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 #include <linux/secretmem.h>
14 
15 #include <linux/sched/signal.h>
16 #include <linux/rwsem.h>
17 #include <linux/hugetlb.h>
18 #include <linux/migrate.h>
19 #include <linux/mm_inline.h>
20 #include <linux/sched/mm.h>
21 #include <linux/shmem_fs.h>
22 
23 #include <asm/mmu_context.h>
24 #include <asm/tlbflush.h>
25 
26 #include "internal.h"
27 
28 struct follow_page_context {
29 	struct dev_pagemap *pgmap;
30 	unsigned int page_mask;
31 };
32 
sanity_check_pinned_pages(struct page ** pages,unsigned long npages)33 static inline void sanity_check_pinned_pages(struct page **pages,
34 					     unsigned long npages)
35 {
36 	if (!IS_ENABLED(CONFIG_DEBUG_VM))
37 		return;
38 
39 	/*
40 	 * We only pin anonymous pages if they are exclusive. Once pinned, we
41 	 * can no longer turn them possibly shared and PageAnonExclusive() will
42 	 * stick around until the page is freed.
43 	 *
44 	 * We'd like to verify that our pinned anonymous pages are still mapped
45 	 * exclusively. The issue with anon THP is that we don't know how
46 	 * they are/were mapped when pinning them. However, for anon
47 	 * THP we can assume that either the given page (PTE-mapped THP) or
48 	 * the head page (PMD-mapped THP) should be PageAnonExclusive(). If
49 	 * neither is the case, there is certainly something wrong.
50 	 */
51 	for (; npages; npages--, pages++) {
52 		struct page *page = *pages;
53 		struct folio *folio = page_folio(page);
54 
55 		if (is_zero_page(page) ||
56 		    !folio_test_anon(folio))
57 			continue;
58 		if (!folio_test_large(folio) || folio_test_hugetlb(folio))
59 			VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page), page);
60 		else
61 			/* Either a PTE-mapped or a PMD-mapped THP. */
62 			VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page) &&
63 				       !PageAnonExclusive(page), page);
64 	}
65 }
66 
67 /*
68  * Return the folio with ref appropriately incremented,
69  * or NULL if that failed.
70  */
try_get_folio(struct page * page,int refs)71 static inline struct folio *try_get_folio(struct page *page, int refs)
72 {
73 	struct folio *folio;
74 
75 retry:
76 	folio = page_folio(page);
77 	if (WARN_ON_ONCE(folio_ref_count(folio) < 0))
78 		return NULL;
79 	if (unlikely(!folio_ref_try_add_rcu(folio, refs)))
80 		return NULL;
81 
82 	/*
83 	 * At this point we have a stable reference to the folio; but it
84 	 * could be that between calling page_folio() and the refcount
85 	 * increment, the folio was split, in which case we'd end up
86 	 * holding a reference on a folio that has nothing to do with the page
87 	 * we were given anymore.
88 	 * So now that the folio is stable, recheck that the page still
89 	 * belongs to this folio.
90 	 */
91 	if (unlikely(page_folio(page) != folio)) {
92 		if (!put_devmap_managed_page_refs(&folio->page, refs))
93 			folio_put_refs(folio, refs);
94 		goto retry;
95 	}
96 
97 	return folio;
98 }
99 
100 /**
101  * try_grab_folio() - Attempt to get or pin a folio.
102  * @page:  pointer to page to be grabbed
103  * @refs:  the value to (effectively) add to the folio's refcount
104  * @flags: gup flags: these are the FOLL_* flag values.
105  *
106  * "grab" names in this file mean, "look at flags to decide whether to use
107  * FOLL_PIN or FOLL_GET behavior, when incrementing the folio's refcount.
108  *
109  * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
110  * same time. (That's true throughout the get_user_pages*() and
111  * pin_user_pages*() APIs.) Cases:
112  *
113  *    FOLL_GET: folio's refcount will be incremented by @refs.
114  *
115  *    FOLL_PIN on large folios: folio's refcount will be incremented by
116  *    @refs, and its pincount will be incremented by @refs.
117  *
118  *    FOLL_PIN on single-page folios: folio's refcount will be incremented by
119  *    @refs * GUP_PIN_COUNTING_BIAS.
120  *
121  * Return: The folio containing @page (with refcount appropriately
122  * incremented) for success, or NULL upon failure. If neither FOLL_GET
123  * nor FOLL_PIN was set, that's considered failure, and furthermore,
124  * a likely bug in the caller, so a warning is also emitted.
125  */
try_grab_folio(struct page * page,int refs,unsigned int flags)126 struct folio *try_grab_folio(struct page *page, int refs, unsigned int flags)
127 {
128 	struct folio *folio;
129 
130 	if (WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == 0))
131 		return NULL;
132 
133 	if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
134 		return NULL;
135 
136 	if (flags & FOLL_GET)
137 		return try_get_folio(page, refs);
138 
139 	/* FOLL_PIN is set */
140 
141 	/*
142 	 * Don't take a pin on the zero page - it's not going anywhere
143 	 * and it is used in a *lot* of places.
144 	 */
145 	if (is_zero_page(page))
146 		return page_folio(page);
147 
148 	folio = try_get_folio(page, refs);
149 	if (!folio)
150 		return NULL;
151 
152 	/*
153 	 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
154 	 * right zone, so fail and let the caller fall back to the slow
155 	 * path.
156 	 */
157 	if (unlikely((flags & FOLL_LONGTERM) &&
158 		     !folio_is_longterm_pinnable(folio))) {
159 		if (!put_devmap_managed_page_refs(&folio->page, refs))
160 			folio_put_refs(folio, refs);
161 		return NULL;
162 	}
163 
164 	/*
165 	 * When pinning a large folio, use an exact count to track it.
166 	 *
167 	 * However, be sure to *also* increment the normal folio
168 	 * refcount field at least once, so that the folio really
169 	 * is pinned.  That's why the refcount from the earlier
170 	 * try_get_folio() is left intact.
171 	 */
172 	if (folio_test_large(folio))
173 		atomic_add(refs, &folio->_pincount);
174 	else
175 		folio_ref_add(folio,
176 				refs * (GUP_PIN_COUNTING_BIAS - 1));
177 	/*
178 	 * Adjust the pincount before re-checking the PTE for changes.
179 	 * This is essentially a smp_mb() and is paired with a memory
180 	 * barrier in page_try_share_anon_rmap().
181 	 */
182 	smp_mb__after_atomic();
183 
184 	node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs);
185 
186 	return folio;
187 }
188 
gup_put_folio(struct folio * folio,int refs,unsigned int flags)189 static void gup_put_folio(struct folio *folio, int refs, unsigned int flags)
190 {
191 	if (flags & FOLL_PIN) {
192 		if (is_zero_folio(folio))
193 			return;
194 		node_stat_mod_folio(folio, NR_FOLL_PIN_RELEASED, refs);
195 		if (folio_test_large(folio))
196 			atomic_sub(refs, &folio->_pincount);
197 		else
198 			refs *= GUP_PIN_COUNTING_BIAS;
199 	}
200 
201 	if (!put_devmap_managed_page_refs(&folio->page, refs))
202 		folio_put_refs(folio, refs);
203 }
204 
205 /**
206  * try_grab_page() - elevate a page's refcount by a flag-dependent amount
207  * @page:    pointer to page to be grabbed
208  * @flags:   gup flags: these are the FOLL_* flag values.
209  *
210  * This might not do anything at all, depending on the flags argument.
211  *
212  * "grab" names in this file mean, "look at flags to decide whether to use
213  * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
214  *
215  * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
216  * time. Cases: please see the try_grab_folio() documentation, with
217  * "refs=1".
218  *
219  * Return: 0 for success, or if no action was required (if neither FOLL_PIN
220  * nor FOLL_GET was set, nothing is done). A negative error code for failure:
221  *
222  *   -ENOMEM		FOLL_GET or FOLL_PIN was set, but the page could not
223  *			be grabbed.
224  */
try_grab_page(struct page * page,unsigned int flags)225 int __must_check try_grab_page(struct page *page, unsigned int flags)
226 {
227 	struct folio *folio = page_folio(page);
228 
229 	if (WARN_ON_ONCE(folio_ref_count(folio) <= 0))
230 		return -ENOMEM;
231 
232 	if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
233 		return -EREMOTEIO;
234 
235 	if (flags & FOLL_GET)
236 		folio_ref_inc(folio);
237 	else if (flags & FOLL_PIN) {
238 		/*
239 		 * Don't take a pin on the zero page - it's not going anywhere
240 		 * and it is used in a *lot* of places.
241 		 */
242 		if (is_zero_page(page))
243 			return 0;
244 
245 		/*
246 		 * Similar to try_grab_folio(): be sure to *also*
247 		 * increment the normal page refcount field at least once,
248 		 * so that the page really is pinned.
249 		 */
250 		if (folio_test_large(folio)) {
251 			folio_ref_add(folio, 1);
252 			atomic_add(1, &folio->_pincount);
253 		} else {
254 			folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
255 		}
256 
257 		node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, 1);
258 	}
259 
260 	return 0;
261 }
262 
263 /**
264  * unpin_user_page() - release a dma-pinned page
265  * @page:            pointer to page to be released
266  *
267  * Pages that were pinned via pin_user_pages*() must be released via either
268  * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
269  * that such pages can be separately tracked and uniquely handled. In
270  * particular, interactions with RDMA and filesystems need special handling.
271  */
unpin_user_page(struct page * page)272 void unpin_user_page(struct page *page)
273 {
274 	sanity_check_pinned_pages(&page, 1);
275 	gup_put_folio(page_folio(page), 1, FOLL_PIN);
276 }
277 EXPORT_SYMBOL(unpin_user_page);
278 
279 /**
280  * folio_add_pin - Try to get an additional pin on a pinned folio
281  * @folio: The folio to be pinned
282  *
283  * Get an additional pin on a folio we already have a pin on.  Makes no change
284  * if the folio is a zero_page.
285  */
folio_add_pin(struct folio * folio)286 void folio_add_pin(struct folio *folio)
287 {
288 	if (is_zero_folio(folio))
289 		return;
290 
291 	/*
292 	 * Similar to try_grab_folio(): be sure to *also* increment the normal
293 	 * page refcount field at least once, so that the page really is
294 	 * pinned.
295 	 */
296 	if (folio_test_large(folio)) {
297 		WARN_ON_ONCE(atomic_read(&folio->_pincount) < 1);
298 		folio_ref_inc(folio);
299 		atomic_inc(&folio->_pincount);
300 	} else {
301 		WARN_ON_ONCE(folio_ref_count(folio) < GUP_PIN_COUNTING_BIAS);
302 		folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
303 	}
304 }
305 
gup_folio_range_next(struct page * start,unsigned long npages,unsigned long i,unsigned int * ntails)306 static inline struct folio *gup_folio_range_next(struct page *start,
307 		unsigned long npages, unsigned long i, unsigned int *ntails)
308 {
309 	struct page *next = nth_page(start, i);
310 	struct folio *folio = page_folio(next);
311 	unsigned int nr = 1;
312 
313 	if (folio_test_large(folio))
314 		nr = min_t(unsigned int, npages - i,
315 			   folio_nr_pages(folio) - folio_page_idx(folio, next));
316 
317 	*ntails = nr;
318 	return folio;
319 }
320 
gup_folio_next(struct page ** list,unsigned long npages,unsigned long i,unsigned int * ntails)321 static inline struct folio *gup_folio_next(struct page **list,
322 		unsigned long npages, unsigned long i, unsigned int *ntails)
323 {
324 	struct folio *folio = page_folio(list[i]);
325 	unsigned int nr;
326 
327 	for (nr = i + 1; nr < npages; nr++) {
328 		if (page_folio(list[nr]) != folio)
329 			break;
330 	}
331 
332 	*ntails = nr - i;
333 	return folio;
334 }
335 
336 /**
337  * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
338  * @pages:  array of pages to be maybe marked dirty, and definitely released.
339  * @npages: number of pages in the @pages array.
340  * @make_dirty: whether to mark the pages dirty
341  *
342  * "gup-pinned page" refers to a page that has had one of the get_user_pages()
343  * variants called on that page.
344  *
345  * For each page in the @pages array, make that page (or its head page, if a
346  * compound page) dirty, if @make_dirty is true, and if the page was previously
347  * listed as clean. In any case, releases all pages using unpin_user_page(),
348  * possibly via unpin_user_pages(), for the non-dirty case.
349  *
350  * Please see the unpin_user_page() documentation for details.
351  *
352  * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
353  * required, then the caller should a) verify that this is really correct,
354  * because _lock() is usually required, and b) hand code it:
355  * set_page_dirty_lock(), unpin_user_page().
356  *
357  */
unpin_user_pages_dirty_lock(struct page ** pages,unsigned long npages,bool make_dirty)358 void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
359 				 bool make_dirty)
360 {
361 	unsigned long i;
362 	struct folio *folio;
363 	unsigned int nr;
364 
365 	if (!make_dirty) {
366 		unpin_user_pages(pages, npages);
367 		return;
368 	}
369 
370 	sanity_check_pinned_pages(pages, npages);
371 	for (i = 0; i < npages; i += nr) {
372 		folio = gup_folio_next(pages, npages, i, &nr);
373 		/*
374 		 * Checking PageDirty at this point may race with
375 		 * clear_page_dirty_for_io(), but that's OK. Two key
376 		 * cases:
377 		 *
378 		 * 1) This code sees the page as already dirty, so it
379 		 * skips the call to set_page_dirty(). That could happen
380 		 * because clear_page_dirty_for_io() called
381 		 * page_mkclean(), followed by set_page_dirty().
382 		 * However, now the page is going to get written back,
383 		 * which meets the original intention of setting it
384 		 * dirty, so all is well: clear_page_dirty_for_io() goes
385 		 * on to call TestClearPageDirty(), and write the page
386 		 * back.
387 		 *
388 		 * 2) This code sees the page as clean, so it calls
389 		 * set_page_dirty(). The page stays dirty, despite being
390 		 * written back, so it gets written back again in the
391 		 * next writeback cycle. This is harmless.
392 		 */
393 		if (!folio_test_dirty(folio)) {
394 			folio_lock(folio);
395 			folio_mark_dirty(folio);
396 			folio_unlock(folio);
397 		}
398 		gup_put_folio(folio, nr, FOLL_PIN);
399 	}
400 }
401 EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
402 
403 /**
404  * unpin_user_page_range_dirty_lock() - release and optionally dirty
405  * gup-pinned page range
406  *
407  * @page:  the starting page of a range maybe marked dirty, and definitely released.
408  * @npages: number of consecutive pages to release.
409  * @make_dirty: whether to mark the pages dirty
410  *
411  * "gup-pinned page range" refers to a range of pages that has had one of the
412  * pin_user_pages() variants called on that page.
413  *
414  * For the page ranges defined by [page .. page+npages], make that range (or
415  * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
416  * page range was previously listed as clean.
417  *
418  * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
419  * required, then the caller should a) verify that this is really correct,
420  * because _lock() is usually required, and b) hand code it:
421  * set_page_dirty_lock(), unpin_user_page().
422  *
423  */
unpin_user_page_range_dirty_lock(struct page * page,unsigned long npages,bool make_dirty)424 void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages,
425 				      bool make_dirty)
426 {
427 	unsigned long i;
428 	struct folio *folio;
429 	unsigned int nr;
430 
431 	for (i = 0; i < npages; i += nr) {
432 		folio = gup_folio_range_next(page, npages, i, &nr);
433 		if (make_dirty && !folio_test_dirty(folio)) {
434 			folio_lock(folio);
435 			folio_mark_dirty(folio);
436 			folio_unlock(folio);
437 		}
438 		gup_put_folio(folio, nr, FOLL_PIN);
439 	}
440 }
441 EXPORT_SYMBOL(unpin_user_page_range_dirty_lock);
442 
unpin_user_pages_lockless(struct page ** pages,unsigned long npages)443 static void unpin_user_pages_lockless(struct page **pages, unsigned long npages)
444 {
445 	unsigned long i;
446 	struct folio *folio;
447 	unsigned int nr;
448 
449 	/*
450 	 * Don't perform any sanity checks because we might have raced with
451 	 * fork() and some anonymous pages might now actually be shared --
452 	 * which is why we're unpinning after all.
453 	 */
454 	for (i = 0; i < npages; i += nr) {
455 		folio = gup_folio_next(pages, npages, i, &nr);
456 		gup_put_folio(folio, nr, FOLL_PIN);
457 	}
458 }
459 
460 /**
461  * unpin_user_pages() - release an array of gup-pinned pages.
462  * @pages:  array of pages to be marked dirty and released.
463  * @npages: number of pages in the @pages array.
464  *
465  * For each page in the @pages array, release the page using unpin_user_page().
466  *
467  * Please see the unpin_user_page() documentation for details.
468  */
unpin_user_pages(struct page ** pages,unsigned long npages)469 void unpin_user_pages(struct page **pages, unsigned long npages)
470 {
471 	unsigned long i;
472 	struct folio *folio;
473 	unsigned int nr;
474 
475 	/*
476 	 * If this WARN_ON() fires, then the system *might* be leaking pages (by
477 	 * leaving them pinned), but probably not. More likely, gup/pup returned
478 	 * a hard -ERRNO error to the caller, who erroneously passed it here.
479 	 */
480 	if (WARN_ON(IS_ERR_VALUE(npages)))
481 		return;
482 
483 	sanity_check_pinned_pages(pages, npages);
484 	for (i = 0; i < npages; i += nr) {
485 		folio = gup_folio_next(pages, npages, i, &nr);
486 		gup_put_folio(folio, nr, FOLL_PIN);
487 	}
488 }
489 EXPORT_SYMBOL(unpin_user_pages);
490 
491 /*
492  * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's
493  * lifecycle.  Avoid setting the bit unless necessary, or it might cause write
494  * cache bouncing on large SMP machines for concurrent pinned gups.
495  */
mm_set_has_pinned_flag(unsigned long * mm_flags)496 static inline void mm_set_has_pinned_flag(unsigned long *mm_flags)
497 {
498 	if (!test_bit(MMF_HAS_PINNED, mm_flags))
499 		set_bit(MMF_HAS_PINNED, mm_flags);
500 }
501 
502 #ifdef CONFIG_MMU
no_page_table(struct vm_area_struct * vma,unsigned int flags)503 static struct page *no_page_table(struct vm_area_struct *vma,
504 		unsigned int flags)
505 {
506 	/*
507 	 * When core dumping an enormous anonymous area that nobody
508 	 * has touched so far, we don't want to allocate unnecessary pages or
509 	 * page tables.  Return error instead of NULL to skip handle_mm_fault,
510 	 * then get_dump_page() will return NULL to leave a hole in the dump.
511 	 * But we can only make this optimization where a hole would surely
512 	 * be zero-filled if handle_mm_fault() actually did handle it.
513 	 */
514 	if ((flags & FOLL_DUMP) &&
515 			(vma_is_anonymous(vma) || !vma->vm_ops->fault))
516 		return ERR_PTR(-EFAULT);
517 	return NULL;
518 }
519 
follow_pfn_pte(struct vm_area_struct * vma,unsigned long address,pte_t * pte,unsigned int flags)520 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
521 		pte_t *pte, unsigned int flags)
522 {
523 	if (flags & FOLL_TOUCH) {
524 		pte_t orig_entry = ptep_get(pte);
525 		pte_t entry = orig_entry;
526 
527 		if (flags & FOLL_WRITE)
528 			entry = pte_mkdirty(entry);
529 		entry = pte_mkyoung(entry);
530 
531 		if (!pte_same(orig_entry, entry)) {
532 			set_pte_at(vma->vm_mm, address, pte, entry);
533 			update_mmu_cache(vma, address, pte);
534 		}
535 	}
536 
537 	/* Proper page table entry exists, but no corresponding struct page */
538 	return -EEXIST;
539 }
540 
541 /* FOLL_FORCE can write to even unwritable PTEs in COW mappings. */
can_follow_write_pte(pte_t pte,struct page * page,struct vm_area_struct * vma,unsigned int flags)542 static inline bool can_follow_write_pte(pte_t pte, struct page *page,
543 					struct vm_area_struct *vma,
544 					unsigned int flags)
545 {
546 	/* If the pte is writable, we can write to the page. */
547 	if (pte_write(pte))
548 		return true;
549 
550 	/* Maybe FOLL_FORCE is set to override it? */
551 	if (!(flags & FOLL_FORCE))
552 		return false;
553 
554 	/* But FOLL_FORCE has no effect on shared mappings */
555 	if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
556 		return false;
557 
558 	/* ... or read-only private ones */
559 	if (!(vma->vm_flags & VM_MAYWRITE))
560 		return false;
561 
562 	/* ... or already writable ones that just need to take a write fault */
563 	if (vma->vm_flags & VM_WRITE)
564 		return false;
565 
566 	/*
567 	 * See can_change_pte_writable(): we broke COW and could map the page
568 	 * writable if we have an exclusive anonymous page ...
569 	 */
570 	if (!page || !PageAnon(page) || !PageAnonExclusive(page))
571 		return false;
572 
573 	/* ... and a write-fault isn't required for other reasons. */
574 	if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte))
575 		return false;
576 	return !userfaultfd_pte_wp(vma, pte);
577 }
578 
follow_page_pte(struct vm_area_struct * vma,unsigned long address,pmd_t * pmd,unsigned int flags,struct dev_pagemap ** pgmap)579 static struct page *follow_page_pte(struct vm_area_struct *vma,
580 		unsigned long address, pmd_t *pmd, unsigned int flags,
581 		struct dev_pagemap **pgmap)
582 {
583 	struct mm_struct *mm = vma->vm_mm;
584 	struct page *page;
585 	spinlock_t *ptl;
586 	pte_t *ptep, pte;
587 	int ret;
588 
589 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
590 	if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
591 			 (FOLL_PIN | FOLL_GET)))
592 		return ERR_PTR(-EINVAL);
593 
594 	ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
595 	if (!ptep)
596 		return no_page_table(vma, flags);
597 	pte = ptep_get(ptep);
598 	if (!pte_present(pte))
599 		goto no_page;
600 	if (pte_protnone(pte) && !gup_can_follow_protnone(vma, flags))
601 		goto no_page;
602 
603 	page = vm_normal_page(vma, address, pte);
604 
605 	/*
606 	 * We only care about anon pages in can_follow_write_pte() and don't
607 	 * have to worry about pte_devmap() because they are never anon.
608 	 */
609 	if ((flags & FOLL_WRITE) &&
610 	    !can_follow_write_pte(pte, page, vma, flags)) {
611 		page = NULL;
612 		goto out;
613 	}
614 
615 	if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
616 		/*
617 		 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
618 		 * case since they are only valid while holding the pgmap
619 		 * reference.
620 		 */
621 		*pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
622 		if (*pgmap)
623 			page = pte_page(pte);
624 		else
625 			goto no_page;
626 	} else if (unlikely(!page)) {
627 		if (flags & FOLL_DUMP) {
628 			/* Avoid special (like zero) pages in core dumps */
629 			page = ERR_PTR(-EFAULT);
630 			goto out;
631 		}
632 
633 		if (is_zero_pfn(pte_pfn(pte))) {
634 			page = pte_page(pte);
635 		} else {
636 			ret = follow_pfn_pte(vma, address, ptep, flags);
637 			page = ERR_PTR(ret);
638 			goto out;
639 		}
640 	}
641 
642 	if (!pte_write(pte) && gup_must_unshare(vma, flags, page)) {
643 		page = ERR_PTR(-EMLINK);
644 		goto out;
645 	}
646 
647 	VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
648 		       !PageAnonExclusive(page), page);
649 
650 	/* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
651 	ret = try_grab_page(page, flags);
652 	if (unlikely(ret)) {
653 		page = ERR_PTR(ret);
654 		goto out;
655 	}
656 
657 	/*
658 	 * We need to make the page accessible if and only if we are going
659 	 * to access its content (the FOLL_PIN case).  Please see
660 	 * Documentation/core-api/pin_user_pages.rst for details.
661 	 */
662 	if (flags & FOLL_PIN) {
663 		ret = arch_make_page_accessible(page);
664 		if (ret) {
665 			unpin_user_page(page);
666 			page = ERR_PTR(ret);
667 			goto out;
668 		}
669 	}
670 	if (flags & FOLL_TOUCH) {
671 		if ((flags & FOLL_WRITE) &&
672 		    !pte_dirty(pte) && !PageDirty(page))
673 			set_page_dirty(page);
674 		/*
675 		 * pte_mkyoung() would be more correct here, but atomic care
676 		 * is needed to avoid losing the dirty bit: it is easier to use
677 		 * mark_page_accessed().
678 		 */
679 		mark_page_accessed(page);
680 	}
681 out:
682 	pte_unmap_unlock(ptep, ptl);
683 	return page;
684 no_page:
685 	pte_unmap_unlock(ptep, ptl);
686 	if (!pte_none(pte))
687 		return NULL;
688 	return no_page_table(vma, flags);
689 }
690 
follow_pmd_mask(struct vm_area_struct * vma,unsigned long address,pud_t * pudp,unsigned int flags,struct follow_page_context * ctx)691 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
692 				    unsigned long address, pud_t *pudp,
693 				    unsigned int flags,
694 				    struct follow_page_context *ctx)
695 {
696 	pmd_t *pmd, pmdval;
697 	spinlock_t *ptl;
698 	struct page *page;
699 	struct mm_struct *mm = vma->vm_mm;
700 
701 	pmd = pmd_offset(pudp, address);
702 	pmdval = pmdp_get_lockless(pmd);
703 	if (pmd_none(pmdval))
704 		return no_page_table(vma, flags);
705 	if (!pmd_present(pmdval))
706 		return no_page_table(vma, flags);
707 	if (pmd_devmap(pmdval)) {
708 		ptl = pmd_lock(mm, pmd);
709 		page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
710 		spin_unlock(ptl);
711 		if (page)
712 			return page;
713 	}
714 	if (likely(!pmd_trans_huge(pmdval)))
715 		return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
716 
717 	if (pmd_protnone(pmdval) && !gup_can_follow_protnone(vma, flags))
718 		return no_page_table(vma, flags);
719 
720 	ptl = pmd_lock(mm, pmd);
721 	if (unlikely(!pmd_present(*pmd))) {
722 		spin_unlock(ptl);
723 		return no_page_table(vma, flags);
724 	}
725 	if (unlikely(!pmd_trans_huge(*pmd))) {
726 		spin_unlock(ptl);
727 		return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
728 	}
729 	if (flags & FOLL_SPLIT_PMD) {
730 		spin_unlock(ptl);
731 		split_huge_pmd(vma, pmd, address);
732 		/* If pmd was left empty, stuff a page table in there quickly */
733 		return pte_alloc(mm, pmd) ? ERR_PTR(-ENOMEM) :
734 			follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
735 	}
736 	page = follow_trans_huge_pmd(vma, address, pmd, flags);
737 	spin_unlock(ptl);
738 	ctx->page_mask = HPAGE_PMD_NR - 1;
739 	return page;
740 }
741 
follow_pud_mask(struct vm_area_struct * vma,unsigned long address,p4d_t * p4dp,unsigned int flags,struct follow_page_context * ctx)742 static struct page *follow_pud_mask(struct vm_area_struct *vma,
743 				    unsigned long address, p4d_t *p4dp,
744 				    unsigned int flags,
745 				    struct follow_page_context *ctx)
746 {
747 	pud_t *pud;
748 	spinlock_t *ptl;
749 	struct page *page;
750 	struct mm_struct *mm = vma->vm_mm;
751 
752 	pud = pud_offset(p4dp, address);
753 	if (pud_none(*pud))
754 		return no_page_table(vma, flags);
755 	if (pud_devmap(*pud)) {
756 		ptl = pud_lock(mm, pud);
757 		page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
758 		spin_unlock(ptl);
759 		if (page)
760 			return page;
761 	}
762 	if (unlikely(pud_bad(*pud)))
763 		return no_page_table(vma, flags);
764 
765 	return follow_pmd_mask(vma, address, pud, flags, ctx);
766 }
767 
follow_p4d_mask(struct vm_area_struct * vma,unsigned long address,pgd_t * pgdp,unsigned int flags,struct follow_page_context * ctx)768 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
769 				    unsigned long address, pgd_t *pgdp,
770 				    unsigned int flags,
771 				    struct follow_page_context *ctx)
772 {
773 	p4d_t *p4d;
774 
775 	p4d = p4d_offset(pgdp, address);
776 	if (p4d_none(*p4d))
777 		return no_page_table(vma, flags);
778 	BUILD_BUG_ON(p4d_huge(*p4d));
779 	if (unlikely(p4d_bad(*p4d)))
780 		return no_page_table(vma, flags);
781 
782 	return follow_pud_mask(vma, address, p4d, flags, ctx);
783 }
784 
785 /**
786  * follow_page_mask - look up a page descriptor from a user-virtual address
787  * @vma: vm_area_struct mapping @address
788  * @address: virtual address to look up
789  * @flags: flags modifying lookup behaviour
790  * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
791  *       pointer to output page_mask
792  *
793  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
794  *
795  * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
796  * the device's dev_pagemap metadata to avoid repeating expensive lookups.
797  *
798  * When getting an anonymous page and the caller has to trigger unsharing
799  * of a shared anonymous page first, -EMLINK is returned. The caller should
800  * trigger a fault with FAULT_FLAG_UNSHARE set. Note that unsharing is only
801  * relevant with FOLL_PIN and !FOLL_WRITE.
802  *
803  * On output, the @ctx->page_mask is set according to the size of the page.
804  *
805  * Return: the mapped (struct page *), %NULL if no mapping exists, or
806  * an error pointer if there is a mapping to something not represented
807  * by a page descriptor (see also vm_normal_page()).
808  */
follow_page_mask(struct vm_area_struct * vma,unsigned long address,unsigned int flags,struct follow_page_context * ctx)809 static struct page *follow_page_mask(struct vm_area_struct *vma,
810 			      unsigned long address, unsigned int flags,
811 			      struct follow_page_context *ctx)
812 {
813 	pgd_t *pgd;
814 	struct mm_struct *mm = vma->vm_mm;
815 
816 	ctx->page_mask = 0;
817 
818 	/*
819 	 * Call hugetlb_follow_page_mask for hugetlb vmas as it will use
820 	 * special hugetlb page table walking code.  This eliminates the
821 	 * need to check for hugetlb entries in the general walking code.
822 	 */
823 	if (is_vm_hugetlb_page(vma))
824 		return hugetlb_follow_page_mask(vma, address, flags,
825 						&ctx->page_mask);
826 
827 	pgd = pgd_offset(mm, address);
828 
829 	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
830 		return no_page_table(vma, flags);
831 
832 	return follow_p4d_mask(vma, address, pgd, flags, ctx);
833 }
834 
follow_page(struct vm_area_struct * vma,unsigned long address,unsigned int foll_flags)835 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
836 			 unsigned int foll_flags)
837 {
838 	struct follow_page_context ctx = { NULL };
839 	struct page *page;
840 
841 	if (vma_is_secretmem(vma))
842 		return NULL;
843 
844 	if (WARN_ON_ONCE(foll_flags & FOLL_PIN))
845 		return NULL;
846 
847 	/*
848 	 * We never set FOLL_HONOR_NUMA_FAULT because callers don't expect
849 	 * to fail on PROT_NONE-mapped pages.
850 	 */
851 	page = follow_page_mask(vma, address, foll_flags, &ctx);
852 	if (ctx.pgmap)
853 		put_dev_pagemap(ctx.pgmap);
854 	return page;
855 }
856 
get_gate_page(struct mm_struct * mm,unsigned long address,unsigned int gup_flags,struct vm_area_struct ** vma,struct page ** page)857 static int get_gate_page(struct mm_struct *mm, unsigned long address,
858 		unsigned int gup_flags, struct vm_area_struct **vma,
859 		struct page **page)
860 {
861 	pgd_t *pgd;
862 	p4d_t *p4d;
863 	pud_t *pud;
864 	pmd_t *pmd;
865 	pte_t *pte;
866 	pte_t entry;
867 	int ret = -EFAULT;
868 
869 	/* user gate pages are read-only */
870 	if (gup_flags & FOLL_WRITE)
871 		return -EFAULT;
872 	if (address > TASK_SIZE)
873 		pgd = pgd_offset_k(address);
874 	else
875 		pgd = pgd_offset_gate(mm, address);
876 	if (pgd_none(*pgd))
877 		return -EFAULT;
878 	p4d = p4d_offset(pgd, address);
879 	if (p4d_none(*p4d))
880 		return -EFAULT;
881 	pud = pud_offset(p4d, address);
882 	if (pud_none(*pud))
883 		return -EFAULT;
884 	pmd = pmd_offset(pud, address);
885 	if (!pmd_present(*pmd))
886 		return -EFAULT;
887 	pte = pte_offset_map(pmd, address);
888 	if (!pte)
889 		return -EFAULT;
890 	entry = ptep_get(pte);
891 	if (pte_none(entry))
892 		goto unmap;
893 	*vma = get_gate_vma(mm);
894 	if (!page)
895 		goto out;
896 	*page = vm_normal_page(*vma, address, entry);
897 	if (!*page) {
898 		if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(entry)))
899 			goto unmap;
900 		*page = pte_page(entry);
901 	}
902 	ret = try_grab_page(*page, gup_flags);
903 	if (unlikely(ret))
904 		goto unmap;
905 out:
906 	ret = 0;
907 unmap:
908 	pte_unmap(pte);
909 	return ret;
910 }
911 
912 /*
913  * mmap_lock must be held on entry.  If @flags has FOLL_UNLOCKABLE but not
914  * FOLL_NOWAIT, the mmap_lock may be released.  If it is, *@locked will be set
915  * to 0 and -EBUSY returned.
916  */
faultin_page(struct vm_area_struct * vma,unsigned long address,unsigned int * flags,bool unshare,int * locked)917 static int faultin_page(struct vm_area_struct *vma,
918 		unsigned long address, unsigned int *flags, bool unshare,
919 		int *locked)
920 {
921 	unsigned int fault_flags = 0;
922 	vm_fault_t ret;
923 
924 	if (*flags & FOLL_NOFAULT)
925 		return -EFAULT;
926 	if (*flags & FOLL_WRITE)
927 		fault_flags |= FAULT_FLAG_WRITE;
928 	if (*flags & FOLL_REMOTE)
929 		fault_flags |= FAULT_FLAG_REMOTE;
930 	if (*flags & FOLL_UNLOCKABLE) {
931 		fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
932 		/*
933 		 * FAULT_FLAG_INTERRUPTIBLE is opt-in. GUP callers must set
934 		 * FOLL_INTERRUPTIBLE to enable FAULT_FLAG_INTERRUPTIBLE.
935 		 * That's because some callers may not be prepared to
936 		 * handle early exits caused by non-fatal signals.
937 		 */
938 		if (*flags & FOLL_INTERRUPTIBLE)
939 			fault_flags |= FAULT_FLAG_INTERRUPTIBLE;
940 	}
941 	if (*flags & FOLL_NOWAIT)
942 		fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
943 	if (*flags & FOLL_TRIED) {
944 		/*
945 		 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
946 		 * can co-exist
947 		 */
948 		fault_flags |= FAULT_FLAG_TRIED;
949 	}
950 	if (unshare) {
951 		fault_flags |= FAULT_FLAG_UNSHARE;
952 		/* FAULT_FLAG_WRITE and FAULT_FLAG_UNSHARE are incompatible */
953 		VM_BUG_ON(fault_flags & FAULT_FLAG_WRITE);
954 	}
955 
956 	ret = handle_mm_fault(vma, address, fault_flags, NULL);
957 
958 	if (ret & VM_FAULT_COMPLETED) {
959 		/*
960 		 * With FAULT_FLAG_RETRY_NOWAIT we'll never release the
961 		 * mmap lock in the page fault handler. Sanity check this.
962 		 */
963 		WARN_ON_ONCE(fault_flags & FAULT_FLAG_RETRY_NOWAIT);
964 		*locked = 0;
965 
966 		/*
967 		 * We should do the same as VM_FAULT_RETRY, but let's not
968 		 * return -EBUSY since that's not reflecting the reality of
969 		 * what has happened - we've just fully completed a page
970 		 * fault, with the mmap lock released.  Use -EAGAIN to show
971 		 * that we want to take the mmap lock _again_.
972 		 */
973 		return -EAGAIN;
974 	}
975 
976 	if (ret & VM_FAULT_ERROR) {
977 		int err = vm_fault_to_errno(ret, *flags);
978 
979 		if (err)
980 			return err;
981 		BUG();
982 	}
983 
984 	if (ret & VM_FAULT_RETRY) {
985 		if (!(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
986 			*locked = 0;
987 		return -EBUSY;
988 	}
989 
990 	return 0;
991 }
992 
993 /*
994  * Writing to file-backed mappings which require folio dirty tracking using GUP
995  * is a fundamentally broken operation, as kernel write access to GUP mappings
996  * do not adhere to the semantics expected by a file system.
997  *
998  * Consider the following scenario:-
999  *
1000  * 1. A folio is written to via GUP which write-faults the memory, notifying
1001  *    the file system and dirtying the folio.
1002  * 2. Later, writeback is triggered, resulting in the folio being cleaned and
1003  *    the PTE being marked read-only.
1004  * 3. The GUP caller writes to the folio, as it is mapped read/write via the
1005  *    direct mapping.
1006  * 4. The GUP caller, now done with the page, unpins it and sets it dirty
1007  *    (though it does not have to).
1008  *
1009  * This results in both data being written to a folio without writenotify, and
1010  * the folio being dirtied unexpectedly (if the caller decides to do so).
1011  */
writable_file_mapping_allowed(struct vm_area_struct * vma,unsigned long gup_flags)1012 static bool writable_file_mapping_allowed(struct vm_area_struct *vma,
1013 					  unsigned long gup_flags)
1014 {
1015 	/*
1016 	 * If we aren't pinning then no problematic write can occur. A long term
1017 	 * pin is the most egregious case so this is the case we disallow.
1018 	 */
1019 	if ((gup_flags & (FOLL_PIN | FOLL_LONGTERM)) !=
1020 	    (FOLL_PIN | FOLL_LONGTERM))
1021 		return true;
1022 
1023 	/*
1024 	 * If the VMA does not require dirty tracking then no problematic write
1025 	 * can occur either.
1026 	 */
1027 	return !vma_needs_dirty_tracking(vma);
1028 }
1029 
check_vma_flags(struct vm_area_struct * vma,unsigned long gup_flags)1030 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
1031 {
1032 	vm_flags_t vm_flags = vma->vm_flags;
1033 	int write = (gup_flags & FOLL_WRITE);
1034 	int foreign = (gup_flags & FOLL_REMOTE);
1035 	bool vma_anon = vma_is_anonymous(vma);
1036 
1037 	if (vm_flags & (VM_IO | VM_PFNMAP))
1038 		return -EFAULT;
1039 
1040 	if ((gup_flags & FOLL_ANON) && !vma_anon)
1041 		return -EFAULT;
1042 
1043 	if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma))
1044 		return -EOPNOTSUPP;
1045 
1046 	if (vma_is_secretmem(vma))
1047 		return -EFAULT;
1048 
1049 	if (write) {
1050 		if (!vma_anon &&
1051 		    !writable_file_mapping_allowed(vma, gup_flags))
1052 			return -EFAULT;
1053 
1054 		if (!(vm_flags & VM_WRITE) || (vm_flags & VM_SHADOW_STACK)) {
1055 			if (!(gup_flags & FOLL_FORCE))
1056 				return -EFAULT;
1057 			/* hugetlb does not support FOLL_FORCE|FOLL_WRITE. */
1058 			if (is_vm_hugetlb_page(vma))
1059 				return -EFAULT;
1060 			/*
1061 			 * We used to let the write,force case do COW in a
1062 			 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
1063 			 * set a breakpoint in a read-only mapping of an
1064 			 * executable, without corrupting the file (yet only
1065 			 * when that file had been opened for writing!).
1066 			 * Anon pages in shared mappings are surprising: now
1067 			 * just reject it.
1068 			 */
1069 			if (!is_cow_mapping(vm_flags))
1070 				return -EFAULT;
1071 		}
1072 	} else if (!(vm_flags & VM_READ)) {
1073 		if (!(gup_flags & FOLL_FORCE))
1074 			return -EFAULT;
1075 		/*
1076 		 * Is there actually any vma we can reach here which does not
1077 		 * have VM_MAYREAD set?
1078 		 */
1079 		if (!(vm_flags & VM_MAYREAD))
1080 			return -EFAULT;
1081 	}
1082 	/*
1083 	 * gups are always data accesses, not instruction
1084 	 * fetches, so execute=false here
1085 	 */
1086 	if (!arch_vma_access_permitted(vma, write, false, foreign))
1087 		return -EFAULT;
1088 	return 0;
1089 }
1090 
1091 /*
1092  * This is "vma_lookup()", but with a warning if we would have
1093  * historically expanded the stack in the GUP code.
1094  */
gup_vma_lookup(struct mm_struct * mm,unsigned long addr)1095 static struct vm_area_struct *gup_vma_lookup(struct mm_struct *mm,
1096 	 unsigned long addr)
1097 {
1098 #ifdef CONFIG_STACK_GROWSUP
1099 	return vma_lookup(mm, addr);
1100 #else
1101 	static volatile unsigned long next_warn;
1102 	struct vm_area_struct *vma;
1103 	unsigned long now, next;
1104 
1105 	vma = find_vma(mm, addr);
1106 	if (!vma || (addr >= vma->vm_start))
1107 		return vma;
1108 
1109 	/* Only warn for half-way relevant accesses */
1110 	if (!(vma->vm_flags & VM_GROWSDOWN))
1111 		return NULL;
1112 	if (vma->vm_start - addr > 65536)
1113 		return NULL;
1114 
1115 	/* Let's not warn more than once an hour.. */
1116 	now = jiffies; next = next_warn;
1117 	if (next && time_before(now, next))
1118 		return NULL;
1119 	next_warn = now + 60*60*HZ;
1120 
1121 	/* Let people know things may have changed. */
1122 	pr_warn("GUP no longer grows the stack in %s (%d): %lx-%lx (%lx)\n",
1123 		current->comm, task_pid_nr(current),
1124 		vma->vm_start, vma->vm_end, addr);
1125 	dump_stack();
1126 	return NULL;
1127 #endif
1128 }
1129 
1130 /**
1131  * __get_user_pages() - pin user pages in memory
1132  * @mm:		mm_struct of target mm
1133  * @start:	starting user address
1134  * @nr_pages:	number of pages from start to pin
1135  * @gup_flags:	flags modifying pin behaviour
1136  * @pages:	array that receives pointers to the pages pinned.
1137  *		Should be at least nr_pages long. Or NULL, if caller
1138  *		only intends to ensure the pages are faulted in.
1139  * @locked:     whether we're still with the mmap_lock held
1140  *
1141  * Returns either number of pages pinned (which may be less than the
1142  * number requested), or an error. Details about the return value:
1143  *
1144  * -- If nr_pages is 0, returns 0.
1145  * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1146  * -- If nr_pages is >0, and some pages were pinned, returns the number of
1147  *    pages pinned. Again, this may be less than nr_pages.
1148  * -- 0 return value is possible when the fault would need to be retried.
1149  *
1150  * The caller is responsible for releasing returned @pages, via put_page().
1151  *
1152  * Must be called with mmap_lock held.  It may be released.  See below.
1153  *
1154  * __get_user_pages walks a process's page tables and takes a reference to
1155  * each struct page that each user address corresponds to at a given
1156  * instant. That is, it takes the page that would be accessed if a user
1157  * thread accesses the given user virtual address at that instant.
1158  *
1159  * This does not guarantee that the page exists in the user mappings when
1160  * __get_user_pages returns, and there may even be a completely different
1161  * page there in some cases (eg. if mmapped pagecache has been invalidated
1162  * and subsequently re-faulted). However it does guarantee that the page
1163  * won't be freed completely. And mostly callers simply care that the page
1164  * contains data that was valid *at some point in time*. Typically, an IO
1165  * or similar operation cannot guarantee anything stronger anyway because
1166  * locks can't be held over the syscall boundary.
1167  *
1168  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1169  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1170  * appropriate) must be called after the page is finished with, and
1171  * before put_page is called.
1172  *
1173  * If FOLL_UNLOCKABLE is set without FOLL_NOWAIT then the mmap_lock may
1174  * be released. If this happens *@locked will be set to 0 on return.
1175  *
1176  * A caller using such a combination of @gup_flags must therefore hold the
1177  * mmap_lock for reading only, and recognize when it's been released. Otherwise,
1178  * it must be held for either reading or writing and will not be released.
1179  *
1180  * In most cases, get_user_pages or get_user_pages_fast should be used
1181  * instead of __get_user_pages. __get_user_pages should be used only if
1182  * you need some special @gup_flags.
1183  */
__get_user_pages(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)1184 static long __get_user_pages(struct mm_struct *mm,
1185 		unsigned long start, unsigned long nr_pages,
1186 		unsigned int gup_flags, struct page **pages,
1187 		int *locked)
1188 {
1189 	long ret = 0, i = 0;
1190 	struct vm_area_struct *vma = NULL;
1191 	struct follow_page_context ctx = { NULL };
1192 
1193 	if (!nr_pages)
1194 		return 0;
1195 
1196 	start = untagged_addr_remote(mm, start);
1197 
1198 	VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1199 
1200 	do {
1201 		struct page *page;
1202 		unsigned int foll_flags = gup_flags;
1203 		unsigned int page_increm;
1204 
1205 		/* first iteration or cross vma bound */
1206 		if (!vma || start >= vma->vm_end) {
1207 			vma = gup_vma_lookup(mm, start);
1208 			if (!vma && in_gate_area(mm, start)) {
1209 				ret = get_gate_page(mm, start & PAGE_MASK,
1210 						gup_flags, &vma,
1211 						pages ? &page : NULL);
1212 				if (ret)
1213 					goto out;
1214 				ctx.page_mask = 0;
1215 				goto next_page;
1216 			}
1217 
1218 			if (!vma) {
1219 				ret = -EFAULT;
1220 				goto out;
1221 			}
1222 			ret = check_vma_flags(vma, gup_flags);
1223 			if (ret)
1224 				goto out;
1225 		}
1226 retry:
1227 		/*
1228 		 * If we have a pending SIGKILL, don't keep faulting pages and
1229 		 * potentially allocating memory.
1230 		 */
1231 		if (fatal_signal_pending(current)) {
1232 			ret = -EINTR;
1233 			goto out;
1234 		}
1235 		cond_resched();
1236 
1237 		page = follow_page_mask(vma, start, foll_flags, &ctx);
1238 		if (!page || PTR_ERR(page) == -EMLINK) {
1239 			ret = faultin_page(vma, start, &foll_flags,
1240 					   PTR_ERR(page) == -EMLINK, locked);
1241 			switch (ret) {
1242 			case 0:
1243 				goto retry;
1244 			case -EBUSY:
1245 			case -EAGAIN:
1246 				ret = 0;
1247 				fallthrough;
1248 			case -EFAULT:
1249 			case -ENOMEM:
1250 			case -EHWPOISON:
1251 				goto out;
1252 			}
1253 			BUG();
1254 		} else if (PTR_ERR(page) == -EEXIST) {
1255 			/*
1256 			 * Proper page table entry exists, but no corresponding
1257 			 * struct page. If the caller expects **pages to be
1258 			 * filled in, bail out now, because that can't be done
1259 			 * for this page.
1260 			 */
1261 			if (pages) {
1262 				ret = PTR_ERR(page);
1263 				goto out;
1264 			}
1265 		} else if (IS_ERR(page)) {
1266 			ret = PTR_ERR(page);
1267 			goto out;
1268 		}
1269 next_page:
1270 		page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1271 		if (page_increm > nr_pages)
1272 			page_increm = nr_pages;
1273 
1274 		if (pages) {
1275 			struct page *subpage;
1276 			unsigned int j;
1277 
1278 			/*
1279 			 * This must be a large folio (and doesn't need to
1280 			 * be the whole folio; it can be part of it), do
1281 			 * the refcount work for all the subpages too.
1282 			 *
1283 			 * NOTE: here the page may not be the head page
1284 			 * e.g. when start addr is not thp-size aligned.
1285 			 * try_grab_folio() should have taken care of tail
1286 			 * pages.
1287 			 */
1288 			if (page_increm > 1) {
1289 				struct folio *folio;
1290 
1291 				/*
1292 				 * Since we already hold refcount on the
1293 				 * large folio, this should never fail.
1294 				 */
1295 				folio = try_grab_folio(page, page_increm - 1,
1296 						       foll_flags);
1297 				if (WARN_ON_ONCE(!folio)) {
1298 					/*
1299 					 * Release the 1st page ref if the
1300 					 * folio is problematic, fail hard.
1301 					 */
1302 					gup_put_folio(page_folio(page), 1,
1303 						      foll_flags);
1304 					ret = -EFAULT;
1305 					goto out;
1306 				}
1307 			}
1308 
1309 			for (j = 0; j < page_increm; j++) {
1310 				subpage = nth_page(page, j);
1311 				pages[i + j] = subpage;
1312 				flush_anon_page(vma, subpage, start + j * PAGE_SIZE);
1313 				flush_dcache_page(subpage);
1314 			}
1315 		}
1316 
1317 		i += page_increm;
1318 		start += page_increm * PAGE_SIZE;
1319 		nr_pages -= page_increm;
1320 	} while (nr_pages);
1321 out:
1322 	if (ctx.pgmap)
1323 		put_dev_pagemap(ctx.pgmap);
1324 	return i ? i : ret;
1325 }
1326 
vma_permits_fault(struct vm_area_struct * vma,unsigned int fault_flags)1327 static bool vma_permits_fault(struct vm_area_struct *vma,
1328 			      unsigned int fault_flags)
1329 {
1330 	bool write   = !!(fault_flags & FAULT_FLAG_WRITE);
1331 	bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1332 	vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1333 
1334 	if (!(vm_flags & vma->vm_flags))
1335 		return false;
1336 
1337 	/*
1338 	 * The architecture might have a hardware protection
1339 	 * mechanism other than read/write that can deny access.
1340 	 *
1341 	 * gup always represents data access, not instruction
1342 	 * fetches, so execute=false here:
1343 	 */
1344 	if (!arch_vma_access_permitted(vma, write, false, foreign))
1345 		return false;
1346 
1347 	return true;
1348 }
1349 
1350 /**
1351  * fixup_user_fault() - manually resolve a user page fault
1352  * @mm:		mm_struct of target mm
1353  * @address:	user address
1354  * @fault_flags:flags to pass down to handle_mm_fault()
1355  * @unlocked:	did we unlock the mmap_lock while retrying, maybe NULL if caller
1356  *		does not allow retry. If NULL, the caller must guarantee
1357  *		that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1358  *
1359  * This is meant to be called in the specific scenario where for locking reasons
1360  * we try to access user memory in atomic context (within a pagefault_disable()
1361  * section), this returns -EFAULT, and we want to resolve the user fault before
1362  * trying again.
1363  *
1364  * Typically this is meant to be used by the futex code.
1365  *
1366  * The main difference with get_user_pages() is that this function will
1367  * unconditionally call handle_mm_fault() which will in turn perform all the
1368  * necessary SW fixup of the dirty and young bits in the PTE, while
1369  * get_user_pages() only guarantees to update these in the struct page.
1370  *
1371  * This is important for some architectures where those bits also gate the
1372  * access permission to the page because they are maintained in software.  On
1373  * such architectures, gup() will not be enough to make a subsequent access
1374  * succeed.
1375  *
1376  * This function will not return with an unlocked mmap_lock. So it has not the
1377  * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1378  */
fixup_user_fault(struct mm_struct * mm,unsigned long address,unsigned int fault_flags,bool * unlocked)1379 int fixup_user_fault(struct mm_struct *mm,
1380 		     unsigned long address, unsigned int fault_flags,
1381 		     bool *unlocked)
1382 {
1383 	struct vm_area_struct *vma;
1384 	vm_fault_t ret;
1385 
1386 	address = untagged_addr_remote(mm, address);
1387 
1388 	if (unlocked)
1389 		fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1390 
1391 retry:
1392 	vma = gup_vma_lookup(mm, address);
1393 	if (!vma)
1394 		return -EFAULT;
1395 
1396 	if (!vma_permits_fault(vma, fault_flags))
1397 		return -EFAULT;
1398 
1399 	if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1400 	    fatal_signal_pending(current))
1401 		return -EINTR;
1402 
1403 	ret = handle_mm_fault(vma, address, fault_flags, NULL);
1404 
1405 	if (ret & VM_FAULT_COMPLETED) {
1406 		/*
1407 		 * NOTE: it's a pity that we need to retake the lock here
1408 		 * to pair with the unlock() in the callers. Ideally we
1409 		 * could tell the callers so they do not need to unlock.
1410 		 */
1411 		mmap_read_lock(mm);
1412 		*unlocked = true;
1413 		return 0;
1414 	}
1415 
1416 	if (ret & VM_FAULT_ERROR) {
1417 		int err = vm_fault_to_errno(ret, 0);
1418 
1419 		if (err)
1420 			return err;
1421 		BUG();
1422 	}
1423 
1424 	if (ret & VM_FAULT_RETRY) {
1425 		mmap_read_lock(mm);
1426 		*unlocked = true;
1427 		fault_flags |= FAULT_FLAG_TRIED;
1428 		goto retry;
1429 	}
1430 
1431 	return 0;
1432 }
1433 EXPORT_SYMBOL_GPL(fixup_user_fault);
1434 
1435 /*
1436  * GUP always responds to fatal signals.  When FOLL_INTERRUPTIBLE is
1437  * specified, it'll also respond to generic signals.  The caller of GUP
1438  * that has FOLL_INTERRUPTIBLE should take care of the GUP interruption.
1439  */
gup_signal_pending(unsigned int flags)1440 static bool gup_signal_pending(unsigned int flags)
1441 {
1442 	if (fatal_signal_pending(current))
1443 		return true;
1444 
1445 	if (!(flags & FOLL_INTERRUPTIBLE))
1446 		return false;
1447 
1448 	return signal_pending(current);
1449 }
1450 
1451 /*
1452  * Locking: (*locked == 1) means that the mmap_lock has already been acquired by
1453  * the caller. This function may drop the mmap_lock. If it does so, then it will
1454  * set (*locked = 0).
1455  *
1456  * (*locked == 0) means that the caller expects this function to acquire and
1457  * drop the mmap_lock. Therefore, the value of *locked will still be zero when
1458  * the function returns, even though it may have changed temporarily during
1459  * function execution.
1460  *
1461  * Please note that this function, unlike __get_user_pages(), will not return 0
1462  * for nr_pages > 0, unless FOLL_NOWAIT is used.
1463  */
__get_user_pages_locked(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,int * locked,unsigned int flags)1464 static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1465 						unsigned long start,
1466 						unsigned long nr_pages,
1467 						struct page **pages,
1468 						int *locked,
1469 						unsigned int flags)
1470 {
1471 	long ret, pages_done;
1472 	bool must_unlock = false;
1473 
1474 	/*
1475 	 * The internal caller expects GUP to manage the lock internally and the
1476 	 * lock must be released when this returns.
1477 	 */
1478 	if (!*locked) {
1479 		if (mmap_read_lock_killable(mm))
1480 			return -EAGAIN;
1481 		must_unlock = true;
1482 		*locked = 1;
1483 	}
1484 	else
1485 		mmap_assert_locked(mm);
1486 
1487 	if (flags & FOLL_PIN)
1488 		mm_set_has_pinned_flag(&mm->flags);
1489 
1490 	/*
1491 	 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1492 	 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1493 	 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1494 	 * for FOLL_GET, not for the newer FOLL_PIN.
1495 	 *
1496 	 * FOLL_PIN always expects pages to be non-null, but no need to assert
1497 	 * that here, as any failures will be obvious enough.
1498 	 */
1499 	if (pages && !(flags & FOLL_PIN))
1500 		flags |= FOLL_GET;
1501 
1502 	pages_done = 0;
1503 	for (;;) {
1504 		ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1505 				       locked);
1506 		if (!(flags & FOLL_UNLOCKABLE)) {
1507 			/* VM_FAULT_RETRY couldn't trigger, bypass */
1508 			pages_done = ret;
1509 			break;
1510 		}
1511 
1512 		/* VM_FAULT_RETRY or VM_FAULT_COMPLETED cannot return errors */
1513 		if (!*locked) {
1514 			BUG_ON(ret < 0);
1515 			BUG_ON(ret >= nr_pages);
1516 		}
1517 
1518 		if (ret > 0) {
1519 			nr_pages -= ret;
1520 			pages_done += ret;
1521 			if (!nr_pages)
1522 				break;
1523 		}
1524 		if (*locked) {
1525 			/*
1526 			 * VM_FAULT_RETRY didn't trigger or it was a
1527 			 * FOLL_NOWAIT.
1528 			 */
1529 			if (!pages_done)
1530 				pages_done = ret;
1531 			break;
1532 		}
1533 		/*
1534 		 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1535 		 * For the prefault case (!pages) we only update counts.
1536 		 */
1537 		if (likely(pages))
1538 			pages += ret;
1539 		start += ret << PAGE_SHIFT;
1540 
1541 		/* The lock was temporarily dropped, so we must unlock later */
1542 		must_unlock = true;
1543 
1544 retry:
1545 		/*
1546 		 * Repeat on the address that fired VM_FAULT_RETRY
1547 		 * with both FAULT_FLAG_ALLOW_RETRY and
1548 		 * FAULT_FLAG_TRIED.  Note that GUP can be interrupted
1549 		 * by fatal signals of even common signals, depending on
1550 		 * the caller's request. So we need to check it before we
1551 		 * start trying again otherwise it can loop forever.
1552 		 */
1553 		if (gup_signal_pending(flags)) {
1554 			if (!pages_done)
1555 				pages_done = -EINTR;
1556 			break;
1557 		}
1558 
1559 		ret = mmap_read_lock_killable(mm);
1560 		if (ret) {
1561 			BUG_ON(ret > 0);
1562 			if (!pages_done)
1563 				pages_done = ret;
1564 			break;
1565 		}
1566 
1567 		*locked = 1;
1568 		ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1569 				       pages, locked);
1570 		if (!*locked) {
1571 			/* Continue to retry until we succeeded */
1572 			BUG_ON(ret != 0);
1573 			goto retry;
1574 		}
1575 		if (ret != 1) {
1576 			BUG_ON(ret > 1);
1577 			if (!pages_done)
1578 				pages_done = ret;
1579 			break;
1580 		}
1581 		nr_pages--;
1582 		pages_done++;
1583 		if (!nr_pages)
1584 			break;
1585 		if (likely(pages))
1586 			pages++;
1587 		start += PAGE_SIZE;
1588 	}
1589 	if (must_unlock && *locked) {
1590 		/*
1591 		 * We either temporarily dropped the lock, or the caller
1592 		 * requested that we both acquire and drop the lock. Either way,
1593 		 * we must now unlock, and notify the caller of that state.
1594 		 */
1595 		mmap_read_unlock(mm);
1596 		*locked = 0;
1597 	}
1598 	return pages_done;
1599 }
1600 
1601 /**
1602  * populate_vma_page_range() -  populate a range of pages in the vma.
1603  * @vma:   target vma
1604  * @start: start address
1605  * @end:   end address
1606  * @locked: whether the mmap_lock is still held
1607  *
1608  * This takes care of mlocking the pages too if VM_LOCKED is set.
1609  *
1610  * Return either number of pages pinned in the vma, or a negative error
1611  * code on error.
1612  *
1613  * vma->vm_mm->mmap_lock must be held.
1614  *
1615  * If @locked is NULL, it may be held for read or write and will
1616  * be unperturbed.
1617  *
1618  * If @locked is non-NULL, it must held for read only and may be
1619  * released.  If it's released, *@locked will be set to 0.
1620  */
populate_vma_page_range(struct vm_area_struct * vma,unsigned long start,unsigned long end,int * locked)1621 long populate_vma_page_range(struct vm_area_struct *vma,
1622 		unsigned long start, unsigned long end, int *locked)
1623 {
1624 	struct mm_struct *mm = vma->vm_mm;
1625 	unsigned long nr_pages = (end - start) / PAGE_SIZE;
1626 	int local_locked = 1;
1627 	int gup_flags;
1628 	long ret;
1629 
1630 	VM_BUG_ON(!PAGE_ALIGNED(start));
1631 	VM_BUG_ON(!PAGE_ALIGNED(end));
1632 	VM_BUG_ON_VMA(start < vma->vm_start, vma);
1633 	VM_BUG_ON_VMA(end   > vma->vm_end, vma);
1634 	mmap_assert_locked(mm);
1635 
1636 	/*
1637 	 * Rightly or wrongly, the VM_LOCKONFAULT case has never used
1638 	 * faultin_page() to break COW, so it has no work to do here.
1639 	 */
1640 	if (vma->vm_flags & VM_LOCKONFAULT)
1641 		return nr_pages;
1642 
1643 	gup_flags = FOLL_TOUCH;
1644 	/*
1645 	 * We want to touch writable mappings with a write fault in order
1646 	 * to break COW, except for shared mappings because these don't COW
1647 	 * and we would not want to dirty them for nothing.
1648 	 */
1649 	if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1650 		gup_flags |= FOLL_WRITE;
1651 
1652 	/*
1653 	 * We want mlock to succeed for regions that have any permissions
1654 	 * other than PROT_NONE.
1655 	 */
1656 	if (vma_is_accessible(vma))
1657 		gup_flags |= FOLL_FORCE;
1658 
1659 	if (locked)
1660 		gup_flags |= FOLL_UNLOCKABLE;
1661 
1662 	/*
1663 	 * We made sure addr is within a VMA, so the following will
1664 	 * not result in a stack expansion that recurses back here.
1665 	 */
1666 	ret = __get_user_pages(mm, start, nr_pages, gup_flags,
1667 			       NULL, locked ? locked : &local_locked);
1668 	lru_add_drain();
1669 	return ret;
1670 }
1671 
1672 /*
1673  * faultin_vma_page_range() - populate (prefault) page tables inside the
1674  *			      given VMA range readable/writable
1675  *
1676  * This takes care of mlocking the pages, too, if VM_LOCKED is set.
1677  *
1678  * @vma: target vma
1679  * @start: start address
1680  * @end: end address
1681  * @write: whether to prefault readable or writable
1682  * @locked: whether the mmap_lock is still held
1683  *
1684  * Returns either number of processed pages in the vma, or a negative error
1685  * code on error (see __get_user_pages()).
1686  *
1687  * vma->vm_mm->mmap_lock must be held. The range must be page-aligned and
1688  * covered by the VMA. If it's released, *@locked will be set to 0.
1689  */
faultin_vma_page_range(struct vm_area_struct * vma,unsigned long start,unsigned long end,bool write,int * locked)1690 long faultin_vma_page_range(struct vm_area_struct *vma, unsigned long start,
1691 			    unsigned long end, bool write, int *locked)
1692 {
1693 	struct mm_struct *mm = vma->vm_mm;
1694 	unsigned long nr_pages = (end - start) / PAGE_SIZE;
1695 	int gup_flags;
1696 	long ret;
1697 
1698 	VM_BUG_ON(!PAGE_ALIGNED(start));
1699 	VM_BUG_ON(!PAGE_ALIGNED(end));
1700 	VM_BUG_ON_VMA(start < vma->vm_start, vma);
1701 	VM_BUG_ON_VMA(end > vma->vm_end, vma);
1702 	mmap_assert_locked(mm);
1703 
1704 	/*
1705 	 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark
1706 	 *	       the page dirty with FOLL_WRITE -- which doesn't make a
1707 	 *	       difference with !FOLL_FORCE, because the page is writable
1708 	 *	       in the page table.
1709 	 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit
1710 	 *		  a poisoned page.
1711 	 * !FOLL_FORCE: Require proper access permissions.
1712 	 */
1713 	gup_flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_UNLOCKABLE;
1714 	if (write)
1715 		gup_flags |= FOLL_WRITE;
1716 
1717 	/*
1718 	 * We want to report -EINVAL instead of -EFAULT for any permission
1719 	 * problems or incompatible mappings.
1720 	 */
1721 	if (check_vma_flags(vma, gup_flags))
1722 		return -EINVAL;
1723 
1724 	ret = __get_user_pages(mm, start, nr_pages, gup_flags,
1725 			       NULL, locked);
1726 	lru_add_drain();
1727 	return ret;
1728 }
1729 
1730 /*
1731  * __mm_populate - populate and/or mlock pages within a range of address space.
1732  *
1733  * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1734  * flags. VMAs must be already marked with the desired vm_flags, and
1735  * mmap_lock must not be held.
1736  */
__mm_populate(unsigned long start,unsigned long len,int ignore_errors)1737 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1738 {
1739 	struct mm_struct *mm = current->mm;
1740 	unsigned long end, nstart, nend;
1741 	struct vm_area_struct *vma = NULL;
1742 	int locked = 0;
1743 	long ret = 0;
1744 
1745 	end = start + len;
1746 
1747 	for (nstart = start; nstart < end; nstart = nend) {
1748 		/*
1749 		 * We want to fault in pages for [nstart; end) address range.
1750 		 * Find first corresponding VMA.
1751 		 */
1752 		if (!locked) {
1753 			locked = 1;
1754 			mmap_read_lock(mm);
1755 			vma = find_vma_intersection(mm, nstart, end);
1756 		} else if (nstart >= vma->vm_end)
1757 			vma = find_vma_intersection(mm, vma->vm_end, end);
1758 
1759 		if (!vma)
1760 			break;
1761 		/*
1762 		 * Set [nstart; nend) to intersection of desired address
1763 		 * range with the first VMA. Also, skip undesirable VMA types.
1764 		 */
1765 		nend = min(end, vma->vm_end);
1766 		if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1767 			continue;
1768 		if (nstart < vma->vm_start)
1769 			nstart = vma->vm_start;
1770 		/*
1771 		 * Now fault in a range of pages. populate_vma_page_range()
1772 		 * double checks the vma flags, so that it won't mlock pages
1773 		 * if the vma was already munlocked.
1774 		 */
1775 		ret = populate_vma_page_range(vma, nstart, nend, &locked);
1776 		if (ret < 0) {
1777 			if (ignore_errors) {
1778 				ret = 0;
1779 				continue;	/* continue at next VMA */
1780 			}
1781 			break;
1782 		}
1783 		nend = nstart + ret * PAGE_SIZE;
1784 		ret = 0;
1785 	}
1786 	if (locked)
1787 		mmap_read_unlock(mm);
1788 	return ret;	/* 0 or negative error code */
1789 }
1790 #else /* CONFIG_MMU */
__get_user_pages_locked(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,int * locked,unsigned int foll_flags)1791 static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1792 		unsigned long nr_pages, struct page **pages,
1793 		int *locked, unsigned int foll_flags)
1794 {
1795 	struct vm_area_struct *vma;
1796 	bool must_unlock = false;
1797 	unsigned long vm_flags;
1798 	long i;
1799 
1800 	if (!nr_pages)
1801 		return 0;
1802 
1803 	/*
1804 	 * The internal caller expects GUP to manage the lock internally and the
1805 	 * lock must be released when this returns.
1806 	 */
1807 	if (!*locked) {
1808 		if (mmap_read_lock_killable(mm))
1809 			return -EAGAIN;
1810 		must_unlock = true;
1811 		*locked = 1;
1812 	}
1813 
1814 	/* calculate required read or write permissions.
1815 	 * If FOLL_FORCE is set, we only require the "MAY" flags.
1816 	 */
1817 	vm_flags  = (foll_flags & FOLL_WRITE) ?
1818 			(VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1819 	vm_flags &= (foll_flags & FOLL_FORCE) ?
1820 			(VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1821 
1822 	for (i = 0; i < nr_pages; i++) {
1823 		vma = find_vma(mm, start);
1824 		if (!vma)
1825 			break;
1826 
1827 		/* protect what we can, including chardevs */
1828 		if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1829 		    !(vm_flags & vma->vm_flags))
1830 			break;
1831 
1832 		if (pages) {
1833 			pages[i] = virt_to_page((void *)start);
1834 			if (pages[i])
1835 				get_page(pages[i]);
1836 		}
1837 
1838 		start = (start + PAGE_SIZE) & PAGE_MASK;
1839 	}
1840 
1841 	if (must_unlock && *locked) {
1842 		mmap_read_unlock(mm);
1843 		*locked = 0;
1844 	}
1845 
1846 	return i ? : -EFAULT;
1847 }
1848 #endif /* !CONFIG_MMU */
1849 
1850 /**
1851  * fault_in_writeable - fault in userspace address range for writing
1852  * @uaddr: start of address range
1853  * @size: size of address range
1854  *
1855  * Returns the number of bytes not faulted in (like copy_to_user() and
1856  * copy_from_user()).
1857  */
fault_in_writeable(char __user * uaddr,size_t size)1858 size_t fault_in_writeable(char __user *uaddr, size_t size)
1859 {
1860 	char __user *start = uaddr, *end;
1861 
1862 	if (unlikely(size == 0))
1863 		return 0;
1864 	if (!user_write_access_begin(uaddr, size))
1865 		return size;
1866 	if (!PAGE_ALIGNED(uaddr)) {
1867 		unsafe_put_user(0, uaddr, out);
1868 		uaddr = (char __user *)PAGE_ALIGN((unsigned long)uaddr);
1869 	}
1870 	end = (char __user *)PAGE_ALIGN((unsigned long)start + size);
1871 	if (unlikely(end < start))
1872 		end = NULL;
1873 	while (uaddr != end) {
1874 		unsafe_put_user(0, uaddr, out);
1875 		uaddr += PAGE_SIZE;
1876 	}
1877 
1878 out:
1879 	user_write_access_end();
1880 	if (size > uaddr - start)
1881 		return size - (uaddr - start);
1882 	return 0;
1883 }
1884 EXPORT_SYMBOL(fault_in_writeable);
1885 
1886 /**
1887  * fault_in_subpage_writeable - fault in an address range for writing
1888  * @uaddr: start of address range
1889  * @size: size of address range
1890  *
1891  * Fault in a user address range for writing while checking for permissions at
1892  * sub-page granularity (e.g. arm64 MTE). This function should be used when
1893  * the caller cannot guarantee forward progress of a copy_to_user() loop.
1894  *
1895  * Returns the number of bytes not faulted in (like copy_to_user() and
1896  * copy_from_user()).
1897  */
fault_in_subpage_writeable(char __user * uaddr,size_t size)1898 size_t fault_in_subpage_writeable(char __user *uaddr, size_t size)
1899 {
1900 	size_t faulted_in;
1901 
1902 	/*
1903 	 * Attempt faulting in at page granularity first for page table
1904 	 * permission checking. The arch-specific probe_subpage_writeable()
1905 	 * functions may not check for this.
1906 	 */
1907 	faulted_in = size - fault_in_writeable(uaddr, size);
1908 	if (faulted_in)
1909 		faulted_in -= probe_subpage_writeable(uaddr, faulted_in);
1910 
1911 	return size - faulted_in;
1912 }
1913 EXPORT_SYMBOL(fault_in_subpage_writeable);
1914 
1915 /*
1916  * fault_in_safe_writeable - fault in an address range for writing
1917  * @uaddr: start of address range
1918  * @size: length of address range
1919  *
1920  * Faults in an address range for writing.  This is primarily useful when we
1921  * already know that some or all of the pages in the address range aren't in
1922  * memory.
1923  *
1924  * Unlike fault_in_writeable(), this function is non-destructive.
1925  *
1926  * Note that we don't pin or otherwise hold the pages referenced that we fault
1927  * in.  There's no guarantee that they'll stay in memory for any duration of
1928  * time.
1929  *
1930  * Returns the number of bytes not faulted in, like copy_to_user() and
1931  * copy_from_user().
1932  */
fault_in_safe_writeable(const char __user * uaddr,size_t size)1933 size_t fault_in_safe_writeable(const char __user *uaddr, size_t size)
1934 {
1935 	unsigned long start = (unsigned long)uaddr, end;
1936 	struct mm_struct *mm = current->mm;
1937 	bool unlocked = false;
1938 
1939 	if (unlikely(size == 0))
1940 		return 0;
1941 	end = PAGE_ALIGN(start + size);
1942 	if (end < start)
1943 		end = 0;
1944 
1945 	mmap_read_lock(mm);
1946 	do {
1947 		if (fixup_user_fault(mm, start, FAULT_FLAG_WRITE, &unlocked))
1948 			break;
1949 		start = (start + PAGE_SIZE) & PAGE_MASK;
1950 	} while (start != end);
1951 	mmap_read_unlock(mm);
1952 
1953 	if (size > (unsigned long)uaddr - start)
1954 		return size - ((unsigned long)uaddr - start);
1955 	return 0;
1956 }
1957 EXPORT_SYMBOL(fault_in_safe_writeable);
1958 
1959 /**
1960  * fault_in_readable - fault in userspace address range for reading
1961  * @uaddr: start of user address range
1962  * @size: size of user address range
1963  *
1964  * Returns the number of bytes not faulted in (like copy_to_user() and
1965  * copy_from_user()).
1966  */
fault_in_readable(const char __user * uaddr,size_t size)1967 size_t fault_in_readable(const char __user *uaddr, size_t size)
1968 {
1969 	const char __user *start = uaddr, *end;
1970 	volatile char c;
1971 
1972 	if (unlikely(size == 0))
1973 		return 0;
1974 	if (!user_read_access_begin(uaddr, size))
1975 		return size;
1976 	if (!PAGE_ALIGNED(uaddr)) {
1977 		unsafe_get_user(c, uaddr, out);
1978 		uaddr = (const char __user *)PAGE_ALIGN((unsigned long)uaddr);
1979 	}
1980 	end = (const char __user *)PAGE_ALIGN((unsigned long)start + size);
1981 	if (unlikely(end < start))
1982 		end = NULL;
1983 	while (uaddr != end) {
1984 		unsafe_get_user(c, uaddr, out);
1985 		uaddr += PAGE_SIZE;
1986 	}
1987 
1988 out:
1989 	user_read_access_end();
1990 	(void)c;
1991 	if (size > uaddr - start)
1992 		return size - (uaddr - start);
1993 	return 0;
1994 }
1995 EXPORT_SYMBOL(fault_in_readable);
1996 
1997 /**
1998  * get_dump_page() - pin user page in memory while writing it to core dump
1999  * @addr: user address
2000  *
2001  * Returns struct page pointer of user page pinned for dump,
2002  * to be freed afterwards by put_page().
2003  *
2004  * Returns NULL on any kind of failure - a hole must then be inserted into
2005  * the corefile, to preserve alignment with its headers; and also returns
2006  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2007  * allowing a hole to be left in the corefile to save disk space.
2008  *
2009  * Called without mmap_lock (takes and releases the mmap_lock by itself).
2010  */
2011 #ifdef CONFIG_ELF_CORE
get_dump_page(unsigned long addr)2012 struct page *get_dump_page(unsigned long addr)
2013 {
2014 	struct page *page;
2015 	int locked = 0;
2016 	int ret;
2017 
2018 	ret = __get_user_pages_locked(current->mm, addr, 1, &page, &locked,
2019 				      FOLL_FORCE | FOLL_DUMP | FOLL_GET);
2020 	return (ret == 1) ? page : NULL;
2021 }
2022 #endif /* CONFIG_ELF_CORE */
2023 
2024 #ifdef CONFIG_MIGRATION
2025 /*
2026  * Returns the number of collected pages. Return value is always >= 0.
2027  */
collect_longterm_unpinnable_pages(struct list_head * movable_page_list,unsigned long nr_pages,struct page ** pages)2028 static unsigned long collect_longterm_unpinnable_pages(
2029 					struct list_head *movable_page_list,
2030 					unsigned long nr_pages,
2031 					struct page **pages)
2032 {
2033 	unsigned long i, collected = 0;
2034 	struct folio *prev_folio = NULL;
2035 	bool drain_allow = true;
2036 
2037 	for (i = 0; i < nr_pages; i++) {
2038 		struct folio *folio = page_folio(pages[i]);
2039 
2040 		if (folio == prev_folio)
2041 			continue;
2042 		prev_folio = folio;
2043 
2044 		if (folio_is_longterm_pinnable(folio))
2045 			continue;
2046 
2047 		collected++;
2048 
2049 		if (folio_is_device_coherent(folio))
2050 			continue;
2051 
2052 		if (folio_test_hugetlb(folio)) {
2053 			isolate_hugetlb(folio, movable_page_list);
2054 			continue;
2055 		}
2056 
2057 		if (!folio_test_lru(folio) && drain_allow) {
2058 			lru_add_drain_all();
2059 			drain_allow = false;
2060 		}
2061 
2062 		if (!folio_isolate_lru(folio))
2063 			continue;
2064 
2065 		list_add_tail(&folio->lru, movable_page_list);
2066 		node_stat_mod_folio(folio,
2067 				    NR_ISOLATED_ANON + folio_is_file_lru(folio),
2068 				    folio_nr_pages(folio));
2069 	}
2070 
2071 	return collected;
2072 }
2073 
2074 /*
2075  * Unpins all pages and migrates device coherent pages and movable_page_list.
2076  * Returns -EAGAIN if all pages were successfully migrated or -errno for failure
2077  * (or partial success).
2078  */
migrate_longterm_unpinnable_pages(struct list_head * movable_page_list,unsigned long nr_pages,struct page ** pages)2079 static int migrate_longterm_unpinnable_pages(
2080 					struct list_head *movable_page_list,
2081 					unsigned long nr_pages,
2082 					struct page **pages)
2083 {
2084 	int ret;
2085 	unsigned long i;
2086 
2087 	for (i = 0; i < nr_pages; i++) {
2088 		struct folio *folio = page_folio(pages[i]);
2089 
2090 		if (folio_is_device_coherent(folio)) {
2091 			/*
2092 			 * Migration will fail if the page is pinned, so convert
2093 			 * the pin on the source page to a normal reference.
2094 			 */
2095 			pages[i] = NULL;
2096 			folio_get(folio);
2097 			gup_put_folio(folio, 1, FOLL_PIN);
2098 
2099 			if (migrate_device_coherent_page(&folio->page)) {
2100 				ret = -EBUSY;
2101 				goto err;
2102 			}
2103 
2104 			continue;
2105 		}
2106 
2107 		/*
2108 		 * We can't migrate pages with unexpected references, so drop
2109 		 * the reference obtained by __get_user_pages_locked().
2110 		 * Migrating pages have been added to movable_page_list after
2111 		 * calling folio_isolate_lru() which takes a reference so the
2112 		 * page won't be freed if it's migrating.
2113 		 */
2114 		unpin_user_page(pages[i]);
2115 		pages[i] = NULL;
2116 	}
2117 
2118 	if (!list_empty(movable_page_list)) {
2119 		struct migration_target_control mtc = {
2120 			.nid = NUMA_NO_NODE,
2121 			.gfp_mask = GFP_USER | __GFP_NOWARN,
2122 		};
2123 
2124 		if (migrate_pages(movable_page_list, alloc_migration_target,
2125 				  NULL, (unsigned long)&mtc, MIGRATE_SYNC,
2126 				  MR_LONGTERM_PIN, NULL)) {
2127 			ret = -ENOMEM;
2128 			goto err;
2129 		}
2130 	}
2131 
2132 	putback_movable_pages(movable_page_list);
2133 
2134 	return -EAGAIN;
2135 
2136 err:
2137 	for (i = 0; i < nr_pages; i++)
2138 		if (pages[i])
2139 			unpin_user_page(pages[i]);
2140 	putback_movable_pages(movable_page_list);
2141 
2142 	return ret;
2143 }
2144 
2145 /*
2146  * Check whether all pages are *allowed* to be pinned. Rather confusingly, all
2147  * pages in the range are required to be pinned via FOLL_PIN, before calling
2148  * this routine.
2149  *
2150  * If any pages in the range are not allowed to be pinned, then this routine
2151  * will migrate those pages away, unpin all the pages in the range and return
2152  * -EAGAIN. The caller should re-pin the entire range with FOLL_PIN and then
2153  * call this routine again.
2154  *
2155  * If an error other than -EAGAIN occurs, this indicates a migration failure.
2156  * The caller should give up, and propagate the error back up the call stack.
2157  *
2158  * If everything is OK and all pages in the range are allowed to be pinned, then
2159  * this routine leaves all pages pinned and returns zero for success.
2160  */
check_and_migrate_movable_pages(unsigned long nr_pages,struct page ** pages)2161 static long check_and_migrate_movable_pages(unsigned long nr_pages,
2162 					    struct page **pages)
2163 {
2164 	unsigned long collected;
2165 	LIST_HEAD(movable_page_list);
2166 
2167 	collected = collect_longterm_unpinnable_pages(&movable_page_list,
2168 						nr_pages, pages);
2169 	if (!collected)
2170 		return 0;
2171 
2172 	return migrate_longterm_unpinnable_pages(&movable_page_list, nr_pages,
2173 						pages);
2174 }
2175 #else
check_and_migrate_movable_pages(unsigned long nr_pages,struct page ** pages)2176 static long check_and_migrate_movable_pages(unsigned long nr_pages,
2177 					    struct page **pages)
2178 {
2179 	return 0;
2180 }
2181 #endif /* CONFIG_MIGRATION */
2182 
2183 /*
2184  * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
2185  * allows us to process the FOLL_LONGTERM flag.
2186  */
__gup_longterm_locked(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,int * locked,unsigned int gup_flags)2187 static long __gup_longterm_locked(struct mm_struct *mm,
2188 				  unsigned long start,
2189 				  unsigned long nr_pages,
2190 				  struct page **pages,
2191 				  int *locked,
2192 				  unsigned int gup_flags)
2193 {
2194 	unsigned int flags;
2195 	long rc, nr_pinned_pages;
2196 
2197 	if (!(gup_flags & FOLL_LONGTERM))
2198 		return __get_user_pages_locked(mm, start, nr_pages, pages,
2199 					       locked, gup_flags);
2200 
2201 	flags = memalloc_pin_save();
2202 	do {
2203 		nr_pinned_pages = __get_user_pages_locked(mm, start, nr_pages,
2204 							  pages, locked,
2205 							  gup_flags);
2206 		if (nr_pinned_pages <= 0) {
2207 			rc = nr_pinned_pages;
2208 			break;
2209 		}
2210 
2211 		/* FOLL_LONGTERM implies FOLL_PIN */
2212 		rc = check_and_migrate_movable_pages(nr_pinned_pages, pages);
2213 	} while (rc == -EAGAIN);
2214 	memalloc_pin_restore(flags);
2215 	return rc ? rc : nr_pinned_pages;
2216 }
2217 
2218 /*
2219  * Check that the given flags are valid for the exported gup/pup interface, and
2220  * update them with the required flags that the caller must have set.
2221  */
is_valid_gup_args(struct page ** pages,int * locked,unsigned int * gup_flags_p,unsigned int to_set)2222 static bool is_valid_gup_args(struct page **pages, int *locked,
2223 			      unsigned int *gup_flags_p, unsigned int to_set)
2224 {
2225 	unsigned int gup_flags = *gup_flags_p;
2226 
2227 	/*
2228 	 * These flags not allowed to be specified externally to the gup
2229 	 * interfaces:
2230 	 * - FOLL_PIN/FOLL_TRIED/FOLL_FAST_ONLY are internal only
2231 	 * - FOLL_REMOTE is internal only and used on follow_page()
2232 	 * - FOLL_UNLOCKABLE is internal only and used if locked is !NULL
2233 	 */
2234 	if (WARN_ON_ONCE(gup_flags & (FOLL_PIN | FOLL_TRIED | FOLL_UNLOCKABLE |
2235 				      FOLL_REMOTE | FOLL_FAST_ONLY)))
2236 		return false;
2237 
2238 	gup_flags |= to_set;
2239 	if (locked) {
2240 		/* At the external interface locked must be set */
2241 		if (WARN_ON_ONCE(*locked != 1))
2242 			return false;
2243 
2244 		gup_flags |= FOLL_UNLOCKABLE;
2245 	}
2246 
2247 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
2248 	if (WARN_ON_ONCE((gup_flags & (FOLL_PIN | FOLL_GET)) ==
2249 			 (FOLL_PIN | FOLL_GET)))
2250 		return false;
2251 
2252 	/* LONGTERM can only be specified when pinning */
2253 	if (WARN_ON_ONCE(!(gup_flags & FOLL_PIN) && (gup_flags & FOLL_LONGTERM)))
2254 		return false;
2255 
2256 	/* Pages input must be given if using GET/PIN */
2257 	if (WARN_ON_ONCE((gup_flags & (FOLL_GET | FOLL_PIN)) && !pages))
2258 		return false;
2259 
2260 	/* We want to allow the pgmap to be hot-unplugged at all times */
2261 	if (WARN_ON_ONCE((gup_flags & FOLL_LONGTERM) &&
2262 			 (gup_flags & FOLL_PCI_P2PDMA)))
2263 		return false;
2264 
2265 	*gup_flags_p = gup_flags;
2266 	return true;
2267 }
2268 
2269 #ifdef CONFIG_MMU
2270 /**
2271  * get_user_pages_remote() - pin user pages in memory
2272  * @mm:		mm_struct of target mm
2273  * @start:	starting user address
2274  * @nr_pages:	number of pages from start to pin
2275  * @gup_flags:	flags modifying lookup behaviour
2276  * @pages:	array that receives pointers to the pages pinned.
2277  *		Should be at least nr_pages long. Or NULL, if caller
2278  *		only intends to ensure the pages are faulted in.
2279  * @locked:	pointer to lock flag indicating whether lock is held and
2280  *		subsequently whether VM_FAULT_RETRY functionality can be
2281  *		utilised. Lock must initially be held.
2282  *
2283  * Returns either number of pages pinned (which may be less than the
2284  * number requested), or an error. Details about the return value:
2285  *
2286  * -- If nr_pages is 0, returns 0.
2287  * -- If nr_pages is >0, but no pages were pinned, returns -errno.
2288  * -- If nr_pages is >0, and some pages were pinned, returns the number of
2289  *    pages pinned. Again, this may be less than nr_pages.
2290  *
2291  * The caller is responsible for releasing returned @pages, via put_page().
2292  *
2293  * Must be called with mmap_lock held for read or write.
2294  *
2295  * get_user_pages_remote walks a process's page tables and takes a reference
2296  * to each struct page that each user address corresponds to at a given
2297  * instant. That is, it takes the page that would be accessed if a user
2298  * thread accesses the given user virtual address at that instant.
2299  *
2300  * This does not guarantee that the page exists in the user mappings when
2301  * get_user_pages_remote returns, and there may even be a completely different
2302  * page there in some cases (eg. if mmapped pagecache has been invalidated
2303  * and subsequently re-faulted). However it does guarantee that the page
2304  * won't be freed completely. And mostly callers simply care that the page
2305  * contains data that was valid *at some point in time*. Typically, an IO
2306  * or similar operation cannot guarantee anything stronger anyway because
2307  * locks can't be held over the syscall boundary.
2308  *
2309  * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
2310  * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
2311  * be called after the page is finished with, and before put_page is called.
2312  *
2313  * get_user_pages_remote is typically used for fewer-copy IO operations,
2314  * to get a handle on the memory by some means other than accesses
2315  * via the user virtual addresses. The pages may be submitted for
2316  * DMA to devices or accessed via their kernel linear mapping (via the
2317  * kmap APIs). Care should be taken to use the correct cache flushing APIs.
2318  *
2319  * See also get_user_pages_fast, for performance critical applications.
2320  *
2321  * get_user_pages_remote should be phased out in favor of
2322  * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
2323  * should use get_user_pages_remote because it cannot pass
2324  * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
2325  */
get_user_pages_remote(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)2326 long get_user_pages_remote(struct mm_struct *mm,
2327 		unsigned long start, unsigned long nr_pages,
2328 		unsigned int gup_flags, struct page **pages,
2329 		int *locked)
2330 {
2331 	int local_locked = 1;
2332 
2333 	if (!is_valid_gup_args(pages, locked, &gup_flags,
2334 			       FOLL_TOUCH | FOLL_REMOTE))
2335 		return -EINVAL;
2336 
2337 	return __get_user_pages_locked(mm, start, nr_pages, pages,
2338 				       locked ? locked : &local_locked,
2339 				       gup_flags);
2340 }
2341 EXPORT_SYMBOL(get_user_pages_remote);
2342 
2343 #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,int * locked)2344 long get_user_pages_remote(struct mm_struct *mm,
2345 			   unsigned long start, unsigned long nr_pages,
2346 			   unsigned int gup_flags, struct page **pages,
2347 			   int *locked)
2348 {
2349 	return 0;
2350 }
2351 #endif /* !CONFIG_MMU */
2352 
2353 /**
2354  * get_user_pages() - pin user pages in memory
2355  * @start:      starting user address
2356  * @nr_pages:   number of pages from start to pin
2357  * @gup_flags:  flags modifying lookup behaviour
2358  * @pages:      array that receives pointers to the pages pinned.
2359  *              Should be at least nr_pages long. Or NULL, if caller
2360  *              only intends to ensure the pages are faulted in.
2361  *
2362  * This is the same as get_user_pages_remote(), just with a less-flexible
2363  * calling convention where we assume that the mm being operated on belongs to
2364  * the current task, and doesn't allow passing of a locked parameter.  We also
2365  * obviously don't pass FOLL_REMOTE in here.
2366  */
get_user_pages(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages)2367 long get_user_pages(unsigned long start, unsigned long nr_pages,
2368 		    unsigned int gup_flags, struct page **pages)
2369 {
2370 	int locked = 1;
2371 
2372 	if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_TOUCH))
2373 		return -EINVAL;
2374 
2375 	return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2376 				       &locked, gup_flags);
2377 }
2378 EXPORT_SYMBOL(get_user_pages);
2379 
2380 /*
2381  * get_user_pages_unlocked() is suitable to replace the form:
2382  *
2383  *      mmap_read_lock(mm);
2384  *      get_user_pages(mm, ..., pages, NULL);
2385  *      mmap_read_unlock(mm);
2386  *
2387  *  with:
2388  *
2389  *      get_user_pages_unlocked(mm, ..., pages);
2390  *
2391  * It is functionally equivalent to get_user_pages_fast so
2392  * get_user_pages_fast should be used instead if specific gup_flags
2393  * (e.g. FOLL_FORCE) are not required.
2394  */
get_user_pages_unlocked(unsigned long start,unsigned long nr_pages,struct page ** pages,unsigned int gup_flags)2395 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2396 			     struct page **pages, unsigned int gup_flags)
2397 {
2398 	int locked = 0;
2399 
2400 	if (!is_valid_gup_args(pages, NULL, &gup_flags,
2401 			       FOLL_TOUCH | FOLL_UNLOCKABLE))
2402 		return -EINVAL;
2403 
2404 	return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2405 				       &locked, gup_flags);
2406 }
2407 EXPORT_SYMBOL(get_user_pages_unlocked);
2408 
2409 /*
2410  * Fast GUP
2411  *
2412  * get_user_pages_fast attempts to pin user pages by walking the page
2413  * tables directly and avoids taking locks. Thus the walker needs to be
2414  * protected from page table pages being freed from under it, and should
2415  * block any THP splits.
2416  *
2417  * One way to achieve this is to have the walker disable interrupts, and
2418  * rely on IPIs from the TLB flushing code blocking before the page table
2419  * pages are freed. This is unsuitable for architectures that do not need
2420  * to broadcast an IPI when invalidating TLBs.
2421  *
2422  * Another way to achieve this is to batch up page table containing pages
2423  * belonging to more than one mm_user, then rcu_sched a callback to free those
2424  * pages. Disabling interrupts will allow the fast_gup walker to both block
2425  * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2426  * (which is a relatively rare event). The code below adopts this strategy.
2427  *
2428  * Before activating this code, please be aware that the following assumptions
2429  * are currently made:
2430  *
2431  *  *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2432  *  free pages containing page tables or TLB flushing requires IPI broadcast.
2433  *
2434  *  *) ptes can be read atomically by the architecture.
2435  *
2436  *  *) access_ok is sufficient to validate userspace address ranges.
2437  *
2438  * The last two assumptions can be relaxed by the addition of helper functions.
2439  *
2440  * This code is based heavily on the PowerPC implementation by Nick Piggin.
2441  */
2442 #ifdef CONFIG_HAVE_FAST_GUP
2443 
2444 /*
2445  * Used in the GUP-fast path to determine whether a pin is permitted for a
2446  * specific folio.
2447  *
2448  * This call assumes the caller has pinned the folio, that the lowest page table
2449  * level still points to this folio, and that interrupts have been disabled.
2450  *
2451  * Writing to pinned file-backed dirty tracked folios is inherently problematic
2452  * (see comment describing the writable_file_mapping_allowed() function). We
2453  * therefore try to avoid the most egregious case of a long-term mapping doing
2454  * so.
2455  *
2456  * This function cannot be as thorough as that one as the VMA is not available
2457  * in the fast path, so instead we whitelist known good cases and if in doubt,
2458  * fall back to the slow path.
2459  */
folio_fast_pin_allowed(struct folio * folio,unsigned int flags)2460 static bool folio_fast_pin_allowed(struct folio *folio, unsigned int flags)
2461 {
2462 	struct address_space *mapping;
2463 	unsigned long mapping_flags;
2464 
2465 	/*
2466 	 * If we aren't pinning then no problematic write can occur. A long term
2467 	 * pin is the most egregious case so this is the one we disallow.
2468 	 */
2469 	if ((flags & (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE)) !=
2470 	    (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE))
2471 		return true;
2472 
2473 	/* The folio is pinned, so we can safely access folio fields. */
2474 
2475 	if (WARN_ON_ONCE(folio_test_slab(folio)))
2476 		return false;
2477 
2478 	/* hugetlb mappings do not require dirty-tracking. */
2479 	if (folio_test_hugetlb(folio))
2480 		return true;
2481 
2482 	/*
2483 	 * GUP-fast disables IRQs. When IRQS are disabled, RCU grace periods
2484 	 * cannot proceed, which means no actions performed under RCU can
2485 	 * proceed either.
2486 	 *
2487 	 * inodes and thus their mappings are freed under RCU, which means the
2488 	 * mapping cannot be freed beneath us and thus we can safely dereference
2489 	 * it.
2490 	 */
2491 	lockdep_assert_irqs_disabled();
2492 
2493 	/*
2494 	 * However, there may be operations which _alter_ the mapping, so ensure
2495 	 * we read it once and only once.
2496 	 */
2497 	mapping = READ_ONCE(folio->mapping);
2498 
2499 	/*
2500 	 * The mapping may have been truncated, in any case we cannot determine
2501 	 * if this mapping is safe - fall back to slow path to determine how to
2502 	 * proceed.
2503 	 */
2504 	if (!mapping)
2505 		return false;
2506 
2507 	/* Anonymous folios pose no problem. */
2508 	mapping_flags = (unsigned long)mapping & PAGE_MAPPING_FLAGS;
2509 	if (mapping_flags)
2510 		return mapping_flags & PAGE_MAPPING_ANON;
2511 
2512 	/*
2513 	 * At this point, we know the mapping is non-null and points to an
2514 	 * address_space object. The only remaining whitelisted file system is
2515 	 * shmem.
2516 	 */
2517 	return shmem_mapping(mapping);
2518 }
2519 
undo_dev_pagemap(int * nr,int nr_start,unsigned int flags,struct page ** pages)2520 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2521 					    unsigned int flags,
2522 					    struct page **pages)
2523 {
2524 	while ((*nr) - nr_start) {
2525 		struct page *page = pages[--(*nr)];
2526 
2527 		ClearPageReferenced(page);
2528 		if (flags & FOLL_PIN)
2529 			unpin_user_page(page);
2530 		else
2531 			put_page(page);
2532 	}
2533 }
2534 
2535 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2536 /*
2537  * Fast-gup relies on pte change detection to avoid concurrent pgtable
2538  * operations.
2539  *
2540  * To pin the page, fast-gup needs to do below in order:
2541  * (1) pin the page (by prefetching pte), then (2) check pte not changed.
2542  *
2543  * For the rest of pgtable operations where pgtable updates can be racy
2544  * with fast-gup, we need to do (1) clear pte, then (2) check whether page
2545  * is pinned.
2546  *
2547  * Above will work for all pte-level operations, including THP split.
2548  *
2549  * For THP collapse, it's a bit more complicated because fast-gup may be
2550  * walking a pgtable page that is being freed (pte is still valid but pmd
2551  * can be cleared already).  To avoid race in such condition, we need to
2552  * also check pmd here to make sure pmd doesn't change (corresponds to
2553  * pmdp_collapse_flush() in the THP collapse code path).
2554  */
gup_pte_range(pmd_t pmd,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2555 static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2556 			 unsigned long end, unsigned int flags,
2557 			 struct page **pages, int *nr)
2558 {
2559 	struct dev_pagemap *pgmap = NULL;
2560 	int nr_start = *nr, ret = 0;
2561 	pte_t *ptep, *ptem;
2562 
2563 	ptem = ptep = pte_offset_map(&pmd, addr);
2564 	if (!ptep)
2565 		return 0;
2566 	do {
2567 		pte_t pte = ptep_get_lockless(ptep);
2568 		struct page *page;
2569 		struct folio *folio;
2570 
2571 		/*
2572 		 * Always fallback to ordinary GUP on PROT_NONE-mapped pages:
2573 		 * pte_access_permitted() better should reject these pages
2574 		 * either way: otherwise, GUP-fast might succeed in
2575 		 * cases where ordinary GUP would fail due to VMA access
2576 		 * permissions.
2577 		 */
2578 		if (pte_protnone(pte))
2579 			goto pte_unmap;
2580 
2581 		if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2582 			goto pte_unmap;
2583 
2584 		if (pte_devmap(pte)) {
2585 			if (unlikely(flags & FOLL_LONGTERM))
2586 				goto pte_unmap;
2587 
2588 			pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2589 			if (unlikely(!pgmap)) {
2590 				undo_dev_pagemap(nr, nr_start, flags, pages);
2591 				goto pte_unmap;
2592 			}
2593 		} else if (pte_special(pte))
2594 			goto pte_unmap;
2595 
2596 		VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2597 		page = pte_page(pte);
2598 
2599 		folio = try_grab_folio(page, 1, flags);
2600 		if (!folio)
2601 			goto pte_unmap;
2602 
2603 		if (unlikely(folio_is_secretmem(folio))) {
2604 			gup_put_folio(folio, 1, flags);
2605 			goto pte_unmap;
2606 		}
2607 
2608 		if (unlikely(pmd_val(pmd) != pmd_val(*pmdp)) ||
2609 		    unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) {
2610 			gup_put_folio(folio, 1, flags);
2611 			goto pte_unmap;
2612 		}
2613 
2614 		if (!folio_fast_pin_allowed(folio, flags)) {
2615 			gup_put_folio(folio, 1, flags);
2616 			goto pte_unmap;
2617 		}
2618 
2619 		if (!pte_write(pte) && gup_must_unshare(NULL, flags, page)) {
2620 			gup_put_folio(folio, 1, flags);
2621 			goto pte_unmap;
2622 		}
2623 
2624 		/*
2625 		 * We need to make the page accessible if and only if we are
2626 		 * going to access its content (the FOLL_PIN case).  Please
2627 		 * see Documentation/core-api/pin_user_pages.rst for
2628 		 * details.
2629 		 */
2630 		if (flags & FOLL_PIN) {
2631 			ret = arch_make_page_accessible(page);
2632 			if (ret) {
2633 				gup_put_folio(folio, 1, flags);
2634 				goto pte_unmap;
2635 			}
2636 		}
2637 		folio_set_referenced(folio);
2638 		pages[*nr] = page;
2639 		(*nr)++;
2640 	} while (ptep++, addr += PAGE_SIZE, addr != end);
2641 
2642 	ret = 1;
2643 
2644 pte_unmap:
2645 	if (pgmap)
2646 		put_dev_pagemap(pgmap);
2647 	pte_unmap(ptem);
2648 	return ret;
2649 }
2650 #else
2651 
2652 /*
2653  * If we can't determine whether or not a pte is special, then fail immediately
2654  * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2655  * to be special.
2656  *
2657  * For a futex to be placed on a THP tail page, get_futex_key requires a
2658  * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2659  * useful to have gup_huge_pmd even if we can't operate on ptes.
2660  */
gup_pte_range(pmd_t pmd,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2661 static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2662 			 unsigned long end, unsigned int flags,
2663 			 struct page **pages, int *nr)
2664 {
2665 	return 0;
2666 }
2667 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2668 
2669 #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)2670 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2671 			     unsigned long end, unsigned int flags,
2672 			     struct page **pages, int *nr)
2673 {
2674 	int nr_start = *nr;
2675 	struct dev_pagemap *pgmap = NULL;
2676 
2677 	do {
2678 		struct page *page = pfn_to_page(pfn);
2679 
2680 		pgmap = get_dev_pagemap(pfn, pgmap);
2681 		if (unlikely(!pgmap)) {
2682 			undo_dev_pagemap(nr, nr_start, flags, pages);
2683 			break;
2684 		}
2685 
2686 		if (!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)) {
2687 			undo_dev_pagemap(nr, nr_start, flags, pages);
2688 			break;
2689 		}
2690 
2691 		SetPageReferenced(page);
2692 		pages[*nr] = page;
2693 		if (unlikely(try_grab_page(page, flags))) {
2694 			undo_dev_pagemap(nr, nr_start, flags, pages);
2695 			break;
2696 		}
2697 		(*nr)++;
2698 		pfn++;
2699 	} while (addr += PAGE_SIZE, addr != end);
2700 
2701 	put_dev_pagemap(pgmap);
2702 	return addr == end;
2703 }
2704 
__gup_device_huge_pmd(pmd_t orig,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2705 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2706 				 unsigned long end, unsigned int flags,
2707 				 struct page **pages, int *nr)
2708 {
2709 	unsigned long fault_pfn;
2710 	int nr_start = *nr;
2711 
2712 	fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2713 	if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2714 		return 0;
2715 
2716 	if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2717 		undo_dev_pagemap(nr, nr_start, flags, pages);
2718 		return 0;
2719 	}
2720 	return 1;
2721 }
2722 
__gup_device_huge_pud(pud_t orig,pud_t * pudp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2723 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2724 				 unsigned long end, unsigned int flags,
2725 				 struct page **pages, int *nr)
2726 {
2727 	unsigned long fault_pfn;
2728 	int nr_start = *nr;
2729 
2730 	fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2731 	if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2732 		return 0;
2733 
2734 	if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2735 		undo_dev_pagemap(nr, nr_start, flags, pages);
2736 		return 0;
2737 	}
2738 	return 1;
2739 }
2740 #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)2741 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2742 				 unsigned long end, unsigned int flags,
2743 				 struct page **pages, int *nr)
2744 {
2745 	BUILD_BUG();
2746 	return 0;
2747 }
2748 
__gup_device_huge_pud(pud_t pud,pud_t * pudp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2749 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2750 				 unsigned long end, unsigned int flags,
2751 				 struct page **pages, int *nr)
2752 {
2753 	BUILD_BUG();
2754 	return 0;
2755 }
2756 #endif
2757 
record_subpages(struct page * page,unsigned long addr,unsigned long end,struct page ** pages)2758 static int record_subpages(struct page *page, unsigned long addr,
2759 			   unsigned long end, struct page **pages)
2760 {
2761 	int nr;
2762 
2763 	for (nr = 0; addr != end; nr++, addr += PAGE_SIZE)
2764 		pages[nr] = nth_page(page, nr);
2765 
2766 	return nr;
2767 }
2768 
2769 #ifdef CONFIG_ARCH_HAS_HUGEPD
hugepte_addr_end(unsigned long addr,unsigned long end,unsigned long sz)2770 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2771 				      unsigned long sz)
2772 {
2773 	unsigned long __boundary = (addr + sz) & ~(sz-1);
2774 	return (__boundary - 1 < end - 1) ? __boundary : end;
2775 }
2776 
gup_hugepte(pte_t * ptep,unsigned long sz,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2777 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2778 		       unsigned long end, unsigned int flags,
2779 		       struct page **pages, int *nr)
2780 {
2781 	unsigned long pte_end;
2782 	struct page *page;
2783 	struct folio *folio;
2784 	pte_t pte;
2785 	int refs;
2786 
2787 	pte_end = (addr + sz) & ~(sz-1);
2788 	if (pte_end < end)
2789 		end = pte_end;
2790 
2791 	pte = huge_ptep_get(ptep);
2792 
2793 	if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2794 		return 0;
2795 
2796 	/* hugepages are never "special" */
2797 	VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2798 
2799 	page = nth_page(pte_page(pte), (addr & (sz - 1)) >> PAGE_SHIFT);
2800 	refs = record_subpages(page, addr, end, pages + *nr);
2801 
2802 	folio = try_grab_folio(page, refs, flags);
2803 	if (!folio)
2804 		return 0;
2805 
2806 	if (unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) {
2807 		gup_put_folio(folio, refs, flags);
2808 		return 0;
2809 	}
2810 
2811 	if (!folio_fast_pin_allowed(folio, flags)) {
2812 		gup_put_folio(folio, refs, flags);
2813 		return 0;
2814 	}
2815 
2816 	if (!pte_write(pte) && gup_must_unshare(NULL, flags, &folio->page)) {
2817 		gup_put_folio(folio, refs, flags);
2818 		return 0;
2819 	}
2820 
2821 	*nr += refs;
2822 	folio_set_referenced(folio);
2823 	return 1;
2824 }
2825 
gup_huge_pd(hugepd_t hugepd,unsigned long addr,unsigned int pdshift,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2826 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2827 		unsigned int pdshift, unsigned long end, unsigned int flags,
2828 		struct page **pages, int *nr)
2829 {
2830 	pte_t *ptep;
2831 	unsigned long sz = 1UL << hugepd_shift(hugepd);
2832 	unsigned long next;
2833 
2834 	ptep = hugepte_offset(hugepd, addr, pdshift);
2835 	do {
2836 		next = hugepte_addr_end(addr, end, sz);
2837 		if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2838 			return 0;
2839 	} while (ptep++, addr = next, addr != end);
2840 
2841 	return 1;
2842 }
2843 #else
gup_huge_pd(hugepd_t hugepd,unsigned long addr,unsigned int pdshift,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2844 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2845 		unsigned int pdshift, unsigned long end, unsigned int flags,
2846 		struct page **pages, int *nr)
2847 {
2848 	return 0;
2849 }
2850 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2851 
gup_huge_pmd(pmd_t orig,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2852 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2853 			unsigned long end, unsigned int flags,
2854 			struct page **pages, int *nr)
2855 {
2856 	struct page *page;
2857 	struct folio *folio;
2858 	int refs;
2859 
2860 	if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2861 		return 0;
2862 
2863 	if (pmd_devmap(orig)) {
2864 		if (unlikely(flags & FOLL_LONGTERM))
2865 			return 0;
2866 		return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2867 					     pages, nr);
2868 	}
2869 
2870 	page = nth_page(pmd_page(orig), (addr & ~PMD_MASK) >> PAGE_SHIFT);
2871 	refs = record_subpages(page, addr, end, pages + *nr);
2872 
2873 	folio = try_grab_folio(page, refs, flags);
2874 	if (!folio)
2875 		return 0;
2876 
2877 	if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2878 		gup_put_folio(folio, refs, flags);
2879 		return 0;
2880 	}
2881 
2882 	if (!folio_fast_pin_allowed(folio, flags)) {
2883 		gup_put_folio(folio, refs, flags);
2884 		return 0;
2885 	}
2886 	if (!pmd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2887 		gup_put_folio(folio, refs, flags);
2888 		return 0;
2889 	}
2890 
2891 	*nr += refs;
2892 	folio_set_referenced(folio);
2893 	return 1;
2894 }
2895 
gup_huge_pud(pud_t orig,pud_t * pudp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2896 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2897 			unsigned long end, unsigned int flags,
2898 			struct page **pages, int *nr)
2899 {
2900 	struct page *page;
2901 	struct folio *folio;
2902 	int refs;
2903 
2904 	if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2905 		return 0;
2906 
2907 	if (pud_devmap(orig)) {
2908 		if (unlikely(flags & FOLL_LONGTERM))
2909 			return 0;
2910 		return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2911 					     pages, nr);
2912 	}
2913 
2914 	page = nth_page(pud_page(orig), (addr & ~PUD_MASK) >> PAGE_SHIFT);
2915 	refs = record_subpages(page, addr, end, pages + *nr);
2916 
2917 	folio = try_grab_folio(page, refs, flags);
2918 	if (!folio)
2919 		return 0;
2920 
2921 	if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2922 		gup_put_folio(folio, refs, flags);
2923 		return 0;
2924 	}
2925 
2926 	if (!folio_fast_pin_allowed(folio, flags)) {
2927 		gup_put_folio(folio, refs, flags);
2928 		return 0;
2929 	}
2930 
2931 	if (!pud_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2932 		gup_put_folio(folio, refs, flags);
2933 		return 0;
2934 	}
2935 
2936 	*nr += refs;
2937 	folio_set_referenced(folio);
2938 	return 1;
2939 }
2940 
gup_huge_pgd(pgd_t orig,pgd_t * pgdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2941 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2942 			unsigned long end, unsigned int flags,
2943 			struct page **pages, int *nr)
2944 {
2945 	int refs;
2946 	struct page *page;
2947 	struct folio *folio;
2948 
2949 	if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2950 		return 0;
2951 
2952 	BUILD_BUG_ON(pgd_devmap(orig));
2953 
2954 	page = nth_page(pgd_page(orig), (addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2955 	refs = record_subpages(page, addr, end, pages + *nr);
2956 
2957 	folio = try_grab_folio(page, refs, flags);
2958 	if (!folio)
2959 		return 0;
2960 
2961 	if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2962 		gup_put_folio(folio, refs, flags);
2963 		return 0;
2964 	}
2965 
2966 	if (!pgd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2967 		gup_put_folio(folio, refs, flags);
2968 		return 0;
2969 	}
2970 
2971 	if (!folio_fast_pin_allowed(folio, flags)) {
2972 		gup_put_folio(folio, refs, flags);
2973 		return 0;
2974 	}
2975 
2976 	*nr += refs;
2977 	folio_set_referenced(folio);
2978 	return 1;
2979 }
2980 
gup_pmd_range(pud_t * pudp,pud_t pud,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2981 static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
2982 		unsigned int flags, struct page **pages, int *nr)
2983 {
2984 	unsigned long next;
2985 	pmd_t *pmdp;
2986 
2987 	pmdp = pmd_offset_lockless(pudp, pud, addr);
2988 	do {
2989 		pmd_t pmd = pmdp_get_lockless(pmdp);
2990 
2991 		next = pmd_addr_end(addr, end);
2992 		if (!pmd_present(pmd))
2993 			return 0;
2994 
2995 		if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2996 			     pmd_devmap(pmd))) {
2997 			/* See gup_pte_range() */
2998 			if (pmd_protnone(pmd))
2999 				return 0;
3000 
3001 			if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
3002 				pages, nr))
3003 				return 0;
3004 
3005 		} else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
3006 			/*
3007 			 * architecture have different format for hugetlbfs
3008 			 * pmd format and THP pmd format
3009 			 */
3010 			if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
3011 					 PMD_SHIFT, next, flags, pages, nr))
3012 				return 0;
3013 		} else if (!gup_pte_range(pmd, pmdp, addr, next, flags, pages, nr))
3014 			return 0;
3015 	} while (pmdp++, addr = next, addr != end);
3016 
3017 	return 1;
3018 }
3019 
gup_pud_range(p4d_t * p4dp,p4d_t p4d,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3020 static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
3021 			 unsigned int flags, struct page **pages, int *nr)
3022 {
3023 	unsigned long next;
3024 	pud_t *pudp;
3025 
3026 	pudp = pud_offset_lockless(p4dp, p4d, addr);
3027 	do {
3028 		pud_t pud = READ_ONCE(*pudp);
3029 
3030 		next = pud_addr_end(addr, end);
3031 		if (unlikely(!pud_present(pud)))
3032 			return 0;
3033 		if (unlikely(pud_huge(pud) || pud_devmap(pud))) {
3034 			if (!gup_huge_pud(pud, pudp, addr, next, flags,
3035 					  pages, nr))
3036 				return 0;
3037 		} else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
3038 			if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
3039 					 PUD_SHIFT, next, flags, pages, nr))
3040 				return 0;
3041 		} else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
3042 			return 0;
3043 	} while (pudp++, addr = next, addr != end);
3044 
3045 	return 1;
3046 }
3047 
gup_p4d_range(pgd_t * pgdp,pgd_t pgd,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3048 static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
3049 			 unsigned int flags, struct page **pages, int *nr)
3050 {
3051 	unsigned long next;
3052 	p4d_t *p4dp;
3053 
3054 	p4dp = p4d_offset_lockless(pgdp, pgd, addr);
3055 	do {
3056 		p4d_t p4d = READ_ONCE(*p4dp);
3057 
3058 		next = p4d_addr_end(addr, end);
3059 		if (p4d_none(p4d))
3060 			return 0;
3061 		BUILD_BUG_ON(p4d_huge(p4d));
3062 		if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
3063 			if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
3064 					 P4D_SHIFT, next, flags, pages, nr))
3065 				return 0;
3066 		} else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
3067 			return 0;
3068 	} while (p4dp++, addr = next, addr != end);
3069 
3070 	return 1;
3071 }
3072 
gup_pgd_range(unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3073 static void gup_pgd_range(unsigned long addr, unsigned long end,
3074 		unsigned int flags, struct page **pages, int *nr)
3075 {
3076 	unsigned long next;
3077 	pgd_t *pgdp;
3078 
3079 	pgdp = pgd_offset(current->mm, addr);
3080 	do {
3081 		pgd_t pgd = READ_ONCE(*pgdp);
3082 
3083 		next = pgd_addr_end(addr, end);
3084 		if (pgd_none(pgd))
3085 			return;
3086 		if (unlikely(pgd_huge(pgd))) {
3087 			if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
3088 					  pages, nr))
3089 				return;
3090 		} else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
3091 			if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
3092 					 PGDIR_SHIFT, next, flags, pages, nr))
3093 				return;
3094 		} else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
3095 			return;
3096 	} while (pgdp++, addr = next, addr != end);
3097 }
3098 #else
gup_pgd_range(unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3099 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
3100 		unsigned int flags, struct page **pages, int *nr)
3101 {
3102 }
3103 #endif /* CONFIG_HAVE_FAST_GUP */
3104 
3105 #ifndef gup_fast_permitted
3106 /*
3107  * Check if it's allowed to use get_user_pages_fast_only() for the range, or
3108  * we need to fall back to the slow version:
3109  */
gup_fast_permitted(unsigned long start,unsigned long end)3110 static bool gup_fast_permitted(unsigned long start, unsigned long end)
3111 {
3112 	return true;
3113 }
3114 #endif
3115 
lockless_pages_from_mm(unsigned long start,unsigned long end,unsigned int gup_flags,struct page ** pages)3116 static unsigned long lockless_pages_from_mm(unsigned long start,
3117 					    unsigned long end,
3118 					    unsigned int gup_flags,
3119 					    struct page **pages)
3120 {
3121 	unsigned long flags;
3122 	int nr_pinned = 0;
3123 	unsigned seq;
3124 
3125 	if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP) ||
3126 	    !gup_fast_permitted(start, end))
3127 		return 0;
3128 
3129 	if (gup_flags & FOLL_PIN) {
3130 		seq = raw_read_seqcount(&current->mm->write_protect_seq);
3131 		if (seq & 1)
3132 			return 0;
3133 	}
3134 
3135 	/*
3136 	 * Disable interrupts. The nested form is used, in order to allow full,
3137 	 * general purpose use of this routine.
3138 	 *
3139 	 * With interrupts disabled, we block page table pages from being freed
3140 	 * from under us. See struct mmu_table_batch comments in
3141 	 * include/asm-generic/tlb.h for more details.
3142 	 *
3143 	 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
3144 	 * that come from THPs splitting.
3145 	 */
3146 	local_irq_save(flags);
3147 	gup_pgd_range(start, end, gup_flags, pages, &nr_pinned);
3148 	local_irq_restore(flags);
3149 
3150 	/*
3151 	 * When pinning pages for DMA there could be a concurrent write protect
3152 	 * from fork() via copy_page_range(), in this case always fail fast GUP.
3153 	 */
3154 	if (gup_flags & FOLL_PIN) {
3155 		if (read_seqcount_retry(&current->mm->write_protect_seq, seq)) {
3156 			unpin_user_pages_lockless(pages, nr_pinned);
3157 			return 0;
3158 		} else {
3159 			sanity_check_pinned_pages(pages, nr_pinned);
3160 		}
3161 	}
3162 	return nr_pinned;
3163 }
3164 
internal_get_user_pages_fast(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages)3165 static int internal_get_user_pages_fast(unsigned long start,
3166 					unsigned long nr_pages,
3167 					unsigned int gup_flags,
3168 					struct page **pages)
3169 {
3170 	unsigned long len, end;
3171 	unsigned long nr_pinned;
3172 	int locked = 0;
3173 	int ret;
3174 
3175 	if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
3176 				       FOLL_FORCE | FOLL_PIN | FOLL_GET |
3177 				       FOLL_FAST_ONLY | FOLL_NOFAULT |
3178 				       FOLL_PCI_P2PDMA | FOLL_HONOR_NUMA_FAULT)))
3179 		return -EINVAL;
3180 
3181 	if (gup_flags & FOLL_PIN)
3182 		mm_set_has_pinned_flag(&current->mm->flags);
3183 
3184 	if (!(gup_flags & FOLL_FAST_ONLY))
3185 		might_lock_read(&current->mm->mmap_lock);
3186 
3187 	start = untagged_addr(start) & PAGE_MASK;
3188 	len = nr_pages << PAGE_SHIFT;
3189 	if (check_add_overflow(start, len, &end))
3190 		return -EOVERFLOW;
3191 	if (end > TASK_SIZE_MAX)
3192 		return -EFAULT;
3193 	if (unlikely(!access_ok((void __user *)start, len)))
3194 		return -EFAULT;
3195 
3196 	nr_pinned = lockless_pages_from_mm(start, end, gup_flags, pages);
3197 	if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
3198 		return nr_pinned;
3199 
3200 	/* Slow path: try to get the remaining pages with get_user_pages */
3201 	start += nr_pinned << PAGE_SHIFT;
3202 	pages += nr_pinned;
3203 	ret = __gup_longterm_locked(current->mm, start, nr_pages - nr_pinned,
3204 				    pages, &locked,
3205 				    gup_flags | FOLL_TOUCH | FOLL_UNLOCKABLE);
3206 	if (ret < 0) {
3207 		/*
3208 		 * The caller has to unpin the pages we already pinned so
3209 		 * returning -errno is not an option
3210 		 */
3211 		if (nr_pinned)
3212 			return nr_pinned;
3213 		return ret;
3214 	}
3215 	return ret + nr_pinned;
3216 }
3217 
3218 /**
3219  * get_user_pages_fast_only() - pin user pages in memory
3220  * @start:      starting user address
3221  * @nr_pages:   number of pages from start to pin
3222  * @gup_flags:  flags modifying pin behaviour
3223  * @pages:      array that receives pointers to the pages pinned.
3224  *              Should be at least nr_pages long.
3225  *
3226  * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
3227  * the regular GUP.
3228  *
3229  * If the architecture does not support this function, simply return with no
3230  * pages pinned.
3231  *
3232  * Careful, careful! COW breaking can go either way, so a non-write
3233  * access can get ambiguous page results. If you call this function without
3234  * 'write' set, you'd better be sure that you're ok with that ambiguity.
3235  */
get_user_pages_fast_only(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)3236 int get_user_pages_fast_only(unsigned long start, int nr_pages,
3237 			     unsigned int gup_flags, struct page **pages)
3238 {
3239 	/*
3240 	 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
3241 	 * because gup fast is always a "pin with a +1 page refcount" request.
3242 	 *
3243 	 * FOLL_FAST_ONLY is required in order to match the API description of
3244 	 * this routine: no fall back to regular ("slow") GUP.
3245 	 */
3246 	if (!is_valid_gup_args(pages, NULL, &gup_flags,
3247 			       FOLL_GET | FOLL_FAST_ONLY))
3248 		return -EINVAL;
3249 
3250 	return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3251 }
3252 EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
3253 
3254 /**
3255  * get_user_pages_fast() - pin user pages in memory
3256  * @start:      starting user address
3257  * @nr_pages:   number of pages from start to pin
3258  * @gup_flags:  flags modifying pin behaviour
3259  * @pages:      array that receives pointers to the pages pinned.
3260  *              Should be at least nr_pages long.
3261  *
3262  * Attempt to pin user pages in memory without taking mm->mmap_lock.
3263  * If not successful, it will fall back to taking the lock and
3264  * calling get_user_pages().
3265  *
3266  * Returns number of pages pinned. This may be fewer than the number requested.
3267  * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
3268  * -errno.
3269  */
get_user_pages_fast(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)3270 int get_user_pages_fast(unsigned long start, int nr_pages,
3271 			unsigned int gup_flags, struct page **pages)
3272 {
3273 	/*
3274 	 * The caller may or may not have explicitly set FOLL_GET; either way is
3275 	 * OK. However, internally (within mm/gup.c), gup fast variants must set
3276 	 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
3277 	 * request.
3278 	 */
3279 	if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_GET))
3280 		return -EINVAL;
3281 	return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3282 }
3283 EXPORT_SYMBOL_GPL(get_user_pages_fast);
3284 
3285 /**
3286  * pin_user_pages_fast() - pin user pages in memory without taking locks
3287  *
3288  * @start:      starting user address
3289  * @nr_pages:   number of pages from start to pin
3290  * @gup_flags:  flags modifying pin behaviour
3291  * @pages:      array that receives pointers to the pages pinned.
3292  *              Should be at least nr_pages long.
3293  *
3294  * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
3295  * get_user_pages_fast() for documentation on the function arguments, because
3296  * the arguments here are identical.
3297  *
3298  * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3299  * see Documentation/core-api/pin_user_pages.rst for further details.
3300  *
3301  * Note that if a zero_page is amongst the returned pages, it will not have
3302  * pins in it and unpin_user_page() will not remove pins from it.
3303  */
pin_user_pages_fast(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)3304 int pin_user_pages_fast(unsigned long start, int nr_pages,
3305 			unsigned int gup_flags, struct page **pages)
3306 {
3307 	if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN))
3308 		return -EINVAL;
3309 	return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3310 }
3311 EXPORT_SYMBOL_GPL(pin_user_pages_fast);
3312 
3313 /**
3314  * pin_user_pages_remote() - pin pages of a remote process
3315  *
3316  * @mm:		mm_struct of target mm
3317  * @start:	starting user address
3318  * @nr_pages:	number of pages from start to pin
3319  * @gup_flags:	flags modifying lookup behaviour
3320  * @pages:	array that receives pointers to the pages pinned.
3321  *		Should be at least nr_pages long.
3322  * @locked:	pointer to lock flag indicating whether lock is held and
3323  *		subsequently whether VM_FAULT_RETRY functionality can be
3324  *		utilised. Lock must initially be held.
3325  *
3326  * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
3327  * get_user_pages_remote() for documentation on the function arguments, because
3328  * the arguments here are identical.
3329  *
3330  * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3331  * see Documentation/core-api/pin_user_pages.rst for details.
3332  *
3333  * Note that if a zero_page is amongst the returned pages, it will not have
3334  * pins in it and unpin_user_page*() will not remove pins from it.
3335  */
pin_user_pages_remote(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)3336 long pin_user_pages_remote(struct mm_struct *mm,
3337 			   unsigned long start, unsigned long nr_pages,
3338 			   unsigned int gup_flags, struct page **pages,
3339 			   int *locked)
3340 {
3341 	int local_locked = 1;
3342 
3343 	if (!is_valid_gup_args(pages, locked, &gup_flags,
3344 			       FOLL_PIN | FOLL_TOUCH | FOLL_REMOTE))
3345 		return 0;
3346 	return __gup_longterm_locked(mm, start, nr_pages, pages,
3347 				     locked ? locked : &local_locked,
3348 				     gup_flags);
3349 }
3350 EXPORT_SYMBOL(pin_user_pages_remote);
3351 
3352 /**
3353  * pin_user_pages() - pin user pages in memory for use by other devices
3354  *
3355  * @start:	starting user address
3356  * @nr_pages:	number of pages from start to pin
3357  * @gup_flags:	flags modifying lookup behaviour
3358  * @pages:	array that receives pointers to the pages pinned.
3359  *		Should be at least nr_pages long.
3360  *
3361  * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
3362  * FOLL_PIN is set.
3363  *
3364  * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3365  * see Documentation/core-api/pin_user_pages.rst for details.
3366  *
3367  * Note that if a zero_page is amongst the returned pages, it will not have
3368  * pins in it and unpin_user_page*() will not remove pins from it.
3369  */
pin_user_pages(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages)3370 long pin_user_pages(unsigned long start, unsigned long nr_pages,
3371 		    unsigned int gup_flags, struct page **pages)
3372 {
3373 	int locked = 1;
3374 
3375 	if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN))
3376 		return 0;
3377 	return __gup_longterm_locked(current->mm, start, nr_pages,
3378 				     pages, &locked, gup_flags);
3379 }
3380 EXPORT_SYMBOL(pin_user_pages);
3381 
3382 /*
3383  * pin_user_pages_unlocked() is the FOLL_PIN variant of
3384  * get_user_pages_unlocked(). Behavior is the same, except that this one sets
3385  * FOLL_PIN and rejects FOLL_GET.
3386  *
3387  * Note that if a zero_page is amongst the returned pages, it will not have
3388  * pins in it and unpin_user_page*() will not remove pins from it.
3389  */
pin_user_pages_unlocked(unsigned long start,unsigned long nr_pages,struct page ** pages,unsigned int gup_flags)3390 long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
3391 			     struct page **pages, unsigned int gup_flags)
3392 {
3393 	int locked = 0;
3394 
3395 	if (!is_valid_gup_args(pages, NULL, &gup_flags,
3396 			       FOLL_PIN | FOLL_TOUCH | FOLL_UNLOCKABLE))
3397 		return 0;
3398 
3399 	return __gup_longterm_locked(current->mm, start, nr_pages, pages,
3400 				     &locked, gup_flags);
3401 }
3402 EXPORT_SYMBOL(pin_user_pages_unlocked);
3403