1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * mm/userfaultfd.c
4 *
5 * Copyright (C) 2015 Red Hat, Inc.
6 */
7
8 #include <linux/mm.h>
9 #include <linux/sched/signal.h>
10 #include <linux/pagemap.h>
11 #include <linux/rmap.h>
12 #include <linux/swap.h>
13 #include <linux/swapops.h>
14 #include <linux/userfaultfd_k.h>
15 #include <linux/mmu_notifier.h>
16 #include <linux/hugetlb.h>
17 #include <linux/shmem_fs.h>
18 #include <asm/tlbflush.h>
19 #include <asm/tlb.h>
20 #include "internal.h"
21
22 static __always_inline
23 bool validate_dst_vma(struct vm_area_struct *dst_vma, unsigned long dst_end)
24 {
25 /* Make sure that the dst range is fully within dst_vma. */
26 if (dst_end > dst_vma->vm_end)
27 return false;
28
29 /*
30 * Check the vma is registered in uffd, this is required to
31 * enforce the VM_MAYWRITE check done at uffd registration
32 * time.
33 */
34 if (!dst_vma->vm_userfaultfd_ctx.ctx)
35 return false;
36
37 return true;
38 }
39
40 static __always_inline
41 struct vm_area_struct *find_vma_and_prepare_anon(struct mm_struct *mm,
42 unsigned long addr)
43 {
44 struct vm_area_struct *vma;
45
46 mmap_assert_locked(mm);
47 vma = vma_lookup(mm, addr);
48 if (!vma)
49 vma = ERR_PTR(-ENOENT);
50 else if (!(vma->vm_flags & VM_SHARED) &&
51 unlikely(anon_vma_prepare(vma)))
52 vma = ERR_PTR(-ENOMEM);
53
54 return vma;
55 }
56
57 #ifdef CONFIG_PER_VMA_LOCK
58 /*
59 * uffd_lock_vma() - Lookup and lock vma corresponding to @address.
60 * @mm: mm to search vma in.
61 * @address: address that the vma should contain.
62 *
63 * Should be called without holding mmap_lock.
64 *
65 * Return: A locked vma containing @address, -ENOENT if no vma is found, or
66 * -ENOMEM if anon_vma couldn't be allocated.
67 */
68 static struct vm_area_struct *uffd_lock_vma(struct mm_struct *mm,
69 unsigned long address)
70 {
71 struct vm_area_struct *vma;
72
73 vma = lock_vma_under_rcu(mm, address);
74 if (vma) {
75 /*
76 * We know we're going to need to use anon_vma, so check
77 * that early.
78 */
79 if (!(vma->vm_flags & VM_SHARED) && unlikely(!vma->anon_vma))
80 vma_end_read(vma);
81 else
82 return vma;
83 }
84
85 mmap_read_lock(mm);
86 vma = find_vma_and_prepare_anon(mm, address);
87 if (!IS_ERR(vma)) {
88 /*
89 * We cannot use vma_start_read() as it may fail due to
90 * false locked (see comment in vma_start_read()). We
91 * can avoid that by directly locking vm_lock under
92 * mmap_lock, which guarantees that nobody can lock the
93 * vma for write (vma_start_write()) under us.
94 */
95 down_read(&vma->vm_lock->lock);
96 }
97
98 mmap_read_unlock(mm);
99 return vma;
100 }
101
102 static struct vm_area_struct *uffd_mfill_lock(struct mm_struct *dst_mm,
103 unsigned long dst_start,
104 unsigned long len)
105 {
106 struct vm_area_struct *dst_vma;
107
108 dst_vma = uffd_lock_vma(dst_mm, dst_start);
109 if (IS_ERR(dst_vma) || validate_dst_vma(dst_vma, dst_start + len))
110 return dst_vma;
111
112 vma_end_read(dst_vma);
113 return ERR_PTR(-ENOENT);
114 }
115
116 static void uffd_mfill_unlock(struct vm_area_struct *vma)
117 {
118 vma_end_read(vma);
119 }
120
121 #else
122
123 static struct vm_area_struct *uffd_mfill_lock(struct mm_struct *dst_mm,
124 unsigned long dst_start,
125 unsigned long len)
126 {
127 struct vm_area_struct *dst_vma;
128
129 mmap_read_lock(dst_mm);
130 dst_vma = find_vma_and_prepare_anon(dst_mm, dst_start);
131 if (IS_ERR(dst_vma))
132 goto out_unlock;
133
134 if (validate_dst_vma(dst_vma, dst_start + len))
135 return dst_vma;
136
137 dst_vma = ERR_PTR(-ENOENT);
138 out_unlock:
139 mmap_read_unlock(dst_mm);
140 return dst_vma;
141 }
142
143 static void uffd_mfill_unlock(struct vm_area_struct *vma)
144 {
145 mmap_read_unlock(vma->vm_mm);
146 }
147 #endif
148
149 /* Check if dst_addr is outside of file's size. Must be called with ptl held. */
150 static bool mfill_file_over_size(struct vm_area_struct *dst_vma,
151 unsigned long dst_addr)
152 {
153 struct inode *inode;
154 pgoff_t offset, max_off;
155
156 if (!dst_vma->vm_file)
157 return false;
158
159 inode = dst_vma->vm_file->f_inode;
160 offset = linear_page_index(dst_vma, dst_addr);
161 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
162 return offset >= max_off;
163 }
164
165 /*
166 * Install PTEs, to map dst_addr (within dst_vma) to page.
167 *
168 * This function handles both MCOPY_ATOMIC_NORMAL and _CONTINUE for both shmem
169 * and anon, and for both shared and private VMAs.
170 */
171 int mfill_atomic_install_pte(pmd_t *dst_pmd,
172 struct vm_area_struct *dst_vma,
173 unsigned long dst_addr, struct page *page,
174 bool newly_allocated, uffd_flags_t flags)
175 {
176 int ret;
177 struct mm_struct *dst_mm = dst_vma->vm_mm;
178 pte_t _dst_pte, *dst_pte;
179 bool writable = dst_vma->vm_flags & VM_WRITE;
180 bool vm_shared = dst_vma->vm_flags & VM_SHARED;
181 spinlock_t *ptl;
182 struct folio *folio = page_folio(page);
183 bool page_in_cache = folio_mapping(folio);
184
185 _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
186 _dst_pte = pte_mkdirty(_dst_pte);
187 if (page_in_cache && !vm_shared)
188 writable = false;
189 if (writable)
190 _dst_pte = pte_mkwrite(_dst_pte, dst_vma);
191 if (flags & MFILL_ATOMIC_WP)
192 _dst_pte = pte_mkuffd_wp(_dst_pte);
193
194 ret = -EAGAIN;
195 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
196 if (!dst_pte)
197 goto out;
198
199 if (mfill_file_over_size(dst_vma, dst_addr)) {
200 ret = -EFAULT;
201 goto out_unlock;
202 }
203
204 ret = -EEXIST;
205 /*
206 * We allow to overwrite a pte marker: consider when both MISSING|WP
207 * registered, we firstly wr-protect a none pte which has no page cache
208 * page backing it, then access the page.
209 */
210 if (!pte_none_mostly(ptep_get(dst_pte)))
211 goto out_unlock;
212
213 if (page_in_cache) {
214 /* Usually, cache pages are already added to LRU */
215 if (newly_allocated)
216 folio_add_lru(folio);
217 folio_add_file_rmap_pte(folio, page, dst_vma);
218 } else {
219 folio_add_new_anon_rmap(folio, dst_vma, dst_addr, RMAP_EXCLUSIVE);
220 folio_add_lru_vma(folio, dst_vma);
221 }
222
223 /*
224 * Must happen after rmap, as mm_counter() checks mapping (via
225 * PageAnon()), which is set by __page_set_anon_rmap().
226 */
227 inc_mm_counter(dst_mm, mm_counter(folio));
228
229 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
230
231 /* No need to invalidate - it was non-present before */
232 update_mmu_cache(dst_vma, dst_addr, dst_pte);
233 ret = 0;
234 out_unlock:
235 pte_unmap_unlock(dst_pte, ptl);
236 out:
237 return ret;
238 }
239
240 static int mfill_atomic_pte_copy(pmd_t *dst_pmd,
241 struct vm_area_struct *dst_vma,
242 unsigned long dst_addr,
243 unsigned long src_addr,
244 uffd_flags_t flags,
245 struct folio **foliop)
246 {
247 void *kaddr;
248 int ret;
249 struct folio *folio;
250
251 if (!*foliop) {
252 ret = -ENOMEM;
253 folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, dst_vma,
254 dst_addr, false);
255 if (!folio)
256 goto out;
257
258 kaddr = kmap_local_folio(folio, 0);
259 /*
260 * The read mmap_lock is held here. Despite the
261 * mmap_lock being read recursive a deadlock is still
262 * possible if a writer has taken a lock. For example:
263 *
264 * process A thread 1 takes read lock on own mmap_lock
265 * process A thread 2 calls mmap, blocks taking write lock
266 * process B thread 1 takes page fault, read lock on own mmap lock
267 * process B thread 2 calls mmap, blocks taking write lock
268 * process A thread 1 blocks taking read lock on process B
269 * process B thread 1 blocks taking read lock on process A
270 *
271 * Disable page faults to prevent potential deadlock
272 * and retry the copy outside the mmap_lock.
273 */
274 pagefault_disable();
275 ret = copy_from_user(kaddr, (const void __user *) src_addr,
276 PAGE_SIZE);
277 pagefault_enable();
278 kunmap_local(kaddr);
279
280 /* fallback to copy_from_user outside mmap_lock */
281 if (unlikely(ret)) {
282 ret = -ENOENT;
283 *foliop = folio;
284 /* don't free the page */
285 goto out;
286 }
287
288 flush_dcache_folio(folio);
289 } else {
290 folio = *foliop;
291 *foliop = NULL;
292 }
293
294 /*
295 * The memory barrier inside __folio_mark_uptodate makes sure that
296 * preceding stores to the page contents become visible before
297 * the set_pte_at() write.
298 */
299 __folio_mark_uptodate(folio);
300
301 ret = -ENOMEM;
302 if (mem_cgroup_charge(folio, dst_vma->vm_mm, GFP_KERNEL))
303 goto out_release;
304
305 ret = mfill_atomic_install_pte(dst_pmd, dst_vma, dst_addr,
306 &folio->page, true, flags);
307 if (ret)
308 goto out_release;
309 out:
310 return ret;
311 out_release:
312 folio_put(folio);
313 goto out;
314 }
315
316 static int mfill_atomic_pte_zeroed_folio(pmd_t *dst_pmd,
317 struct vm_area_struct *dst_vma,
318 unsigned long dst_addr)
319 {
320 struct folio *folio;
321 int ret = -ENOMEM;
322
323 folio = vma_alloc_zeroed_movable_folio(dst_vma, dst_addr);
324 if (!folio)
325 return ret;
326
327 if (mem_cgroup_charge(folio, dst_vma->vm_mm, GFP_KERNEL))
328 goto out_put;
329
330 /*
331 * The memory barrier inside __folio_mark_uptodate makes sure that
332 * zeroing out the folio become visible before mapping the page
333 * using set_pte_at(). See do_anonymous_page().
334 */
335 __folio_mark_uptodate(folio);
336
337 ret = mfill_atomic_install_pte(dst_pmd, dst_vma, dst_addr,
338 &folio->page, true, 0);
339 if (ret)
340 goto out_put;
341
342 return 0;
343 out_put:
344 folio_put(folio);
345 return ret;
346 }
347
348 static int mfill_atomic_pte_zeropage(pmd_t *dst_pmd,
349 struct vm_area_struct *dst_vma,
350 unsigned long dst_addr)
351 {
352 pte_t _dst_pte, *dst_pte;
353 spinlock_t *ptl;
354 int ret;
355
356 if (mm_forbids_zeropage(dst_vma->vm_mm))
357 return mfill_atomic_pte_zeroed_folio(dst_pmd, dst_vma, dst_addr);
358
359 _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
360 dst_vma->vm_page_prot));
361 ret = -EAGAIN;
362 dst_pte = pte_offset_map_lock(dst_vma->vm_mm, dst_pmd, dst_addr, &ptl);
363 if (!dst_pte)
364 goto out;
365 if (mfill_file_over_size(dst_vma, dst_addr)) {
366 ret = -EFAULT;
367 goto out_unlock;
368 }
369 ret = -EEXIST;
370 if (!pte_none(ptep_get(dst_pte)))
371 goto out_unlock;
372 set_pte_at(dst_vma->vm_mm, dst_addr, dst_pte, _dst_pte);
373 /* No need to invalidate - it was non-present before */
374 update_mmu_cache(dst_vma, dst_addr, dst_pte);
375 ret = 0;
376 out_unlock:
377 pte_unmap_unlock(dst_pte, ptl);
378 out:
379 return ret;
380 }
381
382 /* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */
383 static int mfill_atomic_pte_continue(pmd_t *dst_pmd,
384 struct vm_area_struct *dst_vma,
385 unsigned long dst_addr,
386 uffd_flags_t flags)
387 {
388 struct inode *inode = file_inode(dst_vma->vm_file);
389 pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
390 struct folio *folio;
391 struct page *page;
392 int ret;
393
394 ret = shmem_get_folio(inode, pgoff, &folio, SGP_NOALLOC);
395 /* Our caller expects us to return -EFAULT if we failed to find folio */
396 if (ret == -ENOENT)
397 ret = -EFAULT;
398 if (ret)
399 goto out;
400 if (!folio) {
401 ret = -EFAULT;
402 goto out;
403 }
404
405 page = folio_file_page(folio, pgoff);
406 if (PageHWPoison(page)) {
407 ret = -EIO;
408 goto out_release;
409 }
410
411 ret = mfill_atomic_install_pte(dst_pmd, dst_vma, dst_addr,
412 page, false, flags);
413 if (ret)
414 goto out_release;
415
416 folio_unlock(folio);
417 ret = 0;
418 out:
419 return ret;
420 out_release:
421 folio_unlock(folio);
422 folio_put(folio);
423 goto out;
424 }
425
426 /* Handles UFFDIO_POISON for all non-hugetlb VMAs. */
427 static int mfill_atomic_pte_poison(pmd_t *dst_pmd,
428 struct vm_area_struct *dst_vma,
429 unsigned long dst_addr,
430 uffd_flags_t flags)
431 {
432 int ret;
433 struct mm_struct *dst_mm = dst_vma->vm_mm;
434 pte_t _dst_pte, *dst_pte;
435 spinlock_t *ptl;
436
437 _dst_pte = make_pte_marker(PTE_MARKER_POISONED);
438 ret = -EAGAIN;
439 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
440 if (!dst_pte)
441 goto out;
442
443 if (mfill_file_over_size(dst_vma, dst_addr)) {
444 ret = -EFAULT;
445 goto out_unlock;
446 }
447
448 ret = -EEXIST;
449 /* Refuse to overwrite any PTE, even a PTE marker (e.g. UFFD WP). */
450 if (!pte_none(ptep_get(dst_pte)))
451 goto out_unlock;
452
453 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
454
455 /* No need to invalidate - it was non-present before */
456 update_mmu_cache(dst_vma, dst_addr, dst_pte);
457 ret = 0;
458 out_unlock:
459 pte_unmap_unlock(dst_pte, ptl);
460 out:
461 return ret;
462 }
463
464 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
465 {
466 pgd_t *pgd;
467 p4d_t *p4d;
468 pud_t *pud;
469
470 pgd = pgd_offset(mm, address);
471 p4d = p4d_alloc(mm, pgd, address);
472 if (!p4d)
473 return NULL;
474 pud = pud_alloc(mm, p4d, address);
475 if (!pud)
476 return NULL;
477 /*
478 * Note that we didn't run this because the pmd was
479 * missing, the *pmd may be already established and in
480 * turn it may also be a trans_huge_pmd.
481 */
482 return pmd_alloc(mm, pud, address);
483 }
484
485 #ifdef CONFIG_HUGETLB_PAGE
486 /*
487 * mfill_atomic processing for HUGETLB vmas. Note that this routine is
488 * called with either vma-lock or mmap_lock held, it will release the lock
489 * before returning.
490 */
491 static __always_inline ssize_t mfill_atomic_hugetlb(
492 struct userfaultfd_ctx *ctx,
493 struct vm_area_struct *dst_vma,
494 unsigned long dst_start,
495 unsigned long src_start,
496 unsigned long len,
497 uffd_flags_t flags)
498 {
499 struct mm_struct *dst_mm = dst_vma->vm_mm;
500 ssize_t err;
501 pte_t *dst_pte;
502 unsigned long src_addr, dst_addr;
503 long copied;
504 struct folio *folio;
505 unsigned long vma_hpagesize;
506 pgoff_t idx;
507 u32 hash;
508 struct address_space *mapping;
509
510 /*
511 * There is no default zero huge page for all huge page sizes as
512 * supported by hugetlb. A PMD_SIZE huge pages may exist as used
513 * by THP. Since we can not reliably insert a zero page, this
514 * feature is not supported.
515 */
516 if (uffd_flags_mode_is(flags, MFILL_ATOMIC_ZEROPAGE)) {
517 up_read(&ctx->map_changing_lock);
518 uffd_mfill_unlock(dst_vma);
519 return -EINVAL;
520 }
521
522 src_addr = src_start;
523 dst_addr = dst_start;
524 copied = 0;
525 folio = NULL;
526 vma_hpagesize = vma_kernel_pagesize(dst_vma);
527
528 /*
529 * Validate alignment based on huge page size
530 */
531 err = -EINVAL;
532 if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
533 goto out_unlock;
534
535 retry:
536 /*
537 * On routine entry dst_vma is set. If we had to drop mmap_lock and
538 * retry, dst_vma will be set to NULL and we must lookup again.
539 */
540 if (!dst_vma) {
541 dst_vma = uffd_mfill_lock(dst_mm, dst_start, len);
542 if (IS_ERR(dst_vma)) {
543 err = PTR_ERR(dst_vma);
544 goto out;
545 }
546
547 err = -ENOENT;
548 if (!is_vm_hugetlb_page(dst_vma))
549 goto out_unlock_vma;
550
551 err = -EINVAL;
552 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
553 goto out_unlock_vma;
554
555 /*
556 * If memory mappings are changing because of non-cooperative
557 * operation (e.g. mremap) running in parallel, bail out and
558 * request the user to retry later
559 */
560 down_read(&ctx->map_changing_lock);
561 err = -EAGAIN;
562 if (atomic_read(&ctx->mmap_changing))
563 goto out_unlock;
564 }
565
566 while (src_addr < src_start + len) {
567 BUG_ON(dst_addr >= dst_start + len);
568
569 /*
570 * Serialize via vma_lock and hugetlb_fault_mutex.
571 * vma_lock ensures the dst_pte remains valid even
572 * in the case of shared pmds. fault mutex prevents
573 * races with other faulting threads.
574 */
575 idx = linear_page_index(dst_vma, dst_addr);
576 mapping = dst_vma->vm_file->f_mapping;
577 hash = hugetlb_fault_mutex_hash(mapping, idx);
578 mutex_lock(&hugetlb_fault_mutex_table[hash]);
579 hugetlb_vma_lock_read(dst_vma);
580
581 err = -ENOMEM;
582 dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize);
583 if (!dst_pte) {
584 hugetlb_vma_unlock_read(dst_vma);
585 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
586 goto out_unlock;
587 }
588
589 if (!uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE) &&
590 !huge_pte_none_mostly(huge_ptep_get(dst_mm, dst_addr, dst_pte))) {
591 err = -EEXIST;
592 hugetlb_vma_unlock_read(dst_vma);
593 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
594 goto out_unlock;
595 }
596
597 err = hugetlb_mfill_atomic_pte(dst_pte, dst_vma, dst_addr,
598 src_addr, flags, &folio);
599
600 hugetlb_vma_unlock_read(dst_vma);
601 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
602
603 cond_resched();
604
605 if (unlikely(err == -ENOENT)) {
606 up_read(&ctx->map_changing_lock);
607 uffd_mfill_unlock(dst_vma);
608 BUG_ON(!folio);
609
610 err = copy_folio_from_user(folio,
611 (const void __user *)src_addr, true);
612 if (unlikely(err)) {
613 err = -EFAULT;
614 goto out;
615 }
616
617 dst_vma = NULL;
618 goto retry;
619 } else
620 BUG_ON(folio);
621
622 if (!err) {
623 dst_addr += vma_hpagesize;
624 src_addr += vma_hpagesize;
625 copied += vma_hpagesize;
626
627 if (fatal_signal_pending(current))
628 err = -EINTR;
629 }
630 if (err)
631 break;
632 }
633
634 out_unlock:
635 up_read(&ctx->map_changing_lock);
636 out_unlock_vma:
637 uffd_mfill_unlock(dst_vma);
638 out:
639 if (folio)
640 folio_put(folio);
641 BUG_ON(copied < 0);
642 BUG_ON(err > 0);
643 BUG_ON(!copied && !err);
644 return copied ? copied : err;
645 }
646 #else /* !CONFIG_HUGETLB_PAGE */
647 /* fail at build time if gcc attempts to use this */
648 extern ssize_t mfill_atomic_hugetlb(struct userfaultfd_ctx *ctx,
649 struct vm_area_struct *dst_vma,
650 unsigned long dst_start,
651 unsigned long src_start,
652 unsigned long len,
653 uffd_flags_t flags);
654 #endif /* CONFIG_HUGETLB_PAGE */
655
656 static __always_inline ssize_t mfill_atomic_pte(pmd_t *dst_pmd,
657 struct vm_area_struct *dst_vma,
658 unsigned long dst_addr,
659 unsigned long src_addr,
660 uffd_flags_t flags,
661 struct folio **foliop)
662 {
663 ssize_t err;
664
665 if (uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE)) {
666 return mfill_atomic_pte_continue(dst_pmd, dst_vma,
667 dst_addr, flags);
668 } else if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) {
669 return mfill_atomic_pte_poison(dst_pmd, dst_vma,
670 dst_addr, flags);
671 }
672
673 /*
674 * The normal page fault path for a shmem will invoke the
675 * fault, fill the hole in the file and COW it right away. The
676 * result generates plain anonymous memory. So when we are
677 * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
678 * generate anonymous memory directly without actually filling
679 * the hole. For the MAP_PRIVATE case the robustness check
680 * only happens in the pagetable (to verify it's still none)
681 * and not in the radix tree.
682 */
683 if (!(dst_vma->vm_flags & VM_SHARED)) {
684 if (uffd_flags_mode_is(flags, MFILL_ATOMIC_COPY))
685 err = mfill_atomic_pte_copy(dst_pmd, dst_vma,
686 dst_addr, src_addr,
687 flags, foliop);
688 else
689 err = mfill_atomic_pte_zeropage(dst_pmd,
690 dst_vma, dst_addr);
691 } else {
692 err = shmem_mfill_atomic_pte(dst_pmd, dst_vma,
693 dst_addr, src_addr,
694 flags, foliop);
695 }
696
697 return err;
698 }
699
700 static __always_inline ssize_t mfill_atomic(struct userfaultfd_ctx *ctx,
701 unsigned long dst_start,
702 unsigned long src_start,
703 unsigned long len,
704 uffd_flags_t flags)
705 {
706 struct mm_struct *dst_mm = ctx->mm;
707 struct vm_area_struct *dst_vma;
708 ssize_t err;
709 pmd_t *dst_pmd;
710 unsigned long src_addr, dst_addr;
711 long copied;
712 struct folio *folio;
713
714 /*
715 * Sanitize the command parameters:
716 */
717 BUG_ON(dst_start & ~PAGE_MASK);
718 BUG_ON(len & ~PAGE_MASK);
719
720 /* Does the address range wrap, or is the span zero-sized? */
721 BUG_ON(src_start + len <= src_start);
722 BUG_ON(dst_start + len <= dst_start);
723
724 src_addr = src_start;
725 dst_addr = dst_start;
726 copied = 0;
727 folio = NULL;
728 retry:
729 /*
730 * Make sure the vma is not shared, that the dst range is
731 * both valid and fully within a single existing vma.
732 */
733 dst_vma = uffd_mfill_lock(dst_mm, dst_start, len);
734 if (IS_ERR(dst_vma)) {
735 err = PTR_ERR(dst_vma);
736 goto out;
737 }
738
739 /*
740 * If memory mappings are changing because of non-cooperative
741 * operation (e.g. mremap) running in parallel, bail out and
742 * request the user to retry later
743 */
744 down_read(&ctx->map_changing_lock);
745 err = -EAGAIN;
746 if (atomic_read(&ctx->mmap_changing))
747 goto out_unlock;
748
749 err = -EINVAL;
750 /*
751 * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
752 * it will overwrite vm_ops, so vma_is_anonymous must return false.
753 */
754 if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
755 dst_vma->vm_flags & VM_SHARED))
756 goto out_unlock;
757
758 /*
759 * validate 'mode' now that we know the dst_vma: don't allow
760 * a wrprotect copy if the userfaultfd didn't register as WP.
761 */
762 if ((flags & MFILL_ATOMIC_WP) && !(dst_vma->vm_flags & VM_UFFD_WP))
763 goto out_unlock;
764
765 /*
766 * If this is a HUGETLB vma, pass off to appropriate routine
767 */
768 if (is_vm_hugetlb_page(dst_vma))
769 return mfill_atomic_hugetlb(ctx, dst_vma, dst_start,
770 src_start, len, flags);
771
772 if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
773 goto out_unlock;
774 if (!vma_is_shmem(dst_vma) &&
775 uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE))
776 goto out_unlock;
777
778 while (src_addr < src_start + len) {
779 pmd_t dst_pmdval;
780
781 BUG_ON(dst_addr >= dst_start + len);
782
783 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
784 if (unlikely(!dst_pmd)) {
785 err = -ENOMEM;
786 break;
787 }
788
789 dst_pmdval = pmdp_get_lockless(dst_pmd);
790 if (unlikely(pmd_none(dst_pmdval)) &&
791 unlikely(__pte_alloc(dst_mm, dst_pmd))) {
792 err = -ENOMEM;
793 break;
794 }
795 dst_pmdval = pmdp_get_lockless(dst_pmd);
796 /*
797 * If the dst_pmd is THP don't override it and just be strict.
798 * (This includes the case where the PMD used to be THP and
799 * changed back to none after __pte_alloc().)
800 */
801 if (unlikely(!pmd_present(dst_pmdval) || pmd_trans_huge(dst_pmdval) ||
802 pmd_devmap(dst_pmdval))) {
803 err = -EEXIST;
804 break;
805 }
806 if (unlikely(pmd_bad(dst_pmdval))) {
807 err = -EFAULT;
808 break;
809 }
810 /*
811 * For shmem mappings, khugepaged is allowed to remove page
812 * tables under us; pte_offset_map_lock() will deal with that.
813 */
814
815 err = mfill_atomic_pte(dst_pmd, dst_vma, dst_addr,
816 src_addr, flags, &folio);
817 cond_resched();
818
819 if (unlikely(err == -ENOENT)) {
820 void *kaddr;
821
822 up_read(&ctx->map_changing_lock);
823 uffd_mfill_unlock(dst_vma);
824 BUG_ON(!folio);
825
826 kaddr = kmap_local_folio(folio, 0);
827 err = copy_from_user(kaddr,
828 (const void __user *) src_addr,
829 PAGE_SIZE);
830 kunmap_local(kaddr);
831 if (unlikely(err)) {
832 err = -EFAULT;
833 goto out;
834 }
835 flush_dcache_folio(folio);
836 goto retry;
837 } else
838 BUG_ON(folio);
839
840 if (!err) {
841 dst_addr += PAGE_SIZE;
842 src_addr += PAGE_SIZE;
843 copied += PAGE_SIZE;
844
845 if (fatal_signal_pending(current))
846 err = -EINTR;
847 }
848 if (err)
849 break;
850 }
851
852 out_unlock:
853 up_read(&ctx->map_changing_lock);
854 uffd_mfill_unlock(dst_vma);
855 out:
856 if (folio)
857 folio_put(folio);
858 BUG_ON(copied < 0);
859 BUG_ON(err > 0);
860 BUG_ON(!copied && !err);
861 return copied ? copied : err;
862 }
863
864 ssize_t mfill_atomic_copy(struct userfaultfd_ctx *ctx, unsigned long dst_start,
865 unsigned long src_start, unsigned long len,
866 uffd_flags_t flags)
867 {
868 return mfill_atomic(ctx, dst_start, src_start, len,
869 uffd_flags_set_mode(flags, MFILL_ATOMIC_COPY));
870 }
871
872 ssize_t mfill_atomic_zeropage(struct userfaultfd_ctx *ctx,
873 unsigned long start,
874 unsigned long len)
875 {
876 return mfill_atomic(ctx, start, 0, len,
877 uffd_flags_set_mode(0, MFILL_ATOMIC_ZEROPAGE));
878 }
879
880 ssize_t mfill_atomic_continue(struct userfaultfd_ctx *ctx, unsigned long start,
881 unsigned long len, uffd_flags_t flags)
882 {
883
884 /*
885 * A caller might reasonably assume that UFFDIO_CONTINUE contains an
886 * smp_wmb() to ensure that any writes to the about-to-be-mapped page by
887 * the thread doing the UFFDIO_CONTINUE are guaranteed to be visible to
888 * subsequent loads from the page through the newly mapped address range.
889 */
890 smp_wmb();
891
892 return mfill_atomic(ctx, start, 0, len,
893 uffd_flags_set_mode(flags, MFILL_ATOMIC_CONTINUE));
894 }
895
896 ssize_t mfill_atomic_poison(struct userfaultfd_ctx *ctx, unsigned long start,
897 unsigned long len, uffd_flags_t flags)
898 {
899 return mfill_atomic(ctx, start, 0, len,
900 uffd_flags_set_mode(flags, MFILL_ATOMIC_POISON));
901 }
902
903 long uffd_wp_range(struct vm_area_struct *dst_vma,
904 unsigned long start, unsigned long len, bool enable_wp)
905 {
906 unsigned int mm_cp_flags;
907 struct mmu_gather tlb;
908 long ret;
909
910 VM_WARN_ONCE(start < dst_vma->vm_start || start + len > dst_vma->vm_end,
911 "The address range exceeds VMA boundary.\n");
912 if (enable_wp)
913 mm_cp_flags = MM_CP_UFFD_WP;
914 else
915 mm_cp_flags = MM_CP_UFFD_WP_RESOLVE;
916
917 /*
918 * vma->vm_page_prot already reflects that uffd-wp is enabled for this
919 * VMA (see userfaultfd_set_vm_flags()) and that all PTEs are supposed
920 * to be write-protected as default whenever protection changes.
921 * Try upgrading write permissions manually.
922 */
923 if (!enable_wp && vma_wants_manual_pte_write_upgrade(dst_vma))
924 mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
925 tlb_gather_mmu(&tlb, dst_vma->vm_mm);
926 ret = change_protection(&tlb, dst_vma, start, start + len, mm_cp_flags);
927 tlb_finish_mmu(&tlb);
928
929 return ret;
930 }
931
932 int mwriteprotect_range(struct userfaultfd_ctx *ctx, unsigned long start,
933 unsigned long len, bool enable_wp)
934 {
935 struct mm_struct *dst_mm = ctx->mm;
936 unsigned long end = start + len;
937 unsigned long _start, _end;
938 struct vm_area_struct *dst_vma;
939 unsigned long page_mask;
940 long err;
941 VMA_ITERATOR(vmi, dst_mm, start);
942
943 /*
944 * Sanitize the command parameters:
945 */
946 BUG_ON(start & ~PAGE_MASK);
947 BUG_ON(len & ~PAGE_MASK);
948
949 /* Does the address range wrap, or is the span zero-sized? */
950 BUG_ON(start + len <= start);
951
952 mmap_read_lock(dst_mm);
953
954 /*
955 * If memory mappings are changing because of non-cooperative
956 * operation (e.g. mremap) running in parallel, bail out and
957 * request the user to retry later
958 */
959 down_read(&ctx->map_changing_lock);
960 err = -EAGAIN;
961 if (atomic_read(&ctx->mmap_changing))
962 goto out_unlock;
963
964 err = -ENOENT;
965 for_each_vma_range(vmi, dst_vma, end) {
966
967 if (!userfaultfd_wp(dst_vma)) {
968 err = -ENOENT;
969 break;
970 }
971
972 if (is_vm_hugetlb_page(dst_vma)) {
973 err = -EINVAL;
974 page_mask = vma_kernel_pagesize(dst_vma) - 1;
975 if ((start & page_mask) || (len & page_mask))
976 break;
977 }
978
979 _start = max(dst_vma->vm_start, start);
980 _end = min(dst_vma->vm_end, end);
981
982 err = uffd_wp_range(dst_vma, _start, _end - _start, enable_wp);
983
984 /* Return 0 on success, <0 on failures */
985 if (err < 0)
986 break;
987 err = 0;
988 }
989 out_unlock:
990 up_read(&ctx->map_changing_lock);
991 mmap_read_unlock(dst_mm);
992 return err;
993 }
994
995
996 void double_pt_lock(spinlock_t *ptl1,
997 spinlock_t *ptl2)
998 __acquires(ptl1)
999 __acquires(ptl2)
1000 {
1001 if (ptl1 > ptl2)
1002 swap(ptl1, ptl2);
1003 /* lock in virtual address order to avoid lock inversion */
1004 spin_lock(ptl1);
1005 if (ptl1 != ptl2)
1006 spin_lock_nested(ptl2, SINGLE_DEPTH_NESTING);
1007 else
1008 __acquire(ptl2);
1009 }
1010
1011 void double_pt_unlock(spinlock_t *ptl1,
1012 spinlock_t *ptl2)
1013 __releases(ptl1)
1014 __releases(ptl2)
1015 {
1016 spin_unlock(ptl1);
1017 if (ptl1 != ptl2)
1018 spin_unlock(ptl2);
1019 else
1020 __release(ptl2);
1021 }
1022
1023
1024 static int move_present_pte(struct mm_struct *mm,
1025 struct vm_area_struct *dst_vma,
1026 struct vm_area_struct *src_vma,
1027 unsigned long dst_addr, unsigned long src_addr,
1028 pte_t *dst_pte, pte_t *src_pte,
1029 pte_t orig_dst_pte, pte_t orig_src_pte,
1030 spinlock_t *dst_ptl, spinlock_t *src_ptl,
1031 struct folio *src_folio)
1032 {
1033 int err = 0;
1034
1035 double_pt_lock(dst_ptl, src_ptl);
1036
1037 if (!pte_same(ptep_get(src_pte), orig_src_pte) ||
1038 !pte_same(ptep_get(dst_pte), orig_dst_pte)) {
1039 err = -EAGAIN;
1040 goto out;
1041 }
1042 if (folio_test_large(src_folio) ||
1043 folio_maybe_dma_pinned(src_folio) ||
1044 !PageAnonExclusive(&src_folio->page)) {
1045 err = -EBUSY;
1046 goto out;
1047 }
1048
1049 orig_src_pte = ptep_clear_flush(src_vma, src_addr, src_pte);
1050 /* Folio got pinned from under us. Put it back and fail the move. */
1051 if (folio_maybe_dma_pinned(src_folio)) {
1052 set_pte_at(mm, src_addr, src_pte, orig_src_pte);
1053 err = -EBUSY;
1054 goto out;
1055 }
1056
1057 folio_move_anon_rmap(src_folio, dst_vma);
1058 src_folio->index = linear_page_index(dst_vma, dst_addr);
1059
1060 orig_dst_pte = mk_pte(&src_folio->page, dst_vma->vm_page_prot);
1061 /* Follow mremap() behavior and treat the entry dirty after the move */
1062 orig_dst_pte = pte_mkwrite(pte_mkdirty(orig_dst_pte), dst_vma);
1063
1064 set_pte_at(mm, dst_addr, dst_pte, orig_dst_pte);
1065 out:
1066 double_pt_unlock(dst_ptl, src_ptl);
1067 return err;
1068 }
1069
1070 static int move_swap_pte(struct mm_struct *mm,
1071 unsigned long dst_addr, unsigned long src_addr,
1072 pte_t *dst_pte, pte_t *src_pte,
1073 pte_t orig_dst_pte, pte_t orig_src_pte,
1074 spinlock_t *dst_ptl, spinlock_t *src_ptl)
1075 {
1076 if (!pte_swp_exclusive(orig_src_pte))
1077 return -EBUSY;
1078
1079 double_pt_lock(dst_ptl, src_ptl);
1080
1081 if (!pte_same(ptep_get(src_pte), orig_src_pte) ||
1082 !pte_same(ptep_get(dst_pte), orig_dst_pte)) {
1083 double_pt_unlock(dst_ptl, src_ptl);
1084 return -EAGAIN;
1085 }
1086
1087 orig_src_pte = ptep_get_and_clear(mm, src_addr, src_pte);
1088 set_pte_at(mm, dst_addr, dst_pte, orig_src_pte);
1089 double_pt_unlock(dst_ptl, src_ptl);
1090
1091 return 0;
1092 }
1093
1094 static int move_zeropage_pte(struct mm_struct *mm,
1095 struct vm_area_struct *dst_vma,
1096 struct vm_area_struct *src_vma,
1097 unsigned long dst_addr, unsigned long src_addr,
1098 pte_t *dst_pte, pte_t *src_pte,
1099 pte_t orig_dst_pte, pte_t orig_src_pte,
1100 spinlock_t *dst_ptl, spinlock_t *src_ptl)
1101 {
1102 pte_t zero_pte;
1103
1104 double_pt_lock(dst_ptl, src_ptl);
1105 if (!pte_same(ptep_get(src_pte), orig_src_pte) ||
1106 !pte_same(ptep_get(dst_pte), orig_dst_pte)) {
1107 double_pt_unlock(dst_ptl, src_ptl);
1108 return -EAGAIN;
1109 }
1110
1111 zero_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
1112 dst_vma->vm_page_prot));
1113 ptep_clear_flush(src_vma, src_addr, src_pte);
1114 set_pte_at(mm, dst_addr, dst_pte, zero_pte);
1115 double_pt_unlock(dst_ptl, src_ptl);
1116
1117 return 0;
1118 }
1119
1120
1121 /*
1122 * The mmap_lock for reading is held by the caller. Just move the page
1123 * from src_pmd to dst_pmd if possible, and return true if succeeded
1124 * in moving the page.
1125 */
1126 static int move_pages_pte(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd,
1127 struct vm_area_struct *dst_vma,
1128 struct vm_area_struct *src_vma,
1129 unsigned long dst_addr, unsigned long src_addr,
1130 __u64 mode)
1131 {
1132 swp_entry_t entry;
1133 pte_t orig_src_pte, orig_dst_pte;
1134 pte_t src_folio_pte;
1135 spinlock_t *src_ptl, *dst_ptl;
1136 pte_t *src_pte = NULL;
1137 pte_t *dst_pte = NULL;
1138
1139 struct folio *src_folio = NULL;
1140 struct anon_vma *src_anon_vma = NULL;
1141 struct mmu_notifier_range range;
1142 int err = 0;
1143
1144 flush_cache_range(src_vma, src_addr, src_addr + PAGE_SIZE);
1145 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
1146 src_addr, src_addr + PAGE_SIZE);
1147 mmu_notifier_invalidate_range_start(&range);
1148 retry:
1149 dst_pte = pte_offset_map_nolock(mm, dst_pmd, dst_addr, &dst_ptl);
1150
1151 /* Retry if a huge pmd materialized from under us */
1152 if (unlikely(!dst_pte)) {
1153 err = -EAGAIN;
1154 goto out;
1155 }
1156
1157 src_pte = pte_offset_map_nolock(mm, src_pmd, src_addr, &src_ptl);
1158
1159 /*
1160 * We held the mmap_lock for reading so MADV_DONTNEED
1161 * can zap transparent huge pages under us, or the
1162 * transparent huge page fault can establish new
1163 * transparent huge pages under us.
1164 */
1165 if (unlikely(!src_pte)) {
1166 err = -EAGAIN;
1167 goto out;
1168 }
1169
1170 /* Sanity checks before the operation */
1171 if (WARN_ON_ONCE(pmd_none(*dst_pmd)) || WARN_ON_ONCE(pmd_none(*src_pmd)) ||
1172 WARN_ON_ONCE(pmd_trans_huge(*dst_pmd)) || WARN_ON_ONCE(pmd_trans_huge(*src_pmd))) {
1173 err = -EINVAL;
1174 goto out;
1175 }
1176
1177 spin_lock(dst_ptl);
1178 orig_dst_pte = ptep_get(dst_pte);
1179 spin_unlock(dst_ptl);
1180 if (!pte_none(orig_dst_pte)) {
1181 err = -EEXIST;
1182 goto out;
1183 }
1184
1185 spin_lock(src_ptl);
1186 orig_src_pte = ptep_get(src_pte);
1187 spin_unlock(src_ptl);
1188 if (pte_none(orig_src_pte)) {
1189 if (!(mode & UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES))
1190 err = -ENOENT;
1191 else /* nothing to do to move a hole */
1192 err = 0;
1193 goto out;
1194 }
1195
1196 /* If PTE changed after we locked the folio them start over */
1197 if (src_folio && unlikely(!pte_same(src_folio_pte, orig_src_pte))) {
1198 err = -EAGAIN;
1199 goto out;
1200 }
1201
1202 if (pte_present(orig_src_pte)) {
1203 if (is_zero_pfn(pte_pfn(orig_src_pte))) {
1204 err = move_zeropage_pte(mm, dst_vma, src_vma,
1205 dst_addr, src_addr, dst_pte, src_pte,
1206 orig_dst_pte, orig_src_pte,
1207 dst_ptl, src_ptl);
1208 goto out;
1209 }
1210
1211 /*
1212 * Pin and lock both source folio and anon_vma. Since we are in
1213 * RCU read section, we can't block, so on contention have to
1214 * unmap the ptes, obtain the lock and retry.
1215 */
1216 if (!src_folio) {
1217 struct folio *folio;
1218
1219 /*
1220 * Pin the page while holding the lock to be sure the
1221 * page isn't freed under us
1222 */
1223 spin_lock(src_ptl);
1224 if (!pte_same(orig_src_pte, ptep_get(src_pte))) {
1225 spin_unlock(src_ptl);
1226 err = -EAGAIN;
1227 goto out;
1228 }
1229
1230 folio = vm_normal_folio(src_vma, src_addr, orig_src_pte);
1231 if (!folio || !PageAnonExclusive(&folio->page)) {
1232 spin_unlock(src_ptl);
1233 err = -EBUSY;
1234 goto out;
1235 }
1236
1237 folio_get(folio);
1238 src_folio = folio;
1239 src_folio_pte = orig_src_pte;
1240 spin_unlock(src_ptl);
1241
1242 if (!folio_trylock(src_folio)) {
1243 pte_unmap(&orig_src_pte);
1244 pte_unmap(&orig_dst_pte);
1245 src_pte = dst_pte = NULL;
1246 /* now we can block and wait */
1247 folio_lock(src_folio);
1248 goto retry;
1249 }
1250
1251 if (WARN_ON_ONCE(!folio_test_anon(src_folio))) {
1252 err = -EBUSY;
1253 goto out;
1254 }
1255 }
1256
1257 /* at this point we have src_folio locked */
1258 if (folio_test_large(src_folio)) {
1259 /* split_folio() can block */
1260 pte_unmap(&orig_src_pte);
1261 pte_unmap(&orig_dst_pte);
1262 src_pte = dst_pte = NULL;
1263 err = split_folio(src_folio);
1264 if (err)
1265 goto out;
1266 /* have to reacquire the folio after it got split */
1267 folio_unlock(src_folio);
1268 folio_put(src_folio);
1269 src_folio = NULL;
1270 goto retry;
1271 }
1272
1273 if (!src_anon_vma) {
1274 /*
1275 * folio_referenced walks the anon_vma chain
1276 * without the folio lock. Serialize against it with
1277 * the anon_vma lock, the folio lock is not enough.
1278 */
1279 src_anon_vma = folio_get_anon_vma(src_folio);
1280 if (!src_anon_vma) {
1281 /* page was unmapped from under us */
1282 err = -EAGAIN;
1283 goto out;
1284 }
1285 if (!anon_vma_trylock_write(src_anon_vma)) {
1286 pte_unmap(&orig_src_pte);
1287 pte_unmap(&orig_dst_pte);
1288 src_pte = dst_pte = NULL;
1289 /* now we can block and wait */
1290 anon_vma_lock_write(src_anon_vma);
1291 goto retry;
1292 }
1293 }
1294
1295 err = move_present_pte(mm, dst_vma, src_vma,
1296 dst_addr, src_addr, dst_pte, src_pte,
1297 orig_dst_pte, orig_src_pte,
1298 dst_ptl, src_ptl, src_folio);
1299 } else {
1300 entry = pte_to_swp_entry(orig_src_pte);
1301 if (non_swap_entry(entry)) {
1302 if (is_migration_entry(entry)) {
1303 pte_unmap(&orig_src_pte);
1304 pte_unmap(&orig_dst_pte);
1305 src_pte = dst_pte = NULL;
1306 migration_entry_wait(mm, src_pmd, src_addr);
1307 err = -EAGAIN;
1308 } else
1309 err = -EFAULT;
1310 goto out;
1311 }
1312
1313 err = move_swap_pte(mm, dst_addr, src_addr,
1314 dst_pte, src_pte,
1315 orig_dst_pte, orig_src_pte,
1316 dst_ptl, src_ptl);
1317 }
1318
1319 out:
1320 if (src_anon_vma) {
1321 anon_vma_unlock_write(src_anon_vma);
1322 put_anon_vma(src_anon_vma);
1323 }
1324 if (src_folio) {
1325 folio_unlock(src_folio);
1326 folio_put(src_folio);
1327 }
1328 if (dst_pte)
1329 pte_unmap(dst_pte);
1330 if (src_pte)
1331 pte_unmap(src_pte);
1332 mmu_notifier_invalidate_range_end(&range);
1333
1334 return err;
1335 }
1336
1337 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1338 static inline bool move_splits_huge_pmd(unsigned long dst_addr,
1339 unsigned long src_addr,
1340 unsigned long src_end)
1341 {
1342 return (src_addr & ~HPAGE_PMD_MASK) || (dst_addr & ~HPAGE_PMD_MASK) ||
1343 src_end - src_addr < HPAGE_PMD_SIZE;
1344 }
1345 #else
1346 static inline bool move_splits_huge_pmd(unsigned long dst_addr,
1347 unsigned long src_addr,
1348 unsigned long src_end)
1349 {
1350 /* This is unreachable anyway, just to avoid warnings when HPAGE_PMD_SIZE==0 */
1351 return false;
1352 }
1353 #endif
1354
1355 static inline bool vma_move_compatible(struct vm_area_struct *vma)
1356 {
1357 return !(vma->vm_flags & (VM_PFNMAP | VM_IO | VM_HUGETLB |
1358 VM_MIXEDMAP | VM_SHADOW_STACK));
1359 }
1360
1361 static int validate_move_areas(struct userfaultfd_ctx *ctx,
1362 struct vm_area_struct *src_vma,
1363 struct vm_area_struct *dst_vma)
1364 {
1365 /* Only allow moving if both have the same access and protection */
1366 if ((src_vma->vm_flags & VM_ACCESS_FLAGS) != (dst_vma->vm_flags & VM_ACCESS_FLAGS) ||
1367 pgprot_val(src_vma->vm_page_prot) != pgprot_val(dst_vma->vm_page_prot))
1368 return -EINVAL;
1369
1370 /* Only allow moving if both are mlocked or both aren't */
1371 if ((src_vma->vm_flags & VM_LOCKED) != (dst_vma->vm_flags & VM_LOCKED))
1372 return -EINVAL;
1373
1374 /*
1375 * For now, we keep it simple and only move between writable VMAs.
1376 * Access flags are equal, therefore cheching only the source is enough.
1377 */
1378 if (!(src_vma->vm_flags & VM_WRITE))
1379 return -EINVAL;
1380
1381 /* Check if vma flags indicate content which can be moved */
1382 if (!vma_move_compatible(src_vma) || !vma_move_compatible(dst_vma))
1383 return -EINVAL;
1384
1385 /* Ensure dst_vma is registered in uffd we are operating on */
1386 if (!dst_vma->vm_userfaultfd_ctx.ctx ||
1387 dst_vma->vm_userfaultfd_ctx.ctx != ctx)
1388 return -EINVAL;
1389
1390 /* Only allow moving across anonymous vmas */
1391 if (!vma_is_anonymous(src_vma) || !vma_is_anonymous(dst_vma))
1392 return -EINVAL;
1393
1394 return 0;
1395 }
1396
1397 static __always_inline
1398 int find_vmas_mm_locked(struct mm_struct *mm,
1399 unsigned long dst_start,
1400 unsigned long src_start,
1401 struct vm_area_struct **dst_vmap,
1402 struct vm_area_struct **src_vmap)
1403 {
1404 struct vm_area_struct *vma;
1405
1406 mmap_assert_locked(mm);
1407 vma = find_vma_and_prepare_anon(mm, dst_start);
1408 if (IS_ERR(vma))
1409 return PTR_ERR(vma);
1410
1411 *dst_vmap = vma;
1412 /* Skip finding src_vma if src_start is in dst_vma */
1413 if (src_start >= vma->vm_start && src_start < vma->vm_end)
1414 goto out_success;
1415
1416 vma = vma_lookup(mm, src_start);
1417 if (!vma)
1418 return -ENOENT;
1419 out_success:
1420 *src_vmap = vma;
1421 return 0;
1422 }
1423
1424 #ifdef CONFIG_PER_VMA_LOCK
1425 static int uffd_move_lock(struct mm_struct *mm,
1426 unsigned long dst_start,
1427 unsigned long src_start,
1428 struct vm_area_struct **dst_vmap,
1429 struct vm_area_struct **src_vmap)
1430 {
1431 struct vm_area_struct *vma;
1432 int err;
1433
1434 vma = uffd_lock_vma(mm, dst_start);
1435 if (IS_ERR(vma))
1436 return PTR_ERR(vma);
1437
1438 *dst_vmap = vma;
1439 /*
1440 * Skip finding src_vma if src_start is in dst_vma. This also ensures
1441 * that we don't lock the same vma twice.
1442 */
1443 if (src_start >= vma->vm_start && src_start < vma->vm_end) {
1444 *src_vmap = vma;
1445 return 0;
1446 }
1447
1448 /*
1449 * Using uffd_lock_vma() to get src_vma can lead to following deadlock:
1450 *
1451 * Thread1 Thread2
1452 * ------- -------
1453 * vma_start_read(dst_vma)
1454 * mmap_write_lock(mm)
1455 * vma_start_write(src_vma)
1456 * vma_start_read(src_vma)
1457 * mmap_read_lock(mm)
1458 * vma_start_write(dst_vma)
1459 */
1460 *src_vmap = lock_vma_under_rcu(mm, src_start);
1461 if (likely(*src_vmap))
1462 return 0;
1463
1464 /* Undo any locking and retry in mmap_lock critical section */
1465 vma_end_read(*dst_vmap);
1466
1467 mmap_read_lock(mm);
1468 err = find_vmas_mm_locked(mm, dst_start, src_start, dst_vmap, src_vmap);
1469 if (!err) {
1470 /*
1471 * See comment in uffd_lock_vma() as to why not using
1472 * vma_start_read() here.
1473 */
1474 down_read(&(*dst_vmap)->vm_lock->lock);
1475 if (*dst_vmap != *src_vmap)
1476 down_read_nested(&(*src_vmap)->vm_lock->lock,
1477 SINGLE_DEPTH_NESTING);
1478 }
1479 mmap_read_unlock(mm);
1480 return err;
1481 }
1482
1483 static void uffd_move_unlock(struct vm_area_struct *dst_vma,
1484 struct vm_area_struct *src_vma)
1485 {
1486 vma_end_read(src_vma);
1487 if (src_vma != dst_vma)
1488 vma_end_read(dst_vma);
1489 }
1490
1491 #else
1492
1493 static int uffd_move_lock(struct mm_struct *mm,
1494 unsigned long dst_start,
1495 unsigned long src_start,
1496 struct vm_area_struct **dst_vmap,
1497 struct vm_area_struct **src_vmap)
1498 {
1499 int err;
1500
1501 mmap_read_lock(mm);
1502 err = find_vmas_mm_locked(mm, dst_start, src_start, dst_vmap, src_vmap);
1503 if (err)
1504 mmap_read_unlock(mm);
1505 return err;
1506 }
1507
1508 static void uffd_move_unlock(struct vm_area_struct *dst_vma,
1509 struct vm_area_struct *src_vma)
1510 {
1511 mmap_assert_locked(src_vma->vm_mm);
1512 mmap_read_unlock(dst_vma->vm_mm);
1513 }
1514 #endif
1515
1516 /**
1517 * move_pages - move arbitrary anonymous pages of an existing vma
1518 * @ctx: pointer to the userfaultfd context
1519 * @dst_start: start of the destination virtual memory range
1520 * @src_start: start of the source virtual memory range
1521 * @len: length of the virtual memory range
1522 * @mode: flags from uffdio_move.mode
1523 *
1524 * It will either use the mmap_lock in read mode or per-vma locks
1525 *
1526 * move_pages() remaps arbitrary anonymous pages atomically in zero
1527 * copy. It only works on non shared anonymous pages because those can
1528 * be relocated without generating non linear anon_vmas in the rmap
1529 * code.
1530 *
1531 * It provides a zero copy mechanism to handle userspace page faults.
1532 * The source vma pages should have mapcount == 1, which can be
1533 * enforced by using madvise(MADV_DONTFORK) on src vma.
1534 *
1535 * The thread receiving the page during the userland page fault
1536 * will receive the faulting page in the source vma through the network,
1537 * storage or any other I/O device (MADV_DONTFORK in the source vma
1538 * avoids move_pages() to fail with -EBUSY if the process forks before
1539 * move_pages() is called), then it will call move_pages() to map the
1540 * page in the faulting address in the destination vma.
1541 *
1542 * This userfaultfd command works purely via pagetables, so it's the
1543 * most efficient way to move physical non shared anonymous pages
1544 * across different virtual addresses. Unlike mremap()/mmap()/munmap()
1545 * it does not create any new vmas. The mapping in the destination
1546 * address is atomic.
1547 *
1548 * It only works if the vma protection bits are identical from the
1549 * source and destination vma.
1550 *
1551 * It can remap non shared anonymous pages within the same vma too.
1552 *
1553 * If the source virtual memory range has any unmapped holes, or if
1554 * the destination virtual memory range is not a whole unmapped hole,
1555 * move_pages() will fail respectively with -ENOENT or -EEXIST. This
1556 * provides a very strict behavior to avoid any chance of memory
1557 * corruption going unnoticed if there are userland race conditions.
1558 * Only one thread should resolve the userland page fault at any given
1559 * time for any given faulting address. This means that if two threads
1560 * try to both call move_pages() on the same destination address at the
1561 * same time, the second thread will get an explicit error from this
1562 * command.
1563 *
1564 * The command retval will return "len" is successful. The command
1565 * however can be interrupted by fatal signals or errors. If
1566 * interrupted it will return the number of bytes successfully
1567 * remapped before the interruption if any, or the negative error if
1568 * none. It will never return zero. Either it will return an error or
1569 * an amount of bytes successfully moved. If the retval reports a
1570 * "short" remap, the move_pages() command should be repeated by
1571 * userland with src+retval, dst+reval, len-retval if it wants to know
1572 * about the error that interrupted it.
1573 *
1574 * The UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES flag can be specified to
1575 * prevent -ENOENT errors to materialize if there are holes in the
1576 * source virtual range that is being remapped. The holes will be
1577 * accounted as successfully remapped in the retval of the
1578 * command. This is mostly useful to remap hugepage naturally aligned
1579 * virtual regions without knowing if there are transparent hugepage
1580 * in the regions or not, but preventing the risk of having to split
1581 * the hugepmd during the remap.
1582 *
1583 * If there's any rmap walk that is taking the anon_vma locks without
1584 * first obtaining the folio lock (the only current instance is
1585 * folio_referenced), they will have to verify if the folio->mapping
1586 * has changed after taking the anon_vma lock. If it changed they
1587 * should release the lock and retry obtaining a new anon_vma, because
1588 * it means the anon_vma was changed by move_pages() before the lock
1589 * could be obtained. This is the only additional complexity added to
1590 * the rmap code to provide this anonymous page remapping functionality.
1591 */
1592 ssize_t move_pages(struct userfaultfd_ctx *ctx, unsigned long dst_start,
1593 unsigned long src_start, unsigned long len, __u64 mode)
1594 {
1595 struct mm_struct *mm = ctx->mm;
1596 struct vm_area_struct *src_vma, *dst_vma;
1597 unsigned long src_addr, dst_addr;
1598 pmd_t *src_pmd, *dst_pmd;
1599 long err = -EINVAL;
1600 ssize_t moved = 0;
1601
1602 /* Sanitize the command parameters. */
1603 if (WARN_ON_ONCE(src_start & ~PAGE_MASK) ||
1604 WARN_ON_ONCE(dst_start & ~PAGE_MASK) ||
1605 WARN_ON_ONCE(len & ~PAGE_MASK))
1606 goto out;
1607
1608 /* Does the address range wrap, or is the span zero-sized? */
1609 if (WARN_ON_ONCE(src_start + len <= src_start) ||
1610 WARN_ON_ONCE(dst_start + len <= dst_start))
1611 goto out;
1612
1613 err = uffd_move_lock(mm, dst_start, src_start, &dst_vma, &src_vma);
1614 if (err)
1615 goto out;
1616
1617 /* Re-check after taking map_changing_lock */
1618 err = -EAGAIN;
1619 down_read(&ctx->map_changing_lock);
1620 if (likely(atomic_read(&ctx->mmap_changing)))
1621 goto out_unlock;
1622 /*
1623 * Make sure the vma is not shared, that the src and dst remap
1624 * ranges are both valid and fully within a single existing
1625 * vma.
1626 */
1627 err = -EINVAL;
1628 if (src_vma->vm_flags & VM_SHARED)
1629 goto out_unlock;
1630 if (src_start + len > src_vma->vm_end)
1631 goto out_unlock;
1632
1633 if (dst_vma->vm_flags & VM_SHARED)
1634 goto out_unlock;
1635 if (dst_start + len > dst_vma->vm_end)
1636 goto out_unlock;
1637
1638 err = validate_move_areas(ctx, src_vma, dst_vma);
1639 if (err)
1640 goto out_unlock;
1641
1642 for (src_addr = src_start, dst_addr = dst_start;
1643 src_addr < src_start + len;) {
1644 spinlock_t *ptl;
1645 pmd_t dst_pmdval;
1646 unsigned long step_size;
1647
1648 /*
1649 * Below works because anonymous area would not have a
1650 * transparent huge PUD. If file-backed support is added,
1651 * that case would need to be handled here.
1652 */
1653 src_pmd = mm_find_pmd(mm, src_addr);
1654 if (unlikely(!src_pmd)) {
1655 if (!(mode & UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES)) {
1656 err = -ENOENT;
1657 break;
1658 }
1659 src_pmd = mm_alloc_pmd(mm, src_addr);
1660 if (unlikely(!src_pmd)) {
1661 err = -ENOMEM;
1662 break;
1663 }
1664 }
1665 dst_pmd = mm_alloc_pmd(mm, dst_addr);
1666 if (unlikely(!dst_pmd)) {
1667 err = -ENOMEM;
1668 break;
1669 }
1670
1671 dst_pmdval = pmdp_get_lockless(dst_pmd);
1672 /*
1673 * If the dst_pmd is mapped as THP don't override it and just
1674 * be strict. If dst_pmd changes into TPH after this check, the
1675 * move_pages_huge_pmd() will detect the change and retry
1676 * while move_pages_pte() will detect the change and fail.
1677 */
1678 if (unlikely(pmd_trans_huge(dst_pmdval))) {
1679 err = -EEXIST;
1680 break;
1681 }
1682
1683 ptl = pmd_trans_huge_lock(src_pmd, src_vma);
1684 if (ptl) {
1685 if (pmd_devmap(*src_pmd)) {
1686 spin_unlock(ptl);
1687 err = -ENOENT;
1688 break;
1689 }
1690
1691 /* Check if we can move the pmd without splitting it. */
1692 if (move_splits_huge_pmd(dst_addr, src_addr, src_start + len) ||
1693 !pmd_none(dst_pmdval)) {
1694 struct folio *folio = pmd_folio(*src_pmd);
1695
1696 if (!folio || (!is_huge_zero_folio(folio) &&
1697 !PageAnonExclusive(&folio->page))) {
1698 spin_unlock(ptl);
1699 err = -EBUSY;
1700 break;
1701 }
1702
1703 spin_unlock(ptl);
1704 split_huge_pmd(src_vma, src_pmd, src_addr);
1705 /* The folio will be split by move_pages_pte() */
1706 continue;
1707 }
1708
1709 err = move_pages_huge_pmd(mm, dst_pmd, src_pmd,
1710 dst_pmdval, dst_vma, src_vma,
1711 dst_addr, src_addr);
1712 step_size = HPAGE_PMD_SIZE;
1713 } else {
1714 if (pmd_none(*src_pmd)) {
1715 if (!(mode & UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES)) {
1716 err = -ENOENT;
1717 break;
1718 }
1719 if (unlikely(__pte_alloc(mm, src_pmd))) {
1720 err = -ENOMEM;
1721 break;
1722 }
1723 }
1724
1725 if (unlikely(pte_alloc(mm, dst_pmd))) {
1726 err = -ENOMEM;
1727 break;
1728 }
1729
1730 err = move_pages_pte(mm, dst_pmd, src_pmd,
1731 dst_vma, src_vma,
1732 dst_addr, src_addr, mode);
1733 step_size = PAGE_SIZE;
1734 }
1735
1736 cond_resched();
1737
1738 if (fatal_signal_pending(current)) {
1739 /* Do not override an error */
1740 if (!err || err == -EAGAIN)
1741 err = -EINTR;
1742 break;
1743 }
1744
1745 if (err) {
1746 if (err == -EAGAIN)
1747 continue;
1748 break;
1749 }
1750
1751 /* Proceed to the next page */
1752 dst_addr += step_size;
1753 src_addr += step_size;
1754 moved += step_size;
1755 }
1756
1757 out_unlock:
1758 up_read(&ctx->map_changing_lock);
1759 uffd_move_unlock(dst_vma, src_vma);
1760 out:
1761 VM_WARN_ON(moved < 0);
1762 VM_WARN_ON(err > 0);
1763 VM_WARN_ON(!moved && !err);
1764 return moved ? moved : err;
1765 }
1766
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.