Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * mm/mremap.c
4 *
5 * (C) Copyright 1996 Linus Torvalds
6 *
7 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
8 * (C) Copyright 2002 Red Hat Inc, All Rights Reserved
9 */
10
11#include <linux/mm.h>
12#include <linux/mm_inline.h>
13#include <linux/hugetlb.h>
14#include <linux/shm.h>
15#include <linux/ksm.h>
16#include <linux/mman.h>
17#include <linux/swap.h>
18#include <linux/capability.h>
19#include <linux/fs.h>
20#include <linux/swapops.h>
21#include <linux/highmem.h>
22#include <linux/security.h>
23#include <linux/syscalls.h>
24#include <linux/mmu_notifier.h>
25#include <linux/uaccess.h>
26#include <linux/userfaultfd_k.h>
27#include <linux/mempolicy.h>
28
29#include <asm/cacheflush.h>
30#include <asm/tlb.h>
31#include <asm/pgalloc.h>
32
33#include "internal.h"
34
35static pud_t *get_old_pud(struct mm_struct *mm, unsigned long addr)
36{
37 pgd_t *pgd;
38 p4d_t *p4d;
39 pud_t *pud;
40
41 pgd = pgd_offset(mm, addr);
42 if (pgd_none_or_clear_bad(pgd))
43 return NULL;
44
45 p4d = p4d_offset(pgd, addr);
46 if (p4d_none_or_clear_bad(p4d))
47 return NULL;
48
49 pud = pud_offset(p4d, addr);
50 if (pud_none_or_clear_bad(pud))
51 return NULL;
52
53 return pud;
54}
55
56static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
57{
58 pud_t *pud;
59 pmd_t *pmd;
60
61 pud = get_old_pud(mm, addr);
62 if (!pud)
63 return NULL;
64
65 pmd = pmd_offset(pud, addr);
66 if (pmd_none(*pmd))
67 return NULL;
68
69 return pmd;
70}
71
72static pud_t *alloc_new_pud(struct mm_struct *mm, struct vm_area_struct *vma,
73 unsigned long addr)
74{
75 pgd_t *pgd;
76 p4d_t *p4d;
77
78 pgd = pgd_offset(mm, addr);
79 p4d = p4d_alloc(mm, pgd, addr);
80 if (!p4d)
81 return NULL;
82
83 return pud_alloc(mm, p4d, addr);
84}
85
86static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
87 unsigned long addr)
88{
89 pud_t *pud;
90 pmd_t *pmd;
91
92 pud = alloc_new_pud(mm, vma, addr);
93 if (!pud)
94 return NULL;
95
96 pmd = pmd_alloc(mm, pud, addr);
97 if (!pmd)
98 return NULL;
99
100 VM_BUG_ON(pmd_trans_huge(*pmd));
101
102 return pmd;
103}
104
105static void take_rmap_locks(struct vm_area_struct *vma)
106{
107 if (vma->vm_file)
108 i_mmap_lock_write(vma->vm_file->f_mapping);
109 if (vma->anon_vma)
110 anon_vma_lock_write(vma->anon_vma);
111}
112
113static void drop_rmap_locks(struct vm_area_struct *vma)
114{
115 if (vma->anon_vma)
116 anon_vma_unlock_write(vma->anon_vma);
117 if (vma->vm_file)
118 i_mmap_unlock_write(vma->vm_file->f_mapping);
119}
120
121static pte_t move_soft_dirty_pte(pte_t pte)
122{
123 /*
124 * Set soft dirty bit so we can notice
125 * in userspace the ptes were moved.
126 */
127#ifdef CONFIG_MEM_SOFT_DIRTY
128 if (pte_present(pte))
129 pte = pte_mksoft_dirty(pte);
130 else if (is_swap_pte(pte))
131 pte = pte_swp_mksoft_dirty(pte);
132#endif
133 return pte;
134}
135
136static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
137 unsigned long old_addr, unsigned long old_end,
138 struct vm_area_struct *new_vma, pmd_t *new_pmd,
139 unsigned long new_addr, bool need_rmap_locks)
140{
141 struct mm_struct *mm = vma->vm_mm;
142 pte_t *old_pte, *new_pte, pte;
143 spinlock_t *old_ptl, *new_ptl;
144 bool force_flush = false;
145 unsigned long len = old_end - old_addr;
146
147 /*
148 * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
149 * locks to ensure that rmap will always observe either the old or the
150 * new ptes. This is the easiest way to avoid races with
151 * truncate_pagecache(), page migration, etc...
152 *
153 * When need_rmap_locks is false, we use other ways to avoid
154 * such races:
155 *
156 * - During exec() shift_arg_pages(), we use a specially tagged vma
157 * which rmap call sites look for using vma_is_temporary_stack().
158 *
159 * - During mremap(), new_vma is often known to be placed after vma
160 * in rmap traversal order. This ensures rmap will always observe
161 * either the old pte, or the new pte, or both (the page table locks
162 * serialize access to individual ptes, but only rmap traversal
163 * order guarantees that we won't miss both the old and new ptes).
164 */
165 if (need_rmap_locks)
166 take_rmap_locks(vma);
167
168 /*
169 * We don't have to worry about the ordering of src and dst
170 * pte locks because exclusive mmap_lock prevents deadlock.
171 */
172 old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
173 new_pte = pte_offset_map(new_pmd, new_addr);
174 new_ptl = pte_lockptr(mm, new_pmd);
175 if (new_ptl != old_ptl)
176 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
177 flush_tlb_batched_pending(vma->vm_mm);
178 arch_enter_lazy_mmu_mode();
179
180 for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
181 new_pte++, new_addr += PAGE_SIZE) {
182 if (pte_none(*old_pte))
183 continue;
184
185 pte = ptep_get_and_clear(mm, old_addr, old_pte);
186 /*
187 * If we are remapping a valid PTE, make sure
188 * to flush TLB before we drop the PTL for the
189 * PTE.
190 *
191 * NOTE! Both old and new PTL matter: the old one
192 * for racing with page_mkclean(), the new one to
193 * make sure the physical page stays valid until
194 * the TLB entry for the old mapping has been
195 * flushed.
196 */
197 if (pte_present(pte))
198 force_flush = true;
199 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
200 pte = move_soft_dirty_pte(pte);
201 set_pte_at(mm, new_addr, new_pte, pte);
202 }
203
204 arch_leave_lazy_mmu_mode();
205 if (force_flush)
206 flush_tlb_range(vma, old_end - len, old_end);
207 if (new_ptl != old_ptl)
208 spin_unlock(new_ptl);
209 pte_unmap(new_pte - 1);
210 pte_unmap_unlock(old_pte - 1, old_ptl);
211 if (need_rmap_locks)
212 drop_rmap_locks(vma);
213}
214
215#ifndef arch_supports_page_table_move
216#define arch_supports_page_table_move arch_supports_page_table_move
217static inline bool arch_supports_page_table_move(void)
218{
219 return IS_ENABLED(CONFIG_HAVE_MOVE_PMD) ||
220 IS_ENABLED(CONFIG_HAVE_MOVE_PUD);
221}
222#endif
223
224#ifdef CONFIG_HAVE_MOVE_PMD
225static bool move_normal_pmd(struct vm_area_struct *vma, unsigned long old_addr,
226 unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
227{
228 spinlock_t *old_ptl, *new_ptl;
229 struct mm_struct *mm = vma->vm_mm;
230 pmd_t pmd;
231
232 if (!arch_supports_page_table_move())
233 return false;
234 /*
235 * The destination pmd shouldn't be established, free_pgtables()
236 * should have released it.
237 *
238 * However, there's a case during execve() where we use mremap
239 * to move the initial stack, and in that case the target area
240 * may overlap the source area (always moving down).
241 *
242 * If everything is PMD-aligned, that works fine, as moving
243 * each pmd down will clear the source pmd. But if we first
244 * have a few 4kB-only pages that get moved down, and then
245 * hit the "now the rest is PMD-aligned, let's do everything
246 * one pmd at a time", we will still have the old (now empty
247 * of any 4kB pages, but still there) PMD in the page table
248 * tree.
249 *
250 * Warn on it once - because we really should try to figure
251 * out how to do this better - but then say "I won't move
252 * this pmd".
253 *
254 * One alternative might be to just unmap the target pmd at
255 * this point, and verify that it really is empty. We'll see.
256 */
257 if (WARN_ON_ONCE(!pmd_none(*new_pmd)))
258 return false;
259
260 /*
261 * We don't have to worry about the ordering of src and dst
262 * ptlocks because exclusive mmap_lock prevents deadlock.
263 */
264 old_ptl = pmd_lock(vma->vm_mm, old_pmd);
265 new_ptl = pmd_lockptr(mm, new_pmd);
266 if (new_ptl != old_ptl)
267 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
268
269 /* Clear the pmd */
270 pmd = *old_pmd;
271 pmd_clear(old_pmd);
272
273 VM_BUG_ON(!pmd_none(*new_pmd));
274
275 pmd_populate(mm, new_pmd, pmd_pgtable(pmd));
276 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
277 if (new_ptl != old_ptl)
278 spin_unlock(new_ptl);
279 spin_unlock(old_ptl);
280
281 return true;
282}
283#else
284static inline bool move_normal_pmd(struct vm_area_struct *vma,
285 unsigned long old_addr, unsigned long new_addr, pmd_t *old_pmd,
286 pmd_t *new_pmd)
287{
288 return false;
289}
290#endif
291
292#if CONFIG_PGTABLE_LEVELS > 2 && defined(CONFIG_HAVE_MOVE_PUD)
293static bool move_normal_pud(struct vm_area_struct *vma, unsigned long old_addr,
294 unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
295{
296 spinlock_t *old_ptl, *new_ptl;
297 struct mm_struct *mm = vma->vm_mm;
298 pud_t pud;
299
300 if (!arch_supports_page_table_move())
301 return false;
302 /*
303 * The destination pud shouldn't be established, free_pgtables()
304 * should have released it.
305 */
306 if (WARN_ON_ONCE(!pud_none(*new_pud)))
307 return false;
308
309 /*
310 * We don't have to worry about the ordering of src and dst
311 * ptlocks because exclusive mmap_lock prevents deadlock.
312 */
313 old_ptl = pud_lock(vma->vm_mm, old_pud);
314 new_ptl = pud_lockptr(mm, new_pud);
315 if (new_ptl != old_ptl)
316 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
317
318 /* Clear the pud */
319 pud = *old_pud;
320 pud_clear(old_pud);
321
322 VM_BUG_ON(!pud_none(*new_pud));
323
324 pud_populate(mm, new_pud, pud_pgtable(pud));
325 flush_tlb_range(vma, old_addr, old_addr + PUD_SIZE);
326 if (new_ptl != old_ptl)
327 spin_unlock(new_ptl);
328 spin_unlock(old_ptl);
329
330 return true;
331}
332#else
333static inline bool move_normal_pud(struct vm_area_struct *vma,
334 unsigned long old_addr, unsigned long new_addr, pud_t *old_pud,
335 pud_t *new_pud)
336{
337 return false;
338}
339#endif
340
341#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
342static bool move_huge_pud(struct vm_area_struct *vma, unsigned long old_addr,
343 unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
344{
345 spinlock_t *old_ptl, *new_ptl;
346 struct mm_struct *mm = vma->vm_mm;
347 pud_t pud;
348
349 /*
350 * The destination pud shouldn't be established, free_pgtables()
351 * should have released it.
352 */
353 if (WARN_ON_ONCE(!pud_none(*new_pud)))
354 return false;
355
356 /*
357 * We don't have to worry about the ordering of src and dst
358 * ptlocks because exclusive mmap_lock prevents deadlock.
359 */
360 old_ptl = pud_lock(vma->vm_mm, old_pud);
361 new_ptl = pud_lockptr(mm, new_pud);
362 if (new_ptl != old_ptl)
363 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
364
365 /* Clear the pud */
366 pud = *old_pud;
367 pud_clear(old_pud);
368
369 VM_BUG_ON(!pud_none(*new_pud));
370
371 /* Set the new pud */
372 /* mark soft_ditry when we add pud level soft dirty support */
373 set_pud_at(mm, new_addr, new_pud, pud);
374 flush_pud_tlb_range(vma, old_addr, old_addr + HPAGE_PUD_SIZE);
375 if (new_ptl != old_ptl)
376 spin_unlock(new_ptl);
377 spin_unlock(old_ptl);
378
379 return true;
380}
381#else
382static bool move_huge_pud(struct vm_area_struct *vma, unsigned long old_addr,
383 unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
384{
385 WARN_ON_ONCE(1);
386 return false;
387
388}
389#endif
390
391enum pgt_entry {
392 NORMAL_PMD,
393 HPAGE_PMD,
394 NORMAL_PUD,
395 HPAGE_PUD,
396};
397
398/*
399 * Returns an extent of the corresponding size for the pgt_entry specified if
400 * valid. Else returns a smaller extent bounded by the end of the source and
401 * destination pgt_entry.
402 */
403static __always_inline unsigned long get_extent(enum pgt_entry entry,
404 unsigned long old_addr, unsigned long old_end,
405 unsigned long new_addr)
406{
407 unsigned long next, extent, mask, size;
408
409 switch (entry) {
410 case HPAGE_PMD:
411 case NORMAL_PMD:
412 mask = PMD_MASK;
413 size = PMD_SIZE;
414 break;
415 case HPAGE_PUD:
416 case NORMAL_PUD:
417 mask = PUD_MASK;
418 size = PUD_SIZE;
419 break;
420 default:
421 BUILD_BUG();
422 break;
423 }
424
425 next = (old_addr + size) & mask;
426 /* even if next overflowed, extent below will be ok */
427 extent = next - old_addr;
428 if (extent > old_end - old_addr)
429 extent = old_end - old_addr;
430 next = (new_addr + size) & mask;
431 if (extent > next - new_addr)
432 extent = next - new_addr;
433 return extent;
434}
435
436/*
437 * Attempts to speedup the move by moving entry at the level corresponding to
438 * pgt_entry. Returns true if the move was successful, else false.
439 */
440static bool move_pgt_entry(enum pgt_entry entry, struct vm_area_struct *vma,
441 unsigned long old_addr, unsigned long new_addr,
442 void *old_entry, void *new_entry, bool need_rmap_locks)
443{
444 bool moved = false;
445
446 /* See comment in move_ptes() */
447 if (need_rmap_locks)
448 take_rmap_locks(vma);
449
450 switch (entry) {
451 case NORMAL_PMD:
452 moved = move_normal_pmd(vma, old_addr, new_addr, old_entry,
453 new_entry);
454 break;
455 case NORMAL_PUD:
456 moved = move_normal_pud(vma, old_addr, new_addr, old_entry,
457 new_entry);
458 break;
459 case HPAGE_PMD:
460 moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
461 move_huge_pmd(vma, old_addr, new_addr, old_entry,
462 new_entry);
463 break;
464 case HPAGE_PUD:
465 moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
466 move_huge_pud(vma, old_addr, new_addr, old_entry,
467 new_entry);
468 break;
469
470 default:
471 WARN_ON_ONCE(1);
472 break;
473 }
474
475 if (need_rmap_locks)
476 drop_rmap_locks(vma);
477
478 return moved;
479}
480
481unsigned long move_page_tables(struct vm_area_struct *vma,
482 unsigned long old_addr, struct vm_area_struct *new_vma,
483 unsigned long new_addr, unsigned long len,
484 bool need_rmap_locks)
485{
486 unsigned long extent, old_end;
487 struct mmu_notifier_range range;
488 pmd_t *old_pmd, *new_pmd;
489 pud_t *old_pud, *new_pud;
490
491 if (!len)
492 return 0;
493
494 old_end = old_addr + len;
495
496 if (is_vm_hugetlb_page(vma))
497 return move_hugetlb_page_tables(vma, new_vma, old_addr,
498 new_addr, len);
499
500 flush_cache_range(vma, old_addr, old_end);
501 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
502 old_addr, old_end);
503 mmu_notifier_invalidate_range_start(&range);
504
505 for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
506 cond_resched();
507 /*
508 * If extent is PUD-sized try to speed up the move by moving at the
509 * PUD level if possible.
510 */
511 extent = get_extent(NORMAL_PUD, old_addr, old_end, new_addr);
512
513 old_pud = get_old_pud(vma->vm_mm, old_addr);
514 if (!old_pud)
515 continue;
516 new_pud = alloc_new_pud(vma->vm_mm, vma, new_addr);
517 if (!new_pud)
518 break;
519 if (pud_trans_huge(*old_pud) || pud_devmap(*old_pud)) {
520 if (extent == HPAGE_PUD_SIZE) {
521 move_pgt_entry(HPAGE_PUD, vma, old_addr, new_addr,
522 old_pud, new_pud, need_rmap_locks);
523 /* We ignore and continue on error? */
524 continue;
525 }
526 } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PUD) && extent == PUD_SIZE) {
527
528 if (move_pgt_entry(NORMAL_PUD, vma, old_addr, new_addr,
529 old_pud, new_pud, true))
530 continue;
531 }
532
533 extent = get_extent(NORMAL_PMD, old_addr, old_end, new_addr);
534 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
535 if (!old_pmd)
536 continue;
537 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
538 if (!new_pmd)
539 break;
540 if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) ||
541 pmd_devmap(*old_pmd)) {
542 if (extent == HPAGE_PMD_SIZE &&
543 move_pgt_entry(HPAGE_PMD, vma, old_addr, new_addr,
544 old_pmd, new_pmd, need_rmap_locks))
545 continue;
546 split_huge_pmd(vma, old_pmd, old_addr);
547 if (pmd_trans_unstable(old_pmd))
548 continue;
549 } else if (IS_ENABLED(CONFIG_HAVE_MOVE_PMD) &&
550 extent == PMD_SIZE) {
551 /*
552 * If the extent is PMD-sized, try to speed the move by
553 * moving at the PMD level if possible.
554 */
555 if (move_pgt_entry(NORMAL_PMD, vma, old_addr, new_addr,
556 old_pmd, new_pmd, true))
557 continue;
558 }
559
560 if (pte_alloc(new_vma->vm_mm, new_pmd))
561 break;
562 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
563 new_pmd, new_addr, need_rmap_locks);
564 }
565
566 mmu_notifier_invalidate_range_end(&range);
567
568 return len + old_addr - old_end; /* how much done */
569}
570
571static unsigned long move_vma(struct vm_area_struct *vma,
572 unsigned long old_addr, unsigned long old_len,
573 unsigned long new_len, unsigned long new_addr,
574 bool *locked, unsigned long flags,
575 struct vm_userfaultfd_ctx *uf, struct list_head *uf_unmap)
576{
577 long to_account = new_len - old_len;
578 struct mm_struct *mm = vma->vm_mm;
579 struct vm_area_struct *new_vma;
580 unsigned long vm_flags = vma->vm_flags;
581 unsigned long new_pgoff;
582 unsigned long moved_len;
583 unsigned long excess = 0;
584 unsigned long hiwater_vm;
585 int split = 0;
586 int err = 0;
587 bool need_rmap_locks;
588
589 /*
590 * We'd prefer to avoid failure later on in do_munmap:
591 * which may split one vma into three before unmapping.
592 */
593 if (mm->map_count >= sysctl_max_map_count - 3)
594 return -ENOMEM;
595
596 if (unlikely(flags & MREMAP_DONTUNMAP))
597 to_account = new_len;
598
599 if (vma->vm_ops && vma->vm_ops->may_split) {
600 if (vma->vm_start != old_addr)
601 err = vma->vm_ops->may_split(vma, old_addr);
602 if (!err && vma->vm_end != old_addr + old_len)
603 err = vma->vm_ops->may_split(vma, old_addr + old_len);
604 if (err)
605 return err;
606 }
607
608 /*
609 * Advise KSM to break any KSM pages in the area to be moved:
610 * it would be confusing if they were to turn up at the new
611 * location, where they happen to coincide with different KSM
612 * pages recently unmapped. But leave vma->vm_flags as it was,
613 * so KSM can come around to merge on vma and new_vma afterwards.
614 */
615 err = ksm_madvise(vma, old_addr, old_addr + old_len,
616 MADV_UNMERGEABLE, &vm_flags);
617 if (err)
618 return err;
619
620 if (vm_flags & VM_ACCOUNT) {
621 if (security_vm_enough_memory_mm(mm, to_account >> PAGE_SHIFT))
622 return -ENOMEM;
623 }
624
625 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
626 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
627 &need_rmap_locks);
628 if (!new_vma) {
629 if (vm_flags & VM_ACCOUNT)
630 vm_unacct_memory(to_account >> PAGE_SHIFT);
631 return -ENOMEM;
632 }
633
634 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
635 need_rmap_locks);
636 if (moved_len < old_len) {
637 err = -ENOMEM;
638 } else if (vma->vm_ops && vma->vm_ops->mremap) {
639 err = vma->vm_ops->mremap(new_vma);
640 }
641
642 if (unlikely(err)) {
643 /*
644 * On error, move entries back from new area to old,
645 * which will succeed since page tables still there,
646 * and then proceed to unmap new area instead of old.
647 */
648 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
649 true);
650 vma = new_vma;
651 old_len = new_len;
652 old_addr = new_addr;
653 new_addr = err;
654 } else {
655 mremap_userfaultfd_prep(new_vma, uf);
656 }
657
658 if (is_vm_hugetlb_page(vma)) {
659 clear_vma_resv_huge_pages(vma);
660 }
661
662 /* Conceal VM_ACCOUNT so old reservation is not undone */
663 if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP)) {
664 vma->vm_flags &= ~VM_ACCOUNT;
665 excess = vma->vm_end - vma->vm_start - old_len;
666 if (old_addr > vma->vm_start &&
667 old_addr + old_len < vma->vm_end)
668 split = 1;
669 }
670
671 /*
672 * If we failed to move page tables we still do total_vm increment
673 * since do_munmap() will decrement it by old_len == new_len.
674 *
675 * Since total_vm is about to be raised artificially high for a
676 * moment, we need to restore high watermark afterwards: if stats
677 * are taken meanwhile, total_vm and hiwater_vm appear too high.
678 * If this were a serious issue, we'd add a flag to do_munmap().
679 */
680 hiwater_vm = mm->hiwater_vm;
681 vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
682
683 /* Tell pfnmap has moved from this vma */
684 if (unlikely(vma->vm_flags & VM_PFNMAP))
685 untrack_pfn_moved(vma);
686
687 if (unlikely(!err && (flags & MREMAP_DONTUNMAP))) {
688 /* We always clear VM_LOCKED[ONFAULT] on the old vma */
689 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
690
691 /*
692 * anon_vma links of the old vma is no longer needed after its page
693 * table has been moved.
694 */
695 if (new_vma != vma && vma->vm_start == old_addr &&
696 vma->vm_end == (old_addr + old_len))
697 unlink_anon_vmas(vma);
698
699 /* Because we won't unmap we don't need to touch locked_vm */
700 return new_addr;
701 }
702
703 if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
704 /* OOM: unable to split vma, just get accounts right */
705 if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP))
706 vm_acct_memory(old_len >> PAGE_SHIFT);
707 excess = 0;
708 }
709
710 if (vm_flags & VM_LOCKED) {
711 mm->locked_vm += new_len >> PAGE_SHIFT;
712 *locked = true;
713 }
714
715 mm->hiwater_vm = hiwater_vm;
716
717 /* Restore VM_ACCOUNT if one or two pieces of vma left */
718 if (excess) {
719 vma->vm_flags |= VM_ACCOUNT;
720 if (split)
721 find_vma(mm, vma->vm_end)->vm_flags |= VM_ACCOUNT;
722 }
723
724 return new_addr;
725}
726
727static struct vm_area_struct *vma_to_resize(unsigned long addr,
728 unsigned long old_len, unsigned long new_len, unsigned long flags)
729{
730 struct mm_struct *mm = current->mm;
731 struct vm_area_struct *vma;
732 unsigned long pgoff;
733
734 vma = vma_lookup(mm, addr);
735 if (!vma)
736 return ERR_PTR(-EFAULT);
737
738 /*
739 * !old_len is a special case where an attempt is made to 'duplicate'
740 * a mapping. This makes no sense for private mappings as it will
741 * instead create a fresh/new mapping unrelated to the original. This
742 * is contrary to the basic idea of mremap which creates new mappings
743 * based on the original. There are no known use cases for this
744 * behavior. As a result, fail such attempts.
745 */
746 if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
747 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap. This is not supported.\n", current->comm, current->pid);
748 return ERR_PTR(-EINVAL);
749 }
750
751 if ((flags & MREMAP_DONTUNMAP) &&
752 (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)))
753 return ERR_PTR(-EINVAL);
754
755 /* We can't remap across vm area boundaries */
756 if (old_len > vma->vm_end - addr)
757 return ERR_PTR(-EFAULT);
758
759 if (new_len == old_len)
760 return vma;
761
762 /* Need to be careful about a growing mapping */
763 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
764 pgoff += vma->vm_pgoff;
765 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
766 return ERR_PTR(-EINVAL);
767
768 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
769 return ERR_PTR(-EFAULT);
770
771 if (mlock_future_check(mm, vma->vm_flags, new_len - old_len))
772 return ERR_PTR(-EAGAIN);
773
774 if (!may_expand_vm(mm, vma->vm_flags,
775 (new_len - old_len) >> PAGE_SHIFT))
776 return ERR_PTR(-ENOMEM);
777
778 return vma;
779}
780
781static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
782 unsigned long new_addr, unsigned long new_len, bool *locked,
783 unsigned long flags, struct vm_userfaultfd_ctx *uf,
784 struct list_head *uf_unmap_early,
785 struct list_head *uf_unmap)
786{
787 struct mm_struct *mm = current->mm;
788 struct vm_area_struct *vma;
789 unsigned long ret = -EINVAL;
790 unsigned long map_flags = 0;
791
792 if (offset_in_page(new_addr))
793 goto out;
794
795 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
796 goto out;
797
798 /* Ensure the old/new locations do not overlap */
799 if (addr + old_len > new_addr && new_addr + new_len > addr)
800 goto out;
801
802 /*
803 * move_vma() need us to stay 4 maps below the threshold, otherwise
804 * it will bail out at the very beginning.
805 * That is a problem if we have already unmaped the regions here
806 * (new_addr, and old_addr), because userspace will not know the
807 * state of the vma's after it gets -ENOMEM.
808 * So, to avoid such scenario we can pre-compute if the whole
809 * operation has high chances to success map-wise.
810 * Worst-scenario case is when both vma's (new_addr and old_addr) get
811 * split in 3 before unmapping it.
812 * That means 2 more maps (1 for each) to the ones we already hold.
813 * Check whether current map count plus 2 still leads us to 4 maps below
814 * the threshold, otherwise return -ENOMEM here to be more safe.
815 */
816 if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
817 return -ENOMEM;
818
819 if (flags & MREMAP_FIXED) {
820 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
821 if (ret)
822 goto out;
823 }
824
825 if (old_len > new_len) {
826 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
827 if (ret)
828 goto out;
829 old_len = new_len;
830 }
831
832 vma = vma_to_resize(addr, old_len, new_len, flags);
833 if (IS_ERR(vma)) {
834 ret = PTR_ERR(vma);
835 goto out;
836 }
837
838 /* MREMAP_DONTUNMAP expands by old_len since old_len == new_len */
839 if (flags & MREMAP_DONTUNMAP &&
840 !may_expand_vm(mm, vma->vm_flags, old_len >> PAGE_SHIFT)) {
841 ret = -ENOMEM;
842 goto out;
843 }
844
845 if (flags & MREMAP_FIXED)
846 map_flags |= MAP_FIXED;
847
848 if (vma->vm_flags & VM_MAYSHARE)
849 map_flags |= MAP_SHARED;
850
851 ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
852 ((addr - vma->vm_start) >> PAGE_SHIFT),
853 map_flags);
854 if (IS_ERR_VALUE(ret))
855 goto out;
856
857 /* We got a new mapping */
858 if (!(flags & MREMAP_FIXED))
859 new_addr = ret;
860
861 ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, flags, uf,
862 uf_unmap);
863
864out:
865 return ret;
866}
867
868static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
869{
870 unsigned long end = vma->vm_end + delta;
871
872 if (end < vma->vm_end) /* overflow */
873 return 0;
874 if (find_vma_intersection(vma->vm_mm, vma->vm_end, end))
875 return 0;
876 if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
877 0, MAP_FIXED) & ~PAGE_MASK)
878 return 0;
879 return 1;
880}
881
882/*
883 * Expand (or shrink) an existing mapping, potentially moving it at the
884 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
885 *
886 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
887 * This option implies MREMAP_MAYMOVE.
888 */
889SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
890 unsigned long, new_len, unsigned long, flags,
891 unsigned long, new_addr)
892{
893 struct mm_struct *mm = current->mm;
894 struct vm_area_struct *vma;
895 unsigned long ret = -EINVAL;
896 bool locked = false;
897 bool downgraded = false;
898 struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
899 LIST_HEAD(uf_unmap_early);
900 LIST_HEAD(uf_unmap);
901
902 /*
903 * There is a deliberate asymmetry here: we strip the pointer tag
904 * from the old address but leave the new address alone. This is
905 * for consistency with mmap(), where we prevent the creation of
906 * aliasing mappings in userspace by leaving the tag bits of the
907 * mapping address intact. A non-zero tag will cause the subsequent
908 * range checks to reject the address as invalid.
909 *
910 * See Documentation/arm64/tagged-address-abi.rst for more information.
911 */
912 addr = untagged_addr(addr);
913
914 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP))
915 return ret;
916
917 if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
918 return ret;
919
920 /*
921 * MREMAP_DONTUNMAP is always a move and it does not allow resizing
922 * in the process.
923 */
924 if (flags & MREMAP_DONTUNMAP &&
925 (!(flags & MREMAP_MAYMOVE) || old_len != new_len))
926 return ret;
927
928
929 if (offset_in_page(addr))
930 return ret;
931
932 old_len = PAGE_ALIGN(old_len);
933 new_len = PAGE_ALIGN(new_len);
934
935 /*
936 * We allow a zero old-len as a special case
937 * for DOS-emu "duplicate shm area" thing. But
938 * a zero new-len is nonsensical.
939 */
940 if (!new_len)
941 return ret;
942
943 if (mmap_write_lock_killable(current->mm))
944 return -EINTR;
945 vma = vma_lookup(mm, addr);
946 if (!vma) {
947 ret = -EFAULT;
948 goto out;
949 }
950
951 if (is_vm_hugetlb_page(vma)) {
952 struct hstate *h __maybe_unused = hstate_vma(vma);
953
954 old_len = ALIGN(old_len, huge_page_size(h));
955 new_len = ALIGN(new_len, huge_page_size(h));
956
957 /* addrs must be huge page aligned */
958 if (addr & ~huge_page_mask(h))
959 goto out;
960 if (new_addr & ~huge_page_mask(h))
961 goto out;
962
963 /*
964 * Don't allow remap expansion, because the underlying hugetlb
965 * reservation is not yet capable to handle split reservation.
966 */
967 if (new_len > old_len)
968 goto out;
969 }
970
971 if (flags & (MREMAP_FIXED | MREMAP_DONTUNMAP)) {
972 ret = mremap_to(addr, old_len, new_addr, new_len,
973 &locked, flags, &uf, &uf_unmap_early,
974 &uf_unmap);
975 goto out;
976 }
977
978 /*
979 * Always allow a shrinking remap: that just unmaps
980 * the unnecessary pages..
981 * do_mas_munmap does all the needed commit accounting, and
982 * downgrades mmap_lock to read if so directed.
983 */
984 if (old_len >= new_len) {
985 int retval;
986 MA_STATE(mas, &mm->mm_mt, addr + new_len, addr + new_len);
987
988 retval = do_mas_munmap(&mas, mm, addr + new_len,
989 old_len - new_len, &uf_unmap, true);
990 /* Returning 1 indicates mmap_lock is downgraded to read. */
991 if (retval == 1) {
992 downgraded = true;
993 } else if (retval < 0 && old_len != new_len) {
994 ret = retval;
995 goto out;
996 }
997
998 ret = addr;
999 goto out;
1000 }
1001
1002 /*
1003 * Ok, we need to grow..
1004 */
1005 vma = vma_to_resize(addr, old_len, new_len, flags);
1006 if (IS_ERR(vma)) {
1007 ret = PTR_ERR(vma);
1008 goto out;
1009 }
1010
1011 /* old_len exactly to the end of the area..
1012 */
1013 if (old_len == vma->vm_end - addr) {
1014 /* can we just expand the current mapping? */
1015 if (vma_expandable(vma, new_len - old_len)) {
1016 long pages = (new_len - old_len) >> PAGE_SHIFT;
1017 unsigned long extension_start = addr + old_len;
1018 unsigned long extension_end = addr + new_len;
1019 pgoff_t extension_pgoff = vma->vm_pgoff +
1020 ((extension_start - vma->vm_start) >> PAGE_SHIFT);
1021
1022 if (vma->vm_flags & VM_ACCOUNT) {
1023 if (security_vm_enough_memory_mm(mm, pages)) {
1024 ret = -ENOMEM;
1025 goto out;
1026 }
1027 }
1028
1029 /*
1030 * Function vma_merge() is called on the extension we
1031 * are adding to the already existing vma, vma_merge()
1032 * will merge this extension with the already existing
1033 * vma (expand operation itself) and possibly also with
1034 * the next vma if it becomes adjacent to the expanded
1035 * vma and otherwise compatible.
1036 *
1037 * However, vma_merge() can currently fail due to
1038 * is_mergeable_vma() check for vm_ops->close (see the
1039 * comment there). Yet this should not prevent vma
1040 * expanding, so perform a simple expand for such vma.
1041 * Ideally the check for close op should be only done
1042 * when a vma would be actually removed due to a merge.
1043 */
1044 if (!vma->vm_ops || !vma->vm_ops->close) {
1045 vma = vma_merge(mm, vma, extension_start, extension_end,
1046 vma->vm_flags, vma->anon_vma, vma->vm_file,
1047 extension_pgoff, vma_policy(vma),
1048 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
1049 } else if (vma_adjust(vma, vma->vm_start, addr + new_len,
1050 vma->vm_pgoff, NULL)) {
1051 vma = NULL;
1052 }
1053 if (!vma) {
1054 vm_unacct_memory(pages);
1055 ret = -ENOMEM;
1056 goto out;
1057 }
1058
1059 vm_stat_account(mm, vma->vm_flags, pages);
1060 if (vma->vm_flags & VM_LOCKED) {
1061 mm->locked_vm += pages;
1062 locked = true;
1063 new_addr = addr;
1064 }
1065 ret = addr;
1066 goto out;
1067 }
1068 }
1069
1070 /*
1071 * We weren't able to just expand or shrink the area,
1072 * we need to create a new one and move it..
1073 */
1074 ret = -ENOMEM;
1075 if (flags & MREMAP_MAYMOVE) {
1076 unsigned long map_flags = 0;
1077 if (vma->vm_flags & VM_MAYSHARE)
1078 map_flags |= MAP_SHARED;
1079
1080 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
1081 vma->vm_pgoff +
1082 ((addr - vma->vm_start) >> PAGE_SHIFT),
1083 map_flags);
1084 if (IS_ERR_VALUE(new_addr)) {
1085 ret = new_addr;
1086 goto out;
1087 }
1088
1089 ret = move_vma(vma, addr, old_len, new_len, new_addr,
1090 &locked, flags, &uf, &uf_unmap);
1091 }
1092out:
1093 if (offset_in_page(ret))
1094 locked = false;
1095 if (downgraded)
1096 mmap_read_unlock(current->mm);
1097 else
1098 mmap_write_unlock(current->mm);
1099 if (locked && new_len > old_len)
1100 mm_populate(new_addr + old_len, new_len - old_len);
1101 userfaultfd_unmap_complete(mm, &uf_unmap_early);
1102 mremap_userfaultfd_complete(&uf, addr, ret, old_len);
1103 userfaultfd_unmap_complete(mm, &uf_unmap);
1104 return ret;
1105}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * mm/mremap.c
4 *
5 * (C) Copyright 1996 Linus Torvalds
6 *
7 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
8 * (C) Copyright 2002 Red Hat Inc, All Rights Reserved
9 */
10
11#include <linux/mm.h>
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>
18#include <linux/fs.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>
27
28#include <asm/cacheflush.h>
29#include <asm/tlbflush.h>
30
31#include "internal.h"
32
33static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
34{
35 pgd_t *pgd;
36 p4d_t *p4d;
37 pud_t *pud;
38 pmd_t *pmd;
39
40 pgd = pgd_offset(mm, addr);
41 if (pgd_none_or_clear_bad(pgd))
42 return NULL;
43
44 p4d = p4d_offset(pgd, addr);
45 if (p4d_none_or_clear_bad(p4d))
46 return NULL;
47
48 pud = pud_offset(p4d, addr);
49 if (pud_none_or_clear_bad(pud))
50 return NULL;
51
52 pmd = pmd_offset(pud, addr);
53 if (pmd_none(*pmd))
54 return NULL;
55
56 return pmd;
57}
58
59static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
60 unsigned long addr)
61{
62 pgd_t *pgd;
63 p4d_t *p4d;
64 pud_t *pud;
65 pmd_t *pmd;
66
67 pgd = pgd_offset(mm, addr);
68 p4d = p4d_alloc(mm, pgd, addr);
69 if (!p4d)
70 return NULL;
71 pud = pud_alloc(mm, p4d, addr);
72 if (!pud)
73 return NULL;
74
75 pmd = pmd_alloc(mm, pud, addr);
76 if (!pmd)
77 return NULL;
78
79 VM_BUG_ON(pmd_trans_huge(*pmd));
80
81 return pmd;
82}
83
84static void take_rmap_locks(struct vm_area_struct *vma)
85{
86 if (vma->vm_file)
87 i_mmap_lock_write(vma->vm_file->f_mapping);
88 if (vma->anon_vma)
89 anon_vma_lock_write(vma->anon_vma);
90}
91
92static void drop_rmap_locks(struct vm_area_struct *vma)
93{
94 if (vma->anon_vma)
95 anon_vma_unlock_write(vma->anon_vma);
96 if (vma->vm_file)
97 i_mmap_unlock_write(vma->vm_file->f_mapping);
98}
99
100static pte_t move_soft_dirty_pte(pte_t pte)
101{
102 /*
103 * Set soft dirty bit so we can notice
104 * in userspace the ptes were moved.
105 */
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);
111#endif
112 return pte;
113}
114
115static 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)
119{
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;
125
126 /*
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...
131 *
132 * When need_rmap_locks is false, we use other ways to avoid
133 * such races:
134 *
135 * - During exec() shift_arg_pages(), we use a specially tagged vma
136 * which rmap call sites look for using vma_is_temporary_stack().
137 *
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).
143 */
144 if (need_rmap_locks)
145 take_rmap_locks(vma);
146
147 /*
148 * We don't have to worry about the ordering of src and dst
149 * pte locks because exclusive mmap_lock prevents deadlock.
150 */
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();
158
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))
162 continue;
163
164 pte = ptep_get_and_clear(mm, old_addr, old_pte);
165 /*
166 * If we are remapping a valid PTE, make sure
167 * to flush TLB before we drop the PTL for the
168 * PTE.
169 *
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
174 * flushed.
175 */
176 if (pte_present(pte))
177 force_flush = true;
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);
181 }
182
183 arch_leave_lazy_mmu_mode();
184 if (force_flush)
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);
190 if (need_rmap_locks)
191 drop_rmap_locks(vma);
192}
193
194#ifdef CONFIG_HAVE_MOVE_PMD
195static 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)
197{
198 spinlock_t *old_ptl, *new_ptl;
199 struct mm_struct *mm = vma->vm_mm;
200 pmd_t pmd;
201
202 /*
203 * The destination pmd shouldn't be established, free_pgtables()
204 * should have released it.
205 *
206 * However, there's a case during execve() where we use mremap
207 * to move the initial stack, and in that case the target area
208 * may overlap the source area (always moving down).
209 *
210 * If everything is PMD-aligned, that works fine, as moving
211 * each pmd down will clear the source pmd. But if we first
212 * have a few 4kB-only pages that get moved down, and then
213 * hit the "now the rest is PMD-aligned, let's do everything
214 * one pmd at a time", we will still have the old (now empty
215 * of any 4kB pages, but still there) PMD in the page table
216 * tree.
217 *
218 * Warn on it once - because we really should try to figure
219 * out how to do this better - but then say "I won't move
220 * this pmd".
221 *
222 * One alternative might be to just unmap the target pmd at
223 * this point, and verify that it really is empty. We'll see.
224 */
225 if (WARN_ON_ONCE(!pmd_none(*new_pmd)))
226 return false;
227
228 /*
229 * We don't have to worry about the ordering of src and dst
230 * ptlocks because exclusive mmap_lock prevents deadlock.
231 */
232 old_ptl = pmd_lock(vma->vm_mm, old_pmd);
233 new_ptl = pmd_lockptr(mm, new_pmd);
234 if (new_ptl != old_ptl)
235 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
236
237 /* Clear the pmd */
238 pmd = *old_pmd;
239 pmd_clear(old_pmd);
240
241 VM_BUG_ON(!pmd_none(*new_pmd));
242
243 /* Set the new pmd */
244 set_pmd_at(mm, new_addr, new_pmd, pmd);
245 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
246 if (new_ptl != old_ptl)
247 spin_unlock(new_ptl);
248 spin_unlock(old_ptl);
249
250 return true;
251}
252#endif
253
254unsigned long move_page_tables(struct vm_area_struct *vma,
255 unsigned long old_addr, struct vm_area_struct *new_vma,
256 unsigned long new_addr, unsigned long len,
257 bool need_rmap_locks)
258{
259 unsigned long extent, next, old_end;
260 struct mmu_notifier_range range;
261 pmd_t *old_pmd, *new_pmd;
262
263 old_end = old_addr + len;
264 flush_cache_range(vma, old_addr, old_end);
265
266 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
267 old_addr, old_end);
268 mmu_notifier_invalidate_range_start(&range);
269
270 for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
271 cond_resched();
272 next = (old_addr + PMD_SIZE) & PMD_MASK;
273 /* even if next overflowed, extent below will be ok */
274 extent = next - old_addr;
275 if (extent > old_end - old_addr)
276 extent = old_end - old_addr;
277 next = (new_addr + PMD_SIZE) & PMD_MASK;
278 if (extent > next - new_addr)
279 extent = next - new_addr;
280 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
281 if (!old_pmd)
282 continue;
283 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
284 if (!new_pmd)
285 break;
286 if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) || pmd_devmap(*old_pmd)) {
287 if (extent == HPAGE_PMD_SIZE) {
288 bool moved;
289 /* See comment in move_ptes() */
290 if (need_rmap_locks)
291 take_rmap_locks(vma);
292 moved = move_huge_pmd(vma, old_addr, new_addr,
293 old_pmd, new_pmd);
294 if (need_rmap_locks)
295 drop_rmap_locks(vma);
296 if (moved)
297 continue;
298 }
299 split_huge_pmd(vma, old_pmd, old_addr);
300 if (pmd_trans_unstable(old_pmd))
301 continue;
302 } else if (extent == PMD_SIZE) {
303#ifdef CONFIG_HAVE_MOVE_PMD
304 /*
305 * If the extent is PMD-sized, try to speed the move by
306 * moving at the PMD level if possible.
307 */
308 bool moved;
309
310 if (need_rmap_locks)
311 take_rmap_locks(vma);
312 moved = move_normal_pmd(vma, old_addr, new_addr,
313 old_pmd, new_pmd);
314 if (need_rmap_locks)
315 drop_rmap_locks(vma);
316 if (moved)
317 continue;
318#endif
319 }
320
321 if (pte_alloc(new_vma->vm_mm, new_pmd))
322 break;
323 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
324 new_pmd, new_addr, need_rmap_locks);
325 }
326
327 mmu_notifier_invalidate_range_end(&range);
328
329 return len + old_addr - old_end; /* how much done */
330}
331
332static unsigned long move_vma(struct vm_area_struct *vma,
333 unsigned long old_addr, unsigned long old_len,
334 unsigned long new_len, unsigned long new_addr,
335 bool *locked, unsigned long flags,
336 struct vm_userfaultfd_ctx *uf, struct list_head *uf_unmap)
337{
338 struct mm_struct *mm = vma->vm_mm;
339 struct vm_area_struct *new_vma;
340 unsigned long vm_flags = vma->vm_flags;
341 unsigned long new_pgoff;
342 unsigned long moved_len;
343 unsigned long excess = 0;
344 unsigned long hiwater_vm;
345 int split = 0;
346 int err;
347 bool need_rmap_locks;
348
349 /*
350 * We'd prefer to avoid failure later on in do_munmap:
351 * which may split one vma into three before unmapping.
352 */
353 if (mm->map_count >= sysctl_max_map_count - 3)
354 return -ENOMEM;
355
356 /*
357 * Advise KSM to break any KSM pages in the area to be moved:
358 * it would be confusing if they were to turn up at the new
359 * location, where they happen to coincide with different KSM
360 * pages recently unmapped. But leave vma->vm_flags as it was,
361 * so KSM can come around to merge on vma and new_vma afterwards.
362 */
363 err = ksm_madvise(vma, old_addr, old_addr + old_len,
364 MADV_UNMERGEABLE, &vm_flags);
365 if (err)
366 return err;
367
368 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
369 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
370 &need_rmap_locks);
371 if (!new_vma)
372 return -ENOMEM;
373
374 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
375 need_rmap_locks);
376 if (moved_len < old_len) {
377 err = -ENOMEM;
378 } else if (vma->vm_ops && vma->vm_ops->mremap) {
379 err = vma->vm_ops->mremap(new_vma);
380 }
381
382 if (unlikely(err)) {
383 /*
384 * On error, move entries back from new area to old,
385 * which will succeed since page tables still there,
386 * and then proceed to unmap new area instead of old.
387 */
388 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
389 true);
390 vma = new_vma;
391 old_len = new_len;
392 old_addr = new_addr;
393 new_addr = err;
394 } else {
395 mremap_userfaultfd_prep(new_vma, uf);
396 arch_remap(mm, old_addr, old_addr + old_len,
397 new_addr, new_addr + new_len);
398 }
399
400 /* Conceal VM_ACCOUNT so old reservation is not undone */
401 if (vm_flags & VM_ACCOUNT) {
402 vma->vm_flags &= ~VM_ACCOUNT;
403 excess = vma->vm_end - vma->vm_start - old_len;
404 if (old_addr > vma->vm_start &&
405 old_addr + old_len < vma->vm_end)
406 split = 1;
407 }
408
409 /*
410 * If we failed to move page tables we still do total_vm increment
411 * since do_munmap() will decrement it by old_len == new_len.
412 *
413 * Since total_vm is about to be raised artificially high for a
414 * moment, we need to restore high watermark afterwards: if stats
415 * are taken meanwhile, total_vm and hiwater_vm appear too high.
416 * If this were a serious issue, we'd add a flag to do_munmap().
417 */
418 hiwater_vm = mm->hiwater_vm;
419 vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
420
421 /* Tell pfnmap has moved from this vma */
422 if (unlikely(vma->vm_flags & VM_PFNMAP))
423 untrack_pfn_moved(vma);
424
425 if (unlikely(!err && (flags & MREMAP_DONTUNMAP))) {
426 if (vm_flags & VM_ACCOUNT) {
427 /* Always put back VM_ACCOUNT since we won't unmap */
428 vma->vm_flags |= VM_ACCOUNT;
429
430 vm_acct_memory(new_len >> PAGE_SHIFT);
431 }
432
433 /*
434 * VMAs can actually be merged back together in copy_vma
435 * calling merge_vma. This can happen with anonymous vmas
436 * which have not yet been faulted, so if we were to consider
437 * this VMA split we'll end up adding VM_ACCOUNT on the
438 * next VMA, which is completely unrelated if this VMA
439 * was re-merged.
440 */
441 if (split && new_vma == vma)
442 split = 0;
443
444 /* We always clear VM_LOCKED[ONFAULT] on the old vma */
445 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
446
447 /* Because we won't unmap we don't need to touch locked_vm */
448 goto out;
449 }
450
451 if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
452 /* OOM: unable to split vma, just get accounts right */
453 vm_unacct_memory(excess >> PAGE_SHIFT);
454 excess = 0;
455 }
456
457 if (vm_flags & VM_LOCKED) {
458 mm->locked_vm += new_len >> PAGE_SHIFT;
459 *locked = true;
460 }
461out:
462 mm->hiwater_vm = hiwater_vm;
463
464 /* Restore VM_ACCOUNT if one or two pieces of vma left */
465 if (excess) {
466 vma->vm_flags |= VM_ACCOUNT;
467 if (split)
468 vma->vm_next->vm_flags |= VM_ACCOUNT;
469 }
470
471 return new_addr;
472}
473
474static struct vm_area_struct *vma_to_resize(unsigned long addr,
475 unsigned long old_len, unsigned long new_len, unsigned long flags,
476 unsigned long *p)
477{
478 struct mm_struct *mm = current->mm;
479 struct vm_area_struct *vma = find_vma(mm, addr);
480 unsigned long pgoff;
481
482 if (!vma || vma->vm_start > addr)
483 return ERR_PTR(-EFAULT);
484
485 /*
486 * !old_len is a special case where an attempt is made to 'duplicate'
487 * a mapping. This makes no sense for private mappings as it will
488 * instead create a fresh/new mapping unrelated to the original. This
489 * is contrary to the basic idea of mremap which creates new mappings
490 * based on the original. There are no known use cases for this
491 * behavior. As a result, fail such attempts.
492 */
493 if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
494 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap. This is not supported.\n", current->comm, current->pid);
495 return ERR_PTR(-EINVAL);
496 }
497
498 if (flags & MREMAP_DONTUNMAP && (!vma_is_anonymous(vma) ||
499 vma->vm_flags & VM_SHARED))
500 return ERR_PTR(-EINVAL);
501
502 if (is_vm_hugetlb_page(vma))
503 return ERR_PTR(-EINVAL);
504
505 /* We can't remap across vm area boundaries */
506 if (old_len > vma->vm_end - addr)
507 return ERR_PTR(-EFAULT);
508
509 if (new_len == old_len)
510 return vma;
511
512 /* Need to be careful about a growing mapping */
513 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
514 pgoff += vma->vm_pgoff;
515 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
516 return ERR_PTR(-EINVAL);
517
518 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
519 return ERR_PTR(-EFAULT);
520
521 if (vma->vm_flags & VM_LOCKED) {
522 unsigned long locked, lock_limit;
523 locked = mm->locked_vm << PAGE_SHIFT;
524 lock_limit = rlimit(RLIMIT_MEMLOCK);
525 locked += new_len - old_len;
526 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
527 return ERR_PTR(-EAGAIN);
528 }
529
530 if (!may_expand_vm(mm, vma->vm_flags,
531 (new_len - old_len) >> PAGE_SHIFT))
532 return ERR_PTR(-ENOMEM);
533
534 if (vma->vm_flags & VM_ACCOUNT) {
535 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
536 if (security_vm_enough_memory_mm(mm, charged))
537 return ERR_PTR(-ENOMEM);
538 *p = charged;
539 }
540
541 return vma;
542}
543
544static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
545 unsigned long new_addr, unsigned long new_len, bool *locked,
546 unsigned long flags, struct vm_userfaultfd_ctx *uf,
547 struct list_head *uf_unmap_early,
548 struct list_head *uf_unmap)
549{
550 struct mm_struct *mm = current->mm;
551 struct vm_area_struct *vma;
552 unsigned long ret = -EINVAL;
553 unsigned long charged = 0;
554 unsigned long map_flags = 0;
555
556 if (offset_in_page(new_addr))
557 goto out;
558
559 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
560 goto out;
561
562 /* Ensure the old/new locations do not overlap */
563 if (addr + old_len > new_addr && new_addr + new_len > addr)
564 goto out;
565
566 /*
567 * move_vma() need us to stay 4 maps below the threshold, otherwise
568 * it will bail out at the very beginning.
569 * That is a problem if we have already unmaped the regions here
570 * (new_addr, and old_addr), because userspace will not know the
571 * state of the vma's after it gets -ENOMEM.
572 * So, to avoid such scenario we can pre-compute if the whole
573 * operation has high chances to success map-wise.
574 * Worst-scenario case is when both vma's (new_addr and old_addr) get
575 * split in 3 before unmaping it.
576 * That means 2 more maps (1 for each) to the ones we already hold.
577 * Check whether current map count plus 2 still leads us to 4 maps below
578 * the threshold, otherwise return -ENOMEM here to be more safe.
579 */
580 if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
581 return -ENOMEM;
582
583 if (flags & MREMAP_FIXED) {
584 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
585 if (ret)
586 goto out;
587 }
588
589 if (old_len >= new_len) {
590 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
591 if (ret && old_len != new_len)
592 goto out;
593 old_len = new_len;
594 }
595
596 vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
597 if (IS_ERR(vma)) {
598 ret = PTR_ERR(vma);
599 goto out;
600 }
601
602 /* MREMAP_DONTUNMAP expands by old_len since old_len == new_len */
603 if (flags & MREMAP_DONTUNMAP &&
604 !may_expand_vm(mm, vma->vm_flags, old_len >> PAGE_SHIFT)) {
605 ret = -ENOMEM;
606 goto out;
607 }
608
609 if (flags & MREMAP_FIXED)
610 map_flags |= MAP_FIXED;
611
612 if (vma->vm_flags & VM_MAYSHARE)
613 map_flags |= MAP_SHARED;
614
615 ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
616 ((addr - vma->vm_start) >> PAGE_SHIFT),
617 map_flags);
618 if (IS_ERR_VALUE(ret))
619 goto out1;
620
621 /* We got a new mapping */
622 if (!(flags & MREMAP_FIXED))
623 new_addr = ret;
624
625 ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, flags, uf,
626 uf_unmap);
627
628 if (!(offset_in_page(ret)))
629 goto out;
630
631out1:
632 vm_unacct_memory(charged);
633
634out:
635 return ret;
636}
637
638static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
639{
640 unsigned long end = vma->vm_end + delta;
641 if (end < vma->vm_end) /* overflow */
642 return 0;
643 if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
644 return 0;
645 if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
646 0, MAP_FIXED) & ~PAGE_MASK)
647 return 0;
648 return 1;
649}
650
651/*
652 * Expand (or shrink) an existing mapping, potentially moving it at the
653 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
654 *
655 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
656 * This option implies MREMAP_MAYMOVE.
657 */
658SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
659 unsigned long, new_len, unsigned long, flags,
660 unsigned long, new_addr)
661{
662 struct mm_struct *mm = current->mm;
663 struct vm_area_struct *vma;
664 unsigned long ret = -EINVAL;
665 unsigned long charged = 0;
666 bool locked = false;
667 bool downgraded = false;
668 struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
669 LIST_HEAD(uf_unmap_early);
670 LIST_HEAD(uf_unmap);
671
672 /*
673 * There is a deliberate asymmetry here: we strip the pointer tag
674 * from the old address but leave the new address alone. This is
675 * for consistency with mmap(), where we prevent the creation of
676 * aliasing mappings in userspace by leaving the tag bits of the
677 * mapping address intact. A non-zero tag will cause the subsequent
678 * range checks to reject the address as invalid.
679 *
680 * See Documentation/arm64/tagged-address-abi.rst for more information.
681 */
682 addr = untagged_addr(addr);
683
684 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP))
685 return ret;
686
687 if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
688 return ret;
689
690 /*
691 * MREMAP_DONTUNMAP is always a move and it does not allow resizing
692 * in the process.
693 */
694 if (flags & MREMAP_DONTUNMAP &&
695 (!(flags & MREMAP_MAYMOVE) || old_len != new_len))
696 return ret;
697
698
699 if (offset_in_page(addr))
700 return ret;
701
702 old_len = PAGE_ALIGN(old_len);
703 new_len = PAGE_ALIGN(new_len);
704
705 /*
706 * We allow a zero old-len as a special case
707 * for DOS-emu "duplicate shm area" thing. But
708 * a zero new-len is nonsensical.
709 */
710 if (!new_len)
711 return ret;
712
713 if (mmap_write_lock_killable(current->mm))
714 return -EINTR;
715
716 if (flags & (MREMAP_FIXED | MREMAP_DONTUNMAP)) {
717 ret = mremap_to(addr, old_len, new_addr, new_len,
718 &locked, flags, &uf, &uf_unmap_early,
719 &uf_unmap);
720 goto out;
721 }
722
723 /*
724 * Always allow a shrinking remap: that just unmaps
725 * the unnecessary pages..
726 * __do_munmap does all the needed commit accounting, and
727 * downgrades mmap_lock to read if so directed.
728 */
729 if (old_len >= new_len) {
730 int retval;
731
732 retval = __do_munmap(mm, addr+new_len, old_len - new_len,
733 &uf_unmap, true);
734 if (retval < 0 && old_len != new_len) {
735 ret = retval;
736 goto out;
737 /* Returning 1 indicates mmap_lock is downgraded to read. */
738 } else if (retval == 1)
739 downgraded = true;
740 ret = addr;
741 goto out;
742 }
743
744 /*
745 * Ok, we need to grow..
746 */
747 vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
748 if (IS_ERR(vma)) {
749 ret = PTR_ERR(vma);
750 goto out;
751 }
752
753 /* old_len exactly to the end of the area..
754 */
755 if (old_len == vma->vm_end - addr) {
756 /* can we just expand the current mapping? */
757 if (vma_expandable(vma, new_len - old_len)) {
758 int pages = (new_len - old_len) >> PAGE_SHIFT;
759
760 if (vma_adjust(vma, vma->vm_start, addr + new_len,
761 vma->vm_pgoff, NULL)) {
762 ret = -ENOMEM;
763 goto out;
764 }
765
766 vm_stat_account(mm, vma->vm_flags, pages);
767 if (vma->vm_flags & VM_LOCKED) {
768 mm->locked_vm += pages;
769 locked = true;
770 new_addr = addr;
771 }
772 ret = addr;
773 goto out;
774 }
775 }
776
777 /*
778 * We weren't able to just expand or shrink the area,
779 * we need to create a new one and move it..
780 */
781 ret = -ENOMEM;
782 if (flags & MREMAP_MAYMOVE) {
783 unsigned long map_flags = 0;
784 if (vma->vm_flags & VM_MAYSHARE)
785 map_flags |= MAP_SHARED;
786
787 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
788 vma->vm_pgoff +
789 ((addr - vma->vm_start) >> PAGE_SHIFT),
790 map_flags);
791 if (IS_ERR_VALUE(new_addr)) {
792 ret = new_addr;
793 goto out;
794 }
795
796 ret = move_vma(vma, addr, old_len, new_len, new_addr,
797 &locked, flags, &uf, &uf_unmap);
798 }
799out:
800 if (offset_in_page(ret)) {
801 vm_unacct_memory(charged);
802 locked = false;
803 }
804 if (downgraded)
805 mmap_read_unlock(current->mm);
806 else
807 mmap_write_unlock(current->mm);
808 if (locked && new_len > old_len)
809 mm_populate(new_addr + old_len, new_len - old_len);
810 userfaultfd_unmap_complete(mm, &uf_unmap_early);
811 mremap_userfaultfd_complete(&uf, addr, ret, old_len);
812 userfaultfd_unmap_complete(mm, &uf_unmap);
813 return ret;
814}