Loading...
1/*
2 * mm/userfaultfd.c
3 *
4 * Copyright (C) 2015 Red Hat, Inc.
5 *
6 * This work is licensed under the terms of the GNU GPL, version 2. See
7 * the COPYING file in the top-level directory.
8 */
9
10#include <linux/mm.h>
11#include <linux/sched/signal.h>
12#include <linux/pagemap.h>
13#include <linux/rmap.h>
14#include <linux/swap.h>
15#include <linux/swapops.h>
16#include <linux/userfaultfd_k.h>
17#include <linux/mmu_notifier.h>
18#include <linux/hugetlb.h>
19#include <linux/shmem_fs.h>
20#include <asm/tlbflush.h>
21#include "internal.h"
22
23static int mcopy_atomic_pte(struct mm_struct *dst_mm,
24 pmd_t *dst_pmd,
25 struct vm_area_struct *dst_vma,
26 unsigned long dst_addr,
27 unsigned long src_addr,
28 struct page **pagep)
29{
30 struct mem_cgroup *memcg;
31 pte_t _dst_pte, *dst_pte;
32 spinlock_t *ptl;
33 void *page_kaddr;
34 int ret;
35 struct page *page;
36
37 if (!*pagep) {
38 ret = -ENOMEM;
39 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
40 if (!page)
41 goto out;
42
43 page_kaddr = kmap_atomic(page);
44 ret = copy_from_user(page_kaddr,
45 (const void __user *) src_addr,
46 PAGE_SIZE);
47 kunmap_atomic(page_kaddr);
48
49 /* fallback to copy_from_user outside mmap_sem */
50 if (unlikely(ret)) {
51 ret = -EFAULT;
52 *pagep = page;
53 /* don't free the page */
54 goto out;
55 }
56 } else {
57 page = *pagep;
58 *pagep = NULL;
59 }
60
61 /*
62 * The memory barrier inside __SetPageUptodate makes sure that
63 * preceeding stores to the page contents become visible before
64 * the set_pte_at() write.
65 */
66 __SetPageUptodate(page);
67
68 ret = -ENOMEM;
69 if (mem_cgroup_try_charge(page, dst_mm, GFP_KERNEL, &memcg, false))
70 goto out_release;
71
72 _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
73 if (dst_vma->vm_flags & VM_WRITE)
74 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
75
76 ret = -EEXIST;
77 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
78 if (!pte_none(*dst_pte))
79 goto out_release_uncharge_unlock;
80
81 inc_mm_counter(dst_mm, MM_ANONPAGES);
82 page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
83 mem_cgroup_commit_charge(page, memcg, false, false);
84 lru_cache_add_active_or_unevictable(page, dst_vma);
85
86 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
87
88 /* No need to invalidate - it was non-present before */
89 update_mmu_cache(dst_vma, dst_addr, dst_pte);
90
91 pte_unmap_unlock(dst_pte, ptl);
92 ret = 0;
93out:
94 return ret;
95out_release_uncharge_unlock:
96 pte_unmap_unlock(dst_pte, ptl);
97 mem_cgroup_cancel_charge(page, memcg, false);
98out_release:
99 put_page(page);
100 goto out;
101}
102
103static int mfill_zeropage_pte(struct mm_struct *dst_mm,
104 pmd_t *dst_pmd,
105 struct vm_area_struct *dst_vma,
106 unsigned long dst_addr)
107{
108 pte_t _dst_pte, *dst_pte;
109 spinlock_t *ptl;
110 int ret;
111
112 _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
113 dst_vma->vm_page_prot));
114 ret = -EEXIST;
115 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
116 if (!pte_none(*dst_pte))
117 goto out_unlock;
118 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
119 /* No need to invalidate - it was non-present before */
120 update_mmu_cache(dst_vma, dst_addr, dst_pte);
121 ret = 0;
122out_unlock:
123 pte_unmap_unlock(dst_pte, ptl);
124 return ret;
125}
126
127static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
128{
129 pgd_t *pgd;
130 p4d_t *p4d;
131 pud_t *pud;
132
133 pgd = pgd_offset(mm, address);
134 p4d = p4d_alloc(mm, pgd, address);
135 if (!p4d)
136 return NULL;
137 pud = pud_alloc(mm, p4d, address);
138 if (!pud)
139 return NULL;
140 /*
141 * Note that we didn't run this because the pmd was
142 * missing, the *pmd may be already established and in
143 * turn it may also be a trans_huge_pmd.
144 */
145 return pmd_alloc(mm, pud, address);
146}
147
148#ifdef CONFIG_HUGETLB_PAGE
149/*
150 * __mcopy_atomic processing for HUGETLB vmas. Note that this routine is
151 * called with mmap_sem held, it will release mmap_sem before returning.
152 */
153static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
154 struct vm_area_struct *dst_vma,
155 unsigned long dst_start,
156 unsigned long src_start,
157 unsigned long len,
158 bool zeropage)
159{
160 int vm_alloc_shared = dst_vma->vm_flags & VM_SHARED;
161 int vm_shared = dst_vma->vm_flags & VM_SHARED;
162 ssize_t err;
163 pte_t *dst_pte;
164 unsigned long src_addr, dst_addr;
165 long copied;
166 struct page *page;
167 struct hstate *h;
168 unsigned long vma_hpagesize;
169 pgoff_t idx;
170 u32 hash;
171 struct address_space *mapping;
172
173 /*
174 * There is no default zero huge page for all huge page sizes as
175 * supported by hugetlb. A PMD_SIZE huge pages may exist as used
176 * by THP. Since we can not reliably insert a zero page, this
177 * feature is not supported.
178 */
179 if (zeropage) {
180 up_read(&dst_mm->mmap_sem);
181 return -EINVAL;
182 }
183
184 src_addr = src_start;
185 dst_addr = dst_start;
186 copied = 0;
187 page = NULL;
188 vma_hpagesize = vma_kernel_pagesize(dst_vma);
189
190 /*
191 * Validate alignment based on huge page size
192 */
193 err = -EINVAL;
194 if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
195 goto out_unlock;
196
197retry:
198 /*
199 * On routine entry dst_vma is set. If we had to drop mmap_sem and
200 * retry, dst_vma will be set to NULL and we must lookup again.
201 */
202 if (!dst_vma) {
203 err = -ENOENT;
204 dst_vma = find_vma(dst_mm, dst_start);
205 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
206 goto out_unlock;
207 /*
208 * Only allow __mcopy_atomic_hugetlb on userfaultfd
209 * registered ranges.
210 */
211 if (!dst_vma->vm_userfaultfd_ctx.ctx)
212 goto out_unlock;
213
214 if (dst_start < dst_vma->vm_start ||
215 dst_start + len > dst_vma->vm_end)
216 goto out_unlock;
217
218 err = -EINVAL;
219 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
220 goto out_unlock;
221
222 vm_shared = dst_vma->vm_flags & VM_SHARED;
223 }
224
225 if (WARN_ON(dst_addr & (vma_hpagesize - 1) ||
226 (len - copied) & (vma_hpagesize - 1)))
227 goto out_unlock;
228
229 /*
230 * If not shared, ensure the dst_vma has a anon_vma.
231 */
232 err = -ENOMEM;
233 if (!vm_shared) {
234 if (unlikely(anon_vma_prepare(dst_vma)))
235 goto out_unlock;
236 }
237
238 h = hstate_vma(dst_vma);
239
240 while (src_addr < src_start + len) {
241 pte_t dst_pteval;
242
243 BUG_ON(dst_addr >= dst_start + len);
244 VM_BUG_ON(dst_addr & ~huge_page_mask(h));
245
246 /*
247 * Serialize via hugetlb_fault_mutex
248 */
249 idx = linear_page_index(dst_vma, dst_addr);
250 mapping = dst_vma->vm_file->f_mapping;
251 hash = hugetlb_fault_mutex_hash(h, dst_mm, dst_vma, mapping,
252 idx, dst_addr);
253 mutex_lock(&hugetlb_fault_mutex_table[hash]);
254
255 err = -ENOMEM;
256 dst_pte = huge_pte_alloc(dst_mm, dst_addr, huge_page_size(h));
257 if (!dst_pte) {
258 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
259 goto out_unlock;
260 }
261
262 err = -EEXIST;
263 dst_pteval = huge_ptep_get(dst_pte);
264 if (!huge_pte_none(dst_pteval)) {
265 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
266 goto out_unlock;
267 }
268
269 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
270 dst_addr, src_addr, &page);
271
272 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
273 vm_alloc_shared = vm_shared;
274
275 cond_resched();
276
277 if (unlikely(err == -EFAULT)) {
278 up_read(&dst_mm->mmap_sem);
279 BUG_ON(!page);
280
281 err = copy_huge_page_from_user(page,
282 (const void __user *)src_addr,
283 pages_per_huge_page(h), true);
284 if (unlikely(err)) {
285 err = -EFAULT;
286 goto out;
287 }
288 down_read(&dst_mm->mmap_sem);
289
290 dst_vma = NULL;
291 goto retry;
292 } else
293 BUG_ON(page);
294
295 if (!err) {
296 dst_addr += vma_hpagesize;
297 src_addr += vma_hpagesize;
298 copied += vma_hpagesize;
299
300 if (fatal_signal_pending(current))
301 err = -EINTR;
302 }
303 if (err)
304 break;
305 }
306
307out_unlock:
308 up_read(&dst_mm->mmap_sem);
309out:
310 if (page) {
311 /*
312 * We encountered an error and are about to free a newly
313 * allocated huge page.
314 *
315 * Reservation handling is very subtle, and is different for
316 * private and shared mappings. See the routine
317 * restore_reserve_on_error for details. Unfortunately, we
318 * can not call restore_reserve_on_error now as it would
319 * require holding mmap_sem.
320 *
321 * If a reservation for the page existed in the reservation
322 * map of a private mapping, the map was modified to indicate
323 * the reservation was consumed when the page was allocated.
324 * We clear the PagePrivate flag now so that the global
325 * reserve count will not be incremented in free_huge_page.
326 * The reservation map will still indicate the reservation
327 * was consumed and possibly prevent later page allocation.
328 * This is better than leaking a global reservation. If no
329 * reservation existed, it is still safe to clear PagePrivate
330 * as no adjustments to reservation counts were made during
331 * allocation.
332 *
333 * The reservation map for shared mappings indicates which
334 * pages have reservations. When a huge page is allocated
335 * for an address with a reservation, no change is made to
336 * the reserve map. In this case PagePrivate will be set
337 * to indicate that the global reservation count should be
338 * incremented when the page is freed. This is the desired
339 * behavior. However, when a huge page is allocated for an
340 * address without a reservation a reservation entry is added
341 * to the reservation map, and PagePrivate will not be set.
342 * When the page is freed, the global reserve count will NOT
343 * be incremented and it will appear as though we have leaked
344 * reserved page. In this case, set PagePrivate so that the
345 * global reserve count will be incremented to match the
346 * reservation map entry which was created.
347 *
348 * Note that vm_alloc_shared is based on the flags of the vma
349 * for which the page was originally allocated. dst_vma could
350 * be different or NULL on error.
351 */
352 if (vm_alloc_shared)
353 SetPagePrivate(page);
354 else
355 ClearPagePrivate(page);
356 put_page(page);
357 }
358 BUG_ON(copied < 0);
359 BUG_ON(err > 0);
360 BUG_ON(!copied && !err);
361 return copied ? copied : err;
362}
363#else /* !CONFIG_HUGETLB_PAGE */
364/* fail at build time if gcc attempts to use this */
365extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
366 struct vm_area_struct *dst_vma,
367 unsigned long dst_start,
368 unsigned long src_start,
369 unsigned long len,
370 bool zeropage);
371#endif /* CONFIG_HUGETLB_PAGE */
372
373static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
374 pmd_t *dst_pmd,
375 struct vm_area_struct *dst_vma,
376 unsigned long dst_addr,
377 unsigned long src_addr,
378 struct page **page,
379 bool zeropage)
380{
381 ssize_t err;
382
383 if (vma_is_anonymous(dst_vma)) {
384 if (!zeropage)
385 err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
386 dst_addr, src_addr, page);
387 else
388 err = mfill_zeropage_pte(dst_mm, dst_pmd,
389 dst_vma, dst_addr);
390 } else {
391 if (!zeropage)
392 err = shmem_mcopy_atomic_pte(dst_mm, dst_pmd,
393 dst_vma, dst_addr,
394 src_addr, page);
395 else
396 err = shmem_mfill_zeropage_pte(dst_mm, dst_pmd,
397 dst_vma, dst_addr);
398 }
399
400 return err;
401}
402
403static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
404 unsigned long dst_start,
405 unsigned long src_start,
406 unsigned long len,
407 bool zeropage)
408{
409 struct vm_area_struct *dst_vma;
410 ssize_t err;
411 pmd_t *dst_pmd;
412 unsigned long src_addr, dst_addr;
413 long copied;
414 struct page *page;
415
416 /*
417 * Sanitize the command parameters:
418 */
419 BUG_ON(dst_start & ~PAGE_MASK);
420 BUG_ON(len & ~PAGE_MASK);
421
422 /* Does the address range wrap, or is the span zero-sized? */
423 BUG_ON(src_start + len <= src_start);
424 BUG_ON(dst_start + len <= dst_start);
425
426 src_addr = src_start;
427 dst_addr = dst_start;
428 copied = 0;
429 page = NULL;
430retry:
431 down_read(&dst_mm->mmap_sem);
432
433 /*
434 * Make sure the vma is not shared, that the dst range is
435 * both valid and fully within a single existing vma.
436 */
437 err = -ENOENT;
438 dst_vma = find_vma(dst_mm, dst_start);
439 if (!dst_vma)
440 goto out_unlock;
441 /*
442 * Be strict and only allow __mcopy_atomic on userfaultfd
443 * registered ranges to prevent userland errors going
444 * unnoticed. As far as the VM consistency is concerned, it
445 * would be perfectly safe to remove this check, but there's
446 * no useful usage for __mcopy_atomic ouside of userfaultfd
447 * registered ranges. This is after all why these are ioctls
448 * belonging to the userfaultfd and not syscalls.
449 */
450 if (!dst_vma->vm_userfaultfd_ctx.ctx)
451 goto out_unlock;
452
453 if (dst_start < dst_vma->vm_start ||
454 dst_start + len > dst_vma->vm_end)
455 goto out_unlock;
456
457 err = -EINVAL;
458 /*
459 * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
460 * it will overwrite vm_ops, so vma_is_anonymous must return false.
461 */
462 if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
463 dst_vma->vm_flags & VM_SHARED))
464 goto out_unlock;
465
466 /*
467 * If this is a HUGETLB vma, pass off to appropriate routine
468 */
469 if (is_vm_hugetlb_page(dst_vma))
470 return __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
471 src_start, len, zeropage);
472
473 if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
474 goto out_unlock;
475
476 /*
477 * Ensure the dst_vma has a anon_vma or this page
478 * would get a NULL anon_vma when moved in the
479 * dst_vma.
480 */
481 err = -ENOMEM;
482 if (vma_is_anonymous(dst_vma) && unlikely(anon_vma_prepare(dst_vma)))
483 goto out_unlock;
484
485 while (src_addr < src_start + len) {
486 pmd_t dst_pmdval;
487
488 BUG_ON(dst_addr >= dst_start + len);
489
490 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
491 if (unlikely(!dst_pmd)) {
492 err = -ENOMEM;
493 break;
494 }
495
496 dst_pmdval = pmd_read_atomic(dst_pmd);
497 /*
498 * If the dst_pmd is mapped as THP don't
499 * override it and just be strict.
500 */
501 if (unlikely(pmd_trans_huge(dst_pmdval))) {
502 err = -EEXIST;
503 break;
504 }
505 if (unlikely(pmd_none(dst_pmdval)) &&
506 unlikely(__pte_alloc(dst_mm, dst_pmd, dst_addr))) {
507 err = -ENOMEM;
508 break;
509 }
510 /* If an huge pmd materialized from under us fail */
511 if (unlikely(pmd_trans_huge(*dst_pmd))) {
512 err = -EFAULT;
513 break;
514 }
515
516 BUG_ON(pmd_none(*dst_pmd));
517 BUG_ON(pmd_trans_huge(*dst_pmd));
518
519 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
520 src_addr, &page, zeropage);
521 cond_resched();
522
523 if (unlikely(err == -EFAULT)) {
524 void *page_kaddr;
525
526 up_read(&dst_mm->mmap_sem);
527 BUG_ON(!page);
528
529 page_kaddr = kmap(page);
530 err = copy_from_user(page_kaddr,
531 (const void __user *) src_addr,
532 PAGE_SIZE);
533 kunmap(page);
534 if (unlikely(err)) {
535 err = -EFAULT;
536 goto out;
537 }
538 goto retry;
539 } else
540 BUG_ON(page);
541
542 if (!err) {
543 dst_addr += PAGE_SIZE;
544 src_addr += PAGE_SIZE;
545 copied += PAGE_SIZE;
546
547 if (fatal_signal_pending(current))
548 err = -EINTR;
549 }
550 if (err)
551 break;
552 }
553
554out_unlock:
555 up_read(&dst_mm->mmap_sem);
556out:
557 if (page)
558 put_page(page);
559 BUG_ON(copied < 0);
560 BUG_ON(err > 0);
561 BUG_ON(!copied && !err);
562 return copied ? copied : err;
563}
564
565ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
566 unsigned long src_start, unsigned long len)
567{
568 return __mcopy_atomic(dst_mm, dst_start, src_start, len, false);
569}
570
571ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
572 unsigned long len)
573{
574 return __mcopy_atomic(dst_mm, start, 0, len, true);
575}
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
22static __always_inline
23struct vm_area_struct *find_dst_vma(struct mm_struct *dst_mm,
24 unsigned long dst_start,
25 unsigned long len)
26{
27 /*
28 * Make sure that the dst range is both valid and fully within a
29 * single existing vma.
30 */
31 struct vm_area_struct *dst_vma;
32
33 dst_vma = find_vma(dst_mm, dst_start);
34 if (!dst_vma)
35 return NULL;
36
37 if (dst_start < dst_vma->vm_start ||
38 dst_start + len > dst_vma->vm_end)
39 return NULL;
40
41 /*
42 * Check the vma is registered in uffd, this is required to
43 * enforce the VM_MAYWRITE check done at uffd registration
44 * time.
45 */
46 if (!dst_vma->vm_userfaultfd_ctx.ctx)
47 return NULL;
48
49 return dst_vma;
50}
51
52/*
53 * Install PTEs, to map dst_addr (within dst_vma) to page.
54 *
55 * This function handles both MCOPY_ATOMIC_NORMAL and _CONTINUE for both shmem
56 * and anon, and for both shared and private VMAs.
57 */
58int mfill_atomic_install_pte(struct mm_struct *dst_mm, pmd_t *dst_pmd,
59 struct vm_area_struct *dst_vma,
60 unsigned long dst_addr, struct page *page,
61 bool newly_allocated, bool wp_copy)
62{
63 int ret;
64 pte_t _dst_pte, *dst_pte;
65 bool writable = dst_vma->vm_flags & VM_WRITE;
66 bool vm_shared = dst_vma->vm_flags & VM_SHARED;
67 bool page_in_cache = page_mapping(page);
68 spinlock_t *ptl;
69 struct folio *folio;
70 struct inode *inode;
71 pgoff_t offset, max_off;
72
73 _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
74 _dst_pte = pte_mkdirty(_dst_pte);
75 if (page_in_cache && !vm_shared)
76 writable = false;
77
78 /*
79 * Always mark a PTE as write-protected when needed, regardless of
80 * VM_WRITE, which the user might change.
81 */
82 if (wp_copy) {
83 _dst_pte = pte_mkuffd_wp(_dst_pte);
84 writable = false;
85 }
86
87 if (writable)
88 _dst_pte = pte_mkwrite(_dst_pte);
89 else
90 /*
91 * We need this to make sure write bit removed; as mk_pte()
92 * could return a pte with write bit set.
93 */
94 _dst_pte = pte_wrprotect(_dst_pte);
95
96 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
97
98 if (vma_is_shmem(dst_vma)) {
99 /* serialize against truncate with the page table lock */
100 inode = dst_vma->vm_file->f_inode;
101 offset = linear_page_index(dst_vma, dst_addr);
102 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
103 ret = -EFAULT;
104 if (unlikely(offset >= max_off))
105 goto out_unlock;
106 }
107
108 ret = -EEXIST;
109 /*
110 * We allow to overwrite a pte marker: consider when both MISSING|WP
111 * registered, we firstly wr-protect a none pte which has no page cache
112 * page backing it, then access the page.
113 */
114 if (!pte_none_mostly(*dst_pte))
115 goto out_unlock;
116
117 folio = page_folio(page);
118 if (page_in_cache) {
119 /* Usually, cache pages are already added to LRU */
120 if (newly_allocated)
121 folio_add_lru(folio);
122 page_add_file_rmap(page, dst_vma, false);
123 } else {
124 page_add_new_anon_rmap(page, dst_vma, dst_addr);
125 folio_add_lru_vma(folio, dst_vma);
126 }
127
128 /*
129 * Must happen after rmap, as mm_counter() checks mapping (via
130 * PageAnon()), which is set by __page_set_anon_rmap().
131 */
132 inc_mm_counter(dst_mm, mm_counter(page));
133
134 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
135
136 /* No need to invalidate - it was non-present before */
137 update_mmu_cache(dst_vma, dst_addr, dst_pte);
138 ret = 0;
139out_unlock:
140 pte_unmap_unlock(dst_pte, ptl);
141 return ret;
142}
143
144static int mcopy_atomic_pte(struct mm_struct *dst_mm,
145 pmd_t *dst_pmd,
146 struct vm_area_struct *dst_vma,
147 unsigned long dst_addr,
148 unsigned long src_addr,
149 struct page **pagep,
150 bool wp_copy)
151{
152 void *page_kaddr;
153 int ret;
154 struct page *page;
155
156 if (!*pagep) {
157 ret = -ENOMEM;
158 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
159 if (!page)
160 goto out;
161
162 page_kaddr = kmap_local_page(page);
163 /*
164 * The read mmap_lock is held here. Despite the
165 * mmap_lock being read recursive a deadlock is still
166 * possible if a writer has taken a lock. For example:
167 *
168 * process A thread 1 takes read lock on own mmap_lock
169 * process A thread 2 calls mmap, blocks taking write lock
170 * process B thread 1 takes page fault, read lock on own mmap lock
171 * process B thread 2 calls mmap, blocks taking write lock
172 * process A thread 1 blocks taking read lock on process B
173 * process B thread 1 blocks taking read lock on process A
174 *
175 * Disable page faults to prevent potential deadlock
176 * and retry the copy outside the mmap_lock.
177 */
178 pagefault_disable();
179 ret = copy_from_user(page_kaddr,
180 (const void __user *) src_addr,
181 PAGE_SIZE);
182 pagefault_enable();
183 kunmap_local(page_kaddr);
184
185 /* fallback to copy_from_user outside mmap_lock */
186 if (unlikely(ret)) {
187 ret = -ENOENT;
188 *pagep = page;
189 /* don't free the page */
190 goto out;
191 }
192
193 flush_dcache_page(page);
194 } else {
195 page = *pagep;
196 *pagep = NULL;
197 }
198
199 /*
200 * The memory barrier inside __SetPageUptodate makes sure that
201 * preceding stores to the page contents become visible before
202 * the set_pte_at() write.
203 */
204 __SetPageUptodate(page);
205
206 ret = -ENOMEM;
207 if (mem_cgroup_charge(page_folio(page), dst_mm, GFP_KERNEL))
208 goto out_release;
209
210 ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
211 page, true, wp_copy);
212 if (ret)
213 goto out_release;
214out:
215 return ret;
216out_release:
217 put_page(page);
218 goto out;
219}
220
221static int mfill_zeropage_pte(struct mm_struct *dst_mm,
222 pmd_t *dst_pmd,
223 struct vm_area_struct *dst_vma,
224 unsigned long dst_addr)
225{
226 pte_t _dst_pte, *dst_pte;
227 spinlock_t *ptl;
228 int ret;
229 pgoff_t offset, max_off;
230 struct inode *inode;
231
232 _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
233 dst_vma->vm_page_prot));
234 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
235 if (dst_vma->vm_file) {
236 /* the shmem MAP_PRIVATE case requires checking the i_size */
237 inode = dst_vma->vm_file->f_inode;
238 offset = linear_page_index(dst_vma, dst_addr);
239 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
240 ret = -EFAULT;
241 if (unlikely(offset >= max_off))
242 goto out_unlock;
243 }
244 ret = -EEXIST;
245 if (!pte_none(*dst_pte))
246 goto out_unlock;
247 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
248 /* No need to invalidate - it was non-present before */
249 update_mmu_cache(dst_vma, dst_addr, dst_pte);
250 ret = 0;
251out_unlock:
252 pte_unmap_unlock(dst_pte, ptl);
253 return ret;
254}
255
256/* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */
257static int mcontinue_atomic_pte(struct mm_struct *dst_mm,
258 pmd_t *dst_pmd,
259 struct vm_area_struct *dst_vma,
260 unsigned long dst_addr,
261 bool wp_copy)
262{
263 struct inode *inode = file_inode(dst_vma->vm_file);
264 pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
265 struct folio *folio;
266 struct page *page;
267 int ret;
268
269 ret = shmem_get_folio(inode, pgoff, &folio, SGP_NOALLOC);
270 /* Our caller expects us to return -EFAULT if we failed to find folio */
271 if (ret == -ENOENT)
272 ret = -EFAULT;
273 if (ret)
274 goto out;
275 if (!folio) {
276 ret = -EFAULT;
277 goto out;
278 }
279
280 page = folio_file_page(folio, pgoff);
281 if (PageHWPoison(page)) {
282 ret = -EIO;
283 goto out_release;
284 }
285
286 ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
287 page, false, wp_copy);
288 if (ret)
289 goto out_release;
290
291 folio_unlock(folio);
292 ret = 0;
293out:
294 return ret;
295out_release:
296 folio_unlock(folio);
297 folio_put(folio);
298 goto out;
299}
300
301static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
302{
303 pgd_t *pgd;
304 p4d_t *p4d;
305 pud_t *pud;
306
307 pgd = pgd_offset(mm, address);
308 p4d = p4d_alloc(mm, pgd, address);
309 if (!p4d)
310 return NULL;
311 pud = pud_alloc(mm, p4d, address);
312 if (!pud)
313 return NULL;
314 /*
315 * Note that we didn't run this because the pmd was
316 * missing, the *pmd may be already established and in
317 * turn it may also be a trans_huge_pmd.
318 */
319 return pmd_alloc(mm, pud, address);
320}
321
322#ifdef CONFIG_HUGETLB_PAGE
323/*
324 * __mcopy_atomic processing for HUGETLB vmas. Note that this routine is
325 * called with mmap_lock held, it will release mmap_lock before returning.
326 */
327static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
328 struct vm_area_struct *dst_vma,
329 unsigned long dst_start,
330 unsigned long src_start,
331 unsigned long len,
332 enum mcopy_atomic_mode mode,
333 bool wp_copy)
334{
335 int vm_shared = dst_vma->vm_flags & VM_SHARED;
336 ssize_t err;
337 pte_t *dst_pte;
338 unsigned long src_addr, dst_addr;
339 long copied;
340 struct page *page;
341 unsigned long vma_hpagesize;
342 pgoff_t idx;
343 u32 hash;
344 struct address_space *mapping;
345
346 /*
347 * There is no default zero huge page for all huge page sizes as
348 * supported by hugetlb. A PMD_SIZE huge pages may exist as used
349 * by THP. Since we can not reliably insert a zero page, this
350 * feature is not supported.
351 */
352 if (mode == MCOPY_ATOMIC_ZEROPAGE) {
353 mmap_read_unlock(dst_mm);
354 return -EINVAL;
355 }
356
357 src_addr = src_start;
358 dst_addr = dst_start;
359 copied = 0;
360 page = NULL;
361 vma_hpagesize = vma_kernel_pagesize(dst_vma);
362
363 /*
364 * Validate alignment based on huge page size
365 */
366 err = -EINVAL;
367 if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
368 goto out_unlock;
369
370retry:
371 /*
372 * On routine entry dst_vma is set. If we had to drop mmap_lock and
373 * retry, dst_vma will be set to NULL and we must lookup again.
374 */
375 if (!dst_vma) {
376 err = -ENOENT;
377 dst_vma = find_dst_vma(dst_mm, dst_start, len);
378 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
379 goto out_unlock;
380
381 err = -EINVAL;
382 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
383 goto out_unlock;
384
385 vm_shared = dst_vma->vm_flags & VM_SHARED;
386 }
387
388 /*
389 * If not shared, ensure the dst_vma has a anon_vma.
390 */
391 err = -ENOMEM;
392 if (!vm_shared) {
393 if (unlikely(anon_vma_prepare(dst_vma)))
394 goto out_unlock;
395 }
396
397 while (src_addr < src_start + len) {
398 BUG_ON(dst_addr >= dst_start + len);
399
400 /*
401 * Serialize via vma_lock and hugetlb_fault_mutex.
402 * vma_lock ensures the dst_pte remains valid even
403 * in the case of shared pmds. fault mutex prevents
404 * races with other faulting threads.
405 */
406 idx = linear_page_index(dst_vma, dst_addr);
407 mapping = dst_vma->vm_file->f_mapping;
408 hash = hugetlb_fault_mutex_hash(mapping, idx);
409 mutex_lock(&hugetlb_fault_mutex_table[hash]);
410 hugetlb_vma_lock_read(dst_vma);
411
412 err = -ENOMEM;
413 dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize);
414 if (!dst_pte) {
415 hugetlb_vma_unlock_read(dst_vma);
416 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
417 goto out_unlock;
418 }
419
420 if (mode != MCOPY_ATOMIC_CONTINUE &&
421 !huge_pte_none_mostly(huge_ptep_get(dst_pte))) {
422 err = -EEXIST;
423 hugetlb_vma_unlock_read(dst_vma);
424 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
425 goto out_unlock;
426 }
427
428 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
429 dst_addr, src_addr, mode, &page,
430 wp_copy);
431
432 hugetlb_vma_unlock_read(dst_vma);
433 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
434
435 cond_resched();
436
437 if (unlikely(err == -ENOENT)) {
438 mmap_read_unlock(dst_mm);
439 BUG_ON(!page);
440
441 err = copy_huge_page_from_user(page,
442 (const void __user *)src_addr,
443 vma_hpagesize / PAGE_SIZE,
444 true);
445 if (unlikely(err)) {
446 err = -EFAULT;
447 goto out;
448 }
449 mmap_read_lock(dst_mm);
450
451 dst_vma = NULL;
452 goto retry;
453 } else
454 BUG_ON(page);
455
456 if (!err) {
457 dst_addr += vma_hpagesize;
458 src_addr += vma_hpagesize;
459 copied += vma_hpagesize;
460
461 if (fatal_signal_pending(current))
462 err = -EINTR;
463 }
464 if (err)
465 break;
466 }
467
468out_unlock:
469 mmap_read_unlock(dst_mm);
470out:
471 if (page)
472 put_page(page);
473 BUG_ON(copied < 0);
474 BUG_ON(err > 0);
475 BUG_ON(!copied && !err);
476 return copied ? copied : err;
477}
478#else /* !CONFIG_HUGETLB_PAGE */
479/* fail at build time if gcc attempts to use this */
480extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
481 struct vm_area_struct *dst_vma,
482 unsigned long dst_start,
483 unsigned long src_start,
484 unsigned long len,
485 enum mcopy_atomic_mode mode,
486 bool wp_copy);
487#endif /* CONFIG_HUGETLB_PAGE */
488
489static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
490 pmd_t *dst_pmd,
491 struct vm_area_struct *dst_vma,
492 unsigned long dst_addr,
493 unsigned long src_addr,
494 struct page **page,
495 enum mcopy_atomic_mode mode,
496 bool wp_copy)
497{
498 ssize_t err;
499
500 if (mode == MCOPY_ATOMIC_CONTINUE) {
501 return mcontinue_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
502 wp_copy);
503 }
504
505 /*
506 * The normal page fault path for a shmem will invoke the
507 * fault, fill the hole in the file and COW it right away. The
508 * result generates plain anonymous memory. So when we are
509 * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
510 * generate anonymous memory directly without actually filling
511 * the hole. For the MAP_PRIVATE case the robustness check
512 * only happens in the pagetable (to verify it's still none)
513 * and not in the radix tree.
514 */
515 if (!(dst_vma->vm_flags & VM_SHARED)) {
516 if (mode == MCOPY_ATOMIC_NORMAL)
517 err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
518 dst_addr, src_addr, page,
519 wp_copy);
520 else
521 err = mfill_zeropage_pte(dst_mm, dst_pmd,
522 dst_vma, dst_addr);
523 } else {
524 err = shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
525 dst_addr, src_addr,
526 mode != MCOPY_ATOMIC_NORMAL,
527 wp_copy, page);
528 }
529
530 return err;
531}
532
533static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
534 unsigned long dst_start,
535 unsigned long src_start,
536 unsigned long len,
537 enum mcopy_atomic_mode mcopy_mode,
538 atomic_t *mmap_changing,
539 __u64 mode)
540{
541 struct vm_area_struct *dst_vma;
542 ssize_t err;
543 pmd_t *dst_pmd;
544 unsigned long src_addr, dst_addr;
545 long copied;
546 struct page *page;
547 bool wp_copy;
548
549 /*
550 * Sanitize the command parameters:
551 */
552 BUG_ON(dst_start & ~PAGE_MASK);
553 BUG_ON(len & ~PAGE_MASK);
554
555 /* Does the address range wrap, or is the span zero-sized? */
556 BUG_ON(src_start + len <= src_start);
557 BUG_ON(dst_start + len <= dst_start);
558
559 src_addr = src_start;
560 dst_addr = dst_start;
561 copied = 0;
562 page = NULL;
563retry:
564 mmap_read_lock(dst_mm);
565
566 /*
567 * If memory mappings are changing because of non-cooperative
568 * operation (e.g. mremap) running in parallel, bail out and
569 * request the user to retry later
570 */
571 err = -EAGAIN;
572 if (mmap_changing && atomic_read(mmap_changing))
573 goto out_unlock;
574
575 /*
576 * Make sure the vma is not shared, that the dst range is
577 * both valid and fully within a single existing vma.
578 */
579 err = -ENOENT;
580 dst_vma = find_dst_vma(dst_mm, dst_start, len);
581 if (!dst_vma)
582 goto out_unlock;
583
584 err = -EINVAL;
585 /*
586 * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
587 * it will overwrite vm_ops, so vma_is_anonymous must return false.
588 */
589 if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
590 dst_vma->vm_flags & VM_SHARED))
591 goto out_unlock;
592
593 /*
594 * validate 'mode' now that we know the dst_vma: don't allow
595 * a wrprotect copy if the userfaultfd didn't register as WP.
596 */
597 wp_copy = mode & UFFDIO_COPY_MODE_WP;
598 if (wp_copy && !(dst_vma->vm_flags & VM_UFFD_WP))
599 goto out_unlock;
600
601 /*
602 * If this is a HUGETLB vma, pass off to appropriate routine
603 */
604 if (is_vm_hugetlb_page(dst_vma))
605 return __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
606 src_start, len, mcopy_mode,
607 wp_copy);
608
609 if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
610 goto out_unlock;
611 if (!vma_is_shmem(dst_vma) && mcopy_mode == MCOPY_ATOMIC_CONTINUE)
612 goto out_unlock;
613
614 /*
615 * Ensure the dst_vma has a anon_vma or this page
616 * would get a NULL anon_vma when moved in the
617 * dst_vma.
618 */
619 err = -ENOMEM;
620 if (!(dst_vma->vm_flags & VM_SHARED) &&
621 unlikely(anon_vma_prepare(dst_vma)))
622 goto out_unlock;
623
624 while (src_addr < src_start + len) {
625 pmd_t dst_pmdval;
626
627 BUG_ON(dst_addr >= dst_start + len);
628
629 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
630 if (unlikely(!dst_pmd)) {
631 err = -ENOMEM;
632 break;
633 }
634
635 dst_pmdval = pmdp_get_lockless(dst_pmd);
636 /*
637 * If the dst_pmd is mapped as THP don't
638 * override it and just be strict.
639 */
640 if (unlikely(pmd_trans_huge(dst_pmdval))) {
641 err = -EEXIST;
642 break;
643 }
644 if (unlikely(pmd_none(dst_pmdval)) &&
645 unlikely(__pte_alloc(dst_mm, dst_pmd))) {
646 err = -ENOMEM;
647 break;
648 }
649 /* If an huge pmd materialized from under us fail */
650 if (unlikely(pmd_trans_huge(*dst_pmd))) {
651 err = -EFAULT;
652 break;
653 }
654
655 BUG_ON(pmd_none(*dst_pmd));
656 BUG_ON(pmd_trans_huge(*dst_pmd));
657
658 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
659 src_addr, &page, mcopy_mode, wp_copy);
660 cond_resched();
661
662 if (unlikely(err == -ENOENT)) {
663 void *page_kaddr;
664
665 mmap_read_unlock(dst_mm);
666 BUG_ON(!page);
667
668 page_kaddr = kmap_local_page(page);
669 err = copy_from_user(page_kaddr,
670 (const void __user *) src_addr,
671 PAGE_SIZE);
672 kunmap_local(page_kaddr);
673 if (unlikely(err)) {
674 err = -EFAULT;
675 goto out;
676 }
677 flush_dcache_page(page);
678 goto retry;
679 } else
680 BUG_ON(page);
681
682 if (!err) {
683 dst_addr += PAGE_SIZE;
684 src_addr += PAGE_SIZE;
685 copied += PAGE_SIZE;
686
687 if (fatal_signal_pending(current))
688 err = -EINTR;
689 }
690 if (err)
691 break;
692 }
693
694out_unlock:
695 mmap_read_unlock(dst_mm);
696out:
697 if (page)
698 put_page(page);
699 BUG_ON(copied < 0);
700 BUG_ON(err > 0);
701 BUG_ON(!copied && !err);
702 return copied ? copied : err;
703}
704
705ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
706 unsigned long src_start, unsigned long len,
707 atomic_t *mmap_changing, __u64 mode)
708{
709 return __mcopy_atomic(dst_mm, dst_start, src_start, len,
710 MCOPY_ATOMIC_NORMAL, mmap_changing, mode);
711}
712
713ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
714 unsigned long len, atomic_t *mmap_changing)
715{
716 return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_ZEROPAGE,
717 mmap_changing, 0);
718}
719
720ssize_t mcopy_continue(struct mm_struct *dst_mm, unsigned long start,
721 unsigned long len, atomic_t *mmap_changing)
722{
723 return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_CONTINUE,
724 mmap_changing, 0);
725}
726
727void uffd_wp_range(struct mm_struct *dst_mm, struct vm_area_struct *dst_vma,
728 unsigned long start, unsigned long len, bool enable_wp)
729{
730 struct mmu_gather tlb;
731 pgprot_t newprot;
732
733 if (enable_wp)
734 newprot = vm_get_page_prot(dst_vma->vm_flags & ~(VM_WRITE));
735 else
736 newprot = vm_get_page_prot(dst_vma->vm_flags);
737
738 tlb_gather_mmu(&tlb, dst_mm);
739 change_protection(&tlb, dst_vma, start, start + len, newprot,
740 enable_wp ? MM_CP_UFFD_WP : MM_CP_UFFD_WP_RESOLVE);
741 tlb_finish_mmu(&tlb);
742}
743
744int mwriteprotect_range(struct mm_struct *dst_mm, unsigned long start,
745 unsigned long len, bool enable_wp,
746 atomic_t *mmap_changing)
747{
748 struct vm_area_struct *dst_vma;
749 unsigned long page_mask;
750 int err;
751
752 /*
753 * Sanitize the command parameters:
754 */
755 BUG_ON(start & ~PAGE_MASK);
756 BUG_ON(len & ~PAGE_MASK);
757
758 /* Does the address range wrap, or is the span zero-sized? */
759 BUG_ON(start + len <= start);
760
761 mmap_read_lock(dst_mm);
762
763 /*
764 * If memory mappings are changing because of non-cooperative
765 * operation (e.g. mremap) running in parallel, bail out and
766 * request the user to retry later
767 */
768 err = -EAGAIN;
769 if (mmap_changing && atomic_read(mmap_changing))
770 goto out_unlock;
771
772 err = -ENOENT;
773 dst_vma = find_dst_vma(dst_mm, start, len);
774
775 if (!dst_vma)
776 goto out_unlock;
777 if (!userfaultfd_wp(dst_vma))
778 goto out_unlock;
779 if (!vma_can_userfault(dst_vma, dst_vma->vm_flags))
780 goto out_unlock;
781
782 if (is_vm_hugetlb_page(dst_vma)) {
783 err = -EINVAL;
784 page_mask = vma_kernel_pagesize(dst_vma) - 1;
785 if ((start & page_mask) || (len & page_mask))
786 goto out_unlock;
787 }
788
789 uffd_wp_range(dst_mm, dst_vma, start, len, enable_wp);
790
791 err = 0;
792out_unlock:
793 mmap_read_unlock(dst_mm);
794 return err;
795}