1 // SPDX-License-Identifier: GPL-2.0-only
7 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12 #include <linux/kernel.h>
13 #include <linux/slab.h>
14 #include <linux/backing-dev.h>
16 #include <linux/mm_inline.h>
17 #include <linux/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 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
68 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
70 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
71 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
72 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
73 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
76 static bool ignore_rlimit_data;
77 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
79 static void unmap_region(struct mm_struct *mm, 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 file *file, struct address_space *mapping)
110 if (vma->vm_flags & VM_SHARED)
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, file, 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->vm_flags & VM_SHARED)
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 int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
397 VMA_ITERATOR(vmi, mm, 0);
398 struct address_space *mapping = NULL;
400 vma_iter_config(&vmi, vma->vm_start, vma->vm_end);
401 if (vma_iter_prealloc(&vmi, vma))
404 vma_start_write(vma);
406 vma_iter_store(&vmi, vma);
409 mapping = vma->vm_file->f_mapping;
410 i_mmap_lock_write(mapping);
411 __vma_link_file(vma, mapping);
412 i_mmap_unlock_write(mapping);
421 * init_multi_vma_prep() - Initializer for struct vma_prepare
422 * @vp: The vma_prepare struct
423 * @vma: The vma that will be altered once locked
424 * @next: The next vma if it is to be adjusted
425 * @remove: The first vma to be removed
426 * @remove2: The second vma to be removed
428 static inline void init_multi_vma_prep(struct vma_prepare *vp,
429 struct vm_area_struct *vma, struct vm_area_struct *next,
430 struct vm_area_struct *remove, struct vm_area_struct *remove2)
432 memset(vp, 0, sizeof(struct vma_prepare));
434 vp->anon_vma = vma->anon_vma;
436 vp->remove2 = remove2;
438 if (!vp->anon_vma && next)
439 vp->anon_vma = next->anon_vma;
441 vp->file = vma->vm_file;
443 vp->mapping = vma->vm_file->f_mapping;
448 * init_vma_prep() - Initializer wrapper for vma_prepare struct
449 * @vp: The vma_prepare struct
450 * @vma: The vma that will be altered once locked
452 static inline void init_vma_prep(struct vma_prepare *vp,
453 struct vm_area_struct *vma)
455 init_multi_vma_prep(vp, vma, NULL, NULL, NULL);
460 * vma_prepare() - Helper function for handling locking VMAs prior to altering
461 * @vp: The initialized vma_prepare struct
463 static inline void vma_prepare(struct vma_prepare *vp)
466 uprobe_munmap(vp->vma, vp->vma->vm_start, vp->vma->vm_end);
469 uprobe_munmap(vp->adj_next, vp->adj_next->vm_start,
470 vp->adj_next->vm_end);
472 i_mmap_lock_write(vp->mapping);
473 if (vp->insert && vp->insert->vm_file) {
475 * Put into interval tree now, so instantiated pages
476 * are visible to arm/parisc __flush_dcache_page
477 * throughout; but we cannot insert into address
478 * space until vma start or end is updated.
480 __vma_link_file(vp->insert,
481 vp->insert->vm_file->f_mapping);
486 anon_vma_lock_write(vp->anon_vma);
487 anon_vma_interval_tree_pre_update_vma(vp->vma);
489 anon_vma_interval_tree_pre_update_vma(vp->adj_next);
493 flush_dcache_mmap_lock(vp->mapping);
494 vma_interval_tree_remove(vp->vma, &vp->mapping->i_mmap);
496 vma_interval_tree_remove(vp->adj_next,
497 &vp->mapping->i_mmap);
503 * vma_complete- Helper function for handling the unlocking after altering VMAs,
504 * or for inserting a VMA.
506 * @vp: The vma_prepare struct
507 * @vmi: The vma iterator
510 static inline void vma_complete(struct vma_prepare *vp,
511 struct vma_iterator *vmi, struct mm_struct *mm)
515 vma_interval_tree_insert(vp->adj_next,
516 &vp->mapping->i_mmap);
517 vma_interval_tree_insert(vp->vma, &vp->mapping->i_mmap);
518 flush_dcache_mmap_unlock(vp->mapping);
521 if (vp->remove && vp->file) {
522 __remove_shared_vm_struct(vp->remove, vp->file, vp->mapping);
524 __remove_shared_vm_struct(vp->remove2, vp->file,
526 } else if (vp->insert) {
528 * split_vma has split insert from vma, and needs
529 * us to insert it before dropping the locks
530 * (it may either follow vma or precede it).
532 vma_iter_store(vmi, vp->insert);
537 anon_vma_interval_tree_post_update_vma(vp->vma);
539 anon_vma_interval_tree_post_update_vma(vp->adj_next);
540 anon_vma_unlock_write(vp->anon_vma);
544 i_mmap_unlock_write(vp->mapping);
545 uprobe_mmap(vp->vma);
548 uprobe_mmap(vp->adj_next);
553 vma_mark_detached(vp->remove, true);
555 uprobe_munmap(vp->remove, vp->remove->vm_start,
559 if (vp->remove->anon_vma)
560 anon_vma_merge(vp->vma, vp->remove);
562 mpol_put(vma_policy(vp->remove));
564 WARN_ON_ONCE(vp->vma->vm_end < vp->remove->vm_end);
565 vm_area_free(vp->remove);
568 * In mprotect's case 6 (see comments on vma_merge),
569 * we are removing both mid and next vmas
572 vp->remove = vp->remove2;
577 if (vp->insert && vp->file)
578 uprobe_mmap(vp->insert);
583 * dup_anon_vma() - Helper function to duplicate anon_vma
584 * @dst: The destination VMA
585 * @src: The source VMA
586 * @dup: Pointer to the destination VMA when successful.
588 * Returns: 0 on success.
590 static inline int dup_anon_vma(struct vm_area_struct *dst,
591 struct vm_area_struct *src, struct vm_area_struct **dup)
594 * Easily overlooked: when mprotect shifts the boundary, make sure the
595 * expanding vma has anon_vma set if the shrinking vma had, to cover any
596 * anon pages imported.
598 if (src->anon_vma && !dst->anon_vma) {
601 vma_assert_write_locked(dst);
602 dst->anon_vma = src->anon_vma;
603 ret = anon_vma_clone(dst, src);
614 * vma_expand - Expand an existing VMA
616 * @vmi: The vma iterator
617 * @vma: The vma to expand
618 * @start: The start of the vma
619 * @end: The exclusive end of the vma
620 * @pgoff: The page offset of vma
621 * @next: The current of next vma.
623 * Expand @vma to @start and @end. Can expand off the start and end. Will
624 * expand over @next if it's different from @vma and @end == @next->vm_end.
625 * Checking if the @vma can expand and merge with @next needs to be handled by
628 * Returns: 0 on success
630 int vma_expand(struct vma_iterator *vmi, struct vm_area_struct *vma,
631 unsigned long start, unsigned long end, pgoff_t pgoff,
632 struct vm_area_struct *next)
634 struct vm_area_struct *anon_dup = NULL;
635 bool remove_next = false;
636 struct vma_prepare vp;
638 vma_start_write(vma);
639 if (next && (vma != next) && (end == next->vm_end)) {
643 vma_start_write(next);
644 ret = dup_anon_vma(vma, next, &anon_dup);
649 init_multi_vma_prep(&vp, vma, NULL, remove_next ? next : NULL, NULL);
650 /* Not merging but overwriting any part of next is not handled. */
651 VM_WARN_ON(next && !vp.remove &&
652 next != vma && end > next->vm_start);
653 /* Only handles expanding */
654 VM_WARN_ON(vma->vm_start < start || vma->vm_end > end);
656 /* Note: vma iterator must be pointing to 'start' */
657 vma_iter_config(vmi, start, end);
658 if (vma_iter_prealloc(vmi, vma))
662 vma_adjust_trans_huge(vma, start, end, 0);
663 vma->vm_start = start;
665 vma->vm_pgoff = pgoff;
666 vma_iter_store(vmi, vma);
668 vma_complete(&vp, vmi, vma->vm_mm);
673 unlink_anon_vmas(anon_dup);
678 * vma_shrink() - Reduce an existing VMAs memory area
679 * @vmi: The vma iterator
680 * @vma: The VMA to modify
681 * @start: The new start
684 * Returns: 0 on success, -ENOMEM otherwise
686 int vma_shrink(struct vma_iterator *vmi, struct vm_area_struct *vma,
687 unsigned long start, unsigned long end, pgoff_t pgoff)
689 struct vma_prepare vp;
691 WARN_ON((vma->vm_start != start) && (vma->vm_end != end));
693 if (vma->vm_start < start)
694 vma_iter_config(vmi, vma->vm_start, start);
696 vma_iter_config(vmi, end, vma->vm_end);
698 if (vma_iter_prealloc(vmi, NULL))
701 vma_start_write(vma);
703 init_vma_prep(&vp, vma);
705 vma_adjust_trans_huge(vma, start, end, 0);
708 vma->vm_start = start;
710 vma->vm_pgoff = pgoff;
711 vma_complete(&vp, vmi, vma->vm_mm);
716 * If the vma has a ->close operation then the driver probably needs to release
717 * per-vma resources, so we don't attempt to merge those if the caller indicates
718 * the current vma may be removed as part of the merge.
720 static inline bool is_mergeable_vma(struct vm_area_struct *vma,
721 struct file *file, unsigned long vm_flags,
722 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
723 struct anon_vma_name *anon_name, bool may_remove_vma)
726 * VM_SOFTDIRTY should not prevent from VMA merging, if we
727 * match the flags but dirty bit -- the caller should mark
728 * merged VMA as dirty. If dirty bit won't be excluded from
729 * comparison, we increase pressure on the memory system forcing
730 * the kernel to generate new VMAs when old one could be
733 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
735 if (vma->vm_file != file)
737 if (may_remove_vma && vma->vm_ops && vma->vm_ops->close)
739 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
741 if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
746 static inline bool is_mergeable_anon_vma(struct anon_vma *anon_vma1,
747 struct anon_vma *anon_vma2, struct vm_area_struct *vma)
750 * The list_is_singular() test is to avoid merging VMA cloned from
751 * parents. This can improve scalability caused by anon_vma lock.
753 if ((!anon_vma1 || !anon_vma2) && (!vma ||
754 list_is_singular(&vma->anon_vma_chain)))
756 return anon_vma1 == anon_vma2;
760 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
761 * in front of (at a lower virtual address and file offset than) the vma.
763 * We cannot merge two vmas if they have differently assigned (non-NULL)
764 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
766 * We don't check here for the merged mmap wrapping around the end of pagecache
767 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
768 * wrap, nor mmaps which cover the final page at index -1UL.
770 * We assume the vma may be removed as part of the merge.
773 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
774 struct anon_vma *anon_vma, struct file *file,
775 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
776 struct anon_vma_name *anon_name)
778 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, true) &&
779 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
780 if (vma->vm_pgoff == vm_pgoff)
787 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
788 * beyond (at a higher virtual address and file offset than) the vma.
790 * We cannot merge two vmas if they have differently assigned (non-NULL)
791 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
793 * We assume that vma is not removed as part of the merge.
796 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
797 struct anon_vma *anon_vma, struct file *file,
798 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
799 struct anon_vma_name *anon_name)
801 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, false) &&
802 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
804 vm_pglen = vma_pages(vma);
805 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
812 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
813 * figure out whether that can be merged with its predecessor or its
814 * successor. Or both (it neatly fills a hole).
816 * In most cases - when called for mmap, brk or mremap - [addr,end) is
817 * certain not to be mapped by the time vma_merge is called; but when
818 * called for mprotect, it is certain to be already mapped (either at
819 * an offset within prev, or at the start of next), and the flags of
820 * this area are about to be changed to vm_flags - and the no-change
821 * case has already been eliminated.
823 * The following mprotect cases have to be considered, where **** is
824 * the area passed down from mprotect_fixup, never extending beyond one
825 * vma, PPPP is the previous vma, CCCC is a concurrent vma that starts
826 * at the same address as **** and is of the same or larger span, and
827 * NNNN the next vma after ****:
830 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPCCCCCC
831 * cannot merge might become might become
832 * PPNNNNNNNNNN PPPPPPPPPPCC
833 * mmap, brk or case 4 below case 5 below
836 * PPPP NNNN PPPPCCCCNNNN
837 * might become might become
838 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
839 * PPPPPPPPNNNN 2 or PPPPPPPPNNNN 7 or
840 * PPPPNNNNNNNN 3 PPPPNNNNNNNN 8
842 * It is important for case 8 that the vma CCCC overlapping the
843 * region **** is never going to extended over NNNN. Instead NNNN must
844 * be extended in region **** and CCCC must be removed. This way in
845 * all cases where vma_merge succeeds, the moment vma_merge drops the
846 * rmap_locks, the properties of the merged vma will be already
847 * correct for the whole merged range. Some of those properties like
848 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
849 * be correct for the whole merged range immediately after the
850 * rmap_locks are released. Otherwise if NNNN would be removed and
851 * CCCC would be extended over the NNNN range, remove_migration_ptes
852 * or other rmap walkers (if working on addresses beyond the "end"
853 * parameter) may establish ptes with the wrong permissions of CCCC
854 * instead of the right permissions of NNNN.
857 * PPPP is represented by *prev
858 * CCCC is represented by *curr or not represented at all (NULL)
859 * NNNN is represented by *next or not represented at all (NULL)
860 * **** is not represented - it will be merged and the vma containing the
861 * area is returned, or the function will return NULL
863 struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm,
864 struct vm_area_struct *prev, unsigned long addr,
865 unsigned long end, unsigned long vm_flags,
866 struct anon_vma *anon_vma, struct file *file,
867 pgoff_t pgoff, struct mempolicy *policy,
868 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
869 struct anon_vma_name *anon_name)
871 struct vm_area_struct *curr, *next, *res;
872 struct vm_area_struct *vma, *adjust, *remove, *remove2;
873 struct vm_area_struct *anon_dup = NULL;
874 struct vma_prepare vp;
877 bool merge_prev = false;
878 bool merge_next = false;
879 bool vma_expanded = false;
880 unsigned long vma_start = addr;
881 unsigned long vma_end = end;
882 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
886 * We later require that vma->vm_flags == vm_flags,
887 * so this tests vma->vm_flags & VM_SPECIAL, too.
889 if (vm_flags & VM_SPECIAL)
892 /* Does the input range span an existing VMA? (cases 5 - 8) */
893 curr = find_vma_intersection(mm, prev ? prev->vm_end : 0, end);
895 if (!curr || /* cases 1 - 4 */
896 end == curr->vm_end) /* cases 6 - 8, adjacent VMA */
897 next = vma_lookup(mm, end);
899 next = NULL; /* case 5 */
902 vma_start = prev->vm_start;
903 vma_pgoff = prev->vm_pgoff;
905 /* Can we merge the predecessor? */
906 if (addr == prev->vm_end && mpol_equal(vma_policy(prev), policy)
907 && can_vma_merge_after(prev, vm_flags, anon_vma, file,
908 pgoff, vm_userfaultfd_ctx, anon_name)) {
914 /* Can we merge the successor? */
915 if (next && mpol_equal(policy, vma_policy(next)) &&
916 can_vma_merge_before(next, vm_flags, anon_vma, file, pgoff+pglen,
917 vm_userfaultfd_ctx, anon_name)) {
921 /* Verify some invariant that must be enforced by the caller. */
922 VM_WARN_ON(prev && addr <= prev->vm_start);
923 VM_WARN_ON(curr && (addr != curr->vm_start || end > curr->vm_end));
924 VM_WARN_ON(addr >= end);
926 if (!merge_prev && !merge_next)
927 return NULL; /* Not mergeable. */
930 vma_start_write(prev);
933 remove = remove2 = adjust = NULL;
935 /* Can we merge both the predecessor and the successor? */
936 if (merge_prev && merge_next &&
937 is_mergeable_anon_vma(prev->anon_vma, next->anon_vma, NULL)) {
938 vma_start_write(next);
939 remove = next; /* case 1 */
940 vma_end = next->vm_end;
941 err = dup_anon_vma(prev, next, &anon_dup);
942 if (curr) { /* case 6 */
943 vma_start_write(curr);
947 * Note that the dup_anon_vma below cannot overwrite err
948 * since the first caller would do nothing unless next
952 err = dup_anon_vma(prev, curr, &anon_dup);
954 } else if (merge_prev) { /* case 2 */
956 vma_start_write(curr);
957 err = dup_anon_vma(prev, curr, &anon_dup);
958 if (end == curr->vm_end) { /* case 7 */
960 } else { /* case 5 */
962 adj_start = (end - curr->vm_start);
965 } else { /* merge_next */
966 vma_start_write(next);
968 if (prev && addr < prev->vm_end) { /* case 4 */
969 vma_start_write(prev);
972 adj_start = -(prev->vm_end - addr);
973 err = dup_anon_vma(next, prev, &anon_dup);
976 * Note that cases 3 and 8 are the ONLY ones where prev
977 * is permitted to be (but is not necessarily) NULL.
979 vma = next; /* case 3 */
981 vma_end = next->vm_end;
982 vma_pgoff = next->vm_pgoff - pglen;
983 if (curr) { /* case 8 */
984 vma_pgoff = curr->vm_pgoff;
985 vma_start_write(curr);
987 err = dup_anon_vma(next, curr, &anon_dup);
992 /* Error in anon_vma clone. */
996 if (vma_start < vma->vm_start || vma_end > vma->vm_end)
1000 vma_iter_config(vmi, vma_start, vma_end);
1002 vma_iter_config(vmi, adjust->vm_start + adj_start,
1006 if (vma_iter_prealloc(vmi, vma))
1009 init_multi_vma_prep(&vp, vma, adjust, remove, remove2);
1010 VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma &&
1011 vp.anon_vma != adjust->anon_vma);
1014 vma_adjust_trans_huge(vma, vma_start, vma_end, adj_start);
1016 vma->vm_start = vma_start;
1017 vma->vm_end = vma_end;
1018 vma->vm_pgoff = vma_pgoff;
1021 vma_iter_store(vmi, vma);
1024 adjust->vm_start += adj_start;
1025 adjust->vm_pgoff += adj_start >> PAGE_SHIFT;
1026 if (adj_start < 0) {
1027 WARN_ON(vma_expanded);
1028 vma_iter_store(vmi, next);
1032 vma_complete(&vp, vmi, mm);
1033 khugepaged_enter_vma(res, vm_flags);
1038 unlink_anon_vmas(anon_dup);
1041 vma_iter_set(vmi, addr);
1047 * Rough compatibility check to quickly see if it's even worth looking
1048 * at sharing an anon_vma.
1050 * They need to have the same vm_file, and the flags can only differ
1051 * in things that mprotect may change.
1053 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1054 * we can merge the two vma's. For example, we refuse to merge a vma if
1055 * there is a vm_ops->close() function, because that indicates that the
1056 * driver is doing some kind of reference counting. But that doesn't
1057 * really matter for the anon_vma sharing case.
1059 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1061 return a->vm_end == b->vm_start &&
1062 mpol_equal(vma_policy(a), vma_policy(b)) &&
1063 a->vm_file == b->vm_file &&
1064 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1065 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1069 * Do some basic sanity checking to see if we can re-use the anon_vma
1070 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1071 * the same as 'old', the other will be the new one that is trying
1072 * to share the anon_vma.
1074 * NOTE! This runs with mmap_lock held for reading, so it is possible that
1075 * the anon_vma of 'old' is concurrently in the process of being set up
1076 * by another page fault trying to merge _that_. But that's ok: if it
1077 * is being set up, that automatically means that it will be a singleton
1078 * acceptable for merging, so we can do all of this optimistically. But
1079 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1081 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1082 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1083 * is to return an anon_vma that is "complex" due to having gone through
1086 * We also make sure that the two vma's are compatible (adjacent,
1087 * and with the same memory policies). That's all stable, even with just
1088 * a read lock on the mmap_lock.
1090 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1092 if (anon_vma_compatible(a, b)) {
1093 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1095 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1102 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1103 * neighbouring vmas for a suitable anon_vma, before it goes off
1104 * to allocate a new anon_vma. It checks because a repetitive
1105 * sequence of mprotects and faults may otherwise lead to distinct
1106 * anon_vmas being allocated, preventing vma merge in subsequent
1109 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1111 MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end);
1112 struct anon_vma *anon_vma = NULL;
1113 struct vm_area_struct *prev, *next;
1115 /* Try next first. */
1116 next = mas_walk(&mas);
1118 anon_vma = reusable_anon_vma(next, vma, next);
1123 prev = mas_prev(&mas, 0);
1124 VM_BUG_ON_VMA(prev != vma, vma);
1125 prev = mas_prev(&mas, 0);
1126 /* Try prev next. */
1128 anon_vma = reusable_anon_vma(prev, prev, vma);
1131 * We might reach here with anon_vma == NULL if we can't find
1132 * any reusable anon_vma.
1133 * There's no absolute need to look only at touching neighbours:
1134 * we could search further afield for "compatible" anon_vmas.
1135 * But it would probably just be a waste of time searching,
1136 * or lead to too many vmas hanging off the same anon_vma.
1137 * We're trying to allow mprotect remerging later on,
1138 * not trying to minimize memory used for anon_vmas.
1144 * If a hint addr is less than mmap_min_addr change hint to be as
1145 * low as possible but still greater than mmap_min_addr
1147 static inline unsigned long round_hint_to_min(unsigned long hint)
1150 if (((void *)hint != NULL) &&
1151 (hint < mmap_min_addr))
1152 return PAGE_ALIGN(mmap_min_addr);
1156 bool mlock_future_ok(struct mm_struct *mm, unsigned long flags,
1157 unsigned long bytes)
1159 unsigned long locked_pages, limit_pages;
1161 if (!(flags & VM_LOCKED) || capable(CAP_IPC_LOCK))
1164 locked_pages = bytes >> PAGE_SHIFT;
1165 locked_pages += mm->locked_vm;
1167 limit_pages = rlimit(RLIMIT_MEMLOCK);
1168 limit_pages >>= PAGE_SHIFT;
1170 return locked_pages <= limit_pages;
1173 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1175 if (S_ISREG(inode->i_mode))
1176 return MAX_LFS_FILESIZE;
1178 if (S_ISBLK(inode->i_mode))
1179 return MAX_LFS_FILESIZE;
1181 if (S_ISSOCK(inode->i_mode))
1182 return MAX_LFS_FILESIZE;
1184 /* Special "we do even unsigned file positions" case */
1185 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1188 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1192 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1193 unsigned long pgoff, unsigned long len)
1195 u64 maxsize = file_mmap_size_max(file, inode);
1197 if (maxsize && len > maxsize)
1200 if (pgoff > maxsize >> PAGE_SHIFT)
1206 * The caller must write-lock current->mm->mmap_lock.
1208 unsigned long do_mmap(struct file *file, unsigned long addr,
1209 unsigned long len, unsigned long prot,
1210 unsigned long flags, vm_flags_t vm_flags,
1211 unsigned long pgoff, unsigned long *populate,
1212 struct list_head *uf)
1214 struct mm_struct *mm = current->mm;
1223 * Does the application expect PROT_READ to imply PROT_EXEC?
1225 * (the exception is when the underlying filesystem is noexec
1226 * mounted, in which case we dont add PROT_EXEC.)
1228 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1229 if (!(file && path_noexec(&file->f_path)))
1232 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1233 if (flags & MAP_FIXED_NOREPLACE)
1236 if (!(flags & MAP_FIXED))
1237 addr = round_hint_to_min(addr);
1239 /* Careful about overflows.. */
1240 len = PAGE_ALIGN(len);
1244 /* offset overflow? */
1245 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1248 /* Too many mappings? */
1249 if (mm->map_count > sysctl_max_map_count)
1252 /* Obtain the address to map to. we verify (or select) it and ensure
1253 * that it represents a valid section of the address space.
1255 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1256 if (IS_ERR_VALUE(addr))
1259 if (flags & MAP_FIXED_NOREPLACE) {
1260 if (find_vma_intersection(mm, addr, addr + len))
1264 if (prot == PROT_EXEC) {
1265 pkey = execute_only_pkey(mm);
1270 /* Do simple checking here so the lower-level routines won't have
1271 * to. we assume access permissions have been handled by the open
1272 * of the memory object, so we don't do any here.
1274 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1275 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1277 if (flags & MAP_LOCKED)
1278 if (!can_do_mlock())
1281 if (!mlock_future_ok(mm, vm_flags, len))
1285 struct inode *inode = file_inode(file);
1286 unsigned long flags_mask;
1288 if (!file_mmap_ok(file, inode, pgoff, len))
1291 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1293 switch (flags & MAP_TYPE) {
1296 * Force use of MAP_SHARED_VALIDATE with non-legacy
1297 * flags. E.g. MAP_SYNC is dangerous to use with
1298 * MAP_SHARED as you don't know which consistency model
1299 * you will get. We silently ignore unsupported flags
1300 * with MAP_SHARED to preserve backward compatibility.
1302 flags &= LEGACY_MAP_MASK;
1304 case MAP_SHARED_VALIDATE:
1305 if (flags & ~flags_mask)
1307 if (prot & PROT_WRITE) {
1308 if (!(file->f_mode & FMODE_WRITE))
1310 if (IS_SWAPFILE(file->f_mapping->host))
1315 * Make sure we don't allow writing to an append-only
1318 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1321 vm_flags |= VM_SHARED | VM_MAYSHARE;
1322 if (!(file->f_mode & FMODE_WRITE))
1323 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1326 if (!(file->f_mode & FMODE_READ))
1328 if (path_noexec(&file->f_path)) {
1329 if (vm_flags & VM_EXEC)
1331 vm_flags &= ~VM_MAYEXEC;
1334 if (!file->f_op->mmap)
1336 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1344 switch (flags & MAP_TYPE) {
1346 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1352 vm_flags |= VM_SHARED | VM_MAYSHARE;
1356 * Set pgoff according to addr for anon_vma.
1358 pgoff = addr >> PAGE_SHIFT;
1366 * Set 'VM_NORESERVE' if we should not account for the
1367 * memory use of this mapping.
1369 if (flags & MAP_NORESERVE) {
1370 /* We honor MAP_NORESERVE if allowed to overcommit */
1371 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1372 vm_flags |= VM_NORESERVE;
1374 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1375 if (file && is_file_hugepages(file))
1376 vm_flags |= VM_NORESERVE;
1379 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1380 if (!IS_ERR_VALUE(addr) &&
1381 ((vm_flags & VM_LOCKED) ||
1382 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1387 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1388 unsigned long prot, unsigned long flags,
1389 unsigned long fd, unsigned long pgoff)
1391 struct file *file = NULL;
1392 unsigned long retval;
1394 if (!(flags & MAP_ANONYMOUS)) {
1395 audit_mmap_fd(fd, flags);
1399 if (is_file_hugepages(file)) {
1400 len = ALIGN(len, huge_page_size(hstate_file(file)));
1401 } else if (unlikely(flags & MAP_HUGETLB)) {
1405 } else if (flags & MAP_HUGETLB) {
1408 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1412 len = ALIGN(len, huge_page_size(hs));
1414 * VM_NORESERVE is used because the reservations will be
1415 * taken when vm_ops->mmap() is called
1417 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1419 HUGETLB_ANONHUGE_INODE,
1420 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1422 return PTR_ERR(file);
1425 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1432 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1433 unsigned long, prot, unsigned long, flags,
1434 unsigned long, fd, unsigned long, pgoff)
1436 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1439 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1440 struct mmap_arg_struct {
1444 unsigned long flags;
1446 unsigned long offset;
1449 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1451 struct mmap_arg_struct a;
1453 if (copy_from_user(&a, arg, sizeof(a)))
1455 if (offset_in_page(a.offset))
1458 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1459 a.offset >> PAGE_SHIFT);
1461 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1463 static bool vm_ops_needs_writenotify(const struct vm_operations_struct *vm_ops)
1465 return vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite);
1468 static bool vma_is_shared_writable(struct vm_area_struct *vma)
1470 return (vma->vm_flags & (VM_WRITE | VM_SHARED)) ==
1471 (VM_WRITE | VM_SHARED);
1474 static bool vma_fs_can_writeback(struct vm_area_struct *vma)
1476 /* No managed pages to writeback. */
1477 if (vma->vm_flags & VM_PFNMAP)
1480 return vma->vm_file && vma->vm_file->f_mapping &&
1481 mapping_can_writeback(vma->vm_file->f_mapping);
1485 * Does this VMA require the underlying folios to have their dirty state
1488 bool vma_needs_dirty_tracking(struct vm_area_struct *vma)
1490 /* Only shared, writable VMAs require dirty tracking. */
1491 if (!vma_is_shared_writable(vma))
1494 /* Does the filesystem need to be notified? */
1495 if (vm_ops_needs_writenotify(vma->vm_ops))
1499 * Even if the filesystem doesn't indicate a need for writenotify, if it
1500 * can writeback, dirty tracking is still required.
1502 return vma_fs_can_writeback(vma);
1506 * Some shared mappings will want the pages marked read-only
1507 * to track write events. If so, we'll downgrade vm_page_prot
1508 * to the private version (using protection_map[] without the
1511 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1513 /* If it was private or non-writable, the write bit is already clear */
1514 if (!vma_is_shared_writable(vma))
1517 /* The backer wishes to know when pages are first written to? */
1518 if (vm_ops_needs_writenotify(vma->vm_ops))
1521 /* The open routine did something to the protections that pgprot_modify
1522 * won't preserve? */
1523 if (pgprot_val(vm_page_prot) !=
1524 pgprot_val(vm_pgprot_modify(vm_page_prot, vma->vm_flags)))
1528 * Do we need to track softdirty? hugetlb does not support softdirty
1531 if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma))
1534 /* Do we need write faults for uffd-wp tracking? */
1535 if (userfaultfd_wp(vma))
1538 /* Can the mapping track the dirty pages? */
1539 return vma_fs_can_writeback(vma);
1543 * We account for memory if it's a private writeable mapping,
1544 * not hugepages and VM_NORESERVE wasn't set.
1546 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1549 * hugetlb has its own accounting separate from the core VM
1550 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1552 if (file && is_file_hugepages(file))
1555 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1559 * unmapped_area() - Find an area between the low_limit and the high_limit with
1560 * the correct alignment and offset, all from @info. Note: current->mm is used
1563 * @info: The unmapped area information including the range [low_limit -
1564 * high_limit), the alignment offset and mask.
1566 * Return: A memory address or -ENOMEM.
1568 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1570 unsigned long length, gap;
1571 unsigned long low_limit, high_limit;
1572 struct vm_area_struct *tmp;
1574 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1576 /* Adjust search length to account for worst case alignment overhead */
1577 length = info->length + info->align_mask;
1578 if (length < info->length)
1581 low_limit = info->low_limit;
1582 if (low_limit < mmap_min_addr)
1583 low_limit = mmap_min_addr;
1584 high_limit = info->high_limit;
1586 if (mas_empty_area(&mas, low_limit, high_limit - 1, length))
1590 gap += (info->align_offset - gap) & info->align_mask;
1591 tmp = mas_next(&mas, ULONG_MAX);
1592 if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1593 if (vm_start_gap(tmp) < gap + length - 1) {
1594 low_limit = tmp->vm_end;
1599 tmp = mas_prev(&mas, 0);
1600 if (tmp && vm_end_gap(tmp) > gap) {
1601 low_limit = vm_end_gap(tmp);
1611 * unmapped_area_topdown() - Find an area between the low_limit and the
1612 * high_limit with the correct alignment and offset at the highest available
1613 * address, all from @info. Note: current->mm is used for the search.
1615 * @info: The unmapped area information including the range [low_limit -
1616 * high_limit), the alignment offset and mask.
1618 * Return: A memory address or -ENOMEM.
1620 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1622 unsigned long length, gap, gap_end;
1623 unsigned long low_limit, high_limit;
1624 struct vm_area_struct *tmp;
1626 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1627 /* Adjust search length to account for worst case alignment overhead */
1628 length = info->length + info->align_mask;
1629 if (length < info->length)
1632 low_limit = info->low_limit;
1633 if (low_limit < mmap_min_addr)
1634 low_limit = mmap_min_addr;
1635 high_limit = info->high_limit;
1637 if (mas_empty_area_rev(&mas, low_limit, high_limit - 1, length))
1640 gap = mas.last + 1 - info->length;
1641 gap -= (gap - info->align_offset) & info->align_mask;
1643 tmp = mas_next(&mas, ULONG_MAX);
1644 if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1645 if (vm_start_gap(tmp) <= gap_end) {
1646 high_limit = vm_start_gap(tmp);
1651 tmp = mas_prev(&mas, 0);
1652 if (tmp && vm_end_gap(tmp) > gap) {
1653 high_limit = tmp->vm_start;
1663 * Search for an unmapped address range.
1665 * We are looking for a range that:
1666 * - does not intersect with any VMA;
1667 * - is contained within the [low_limit, high_limit) interval;
1668 * - is at least the desired size.
1669 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
1671 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
1675 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
1676 addr = unmapped_area_topdown(info);
1678 addr = unmapped_area(info);
1680 trace_vm_unmapped_area(addr, info);
1684 /* Get an address range which is currently unmapped.
1685 * For shmat() with addr=0.
1687 * Ugly calling convention alert:
1688 * Return value with the low bits set means error value,
1690 * if (ret & ~PAGE_MASK)
1693 * This function "knows" that -ENOMEM has the bits set.
1696 generic_get_unmapped_area(struct file *filp, unsigned long addr,
1697 unsigned long len, unsigned long pgoff,
1698 unsigned long flags)
1700 struct mm_struct *mm = current->mm;
1701 struct vm_area_struct *vma, *prev;
1702 struct vm_unmapped_area_info info;
1703 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1705 if (len > mmap_end - mmap_min_addr)
1708 if (flags & MAP_FIXED)
1712 addr = PAGE_ALIGN(addr);
1713 vma = find_vma_prev(mm, addr, &prev);
1714 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1715 (!vma || addr + len <= vm_start_gap(vma)) &&
1716 (!prev || addr >= vm_end_gap(prev)))
1722 info.low_limit = mm->mmap_base;
1723 info.high_limit = mmap_end;
1724 info.align_mask = 0;
1725 info.align_offset = 0;
1726 return vm_unmapped_area(&info);
1729 #ifndef HAVE_ARCH_UNMAPPED_AREA
1731 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1732 unsigned long len, unsigned long pgoff,
1733 unsigned long flags)
1735 return generic_get_unmapped_area(filp, addr, len, pgoff, flags);
1740 * This mmap-allocator allocates new areas top-down from below the
1741 * stack's low limit (the base):
1744 generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1745 unsigned long len, unsigned long pgoff,
1746 unsigned long flags)
1748 struct vm_area_struct *vma, *prev;
1749 struct mm_struct *mm = current->mm;
1750 struct vm_unmapped_area_info info;
1751 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1753 /* requested length too big for entire address space */
1754 if (len > mmap_end - mmap_min_addr)
1757 if (flags & MAP_FIXED)
1760 /* requesting a specific address */
1762 addr = PAGE_ALIGN(addr);
1763 vma = find_vma_prev(mm, addr, &prev);
1764 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1765 (!vma || addr + len <= vm_start_gap(vma)) &&
1766 (!prev || addr >= vm_end_gap(prev)))
1770 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1772 info.low_limit = PAGE_SIZE;
1773 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
1774 info.align_mask = 0;
1775 info.align_offset = 0;
1776 addr = vm_unmapped_area(&info);
1779 * A failed mmap() very likely causes application failure,
1780 * so fall back to the bottom-up function here. This scenario
1781 * can happen with large stack limits and large mmap()
1784 if (offset_in_page(addr)) {
1785 VM_BUG_ON(addr != -ENOMEM);
1787 info.low_limit = TASK_UNMAPPED_BASE;
1788 info.high_limit = mmap_end;
1789 addr = vm_unmapped_area(&info);
1795 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1797 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1798 unsigned long len, unsigned long pgoff,
1799 unsigned long flags)
1801 return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags);
1806 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1807 unsigned long pgoff, unsigned long flags)
1809 unsigned long (*get_area)(struct file *, unsigned long,
1810 unsigned long, unsigned long, unsigned long);
1812 unsigned long error = arch_mmap_check(addr, len, flags);
1816 /* Careful about overflows.. */
1817 if (len > TASK_SIZE)
1820 get_area = current->mm->get_unmapped_area;
1822 if (file->f_op->get_unmapped_area)
1823 get_area = file->f_op->get_unmapped_area;
1824 } else if (flags & MAP_SHARED) {
1826 * mmap_region() will call shmem_zero_setup() to create a file,
1827 * so use shmem's get_unmapped_area in case it can be huge.
1828 * do_mmap() will clear pgoff, so match alignment.
1831 get_area = shmem_get_unmapped_area;
1834 addr = get_area(file, addr, len, pgoff, flags);
1835 if (IS_ERR_VALUE(addr))
1838 if (addr > TASK_SIZE - len)
1840 if (offset_in_page(addr))
1843 error = security_mmap_addr(addr);
1844 return error ? error : addr;
1847 EXPORT_SYMBOL(get_unmapped_area);
1850 * find_vma_intersection() - Look up the first VMA which intersects the interval
1851 * @mm: The process address space.
1852 * @start_addr: The inclusive start user address.
1853 * @end_addr: The exclusive end user address.
1855 * Returns: The first VMA within the provided range, %NULL otherwise. Assumes
1856 * start_addr < end_addr.
1858 struct vm_area_struct *find_vma_intersection(struct mm_struct *mm,
1859 unsigned long start_addr,
1860 unsigned long end_addr)
1862 unsigned long index = start_addr;
1864 mmap_assert_locked(mm);
1865 return mt_find(&mm->mm_mt, &index, end_addr - 1);
1867 EXPORT_SYMBOL(find_vma_intersection);
1870 * find_vma() - Find the VMA for a given address, or the next VMA.
1871 * @mm: The mm_struct to check
1872 * @addr: The address
1874 * Returns: The VMA associated with addr, or the next VMA.
1875 * May return %NULL in the case of no VMA at addr or above.
1877 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1879 unsigned long index = addr;
1881 mmap_assert_locked(mm);
1882 return mt_find(&mm->mm_mt, &index, ULONG_MAX);
1884 EXPORT_SYMBOL(find_vma);
1887 * find_vma_prev() - Find the VMA for a given address, or the next vma and
1888 * set %pprev to the previous VMA, if any.
1889 * @mm: The mm_struct to check
1890 * @addr: The address
1891 * @pprev: The pointer to set to the previous VMA
1893 * Note that RCU lock is missing here since the external mmap_lock() is used
1896 * Returns: The VMA associated with @addr, or the next vma.
1897 * May return %NULL in the case of no vma at addr or above.
1899 struct vm_area_struct *
1900 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1901 struct vm_area_struct **pprev)
1903 struct vm_area_struct *vma;
1904 MA_STATE(mas, &mm->mm_mt, addr, addr);
1906 vma = mas_walk(&mas);
1907 *pprev = mas_prev(&mas, 0);
1909 vma = mas_next(&mas, ULONG_MAX);
1914 * Verify that the stack growth is acceptable and
1915 * update accounting. This is shared with both the
1916 * grow-up and grow-down cases.
1918 static int acct_stack_growth(struct vm_area_struct *vma,
1919 unsigned long size, unsigned long grow)
1921 struct mm_struct *mm = vma->vm_mm;
1922 unsigned long new_start;
1924 /* address space limit tests */
1925 if (!may_expand_vm(mm, vma->vm_flags, grow))
1928 /* Stack limit test */
1929 if (size > rlimit(RLIMIT_STACK))
1932 /* mlock limit tests */
1933 if (!mlock_future_ok(mm, vma->vm_flags, grow << PAGE_SHIFT))
1936 /* Check to ensure the stack will not grow into a hugetlb-only region */
1937 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1939 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1943 * Overcommit.. This must be the final test, as it will
1944 * update security statistics.
1946 if (security_vm_enough_memory_mm(mm, grow))
1952 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1954 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1955 * vma is the last one with address > vma->vm_end. Have to extend vma.
1957 static int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1959 struct mm_struct *mm = vma->vm_mm;
1960 struct vm_area_struct *next;
1961 unsigned long gap_addr;
1963 MA_STATE(mas, &mm->mm_mt, vma->vm_start, address);
1965 if (!(vma->vm_flags & VM_GROWSUP))
1968 /* Guard against exceeding limits of the address space. */
1969 address &= PAGE_MASK;
1970 if (address >= (TASK_SIZE & PAGE_MASK))
1972 address += PAGE_SIZE;
1974 /* Enforce stack_guard_gap */
1975 gap_addr = address + stack_guard_gap;
1977 /* Guard against overflow */
1978 if (gap_addr < address || gap_addr > TASK_SIZE)
1979 gap_addr = TASK_SIZE;
1981 next = find_vma_intersection(mm, vma->vm_end, gap_addr);
1982 if (next && vma_is_accessible(next)) {
1983 if (!(next->vm_flags & VM_GROWSUP))
1985 /* Check that both stack segments have the same anon_vma? */
1989 mas_prev_range(&mas, address);
1991 __mas_set_range(&mas, vma->vm_start, address - 1);
1992 if (mas_preallocate(&mas, vma, GFP_KERNEL))
1995 /* We must make sure the anon_vma is allocated. */
1996 if (unlikely(anon_vma_prepare(vma))) {
2001 /* Lock the VMA before expanding to prevent concurrent page faults */
2002 vma_start_write(vma);
2004 * vma->vm_start/vm_end cannot change under us because the caller
2005 * is required to hold the mmap_lock in read mode. We need the
2006 * anon_vma lock to serialize against concurrent expand_stacks.
2008 anon_vma_lock_write(vma->anon_vma);
2010 /* Somebody else might have raced and expanded it already */
2011 if (address > vma->vm_end) {
2012 unsigned long size, grow;
2014 size = address - vma->vm_start;
2015 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2018 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2019 error = acct_stack_growth(vma, size, grow);
2022 * We only hold a shared mmap_lock lock here, so
2023 * we need to protect against concurrent vma
2024 * expansions. anon_vma_lock_write() doesn't
2025 * help here, as we don't guarantee that all
2026 * growable vmas in a mm share the same root
2027 * anon vma. So, we reuse mm->page_table_lock
2028 * to guard against concurrent vma expansions.
2030 spin_lock(&mm->page_table_lock);
2031 if (vma->vm_flags & VM_LOCKED)
2032 mm->locked_vm += grow;
2033 vm_stat_account(mm, vma->vm_flags, grow);
2034 anon_vma_interval_tree_pre_update_vma(vma);
2035 vma->vm_end = address;
2036 /* Overwrite old entry in mtree. */
2037 mas_store_prealloc(&mas, vma);
2038 anon_vma_interval_tree_post_update_vma(vma);
2039 spin_unlock(&mm->page_table_lock);
2041 perf_event_mmap(vma);
2045 anon_vma_unlock_write(vma->anon_vma);
2046 khugepaged_enter_vma(vma, vma->vm_flags);
2051 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2054 * vma is the first one with address < vma->vm_start. Have to extend vma.
2055 * mmap_lock held for writing.
2057 int expand_downwards(struct vm_area_struct *vma, unsigned long address)
2059 struct mm_struct *mm = vma->vm_mm;
2060 MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start);
2061 struct vm_area_struct *prev;
2064 if (!(vma->vm_flags & VM_GROWSDOWN))
2067 address &= PAGE_MASK;
2068 if (address < mmap_min_addr || address < FIRST_USER_ADDRESS)
2071 /* Enforce stack_guard_gap */
2072 prev = mas_prev(&mas, 0);
2073 /* Check that both stack segments have the same anon_vma? */
2075 if (!(prev->vm_flags & VM_GROWSDOWN) &&
2076 vma_is_accessible(prev) &&
2077 (address - prev->vm_end < stack_guard_gap))
2082 mas_next_range(&mas, vma->vm_start);
2084 __mas_set_range(&mas, address, vma->vm_end - 1);
2085 if (mas_preallocate(&mas, vma, GFP_KERNEL))
2088 /* We must make sure the anon_vma is allocated. */
2089 if (unlikely(anon_vma_prepare(vma))) {
2094 /* Lock the VMA before expanding to prevent concurrent page faults */
2095 vma_start_write(vma);
2097 * vma->vm_start/vm_end cannot change under us because the caller
2098 * is required to hold the mmap_lock in read mode. We need the
2099 * anon_vma lock to serialize against concurrent expand_stacks.
2101 anon_vma_lock_write(vma->anon_vma);
2103 /* Somebody else might have raced and expanded it already */
2104 if (address < vma->vm_start) {
2105 unsigned long size, grow;
2107 size = vma->vm_end - address;
2108 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2111 if (grow <= vma->vm_pgoff) {
2112 error = acct_stack_growth(vma, size, grow);
2115 * We only hold a shared mmap_lock lock here, so
2116 * we need to protect against concurrent vma
2117 * expansions. anon_vma_lock_write() doesn't
2118 * help here, as we don't guarantee that all
2119 * growable vmas in a mm share the same root
2120 * anon vma. So, we reuse mm->page_table_lock
2121 * to guard against concurrent vma expansions.
2123 spin_lock(&mm->page_table_lock);
2124 if (vma->vm_flags & VM_LOCKED)
2125 mm->locked_vm += grow;
2126 vm_stat_account(mm, vma->vm_flags, grow);
2127 anon_vma_interval_tree_pre_update_vma(vma);
2128 vma->vm_start = address;
2129 vma->vm_pgoff -= grow;
2130 /* Overwrite old entry in mtree. */
2131 mas_store_prealloc(&mas, vma);
2132 anon_vma_interval_tree_post_update_vma(vma);
2133 spin_unlock(&mm->page_table_lock);
2135 perf_event_mmap(vma);
2139 anon_vma_unlock_write(vma->anon_vma);
2140 khugepaged_enter_vma(vma, vma->vm_flags);
2146 /* enforced gap between the expanding stack and other mappings. */
2147 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2149 static int __init cmdline_parse_stack_guard_gap(char *p)
2154 val = simple_strtoul(p, &endptr, 10);
2156 stack_guard_gap = val << PAGE_SHIFT;
2160 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2162 #ifdef CONFIG_STACK_GROWSUP
2163 int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2165 return expand_upwards(vma, address);
2168 struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2170 struct vm_area_struct *vma, *prev;
2173 vma = find_vma_prev(mm, addr, &prev);
2174 if (vma && (vma->vm_start <= addr))
2178 if (expand_stack_locked(prev, addr))
2180 if (prev->vm_flags & VM_LOCKED)
2181 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2185 int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2187 return expand_downwards(vma, address);
2190 struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2192 struct vm_area_struct *vma;
2193 unsigned long start;
2196 vma = find_vma(mm, addr);
2199 if (vma->vm_start <= addr)
2201 start = vma->vm_start;
2202 if (expand_stack_locked(vma, addr))
2204 if (vma->vm_flags & VM_LOCKED)
2205 populate_vma_page_range(vma, addr, start, NULL);
2211 * IA64 has some horrid mapping rules: it can expand both up and down,
2212 * but with various special rules.
2214 * We'll get rid of this architecture eventually, so the ugliness is
2218 static inline bool vma_expand_ok(struct vm_area_struct *vma, unsigned long addr)
2220 return REGION_NUMBER(addr) == REGION_NUMBER(vma->vm_start) &&
2221 REGION_OFFSET(addr) < RGN_MAP_LIMIT;
2225 * IA64 stacks grow down, but there's a special register backing store
2226 * that can grow up. Only sequentially, though, so the new address must
2229 static inline int vma_expand_up(struct vm_area_struct *vma, unsigned long addr)
2231 if (!vma_expand_ok(vma, addr))
2233 if (vma->vm_end != (addr & PAGE_MASK))
2235 return expand_upwards(vma, addr);
2238 static inline bool vma_expand_down(struct vm_area_struct *vma, unsigned long addr)
2240 if (!vma_expand_ok(vma, addr))
2242 return expand_downwards(vma, addr);
2245 #elif defined(CONFIG_STACK_GROWSUP)
2247 #define vma_expand_up(vma,addr) expand_upwards(vma, addr)
2248 #define vma_expand_down(vma, addr) (-EFAULT)
2252 #define vma_expand_up(vma,addr) (-EFAULT)
2253 #define vma_expand_down(vma, addr) expand_downwards(vma, addr)
2258 * expand_stack(): legacy interface for page faulting. Don't use unless
2261 * This is called with the mm locked for reading, drops the lock, takes
2262 * the lock for writing, tries to look up a vma again, expands it if
2263 * necessary, and downgrades the lock to reading again.
2265 * If no vma is found or it can't be expanded, it returns NULL and has
2268 struct vm_area_struct *expand_stack(struct mm_struct *mm, unsigned long addr)
2270 struct vm_area_struct *vma, *prev;
2272 mmap_read_unlock(mm);
2273 if (mmap_write_lock_killable(mm))
2276 vma = find_vma_prev(mm, addr, &prev);
2277 if (vma && vma->vm_start <= addr)
2280 if (prev && !vma_expand_up(prev, addr)) {
2285 if (vma && !vma_expand_down(vma, addr))
2288 mmap_write_unlock(mm);
2292 mmap_write_downgrade(mm);
2297 * Ok - we have the memory areas we should free on a maple tree so release them,
2298 * and do the vma updates.
2300 * Called with the mm semaphore held.
2302 static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas)
2304 unsigned long nr_accounted = 0;
2305 struct vm_area_struct *vma;
2307 /* Update high watermark before we lower total_vm */
2308 update_hiwater_vm(mm);
2309 mas_for_each(mas, vma, ULONG_MAX) {
2310 long nrpages = vma_pages(vma);
2312 if (vma->vm_flags & VM_ACCOUNT)
2313 nr_accounted += nrpages;
2314 vm_stat_account(mm, vma->vm_flags, -nrpages);
2315 remove_vma(vma, false);
2317 vm_unacct_memory(nr_accounted);
2321 * Get rid of page table information in the indicated region.
2323 * Called with the mm semaphore held.
2325 static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
2326 struct vm_area_struct *vma, struct vm_area_struct *prev,
2327 struct vm_area_struct *next, unsigned long start,
2328 unsigned long end, unsigned long tree_end, bool mm_wr_locked)
2330 struct mmu_gather tlb;
2331 unsigned long mt_start = mas->index;
2334 tlb_gather_mmu(&tlb, mm);
2335 update_hiwater_rss(mm);
2336 unmap_vmas(&tlb, mas, vma, start, end, tree_end, mm_wr_locked);
2337 mas_set(mas, mt_start);
2338 free_pgtables(&tlb, mas, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2339 next ? next->vm_start : USER_PGTABLES_CEILING,
2341 tlb_finish_mmu(&tlb);
2345 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2346 * has already been checked or doesn't make sense to fail.
2347 * VMA Iterator will point to the end VMA.
2349 static int __split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2350 unsigned long addr, int new_below)
2352 struct vma_prepare vp;
2353 struct vm_area_struct *new;
2356 WARN_ON(vma->vm_start >= addr);
2357 WARN_ON(vma->vm_end <= addr);
2359 if (vma->vm_ops && vma->vm_ops->may_split) {
2360 err = vma->vm_ops->may_split(vma, addr);
2365 new = vm_area_dup(vma);
2372 new->vm_start = addr;
2373 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2377 vma_iter_config(vmi, new->vm_start, new->vm_end);
2378 if (vma_iter_prealloc(vmi, new))
2381 err = vma_dup_policy(vma, new);
2385 err = anon_vma_clone(new, vma);
2390 get_file(new->vm_file);
2392 if (new->vm_ops && new->vm_ops->open)
2393 new->vm_ops->open(new);
2395 vma_start_write(vma);
2396 vma_start_write(new);
2398 init_vma_prep(&vp, vma);
2401 vma_adjust_trans_huge(vma, vma->vm_start, addr, 0);
2404 vma->vm_start = addr;
2405 vma->vm_pgoff += (addr - new->vm_start) >> PAGE_SHIFT;
2410 /* vma_complete stores the new vma */
2411 vma_complete(&vp, vmi, vma->vm_mm);
2419 mpol_put(vma_policy(new));
2428 * Split a vma into two pieces at address 'addr', a new vma is allocated
2429 * either for the first part or the tail.
2431 static int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2432 unsigned long addr, int new_below)
2434 if (vma->vm_mm->map_count >= sysctl_max_map_count)
2437 return __split_vma(vmi, vma, addr, new_below);
2441 * We are about to modify one or multiple of a VMA's flags, policy, userfaultfd
2442 * context and anonymous VMA name within the range [start, end).
2444 * As a result, we might be able to merge the newly modified VMA range with an
2445 * adjacent VMA with identical properties.
2447 * If no merge is possible and the range does not span the entirety of the VMA,
2448 * we then need to split the VMA to accommodate the change.
2450 * The function returns either the merged VMA, the original VMA if a split was
2451 * required instead, or an error if the split failed.
2453 struct vm_area_struct *vma_modify(struct vma_iterator *vmi,
2454 struct vm_area_struct *prev,
2455 struct vm_area_struct *vma,
2456 unsigned long start, unsigned long end,
2457 unsigned long vm_flags,
2458 struct mempolicy *policy,
2459 struct vm_userfaultfd_ctx uffd_ctx,
2460 struct anon_vma_name *anon_name)
2462 pgoff_t pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
2463 struct vm_area_struct *merged;
2465 merged = vma_merge(vmi, vma->vm_mm, prev, start, end, vm_flags,
2466 vma->anon_vma, vma->vm_file, pgoff, policy,
2467 uffd_ctx, anon_name);
2471 if (vma->vm_start < start) {
2472 int err = split_vma(vmi, vma, start, 1);
2475 return ERR_PTR(err);
2478 if (vma->vm_end > end) {
2479 int err = split_vma(vmi, vma, end, 0);
2482 return ERR_PTR(err);
2489 * do_vmi_align_munmap() - munmap the aligned region from @start to @end.
2490 * @vmi: The vma iterator
2491 * @vma: The starting vm_area_struct
2492 * @mm: The mm_struct
2493 * @start: The aligned start address to munmap.
2494 * @end: The aligned end address to munmap.
2495 * @uf: The userfaultfd list_head
2496 * @unlock: Set to true to drop the mmap_lock. unlocking only happens on
2499 * Return: 0 on success and drops the lock if so directed, error and leaves the
2500 * lock held otherwise.
2503 do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
2504 struct mm_struct *mm, unsigned long start,
2505 unsigned long end, struct list_head *uf, bool unlock)
2507 struct vm_area_struct *prev, *next = NULL;
2508 struct maple_tree mt_detach;
2510 int error = -ENOMEM;
2511 unsigned long locked_vm = 0;
2512 MA_STATE(mas_detach, &mt_detach, 0, 0);
2513 mt_init_flags(&mt_detach, vmi->mas.tree->ma_flags & MT_FLAGS_LOCK_MASK);
2514 mt_on_stack(mt_detach);
2517 * If we need to split any vma, do it now to save pain later.
2519 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2520 * unmapped vm_area_struct will remain in use: so lower split_vma
2521 * places tmp vma above, and higher split_vma places tmp vma below.
2524 /* Does it split the first one? */
2525 if (start > vma->vm_start) {
2528 * Make sure that map_count on return from munmap() will
2529 * not exceed its limit; but let map_count go just above
2530 * its limit temporarily, to help free resources as expected.
2532 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2533 goto map_count_exceeded;
2535 error = __split_vma(vmi, vma, start, 1);
2537 goto start_split_failed;
2541 * Detach a range of VMAs from the mm. Using next as a temp variable as
2542 * it is always overwritten.
2546 /* Does it split the end? */
2547 if (next->vm_end > end) {
2548 error = __split_vma(vmi, next, end, 0);
2550 goto end_split_failed;
2552 vma_start_write(next);
2553 mas_set(&mas_detach, count);
2554 error = mas_store_gfp(&mas_detach, next, GFP_KERNEL);
2556 goto munmap_gather_failed;
2557 vma_mark_detached(next, true);
2558 if (next->vm_flags & VM_LOCKED)
2559 locked_vm += vma_pages(next);
2564 * If userfaultfd_unmap_prep returns an error the vmas
2565 * will remain split, but userland will get a
2566 * highly unexpected error anyway. This is no
2567 * different than the case where the first of the two
2568 * __split_vma fails, but we don't undo the first
2569 * split, despite we could. This is unlikely enough
2570 * failure that it's not worth optimizing it for.
2572 error = userfaultfd_unmap_prep(next, start, end, uf);
2575 goto userfaultfd_error;
2577 #ifdef CONFIG_DEBUG_VM_MAPLE_TREE
2578 BUG_ON(next->vm_start < start);
2579 BUG_ON(next->vm_start > end);
2581 } for_each_vma_range(*vmi, next, end);
2583 #if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
2584 /* Make sure no VMAs are about to be lost. */
2586 MA_STATE(test, &mt_detach, 0, 0);
2587 struct vm_area_struct *vma_mas, *vma_test;
2590 vma_iter_set(vmi, start);
2592 vma_test = mas_find(&test, count - 1);
2593 for_each_vma_range(*vmi, vma_mas, end) {
2594 BUG_ON(vma_mas != vma_test);
2596 vma_test = mas_next(&test, count - 1);
2599 BUG_ON(count != test_count);
2603 while (vma_iter_addr(vmi) > start)
2604 vma_iter_prev_range(vmi);
2606 error = vma_iter_clear_gfp(vmi, start, end, GFP_KERNEL);
2608 goto clear_tree_failed;
2610 /* Point of no return */
2611 mm->locked_vm -= locked_vm;
2612 mm->map_count -= count;
2614 mmap_write_downgrade(mm);
2616 prev = vma_iter_prev_range(vmi);
2617 next = vma_next(vmi);
2619 vma_iter_prev_range(vmi);
2622 * We can free page tables without write-locking mmap_lock because VMAs
2623 * were isolated before we downgraded mmap_lock.
2625 mas_set(&mas_detach, 1);
2626 unmap_region(mm, &mas_detach, vma, prev, next, start, end, count,
2628 /* Statistics and freeing VMAs */
2629 mas_set(&mas_detach, 0);
2630 remove_mt(mm, &mas_detach);
2633 mmap_read_unlock(mm);
2635 __mt_destroy(&mt_detach);
2640 munmap_gather_failed:
2642 mas_set(&mas_detach, 0);
2643 mas_for_each(&mas_detach, next, end)
2644 vma_mark_detached(next, false);
2646 __mt_destroy(&mt_detach);
2654 * do_vmi_munmap() - munmap a given range.
2655 * @vmi: The vma iterator
2656 * @mm: The mm_struct
2657 * @start: The start address to munmap
2658 * @len: The length of the range to munmap
2659 * @uf: The userfaultfd list_head
2660 * @unlock: set to true if the user wants to drop the mmap_lock on success
2662 * This function takes a @mas that is either pointing to the previous VMA or set
2663 * to MA_START and sets it up to remove the mapping(s). The @len will be
2664 * aligned and any arch_unmap work will be preformed.
2666 * Return: 0 on success and drops the lock if so directed, error and leaves the
2667 * lock held otherwise.
2669 int do_vmi_munmap(struct vma_iterator *vmi, struct mm_struct *mm,
2670 unsigned long start, size_t len, struct list_head *uf,
2674 struct vm_area_struct *vma;
2676 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2679 end = start + PAGE_ALIGN(len);
2683 /* arch_unmap() might do unmaps itself. */
2684 arch_unmap(mm, start, end);
2686 /* Find the first overlapping VMA */
2687 vma = vma_find(vmi, end);
2690 mmap_write_unlock(mm);
2694 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
2697 /* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls.
2698 * @mm: The mm_struct
2699 * @start: The start address to munmap
2700 * @len: The length to be munmapped.
2701 * @uf: The userfaultfd list_head
2703 * Return: 0 on success, error otherwise.
2705 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2706 struct list_head *uf)
2708 VMA_ITERATOR(vmi, mm, start);
2710 return do_vmi_munmap(&vmi, mm, start, len, uf, false);
2713 unsigned long mmap_region(struct file *file, unsigned long addr,
2714 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
2715 struct list_head *uf)
2717 struct mm_struct *mm = current->mm;
2718 struct vm_area_struct *vma = NULL;
2719 struct vm_area_struct *next, *prev, *merge;
2720 pgoff_t pglen = len >> PAGE_SHIFT;
2721 unsigned long charged = 0;
2722 unsigned long end = addr + len;
2723 unsigned long merge_start = addr, merge_end = end;
2726 VMA_ITERATOR(vmi, mm, addr);
2728 /* Check against address space limit. */
2729 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
2730 unsigned long nr_pages;
2733 * MAP_FIXED may remove pages of mappings that intersects with
2734 * requested mapping. Account for the pages it would unmap.
2736 nr_pages = count_vma_pages_range(mm, addr, end);
2738 if (!may_expand_vm(mm, vm_flags,
2739 (len >> PAGE_SHIFT) - nr_pages))
2743 /* Unmap any existing mapping in the area */
2744 if (do_vmi_munmap(&vmi, mm, addr, len, uf, false))
2748 * Private writable mapping: check memory availability
2750 if (accountable_mapping(file, vm_flags)) {
2751 charged = len >> PAGE_SHIFT;
2752 if (security_vm_enough_memory_mm(mm, charged))
2754 vm_flags |= VM_ACCOUNT;
2757 next = vma_next(&vmi);
2758 prev = vma_prev(&vmi);
2759 if (vm_flags & VM_SPECIAL) {
2761 vma_iter_next_range(&vmi);
2765 /* Attempt to expand an old mapping */
2767 if (next && next->vm_start == end && !vma_policy(next) &&
2768 can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen,
2769 NULL_VM_UFFD_CTX, NULL)) {
2770 merge_end = next->vm_end;
2772 vm_pgoff = next->vm_pgoff - pglen;
2776 if (prev && prev->vm_end == addr && !vma_policy(prev) &&
2777 (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file,
2778 pgoff, vma->vm_userfaultfd_ctx, NULL) :
2779 can_vma_merge_after(prev, vm_flags, NULL, file, pgoff,
2780 NULL_VM_UFFD_CTX, NULL))) {
2781 merge_start = prev->vm_start;
2783 vm_pgoff = prev->vm_pgoff;
2785 vma_iter_next_range(&vmi);
2788 /* Actually expand, if possible */
2790 !vma_expand(&vmi, vma, merge_start, merge_end, vm_pgoff, next)) {
2791 khugepaged_enter_vma(vma, vm_flags);
2796 vma_iter_set(&vmi, addr);
2800 * Determine the object being mapped and call the appropriate
2801 * specific mapper. the address has already been validated, but
2802 * not unmapped, but the maps are removed from the list.
2804 vma = vm_area_alloc(mm);
2810 vma_iter_config(&vmi, addr, end);
2811 vma->vm_start = addr;
2813 vm_flags_init(vma, vm_flags);
2814 vma->vm_page_prot = vm_get_page_prot(vm_flags);
2815 vma->vm_pgoff = pgoff;
2818 if (vm_flags & VM_SHARED) {
2819 error = mapping_map_writable(file->f_mapping);
2824 vma->vm_file = get_file(file);
2825 error = call_mmap(file, vma);
2827 goto unmap_and_free_vma;
2830 * Expansion is handled above, merging is handled below.
2831 * Drivers should not alter the address of the VMA.
2834 if (WARN_ON((addr != vma->vm_start)))
2835 goto close_and_free_vma;
2837 vma_iter_config(&vmi, addr, end);
2839 * If vm_flags changed after call_mmap(), we should try merge
2840 * vma again as we may succeed this time.
2842 if (unlikely(vm_flags != vma->vm_flags && prev)) {
2843 merge = vma_merge(&vmi, mm, prev, vma->vm_start,
2844 vma->vm_end, vma->vm_flags, NULL,
2845 vma->vm_file, vma->vm_pgoff, NULL,
2846 NULL_VM_UFFD_CTX, NULL);
2849 * ->mmap() can change vma->vm_file and fput
2850 * the original file. So fput the vma->vm_file
2851 * here or we would add an extra fput for file
2852 * and cause general protection fault
2858 /* Update vm_flags to pick up the change. */
2859 vm_flags = vma->vm_flags;
2860 goto unmap_writable;
2864 vm_flags = vma->vm_flags;
2865 } else if (vm_flags & VM_SHARED) {
2866 error = shmem_zero_setup(vma);
2870 vma_set_anonymous(vma);
2873 if (map_deny_write_exec(vma, vma->vm_flags)) {
2875 goto close_and_free_vma;
2878 /* Allow architectures to sanity-check the vm_flags */
2880 if (!arch_validate_flags(vma->vm_flags))
2881 goto close_and_free_vma;
2884 if (vma_iter_prealloc(&vmi, vma))
2885 goto close_and_free_vma;
2887 /* Lock the VMA since it is modified after insertion into VMA tree */
2888 vma_start_write(vma);
2889 vma_iter_store(&vmi, vma);
2892 i_mmap_lock_write(vma->vm_file->f_mapping);
2893 if (vma->vm_flags & VM_SHARED)
2894 mapping_allow_writable(vma->vm_file->f_mapping);
2896 flush_dcache_mmap_lock(vma->vm_file->f_mapping);
2897 vma_interval_tree_insert(vma, &vma->vm_file->f_mapping->i_mmap);
2898 flush_dcache_mmap_unlock(vma->vm_file->f_mapping);
2899 i_mmap_unlock_write(vma->vm_file->f_mapping);
2903 * vma_merge() calls khugepaged_enter_vma() either, the below
2904 * call covers the non-merge case.
2906 khugepaged_enter_vma(vma, vma->vm_flags);
2908 /* Once vma denies write, undo our temporary denial count */
2910 if (file && vm_flags & VM_SHARED)
2911 mapping_unmap_writable(file->f_mapping);
2912 file = vma->vm_file;
2915 perf_event_mmap(vma);
2917 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
2918 if (vm_flags & VM_LOCKED) {
2919 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
2920 is_vm_hugetlb_page(vma) ||
2921 vma == get_gate_vma(current->mm))
2922 vm_flags_clear(vma, VM_LOCKED_MASK);
2924 mm->locked_vm += (len >> PAGE_SHIFT);
2931 * New (or expanded) vma always get soft dirty status.
2932 * Otherwise user-space soft-dirty page tracker won't
2933 * be able to distinguish situation when vma area unmapped,
2934 * then new mapped in-place (which must be aimed as
2935 * a completely new data area).
2937 vm_flags_set(vma, VM_SOFTDIRTY);
2939 vma_set_page_prot(vma);
2945 if (file && vma->vm_ops && vma->vm_ops->close)
2946 vma->vm_ops->close(vma);
2948 if (file || vma->vm_file) {
2951 vma->vm_file = NULL;
2953 vma_iter_set(&vmi, vma->vm_end);
2954 /* Undo any partial mapping done by a device driver. */
2955 unmap_region(mm, &vmi.mas, vma, prev, next, vma->vm_start,
2956 vma->vm_end, vma->vm_end, true);
2958 if (file && (vm_flags & VM_SHARED))
2959 mapping_unmap_writable(file->f_mapping);
2964 vm_unacct_memory(charged);
2969 static int __vm_munmap(unsigned long start, size_t len, bool unlock)
2972 struct mm_struct *mm = current->mm;
2974 VMA_ITERATOR(vmi, mm, start);
2976 if (mmap_write_lock_killable(mm))
2979 ret = do_vmi_munmap(&vmi, mm, start, len, &uf, unlock);
2981 mmap_write_unlock(mm);
2983 userfaultfd_unmap_complete(mm, &uf);
2987 int vm_munmap(unsigned long start, size_t len)
2989 return __vm_munmap(start, len, false);
2991 EXPORT_SYMBOL(vm_munmap);
2993 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2995 addr = untagged_addr(addr);
2996 return __vm_munmap(addr, len, true);
3001 * Emulation of deprecated remap_file_pages() syscall.
3003 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
3004 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
3007 struct mm_struct *mm = current->mm;
3008 struct vm_area_struct *vma;
3009 unsigned long populate = 0;
3010 unsigned long ret = -EINVAL;
3013 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
3014 current->comm, current->pid);
3018 start = start & PAGE_MASK;
3019 size = size & PAGE_MASK;
3021 if (start + size <= start)
3024 /* Does pgoff wrap? */
3025 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
3028 if (mmap_write_lock_killable(mm))
3031 vma = vma_lookup(mm, start);
3033 if (!vma || !(vma->vm_flags & VM_SHARED))
3036 if (start + size > vma->vm_end) {
3037 VMA_ITERATOR(vmi, mm, vma->vm_end);
3038 struct vm_area_struct *next, *prev = vma;
3040 for_each_vma_range(vmi, next, start + size) {
3041 /* hole between vmas ? */
3042 if (next->vm_start != prev->vm_end)
3045 if (next->vm_file != vma->vm_file)
3048 if (next->vm_flags != vma->vm_flags)
3051 if (start + size <= next->vm_end)
3061 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3062 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3063 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3065 flags &= MAP_NONBLOCK;
3066 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3067 if (vma->vm_flags & VM_LOCKED)
3068 flags |= MAP_LOCKED;
3070 file = get_file(vma->vm_file);
3071 ret = do_mmap(vma->vm_file, start, size,
3072 prot, flags, 0, pgoff, &populate, NULL);
3075 mmap_write_unlock(mm);
3077 mm_populate(ret, populate);
3078 if (!IS_ERR_VALUE(ret))
3084 * do_vma_munmap() - Unmap a full or partial vma.
3085 * @vmi: The vma iterator pointing at the vma
3086 * @vma: The first vma to be munmapped
3087 * @start: the start of the address to unmap
3088 * @end: The end of the address to unmap
3089 * @uf: The userfaultfd list_head
3090 * @unlock: Drop the lock on success
3092 * unmaps a VMA mapping when the vma iterator is already in position.
3093 * Does not handle alignment.
3095 * Return: 0 on success drops the lock of so directed, error on failure and will
3096 * still hold the lock.
3098 int do_vma_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
3099 unsigned long start, unsigned long end, struct list_head *uf,
3102 struct mm_struct *mm = vma->vm_mm;
3104 arch_unmap(mm, start, end);
3105 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
3109 * do_brk_flags() - Increase the brk vma if the flags match.
3110 * @vmi: The vma iterator
3111 * @addr: The start address
3112 * @len: The length of the increase
3114 * @flags: The VMA Flags
3116 * Extend the brk VMA from addr to addr + len. If the VMA is NULL or the flags
3117 * do not match then create a new anonymous VMA. Eventually we may be able to
3118 * do some brk-specific accounting here.
3120 static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *vma,
3121 unsigned long addr, unsigned long len, unsigned long flags)
3123 struct mm_struct *mm = current->mm;
3124 struct vma_prepare vp;
3127 * Check against address space limits by the changed size
3128 * Note: This happens *after* clearing old mappings in some code paths.
3130 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3131 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3134 if (mm->map_count > sysctl_max_map_count)
3137 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3141 * Expand the existing vma if possible; Note that singular lists do not
3142 * occur after forking, so the expand will only happen on new VMAs.
3144 if (vma && vma->vm_end == addr && !vma_policy(vma) &&
3145 can_vma_merge_after(vma, flags, NULL, NULL,
3146 addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) {
3147 vma_iter_config(vmi, vma->vm_start, addr + len);
3148 if (vma_iter_prealloc(vmi, vma))
3151 vma_start_write(vma);
3153 init_vma_prep(&vp, vma);
3155 vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0);
3156 vma->vm_end = addr + len;
3157 vm_flags_set(vma, VM_SOFTDIRTY);
3158 vma_iter_store(vmi, vma);
3160 vma_complete(&vp, vmi, mm);
3161 khugepaged_enter_vma(vma, flags);
3166 vma_iter_next_range(vmi);
3167 /* create a vma struct for an anonymous mapping */
3168 vma = vm_area_alloc(mm);
3172 vma_set_anonymous(vma);
3173 vma->vm_start = addr;
3174 vma->vm_end = addr + len;
3175 vma->vm_pgoff = addr >> PAGE_SHIFT;
3176 vm_flags_init(vma, flags);
3177 vma->vm_page_prot = vm_get_page_prot(flags);
3178 vma_start_write(vma);
3179 if (vma_iter_store_gfp(vmi, vma, GFP_KERNEL))
3180 goto mas_store_fail;
3186 perf_event_mmap(vma);
3187 mm->total_vm += len >> PAGE_SHIFT;
3188 mm->data_vm += len >> PAGE_SHIFT;
3189 if (flags & VM_LOCKED)
3190 mm->locked_vm += (len >> PAGE_SHIFT);
3191 vm_flags_set(vma, VM_SOFTDIRTY);
3197 vm_unacct_memory(len >> PAGE_SHIFT);
3201 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3203 struct mm_struct *mm = current->mm;
3204 struct vm_area_struct *vma = NULL;
3209 VMA_ITERATOR(vmi, mm, addr);
3211 len = PAGE_ALIGN(request);
3217 /* Until we need other flags, refuse anything except VM_EXEC. */
3218 if ((flags & (~VM_EXEC)) != 0)
3221 if (mmap_write_lock_killable(mm))
3224 ret = check_brk_limits(addr, len);
3228 ret = do_vmi_munmap(&vmi, mm, addr, len, &uf, 0);
3232 vma = vma_prev(&vmi);
3233 ret = do_brk_flags(&vmi, vma, addr, len, flags);
3234 populate = ((mm->def_flags & VM_LOCKED) != 0);
3235 mmap_write_unlock(mm);
3236 userfaultfd_unmap_complete(mm, &uf);
3237 if (populate && !ret)
3238 mm_populate(addr, len);
3243 mmap_write_unlock(mm);
3246 EXPORT_SYMBOL(vm_brk_flags);
3248 int vm_brk(unsigned long addr, unsigned long len)
3250 return vm_brk_flags(addr, len, 0);
3252 EXPORT_SYMBOL(vm_brk);
3254 /* Release all mmaps. */
3255 void exit_mmap(struct mm_struct *mm)
3257 struct mmu_gather tlb;
3258 struct vm_area_struct *vma;
3259 unsigned long nr_accounted = 0;
3260 MA_STATE(mas, &mm->mm_mt, 0, 0);
3263 /* mm's last user has gone, and its about to be pulled down */
3264 mmu_notifier_release(mm);
3269 vma = mas_find(&mas, ULONG_MAX);
3271 /* Can happen if dup_mmap() received an OOM */
3272 mmap_read_unlock(mm);
3278 tlb_gather_mmu_fullmm(&tlb, mm);
3279 /* update_hiwater_rss(mm) here? but nobody should be looking */
3280 /* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */
3281 unmap_vmas(&tlb, &mas, vma, 0, ULONG_MAX, ULONG_MAX, false);
3282 mmap_read_unlock(mm);
3285 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
3286 * because the memory has been already freed.
3288 set_bit(MMF_OOM_SKIP, &mm->flags);
3289 mmap_write_lock(mm);
3290 mt_clear_in_rcu(&mm->mm_mt);
3291 mas_set(&mas, vma->vm_end);
3292 free_pgtables(&tlb, &mas, vma, FIRST_USER_ADDRESS,
3293 USER_PGTABLES_CEILING, true);
3294 tlb_finish_mmu(&tlb);
3297 * Walk the list again, actually closing and freeing it, with preemption
3298 * enabled, without holding any MM locks besides the unreachable
3301 mas_set(&mas, vma->vm_end);
3303 if (vma->vm_flags & VM_ACCOUNT)
3304 nr_accounted += vma_pages(vma);
3305 remove_vma(vma, true);
3308 } while ((vma = mas_find(&mas, ULONG_MAX)) != NULL);
3310 BUG_ON(count != mm->map_count);
3312 trace_exit_mmap(mm);
3313 __mt_destroy(&mm->mm_mt);
3314 mmap_write_unlock(mm);
3315 vm_unacct_memory(nr_accounted);
3318 /* Insert vm structure into process list sorted by address
3319 * and into the inode's i_mmap tree. If vm_file is non-NULL
3320 * then i_mmap_rwsem is taken here.
3322 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3324 unsigned long charged = vma_pages(vma);
3327 if (find_vma_intersection(mm, vma->vm_start, vma->vm_end))
3330 if ((vma->vm_flags & VM_ACCOUNT) &&
3331 security_vm_enough_memory_mm(mm, charged))
3335 * The vm_pgoff of a purely anonymous vma should be irrelevant
3336 * until its first write fault, when page's anon_vma and index
3337 * are set. But now set the vm_pgoff it will almost certainly
3338 * end up with (unless mremap moves it elsewhere before that
3339 * first wfault), so /proc/pid/maps tells a consistent story.
3341 * By setting it to reflect the virtual start address of the
3342 * vma, merges and splits can happen in a seamless way, just
3343 * using the existing file pgoff checks and manipulations.
3344 * Similarly in do_mmap and in do_brk_flags.
3346 if (vma_is_anonymous(vma)) {
3347 BUG_ON(vma->anon_vma);
3348 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3351 if (vma_link(mm, vma)) {
3352 if (vma->vm_flags & VM_ACCOUNT)
3353 vm_unacct_memory(charged);
3361 * Copy the vma structure to a new location in the same mm,
3362 * prior to moving page table entries, to effect an mremap move.
3364 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3365 unsigned long addr, unsigned long len, pgoff_t pgoff,
3366 bool *need_rmap_locks)
3368 struct vm_area_struct *vma = *vmap;
3369 unsigned long vma_start = vma->vm_start;
3370 struct mm_struct *mm = vma->vm_mm;
3371 struct vm_area_struct *new_vma, *prev;
3372 bool faulted_in_anon_vma = true;
3373 VMA_ITERATOR(vmi, mm, addr);
3376 * If anonymous vma has not yet been faulted, update new pgoff
3377 * to match new location, to increase its chance of merging.
3379 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3380 pgoff = addr >> PAGE_SHIFT;
3381 faulted_in_anon_vma = false;
3384 new_vma = find_vma_prev(mm, addr, &prev);
3385 if (new_vma && new_vma->vm_start < addr + len)
3386 return NULL; /* should never get here */
3388 new_vma = vma_merge(&vmi, mm, prev, addr, addr + len, vma->vm_flags,
3389 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3390 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
3393 * Source vma may have been merged into new_vma
3395 if (unlikely(vma_start >= new_vma->vm_start &&
3396 vma_start < new_vma->vm_end)) {
3398 * The only way we can get a vma_merge with
3399 * self during an mremap is if the vma hasn't
3400 * been faulted in yet and we were allowed to
3401 * reset the dst vma->vm_pgoff to the
3402 * destination address of the mremap to allow
3403 * the merge to happen. mremap must change the
3404 * vm_pgoff linearity between src and dst vmas
3405 * (in turn preventing a vma_merge) to be
3406 * safe. It is only safe to keep the vm_pgoff
3407 * linear if there are no pages mapped yet.
3409 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3410 *vmap = vma = new_vma;
3412 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3414 new_vma = vm_area_dup(vma);
3417 new_vma->vm_start = addr;
3418 new_vma->vm_end = addr + len;
3419 new_vma->vm_pgoff = pgoff;
3420 if (vma_dup_policy(vma, new_vma))
3422 if (anon_vma_clone(new_vma, vma))
3423 goto out_free_mempol;
3424 if (new_vma->vm_file)
3425 get_file(new_vma->vm_file);
3426 if (new_vma->vm_ops && new_vma->vm_ops->open)
3427 new_vma->vm_ops->open(new_vma);
3428 if (vma_link(mm, new_vma))
3430 *need_rmap_locks = false;
3435 if (new_vma->vm_ops && new_vma->vm_ops->close)
3436 new_vma->vm_ops->close(new_vma);
3438 if (new_vma->vm_file)
3439 fput(new_vma->vm_file);
3441 unlink_anon_vmas(new_vma);
3443 mpol_put(vma_policy(new_vma));
3445 vm_area_free(new_vma);
3451 * Return true if the calling process may expand its vm space by the passed
3454 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3456 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3459 if (is_data_mapping(flags) &&
3460 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3461 /* Workaround for Valgrind */
3462 if (rlimit(RLIMIT_DATA) == 0 &&
3463 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3466 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3467 current->comm, current->pid,
3468 (mm->data_vm + npages) << PAGE_SHIFT,
3469 rlimit(RLIMIT_DATA),
3470 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3472 if (!ignore_rlimit_data)
3479 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3481 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3483 if (is_exec_mapping(flags))
3484 mm->exec_vm += npages;
3485 else if (is_stack_mapping(flags))
3486 mm->stack_vm += npages;
3487 else if (is_data_mapping(flags))
3488 mm->data_vm += npages;
3491 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3494 * Having a close hook prevents vma merging regardless of flags.
3496 static void special_mapping_close(struct vm_area_struct *vma)
3500 static const char *special_mapping_name(struct vm_area_struct *vma)
3502 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3505 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3507 struct vm_special_mapping *sm = new_vma->vm_private_data;
3509 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3513 return sm->mremap(sm, new_vma);
3518 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3521 * Forbid splitting special mappings - kernel has expectations over
3522 * the number of pages in mapping. Together with VM_DONTEXPAND
3523 * the size of vma should stay the same over the special mapping's
3529 static const struct vm_operations_struct special_mapping_vmops = {
3530 .close = special_mapping_close,
3531 .fault = special_mapping_fault,
3532 .mremap = special_mapping_mremap,
3533 .name = special_mapping_name,
3534 /* vDSO code relies that VVAR can't be accessed remotely */
3536 .may_split = special_mapping_split,
3539 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3540 .close = special_mapping_close,
3541 .fault = special_mapping_fault,
3544 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3546 struct vm_area_struct *vma = vmf->vma;
3548 struct page **pages;
3550 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3551 pages = vma->vm_private_data;
3553 struct vm_special_mapping *sm = vma->vm_private_data;
3556 return sm->fault(sm, vmf->vma, vmf);
3561 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3565 struct page *page = *pages;
3571 return VM_FAULT_SIGBUS;
3574 static struct vm_area_struct *__install_special_mapping(
3575 struct mm_struct *mm,
3576 unsigned long addr, unsigned long len,
3577 unsigned long vm_flags, void *priv,
3578 const struct vm_operations_struct *ops)
3581 struct vm_area_struct *vma;
3583 vma = vm_area_alloc(mm);
3584 if (unlikely(vma == NULL))
3585 return ERR_PTR(-ENOMEM);
3587 vma->vm_start = addr;
3588 vma->vm_end = addr + len;
3590 vm_flags_init(vma, (vm_flags | mm->def_flags |
3591 VM_DONTEXPAND | VM_SOFTDIRTY) & ~VM_LOCKED_MASK);
3592 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3595 vma->vm_private_data = priv;
3597 ret = insert_vm_struct(mm, vma);
3601 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3603 perf_event_mmap(vma);
3609 return ERR_PTR(ret);
3612 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3613 const struct vm_special_mapping *sm)
3615 return vma->vm_private_data == sm &&
3616 (vma->vm_ops == &special_mapping_vmops ||
3617 vma->vm_ops == &legacy_special_mapping_vmops);
3621 * Called with mm->mmap_lock held for writing.
3622 * Insert a new vma covering the given region, with the given flags.
3623 * Its pages are supplied by the given array of struct page *.
3624 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3625 * The region past the last page supplied will always produce SIGBUS.
3626 * The array pointer and the pages it points to are assumed to stay alive
3627 * for as long as this mapping might exist.
3629 struct vm_area_struct *_install_special_mapping(
3630 struct mm_struct *mm,
3631 unsigned long addr, unsigned long len,
3632 unsigned long vm_flags, const struct vm_special_mapping *spec)
3634 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3635 &special_mapping_vmops);
3638 int install_special_mapping(struct mm_struct *mm,
3639 unsigned long addr, unsigned long len,
3640 unsigned long vm_flags, struct page **pages)
3642 struct vm_area_struct *vma = __install_special_mapping(
3643 mm, addr, len, vm_flags, (void *)pages,
3644 &legacy_special_mapping_vmops);
3646 return PTR_ERR_OR_ZERO(vma);
3649 static DEFINE_MUTEX(mm_all_locks_mutex);
3651 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3653 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3655 * The LSB of head.next can't change from under us
3656 * because we hold the mm_all_locks_mutex.
3658 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3660 * We can safely modify head.next after taking the
3661 * anon_vma->root->rwsem. If some other vma in this mm shares
3662 * the same anon_vma we won't take it again.
3664 * No need of atomic instructions here, head.next
3665 * can't change from under us thanks to the
3666 * anon_vma->root->rwsem.
3668 if (__test_and_set_bit(0, (unsigned long *)
3669 &anon_vma->root->rb_root.rb_root.rb_node))
3674 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3676 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3678 * AS_MM_ALL_LOCKS can't change from under us because
3679 * we hold the mm_all_locks_mutex.
3681 * Operations on ->flags have to be atomic because
3682 * even if AS_MM_ALL_LOCKS is stable thanks to the
3683 * mm_all_locks_mutex, there may be other cpus
3684 * changing other bitflags in parallel to us.
3686 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3688 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3693 * This operation locks against the VM for all pte/vma/mm related
3694 * operations that could ever happen on a certain mm. This includes
3695 * vmtruncate, try_to_unmap, and all page faults.
3697 * The caller must take the mmap_lock in write mode before calling
3698 * mm_take_all_locks(). The caller isn't allowed to release the
3699 * mmap_lock until mm_drop_all_locks() returns.
3701 * mmap_lock in write mode is required in order to block all operations
3702 * that could modify pagetables and free pages without need of
3703 * altering the vma layout. It's also needed in write mode to avoid new
3704 * anon_vmas to be associated with existing vmas.
3706 * A single task can't take more than one mm_take_all_locks() in a row
3707 * or it would deadlock.
3709 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3710 * mapping->flags avoid to take the same lock twice, if more than one
3711 * vma in this mm is backed by the same anon_vma or address_space.
3713 * We take locks in following order, accordingly to comment at beginning
3715 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3717 * - all vmas marked locked
3718 * - all i_mmap_rwsem locks;
3719 * - all anon_vma->rwseml
3721 * We can take all locks within these types randomly because the VM code
3722 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3723 * mm_all_locks_mutex.
3725 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3726 * that may have to take thousand of locks.
3728 * mm_take_all_locks() can fail if it's interrupted by signals.
3730 int mm_take_all_locks(struct mm_struct *mm)
3732 struct vm_area_struct *vma;
3733 struct anon_vma_chain *avc;
3734 MA_STATE(mas, &mm->mm_mt, 0, 0);
3736 mmap_assert_write_locked(mm);
3738 mutex_lock(&mm_all_locks_mutex);
3741 * vma_start_write() does not have a complement in mm_drop_all_locks()
3742 * because vma_start_write() is always asymmetrical; it marks a VMA as
3743 * being written to until mmap_write_unlock() or mmap_write_downgrade()
3746 mas_for_each(&mas, vma, ULONG_MAX) {
3747 if (signal_pending(current))
3749 vma_start_write(vma);
3753 mas_for_each(&mas, vma, ULONG_MAX) {
3754 if (signal_pending(current))
3756 if (vma->vm_file && vma->vm_file->f_mapping &&
3757 is_vm_hugetlb_page(vma))
3758 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3762 mas_for_each(&mas, vma, ULONG_MAX) {
3763 if (signal_pending(current))
3765 if (vma->vm_file && vma->vm_file->f_mapping &&
3766 !is_vm_hugetlb_page(vma))
3767 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3771 mas_for_each(&mas, vma, ULONG_MAX) {
3772 if (signal_pending(current))
3775 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3776 vm_lock_anon_vma(mm, avc->anon_vma);
3782 mm_drop_all_locks(mm);
3786 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3788 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3790 * The LSB of head.next can't change to 0 from under
3791 * us because we hold the mm_all_locks_mutex.
3793 * We must however clear the bitflag before unlocking
3794 * the vma so the users using the anon_vma->rb_root will
3795 * never see our bitflag.
3797 * No need of atomic instructions here, head.next
3798 * can't change from under us until we release the
3799 * anon_vma->root->rwsem.
3801 if (!__test_and_clear_bit(0, (unsigned long *)
3802 &anon_vma->root->rb_root.rb_root.rb_node))
3804 anon_vma_unlock_write(anon_vma);
3808 static void vm_unlock_mapping(struct address_space *mapping)
3810 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3812 * AS_MM_ALL_LOCKS can't change to 0 from under us
3813 * because we hold the mm_all_locks_mutex.
3815 i_mmap_unlock_write(mapping);
3816 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3823 * The mmap_lock cannot be released by the caller until
3824 * mm_drop_all_locks() returns.
3826 void mm_drop_all_locks(struct mm_struct *mm)
3828 struct vm_area_struct *vma;
3829 struct anon_vma_chain *avc;
3830 MA_STATE(mas, &mm->mm_mt, 0, 0);
3832 mmap_assert_write_locked(mm);
3833 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3835 mas_for_each(&mas, vma, ULONG_MAX) {
3837 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3838 vm_unlock_anon_vma(avc->anon_vma);
3839 if (vma->vm_file && vma->vm_file->f_mapping)
3840 vm_unlock_mapping(vma->vm_file->f_mapping);
3843 mutex_unlock(&mm_all_locks_mutex);
3847 * initialise the percpu counter for VM
3849 void __init mmap_init(void)
3853 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3858 * Initialise sysctl_user_reserve_kbytes.
3860 * This is intended to prevent a user from starting a single memory hogging
3861 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3864 * The default value is min(3% of free memory, 128MB)
3865 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3867 static int init_user_reserve(void)
3869 unsigned long free_kbytes;
3871 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3873 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3876 subsys_initcall(init_user_reserve);
3879 * Initialise sysctl_admin_reserve_kbytes.
3881 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3882 * to log in and kill a memory hogging process.
3884 * Systems with more than 256MB will reserve 8MB, enough to recover
3885 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3886 * only reserve 3% of free pages by default.
3888 static int init_admin_reserve(void)
3890 unsigned long free_kbytes;
3892 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3894 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3897 subsys_initcall(init_admin_reserve);
3900 * Reinititalise user and admin reserves if memory is added or removed.
3902 * The default user reserve max is 128MB, and the default max for the
3903 * admin reserve is 8MB. These are usually, but not always, enough to
3904 * enable recovery from a memory hogging process using login/sshd, a shell,
3905 * and tools like top. It may make sense to increase or even disable the
3906 * reserve depending on the existence of swap or variations in the recovery
3907 * tools. So, the admin may have changed them.
3909 * If memory is added and the reserves have been eliminated or increased above
3910 * the default max, then we'll trust the admin.
3912 * If memory is removed and there isn't enough free memory, then we
3913 * need to reset the reserves.
3915 * Otherwise keep the reserve set by the admin.
3917 static int reserve_mem_notifier(struct notifier_block *nb,
3918 unsigned long action, void *data)
3920 unsigned long tmp, free_kbytes;
3924 /* Default max is 128MB. Leave alone if modified by operator. */
3925 tmp = sysctl_user_reserve_kbytes;
3926 if (0 < tmp && tmp < (1UL << 17))
3927 init_user_reserve();
3929 /* Default max is 8MB. Leave alone if modified by operator. */
3930 tmp = sysctl_admin_reserve_kbytes;
3931 if (0 < tmp && tmp < (1UL << 13))
3932 init_admin_reserve();
3936 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3938 if (sysctl_user_reserve_kbytes > free_kbytes) {
3939 init_user_reserve();
3940 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3941 sysctl_user_reserve_kbytes);
3944 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3945 init_admin_reserve();
3946 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3947 sysctl_admin_reserve_kbytes);
3956 static int __meminit init_reserve_notifier(void)
3958 if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI))
3959 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3963 subsys_initcall(init_reserve_notifier);