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 "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
51/*
52 * Install PTEs, to map dst_addr (within dst_vma) to page.
53 *
54 * This function handles both MCOPY_ATOMIC_NORMAL and _CONTINUE for both shmem
55 * and anon, and for both shared and private VMAs.
56 */
57int mfill_atomic_install_pte(struct mm_struct *dst_mm, pmd_t *dst_pmd,
58 struct vm_area_struct *dst_vma,
59 unsigned long dst_addr, struct page *page,
60 bool newly_allocated, bool wp_copy)
61{
62 int ret;
63 pte_t _dst_pte, *dst_pte;
64 bool writable = dst_vma->vm_flags & VM_WRITE;
65 bool vm_shared = dst_vma->vm_flags & VM_SHARED;
66 bool page_in_cache = page->mapping;
67 spinlock_t *ptl;
68 struct inode *inode;
69 pgoff_t offset, max_off;
70
71 _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
72 if (page_in_cache && !vm_shared)
73 writable = false;
74 if (writable || !page_in_cache)
75 _dst_pte = pte_mkdirty(_dst_pte);
76 if (writable) {
77 if (wp_copy)
78 _dst_pte = pte_mkuffd_wp(_dst_pte);
79 else
80 _dst_pte = pte_mkwrite(_dst_pte);
81 }
82
83 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
84
85 if (vma_is_shmem(dst_vma)) {
86 /* serialize against truncate with the page table lock */
87 inode = dst_vma->vm_file->f_inode;
88 offset = linear_page_index(dst_vma, dst_addr);
89 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
90 ret = -EFAULT;
91 if (unlikely(offset >= max_off))
92 goto out_unlock;
93 }
94
95 ret = -EEXIST;
96 if (!pte_none(*dst_pte))
97 goto out_unlock;
98
99 if (page_in_cache)
100 page_add_file_rmap(page, false);
101 else
102 page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
103
104 /*
105 * Must happen after rmap, as mm_counter() checks mapping (via
106 * PageAnon()), which is set by __page_set_anon_rmap().
107 */
108 inc_mm_counter(dst_mm, mm_counter(page));
109
110 if (newly_allocated)
111 lru_cache_add_inactive_or_unevictable(page, dst_vma);
112
113 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
114
115 /* No need to invalidate - it was non-present before */
116 update_mmu_cache(dst_vma, dst_addr, dst_pte);
117 ret = 0;
118out_unlock:
119 pte_unmap_unlock(dst_pte, ptl);
120 return ret;
121}
122
123static int mcopy_atomic_pte(struct mm_struct *dst_mm,
124 pmd_t *dst_pmd,
125 struct vm_area_struct *dst_vma,
126 unsigned long dst_addr,
127 unsigned long src_addr,
128 struct page **pagep,
129 bool wp_copy)
130{
131 void *page_kaddr;
132 int ret;
133 struct page *page;
134
135 if (!*pagep) {
136 ret = -ENOMEM;
137 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
138 if (!page)
139 goto out;
140
141 page_kaddr = kmap_atomic(page);
142 ret = copy_from_user(page_kaddr,
143 (const void __user *) src_addr,
144 PAGE_SIZE);
145 kunmap_atomic(page_kaddr);
146
147 /* fallback to copy_from_user outside mmap_lock */
148 if (unlikely(ret)) {
149 ret = -ENOENT;
150 *pagep = page;
151 /* don't free the page */
152 goto out;
153 }
154 } else {
155 page = *pagep;
156 *pagep = NULL;
157 }
158
159 /*
160 * The memory barrier inside __SetPageUptodate makes sure that
161 * preceding stores to the page contents become visible before
162 * the set_pte_at() write.
163 */
164 __SetPageUptodate(page);
165
166 ret = -ENOMEM;
167 if (mem_cgroup_charge(page, dst_mm, GFP_KERNEL))
168 goto out_release;
169
170 ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
171 page, true, wp_copy);
172 if (ret)
173 goto out_release;
174out:
175 return ret;
176out_release:
177 put_page(page);
178 goto out;
179}
180
181static int mfill_zeropage_pte(struct mm_struct *dst_mm,
182 pmd_t *dst_pmd,
183 struct vm_area_struct *dst_vma,
184 unsigned long dst_addr)
185{
186 pte_t _dst_pte, *dst_pte;
187 spinlock_t *ptl;
188 int ret;
189 pgoff_t offset, max_off;
190 struct inode *inode;
191
192 _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
193 dst_vma->vm_page_prot));
194 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
195 if (dst_vma->vm_file) {
196 /* the shmem MAP_PRIVATE case requires checking the i_size */
197 inode = dst_vma->vm_file->f_inode;
198 offset = linear_page_index(dst_vma, dst_addr);
199 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
200 ret = -EFAULT;
201 if (unlikely(offset >= max_off))
202 goto out_unlock;
203 }
204 ret = -EEXIST;
205 if (!pte_none(*dst_pte))
206 goto out_unlock;
207 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
208 /* No need to invalidate - it was non-present before */
209 update_mmu_cache(dst_vma, dst_addr, dst_pte);
210 ret = 0;
211out_unlock:
212 pte_unmap_unlock(dst_pte, ptl);
213 return ret;
214}
215
216/* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */
217static int mcontinue_atomic_pte(struct mm_struct *dst_mm,
218 pmd_t *dst_pmd,
219 struct vm_area_struct *dst_vma,
220 unsigned long dst_addr,
221 bool wp_copy)
222{
223 struct inode *inode = file_inode(dst_vma->vm_file);
224 pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
225 struct page *page;
226 int ret;
227
228 ret = shmem_getpage(inode, pgoff, &page, SGP_READ);
229 if (ret)
230 goto out;
231 if (!page) {
232 ret = -EFAULT;
233 goto out;
234 }
235
236 ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
237 page, false, wp_copy);
238 if (ret)
239 goto out_release;
240
241 unlock_page(page);
242 ret = 0;
243out:
244 return ret;
245out_release:
246 unlock_page(page);
247 put_page(page);
248 goto out;
249}
250
251static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
252{
253 pgd_t *pgd;
254 p4d_t *p4d;
255 pud_t *pud;
256
257 pgd = pgd_offset(mm, address);
258 p4d = p4d_alloc(mm, pgd, address);
259 if (!p4d)
260 return NULL;
261 pud = pud_alloc(mm, p4d, address);
262 if (!pud)
263 return NULL;
264 /*
265 * Note that we didn't run this because the pmd was
266 * missing, the *pmd may be already established and in
267 * turn it may also be a trans_huge_pmd.
268 */
269 return pmd_alloc(mm, pud, address);
270}
271
272#ifdef CONFIG_HUGETLB_PAGE
273/*
274 * __mcopy_atomic processing for HUGETLB vmas. Note that this routine is
275 * called with mmap_lock held, it will release mmap_lock before returning.
276 */
277static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
278 struct vm_area_struct *dst_vma,
279 unsigned long dst_start,
280 unsigned long src_start,
281 unsigned long len,
282 enum mcopy_atomic_mode mode)
283{
284 int vm_shared = dst_vma->vm_flags & VM_SHARED;
285 ssize_t err;
286 pte_t *dst_pte;
287 unsigned long src_addr, dst_addr;
288 long copied;
289 struct page *page;
290 unsigned long vma_hpagesize;
291 pgoff_t idx;
292 u32 hash;
293 struct address_space *mapping;
294
295 /*
296 * There is no default zero huge page for all huge page sizes as
297 * supported by hugetlb. A PMD_SIZE huge pages may exist as used
298 * by THP. Since we can not reliably insert a zero page, this
299 * feature is not supported.
300 */
301 if (mode == MCOPY_ATOMIC_ZEROPAGE) {
302 mmap_read_unlock(dst_mm);
303 return -EINVAL;
304 }
305
306 src_addr = src_start;
307 dst_addr = dst_start;
308 copied = 0;
309 page = NULL;
310 vma_hpagesize = vma_kernel_pagesize(dst_vma);
311
312 /*
313 * Validate alignment based on huge page size
314 */
315 err = -EINVAL;
316 if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
317 goto out_unlock;
318
319retry:
320 /*
321 * On routine entry dst_vma is set. If we had to drop mmap_lock and
322 * retry, dst_vma will be set to NULL and we must lookup again.
323 */
324 if (!dst_vma) {
325 err = -ENOENT;
326 dst_vma = find_dst_vma(dst_mm, dst_start, len);
327 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
328 goto out_unlock;
329
330 err = -EINVAL;
331 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
332 goto out_unlock;
333
334 vm_shared = dst_vma->vm_flags & VM_SHARED;
335 }
336
337 /*
338 * If not shared, ensure the dst_vma has a anon_vma.
339 */
340 err = -ENOMEM;
341 if (!vm_shared) {
342 if (unlikely(anon_vma_prepare(dst_vma)))
343 goto out_unlock;
344 }
345
346 while (src_addr < src_start + len) {
347 BUG_ON(dst_addr >= dst_start + len);
348
349 /*
350 * Serialize via i_mmap_rwsem and hugetlb_fault_mutex.
351 * i_mmap_rwsem ensures the dst_pte remains valid even
352 * in the case of shared pmds. fault mutex prevents
353 * races with other faulting threads.
354 */
355 mapping = dst_vma->vm_file->f_mapping;
356 i_mmap_lock_read(mapping);
357 idx = linear_page_index(dst_vma, dst_addr);
358 hash = hugetlb_fault_mutex_hash(mapping, idx);
359 mutex_lock(&hugetlb_fault_mutex_table[hash]);
360
361 err = -ENOMEM;
362 dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize);
363 if (!dst_pte) {
364 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
365 i_mmap_unlock_read(mapping);
366 goto out_unlock;
367 }
368
369 if (mode != MCOPY_ATOMIC_CONTINUE &&
370 !huge_pte_none(huge_ptep_get(dst_pte))) {
371 err = -EEXIST;
372 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
373 i_mmap_unlock_read(mapping);
374 goto out_unlock;
375 }
376
377 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
378 dst_addr, src_addr, mode, &page);
379
380 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
381 i_mmap_unlock_read(mapping);
382
383 cond_resched();
384
385 if (unlikely(err == -ENOENT)) {
386 mmap_read_unlock(dst_mm);
387 BUG_ON(!page);
388
389 err = copy_huge_page_from_user(page,
390 (const void __user *)src_addr,
391 vma_hpagesize / PAGE_SIZE,
392 true);
393 if (unlikely(err)) {
394 err = -EFAULT;
395 goto out;
396 }
397 mmap_read_lock(dst_mm);
398
399 dst_vma = NULL;
400 goto retry;
401 } else
402 BUG_ON(page);
403
404 if (!err) {
405 dst_addr += vma_hpagesize;
406 src_addr += vma_hpagesize;
407 copied += vma_hpagesize;
408
409 if (fatal_signal_pending(current))
410 err = -EINTR;
411 }
412 if (err)
413 break;
414 }
415
416out_unlock:
417 mmap_read_unlock(dst_mm);
418out:
419 if (page)
420 put_page(page);
421 BUG_ON(copied < 0);
422 BUG_ON(err > 0);
423 BUG_ON(!copied && !err);
424 return copied ? copied : err;
425}
426#else /* !CONFIG_HUGETLB_PAGE */
427/* fail at build time if gcc attempts to use this */
428extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
429 struct vm_area_struct *dst_vma,
430 unsigned long dst_start,
431 unsigned long src_start,
432 unsigned long len,
433 enum mcopy_atomic_mode mode);
434#endif /* CONFIG_HUGETLB_PAGE */
435
436static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
437 pmd_t *dst_pmd,
438 struct vm_area_struct *dst_vma,
439 unsigned long dst_addr,
440 unsigned long src_addr,
441 struct page **page,
442 enum mcopy_atomic_mode mode,
443 bool wp_copy)
444{
445 ssize_t err;
446
447 if (mode == MCOPY_ATOMIC_CONTINUE) {
448 return mcontinue_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
449 wp_copy);
450 }
451
452 /*
453 * The normal page fault path for a shmem will invoke the
454 * fault, fill the hole in the file and COW it right away. The
455 * result generates plain anonymous memory. So when we are
456 * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
457 * generate anonymous memory directly without actually filling
458 * the hole. For the MAP_PRIVATE case the robustness check
459 * only happens in the pagetable (to verify it's still none)
460 * and not in the radix tree.
461 */
462 if (!(dst_vma->vm_flags & VM_SHARED)) {
463 if (mode == MCOPY_ATOMIC_NORMAL)
464 err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
465 dst_addr, src_addr, page,
466 wp_copy);
467 else
468 err = mfill_zeropage_pte(dst_mm, dst_pmd,
469 dst_vma, dst_addr);
470 } else {
471 VM_WARN_ON_ONCE(wp_copy);
472 err = shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
473 dst_addr, src_addr,
474 mode != MCOPY_ATOMIC_NORMAL,
475 page);
476 }
477
478 return err;
479}
480
481static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
482 unsigned long dst_start,
483 unsigned long src_start,
484 unsigned long len,
485 enum mcopy_atomic_mode mcopy_mode,
486 bool *mmap_changing,
487 __u64 mode)
488{
489 struct vm_area_struct *dst_vma;
490 ssize_t err;
491 pmd_t *dst_pmd;
492 unsigned long src_addr, dst_addr;
493 long copied;
494 struct page *page;
495 bool wp_copy;
496
497 /*
498 * Sanitize the command parameters:
499 */
500 BUG_ON(dst_start & ~PAGE_MASK);
501 BUG_ON(len & ~PAGE_MASK);
502
503 /* Does the address range wrap, or is the span zero-sized? */
504 BUG_ON(src_start + len <= src_start);
505 BUG_ON(dst_start + len <= dst_start);
506
507 src_addr = src_start;
508 dst_addr = dst_start;
509 copied = 0;
510 page = NULL;
511retry:
512 mmap_read_lock(dst_mm);
513
514 /*
515 * If memory mappings are changing because of non-cooperative
516 * operation (e.g. mremap) running in parallel, bail out and
517 * request the user to retry later
518 */
519 err = -EAGAIN;
520 if (mmap_changing && READ_ONCE(*mmap_changing))
521 goto out_unlock;
522
523 /*
524 * Make sure the vma is not shared, that the dst range is
525 * both valid and fully within a single existing vma.
526 */
527 err = -ENOENT;
528 dst_vma = find_dst_vma(dst_mm, dst_start, len);
529 if (!dst_vma)
530 goto out_unlock;
531
532 err = -EINVAL;
533 /*
534 * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
535 * it will overwrite vm_ops, so vma_is_anonymous must return false.
536 */
537 if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
538 dst_vma->vm_flags & VM_SHARED))
539 goto out_unlock;
540
541 /*
542 * validate 'mode' now that we know the dst_vma: don't allow
543 * a wrprotect copy if the userfaultfd didn't register as WP.
544 */
545 wp_copy = mode & UFFDIO_COPY_MODE_WP;
546 if (wp_copy && !(dst_vma->vm_flags & VM_UFFD_WP))
547 goto out_unlock;
548
549 /*
550 * If this is a HUGETLB vma, pass off to appropriate routine
551 */
552 if (is_vm_hugetlb_page(dst_vma))
553 return __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
554 src_start, len, mcopy_mode);
555
556 if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
557 goto out_unlock;
558 if (!vma_is_shmem(dst_vma) && mcopy_mode == MCOPY_ATOMIC_CONTINUE)
559 goto out_unlock;
560
561 /*
562 * Ensure the dst_vma has a anon_vma or this page
563 * would get a NULL anon_vma when moved in the
564 * dst_vma.
565 */
566 err = -ENOMEM;
567 if (!(dst_vma->vm_flags & VM_SHARED) &&
568 unlikely(anon_vma_prepare(dst_vma)))
569 goto out_unlock;
570
571 while (src_addr < src_start + len) {
572 pmd_t dst_pmdval;
573
574 BUG_ON(dst_addr >= dst_start + len);
575
576 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
577 if (unlikely(!dst_pmd)) {
578 err = -ENOMEM;
579 break;
580 }
581
582 dst_pmdval = pmd_read_atomic(dst_pmd);
583 /*
584 * If the dst_pmd is mapped as THP don't
585 * override it and just be strict.
586 */
587 if (unlikely(pmd_trans_huge(dst_pmdval))) {
588 err = -EEXIST;
589 break;
590 }
591 if (unlikely(pmd_none(dst_pmdval)) &&
592 unlikely(__pte_alloc(dst_mm, dst_pmd))) {
593 err = -ENOMEM;
594 break;
595 }
596 /* If an huge pmd materialized from under us fail */
597 if (unlikely(pmd_trans_huge(*dst_pmd))) {
598 err = -EFAULT;
599 break;
600 }
601
602 BUG_ON(pmd_none(*dst_pmd));
603 BUG_ON(pmd_trans_huge(*dst_pmd));
604
605 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
606 src_addr, &page, mcopy_mode, wp_copy);
607 cond_resched();
608
609 if (unlikely(err == -ENOENT)) {
610 void *page_kaddr;
611
612 mmap_read_unlock(dst_mm);
613 BUG_ON(!page);
614
615 page_kaddr = kmap(page);
616 err = copy_from_user(page_kaddr,
617 (const void __user *) src_addr,
618 PAGE_SIZE);
619 kunmap(page);
620 if (unlikely(err)) {
621 err = -EFAULT;
622 goto out;
623 }
624 goto retry;
625 } else
626 BUG_ON(page);
627
628 if (!err) {
629 dst_addr += PAGE_SIZE;
630 src_addr += PAGE_SIZE;
631 copied += PAGE_SIZE;
632
633 if (fatal_signal_pending(current))
634 err = -EINTR;
635 }
636 if (err)
637 break;
638 }
639
640out_unlock:
641 mmap_read_unlock(dst_mm);
642out:
643 if (page)
644 put_page(page);
645 BUG_ON(copied < 0);
646 BUG_ON(err > 0);
647 BUG_ON(!copied && !err);
648 return copied ? copied : err;
649}
650
651ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
652 unsigned long src_start, unsigned long len,
653 bool *mmap_changing, __u64 mode)
654{
655 return __mcopy_atomic(dst_mm, dst_start, src_start, len,
656 MCOPY_ATOMIC_NORMAL, mmap_changing, mode);
657}
658
659ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
660 unsigned long len, bool *mmap_changing)
661{
662 return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_ZEROPAGE,
663 mmap_changing, 0);
664}
665
666ssize_t mcopy_continue(struct mm_struct *dst_mm, unsigned long start,
667 unsigned long len, bool *mmap_changing)
668{
669 return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_CONTINUE,
670 mmap_changing, 0);
671}
672
673int mwriteprotect_range(struct mm_struct *dst_mm, unsigned long start,
674 unsigned long len, bool enable_wp, bool *mmap_changing)
675{
676 struct vm_area_struct *dst_vma;
677 pgprot_t newprot;
678 int err;
679
680 /*
681 * Sanitize the command parameters:
682 */
683 BUG_ON(start & ~PAGE_MASK);
684 BUG_ON(len & ~PAGE_MASK);
685
686 /* Does the address range wrap, or is the span zero-sized? */
687 BUG_ON(start + len <= start);
688
689 mmap_read_lock(dst_mm);
690
691 /*
692 * If memory mappings are changing because of non-cooperative
693 * operation (e.g. mremap) running in parallel, bail out and
694 * request the user to retry later
695 */
696 err = -EAGAIN;
697 if (mmap_changing && READ_ONCE(*mmap_changing))
698 goto out_unlock;
699
700 err = -ENOENT;
701 dst_vma = find_dst_vma(dst_mm, start, len);
702 /*
703 * Make sure the vma is not shared, that the dst range is
704 * both valid and fully within a single existing vma.
705 */
706 if (!dst_vma || (dst_vma->vm_flags & VM_SHARED))
707 goto out_unlock;
708 if (!userfaultfd_wp(dst_vma))
709 goto out_unlock;
710 if (!vma_is_anonymous(dst_vma))
711 goto out_unlock;
712
713 if (enable_wp)
714 newprot = vm_get_page_prot(dst_vma->vm_flags & ~(VM_WRITE));
715 else
716 newprot = vm_get_page_prot(dst_vma->vm_flags);
717
718 change_protection(dst_vma, start, start + len, newprot,
719 enable_wp ? MM_CP_UFFD_WP : MM_CP_UFFD_WP_RESOLVE);
720
721 err = 0;
722out_unlock:
723 mmap_read_unlock(dst_mm);
724 return err;
725}