1 // SPDX-License-Identifier: GPL-2.0
5 * (C) Copyright 1996 Linus Torvalds
7 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
8 * (C) Copyright 2002 Red Hat Inc, All Rights Reserved
12 #include <linux/hugetlb.h>
13 #include <linux/shm.h>
14 #include <linux/ksm.h>
15 #include <linux/mman.h>
16 #include <linux/swap.h>
17 #include <linux/capability.h>
19 #include <linux/swapops.h>
20 #include <linux/highmem.h>
21 #include <linux/security.h>
22 #include <linux/syscalls.h>
23 #include <linux/mmu_notifier.h>
24 #include <linux/uaccess.h>
25 #include <linux/mm-arch-hooks.h>
26 #include <linux/userfaultfd_k.h>
28 #include <asm/cacheflush.h>
29 #include <asm/tlbflush.h>
33 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
40 pgd = pgd_offset(mm, addr);
41 if (pgd_none_or_clear_bad(pgd))
44 p4d = p4d_offset(pgd, addr);
45 if (p4d_none_or_clear_bad(p4d))
48 pud = pud_offset(p4d, addr);
49 if (pud_none_or_clear_bad(pud))
52 pmd = pmd_offset(pud, addr);
59 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
67 pgd = pgd_offset(mm, addr);
68 p4d = p4d_alloc(mm, pgd, addr);
71 pud = pud_alloc(mm, p4d, addr);
75 pmd = pmd_alloc(mm, pud, addr);
79 VM_BUG_ON(pmd_trans_huge(*pmd));
84 static void take_rmap_locks(struct vm_area_struct *vma)
87 i_mmap_lock_write(vma->vm_file->f_mapping);
89 anon_vma_lock_write(vma->anon_vma);
92 static void drop_rmap_locks(struct vm_area_struct *vma)
95 anon_vma_unlock_write(vma->anon_vma);
97 i_mmap_unlock_write(vma->vm_file->f_mapping);
100 static pte_t move_soft_dirty_pte(pte_t pte)
103 * Set soft dirty bit so we can notice
104 * in userspace the ptes were moved.
106 #ifdef CONFIG_MEM_SOFT_DIRTY
107 if (pte_present(pte))
108 pte = pte_mksoft_dirty(pte);
109 else if (is_swap_pte(pte))
110 pte = pte_swp_mksoft_dirty(pte);
115 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
116 unsigned long old_addr, unsigned long old_end,
117 struct vm_area_struct *new_vma, pmd_t *new_pmd,
118 unsigned long new_addr, bool need_rmap_locks)
120 struct mm_struct *mm = vma->vm_mm;
121 pte_t *old_pte, *new_pte, pte;
122 spinlock_t *old_ptl, *new_ptl;
123 bool force_flush = false;
124 unsigned long len = old_end - old_addr;
127 * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
128 * locks to ensure that rmap will always observe either the old or the
129 * new ptes. This is the easiest way to avoid races with
130 * truncate_pagecache(), page migration, etc...
132 * When need_rmap_locks is false, we use other ways to avoid
135 * - During exec() shift_arg_pages(), we use a specially tagged vma
136 * which rmap call sites look for using vma_is_temporary_stack().
138 * - During mremap(), new_vma is often known to be placed after vma
139 * in rmap traversal order. This ensures rmap will always observe
140 * either the old pte, or the new pte, or both (the page table locks
141 * serialize access to individual ptes, but only rmap traversal
142 * order guarantees that we won't miss both the old and new ptes).
145 take_rmap_locks(vma);
148 * We don't have to worry about the ordering of src and dst
149 * pte locks because exclusive mmap_lock prevents deadlock.
151 old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
152 new_pte = pte_offset_map(new_pmd, new_addr);
153 new_ptl = pte_lockptr(mm, new_pmd);
154 if (new_ptl != old_ptl)
155 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
156 flush_tlb_batched_pending(vma->vm_mm);
157 arch_enter_lazy_mmu_mode();
159 for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
160 new_pte++, new_addr += PAGE_SIZE) {
161 if (pte_none(*old_pte))
164 pte = ptep_get_and_clear(mm, old_addr, old_pte);
166 * If we are remapping a valid PTE, make sure
167 * to flush TLB before we drop the PTL for the
170 * NOTE! Both old and new PTL matter: the old one
171 * for racing with page_mkclean(), the new one to
172 * make sure the physical page stays valid until
173 * the TLB entry for the old mapping has been
176 if (pte_present(pte))
178 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
179 pte = move_soft_dirty_pte(pte);
180 set_pte_at(mm, new_addr, new_pte, pte);
183 arch_leave_lazy_mmu_mode();
185 flush_tlb_range(vma, old_end - len, old_end);
186 if (new_ptl != old_ptl)
187 spin_unlock(new_ptl);
188 pte_unmap(new_pte - 1);
189 pte_unmap_unlock(old_pte - 1, old_ptl);
191 drop_rmap_locks(vma);
194 #ifdef CONFIG_HAVE_MOVE_PMD
195 static bool move_normal_pmd(struct vm_area_struct *vma, unsigned long old_addr,
196 unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
198 spinlock_t *old_ptl, *new_ptl;
199 struct mm_struct *mm = vma->vm_mm;
202 if ((old_addr & ~PMD_MASK) || (new_addr & ~PMD_MASK))
206 * The destination pmd shouldn't be established, free_pgtables()
207 * should have released it.
209 * However, there's a case during execve() where we use mremap
210 * to move the initial stack, and in that case the target area
211 * may overlap the source area (always moving down).
213 * If everything is PMD-aligned, that works fine, as moving
214 * each pmd down will clear the source pmd. But if we first
215 * have a few 4kB-only pages that get moved down, and then
216 * hit the "now the rest is PMD-aligned, let's do everything
217 * one pmd at a time", we will still have the old (now empty
218 * of any 4kB pages, but still there) PMD in the page table
221 * Warn on it once - because we really should try to figure
222 * out how to do this better - but then say "I won't move
225 * One alternative might be to just unmap the target pmd at
226 * this point, and verify that it really is empty. We'll see.
228 if (WARN_ON_ONCE(!pmd_none(*new_pmd)))
232 * We don't have to worry about the ordering of src and dst
233 * ptlocks because exclusive mmap_lock prevents deadlock.
235 old_ptl = pmd_lock(vma->vm_mm, old_pmd);
236 new_ptl = pmd_lockptr(mm, new_pmd);
237 if (new_ptl != old_ptl)
238 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
244 VM_BUG_ON(!pmd_none(*new_pmd));
246 /* Set the new pmd */
247 set_pmd_at(mm, new_addr, new_pmd, pmd);
248 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
249 if (new_ptl != old_ptl)
250 spin_unlock(new_ptl);
251 spin_unlock(old_ptl);
257 unsigned long move_page_tables(struct vm_area_struct *vma,
258 unsigned long old_addr, struct vm_area_struct *new_vma,
259 unsigned long new_addr, unsigned long len,
260 bool need_rmap_locks)
262 unsigned long extent, next, old_end;
263 struct mmu_notifier_range range;
264 pmd_t *old_pmd, *new_pmd;
266 old_end = old_addr + len;
267 flush_cache_range(vma, old_addr, old_end);
269 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
271 mmu_notifier_invalidate_range_start(&range);
273 for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
275 next = (old_addr + PMD_SIZE) & PMD_MASK;
276 /* even if next overflowed, extent below will be ok */
277 extent = next - old_addr;
278 if (extent > old_end - old_addr)
279 extent = old_end - old_addr;
280 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
283 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
286 if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) || pmd_devmap(*old_pmd)) {
287 if (extent == HPAGE_PMD_SIZE) {
289 /* See comment in move_ptes() */
291 take_rmap_locks(vma);
292 moved = move_huge_pmd(vma, old_addr, new_addr,
295 drop_rmap_locks(vma);
299 split_huge_pmd(vma, old_pmd, old_addr);
300 if (pmd_trans_unstable(old_pmd))
302 } else if (extent == PMD_SIZE) {
303 #ifdef CONFIG_HAVE_MOVE_PMD
305 * If the extent is PMD-sized, try to speed the move by
306 * moving at the PMD level if possible.
311 take_rmap_locks(vma);
312 moved = move_normal_pmd(vma, old_addr, new_addr,
315 drop_rmap_locks(vma);
321 if (pte_alloc(new_vma->vm_mm, new_pmd))
323 next = (new_addr + PMD_SIZE) & PMD_MASK;
324 if (extent > next - new_addr)
325 extent = next - new_addr;
326 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
327 new_pmd, new_addr, need_rmap_locks);
330 mmu_notifier_invalidate_range_end(&range);
332 return len + old_addr - old_end; /* how much done */
335 static unsigned long move_vma(struct vm_area_struct *vma,
336 unsigned long old_addr, unsigned long old_len,
337 unsigned long new_len, unsigned long new_addr,
338 bool *locked, unsigned long flags,
339 struct vm_userfaultfd_ctx *uf, struct list_head *uf_unmap)
341 struct mm_struct *mm = vma->vm_mm;
342 struct vm_area_struct *new_vma;
343 unsigned long vm_flags = vma->vm_flags;
344 unsigned long new_pgoff;
345 unsigned long moved_len;
346 unsigned long excess = 0;
347 unsigned long hiwater_vm;
350 bool need_rmap_locks;
353 * We'd prefer to avoid failure later on in do_munmap:
354 * which may split one vma into three before unmapping.
356 if (mm->map_count >= sysctl_max_map_count - 3)
360 * Advise KSM to break any KSM pages in the area to be moved:
361 * it would be confusing if they were to turn up at the new
362 * location, where they happen to coincide with different KSM
363 * pages recently unmapped. But leave vma->vm_flags as it was,
364 * so KSM can come around to merge on vma and new_vma afterwards.
366 err = ksm_madvise(vma, old_addr, old_addr + old_len,
367 MADV_UNMERGEABLE, &vm_flags);
371 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
372 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
377 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
379 if (moved_len < old_len) {
381 } else if (vma->vm_ops && vma->vm_ops->mremap) {
382 err = vma->vm_ops->mremap(new_vma);
387 * On error, move entries back from new area to old,
388 * which will succeed since page tables still there,
389 * and then proceed to unmap new area instead of old.
391 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
398 mremap_userfaultfd_prep(new_vma, uf);
399 arch_remap(mm, old_addr, old_addr + old_len,
400 new_addr, new_addr + new_len);
403 /* Conceal VM_ACCOUNT so old reservation is not undone */
404 if (vm_flags & VM_ACCOUNT) {
405 vma->vm_flags &= ~VM_ACCOUNT;
406 excess = vma->vm_end - vma->vm_start - old_len;
407 if (old_addr > vma->vm_start &&
408 old_addr + old_len < vma->vm_end)
413 * If we failed to move page tables we still do total_vm increment
414 * since do_munmap() will decrement it by old_len == new_len.
416 * Since total_vm is about to be raised artificially high for a
417 * moment, we need to restore high watermark afterwards: if stats
418 * are taken meanwhile, total_vm and hiwater_vm appear too high.
419 * If this were a serious issue, we'd add a flag to do_munmap().
421 hiwater_vm = mm->hiwater_vm;
422 vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
424 /* Tell pfnmap has moved from this vma */
425 if (unlikely(vma->vm_flags & VM_PFNMAP))
426 untrack_pfn_moved(vma);
428 if (unlikely(!err && (flags & MREMAP_DONTUNMAP))) {
429 if (vm_flags & VM_ACCOUNT) {
430 /* Always put back VM_ACCOUNT since we won't unmap */
431 vma->vm_flags |= VM_ACCOUNT;
433 vm_acct_memory(new_len >> PAGE_SHIFT);
437 * VMAs can actually be merged back together in copy_vma
438 * calling merge_vma. This can happen with anonymous vmas
439 * which have not yet been faulted, so if we were to consider
440 * this VMA split we'll end up adding VM_ACCOUNT on the
441 * next VMA, which is completely unrelated if this VMA
444 if (split && new_vma == vma)
447 /* We always clear VM_LOCKED[ONFAULT] on the old vma */
448 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
450 /* Because we won't unmap we don't need to touch locked_vm */
454 if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
455 /* OOM: unable to split vma, just get accounts right */
456 vm_unacct_memory(excess >> PAGE_SHIFT);
460 if (vm_flags & VM_LOCKED) {
461 mm->locked_vm += new_len >> PAGE_SHIFT;
465 mm->hiwater_vm = hiwater_vm;
467 /* Restore VM_ACCOUNT if one or two pieces of vma left */
469 vma->vm_flags |= VM_ACCOUNT;
471 vma->vm_next->vm_flags |= VM_ACCOUNT;
477 static struct vm_area_struct *vma_to_resize(unsigned long addr,
478 unsigned long old_len, unsigned long new_len, unsigned long flags,
481 struct mm_struct *mm = current->mm;
482 struct vm_area_struct *vma = find_vma(mm, addr);
485 if (!vma || vma->vm_start > addr)
486 return ERR_PTR(-EFAULT);
489 * !old_len is a special case where an attempt is made to 'duplicate'
490 * a mapping. This makes no sense for private mappings as it will
491 * instead create a fresh/new mapping unrelated to the original. This
492 * is contrary to the basic idea of mremap which creates new mappings
493 * based on the original. There are no known use cases for this
494 * behavior. As a result, fail such attempts.
496 if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
497 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap. This is not supported.\n", current->comm, current->pid);
498 return ERR_PTR(-EINVAL);
501 if (flags & MREMAP_DONTUNMAP && (!vma_is_anonymous(vma) ||
502 vma->vm_flags & VM_SHARED))
503 return ERR_PTR(-EINVAL);
505 if (is_vm_hugetlb_page(vma))
506 return ERR_PTR(-EINVAL);
508 /* We can't remap across vm area boundaries */
509 if (old_len > vma->vm_end - addr)
510 return ERR_PTR(-EFAULT);
512 if (new_len == old_len)
515 /* Need to be careful about a growing mapping */
516 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
517 pgoff += vma->vm_pgoff;
518 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
519 return ERR_PTR(-EINVAL);
521 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
522 return ERR_PTR(-EFAULT);
524 if (vma->vm_flags & VM_LOCKED) {
525 unsigned long locked, lock_limit;
526 locked = mm->locked_vm << PAGE_SHIFT;
527 lock_limit = rlimit(RLIMIT_MEMLOCK);
528 locked += new_len - old_len;
529 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
530 return ERR_PTR(-EAGAIN);
533 if (!may_expand_vm(mm, vma->vm_flags,
534 (new_len - old_len) >> PAGE_SHIFT))
535 return ERR_PTR(-ENOMEM);
537 if (vma->vm_flags & VM_ACCOUNT) {
538 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
539 if (security_vm_enough_memory_mm(mm, charged))
540 return ERR_PTR(-ENOMEM);
547 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
548 unsigned long new_addr, unsigned long new_len, bool *locked,
549 unsigned long flags, struct vm_userfaultfd_ctx *uf,
550 struct list_head *uf_unmap_early,
551 struct list_head *uf_unmap)
553 struct mm_struct *mm = current->mm;
554 struct vm_area_struct *vma;
555 unsigned long ret = -EINVAL;
556 unsigned long charged = 0;
557 unsigned long map_flags = 0;
559 if (offset_in_page(new_addr))
562 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
565 /* Ensure the old/new locations do not overlap */
566 if (addr + old_len > new_addr && new_addr + new_len > addr)
570 * move_vma() need us to stay 4 maps below the threshold, otherwise
571 * it will bail out at the very beginning.
572 * That is a problem if we have already unmaped the regions here
573 * (new_addr, and old_addr), because userspace will not know the
574 * state of the vma's after it gets -ENOMEM.
575 * So, to avoid such scenario we can pre-compute if the whole
576 * operation has high chances to success map-wise.
577 * Worst-scenario case is when both vma's (new_addr and old_addr) get
578 * split in 3 before unmaping it.
579 * That means 2 more maps (1 for each) to the ones we already hold.
580 * Check whether current map count plus 2 still leads us to 4 maps below
581 * the threshold, otherwise return -ENOMEM here to be more safe.
583 if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
586 if (flags & MREMAP_FIXED) {
587 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
592 if (old_len >= new_len) {
593 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
594 if (ret && old_len != new_len)
599 vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
605 /* MREMAP_DONTUNMAP expands by old_len since old_len == new_len */
606 if (flags & MREMAP_DONTUNMAP &&
607 !may_expand_vm(mm, vma->vm_flags, old_len >> PAGE_SHIFT)) {
612 if (flags & MREMAP_FIXED)
613 map_flags |= MAP_FIXED;
615 if (vma->vm_flags & VM_MAYSHARE)
616 map_flags |= MAP_SHARED;
618 ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
619 ((addr - vma->vm_start) >> PAGE_SHIFT),
621 if (IS_ERR_VALUE(ret))
624 /* We got a new mapping */
625 if (!(flags & MREMAP_FIXED))
628 ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, flags, uf,
631 if (!(offset_in_page(ret)))
635 vm_unacct_memory(charged);
641 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
643 unsigned long end = vma->vm_end + delta;
644 if (end < vma->vm_end) /* overflow */
646 if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
648 if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
649 0, MAP_FIXED) & ~PAGE_MASK)
655 * Expand (or shrink) an existing mapping, potentially moving it at the
656 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
658 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
659 * This option implies MREMAP_MAYMOVE.
661 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
662 unsigned long, new_len, unsigned long, flags,
663 unsigned long, new_addr)
665 struct mm_struct *mm = current->mm;
666 struct vm_area_struct *vma;
667 unsigned long ret = -EINVAL;
668 unsigned long charged = 0;
670 bool downgraded = false;
671 struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
672 LIST_HEAD(uf_unmap_early);
676 * There is a deliberate asymmetry here: we strip the pointer tag
677 * from the old address but leave the new address alone. This is
678 * for consistency with mmap(), where we prevent the creation of
679 * aliasing mappings in userspace by leaving the tag bits of the
680 * mapping address intact. A non-zero tag will cause the subsequent
681 * range checks to reject the address as invalid.
683 * See Documentation/arm64/tagged-address-abi.rst for more information.
685 addr = untagged_addr(addr);
687 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP))
690 if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
694 * MREMAP_DONTUNMAP is always a move and it does not allow resizing
697 if (flags & MREMAP_DONTUNMAP &&
698 (!(flags & MREMAP_MAYMOVE) || old_len != new_len))
702 if (offset_in_page(addr))
705 old_len = PAGE_ALIGN(old_len);
706 new_len = PAGE_ALIGN(new_len);
709 * We allow a zero old-len as a special case
710 * for DOS-emu "duplicate shm area" thing. But
711 * a zero new-len is nonsensical.
716 if (mmap_write_lock_killable(current->mm))
719 if (flags & (MREMAP_FIXED | MREMAP_DONTUNMAP)) {
720 ret = mremap_to(addr, old_len, new_addr, new_len,
721 &locked, flags, &uf, &uf_unmap_early,
727 * Always allow a shrinking remap: that just unmaps
728 * the unnecessary pages..
729 * __do_munmap does all the needed commit accounting, and
730 * downgrades mmap_lock to read if so directed.
732 if (old_len >= new_len) {
735 retval = __do_munmap(mm, addr+new_len, old_len - new_len,
737 if (retval < 0 && old_len != new_len) {
740 /* Returning 1 indicates mmap_lock is downgraded to read. */
741 } else if (retval == 1)
748 * Ok, we need to grow..
750 vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
756 /* old_len exactly to the end of the area..
758 if (old_len == vma->vm_end - addr) {
759 /* can we just expand the current mapping? */
760 if (vma_expandable(vma, new_len - old_len)) {
761 int pages = (new_len - old_len) >> PAGE_SHIFT;
763 if (vma_adjust(vma, vma->vm_start, addr + new_len,
764 vma->vm_pgoff, NULL)) {
769 vm_stat_account(mm, vma->vm_flags, pages);
770 if (vma->vm_flags & VM_LOCKED) {
771 mm->locked_vm += pages;
781 * We weren't able to just expand or shrink the area,
782 * we need to create a new one and move it..
785 if (flags & MREMAP_MAYMOVE) {
786 unsigned long map_flags = 0;
787 if (vma->vm_flags & VM_MAYSHARE)
788 map_flags |= MAP_SHARED;
790 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
792 ((addr - vma->vm_start) >> PAGE_SHIFT),
794 if (IS_ERR_VALUE(new_addr)) {
799 ret = move_vma(vma, addr, old_len, new_len, new_addr,
800 &locked, flags, &uf, &uf_unmap);
803 if (offset_in_page(ret)) {
804 vm_unacct_memory(charged);
808 mmap_read_unlock(current->mm);
810 mmap_write_unlock(current->mm);
811 if (locked && new_len > old_len)
812 mm_populate(new_addr + old_len, new_len - old_len);
813 userfaultfd_unmap_complete(mm, &uf_unmap_early);
814 mremap_userfaultfd_complete(&uf, addr, ret, old_len);
815 userfaultfd_unmap_complete(mm, &uf_unmap);