projects/linux·blob v5.7 · refs

694 lines · 17.8 KiB

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 struct mem_cgroup *memcg;
60 pte_t _dst_pte, *dst_pte;
61 spinlock_t *ptl;
62 void *page_kaddr;
63 int ret;
64 struct page *page;
65 pgoff_t offset, max_off;
66 struct inode *inode;
67
68 if (!*pagep) {
69 ret = -ENOMEM;
70 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
71 if (!page)
72 goto out;
73
74 page_kaddr = kmap_atomic(page);
75 ret = copy_from_user(page_kaddr,
76 (const void __user *) src_addr,
77 PAGE_SIZE);
78 kunmap_atomic(page_kaddr);
79
80 /* fallback to copy_from_user outside mmap_sem */
81 if (unlikely(ret)) {
82 ret = -ENOENT;
83 *pagep = page;
84 /* don't free the page */
85 goto out;
86 }
87 } else {
88 page = *pagep;
89 *pagep = NULL;
90 }
91
92 /*
93 * The memory barrier inside __SetPageUptodate makes sure that
94 * preceding stores to the page contents become visible before
95 * the set_pte_at() write.
96 */
97 __SetPageUptodate(page);
98
99 ret = -ENOMEM;
100 if (mem_cgroup_try_charge(page, dst_mm, GFP_KERNEL, &memcg, false))
101 goto out_release;
102
103 _dst_pte = pte_mkdirty(mk_pte(page, dst_vma->vm_page_prot));
104 if (dst_vma->vm_flags & VM_WRITE) {
105 if (wp_copy)
106 _dst_pte = pte_mkuffd_wp(_dst_pte);
107 else
108 _dst_pte = pte_mkwrite(_dst_pte);
109 }
110
111 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
112 if (dst_vma->vm_file) {
113 /* the shmem MAP_PRIVATE case requires checking the i_size */
114 inode = dst_vma->vm_file->f_inode;
115 offset = linear_page_index(dst_vma, dst_addr);
116 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
117 ret = -EFAULT;
118 if (unlikely(offset >= max_off))
119 goto out_release_uncharge_unlock;
120 }
121 ret = -EEXIST;
122 if (!pte_none(*dst_pte))
123 goto out_release_uncharge_unlock;
124
125 inc_mm_counter(dst_mm, MM_ANONPAGES);
126 page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
127 mem_cgroup_commit_charge(page, memcg, false, false);
128 lru_cache_add_active_or_unevictable(page, dst_vma);
129
130 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
131
132 /* No need to invalidate - it was non-present before */
133 update_mmu_cache(dst_vma, dst_addr, dst_pte);
134
135 pte_unmap_unlock(dst_pte, ptl);
136 ret = 0;
137out:
138 return ret;
139out_release_uncharge_unlock:
140 pte_unmap_unlock(dst_pte, ptl);
141 mem_cgroup_cancel_charge(page, memcg, false);
142out_release:
143 put_page(page);
144 goto out;
145}
146
147static int mfill_zeropage_pte(struct mm_struct *dst_mm,
148 pmd_t *dst_pmd,
149 struct vm_area_struct *dst_vma,
150 unsigned long dst_addr)
151{
152 pte_t _dst_pte, *dst_pte;
153 spinlock_t *ptl;
154 int ret;
155 pgoff_t offset, max_off;
156 struct inode *inode;
157
158 _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
159 dst_vma->vm_page_prot));
160 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
161 if (dst_vma->vm_file) {
162 /* the shmem MAP_PRIVATE case requires checking the i_size */
163 inode = dst_vma->vm_file->f_inode;
164 offset = linear_page_index(dst_vma, dst_addr);
165 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
166 ret = -EFAULT;
167 if (unlikely(offset >= max_off))
168 goto out_unlock;
169 }
170 ret = -EEXIST;
171 if (!pte_none(*dst_pte))
172 goto out_unlock;
173 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
174 /* No need to invalidate - it was non-present before */
175 update_mmu_cache(dst_vma, dst_addr, dst_pte);
176 ret = 0;
177out_unlock:
178 pte_unmap_unlock(dst_pte, ptl);
179 return ret;
180}
181
182static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
183{
184 pgd_t *pgd;
185 p4d_t *p4d;
186 pud_t *pud;
187
188 pgd = pgd_offset(mm, address);
189 p4d = p4d_alloc(mm, pgd, address);
190 if (!p4d)
191 return NULL;
192 pud = pud_alloc(mm, p4d, address);
193 if (!pud)
194 return NULL;
195 /*
196 * Note that we didn't run this because the pmd was
197 * missing, the *pmd may be already established and in
198 * turn it may also be a trans_huge_pmd.
199 */
200 return pmd_alloc(mm, pud, address);
201}
202
203#ifdef CONFIG_HUGETLB_PAGE
204/*
205 * __mcopy_atomic processing for HUGETLB vmas. Note that this routine is
206 * called with mmap_sem held, it will release mmap_sem before returning.
207 */
208static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
209 struct vm_area_struct *dst_vma,
210 unsigned long dst_start,
211 unsigned long src_start,
212 unsigned long len,
213 bool zeropage)
214{
215 int vm_alloc_shared = dst_vma->vm_flags & VM_SHARED;
216 int vm_shared = dst_vma->vm_flags & VM_SHARED;
217 ssize_t err;
218 pte_t *dst_pte;
219 unsigned long src_addr, dst_addr;
220 long copied;
221 struct page *page;
222 unsigned long vma_hpagesize;
223 pgoff_t idx;
224 u32 hash;
225 struct address_space *mapping;
226
227 /*
228 * There is no default zero huge page for all huge page sizes as
229 * supported by hugetlb. A PMD_SIZE huge pages may exist as used
230 * by THP. Since we can not reliably insert a zero page, this
231 * feature is not supported.
232 */
233 if (zeropage) {
234 up_read(&dst_mm->mmap_sem);
235 return -EINVAL;
236 }
237
238 src_addr = src_start;
239 dst_addr = dst_start;
240 copied = 0;
241 page = NULL;
242 vma_hpagesize = vma_kernel_pagesize(dst_vma);
243
244 /*
245 * Validate alignment based on huge page size
246 */
247 err = -EINVAL;
248 if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
249 goto out_unlock;
250
251retry:
252 /*
253 * On routine entry dst_vma is set. If we had to drop mmap_sem and
254 * retry, dst_vma will be set to NULL and we must lookup again.
255 */
256 if (!dst_vma) {
257 err = -ENOENT;
258 dst_vma = find_dst_vma(dst_mm, dst_start, len);
259 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
260 goto out_unlock;
261
262 err = -EINVAL;
263 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
264 goto out_unlock;
265
266 vm_shared = dst_vma->vm_flags & VM_SHARED;
267 }
268
269 /*
270 * If not shared, ensure the dst_vma has a anon_vma.
271 */
272 err = -ENOMEM;
273 if (!vm_shared) {
274 if (unlikely(anon_vma_prepare(dst_vma)))
275 goto out_unlock;
276 }
277
278 while (src_addr < src_start + len) {
279 pte_t dst_pteval;
280
281 BUG_ON(dst_addr >= dst_start + len);
282
283 /*
284 * Serialize via i_mmap_rwsem and hugetlb_fault_mutex.
285 * i_mmap_rwsem ensures the dst_pte remains valid even
286 * in the case of shared pmds. fault mutex prevents
287 * races with other faulting threads.
288 */
289 mapping = dst_vma->vm_file->f_mapping;
290 i_mmap_lock_read(mapping);
291 idx = linear_page_index(dst_vma, dst_addr);
292 hash = hugetlb_fault_mutex_hash(mapping, idx);
293 mutex_lock(&hugetlb_fault_mutex_table[hash]);
294
295 err = -ENOMEM;
296 dst_pte = huge_pte_alloc(dst_mm, dst_addr, vma_hpagesize);
297 if (!dst_pte) {
298 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
299 i_mmap_unlock_read(mapping);
300 goto out_unlock;
301 }
302
303 err = -EEXIST;
304 dst_pteval = huge_ptep_get(dst_pte);
305 if (!huge_pte_none(dst_pteval)) {
306 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
307 i_mmap_unlock_read(mapping);
308 goto out_unlock;
309 }
310
311 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
312 dst_addr, src_addr, &page);
313
314 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
315 i_mmap_unlock_read(mapping);
316 vm_alloc_shared = vm_shared;
317
318 cond_resched();
319
320 if (unlikely(err == -ENOENT)) {
321 up_read(&dst_mm->mmap_sem);
322 BUG_ON(!page);
323
324 err = copy_huge_page_from_user(page,
325 (const void __user *)src_addr,
326 vma_hpagesize / PAGE_SIZE,
327 true);
328 if (unlikely(err)) {
329 err = -EFAULT;
330 goto out;
331 }
332 down_read(&dst_mm->mmap_sem);
333
334 dst_vma = NULL;
335 goto retry;
336 } else
337 BUG_ON(page);
338
339 if (!err) {
340 dst_addr += vma_hpagesize;
341 src_addr += vma_hpagesize;
342 copied += vma_hpagesize;
343
344 if (fatal_signal_pending(current))
345 err = -EINTR;
346 }
347 if (err)
348 break;
349 }
350
351out_unlock:
352 up_read(&dst_mm->mmap_sem);
353out:
354 if (page) {
355 /*
356 * We encountered an error and are about to free a newly
357 * allocated huge page.
358 *
359 * Reservation handling is very subtle, and is different for
360 * private and shared mappings. See the routine
361 * restore_reserve_on_error for details. Unfortunately, we
362 * can not call restore_reserve_on_error now as it would
363 * require holding mmap_sem.
364 *
365 * If a reservation for the page existed in the reservation
366 * map of a private mapping, the map was modified to indicate
367 * the reservation was consumed when the page was allocated.
368 * We clear the PagePrivate flag now so that the global
369 * reserve count will not be incremented in free_huge_page.
370 * The reservation map will still indicate the reservation
371 * was consumed and possibly prevent later page allocation.
372 * This is better than leaking a global reservation. If no
373 * reservation existed, it is still safe to clear PagePrivate
374 * as no adjustments to reservation counts were made during
375 * allocation.
376 *
377 * The reservation map for shared mappings indicates which
378 * pages have reservations. When a huge page is allocated
379 * for an address with a reservation, no change is made to
380 * the reserve map. In this case PagePrivate will be set
381 * to indicate that the global reservation count should be
382 * incremented when the page is freed. This is the desired
383 * behavior. However, when a huge page is allocated for an
384 * address without a reservation a reservation entry is added
385 * to the reservation map, and PagePrivate will not be set.
386 * When the page is freed, the global reserve count will NOT
387 * be incremented and it will appear as though we have leaked
388 * reserved page. In this case, set PagePrivate so that the
389 * global reserve count will be incremented to match the
390 * reservation map entry which was created.
391 *
392 * Note that vm_alloc_shared is based on the flags of the vma
393 * for which the page was originally allocated. dst_vma could
394 * be different or NULL on error.
395 */
396 if (vm_alloc_shared)
397 SetPagePrivate(page);
398 else
399 ClearPagePrivate(page);
400 put_page(page);
401 }
402 BUG_ON(copied < 0);
403 BUG_ON(err > 0);
404 BUG_ON(!copied && !err);
405 return copied ? copied : err;
406}
407#else /* !CONFIG_HUGETLB_PAGE */
408/* fail at build time if gcc attempts to use this */
409extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
410 struct vm_area_struct *dst_vma,
411 unsigned long dst_start,
412 unsigned long src_start,
413 unsigned long len,
414 bool zeropage);
415#endif /* CONFIG_HUGETLB_PAGE */
416
417static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
418 pmd_t *dst_pmd,
419 struct vm_area_struct *dst_vma,
420 unsigned long dst_addr,
421 unsigned long src_addr,
422 struct page **page,
423 bool zeropage,
424 bool wp_copy)
425{
426 ssize_t err;
427
428 /*
429 * The normal page fault path for a shmem will invoke the
430 * fault, fill the hole in the file and COW it right away. The
431 * result generates plain anonymous memory. So when we are
432 * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
433 * generate anonymous memory directly without actually filling
434 * the hole. For the MAP_PRIVATE case the robustness check
435 * only happens in the pagetable (to verify it's still none)
436 * and not in the radix tree.
437 */
438 if (!(dst_vma->vm_flags & VM_SHARED)) {
439 if (!zeropage)
440 err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
441 dst_addr, src_addr, page,
442 wp_copy);
443 else
444 err = mfill_zeropage_pte(dst_mm, dst_pmd,
445 dst_vma, dst_addr);
446 } else {
447 VM_WARN_ON_ONCE(wp_copy);
448 if (!zeropage)
449 err = shmem_mcopy_atomic_pte(dst_mm, dst_pmd,
450 dst_vma, dst_addr,
451 src_addr, page);
452 else
453 err = shmem_mfill_zeropage_pte(dst_mm, dst_pmd,
454 dst_vma, dst_addr);
455 }
456
457 return err;
458}
459
460static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
461 unsigned long dst_start,
462 unsigned long src_start,
463 unsigned long len,
464 bool zeropage,
465 bool *mmap_changing,
466 __u64 mode)
467{
468 struct vm_area_struct *dst_vma;
469 ssize_t err;
470 pmd_t *dst_pmd;
471 unsigned long src_addr, dst_addr;
472 long copied;
473 struct page *page;
474 bool wp_copy;
475
476 /*
477 * Sanitize the command parameters:
478 */
479 BUG_ON(dst_start & ~PAGE_MASK);
480 BUG_ON(len & ~PAGE_MASK);
481
482 /* Does the address range wrap, or is the span zero-sized? */
483 BUG_ON(src_start + len <= src_start);
484 BUG_ON(dst_start + len <= dst_start);
485
486 src_addr = src_start;
487 dst_addr = dst_start;
488 copied = 0;
489 page = NULL;
490retry:
491 down_read(&dst_mm->mmap_sem);
492
493 /*
494 * If memory mappings are changing because of non-cooperative
495 * operation (e.g. mremap) running in parallel, bail out and
496 * request the user to retry later
497 */
498 err = -EAGAIN;
499 if (mmap_changing && READ_ONCE(*mmap_changing))
500 goto out_unlock;
501
502 /*
503 * Make sure the vma is not shared, that the dst range is
504 * both valid and fully within a single existing vma.
505 */
506 err = -ENOENT;
507 dst_vma = find_dst_vma(dst_mm, dst_start, len);
508 if (!dst_vma)
509 goto out_unlock;
510
511 err = -EINVAL;
512 /*
513 * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
514 * it will overwrite vm_ops, so vma_is_anonymous must return false.
515 */
516 if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
517 dst_vma->vm_flags & VM_SHARED))
518 goto out_unlock;
519
520 /*
521 * validate 'mode' now that we know the dst_vma: don't allow
522 * a wrprotect copy if the userfaultfd didn't register as WP.
523 */
524 wp_copy = mode & UFFDIO_COPY_MODE_WP;
525 if (wp_copy && !(dst_vma->vm_flags & VM_UFFD_WP))
526 goto out_unlock;
527
528 /*
529 * If this is a HUGETLB vma, pass off to appropriate routine
530 */
531 if (is_vm_hugetlb_page(dst_vma))
532 return __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
533 src_start, len, zeropage);
534
535 if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
536 goto out_unlock;
537
538 /*
539 * Ensure the dst_vma has a anon_vma or this page
540 * would get a NULL anon_vma when moved in the
541 * dst_vma.
542 */
543 err = -ENOMEM;
544 if (!(dst_vma->vm_flags & VM_SHARED) &&
545 unlikely(anon_vma_prepare(dst_vma)))
546 goto out_unlock;
547
548 while (src_addr < src_start + len) {
549 pmd_t dst_pmdval;
550
551 BUG_ON(dst_addr >= dst_start + len);
552
553 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
554 if (unlikely(!dst_pmd)) {
555 err = -ENOMEM;
556 break;
557 }
558
559 dst_pmdval = pmd_read_atomic(dst_pmd);
560 /*
561 * If the dst_pmd is mapped as THP don't
562 * override it and just be strict.
563 */
564 if (unlikely(pmd_trans_huge(dst_pmdval))) {
565 err = -EEXIST;
566 break;
567 }
568 if (unlikely(pmd_none(dst_pmdval)) &&
569 unlikely(__pte_alloc(dst_mm, dst_pmd))) {
570 err = -ENOMEM;
571 break;
572 }
573 /* If an huge pmd materialized from under us fail */
574 if (unlikely(pmd_trans_huge(*dst_pmd))) {
575 err = -EFAULT;
576 break;
577 }
578
579 BUG_ON(pmd_none(*dst_pmd));
580 BUG_ON(pmd_trans_huge(*dst_pmd));
581
582 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
583 src_addr, &page, zeropage, wp_copy);
584 cond_resched();
585
586 if (unlikely(err == -ENOENT)) {
587 void *page_kaddr;
588
589 up_read(&dst_mm->mmap_sem);
590 BUG_ON(!page);
591
592 page_kaddr = kmap(page);
593 err = copy_from_user(page_kaddr,
594 (const void __user *) src_addr,
595 PAGE_SIZE);
596 kunmap(page);
597 if (unlikely(err)) {
598 err = -EFAULT;
599 goto out;
600 }
601 goto retry;
602 } else
603 BUG_ON(page);
604
605 if (!err) {
606 dst_addr += PAGE_SIZE;
607 src_addr += PAGE_SIZE;
608 copied += PAGE_SIZE;
609
610 if (fatal_signal_pending(current))
611 err = -EINTR;
612 }
613 if (err)
614 break;
615 }
616
617out_unlock:
618 up_read(&dst_mm->mmap_sem);
619out:
620 if (page)
621 put_page(page);
622 BUG_ON(copied < 0);
623 BUG_ON(err > 0);
624 BUG_ON(!copied && !err);
625 return copied ? copied : err;
626}
627
628ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
629 unsigned long src_start, unsigned long len,
630 bool *mmap_changing, __u64 mode)
631{
632 return __mcopy_atomic(dst_mm, dst_start, src_start, len, false,
633 mmap_changing, mode);
634}
635
636ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
637 unsigned long len, bool *mmap_changing)
638{
639 return __mcopy_atomic(dst_mm, start, 0, len, true, mmap_changing, 0);
640}
641
642int mwriteprotect_range(struct mm_struct *dst_mm, unsigned long start,
643 unsigned long len, bool enable_wp, bool *mmap_changing)
644{
645 struct vm_area_struct *dst_vma;
646 pgprot_t newprot;
647 int err;
648
649 /*
650 * Sanitize the command parameters:
651 */
652 BUG_ON(start & ~PAGE_MASK);
653 BUG_ON(len & ~PAGE_MASK);
654
655 /* Does the address range wrap, or is the span zero-sized? */
656 BUG_ON(start + len <= start);
657
658 down_read(&dst_mm->mmap_sem);
659
660 /*
661 * If memory mappings are changing because of non-cooperative
662 * operation (e.g. mremap) running in parallel, bail out and
663 * request the user to retry later
664 */
665 err = -EAGAIN;
666 if (mmap_changing && READ_ONCE(*mmap_changing))
667 goto out_unlock;
668
669 err = -ENOENT;
670 dst_vma = find_dst_vma(dst_mm, start, len);
671 /*
672 * Make sure the vma is not shared, that the dst range is
673 * both valid and fully within a single existing vma.
674 */
675 if (!dst_vma || (dst_vma->vm_flags & VM_SHARED))
676 goto out_unlock;
677 if (!userfaultfd_wp(dst_vma))
678 goto out_unlock;
679 if (!vma_is_anonymous(dst_vma))
680 goto out_unlock;
681
682 if (enable_wp)
683 newprot = vm_get_page_prot(dst_vma->vm_flags & ~(VM_WRITE));
684 else
685 newprot = vm_get_page_prot(dst_vma->vm_flags);
686
687 change_protection(dst_vma, start, start + len, newprot,
688 enable_wp ? MM_CP_UFFD_WP : MM_CP_UFFD_WP_RESOLVE);
689
690 err = 0;
691out_unlock:
692 up_read(&dst_mm->mmap_sem);
693 return err;
694}