,
1059 struct vm_area_struct *vma)
1062 * The list_is_singular() test is to avoid merging VMA cloned from
1063 * parents. This can improve scalability caused by anon_vma lock.
1065 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1066 list_is_singular(&vma->anon_vma_chain)))
1068 return anon_vma1 == anon_vma2;
1072 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1073 * in front of (at a lower virtual address and file offset than) the vma.
1075 * We cannot merge two vmas if they have differently assigned (non-NULL)
1076 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1078 * We don't check here for the merged mmap wrapping around the end of pagecache
1079 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
1080 * wrap, nor mmaps which cover the final page at index -1UL.
1083 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1084 struct anon_vma *anon_vma, struct file *file,
1086 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1087 const char *anon_name)
1089 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
1090 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1091 if (vma->vm_pgoff == vm_pgoff)
1098 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1099 * beyond (at a higher virtual address and file offset than) the vma.
1101 * We cannot merge two vmas if they have differently assigned (non-NULL)
1102 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1105 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1106 struct anon_vma *anon_vma, struct file *file,
1108 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1109 const char *anon_name)
1111 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
1112 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1114 vm_pglen = vma_pages(vma);
1115 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1122 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
1123 * figure out whether that can be merged with its predecessor or its
1124 * successor. Or both (it neatly fills a hole).
1126 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1127 * certain not to be mapped by the time vma_merge is called; but when
1128 * called for mprotect, it is certain to be already mapped (either at
1129 * an offset within prev, or at the start of next), and the flags of
1130 * this area are about to be changed to vm_flags - and the no-change
1131 * case has already been eliminated.
1133 * The following mprotect cases have to be considered, where AAAA is
1134 * the area passed down from mprotect_fixup, never extending beyond one
1135 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1138 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN
1139 * cannot merge might become might become
1140 * PPNNNNNNNNNN PPPPPPPPPPNN
1141 * mmap, brk or case 4 below case 5 below
1144 * PPPP NNNN PPPPNNNNXXXX
1145 * might become might become
1146 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
1147 * PPPPPPPPNNNN 2 or PPPPPPPPXXXX 7 or
1148 * PPPPNNNNNNNN 3 PPPPXXXXXXXX 8
1150 * It is important for case 8 that the vma NNNN overlapping the
1151 * region AAAA is never going to extended over XXXX. Instead XXXX must
1152 * be extended in region AAAA and NNNN must be removed. This way in
1153 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1154 * rmap_locks, the properties of the merged vma will be already
1155 * correct for the whole merged range. Some of those properties like
1156 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1157 * be correct for the whole merged range immediately after the
1158 * rmap_locks are released. Otherwise if XXXX would be removed and
1159 * NNNN would be extended over the XXXX range, remove_migration_ptes
1160 * or other rmap walkers (if working on addresses beyond the "end"
1161 * parameter) may establish ptes with the wrong permissions of NNNN
1162 * instead of the right permissions of XXXX.
1164 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1165 struct vm_area_struct *prev, unsigned long addr,
1166 unsigned long end, unsigned long vm_flags,
1167 struct anon_vma *anon_vma, struct file *file,
1168 pgoff_t pgoff, struct mempolicy *policy,
1169 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1170 const char *anon_name)
1172 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1173 struct vm_area_struct *area, *next;
1177 * We later require that vma->vm_flags == vm_flags,
1178 * so this tests vma->vm_flags & VM_SPECIAL, too.
1180 if (vm_flags & VM_SPECIAL)
1183 next = vma_next(mm, prev);
1185 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1186 next = next->vm_next;
1188 /* verify some invariant that must be enforced by the caller */
1189 VM_WARN_ON(prev && addr <= prev->vm_start);
1190 VM_WARN_ON(area && end > area->vm_end);
1191 VM_WARN_ON(addr >= end);
1194 * Can it merge with the predecessor?
1196 if (prev && prev->vm_end == addr &&
1197 mpol_equal(vma_policy(prev), policy) &&
1198 can_vma_merge_after(prev, vm_flags,
1199 anon_vma, file, pgoff,
1200 vm_userfaultfd_ctx, anon_name)) {
1202 * OK, it can. Can we now merge in the successor as well?
1204 if (next && end == next->vm_start &&
1205 mpol_equal(policy, vma_policy(next)) &&
1206 can_vma_merge_before(next, vm_flags,
1209 vm_userfaultfd_ctx, anon_name) &&
1210 is_mergeable_anon_vma(prev->anon_vma,
1211 next->anon_vma, NULL)) {
1213 err = __vma_adjust(prev, prev->vm_start,
1214 next->vm_end, prev->vm_pgoff, NULL,
1216 } else /* cases 2, 5, 7 */
1217 err = __vma_adjust(prev, prev->vm_start,
1218 end, prev->vm_pgoff, NULL, prev);
1221 khugepaged_enter_vma_merge(prev, vm_flags);
1226 * Can this new request be merged in front of next?
1228 if (next && end == next->vm_start &&
1229 mpol_equal(policy, vma_policy(next)) &&
1230 can_vma_merge_before(next, vm_flags,
1231 anon_vma, file, pgoff+pglen,
1232 vm_userfaultfd_ctx, anon_name)) {
1233 if (prev && addr < prev->vm_end) /* case 4 */
1234 err = __vma_adjust(prev, prev->vm_start,
1235 addr, prev->vm_pgoff, NULL, next);
1236 else { /* cases 3, 8 */
1237 err = __vma_adjust(area, addr, next->vm_end,
1238 next->vm_pgoff - pglen, NULL, next);
1240 * In case 3 area is already equal to next and
1241 * this is a noop, but in case 8 "area" has
1242 * been removed and next was expanded over it.
1248 khugepaged_enter_vma_merge(area, vm_flags);
1256 * Rough compatibility check to quickly see if it's even worth looking
1257 * at sharing an anon_vma.
1259 * They need to have the same vm_file, and the flags can only differ
1260 * in things that mprotect may change.
1262 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1263 * we can merge the two vma's. For example, we refuse to merge a vma if
1264 * there is a vm_ops->close() function, because that indicates that the
1265 * driver is doing some kind of reference counting. But that doesn't
1266 * really matter for the anon_vma sharing case.
1268 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1270 return a->vm_end == b->vm_start &&
1271 mpol_equal(vma_policy(a), vma_policy(b)) &&
1272 a->vm_file == b->vm_file &&
1273 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1274 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1278 * Do some basic sanity checking to see if we can re-use the anon_vma
1279 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1280 * the same as 'old', the other will be the new one that is trying
1281 * to share the anon_vma.
1283 * NOTE! This runs with mm_sem held for reading, so it is possible that
1284 * the anon_vma of 'old' is concurrently in the process of being set up
1285 * by another page fault trying to merge _that_. But that's ok: if it
1286 * is being set up, that automatically means that it will be a singleton
1287 * acceptable for merging, so we can do all of this optimistically. But
1288 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1290 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1291 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1292 * is to return an anon_vma that is "complex" due to having gone through
1295 * We also make sure that the two vma's are compatible (adjacent,
1296 * and with the same memory policies). That's all stable, even with just
1297 * a read lock on the mm_sem.
1299 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1301 if (anon_vma_compatible(a, b)) {
1302 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1304 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1311 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1312 * neighbouring vmas for a suitable anon_vma, before it goes off
1313 * to allocate a new anon_vma. It checks because a repetitive
1314 * sequence of mprotects and faults may otherwise lead to distinct
1315 * anon_vmas being allocated, preventing vma merge in subsequent
1318 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1320 struct anon_vma *anon_vma = NULL;
1322 /* Try next first. */
1324 anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next);
1329 /* Try prev next. */
1331 anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma);
1334 * We might reach here with anon_vma == NULL if we can't find
1335 * any reusable anon_vma.
1336 * There's no absolute need to look only at touching neighbours:
1337 * we could search further afield for "compatible" anon_vmas.
1338 * But it would probably just be a waste of time searching,
1339 * or lead to too many vmas hanging off the same anon_vma.
1340 * We're trying to allow mprotect remerging later on,
1341 * not trying to minimize memory used for anon_vmas.
1347 * If a hint addr is less than mmap_min_addr change hint to be as
1348 * low as possible but still greater than mmap_min_addr
1350 static inline unsigned long round_hint_to_min(unsigned long hint)
1353 if (((void *)hint != NULL) &&
1354 (hint < mmap_min_addr))
1355 return PAGE_ALIGN(mmap_min_addr);
1359 int mlock_future_check(struct mm_struct *mm, unsigned long flags,
1362 unsigned long locked, lock_limit;
1364 /* mlock MCL_FUTURE? */
1365 if (flags & VM_LOCKED) {
1366 locked = len >> PAGE_SHIFT;
1367 locked += mm->locked_vm;
1368 lock_limit = rlimit(RLIMIT_MEMLOCK);
1369 lock_limit >>= PAGE_SHIFT;
1370 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1376 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1378 if (S_ISREG(inode->i_mode))
1379 return MAX_LFS_FILESIZE;
1381 if (S_ISBLK(inode->i_mode))
1382 return MAX_LFS_FILESIZE;
1384 if (S_ISSOCK(inode->i_mode))
1385 return MAX_LFS_FILESIZE;
1387 /* Special "we do even unsigned file positions" case */
1388 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1391 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1395 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1396 unsigned long pgoff, unsigned long len)
1398 u64 maxsize = file_mmap_size_max(file, inode);
1400 if (maxsize && len > maxsize)
1403 if (pgoff > maxsize >> PAGE_SHIFT)
1409 * The caller must write-lock current->mm->mmap_lock.
1411 unsigned long do_mmap(struct file *file, unsigned long addr,
1412 unsigned long len, unsigned long prot,
1413 unsigned long flags, unsigned long pgoff,
1414 unsigned long *populate, struct list_head *uf)
1416 struct mm_struct *mm = current->mm;
1417 vm_flags_t vm_flags;
1426 * Does the application expect PROT_READ to imply PROT_EXEC?
1428 * (the exception is when the underlying filesystem is noexec
1429 * mounted, in which case we dont add PROT_EXEC.)
1431 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1432 if (!(file && path_noexec(&file->f_path)))
1435 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1436 if (flags & MAP_FIXED_NOREPLACE)
1439 if (!(flags & MAP_FIXED))
1440 addr = round_hint_to_min(addr);
1442 /* Careful about overflows.. */
1443 len = PAGE_ALIGN(len);
1447 /* offset overflow? */
1448 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1451 /* Too many mappings? */
1452 if (mm->map_count > sysctl_max_map_count)
1455 /* Obtain the address to map to. we verify (or select) it and ensure
1456 * that it represents a valid section of the address space.
1458 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1459 if (IS_ERR_VALUE(addr))
1462 if (flags & MAP_FIXED_NOREPLACE) {
1463 if (find_vma_intersection(mm, addr, addr + len))
1467 if (prot == PROT_EXEC) {
1468 pkey = execute_only_pkey(mm);
1473 /* Do simple checking here so the lower-level routines won't have
1474 * to. we assume access permissions have been handled by the open
1475 * of the memory object, so we don't do any here.
1477 vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1478 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1480 if (flags & MAP_LOCKED)
1481 if (!can_do_mlock())
1484 if (mlock_future_check(mm, vm_flags, len))
1488 struct inode *inode = file_inode(file);
1489 unsigned long flags_mask;
1491 if (!file_mmap_ok(file, inode, pgoff, len))
1494 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1496 switch (flags & MAP_TYPE) {
1499 * Force use of MAP_SHARED_VALIDATE with non-legacy
1500 * flags. E.g. MAP_SYNC is dangerous to use with
1501 * MAP_SHARED as you don't know which consistency model
1502 * you will get. We silently ignore unsupported flags
1503 * with MAP_SHARED to preserve backward compatibility.
1505 flags &= LEGACY_MAP_MASK;
1507 case MAP_SHARED_VALIDATE:
1508 if (flags & ~flags_mask)
1510 if (prot & PROT_WRITE) {
1511 if (!(file->f_mode & FMODE_WRITE))
1513 if (IS_SWAPFILE(file->f_mapping->host))
1518 * Make sure we don't allow writing to an append-only
1521 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1524 vm_flags |= VM_SHARED | VM_MAYSHARE;
1525 if (!(file->f_mode & FMODE_WRITE))
1526 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1529 if (!(file->f_mode & FMODE_READ))
1531 if (path_noexec(&file->f_path)) {
1532 if (vm_flags & VM_EXEC)
1534 vm_flags &= ~VM_MAYEXEC;
1537 if (!file->f_op->mmap)
1539 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1547 switch (flags & MAP_TYPE) {
1549 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1555 vm_flags |= VM_SHARED | VM_MAYSHARE;
1559 * Set pgoff according to addr for anon_vma.
1561 pgoff = addr >> PAGE_SHIFT;
1569 * Set 'VM_NORESERVE' if we should not account for the
1570 * memory use of this mapping.
1572 if (flags & MAP_NORESERVE) {
1573 /* We honor MAP_NORESERVE if allowed to overcommit */
1574 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1575 vm_flags |= VM_NORESERVE;
1577 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1578 if (file && is_file_hugepages(file))
1579 vm_flags |= VM_NORESERVE;
1582 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1583 if (!IS_ERR_VALUE(addr) &&
1584 ((vm_flags & VM_LOCKED) ||
1585 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1590 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1591 unsigned long prot, unsigned long flags,
1592 unsigned long fd, unsigned long pgoff)
1594 struct file *file = NULL;
1595 unsigned long retval;
1597 if (!(flags & MAP_ANONYMOUS)) {
1598 audit_mmap_fd(fd, flags);
1602 if (is_file_hugepages(file)) {
1603 len = ALIGN(len, huge_page_size(hstate_file(file)));
1604 } else if (unlikely(flags & MAP_HUGETLB)) {
1608 } else if (flags & MAP_HUGETLB) {
1611 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1615 len = ALIGN(len, huge_page_size(hs));
1617 * VM_NORESERVE is used because the reservations will be
1618 * taken when vm_ops->mmap() is called
1619 * A dummy user value is used because we are not locking
1620 * memory so no accounting is necessary
1622 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1624 HUGETLB_ANONHUGE_INODE,
1625 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1627 return PTR_ERR(file);
1630 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1637 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1638 unsigned long, prot, unsigned long, flags,
1639 unsigned long, fd, unsigned long, pgoff)
1641 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1644 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1645 struct mmap_arg_struct {
1649 unsigned long flags;
1651 unsigned long offset;
1654 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1656 struct mmap_arg_struct a;
1658 if (copy_from_user(&a, arg, sizeof(a)))
1660 if (offset_in_page(a.offset))
1663 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1664 a.offset >> PAGE_SHIFT);
1666 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1669 * Some shared mappings will want the pages marked read-only
1670 * to track write events. If so, we'll downgrade vm_page_prot
1671 * to the private version (using protection_map[] without the
1674 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1676 vm_flags_t vm_flags = vma->vm_flags;
1677 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1679 /* If it was private or non-writable, the write bit is already clear */
1680 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1683 /* The backer wishes to know when pages are first written to? */
1684 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1687 /* The open routine did something to the protections that pgprot_modify
1688 * won't preserve? */
1689 if (pgprot_val(vm_page_prot) !=
1690 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1693 /* Do we need to track softdirty? */
1694 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1697 /* Specialty mapping? */
1698 if (vm_flags & VM_PFNMAP)
1701 /* Can the mapping track the dirty pages? */
1702 return vma->vm_file && vma->vm_file->f_mapping &&
1703 mapping_can_writeback(vma->vm_file->f_mapping);
1707 * We account for memory if it's a private writeable mapping,
1708 * not hugepages and VM_NORESERVE wasn't set.
1710 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1713 * hugetlb has its own accounting separate from the core VM
1714 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1716 if (file && is_file_hugepages(file))
1719 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1722 unsigned long mmap_region(struct file *file, unsigned long addr,
1723 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1724 struct list_head *uf)
1726 struct mm_struct *mm = current->mm;
1727 struct vm_area_struct *vma, *prev, *merge;
1729 struct rb_node **rb_link, *rb_parent;
1730 unsigned long charged = 0;
1732 /* Check against address space limit. */
1733 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1734 unsigned long nr_pages;
1737 * MAP_FIXED may remove pages of mappings that intersects with
1738 * requested mapping. Account for the pages it would unmap.
1740 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1742 if (!may_expand_vm(mm, vm_flags,
1743 (len >> PAGE_SHIFT) - nr_pages))
1747 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
1748 if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
1751 * Private writable mapping: check memory availability
1753 if (accountable_mapping(file, vm_flags)) {
1754 charged = len >> PAGE_SHIFT;
1755 if (security_vm_enough_memory_mm(mm, charged))
1757 vm_flags |= VM_ACCOUNT;
1761 * Can we just expand an old mapping?
1763 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1764 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
1769 * Determine the object being mapped and call the appropriate
1770 * specific mapper. the address has already been validated, but
1771 * not unmapped, but the maps are removed from the list.
1773 vma = vm_area_alloc(mm);
1779 vma->vm_start = addr;
1780 vma->vm_end = addr + len;
1781 vma->vm_flags = vm_flags;
1782 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1783 vma->vm_pgoff = pgoff;
1786 if (vm_flags & VM_SHARED) {
1787 error = mapping_map_writable(file->f_mapping);
1792 vma->vm_file = get_file(file);
1793 error = call_mmap(file, vma);
1795 goto unmap_and_free_vma;
1797 /* Can addr have changed??
1799 * Answer: Yes, several device drivers can do it in their
1800 * f_op->mmap method. -DaveM
1801 * Bug: If addr is changed, prev, rb_link, rb_parent should
1802 * be updated for vma_link()
1804 WARN_ON_ONCE(addr != vma->vm_start);
1806 addr = vma->vm_start;
1808 /* If vm_flags changed after call_mmap(), we should try merge vma again
1809 * as we may succeed this time.
1811 if (unlikely(vm_flags != vma->vm_flags && prev)) {
1812 merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
1813 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
1815 /* ->mmap() can change vma->vm_file and fput the original file. So
1816 * fput the vma->vm_file here or we would add an extra fput for file
1817 * and cause general protection fault ultimately.
1822 /* Update vm_flags to pick up the change. */
1823 vm_flags = vma->vm_flags;
1824 goto unmap_writable;
1828 vm_flags = vma->vm_flags;
1829 } else if (vm_flags & VM_SHARED) {
1830 error = shmem_zero_setup(vma);
1834 vma_set_anonymous(vma);
1837 /* Allow architectures to sanity-check the vm_flags */
1838 if (!arch_validate_flags(vma->vm_flags)) {
1841 goto unmap_and_free_vma;
1846 vma_link(mm, vma, prev, rb_link, rb_parent);
1847 /* Once vma denies write, undo our temporary denial count */
1849 if (file && vm_flags & VM_SHARED)
1850 mapping_unmap_writable(file->f_mapping);
1851 file = vma->vm_file;
1853 perf_event_mmap(vma);
1855 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1856 if (vm_flags & VM_LOCKED) {
1857 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1858 is_vm_hugetlb_page(vma) ||
1859 vma == get_gate_vma(current->mm))
1860 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1862 mm->locked_vm += (len >> PAGE_SHIFT);
1869 * New (or expanded) vma always get soft dirty status.
1870 * Otherwise user-space soft-dirty page tracker won't
1871 * be able to distinguish situation when vma area unmapped,
1872 * then new mapped in-place (which must be aimed as
1873 * a completely new data area).
1875 vma->vm_flags |= VM_SOFTDIRTY;
1877 vma_set_page_prot(vma);
1883 vma->vm_file = NULL;
1885 /* Undo any partial mapping done by a device driver. */
1886 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1888 if (vm_flags & VM_SHARED)
1889 mapping_unmap_writable(file->f_mapping);
1894 vm_unacct_memory(charged);
1898 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1901 * We implement the search by looking for an rbtree node that
1902 * immediately follows a suitable gap. That is,
1903 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1904 * - gap_end = vma->vm_start >= info->low_limit + length;
1905 * - gap_end - gap_start >= length
1908 struct mm_struct *mm = current->mm;
1909 struct vm_area_struct *vma;
1910 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1912 /* Adjust search length to account for worst case alignment overhead */
1913 length = info->length + info->align_mask;
1914 if (length < info->length)
1917 /* Adjust search limits by the desired length */
1918 if (info->high_limit < length)
1920 high_limit = info->high_limit - length;
1922 if (info->low_limit > high_limit)
1924 low_limit = info->low_limit + length;
1926 /* Check if rbtree root looks promising */
1927 if (RB_EMPTY_ROOT(&mm->mm_rb))
1929 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1930 if (vma->rb_subtree_gap < length)
1934 /* Visit left subtree if it looks promising */
1935 gap_end = vm_start_gap(vma);
1936 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1937 struct vm_area_struct *left =
1938 rb_entry(vma->vm_rb.rb_left,
1939 struct vm_area_struct, vm_rb);
1940 if (left->rb_subtree_gap >= length) {
1946 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1948 /* Check if current node has a suitable gap */
1949 if (gap_start > high_limit)
1951 if (gap_end >= low_limit &&
1952 gap_end > gap_start && gap_end - gap_start >= length)
1955 /* Visit right subtree if it looks promising */
1956 if (vma->vm_rb.rb_right) {
1957 struct vm_area_struct *right =
1958 rb_entry(vma->vm_rb.rb_right,
1959 struct vm_area_struct, vm_rb);
1960 if (right->rb_subtree_gap >= length) {
1966 /* Go back up the rbtree to find next candidate node */
1968 struct rb_node *prev = &vma->vm_rb;
1969 if (!rb_parent(prev))
1971 vma = rb_entry(rb_parent(prev),
1972 struct vm_area_struct, vm_rb);
1973 if (prev == vma->vm_rb.rb_left) {
1974 gap_start = vm_end_gap(vma->vm_prev);
1975 gap_end = vm_start_gap(vma);
1982 /* Check highest gap, which does not precede any rbtree node */
1983 gap_start = mm->highest_vm_end;
1984 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
1985 if (gap_start > high_limit)
1989 /* We found a suitable gap. Clip it with the original low_limit. */
1990 if (gap_start < info->low_limit)
1991 gap_start = info->low_limit;
1993 /* Adjust gap address to the desired alignment */
1994 gap_start += (info->align_offset - gap_start) & info->align_mask;
1996 VM_BUG_ON(gap_start + info->length > info->high_limit);
1997 VM_BUG_ON(gap_start + info->length > gap_end);
2001 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
2003 struct mm_struct *mm = current->mm;
2004 struct vm_area_struct *vma;
2005 unsigned long length, low_limit, high_limit, gap_start, gap_end;
2007 /* Adjust search length to account for worst case alignment overhead */
2008 length = info->length + info->align_mask;
2009 if (length < info->length)
2013 * Adjust search limits by the desired length.
2014 * See implementation comment at top of unmapped_area().
2016 gap_end = info->high_limit;
2017 if (gap_end < length)
2019 high_limit = gap_end - length;
2021 if (info->low_limit > high_limit)
2023 low_limit = info->low_limit + length;
2025 /* Check highest gap, which does not precede any rbtree node */
2026 gap_start = mm->highest_vm_end;
2027 if (gap_start <= high_limit)
2030 /* Check if rbtree root looks promising */
2031 if (RB_EMPTY_ROOT(&mm->mm_rb))
2033 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
2034 if (vma->rb_subtree_gap < length)
2038 /* Visit right subtree if it looks promising */
2039 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
2040 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
2041 struct vm_area_struct *right =
2042 rb_entry(vma->vm_rb.rb_right,
2043 struct vm_area_struct, vm_rb);
2044 if (right->rb_subtree_gap >= length) {
2051 /* Check if current node has a suitable gap */
2052 gap_end = vm_start_gap(vma);
2053 if (gap_end < low_limit)
2055 if (gap_start <= high_limit &&
2056 gap_end > gap_start && gap_end - gap_start >= length)
2059 /* Visit left subtree if it looks promising */
2060 if (vma->vm_rb.rb_left) {
2061 struct vm_area_struct *left =
2062 rb_entry(vma->vm_rb.rb_left,
2063 struct vm_area_struct, vm_rb);
2064 if (left->rb_subtree_gap >= length) {
2070 /* Go back up the rbtree to find next candidate node */
2072 struct rb_node *prev = &vma->vm_rb;
2073 if (!rb_parent(prev))
2075 vma = rb_entry(rb_parent(prev),
2076 struct vm_area_struct, vm_rb);
2077 if (prev == vma->vm_rb.rb_right) {
2078 gap_start = vma->vm_prev ?
2079 vm_end_gap(vma->vm_prev) : 0;
2086 /* We found a suitable gap. Clip it with the original high_limit. */
2087 if (gap_end > info->high_limit)
2088 gap_end = info->high_limit;
2091 /* Compute highest gap address at the desired alignment */
2092 gap_end -= info->length;
2093 gap_end -= (gap_end - info->align_offset) & info->align_mask;
2095 VM_BUG_ON(gap_end < info->low_limit);
2096 VM_BUG_ON(gap_end < gap_start);
2101 * Search for an unmapped address range.
2103 * We are looking for a range that:
2104 * - does not intersect with any VMA;
2105 * - is contained within the [low_limit, high_limit) interval;
2106 * - is at least the desired size.
2107 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
2109 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
2113 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
2114 addr = unmapped_area_topdown(info);
2116 addr = unmapped_area(info);
2118 trace_vm_unmapped_area(addr, info);
2122 #ifndef arch_get_mmap_end
2123 #define arch_get_mmap_end(addr) (TASK_SIZE)
2126 #ifndef arch_get_mmap_base
2127 #define arch_get_mmap_base(addr, base) (base)
2130 /* Get an address range which is currently unmapped.
2131 * For shmat() with addr=0.
2133 * Ugly calling convention alert:
2134 * Return value with the low bits set means error value,
2136 * if (ret & ~PAGE_MASK)
2139 * This function "knows" that -ENOMEM has the bits set.
2141 #ifndef HAVE_ARCH_UNMAPPED_AREA
2143 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2144 unsigned long len, unsigned long pgoff, unsigned long flags)
2146 struct mm_struct *mm = current->mm;
2147 struct vm_area_struct *vma, *prev;
2148 struct vm_unmapped_area_info info;
2149 const unsigned long mmap_end = arch_get_mmap_end(addr);
2151 if (len > mmap_end - mmap_min_addr)
2154 if (flags & MAP_FIXED)
2158 addr = PAGE_ALIGN(addr);
2159 vma = find_vma_prev(mm, addr, &prev);
2160 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2161 (!vma || addr + len <= vm_start_gap(vma)) &&
2162 (!prev || addr >= vm_end_gap(prev)))
2168 info.low_limit = mm->mmap_base;
2169 info.high_limit = mmap_end;
2170 info.align_mask = 0;
2171 info.align_offset = 0;
2172 return vm_unmapped_area(&info);
2177 * This mmap-allocator allocates new areas top-down from below the
2178 * stack's low limit (the base):
2180 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2182 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2183 unsigned long len, unsigned long pgoff,
2184 unsigned long flags)
2186 struct vm_area_struct *vma, *prev;
2187 struct mm_struct *mm = current->mm;
2188 struct vm_unmapped_area_info info;
2189 const unsigned long mmap_end = arch_get_mmap_end(addr);
2191 /* requested length too big for entire address space */
2192 if (len > mmap_end - mmap_min_addr)
2195 if (flags & MAP_FIXED)
2198 /* requesting a specific address */
2200 addr = PAGE_ALIGN(addr);
2201 vma = find_vma_prev(mm, addr, &prev);
2202 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2203 (!vma || addr + len <= vm_start_gap(vma)) &&
2204 (!prev || addr >= vm_end_gap(prev)))
2208 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2210 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2211 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2212 info.align_mask = 0;
2213 info.align_offset = 0;
2214 addr = vm_unmapped_area(&info);
2217 * A failed mmap() very likely causes application failure,
2218 * so fall back to the bottom-up function here. This scenario
2219 * can happen with large stack limits and large mmap()
2222 if (offset_in_page(addr)) {
2223 VM_BUG_ON(addr != -ENOMEM);
2225 info.low_limit = TASK_UNMAPPED_BASE;
2226 info.high_limit = mmap_end;
2227 addr = vm_unmapped_area(&info);
2235 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2236 unsigned long pgoff, unsigned long flags)
2238 unsigned long (*get_area)(struct file *, unsigned long,
2239 unsigned long, unsigned long, unsigned long);
2241 unsigned long error = arch_mmap_check(addr, len, flags);
2245 /* Careful about overflows.. */
2246 if (len > TASK_SIZE)
2249 get_area = current->mm->get_unmapped_area;
2251 if (file->f_op->get_unmapped_area)
2252 get_area = file->f_op->get_unmapped_area;
2253 } else if (flags & MAP_SHARED) {
2255 * mmap_region() will call shmem_zero_setup() to create a file,
2256 * so use shmem's get_unmapped_area in case it can be huge.
2257 * do_mmap() will clear pgoff, so match alignment.
2260 get_area = shmem_get_unmapped_area;
2263 addr = get_area(file, addr, len, pgoff, flags);
2264 if (IS_ERR_VALUE(addr))
2267 if (addr > TASK_SIZE - len)
2269 if (offset_in_page(addr))
2272 error = security_mmap_addr(addr);
2273 return error ? error : addr;
2276 EXPORT_SYMBOL(get_unmapped_area);
2278 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2279 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2281 struct rb_node *rb_node;
2282 struct vm_area_struct *vma;
2284 mmap_assert_locked(mm);
2285 /* Check the cache first. */
2286 vma = vmacache_find(mm, addr);
2290 rb_node = mm->mm_rb.rb_node;
2293 struct vm_area_struct *tmp;
2295 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2297 if (tmp->vm_end > addr) {
2299 if (tmp->vm_start <= addr)
2301 rb_node = rb_node->rb_left;
2303 rb_node = rb_node->rb_right;
2307 vmacache_update(addr, vma);
2311 EXPORT_SYMBOL(find_vma);
2314 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2316 struct vm_area_struct *
2317 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2318 struct vm_area_struct **pprev)
2320 struct vm_area_struct *vma;
2322 vma = find_vma(mm, addr);
2324 *pprev = vma->vm_prev;
2326 struct rb_node *rb_node = rb_last(&mm->mm_rb);
2328 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
2334 * Verify that the stack growth is acceptable and
2335 * update accounting. This is shared with both the
2336 * grow-up and grow-down cases.
2338 static int acct_stack_growth(struct vm_area_struct *vma,
2339 unsigned long size, unsigned long grow)
2341 struct mm_struct *mm = vma->vm_mm;
2342 unsigned long new_start;
2344 /* address space limit tests */
2345 if (!may_expand_vm(mm, vma->vm_flags, grow))
2348 /* Stack limit test */
2349 if (size > rlimit(RLIMIT_STACK))
2352 /* mlock limit tests */
2353 if (vma->vm_flags & VM_LOCKED) {
2354 unsigned long locked;
2355 unsigned long limit;
2356 locked = mm->locked_vm + grow;
2357 limit = rlimit(RLIMIT_MEMLOCK);
2358 limit >>= PAGE_SHIFT;
2359 if (locked > limit && !capable(CAP_IPC_LOCK))
2363 /* Check to ensure the stack will not grow into a hugetlb-only region */
2364 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2366 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2370 * Overcommit.. This must be the final test, as it will
2371 * update security statistics.
2373 if (security_vm_enough_memory_mm(mm, grow))
2379 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2381 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2382 * vma is the last one with address > vma->vm_end. Have to extend vma.
2384 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2386 struct mm_struct *mm = vma->vm_mm;
2387 struct vm_area_struct *next;
2388 unsigned long gap_addr;
2391 if (!(vma->vm_flags & VM_GROWSUP))
2394 /* Guard against exceeding limits of the address space. */
2395 address &= PAGE_MASK;
2396 if (address >= (TASK_SIZE & PAGE_MASK))
2398 address += PAGE_SIZE;
2400 /* Enforce stack_guard_gap */
2401 gap_addr = address + stack_guard_gap;
2403 /* Guard against overflow */
2404 if (gap_addr < address || gap_addr > TASK_SIZE)
2405 gap_addr = TASK_SIZE;
2407 next = vma->vm_next;
2408 if (next && next->vm_start < gap_addr && vma_is_accessible(next)) {
2409 if (!(next->vm_flags & VM_GROWSUP))
2411 /* Check that both stack segments have the same anon_vma? */
2414 /* We must make sure the anon_vma is allocated. */
2415 if (unlikely(anon_vma_prepare(vma)))
2419 * vma->vm_start/vm_end cannot change under us because the caller
2420 * is required to hold the mmap_lock in read mode. We need the
2421 * anon_vma lock to serialize against concurrent expand_stacks.
2423 anon_vma_lock_write(vma->anon_vma);
2425 /* Somebody else might have raced and expanded it already */
2426 if (address > vma->vm_end) {
2427 unsigned long size, grow;
2429 size = address - vma->vm_start;
2430 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2433 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2434 error = acct_stack_growth(vma, size, grow);
2437 * vma_gap_update() doesn't support concurrent
2438 * updates, but we only hold a shared mmap_lock
2439 * lock here, so we need to protect against
2440 * concurrent vma expansions.
2441 * anon_vma_lock_write() doesn't help here, as
2442 * we don't guarantee that all growable vmas
2443 * in a mm share the same root anon vma.
2444 * So, we reuse mm->page_table_lock to guard
2445 * against concurrent vma expansions.
2447 spin_lock(&mm->page_table_lock);
2448 if (vma->vm_flags & VM_LOCKED)
2449 mm->locked_vm += grow;
2450 vm_stat_account(mm, vma->vm_flags, grow);
2451 anon_vma_interval_tree_pre_update_vma(vma);
2452 vma->vm_end = address;
2453 anon_vma_interval_tree_post_update_vma(vma);
2455 vma_gap_update(vma->vm_next);
2457 mm->highest_vm_end = vm_end_gap(vma);
2458 spin_unlock(&mm->page_table_lock);
2460 perf_event_mmap(vma);
2464 anon_vma_unlock_write(vma->anon_vma);
2465 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2469 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2472 * vma is the first one with address < vma->vm_start. Have to extend vma.
2474 int expand_downwards(struct vm_area_struct *vma,
2475 unsigned long address)
2477 struct mm_struct *mm = vma->vm_mm;
2478 struct vm_area_struct *prev;
2481 address &= PAGE_MASK;
2482 if (address < mmap_min_addr)
2485 /* Enforce stack_guard_gap */
2486 prev = vma->vm_prev;
2487 /* Check that both stack segments have the same anon_vma? */
2488 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2489 vma_is_accessible(prev)) {
2490 if (address - prev->vm_end < stack_guard_gap)
2494 /* We must make sure the anon_vma is allocated. */
2495 if (unlikely(anon_vma_prepare(vma)))
2499 * vma->vm_start/vm_end cannot change under us because the caller
2500 * is required to hold the mmap_lock in read mode. We need the
2501 * anon_vma lock to serialize against concurrent expand_stacks.
2503 anon_vma_lock_write(vma->anon_vma);
2505 /* Somebody else might have raced and expanded it already */
2506 if (address < vma->vm_start) {
2507 unsigned long size, grow;
2509 size = vma->vm_end - address;
2510 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2513 if (grow <= vma->vm_pgoff) {
2514 error = acct_stack_growth(vma, size, grow);
2517 * vma_gap_update() doesn't support concurrent
2518 * updates, but we only hold a shared mmap_lock
2519 * lock here, so we need to protect against
2520 * concurrent vma expansions.
2521 * anon_vma_lock_write() doesn't help here, as
2522 * we don't guarantee that all growable vmas
2523 * in a mm share the same root anon vma.
2524 * So, we reuse mm->page_table_lock to guard
2525 * against concurrent vma expansions.
2527 spin_lock(&mm->page_table_lock);
2528 if (vma->vm_flags & VM_LOCKED)
2529 mm->locked_vm += grow;
2530 vm_stat_account(mm, vma->vm_flags, grow);
2531 anon_vma_interval_tree_pre_update_vma(vma);
2532 vma->vm_start = address;
2533 vma->vm_pgoff -= grow;
2534 anon_vma_interval_tree_post_update_vma(vma);
2535 vma_gap_update(vma);
2536 spin_unlock(&mm->page_table_lock);
2538 perf_event_mmap(vma);
2542 anon_vma_unlock_write(vma->anon_vma);
2543 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2548 /* enforced gap between the expanding stack and other mappings. */
2549 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2551 static int __init cmdline_parse_stack_guard_gap(char *p)
2556 val = simple_strtoul(p, &endptr, 10);
2558 stack_guard_gap = val << PAGE_SHIFT;
2562 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2564 #ifdef CONFIG_STACK_GROWSUP
2565 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2567 return expand_upwards(vma, address);
2570 struct vm_area_struct *
2571 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2573 struct vm_area_struct *vma, *prev;
2576 vma = find_vma_prev(mm, addr, &prev);
2577 if (vma && (vma->vm_start <= addr))
2579 /* don't alter vm_end if the coredump is running */
2580 if (!prev || expand_stack(prev, addr))
2582 if (prev->vm_flags & VM_LOCKED)
2583 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2587 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2589 return expand_downwards(vma, address);
2592 struct vm_area_struct *
2593 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2595 struct vm_area_struct *vma;
2596 unsigned long start;
2599 vma = find_vma(mm, addr);
2602 if (vma->vm_start <= addr)
2604 if (!(vma->vm_flags & VM_GROWSDOWN))
2606 start = vma->vm_start;
2607 if (expand_stack(vma, addr))
2609 if (vma->vm_flags & VM_LOCKED)
2610 populate_vma_page_range(vma, addr, start, NULL);
2615 EXPORT_SYMBOL_GPL(find_extend_vma);
2618 * Ok - we have the memory areas we should free on the vma list,
2619 * so release them, and do the vma updates.
2621 * Called with the mm semaphore held.
2623 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2625 unsigned long nr_accounted = 0;
2627 /* Update high watermark before we lower total_vm */
2628 update_hiwater_vm(mm);
2630 long nrpages = vma_pages(vma);
2632 if (vma->vm_flags & VM_ACCOUNT)
2633 nr_accounted += nrpages;
2634 vm_stat_account(mm, vma->vm_flags, -nrpages);
2635 vma = remove_vma(vma);
2637 vm_unacct_memory(nr_accounted);
2642 * Get rid of page table information in the indicated region.
2644 * Called with the mm semaphore held.
2646 static void unmap_region(struct mm_struct *mm,
2647 struct vm_area_struct *vma, struct vm_area_struct *prev,
2648 unsigned long start, unsigned long end)
2650 struct vm_area_struct *next = vma_next(mm, prev);
2651 struct mmu_gather tlb;
2654 tlb_gather_mmu(&tlb, mm);
2655 update_hiwater_rss(mm);
2656 unmap_vmas(&tlb, vma, start, end);
2657 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2658 next ? next->vm_start : USER_PGTABLES_CEILING);
2659 tlb_finish_mmu(&tlb);
2663 * Create a list of vma's touched by the unmap, removing them from the mm's
2664 * vma list as we go..
2667 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2668 struct vm_area_struct *prev, unsigned long end)
2670 struct vm_area_struct **insertion_point;
2671 struct vm_area_struct *tail_vma = NULL;
2673 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2674 vma->vm_prev = NULL;
2676 vma_rb_erase(vma, &mm->mm_rb);
2680 } while (vma && vma->vm_start < end);
2681 *insertion_point = vma;
2683 vma->vm_prev = prev;
2684 vma_gap_update(vma);
2686 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2687 tail_vma->vm_next = NULL;
2689 /* Kill the cache */
2690 vmacache_invalidate(mm);
2693 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2694 * VM_GROWSUP VMA. Such VMAs can change their size under
2695 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2697 if (vma && (vma->vm_flags & VM_GROWSDOWN))
2699 if (prev && (prev->vm_flags & VM_GROWSUP))
2705 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2706 * has already been checked or doesn't make sense to fail.
2708 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2709 unsigned long addr, int new_below)
2711 struct vm_area_struct *new;
2714 if (vma->vm_ops && vma->vm_ops->may_split) {
2715 err = vma->vm_ops->may_split(vma, addr);
2720 new = vm_area_dup(vma);
2727 new->vm_start = addr;
2728 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2731 err = vma_dup_policy(vma, new);
2735 err = anon_vma_clone(new, vma);
2740 get_file(new->vm_file);
2742 if (new->vm_ops && new->vm_ops->open)
2743 new->vm_ops->open(new);
2746 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2747 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2749 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2755 /* Clean everything up if vma_adjust failed. */
2756 if (new->vm_ops && new->vm_ops->close)
2757 new->vm_ops->close(new);
2760 unlink_anon_vmas(new);
2762 mpol_put(vma_policy(new));
2769 * Split a vma into two pieces at address 'addr', a new vma is allocated
2770 * either for the first part or the tail.
2772 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2773 unsigned long addr, int new_below)
2775 if (mm->map_count >= sysctl_max_map_count)
2778 return __split_vma(mm, vma, addr, new_below);
2782 unlock_range(struct vm_area_struct *start, unsigned long limit)
2784 struct mm_struct *mm = start->vm_mm;
2785 struct vm_area_struct *tmp = start;
2787 while (tmp && tmp->vm_start < limit) {
2788 if (tmp->vm_flags & VM_LOCKED) {
2789 mm->locked_vm -= vma_pages(tmp);
2790 munlock_vma_pages_all(tmp);
2797 /* Munmap is split into 2 main parts -- this part which finds
2798 * what needs doing, and the areas themselves, which do the
2799 * work. This now handles partial unmappings.
2800 * Jeremy Fitzhardinge <jeremy@goop.org>
2802 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2803 struct list_head *uf, bool downgrade)
2806 struct vm_area_struct *vma, *prev, *last;
2808 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2811 len = PAGE_ALIGN(len);
2817 * arch_unmap() might do unmaps itself. It must be called
2818 * and finish any rbtree manipulation before this code
2819 * runs and also starts to manipulate the rbtree.
2821 arch_unmap(mm, start, end);
2823 /* Find the first overlapping VMA where start < vma->vm_end */
2824 vma = find_vma_intersection(mm, start, end);
2827 prev = vma->vm_prev;
2830 * If we need to split any vma, do it now to save pain later.
2832 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2833 * unmapped vm_area_struct will remain in use: so lower split_vma
2834 * places tmp vma above, and higher split_vma places tmp vma below.
2836 if (start > vma->vm_start) {
2840 * Make sure that map_count on return from munmap() will
2841 * not exceed its limit; but let map_count go just above
2842 * its limit temporarily, to help free resources as expected.
2844 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2847 error = __split_vma(mm, vma, start, 0);
2853 /* Does it split the last one? */
2854 last = find_vma(mm, end);
2855 if (last && end > last->vm_start) {
2856 int error = __split_vma(mm, last, end, 1);
2860 vma = vma_next(mm, prev);
2864 * If userfaultfd_unmap_prep returns an error the vmas
2865 * will remain split, but userland will get a
2866 * highly unexpected error anyway. This is no
2867 * different than the case where the first of the two
2868 * __split_vma fails, but we don't undo the first
2869 * split, despite we could. This is unlikely enough
2870 * failure that it's not worth optimizing it for.
2872 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2878 * unlock any mlock()ed ranges before detaching vmas
2881 unlock_range(vma, end);
2883 /* Detach vmas from rbtree */
2884 if (!detach_vmas_to_be_unmapped(mm, vma, prev, end))
2888 mmap_write_downgrade(mm);
2890 unmap_region(mm, vma, prev, start, end);
2892 /* Fix up all other VM information */
2893 remove_vma_list(mm, vma);
2895 return downgrade ? 1 : 0;
2898 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2899 struct list_head *uf)
2901 return __do_munmap(mm, start, len, uf, false);
2904 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2907 struct mm_struct *mm = current->mm;
2910 if (mmap_write_lock_killable(mm))
2913 ret = __do_munmap(mm, start, len, &uf, downgrade);
2915 * Returning 1 indicates mmap_lock is downgraded.
2916 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2917 * it to 0 before return.
2920 mmap_read_unlock(mm);
2923 mmap_write_unlock(mm);
2925 userfaultfd_unmap_complete(mm, &uf);
2929 int vm_munmap(unsigned long start, size_t len)
2931 return __vm_munmap(start, len, false);
2933 EXPORT_SYMBOL(vm_munmap);
2935 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2937 addr = untagged_addr(addr);
2938 return __vm_munmap(addr, len, true);
2943 * Emulation of deprecated remap_file_pages() syscall.
2945 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2946 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2949 struct mm_struct *mm = current->mm;
2950 struct vm_area_struct *vma;
2951 unsigned long populate = 0;
2952 unsigned long ret = -EINVAL;
2955 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2956 current->comm, current->pid);
2960 start = start & PAGE_MASK;
2961 size = size & PAGE_MASK;
2963 if (start + size <= start)
2966 /* Does pgoff wrap? */
2967 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2970 if (mmap_write_lock_killable(mm))
2973 vma = vma_lookup(mm, start);
2975 if (!vma || !(vma->vm_flags & VM_SHARED))
2978 if (start + size > vma->vm_end) {
2979 struct vm_area_struct *next;
2981 for (next = vma->vm_next; next; next = next->vm_next) {
2982 /* hole between vmas ? */
2983 if (next->vm_start != next->vm_prev->vm_end)
2986 if (next->vm_file != vma->vm_file)
2989 if (next->vm_flags != vma->vm_flags)
2992 if (start + size <= next->vm_end)
3000 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3001 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3002 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3004 flags &= MAP_NONBLOCK;
3005 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3006 if (vma->vm_flags & VM_LOCKED)
3007 flags |= MAP_LOCKED;
3009 file = get_file(vma->vm_file);
3010 ret = do_mmap(vma->vm_file, start, size,
3011 prot, flags, pgoff, &populate, NULL);
3014 mmap_write_unlock(mm);
3016 mm_populate(ret, populate);
3017 if (!IS_ERR_VALUE(ret))
3023 * this is really a simplified "do_mmap". it only handles
3024 * anonymous maps. eventually we may be able to do some
3025 * brk-specific accounting here.
3027 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
3029 struct mm_struct *mm = current->mm;
3030 struct vm_area_struct *vma, *prev;
3031 struct rb_node **rb_link, *rb_parent;
3032 pgoff_t pgoff = addr >> PAGE_SHIFT;
3034 unsigned long mapped_addr;
3036 /* Until we need other flags, refuse anything except VM_EXEC. */
3037 if ((flags & (~VM_EXEC)) != 0)
3039 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3041 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
3042 if (IS_ERR_VALUE(mapped_addr))
3045 error = mlock_future_check(mm, mm->def_flags, len);
3049 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
3050 if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
3053 /* Check against address space limits *after* clearing old maps... */
3054 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3057 if (mm->map_count > sysctl_max_map_count)
3060 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3063 /* Can we just expand an old private anonymous mapping? */
3064 vma = vma_merge(mm, prev, addr, addr + len, flags,
3065 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
3070 * create a vma struct for an anonymous mapping
3072 vma = vm_area_alloc(mm);
3074 vm_unacct_memory(len >> PAGE_SHIFT);
3078 vma_set_anonymous(vma);
3079 vma->vm_start = addr;
3080 vma->vm_end = addr + len;
3081 vma->vm_pgoff = pgoff;
3082 vma->vm_flags = flags;
3083 vma->vm_page_prot = vm_get_page_prot(flags);
3084 vma_link(mm, vma, prev, rb_link, rb_parent);
3086 perf_event_mmap(vma);
3087 mm->total_vm += len >> PAGE_SHIFT;
3088 mm->data_vm += len >> PAGE_SHIFT;
3089 if (flags & VM_LOCKED)
3090 mm->locked_vm += (len >> PAGE_SHIFT);
3091 vma->vm_flags |= VM_SOFTDIRTY;
3095 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3097 struct mm_struct *mm = current->mm;
3103 len = PAGE_ALIGN(request);
3109 if (mmap_write_lock_killable(mm))
3112 ret = do_brk_flags(addr, len, flags, &uf);
3113 populate = ((mm->def_flags & VM_LOCKED) != 0);
3114 mmap_write_unlock(mm);
3115 userfaultfd_unmap_complete(mm, &uf);
3116 if (populate && !ret)
3117 mm_populate(addr, len);
3120 EXPORT_SYMBOL(vm_brk_flags);
3122 int vm_brk(unsigned long addr, unsigned long len)
3124 return vm_brk_flags(addr, len, 0);
3126 EXPORT_SYMBOL(vm_brk);
3128 /* Release all mmaps. */
3129 void exit_mmap(struct mm_struct *mm)
3131 struct mmu_gather tlb;
3132 struct vm_area_struct *vma;
3133 unsigned long nr_accounted = 0;
3135 /* mm's last user has gone, and its about to be pulled down */
3136 mmu_notifier_release(mm);
3138 if (unlikely(mm_is_oom_victim(mm))) {
3140 * Manually reap the mm to free as much memory as possible.
3141 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3142 * this mm from further consideration. Taking mm->mmap_lock for
3143 * write after setting MMF_OOM_SKIP will guarantee that the oom
3144 * reaper will not run on this mm again after mmap_lock is
3147 * Nothing can be holding mm->mmap_lock here and the above call
3148 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3149 * __oom_reap_task_mm() will not block.
3151 * This needs to be done before calling unlock_range(),
3152 * which clears VM_LOCKED, otherwise the oom reaper cannot
3155 (void)__oom_reap_task_mm(mm);
3157 set_bit(MMF_OOM_SKIP, &mm->flags);
3160 mmap_write_lock(mm);
3162 unlock_range(mm->mmap, ULONG_MAX);
3168 /* Can happen if dup_mmap() received an OOM */
3169 mmap_write_unlock(mm);
3175 tlb_gather_mmu_fullmm(&tlb, mm);
3176 /* update_hiwater_rss(mm) here? but nobody should be looking */
3177 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3178 unmap_vmas(&tlb, vma, 0, -1);
3179 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3180 tlb_finish_mmu(&tlb);
3182 /* Walk the list again, actually closing and freeing it. */
3184 if (vma->vm_flags & VM_ACCOUNT)
3185 nr_accounted += vma_pages(vma);
3186 vma = remove_vma(vma);
3189 mmap_write_unlock(mm);
3190 vm_unacct_memory(nr_accounted);
3193 /* Insert vm structure into process list sorted by address
3194 * and into the inode's i_mmap tree. If vm_file is non-NULL
3195 * then i_mmap_rwsem is taken here.
3197 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3199 struct vm_area_struct *prev;
3200 struct rb_node **rb_link, *rb_parent;
3202 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3203 &prev, &rb_link, &rb_parent))
3205 if ((vma->vm_flags & VM_ACCOUNT) &&
3206 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3210 * The vm_pgoff of a purely anonymous vma should be irrelevant
3211 * until its first write fault, when page's anon_vma and index
3212 * are set. But now set the vm_pgoff it will almost certainly
3213 * end up with (unless mremap moves it elsewhere before that
3214 * first wfault), so /proc/pid/maps tells a consistent story.
3216 * By setting it to reflect the virtual start address of the
3217 * vma, merges and splits can happen in a seamless way, just
3218 * using the existing file pgoff checks and manipulations.
3219 * Similarly in do_mmap and in do_brk_flags.
3221 if (vma_is_anonymous(vma)) {
3222 BUG_ON(vma->anon_vma);
3223 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3226 vma_link(mm, vma, prev, rb_link, rb_parent);
3231 * Copy the vma structure to a new location in the same mm,
3232 * prior to moving page table entries, to effect an mremap move.
3234 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3235 unsigned long addr, unsigned long len, pgoff_t pgoff,
3236 bool *need_rmap_locks)
3238 struct vm_area_struct *vma = *vmap;
3239 unsigned long vma_start = vma->vm_start;
3240 struct mm_struct *mm = vma->vm_mm;
3241 struct vm_area_struct *new_vma, *prev;
3242 struct rb_node **rb_link, *rb_parent;
3243 bool faulted_in_anon_vma = true;
3246 * If anonymous vma has not yet been faulted, update new pgoff
3247 * to match new location, to increase its chance of merging.
3249 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3250 pgoff = addr >> PAGE_SHIFT;
3251 faulted_in_anon_vma = false;
3254 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3255 return NULL; /* should never get here */
3256 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3257 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3258 vma->vm_userfaultfd_ctx, vma_anon_name(vma));
3261 * Source vma may have been merged into new_vma
3263 if (unlikely(vma_start >= new_vma->vm_start &&
3264 vma_start < new_vma->vm_end)) {
3266 * The only way we can get a vma_merge with
3267 * self during an mremap is if the vma hasn't
3268 * been faulted in yet and we were allowed to
3269 * reset the dst vma->vm_pgoff to the
3270 * destination address of the mremap to allow
3271 * the merge to happen. mremap must change the
3272 * vm_pgoff linearity between src and dst vmas
3273 * (in turn preventing a vma_merge) to be
3274 * safe. It is only safe to keep the vm_pgoff
3275 * linear if there are no pages mapped yet.
3277 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3278 *vmap = vma = new_vma;
3280 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3282 new_vma = vm_area_dup(vma);
3285 new_vma->vm_start = addr;
3286 new_vma->vm_end = addr + len;
3287 new_vma->vm_pgoff = pgoff;
3288 if (vma_dup_policy(vma, new_vma))
3290 if (anon_vma_clone(new_vma, vma))
3291 goto out_free_mempol;
3292 if (new_vma->vm_file)
3293 get_file(new_vma->vm_file);
3294 if (new_vma->vm_ops && new_vma->vm_ops->open)
3295 new_vma->vm_ops->open(new_vma);
3296 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3297 *need_rmap_locks = false;
3302 mpol_put(vma_policy(new_vma));
3304 vm_area_free(new_vma);
3310 * Return true if the calling process may expand its vm space by the passed
3313 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3315 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3318 if (is_data_mapping(flags) &&
3319 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3320 /* Workaround for Valgrind */
3321 if (rlimit(RLIMIT_DATA) == 0 &&
3322 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3325 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3326 current->comm, current->pid,
3327 (mm->data_vm + npages) << PAGE_SHIFT,
3328 rlimit(RLIMIT_DATA),
3329 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3331 if (!ignore_rlimit_data)
3338 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3340 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3342 if (is_exec_mapping(flags))
3343 mm->exec_vm += npages;
3344 else if (is_stack_mapping(flags))
3345 mm->stack_vm += npages;
3346 else if (is_data_mapping(flags))
3347 mm->data_vm += npages;
3350 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3353 * Having a close hook prevents vma merging regardless of flags.
3355 static void special_mapping_close(struct vm_area_struct *vma)
3359 static const char *special_mapping_name(struct vm_area_struct *vma)
3361 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3364 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3366 struct vm_special_mapping *sm = new_vma->vm_private_data;
3368 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3372 return sm->mremap(sm, new_vma);
3377 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3380 * Forbid splitting special mappings - kernel has expectations over
3381 * the number of pages in mapping. Together with VM_DONTEXPAND
3382 * the size of vma should stay the same over the special mapping's
3388 static const struct vm_operations_struct special_mapping_vmops = {
3389 .close = special_mapping_close,
3390 .fault = special_mapping_fault,
3391 .mremap = special_mapping_mremap,
3392 .name = special_mapping_name,
3393 /* vDSO code relies that VVAR can't be accessed remotely */
3395 .may_split = special_mapping_split,
3398 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3399 .close = special_mapping_close,
3400 .fault = special_mapping_fault,
3403 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3405 struct vm_area_struct *vma = vmf->vma;
3407 struct page **pages;
3409 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3410 pages = vma->vm_private_data;
3412 struct vm_special_mapping *sm = vma->vm_private_data;
3415 return sm->fault(sm, vmf->vma, vmf);
3420 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3424 struct page *page = *pages;
3430 return VM_FAULT_SIGBUS;
3433 static struct vm_area_struct *__install_special_mapping(
3434 struct mm_struct *mm,
3435 unsigned long addr, unsigned long len,
3436 unsigned long vm_flags, void *priv,
3437 const struct vm_operations_struct *ops)
3440 struct vm_area_struct *vma;
3442 vma = vm_area_alloc(mm);
3443 if (unlikely(vma == NULL))
3444 return ERR_PTR(-ENOMEM);
3446 vma->vm_start = addr;
3447 vma->vm_end = addr + len;
3449 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3450 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3453 vma->vm_private_data = priv;
3455 ret = insert_vm_struct(mm, vma);
3459 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3461 perf_event_mmap(vma);
3467 return ERR_PTR(ret);
3470 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3471 const struct vm_special_mapping *sm)
3473 return vma->vm_private_data == sm &&
3474 (vma->vm_ops == &special_mapping_vmops ||
3475 vma->vm_ops == &legacy_special_mapping_vmops);
3479 * Called with mm->mmap_lock held for writing.
3480 * Insert a new vma covering the given region, with the given flags.
3481 * Its pages are supplied by the given array of struct page *.
3482 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3483 * The region past the last page supplied will always produce SIGBUS.
3484 * The array pointer and the pages it points to are assumed to stay alive
3485 * for as long as this mapping might exist.
3487 struct vm_area_struct *_install_special_mapping(
3488 struct mm_struct *mm,
3489 unsigned long addr, unsigned long len,
3490 unsigned long vm_flags, const struct vm_special_mapping *spec)
3492 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3493 &special_mapping_vmops);
3496 int install_special_mapping(struct mm_struct *mm,
3497 unsigned long addr, unsigned long len,
3498 unsigned long vm_flags, struct page **pages)
3500 struct vm_area_struct *vma = __install_special_mapping(
3501 mm, addr, len, vm_flags, (void *)pages,
3502 &legacy_special_mapping_vmops);
3504 return PTR_ERR_OR_ZERO(vma);
3507 static DEFINE_MUTEX(mm_all_locks_mutex);
3509 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3511 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3513 * The LSB of head.next can't change from under us
3514 * because we hold the mm_all_locks_mutex.
3516 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3518 * We can safely modify head.next after taking the
3519 * anon_vma->root->rwsem. If some other vma in this mm shares
3520 * the same anon_vma we won't take it again.
3522 * No need of atomic instructions here, head.next
3523 * can't change from under us thanks to the
3524 * anon_vma->root->rwsem.
3526 if (__test_and_set_bit(0, (unsigned long *)
3527 &anon_vma->root->rb_root.rb_root.rb_node))
3532 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3534 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3536 * AS_MM_ALL_LOCKS can't change from under us because
3537 * we hold the mm_all_locks_mutex.
3539 * Operations on ->flags have to be atomic because
3540 * even if AS_MM_ALL_LOCKS is stable thanks to the
3541 * mm_all_locks_mutex, there may be other cpus
3542 * changing other bitflags in parallel to us.
3544 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3546 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3551 * This operation locks against the VM for all pte/vma/mm related
3552 * operations that could ever happen on a certain mm. This includes
3553 * vmtruncate, try_to_unmap, and all page faults.
3555 * The caller must take the mmap_lock in write mode before calling
3556 * mm_take_all_locks(). The caller isn't allowed to release the
3557 * mmap_lock until mm_drop_all_locks() returns.
3559 * mmap_lock in write mode is required in order to block all operations
3560 * that could modify pagetables and free pages without need of
3561 * altering the vma layout. It's also needed in write mode to avoid new
3562 * anon_vmas to be associated with existing vmas.
3564 * A single task can't take more than one mm_take_all_locks() in a row
3565 * or it would deadlock.
3567 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3568 * mapping->flags avoid to take the same lock twice, if more than one
3569 * vma in this mm is backed by the same anon_vma or address_space.
3571 * We take locks in following order, accordingly to comment at beginning
3573 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3575 * - all i_mmap_rwsem locks;
3576 * - all anon_vma->rwseml
3578 * We can take all locks within these types randomly because the VM code
3579 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3580 * mm_all_locks_mutex.
3582 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3583 * that may have to take thousand of locks.
3585 * mm_take_all_locks() can fail if it's interrupted by signals.
3587 int mm_take_all_locks(struct mm_struct *mm)
3589 struct vm_area_struct *vma;
3590 struct anon_vma_chain *avc;
3592 BUG_ON(mmap_read_trylock(mm));
3594 mutex_lock(&mm_all_locks_mutex);
3596 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3597 if (signal_pending(current))
3599 if (vma->vm_file && vma->vm_file->f_mapping &&
3600 is_vm_hugetlb_page(vma))
3601 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3604 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3605 if (signal_pending(current))
3607 if (vma->vm_file && vma->vm_file->f_mapping &&
3608 !is_vm_hugetlb_page(vma))
3609 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3612 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3613 if (signal_pending(current))
3616 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3617 vm_lock_anon_vma(mm, avc->anon_vma);
3623 mm_drop_all_locks(mm);
3627 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3629 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3631 * The LSB of head.next can't change to 0 from under
3632 * us because we hold the mm_all_locks_mutex.
3634 * We must however clear the bitflag before unlocking
3635 * the vma so the users using the anon_vma->rb_root will
3636 * never see our bitflag.
3638 * No need of atomic instructions here, head.next
3639 * can't change from under us until we release the
3640 * anon_vma->root->rwsem.
3642 if (!__test_and_clear_bit(0, (unsigned long *)
3643 &anon_vma->root->rb_root.rb_root.rb_node))
3645 anon_vma_unlock_write(anon_vma);
3649 static void vm_unlock_mapping(struct address_space *mapping)
3651 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3653 * AS_MM_ALL_LOCKS can't change to 0 from under us
3654 * because we hold the mm_all_locks_mutex.
3656 i_mmap_unlock_write(mapping);
3657 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3664 * The mmap_lock cannot be released by the caller until
3665 * mm_drop_all_locks() returns.
3667 void mm_drop_all_locks(struct mm_struct *mm)
3669 struct vm_area_struct *vma;
3670 struct anon_vma_chain *avc;
3672 BUG_ON(mmap_read_trylock(mm));
3673 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3675 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3677 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3678 vm_unlock_anon_vma(avc->anon_vma);
3679 if (vma->vm_file && vma->vm_file->f_mapping)
3680 vm_unlock_mapping(vma->vm_file->f_mapping);
3683 mutex_unlock(&mm_all_locks_mutex);
3687 * initialise the percpu counter for VM
3689 void __init mmap_init(void)
3693 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3698 * Initialise sysctl_user_reserve_kbytes.
3700 * This is intended to prevent a user from starting a single memory hogging
3701 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3704 * The default value is min(3% of free memory, 128MB)
3705 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3707 static int init_user_reserve(void)
3709 unsigned long free_kbytes;
3711 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3713 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3716 subsys_initcall(init_user_reserve);
3719 * Initialise sysctl_admin_reserve_kbytes.
3721 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3722 * to log in and kill a memory hogging process.
3724 * Systems with more than 256MB will reserve 8MB, enough to recover
3725 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3726 * only reserve 3% of free pages by default.
3728 static int init_admin_reserve(void)
3730 unsigned long free_kbytes;
3732 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3734 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3737 subsys_initcall(init_admin_reserve);
3740 * Reinititalise user and admin reserves if memory is added or removed.
3742 * The default user reserve max is 128MB, and the default max for the
3743 * admin reserve is 8MB. These are usually, but not always, enough to
3744 * enable recovery from a memory hogging process using login/sshd, a shell,
3745 * and tools like top. It may make sense to increase or even disable the
3746 * reserve depending on the existence of swap or variations in the recovery
3747 * tools. So, the admin may have changed them.
3749 * If memory is added and the reserves have been eliminated or increased above
3750 * the default max, then we'll trust the admin.
3752 * If memory is removed and there isn't enough free memory, then we
3753 * need to reset the reserves.
3755 * Otherwise keep the reserve set by the admin.
3757 static int reserve_mem_notifier(struct notifier_block *nb,
3758 unsigned long action, void *data)
3760 unsigned long tmp, free_kbytes;
3764 /* Default max is 128MB. Leave alone if modified by operator. */
3765 tmp = sysctl_user_reserve_kbytes;
3766 if (0 < tmp && tmp < (1UL << 17))
3767 init_user_reserve();
3769 /* Default max is 8MB. Leave alone if modified by operator. */
3770 tmp = sysctl_admin_reserve_kbytes;
3771 if (0 < tmp && tmp < (1UL << 13))
3772 init_admin_reserve();
3776 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3778 if (sysctl_user_reserve_kbytes > free_kbytes) {
3779 init_user_reserve();
3780 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3781 sysctl_user_reserve_kbytes);
3784 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3785 init_admin_reserve();
3786 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3787 sysctl_admin_reserve_kbytes);
3796 static struct notifier_block reserve_mem_nb = {
3797 .notifier_call = reserve_mem_notifier,
3800 static int __meminit init_reserve_notifier(void)
3802 if (register_hotmemory_notifier(&reserve_mem_nb))
3803 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3807 subsys_initcall(init_reserve_notifier);
projects / linux-2.6-microblaze.git / blob