1 // SPDX-License-Identifier: GPL-2.0-only
7 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12 #include <linux/kernel.h>
13 #include <linux/slab.h>
14 #include <linux/backing-dev.h>
16 #include <linux/vmacache.h>
17 #include <linux/shm.h>
18 #include <linux/mman.h>
19 #include <linux/pagemap.h>
20 #include <linux/swap.h>
21 #include <linux/syscalls.h>
22 #include <linux/capability.h>
23 #include <linux/init.h>
24 #include <linux/file.h>
26 #include <linux/personality.h>
27 #include <linux/security.h>
28 #include <linux/hugetlb.h>
29 #include <linux/shmem_fs.h>
30 #include <linux/profile.h>
31 #include <linux/export.h>
32 #include <linux/mount.h>
33 #include <linux/mempolicy.h>
34 #include <linux/rmap.h>
35 #include <linux/mmu_notifier.h>
36 #include <linux/mmdebug.h>
37 #include <linux/perf_event.h>
38 #include <linux/audit.h>
39 #include <linux/khugepaged.h>
40 #include <linux/uprobes.h>
41 #include <linux/rbtree_augmented.h>
42 #include <linux/notifier.h>
43 #include <linux/memory.h>
44 #include <linux/printk.h>
45 #include <linux/userfaultfd_k.h>
46 #include <linux/moduleparam.h>
47 #include <linux/pkeys.h>
48 #include <linux/oom.h>
49 #include <linux/sched/mm.h>
51 #include <linux/uaccess.h>
52 #include <asm/cacheflush.h>
54 #include <asm/mmu_context.h>
55 #ifdef CONFIG_FINEGRAINED_THP
56 #include <asm/finegrained_thp.h>
58 #include <asm-generic/finegrained_thp.h>
61 #define CREATE_TRACE_POINTS
62 #include <trace/events/mmap.h>
66 #ifndef arch_mmap_check
67 #define arch_mmap_check(addr, len, flags) (0)
70 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
71 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
72 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
73 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
75 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
76 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
77 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
78 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
81 static bool ignore_rlimit_data;
82 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
84 static void unmap_region(struct mm_struct *mm,
85 struct vm_area_struct *vma, struct vm_area_struct *prev,
86 unsigned long start, unsigned long end);
88 /* description of effects of mapping type and prot in current implementation.
89 * this is due to the limited x86 page protection hardware. The expected
90 * behavior is in parens:
93 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
94 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
95 * w: (no) no w: (no) no w: (yes) yes w: (no) no
96 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
98 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
99 * w: (no) no w: (no) no w: (copy) copy w: (no) no
100 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
102 pgprot_t protection_map[16] __ro_after_init = {
103 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
104 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
107 #ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
108 static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
114 pgprot_t vm_get_page_prot(unsigned long vm_flags)
116 pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
117 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
118 pgprot_val(arch_vm_get_page_prot(vm_flags)));
120 return arch_filter_pgprot(ret);
122 EXPORT_SYMBOL(vm_get_page_prot);
124 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
126 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
129 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
130 void vma_set_page_prot(struct vm_area_struct *vma)
132 unsigned long vm_flags = vma->vm_flags;
133 pgprot_t vm_page_prot;
135 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
136 if (vma_wants_writenotify(vma, vm_page_prot)) {
137 vm_flags &= ~VM_SHARED;
138 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
140 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
141 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
145 * Requires inode->i_mapping->i_mmap_rwsem
147 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
148 struct file *file, struct address_space *mapping)
150 if (vma->vm_flags & VM_DENYWRITE)
151 allow_write_access(file);
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 retval;
198 unsigned long newbrk, oldbrk, origbrk;
199 struct mm_struct *mm = current->mm;
200 struct vm_area_struct *next;
201 unsigned long min_brk;
203 bool downgraded = false;
206 if (mmap_write_lock_killable(mm))
211 #ifdef CONFIG_COMPAT_BRK
213 * CONFIG_COMPAT_BRK can still be overridden by setting
214 * randomize_va_space to 2, which will still cause mm->start_brk
215 * to be arbitrarily shifted
217 if (current->brk_randomized)
218 min_brk = mm->start_brk;
220 min_brk = mm->end_data;
222 min_brk = mm->start_brk;
228 * Check against rlimit here. If this check is done later after the test
229 * of oldbrk with newbrk then it can escape the test and let the data
230 * segment grow beyond its set limit the in case where the limit is
231 * not page aligned -Ram Gupta
233 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
234 mm->end_data, mm->start_data))
237 newbrk = PAGE_ALIGN(brk);
238 oldbrk = PAGE_ALIGN(mm->brk);
239 if (oldbrk == newbrk) {
245 * Always allow shrinking brk.
246 * __do_munmap() may downgrade mmap_lock to read.
248 if (brk <= mm->brk) {
252 * mm->brk must to be protected by write mmap_lock so update it
253 * before downgrading mmap_lock. When __do_munmap() fails,
254 * mm->brk will be restored from origbrk.
257 ret = __do_munmap(mm, newbrk, oldbrk-newbrk, &uf, true);
261 } else if (ret == 1) {
267 /* Check against existing mmap mappings. */
268 next = find_vma(mm, oldbrk);
269 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
272 /* Ok, looks good - let it rip. */
273 if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
278 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
280 khugepaged_mem_hook(mm, origbrk, newbrk - oldbrk, __func__);
282 mmap_read_unlock(mm);
284 mmap_write_unlock(mm);
285 userfaultfd_unmap_complete(mm, &uf);
287 mm_populate(oldbrk, newbrk - oldbrk);
292 mmap_write_unlock(mm);
296 static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
298 unsigned long gap, prev_end;
301 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
302 * allow two stack_guard_gaps between them here, and when choosing
303 * an unmapped area; whereas when expanding we only require one.
304 * That's a little inconsistent, but keeps the code here simpler.
306 gap = vm_start_gap(vma);
308 prev_end = vm_end_gap(vma->vm_prev);
317 #ifdef CONFIG_DEBUG_VM_RB
318 static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma)
320 unsigned long max = vma_compute_gap(vma), subtree_gap;
321 if (vma->vm_rb.rb_left) {
322 subtree_gap = rb_entry(vma->vm_rb.rb_left,
323 struct vm_area_struct, vm_rb)->rb_subtree_gap;
324 if (subtree_gap > max)
327 if (vma->vm_rb.rb_right) {
328 subtree_gap = rb_entry(vma->vm_rb.rb_right,
329 struct vm_area_struct, vm_rb)->rb_subtree_gap;
330 if (subtree_gap > max)
336 static int browse_rb(struct mm_struct *mm)
338 struct rb_root *root = &mm->mm_rb;
339 int i = 0, j, bug = 0;
340 struct rb_node *nd, *pn = NULL;
341 unsigned long prev = 0, pend = 0;
343 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
344 struct vm_area_struct *vma;
345 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
346 if (vma->vm_start < prev) {
347 pr_emerg("vm_start %lx < prev %lx\n",
348 vma->vm_start, prev);
351 if (vma->vm_start < pend) {
352 pr_emerg("vm_start %lx < pend %lx\n",
353 vma->vm_start, pend);
356 if (vma->vm_start > vma->vm_end) {
357 pr_emerg("vm_start %lx > vm_end %lx\n",
358 vma->vm_start, vma->vm_end);
361 spin_lock(&mm->page_table_lock);
362 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
363 pr_emerg("free gap %lx, correct %lx\n",
365 vma_compute_subtree_gap(vma));
368 spin_unlock(&mm->page_table_lock);
371 prev = vma->vm_start;
375 for (nd = pn; nd; nd = rb_prev(nd))
378 pr_emerg("backwards %d, forwards %d\n", j, i);
384 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
388 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
389 struct vm_area_struct *vma;
390 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
391 VM_BUG_ON_VMA(vma != ignore &&
392 vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
397 static void validate_mm(struct mm_struct *mm)
401 unsigned long highest_address = 0;
402 struct vm_area_struct *vma = mm->mmap;
405 struct anon_vma *anon_vma = vma->anon_vma;
406 struct anon_vma_chain *avc;
409 anon_vma_lock_read(anon_vma);
410 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
411 anon_vma_interval_tree_verify(avc);
412 anon_vma_unlock_read(anon_vma);
415 highest_address = vm_end_gap(vma);
419 if (i != mm->map_count) {
420 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
423 if (highest_address != mm->highest_vm_end) {
424 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
425 mm->highest_vm_end, highest_address);
429 if (i != mm->map_count) {
431 pr_emerg("map_count %d rb %d\n", mm->map_count, i);
434 VM_BUG_ON_MM(bug, mm);
437 #define validate_mm_rb(root, ignore) do { } while (0)
438 #define validate_mm(mm) do { } while (0)
441 RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks,
442 struct vm_area_struct, vm_rb,
443 unsigned long, rb_subtree_gap, vma_compute_gap)
446 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
447 * vma->vm_prev->vm_end values changed, without modifying the vma's position
450 static void vma_gap_update(struct vm_area_struct *vma)
453 * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
454 * a callback function that does exactly what we want.
456 vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
459 static inline void vma_rb_insert(struct vm_area_struct *vma,
460 struct rb_root *root)
462 /* All rb_subtree_gap values must be consistent prior to insertion */
463 validate_mm_rb(root, NULL);
465 rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
468 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
471 * Note rb_erase_augmented is a fairly large inline function,
472 * so make sure we instantiate it only once with our desired
473 * augmented rbtree callbacks.
475 rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
478 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
479 struct rb_root *root,
480 struct vm_area_struct *ignore)
483 * All rb_subtree_gap values must be consistent prior to erase,
484 * with the possible exception of
486 * a. the "next" vma being erased if next->vm_start was reduced in
487 * __vma_adjust() -> __vma_unlink()
488 * b. the vma being erased in detach_vmas_to_be_unmapped() ->
491 validate_mm_rb(root, ignore);
493 __vma_rb_erase(vma, root);
496 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
497 struct rb_root *root)
499 vma_rb_erase_ignore(vma, root, vma);
503 * vma has some anon_vma assigned, and is already inserted on that
504 * anon_vma's interval trees.
506 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
507 * vma must be removed from the anon_vma's interval trees using
508 * anon_vma_interval_tree_pre_update_vma().
510 * After the update, the vma will be reinserted using
511 * anon_vma_interval_tree_post_update_vma().
513 * The entire update must be protected by exclusive mmap_lock and by
514 * the root anon_vma's mutex.
517 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
519 struct anon_vma_chain *avc;
521 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
522 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
526 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
528 struct anon_vma_chain *avc;
530 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
531 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
534 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
535 unsigned long end, struct vm_area_struct **pprev,
536 struct rb_node ***rb_link, struct rb_node **rb_parent)
538 struct rb_node **__rb_link, *__rb_parent, *rb_prev;
540 __rb_link = &mm->mm_rb.rb_node;
541 rb_prev = __rb_parent = NULL;
544 struct vm_area_struct *vma_tmp;
546 __rb_parent = *__rb_link;
547 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
549 if (vma_tmp->vm_end > addr) {
550 /* Fail if an existing vma overlaps the area */
551 if (vma_tmp->vm_start < end)
553 __rb_link = &__rb_parent->rb_left;
555 rb_prev = __rb_parent;
556 __rb_link = &__rb_parent->rb_right;
562 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
563 *rb_link = __rb_link;
564 *rb_parent = __rb_parent;
569 * vma_next() - Get the next VMA.
570 * @mm: The mm_struct.
571 * @vma: The current vma.
573 * If @vma is NULL, return the first vma in the mm.
575 * Returns: The next VMA after @vma.
577 static inline struct vm_area_struct *vma_next(struct mm_struct *mm,
578 struct vm_area_struct *vma)
587 * munmap_vma_range() - munmap VMAs that overlap a range.
589 * @start: The start of the range.
590 * @len: The length of the range.
591 * @pprev: pointer to the pointer that will be set to previous vm_area_struct
592 * @rb_link: the rb_node
593 * @rb_parent: the parent rb_node
595 * Find all the vm_area_struct that overlap from @start to
596 * @end and munmap them. Set @pprev to the previous vm_area_struct.
598 * Returns: -ENOMEM on munmap failure or 0 on success.
601 munmap_vma_range(struct mm_struct *mm, unsigned long start, unsigned long len,
602 struct vm_area_struct **pprev, struct rb_node ***link,
603 struct rb_node **parent, struct list_head *uf)
606 while (find_vma_links(mm, start, start + len, pprev, link, parent))
607 if (do_munmap(mm, start, len, uf))
612 static unsigned long count_vma_pages_range(struct mm_struct *mm,
613 unsigned long addr, unsigned long end)
615 unsigned long nr_pages = 0;
616 struct vm_area_struct *vma;
618 /* Find first overlaping mapping */
619 vma = find_vma_intersection(mm, addr, end);
623 nr_pages = (min(end, vma->vm_end) -
624 max(addr, vma->vm_start)) >> PAGE_SHIFT;
626 /* Iterate over the rest of the overlaps */
627 for (vma = vma->vm_next; vma; vma = vma->vm_next) {
628 unsigned long overlap_len;
630 if (vma->vm_start > end)
633 overlap_len = min(end, vma->vm_end) - vma->vm_start;
634 nr_pages += overlap_len >> PAGE_SHIFT;
640 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
641 struct rb_node **rb_link, struct rb_node *rb_parent)
643 /* Update tracking information for the gap following the new vma. */
645 vma_gap_update(vma->vm_next);
647 mm->highest_vm_end = vm_end_gap(vma);
650 * vma->vm_prev wasn't known when we followed the rbtree to find the
651 * correct insertion point for that vma. As a result, we could not
652 * update the vma vm_rb parents rb_subtree_gap values on the way down.
653 * So, we first insert the vma with a zero rb_subtree_gap value
654 * (to be consistent with what we did on the way down), and then
655 * immediately update the gap to the correct value. Finally we
656 * rebalance the rbtree after all augmented values have been set.
658 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
659 vma->rb_subtree_gap = 0;
661 vma_rb_insert(vma, &mm->mm_rb);
664 static void __vma_link_file(struct vm_area_struct *vma)
670 struct address_space *mapping = file->f_mapping;
672 if (vma->vm_flags & VM_DENYWRITE)
673 put_write_access(file_inode(file));
674 if (vma->vm_flags & VM_SHARED)
675 mapping_allow_writable(mapping);
677 flush_dcache_mmap_lock(mapping);
678 vma_interval_tree_insert(vma, &mapping->i_mmap);
679 flush_dcache_mmap_unlock(mapping);
684 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
685 struct vm_area_struct *prev, struct rb_node **rb_link,
686 struct rb_node *rb_parent)
688 __vma_link_list(mm, vma, prev);
689 __vma_link_rb(mm, vma, rb_link, rb_parent);
692 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
693 struct vm_area_struct *prev, struct rb_node **rb_link,
694 struct rb_node *rb_parent)
696 struct address_space *mapping = NULL;
699 mapping = vma->vm_file->f_mapping;
700 i_mmap_lock_write(mapping);
703 __vma_link(mm, vma, prev, rb_link, rb_parent);
704 __vma_link_file(vma);
707 i_mmap_unlock_write(mapping);
714 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
715 * mm's list and rbtree. It has already been inserted into the interval tree.
717 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
719 struct vm_area_struct *prev;
720 struct rb_node **rb_link, *rb_parent;
722 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
723 &prev, &rb_link, &rb_parent))
725 __vma_link(mm, vma, prev, rb_link, rb_parent);
729 static __always_inline void __vma_unlink(struct mm_struct *mm,
730 struct vm_area_struct *vma,
731 struct vm_area_struct *ignore)
733 vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
734 __vma_unlink_list(mm, vma);
736 vmacache_invalidate(mm);
740 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
741 * is already present in an i_mmap tree without adjusting the tree.
742 * The following helper function should be used when such adjustments
743 * are necessary. The "insert" vma (if any) is to be inserted
744 * before we drop the necessary locks.
746 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
747 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
748 struct vm_area_struct *expand)
750 struct mm_struct *mm = vma->vm_mm;
751 struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
752 struct address_space *mapping = NULL;
753 struct rb_root_cached *root = NULL;
754 struct anon_vma *anon_vma = NULL;
755 struct file *file = vma->vm_file;
756 bool start_changed = false, end_changed = false;
757 long adjust_next = 0;
760 if (next && !insert) {
761 struct vm_area_struct *exporter = NULL, *importer = NULL;
763 if (end >= next->vm_end) {
765 * vma expands, overlapping all the next, and
766 * perhaps the one after too (mprotect case 6).
767 * The only other cases that gets here are
768 * case 1, case 7 and case 8.
770 if (next == expand) {
772 * The only case where we don't expand "vma"
773 * and we expand "next" instead is case 8.
775 VM_WARN_ON(end != next->vm_end);
777 * remove_next == 3 means we're
778 * removing "vma" and that to do so we
779 * swapped "vma" and "next".
782 VM_WARN_ON(file != next->vm_file);
785 VM_WARN_ON(expand != vma);
787 * case 1, 6, 7, remove_next == 2 is case 6,
788 * remove_next == 1 is case 1 or 7.
790 remove_next = 1 + (end > next->vm_end);
791 VM_WARN_ON(remove_next == 2 &&
792 end != next->vm_next->vm_end);
793 /* trim end to next, for case 6 first pass */
801 * If next doesn't have anon_vma, import from vma after
802 * next, if the vma overlaps with it.
804 if (remove_next == 2 && !next->anon_vma)
805 exporter = next->vm_next;
807 } else if (end > next->vm_start) {
809 * vma expands, overlapping part of the next:
810 * mprotect case 5 shifting the boundary up.
812 adjust_next = (end - next->vm_start);
815 VM_WARN_ON(expand != importer);
816 } else if (end < vma->vm_end) {
818 * vma shrinks, and !insert tells it's not
819 * split_vma inserting another: so it must be
820 * mprotect case 4 shifting the boundary down.
822 adjust_next = -(vma->vm_end - end);
825 VM_WARN_ON(expand != importer);
829 * Easily overlooked: when mprotect shifts the boundary,
830 * make sure the expanding vma has anon_vma set if the
831 * shrinking vma had, to cover any anon pages imported.
833 if (exporter && exporter->anon_vma && !importer->anon_vma) {
836 importer->anon_vma = exporter->anon_vma;
837 error = anon_vma_clone(importer, exporter);
843 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
846 mapping = file->f_mapping;
847 root = &mapping->i_mmap;
848 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
851 uprobe_munmap(next, next->vm_start, next->vm_end);
853 i_mmap_lock_write(mapping);
856 * Put into interval tree now, so instantiated pages
857 * are visible to arm/parisc __flush_dcache_page
858 * throughout; but we cannot insert into address
859 * space until vma start or end is updated.
861 __vma_link_file(insert);
865 anon_vma = vma->anon_vma;
866 if (!anon_vma && adjust_next)
867 anon_vma = next->anon_vma;
869 VM_WARN_ON(adjust_next && next->anon_vma &&
870 anon_vma != next->anon_vma);
871 anon_vma_lock_write(anon_vma);
872 anon_vma_interval_tree_pre_update_vma(vma);
874 anon_vma_interval_tree_pre_update_vma(next);
878 flush_dcache_mmap_lock(mapping);
879 vma_interval_tree_remove(vma, root);
881 vma_interval_tree_remove(next, root);
884 if (start != vma->vm_start) {
885 vma->vm_start = start;
886 start_changed = true;
888 if (end != vma->vm_end) {
892 vma->vm_pgoff = pgoff;
894 next->vm_start += adjust_next;
895 next->vm_pgoff += adjust_next >> PAGE_SHIFT;
900 vma_interval_tree_insert(next, root);
901 vma_interval_tree_insert(vma, root);
902 flush_dcache_mmap_unlock(mapping);
907 * vma_merge has merged next into vma, and needs
908 * us to remove next before dropping the locks.
910 if (remove_next != 3)
911 __vma_unlink(mm, next, next);
914 * vma is not before next if they've been
917 * pre-swap() next->vm_start was reduced so
918 * tell validate_mm_rb to ignore pre-swap()
919 * "next" (which is stored in post-swap()
922 __vma_unlink(mm, next, vma);
924 __remove_shared_vm_struct(next, file, mapping);
927 * split_vma has split insert from vma, and needs
928 * us to insert it before dropping the locks
929 * (it may either follow vma or precede it).
931 __insert_vm_struct(mm, insert);
937 mm->highest_vm_end = vm_end_gap(vma);
938 else if (!adjust_next)
939 vma_gap_update(next);
944 anon_vma_interval_tree_post_update_vma(vma);
946 anon_vma_interval_tree_post_update_vma(next);
947 anon_vma_unlock_write(anon_vma);
951 i_mmap_unlock_write(mapping);
960 uprobe_munmap(next, next->vm_start, next->vm_end);
964 anon_vma_merge(vma, next);
966 mpol_put(vma_policy(next));
969 * In mprotect's case 6 (see comments on vma_merge),
970 * we must remove another next too. It would clutter
971 * up the code too much to do both in one go.
973 if (remove_next != 3) {
975 * If "next" was removed and vma->vm_end was
976 * expanded (up) over it, in turn
977 * "next->vm_prev->vm_end" changed and the
978 * "vma->vm_next" gap must be updated.
983 * For the scope of the comment "next" and
984 * "vma" considered pre-swap(): if "vma" was
985 * removed, next->vm_start was expanded (down)
986 * over it and the "next" gap must be updated.
987 * Because of the swap() the post-swap() "vma"
988 * actually points to pre-swap() "next"
989 * (post-swap() "next" as opposed is now a
994 if (remove_next == 2) {
1000 vma_gap_update(next);
1003 * If remove_next == 2 we obviously can't
1006 * If remove_next == 3 we can't reach this
1007 * path because pre-swap() next is always not
1008 * NULL. pre-swap() "next" is not being
1009 * removed and its next->vm_end is not altered
1010 * (and furthermore "end" already matches
1011 * next->vm_end in remove_next == 3).
1013 * We reach this only in the remove_next == 1
1014 * case if the "next" vma that was removed was
1015 * the highest vma of the mm. However in such
1016 * case next->vm_end == "end" and the extended
1017 * "vma" has vma->vm_end == next->vm_end so
1018 * mm->highest_vm_end doesn't need any update
1019 * in remove_next == 1 case.
1021 VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
1025 uprobe_mmap(insert);
1033 * If the vma has a ->close operation then the driver probably needs to release
1034 * per-vma resources, so we don't attempt to merge those.
1036 static inline int is_mergeable_vma(struct vm_area_struct *vma,
1037 struct file *file, unsigned long vm_flags,
1038 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1041 * VM_SOFTDIRTY should not prevent from VMA merging, if we
1042 * match the flags but dirty bit -- the caller should mark
1043 * merged VMA as dirty. If dirty bit won't be excluded from
1044 * comparison, we increase pressure on the memory system forcing
1045 * the kernel to generate new VMAs when old one could be
1048 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
1050 if (vma->vm_file != file)
1052 if (vma->vm_ops && vma->vm_ops->close)
1054 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1059 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1060 struct anon_vma *anon_vma2,
1061 struct vm_area_struct *vma)
1064 * The list_is_singular() test is to avoid merging VMA cloned from
1065 * parents. This can improve scalability caused by anon_vma lock.
1067 if ((!anon_vma1 || !anon_vma2) && (!vma ||
1068 list_is_singular(&vma->anon_vma_chain)))
1070 return anon_vma1 == anon_vma2;
1074 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1075 * in front of (at a lower virtual address and file offset than) the vma.
1077 * We cannot merge two vmas if they have differently assigned (non-NULL)
1078 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1080 * We don't check here for the merged mmap wrapping around the end of pagecache
1081 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
1082 * wrap, nor mmaps which cover the final page at index -1UL.
1085 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1086 struct anon_vma *anon_vma, struct file *file,
1088 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1090 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1091 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1092 if (vma->vm_pgoff == vm_pgoff)
1099 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1100 * beyond (at a higher virtual address and file offset than) the vma.
1102 * We cannot merge two vmas if they have differently assigned (non-NULL)
1103 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1106 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1107 struct anon_vma *anon_vma, struct file *file,
1109 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1111 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
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), figure out
1123 * whether that can be merged with its predecessor or its successor.
1124 * 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)
1171 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1172 struct vm_area_struct *area, *next;
1176 * We later require that vma->vm_flags == vm_flags,
1177 * so this tests vma->vm_flags & VM_SPECIAL, too.
1179 if (vm_flags & VM_SPECIAL)
1182 next = vma_next(mm, prev);
1184 if (area && area->vm_end == end) /* cases 6, 7, 8 */
1185 next = next->vm_next;
1187 /* verify some invariant that must be enforced by the caller */
1188 VM_WARN_ON(prev && addr <= prev->vm_start);
1189 VM_WARN_ON(area && end > area->vm_end);
1190 VM_WARN_ON(addr >= end);
1193 * Can it merge with the predecessor?
1195 if (prev && prev->vm_end == addr &&
1196 mpol_equal(vma_policy(prev), policy) &&
1197 can_vma_merge_after(prev, vm_flags,
1198 anon_vma, file, pgoff,
1199 vm_userfaultfd_ctx)) {
1201 * OK, it can. Can we now merge in the successor as well?
1203 if (next && end == next->vm_start &&
1204 mpol_equal(policy, vma_policy(next)) &&
1205 can_vma_merge_before(next, vm_flags,
1208 vm_userfaultfd_ctx) &&
1209 is_mergeable_anon_vma(prev->anon_vma,
1210 next->anon_vma, NULL)) {
1212 err = __vma_adjust(prev, prev->vm_start,
1213 next->vm_end, prev->vm_pgoff, NULL,
1215 } else /* cases 2, 5, 7 */
1216 err = __vma_adjust(prev, prev->vm_start,
1217 end, prev->vm_pgoff, NULL, prev);
1220 khugepaged_enter_vma_merge(prev, vm_flags);
1225 * Can this new request be merged in front of next?
1227 if (next && end == next->vm_start &&
1228 mpol_equal(policy, vma_policy(next)) &&
1229 can_vma_merge_before(next, vm_flags,
1230 anon_vma, file, pgoff+pglen,
1231 vm_userfaultfd_ctx)) {
1232 if (prev && addr < prev->vm_end) /* case 4 */
1233 err = __vma_adjust(prev, prev->vm_start,
1234 addr, prev->vm_pgoff, NULL, next);
1235 else { /* cases 3, 8 */
1236 err = __vma_adjust(area, addr, next->vm_end,
1237 next->vm_pgoff - pglen, NULL, next);
1239 * In case 3 area is already equal to next and
1240 * this is a noop, but in case 8 "area" has
1241 * been removed and next was expanded over it.
1247 khugepaged_enter_vma_merge(area, vm_flags);
1255 * Rough compatibility check to quickly see if it's even worth looking
1256 * at sharing an anon_vma.
1258 * They need to have the same vm_file, and the flags can only differ
1259 * in things that mprotect may change.
1261 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1262 * we can merge the two vma's. For example, we refuse to merge a vma if
1263 * there is a vm_ops->close() function, because that indicates that the
1264 * driver is doing some kind of reference counting. But that doesn't
1265 * really matter for the anon_vma sharing case.
1267 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1269 return a->vm_end == b->vm_start &&
1270 mpol_equal(vma_policy(a), vma_policy(b)) &&
1271 a->vm_file == b->vm_file &&
1272 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1273 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1277 * Do some basic sanity checking to see if we can re-use the anon_vma
1278 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1279 * the same as 'old', the other will be the new one that is trying
1280 * to share the anon_vma.
1282 * NOTE! This runs with mm_sem held for reading, so it is possible that
1283 * the anon_vma of 'old' is concurrently in the process of being set up
1284 * by another page fault trying to merge _that_. But that's ok: if it
1285 * is being set up, that automatically means that it will be a singleton
1286 * acceptable for merging, so we can do all of this optimistically. But
1287 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1289 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1290 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1291 * is to return an anon_vma that is "complex" due to having gone through
1294 * We also make sure that the two vma's are compatible (adjacent,
1295 * and with the same memory policies). That's all stable, even with just
1296 * a read lock on the mm_sem.
1298 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1300 if (anon_vma_compatible(a, b)) {
1301 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1303 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1310 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1311 * neighbouring vmas for a suitable anon_vma, before it goes off
1312 * to allocate a new anon_vma. It checks because a repetitive
1313 * sequence of mprotects and faults may otherwise lead to distinct
1314 * anon_vmas being allocated, preventing vma merge in subsequent
1317 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1319 struct anon_vma *anon_vma = NULL;
1321 /* Try next first. */
1323 anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next);
1328 /* Try prev next. */
1330 anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma);
1333 * We might reach here with anon_vma == NULL if we can't find
1334 * any reusable anon_vma.
1335 * There's no absolute need to look only at touching neighbours:
1336 * we could search further afield for "compatible" anon_vmas.
1337 * But it would probably just be a waste of time searching,
1338 * or lead to too many vmas hanging off the same anon_vma.
1339 * We're trying to allow mprotect remerging later on,
1340 * not trying to minimize memory used for anon_vmas.
1346 * If a hint addr is less than mmap_min_addr change hint to be as
1347 * low as possible but still greater than mmap_min_addr
1349 static inline unsigned long round_hint_to_min(unsigned long hint)
1352 if (((void *)hint != NULL) &&
1353 (hint < mmap_min_addr))
1354 return PAGE_ALIGN(mmap_min_addr);
1358 static inline int mlock_future_check(struct mm_struct *mm,
1359 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 #ifdef CONFIG_FINEGRAINED_THP
1456 if ((len >> PAGE_SHIFT) >= HPAGE_CONT_PTE_NR &&
1458 flags |= MAP_FILE_THP;
1461 /* Obtain the address to map to. we verify (or select) it and ensure
1462 * that it represents a valid section of the address space.
1464 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1465 if (IS_ERR_VALUE(addr))
1468 if (flags & MAP_FIXED_NOREPLACE) {
1469 struct vm_area_struct *vma = find_vma(mm, addr);
1471 if (vma && vma->vm_start < addr + len)
1475 if (prot == PROT_EXEC) {
1476 pkey = execute_only_pkey(mm);
1481 /* Do simple checking here so the lower-level routines won't have
1482 * to. we assume access permissions have been handled by the open
1483 * of the memory object, so we don't do any here.
1485 vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1486 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1488 if (flags & MAP_LOCKED)
1489 if (!can_do_mlock())
1492 if (mlock_future_check(mm, vm_flags, len))
1496 struct inode *inode = file_inode(file);
1497 unsigned long flags_mask;
1499 if (!file_mmap_ok(file, inode, pgoff, len))
1502 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1504 switch (flags & MAP_TYPE) {
1507 * Force use of MAP_SHARED_VALIDATE with non-legacy
1508 * flags. E.g. MAP_SYNC is dangerous to use with
1509 * MAP_SHARED as you don't know which consistency model
1510 * you will get. We silently ignore unsupported flags
1511 * with MAP_SHARED to preserve backward compatibility.
1513 flags &= LEGACY_MAP_MASK;
1515 case MAP_SHARED_VALIDATE:
1516 if (flags & ~flags_mask)
1518 if (prot & PROT_WRITE) {
1519 if (!(file->f_mode & FMODE_WRITE))
1521 if (IS_SWAPFILE(file->f_mapping->host))
1526 * Make sure we don't allow writing to an append-only
1529 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1533 * Make sure there are no mandatory locks on the file.
1535 if (locks_verify_locked(file))
1538 vm_flags |= VM_SHARED | VM_MAYSHARE;
1539 if (!(file->f_mode & FMODE_WRITE))
1540 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1543 if (!(file->f_mode & FMODE_READ))
1545 if (path_noexec(&file->f_path)) {
1546 if (vm_flags & VM_EXEC)
1548 vm_flags &= ~VM_MAYEXEC;
1551 if (!file->f_op->mmap)
1553 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1561 switch (flags & MAP_TYPE) {
1563 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1569 vm_flags |= VM_SHARED | VM_MAYSHARE;
1573 * Set pgoff according to addr for anon_vma.
1575 pgoff = addr >> PAGE_SHIFT;
1583 * Set 'VM_NORESERVE' if we should not account for the
1584 * memory use of this mapping.
1586 if (flags & MAP_NORESERVE) {
1587 /* We honor MAP_NORESERVE if allowed to overcommit */
1588 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1589 vm_flags |= VM_NORESERVE;
1591 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1592 if (file && is_file_hugepages(file))
1593 vm_flags |= VM_NORESERVE;
1596 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1597 if (!IS_ERR_VALUE(addr) &&
1598 ((vm_flags & VM_LOCKED) ||
1599 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1604 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1605 unsigned long prot, unsigned long flags,
1606 unsigned long fd, unsigned long pgoff)
1608 struct file *file = NULL;
1609 unsigned long retval;
1611 if (!(flags & MAP_ANONYMOUS)) {
1612 audit_mmap_fd(fd, flags);
1616 if (is_file_hugepages(file)) {
1617 len = ALIGN(len, huge_page_size(hstate_file(file)));
1618 } else if (unlikely(flags & MAP_HUGETLB)) {
1622 } else if (flags & MAP_HUGETLB) {
1623 struct user_struct *user = NULL;
1626 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1630 len = ALIGN(len, huge_page_size(hs));
1632 * VM_NORESERVE is used because the reservations will be
1633 * taken when vm_ops->mmap() is called
1634 * A dummy user value is used because we are not locking
1635 * memory so no accounting is necessary
1637 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1639 &user, HUGETLB_ANONHUGE_INODE,
1640 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1642 return PTR_ERR(file);
1645 flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1647 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1654 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1655 unsigned long, prot, unsigned long, flags,
1656 unsigned long, fd, unsigned long, pgoff)
1658 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1661 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1662 struct mmap_arg_struct {
1666 unsigned long flags;
1668 unsigned long offset;
1671 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1673 struct mmap_arg_struct a;
1675 if (copy_from_user(&a, arg, sizeof(a)))
1677 if (offset_in_page(a.offset))
1680 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1681 a.offset >> PAGE_SHIFT);
1683 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1686 * Some shared mappings will want the pages marked read-only
1687 * to track write events. If so, we'll downgrade vm_page_prot
1688 * to the private version (using protection_map[] without the
1691 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1693 vm_flags_t vm_flags = vma->vm_flags;
1694 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1696 /* If it was private or non-writable, the write bit is already clear */
1697 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1700 /* The backer wishes to know when pages are first written to? */
1701 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1704 /* The open routine did something to the protections that pgprot_modify
1705 * won't preserve? */
1706 if (pgprot_val(vm_page_prot) !=
1707 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1710 /* Do we need to track softdirty? */
1711 if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1714 /* Specialty mapping? */
1715 if (vm_flags & VM_PFNMAP)
1718 /* Can the mapping track the dirty pages? */
1719 return vma->vm_file && vma->vm_file->f_mapping &&
1720 mapping_can_writeback(vma->vm_file->f_mapping);
1724 * We account for memory if it's a private writeable mapping,
1725 * not hugepages and VM_NORESERVE wasn't set.
1727 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1730 * hugetlb has its own accounting separate from the core VM
1731 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1733 if (file && is_file_hugepages(file))
1736 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1739 unsigned long mmap_region(struct file *file, unsigned long addr,
1740 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1741 struct list_head *uf)
1743 struct mm_struct *mm = current->mm;
1744 struct vm_area_struct *vma, *prev, *merge;
1746 struct rb_node **rb_link, *rb_parent;
1747 unsigned long charged = 0;
1749 /* Check against address space limit. */
1750 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1751 unsigned long nr_pages;
1754 * MAP_FIXED may remove pages of mappings that intersects with
1755 * requested mapping. Account for the pages it would unmap.
1757 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1759 if (!may_expand_vm(mm, vm_flags,
1760 (len >> PAGE_SHIFT) - nr_pages))
1764 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
1765 if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
1768 * Private writable mapping: check memory availability
1770 if (accountable_mapping(file, vm_flags)) {
1771 charged = len >> PAGE_SHIFT;
1772 if (security_vm_enough_memory_mm(mm, charged))
1774 vm_flags |= VM_ACCOUNT;
1778 * Can we just expand an old mapping?
1780 vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1781 NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1786 * Determine the object being mapped and call the appropriate
1787 * specific mapper. the address has already been validated, but
1788 * not unmapped, but the maps are removed from the list.
1790 vma = vm_area_alloc(mm);
1796 vma->vm_start = addr;
1797 vma->vm_end = addr + len;
1798 vma->vm_flags = vm_flags;
1799 vma->vm_page_prot = vm_get_page_prot(vm_flags);
1800 vma->vm_pgoff = pgoff;
1803 if (vm_flags & VM_DENYWRITE) {
1804 error = deny_write_access(file);
1808 if (vm_flags & VM_SHARED) {
1809 error = mapping_map_writable(file->f_mapping);
1811 goto allow_write_and_free_vma;
1814 /* ->mmap() can change vma->vm_file, but must guarantee that
1815 * vma_link() below can deny write-access if VM_DENYWRITE is set
1816 * and map writably if VM_SHARED is set. This usually means the
1817 * new file must not have been exposed to user-space, yet.
1819 vma->vm_file = get_file(file);
1820 error = call_mmap(file, vma);
1822 goto unmap_and_free_vma;
1824 /* Can addr have changed??
1826 * Answer: Yes, several device drivers can do it in their
1827 * f_op->mmap method. -DaveM
1828 * Bug: If addr is changed, prev, rb_link, rb_parent should
1829 * be updated for vma_link()
1831 WARN_ON_ONCE(addr != vma->vm_start);
1833 addr = vma->vm_start;
1835 /* If vm_flags changed after call_mmap(), we should try merge vma again
1836 * as we may succeed this time.
1838 if (unlikely(vm_flags != vma->vm_flags && prev)) {
1839 merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
1840 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX);
1842 /* ->mmap() can change vma->vm_file and fput the original file. So
1843 * fput the vma->vm_file here or we would add an extra fput for file
1844 * and cause general protection fault ultimately.
1849 /* Update vm_flags to pick up the change. */
1850 vm_flags = vma->vm_flags;
1851 goto unmap_writable;
1855 vm_flags = vma->vm_flags;
1856 } else if (vm_flags & VM_SHARED) {
1857 error = shmem_zero_setup(vma);
1861 vma_set_anonymous(vma);
1864 /* Allow architectures to sanity-check the vm_flags */
1865 if (!arch_validate_flags(vma->vm_flags)) {
1868 goto unmap_and_free_vma;
1873 vma_link(mm, vma, prev, rb_link, rb_parent);
1874 /* Once vma denies write, undo our temporary denial count */
1877 if (vm_flags & VM_SHARED)
1878 mapping_unmap_writable(file->f_mapping);
1879 if (vm_flags & VM_DENYWRITE)
1880 allow_write_access(file);
1882 file = vma->vm_file;
1883 if (file && (vm_flags & VM_DENYWRITE))
1884 /* read-only file pages */
1885 khugepaged_mem_hook(mm, addr, len, __func__);
1886 else if (!file && !vma->vm_ops)
1887 /* anonymous pages */
1888 khugepaged_mem_hook(mm, addr, len, __func__);
1890 perf_event_mmap(vma);
1892 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1893 if (vm_flags & VM_LOCKED) {
1894 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1895 is_vm_hugetlb_page(vma) ||
1896 vma == get_gate_vma(current->mm))
1897 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1899 mm->locked_vm += (len >> PAGE_SHIFT);
1906 * New (or expanded) vma always get soft dirty status.
1907 * Otherwise user-space soft-dirty page tracker won't
1908 * be able to distinguish situation when vma area unmapped,
1909 * then new mapped in-place (which must be aimed as
1910 * a completely new data area).
1912 vma->vm_flags |= VM_SOFTDIRTY;
1914 vma_set_page_prot(vma);
1919 vma->vm_file = NULL;
1922 /* Undo any partial mapping done by a device driver. */
1923 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1925 if (vm_flags & VM_SHARED)
1926 mapping_unmap_writable(file->f_mapping);
1927 allow_write_and_free_vma:
1928 if (vm_flags & VM_DENYWRITE)
1929 allow_write_access(file);
1934 vm_unacct_memory(charged);
1938 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1941 * We implement the search by looking for an rbtree node that
1942 * immediately follows a suitable gap. That is,
1943 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1944 * - gap_end = vma->vm_start >= info->low_limit + length;
1945 * - gap_end - gap_start >= length
1948 struct mm_struct *mm = current->mm;
1949 struct vm_area_struct *vma;
1950 unsigned long length, low_limit, high_limit, gap_start, gap_end;
1952 /* Adjust search length to account for worst case alignment overhead */
1953 length = info->length + info->align_mask;
1954 if (length < info->length)
1957 /* Adjust search limits by the desired length */
1958 if (info->high_limit < length)
1960 high_limit = info->high_limit - length;
1962 if (info->low_limit > high_limit)
1964 low_limit = info->low_limit + length;
1966 /* Check if rbtree root looks promising */
1967 if (RB_EMPTY_ROOT(&mm->mm_rb))
1969 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1970 if (vma->rb_subtree_gap < length)
1974 /* Visit left subtree if it looks promising */
1975 gap_end = vm_start_gap(vma);
1976 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1977 struct vm_area_struct *left =
1978 rb_entry(vma->vm_rb.rb_left,
1979 struct vm_area_struct, vm_rb);
1980 if (left->rb_subtree_gap >= length) {
1986 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1988 /* Check if current node has a suitable gap */
1989 if (gap_start > high_limit)
1991 if (gap_end >= low_limit &&
1992 gap_end > gap_start && gap_end - gap_start >= length)
1995 /* Visit right subtree if it looks promising */
1996 if (vma->vm_rb.rb_right) {
1997 struct vm_area_struct *right =
1998 rb_entry(vma->vm_rb.rb_right,
1999 struct vm_area_struct, vm_rb);
2000 if (right->rb_subtree_gap >= length) {
2006 /* Go back up the rbtree to find next candidate node */
2008 struct rb_node *prev = &vma->vm_rb;
2009 if (!rb_parent(prev))
2011 vma = rb_entry(rb_parent(prev),
2012 struct vm_area_struct, vm_rb);
2013 if (prev == vma->vm_rb.rb_left) {
2014 gap_start = vm_end_gap(vma->vm_prev);
2015 gap_end = vm_start_gap(vma);
2022 /* Check highest gap, which does not precede any rbtree node */
2023 gap_start = mm->highest_vm_end;
2024 gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
2025 if (gap_start > high_limit)
2029 /* We found a suitable gap. Clip it with the original low_limit. */
2030 if (gap_start < info->low_limit)
2031 gap_start = info->low_limit;
2033 /* Adjust gap address to the desired alignment */
2034 gap_start += (info->align_offset - gap_start) & info->align_mask;
2036 VM_BUG_ON(gap_start + info->length > info->high_limit);
2037 VM_BUG_ON(gap_start + info->length > gap_end);
2041 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
2043 struct mm_struct *mm = current->mm;
2044 struct vm_area_struct *vma;
2045 unsigned long length, low_limit, high_limit, gap_start, gap_end;
2047 /* Adjust search length to account for worst case alignment overhead */
2048 length = info->length + info->align_mask;
2049 if (length < info->length)
2053 * Adjust search limits by the desired length.
2054 * See implementation comment at top of unmapped_area().
2056 gap_end = info->high_limit;
2057 if (gap_end < length)
2059 high_limit = gap_end - length;
2061 if (info->low_limit > high_limit)
2063 low_limit = info->low_limit + length;
2065 /* Check highest gap, which does not precede any rbtree node */
2066 gap_start = mm->highest_vm_end;
2067 if (gap_start <= high_limit)
2070 /* Check if rbtree root looks promising */
2071 if (RB_EMPTY_ROOT(&mm->mm_rb))
2073 vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
2074 if (vma->rb_subtree_gap < length)
2078 /* Visit right subtree if it looks promising */
2079 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
2080 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
2081 struct vm_area_struct *right =
2082 rb_entry(vma->vm_rb.rb_right,
2083 struct vm_area_struct, vm_rb);
2084 if (right->rb_subtree_gap >= length) {
2091 /* Check if current node has a suitable gap */
2092 gap_end = vm_start_gap(vma);
2093 if (gap_end < low_limit)
2095 if (gap_start <= high_limit &&
2096 gap_end > gap_start && gap_end - gap_start >= length)
2099 /* Visit left subtree if it looks promising */
2100 if (vma->vm_rb.rb_left) {
2101 struct vm_area_struct *left =
2102 rb_entry(vma->vm_rb.rb_left,
2103 struct vm_area_struct, vm_rb);
2104 if (left->rb_subtree_gap >= length) {
2110 /* Go back up the rbtree to find next candidate node */
2112 struct rb_node *prev = &vma->vm_rb;
2113 if (!rb_parent(prev))
2115 vma = rb_entry(rb_parent(prev),
2116 struct vm_area_struct, vm_rb);
2117 if (prev == vma->vm_rb.rb_right) {
2118 gap_start = vma->vm_prev ?
2119 vm_end_gap(vma->vm_prev) : 0;
2126 /* We found a suitable gap. Clip it with the original high_limit. */
2127 if (gap_end > info->high_limit)
2128 gap_end = info->high_limit;
2131 /* Compute highest gap address at the desired alignment */
2132 gap_end -= info->length;
2133 gap_end -= (gap_end - info->align_offset) & info->align_mask;
2135 VM_BUG_ON(gap_end < info->low_limit);
2136 VM_BUG_ON(gap_end < gap_start);
2141 * Search for an unmapped address range.
2143 * We are looking for a range that:
2144 * - does not intersect with any VMA;
2145 * - is contained within the [low_limit, high_limit) interval;
2146 * - is at least the desired size.
2147 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
2149 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
2153 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
2154 addr = unmapped_area_topdown(info);
2156 addr = unmapped_area(info);
2158 trace_vm_unmapped_area(addr, info);
2162 #ifndef arch_get_mmap_end
2163 #define arch_get_mmap_end(addr) (TASK_SIZE)
2166 #ifndef arch_get_mmap_base
2167 #define arch_get_mmap_base(addr, base) (base)
2170 /* Get an address range which is currently unmapped.
2171 * For shmat() with addr=0.
2173 * Ugly calling convention alert:
2174 * Return value with the low bits set means error value,
2176 * if (ret & ~PAGE_MASK)
2179 * This function "knows" that -ENOMEM has the bits set.
2181 #ifndef HAVE_ARCH_UNMAPPED_AREA
2183 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2184 unsigned long len, unsigned long pgoff, unsigned long flags)
2186 struct mm_struct *mm = current->mm;
2187 struct vm_area_struct *vma, *prev;
2188 struct vm_unmapped_area_info info;
2189 const unsigned long mmap_end = arch_get_mmap_end(addr);
2191 if (len > mmap_end - mmap_min_addr)
2194 if (flags & MAP_FIXED)
2198 addr = PAGE_ALIGN(addr);
2199 vma = find_vma_prev(mm, addr, &prev);
2200 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2201 (!vma || addr + len <= vm_start_gap(vma)) &&
2202 (!prev || addr >= vm_end_gap(prev)))
2208 info.low_limit = mm->mmap_base;
2209 info.high_limit = mmap_end;
2210 info.align_mask = 0;
2211 info.align_offset = 0;
2213 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2214 if (!addr && len >= HPAGE_PMD_SIZE) {
2215 info.align_mask = HPAGE_PMD_SIZE - 1;
2216 info.align_offset = HPAGE_PMD_SIZE;
2217 #ifdef CONFIG_FINEGRAINED_THP
2218 } else if (!addr && len >= HPAGE_CONT_PTE_SIZE) {
2219 info.align_mask = HPAGE_CONT_PTE_SIZE - 1;
2220 info.align_offset = HPAGE_CONT_PTE_SIZE;
2225 return vm_unmapped_area(&info);
2230 * This mmap-allocator allocates new areas top-down from below the
2231 * stack's low limit (the base):
2233 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2235 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2236 unsigned long len, unsigned long pgoff,
2237 unsigned long flags)
2239 struct vm_area_struct *vma, *prev;
2240 struct mm_struct *mm = current->mm;
2241 struct vm_unmapped_area_info info;
2242 const unsigned long mmap_end = arch_get_mmap_end(addr);
2244 /* requested length too big for entire address space */
2245 if (len > mmap_end - mmap_min_addr)
2248 if (flags & MAP_FIXED)
2251 /* requesting a specific address */
2253 addr = PAGE_ALIGN(addr);
2254 vma = find_vma_prev(mm, addr, &prev);
2255 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2256 (!vma || addr + len <= vm_start_gap(vma)) &&
2257 (!prev || addr >= vm_end_gap(prev)))
2261 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2263 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2264 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2265 info.align_mask = 0;
2266 info.align_offset = 0;
2268 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2269 if (!addr && len >= HPAGE_PMD_SIZE) {
2270 info.align_mask = HPAGE_PMD_SIZE - 1;
2271 info.align_offset = HPAGE_PMD_SIZE;
2272 #ifdef CONFIG_FINEGRAINED_THP
2273 } else if (!addr && len >= HPAGE_CONT_PTE_SIZE) {
2274 info.align_mask = HPAGE_CONT_PTE_SIZE - 1;
2275 info.align_offset = HPAGE_CONT_PTE_SIZE;
2280 addr = vm_unmapped_area(&info);
2283 * A failed mmap() very likely causes application failure,
2284 * so fall back to the bottom-up function here. This scenario
2285 * can happen with large stack limits and large mmap()
2288 if (offset_in_page(addr)) {
2289 VM_BUG_ON(addr != -ENOMEM);
2291 info.low_limit = TASK_UNMAPPED_BASE;
2292 info.high_limit = mmap_end;
2293 addr = vm_unmapped_area(&info);
2301 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2302 unsigned long pgoff, unsigned long flags)
2304 unsigned long (*get_area)(struct file *, unsigned long,
2305 unsigned long, unsigned long, unsigned long);
2307 unsigned long error = arch_mmap_check(addr, len, flags);
2311 /* Careful about overflows.. */
2312 if (len > TASK_SIZE)
2315 get_area = current->mm->get_unmapped_area;
2317 if (file->f_op->get_unmapped_area)
2318 get_area = file->f_op->get_unmapped_area;
2319 } else if (flags & MAP_SHARED) {
2321 * mmap_region() will call shmem_zero_setup() to create a file,
2322 * so use shmem's get_unmapped_area in case it can be huge.
2323 * do_mmap() will clear pgoff, so match alignment.
2326 get_area = shmem_get_unmapped_area;
2329 addr = get_area(file, addr, len, pgoff, flags);
2330 if (IS_ERR_VALUE(addr))
2333 if (addr > TASK_SIZE - len)
2335 if (offset_in_page(addr))
2338 error = security_mmap_addr(addr);
2339 return error ? error : addr;
2342 EXPORT_SYMBOL(get_unmapped_area);
2344 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2345 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2347 struct rb_node *rb_node;
2348 struct vm_area_struct *vma;
2350 /* Check the cache first. */
2351 vma = vmacache_find(mm, addr);
2355 rb_node = mm->mm_rb.rb_node;
2358 struct vm_area_struct *tmp;
2360 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2362 if (tmp->vm_end > addr) {
2364 if (tmp->vm_start <= addr)
2366 rb_node = rb_node->rb_left;
2368 rb_node = rb_node->rb_right;
2372 vmacache_update(addr, vma);
2376 EXPORT_SYMBOL(find_vma);
2379 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2381 struct vm_area_struct *
2382 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2383 struct vm_area_struct **pprev)
2385 struct vm_area_struct *vma;
2387 vma = find_vma(mm, addr);
2389 *pprev = vma->vm_prev;
2391 struct rb_node *rb_node = rb_last(&mm->mm_rb);
2393 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
2399 * Verify that the stack growth is acceptable and
2400 * update accounting. This is shared with both the
2401 * grow-up and grow-down cases.
2403 static int acct_stack_growth(struct vm_area_struct *vma,
2404 unsigned long size, unsigned long grow)
2406 struct mm_struct *mm = vma->vm_mm;
2407 unsigned long new_start;
2409 /* address space limit tests */
2410 if (!may_expand_vm(mm, vma->vm_flags, grow))
2413 /* Stack limit test */
2414 if (size > rlimit(RLIMIT_STACK))
2417 /* mlock limit tests */
2418 if (vma->vm_flags & VM_LOCKED) {
2419 unsigned long locked;
2420 unsigned long limit;
2421 locked = mm->locked_vm + grow;
2422 limit = rlimit(RLIMIT_MEMLOCK);
2423 limit >>= PAGE_SHIFT;
2424 if (locked > limit && !capable(CAP_IPC_LOCK))
2428 /* Check to ensure the stack will not grow into a hugetlb-only region */
2429 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2431 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2435 * Overcommit.. This must be the final test, as it will
2436 * update security statistics.
2438 if (security_vm_enough_memory_mm(mm, grow))
2444 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2446 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2447 * vma is the last one with address > vma->vm_end. Have to extend vma.
2449 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2451 struct mm_struct *mm = vma->vm_mm;
2452 struct vm_area_struct *next;
2453 unsigned long gap_addr;
2456 if (!(vma->vm_flags & VM_GROWSUP))
2459 /* Guard against exceeding limits of the address space. */
2460 address &= PAGE_MASK;
2461 if (address >= (TASK_SIZE & PAGE_MASK))
2463 address += PAGE_SIZE;
2465 /* Enforce stack_guard_gap */
2466 gap_addr = address + stack_guard_gap;
2468 /* Guard against overflow */
2469 if (gap_addr < address || gap_addr > TASK_SIZE)
2470 gap_addr = TASK_SIZE;
2472 next = vma->vm_next;
2473 if (next && next->vm_start < gap_addr && vma_is_accessible(next)) {
2474 if (!(next->vm_flags & VM_GROWSUP))
2476 /* Check that both stack segments have the same anon_vma? */
2479 /* We must make sure the anon_vma is allocated. */
2480 if (unlikely(anon_vma_prepare(vma)))
2484 * vma->vm_start/vm_end cannot change under us because the caller
2485 * is required to hold the mmap_lock in read mode. We need the
2486 * anon_vma lock to serialize against concurrent expand_stacks.
2488 anon_vma_lock_write(vma->anon_vma);
2490 /* Somebody else might have raced and expanded it already */
2491 if (address > vma->vm_end) {
2492 unsigned long size, grow;
2494 size = address - vma->vm_start;
2495 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2498 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2499 error = acct_stack_growth(vma, size, grow);
2502 * vma_gap_update() doesn't support concurrent
2503 * updates, but we only hold a shared mmap_lock
2504 * lock here, so we need to protect against
2505 * concurrent vma expansions.
2506 * anon_vma_lock_write() doesn't help here, as
2507 * we don't guarantee that all growable vmas
2508 * in a mm share the same root anon vma.
2509 * So, we reuse mm->page_table_lock to guard
2510 * against concurrent vma expansions.
2512 spin_lock(&mm->page_table_lock);
2513 if (vma->vm_flags & VM_LOCKED)
2514 mm->locked_vm += grow;
2515 vm_stat_account(mm, vma->vm_flags, grow);
2516 anon_vma_interval_tree_pre_update_vma(vma);
2517 vma->vm_end = address;
2518 anon_vma_interval_tree_post_update_vma(vma);
2520 vma_gap_update(vma->vm_next);
2522 mm->highest_vm_end = vm_end_gap(vma);
2523 spin_unlock(&mm->page_table_lock);
2525 perf_event_mmap(vma);
2529 anon_vma_unlock_write(vma->anon_vma);
2530 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2534 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2537 * vma is the first one with address < vma->vm_start. Have to extend vma.
2539 int expand_downwards(struct vm_area_struct *vma,
2540 unsigned long address)
2542 struct mm_struct *mm = vma->vm_mm;
2543 struct vm_area_struct *prev;
2546 address &= PAGE_MASK;
2547 if (address < mmap_min_addr)
2550 /* Enforce stack_guard_gap */
2551 prev = vma->vm_prev;
2552 /* Check that both stack segments have the same anon_vma? */
2553 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2554 vma_is_accessible(prev)) {
2555 if (address - prev->vm_end < stack_guard_gap)
2559 /* We must make sure the anon_vma is allocated. */
2560 if (unlikely(anon_vma_prepare(vma)))
2564 * vma->vm_start/vm_end cannot change under us because the caller
2565 * is required to hold the mmap_lock in read mode. We need the
2566 * anon_vma lock to serialize against concurrent expand_stacks.
2568 anon_vma_lock_write(vma->anon_vma);
2570 /* Somebody else might have raced and expanded it already */
2571 if (address < vma->vm_start) {
2572 unsigned long size, grow;
2574 size = vma->vm_end - address;
2575 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2578 if (grow <= vma->vm_pgoff) {
2579 error = acct_stack_growth(vma, size, grow);
2582 * vma_gap_update() doesn't support concurrent
2583 * updates, but we only hold a shared mmap_lock
2584 * lock here, so we need to protect against
2585 * concurrent vma expansions.
2586 * anon_vma_lock_write() doesn't help here, as
2587 * we don't guarantee that all growable vmas
2588 * in a mm share the same root anon vma.
2589 * So, we reuse mm->page_table_lock to guard
2590 * against concurrent vma expansions.
2592 spin_lock(&mm->page_table_lock);
2593 if (vma->vm_flags & VM_LOCKED)
2594 mm->locked_vm += grow;
2595 vm_stat_account(mm, vma->vm_flags, grow);
2596 anon_vma_interval_tree_pre_update_vma(vma);
2597 vma->vm_start = address;
2598 vma->vm_pgoff -= grow;
2599 anon_vma_interval_tree_post_update_vma(vma);
2600 vma_gap_update(vma);
2601 spin_unlock(&mm->page_table_lock);
2603 perf_event_mmap(vma);
2607 anon_vma_unlock_write(vma->anon_vma);
2608 khugepaged_enter_vma_merge(vma, vma->vm_flags);
2613 /* enforced gap between the expanding stack and other mappings. */
2614 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2616 static int __init cmdline_parse_stack_guard_gap(char *p)
2621 val = simple_strtoul(p, &endptr, 10);
2623 stack_guard_gap = val << PAGE_SHIFT;
2627 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2629 #ifdef CONFIG_STACK_GROWSUP
2630 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2632 return expand_upwards(vma, address);
2635 struct vm_area_struct *
2636 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2638 struct vm_area_struct *vma, *prev;
2641 vma = find_vma_prev(mm, addr, &prev);
2642 if (vma && (vma->vm_start <= addr))
2644 /* don't alter vm_end if the coredump is running */
2645 if (!prev || expand_stack(prev, addr))
2647 if (prev->vm_flags & VM_LOCKED)
2648 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2652 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2654 return expand_downwards(vma, address);
2657 struct vm_area_struct *
2658 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2660 struct vm_area_struct *vma;
2661 unsigned long start;
2664 vma = find_vma(mm, addr);
2667 if (vma->vm_start <= addr)
2669 if (!(vma->vm_flags & VM_GROWSDOWN))
2671 start = vma->vm_start;
2672 if (expand_stack(vma, addr))
2674 if (vma->vm_flags & VM_LOCKED)
2675 populate_vma_page_range(vma, addr, start, NULL);
2680 EXPORT_SYMBOL_GPL(find_extend_vma);
2683 * Ok - we have the memory areas we should free on the vma list,
2684 * so release them, and do the vma updates.
2686 * Called with the mm semaphore held.
2688 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2690 unsigned long nr_accounted = 0;
2692 /* Update high watermark before we lower total_vm */
2693 update_hiwater_vm(mm);
2695 long nrpages = vma_pages(vma);
2697 if (vma->vm_flags & VM_ACCOUNT)
2698 nr_accounted += nrpages;
2699 vm_stat_account(mm, vma->vm_flags, -nrpages);
2700 vma = remove_vma(vma);
2702 vm_unacct_memory(nr_accounted);
2707 * Get rid of page table information in the indicated region.
2709 * Called with the mm semaphore held.
2711 static void unmap_region(struct mm_struct *mm,
2712 struct vm_area_struct *vma, struct vm_area_struct *prev,
2713 unsigned long start, unsigned long end)
2715 struct vm_area_struct *next = vma_next(mm, prev);
2716 struct mmu_gather tlb;
2719 tlb_gather_mmu(&tlb, mm, start, end);
2720 update_hiwater_rss(mm);
2721 unmap_vmas(&tlb, vma, start, end);
2722 free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2723 next ? next->vm_start : USER_PGTABLES_CEILING);
2724 tlb_finish_mmu(&tlb, start, end);
2728 * Create a list of vma's touched by the unmap, removing them from the mm's
2729 * vma list as we go..
2732 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2733 struct vm_area_struct *prev, unsigned long end)
2735 struct vm_area_struct **insertion_point;
2736 struct vm_area_struct *tail_vma = NULL;
2738 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2739 vma->vm_prev = NULL;
2741 vma_rb_erase(vma, &mm->mm_rb);
2745 } while (vma && vma->vm_start < end);
2746 *insertion_point = vma;
2748 vma->vm_prev = prev;
2749 vma_gap_update(vma);
2751 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2752 tail_vma->vm_next = NULL;
2754 /* Kill the cache */
2755 vmacache_invalidate(mm);
2758 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2759 * VM_GROWSUP VMA. Such VMAs can change their size under
2760 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2762 if (vma && (vma->vm_flags & VM_GROWSDOWN))
2764 if (prev && (prev->vm_flags & VM_GROWSUP))
2770 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2771 * has already been checked or doesn't make sense to fail.
2773 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2774 unsigned long addr, int new_below)
2776 struct vm_area_struct *new;
2779 if (vma->vm_ops && vma->vm_ops->split) {
2780 err = vma->vm_ops->split(vma, addr);
2785 new = vm_area_dup(vma);
2792 new->vm_start = addr;
2793 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2796 err = vma_dup_policy(vma, new);
2800 err = anon_vma_clone(new, vma);
2805 get_file(new->vm_file);
2807 if (new->vm_ops && new->vm_ops->open)
2808 new->vm_ops->open(new);
2811 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2812 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2814 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2820 /* Clean everything up if vma_adjust failed. */
2821 if (new->vm_ops && new->vm_ops->close)
2822 new->vm_ops->close(new);
2825 unlink_anon_vmas(new);
2827 mpol_put(vma_policy(new));
2834 * Split a vma into two pieces at address 'addr', a new vma is allocated
2835 * either for the first part or the tail.
2837 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2838 unsigned long addr, int new_below)
2840 if (mm->map_count >= sysctl_max_map_count)
2843 return __split_vma(mm, vma, addr, new_below);
2846 /* Munmap is split into 2 main parts -- this part which finds
2847 * what needs doing, and the areas themselves, which do the
2848 * work. This now handles partial unmappings.
2849 * Jeremy Fitzhardinge <jeremy@goop.org>
2851 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2852 struct list_head *uf, bool downgrade)
2855 struct vm_area_struct *vma, *prev, *last;
2857 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2860 len = PAGE_ALIGN(len);
2866 * arch_unmap() might do unmaps itself. It must be called
2867 * and finish any rbtree manipulation before this code
2868 * runs and also starts to manipulate the rbtree.
2870 arch_unmap(mm, start, end);
2872 /* Find the first overlapping VMA */
2873 vma = find_vma(mm, start);
2876 prev = vma->vm_prev;
2877 /* we have start < vma->vm_end */
2879 /* if it doesn't overlap, we have nothing.. */
2880 if (vma->vm_start >= end)
2884 * If we need to split any vma, do it now to save pain later.
2886 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2887 * unmapped vm_area_struct will remain in use: so lower split_vma
2888 * places tmp vma above, and higher split_vma places tmp vma below.
2890 if (start > vma->vm_start) {
2894 * Make sure that map_count on return from munmap() will
2895 * not exceed its limit; but let map_count go just above
2896 * its limit temporarily, to help free resources as expected.
2898 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2901 error = __split_vma(mm, vma, start, 0);
2907 /* Does it split the last one? */
2908 last = find_vma(mm, end);
2909 if (last && end > last->vm_start) {
2910 int error = __split_vma(mm, last, end, 1);
2914 vma = vma_next(mm, prev);
2918 * If userfaultfd_unmap_prep returns an error the vmas
2919 * will remain splitted, but userland will get a
2920 * highly unexpected error anyway. This is no
2921 * different than the case where the first of the two
2922 * __split_vma fails, but we don't undo the first
2923 * split, despite we could. This is unlikely enough
2924 * failure that it's not worth optimizing it for.
2926 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2932 * unlock any mlock()ed ranges before detaching vmas
2934 if (mm->locked_vm) {
2935 struct vm_area_struct *tmp = vma;
2936 while (tmp && tmp->vm_start < end) {
2937 if (tmp->vm_flags & VM_LOCKED) {
2938 mm->locked_vm -= vma_pages(tmp);
2939 munlock_vma_pages_all(tmp);
2946 /* Detach vmas from rbtree */
2947 if (!detach_vmas_to_be_unmapped(mm, vma, prev, end))
2951 mmap_write_downgrade(mm);
2953 unmap_region(mm, vma, prev, start, end);
2955 /* Fix up all other VM information */
2956 remove_vma_list(mm, vma);
2958 return downgrade ? 1 : 0;
2961 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2962 struct list_head *uf)
2964 return __do_munmap(mm, start, len, uf, false);
2967 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2970 struct mm_struct *mm = current->mm;
2973 if (mmap_write_lock_killable(mm))
2976 ret = __do_munmap(mm, start, len, &uf, downgrade);
2978 * Returning 1 indicates mmap_lock is downgraded.
2979 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2980 * it to 0 before return.
2983 mmap_read_unlock(mm);
2986 mmap_write_unlock(mm);
2988 userfaultfd_unmap_complete(mm, &uf);
2992 int vm_munmap(unsigned long start, size_t len)
2994 return __vm_munmap(start, len, false);
2996 EXPORT_SYMBOL(vm_munmap);
2998 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
3000 addr = untagged_addr(addr);
3001 profile_munmap(addr);
3002 return __vm_munmap(addr, len, true);
3007 * Emulation of deprecated remap_file_pages() syscall.
3009 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
3010 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
3013 struct mm_struct *mm = current->mm;
3014 struct vm_area_struct *vma;
3015 unsigned long populate = 0;
3016 unsigned long ret = -EINVAL;
3019 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
3020 current->comm, current->pid);
3024 start = start & PAGE_MASK;
3025 size = size & PAGE_MASK;
3027 if (start + size <= start)
3030 /* Does pgoff wrap? */
3031 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
3034 if (mmap_write_lock_killable(mm))
3037 vma = find_vma(mm, start);
3039 if (!vma || !(vma->vm_flags & VM_SHARED))
3042 if (start < vma->vm_start)
3045 if (start + size > vma->vm_end) {
3046 struct vm_area_struct *next;
3048 for (next = vma->vm_next; next; next = next->vm_next) {
3049 /* hole between vmas ? */
3050 if (next->vm_start != next->vm_prev->vm_end)
3053 if (next->vm_file != vma->vm_file)
3056 if (next->vm_flags != vma->vm_flags)
3059 if (start + size <= next->vm_end)
3067 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3068 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3069 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3071 flags &= MAP_NONBLOCK;
3072 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3073 if (vma->vm_flags & VM_LOCKED) {
3074 struct vm_area_struct *tmp;
3075 flags |= MAP_LOCKED;
3077 /* drop PG_Mlocked flag for over-mapped range */
3078 for (tmp = vma; tmp->vm_start >= start + size;
3079 tmp = tmp->vm_next) {
3081 * Split pmd and munlock page on the border
3084 vma_adjust_trans_huge(tmp, start, start + size, 0);
3086 munlock_vma_pages_range(tmp,
3087 max(tmp->vm_start, start),
3088 min(tmp->vm_end, start + size));
3092 file = get_file(vma->vm_file);
3093 ret = do_mmap(vma->vm_file, start, size,
3094 prot, flags, pgoff, &populate, NULL);
3097 mmap_write_unlock(mm);
3099 mm_populate(ret, populate);
3100 if (!IS_ERR_VALUE(ret))
3106 * this is really a simplified "do_mmap". it only handles
3107 * anonymous maps. eventually we may be able to do some
3108 * brk-specific accounting here.
3110 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
3112 struct mm_struct *mm = current->mm;
3113 struct vm_area_struct *vma, *prev;
3114 struct rb_node **rb_link, *rb_parent;
3115 pgoff_t pgoff = addr >> PAGE_SHIFT;
3117 unsigned long mapped_addr;
3119 /* Until we need other flags, refuse anything except VM_EXEC. */
3120 if ((flags & (~VM_EXEC)) != 0)
3122 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3124 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
3125 if (IS_ERR_VALUE(mapped_addr))
3128 error = mlock_future_check(mm, mm->def_flags, len);
3132 /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
3133 if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
3136 /* Check against address space limits *after* clearing old maps... */
3137 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3140 if (mm->map_count > sysctl_max_map_count)
3143 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3146 /* Can we just expand an old private anonymous mapping? */
3147 vma = vma_merge(mm, prev, addr, addr + len, flags,
3148 NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
3153 * create a vma struct for an anonymous mapping
3155 vma = vm_area_alloc(mm);
3157 vm_unacct_memory(len >> PAGE_SHIFT);
3161 vma_set_anonymous(vma);
3162 vma->vm_start = addr;
3163 vma->vm_end = addr + len;
3164 vma->vm_pgoff = pgoff;
3165 vma->vm_flags = flags;
3166 vma->vm_page_prot = vm_get_page_prot(flags);
3167 vma_link(mm, vma, prev, rb_link, rb_parent);
3169 perf_event_mmap(vma);
3170 mm->total_vm += len >> PAGE_SHIFT;
3171 mm->data_vm += len >> PAGE_SHIFT;
3172 if (flags & VM_LOCKED)
3173 mm->locked_vm += (len >> PAGE_SHIFT);
3174 vma->vm_flags |= VM_SOFTDIRTY;
3178 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3180 struct mm_struct *mm = current->mm;
3186 len = PAGE_ALIGN(request);
3192 if (mmap_write_lock_killable(mm))
3195 ret = do_brk_flags(addr, len, flags, &uf);
3196 populate = ((mm->def_flags & VM_LOCKED) != 0);
3197 mmap_write_unlock(mm);
3198 userfaultfd_unmap_complete(mm, &uf);
3199 if (populate && !ret)
3200 mm_populate(addr, len);
3203 EXPORT_SYMBOL(vm_brk_flags);
3205 int vm_brk(unsigned long addr, unsigned long len)
3207 return vm_brk_flags(addr, len, 0);
3209 EXPORT_SYMBOL(vm_brk);
3211 /* Release all mmaps. */
3212 void exit_mmap(struct mm_struct *mm)
3214 struct mmu_gather tlb;
3215 struct vm_area_struct *vma;
3216 unsigned long nr_accounted = 0;
3218 /* mm's last user has gone, and its about to be pulled down */
3219 mmu_notifier_release(mm);
3221 if (unlikely(mm_is_oom_victim(mm))) {
3223 * Manually reap the mm to free as much memory as possible.
3224 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3225 * this mm from further consideration. Taking mm->mmap_lock for
3226 * write after setting MMF_OOM_SKIP will guarantee that the oom
3227 * reaper will not run on this mm again after mmap_lock is
3230 * Nothing can be holding mm->mmap_lock here and the above call
3231 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3232 * __oom_reap_task_mm() will not block.
3234 * This needs to be done before calling munlock_vma_pages_all(),
3235 * which clears VM_LOCKED, otherwise the oom reaper cannot
3238 (void)__oom_reap_task_mm(mm);
3240 set_bit(MMF_OOM_SKIP, &mm->flags);
3241 mmap_write_lock(mm);
3242 mmap_write_unlock(mm);
3245 if (mm->locked_vm) {
3248 if (vma->vm_flags & VM_LOCKED)
3249 munlock_vma_pages_all(vma);
3257 if (!vma) /* Can happen if dup_mmap() received an OOM */
3262 tlb_gather_mmu(&tlb, mm, 0, -1);
3263 /* update_hiwater_rss(mm) here? but nobody should be looking */
3264 /* Use -1 here to ensure all VMAs in the mm are unmapped */
3265 unmap_vmas(&tlb, vma, 0, -1);
3266 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3267 tlb_finish_mmu(&tlb, 0, -1);
3270 * Walk the list again, actually closing and freeing it,
3271 * with preemption enabled, without holding any MM locks.
3274 if (vma->vm_flags & VM_ACCOUNT)
3275 nr_accounted += vma_pages(vma);
3276 vma = remove_vma(vma);
3279 vm_unacct_memory(nr_accounted);
3282 /* Insert vm structure into process list sorted by address
3283 * and into the inode's i_mmap tree. If vm_file is non-NULL
3284 * then i_mmap_rwsem is taken here.
3286 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3288 struct vm_area_struct *prev;
3289 struct rb_node **rb_link, *rb_parent;
3291 if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3292 &prev, &rb_link, &rb_parent))
3294 if ((vma->vm_flags & VM_ACCOUNT) &&
3295 security_vm_enough_memory_mm(mm, vma_pages(vma)))
3299 * The vm_pgoff of a purely anonymous vma should be irrelevant
3300 * until its first write fault, when page's anon_vma and index
3301 * are set. But now set the vm_pgoff it will almost certainly
3302 * end up with (unless mremap moves it elsewhere before that
3303 * first wfault), so /proc/pid/maps tells a consistent story.
3305 * By setting it to reflect the virtual start address of the
3306 * vma, merges and splits can happen in a seamless way, just
3307 * using the existing file pgoff checks and manipulations.
3308 * Similarly in do_mmap and in do_brk_flags.
3310 if (vma_is_anonymous(vma)) {
3311 BUG_ON(vma->anon_vma);
3312 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3315 vma_link(mm, vma, prev, rb_link, rb_parent);
3320 * Copy the vma structure to a new location in the same mm,
3321 * prior to moving page table entries, to effect an mremap move.
3323 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3324 unsigned long addr, unsigned long len, pgoff_t pgoff,
3325 bool *need_rmap_locks)
3327 struct vm_area_struct *vma = *vmap;
3328 unsigned long vma_start = vma->vm_start;
3329 struct mm_struct *mm = vma->vm_mm;
3330 struct vm_area_struct *new_vma, *prev;
3331 struct rb_node **rb_link, *rb_parent;
3332 bool faulted_in_anon_vma = true;
3335 * If anonymous vma has not yet been faulted, update new pgoff
3336 * to match new location, to increase its chance of merging.
3338 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3339 pgoff = addr >> PAGE_SHIFT;
3340 faulted_in_anon_vma = false;
3343 if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3344 return NULL; /* should never get here */
3345 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3346 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3347 vma->vm_userfaultfd_ctx);
3350 * Source vma may have been merged into new_vma
3352 if (unlikely(vma_start >= new_vma->vm_start &&
3353 vma_start < new_vma->vm_end)) {
3355 * The only way we can get a vma_merge with
3356 * self during an mremap is if the vma hasn't
3357 * been faulted in yet and we were allowed to
3358 * reset the dst vma->vm_pgoff to the
3359 * destination address of the mremap to allow
3360 * the merge to happen. mremap must change the
3361 * vm_pgoff linearity between src and dst vmas
3362 * (in turn preventing a vma_merge) to be
3363 * safe. It is only safe to keep the vm_pgoff
3364 * linear if there are no pages mapped yet.
3366 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3367 *vmap = vma = new_vma;
3369 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3371 new_vma = vm_area_dup(vma);
3374 new_vma->vm_start = addr;
3375 new_vma->vm_end = addr + len;
3376 new_vma->vm_pgoff = pgoff;
3377 if (vma_dup_policy(vma, new_vma))
3379 if (anon_vma_clone(new_vma, vma))
3380 goto out_free_mempol;
3381 if (new_vma->vm_file)
3382 get_file(new_vma->vm_file);
3383 if (new_vma->vm_ops && new_vma->vm_ops->open)
3384 new_vma->vm_ops->open(new_vma);
3385 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3386 *need_rmap_locks = false;
3391 mpol_put(vma_policy(new_vma));
3393 vm_area_free(new_vma);
3399 * Return true if the calling process may expand its vm space by the passed
3402 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3404 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3407 if (is_data_mapping(flags) &&
3408 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3409 /* Workaround for Valgrind */
3410 if (rlimit(RLIMIT_DATA) == 0 &&
3411 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3414 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3415 current->comm, current->pid,
3416 (mm->data_vm + npages) << PAGE_SHIFT,
3417 rlimit(RLIMIT_DATA),
3418 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3420 if (!ignore_rlimit_data)
3427 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3429 mm->total_vm += npages;
3431 if (is_exec_mapping(flags))
3432 mm->exec_vm += npages;
3433 else if (is_stack_mapping(flags))
3434 mm->stack_vm += npages;
3435 else if (is_data_mapping(flags))
3436 mm->data_vm += npages;
3439 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3442 * Having a close hook prevents vma merging regardless of flags.
3444 static void special_mapping_close(struct vm_area_struct *vma)
3448 static const char *special_mapping_name(struct vm_area_struct *vma)
3450 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3453 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3455 struct vm_special_mapping *sm = new_vma->vm_private_data;
3457 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3461 return sm->mremap(sm, new_vma);
3466 static const struct vm_operations_struct special_mapping_vmops = {
3467 .close = special_mapping_close,
3468 .fault = special_mapping_fault,
3469 .mremap = special_mapping_mremap,
3470 .name = special_mapping_name,
3471 /* vDSO code relies that VVAR can't be accessed remotely */
3475 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3476 .close = special_mapping_close,
3477 .fault = special_mapping_fault,
3480 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3482 struct vm_area_struct *vma = vmf->vma;
3484 struct page **pages;
3486 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3487 pages = vma->vm_private_data;
3489 struct vm_special_mapping *sm = vma->vm_private_data;
3492 return sm->fault(sm, vmf->vma, vmf);
3497 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3501 struct page *page = *pages;
3507 return VM_FAULT_SIGBUS;
3510 static struct vm_area_struct *__install_special_mapping(
3511 struct mm_struct *mm,
3512 unsigned long addr, unsigned long len,
3513 unsigned long vm_flags, void *priv,
3514 const struct vm_operations_struct *ops)
3517 struct vm_area_struct *vma;
3519 vma = vm_area_alloc(mm);
3520 if (unlikely(vma == NULL))
3521 return ERR_PTR(-ENOMEM);
3523 vma->vm_start = addr;
3524 vma->vm_end = addr + len;
3526 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3527 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3530 vma->vm_private_data = priv;
3532 ret = insert_vm_struct(mm, vma);
3536 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3538 perf_event_mmap(vma);
3544 return ERR_PTR(ret);
3547 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3548 const struct vm_special_mapping *sm)
3550 return vma->vm_private_data == sm &&
3551 (vma->vm_ops == &special_mapping_vmops ||
3552 vma->vm_ops == &legacy_special_mapping_vmops);
3556 * Called with mm->mmap_lock held for writing.
3557 * Insert a new vma covering the given region, with the given flags.
3558 * Its pages are supplied by the given array of struct page *.
3559 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3560 * The region past the last page supplied will always produce SIGBUS.
3561 * The array pointer and the pages it points to are assumed to stay alive
3562 * for as long as this mapping might exist.
3564 struct vm_area_struct *_install_special_mapping(
3565 struct mm_struct *mm,
3566 unsigned long addr, unsigned long len,
3567 unsigned long vm_flags, const struct vm_special_mapping *spec)
3569 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3570 &special_mapping_vmops);
3573 int install_special_mapping(struct mm_struct *mm,
3574 unsigned long addr, unsigned long len,
3575 unsigned long vm_flags, struct page **pages)
3577 struct vm_area_struct *vma = __install_special_mapping(
3578 mm, addr, len, vm_flags, (void *)pages,
3579 &legacy_special_mapping_vmops);
3581 return PTR_ERR_OR_ZERO(vma);
3584 static DEFINE_MUTEX(mm_all_locks_mutex);
3586 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3588 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3590 * The LSB of head.next can't change from under us
3591 * because we hold the mm_all_locks_mutex.
3593 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3595 * We can safely modify head.next after taking the
3596 * anon_vma->root->rwsem. If some other vma in this mm shares
3597 * the same anon_vma we won't take it again.
3599 * No need of atomic instructions here, head.next
3600 * can't change from under us thanks to the
3601 * anon_vma->root->rwsem.
3603 if (__test_and_set_bit(0, (unsigned long *)
3604 &anon_vma->root->rb_root.rb_root.rb_node))
3609 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3611 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3613 * AS_MM_ALL_LOCKS can't change from under us because
3614 * we hold the mm_all_locks_mutex.
3616 * Operations on ->flags have to be atomic because
3617 * even if AS_MM_ALL_LOCKS is stable thanks to the
3618 * mm_all_locks_mutex, there may be other cpus
3619 * changing other bitflags in parallel to us.
3621 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3623 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3628 * This operation locks against the VM for all pte/vma/mm related
3629 * operations that could ever happen on a certain mm. This includes
3630 * vmtruncate, try_to_unmap, and all page faults.
3632 * The caller must take the mmap_lock in write mode before calling
3633 * mm_take_all_locks(). The caller isn't allowed to release the
3634 * mmap_lock until mm_drop_all_locks() returns.
3636 * mmap_lock in write mode is required in order to block all operations
3637 * that could modify pagetables and free pages without need of
3638 * altering the vma layout. It's also needed in write mode to avoid new
3639 * anon_vmas to be associated with existing vmas.
3641 * A single task can't take more than one mm_take_all_locks() in a row
3642 * or it would deadlock.
3644 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3645 * mapping->flags avoid to take the same lock twice, if more than one
3646 * vma in this mm is backed by the same anon_vma or address_space.
3648 * We take locks in following order, accordingly to comment at beginning
3650 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3652 * - all i_mmap_rwsem locks;
3653 * - all anon_vma->rwseml
3655 * We can take all locks within these types randomly because the VM code
3656 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3657 * mm_all_locks_mutex.
3659 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3660 * that may have to take thousand of locks.
3662 * mm_take_all_locks() can fail if it's interrupted by signals.
3664 int mm_take_all_locks(struct mm_struct *mm)
3666 struct vm_area_struct *vma;
3667 struct anon_vma_chain *avc;
3669 BUG_ON(mmap_read_trylock(mm));
3671 mutex_lock(&mm_all_locks_mutex);
3673 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3674 if (signal_pending(current))
3676 if (vma->vm_file && vma->vm_file->f_mapping &&
3677 is_vm_hugetlb_page(vma))
3678 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3681 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3682 if (signal_pending(current))
3684 if (vma->vm_file && vma->vm_file->f_mapping &&
3685 !is_vm_hugetlb_page(vma))
3686 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3689 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3690 if (signal_pending(current))
3693 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3694 vm_lock_anon_vma(mm, avc->anon_vma);
3700 mm_drop_all_locks(mm);
3704 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3706 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3708 * The LSB of head.next can't change to 0 from under
3709 * us because we hold the mm_all_locks_mutex.
3711 * We must however clear the bitflag before unlocking
3712 * the vma so the users using the anon_vma->rb_root will
3713 * never see our bitflag.
3715 * No need of atomic instructions here, head.next
3716 * can't change from under us until we release the
3717 * anon_vma->root->rwsem.
3719 if (!__test_and_clear_bit(0, (unsigned long *)
3720 &anon_vma->root->rb_root.rb_root.rb_node))
3722 anon_vma_unlock_write(anon_vma);
3726 static void vm_unlock_mapping(struct address_space *mapping)
3728 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3730 * AS_MM_ALL_LOCKS can't change to 0 from under us
3731 * because we hold the mm_all_locks_mutex.
3733 i_mmap_unlock_write(mapping);
3734 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3741 * The mmap_lock cannot be released by the caller until
3742 * mm_drop_all_locks() returns.
3744 void mm_drop_all_locks(struct mm_struct *mm)
3746 struct vm_area_struct *vma;
3747 struct anon_vma_chain *avc;
3749 BUG_ON(mmap_read_trylock(mm));
3750 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3752 for (vma = mm->mmap; vma; vma = vma->vm_next) {
3754 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3755 vm_unlock_anon_vma(avc->anon_vma);
3756 if (vma->vm_file && vma->vm_file->f_mapping)
3757 vm_unlock_mapping(vma->vm_file->f_mapping);
3760 mutex_unlock(&mm_all_locks_mutex);
3764 * initialise the percpu counter for VM
3766 void __init mmap_init(void)
3770 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3775 * Initialise sysctl_user_reserve_kbytes.
3777 * This is intended to prevent a user from starting a single memory hogging
3778 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3781 * The default value is min(3% of free memory, 128MB)
3782 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3784 static int init_user_reserve(void)
3786 unsigned long free_kbytes;
3788 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3790 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3793 subsys_initcall(init_user_reserve);
3796 * Initialise sysctl_admin_reserve_kbytes.
3798 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3799 * to log in and kill a memory hogging process.
3801 * Systems with more than 256MB will reserve 8MB, enough to recover
3802 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3803 * only reserve 3% of free pages by default.
3805 static int init_admin_reserve(void)
3807 unsigned long free_kbytes;
3809 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3811 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3814 subsys_initcall(init_admin_reserve);
3817 * Reinititalise user and admin reserves if memory is added or removed.
3819 * The default user reserve max is 128MB, and the default max for the
3820 * admin reserve is 8MB. These are usually, but not always, enough to
3821 * enable recovery from a memory hogging process using login/sshd, a shell,
3822 * and tools like top. It may make sense to increase or even disable the
3823 * reserve depending on the existence of swap or variations in the recovery
3824 * tools. So, the admin may have changed them.
3826 * If memory is added and the reserves have been eliminated or increased above
3827 * the default max, then we'll trust the admin.
3829 * If memory is removed and there isn't enough free memory, then we
3830 * need to reset the reserves.
3832 * Otherwise keep the reserve set by the admin.
3834 static int reserve_mem_notifier(struct notifier_block *nb,
3835 unsigned long action, void *data)
3837 unsigned long tmp, free_kbytes;
3841 /* Default max is 128MB. Leave alone if modified by operator. */
3842 tmp = sysctl_user_reserve_kbytes;
3843 if (0 < tmp && tmp < (1UL << 17))
3844 init_user_reserve();
3846 /* Default max is 8MB. Leave alone if modified by operator. */
3847 tmp = sysctl_admin_reserve_kbytes;
3848 if (0 < tmp && tmp < (1UL << 13))
3849 init_admin_reserve();
3853 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3855 if (sysctl_user_reserve_kbytes > free_kbytes) {
3856 init_user_reserve();
3857 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3858 sysctl_user_reserve_kbytes);
3861 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3862 init_admin_reserve();
3863 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3864 sysctl_admin_reserve_kbytes);
3873 static struct notifier_block reserve_mem_nb = {
3874 .notifier_call = reserve_mem_notifier,
3877 static int __meminit init_reserve_notifier(void)
3879 if (register_hotmemory_notifier(&reserve_mem_nb))
3880 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3884 subsys_initcall(init_reserve_notifier);