Loading...
Note: File does not exist in v3.1.
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 "internal.h"
20
21static __always_inline
22struct vm_area_struct *find_dst_vma(struct mm_struct *dst_mm,
23 unsigned long dst_start,
24 unsigned long len)
25{
26 /*
27 * Make sure that the dst range is both valid and fully within a
28 * single existing vma.
29 */
30 struct vm_area_struct *dst_vma;
31
32 dst_vma = find_vma(dst_mm, dst_start);
33 if (!dst_vma)
34 return NULL;
35
36 if (dst_start < dst_vma->vm_start ||
37 dst_start + len > dst_vma->vm_end)
38 return NULL;
39
40 /*
41 * Check the vma is registered in uffd, this is required to
42 * enforce the VM_MAYWRITE check done at uffd registration
43 * time.
44 */
45 if (!dst_vma->vm_userfaultfd_ctx.ctx)
46 return NULL;
47
48 return dst_vma;
49}
50
51static int mcopy_atomic_pte(struct mm_struct *dst_mm,
52 pmd_t *dst_pmd,
53 struct vm_area_struct *dst_vma,
54 unsigned long dst_addr,
55 unsigned long src_addr,
56 struct page **pagep,
57 bool wp_copy)
58{
59 pte_t _dst_pte, *dst_pte;
60 spinlock_t *ptl;
61 void *page_kaddr;
62 int ret;
63 struct page *page;
64 pgoff_t offset, max_off;
65 struct inode *inode;
66
67 if (!*pagep) {
68 ret = -ENOMEM;
69 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
70 if (!page)
71 goto out;
72
73 page_kaddr = kmap_atomic(page);
74 ret = copy_from_user(page_kaddr,
75 (const void __user *) src_addr,
76 PAGE_SIZE);
77 kunmap_atomic(page_kaddr);
78
79 /* fallback to copy_from_user outside mmap_lock */
80 if (unlikely(ret)) {
81 ret = -ENOENT;
82 *pagep = page;
83 /* don't free the page */
84 goto out;
85 }
86 } else {
87 page = *pagep;
88 *pagep = NULL;
89 }
90
91 /*
92 * The memory barrier inside __SetPageUptodate makes sure that
93 * preceding stores to the page contents become visible before
94 * the set_pte_at() write.
95 */
96 __SetPageUptodate(page);
97
98 ret = -ENOMEM;
99 if (mem_cgroup_charge(page, dst_mm, GFP_KERNEL))
100 goto out_release;
101
102 _dst_pte = pte_mkdirty(mk_pte(page, dst_vma->vm_page_prot));
103 if (dst_vma->vm_flags & VM_WRITE) {
104 if (wp_copy)
105 _dst_pte = pte_mkuffd_wp(_dst_pte);
106 else
107 _dst_pte = pte_mkwrite(_dst_pte);
108 }
109
110 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
111 if (dst_vma->vm_file) {
112 /* the shmem MAP_PRIVATE case requires checking the i_size */
113 inode = dst_vma->vm_file->f_inode;
114 offset = linear_page_index(dst_vma, dst_addr);
115 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
116 ret = -EFAULT;
117 if (unlikely(offset >= max_off))
118 goto out_release_uncharge_unlock;
119 }
120 ret = -EEXIST;
121 if (!pte_none(*dst_pte))
122 goto out_release_uncharge_unlock;
123
124 inc_mm_counter(dst_mm, MM_ANONPAGES);
125 page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
126 lru_cache_add_inactive_or_unevictable(page, dst_vma);
127
128 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
129
130 /* No need to invalidate - it was non-present before */
131 update_mmu_cache(dst_vma, dst_addr, dst_pte);
132
133 pte_unmap_unlock(dst_pte, ptl);
134 ret = 0;
135out:
136 return ret;
137out_release_uncharge_unlock:
138 pte_unmap_unlock(dst_pte, ptl);
139out_release:
140 put_page(page);
141 goto out;
142}
143
144static int mfill_zeropage_pte(struct mm_struct *dst_mm,
145 pmd_t *dst_pmd,
146 struct vm_area_struct *dst_vma,
147 unsigned long dst_addr)
148{
149 pte_t _dst_pte, *dst_pte;
150 spinlock_t *ptl;
151 int ret;
152 pgoff_t offset, max_off;
153 struct inode *inode;
154
155 _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
156 dst_vma->vm_page_prot));
157 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
158 if (dst_vma->vm_file) {
159 /* the shmem MAP_PRIVATE case requires checking the i_size */
160 inode = dst_vma->vm_file->f_inode;
161 offset = linear_page_index(dst_vma, dst_addr);
162 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
163 ret = -EFAULT;
164 if (unlikely(offset >= max_off))
165 goto out_unlock;
166 }
167 ret = -EEXIST;
168 if (!pte_none(*dst_pte))
169 goto out_unlock;
170 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
171 /* No need to invalidate - it was non-present before */
172 update_mmu_cache(dst_vma, dst_addr, dst_pte);
173 ret = 0;
174out_unlock:
175 pte_unmap_unlock(dst_pte, ptl);
176 return ret;
177}
178
179static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
180{
181 pgd_t *pgd;
182 p4d_t *p4d;
183 pud_t *pud;
184
185 pgd = pgd_offset(mm, address);
186 p4d = p4d_alloc(mm, pgd, address);
187 if (!p4d)
188 return NULL;
189 pud = pud_alloc(mm, p4d, address);
190 if (!pud)
191 return NULL;
192 /*
193 * Note that we didn't run this because the pmd was
194 * missing, the *pmd may be already established and in
195 * turn it may also be a trans_huge_pmd.
196 */
197 return pmd_alloc(mm, pud, address);
198}
199
200#ifdef CONFIG_HUGETLB_PAGE
201/*
202 * __mcopy_atomic processing for HUGETLB vmas. Note that this routine is
203 * called with mmap_lock held, it will release mmap_lock before returning.
204 */
205static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
206 struct vm_area_struct *dst_vma,
207 unsigned long dst_start,
208 unsigned long src_start,
209 unsigned long len,
210 bool zeropage)
211{
212 int vm_alloc_shared = dst_vma->vm_flags & VM_SHARED;
213 int vm_shared = dst_vma->vm_flags & VM_SHARED;
214 ssize_t err;
215 pte_t *dst_pte;
216 unsigned long src_addr, dst_addr;
217 long copied;
218 struct page *page;
219 unsigned long vma_hpagesize;
220 pgoff_t idx;
221 u32 hash;
222 struct address_space *mapping;
223
224 /*
225 * There is no default zero huge page for all huge page sizes as
226 * supported by hugetlb. A PMD_SIZE huge pages may exist as used
227 * by THP. Since we can not reliably insert a zero page, this
228 * feature is not supported.
229 */
230 if (zeropage) {
231 mmap_read_unlock(dst_mm);
232 return -EINVAL;
233 }
234
235 src_addr = src_start;
236 dst_addr = dst_start;
237 copied = 0;
238 page = NULL;
239 vma_hpagesize = vma_kernel_pagesize(dst_vma);
240
241 /*
242 * Validate alignment based on huge page size
243 */
244 err = -EINVAL;
245 if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
246 goto out_unlock;
247
248retry:
249 /*
250 * On routine entry dst_vma is set. If we had to drop mmap_lock and
251 * retry, dst_vma will be set to NULL and we must lookup again.
252 */
253 if (!dst_vma) {
254 err = -ENOENT;
255 dst_vma = find_dst_vma(dst_mm, dst_start, len);
256 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
257 goto out_unlock;
258
259 err = -EINVAL;
260 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
261 goto out_unlock;
262
263 vm_shared = dst_vma->vm_flags & VM_SHARED;
264 }
265
266 /*
267 * If not shared, ensure the dst_vma has a anon_vma.
268 */
269 err = -ENOMEM;
270 if (!vm_shared) {
271 if (unlikely(anon_vma_prepare(dst_vma)))
272 goto out_unlock;
273 }
274
275 while (src_addr < src_start + len) {
276 pte_t dst_pteval;
277
278 BUG_ON(dst_addr >= dst_start + len);
279
280 /*
281 * Serialize via i_mmap_rwsem and hugetlb_fault_mutex.
282 * i_mmap_rwsem ensures the dst_pte remains valid even
283 * in the case of shared pmds. fault mutex prevents
284 * races with other faulting threads.
285 */
286 mapping = dst_vma->vm_file->f_mapping;
287 i_mmap_lock_read(mapping);
288 idx = linear_page_index(dst_vma, dst_addr);
289 hash = hugetlb_fault_mutex_hash(mapping, idx);
290 mutex_lock(&hugetlb_fault_mutex_table[hash]);
291
292 err = -ENOMEM;
293 dst_pte = huge_pte_alloc(dst_mm, dst_addr, vma_hpagesize);
294 if (!dst_pte) {
295 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
296 i_mmap_unlock_read(mapping);
297 goto out_unlock;
298 }
299
300 err = -EEXIST;
301 dst_pteval = huge_ptep_get(dst_pte);
302 if (!huge_pte_none(dst_pteval)) {
303 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
304 i_mmap_unlock_read(mapping);
305 goto out_unlock;
306 }
307
308 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
309 dst_addr, src_addr, &page);
310
311 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
312 i_mmap_unlock_read(mapping);
313 vm_alloc_shared = vm_shared;
314
315 cond_resched();
316
317 if (unlikely(err == -ENOENT)) {
318 mmap_read_unlock(dst_mm);
319 BUG_ON(!page);
320
321 err = copy_huge_page_from_user(page,
322 (const void __user *)src_addr,
323 vma_hpagesize / PAGE_SIZE,
324 true);
325 if (unlikely(err)) {
326 err = -EFAULT;
327 goto out;
328 }
329 mmap_read_lock(dst_mm);
330
331 dst_vma = NULL;
332 goto retry;
333 } else
334 BUG_ON(page);
335
336 if (!err) {
337 dst_addr += vma_hpagesize;
338 src_addr += vma_hpagesize;
339 copied += vma_hpagesize;
340
341 if (fatal_signal_pending(current))
342 err = -EINTR;
343 }
344 if (err)
345 break;
346 }
347
348out_unlock:
349 mmap_read_unlock(dst_mm);
350out:
351 if (page) {
352 /*
353 * We encountered an error and are about to free a newly
354 * allocated huge page.
355 *
356 * Reservation handling is very subtle, and is different for
357 * private and shared mappings. See the routine
358 * restore_reserve_on_error for details. Unfortunately, we
359 * can not call restore_reserve_on_error now as it would
360 * require holding mmap_lock.
361 *
362 * If a reservation for the page existed in the reservation
363 * map of a private mapping, the map was modified to indicate
364 * the reservation was consumed when the page was allocated.
365 * We clear the PagePrivate flag now so that the global
366 * reserve count will not be incremented in free_huge_page.
367 * The reservation map will still indicate the reservation
368 * was consumed and possibly prevent later page allocation.
369 * This is better than leaking a global reservation. If no
370 * reservation existed, it is still safe to clear PagePrivate
371 * as no adjustments to reservation counts were made during
372 * allocation.
373 *
374 * The reservation map for shared mappings indicates which
375 * pages have reservations. When a huge page is allocated
376 * for an address with a reservation, no change is made to
377 * the reserve map. In this case PagePrivate will be set
378 * to indicate that the global reservation count should be
379 * incremented when the page is freed. This is the desired
380 * behavior. However, when a huge page is allocated for an
381 * address without a reservation a reservation entry is added
382 * to the reservation map, and PagePrivate will not be set.
383 * When the page is freed, the global reserve count will NOT
384 * be incremented and it will appear as though we have leaked
385 * reserved page. In this case, set PagePrivate so that the
386 * global reserve count will be incremented to match the
387 * reservation map entry which was created.
388 *
389 * Note that vm_alloc_shared is based on the flags of the vma
390 * for which the page was originally allocated. dst_vma could
391 * be different or NULL on error.
392 */
393 if (vm_alloc_shared)
394 SetPagePrivate(page);
395 else
396 ClearPagePrivate(page);
397 put_page(page);
398 }
399 BUG_ON(copied < 0);
400 BUG_ON(err > 0);
401 BUG_ON(!copied && !err);
402 return copied ? copied : err;
403}
404#else /* !CONFIG_HUGETLB_PAGE */
405/* fail at build time if gcc attempts to use this */
406extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
407 struct vm_area_struct *dst_vma,
408 unsigned long dst_start,
409 unsigned long src_start,
410 unsigned long len,
411 bool zeropage);
412#endif /* CONFIG_HUGETLB_PAGE */
413
414static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
415 pmd_t *dst_pmd,
416 struct vm_area_struct *dst_vma,
417 unsigned long dst_addr,
418 unsigned long src_addr,
419 struct page **page,
420 bool zeropage,
421 bool wp_copy)
422{
423 ssize_t err;
424
425 /*
426 * The normal page fault path for a shmem will invoke the
427 * fault, fill the hole in the file and COW it right away. The
428 * result generates plain anonymous memory. So when we are
429 * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
430 * generate anonymous memory directly without actually filling
431 * the hole. For the MAP_PRIVATE case the robustness check
432 * only happens in the pagetable (to verify it's still none)
433 * and not in the radix tree.
434 */
435 if (!(dst_vma->vm_flags & VM_SHARED)) {
436 if (!zeropage)
437 err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
438 dst_addr, src_addr, page,
439 wp_copy);
440 else
441 err = mfill_zeropage_pte(dst_mm, dst_pmd,
442 dst_vma, dst_addr);
443 } else {
444 VM_WARN_ON_ONCE(wp_copy);
445 if (!zeropage)
446 err = shmem_mcopy_atomic_pte(dst_mm, dst_pmd,
447 dst_vma, dst_addr,
448 src_addr, page);
449 else
450 err = shmem_mfill_zeropage_pte(dst_mm, dst_pmd,
451 dst_vma, dst_addr);
452 }
453
454 return err;
455}
456
457static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
458 unsigned long dst_start,
459 unsigned long src_start,
460 unsigned long len,
461 bool zeropage,
462 bool *mmap_changing,
463 __u64 mode)
464{
465 struct vm_area_struct *dst_vma;
466 ssize_t err;
467 pmd_t *dst_pmd;
468 unsigned long src_addr, dst_addr;
469 long copied;
470 struct page *page;
471 bool wp_copy;
472
473 /*
474 * Sanitize the command parameters:
475 */
476 BUG_ON(dst_start & ~PAGE_MASK);
477 BUG_ON(len & ~PAGE_MASK);
478
479 /* Does the address range wrap, or is the span zero-sized? */
480 BUG_ON(src_start + len <= src_start);
481 BUG_ON(dst_start + len <= dst_start);
482
483 src_addr = src_start;
484 dst_addr = dst_start;
485 copied = 0;
486 page = NULL;
487retry:
488 mmap_read_lock(dst_mm);
489
490 /*
491 * If memory mappings are changing because of non-cooperative
492 * operation (e.g. mremap) running in parallel, bail out and
493 * request the user to retry later
494 */
495 err = -EAGAIN;
496 if (mmap_changing && READ_ONCE(*mmap_changing))
497 goto out_unlock;
498
499 /*
500 * Make sure the vma is not shared, that the dst range is
501 * both valid and fully within a single existing vma.
502 */
503 err = -ENOENT;
504 dst_vma = find_dst_vma(dst_mm, dst_start, len);
505 if (!dst_vma)
506 goto out_unlock;
507
508 err = -EINVAL;
509 /*
510 * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
511 * it will overwrite vm_ops, so vma_is_anonymous must return false.
512 */
513 if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
514 dst_vma->vm_flags & VM_SHARED))
515 goto out_unlock;
516
517 /*
518 * validate 'mode' now that we know the dst_vma: don't allow
519 * a wrprotect copy if the userfaultfd didn't register as WP.
520 */
521 wp_copy = mode & UFFDIO_COPY_MODE_WP;
522 if (wp_copy && !(dst_vma->vm_flags & VM_UFFD_WP))
523 goto out_unlock;
524
525 /*
526 * If this is a HUGETLB vma, pass off to appropriate routine
527 */
528 if (is_vm_hugetlb_page(dst_vma))
529 return __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
530 src_start, len, zeropage);
531
532 if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
533 goto out_unlock;
534
535 /*
536 * Ensure the dst_vma has a anon_vma or this page
537 * would get a NULL anon_vma when moved in the
538 * dst_vma.
539 */
540 err = -ENOMEM;
541 if (!(dst_vma->vm_flags & VM_SHARED) &&
542 unlikely(anon_vma_prepare(dst_vma)))
543 goto out_unlock;
544
545 while (src_addr < src_start + len) {
546 pmd_t dst_pmdval;
547
548 BUG_ON(dst_addr >= dst_start + len);
549
550 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
551 if (unlikely(!dst_pmd)) {
552 err = -ENOMEM;
553 break;
554 }
555
556 dst_pmdval = pmd_read_atomic(dst_pmd);
557 /*
558 * If the dst_pmd is mapped as THP don't
559 * override it and just be strict.
560 */
561 if (unlikely(pmd_trans_huge(dst_pmdval))) {
562 err = -EEXIST;
563 break;
564 }
565 if (unlikely(pmd_none(dst_pmdval)) &&
566 unlikely(__pte_alloc(dst_mm, dst_pmd))) {
567 err = -ENOMEM;
568 break;
569 }
570 /* If an huge pmd materialized from under us fail */
571 if (unlikely(pmd_trans_huge(*dst_pmd))) {
572 err = -EFAULT;
573 break;
574 }
575
576 BUG_ON(pmd_none(*dst_pmd));
577 BUG_ON(pmd_trans_huge(*dst_pmd));
578
579 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
580 src_addr, &page, zeropage, wp_copy);
581 cond_resched();
582
583 if (unlikely(err == -ENOENT)) {
584 void *page_kaddr;
585
586 mmap_read_unlock(dst_mm);
587 BUG_ON(!page);
588
589 page_kaddr = kmap(page);
590 err = copy_from_user(page_kaddr,
591 (const void __user *) src_addr,
592 PAGE_SIZE);
593 kunmap(page);
594 if (unlikely(err)) {
595 err = -EFAULT;
596 goto out;
597 }
598 goto retry;
599 } else
600 BUG_ON(page);
601
602 if (!err) {
603 dst_addr += PAGE_SIZE;
604 src_addr += PAGE_SIZE;
605 copied += PAGE_SIZE;
606
607 if (fatal_signal_pending(current))
608 err = -EINTR;
609 }
610 if (err)
611 break;
612 }
613
614out_unlock:
615 mmap_read_unlock(dst_mm);
616out:
617 if (page)
618 put_page(page);
619 BUG_ON(copied < 0);
620 BUG_ON(err > 0);
621 BUG_ON(!copied && !err);
622 return copied ? copied : err;
623}
624
625ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
626 unsigned long src_start, unsigned long len,
627 bool *mmap_changing, __u64 mode)
628{
629 return __mcopy_atomic(dst_mm, dst_start, src_start, len, false,
630 mmap_changing, mode);
631}
632
633ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
634 unsigned long len, bool *mmap_changing)
635{
636 return __mcopy_atomic(dst_mm, start, 0, len, true, mmap_changing, 0);
637}
638
639int mwriteprotect_range(struct mm_struct *dst_mm, unsigned long start,
640 unsigned long len, bool enable_wp, bool *mmap_changing)
641{
642 struct vm_area_struct *dst_vma;
643 pgprot_t newprot;
644 int err;
645
646 /*
647 * Sanitize the command parameters:
648 */
649 BUG_ON(start & ~PAGE_MASK);
650 BUG_ON(len & ~PAGE_MASK);
651
652 /* Does the address range wrap, or is the span zero-sized? */
653 BUG_ON(start + len <= start);
654
655 mmap_read_lock(dst_mm);
656
657 /*
658 * If memory mappings are changing because of non-cooperative
659 * operation (e.g. mremap) running in parallel, bail out and
660 * request the user to retry later
661 */
662 err = -EAGAIN;
663 if (mmap_changing && READ_ONCE(*mmap_changing))
664 goto out_unlock;
665
666 err = -ENOENT;
667 dst_vma = find_dst_vma(dst_mm, start, len);
668 /*
669 * Make sure the vma is not shared, that the dst range is
670 * both valid and fully within a single existing vma.
671 */
672 if (!dst_vma || (dst_vma->vm_flags & VM_SHARED))
673 goto out_unlock;
674 if (!userfaultfd_wp(dst_vma))
675 goto out_unlock;
676 if (!vma_is_anonymous(dst_vma))
677 goto out_unlock;
678
679 if (enable_wp)
680 newprot = vm_get_page_prot(dst_vma->vm_flags & ~(VM_WRITE));
681 else
682 newprot = vm_get_page_prot(dst_vma->vm_flags);
683
684 change_protection(dst_vma, start, start + len, newprot,
685 enable_wp ? MM_CP_UFFD_WP : MM_CP_UFFD_WP_RESOLVE);
686
687 err = 0;
688out_unlock:
689 mmap_read_unlock(dst_mm);
690 return err;
691}