xref: /openbmc/linux/mm/userfaultfd.c (revision 3c6b4bcf)
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 
22 static __always_inline
find_dst_vma(struct mm_struct * dst_mm,unsigned long dst_start,unsigned long len)23 struct 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 (!range_in_vma(dst_vma, dst_start, dst_start + len))
35 		return NULL;
36 
37 	/*
38 	 * Check the vma is registered in uffd, this is required to
39 	 * enforce the VM_MAYWRITE check done at uffd registration
40 	 * time.
41 	 */
42 	if (!dst_vma->vm_userfaultfd_ctx.ctx)
43 		return NULL;
44 
45 	return dst_vma;
46 }
47 
48 /* Check if dst_addr is outside of file's size. Must be called with ptl held. */
mfill_file_over_size(struct vm_area_struct * dst_vma,unsigned long dst_addr)49 static bool mfill_file_over_size(struct vm_area_struct *dst_vma,
50 				 unsigned long dst_addr)
51 {
52 	struct inode *inode;
53 	pgoff_t offset, max_off;
54 
55 	if (!dst_vma->vm_file)
56 		return false;
57 
58 	inode = dst_vma->vm_file->f_inode;
59 	offset = linear_page_index(dst_vma, dst_addr);
60 	max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
61 	return offset >= max_off;
62 }
63 
64 /*
65  * Install PTEs, to map dst_addr (within dst_vma) to page.
66  *
67  * This function handles both MCOPY_ATOMIC_NORMAL and _CONTINUE for both shmem
68  * and anon, and for both shared and private VMAs.
69  */
mfill_atomic_install_pte(pmd_t * dst_pmd,struct vm_area_struct * dst_vma,unsigned long dst_addr,struct page * page,bool newly_allocated,uffd_flags_t flags)70 int mfill_atomic_install_pte(pmd_t *dst_pmd,
71 			     struct vm_area_struct *dst_vma,
72 			     unsigned long dst_addr, struct page *page,
73 			     bool newly_allocated, uffd_flags_t flags)
74 {
75 	int ret;
76 	struct mm_struct *dst_mm = dst_vma->vm_mm;
77 	pte_t _dst_pte, *dst_pte;
78 	bool writable = dst_vma->vm_flags & VM_WRITE;
79 	bool vm_shared = dst_vma->vm_flags & VM_SHARED;
80 	bool page_in_cache = page_mapping(page);
81 	spinlock_t *ptl;
82 	struct folio *folio;
83 
84 	_dst_pte = mk_pte(page, dst_vma->vm_page_prot);
85 	_dst_pte = pte_mkdirty(_dst_pte);
86 	if (page_in_cache && !vm_shared)
87 		writable = false;
88 	if (writable)
89 		_dst_pte = pte_mkwrite(_dst_pte, dst_vma);
90 	if (flags & MFILL_ATOMIC_WP)
91 		_dst_pte = pte_mkuffd_wp(_dst_pte);
92 
93 	ret = -EAGAIN;
94 	dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
95 	if (!dst_pte)
96 		goto out;
97 
98 	if (mfill_file_over_size(dst_vma, dst_addr)) {
99 		ret = -EFAULT;
100 		goto out_unlock;
101 	}
102 
103 	ret = -EEXIST;
104 	/*
105 	 * We allow to overwrite a pte marker: consider when both MISSING|WP
106 	 * registered, we firstly wr-protect a none pte which has no page cache
107 	 * page backing it, then access the page.
108 	 */
109 	if (!pte_none_mostly(ptep_get(dst_pte)))
110 		goto out_unlock;
111 
112 	folio = page_folio(page);
113 	if (page_in_cache) {
114 		/* Usually, cache pages are already added to LRU */
115 		if (newly_allocated)
116 			folio_add_lru(folio);
117 		page_add_file_rmap(page, dst_vma, false);
118 	} else {
119 		page_add_new_anon_rmap(page, dst_vma, dst_addr);
120 		folio_add_lru_vma(folio, dst_vma);
121 	}
122 
123 	/*
124 	 * Must happen after rmap, as mm_counter() checks mapping (via
125 	 * PageAnon()), which is set by __page_set_anon_rmap().
126 	 */
127 	inc_mm_counter(dst_mm, mm_counter(page));
128 
129 	set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
130 
131 	/* No need to invalidate - it was non-present before */
132 	update_mmu_cache(dst_vma, dst_addr, dst_pte);
133 	ret = 0;
134 out_unlock:
135 	pte_unmap_unlock(dst_pte, ptl);
136 out:
137 	return ret;
138 }
139 
mfill_atomic_pte_copy(pmd_t * dst_pmd,struct vm_area_struct * dst_vma,unsigned long dst_addr,unsigned long src_addr,uffd_flags_t flags,struct folio ** foliop)140 static int mfill_atomic_pte_copy(pmd_t *dst_pmd,
141 				 struct vm_area_struct *dst_vma,
142 				 unsigned long dst_addr,
143 				 unsigned long src_addr,
144 				 uffd_flags_t flags,
145 				 struct folio **foliop)
146 {
147 	void *kaddr;
148 	int ret;
149 	struct folio *folio;
150 
151 	if (!*foliop) {
152 		ret = -ENOMEM;
153 		folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, dst_vma,
154 					dst_addr, false);
155 		if (!folio)
156 			goto out;
157 
158 		kaddr = kmap_local_folio(folio, 0);
159 		/*
160 		 * The read mmap_lock is held here.  Despite the
161 		 * mmap_lock being read recursive a deadlock is still
162 		 * possible if a writer has taken a lock.  For example:
163 		 *
164 		 * process A thread 1 takes read lock on own mmap_lock
165 		 * process A thread 2 calls mmap, blocks taking write lock
166 		 * process B thread 1 takes page fault, read lock on own mmap lock
167 		 * process B thread 2 calls mmap, blocks taking write lock
168 		 * process A thread 1 blocks taking read lock on process B
169 		 * process B thread 1 blocks taking read lock on process A
170 		 *
171 		 * Disable page faults to prevent potential deadlock
172 		 * and retry the copy outside the mmap_lock.
173 		 */
174 		pagefault_disable();
175 		ret = copy_from_user(kaddr, (const void __user *) src_addr,
176 				     PAGE_SIZE);
177 		pagefault_enable();
178 		kunmap_local(kaddr);
179 
180 		/* fallback to copy_from_user outside mmap_lock */
181 		if (unlikely(ret)) {
182 			ret = -ENOENT;
183 			*foliop = folio;
184 			/* don't free the page */
185 			goto out;
186 		}
187 
188 		flush_dcache_folio(folio);
189 	} else {
190 		folio = *foliop;
191 		*foliop = NULL;
192 	}
193 
194 	/*
195 	 * The memory barrier inside __folio_mark_uptodate makes sure that
196 	 * preceding stores to the page contents become visible before
197 	 * the set_pte_at() write.
198 	 */
199 	__folio_mark_uptodate(folio);
200 
201 	ret = -ENOMEM;
202 	if (mem_cgroup_charge(folio, dst_vma->vm_mm, GFP_KERNEL))
203 		goto out_release;
204 
205 	ret = mfill_atomic_install_pte(dst_pmd, dst_vma, dst_addr,
206 				       &folio->page, true, flags);
207 	if (ret)
208 		goto out_release;
209 out:
210 	return ret;
211 out_release:
212 	folio_put(folio);
213 	goto out;
214 }
215 
mfill_atomic_pte_zeroed_folio(pmd_t * dst_pmd,struct vm_area_struct * dst_vma,unsigned long dst_addr)216 static int mfill_atomic_pte_zeroed_folio(pmd_t *dst_pmd,
217 					 struct vm_area_struct *dst_vma,
218 					 unsigned long dst_addr)
219 {
220 	struct folio *folio;
221 	int ret = -ENOMEM;
222 
223 	folio = vma_alloc_zeroed_movable_folio(dst_vma, dst_addr);
224 	if (!folio)
225 		return ret;
226 
227 	if (mem_cgroup_charge(folio, dst_vma->vm_mm, GFP_KERNEL))
228 		goto out_put;
229 
230 	/*
231 	 * The memory barrier inside __folio_mark_uptodate makes sure that
232 	 * zeroing out the folio become visible before mapping the page
233 	 * using set_pte_at(). See do_anonymous_page().
234 	 */
235 	__folio_mark_uptodate(folio);
236 
237 	ret = mfill_atomic_install_pte(dst_pmd, dst_vma, dst_addr,
238 				       &folio->page, true, 0);
239 	if (ret)
240 		goto out_put;
241 
242 	return 0;
243 out_put:
244 	folio_put(folio);
245 	return ret;
246 }
247 
mfill_atomic_pte_zeropage(pmd_t * dst_pmd,struct vm_area_struct * dst_vma,unsigned long dst_addr)248 static int mfill_atomic_pte_zeropage(pmd_t *dst_pmd,
249 				     struct vm_area_struct *dst_vma,
250 				     unsigned long dst_addr)
251 {
252 	pte_t _dst_pte, *dst_pte;
253 	spinlock_t *ptl;
254 	int ret;
255 
256 	if (mm_forbids_zeropage(dst_vma->vm_mm))
257 		return mfill_atomic_pte_zeroed_folio(dst_pmd, dst_vma, dst_addr);
258 
259 	_dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
260 					 dst_vma->vm_page_prot));
261 	ret = -EAGAIN;
262 	dst_pte = pte_offset_map_lock(dst_vma->vm_mm, dst_pmd, dst_addr, &ptl);
263 	if (!dst_pte)
264 		goto out;
265 	if (mfill_file_over_size(dst_vma, dst_addr)) {
266 		ret = -EFAULT;
267 		goto out_unlock;
268 	}
269 	ret = -EEXIST;
270 	if (!pte_none(ptep_get(dst_pte)))
271 		goto out_unlock;
272 	set_pte_at(dst_vma->vm_mm, dst_addr, dst_pte, _dst_pte);
273 	/* No need to invalidate - it was non-present before */
274 	update_mmu_cache(dst_vma, dst_addr, dst_pte);
275 	ret = 0;
276 out_unlock:
277 	pte_unmap_unlock(dst_pte, ptl);
278 out:
279 	return ret;
280 }
281 
282 /* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */
mfill_atomic_pte_continue(pmd_t * dst_pmd,struct vm_area_struct * dst_vma,unsigned long dst_addr,uffd_flags_t flags)283 static int mfill_atomic_pte_continue(pmd_t *dst_pmd,
284 				     struct vm_area_struct *dst_vma,
285 				     unsigned long dst_addr,
286 				     uffd_flags_t flags)
287 {
288 	struct inode *inode = file_inode(dst_vma->vm_file);
289 	pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
290 	struct folio *folio;
291 	struct page *page;
292 	int ret;
293 
294 	ret = shmem_get_folio(inode, pgoff, &folio, SGP_NOALLOC);
295 	/* Our caller expects us to return -EFAULT if we failed to find folio */
296 	if (ret == -ENOENT)
297 		ret = -EFAULT;
298 	if (ret)
299 		goto out;
300 	if (!folio) {
301 		ret = -EFAULT;
302 		goto out;
303 	}
304 
305 	page = folio_file_page(folio, pgoff);
306 	if (PageHWPoison(page)) {
307 		ret = -EIO;
308 		goto out_release;
309 	}
310 
311 	ret = mfill_atomic_install_pte(dst_pmd, dst_vma, dst_addr,
312 				       page, false, flags);
313 	if (ret)
314 		goto out_release;
315 
316 	folio_unlock(folio);
317 	ret = 0;
318 out:
319 	return ret;
320 out_release:
321 	folio_unlock(folio);
322 	folio_put(folio);
323 	goto out;
324 }
325 
326 /* Handles UFFDIO_POISON for all non-hugetlb VMAs. */
mfill_atomic_pte_poison(pmd_t * dst_pmd,struct vm_area_struct * dst_vma,unsigned long dst_addr,uffd_flags_t flags)327 static int mfill_atomic_pte_poison(pmd_t *dst_pmd,
328 				   struct vm_area_struct *dst_vma,
329 				   unsigned long dst_addr,
330 				   uffd_flags_t flags)
331 {
332 	int ret;
333 	struct mm_struct *dst_mm = dst_vma->vm_mm;
334 	pte_t _dst_pte, *dst_pte;
335 	spinlock_t *ptl;
336 
337 	_dst_pte = make_pte_marker(PTE_MARKER_POISONED);
338 	ret = -EAGAIN;
339 	dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
340 	if (!dst_pte)
341 		goto out;
342 
343 	if (mfill_file_over_size(dst_vma, dst_addr)) {
344 		ret = -EFAULT;
345 		goto out_unlock;
346 	}
347 
348 	ret = -EEXIST;
349 	/* Refuse to overwrite any PTE, even a PTE marker (e.g. UFFD WP). */
350 	if (!pte_none(*dst_pte))
351 		goto out_unlock;
352 
353 	set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
354 
355 	/* No need to invalidate - it was non-present before */
356 	update_mmu_cache(dst_vma, dst_addr, dst_pte);
357 	ret = 0;
358 out_unlock:
359 	pte_unmap_unlock(dst_pte, ptl);
360 out:
361 	return ret;
362 }
363 
mm_alloc_pmd(struct mm_struct * mm,unsigned long address)364 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
365 {
366 	pgd_t *pgd;
367 	p4d_t *p4d;
368 	pud_t *pud;
369 
370 	pgd = pgd_offset(mm, address);
371 	p4d = p4d_alloc(mm, pgd, address);
372 	if (!p4d)
373 		return NULL;
374 	pud = pud_alloc(mm, p4d, address);
375 	if (!pud)
376 		return NULL;
377 	/*
378 	 * Note that we didn't run this because the pmd was
379 	 * missing, the *pmd may be already established and in
380 	 * turn it may also be a trans_huge_pmd.
381 	 */
382 	return pmd_alloc(mm, pud, address);
383 }
384 
385 #ifdef CONFIG_HUGETLB_PAGE
386 /*
387  * mfill_atomic processing for HUGETLB vmas.  Note that this routine is
388  * called with mmap_lock held, it will release mmap_lock before returning.
389  */
mfill_atomic_hugetlb(struct vm_area_struct * dst_vma,unsigned long dst_start,unsigned long src_start,unsigned long len,atomic_t * mmap_changing,uffd_flags_t flags)390 static __always_inline ssize_t mfill_atomic_hugetlb(
391 					      struct vm_area_struct *dst_vma,
392 					      unsigned long dst_start,
393 					      unsigned long src_start,
394 					      unsigned long len,
395 					      atomic_t *mmap_changing,
396 					      uffd_flags_t flags)
397 {
398 	struct mm_struct *dst_mm = dst_vma->vm_mm;
399 	int vm_shared = dst_vma->vm_flags & VM_SHARED;
400 	ssize_t err;
401 	pte_t *dst_pte;
402 	unsigned long src_addr, dst_addr;
403 	long copied;
404 	struct folio *folio;
405 	unsigned long vma_hpagesize;
406 	pgoff_t idx;
407 	u32 hash;
408 	struct address_space *mapping;
409 
410 	/*
411 	 * There is no default zero huge page for all huge page sizes as
412 	 * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
413 	 * by THP.  Since we can not reliably insert a zero page, this
414 	 * feature is not supported.
415 	 */
416 	if (uffd_flags_mode_is(flags, MFILL_ATOMIC_ZEROPAGE)) {
417 		mmap_read_unlock(dst_mm);
418 		return -EINVAL;
419 	}
420 
421 	src_addr = src_start;
422 	dst_addr = dst_start;
423 	copied = 0;
424 	folio = NULL;
425 	vma_hpagesize = vma_kernel_pagesize(dst_vma);
426 
427 	/*
428 	 * Validate alignment based on huge page size
429 	 */
430 	err = -EINVAL;
431 	if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
432 		goto out_unlock;
433 
434 retry:
435 	/*
436 	 * On routine entry dst_vma is set.  If we had to drop mmap_lock and
437 	 * retry, dst_vma will be set to NULL and we must lookup again.
438 	 */
439 	if (!dst_vma) {
440 		err = -ENOENT;
441 		dst_vma = find_dst_vma(dst_mm, dst_start, len);
442 		if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
443 			goto out_unlock;
444 
445 		err = -EINVAL;
446 		if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
447 			goto out_unlock;
448 
449 		vm_shared = dst_vma->vm_flags & VM_SHARED;
450 	}
451 
452 	/*
453 	 * If not shared, ensure the dst_vma has a anon_vma.
454 	 */
455 	err = -ENOMEM;
456 	if (!vm_shared) {
457 		if (unlikely(anon_vma_prepare(dst_vma)))
458 			goto out_unlock;
459 	}
460 
461 	while (src_addr < src_start + len) {
462 		BUG_ON(dst_addr >= dst_start + len);
463 
464 		/*
465 		 * Serialize via vma_lock and hugetlb_fault_mutex.
466 		 * vma_lock ensures the dst_pte remains valid even
467 		 * in the case of shared pmds.  fault mutex prevents
468 		 * races with other faulting threads.
469 		 */
470 		idx = linear_page_index(dst_vma, dst_addr);
471 		mapping = dst_vma->vm_file->f_mapping;
472 		hash = hugetlb_fault_mutex_hash(mapping, idx);
473 		mutex_lock(&hugetlb_fault_mutex_table[hash]);
474 		hugetlb_vma_lock_read(dst_vma);
475 
476 		err = -ENOMEM;
477 		dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize);
478 		if (!dst_pte) {
479 			hugetlb_vma_unlock_read(dst_vma);
480 			mutex_unlock(&hugetlb_fault_mutex_table[hash]);
481 			goto out_unlock;
482 		}
483 
484 		if (!uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE) &&
485 		    !huge_pte_none_mostly(huge_ptep_get(dst_pte))) {
486 			err = -EEXIST;
487 			hugetlb_vma_unlock_read(dst_vma);
488 			mutex_unlock(&hugetlb_fault_mutex_table[hash]);
489 			goto out_unlock;
490 		}
491 
492 		err = hugetlb_mfill_atomic_pte(dst_pte, dst_vma, dst_addr,
493 					       src_addr, flags, &folio);
494 
495 		hugetlb_vma_unlock_read(dst_vma);
496 		mutex_unlock(&hugetlb_fault_mutex_table[hash]);
497 
498 		cond_resched();
499 
500 		if (unlikely(err == -ENOENT)) {
501 			mmap_read_unlock(dst_mm);
502 			BUG_ON(!folio);
503 
504 			err = copy_folio_from_user(folio,
505 						   (const void __user *)src_addr, true);
506 			if (unlikely(err)) {
507 				err = -EFAULT;
508 				goto out;
509 			}
510 			mmap_read_lock(dst_mm);
511 			/*
512 			 * If memory mappings are changing because of non-cooperative
513 			 * operation (e.g. mremap) running in parallel, bail out and
514 			 * request the user to retry later
515 			 */
516 			if (mmap_changing && atomic_read(mmap_changing)) {
517 				err = -EAGAIN;
518 				break;
519 			}
520 
521 			dst_vma = NULL;
522 			goto retry;
523 		} else
524 			BUG_ON(folio);
525 
526 		if (!err) {
527 			dst_addr += vma_hpagesize;
528 			src_addr += vma_hpagesize;
529 			copied += vma_hpagesize;
530 
531 			if (fatal_signal_pending(current))
532 				err = -EINTR;
533 		}
534 		if (err)
535 			break;
536 	}
537 
538 out_unlock:
539 	mmap_read_unlock(dst_mm);
540 out:
541 	if (folio)
542 		folio_put(folio);
543 	BUG_ON(copied < 0);
544 	BUG_ON(err > 0);
545 	BUG_ON(!copied && !err);
546 	return copied ? copied : err;
547 }
548 #else /* !CONFIG_HUGETLB_PAGE */
549 /* fail at build time if gcc attempts to use this */
550 extern ssize_t mfill_atomic_hugetlb(struct vm_area_struct *dst_vma,
551 				    unsigned long dst_start,
552 				    unsigned long src_start,
553 				    unsigned long len,
554 				    atomic_t *mmap_changing,
555 				    uffd_flags_t flags);
556 #endif /* CONFIG_HUGETLB_PAGE */
557 
mfill_atomic_pte(pmd_t * dst_pmd,struct vm_area_struct * dst_vma,unsigned long dst_addr,unsigned long src_addr,uffd_flags_t flags,struct folio ** foliop)558 static __always_inline ssize_t mfill_atomic_pte(pmd_t *dst_pmd,
559 						struct vm_area_struct *dst_vma,
560 						unsigned long dst_addr,
561 						unsigned long src_addr,
562 						uffd_flags_t flags,
563 						struct folio **foliop)
564 {
565 	ssize_t err;
566 
567 	if (uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE)) {
568 		return mfill_atomic_pte_continue(dst_pmd, dst_vma,
569 						 dst_addr, flags);
570 	} else if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) {
571 		return mfill_atomic_pte_poison(dst_pmd, dst_vma,
572 					       dst_addr, flags);
573 	}
574 
575 	/*
576 	 * The normal page fault path for a shmem will invoke the
577 	 * fault, fill the hole in the file and COW it right away. The
578 	 * result generates plain anonymous memory. So when we are
579 	 * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
580 	 * generate anonymous memory directly without actually filling
581 	 * the hole. For the MAP_PRIVATE case the robustness check
582 	 * only happens in the pagetable (to verify it's still none)
583 	 * and not in the radix tree.
584 	 */
585 	if (!(dst_vma->vm_flags & VM_SHARED)) {
586 		if (uffd_flags_mode_is(flags, MFILL_ATOMIC_COPY))
587 			err = mfill_atomic_pte_copy(dst_pmd, dst_vma,
588 						    dst_addr, src_addr,
589 						    flags, foliop);
590 		else
591 			err = mfill_atomic_pte_zeropage(dst_pmd,
592 						 dst_vma, dst_addr);
593 	} else {
594 		err = shmem_mfill_atomic_pte(dst_pmd, dst_vma,
595 					     dst_addr, src_addr,
596 					     flags, foliop);
597 	}
598 
599 	return err;
600 }
601 
mfill_atomic(struct mm_struct * dst_mm,unsigned long dst_start,unsigned long src_start,unsigned long len,atomic_t * mmap_changing,uffd_flags_t flags)602 static __always_inline ssize_t mfill_atomic(struct mm_struct *dst_mm,
603 					    unsigned long dst_start,
604 					    unsigned long src_start,
605 					    unsigned long len,
606 					    atomic_t *mmap_changing,
607 					    uffd_flags_t flags)
608 {
609 	struct vm_area_struct *dst_vma;
610 	ssize_t err;
611 	pmd_t *dst_pmd;
612 	unsigned long src_addr, dst_addr;
613 	long copied;
614 	struct folio *folio;
615 
616 	/*
617 	 * Sanitize the command parameters:
618 	 */
619 	BUG_ON(dst_start & ~PAGE_MASK);
620 	BUG_ON(len & ~PAGE_MASK);
621 
622 	/* Does the address range wrap, or is the span zero-sized? */
623 	BUG_ON(src_start + len <= src_start);
624 	BUG_ON(dst_start + len <= dst_start);
625 
626 	src_addr = src_start;
627 	dst_addr = dst_start;
628 	copied = 0;
629 	folio = NULL;
630 retry:
631 	mmap_read_lock(dst_mm);
632 
633 	/*
634 	 * If memory mappings are changing because of non-cooperative
635 	 * operation (e.g. mremap) running in parallel, bail out and
636 	 * request the user to retry later
637 	 */
638 	err = -EAGAIN;
639 	if (mmap_changing && atomic_read(mmap_changing))
640 		goto out_unlock;
641 
642 	/*
643 	 * Make sure the vma is not shared, that the dst range is
644 	 * both valid and fully within a single existing vma.
645 	 */
646 	err = -ENOENT;
647 	dst_vma = find_dst_vma(dst_mm, dst_start, len);
648 	if (!dst_vma)
649 		goto out_unlock;
650 
651 	err = -EINVAL;
652 	/*
653 	 * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
654 	 * it will overwrite vm_ops, so vma_is_anonymous must return false.
655 	 */
656 	if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
657 	    dst_vma->vm_flags & VM_SHARED))
658 		goto out_unlock;
659 
660 	/*
661 	 * validate 'mode' now that we know the dst_vma: don't allow
662 	 * a wrprotect copy if the userfaultfd didn't register as WP.
663 	 */
664 	if ((flags & MFILL_ATOMIC_WP) && !(dst_vma->vm_flags & VM_UFFD_WP))
665 		goto out_unlock;
666 
667 	/*
668 	 * If this is a HUGETLB vma, pass off to appropriate routine
669 	 */
670 	if (is_vm_hugetlb_page(dst_vma))
671 		return  mfill_atomic_hugetlb(dst_vma, dst_start, src_start,
672 					     len, mmap_changing, flags);
673 
674 	if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
675 		goto out_unlock;
676 	if (!vma_is_shmem(dst_vma) &&
677 	    uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE))
678 		goto out_unlock;
679 
680 	/*
681 	 * Ensure the dst_vma has a anon_vma or this page
682 	 * would get a NULL anon_vma when moved in the
683 	 * dst_vma.
684 	 */
685 	err = -ENOMEM;
686 	if (!(dst_vma->vm_flags & VM_SHARED) &&
687 	    unlikely(anon_vma_prepare(dst_vma)))
688 		goto out_unlock;
689 
690 	while (src_addr < src_start + len) {
691 		pmd_t dst_pmdval;
692 
693 		BUG_ON(dst_addr >= dst_start + len);
694 
695 		dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
696 		if (unlikely(!dst_pmd)) {
697 			err = -ENOMEM;
698 			break;
699 		}
700 
701 		dst_pmdval = pmdp_get_lockless(dst_pmd);
702 		if (unlikely(pmd_none(dst_pmdval)) &&
703 		    unlikely(__pte_alloc(dst_mm, dst_pmd))) {
704 			err = -ENOMEM;
705 			break;
706 		}
707 		dst_pmdval = pmdp_get_lockless(dst_pmd);
708 		/*
709 		 * If the dst_pmd is THP don't override it and just be strict.
710 		 * (This includes the case where the PMD used to be THP and
711 		 * changed back to none after __pte_alloc().)
712 		 */
713 		if (unlikely(!pmd_present(dst_pmdval) || pmd_trans_huge(dst_pmdval) ||
714 			     pmd_devmap(dst_pmdval))) {
715 			err = -EEXIST;
716 			break;
717 		}
718 		if (unlikely(pmd_bad(dst_pmdval))) {
719 			err = -EFAULT;
720 			break;
721 		}
722 		/*
723 		 * For shmem mappings, khugepaged is allowed to remove page
724 		 * tables under us; pte_offset_map_lock() will deal with that.
725 		 */
726 
727 		err = mfill_atomic_pte(dst_pmd, dst_vma, dst_addr,
728 				       src_addr, flags, &folio);
729 		cond_resched();
730 
731 		if (unlikely(err == -ENOENT)) {
732 			void *kaddr;
733 
734 			mmap_read_unlock(dst_mm);
735 			BUG_ON(!folio);
736 
737 			kaddr = kmap_local_folio(folio, 0);
738 			err = copy_from_user(kaddr,
739 					     (const void __user *) src_addr,
740 					     PAGE_SIZE);
741 			kunmap_local(kaddr);
742 			if (unlikely(err)) {
743 				err = -EFAULT;
744 				goto out;
745 			}
746 			flush_dcache_folio(folio);
747 			goto retry;
748 		} else
749 			BUG_ON(folio);
750 
751 		if (!err) {
752 			dst_addr += PAGE_SIZE;
753 			src_addr += PAGE_SIZE;
754 			copied += PAGE_SIZE;
755 
756 			if (fatal_signal_pending(current))
757 				err = -EINTR;
758 		}
759 		if (err)
760 			break;
761 	}
762 
763 out_unlock:
764 	mmap_read_unlock(dst_mm);
765 out:
766 	if (folio)
767 		folio_put(folio);
768 	BUG_ON(copied < 0);
769 	BUG_ON(err > 0);
770 	BUG_ON(!copied && !err);
771 	return copied ? copied : err;
772 }
773 
mfill_atomic_copy(struct mm_struct * dst_mm,unsigned long dst_start,unsigned long src_start,unsigned long len,atomic_t * mmap_changing,uffd_flags_t flags)774 ssize_t mfill_atomic_copy(struct mm_struct *dst_mm, unsigned long dst_start,
775 			  unsigned long src_start, unsigned long len,
776 			  atomic_t *mmap_changing, uffd_flags_t flags)
777 {
778 	return mfill_atomic(dst_mm, dst_start, src_start, len, mmap_changing,
779 			    uffd_flags_set_mode(flags, MFILL_ATOMIC_COPY));
780 }
781 
mfill_atomic_zeropage(struct mm_struct * dst_mm,unsigned long start,unsigned long len,atomic_t * mmap_changing)782 ssize_t mfill_atomic_zeropage(struct mm_struct *dst_mm, unsigned long start,
783 			      unsigned long len, atomic_t *mmap_changing)
784 {
785 	return mfill_atomic(dst_mm, start, 0, len, mmap_changing,
786 			    uffd_flags_set_mode(0, MFILL_ATOMIC_ZEROPAGE));
787 }
788 
mfill_atomic_continue(struct mm_struct * dst_mm,unsigned long start,unsigned long len,atomic_t * mmap_changing,uffd_flags_t flags)789 ssize_t mfill_atomic_continue(struct mm_struct *dst_mm, unsigned long start,
790 			      unsigned long len, atomic_t *mmap_changing,
791 			      uffd_flags_t flags)
792 {
793 	return mfill_atomic(dst_mm, start, 0, len, mmap_changing,
794 			    uffd_flags_set_mode(flags, MFILL_ATOMIC_CONTINUE));
795 }
796 
mfill_atomic_poison(struct mm_struct * dst_mm,unsigned long start,unsigned long len,atomic_t * mmap_changing,uffd_flags_t flags)797 ssize_t mfill_atomic_poison(struct mm_struct *dst_mm, unsigned long start,
798 			    unsigned long len, atomic_t *mmap_changing,
799 			    uffd_flags_t flags)
800 {
801 	return mfill_atomic(dst_mm, start, 0, len, mmap_changing,
802 			    uffd_flags_set_mode(flags, MFILL_ATOMIC_POISON));
803 }
804 
uffd_wp_range(struct vm_area_struct * dst_vma,unsigned long start,unsigned long len,bool enable_wp)805 long uffd_wp_range(struct vm_area_struct *dst_vma,
806 		   unsigned long start, unsigned long len, bool enable_wp)
807 {
808 	unsigned int mm_cp_flags;
809 	struct mmu_gather tlb;
810 	long ret;
811 
812 	VM_WARN_ONCE(start < dst_vma->vm_start || start + len > dst_vma->vm_end,
813 			"The address range exceeds VMA boundary.\n");
814 	if (enable_wp)
815 		mm_cp_flags = MM_CP_UFFD_WP;
816 	else
817 		mm_cp_flags = MM_CP_UFFD_WP_RESOLVE;
818 
819 	/*
820 	 * vma->vm_page_prot already reflects that uffd-wp is enabled for this
821 	 * VMA (see userfaultfd_set_vm_flags()) and that all PTEs are supposed
822 	 * to be write-protected as default whenever protection changes.
823 	 * Try upgrading write permissions manually.
824 	 */
825 	if (!enable_wp && vma_wants_manual_pte_write_upgrade(dst_vma))
826 		mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
827 	tlb_gather_mmu(&tlb, dst_vma->vm_mm);
828 	ret = change_protection(&tlb, dst_vma, start, start + len, mm_cp_flags);
829 	tlb_finish_mmu(&tlb);
830 
831 	return ret;
832 }
833 
mwriteprotect_range(struct mm_struct * dst_mm,unsigned long start,unsigned long len,bool enable_wp,atomic_t * mmap_changing)834 int mwriteprotect_range(struct mm_struct *dst_mm, unsigned long start,
835 			unsigned long len, bool enable_wp,
836 			atomic_t *mmap_changing)
837 {
838 	unsigned long end = start + len;
839 	unsigned long _start, _end;
840 	struct vm_area_struct *dst_vma;
841 	unsigned long page_mask;
842 	long err;
843 	VMA_ITERATOR(vmi, dst_mm, start);
844 
845 	/*
846 	 * Sanitize the command parameters:
847 	 */
848 	BUG_ON(start & ~PAGE_MASK);
849 	BUG_ON(len & ~PAGE_MASK);
850 
851 	/* Does the address range wrap, or is the span zero-sized? */
852 	BUG_ON(start + len <= start);
853 
854 	mmap_read_lock(dst_mm);
855 
856 	/*
857 	 * If memory mappings are changing because of non-cooperative
858 	 * operation (e.g. mremap) running in parallel, bail out and
859 	 * request the user to retry later
860 	 */
861 	err = -EAGAIN;
862 	if (mmap_changing && atomic_read(mmap_changing))
863 		goto out_unlock;
864 
865 	err = -ENOENT;
866 	for_each_vma_range(vmi, dst_vma, end) {
867 
868 		if (!userfaultfd_wp(dst_vma)) {
869 			err = -ENOENT;
870 			break;
871 		}
872 
873 		if (is_vm_hugetlb_page(dst_vma)) {
874 			err = -EINVAL;
875 			page_mask = vma_kernel_pagesize(dst_vma) - 1;
876 			if ((start & page_mask) || (len & page_mask))
877 				break;
878 		}
879 
880 		_start = max(dst_vma->vm_start, start);
881 		_end = min(dst_vma->vm_end, end);
882 
883 		err = uffd_wp_range(dst_vma, _start, _end - _start, enable_wp);
884 
885 		/* Return 0 on success, <0 on failures */
886 		if (err < 0)
887 			break;
888 		err = 0;
889 	}
890 out_unlock:
891 	mmap_read_unlock(dst_mm);
892 	return err;
893 }
894