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 is_vma_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_sem 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, unsigned long old_end,
197 pmd_t *old_pmd, pmd_t *new_pmd)
199 spinlock_t *old_ptl, *new_ptl;
200 struct mm_struct *mm = vma->vm_mm;
203 if ((old_addr & ~PMD_MASK) || (new_addr & ~PMD_MASK)
204 || old_end - old_addr < PMD_SIZE)
208 * The destination pmd shouldn't be established, free_pgtables()
209 * should have release it.
211 if (WARN_ON(!pmd_none(*new_pmd)))
215 * We don't have to worry about the ordering of src and dst
216 * ptlocks because exclusive mmap_sem prevents deadlock.
218 old_ptl = pmd_lock(vma->vm_mm, old_pmd);
219 new_ptl = pmd_lockptr(mm, new_pmd);
220 if (new_ptl != old_ptl)
221 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
227 VM_BUG_ON(!pmd_none(*new_pmd));
229 /* Set the new pmd */
230 set_pmd_at(mm, new_addr, new_pmd, pmd);
231 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
232 if (new_ptl != old_ptl)
233 spin_unlock(new_ptl);
234 spin_unlock(old_ptl);
240 unsigned long move_page_tables(struct vm_area_struct *vma,
241 unsigned long old_addr, struct vm_area_struct *new_vma,
242 unsigned long new_addr, unsigned long len,
243 bool need_rmap_locks)
245 unsigned long extent, next, old_end;
246 struct mmu_notifier_range range;
247 pmd_t *old_pmd, *new_pmd;
252 old_end = old_addr + len;
253 flush_cache_range(vma, old_addr, old_end);
255 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
257 mmu_notifier_invalidate_range_start(&range);
259 for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
261 next = (old_addr + PMD_SIZE) & PMD_MASK;
262 /* even if next overflowed, extent below will be ok */
263 extent = next - old_addr;
264 if (extent > old_end - old_addr)
265 extent = old_end - old_addr;
266 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
269 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
272 if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) || pmd_devmap(*old_pmd)) {
273 if (extent == HPAGE_PMD_SIZE) {
275 /* See comment in move_ptes() */
277 take_rmap_locks(vma);
278 moved = move_huge_pmd(vma, old_addr, new_addr,
279 old_end, old_pmd, new_pmd);
281 drop_rmap_locks(vma);
285 split_huge_pmd(vma, old_pmd, old_addr);
286 if (pmd_trans_unstable(old_pmd))
288 } else if (extent == PMD_SIZE) {
289 #ifdef CONFIG_HAVE_MOVE_PMD
291 * If the extent is PMD-sized, try to speed the move by
292 * moving at the PMD level if possible.
296 take_rmap_locks(vma);
297 moved = move_normal_pmd(vma, old_addr, new_addr,
298 old_end, old_pmd, new_pmd);
299 drop_rmap_locks(vma);
305 if (pte_alloc(new_vma->vm_mm, new_pmd))
307 next = (new_addr + PMD_SIZE) & PMD_MASK;
308 if (extent > next - new_addr)
309 extent = next - new_addr;
310 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
311 new_pmd, new_addr, need_rmap_locks);
314 mmu_notifier_invalidate_range_end(&range);
316 return len + old_addr - old_end; /* how much done */
319 static unsigned long move_vma(struct vm_area_struct *vma,
320 unsigned long old_addr, unsigned long old_len,
321 unsigned long new_len, unsigned long new_addr,
322 bool *locked, struct vm_userfaultfd_ctx *uf,
323 struct list_head *uf_unmap)
325 struct mm_struct *mm = vma->vm_mm;
326 struct vm_area_struct *new_vma;
327 unsigned long vm_flags = vma->vm_flags;
328 unsigned long new_pgoff;
329 unsigned long moved_len;
330 unsigned long excess = 0;
331 unsigned long hiwater_vm;
334 bool need_rmap_locks;
337 * We'd prefer to avoid failure later on in do_munmap:
338 * which may split one vma into three before unmapping.
340 if (mm->map_count >= sysctl_max_map_count - 3)
344 * Advise KSM to break any KSM pages in the area to be moved:
345 * it would be confusing if they were to turn up at the new
346 * location, where they happen to coincide with different KSM
347 * pages recently unmapped. But leave vma->vm_flags as it was,
348 * so KSM can come around to merge on vma and new_vma afterwards.
350 err = ksm_madvise(vma, old_addr, old_addr + old_len,
351 MADV_UNMERGEABLE, &vm_flags);
355 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
356 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
361 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
363 if (moved_len < old_len) {
365 } else if (vma->vm_ops && vma->vm_ops->mremap) {
366 err = vma->vm_ops->mremap(new_vma);
371 * On error, move entries back from new area to old,
372 * which will succeed since page tables still there,
373 * and then proceed to unmap new area instead of old.
375 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
382 mremap_userfaultfd_prep(new_vma, uf);
383 arch_remap(mm, old_addr, old_addr + old_len,
384 new_addr, new_addr + new_len);
387 /* Conceal VM_ACCOUNT so old reservation is not undone */
388 if (vm_flags & VM_ACCOUNT) {
389 vma->vm_flags &= ~VM_ACCOUNT;
390 excess = vma->vm_end - vma->vm_start - old_len;
391 if (old_addr > vma->vm_start &&
392 old_addr + old_len < vma->vm_end)
397 * If we failed to move page tables we still do total_vm increment
398 * since do_munmap() will decrement it by old_len == new_len.
400 * Since total_vm is about to be raised artificially high for a
401 * moment, we need to restore high watermark afterwards: if stats
402 * are taken meanwhile, total_vm and hiwater_vm appear too high.
403 * If this were a serious issue, we'd add a flag to do_munmap().
405 hiwater_vm = mm->hiwater_vm;
406 vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
408 /* Tell pfnmap has moved from this vma */
409 if (unlikely(vma->vm_flags & VM_PFNMAP))
410 untrack_pfn_moved(vma);
412 if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
413 /* OOM: unable to split vma, just get accounts right */
414 vm_unacct_memory(excess >> PAGE_SHIFT);
417 mm->hiwater_vm = hiwater_vm;
419 /* Restore VM_ACCOUNT if one or two pieces of vma left */
421 vma->vm_flags |= VM_ACCOUNT;
423 vma->vm_next->vm_flags |= VM_ACCOUNT;
426 if (vm_flags & VM_LOCKED) {
427 mm->locked_vm += new_len >> PAGE_SHIFT;
434 static struct vm_area_struct *vma_to_resize(unsigned long addr,
435 unsigned long old_len, unsigned long new_len, unsigned long *p)
437 struct mm_struct *mm = current->mm;
438 struct vm_area_struct *vma = find_vma(mm, addr);
441 if (!vma || vma->vm_start > addr)
442 return ERR_PTR(-EFAULT);
445 * !old_len is a special case where an attempt is made to 'duplicate'
446 * a mapping. This makes no sense for private mappings as it will
447 * instead create a fresh/new mapping unrelated to the original. This
448 * is contrary to the basic idea of mremap which creates new mappings
449 * based on the original. There are no known use cases for this
450 * behavior. As a result, fail such attempts.
452 if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
453 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap. This is not supported.\n", current->comm, current->pid);
454 return ERR_PTR(-EINVAL);
457 if (is_vm_hugetlb_page(vma))
458 return ERR_PTR(-EINVAL);
460 /* We can't remap across vm area boundaries */
461 if (old_len > vma->vm_end - addr)
462 return ERR_PTR(-EFAULT);
464 if (new_len == old_len)
467 /* Need to be careful about a growing mapping */
468 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
469 pgoff += vma->vm_pgoff;
470 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
471 return ERR_PTR(-EINVAL);
473 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
474 return ERR_PTR(-EFAULT);
476 if (vma->vm_flags & VM_LOCKED) {
477 unsigned long locked, lock_limit;
478 locked = mm->locked_vm << PAGE_SHIFT;
479 lock_limit = rlimit(RLIMIT_MEMLOCK);
480 locked += new_len - old_len;
481 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
482 return ERR_PTR(-EAGAIN);
485 if (!may_expand_vm(mm, vma->vm_flags,
486 (new_len - old_len) >> PAGE_SHIFT))
487 return ERR_PTR(-ENOMEM);
489 if (vma->vm_flags & VM_ACCOUNT) {
490 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
491 if (security_vm_enough_memory_mm(mm, charged))
492 return ERR_PTR(-ENOMEM);
499 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
500 unsigned long new_addr, unsigned long new_len, bool *locked,
501 struct vm_userfaultfd_ctx *uf,
502 struct list_head *uf_unmap_early,
503 struct list_head *uf_unmap)
505 struct mm_struct *mm = current->mm;
506 struct vm_area_struct *vma;
507 unsigned long ret = -EINVAL;
508 unsigned long charged = 0;
509 unsigned long map_flags;
511 if (offset_in_page(new_addr))
514 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
517 /* Ensure the old/new locations do not overlap */
518 if (addr + old_len > new_addr && new_addr + new_len > addr)
522 * move_vma() need us to stay 4 maps below the threshold, otherwise
523 * it will bail out at the very beginning.
524 * That is a problem if we have already unmaped the regions here
525 * (new_addr, and old_addr), because userspace will not know the
526 * state of the vma's after it gets -ENOMEM.
527 * So, to avoid such scenario we can pre-compute if the whole
528 * operation has high chances to success map-wise.
529 * Worst-scenario case is when both vma's (new_addr and old_addr) get
530 * split in 3 before unmaping it.
531 * That means 2 more maps (1 for each) to the ones we already hold.
532 * Check whether current map count plus 2 still leads us to 4 maps below
533 * the threshold, otherwise return -ENOMEM here to be more safe.
535 if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
538 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
542 if (old_len >= new_len) {
543 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
544 if (ret && old_len != new_len)
549 vma = vma_to_resize(addr, old_len, new_len, &charged);
555 map_flags = MAP_FIXED;
556 if (vma->vm_flags & VM_MAYSHARE)
557 map_flags |= MAP_SHARED;
559 ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
560 ((addr - vma->vm_start) >> PAGE_SHIFT),
562 if (offset_in_page(ret))
565 ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, uf,
567 if (!(offset_in_page(ret)))
570 vm_unacct_memory(charged);
576 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
578 unsigned long end = vma->vm_end + delta;
579 if (end < vma->vm_end) /* overflow */
581 if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
583 if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
584 0, MAP_FIXED) & ~PAGE_MASK)
590 * Expand (or shrink) an existing mapping, potentially moving it at the
591 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
593 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
594 * This option implies MREMAP_MAYMOVE.
596 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
597 unsigned long, new_len, unsigned long, flags,
598 unsigned long, new_addr)
600 struct mm_struct *mm = current->mm;
601 struct vm_area_struct *vma;
602 unsigned long ret = -EINVAL;
603 unsigned long charged = 0;
605 bool downgraded = false;
606 struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
607 LIST_HEAD(uf_unmap_early);
611 * There is a deliberate asymmetry here: we strip the pointer tag
612 * from the old address but leave the new address alone. This is
613 * for consistency with mmap(), where we prevent the creation of
614 * aliasing mappings in userspace by leaving the tag bits of the
615 * mapping address intact. A non-zero tag will cause the subsequent
616 * range checks to reject the address as invalid.
618 * See Documentation/arm64/tagged-address-abi.rst for more information.
620 addr = untagged_addr(addr);
622 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
625 if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
628 if (offset_in_page(addr))
631 old_len = PAGE_ALIGN(old_len);
632 new_len = PAGE_ALIGN(new_len);
635 * We allow a zero old-len as a special case
636 * for DOS-emu "duplicate shm area" thing. But
637 * a zero new-len is nonsensical.
642 if (down_write_killable(¤t->mm->mmap_sem))
645 if (flags & MREMAP_FIXED) {
646 ret = mremap_to(addr, old_len, new_addr, new_len,
647 &locked, &uf, &uf_unmap_early, &uf_unmap);
652 * Always allow a shrinking remap: that just unmaps
653 * the unnecessary pages..
654 * __do_munmap does all the needed commit accounting, and
655 * downgrades mmap_sem to read if so directed.
657 if (old_len >= new_len) {
660 retval = __do_munmap(mm, addr+new_len, old_len - new_len,
662 if (retval < 0 && old_len != new_len) {
665 /* Returning 1 indicates mmap_sem is downgraded to read. */
666 } else if (retval == 1)
673 * Ok, we need to grow..
675 vma = vma_to_resize(addr, old_len, new_len, &charged);
681 /* old_len exactly to the end of the area..
683 if (old_len == vma->vm_end - addr) {
684 /* can we just expand the current mapping? */
685 if (vma_expandable(vma, new_len - old_len)) {
686 int pages = (new_len - old_len) >> PAGE_SHIFT;
688 if (vma_adjust(vma, vma->vm_start, addr + new_len,
689 vma->vm_pgoff, NULL)) {
694 vm_stat_account(mm, vma->vm_flags, pages);
695 if (vma->vm_flags & VM_LOCKED) {
696 mm->locked_vm += pages;
706 * We weren't able to just expand or shrink the area,
707 * we need to create a new one and move it..
710 if (flags & MREMAP_MAYMOVE) {
711 unsigned long map_flags = 0;
712 if (vma->vm_flags & VM_MAYSHARE)
713 map_flags |= MAP_SHARED;
715 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
717 ((addr - vma->vm_start) >> PAGE_SHIFT),
719 if (offset_in_page(new_addr)) {
724 ret = move_vma(vma, addr, old_len, new_len, new_addr,
725 &locked, &uf, &uf_unmap);
728 if (offset_in_page(ret)) {
729 vm_unacct_memory(charged);
733 up_read(¤t->mm->mmap_sem);
735 up_write(¤t->mm->mmap_sem);
736 if (locked && new_len > old_len)
737 mm_populate(new_addr + old_len, new_len - old_len);
738 userfaultfd_unmap_complete(mm, &uf_unmap_early);
739 mremap_userfaultfd_complete(&uf, addr, new_addr, old_len);
740 userfaultfd_unmap_complete(mm, &uf_unmap);