1 // SPDX-License-Identifier: GPL-2.0-only
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12 #include <linux/kernel.h>
13 #include <linux/slab.h>
14 #include <linux/backing-dev.h>
16 #include <linux/mm_inline.h>
17 #include <linux/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/notifier.h>
42 #include <linux/memory.h>
43 #include <linux/printk.h>
44 #include <linux/userfaultfd_k.h>
45 #include <linux/moduleparam.h>
46 #include <linux/pkeys.h>
47 #include <linux/oom.h>
48 #include <linux/sched/mm.h>
49 #include <linux/ksm.h>
51 #include <linux/uaccess.h>
52 #include <asm/cacheflush.h>
54 #include <asm/mmu_context.h>
56 #define CREATE_TRACE_POINTS
57 #include <trace/events/mmap.h>
61 #ifndef arch_mmap_check
62 #define arch_mmap_check(addr, len, flags) (0)
65 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
66 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
67 int mmap_rnd_bits_max __ro_after_init = CONFIG_ARCH_MMAP_RND_BITS_MAX;
68 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
70 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
71 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
72 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
73 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
76 static bool ignore_rlimit_data;
77 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
79 static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
80 struct vm_area_struct *vma, struct vm_area_struct *prev,
81 struct vm_area_struct *next, unsigned long start,
82 unsigned long end, unsigned long tree_end, bool mm_wr_locked);
84 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
86 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
89 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
90 void vma_set_page_prot(struct vm_area_struct *vma)
92 unsigned long vm_flags = vma->vm_flags;
93 pgprot_t vm_page_prot;
95 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
96 if (vma_wants_writenotify(vma, vm_page_prot)) {
97 vm_flags &= ~VM_SHARED;
98 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
100 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
101 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
105 * Requires inode->i_mapping->i_mmap_rwsem
107 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
108 struct address_space *mapping)
110 if (vma_is_shared_maywrite(vma))
111 mapping_unmap_writable(mapping);
113 flush_dcache_mmap_lock(mapping);
114 vma_interval_tree_remove(vma, &mapping->i_mmap);
115 flush_dcache_mmap_unlock(mapping);
119 * Unlink a file-based vm structure from its interval tree, to hide
120 * vma from rmap and vmtruncate before freeing its page tables.
122 void unlink_file_vma(struct vm_area_struct *vma)
124 struct file *file = vma->vm_file;
127 struct address_space *mapping = file->f_mapping;
128 i_mmap_lock_write(mapping);
129 __remove_shared_vm_struct(vma, mapping);
130 i_mmap_unlock_write(mapping);
135 * Close a vm structure and free it.
137 static void remove_vma(struct vm_area_struct *vma, bool unreachable)
140 if (vma->vm_ops && vma->vm_ops->close)
141 vma->vm_ops->close(vma);
144 mpol_put(vma_policy(vma));
151 static inline struct vm_area_struct *vma_prev_limit(struct vma_iterator *vmi,
154 return mas_prev(&vmi->mas, min);
158 * check_brk_limits() - Use platform specific check of range & verify mlock
160 * @addr: The address to check
161 * @len: The size of increase.
163 * Return: 0 on success.
165 static int check_brk_limits(unsigned long addr, unsigned long len)
167 unsigned long mapped_addr;
169 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
170 if (IS_ERR_VALUE(mapped_addr))
173 return mlock_future_ok(current->mm, current->mm->def_flags, len)
176 static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *brkvma,
177 unsigned long addr, unsigned long request, unsigned long flags);
178 SYSCALL_DEFINE1(brk, unsigned long, brk)
180 unsigned long newbrk, oldbrk, origbrk;
181 struct mm_struct *mm = current->mm;
182 struct vm_area_struct *brkvma, *next = NULL;
183 unsigned long min_brk;
184 bool populate = false;
186 struct vma_iterator vmi;
188 if (mmap_write_lock_killable(mm))
193 #ifdef CONFIG_COMPAT_BRK
195 * CONFIG_COMPAT_BRK can still be overridden by setting
196 * randomize_va_space to 2, which will still cause mm->start_brk
197 * to be arbitrarily shifted
199 if (current->brk_randomized)
200 min_brk = mm->start_brk;
202 min_brk = mm->end_data;
204 min_brk = mm->start_brk;
210 * Check against rlimit here. If this check is done later after the test
211 * of oldbrk with newbrk then it can escape the test and let the data
212 * segment grow beyond its set limit the in case where the limit is
213 * not page aligned -Ram Gupta
215 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
216 mm->end_data, mm->start_data))
219 newbrk = PAGE_ALIGN(brk);
220 oldbrk = PAGE_ALIGN(mm->brk);
221 if (oldbrk == newbrk) {
226 /* Always allow shrinking brk. */
227 if (brk <= mm->brk) {
228 /* Search one past newbrk */
229 vma_iter_init(&vmi, mm, newbrk);
230 brkvma = vma_find(&vmi, oldbrk);
231 if (!brkvma || brkvma->vm_start >= oldbrk)
232 goto out; /* mapping intersects with an existing non-brk vma. */
234 * mm->brk must be protected by write mmap_lock.
235 * do_vma_munmap() will drop the lock on success, so update it
236 * before calling do_vma_munmap().
239 if (do_vma_munmap(&vmi, brkvma, newbrk, oldbrk, &uf, true))
242 goto success_unlocked;
245 if (check_brk_limits(oldbrk, newbrk - oldbrk))
249 * Only check if the next VMA is within the stack_guard_gap of the
252 vma_iter_init(&vmi, mm, oldbrk);
253 next = vma_find(&vmi, newbrk + PAGE_SIZE + stack_guard_gap);
254 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
257 brkvma = vma_prev_limit(&vmi, mm->start_brk);
258 /* Ok, looks good - let it rip. */
259 if (do_brk_flags(&vmi, brkvma, oldbrk, newbrk - oldbrk, 0) < 0)
263 if (mm->def_flags & VM_LOCKED)
267 mmap_write_unlock(mm);
269 userfaultfd_unmap_complete(mm, &uf);
271 mm_populate(oldbrk, newbrk - oldbrk);
276 mmap_write_unlock(mm);
280 #if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
281 static void validate_mm(struct mm_struct *mm)
285 struct vm_area_struct *vma;
286 VMA_ITERATOR(vmi, mm, 0);
288 mt_validate(&mm->mm_mt);
289 for_each_vma(vmi, vma) {
290 #ifdef CONFIG_DEBUG_VM_RB
291 struct anon_vma *anon_vma = vma->anon_vma;
292 struct anon_vma_chain *avc;
294 unsigned long vmi_start, vmi_end;
297 vmi_start = vma_iter_addr(&vmi);
298 vmi_end = vma_iter_end(&vmi);
299 if (VM_WARN_ON_ONCE_MM(vma->vm_end != vmi_end, mm))
302 if (VM_WARN_ON_ONCE_MM(vma->vm_start != vmi_start, mm))
306 pr_emerg("issue in %s\n", current->comm);
309 pr_emerg("tree range: %px start %lx end %lx\n", vma,
310 vmi_start, vmi_end - 1);
311 vma_iter_dump_tree(&vmi);
314 #ifdef CONFIG_DEBUG_VM_RB
316 anon_vma_lock_read(anon_vma);
317 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
318 anon_vma_interval_tree_verify(avc);
319 anon_vma_unlock_read(anon_vma);
324 if (i != mm->map_count) {
325 pr_emerg("map_count %d vma iterator %d\n", mm->map_count, i);
328 VM_BUG_ON_MM(bug, mm);
331 #else /* !CONFIG_DEBUG_VM_MAPLE_TREE */
332 #define validate_mm(mm) do { } while (0)
333 #endif /* CONFIG_DEBUG_VM_MAPLE_TREE */
336 * vma has some anon_vma assigned, and is already inserted on that
337 * anon_vma's interval trees.
339 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
340 * vma must be removed from the anon_vma's interval trees using
341 * anon_vma_interval_tree_pre_update_vma().
343 * After the update, the vma will be reinserted using
344 * anon_vma_interval_tree_post_update_vma().
346 * The entire update must be protected by exclusive mmap_lock and by
347 * the root anon_vma's mutex.
350 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
352 struct anon_vma_chain *avc;
354 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
355 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
359 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
361 struct anon_vma_chain *avc;
363 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
364 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
367 static unsigned long count_vma_pages_range(struct mm_struct *mm,
368 unsigned long addr, unsigned long end)
370 VMA_ITERATOR(vmi, mm, addr);
371 struct vm_area_struct *vma;
372 unsigned long nr_pages = 0;
374 for_each_vma_range(vmi, vma, end) {
375 unsigned long vm_start = max(addr, vma->vm_start);
376 unsigned long vm_end = min(end, vma->vm_end);
378 nr_pages += PHYS_PFN(vm_end - vm_start);
384 static void __vma_link_file(struct vm_area_struct *vma,
385 struct address_space *mapping)
387 if (vma_is_shared_maywrite(vma))
388 mapping_allow_writable(mapping);
390 flush_dcache_mmap_lock(mapping);
391 vma_interval_tree_insert(vma, &mapping->i_mmap);
392 flush_dcache_mmap_unlock(mapping);
395 static void vma_link_file(struct vm_area_struct *vma)
397 struct file *file = vma->vm_file;
398 struct address_space *mapping;
401 mapping = file->f_mapping;
402 i_mmap_lock_write(mapping);
403 __vma_link_file(vma, mapping);
404 i_mmap_unlock_write(mapping);
408 static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
410 VMA_ITERATOR(vmi, mm, 0);
412 vma_iter_config(&vmi, vma->vm_start, vma->vm_end);
413 if (vma_iter_prealloc(&vmi, vma))
416 vma_start_write(vma);
417 vma_iter_store(&vmi, vma);
425 * init_multi_vma_prep() - Initializer for struct vma_prepare
426 * @vp: The vma_prepare struct
427 * @vma: The vma that will be altered once locked
428 * @next: The next vma if it is to be adjusted
429 * @remove: The first vma to be removed
430 * @remove2: The second vma to be removed
432 static inline void init_multi_vma_prep(struct vma_prepare *vp,
433 struct vm_area_struct *vma, struct vm_area_struct *next,
434 struct vm_area_struct *remove, struct vm_area_struct *remove2)
436 memset(vp, 0, sizeof(struct vma_prepare));
438 vp->anon_vma = vma->anon_vma;
440 vp->remove2 = remove2;
442 if (!vp->anon_vma && next)
443 vp->anon_vma = next->anon_vma;
445 vp->file = vma->vm_file;
447 vp->mapping = vma->vm_file->f_mapping;
452 * init_vma_prep() - Initializer wrapper for vma_prepare struct
453 * @vp: The vma_prepare struct
454 * @vma: The vma that will be altered once locked
456 static inline void init_vma_prep(struct vma_prepare *vp,
457 struct vm_area_struct *vma)
459 init_multi_vma_prep(vp, vma, NULL, NULL, NULL);
464 * vma_prepare() - Helper function for handling locking VMAs prior to altering
465 * @vp: The initialized vma_prepare struct
467 static inline void vma_prepare(struct vma_prepare *vp)
470 uprobe_munmap(vp->vma, vp->vma->vm_start, vp->vma->vm_end);
473 uprobe_munmap(vp->adj_next, vp->adj_next->vm_start,
474 vp->adj_next->vm_end);
476 i_mmap_lock_write(vp->mapping);
477 if (vp->insert && vp->insert->vm_file) {
479 * Put into interval tree now, so instantiated pages
480 * are visible to arm/parisc __flush_dcache_page
481 * throughout; but we cannot insert into address
482 * space until vma start or end is updated.
484 __vma_link_file(vp->insert,
485 vp->insert->vm_file->f_mapping);
490 anon_vma_lock_write(vp->anon_vma);
491 anon_vma_interval_tree_pre_update_vma(vp->vma);
493 anon_vma_interval_tree_pre_update_vma(vp->adj_next);
497 flush_dcache_mmap_lock(vp->mapping);
498 vma_interval_tree_remove(vp->vma, &vp->mapping->i_mmap);
500 vma_interval_tree_remove(vp->adj_next,
501 &vp->mapping->i_mmap);
507 * vma_complete- Helper function for handling the unlocking after altering VMAs,
508 * or for inserting a VMA.
510 * @vp: The vma_prepare struct
511 * @vmi: The vma iterator
514 static inline void vma_complete(struct vma_prepare *vp,
515 struct vma_iterator *vmi, struct mm_struct *mm)
519 vma_interval_tree_insert(vp->adj_next,
520 &vp->mapping->i_mmap);
521 vma_interval_tree_insert(vp->vma, &vp->mapping->i_mmap);
522 flush_dcache_mmap_unlock(vp->mapping);
525 if (vp->remove && vp->file) {
526 __remove_shared_vm_struct(vp->remove, vp->mapping);
528 __remove_shared_vm_struct(vp->remove2, vp->mapping);
529 } else if (vp->insert) {
531 * split_vma has split insert from vma, and needs
532 * us to insert it before dropping the locks
533 * (it may either follow vma or precede it).
535 vma_iter_store(vmi, vp->insert);
540 anon_vma_interval_tree_post_update_vma(vp->vma);
542 anon_vma_interval_tree_post_update_vma(vp->adj_next);
543 anon_vma_unlock_write(vp->anon_vma);
547 i_mmap_unlock_write(vp->mapping);
548 uprobe_mmap(vp->vma);
551 uprobe_mmap(vp->adj_next);
556 vma_mark_detached(vp->remove, true);
558 uprobe_munmap(vp->remove, vp->remove->vm_start,
562 if (vp->remove->anon_vma)
563 anon_vma_merge(vp->vma, vp->remove);
565 mpol_put(vma_policy(vp->remove));
567 WARN_ON_ONCE(vp->vma->vm_end < vp->remove->vm_end);
568 vm_area_free(vp->remove);
571 * In mprotect's case 6 (see comments on vma_merge),
572 * we are removing both mid and next vmas
575 vp->remove = vp->remove2;
580 if (vp->insert && vp->file)
581 uprobe_mmap(vp->insert);
586 * dup_anon_vma() - Helper function to duplicate anon_vma
587 * @dst: The destination VMA
588 * @src: The source VMA
589 * @dup: Pointer to the destination VMA when successful.
591 * Returns: 0 on success.
593 static inline int dup_anon_vma(struct vm_area_struct *dst,
594 struct vm_area_struct *src, struct vm_area_struct **dup)
597 * Easily overlooked: when mprotect shifts the boundary, make sure the
598 * expanding vma has anon_vma set if the shrinking vma had, to cover any
599 * anon pages imported.
601 if (src->anon_vma && !dst->anon_vma) {
604 vma_assert_write_locked(dst);
605 dst->anon_vma = src->anon_vma;
606 ret = anon_vma_clone(dst, src);
617 * vma_expand - Expand an existing VMA
619 * @vmi: The vma iterator
620 * @vma: The vma to expand
621 * @start: The start of the vma
622 * @end: The exclusive end of the vma
623 * @pgoff: The page offset of vma
624 * @next: The current of next vma.
626 * Expand @vma to @start and @end. Can expand off the start and end. Will
627 * expand over @next if it's different from @vma and @end == @next->vm_end.
628 * Checking if the @vma can expand and merge with @next needs to be handled by
631 * Returns: 0 on success
633 int vma_expand(struct vma_iterator *vmi, struct vm_area_struct *vma,
634 unsigned long start, unsigned long end, pgoff_t pgoff,
635 struct vm_area_struct *next)
637 struct vm_area_struct *anon_dup = NULL;
638 bool remove_next = false;
639 struct vma_prepare vp;
641 vma_start_write(vma);
642 if (next && (vma != next) && (end == next->vm_end)) {
646 vma_start_write(next);
647 ret = dup_anon_vma(vma, next, &anon_dup);
652 init_multi_vma_prep(&vp, vma, NULL, remove_next ? next : NULL, NULL);
653 /* Not merging but overwriting any part of next is not handled. */
654 VM_WARN_ON(next && !vp.remove &&
655 next != vma && end > next->vm_start);
656 /* Only handles expanding */
657 VM_WARN_ON(vma->vm_start < start || vma->vm_end > end);
659 /* Note: vma iterator must be pointing to 'start' */
660 vma_iter_config(vmi, start, end);
661 if (vma_iter_prealloc(vmi, vma))
665 vma_adjust_trans_huge(vma, start, end, 0);
666 vma_set_range(vma, start, end, pgoff);
667 vma_iter_store(vmi, vma);
669 vma_complete(&vp, vmi, vma->vm_mm);
674 unlink_anon_vmas(anon_dup);
679 * vma_shrink() - Reduce an existing VMAs memory area
680 * @vmi: The vma iterator
681 * @vma: The VMA to modify
682 * @start: The new start
685 * Returns: 0 on success, -ENOMEM otherwise
687 int vma_shrink(struct vma_iterator *vmi, struct vm_area_struct *vma,
688 unsigned long start, unsigned long end, pgoff_t pgoff)
690 struct vma_prepare vp;
692 WARN_ON((vma->vm_start != start) && (vma->vm_end != end));
694 if (vma->vm_start < start)
695 vma_iter_config(vmi, vma->vm_start, start);
697 vma_iter_config(vmi, end, vma->vm_end);
699 if (vma_iter_prealloc(vmi, NULL))
702 vma_start_write(vma);
704 init_vma_prep(&vp, vma);
706 vma_adjust_trans_huge(vma, start, end, 0);
709 vma_set_range(vma, start, end, pgoff);
710 vma_complete(&vp, vmi, vma->vm_mm);
715 * If the vma has a ->close operation then the driver probably needs to release
716 * per-vma resources, so we don't attempt to merge those if the caller indicates
717 * the current vma may be removed as part of the merge.
719 static inline bool is_mergeable_vma(struct vm_area_struct *vma,
720 struct file *file, unsigned long vm_flags,
721 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
722 struct anon_vma_name *anon_name, bool may_remove_vma)
725 * VM_SOFTDIRTY should not prevent from VMA merging, if we
726 * match the flags but dirty bit -- the caller should mark
727 * merged VMA as dirty. If dirty bit won't be excluded from
728 * comparison, we increase pressure on the memory system forcing
729 * the kernel to generate new VMAs when old one could be
732 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
734 if (vma->vm_file != file)
736 if (may_remove_vma && vma->vm_ops && vma->vm_ops->close)
738 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
740 if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
745 static inline bool is_mergeable_anon_vma(struct anon_vma *anon_vma1,
746 struct anon_vma *anon_vma2, struct vm_area_struct *vma)
749 * The list_is_singular() test is to avoid merging VMA cloned from
750 * parents. This can improve scalability caused by anon_vma lock.
752 if ((!anon_vma1 || !anon_vma2) && (!vma ||
753 list_is_singular(&vma->anon_vma_chain)))
755 return anon_vma1 == anon_vma2;
759 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
760 * in front of (at a lower virtual address and file offset than) the vma.
762 * We cannot merge two vmas if they have differently assigned (non-NULL)
763 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
765 * We don't check here for the merged mmap wrapping around the end of pagecache
766 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
767 * wrap, nor mmaps which cover the final page at index -1UL.
769 * We assume the vma may be removed as part of the merge.
772 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
773 struct anon_vma *anon_vma, struct file *file,
774 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
775 struct anon_vma_name *anon_name)
777 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, true) &&
778 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
779 if (vma->vm_pgoff == vm_pgoff)
786 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
787 * beyond (at a higher virtual address and file offset than) the vma.
789 * We cannot merge two vmas if they have differently assigned (non-NULL)
790 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
792 * We assume that vma is not removed as part of the merge.
795 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
796 struct anon_vma *anon_vma, struct file *file,
797 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
798 struct anon_vma_name *anon_name)
800 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, false) &&
801 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
803 vm_pglen = vma_pages(vma);
804 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
811 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
812 * figure out whether that can be merged with its predecessor or its
813 * successor. Or both (it neatly fills a hole).
815 * In most cases - when called for mmap, brk or mremap - [addr,end) is
816 * certain not to be mapped by the time vma_merge is called; but when
817 * called for mprotect, it is certain to be already mapped (either at
818 * an offset within prev, or at the start of next), and the flags of
819 * this area are about to be changed to vm_flags - and the no-change
820 * case has already been eliminated.
822 * The following mprotect cases have to be considered, where **** is
823 * the area passed down from mprotect_fixup, never extending beyond one
824 * vma, PPPP is the previous vma, CCCC is a concurrent vma that starts
825 * at the same address as **** and is of the same or larger span, and
826 * NNNN the next vma after ****:
829 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPCCCCCC
830 * cannot merge might become might become
831 * PPNNNNNNNNNN PPPPPPPPPPCC
832 * mmap, brk or case 4 below case 5 below
835 * PPPP NNNN PPPPCCCCNNNN
836 * might become might become
837 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
838 * PPPPPPPPNNNN 2 or PPPPPPPPNNNN 7 or
839 * PPPPNNNNNNNN 3 PPPPNNNNNNNN 8
841 * It is important for case 8 that the vma CCCC overlapping the
842 * region **** is never going to extended over NNNN. Instead NNNN must
843 * be extended in region **** and CCCC must be removed. This way in
844 * all cases where vma_merge succeeds, the moment vma_merge drops the
845 * rmap_locks, the properties of the merged vma will be already
846 * correct for the whole merged range. Some of those properties like
847 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
848 * be correct for the whole merged range immediately after the
849 * rmap_locks are released. Otherwise if NNNN would be removed and
850 * CCCC would be extended over the NNNN range, remove_migration_ptes
851 * or other rmap walkers (if working on addresses beyond the "end"
852 * parameter) may establish ptes with the wrong permissions of CCCC
853 * instead of the right permissions of NNNN.
856 * PPPP is represented by *prev
857 * CCCC is represented by *curr or not represented at all (NULL)
858 * NNNN is represented by *next or not represented at all (NULL)
859 * **** is not represented - it will be merged and the vma containing the
860 * area is returned, or the function will return NULL
862 static struct vm_area_struct
863 *vma_merge(struct vma_iterator *vmi, struct vm_area_struct *prev,
864 struct vm_area_struct *src, unsigned long addr, unsigned long end,
865 unsigned long vm_flags, pgoff_t pgoff, struct mempolicy *policy,
866 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
867 struct anon_vma_name *anon_name)
869 struct mm_struct *mm = src->vm_mm;
870 struct anon_vma *anon_vma = src->anon_vma;
871 struct file *file = src->vm_file;
872 struct vm_area_struct *curr, *next, *res;
873 struct vm_area_struct *vma, *adjust, *remove, *remove2;
874 struct vm_area_struct *anon_dup = NULL;
875 struct vma_prepare vp;
878 bool merge_prev = false;
879 bool merge_next = false;
880 bool vma_expanded = false;
881 unsigned long vma_start = addr;
882 unsigned long vma_end = end;
883 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
887 * We later require that vma->vm_flags == vm_flags,
888 * so this tests vma->vm_flags & VM_SPECIAL, too.
890 if (vm_flags & VM_SPECIAL)
893 /* Does the input range span an existing VMA? (cases 5 - 8) */
894 curr = find_vma_intersection(mm, prev ? prev->vm_end : 0, end);
896 if (!curr || /* cases 1 - 4 */
897 end == curr->vm_end) /* cases 6 - 8, adjacent VMA */
898 next = vma_lookup(mm, end);
900 next = NULL; /* case 5 */
903 vma_start = prev->vm_start;
904 vma_pgoff = prev->vm_pgoff;
906 /* Can we merge the predecessor? */
907 if (addr == prev->vm_end && mpol_equal(vma_policy(prev), policy)
908 && can_vma_merge_after(prev, vm_flags, anon_vma, file,
909 pgoff, vm_userfaultfd_ctx, anon_name)) {
915 /* Can we merge the successor? */
916 if (next && mpol_equal(policy, vma_policy(next)) &&
917 can_vma_merge_before(next, vm_flags, anon_vma, file, pgoff+pglen,
918 vm_userfaultfd_ctx, anon_name)) {
922 /* Verify some invariant that must be enforced by the caller. */
923 VM_WARN_ON(prev && addr <= prev->vm_start);
924 VM_WARN_ON(curr && (addr != curr->vm_start || end > curr->vm_end));
925 VM_WARN_ON(addr >= end);
927 if (!merge_prev && !merge_next)
928 return NULL; /* Not mergeable. */
931 vma_start_write(prev);
934 remove = remove2 = adjust = NULL;
936 /* Can we merge both the predecessor and the successor? */
937 if (merge_prev && merge_next &&
938 is_mergeable_anon_vma(prev->anon_vma, next->anon_vma, NULL)) {
939 vma_start_write(next);
940 remove = next; /* case 1 */
941 vma_end = next->vm_end;
942 err = dup_anon_vma(prev, next, &anon_dup);
943 if (curr) { /* case 6 */
944 vma_start_write(curr);
948 * Note that the dup_anon_vma below cannot overwrite err
949 * since the first caller would do nothing unless next
953 err = dup_anon_vma(prev, curr, &anon_dup);
955 } else if (merge_prev) { /* case 2 */
957 vma_start_write(curr);
958 if (end == curr->vm_end) { /* case 7 */
960 * can_vma_merge_after() assumed we would not be
961 * removing prev vma, so it skipped the check
962 * for vm_ops->close, but we are removing curr
964 if (curr->vm_ops && curr->vm_ops->close)
967 } else { /* case 5 */
969 adj_start = (end - curr->vm_start);
972 err = dup_anon_vma(prev, curr, &anon_dup);
974 } else { /* merge_next */
975 vma_start_write(next);
977 if (prev && addr < prev->vm_end) { /* case 4 */
978 vma_start_write(prev);
981 adj_start = -(prev->vm_end - addr);
982 err = dup_anon_vma(next, prev, &anon_dup);
985 * Note that cases 3 and 8 are the ONLY ones where prev
986 * is permitted to be (but is not necessarily) NULL.
988 vma = next; /* case 3 */
990 vma_end = next->vm_end;
991 vma_pgoff = next->vm_pgoff - pglen;
992 if (curr) { /* case 8 */
993 vma_pgoff = curr->vm_pgoff;
994 vma_start_write(curr);
996 err = dup_anon_vma(next, curr, &anon_dup);
1001 /* Error in anon_vma clone. */
1005 if (vma_start < vma->vm_start || vma_end > vma->vm_end)
1006 vma_expanded = true;
1009 vma_iter_config(vmi, vma_start, vma_end);
1011 vma_iter_config(vmi, adjust->vm_start + adj_start,
1015 if (vma_iter_prealloc(vmi, vma))
1018 init_multi_vma_prep(&vp, vma, adjust, remove, remove2);
1019 VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma &&
1020 vp.anon_vma != adjust->anon_vma);
1023 vma_adjust_trans_huge(vma, vma_start, vma_end, adj_start);
1024 vma_set_range(vma, vma_start, vma_end, vma_pgoff);
1027 vma_iter_store(vmi, vma);
1030 adjust->vm_start += adj_start;
1031 adjust->vm_pgoff += adj_start >> PAGE_SHIFT;
1032 if (adj_start < 0) {
1033 WARN_ON(vma_expanded);
1034 vma_iter_store(vmi, next);
1038 vma_complete(&vp, vmi, mm);
1039 khugepaged_enter_vma(res, vm_flags);
1044 unlink_anon_vmas(anon_dup);
1047 vma_iter_set(vmi, addr);
1053 * Rough compatibility check to quickly see if it's even worth looking
1054 * at sharing an anon_vma.
1056 * They need to have the same vm_file, and the flags can only differ
1057 * in things that mprotect may change.
1059 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1060 * we can merge the two vma's. For example, we refuse to merge a vma if
1061 * there is a vm_ops->close() function, because that indicates that the
1062 * driver is doing some kind of reference counting. But that doesn't
1063 * really matter for the anon_vma sharing case.
1065 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1067 return a->vm_end == b->vm_start &&
1068 mpol_equal(vma_policy(a), vma_policy(b)) &&
1069 a->vm_file == b->vm_file &&
1070 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1071 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1075 * Do some basic sanity checking to see if we can re-use the anon_vma
1076 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1077 * the same as 'old', the other will be the new one that is trying
1078 * to share the anon_vma.
1080 * NOTE! This runs with mmap_lock held for reading, so it is possible that
1081 * the anon_vma of 'old' is concurrently in the process of being set up
1082 * by another page fault trying to merge _that_. But that's ok: if it
1083 * is being set up, that automatically means that it will be a singleton
1084 * acceptable for merging, so we can do all of this optimistically. But
1085 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1087 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1088 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1089 * is to return an anon_vma that is "complex" due to having gone through
1092 * We also make sure that the two vma's are compatible (adjacent,
1093 * and with the same memory policies). That's all stable, even with just
1094 * a read lock on the mmap_lock.
1096 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1098 if (anon_vma_compatible(a, b)) {
1099 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1101 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1108 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1109 * neighbouring vmas for a suitable anon_vma, before it goes off
1110 * to allocate a new anon_vma. It checks because a repetitive
1111 * sequence of mprotects and faults may otherwise lead to distinct
1112 * anon_vmas being allocated, preventing vma merge in subsequent
1115 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1117 MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end);
1118 struct anon_vma *anon_vma = NULL;
1119 struct vm_area_struct *prev, *next;
1121 /* Try next first. */
1122 next = mas_walk(&mas);
1124 anon_vma = reusable_anon_vma(next, vma, next);
1129 prev = mas_prev(&mas, 0);
1130 VM_BUG_ON_VMA(prev != vma, vma);
1131 prev = mas_prev(&mas, 0);
1132 /* Try prev next. */
1134 anon_vma = reusable_anon_vma(prev, prev, vma);
1137 * We might reach here with anon_vma == NULL if we can't find
1138 * any reusable anon_vma.
1139 * There's no absolute need to look only at touching neighbours:
1140 * we could search further afield for "compatible" anon_vmas.
1141 * But it would probably just be a waste of time searching,
1142 * or lead to too many vmas hanging off the same anon_vma.
1143 * We're trying to allow mprotect remerging later on,
1144 * not trying to minimize memory used for anon_vmas.
1150 * If a hint addr is less than mmap_min_addr change hint to be as
1151 * low as possible but still greater than mmap_min_addr
1153 static inline unsigned long round_hint_to_min(unsigned long hint)
1156 if (((void *)hint != NULL) &&
1157 (hint < mmap_min_addr))
1158 return PAGE_ALIGN(mmap_min_addr);
1162 bool mlock_future_ok(struct mm_struct *mm, unsigned long flags,
1163 unsigned long bytes)
1165 unsigned long locked_pages, limit_pages;
1167 if (!(flags & VM_LOCKED) || capable(CAP_IPC_LOCK))
1170 locked_pages = bytes >> PAGE_SHIFT;
1171 locked_pages += mm->locked_vm;
1173 limit_pages = rlimit(RLIMIT_MEMLOCK);
1174 limit_pages >>= PAGE_SHIFT;
1176 return locked_pages <= limit_pages;
1179 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1181 if (S_ISREG(inode->i_mode))
1182 return MAX_LFS_FILESIZE;
1184 if (S_ISBLK(inode->i_mode))
1185 return MAX_LFS_FILESIZE;
1187 if (S_ISSOCK(inode->i_mode))
1188 return MAX_LFS_FILESIZE;
1190 /* Special "we do even unsigned file positions" case */
1191 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1194 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1198 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1199 unsigned long pgoff, unsigned long len)
1201 u64 maxsize = file_mmap_size_max(file, inode);
1203 if (maxsize && len > maxsize)
1206 if (pgoff > maxsize >> PAGE_SHIFT)
1212 * The caller must write-lock current->mm->mmap_lock.
1214 unsigned long do_mmap(struct file *file, unsigned long addr,
1215 unsigned long len, unsigned long prot,
1216 unsigned long flags, vm_flags_t vm_flags,
1217 unsigned long pgoff, unsigned long *populate,
1218 struct list_head *uf)
1220 struct mm_struct *mm = current->mm;
1229 * Does the application expect PROT_READ to imply PROT_EXEC?
1231 * (the exception is when the underlying filesystem is noexec
1232 * mounted, in which case we don't add PROT_EXEC.)
1234 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1235 if (!(file && path_noexec(&file->f_path)))
1238 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1239 if (flags & MAP_FIXED_NOREPLACE)
1242 if (!(flags & MAP_FIXED))
1243 addr = round_hint_to_min(addr);
1245 /* Careful about overflows.. */
1246 len = PAGE_ALIGN(len);
1250 /* offset overflow? */
1251 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1254 /* Too many mappings? */
1255 if (mm->map_count > sysctl_max_map_count)
1258 /* Obtain the address to map to. we verify (or select) it and ensure
1259 * that it represents a valid section of the address space.
1261 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1262 if (IS_ERR_VALUE(addr))
1265 if (flags & MAP_FIXED_NOREPLACE) {
1266 if (find_vma_intersection(mm, addr, addr + len))
1270 if (prot == PROT_EXEC) {
1271 pkey = execute_only_pkey(mm);
1276 /* Do simple checking here so the lower-level routines won't have
1277 * to. we assume access permissions have been handled by the open
1278 * of the memory object, so we don't do any here.
1280 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1281 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1283 if (flags & MAP_LOCKED)
1284 if (!can_do_mlock())
1287 if (!mlock_future_ok(mm, vm_flags, len))
1291 struct inode *inode = file_inode(file);
1292 unsigned long flags_mask;
1294 if (!file_mmap_ok(file, inode, pgoff, len))
1297 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1299 switch (flags & MAP_TYPE) {
1302 * Force use of MAP_SHARED_VALIDATE with non-legacy
1303 * flags. E.g. MAP_SYNC is dangerous to use with
1304 * MAP_SHARED as you don't know which consistency model
1305 * you will get. We silently ignore unsupported flags
1306 * with MAP_SHARED to preserve backward compatibility.
1308 flags &= LEGACY_MAP_MASK;
1310 case MAP_SHARED_VALIDATE:
1311 if (flags & ~flags_mask)
1313 if (prot & PROT_WRITE) {
1314 if (!(file->f_mode & FMODE_WRITE))
1316 if (IS_SWAPFILE(file->f_mapping->host))
1321 * Make sure we don't allow writing to an append-only
1324 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1327 vm_flags |= VM_SHARED | VM_MAYSHARE;
1328 if (!(file->f_mode & FMODE_WRITE))
1329 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1332 if (!(file->f_mode & FMODE_READ))
1334 if (path_noexec(&file->f_path)) {
1335 if (vm_flags & VM_EXEC)
1337 vm_flags &= ~VM_MAYEXEC;
1340 if (!file->f_op->mmap)
1342 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1350 switch (flags & MAP_TYPE) {
1352 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1358 vm_flags |= VM_SHARED | VM_MAYSHARE;
1362 * Set pgoff according to addr for anon_vma.
1364 pgoff = addr >> PAGE_SHIFT;
1372 * Set 'VM_NORESERVE' if we should not account for the
1373 * memory use of this mapping.
1375 if (flags & MAP_NORESERVE) {
1376 /* We honor MAP_NORESERVE if allowed to overcommit */
1377 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1378 vm_flags |= VM_NORESERVE;
1380 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1381 if (file && is_file_hugepages(file))
1382 vm_flags |= VM_NORESERVE;
1385 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1386 if (!IS_ERR_VALUE(addr) &&
1387 ((vm_flags & VM_LOCKED) ||
1388 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1393 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1394 unsigned long prot, unsigned long flags,
1395 unsigned long fd, unsigned long pgoff)
1397 struct file *file = NULL;
1398 unsigned long retval;
1400 if (!(flags & MAP_ANONYMOUS)) {
1401 audit_mmap_fd(fd, flags);
1405 if (is_file_hugepages(file)) {
1406 len = ALIGN(len, huge_page_size(hstate_file(file)));
1407 } else if (unlikely(flags & MAP_HUGETLB)) {
1411 } else if (flags & MAP_HUGETLB) {
1414 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1418 len = ALIGN(len, huge_page_size(hs));
1420 * VM_NORESERVE is used because the reservations will be
1421 * taken when vm_ops->mmap() is called
1423 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1425 HUGETLB_ANONHUGE_INODE,
1426 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1428 return PTR_ERR(file);
1431 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1438 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1439 unsigned long, prot, unsigned long, flags,
1440 unsigned long, fd, unsigned long, pgoff)
1442 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1445 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1446 struct mmap_arg_struct {
1450 unsigned long flags;
1452 unsigned long offset;
1455 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1457 struct mmap_arg_struct a;
1459 if (copy_from_user(&a, arg, sizeof(a)))
1461 if (offset_in_page(a.offset))
1464 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1465 a.offset >> PAGE_SHIFT);
1467 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1469 static bool vm_ops_needs_writenotify(const struct vm_operations_struct *vm_ops)
1471 return vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite);
1474 static bool vma_is_shared_writable(struct vm_area_struct *vma)
1476 return (vma->vm_flags & (VM_WRITE | VM_SHARED)) ==
1477 (VM_WRITE | VM_SHARED);
1480 static bool vma_fs_can_writeback(struct vm_area_struct *vma)
1482 /* No managed pages to writeback. */
1483 if (vma->vm_flags & VM_PFNMAP)
1486 return vma->vm_file && vma->vm_file->f_mapping &&
1487 mapping_can_writeback(vma->vm_file->f_mapping);
1491 * Does this VMA require the underlying folios to have their dirty state
1494 bool vma_needs_dirty_tracking(struct vm_area_struct *vma)
1496 /* Only shared, writable VMAs require dirty tracking. */
1497 if (!vma_is_shared_writable(vma))
1500 /* Does the filesystem need to be notified? */
1501 if (vm_ops_needs_writenotify(vma->vm_ops))
1505 * Even if the filesystem doesn't indicate a need for writenotify, if it
1506 * can writeback, dirty tracking is still required.
1508 return vma_fs_can_writeback(vma);
1512 * Some shared mappings will want the pages marked read-only
1513 * to track write events. If so, we'll downgrade vm_page_prot
1514 * to the private version (using protection_map[] without the
1517 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1519 /* If it was private or non-writable, the write bit is already clear */
1520 if (!vma_is_shared_writable(vma))
1523 /* The backer wishes to know when pages are first written to? */
1524 if (vm_ops_needs_writenotify(vma->vm_ops))
1527 /* The open routine did something to the protections that pgprot_modify
1528 * won't preserve? */
1529 if (pgprot_val(vm_page_prot) !=
1530 pgprot_val(vm_pgprot_modify(vm_page_prot, vma->vm_flags)))
1534 * Do we need to track softdirty? hugetlb does not support softdirty
1537 if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma))
1540 /* Do we need write faults for uffd-wp tracking? */
1541 if (userfaultfd_wp(vma))
1544 /* Can the mapping track the dirty pages? */
1545 return vma_fs_can_writeback(vma);
1549 * We account for memory if it's a private writeable mapping,
1550 * not hugepages and VM_NORESERVE wasn't set.
1552 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1555 * hugetlb has its own accounting separate from the core VM
1556 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1558 if (file && is_file_hugepages(file))
1561 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1565 * unmapped_area() - Find an area between the low_limit and the high_limit with
1566 * the correct alignment and offset, all from @info. Note: current->mm is used
1569 * @info: The unmapped area information including the range [low_limit -
1570 * high_limit), the alignment offset and mask.
1572 * Return: A memory address or -ENOMEM.
1574 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1576 unsigned long length, gap;
1577 unsigned long low_limit, high_limit;
1578 struct vm_area_struct *tmp;
1580 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1582 /* Adjust search length to account for worst case alignment overhead */
1583 length = info->length + info->align_mask;
1584 if (length < info->length)
1587 low_limit = info->low_limit;
1588 if (low_limit < mmap_min_addr)
1589 low_limit = mmap_min_addr;
1590 high_limit = info->high_limit;
1592 if (mas_empty_area(&mas, low_limit, high_limit - 1, length))
1596 gap += (info->align_offset - gap) & info->align_mask;
1597 tmp = mas_next(&mas, ULONG_MAX);
1598 if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1599 if (vm_start_gap(tmp) < gap + length - 1) {
1600 low_limit = tmp->vm_end;
1605 tmp = mas_prev(&mas, 0);
1606 if (tmp && vm_end_gap(tmp) > gap) {
1607 low_limit = vm_end_gap(tmp);
1617 * unmapped_area_topdown() - Find an area between the low_limit and the
1618 * high_limit with the correct alignment and offset at the highest available
1619 * address, all from @info. Note: current->mm is used for the search.
1621 * @info: The unmapped area information including the range [low_limit -
1622 * high_limit), the alignment offset and mask.
1624 * Return: A memory address or -ENOMEM.
1626 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1628 unsigned long length, gap, gap_end;
1629 unsigned long low_limit, high_limit;
1630 struct vm_area_struct *tmp;
1632 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1633 /* Adjust search length to account for worst case alignment overhead */
1634 length = info->length + info->align_mask;
1635 if (length < info->length)
1638 low_limit = info->low_limit;
1639 if (low_limit < mmap_min_addr)
1640 low_limit = mmap_min_addr;
1641 high_limit = info->high_limit;
1643 if (mas_empty_area_rev(&mas, low_limit, high_limit - 1, length))
1646 gap = mas.last + 1 - info->length;
1647 gap -= (gap - info->align_offset) & info->align_mask;
1649 tmp = mas_next(&mas, ULONG_MAX);
1650 if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1651 if (vm_start_gap(tmp) <= gap_end) {
1652 high_limit = vm_start_gap(tmp);
1657 tmp = mas_prev(&mas, 0);
1658 if (tmp && vm_end_gap(tmp) > gap) {
1659 high_limit = tmp->vm_start;
1669 * Search for an unmapped address range.
1671 * We are looking for a range that:
1672 * - does not intersect with any VMA;
1673 * - is contained within the [low_limit, high_limit) interval;
1674 * - is at least the desired size.
1675 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
1677 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
1681 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
1682 addr = unmapped_area_topdown(info);
1684 addr = unmapped_area(info);
1686 trace_vm_unmapped_area(addr, info);
1690 /* Get an address range which is currently unmapped.
1691 * For shmat() with addr=0.
1693 * Ugly calling convention alert:
1694 * Return value with the low bits set means error value,
1696 * if (ret & ~PAGE_MASK)
1699 * This function "knows" that -ENOMEM has the bits set.
1702 generic_get_unmapped_area(struct file *filp, unsigned long addr,
1703 unsigned long len, unsigned long pgoff,
1704 unsigned long flags)
1706 struct mm_struct *mm = current->mm;
1707 struct vm_area_struct *vma, *prev;
1708 struct vm_unmapped_area_info info;
1709 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1711 if (len > mmap_end - mmap_min_addr)
1714 if (flags & MAP_FIXED)
1718 addr = PAGE_ALIGN(addr);
1719 vma = find_vma_prev(mm, addr, &prev);
1720 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1721 (!vma || addr + len <= vm_start_gap(vma)) &&
1722 (!prev || addr >= vm_end_gap(prev)))
1728 info.low_limit = mm->mmap_base;
1729 info.high_limit = mmap_end;
1730 info.align_mask = 0;
1731 info.align_offset = 0;
1732 return vm_unmapped_area(&info);
1735 #ifndef HAVE_ARCH_UNMAPPED_AREA
1737 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1738 unsigned long len, unsigned long pgoff,
1739 unsigned long flags)
1741 return generic_get_unmapped_area(filp, addr, len, pgoff, flags);
1746 * This mmap-allocator allocates new areas top-down from below the
1747 * stack's low limit (the base):
1750 generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1751 unsigned long len, unsigned long pgoff,
1752 unsigned long flags)
1754 struct vm_area_struct *vma, *prev;
1755 struct mm_struct *mm = current->mm;
1756 struct vm_unmapped_area_info info;
1757 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1759 /* requested length too big for entire address space */
1760 if (len > mmap_end - mmap_min_addr)
1763 if (flags & MAP_FIXED)
1766 /* requesting a specific address */
1768 addr = PAGE_ALIGN(addr);
1769 vma = find_vma_prev(mm, addr, &prev);
1770 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1771 (!vma || addr + len <= vm_start_gap(vma)) &&
1772 (!prev || addr >= vm_end_gap(prev)))
1776 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1778 info.low_limit = PAGE_SIZE;
1779 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
1780 info.align_mask = 0;
1781 info.align_offset = 0;
1782 addr = vm_unmapped_area(&info);
1785 * A failed mmap() very likely causes application failure,
1786 * so fall back to the bottom-up function here. This scenario
1787 * can happen with large stack limits and large mmap()
1790 if (offset_in_page(addr)) {
1791 VM_BUG_ON(addr != -ENOMEM);
1793 info.low_limit = TASK_UNMAPPED_BASE;
1794 info.high_limit = mmap_end;
1795 addr = vm_unmapped_area(&info);
1801 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1803 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1804 unsigned long len, unsigned long pgoff,
1805 unsigned long flags)
1807 return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags);
1812 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1813 unsigned long pgoff, unsigned long flags)
1815 unsigned long (*get_area)(struct file *, unsigned long,
1816 unsigned long, unsigned long, unsigned long);
1818 unsigned long error = arch_mmap_check(addr, len, flags);
1822 /* Careful about overflows.. */
1823 if (len > TASK_SIZE)
1826 get_area = current->mm->get_unmapped_area;
1828 if (file->f_op->get_unmapped_area)
1829 get_area = file->f_op->get_unmapped_area;
1830 } else if (flags & MAP_SHARED) {
1832 * mmap_region() will call shmem_zero_setup() to create a file,
1833 * so use shmem's get_unmapped_area in case it can be huge.
1835 get_area = shmem_get_unmapped_area;
1836 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
1837 /* Ensures that larger anonymous mappings are THP aligned. */
1838 get_area = thp_get_unmapped_area;
1841 /* Always treat pgoff as zero for anonymous memory. */
1845 addr = get_area(file, addr, len, pgoff, flags);
1846 if (IS_ERR_VALUE(addr))
1849 if (addr > TASK_SIZE - len)
1851 if (offset_in_page(addr))
1854 error = security_mmap_addr(addr);
1855 return error ? error : addr;
1858 EXPORT_SYMBOL(get_unmapped_area);
1861 * find_vma_intersection() - Look up the first VMA which intersects the interval
1862 * @mm: The process address space.
1863 * @start_addr: The inclusive start user address.
1864 * @end_addr: The exclusive end user address.
1866 * Returns: The first VMA within the provided range, %NULL otherwise. Assumes
1867 * start_addr < end_addr.
1869 struct vm_area_struct *find_vma_intersection(struct mm_struct *mm,
1870 unsigned long start_addr,
1871 unsigned long end_addr)
1873 unsigned long index = start_addr;
1875 mmap_assert_locked(mm);
1876 return mt_find(&mm->mm_mt, &index, end_addr - 1);
1878 EXPORT_SYMBOL(find_vma_intersection);
1881 * find_vma() - Find the VMA for a given address, or the next VMA.
1882 * @mm: The mm_struct to check
1883 * @addr: The address
1885 * Returns: The VMA associated with addr, or the next VMA.
1886 * May return %NULL in the case of no VMA at addr or above.
1888 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1890 unsigned long index = addr;
1892 mmap_assert_locked(mm);
1893 return mt_find(&mm->mm_mt, &index, ULONG_MAX);
1895 EXPORT_SYMBOL(find_vma);
1898 * find_vma_prev() - Find the VMA for a given address, or the next vma and
1899 * set %pprev to the previous VMA, if any.
1900 * @mm: The mm_struct to check
1901 * @addr: The address
1902 * @pprev: The pointer to set to the previous VMA
1904 * Note that RCU lock is missing here since the external mmap_lock() is used
1907 * Returns: The VMA associated with @addr, or the next vma.
1908 * May return %NULL in the case of no vma at addr or above.
1910 struct vm_area_struct *
1911 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1912 struct vm_area_struct **pprev)
1914 struct vm_area_struct *vma;
1915 MA_STATE(mas, &mm->mm_mt, addr, addr);
1917 vma = mas_walk(&mas);
1918 *pprev = mas_prev(&mas, 0);
1920 vma = mas_next(&mas, ULONG_MAX);
1925 * Verify that the stack growth is acceptable and
1926 * update accounting. This is shared with both the
1927 * grow-up and grow-down cases.
1929 static int acct_stack_growth(struct vm_area_struct *vma,
1930 unsigned long size, unsigned long grow)
1932 struct mm_struct *mm = vma->vm_mm;
1933 unsigned long new_start;
1935 /* address space limit tests */
1936 if (!may_expand_vm(mm, vma->vm_flags, grow))
1939 /* Stack limit test */
1940 if (size > rlimit(RLIMIT_STACK))
1943 /* mlock limit tests */
1944 if (!mlock_future_ok(mm, vma->vm_flags, grow << PAGE_SHIFT))
1947 /* Check to ensure the stack will not grow into a hugetlb-only region */
1948 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1950 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1954 * Overcommit.. This must be the final test, as it will
1955 * update security statistics.
1957 if (security_vm_enough_memory_mm(mm, grow))
1963 #if defined(CONFIG_STACK_GROWSUP)
1965 * PA-RISC uses this for its stack.
1966 * vma is the last one with address > vma->vm_end. Have to extend vma.
1968 static int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1970 struct mm_struct *mm = vma->vm_mm;
1971 struct vm_area_struct *next;
1972 unsigned long gap_addr;
1974 MA_STATE(mas, &mm->mm_mt, vma->vm_start, address);
1976 if (!(vma->vm_flags & VM_GROWSUP))
1979 /* Guard against exceeding limits of the address space. */
1980 address &= PAGE_MASK;
1981 if (address >= (TASK_SIZE & PAGE_MASK))
1983 address += PAGE_SIZE;
1985 /* Enforce stack_guard_gap */
1986 gap_addr = address + stack_guard_gap;
1988 /* Guard against overflow */
1989 if (gap_addr < address || gap_addr > TASK_SIZE)
1990 gap_addr = TASK_SIZE;
1992 next = find_vma_intersection(mm, vma->vm_end, gap_addr);
1993 if (next && vma_is_accessible(next)) {
1994 if (!(next->vm_flags & VM_GROWSUP))
1996 /* Check that both stack segments have the same anon_vma? */
2000 mas_prev_range(&mas, address);
2002 __mas_set_range(&mas, vma->vm_start, address - 1);
2003 if (mas_preallocate(&mas, vma, GFP_KERNEL))
2006 /* We must make sure the anon_vma is allocated. */
2007 if (unlikely(anon_vma_prepare(vma))) {
2012 /* Lock the VMA before expanding to prevent concurrent page faults */
2013 vma_start_write(vma);
2015 * vma->vm_start/vm_end cannot change under us because the caller
2016 * is required to hold the mmap_lock in read mode. We need the
2017 * anon_vma lock to serialize against concurrent expand_stacks.
2019 anon_vma_lock_write(vma->anon_vma);
2021 /* Somebody else might have raced and expanded it already */
2022 if (address > vma->vm_end) {
2023 unsigned long size, grow;
2025 size = address - vma->vm_start;
2026 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2029 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2030 error = acct_stack_growth(vma, size, grow);
2033 * We only hold a shared mmap_lock lock here, so
2034 * we need to protect against concurrent vma
2035 * expansions. anon_vma_lock_write() doesn't
2036 * help here, as we don't guarantee that all
2037 * growable vmas in a mm share the same root
2038 * anon vma. So, we reuse mm->page_table_lock
2039 * to guard against concurrent vma expansions.
2041 spin_lock(&mm->page_table_lock);
2042 if (vma->vm_flags & VM_LOCKED)
2043 mm->locked_vm += grow;
2044 vm_stat_account(mm, vma->vm_flags, grow);
2045 anon_vma_interval_tree_pre_update_vma(vma);
2046 vma->vm_end = address;
2047 /* Overwrite old entry in mtree. */
2048 mas_store_prealloc(&mas, vma);
2049 anon_vma_interval_tree_post_update_vma(vma);
2050 spin_unlock(&mm->page_table_lock);
2052 perf_event_mmap(vma);
2056 anon_vma_unlock_write(vma->anon_vma);
2061 #endif /* CONFIG_STACK_GROWSUP */
2064 * vma is the first one with address < vma->vm_start. Have to extend vma.
2065 * mmap_lock held for writing.
2067 int expand_downwards(struct vm_area_struct *vma, unsigned long address)
2069 struct mm_struct *mm = vma->vm_mm;
2070 MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start);
2071 struct vm_area_struct *prev;
2074 if (!(vma->vm_flags & VM_GROWSDOWN))
2077 address &= PAGE_MASK;
2078 if (address < mmap_min_addr || address < FIRST_USER_ADDRESS)
2081 /* Enforce stack_guard_gap */
2082 prev = mas_prev(&mas, 0);
2083 /* Check that both stack segments have the same anon_vma? */
2085 if (!(prev->vm_flags & VM_GROWSDOWN) &&
2086 vma_is_accessible(prev) &&
2087 (address - prev->vm_end < stack_guard_gap))
2092 mas_next_range(&mas, vma->vm_start);
2094 __mas_set_range(&mas, address, vma->vm_end - 1);
2095 if (mas_preallocate(&mas, vma, GFP_KERNEL))
2098 /* We must make sure the anon_vma is allocated. */
2099 if (unlikely(anon_vma_prepare(vma))) {
2104 /* Lock the VMA before expanding to prevent concurrent page faults */
2105 vma_start_write(vma);
2107 * vma->vm_start/vm_end cannot change under us because the caller
2108 * is required to hold the mmap_lock in read mode. We need the
2109 * anon_vma lock to serialize against concurrent expand_stacks.
2111 anon_vma_lock_write(vma->anon_vma);
2113 /* Somebody else might have raced and expanded it already */
2114 if (address < vma->vm_start) {
2115 unsigned long size, grow;
2117 size = vma->vm_end - address;
2118 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2121 if (grow <= vma->vm_pgoff) {
2122 error = acct_stack_growth(vma, size, grow);
2125 * We only hold a shared mmap_lock lock here, so
2126 * we need to protect against concurrent vma
2127 * expansions. anon_vma_lock_write() doesn't
2128 * help here, as we don't guarantee that all
2129 * growable vmas in a mm share the same root
2130 * anon vma. So, we reuse mm->page_table_lock
2131 * to guard against concurrent vma expansions.
2133 spin_lock(&mm->page_table_lock);
2134 if (vma->vm_flags & VM_LOCKED)
2135 mm->locked_vm += grow;
2136 vm_stat_account(mm, vma->vm_flags, grow);
2137 anon_vma_interval_tree_pre_update_vma(vma);
2138 vma->vm_start = address;
2139 vma->vm_pgoff -= grow;
2140 /* Overwrite old entry in mtree. */
2141 mas_store_prealloc(&mas, vma);
2142 anon_vma_interval_tree_post_update_vma(vma);
2143 spin_unlock(&mm->page_table_lock);
2145 perf_event_mmap(vma);
2149 anon_vma_unlock_write(vma->anon_vma);
2155 /* enforced gap between the expanding stack and other mappings. */
2156 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2158 static int __init cmdline_parse_stack_guard_gap(char *p)
2163 val = simple_strtoul(p, &endptr, 10);
2165 stack_guard_gap = val << PAGE_SHIFT;
2169 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2171 #ifdef CONFIG_STACK_GROWSUP
2172 int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2174 return expand_upwards(vma, address);
2177 struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2179 struct vm_area_struct *vma, *prev;
2182 vma = find_vma_prev(mm, addr, &prev);
2183 if (vma && (vma->vm_start <= addr))
2187 if (expand_stack_locked(prev, addr))
2189 if (prev->vm_flags & VM_LOCKED)
2190 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2194 int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2196 return expand_downwards(vma, address);
2199 struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2201 struct vm_area_struct *vma;
2202 unsigned long start;
2205 vma = find_vma(mm, addr);
2208 if (vma->vm_start <= addr)
2210 start = vma->vm_start;
2211 if (expand_stack_locked(vma, addr))
2213 if (vma->vm_flags & VM_LOCKED)
2214 populate_vma_page_range(vma, addr, start, NULL);
2219 #if defined(CONFIG_STACK_GROWSUP)
2221 #define vma_expand_up(vma,addr) expand_upwards(vma, addr)
2222 #define vma_expand_down(vma, addr) (-EFAULT)
2226 #define vma_expand_up(vma,addr) (-EFAULT)
2227 #define vma_expand_down(vma, addr) expand_downwards(vma, addr)
2232 * expand_stack(): legacy interface for page faulting. Don't use unless
2235 * This is called with the mm locked for reading, drops the lock, takes
2236 * the lock for writing, tries to look up a vma again, expands it if
2237 * necessary, and downgrades the lock to reading again.
2239 * If no vma is found or it can't be expanded, it returns NULL and has
2242 struct vm_area_struct *expand_stack(struct mm_struct *mm, unsigned long addr)
2244 struct vm_area_struct *vma, *prev;
2246 mmap_read_unlock(mm);
2247 if (mmap_write_lock_killable(mm))
2250 vma = find_vma_prev(mm, addr, &prev);
2251 if (vma && vma->vm_start <= addr)
2254 if (prev && !vma_expand_up(prev, addr)) {
2259 if (vma && !vma_expand_down(vma, addr))
2262 mmap_write_unlock(mm);
2266 mmap_write_downgrade(mm);
2271 * Ok - we have the memory areas we should free on a maple tree so release them,
2272 * and do the vma updates.
2274 * Called with the mm semaphore held.
2276 static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas)
2278 unsigned long nr_accounted = 0;
2279 struct vm_area_struct *vma;
2281 /* Update high watermark before we lower total_vm */
2282 update_hiwater_vm(mm);
2283 mas_for_each(mas, vma, ULONG_MAX) {
2284 long nrpages = vma_pages(vma);
2286 if (vma->vm_flags & VM_ACCOUNT)
2287 nr_accounted += nrpages;
2288 vm_stat_account(mm, vma->vm_flags, -nrpages);
2289 remove_vma(vma, false);
2291 vm_unacct_memory(nr_accounted);
2295 * Get rid of page table information in the indicated region.
2297 * Called with the mm semaphore held.
2299 static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
2300 struct vm_area_struct *vma, struct vm_area_struct *prev,
2301 struct vm_area_struct *next, unsigned long start,
2302 unsigned long end, unsigned long tree_end, bool mm_wr_locked)
2304 struct mmu_gather tlb;
2305 unsigned long mt_start = mas->index;
2308 tlb_gather_mmu(&tlb, mm);
2309 update_hiwater_rss(mm);
2310 unmap_vmas(&tlb, mas, vma, start, end, tree_end, mm_wr_locked);
2311 mas_set(mas, mt_start);
2312 free_pgtables(&tlb, mas, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2313 next ? next->vm_start : USER_PGTABLES_CEILING,
2315 tlb_finish_mmu(&tlb);
2319 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2320 * has already been checked or doesn't make sense to fail.
2321 * VMA Iterator will point to the end VMA.
2323 static int __split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2324 unsigned long addr, int new_below)
2326 struct vma_prepare vp;
2327 struct vm_area_struct *new;
2330 WARN_ON(vma->vm_start >= addr);
2331 WARN_ON(vma->vm_end <= addr);
2333 if (vma->vm_ops && vma->vm_ops->may_split) {
2334 err = vma->vm_ops->may_split(vma, addr);
2339 new = vm_area_dup(vma);
2346 new->vm_start = addr;
2347 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2351 vma_iter_config(vmi, new->vm_start, new->vm_end);
2352 if (vma_iter_prealloc(vmi, new))
2355 err = vma_dup_policy(vma, new);
2359 err = anon_vma_clone(new, vma);
2364 get_file(new->vm_file);
2366 if (new->vm_ops && new->vm_ops->open)
2367 new->vm_ops->open(new);
2369 vma_start_write(vma);
2370 vma_start_write(new);
2372 init_vma_prep(&vp, vma);
2375 vma_adjust_trans_huge(vma, vma->vm_start, addr, 0);
2378 vma->vm_start = addr;
2379 vma->vm_pgoff += (addr - new->vm_start) >> PAGE_SHIFT;
2384 /* vma_complete stores the new vma */
2385 vma_complete(&vp, vmi, vma->vm_mm);
2393 mpol_put(vma_policy(new));
2402 * Split a vma into two pieces at address 'addr', a new vma is allocated
2403 * either for the first part or the tail.
2405 static int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2406 unsigned long addr, int new_below)
2408 if (vma->vm_mm->map_count >= sysctl_max_map_count)
2411 return __split_vma(vmi, vma, addr, new_below);
2415 * We are about to modify one or multiple of a VMA's flags, policy, userfaultfd
2416 * context and anonymous VMA name within the range [start, end).
2418 * As a result, we might be able to merge the newly modified VMA range with an
2419 * adjacent VMA with identical properties.
2421 * If no merge is possible and the range does not span the entirety of the VMA,
2422 * we then need to split the VMA to accommodate the change.
2424 * The function returns either the merged VMA, the original VMA if a split was
2425 * required instead, or an error if the split failed.
2427 struct vm_area_struct *vma_modify(struct vma_iterator *vmi,
2428 struct vm_area_struct *prev,
2429 struct vm_area_struct *vma,
2430 unsigned long start, unsigned long end,
2431 unsigned long vm_flags,
2432 struct mempolicy *policy,
2433 struct vm_userfaultfd_ctx uffd_ctx,
2434 struct anon_vma_name *anon_name)
2436 pgoff_t pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
2437 struct vm_area_struct *merged;
2439 merged = vma_merge(vmi, prev, vma, start, end, vm_flags,
2440 pgoff, policy, uffd_ctx, anon_name);
2444 if (vma->vm_start < start) {
2445 int err = split_vma(vmi, vma, start, 1);
2448 return ERR_PTR(err);
2451 if (vma->vm_end > end) {
2452 int err = split_vma(vmi, vma, end, 0);
2455 return ERR_PTR(err);
2462 * Attempt to merge a newly mapped VMA with those adjacent to it. The caller
2463 * must ensure that [start, end) does not overlap any existing VMA.
2465 static struct vm_area_struct
2466 *vma_merge_new_vma(struct vma_iterator *vmi, struct vm_area_struct *prev,
2467 struct vm_area_struct *vma, unsigned long start,
2468 unsigned long end, pgoff_t pgoff)
2470 return vma_merge(vmi, prev, vma, start, end, vma->vm_flags, pgoff,
2471 vma_policy(vma), vma->vm_userfaultfd_ctx, anon_vma_name(vma));
2475 * Expand vma by delta bytes, potentially merging with an immediately adjacent
2476 * VMA with identical properties.
2478 struct vm_area_struct *vma_merge_extend(struct vma_iterator *vmi,
2479 struct vm_area_struct *vma,
2480 unsigned long delta)
2482 pgoff_t pgoff = vma->vm_pgoff + vma_pages(vma);
2484 /* vma is specified as prev, so case 1 or 2 will apply. */
2485 return vma_merge(vmi, vma, vma, vma->vm_end, vma->vm_end + delta,
2486 vma->vm_flags, pgoff, vma_policy(vma),
2487 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
2491 * do_vmi_align_munmap() - munmap the aligned region from @start to @end.
2492 * @vmi: The vma iterator
2493 * @vma: The starting vm_area_struct
2494 * @mm: The mm_struct
2495 * @start: The aligned start address to munmap.
2496 * @end: The aligned end address to munmap.
2497 * @uf: The userfaultfd list_head
2498 * @unlock: Set to true to drop the mmap_lock. unlocking only happens on
2501 * Return: 0 on success and drops the lock if so directed, error and leaves the
2502 * lock held otherwise.
2505 do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
2506 struct mm_struct *mm, unsigned long start,
2507 unsigned long end, struct list_head *uf, bool unlock)
2509 struct vm_area_struct *prev, *next = NULL;
2510 struct maple_tree mt_detach;
2512 int error = -ENOMEM;
2513 unsigned long locked_vm = 0;
2514 MA_STATE(mas_detach, &mt_detach, 0, 0);
2515 mt_init_flags(&mt_detach, vmi->mas.tree->ma_flags & MT_FLAGS_LOCK_MASK);
2516 mt_on_stack(mt_detach);
2519 * If we need to split any vma, do it now to save pain later.
2521 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2522 * unmapped vm_area_struct will remain in use: so lower split_vma
2523 * places tmp vma above, and higher split_vma places tmp vma below.
2526 /* Does it split the first one? */
2527 if (start > vma->vm_start) {
2530 * Make sure that map_count on return from munmap() will
2531 * not exceed its limit; but let map_count go just above
2532 * its limit temporarily, to help free resources as expected.
2534 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2535 goto map_count_exceeded;
2537 error = __split_vma(vmi, vma, start, 1);
2539 goto start_split_failed;
2543 * Detach a range of VMAs from the mm. Using next as a temp variable as
2544 * it is always overwritten.
2548 /* Does it split the end? */
2549 if (next->vm_end > end) {
2550 error = __split_vma(vmi, next, end, 0);
2552 goto end_split_failed;
2554 vma_start_write(next);
2555 mas_set(&mas_detach, count);
2556 error = mas_store_gfp(&mas_detach, next, GFP_KERNEL);
2558 goto munmap_gather_failed;
2559 vma_mark_detached(next, true);
2560 if (next->vm_flags & VM_LOCKED)
2561 locked_vm += vma_pages(next);
2566 * If userfaultfd_unmap_prep returns an error the vmas
2567 * will remain split, but userland will get a
2568 * highly unexpected error anyway. This is no
2569 * different than the case where the first of the two
2570 * __split_vma fails, but we don't undo the first
2571 * split, despite we could. This is unlikely enough
2572 * failure that it's not worth optimizing it for.
2574 error = userfaultfd_unmap_prep(next, start, end, uf);
2577 goto userfaultfd_error;
2579 #ifdef CONFIG_DEBUG_VM_MAPLE_TREE
2580 BUG_ON(next->vm_start < start);
2581 BUG_ON(next->vm_start > end);
2583 } for_each_vma_range(*vmi, next, end);
2585 #if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
2586 /* Make sure no VMAs are about to be lost. */
2588 MA_STATE(test, &mt_detach, 0, 0);
2589 struct vm_area_struct *vma_mas, *vma_test;
2592 vma_iter_set(vmi, start);
2594 vma_test = mas_find(&test, count - 1);
2595 for_each_vma_range(*vmi, vma_mas, end) {
2596 BUG_ON(vma_mas != vma_test);
2598 vma_test = mas_next(&test, count - 1);
2601 BUG_ON(count != test_count);
2605 while (vma_iter_addr(vmi) > start)
2606 vma_iter_prev_range(vmi);
2608 error = vma_iter_clear_gfp(vmi, start, end, GFP_KERNEL);
2610 goto clear_tree_failed;
2612 /* Point of no return */
2613 mm->locked_vm -= locked_vm;
2614 mm->map_count -= count;
2616 mmap_write_downgrade(mm);
2618 prev = vma_iter_prev_range(vmi);
2619 next = vma_next(vmi);
2621 vma_iter_prev_range(vmi);
2624 * We can free page tables without write-locking mmap_lock because VMAs
2625 * were isolated before we downgraded mmap_lock.
2627 mas_set(&mas_detach, 1);
2628 unmap_region(mm, &mas_detach, vma, prev, next, start, end, count,
2630 /* Statistics and freeing VMAs */
2631 mas_set(&mas_detach, 0);
2632 remove_mt(mm, &mas_detach);
2635 mmap_read_unlock(mm);
2637 __mt_destroy(&mt_detach);
2642 munmap_gather_failed:
2644 mas_set(&mas_detach, 0);
2645 mas_for_each(&mas_detach, next, end)
2646 vma_mark_detached(next, false);
2648 __mt_destroy(&mt_detach);
2656 * do_vmi_munmap() - munmap a given range.
2657 * @vmi: The vma iterator
2658 * @mm: The mm_struct
2659 * @start: The start address to munmap
2660 * @len: The length of the range to munmap
2661 * @uf: The userfaultfd list_head
2662 * @unlock: set to true if the user wants to drop the mmap_lock on success
2664 * This function takes a @mas that is either pointing to the previous VMA or set
2665 * to MA_START and sets it up to remove the mapping(s). The @len will be
2666 * aligned and any arch_unmap work will be preformed.
2668 * Return: 0 on success and drops the lock if so directed, error and leaves the
2669 * lock held otherwise.
2671 int do_vmi_munmap(struct vma_iterator *vmi, struct mm_struct *mm,
2672 unsigned long start, size_t len, struct list_head *uf,
2676 struct vm_area_struct *vma;
2678 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2681 end = start + PAGE_ALIGN(len);
2685 /* arch_unmap() might do unmaps itself. */
2686 arch_unmap(mm, start, end);
2688 /* Find the first overlapping VMA */
2689 vma = vma_find(vmi, end);
2692 mmap_write_unlock(mm);
2696 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
2699 /* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls.
2700 * @mm: The mm_struct
2701 * @start: The start address to munmap
2702 * @len: The length to be munmapped.
2703 * @uf: The userfaultfd list_head
2705 * Return: 0 on success, error otherwise.
2707 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2708 struct list_head *uf)
2710 VMA_ITERATOR(vmi, mm, start);
2712 return do_vmi_munmap(&vmi, mm, start, len, uf, false);
2715 unsigned long mmap_region(struct file *file, unsigned long addr,
2716 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
2717 struct list_head *uf)
2719 struct mm_struct *mm = current->mm;
2720 struct vm_area_struct *vma = NULL;
2721 struct vm_area_struct *next, *prev, *merge;
2722 pgoff_t pglen = len >> PAGE_SHIFT;
2723 unsigned long charged = 0;
2724 unsigned long end = addr + len;
2725 unsigned long merge_start = addr, merge_end = end;
2726 bool writable_file_mapping = false;
2729 VMA_ITERATOR(vmi, mm, addr);
2731 /* Check against address space limit. */
2732 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
2733 unsigned long nr_pages;
2736 * MAP_FIXED may remove pages of mappings that intersects with
2737 * requested mapping. Account for the pages it would unmap.
2739 nr_pages = count_vma_pages_range(mm, addr, end);
2741 if (!may_expand_vm(mm, vm_flags,
2742 (len >> PAGE_SHIFT) - nr_pages))
2746 /* Unmap any existing mapping in the area */
2747 if (do_vmi_munmap(&vmi, mm, addr, len, uf, false))
2751 * Private writable mapping: check memory availability
2753 if (accountable_mapping(file, vm_flags)) {
2754 charged = len >> PAGE_SHIFT;
2755 if (security_vm_enough_memory_mm(mm, charged))
2757 vm_flags |= VM_ACCOUNT;
2760 next = vma_next(&vmi);
2761 prev = vma_prev(&vmi);
2762 if (vm_flags & VM_SPECIAL) {
2764 vma_iter_next_range(&vmi);
2768 /* Attempt to expand an old mapping */
2770 if (next && next->vm_start == end && !vma_policy(next) &&
2771 can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen,
2772 NULL_VM_UFFD_CTX, NULL)) {
2773 merge_end = next->vm_end;
2775 vm_pgoff = next->vm_pgoff - pglen;
2779 if (prev && prev->vm_end == addr && !vma_policy(prev) &&
2780 (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file,
2781 pgoff, vma->vm_userfaultfd_ctx, NULL) :
2782 can_vma_merge_after(prev, vm_flags, NULL, file, pgoff,
2783 NULL_VM_UFFD_CTX, NULL))) {
2784 merge_start = prev->vm_start;
2786 vm_pgoff = prev->vm_pgoff;
2788 vma_iter_next_range(&vmi);
2791 /* Actually expand, if possible */
2793 !vma_expand(&vmi, vma, merge_start, merge_end, vm_pgoff, next)) {
2794 khugepaged_enter_vma(vma, vm_flags);
2799 vma_iter_set(&vmi, addr);
2803 * Determine the object being mapped and call the appropriate
2804 * specific mapper. the address has already been validated, but
2805 * not unmapped, but the maps are removed from the list.
2807 vma = vm_area_alloc(mm);
2813 vma_iter_config(&vmi, addr, end);
2814 vma_set_range(vma, addr, end, pgoff);
2815 vm_flags_init(vma, vm_flags);
2816 vma->vm_page_prot = vm_get_page_prot(vm_flags);
2819 vma->vm_file = get_file(file);
2820 error = call_mmap(file, vma);
2822 goto unmap_and_free_vma;
2824 if (vma_is_shared_maywrite(vma)) {
2825 error = mapping_map_writable(file->f_mapping);
2827 goto close_and_free_vma;
2829 writable_file_mapping = true;
2833 * Expansion is handled above, merging is handled below.
2834 * Drivers should not alter the address of the VMA.
2837 if (WARN_ON((addr != vma->vm_start)))
2838 goto close_and_free_vma;
2840 vma_iter_config(&vmi, addr, end);
2842 * If vm_flags changed after call_mmap(), we should try merge
2843 * vma again as we may succeed this time.
2845 if (unlikely(vm_flags != vma->vm_flags && prev)) {
2846 merge = vma_merge_new_vma(&vmi, prev, vma,
2847 vma->vm_start, vma->vm_end,
2851 * ->mmap() can change vma->vm_file and fput
2852 * the original file. So fput the vma->vm_file
2853 * here or we would add an extra fput for file
2854 * and cause general protection fault
2860 /* Update vm_flags to pick up the change. */
2861 vm_flags = vma->vm_flags;
2862 goto unmap_writable;
2866 vm_flags = vma->vm_flags;
2867 } else if (vm_flags & VM_SHARED) {
2868 error = shmem_zero_setup(vma);
2872 vma_set_anonymous(vma);
2875 if (map_deny_write_exec(vma, vma->vm_flags)) {
2877 goto close_and_free_vma;
2880 /* Allow architectures to sanity-check the vm_flags */
2882 if (!arch_validate_flags(vma->vm_flags))
2883 goto close_and_free_vma;
2886 if (vma_iter_prealloc(&vmi, vma))
2887 goto close_and_free_vma;
2889 /* Lock the VMA since it is modified after insertion into VMA tree */
2890 vma_start_write(vma);
2891 vma_iter_store(&vmi, vma);
2896 * vma_merge() calls khugepaged_enter_vma() either, the below
2897 * call covers the non-merge case.
2899 khugepaged_enter_vma(vma, vma->vm_flags);
2901 /* Once vma denies write, undo our temporary denial count */
2903 if (writable_file_mapping)
2904 mapping_unmap_writable(file->f_mapping);
2905 file = vma->vm_file;
2908 perf_event_mmap(vma);
2910 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
2911 if (vm_flags & VM_LOCKED) {
2912 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
2913 is_vm_hugetlb_page(vma) ||
2914 vma == get_gate_vma(current->mm))
2915 vm_flags_clear(vma, VM_LOCKED_MASK);
2917 mm->locked_vm += (len >> PAGE_SHIFT);
2924 * New (or expanded) vma always get soft dirty status.
2925 * Otherwise user-space soft-dirty page tracker won't
2926 * be able to distinguish situation when vma area unmapped,
2927 * then new mapped in-place (which must be aimed as
2928 * a completely new data area).
2930 vm_flags_set(vma, VM_SOFTDIRTY);
2932 vma_set_page_prot(vma);
2938 if (file && vma->vm_ops && vma->vm_ops->close)
2939 vma->vm_ops->close(vma);
2941 if (file || vma->vm_file) {
2944 vma->vm_file = NULL;
2946 vma_iter_set(&vmi, vma->vm_end);
2947 /* Undo any partial mapping done by a device driver. */
2948 unmap_region(mm, &vmi.mas, vma, prev, next, vma->vm_start,
2949 vma->vm_end, vma->vm_end, true);
2951 if (writable_file_mapping)
2952 mapping_unmap_writable(file->f_mapping);
2957 vm_unacct_memory(charged);
2962 static int __vm_munmap(unsigned long start, size_t len, bool unlock)
2965 struct mm_struct *mm = current->mm;
2967 VMA_ITERATOR(vmi, mm, start);
2969 if (mmap_write_lock_killable(mm))
2972 ret = do_vmi_munmap(&vmi, mm, start, len, &uf, unlock);
2974 mmap_write_unlock(mm);
2976 userfaultfd_unmap_complete(mm, &uf);
2980 int vm_munmap(unsigned long start, size_t len)
2982 return __vm_munmap(start, len, false);
2984 EXPORT_SYMBOL(vm_munmap);
2986 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2988 addr = untagged_addr(addr);
2989 return __vm_munmap(addr, len, true);
2994 * Emulation of deprecated remap_file_pages() syscall.
2996 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2997 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
3000 struct mm_struct *mm = current->mm;
3001 struct vm_area_struct *vma;
3002 unsigned long populate = 0;
3003 unsigned long ret = -EINVAL;
3006 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
3007 current->comm, current->pid);
3011 start = start & PAGE_MASK;
3012 size = size & PAGE_MASK;
3014 if (start + size <= start)
3017 /* Does pgoff wrap? */
3018 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
3021 if (mmap_write_lock_killable(mm))
3024 vma = vma_lookup(mm, start);
3026 if (!vma || !(vma->vm_flags & VM_SHARED))
3029 if (start + size > vma->vm_end) {
3030 VMA_ITERATOR(vmi, mm, vma->vm_end);
3031 struct vm_area_struct *next, *prev = vma;
3033 for_each_vma_range(vmi, next, start + size) {
3034 /* hole between vmas ? */
3035 if (next->vm_start != prev->vm_end)
3038 if (next->vm_file != vma->vm_file)
3041 if (next->vm_flags != vma->vm_flags)
3044 if (start + size <= next->vm_end)
3054 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3055 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3056 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3058 flags &= MAP_NONBLOCK;
3059 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3060 if (vma->vm_flags & VM_LOCKED)
3061 flags |= MAP_LOCKED;
3063 file = get_file(vma->vm_file);
3064 ret = do_mmap(vma->vm_file, start, size,
3065 prot, flags, 0, pgoff, &populate, NULL);
3068 mmap_write_unlock(mm);
3070 mm_populate(ret, populate);
3071 if (!IS_ERR_VALUE(ret))
3077 * do_vma_munmap() - Unmap a full or partial vma.
3078 * @vmi: The vma iterator pointing at the vma
3079 * @vma: The first vma to be munmapped
3080 * @start: the start of the address to unmap
3081 * @end: The end of the address to unmap
3082 * @uf: The userfaultfd list_head
3083 * @unlock: Drop the lock on success
3085 * unmaps a VMA mapping when the vma iterator is already in position.
3086 * Does not handle alignment.
3088 * Return: 0 on success drops the lock of so directed, error on failure and will
3089 * still hold the lock.
3091 int do_vma_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
3092 unsigned long start, unsigned long end, struct list_head *uf,
3095 struct mm_struct *mm = vma->vm_mm;
3097 arch_unmap(mm, start, end);
3098 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
3102 * do_brk_flags() - Increase the brk vma if the flags match.
3103 * @vmi: The vma iterator
3104 * @addr: The start address
3105 * @len: The length of the increase
3107 * @flags: The VMA Flags
3109 * Extend the brk VMA from addr to addr + len. If the VMA is NULL or the flags
3110 * do not match then create a new anonymous VMA. Eventually we may be able to
3111 * do some brk-specific accounting here.
3113 static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *vma,
3114 unsigned long addr, unsigned long len, unsigned long flags)
3116 struct mm_struct *mm = current->mm;
3117 struct vma_prepare vp;
3120 * Check against address space limits by the changed size
3121 * Note: This happens *after* clearing old mappings in some code paths.
3123 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3124 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3127 if (mm->map_count > sysctl_max_map_count)
3130 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3134 * Expand the existing vma if possible; Note that singular lists do not
3135 * occur after forking, so the expand will only happen on new VMAs.
3137 if (vma && vma->vm_end == addr && !vma_policy(vma) &&
3138 can_vma_merge_after(vma, flags, NULL, NULL,
3139 addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) {
3140 vma_iter_config(vmi, vma->vm_start, addr + len);
3141 if (vma_iter_prealloc(vmi, vma))
3144 vma_start_write(vma);
3146 init_vma_prep(&vp, vma);
3148 vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0);
3149 vma->vm_end = addr + len;
3150 vm_flags_set(vma, VM_SOFTDIRTY);
3151 vma_iter_store(vmi, vma);
3153 vma_complete(&vp, vmi, mm);
3154 khugepaged_enter_vma(vma, flags);
3159 vma_iter_next_range(vmi);
3160 /* create a vma struct for an anonymous mapping */
3161 vma = vm_area_alloc(mm);
3165 vma_set_anonymous(vma);
3166 vma_set_range(vma, addr, addr + len, addr >> PAGE_SHIFT);
3167 vm_flags_init(vma, flags);
3168 vma->vm_page_prot = vm_get_page_prot(flags);
3169 vma_start_write(vma);
3170 if (vma_iter_store_gfp(vmi, vma, GFP_KERNEL))
3171 goto mas_store_fail;
3177 perf_event_mmap(vma);
3178 mm->total_vm += len >> PAGE_SHIFT;
3179 mm->data_vm += len >> PAGE_SHIFT;
3180 if (flags & VM_LOCKED)
3181 mm->locked_vm += (len >> PAGE_SHIFT);
3182 vm_flags_set(vma, VM_SOFTDIRTY);
3188 vm_unacct_memory(len >> PAGE_SHIFT);
3192 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3194 struct mm_struct *mm = current->mm;
3195 struct vm_area_struct *vma = NULL;
3200 VMA_ITERATOR(vmi, mm, addr);
3202 len = PAGE_ALIGN(request);
3208 /* Until we need other flags, refuse anything except VM_EXEC. */
3209 if ((flags & (~VM_EXEC)) != 0)
3212 if (mmap_write_lock_killable(mm))
3215 ret = check_brk_limits(addr, len);
3219 ret = do_vmi_munmap(&vmi, mm, addr, len, &uf, 0);
3223 vma = vma_prev(&vmi);
3224 ret = do_brk_flags(&vmi, vma, addr, len, flags);
3225 populate = ((mm->def_flags & VM_LOCKED) != 0);
3226 mmap_write_unlock(mm);
3227 userfaultfd_unmap_complete(mm, &uf);
3228 if (populate && !ret)
3229 mm_populate(addr, len);
3234 mmap_write_unlock(mm);
3237 EXPORT_SYMBOL(vm_brk_flags);
3239 /* Release all mmaps. */
3240 void exit_mmap(struct mm_struct *mm)
3242 struct mmu_gather tlb;
3243 struct vm_area_struct *vma;
3244 unsigned long nr_accounted = 0;
3245 MA_STATE(mas, &mm->mm_mt, 0, 0);
3248 /* mm's last user has gone, and its about to be pulled down */
3249 mmu_notifier_release(mm);
3254 vma = mas_find(&mas, ULONG_MAX);
3255 if (!vma || unlikely(xa_is_zero(vma))) {
3256 /* Can happen if dup_mmap() received an OOM */
3257 mmap_read_unlock(mm);
3258 mmap_write_lock(mm);
3264 tlb_gather_mmu_fullmm(&tlb, mm);
3265 /* update_hiwater_rss(mm) here? but nobody should be looking */
3266 /* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */
3267 unmap_vmas(&tlb, &mas, vma, 0, ULONG_MAX, ULONG_MAX, false);
3268 mmap_read_unlock(mm);
3271 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
3272 * because the memory has been already freed.
3274 set_bit(MMF_OOM_SKIP, &mm->flags);
3275 mmap_write_lock(mm);
3276 mt_clear_in_rcu(&mm->mm_mt);
3277 mas_set(&mas, vma->vm_end);
3278 free_pgtables(&tlb, &mas, vma, FIRST_USER_ADDRESS,
3279 USER_PGTABLES_CEILING, true);
3280 tlb_finish_mmu(&tlb);
3283 * Walk the list again, actually closing and freeing it, with preemption
3284 * enabled, without holding any MM locks besides the unreachable
3287 mas_set(&mas, vma->vm_end);
3289 if (vma->vm_flags & VM_ACCOUNT)
3290 nr_accounted += vma_pages(vma);
3291 remove_vma(vma, true);
3294 vma = mas_find(&mas, ULONG_MAX);
3295 } while (vma && likely(!xa_is_zero(vma)));
3297 BUG_ON(count != mm->map_count);
3299 trace_exit_mmap(mm);
3301 __mt_destroy(&mm->mm_mt);
3302 mmap_write_unlock(mm);
3303 vm_unacct_memory(nr_accounted);
3306 /* Insert vm structure into process list sorted by address
3307 * and into the inode's i_mmap tree. If vm_file is non-NULL
3308 * then i_mmap_rwsem is taken here.
3310 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3312 unsigned long charged = vma_pages(vma);
3315 if (find_vma_intersection(mm, vma->vm_start, vma->vm_end))
3318 if ((vma->vm_flags & VM_ACCOUNT) &&
3319 security_vm_enough_memory_mm(mm, charged))
3323 * The vm_pgoff of a purely anonymous vma should be irrelevant
3324 * until its first write fault, when page's anon_vma and index
3325 * are set. But now set the vm_pgoff it will almost certainly
3326 * end up with (unless mremap moves it elsewhere before that
3327 * first wfault), so /proc/pid/maps tells a consistent story.
3329 * By setting it to reflect the virtual start address of the
3330 * vma, merges and splits can happen in a seamless way, just
3331 * using the existing file pgoff checks and manipulations.
3332 * Similarly in do_mmap and in do_brk_flags.
3334 if (vma_is_anonymous(vma)) {
3335 BUG_ON(vma->anon_vma);
3336 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3339 if (vma_link(mm, vma)) {
3340 if (vma->vm_flags & VM_ACCOUNT)
3341 vm_unacct_memory(charged);
3349 * Copy the vma structure to a new location in the same mm,
3350 * prior to moving page table entries, to effect an mremap move.
3352 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3353 unsigned long addr, unsigned long len, pgoff_t pgoff,
3354 bool *need_rmap_locks)
3356 struct vm_area_struct *vma = *vmap;
3357 unsigned long vma_start = vma->vm_start;
3358 struct mm_struct *mm = vma->vm_mm;
3359 struct vm_area_struct *new_vma, *prev;
3360 bool faulted_in_anon_vma = true;
3361 VMA_ITERATOR(vmi, mm, addr);
3364 * If anonymous vma has not yet been faulted, update new pgoff
3365 * to match new location, to increase its chance of merging.
3367 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3368 pgoff = addr >> PAGE_SHIFT;
3369 faulted_in_anon_vma = false;
3372 new_vma = find_vma_prev(mm, addr, &prev);
3373 if (new_vma && new_vma->vm_start < addr + len)
3374 return NULL; /* should never get here */
3376 new_vma = vma_merge_new_vma(&vmi, prev, vma, addr, addr + len, pgoff);
3379 * Source vma may have been merged into new_vma
3381 if (unlikely(vma_start >= new_vma->vm_start &&
3382 vma_start < new_vma->vm_end)) {
3384 * The only way we can get a vma_merge with
3385 * self during an mremap is if the vma hasn't
3386 * been faulted in yet and we were allowed to
3387 * reset the dst vma->vm_pgoff to the
3388 * destination address of the mremap to allow
3389 * the merge to happen. mremap must change the
3390 * vm_pgoff linearity between src and dst vmas
3391 * (in turn preventing a vma_merge) to be
3392 * safe. It is only safe to keep the vm_pgoff
3393 * linear if there are no pages mapped yet.
3395 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3396 *vmap = vma = new_vma;
3398 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3400 new_vma = vm_area_dup(vma);
3403 vma_set_range(new_vma, addr, addr + len, pgoff);
3404 if (vma_dup_policy(vma, new_vma))
3406 if (anon_vma_clone(new_vma, vma))
3407 goto out_free_mempol;
3408 if (new_vma->vm_file)
3409 get_file(new_vma->vm_file);
3410 if (new_vma->vm_ops && new_vma->vm_ops->open)
3411 new_vma->vm_ops->open(new_vma);
3412 if (vma_link(mm, new_vma))
3414 *need_rmap_locks = false;
3419 if (new_vma->vm_ops && new_vma->vm_ops->close)
3420 new_vma->vm_ops->close(new_vma);
3422 if (new_vma->vm_file)
3423 fput(new_vma->vm_file);
3425 unlink_anon_vmas(new_vma);
3427 mpol_put(vma_policy(new_vma));
3429 vm_area_free(new_vma);
3435 * Return true if the calling process may expand its vm space by the passed
3438 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3440 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3443 if (is_data_mapping(flags) &&
3444 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3445 /* Workaround for Valgrind */
3446 if (rlimit(RLIMIT_DATA) == 0 &&
3447 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3450 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3451 current->comm, current->pid,
3452 (mm->data_vm + npages) << PAGE_SHIFT,
3453 rlimit(RLIMIT_DATA),
3454 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3456 if (!ignore_rlimit_data)
3463 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3465 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3467 if (is_exec_mapping(flags))
3468 mm->exec_vm += npages;
3469 else if (is_stack_mapping(flags))
3470 mm->stack_vm += npages;
3471 else if (is_data_mapping(flags))
3472 mm->data_vm += npages;
3475 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3478 * Having a close hook prevents vma merging regardless of flags.
3480 static void special_mapping_close(struct vm_area_struct *vma)
3484 static const char *special_mapping_name(struct vm_area_struct *vma)
3486 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3489 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3491 struct vm_special_mapping *sm = new_vma->vm_private_data;
3493 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3497 return sm->mremap(sm, new_vma);
3502 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3505 * Forbid splitting special mappings - kernel has expectations over
3506 * the number of pages in mapping. Together with VM_DONTEXPAND
3507 * the size of vma should stay the same over the special mapping's
3513 static const struct vm_operations_struct special_mapping_vmops = {
3514 .close = special_mapping_close,
3515 .fault = special_mapping_fault,
3516 .mremap = special_mapping_mremap,
3517 .name = special_mapping_name,
3518 /* vDSO code relies that VVAR can't be accessed remotely */
3520 .may_split = special_mapping_split,
3523 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3524 .close = special_mapping_close,
3525 .fault = special_mapping_fault,
3528 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3530 struct vm_area_struct *vma = vmf->vma;
3532 struct page **pages;
3534 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3535 pages = vma->vm_private_data;
3537 struct vm_special_mapping *sm = vma->vm_private_data;
3540 return sm->fault(sm, vmf->vma, vmf);
3545 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3549 struct page *page = *pages;
3555 return VM_FAULT_SIGBUS;
3558 static struct vm_area_struct *__install_special_mapping(
3559 struct mm_struct *mm,
3560 unsigned long addr, unsigned long len,
3561 unsigned long vm_flags, void *priv,
3562 const struct vm_operations_struct *ops)
3565 struct vm_area_struct *vma;
3567 vma = vm_area_alloc(mm);
3568 if (unlikely(vma == NULL))
3569 return ERR_PTR(-ENOMEM);
3571 vma_set_range(vma, addr, addr + len, 0);
3572 vm_flags_init(vma, (vm_flags | mm->def_flags |
3573 VM_DONTEXPAND | VM_SOFTDIRTY) & ~VM_LOCKED_MASK);
3574 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3577 vma->vm_private_data = priv;
3579 ret = insert_vm_struct(mm, vma);
3583 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3585 perf_event_mmap(vma);
3591 return ERR_PTR(ret);
3594 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3595 const struct vm_special_mapping *sm)
3597 return vma->vm_private_data == sm &&
3598 (vma->vm_ops == &special_mapping_vmops ||
3599 vma->vm_ops == &legacy_special_mapping_vmops);
3603 * Called with mm->mmap_lock held for writing.
3604 * Insert a new vma covering the given region, with the given flags.
3605 * Its pages are supplied by the given array of struct page *.
3606 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3607 * The region past the last page supplied will always produce SIGBUS.
3608 * The array pointer and the pages it points to are assumed to stay alive
3609 * for as long as this mapping might exist.
3611 struct vm_area_struct *_install_special_mapping(
3612 struct mm_struct *mm,
3613 unsigned long addr, unsigned long len,
3614 unsigned long vm_flags, const struct vm_special_mapping *spec)
3616 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3617 &special_mapping_vmops);
3620 int install_special_mapping(struct mm_struct *mm,
3621 unsigned long addr, unsigned long len,
3622 unsigned long vm_flags, struct page **pages)
3624 struct vm_area_struct *vma = __install_special_mapping(
3625 mm, addr, len, vm_flags, (void *)pages,
3626 &legacy_special_mapping_vmops);
3628 return PTR_ERR_OR_ZERO(vma);
3631 static DEFINE_MUTEX(mm_all_locks_mutex);
3633 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3635 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3637 * The LSB of head.next can't change from under us
3638 * because we hold the mm_all_locks_mutex.
3640 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3642 * We can safely modify head.next after taking the
3643 * anon_vma->root->rwsem. If some other vma in this mm shares
3644 * the same anon_vma we won't take it again.
3646 * No need of atomic instructions here, head.next
3647 * can't change from under us thanks to the
3648 * anon_vma->root->rwsem.
3650 if (__test_and_set_bit(0, (unsigned long *)
3651 &anon_vma->root->rb_root.rb_root.rb_node))
3656 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3658 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3660 * AS_MM_ALL_LOCKS can't change from under us because
3661 * we hold the mm_all_locks_mutex.
3663 * Operations on ->flags have to be atomic because
3664 * even if AS_MM_ALL_LOCKS is stable thanks to the
3665 * mm_all_locks_mutex, there may be other cpus
3666 * changing other bitflags in parallel to us.
3668 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3670 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3675 * This operation locks against the VM for all pte/vma/mm related
3676 * operations that could ever happen on a certain mm. This includes
3677 * vmtruncate, try_to_unmap, and all page faults.
3679 * The caller must take the mmap_lock in write mode before calling
3680 * mm_take_all_locks(). The caller isn't allowed to release the
3681 * mmap_lock until mm_drop_all_locks() returns.
3683 * mmap_lock in write mode is required in order to block all operations
3684 * that could modify pagetables and free pages without need of
3685 * altering the vma layout. It's also needed in write mode to avoid new
3686 * anon_vmas to be associated with existing vmas.
3688 * A single task can't take more than one mm_take_all_locks() in a row
3689 * or it would deadlock.
3691 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3692 * mapping->flags avoid to take the same lock twice, if more than one
3693 * vma in this mm is backed by the same anon_vma or address_space.
3695 * We take locks in following order, accordingly to comment at beginning
3697 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3699 * - all vmas marked locked
3700 * - all i_mmap_rwsem locks;
3701 * - all anon_vma->rwseml
3703 * We can take all locks within these types randomly because the VM code
3704 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3705 * mm_all_locks_mutex.
3707 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3708 * that may have to take thousand of locks.
3710 * mm_take_all_locks() can fail if it's interrupted by signals.
3712 int mm_take_all_locks(struct mm_struct *mm)
3714 struct vm_area_struct *vma;
3715 struct anon_vma_chain *avc;
3716 MA_STATE(mas, &mm->mm_mt, 0, 0);
3718 mmap_assert_write_locked(mm);
3720 mutex_lock(&mm_all_locks_mutex);
3723 * vma_start_write() does not have a complement in mm_drop_all_locks()
3724 * because vma_start_write() is always asymmetrical; it marks a VMA as
3725 * being written to until mmap_write_unlock() or mmap_write_downgrade()
3728 mas_for_each(&mas, vma, ULONG_MAX) {
3729 if (signal_pending(current))
3731 vma_start_write(vma);
3735 mas_for_each(&mas, vma, ULONG_MAX) {
3736 if (signal_pending(current))
3738 if (vma->vm_file && vma->vm_file->f_mapping &&
3739 is_vm_hugetlb_page(vma))
3740 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3744 mas_for_each(&mas, vma, ULONG_MAX) {
3745 if (signal_pending(current))
3747 if (vma->vm_file && vma->vm_file->f_mapping &&
3748 !is_vm_hugetlb_page(vma))
3749 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3753 mas_for_each(&mas, vma, ULONG_MAX) {
3754 if (signal_pending(current))
3757 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3758 vm_lock_anon_vma(mm, avc->anon_vma);
3764 mm_drop_all_locks(mm);
3768 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3770 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3772 * The LSB of head.next can't change to 0 from under
3773 * us because we hold the mm_all_locks_mutex.
3775 * We must however clear the bitflag before unlocking
3776 * the vma so the users using the anon_vma->rb_root will
3777 * never see our bitflag.
3779 * No need of atomic instructions here, head.next
3780 * can't change from under us until we release the
3781 * anon_vma->root->rwsem.
3783 if (!__test_and_clear_bit(0, (unsigned long *)
3784 &anon_vma->root->rb_root.rb_root.rb_node))
3786 anon_vma_unlock_write(anon_vma);
3790 static void vm_unlock_mapping(struct address_space *mapping)
3792 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3794 * AS_MM_ALL_LOCKS can't change to 0 from under us
3795 * because we hold the mm_all_locks_mutex.
3797 i_mmap_unlock_write(mapping);
3798 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3805 * The mmap_lock cannot be released by the caller until
3806 * mm_drop_all_locks() returns.
3808 void mm_drop_all_locks(struct mm_struct *mm)
3810 struct vm_area_struct *vma;
3811 struct anon_vma_chain *avc;
3812 MA_STATE(mas, &mm->mm_mt, 0, 0);
3814 mmap_assert_write_locked(mm);
3815 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3817 mas_for_each(&mas, vma, ULONG_MAX) {
3819 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3820 vm_unlock_anon_vma(avc->anon_vma);
3821 if (vma->vm_file && vma->vm_file->f_mapping)
3822 vm_unlock_mapping(vma->vm_file->f_mapping);
3825 mutex_unlock(&mm_all_locks_mutex);
3829 * initialise the percpu counter for VM
3831 void __init mmap_init(void)
3835 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3840 * Initialise sysctl_user_reserve_kbytes.
3842 * This is intended to prevent a user from starting a single memory hogging
3843 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3846 * The default value is min(3% of free memory, 128MB)
3847 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3849 static int init_user_reserve(void)
3851 unsigned long free_kbytes;
3853 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3855 sysctl_user_reserve_kbytes = min(free_kbytes / 32, SZ_128K);
3858 subsys_initcall(init_user_reserve);
3861 * Initialise sysctl_admin_reserve_kbytes.
3863 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3864 * to log in and kill a memory hogging process.
3866 * Systems with more than 256MB will reserve 8MB, enough to recover
3867 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3868 * only reserve 3% of free pages by default.
3870 static int init_admin_reserve(void)
3872 unsigned long free_kbytes;
3874 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3876 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, SZ_8K);
3879 subsys_initcall(init_admin_reserve);
3882 * Reinititalise user and admin reserves if memory is added or removed.
3884 * The default user reserve max is 128MB, and the default max for the
3885 * admin reserve is 8MB. These are usually, but not always, enough to
3886 * enable recovery from a memory hogging process using login/sshd, a shell,
3887 * and tools like top. It may make sense to increase or even disable the
3888 * reserve depending on the existence of swap or variations in the recovery
3889 * tools. So, the admin may have changed them.
3891 * If memory is added and the reserves have been eliminated or increased above
3892 * the default max, then we'll trust the admin.
3894 * If memory is removed and there isn't enough free memory, then we
3895 * need to reset the reserves.
3897 * Otherwise keep the reserve set by the admin.
3899 static int reserve_mem_notifier(struct notifier_block *nb,
3900 unsigned long action, void *data)
3902 unsigned long tmp, free_kbytes;
3906 /* Default max is 128MB. Leave alone if modified by operator. */
3907 tmp = sysctl_user_reserve_kbytes;
3908 if (tmp > 0 && tmp < SZ_128K)
3909 init_user_reserve();
3911 /* Default max is 8MB. Leave alone if modified by operator. */
3912 tmp = sysctl_admin_reserve_kbytes;
3913 if (tmp > 0 && tmp < SZ_8K)
3914 init_admin_reserve();
3918 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3920 if (sysctl_user_reserve_kbytes > free_kbytes) {
3921 init_user_reserve();
3922 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3923 sysctl_user_reserve_kbytes);
3926 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3927 init_admin_reserve();
3928 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3929 sysctl_admin_reserve_kbytes);
3938 static int __meminit init_reserve_notifier(void)
3940 if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI))
3941 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3945 subsys_initcall(init_reserve_notifier);