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;
249 old_end = old_addr + len;
250 flush_cache_range(vma, old_addr, old_end);
252 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
254 mmu_notifier_invalidate_range_start(&range);
256 for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
258 next = (old_addr + PMD_SIZE) & PMD_MASK;
259 /* even if next overflowed, extent below will be ok */
260 extent = next - old_addr;
261 if (extent > old_end - old_addr)
262 extent = old_end - old_addr;
263 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
266 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
269 if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd)) {
270 if (extent == HPAGE_PMD_SIZE) {
272 /* See comment in move_ptes() */
274 take_rmap_locks(vma);
275 moved = move_huge_pmd(vma, old_addr, new_addr,
276 old_end, old_pmd, new_pmd);
278 drop_rmap_locks(vma);
282 split_huge_pmd(vma, old_pmd, old_addr);
283 if (pmd_trans_unstable(old_pmd))
285 } else if (extent == PMD_SIZE) {
286 #ifdef CONFIG_HAVE_MOVE_PMD
288 * If the extent is PMD-sized, try to speed the move by
289 * moving at the PMD level if possible.
294 take_rmap_locks(vma);
295 moved = move_normal_pmd(vma, old_addr, new_addr,
296 old_end, old_pmd, new_pmd);
298 drop_rmap_locks(vma);
304 if (pte_alloc(new_vma->vm_mm, new_pmd))
306 next = (new_addr + PMD_SIZE) & PMD_MASK;
307 if (extent > next - new_addr)
308 extent = next - new_addr;
309 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
310 new_pmd, new_addr, need_rmap_locks);
313 mmu_notifier_invalidate_range_end(&range);
315 return len + old_addr - old_end; /* how much done */
318 static unsigned long move_vma(struct vm_area_struct *vma,
319 unsigned long old_addr, unsigned long old_len,
320 unsigned long new_len, unsigned long new_addr,
321 bool *locked, struct vm_userfaultfd_ctx *uf,
322 struct list_head *uf_unmap)
324 struct mm_struct *mm = vma->vm_mm;
325 struct vm_area_struct *new_vma;
326 unsigned long vm_flags = vma->vm_flags;
327 unsigned long new_pgoff;
328 unsigned long moved_len;
329 unsigned long excess = 0;
330 unsigned long hiwater_vm;
333 bool need_rmap_locks;
336 * We'd prefer to avoid failure later on in do_munmap:
337 * which may split one vma into three before unmapping.
339 if (mm->map_count >= sysctl_max_map_count - 3)
343 * Advise KSM to break any KSM pages in the area to be moved:
344 * it would be confusing if they were to turn up at the new
345 * location, where they happen to coincide with different KSM
346 * pages recently unmapped. But leave vma->vm_flags as it was,
347 * so KSM can come around to merge on vma and new_vma afterwards.
349 err = ksm_madvise(vma, old_addr, old_addr + old_len,
350 MADV_UNMERGEABLE, &vm_flags);
354 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
355 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
360 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
362 if (moved_len < old_len) {
364 } else if (vma->vm_ops && vma->vm_ops->mremap) {
365 err = vma->vm_ops->mremap(new_vma);
370 * On error, move entries back from new area to old,
371 * which will succeed since page tables still there,
372 * and then proceed to unmap new area instead of old.
374 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
381 mremap_userfaultfd_prep(new_vma, uf);
382 arch_remap(mm, old_addr, old_addr + old_len,
383 new_addr, new_addr + new_len);
386 /* Conceal VM_ACCOUNT so old reservation is not undone */
387 if (vm_flags & VM_ACCOUNT) {
388 vma->vm_flags &= ~VM_ACCOUNT;
389 excess = vma->vm_end - vma->vm_start - old_len;
390 if (old_addr > vma->vm_start &&
391 old_addr + old_len < vma->vm_end)
396 * If we failed to move page tables we still do total_vm increment
397 * since do_munmap() will decrement it by old_len == new_len.
399 * Since total_vm is about to be raised artificially high for a
400 * moment, we need to restore high watermark afterwards: if stats
401 * are taken meanwhile, total_vm and hiwater_vm appear too high.
402 * If this were a serious issue, we'd add a flag to do_munmap().
404 hiwater_vm = mm->hiwater_vm;
405 vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
407 /* Tell pfnmap has moved from this vma */
408 if (unlikely(vma->vm_flags & VM_PFNMAP))
409 untrack_pfn_moved(vma);
411 if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
412 /* OOM: unable to split vma, just get accounts right */
413 vm_unacct_memory(excess >> PAGE_SHIFT);
416 mm->hiwater_vm = hiwater_vm;
418 /* Restore VM_ACCOUNT if one or two pieces of vma left */
420 vma->vm_flags |= VM_ACCOUNT;
422 vma->vm_next->vm_flags |= VM_ACCOUNT;
425 if (vm_flags & VM_LOCKED) {
426 mm->locked_vm += new_len >> PAGE_SHIFT;
433 static struct vm_area_struct *vma_to_resize(unsigned long addr,
434 unsigned long old_len, unsigned long new_len, unsigned long *p)
436 struct mm_struct *mm = current->mm;
437 struct vm_area_struct *vma = find_vma(mm, addr);
440 if (!vma || vma->vm_start > addr)
441 return ERR_PTR(-EFAULT);
444 * !old_len is a special case where an attempt is made to 'duplicate'
445 * a mapping. This makes no sense for private mappings as it will
446 * instead create a fresh/new mapping unrelated to the original. This
447 * is contrary to the basic idea of mremap which creates new mappings
448 * based on the original. There are no known use cases for this
449 * behavior. As a result, fail such attempts.
451 if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
452 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap. This is not supported.\n", current->comm, current->pid);
453 return ERR_PTR(-EINVAL);
456 if (is_vm_hugetlb_page(vma))
457 return ERR_PTR(-EINVAL);
459 /* We can't remap across vm area boundaries */
460 if (old_len > vma->vm_end - addr)
461 return ERR_PTR(-EFAULT);
463 if (new_len == old_len)
466 /* Need to be careful about a growing mapping */
467 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
468 pgoff += vma->vm_pgoff;
469 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
470 return ERR_PTR(-EINVAL);
472 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
473 return ERR_PTR(-EFAULT);
475 if (vma->vm_flags & VM_LOCKED) {
476 unsigned long locked, lock_limit;
477 locked = mm->locked_vm << PAGE_SHIFT;
478 lock_limit = rlimit(RLIMIT_MEMLOCK);
479 locked += new_len - old_len;
480 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
481 return ERR_PTR(-EAGAIN);
484 if (!may_expand_vm(mm, vma->vm_flags,
485 (new_len - old_len) >> PAGE_SHIFT))
486 return ERR_PTR(-ENOMEM);
488 if (vma->vm_flags & VM_ACCOUNT) {
489 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
490 if (security_vm_enough_memory_mm(mm, charged))
491 return ERR_PTR(-ENOMEM);
498 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
499 unsigned long new_addr, unsigned long new_len, bool *locked,
500 struct vm_userfaultfd_ctx *uf,
501 struct list_head *uf_unmap_early,
502 struct list_head *uf_unmap)
504 struct mm_struct *mm = current->mm;
505 struct vm_area_struct *vma;
506 unsigned long ret = -EINVAL;
507 unsigned long charged = 0;
508 unsigned long map_flags;
510 if (offset_in_page(new_addr))
513 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
516 /* Ensure the old/new locations do not overlap */
517 if (addr + old_len > new_addr && new_addr + new_len > addr)
521 * move_vma() need us to stay 4 maps below the threshold, otherwise
522 * it will bail out at the very beginning.
523 * That is a problem if we have already unmaped the regions here
524 * (new_addr, and old_addr), because userspace will not know the
525 * state of the vma's after it gets -ENOMEM.
526 * So, to avoid such scenario we can pre-compute if the whole
527 * operation has high chances to success map-wise.
528 * Worst-scenario case is when both vma's (new_addr and old_addr) get
529 * split in 3 before unmaping it.
530 * That means 2 more maps (1 for each) to the ones we already hold.
531 * Check whether current map count plus 2 still leads us to 4 maps below
532 * the threshold, otherwise return -ENOMEM here to be more safe.
534 if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
537 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
541 if (old_len >= new_len) {
542 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
543 if (ret && old_len != new_len)
548 vma = vma_to_resize(addr, old_len, new_len, &charged);
554 map_flags = MAP_FIXED;
555 if (vma->vm_flags & VM_MAYSHARE)
556 map_flags |= MAP_SHARED;
558 ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
559 ((addr - vma->vm_start) >> PAGE_SHIFT),
561 if (offset_in_page(ret))
564 ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, uf,
566 if (!(offset_in_page(ret)))
569 vm_unacct_memory(charged);
575 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
577 unsigned long end = vma->vm_end + delta;
578 if (end < vma->vm_end) /* overflow */
580 if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
582 if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
583 0, MAP_FIXED) & ~PAGE_MASK)
589 * Expand (or shrink) an existing mapping, potentially moving it at the
590 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
592 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
593 * This option implies MREMAP_MAYMOVE.
595 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
596 unsigned long, new_len, unsigned long, flags,
597 unsigned long, new_addr)
599 struct mm_struct *mm = current->mm;
600 struct vm_area_struct *vma;
601 unsigned long ret = -EINVAL;
602 unsigned long charged = 0;
604 bool downgraded = false;
605 struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
606 LIST_HEAD(uf_unmap_early);
609 addr = untagged_addr(addr);
610 new_addr = untagged_addr(new_addr);
612 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
615 if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
618 if (offset_in_page(addr))
621 old_len = PAGE_ALIGN(old_len);
622 new_len = PAGE_ALIGN(new_len);
625 * We allow a zero old-len as a special case
626 * for DOS-emu "duplicate shm area" thing. But
627 * a zero new-len is nonsensical.
632 if (down_write_killable(¤t->mm->mmap_sem))
635 if (flags & MREMAP_FIXED) {
636 ret = mremap_to(addr, old_len, new_addr, new_len,
637 &locked, &uf, &uf_unmap_early, &uf_unmap);
642 * Always allow a shrinking remap: that just unmaps
643 * the unnecessary pages..
644 * __do_munmap does all the needed commit accounting, and
645 * downgrades mmap_sem to read if so directed.
647 if (old_len >= new_len) {
650 retval = __do_munmap(mm, addr+new_len, old_len - new_len,
652 if (retval < 0 && old_len != new_len) {
655 /* Returning 1 indicates mmap_sem is downgraded to read. */
656 } else if (retval == 1)
663 * Ok, we need to grow..
665 vma = vma_to_resize(addr, old_len, new_len, &charged);
671 /* old_len exactly to the end of the area..
673 if (old_len == vma->vm_end - addr) {
674 /* can we just expand the current mapping? */
675 if (vma_expandable(vma, new_len - old_len)) {
676 int pages = (new_len - old_len) >> PAGE_SHIFT;
678 if (vma_adjust(vma, vma->vm_start, addr + new_len,
679 vma->vm_pgoff, NULL)) {
684 vm_stat_account(mm, vma->vm_flags, pages);
685 if (vma->vm_flags & VM_LOCKED) {
686 mm->locked_vm += pages;
696 * We weren't able to just expand or shrink the area,
697 * we need to create a new one and move it..
700 if (flags & MREMAP_MAYMOVE) {
701 unsigned long map_flags = 0;
702 if (vma->vm_flags & VM_MAYSHARE)
703 map_flags |= MAP_SHARED;
705 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
707 ((addr - vma->vm_start) >> PAGE_SHIFT),
709 if (offset_in_page(new_addr)) {
714 ret = move_vma(vma, addr, old_len, new_len, new_addr,
715 &locked, &uf, &uf_unmap);
718 if (offset_in_page(ret)) {
719 vm_unacct_memory(charged);
723 up_read(¤t->mm->mmap_sem);
725 up_write(¤t->mm->mmap_sem);
726 if (locked && new_len > old_len)
727 mm_populate(new_addr + old_len, new_len - old_len);
728 userfaultfd_unmap_complete(mm, &uf_unmap_early);
729 mremap_userfaultfd_complete(&uf, addr, new_addr, old_len);
730 userfaultfd_unmap_complete(mm, &uf_unmap);