Commit | Line | Data |
---|---|---|
20c8ccb1 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
c1a4de99 AA |
2 | /* |
3 | * mm/userfaultfd.c | |
4 | * | |
5 | * Copyright (C) 2015 Red Hat, Inc. | |
c1a4de99 AA |
6 | */ |
7 | ||
8 | #include <linux/mm.h> | |
174cd4b1 | 9 | #include <linux/sched/signal.h> |
c1a4de99 AA |
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> | |
60d4d2d2 | 16 | #include <linux/hugetlb.h> |
26071ced | 17 | #include <linux/shmem_fs.h> |
c1a4de99 AA |
18 | #include <asm/tlbflush.h> |
19 | #include "internal.h" | |
20 | ||
643aa36e WY |
21 | static __always_inline |
22 | struct 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 | ||
15313257 AR |
51 | /* |
52 | * Install PTEs, to map dst_addr (within dst_vma) to page. | |
53 | * | |
7d64ae3a AR |
54 | * This function handles both MCOPY_ATOMIC_NORMAL and _CONTINUE for both shmem |
55 | * and anon, and for both shared and private VMAs. | |
15313257 | 56 | */ |
7d64ae3a AR |
57 | int 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) | |
15313257 AR |
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; | |
118 | out_unlock: | |
119 | pte_unmap_unlock(dst_pte, ptl); | |
120 | return ret; | |
121 | } | |
122 | ||
c1a4de99 AA |
123 | static 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, | |
b6ebaedb | 127 | unsigned long src_addr, |
72981e0e AA |
128 | struct page **pagep, |
129 | bool wp_copy) | |
c1a4de99 | 130 | { |
c1a4de99 AA |
131 | void *page_kaddr; |
132 | int ret; | |
b6ebaedb | 133 | struct page *page; |
c1a4de99 | 134 | |
b6ebaedb AA |
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 | ||
c1e8d7c6 | 147 | /* fallback to copy_from_user outside mmap_lock */ |
b6ebaedb | 148 | if (unlikely(ret)) { |
9e368259 | 149 | ret = -ENOENT; |
b6ebaedb AA |
150 | *pagep = page; |
151 | /* don't free the page */ | |
152 | goto out; | |
153 | } | |
154 | } else { | |
155 | page = *pagep; | |
156 | *pagep = NULL; | |
157 | } | |
c1a4de99 AA |
158 | |
159 | /* | |
160 | * The memory barrier inside __SetPageUptodate makes sure that | |
f4f5329d | 161 | * preceding stores to the page contents become visible before |
c1a4de99 AA |
162 | * the set_pte_at() write. |
163 | */ | |
164 | __SetPageUptodate(page); | |
165 | ||
166 | ret = -ENOMEM; | |
d9eb1ea2 | 167 | if (mem_cgroup_charge(page, dst_mm, GFP_KERNEL)) |
c1a4de99 AA |
168 | goto out_release; |
169 | ||
15313257 AR |
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; | |
c1a4de99 AA |
174 | out: |
175 | return ret; | |
c1a4de99 | 176 | out_release: |
09cbfeaf | 177 | put_page(page); |
c1a4de99 | 178 | goto out; |
c1a4de99 AA |
179 | } |
180 | ||
181 | static 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; | |
e2a50c1f AA |
189 | pgoff_t offset, max_off; |
190 | struct inode *inode; | |
c1a4de99 AA |
191 | |
192 | _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr), | |
193 | dst_vma->vm_page_prot)); | |
c1a4de99 | 194 | dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl); |
e2a50c1f AA |
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; | |
c1a4de99 AA |
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; | |
211 | out_unlock: | |
212 | pte_unmap_unlock(dst_pte, ptl); | |
213 | return ret; | |
214 | } | |
215 | ||
15313257 AR |
216 | /* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */ |
217 | static 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; | |
243 | out: | |
244 | return ret; | |
245 | out_release: | |
246 | unlock_page(page); | |
247 | put_page(page); | |
248 | goto out; | |
249 | } | |
250 | ||
c1a4de99 AA |
251 | static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address) |
252 | { | |
253 | pgd_t *pgd; | |
c2febafc | 254 | p4d_t *p4d; |
c1a4de99 | 255 | pud_t *pud; |
c1a4de99 AA |
256 | |
257 | pgd = pgd_offset(mm, address); | |
c2febafc KS |
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); | |
c1a4de99 AA |
270 | } |
271 | ||
60d4d2d2 MK |
272 | #ifdef CONFIG_HUGETLB_PAGE |
273 | /* | |
274 | * __mcopy_atomic processing for HUGETLB vmas. Note that this routine is | |
c1e8d7c6 | 275 | * called with mmap_lock held, it will release mmap_lock before returning. |
60d4d2d2 MK |
276 | */ |
277 | static __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, | |
f6191471 | 282 | enum mcopy_atomic_mode mode) |
60d4d2d2 | 283 | { |
1c9e8def | 284 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
60d4d2d2 MK |
285 | ssize_t err; |
286 | pte_t *dst_pte; | |
287 | unsigned long src_addr, dst_addr; | |
288 | long copied; | |
289 | struct page *page; | |
60d4d2d2 MK |
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 | */ | |
f6191471 | 301 | if (mode == MCOPY_ATOMIC_ZEROPAGE) { |
d8ed45c5 | 302 | mmap_read_unlock(dst_mm); |
60d4d2d2 MK |
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 | ||
319 | retry: | |
320 | /* | |
c1e8d7c6 | 321 | * On routine entry dst_vma is set. If we had to drop mmap_lock and |
60d4d2d2 MK |
322 | * retry, dst_vma will be set to NULL and we must lookup again. |
323 | */ | |
324 | if (!dst_vma) { | |
27d02568 | 325 | err = -ENOENT; |
643aa36e | 326 | dst_vma = find_dst_vma(dst_mm, dst_start, len); |
60d4d2d2 MK |
327 | if (!dst_vma || !is_vm_hugetlb_page(dst_vma)) |
328 | goto out_unlock; | |
1c9e8def | 329 | |
27d02568 MR |
330 | err = -EINVAL; |
331 | if (vma_hpagesize != vma_kernel_pagesize(dst_vma)) | |
332 | goto out_unlock; | |
333 | ||
1c9e8def | 334 | vm_shared = dst_vma->vm_flags & VM_SHARED; |
60d4d2d2 MK |
335 | } |
336 | ||
60d4d2d2 | 337 | /* |
1c9e8def | 338 | * If not shared, ensure the dst_vma has a anon_vma. |
60d4d2d2 MK |
339 | */ |
340 | err = -ENOMEM; | |
1c9e8def MK |
341 | if (!vm_shared) { |
342 | if (unlikely(anon_vma_prepare(dst_vma))) | |
343 | goto out_unlock; | |
344 | } | |
60d4d2d2 | 345 | |
60d4d2d2 | 346 | while (src_addr < src_start + len) { |
60d4d2d2 | 347 | BUG_ON(dst_addr >= dst_start + len); |
60d4d2d2 MK |
348 | |
349 | /* | |
c0d0381a MK |
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. | |
60d4d2d2 | 354 | */ |
ddeaab32 | 355 | mapping = dst_vma->vm_file->f_mapping; |
c0d0381a MK |
356 | i_mmap_lock_read(mapping); |
357 | idx = linear_page_index(dst_vma, dst_addr); | |
188b04a7 | 358 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
60d4d2d2 MK |
359 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
360 | ||
361 | err = -ENOMEM; | |
aec44e0f | 362 | dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize); |
60d4d2d2 MK |
363 | if (!dst_pte) { |
364 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
c0d0381a | 365 | i_mmap_unlock_read(mapping); |
60d4d2d2 MK |
366 | goto out_unlock; |
367 | } | |
368 | ||
f6191471 AR |
369 | if (mode != MCOPY_ATOMIC_CONTINUE && |
370 | !huge_pte_none(huge_ptep_get(dst_pte))) { | |
371 | err = -EEXIST; | |
60d4d2d2 | 372 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 373 | i_mmap_unlock_read(mapping); |
60d4d2d2 MK |
374 | goto out_unlock; |
375 | } | |
376 | ||
377 | err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma, | |
f6191471 | 378 | dst_addr, src_addr, mode, &page); |
60d4d2d2 MK |
379 | |
380 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
c0d0381a | 381 | i_mmap_unlock_read(mapping); |
60d4d2d2 MK |
382 | |
383 | cond_resched(); | |
384 | ||
9e368259 | 385 | if (unlikely(err == -ENOENT)) { |
d8ed45c5 | 386 | mmap_read_unlock(dst_mm); |
60d4d2d2 MK |
387 | BUG_ON(!page); |
388 | ||
389 | err = copy_huge_page_from_user(page, | |
390 | (const void __user *)src_addr, | |
4fb07ee6 WY |
391 | vma_hpagesize / PAGE_SIZE, |
392 | true); | |
60d4d2d2 MK |
393 | if (unlikely(err)) { |
394 | err = -EFAULT; | |
395 | goto out; | |
396 | } | |
d8ed45c5 | 397 | mmap_read_lock(dst_mm); |
60d4d2d2 MK |
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 | ||
416 | out_unlock: | |
d8ed45c5 | 417 | mmap_read_unlock(dst_mm); |
60d4d2d2 | 418 | out: |
8cc5fcbb | 419 | if (page) |
60d4d2d2 MK |
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 */ | |
428 | extern 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, | |
f6191471 | 433 | enum mcopy_atomic_mode mode); |
60d4d2d2 MK |
434 | #endif /* CONFIG_HUGETLB_PAGE */ |
435 | ||
3217d3c7 MR |
436 | static __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, | |
15313257 | 442 | enum mcopy_atomic_mode mode, |
72981e0e | 443 | bool wp_copy) |
3217d3c7 MR |
444 | { |
445 | ssize_t err; | |
446 | ||
15313257 AR |
447 | if (mode == MCOPY_ATOMIC_CONTINUE) { |
448 | return mcontinue_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr, | |
449 | wp_copy); | |
450 | } | |
451 | ||
5b51072e AA |
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)) { | |
15313257 | 463 | if (mode == MCOPY_ATOMIC_NORMAL) |
3217d3c7 | 464 | err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma, |
72981e0e AA |
465 | dst_addr, src_addr, page, |
466 | wp_copy); | |
3217d3c7 MR |
467 | else |
468 | err = mfill_zeropage_pte(dst_mm, dst_pmd, | |
469 | dst_vma, dst_addr); | |
470 | } else { | |
72981e0e | 471 | VM_WARN_ON_ONCE(wp_copy); |
3460f6e5 | 472 | err = shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, |
15313257 AR |
473 | dst_addr, src_addr, |
474 | mode != MCOPY_ATOMIC_NORMAL, | |
3460f6e5 | 475 | page); |
3217d3c7 MR |
476 | } |
477 | ||
478 | return err; | |
479 | } | |
480 | ||
c1a4de99 AA |
481 | static __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, | |
f6191471 | 485 | enum mcopy_atomic_mode mcopy_mode, |
72981e0e AA |
486 | bool *mmap_changing, |
487 | __u64 mode) | |
c1a4de99 AA |
488 | { |
489 | struct vm_area_struct *dst_vma; | |
490 | ssize_t err; | |
491 | pmd_t *dst_pmd; | |
492 | unsigned long src_addr, dst_addr; | |
b6ebaedb AA |
493 | long copied; |
494 | struct page *page; | |
72981e0e | 495 | bool wp_copy; |
c1a4de99 AA |
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 | ||
b6ebaedb AA |
507 | src_addr = src_start; |
508 | dst_addr = dst_start; | |
509 | copied = 0; | |
510 | page = NULL; | |
511 | retry: | |
d8ed45c5 | 512 | mmap_read_lock(dst_mm); |
c1a4de99 | 513 | |
df2cc96e MR |
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 | ||
c1a4de99 AA |
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 | */ | |
27d02568 | 527 | err = -ENOENT; |
643aa36e | 528 | dst_vma = find_dst_vma(dst_mm, dst_start, len); |
26071ced MR |
529 | if (!dst_vma) |
530 | goto out_unlock; | |
c1a4de99 | 531 | |
27d02568 MR |
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 | ||
72981e0e AA |
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 | ||
60d4d2d2 MK |
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, | |
f6191471 | 554 | src_start, len, mcopy_mode); |
60d4d2d2 | 555 | |
26071ced | 556 | if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma)) |
b6ebaedb | 557 | goto out_unlock; |
15313257 | 558 | if (!vma_is_shmem(dst_vma) && mcopy_mode == MCOPY_ATOMIC_CONTINUE) |
f6191471 | 559 | goto out_unlock; |
c1a4de99 AA |
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; | |
5b51072e AA |
567 | if (!(dst_vma->vm_flags & VM_SHARED) && |
568 | unlikely(anon_vma_prepare(dst_vma))) | |
b6ebaedb | 569 | goto out_unlock; |
c1a4de99 | 570 | |
b6ebaedb | 571 | while (src_addr < src_start + len) { |
c1a4de99 | 572 | pmd_t dst_pmdval; |
b6ebaedb | 573 | |
c1a4de99 | 574 | BUG_ON(dst_addr >= dst_start + len); |
b6ebaedb | 575 | |
c1a4de99 AA |
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)) && | |
4cf58924 | 592 | unlikely(__pte_alloc(dst_mm, dst_pmd))) { |
c1a4de99 AA |
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 | ||
3217d3c7 | 605 | err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr, |
15313257 | 606 | src_addr, &page, mcopy_mode, wp_copy); |
c1a4de99 AA |
607 | cond_resched(); |
608 | ||
9e368259 | 609 | if (unlikely(err == -ENOENT)) { |
b6ebaedb AA |
610 | void *page_kaddr; |
611 | ||
d8ed45c5 | 612 | mmap_read_unlock(dst_mm); |
b6ebaedb AA |
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 | ||
c1a4de99 AA |
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 | ||
b6ebaedb | 640 | out_unlock: |
d8ed45c5 | 641 | mmap_read_unlock(dst_mm); |
b6ebaedb AA |
642 | out: |
643 | if (page) | |
09cbfeaf | 644 | put_page(page); |
c1a4de99 AA |
645 | BUG_ON(copied < 0); |
646 | BUG_ON(err > 0); | |
647 | BUG_ON(!copied && !err); | |
648 | return copied ? copied : err; | |
649 | } | |
650 | ||
651 | ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start, | |
df2cc96e | 652 | unsigned long src_start, unsigned long len, |
72981e0e | 653 | bool *mmap_changing, __u64 mode) |
c1a4de99 | 654 | { |
f6191471 AR |
655 | return __mcopy_atomic(dst_mm, dst_start, src_start, len, |
656 | MCOPY_ATOMIC_NORMAL, mmap_changing, mode); | |
c1a4de99 AA |
657 | } |
658 | ||
659 | ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start, | |
df2cc96e | 660 | unsigned long len, bool *mmap_changing) |
c1a4de99 | 661 | { |
f6191471 AR |
662 | return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_ZEROPAGE, |
663 | mmap_changing, 0); | |
664 | } | |
665 | ||
666 | ssize_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); | |
c1a4de99 | 671 | } |
ffd05793 SL |
672 | |
673 | int 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 | ||
d8ed45c5 | 689 | mmap_read_lock(dst_mm); |
ffd05793 SL |
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; | |
722 | out_unlock: | |
d8ed45c5 | 723 | mmap_read_unlock(dst_mm); |
ffd05793 SL |
724 | return err; |
725 | } |