xref: /linux/mm/gup.c (revision a00cda3f)
1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
4 #include <linux/err.h>
5 #include <linux/spinlock.h>
6 
7 #include <linux/mm.h>
8 #include <linux/memremap.h>
9 #include <linux/pagemap.h>
10 #include <linux/rmap.h>
11 #include <linux/swap.h>
12 #include <linux/swapops.h>
13 
14 #include <linux/sched/signal.h>
15 #include <linux/rwsem.h>
16 #include <linux/hugetlb.h>
17 #include <linux/migrate.h>
18 #include <linux/mm_inline.h>
19 #include <linux/sched/mm.h>
20 
21 #include <asm/mmu_context.h>
22 #include <asm/tlbflush.h>
23 
24 #include "internal.h"
25 
26 struct follow_page_context {
27 	struct dev_pagemap *pgmap;
28 	unsigned int page_mask;
29 };
30 
hpage_pincount_add(struct page * page,int refs)31 static void hpage_pincount_add(struct page *page, int refs)
32 {
33 	VM_BUG_ON_PAGE(!hpage_pincount_available(page), page);
34 	VM_BUG_ON_PAGE(page != compound_head(page), page);
35 
36 	atomic_add(refs, compound_pincount_ptr(page));
37 }
38 
hpage_pincount_sub(struct page * page,int refs)39 static void hpage_pincount_sub(struct page *page, int refs)
40 {
41 	VM_BUG_ON_PAGE(!hpage_pincount_available(page), page);
42 	VM_BUG_ON_PAGE(page != compound_head(page), page);
43 
44 	atomic_sub(refs, compound_pincount_ptr(page));
45 }
46 
47 /*
48  * Return the compound head page with ref appropriately incremented,
49  * or NULL if that failed.
50  */
try_get_compound_head(struct page * page,int refs)51 static inline struct page *try_get_compound_head(struct page *page, int refs)
52 {
53 	struct page *head = compound_head(page);
54 
55 	if (WARN_ON_ONCE(page_ref_count(head) < 0))
56 		return NULL;
57 	if (unlikely(!page_cache_add_speculative(head, refs)))
58 		return NULL;
59 	return head;
60 }
61 
62 /*
63  * try_grab_compound_head() - attempt to elevate a page's refcount, by a
64  * flags-dependent amount.
65  *
66  * "grab" names in this file mean, "look at flags to decide whether to use
67  * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
68  *
69  * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
70  * same time. (That's true throughout the get_user_pages*() and
71  * pin_user_pages*() APIs.) Cases:
72  *
73  *    FOLL_GET: page's refcount will be incremented by 1.
74  *    FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
75  *
76  * Return: head page (with refcount appropriately incremented) for success, or
77  * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's
78  * considered failure, and furthermore, a likely bug in the caller, so a warning
79  * is also emitted.
80  */
try_grab_compound_head(struct page * page,int refs,unsigned int flags)81 static __maybe_unused struct page *try_grab_compound_head(struct page *page,
82 							  int refs,
83 							  unsigned int flags)
84 {
85 	if (flags & FOLL_GET)
86 		return try_get_compound_head(page, refs);
87 	else if (flags & FOLL_PIN) {
88 		int orig_refs = refs;
89 
90 		/*
91 		 * Can't do FOLL_LONGTERM + FOLL_PIN with CMA in the gup fast
92 		 * path, so fail and let the caller fall back to the slow path.
93 		 */
94 		if (unlikely(flags & FOLL_LONGTERM) &&
95 				is_migrate_cma_page(page))
96 			return NULL;
97 
98 		/*
99 		 * When pinning a compound page of order > 1 (which is what
100 		 * hpage_pincount_available() checks for), use an exact count to
101 		 * track it, via hpage_pincount_add/_sub().
102 		 *
103 		 * However, be sure to *also* increment the normal page refcount
104 		 * field at least once, so that the page really is pinned.
105 		 */
106 		if (!hpage_pincount_available(page))
107 			refs *= GUP_PIN_COUNTING_BIAS;
108 
109 		page = try_get_compound_head(page, refs);
110 		if (!page)
111 			return NULL;
112 
113 		if (hpage_pincount_available(page))
114 			hpage_pincount_add(page, refs);
115 
116 		mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED,
117 				    orig_refs);
118 
119 		return page;
120 	}
121 
122 	WARN_ON_ONCE(1);
123 	return NULL;
124 }
125 
put_compound_head(struct page * page,int refs,unsigned int flags)126 static void put_compound_head(struct page *page, int refs, unsigned int flags)
127 {
128 	if (flags & FOLL_PIN) {
129 		mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED,
130 				    refs);
131 
132 		if (hpage_pincount_available(page))
133 			hpage_pincount_sub(page, refs);
134 		else
135 			refs *= GUP_PIN_COUNTING_BIAS;
136 	}
137 
138 	VM_BUG_ON_PAGE(page_ref_count(page) < refs, page);
139 	/*
140 	 * Calling put_page() for each ref is unnecessarily slow. Only the last
141 	 * ref needs a put_page().
142 	 */
143 	if (refs > 1)
144 		page_ref_sub(page, refs - 1);
145 	put_page(page);
146 }
147 
148 /**
149  * try_grab_page() - elevate a page's refcount by a flag-dependent amount
150  *
151  * This might not do anything at all, depending on the flags argument.
152  *
153  * "grab" names in this file mean, "look at flags to decide whether to use
154  * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
155  *
156  * @page:    pointer to page to be grabbed
157  * @flags:   gup flags: these are the FOLL_* flag values.
158  *
159  * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
160  * time. Cases:
161  *
162  *    FOLL_GET: page's refcount will be incremented by 1.
163  *    FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
164  *
165  * Return: true for success, or if no action was required (if neither FOLL_PIN
166  * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
167  * FOLL_PIN was set, but the page could not be grabbed.
168  */
try_grab_page(struct page * page,unsigned int flags)169 bool __must_check try_grab_page(struct page *page, unsigned int flags)
170 {
171 	WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == (FOLL_GET | FOLL_PIN));
172 
173 	if (flags & FOLL_GET)
174 		return try_get_page(page);
175 	else if (flags & FOLL_PIN) {
176 		int refs = 1;
177 
178 		page = compound_head(page);
179 
180 		if (WARN_ON_ONCE(page_ref_count(page) <= 0))
181 			return false;
182 
183 		if (hpage_pincount_available(page))
184 			hpage_pincount_add(page, 1);
185 		else
186 			refs = GUP_PIN_COUNTING_BIAS;
187 
188 		/*
189 		 * Similar to try_grab_compound_head(): even if using the
190 		 * hpage_pincount_add/_sub() routines, be sure to
191 		 * *also* increment the normal page refcount field at least
192 		 * once, so that the page really is pinned.
193 		 */
194 		page_ref_add(page, refs);
195 
196 		mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED, 1);
197 	}
198 
199 	return true;
200 }
201 
202 /**
203  * unpin_user_page() - release a dma-pinned page
204  * @page:            pointer to page to be released
205  *
206  * Pages that were pinned via pin_user_pages*() must be released via either
207  * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
208  * that such pages can be separately tracked and uniquely handled. In
209  * particular, interactions with RDMA and filesystems need special handling.
210  */
unpin_user_page(struct page * page)211 void unpin_user_page(struct page *page)
212 {
213 	put_compound_head(compound_head(page), 1, FOLL_PIN);
214 }
215 EXPORT_SYMBOL(unpin_user_page);
216 
217 /**
218  * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
219  * @pages:  array of pages to be maybe marked dirty, and definitely released.
220  * @npages: number of pages in the @pages array.
221  * @make_dirty: whether to mark the pages dirty
222  *
223  * "gup-pinned page" refers to a page that has had one of the get_user_pages()
224  * variants called on that page.
225  *
226  * For each page in the @pages array, make that page (or its head page, if a
227  * compound page) dirty, if @make_dirty is true, and if the page was previously
228  * listed as clean. In any case, releases all pages using unpin_user_page(),
229  * possibly via unpin_user_pages(), for the non-dirty case.
230  *
231  * Please see the unpin_user_page() documentation for details.
232  *
233  * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
234  * required, then the caller should a) verify that this is really correct,
235  * because _lock() is usually required, and b) hand code it:
236  * set_page_dirty_lock(), unpin_user_page().
237  *
238  */
unpin_user_pages_dirty_lock(struct page ** pages,unsigned long npages,bool make_dirty)239 void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
240 				 bool make_dirty)
241 {
242 	unsigned long index;
243 
244 	/*
245 	 * TODO: this can be optimized for huge pages: if a series of pages is
246 	 * physically contiguous and part of the same compound page, then a
247 	 * single operation to the head page should suffice.
248 	 */
249 
250 	if (!make_dirty) {
251 		unpin_user_pages(pages, npages);
252 		return;
253 	}
254 
255 	for (index = 0; index < npages; index++) {
256 		struct page *page = compound_head(pages[index]);
257 		/*
258 		 * Checking PageDirty at this point may race with
259 		 * clear_page_dirty_for_io(), but that's OK. Two key
260 		 * cases:
261 		 *
262 		 * 1) This code sees the page as already dirty, so it
263 		 * skips the call to set_page_dirty(). That could happen
264 		 * because clear_page_dirty_for_io() called
265 		 * page_mkclean(), followed by set_page_dirty().
266 		 * However, now the page is going to get written back,
267 		 * which meets the original intention of setting it
268 		 * dirty, so all is well: clear_page_dirty_for_io() goes
269 		 * on to call TestClearPageDirty(), and write the page
270 		 * back.
271 		 *
272 		 * 2) This code sees the page as clean, so it calls
273 		 * set_page_dirty(). The page stays dirty, despite being
274 		 * written back, so it gets written back again in the
275 		 * next writeback cycle. This is harmless.
276 		 */
277 		if (!PageDirty(page))
278 			set_page_dirty_lock(page);
279 		unpin_user_page(page);
280 	}
281 }
282 EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
283 
284 /**
285  * unpin_user_pages() - release an array of gup-pinned pages.
286  * @pages:  array of pages to be marked dirty and released.
287  * @npages: number of pages in the @pages array.
288  *
289  * For each page in the @pages array, release the page using unpin_user_page().
290  *
291  * Please see the unpin_user_page() documentation for details.
292  */
unpin_user_pages(struct page ** pages,unsigned long npages)293 void unpin_user_pages(struct page **pages, unsigned long npages)
294 {
295 	unsigned long index;
296 
297 	/*
298 	 * If this WARN_ON() fires, then the system *might* be leaking pages (by
299 	 * leaving them pinned), but probably not. More likely, gup/pup returned
300 	 * a hard -ERRNO error to the caller, who erroneously passed it here.
301 	 */
302 	if (WARN_ON(IS_ERR_VALUE(npages)))
303 		return;
304 	/*
305 	 * TODO: this can be optimized for huge pages: if a series of pages is
306 	 * physically contiguous and part of the same compound page, then a
307 	 * single operation to the head page should suffice.
308 	 */
309 	for (index = 0; index < npages; index++)
310 		unpin_user_page(pages[index]);
311 }
312 EXPORT_SYMBOL(unpin_user_pages);
313 
314 #ifdef CONFIG_MMU
no_page_table(struct vm_area_struct * vma,unsigned int flags)315 static struct page *no_page_table(struct vm_area_struct *vma,
316 		unsigned int flags)
317 {
318 	/*
319 	 * When core dumping an enormous anonymous area that nobody
320 	 * has touched so far, we don't want to allocate unnecessary pages or
321 	 * page tables.  Return error instead of NULL to skip handle_mm_fault,
322 	 * then get_dump_page() will return NULL to leave a hole in the dump.
323 	 * But we can only make this optimization where a hole would surely
324 	 * be zero-filled if handle_mm_fault() actually did handle it.
325 	 */
326 	if ((flags & FOLL_DUMP) &&
327 			(vma_is_anonymous(vma) || !vma->vm_ops->fault))
328 		return ERR_PTR(-EFAULT);
329 	return NULL;
330 }
331 
follow_pfn_pte(struct vm_area_struct * vma,unsigned long address,pte_t * pte,unsigned int flags)332 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
333 		pte_t *pte, unsigned int flags)
334 {
335 	/* No page to get reference */
336 	if (flags & FOLL_GET)
337 		return -EFAULT;
338 
339 	if (flags & FOLL_TOUCH) {
340 		pte_t entry = *pte;
341 
342 		if (flags & FOLL_WRITE)
343 			entry = pte_mkdirty(entry);
344 		entry = pte_mkyoung(entry);
345 
346 		if (!pte_same(*pte, entry)) {
347 			set_pte_at(vma->vm_mm, address, pte, entry);
348 			update_mmu_cache(vma, address, pte);
349 		}
350 	}
351 
352 	/* Proper page table entry exists, but no corresponding struct page */
353 	return -EEXIST;
354 }
355 
356 /*
357  * FOLL_FORCE can write to even unwritable pte's, but only
358  * after we've gone through a COW cycle and they are dirty.
359  */
can_follow_write_pte(pte_t pte,unsigned int flags)360 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
361 {
362 	return pte_write(pte) ||
363 		((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
364 }
365 
follow_page_pte(struct vm_area_struct * vma,unsigned long address,pmd_t * pmd,unsigned int flags,struct dev_pagemap ** pgmap)366 static struct page *follow_page_pte(struct vm_area_struct *vma,
367 		unsigned long address, pmd_t *pmd, unsigned int flags,
368 		struct dev_pagemap **pgmap)
369 {
370 	struct mm_struct *mm = vma->vm_mm;
371 	struct page *page;
372 	spinlock_t *ptl;
373 	pte_t *ptep, pte;
374 	int ret;
375 
376 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
377 	if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
378 			 (FOLL_PIN | FOLL_GET)))
379 		return ERR_PTR(-EINVAL);
380 retry:
381 	if (unlikely(pmd_bad(*pmd)))
382 		return no_page_table(vma, flags);
383 
384 	ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
385 	pte = *ptep;
386 	if (!pte_present(pte)) {
387 		swp_entry_t entry;
388 		/*
389 		 * KSM's break_ksm() relies upon recognizing a ksm page
390 		 * even while it is being migrated, so for that case we
391 		 * need migration_entry_wait().
392 		 */
393 		if (likely(!(flags & FOLL_MIGRATION)))
394 			goto no_page;
395 		if (pte_none(pte))
396 			goto no_page;
397 		entry = pte_to_swp_entry(pte);
398 		if (!is_migration_entry(entry))
399 			goto no_page;
400 		pte_unmap_unlock(ptep, ptl);
401 		migration_entry_wait(mm, pmd, address);
402 		goto retry;
403 	}
404 	if ((flags & FOLL_NUMA) && pte_protnone(pte))
405 		goto no_page;
406 	if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
407 		pte_unmap_unlock(ptep, ptl);
408 		return NULL;
409 	}
410 
411 	page = vm_normal_page(vma, address, pte);
412 	if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
413 		/*
414 		 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
415 		 * case since they are only valid while holding the pgmap
416 		 * reference.
417 		 */
418 		*pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
419 		if (*pgmap)
420 			page = pte_page(pte);
421 		else
422 			goto no_page;
423 	} else if (unlikely(!page)) {
424 		if (flags & FOLL_DUMP) {
425 			/* Avoid special (like zero) pages in core dumps */
426 			page = ERR_PTR(-EFAULT);
427 			goto out;
428 		}
429 
430 		if (is_zero_pfn(pte_pfn(pte))) {
431 			page = pte_page(pte);
432 		} else {
433 			ret = follow_pfn_pte(vma, address, ptep, flags);
434 			page = ERR_PTR(ret);
435 			goto out;
436 		}
437 	}
438 
439 	if (flags & FOLL_SPLIT && PageTransCompound(page)) {
440 		get_page(page);
441 		pte_unmap_unlock(ptep, ptl);
442 		lock_page(page);
443 		ret = split_huge_page(page);
444 		unlock_page(page);
445 		put_page(page);
446 		if (ret)
447 			return ERR_PTR(ret);
448 		goto retry;
449 	}
450 
451 	/* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
452 	if (unlikely(!try_grab_page(page, flags))) {
453 		page = ERR_PTR(-ENOMEM);
454 		goto out;
455 	}
456 	/*
457 	 * We need to make the page accessible if and only if we are going
458 	 * to access its content (the FOLL_PIN case).  Please see
459 	 * Documentation/core-api/pin_user_pages.rst for details.
460 	 */
461 	if (flags & FOLL_PIN) {
462 		ret = arch_make_page_accessible(page);
463 		if (ret) {
464 			unpin_user_page(page);
465 			page = ERR_PTR(ret);
466 			goto out;
467 		}
468 	}
469 	if (flags & FOLL_TOUCH) {
470 		if ((flags & FOLL_WRITE) &&
471 		    !pte_dirty(pte) && !PageDirty(page))
472 			set_page_dirty(page);
473 		/*
474 		 * pte_mkyoung() would be more correct here, but atomic care
475 		 * is needed to avoid losing the dirty bit: it is easier to use
476 		 * mark_page_accessed().
477 		 */
478 		mark_page_accessed(page);
479 	}
480 	if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
481 		/* Do not mlock pte-mapped THP */
482 		if (PageTransCompound(page))
483 			goto out;
484 
485 		/*
486 		 * The preliminary mapping check is mainly to avoid the
487 		 * pointless overhead of lock_page on the ZERO_PAGE
488 		 * which might bounce very badly if there is contention.
489 		 *
490 		 * If the page is already locked, we don't need to
491 		 * handle it now - vmscan will handle it later if and
492 		 * when it attempts to reclaim the page.
493 		 */
494 		if (page->mapping && trylock_page(page)) {
495 			lru_add_drain();  /* push cached pages to LRU */
496 			/*
497 			 * Because we lock page here, and migration is
498 			 * blocked by the pte's page reference, and we
499 			 * know the page is still mapped, we don't even
500 			 * need to check for file-cache page truncation.
501 			 */
502 			mlock_vma_page(page);
503 			unlock_page(page);
504 		}
505 	}
506 out:
507 	pte_unmap_unlock(ptep, ptl);
508 	return page;
509 no_page:
510 	pte_unmap_unlock(ptep, ptl);
511 	if (!pte_none(pte))
512 		return NULL;
513 	return no_page_table(vma, flags);
514 }
515 
follow_pmd_mask(struct vm_area_struct * vma,unsigned long address,pud_t * pudp,unsigned int flags,struct follow_page_context * ctx)516 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
517 				    unsigned long address, pud_t *pudp,
518 				    unsigned int flags,
519 				    struct follow_page_context *ctx)
520 {
521 	pmd_t *pmd, pmdval;
522 	spinlock_t *ptl;
523 	struct page *page;
524 	struct mm_struct *mm = vma->vm_mm;
525 
526 	pmd = pmd_offset(pudp, address);
527 	/*
528 	 * The READ_ONCE() will stabilize the pmdval in a register or
529 	 * on the stack so that it will stop changing under the code.
530 	 */
531 	pmdval = READ_ONCE(*pmd);
532 	if (pmd_none(pmdval))
533 		return no_page_table(vma, flags);
534 	if (pmd_huge(pmdval) && is_vm_hugetlb_page(vma)) {
535 		page = follow_huge_pmd(mm, address, pmd, flags);
536 		if (page)
537 			return page;
538 		return no_page_table(vma, flags);
539 	}
540 	if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
541 		page = follow_huge_pd(vma, address,
542 				      __hugepd(pmd_val(pmdval)), flags,
543 				      PMD_SHIFT);
544 		if (page)
545 			return page;
546 		return no_page_table(vma, flags);
547 	}
548 retry:
549 	if (!pmd_present(pmdval)) {
550 		if (likely(!(flags & FOLL_MIGRATION)))
551 			return no_page_table(vma, flags);
552 		VM_BUG_ON(thp_migration_supported() &&
553 				  !is_pmd_migration_entry(pmdval));
554 		if (is_pmd_migration_entry(pmdval))
555 			pmd_migration_entry_wait(mm, pmd);
556 		pmdval = READ_ONCE(*pmd);
557 		/*
558 		 * MADV_DONTNEED may convert the pmd to null because
559 		 * mmap_lock is held in read mode
560 		 */
561 		if (pmd_none(pmdval))
562 			return no_page_table(vma, flags);
563 		goto retry;
564 	}
565 	if (pmd_devmap(pmdval)) {
566 		ptl = pmd_lock(mm, pmd);
567 		page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
568 		spin_unlock(ptl);
569 		if (page)
570 			return page;
571 	}
572 	if (likely(!pmd_trans_huge(pmdval)))
573 		return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
574 
575 	if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
576 		return no_page_table(vma, flags);
577 
578 retry_locked:
579 	ptl = pmd_lock(mm, pmd);
580 	if (unlikely(pmd_none(*pmd))) {
581 		spin_unlock(ptl);
582 		return no_page_table(vma, flags);
583 	}
584 	if (unlikely(!pmd_present(*pmd))) {
585 		spin_unlock(ptl);
586 		if (likely(!(flags & FOLL_MIGRATION)))
587 			return no_page_table(vma, flags);
588 		pmd_migration_entry_wait(mm, pmd);
589 		goto retry_locked;
590 	}
591 	if (unlikely(!pmd_trans_huge(*pmd))) {
592 		spin_unlock(ptl);
593 		return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
594 	}
595 	if (flags & (FOLL_SPLIT | FOLL_SPLIT_PMD)) {
596 		int ret;
597 		page = pmd_page(*pmd);
598 		if (is_huge_zero_page(page)) {
599 			spin_unlock(ptl);
600 			ret = 0;
601 			split_huge_pmd(vma, pmd, address);
602 			if (pmd_trans_unstable(pmd))
603 				ret = -EBUSY;
604 		} else if (flags & FOLL_SPLIT) {
605 			if (unlikely(!try_get_page(page))) {
606 				spin_unlock(ptl);
607 				return ERR_PTR(-ENOMEM);
608 			}
609 			spin_unlock(ptl);
610 			lock_page(page);
611 			ret = split_huge_page(page);
612 			unlock_page(page);
613 			put_page(page);
614 			if (pmd_none(*pmd))
615 				return no_page_table(vma, flags);
616 		} else {  /* flags & FOLL_SPLIT_PMD */
617 			spin_unlock(ptl);
618 			split_huge_pmd(vma, pmd, address);
619 			ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
620 		}
621 
622 		return ret ? ERR_PTR(ret) :
623 			follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
624 	}
625 	page = follow_trans_huge_pmd(vma, address, pmd, flags);
626 	spin_unlock(ptl);
627 	ctx->page_mask = HPAGE_PMD_NR - 1;
628 	return page;
629 }
630 
follow_pud_mask(struct vm_area_struct * vma,unsigned long address,p4d_t * p4dp,unsigned int flags,struct follow_page_context * ctx)631 static struct page *follow_pud_mask(struct vm_area_struct *vma,
632 				    unsigned long address, p4d_t *p4dp,
633 				    unsigned int flags,
634 				    struct follow_page_context *ctx)
635 {
636 	pud_t *pud;
637 	spinlock_t *ptl;
638 	struct page *page;
639 	struct mm_struct *mm = vma->vm_mm;
640 
641 	pud = pud_offset(p4dp, address);
642 	if (pud_none(*pud))
643 		return no_page_table(vma, flags);
644 	if (pud_huge(*pud) && is_vm_hugetlb_page(vma)) {
645 		page = follow_huge_pud(mm, address, pud, flags);
646 		if (page)
647 			return page;
648 		return no_page_table(vma, flags);
649 	}
650 	if (is_hugepd(__hugepd(pud_val(*pud)))) {
651 		page = follow_huge_pd(vma, address,
652 				      __hugepd(pud_val(*pud)), flags,
653 				      PUD_SHIFT);
654 		if (page)
655 			return page;
656 		return no_page_table(vma, flags);
657 	}
658 	if (pud_devmap(*pud)) {
659 		ptl = pud_lock(mm, pud);
660 		page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
661 		spin_unlock(ptl);
662 		if (page)
663 			return page;
664 	}
665 	if (unlikely(pud_bad(*pud)))
666 		return no_page_table(vma, flags);
667 
668 	return follow_pmd_mask(vma, address, pud, flags, ctx);
669 }
670 
follow_p4d_mask(struct vm_area_struct * vma,unsigned long address,pgd_t * pgdp,unsigned int flags,struct follow_page_context * ctx)671 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
672 				    unsigned long address, pgd_t *pgdp,
673 				    unsigned int flags,
674 				    struct follow_page_context *ctx)
675 {
676 	p4d_t *p4d;
677 	struct page *page;
678 
679 	p4d = p4d_offset(pgdp, address);
680 	if (p4d_none(*p4d))
681 		return no_page_table(vma, flags);
682 	BUILD_BUG_ON(p4d_huge(*p4d));
683 	if (unlikely(p4d_bad(*p4d)))
684 		return no_page_table(vma, flags);
685 
686 	if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
687 		page = follow_huge_pd(vma, address,
688 				      __hugepd(p4d_val(*p4d)), flags,
689 				      P4D_SHIFT);
690 		if (page)
691 			return page;
692 		return no_page_table(vma, flags);
693 	}
694 	return follow_pud_mask(vma, address, p4d, flags, ctx);
695 }
696 
697 /**
698  * follow_page_mask - look up a page descriptor from a user-virtual address
699  * @vma: vm_area_struct mapping @address
700  * @address: virtual address to look up
701  * @flags: flags modifying lookup behaviour
702  * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
703  *       pointer to output page_mask
704  *
705  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
706  *
707  * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
708  * the device's dev_pagemap metadata to avoid repeating expensive lookups.
709  *
710  * On output, the @ctx->page_mask is set according to the size of the page.
711  *
712  * Return: the mapped (struct page *), %NULL if no mapping exists, or
713  * an error pointer if there is a mapping to something not represented
714  * by a page descriptor (see also vm_normal_page()).
715  */
follow_page_mask(struct vm_area_struct * vma,unsigned long address,unsigned int flags,struct follow_page_context * ctx)716 static struct page *follow_page_mask(struct vm_area_struct *vma,
717 			      unsigned long address, unsigned int flags,
718 			      struct follow_page_context *ctx)
719 {
720 	pgd_t *pgd;
721 	struct page *page;
722 	struct mm_struct *mm = vma->vm_mm;
723 
724 	ctx->page_mask = 0;
725 
726 	/* make this handle hugepd */
727 	page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
728 	if (!IS_ERR(page)) {
729 		WARN_ON_ONCE(flags & (FOLL_GET | FOLL_PIN));
730 		return page;
731 	}
732 
733 	pgd = pgd_offset(mm, address);
734 
735 	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
736 		return no_page_table(vma, flags);
737 
738 	if (pgd_huge(*pgd)) {
739 		page = follow_huge_pgd(mm, address, pgd, flags);
740 		if (page)
741 			return page;
742 		return no_page_table(vma, flags);
743 	}
744 	if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
745 		page = follow_huge_pd(vma, address,
746 				      __hugepd(pgd_val(*pgd)), flags,
747 				      PGDIR_SHIFT);
748 		if (page)
749 			return page;
750 		return no_page_table(vma, flags);
751 	}
752 
753 	return follow_p4d_mask(vma, address, pgd, flags, ctx);
754 }
755 
follow_page(struct vm_area_struct * vma,unsigned long address,unsigned int foll_flags)756 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
757 			 unsigned int foll_flags)
758 {
759 	struct follow_page_context ctx = { NULL };
760 	struct page *page;
761 
762 	page = follow_page_mask(vma, address, foll_flags, &ctx);
763 	if (ctx.pgmap)
764 		put_dev_pagemap(ctx.pgmap);
765 	return page;
766 }
767 
get_gate_page(struct mm_struct * mm,unsigned long address,unsigned int gup_flags,struct vm_area_struct ** vma,struct page ** page)768 static int get_gate_page(struct mm_struct *mm, unsigned long address,
769 		unsigned int gup_flags, struct vm_area_struct **vma,
770 		struct page **page)
771 {
772 	pgd_t *pgd;
773 	p4d_t *p4d;
774 	pud_t *pud;
775 	pmd_t *pmd;
776 	pte_t *pte;
777 	int ret = -EFAULT;
778 
779 	/* user gate pages are read-only */
780 	if (gup_flags & FOLL_WRITE)
781 		return -EFAULT;
782 	if (address > TASK_SIZE)
783 		pgd = pgd_offset_k(address);
784 	else
785 		pgd = pgd_offset_gate(mm, address);
786 	if (pgd_none(*pgd))
787 		return -EFAULT;
788 	p4d = p4d_offset(pgd, address);
789 	if (p4d_none(*p4d))
790 		return -EFAULT;
791 	pud = pud_offset(p4d, address);
792 	if (pud_none(*pud))
793 		return -EFAULT;
794 	pmd = pmd_offset(pud, address);
795 	if (!pmd_present(*pmd))
796 		return -EFAULT;
797 	VM_BUG_ON(pmd_trans_huge(*pmd));
798 	pte = pte_offset_map(pmd, address);
799 	if (pte_none(*pte))
800 		goto unmap;
801 	*vma = get_gate_vma(mm);
802 	if (!page)
803 		goto out;
804 	*page = vm_normal_page(*vma, address, *pte);
805 	if (!*page) {
806 		if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
807 			goto unmap;
808 		*page = pte_page(*pte);
809 	}
810 	if (unlikely(!try_grab_page(*page, gup_flags))) {
811 		ret = -ENOMEM;
812 		goto unmap;
813 	}
814 out:
815 	ret = 0;
816 unmap:
817 	pte_unmap(pte);
818 	return ret;
819 }
820 
821 /*
822  * mmap_lock must be held on entry.  If @locked != NULL and *@flags
823  * does not include FOLL_NOWAIT, the mmap_lock may be released.  If it
824  * is, *@locked will be set to 0 and -EBUSY returned.
825  */
faultin_page(struct vm_area_struct * vma,unsigned long address,unsigned int * flags,int * locked)826 static int faultin_page(struct vm_area_struct *vma,
827 		unsigned long address, unsigned int *flags, int *locked)
828 {
829 	unsigned int fault_flags = 0;
830 	vm_fault_t ret;
831 
832 	/* mlock all present pages, but do not fault in new pages */
833 	if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
834 		return -ENOENT;
835 	if (*flags & FOLL_WRITE)
836 		fault_flags |= FAULT_FLAG_WRITE;
837 	if (*flags & FOLL_REMOTE)
838 		fault_flags |= FAULT_FLAG_REMOTE;
839 	if (locked)
840 		fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
841 	if (*flags & FOLL_NOWAIT)
842 		fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
843 	if (*flags & FOLL_TRIED) {
844 		/*
845 		 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
846 		 * can co-exist
847 		 */
848 		fault_flags |= FAULT_FLAG_TRIED;
849 	}
850 
851 	ret = handle_mm_fault(vma, address, fault_flags, NULL);
852 	if (ret & VM_FAULT_ERROR) {
853 		int err = vm_fault_to_errno(ret, *flags);
854 
855 		if (err)
856 			return err;
857 		BUG();
858 	}
859 
860 	if (ret & VM_FAULT_RETRY) {
861 		if (locked && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
862 			*locked = 0;
863 		return -EBUSY;
864 	}
865 
866 	/*
867 	 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
868 	 * necessary, even if maybe_mkwrite decided not to set pte_write. We
869 	 * can thus safely do subsequent page lookups as if they were reads.
870 	 * But only do so when looping for pte_write is futile: in some cases
871 	 * userspace may also be wanting to write to the gotten user page,
872 	 * which a read fault here might prevent (a readonly page might get
873 	 * reCOWed by userspace write).
874 	 */
875 	if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
876 		*flags |= FOLL_COW;
877 	return 0;
878 }
879 
check_vma_flags(struct vm_area_struct * vma,unsigned long gup_flags)880 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
881 {
882 	vm_flags_t vm_flags = vma->vm_flags;
883 	int write = (gup_flags & FOLL_WRITE);
884 	int foreign = (gup_flags & FOLL_REMOTE);
885 
886 	if (vm_flags & (VM_IO | VM_PFNMAP))
887 		return -EFAULT;
888 
889 	if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
890 		return -EFAULT;
891 
892 	if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma))
893 		return -EOPNOTSUPP;
894 
895 	if (write) {
896 		if (!(vm_flags & VM_WRITE)) {
897 			if (!(gup_flags & FOLL_FORCE))
898 				return -EFAULT;
899 			/*
900 			 * We used to let the write,force case do COW in a
901 			 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
902 			 * set a breakpoint in a read-only mapping of an
903 			 * executable, without corrupting the file (yet only
904 			 * when that file had been opened for writing!).
905 			 * Anon pages in shared mappings are surprising: now
906 			 * just reject it.
907 			 */
908 			if (!is_cow_mapping(vm_flags))
909 				return -EFAULT;
910 		}
911 	} else if (!(vm_flags & VM_READ)) {
912 		if (!(gup_flags & FOLL_FORCE))
913 			return -EFAULT;
914 		/*
915 		 * Is there actually any vma we can reach here which does not
916 		 * have VM_MAYREAD set?
917 		 */
918 		if (!(vm_flags & VM_MAYREAD))
919 			return -EFAULT;
920 	}
921 	/*
922 	 * gups are always data accesses, not instruction
923 	 * fetches, so execute=false here
924 	 */
925 	if (!arch_vma_access_permitted(vma, write, false, foreign))
926 		return -EFAULT;
927 	return 0;
928 }
929 
930 /**
931  * __get_user_pages() - pin user pages in memory
932  * @mm:		mm_struct of target mm
933  * @start:	starting user address
934  * @nr_pages:	number of pages from start to pin
935  * @gup_flags:	flags modifying pin behaviour
936  * @pages:	array that receives pointers to the pages pinned.
937  *		Should be at least nr_pages long. Or NULL, if caller
938  *		only intends to ensure the pages are faulted in.
939  * @vmas:	array of pointers to vmas corresponding to each page.
940  *		Or NULL if the caller does not require them.
941  * @locked:     whether we're still with the mmap_lock held
942  *
943  * Returns either number of pages pinned (which may be less than the
944  * number requested), or an error. Details about the return value:
945  *
946  * -- If nr_pages is 0, returns 0.
947  * -- If nr_pages is >0, but no pages were pinned, returns -errno.
948  * -- If nr_pages is >0, and some pages were pinned, returns the number of
949  *    pages pinned. Again, this may be less than nr_pages.
950  * -- 0 return value is possible when the fault would need to be retried.
951  *
952  * The caller is responsible for releasing returned @pages, via put_page().
953  *
954  * @vmas are valid only as long as mmap_lock is held.
955  *
956  * Must be called with mmap_lock held.  It may be released.  See below.
957  *
958  * __get_user_pages walks a process's page tables and takes a reference to
959  * each struct page that each user address corresponds to at a given
960  * instant. That is, it takes the page that would be accessed if a user
961  * thread accesses the given user virtual address at that instant.
962  *
963  * This does not guarantee that the page exists in the user mappings when
964  * __get_user_pages returns, and there may even be a completely different
965  * page there in some cases (eg. if mmapped pagecache has been invalidated
966  * and subsequently re faulted). However it does guarantee that the page
967  * won't be freed completely. And mostly callers simply care that the page
968  * contains data that was valid *at some point in time*. Typically, an IO
969  * or similar operation cannot guarantee anything stronger anyway because
970  * locks can't be held over the syscall boundary.
971  *
972  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
973  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
974  * appropriate) must be called after the page is finished with, and
975  * before put_page is called.
976  *
977  * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
978  * released by an up_read().  That can happen if @gup_flags does not
979  * have FOLL_NOWAIT.
980  *
981  * A caller using such a combination of @locked and @gup_flags
982  * must therefore hold the mmap_lock for reading only, and recognize
983  * when it's been released.  Otherwise, it must be held for either
984  * reading or writing and will not be released.
985  *
986  * In most cases, get_user_pages or get_user_pages_fast should be used
987  * instead of __get_user_pages. __get_user_pages should be used only if
988  * you need some special @gup_flags.
989  */
__get_user_pages(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,struct vm_area_struct ** vmas,int * locked)990 static long __get_user_pages(struct mm_struct *mm,
991 		unsigned long start, unsigned long nr_pages,
992 		unsigned int gup_flags, struct page **pages,
993 		struct vm_area_struct **vmas, int *locked)
994 {
995 	long ret = 0, i = 0;
996 	struct vm_area_struct *vma = NULL;
997 	struct follow_page_context ctx = { NULL };
998 
999 	if (!nr_pages)
1000 		return 0;
1001 
1002 	start = untagged_addr(start);
1003 
1004 	VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1005 
1006 	/*
1007 	 * If FOLL_FORCE is set then do not force a full fault as the hinting
1008 	 * fault information is unrelated to the reference behaviour of a task
1009 	 * using the address space
1010 	 */
1011 	if (!(gup_flags & FOLL_FORCE))
1012 		gup_flags |= FOLL_NUMA;
1013 
1014 	do {
1015 		struct page *page;
1016 		unsigned int foll_flags = gup_flags;
1017 		unsigned int page_increm;
1018 
1019 		/* first iteration or cross vma bound */
1020 		if (!vma || start >= vma->vm_end) {
1021 			vma = find_extend_vma(mm, start);
1022 			if (!vma && in_gate_area(mm, start)) {
1023 				ret = get_gate_page(mm, start & PAGE_MASK,
1024 						gup_flags, &vma,
1025 						pages ? &pages[i] : NULL);
1026 				if (ret)
1027 					goto out;
1028 				ctx.page_mask = 0;
1029 				goto next_page;
1030 			}
1031 
1032 			if (!vma) {
1033 				ret = -EFAULT;
1034 				goto out;
1035 			}
1036 			ret = check_vma_flags(vma, gup_flags);
1037 			if (ret)
1038 				goto out;
1039 
1040 			if (is_vm_hugetlb_page(vma)) {
1041 				i = follow_hugetlb_page(mm, vma, pages, vmas,
1042 						&start, &nr_pages, i,
1043 						gup_flags, locked);
1044 				if (locked && *locked == 0) {
1045 					/*
1046 					 * We've got a VM_FAULT_RETRY
1047 					 * and we've lost mmap_lock.
1048 					 * We must stop here.
1049 					 */
1050 					BUG_ON(gup_flags & FOLL_NOWAIT);
1051 					BUG_ON(ret != 0);
1052 					goto out;
1053 				}
1054 				continue;
1055 			}
1056 		}
1057 retry:
1058 		/*
1059 		 * If we have a pending SIGKILL, don't keep faulting pages and
1060 		 * potentially allocating memory.
1061 		 */
1062 		if (fatal_signal_pending(current)) {
1063 			ret = -EINTR;
1064 			goto out;
1065 		}
1066 		cond_resched();
1067 
1068 		page = follow_page_mask(vma, start, foll_flags, &ctx);
1069 		if (!page) {
1070 			ret = faultin_page(vma, start, &foll_flags, locked);
1071 			switch (ret) {
1072 			case 0:
1073 				goto retry;
1074 			case -EBUSY:
1075 				ret = 0;
1076 				fallthrough;
1077 			case -EFAULT:
1078 			case -ENOMEM:
1079 			case -EHWPOISON:
1080 				goto out;
1081 			case -ENOENT:
1082 				goto next_page;
1083 			}
1084 			BUG();
1085 		} else if (PTR_ERR(page) == -EEXIST) {
1086 			/*
1087 			 * Proper page table entry exists, but no corresponding
1088 			 * struct page.
1089 			 */
1090 			goto next_page;
1091 		} else if (IS_ERR(page)) {
1092 			ret = PTR_ERR(page);
1093 			goto out;
1094 		}
1095 		if (pages) {
1096 			pages[i] = page;
1097 			flush_anon_page(vma, page, start);
1098 			flush_dcache_page(page);
1099 			ctx.page_mask = 0;
1100 		}
1101 next_page:
1102 		if (vmas) {
1103 			vmas[i] = vma;
1104 			ctx.page_mask = 0;
1105 		}
1106 		page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1107 		if (page_increm > nr_pages)
1108 			page_increm = nr_pages;
1109 		i += page_increm;
1110 		start += page_increm * PAGE_SIZE;
1111 		nr_pages -= page_increm;
1112 	} while (nr_pages);
1113 out:
1114 	if (ctx.pgmap)
1115 		put_dev_pagemap(ctx.pgmap);
1116 	return i ? i : ret;
1117 }
1118 
vma_permits_fault(struct vm_area_struct * vma,unsigned int fault_flags)1119 static bool vma_permits_fault(struct vm_area_struct *vma,
1120 			      unsigned int fault_flags)
1121 {
1122 	bool write   = !!(fault_flags & FAULT_FLAG_WRITE);
1123 	bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1124 	vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1125 
1126 	if (!(vm_flags & vma->vm_flags))
1127 		return false;
1128 
1129 	/*
1130 	 * The architecture might have a hardware protection
1131 	 * mechanism other than read/write that can deny access.
1132 	 *
1133 	 * gup always represents data access, not instruction
1134 	 * fetches, so execute=false here:
1135 	 */
1136 	if (!arch_vma_access_permitted(vma, write, false, foreign))
1137 		return false;
1138 
1139 	return true;
1140 }
1141 
1142 /**
1143  * fixup_user_fault() - manually resolve a user page fault
1144  * @mm:		mm_struct of target mm
1145  * @address:	user address
1146  * @fault_flags:flags to pass down to handle_mm_fault()
1147  * @unlocked:	did we unlock the mmap_lock while retrying, maybe NULL if caller
1148  *		does not allow retry. If NULL, the caller must guarantee
1149  *		that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1150  *
1151  * This is meant to be called in the specific scenario where for locking reasons
1152  * we try to access user memory in atomic context (within a pagefault_disable()
1153  * section), this returns -EFAULT, and we want to resolve the user fault before
1154  * trying again.
1155  *
1156  * Typically this is meant to be used by the futex code.
1157  *
1158  * The main difference with get_user_pages() is that this function will
1159  * unconditionally call handle_mm_fault() which will in turn perform all the
1160  * necessary SW fixup of the dirty and young bits in the PTE, while
1161  * get_user_pages() only guarantees to update these in the struct page.
1162  *
1163  * This is important for some architectures where those bits also gate the
1164  * access permission to the page because they are maintained in software.  On
1165  * such architectures, gup() will not be enough to make a subsequent access
1166  * succeed.
1167  *
1168  * This function will not return with an unlocked mmap_lock. So it has not the
1169  * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1170  */
fixup_user_fault(struct mm_struct * mm,unsigned long address,unsigned int fault_flags,bool * unlocked)1171 int fixup_user_fault(struct mm_struct *mm,
1172 		     unsigned long address, unsigned int fault_flags,
1173 		     bool *unlocked)
1174 {
1175 	struct vm_area_struct *vma;
1176 	vm_fault_t ret, major = 0;
1177 
1178 	address = untagged_addr(address);
1179 
1180 	if (unlocked)
1181 		fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1182 
1183 retry:
1184 	vma = find_extend_vma(mm, address);
1185 	if (!vma || address < vma->vm_start)
1186 		return -EFAULT;
1187 
1188 	if (!vma_permits_fault(vma, fault_flags))
1189 		return -EFAULT;
1190 
1191 	if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1192 	    fatal_signal_pending(current))
1193 		return -EINTR;
1194 
1195 	ret = handle_mm_fault(vma, address, fault_flags, NULL);
1196 	major |= ret & VM_FAULT_MAJOR;
1197 	if (ret & VM_FAULT_ERROR) {
1198 		int err = vm_fault_to_errno(ret, 0);
1199 
1200 		if (err)
1201 			return err;
1202 		BUG();
1203 	}
1204 
1205 	if (ret & VM_FAULT_RETRY) {
1206 		mmap_read_lock(mm);
1207 		*unlocked = true;
1208 		fault_flags |= FAULT_FLAG_TRIED;
1209 		goto retry;
1210 	}
1211 
1212 	return 0;
1213 }
1214 EXPORT_SYMBOL_GPL(fixup_user_fault);
1215 
1216 /*
1217  * Please note that this function, unlike __get_user_pages will not
1218  * return 0 for nr_pages > 0 without FOLL_NOWAIT
1219  */
__get_user_pages_locked(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,struct vm_area_struct ** vmas,int * locked,unsigned int flags)1220 static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1221 						unsigned long start,
1222 						unsigned long nr_pages,
1223 						struct page **pages,
1224 						struct vm_area_struct **vmas,
1225 						int *locked,
1226 						unsigned int flags)
1227 {
1228 	long ret, pages_done;
1229 	bool lock_dropped;
1230 
1231 	if (locked) {
1232 		/* if VM_FAULT_RETRY can be returned, vmas become invalid */
1233 		BUG_ON(vmas);
1234 		/* check caller initialized locked */
1235 		BUG_ON(*locked != 1);
1236 	}
1237 
1238 	if (flags & FOLL_PIN)
1239 		atomic_set(&mm->has_pinned, 1);
1240 
1241 	/*
1242 	 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1243 	 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1244 	 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1245 	 * for FOLL_GET, not for the newer FOLL_PIN.
1246 	 *
1247 	 * FOLL_PIN always expects pages to be non-null, but no need to assert
1248 	 * that here, as any failures will be obvious enough.
1249 	 */
1250 	if (pages && !(flags & FOLL_PIN))
1251 		flags |= FOLL_GET;
1252 
1253 	pages_done = 0;
1254 	lock_dropped = false;
1255 	for (;;) {
1256 		ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1257 				       vmas, locked);
1258 		if (!locked)
1259 			/* VM_FAULT_RETRY couldn't trigger, bypass */
1260 			return ret;
1261 
1262 		/* VM_FAULT_RETRY cannot return errors */
1263 		if (!*locked) {
1264 			BUG_ON(ret < 0);
1265 			BUG_ON(ret >= nr_pages);
1266 		}
1267 
1268 		if (ret > 0) {
1269 			nr_pages -= ret;
1270 			pages_done += ret;
1271 			if (!nr_pages)
1272 				break;
1273 		}
1274 		if (*locked) {
1275 			/*
1276 			 * VM_FAULT_RETRY didn't trigger or it was a
1277 			 * FOLL_NOWAIT.
1278 			 */
1279 			if (!pages_done)
1280 				pages_done = ret;
1281 			break;
1282 		}
1283 		/*
1284 		 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1285 		 * For the prefault case (!pages) we only update counts.
1286 		 */
1287 		if (likely(pages))
1288 			pages += ret;
1289 		start += ret << PAGE_SHIFT;
1290 		lock_dropped = true;
1291 
1292 retry:
1293 		/*
1294 		 * Repeat on the address that fired VM_FAULT_RETRY
1295 		 * with both FAULT_FLAG_ALLOW_RETRY and
1296 		 * FAULT_FLAG_TRIED.  Note that GUP can be interrupted
1297 		 * by fatal signals, so we need to check it before we
1298 		 * start trying again otherwise it can loop forever.
1299 		 */
1300 
1301 		if (fatal_signal_pending(current)) {
1302 			if (!pages_done)
1303 				pages_done = -EINTR;
1304 			break;
1305 		}
1306 
1307 		ret = mmap_read_lock_killable(mm);
1308 		if (ret) {
1309 			BUG_ON(ret > 0);
1310 			if (!pages_done)
1311 				pages_done = ret;
1312 			break;
1313 		}
1314 
1315 		*locked = 1;
1316 		ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1317 				       pages, NULL, locked);
1318 		if (!*locked) {
1319 			/* Continue to retry until we succeeded */
1320 			BUG_ON(ret != 0);
1321 			goto retry;
1322 		}
1323 		if (ret != 1) {
1324 			BUG_ON(ret > 1);
1325 			if (!pages_done)
1326 				pages_done = ret;
1327 			break;
1328 		}
1329 		nr_pages--;
1330 		pages_done++;
1331 		if (!nr_pages)
1332 			break;
1333 		if (likely(pages))
1334 			pages++;
1335 		start += PAGE_SIZE;
1336 	}
1337 	if (lock_dropped && *locked) {
1338 		/*
1339 		 * We must let the caller know we temporarily dropped the lock
1340 		 * and so the critical section protected by it was lost.
1341 		 */
1342 		mmap_read_unlock(mm);
1343 		*locked = 0;
1344 	}
1345 	return pages_done;
1346 }
1347 
1348 /**
1349  * populate_vma_page_range() -  populate a range of pages in the vma.
1350  * @vma:   target vma
1351  * @start: start address
1352  * @end:   end address
1353  * @locked: whether the mmap_lock is still held
1354  *
1355  * This takes care of mlocking the pages too if VM_LOCKED is set.
1356  *
1357  * Return either number of pages pinned in the vma, or a negative error
1358  * code on error.
1359  *
1360  * vma->vm_mm->mmap_lock must be held.
1361  *
1362  * If @locked is NULL, it may be held for read or write and will
1363  * be unperturbed.
1364  *
1365  * If @locked is non-NULL, it must held for read only and may be
1366  * released.  If it's released, *@locked will be set to 0.
1367  */
populate_vma_page_range(struct vm_area_struct * vma,unsigned long start,unsigned long end,int * locked)1368 long populate_vma_page_range(struct vm_area_struct *vma,
1369 		unsigned long start, unsigned long end, int *locked)
1370 {
1371 	struct mm_struct *mm = vma->vm_mm;
1372 	unsigned long nr_pages = (end - start) / PAGE_SIZE;
1373 	int gup_flags;
1374 
1375 	VM_BUG_ON(start & ~PAGE_MASK);
1376 	VM_BUG_ON(end   & ~PAGE_MASK);
1377 	VM_BUG_ON_VMA(start < vma->vm_start, vma);
1378 	VM_BUG_ON_VMA(end   > vma->vm_end, vma);
1379 	mmap_assert_locked(mm);
1380 
1381 	gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1382 	if (vma->vm_flags & VM_LOCKONFAULT)
1383 		gup_flags &= ~FOLL_POPULATE;
1384 	/*
1385 	 * We want to touch writable mappings with a write fault in order
1386 	 * to break COW, except for shared mappings because these don't COW
1387 	 * and we would not want to dirty them for nothing.
1388 	 */
1389 	if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1390 		gup_flags |= FOLL_WRITE;
1391 
1392 	/*
1393 	 * We want mlock to succeed for regions that have any permissions
1394 	 * other than PROT_NONE.
1395 	 */
1396 	if (vma_is_accessible(vma))
1397 		gup_flags |= FOLL_FORCE;
1398 
1399 	/*
1400 	 * We made sure addr is within a VMA, so the following will
1401 	 * not result in a stack expansion that recurses back here.
1402 	 */
1403 	return __get_user_pages(mm, start, nr_pages, gup_flags,
1404 				NULL, NULL, locked);
1405 }
1406 
1407 /*
1408  * __mm_populate - populate and/or mlock pages within a range of address space.
1409  *
1410  * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1411  * flags. VMAs must be already marked with the desired vm_flags, and
1412  * mmap_lock must not be held.
1413  */
__mm_populate(unsigned long start,unsigned long len,int ignore_errors)1414 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1415 {
1416 	struct mm_struct *mm = current->mm;
1417 	unsigned long end, nstart, nend;
1418 	struct vm_area_struct *vma = NULL;
1419 	int locked = 0;
1420 	long ret = 0;
1421 
1422 	end = start + len;
1423 
1424 	for (nstart = start; nstart < end; nstart = nend) {
1425 		/*
1426 		 * We want to fault in pages for [nstart; end) address range.
1427 		 * Find first corresponding VMA.
1428 		 */
1429 		if (!locked) {
1430 			locked = 1;
1431 			mmap_read_lock(mm);
1432 			vma = find_vma(mm, nstart);
1433 		} else if (nstart >= vma->vm_end)
1434 			vma = vma->vm_next;
1435 		if (!vma || vma->vm_start >= end)
1436 			break;
1437 		/*
1438 		 * Set [nstart; nend) to intersection of desired address
1439 		 * range with the first VMA. Also, skip undesirable VMA types.
1440 		 */
1441 		nend = min(end, vma->vm_end);
1442 		if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1443 			continue;
1444 		if (nstart < vma->vm_start)
1445 			nstart = vma->vm_start;
1446 		/*
1447 		 * Now fault in a range of pages. populate_vma_page_range()
1448 		 * double checks the vma flags, so that it won't mlock pages
1449 		 * if the vma was already munlocked.
1450 		 */
1451 		ret = populate_vma_page_range(vma, nstart, nend, &locked);
1452 		if (ret < 0) {
1453 			if (ignore_errors) {
1454 				ret = 0;
1455 				continue;	/* continue at next VMA */
1456 			}
1457 			break;
1458 		}
1459 		nend = nstart + ret * PAGE_SIZE;
1460 		ret = 0;
1461 	}
1462 	if (locked)
1463 		mmap_read_unlock(mm);
1464 	return ret;	/* 0 or negative error code */
1465 }
1466 #else /* CONFIG_MMU */
__get_user_pages_locked(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,struct vm_area_struct ** vmas,int * locked,unsigned int foll_flags)1467 static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1468 		unsigned long nr_pages, struct page **pages,
1469 		struct vm_area_struct **vmas, int *locked,
1470 		unsigned int foll_flags)
1471 {
1472 	struct vm_area_struct *vma;
1473 	unsigned long vm_flags;
1474 	int i;
1475 
1476 	/* calculate required read or write permissions.
1477 	 * If FOLL_FORCE is set, we only require the "MAY" flags.
1478 	 */
1479 	vm_flags  = (foll_flags & FOLL_WRITE) ?
1480 			(VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1481 	vm_flags &= (foll_flags & FOLL_FORCE) ?
1482 			(VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1483 
1484 	for (i = 0; i < nr_pages; i++) {
1485 		vma = find_vma(mm, start);
1486 		if (!vma)
1487 			goto finish_or_fault;
1488 
1489 		/* protect what we can, including chardevs */
1490 		if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1491 		    !(vm_flags & vma->vm_flags))
1492 			goto finish_or_fault;
1493 
1494 		if (pages) {
1495 			pages[i] = virt_to_page(start);
1496 			if (pages[i])
1497 				get_page(pages[i]);
1498 		}
1499 		if (vmas)
1500 			vmas[i] = vma;
1501 		start = (start + PAGE_SIZE) & PAGE_MASK;
1502 	}
1503 
1504 	return i;
1505 
1506 finish_or_fault:
1507 	return i ? : -EFAULT;
1508 }
1509 #endif /* !CONFIG_MMU */
1510 
1511 /**
1512  * get_dump_page() - pin user page in memory while writing it to core dump
1513  * @addr: user address
1514  *
1515  * Returns struct page pointer of user page pinned for dump,
1516  * to be freed afterwards by put_page().
1517  *
1518  * Returns NULL on any kind of failure - a hole must then be inserted into
1519  * the corefile, to preserve alignment with its headers; and also returns
1520  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1521  * allowing a hole to be left in the corefile to save diskspace.
1522  *
1523  * Called without mmap_lock (takes and releases the mmap_lock by itself).
1524  */
1525 #ifdef CONFIG_ELF_CORE
get_dump_page(unsigned long addr)1526 struct page *get_dump_page(unsigned long addr)
1527 {
1528 	struct mm_struct *mm = current->mm;
1529 	struct page *page;
1530 	int locked = 1;
1531 	int ret;
1532 
1533 	if (mmap_read_lock_killable(mm))
1534 		return NULL;
1535 	ret = __get_user_pages_locked(mm, addr, 1, &page, NULL, &locked,
1536 				      FOLL_FORCE | FOLL_DUMP | FOLL_GET);
1537 	if (locked)
1538 		mmap_read_unlock(mm);
1539 	return (ret == 1) ? page : NULL;
1540 }
1541 #endif /* CONFIG_ELF_CORE */
1542 
1543 #ifdef CONFIG_CMA
check_and_migrate_cma_pages(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,struct vm_area_struct ** vmas,unsigned int gup_flags)1544 static long check_and_migrate_cma_pages(struct mm_struct *mm,
1545 					unsigned long start,
1546 					unsigned long nr_pages,
1547 					struct page **pages,
1548 					struct vm_area_struct **vmas,
1549 					unsigned int gup_flags)
1550 {
1551 	unsigned long i;
1552 	unsigned long step;
1553 	bool drain_allow = true;
1554 	bool migrate_allow = true;
1555 	LIST_HEAD(cma_page_list);
1556 	long ret = nr_pages;
1557 	struct migration_target_control mtc = {
1558 		.nid = NUMA_NO_NODE,
1559 		.gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_NOWARN,
1560 	};
1561 
1562 check_again:
1563 	for (i = 0; i < nr_pages;) {
1564 
1565 		struct page *head = compound_head(pages[i]);
1566 
1567 		/*
1568 		 * gup may start from a tail page. Advance step by the left
1569 		 * part.
1570 		 */
1571 		step = compound_nr(head) - (pages[i] - head);
1572 		/*
1573 		 * If we get a page from the CMA zone, since we are going to
1574 		 * be pinning these entries, we might as well move them out
1575 		 * of the CMA zone if possible.
1576 		 */
1577 		if (is_migrate_cma_page(head)) {
1578 			if (PageHuge(head))
1579 				isolate_huge_page(head, &cma_page_list);
1580 			else {
1581 				if (!PageLRU(head) && drain_allow) {
1582 					lru_add_drain_all();
1583 					drain_allow = false;
1584 				}
1585 
1586 				if (!isolate_lru_page(head)) {
1587 					list_add_tail(&head->lru, &cma_page_list);
1588 					mod_node_page_state(page_pgdat(head),
1589 							    NR_ISOLATED_ANON +
1590 							    page_is_file_lru(head),
1591 							    thp_nr_pages(head));
1592 				}
1593 			}
1594 		}
1595 
1596 		i += step;
1597 	}
1598 
1599 	if (!list_empty(&cma_page_list)) {
1600 		/*
1601 		 * drop the above get_user_pages reference.
1602 		 */
1603 		if (gup_flags & FOLL_PIN)
1604 			unpin_user_pages(pages, nr_pages);
1605 		else
1606 			for (i = 0; i < nr_pages; i++)
1607 				put_page(pages[i]);
1608 
1609 		if (migrate_pages(&cma_page_list, alloc_migration_target, NULL,
1610 			(unsigned long)&mtc, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
1611 			/*
1612 			 * some of the pages failed migration. Do get_user_pages
1613 			 * without migration.
1614 			 */
1615 			migrate_allow = false;
1616 
1617 			if (!list_empty(&cma_page_list))
1618 				putback_movable_pages(&cma_page_list);
1619 		}
1620 		/*
1621 		 * We did migrate all the pages, Try to get the page references
1622 		 * again migrating any new CMA pages which we failed to isolate
1623 		 * earlier.
1624 		 */
1625 		ret = __get_user_pages_locked(mm, start, nr_pages,
1626 						   pages, vmas, NULL,
1627 						   gup_flags);
1628 
1629 		if ((ret > 0) && migrate_allow) {
1630 			nr_pages = ret;
1631 			drain_allow = true;
1632 			goto check_again;
1633 		}
1634 	}
1635 
1636 	return ret;
1637 }
1638 #else
check_and_migrate_cma_pages(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,struct vm_area_struct ** vmas,unsigned int gup_flags)1639 static long check_and_migrate_cma_pages(struct mm_struct *mm,
1640 					unsigned long start,
1641 					unsigned long nr_pages,
1642 					struct page **pages,
1643 					struct vm_area_struct **vmas,
1644 					unsigned int gup_flags)
1645 {
1646 	return nr_pages;
1647 }
1648 #endif /* CONFIG_CMA */
1649 
1650 /*
1651  * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1652  * allows us to process the FOLL_LONGTERM flag.
1653  */
__gup_longterm_locked(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,struct vm_area_struct ** vmas,unsigned int gup_flags)1654 static long __gup_longterm_locked(struct mm_struct *mm,
1655 				  unsigned long start,
1656 				  unsigned long nr_pages,
1657 				  struct page **pages,
1658 				  struct vm_area_struct **vmas,
1659 				  unsigned int gup_flags)
1660 {
1661 	unsigned long flags = 0;
1662 	long rc;
1663 
1664 	if (gup_flags & FOLL_LONGTERM)
1665 		flags = memalloc_nocma_save();
1666 
1667 	rc = __get_user_pages_locked(mm, start, nr_pages, pages, vmas, NULL,
1668 				     gup_flags);
1669 
1670 	if (gup_flags & FOLL_LONGTERM) {
1671 		if (rc > 0)
1672 			rc = check_and_migrate_cma_pages(mm, start, rc, pages,
1673 							 vmas, gup_flags);
1674 		memalloc_nocma_restore(flags);
1675 	}
1676 	return rc;
1677 }
1678 
is_valid_gup_flags(unsigned int gup_flags)1679 static bool is_valid_gup_flags(unsigned int gup_flags)
1680 {
1681 	/*
1682 	 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1683 	 * never directly by the caller, so enforce that with an assertion:
1684 	 */
1685 	if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1686 		return false;
1687 	/*
1688 	 * FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying
1689 	 * that is, FOLL_LONGTERM is a specific case, more restrictive case of
1690 	 * FOLL_PIN.
1691 	 */
1692 	if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1693 		return false;
1694 
1695 	return true;
1696 }
1697 
1698 #ifdef CONFIG_MMU
__get_user_pages_remote(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,struct vm_area_struct ** vmas,int * locked)1699 static long __get_user_pages_remote(struct mm_struct *mm,
1700 				    unsigned long start, unsigned long nr_pages,
1701 				    unsigned int gup_flags, struct page **pages,
1702 				    struct vm_area_struct **vmas, int *locked)
1703 {
1704 	/*
1705 	 * Parts of FOLL_LONGTERM behavior are incompatible with
1706 	 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1707 	 * vmas. However, this only comes up if locked is set, and there are
1708 	 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1709 	 * allow what we can.
1710 	 */
1711 	if (gup_flags & FOLL_LONGTERM) {
1712 		if (WARN_ON_ONCE(locked))
1713 			return -EINVAL;
1714 		/*
1715 		 * This will check the vmas (even if our vmas arg is NULL)
1716 		 * and return -ENOTSUPP if DAX isn't allowed in this case:
1717 		 */
1718 		return __gup_longterm_locked(mm, start, nr_pages, pages,
1719 					     vmas, gup_flags | FOLL_TOUCH |
1720 					     FOLL_REMOTE);
1721 	}
1722 
1723 	return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1724 				       locked,
1725 				       gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1726 }
1727 
1728 /**
1729  * get_user_pages_remote() - pin user pages in memory
1730  * @mm:		mm_struct of target mm
1731  * @start:	starting user address
1732  * @nr_pages:	number of pages from start to pin
1733  * @gup_flags:	flags modifying lookup behaviour
1734  * @pages:	array that receives pointers to the pages pinned.
1735  *		Should be at least nr_pages long. Or NULL, if caller
1736  *		only intends to ensure the pages are faulted in.
1737  * @vmas:	array of pointers to vmas corresponding to each page.
1738  *		Or NULL if the caller does not require them.
1739  * @locked:	pointer to lock flag indicating whether lock is held and
1740  *		subsequently whether VM_FAULT_RETRY functionality can be
1741  *		utilised. Lock must initially be held.
1742  *
1743  * Returns either number of pages pinned (which may be less than the
1744  * number requested), or an error. Details about the return value:
1745  *
1746  * -- If nr_pages is 0, returns 0.
1747  * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1748  * -- If nr_pages is >0, and some pages were pinned, returns the number of
1749  *    pages pinned. Again, this may be less than nr_pages.
1750  *
1751  * The caller is responsible for releasing returned @pages, via put_page().
1752  *
1753  * @vmas are valid only as long as mmap_lock is held.
1754  *
1755  * Must be called with mmap_lock held for read or write.
1756  *
1757  * get_user_pages_remote walks a process's page tables and takes a reference
1758  * to each struct page that each user address corresponds to at a given
1759  * instant. That is, it takes the page that would be accessed if a user
1760  * thread accesses the given user virtual address at that instant.
1761  *
1762  * This does not guarantee that the page exists in the user mappings when
1763  * get_user_pages_remote returns, and there may even be a completely different
1764  * page there in some cases (eg. if mmapped pagecache has been invalidated
1765  * and subsequently re faulted). However it does guarantee that the page
1766  * won't be freed completely. And mostly callers simply care that the page
1767  * contains data that was valid *at some point in time*. Typically, an IO
1768  * or similar operation cannot guarantee anything stronger anyway because
1769  * locks can't be held over the syscall boundary.
1770  *
1771  * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1772  * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1773  * be called after the page is finished with, and before put_page is called.
1774  *
1775  * get_user_pages_remote is typically used for fewer-copy IO operations,
1776  * to get a handle on the memory by some means other than accesses
1777  * via the user virtual addresses. The pages may be submitted for
1778  * DMA to devices or accessed via their kernel linear mapping (via the
1779  * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1780  *
1781  * See also get_user_pages_fast, for performance critical applications.
1782  *
1783  * get_user_pages_remote should be phased out in favor of
1784  * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1785  * should use get_user_pages_remote because it cannot pass
1786  * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1787  */
get_user_pages_remote(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,struct vm_area_struct ** vmas,int * locked)1788 long get_user_pages_remote(struct mm_struct *mm,
1789 		unsigned long start, unsigned long nr_pages,
1790 		unsigned int gup_flags, struct page **pages,
1791 		struct vm_area_struct **vmas, int *locked)
1792 {
1793 	if (!is_valid_gup_flags(gup_flags))
1794 		return -EINVAL;
1795 
1796 	return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
1797 				       pages, vmas, locked);
1798 }
1799 EXPORT_SYMBOL(get_user_pages_remote);
1800 
1801 #else /* CONFIG_MMU */
get_user_pages_remote(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,struct vm_area_struct ** vmas,int * locked)1802 long get_user_pages_remote(struct mm_struct *mm,
1803 			   unsigned long start, unsigned long nr_pages,
1804 			   unsigned int gup_flags, struct page **pages,
1805 			   struct vm_area_struct **vmas, int *locked)
1806 {
1807 	return 0;
1808 }
1809 
__get_user_pages_remote(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,struct vm_area_struct ** vmas,int * locked)1810 static long __get_user_pages_remote(struct mm_struct *mm,
1811 				    unsigned long start, unsigned long nr_pages,
1812 				    unsigned int gup_flags, struct page **pages,
1813 				    struct vm_area_struct **vmas, int *locked)
1814 {
1815 	return 0;
1816 }
1817 #endif /* !CONFIG_MMU */
1818 
1819 /**
1820  * get_user_pages() - pin user pages in memory
1821  * @start:      starting user address
1822  * @nr_pages:   number of pages from start to pin
1823  * @gup_flags:  flags modifying lookup behaviour
1824  * @pages:      array that receives pointers to the pages pinned.
1825  *              Should be at least nr_pages long. Or NULL, if caller
1826  *              only intends to ensure the pages are faulted in.
1827  * @vmas:       array of pointers to vmas corresponding to each page.
1828  *              Or NULL if the caller does not require them.
1829  *
1830  * This is the same as get_user_pages_remote(), just with a less-flexible
1831  * calling convention where we assume that the mm being operated on belongs to
1832  * the current task, and doesn't allow passing of a locked parameter.  We also
1833  * obviously don't pass FOLL_REMOTE in here.
1834  */
get_user_pages(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,struct vm_area_struct ** vmas)1835 long get_user_pages(unsigned long start, unsigned long nr_pages,
1836 		unsigned int gup_flags, struct page **pages,
1837 		struct vm_area_struct **vmas)
1838 {
1839 	if (!is_valid_gup_flags(gup_flags))
1840 		return -EINVAL;
1841 
1842 	return __gup_longterm_locked(current->mm, start, nr_pages,
1843 				     pages, vmas, gup_flags | FOLL_TOUCH);
1844 }
1845 EXPORT_SYMBOL(get_user_pages);
1846 
1847 /**
1848  * get_user_pages_locked() - variant of get_user_pages()
1849  *
1850  * @start:      starting user address
1851  * @nr_pages:   number of pages from start to pin
1852  * @gup_flags:  flags modifying lookup behaviour
1853  * @pages:      array that receives pointers to the pages pinned.
1854  *              Should be at least nr_pages long. Or NULL, if caller
1855  *              only intends to ensure the pages are faulted in.
1856  * @locked:     pointer to lock flag indicating whether lock is held and
1857  *              subsequently whether VM_FAULT_RETRY functionality can be
1858  *              utilised. Lock must initially be held.
1859  *
1860  * It is suitable to replace the form:
1861  *
1862  *      mmap_read_lock(mm);
1863  *      do_something()
1864  *      get_user_pages(mm, ..., pages, NULL);
1865  *      mmap_read_unlock(mm);
1866  *
1867  *  to:
1868  *
1869  *      int locked = 1;
1870  *      mmap_read_lock(mm);
1871  *      do_something()
1872  *      get_user_pages_locked(mm, ..., pages, &locked);
1873  *      if (locked)
1874  *          mmap_read_unlock(mm);
1875  *
1876  * We can leverage the VM_FAULT_RETRY functionality in the page fault
1877  * paths better by using either get_user_pages_locked() or
1878  * get_user_pages_unlocked().
1879  *
1880  */
get_user_pages_locked(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)1881 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1882 			   unsigned int gup_flags, struct page **pages,
1883 			   int *locked)
1884 {
1885 	/*
1886 	 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1887 	 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1888 	 * vmas.  As there are no users of this flag in this call we simply
1889 	 * disallow this option for now.
1890 	 */
1891 	if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1892 		return -EINVAL;
1893 	/*
1894 	 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1895 	 * never directly by the caller, so enforce that:
1896 	 */
1897 	if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1898 		return -EINVAL;
1899 
1900 	return __get_user_pages_locked(current->mm, start, nr_pages,
1901 				       pages, NULL, locked,
1902 				       gup_flags | FOLL_TOUCH);
1903 }
1904 EXPORT_SYMBOL(get_user_pages_locked);
1905 
1906 /*
1907  * get_user_pages_unlocked() is suitable to replace the form:
1908  *
1909  *      mmap_read_lock(mm);
1910  *      get_user_pages(mm, ..., pages, NULL);
1911  *      mmap_read_unlock(mm);
1912  *
1913  *  with:
1914  *
1915  *      get_user_pages_unlocked(mm, ..., pages);
1916  *
1917  * It is functionally equivalent to get_user_pages_fast so
1918  * get_user_pages_fast should be used instead if specific gup_flags
1919  * (e.g. FOLL_FORCE) are not required.
1920  */
get_user_pages_unlocked(unsigned long start,unsigned long nr_pages,struct page ** pages,unsigned int gup_flags)1921 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1922 			     struct page **pages, unsigned int gup_flags)
1923 {
1924 	struct mm_struct *mm = current->mm;
1925 	int locked = 1;
1926 	long ret;
1927 
1928 	/*
1929 	 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1930 	 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1931 	 * vmas.  As there are no users of this flag in this call we simply
1932 	 * disallow this option for now.
1933 	 */
1934 	if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1935 		return -EINVAL;
1936 
1937 	mmap_read_lock(mm);
1938 	ret = __get_user_pages_locked(mm, start, nr_pages, pages, NULL,
1939 				      &locked, gup_flags | FOLL_TOUCH);
1940 	if (locked)
1941 		mmap_read_unlock(mm);
1942 	return ret;
1943 }
1944 EXPORT_SYMBOL(get_user_pages_unlocked);
1945 
1946 /*
1947  * Fast GUP
1948  *
1949  * get_user_pages_fast attempts to pin user pages by walking the page
1950  * tables directly and avoids taking locks. Thus the walker needs to be
1951  * protected from page table pages being freed from under it, and should
1952  * block any THP splits.
1953  *
1954  * One way to achieve this is to have the walker disable interrupts, and
1955  * rely on IPIs from the TLB flushing code blocking before the page table
1956  * pages are freed. This is unsuitable for architectures that do not need
1957  * to broadcast an IPI when invalidating TLBs.
1958  *
1959  * Another way to achieve this is to batch up page table containing pages
1960  * belonging to more than one mm_user, then rcu_sched a callback to free those
1961  * pages. Disabling interrupts will allow the fast_gup walker to both block
1962  * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1963  * (which is a relatively rare event). The code below adopts this strategy.
1964  *
1965  * Before activating this code, please be aware that the following assumptions
1966  * are currently made:
1967  *
1968  *  *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1969  *  free pages containing page tables or TLB flushing requires IPI broadcast.
1970  *
1971  *  *) ptes can be read atomically by the architecture.
1972  *
1973  *  *) access_ok is sufficient to validate userspace address ranges.
1974  *
1975  * The last two assumptions can be relaxed by the addition of helper functions.
1976  *
1977  * This code is based heavily on the PowerPC implementation by Nick Piggin.
1978  */
1979 #ifdef CONFIG_HAVE_FAST_GUP
1980 
undo_dev_pagemap(int * nr,int nr_start,unsigned int flags,struct page ** pages)1981 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
1982 					    unsigned int flags,
1983 					    struct page **pages)
1984 {
1985 	while ((*nr) - nr_start) {
1986 		struct page *page = pages[--(*nr)];
1987 
1988 		ClearPageReferenced(page);
1989 		if (flags & FOLL_PIN)
1990 			unpin_user_page(page);
1991 		else
1992 			put_page(page);
1993 	}
1994 }
1995 
1996 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
gup_pte_range(pmd_t pmd,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)1997 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1998 			 unsigned int flags, struct page **pages, int *nr)
1999 {
2000 	struct dev_pagemap *pgmap = NULL;
2001 	int nr_start = *nr, ret = 0;
2002 	pte_t *ptep, *ptem;
2003 
2004 	ptem = ptep = pte_offset_map(&pmd, addr);
2005 	do {
2006 		pte_t pte = ptep_get_lockless(ptep);
2007 		struct page *head, *page;
2008 
2009 		/*
2010 		 * Similar to the PMD case below, NUMA hinting must take slow
2011 		 * path using the pte_protnone check.
2012 		 */
2013 		if (pte_protnone(pte))
2014 			goto pte_unmap;
2015 
2016 		if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2017 			goto pte_unmap;
2018 
2019 		if (pte_devmap(pte)) {
2020 			if (unlikely(flags & FOLL_LONGTERM))
2021 				goto pte_unmap;
2022 
2023 			pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2024 			if (unlikely(!pgmap)) {
2025 				undo_dev_pagemap(nr, nr_start, flags, pages);
2026 				goto pte_unmap;
2027 			}
2028 		} else if (pte_special(pte))
2029 			goto pte_unmap;
2030 
2031 		VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2032 		page = pte_page(pte);
2033 
2034 		head = try_grab_compound_head(page, 1, flags);
2035 		if (!head)
2036 			goto pte_unmap;
2037 
2038 		if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2039 			put_compound_head(head, 1, flags);
2040 			goto pte_unmap;
2041 		}
2042 
2043 		VM_BUG_ON_PAGE(compound_head(page) != head, page);
2044 
2045 		/*
2046 		 * We need to make the page accessible if and only if we are
2047 		 * going to access its content (the FOLL_PIN case).  Please
2048 		 * see Documentation/core-api/pin_user_pages.rst for
2049 		 * details.
2050 		 */
2051 		if (flags & FOLL_PIN) {
2052 			ret = arch_make_page_accessible(page);
2053 			if (ret) {
2054 				unpin_user_page(page);
2055 				goto pte_unmap;
2056 			}
2057 		}
2058 		SetPageReferenced(page);
2059 		pages[*nr] = page;
2060 		(*nr)++;
2061 
2062 	} while (ptep++, addr += PAGE_SIZE, addr != end);
2063 
2064 	ret = 1;
2065 
2066 pte_unmap:
2067 	if (pgmap)
2068 		put_dev_pagemap(pgmap);
2069 	pte_unmap(ptem);
2070 	return ret;
2071 }
2072 #else
2073 
2074 /*
2075  * If we can't determine whether or not a pte is special, then fail immediately
2076  * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2077  * to be special.
2078  *
2079  * For a futex to be placed on a THP tail page, get_futex_key requires a
2080  * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2081  * useful to have gup_huge_pmd even if we can't operate on ptes.
2082  */
gup_pte_range(pmd_t pmd,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2083 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2084 			 unsigned int flags, struct page **pages, int *nr)
2085 {
2086 	return 0;
2087 }
2088 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2089 
2090 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
__gup_device_huge(unsigned long pfn,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2091 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2092 			     unsigned long end, unsigned int flags,
2093 			     struct page **pages, int *nr)
2094 {
2095 	int nr_start = *nr;
2096 	struct dev_pagemap *pgmap = NULL;
2097 
2098 	do {
2099 		struct page *page = pfn_to_page(pfn);
2100 
2101 		pgmap = get_dev_pagemap(pfn, pgmap);
2102 		if (unlikely(!pgmap)) {
2103 			undo_dev_pagemap(nr, nr_start, flags, pages);
2104 			return 0;
2105 		}
2106 		SetPageReferenced(page);
2107 		pages[*nr] = page;
2108 		if (unlikely(!try_grab_page(page, flags))) {
2109 			undo_dev_pagemap(nr, nr_start, flags, pages);
2110 			return 0;
2111 		}
2112 		(*nr)++;
2113 		pfn++;
2114 	} while (addr += PAGE_SIZE, addr != end);
2115 
2116 	if (pgmap)
2117 		put_dev_pagemap(pgmap);
2118 	return 1;
2119 }
2120 
__gup_device_huge_pmd(pmd_t orig,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2121 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2122 				 unsigned long end, unsigned int flags,
2123 				 struct page **pages, int *nr)
2124 {
2125 	unsigned long fault_pfn;
2126 	int nr_start = *nr;
2127 
2128 	fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2129 	if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2130 		return 0;
2131 
2132 	if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2133 		undo_dev_pagemap(nr, nr_start, flags, pages);
2134 		return 0;
2135 	}
2136 	return 1;
2137 }
2138 
__gup_device_huge_pud(pud_t orig,pud_t * pudp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2139 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2140 				 unsigned long end, unsigned int flags,
2141 				 struct page **pages, int *nr)
2142 {
2143 	unsigned long fault_pfn;
2144 	int nr_start = *nr;
2145 
2146 	fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2147 	if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2148 		return 0;
2149 
2150 	if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2151 		undo_dev_pagemap(nr, nr_start, flags, pages);
2152 		return 0;
2153 	}
2154 	return 1;
2155 }
2156 #else
__gup_device_huge_pmd(pmd_t orig,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2157 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2158 				 unsigned long end, unsigned int flags,
2159 				 struct page **pages, int *nr)
2160 {
2161 	BUILD_BUG();
2162 	return 0;
2163 }
2164 
__gup_device_huge_pud(pud_t pud,pud_t * pudp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2165 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2166 				 unsigned long end, unsigned int flags,
2167 				 struct page **pages, int *nr)
2168 {
2169 	BUILD_BUG();
2170 	return 0;
2171 }
2172 #endif
2173 
record_subpages(struct page * page,unsigned long addr,unsigned long end,struct page ** pages)2174 static int record_subpages(struct page *page, unsigned long addr,
2175 			   unsigned long end, struct page **pages)
2176 {
2177 	int nr;
2178 
2179 	for (nr = 0; addr != end; addr += PAGE_SIZE)
2180 		pages[nr++] = page++;
2181 
2182 	return nr;
2183 }
2184 
2185 #ifdef CONFIG_ARCH_HAS_HUGEPD
hugepte_addr_end(unsigned long addr,unsigned long end,unsigned long sz)2186 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2187 				      unsigned long sz)
2188 {
2189 	unsigned long __boundary = (addr + sz) & ~(sz-1);
2190 	return (__boundary - 1 < end - 1) ? __boundary : end;
2191 }
2192 
gup_hugepte(pte_t * ptep,unsigned long sz,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2193 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2194 		       unsigned long end, unsigned int flags,
2195 		       struct page **pages, int *nr)
2196 {
2197 	unsigned long pte_end;
2198 	struct page *head, *page;
2199 	pte_t pte;
2200 	int refs;
2201 
2202 	pte_end = (addr + sz) & ~(sz-1);
2203 	if (pte_end < end)
2204 		end = pte_end;
2205 
2206 	pte = huge_ptep_get(ptep);
2207 
2208 	if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2209 		return 0;
2210 
2211 	/* hugepages are never "special" */
2212 	VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2213 
2214 	head = pte_page(pte);
2215 	page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2216 	refs = record_subpages(page, addr, end, pages + *nr);
2217 
2218 	head = try_grab_compound_head(head, refs, flags);
2219 	if (!head)
2220 		return 0;
2221 
2222 	if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2223 		put_compound_head(head, refs, flags);
2224 		return 0;
2225 	}
2226 
2227 	*nr += refs;
2228 	SetPageReferenced(head);
2229 	return 1;
2230 }
2231 
gup_huge_pd(hugepd_t hugepd,unsigned long addr,unsigned int pdshift,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2232 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2233 		unsigned int pdshift, unsigned long end, unsigned int flags,
2234 		struct page **pages, int *nr)
2235 {
2236 	pte_t *ptep;
2237 	unsigned long sz = 1UL << hugepd_shift(hugepd);
2238 	unsigned long next;
2239 
2240 	ptep = hugepte_offset(hugepd, addr, pdshift);
2241 	do {
2242 		next = hugepte_addr_end(addr, end, sz);
2243 		if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2244 			return 0;
2245 	} while (ptep++, addr = next, addr != end);
2246 
2247 	return 1;
2248 }
2249 #else
gup_huge_pd(hugepd_t hugepd,unsigned long addr,unsigned int pdshift,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2250 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2251 		unsigned int pdshift, unsigned long end, unsigned int flags,
2252 		struct page **pages, int *nr)
2253 {
2254 	return 0;
2255 }
2256 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2257 
gup_huge_pmd(pmd_t orig,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2258 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2259 			unsigned long end, unsigned int flags,
2260 			struct page **pages, int *nr)
2261 {
2262 	struct page *head, *page;
2263 	int refs;
2264 
2265 	if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2266 		return 0;
2267 
2268 	if (pmd_devmap(orig)) {
2269 		if (unlikely(flags & FOLL_LONGTERM))
2270 			return 0;
2271 		return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2272 					     pages, nr);
2273 	}
2274 
2275 	page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2276 	refs = record_subpages(page, addr, end, pages + *nr);
2277 
2278 	head = try_grab_compound_head(pmd_page(orig), refs, flags);
2279 	if (!head)
2280 		return 0;
2281 
2282 	if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2283 		put_compound_head(head, refs, flags);
2284 		return 0;
2285 	}
2286 
2287 	*nr += refs;
2288 	SetPageReferenced(head);
2289 	return 1;
2290 }
2291 
gup_huge_pud(pud_t orig,pud_t * pudp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2292 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2293 			unsigned long end, unsigned int flags,
2294 			struct page **pages, int *nr)
2295 {
2296 	struct page *head, *page;
2297 	int refs;
2298 
2299 	if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2300 		return 0;
2301 
2302 	if (pud_devmap(orig)) {
2303 		if (unlikely(flags & FOLL_LONGTERM))
2304 			return 0;
2305 		return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2306 					     pages, nr);
2307 	}
2308 
2309 	page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2310 	refs = record_subpages(page, addr, end, pages + *nr);
2311 
2312 	head = try_grab_compound_head(pud_page(orig), refs, flags);
2313 	if (!head)
2314 		return 0;
2315 
2316 	if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2317 		put_compound_head(head, refs, flags);
2318 		return 0;
2319 	}
2320 
2321 	*nr += refs;
2322 	SetPageReferenced(head);
2323 	return 1;
2324 }
2325 
gup_huge_pgd(pgd_t orig,pgd_t * pgdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2326 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2327 			unsigned long end, unsigned int flags,
2328 			struct page **pages, int *nr)
2329 {
2330 	int refs;
2331 	struct page *head, *page;
2332 
2333 	if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2334 		return 0;
2335 
2336 	BUILD_BUG_ON(pgd_devmap(orig));
2337 
2338 	page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2339 	refs = record_subpages(page, addr, end, pages + *nr);
2340 
2341 	head = try_grab_compound_head(pgd_page(orig), refs, flags);
2342 	if (!head)
2343 		return 0;
2344 
2345 	if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2346 		put_compound_head(head, refs, flags);
2347 		return 0;
2348 	}
2349 
2350 	*nr += refs;
2351 	SetPageReferenced(head);
2352 	return 1;
2353 }
2354 
gup_pmd_range(pud_t * pudp,pud_t pud,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2355 static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
2356 		unsigned int flags, struct page **pages, int *nr)
2357 {
2358 	unsigned long next;
2359 	pmd_t *pmdp;
2360 
2361 	pmdp = pmd_offset_lockless(pudp, pud, addr);
2362 	do {
2363 		pmd_t pmd = READ_ONCE(*pmdp);
2364 
2365 		next = pmd_addr_end(addr, end);
2366 		if (!pmd_present(pmd))
2367 			return 0;
2368 
2369 		if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2370 			     pmd_devmap(pmd))) {
2371 			/*
2372 			 * NUMA hinting faults need to be handled in the GUP
2373 			 * slowpath for accounting purposes and so that they
2374 			 * can be serialised against THP migration.
2375 			 */
2376 			if (pmd_protnone(pmd))
2377 				return 0;
2378 
2379 			if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2380 				pages, nr))
2381 				return 0;
2382 
2383 		} else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2384 			/*
2385 			 * architecture have different format for hugetlbfs
2386 			 * pmd format and THP pmd format
2387 			 */
2388 			if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2389 					 PMD_SHIFT, next, flags, pages, nr))
2390 				return 0;
2391 		} else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2392 			return 0;
2393 	} while (pmdp++, addr = next, addr != end);
2394 
2395 	return 1;
2396 }
2397 
gup_pud_range(p4d_t * p4dp,p4d_t p4d,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2398 static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
2399 			 unsigned int flags, struct page **pages, int *nr)
2400 {
2401 	unsigned long next;
2402 	pud_t *pudp;
2403 
2404 	pudp = pud_offset_lockless(p4dp, p4d, addr);
2405 	do {
2406 		pud_t pud = READ_ONCE(*pudp);
2407 
2408 		next = pud_addr_end(addr, end);
2409 		if (unlikely(!pud_present(pud)))
2410 			return 0;
2411 		if (unlikely(pud_huge(pud))) {
2412 			if (!gup_huge_pud(pud, pudp, addr, next, flags,
2413 					  pages, nr))
2414 				return 0;
2415 		} else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2416 			if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2417 					 PUD_SHIFT, next, flags, pages, nr))
2418 				return 0;
2419 		} else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
2420 			return 0;
2421 	} while (pudp++, addr = next, addr != end);
2422 
2423 	return 1;
2424 }
2425 
gup_p4d_range(pgd_t * pgdp,pgd_t pgd,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2426 static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
2427 			 unsigned int flags, struct page **pages, int *nr)
2428 {
2429 	unsigned long next;
2430 	p4d_t *p4dp;
2431 
2432 	p4dp = p4d_offset_lockless(pgdp, pgd, addr);
2433 	do {
2434 		p4d_t p4d = READ_ONCE(*p4dp);
2435 
2436 		next = p4d_addr_end(addr, end);
2437 		if (p4d_none(p4d))
2438 			return 0;
2439 		BUILD_BUG_ON(p4d_huge(p4d));
2440 		if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2441 			if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2442 					 P4D_SHIFT, next, flags, pages, nr))
2443 				return 0;
2444 		} else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
2445 			return 0;
2446 	} while (p4dp++, addr = next, addr != end);
2447 
2448 	return 1;
2449 }
2450 
gup_pgd_range(unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2451 static void gup_pgd_range(unsigned long addr, unsigned long end,
2452 		unsigned int flags, struct page **pages, int *nr)
2453 {
2454 	unsigned long next;
2455 	pgd_t *pgdp;
2456 
2457 	pgdp = pgd_offset(current->mm, addr);
2458 	do {
2459 		pgd_t pgd = READ_ONCE(*pgdp);
2460 
2461 		next = pgd_addr_end(addr, end);
2462 		if (pgd_none(pgd))
2463 			return;
2464 		if (unlikely(pgd_huge(pgd))) {
2465 			if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2466 					  pages, nr))
2467 				return;
2468 		} else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2469 			if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2470 					 PGDIR_SHIFT, next, flags, pages, nr))
2471 				return;
2472 		} else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
2473 			return;
2474 	} while (pgdp++, addr = next, addr != end);
2475 }
2476 #else
gup_pgd_range(unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2477 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2478 		unsigned int flags, struct page **pages, int *nr)
2479 {
2480 }
2481 #endif /* CONFIG_HAVE_FAST_GUP */
2482 
2483 #ifndef gup_fast_permitted
2484 /*
2485  * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2486  * we need to fall back to the slow version:
2487  */
gup_fast_permitted(unsigned long start,unsigned long end)2488 static bool gup_fast_permitted(unsigned long start, unsigned long end)
2489 {
2490 	return true;
2491 }
2492 #endif
2493 
__gup_longterm_unlocked(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)2494 static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2495 				   unsigned int gup_flags, struct page **pages)
2496 {
2497 	int ret;
2498 
2499 	/*
2500 	 * FIXME: FOLL_LONGTERM does not work with
2501 	 * get_user_pages_unlocked() (see comments in that function)
2502 	 */
2503 	if (gup_flags & FOLL_LONGTERM) {
2504 		mmap_read_lock(current->mm);
2505 		ret = __gup_longterm_locked(current->mm,
2506 					    start, nr_pages,
2507 					    pages, NULL, gup_flags);
2508 		mmap_read_unlock(current->mm);
2509 	} else {
2510 		ret = get_user_pages_unlocked(start, nr_pages,
2511 					      pages, gup_flags);
2512 	}
2513 
2514 	return ret;
2515 }
2516 
lockless_pages_from_mm(unsigned long start,unsigned long end,unsigned int gup_flags,struct page ** pages)2517 static unsigned long lockless_pages_from_mm(unsigned long start,
2518 					    unsigned long end,
2519 					    unsigned int gup_flags,
2520 					    struct page **pages)
2521 {
2522 	unsigned long flags;
2523 	int nr_pinned = 0;
2524 	unsigned seq;
2525 
2526 	if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP) ||
2527 	    !gup_fast_permitted(start, end))
2528 		return 0;
2529 
2530 	if (gup_flags & FOLL_PIN) {
2531 		seq = raw_read_seqcount(&current->mm->write_protect_seq);
2532 		if (seq & 1)
2533 			return 0;
2534 	}
2535 
2536 	/*
2537 	 * Disable interrupts. The nested form is used, in order to allow full,
2538 	 * general purpose use of this routine.
2539 	 *
2540 	 * With interrupts disabled, we block page table pages from being freed
2541 	 * from under us. See struct mmu_table_batch comments in
2542 	 * include/asm-generic/tlb.h for more details.
2543 	 *
2544 	 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
2545 	 * that come from THPs splitting.
2546 	 */
2547 	local_irq_save(flags);
2548 	gup_pgd_range(start, end, gup_flags, pages, &nr_pinned);
2549 	local_irq_restore(flags);
2550 
2551 	/*
2552 	 * When pinning pages for DMA there could be a concurrent write protect
2553 	 * from fork() via copy_page_range(), in this case always fail fast GUP.
2554 	 */
2555 	if (gup_flags & FOLL_PIN) {
2556 		if (read_seqcount_retry(&current->mm->write_protect_seq, seq)) {
2557 			unpin_user_pages(pages, nr_pinned);
2558 			return 0;
2559 		}
2560 	}
2561 	return nr_pinned;
2562 }
2563 
internal_get_user_pages_fast(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages)2564 static int internal_get_user_pages_fast(unsigned long start,
2565 					unsigned long nr_pages,
2566 					unsigned int gup_flags,
2567 					struct page **pages)
2568 {
2569 	unsigned long len, end;
2570 	unsigned long nr_pinned;
2571 	int ret;
2572 
2573 	if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
2574 				       FOLL_FORCE | FOLL_PIN | FOLL_GET |
2575 				       FOLL_FAST_ONLY)))
2576 		return -EINVAL;
2577 
2578 	if (gup_flags & FOLL_PIN)
2579 		atomic_set(&current->mm->has_pinned, 1);
2580 
2581 	if (!(gup_flags & FOLL_FAST_ONLY))
2582 		might_lock_read(&current->mm->mmap_lock);
2583 
2584 	start = untagged_addr(start) & PAGE_MASK;
2585 	len = nr_pages << PAGE_SHIFT;
2586 	if (check_add_overflow(start, len, &end))
2587 		return 0;
2588 	if (unlikely(!access_ok((void __user *)start, len)))
2589 		return -EFAULT;
2590 
2591 	nr_pinned = lockless_pages_from_mm(start, end, gup_flags, pages);
2592 	if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
2593 		return nr_pinned;
2594 
2595 	/* Slow path: try to get the remaining pages with get_user_pages */
2596 	start += nr_pinned << PAGE_SHIFT;
2597 	pages += nr_pinned;
2598 	ret = __gup_longterm_unlocked(start, nr_pages - nr_pinned, gup_flags,
2599 				      pages);
2600 	if (ret < 0) {
2601 		/*
2602 		 * The caller has to unpin the pages we already pinned so
2603 		 * returning -errno is not an option
2604 		 */
2605 		if (nr_pinned)
2606 			return nr_pinned;
2607 		return ret;
2608 	}
2609 	return ret + nr_pinned;
2610 }
2611 
2612 /**
2613  * get_user_pages_fast_only() - pin user pages in memory
2614  * @start:      starting user address
2615  * @nr_pages:   number of pages from start to pin
2616  * @gup_flags:  flags modifying pin behaviour
2617  * @pages:      array that receives pointers to the pages pinned.
2618  *              Should be at least nr_pages long.
2619  *
2620  * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2621  * the regular GUP.
2622  * Note a difference with get_user_pages_fast: this always returns the
2623  * number of pages pinned, 0 if no pages were pinned.
2624  *
2625  * If the architecture does not support this function, simply return with no
2626  * pages pinned.
2627  *
2628  * Careful, careful! COW breaking can go either way, so a non-write
2629  * access can get ambiguous page results. If you call this function without
2630  * 'write' set, you'd better be sure that you're ok with that ambiguity.
2631  */
get_user_pages_fast_only(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)2632 int get_user_pages_fast_only(unsigned long start, int nr_pages,
2633 			     unsigned int gup_flags, struct page **pages)
2634 {
2635 	int nr_pinned;
2636 	/*
2637 	 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2638 	 * because gup fast is always a "pin with a +1 page refcount" request.
2639 	 *
2640 	 * FOLL_FAST_ONLY is required in order to match the API description of
2641 	 * this routine: no fall back to regular ("slow") GUP.
2642 	 */
2643 	gup_flags |= FOLL_GET | FOLL_FAST_ONLY;
2644 
2645 	nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2646 						 pages);
2647 
2648 	/*
2649 	 * As specified in the API description above, this routine is not
2650 	 * allowed to return negative values. However, the common core
2651 	 * routine internal_get_user_pages_fast() *can* return -errno.
2652 	 * Therefore, correct for that here:
2653 	 */
2654 	if (nr_pinned < 0)
2655 		nr_pinned = 0;
2656 
2657 	return nr_pinned;
2658 }
2659 EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
2660 
2661 /**
2662  * get_user_pages_fast() - pin user pages in memory
2663  * @start:      starting user address
2664  * @nr_pages:   number of pages from start to pin
2665  * @gup_flags:  flags modifying pin behaviour
2666  * @pages:      array that receives pointers to the pages pinned.
2667  *              Should be at least nr_pages long.
2668  *
2669  * Attempt to pin user pages in memory without taking mm->mmap_lock.
2670  * If not successful, it will fall back to taking the lock and
2671  * calling get_user_pages().
2672  *
2673  * Returns number of pages pinned. This may be fewer than the number requested.
2674  * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2675  * -errno.
2676  */
get_user_pages_fast(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)2677 int get_user_pages_fast(unsigned long start, int nr_pages,
2678 			unsigned int gup_flags, struct page **pages)
2679 {
2680 	if (!is_valid_gup_flags(gup_flags))
2681 		return -EINVAL;
2682 
2683 	/*
2684 	 * The caller may or may not have explicitly set FOLL_GET; either way is
2685 	 * OK. However, internally (within mm/gup.c), gup fast variants must set
2686 	 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2687 	 * request.
2688 	 */
2689 	gup_flags |= FOLL_GET;
2690 	return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2691 }
2692 EXPORT_SYMBOL_GPL(get_user_pages_fast);
2693 
2694 /**
2695  * pin_user_pages_fast() - pin user pages in memory without taking locks
2696  *
2697  * @start:      starting user address
2698  * @nr_pages:   number of pages from start to pin
2699  * @gup_flags:  flags modifying pin behaviour
2700  * @pages:      array that receives pointers to the pages pinned.
2701  *              Should be at least nr_pages long.
2702  *
2703  * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2704  * get_user_pages_fast() for documentation on the function arguments, because
2705  * the arguments here are identical.
2706  *
2707  * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2708  * see Documentation/core-api/pin_user_pages.rst for further details.
2709  */
pin_user_pages_fast(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)2710 int pin_user_pages_fast(unsigned long start, int nr_pages,
2711 			unsigned int gup_flags, struct page **pages)
2712 {
2713 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
2714 	if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2715 		return -EINVAL;
2716 
2717 	gup_flags |= FOLL_PIN;
2718 	return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2719 }
2720 EXPORT_SYMBOL_GPL(pin_user_pages_fast);
2721 
2722 /*
2723  * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
2724  * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
2725  *
2726  * The API rules are the same, too: no negative values may be returned.
2727  */
pin_user_pages_fast_only(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)2728 int pin_user_pages_fast_only(unsigned long start, int nr_pages,
2729 			     unsigned int gup_flags, struct page **pages)
2730 {
2731 	int nr_pinned;
2732 
2733 	/*
2734 	 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
2735 	 * rules require returning 0, rather than -errno:
2736 	 */
2737 	if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2738 		return 0;
2739 	/*
2740 	 * FOLL_FAST_ONLY is required in order to match the API description of
2741 	 * this routine: no fall back to regular ("slow") GUP.
2742 	 */
2743 	gup_flags |= (FOLL_PIN | FOLL_FAST_ONLY);
2744 	nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2745 						 pages);
2746 	/*
2747 	 * This routine is not allowed to return negative values. However,
2748 	 * internal_get_user_pages_fast() *can* return -errno. Therefore,
2749 	 * correct for that here:
2750 	 */
2751 	if (nr_pinned < 0)
2752 		nr_pinned = 0;
2753 
2754 	return nr_pinned;
2755 }
2756 EXPORT_SYMBOL_GPL(pin_user_pages_fast_only);
2757 
2758 /**
2759  * pin_user_pages_remote() - pin pages of a remote process
2760  *
2761  * @mm:		mm_struct of target mm
2762  * @start:	starting user address
2763  * @nr_pages:	number of pages from start to pin
2764  * @gup_flags:	flags modifying lookup behaviour
2765  * @pages:	array that receives pointers to the pages pinned.
2766  *		Should be at least nr_pages long. Or NULL, if caller
2767  *		only intends to ensure the pages are faulted in.
2768  * @vmas:	array of pointers to vmas corresponding to each page.
2769  *		Or NULL if the caller does not require them.
2770  * @locked:	pointer to lock flag indicating whether lock is held and
2771  *		subsequently whether VM_FAULT_RETRY functionality can be
2772  *		utilised. Lock must initially be held.
2773  *
2774  * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2775  * get_user_pages_remote() for documentation on the function arguments, because
2776  * the arguments here are identical.
2777  *
2778  * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2779  * see Documentation/core-api/pin_user_pages.rst for details.
2780  */
pin_user_pages_remote(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,struct vm_area_struct ** vmas,int * locked)2781 long pin_user_pages_remote(struct mm_struct *mm,
2782 			   unsigned long start, unsigned long nr_pages,
2783 			   unsigned int gup_flags, struct page **pages,
2784 			   struct vm_area_struct **vmas, int *locked)
2785 {
2786 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
2787 	if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2788 		return -EINVAL;
2789 
2790 	gup_flags |= FOLL_PIN;
2791 	return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
2792 				       pages, vmas, locked);
2793 }
2794 EXPORT_SYMBOL(pin_user_pages_remote);
2795 
2796 /**
2797  * pin_user_pages() - pin user pages in memory for use by other devices
2798  *
2799  * @start:	starting user address
2800  * @nr_pages:	number of pages from start to pin
2801  * @gup_flags:	flags modifying lookup behaviour
2802  * @pages:	array that receives pointers to the pages pinned.
2803  *		Should be at least nr_pages long. Or NULL, if caller
2804  *		only intends to ensure the pages are faulted in.
2805  * @vmas:	array of pointers to vmas corresponding to each page.
2806  *		Or NULL if the caller does not require them.
2807  *
2808  * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2809  * FOLL_PIN is set.
2810  *
2811  * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2812  * see Documentation/core-api/pin_user_pages.rst for details.
2813  */
pin_user_pages(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,struct vm_area_struct ** vmas)2814 long pin_user_pages(unsigned long start, unsigned long nr_pages,
2815 		    unsigned int gup_flags, struct page **pages,
2816 		    struct vm_area_struct **vmas)
2817 {
2818 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
2819 	if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2820 		return -EINVAL;
2821 
2822 	gup_flags |= FOLL_PIN;
2823 	return __gup_longterm_locked(current->mm, start, nr_pages,
2824 				     pages, vmas, gup_flags);
2825 }
2826 EXPORT_SYMBOL(pin_user_pages);
2827 
2828 /*
2829  * pin_user_pages_unlocked() is the FOLL_PIN variant of
2830  * get_user_pages_unlocked(). Behavior is the same, except that this one sets
2831  * FOLL_PIN and rejects FOLL_GET.
2832  */
pin_user_pages_unlocked(unsigned long start,unsigned long nr_pages,struct page ** pages,unsigned int gup_flags)2833 long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2834 			     struct page **pages, unsigned int gup_flags)
2835 {
2836 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
2837 	if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2838 		return -EINVAL;
2839 
2840 	gup_flags |= FOLL_PIN;
2841 	return get_user_pages_unlocked(start, nr_pages, pages, gup_flags);
2842 }
2843 EXPORT_SYMBOL(pin_user_pages_unlocked);
2844 
2845 /*
2846  * pin_user_pages_locked() is the FOLL_PIN variant of get_user_pages_locked().
2847  * Behavior is the same, except that this one sets FOLL_PIN and rejects
2848  * FOLL_GET.
2849  */
pin_user_pages_locked(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)2850 long pin_user_pages_locked(unsigned long start, unsigned long nr_pages,
2851 			   unsigned int gup_flags, struct page **pages,
2852 			   int *locked)
2853 {
2854 	/*
2855 	 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2856 	 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2857 	 * vmas.  As there are no users of this flag in this call we simply
2858 	 * disallow this option for now.
2859 	 */
2860 	if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2861 		return -EINVAL;
2862 
2863 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
2864 	if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2865 		return -EINVAL;
2866 
2867 	gup_flags |= FOLL_PIN;
2868 	return __get_user_pages_locked(current->mm, start, nr_pages,
2869 				       pages, NULL, locked,
2870 				       gup_flags | FOLL_TOUCH);
2871 }
2872 EXPORT_SYMBOL(pin_user_pages_locked);
2873