xref: /linux/mm/gup.c (revision 2c1ed907520c50326b8f604907a8478b27881a2e)
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/memfd.h>
9 #include <linux/memremap.h>
10 #include <linux/pagemap.h>
11 #include <linux/rmap.h>
12 #include <linux/swap.h>
13 #include <linux/swapops.h>
14 #include <linux/secretmem.h>
15 
16 #include <linux/sched/signal.h>
17 #include <linux/rwsem.h>
18 #include <linux/hugetlb.h>
19 #include <linux/migrate.h>
20 #include <linux/mm_inline.h>
21 #include <linux/pagevec.h>
22 #include <linux/sched/mm.h>
23 #include <linux/shmem_fs.h>
24 
25 #include <asm/mmu_context.h>
26 #include <asm/tlbflush.h>
27 
28 #include "internal.h"
29 
30 struct follow_page_context {
31 	struct dev_pagemap *pgmap;
32 	unsigned int page_mask;
33 };
34 
sanity_check_pinned_pages(struct page ** pages,unsigned long npages)35 static inline void sanity_check_pinned_pages(struct page **pages,
36 					     unsigned long npages)
37 {
38 	if (!IS_ENABLED(CONFIG_DEBUG_VM))
39 		return;
40 
41 	/*
42 	 * We only pin anonymous pages if they are exclusive. Once pinned, we
43 	 * can no longer turn them possibly shared and PageAnonExclusive() will
44 	 * stick around until the page is freed.
45 	 *
46 	 * We'd like to verify that our pinned anonymous pages are still mapped
47 	 * exclusively. The issue with anon THP is that we don't know how
48 	 * they are/were mapped when pinning them. However, for anon
49 	 * THP we can assume that either the given page (PTE-mapped THP) or
50 	 * the head page (PMD-mapped THP) should be PageAnonExclusive(). If
51 	 * neither is the case, there is certainly something wrong.
52 	 */
53 	for (; npages; npages--, pages++) {
54 		struct page *page = *pages;
55 		struct folio *folio;
56 
57 		if (!page)
58 			continue;
59 
60 		folio = page_folio(page);
61 
62 		if (is_zero_page(page) ||
63 		    !folio_test_anon(folio))
64 			continue;
65 		if (!folio_test_large(folio) || folio_test_hugetlb(folio))
66 			VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page), page);
67 		else
68 			/* Either a PTE-mapped or a PMD-mapped THP. */
69 			VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page) &&
70 				       !PageAnonExclusive(page), page);
71 	}
72 }
73 
74 /*
75  * Return the folio with ref appropriately incremented,
76  * or NULL if that failed.
77  */
try_get_folio(struct page * page,int refs)78 static inline struct folio *try_get_folio(struct page *page, int refs)
79 {
80 	struct folio *folio;
81 
82 retry:
83 	folio = page_folio(page);
84 	if (WARN_ON_ONCE(folio_ref_count(folio) < 0))
85 		return NULL;
86 	if (unlikely(!folio_ref_try_add(folio, refs)))
87 		return NULL;
88 
89 	/*
90 	 * At this point we have a stable reference to the folio; but it
91 	 * could be that between calling page_folio() and the refcount
92 	 * increment, the folio was split, in which case we'd end up
93 	 * holding a reference on a folio that has nothing to do with the page
94 	 * we were given anymore.
95 	 * So now that the folio is stable, recheck that the page still
96 	 * belongs to this folio.
97 	 */
98 	if (unlikely(page_folio(page) != folio)) {
99 		if (!put_devmap_managed_folio_refs(folio, refs))
100 			folio_put_refs(folio, refs);
101 		goto retry;
102 	}
103 
104 	return folio;
105 }
106 
gup_put_folio(struct folio * folio,int refs,unsigned int flags)107 static void gup_put_folio(struct folio *folio, int refs, unsigned int flags)
108 {
109 	if (flags & FOLL_PIN) {
110 		if (is_zero_folio(folio))
111 			return;
112 		node_stat_mod_folio(folio, NR_FOLL_PIN_RELEASED, refs);
113 		if (folio_test_large(folio))
114 			atomic_sub(refs, &folio->_pincount);
115 		else
116 			refs *= GUP_PIN_COUNTING_BIAS;
117 	}
118 
119 	if (!put_devmap_managed_folio_refs(folio, refs))
120 		folio_put_refs(folio, refs);
121 }
122 
123 /**
124  * try_grab_folio() - add a folio's refcount by a flag-dependent amount
125  * @folio:    pointer to folio to be grabbed
126  * @refs:     the value to (effectively) add to the folio's refcount
127  * @flags:    gup flags: these are the FOLL_* flag values
128  *
129  * This might not do anything at all, depending on the flags argument.
130  *
131  * "grab" names in this file mean, "look at flags to decide whether to use
132  * FOLL_PIN or FOLL_GET behavior, when incrementing the folio's refcount.
133  *
134  * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
135  * time.
136  *
137  * Return: 0 for success, or if no action was required (if neither FOLL_PIN
138  * nor FOLL_GET was set, nothing is done). A negative error code for failure:
139  *
140  *   -ENOMEM		FOLL_GET or FOLL_PIN was set, but the folio could not
141  *			be grabbed.
142  *
143  * It is called when we have a stable reference for the folio, typically in
144  * GUP slow path.
145  */
try_grab_folio(struct folio * folio,int refs,unsigned int flags)146 int __must_check try_grab_folio(struct folio *folio, int refs,
147 				unsigned int flags)
148 {
149 	if (WARN_ON_ONCE(folio_ref_count(folio) <= 0))
150 		return -ENOMEM;
151 
152 	if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(&folio->page)))
153 		return -EREMOTEIO;
154 
155 	if (flags & FOLL_GET)
156 		folio_ref_add(folio, refs);
157 	else if (flags & FOLL_PIN) {
158 		/*
159 		 * Don't take a pin on the zero page - it's not going anywhere
160 		 * and it is used in a *lot* of places.
161 		 */
162 		if (is_zero_folio(folio))
163 			return 0;
164 
165 		/*
166 		 * Increment the normal page refcount field at least once,
167 		 * so that the page really is pinned.
168 		 */
169 		if (folio_test_large(folio)) {
170 			folio_ref_add(folio, refs);
171 			atomic_add(refs, &folio->_pincount);
172 		} else {
173 			folio_ref_add(folio, refs * GUP_PIN_COUNTING_BIAS);
174 		}
175 
176 		node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs);
177 	}
178 
179 	return 0;
180 }
181 
182 /**
183  * unpin_user_page() - release a dma-pinned page
184  * @page:            pointer to page to be released
185  *
186  * Pages that were pinned via pin_user_pages*() must be released via either
187  * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
188  * that such pages can be separately tracked and uniquely handled. In
189  * particular, interactions with RDMA and filesystems need special handling.
190  */
unpin_user_page(struct page * page)191 void unpin_user_page(struct page *page)
192 {
193 	sanity_check_pinned_pages(&page, 1);
194 	gup_put_folio(page_folio(page), 1, FOLL_PIN);
195 }
196 EXPORT_SYMBOL(unpin_user_page);
197 
198 /**
199  * unpin_folio() - release a dma-pinned folio
200  * @folio:         pointer to folio to be released
201  *
202  * Folios that were pinned via memfd_pin_folios() or other similar routines
203  * must be released either using unpin_folio() or unpin_folios().
204  */
unpin_folio(struct folio * folio)205 void unpin_folio(struct folio *folio)
206 {
207 	gup_put_folio(folio, 1, FOLL_PIN);
208 }
209 EXPORT_SYMBOL_GPL(unpin_folio);
210 
211 /**
212  * folio_add_pin - Try to get an additional pin on a pinned folio
213  * @folio: The folio to be pinned
214  *
215  * Get an additional pin on a folio we already have a pin on.  Makes no change
216  * if the folio is a zero_page.
217  */
folio_add_pin(struct folio * folio)218 void folio_add_pin(struct folio *folio)
219 {
220 	if (is_zero_folio(folio))
221 		return;
222 
223 	/*
224 	 * Similar to try_grab_folio(): be sure to *also* increment the normal
225 	 * page refcount field at least once, so that the page really is
226 	 * pinned.
227 	 */
228 	if (folio_test_large(folio)) {
229 		WARN_ON_ONCE(atomic_read(&folio->_pincount) < 1);
230 		folio_ref_inc(folio);
231 		atomic_inc(&folio->_pincount);
232 	} else {
233 		WARN_ON_ONCE(folio_ref_count(folio) < GUP_PIN_COUNTING_BIAS);
234 		folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
235 	}
236 }
237 
gup_folio_range_next(struct page * start,unsigned long npages,unsigned long i,unsigned int * ntails)238 static inline struct folio *gup_folio_range_next(struct page *start,
239 		unsigned long npages, unsigned long i, unsigned int *ntails)
240 {
241 	struct page *next = nth_page(start, i);
242 	struct folio *folio = page_folio(next);
243 	unsigned int nr = 1;
244 
245 	if (folio_test_large(folio))
246 		nr = min_t(unsigned int, npages - i,
247 			   folio_nr_pages(folio) - folio_page_idx(folio, next));
248 
249 	*ntails = nr;
250 	return folio;
251 }
252 
gup_folio_next(struct page ** list,unsigned long npages,unsigned long i,unsigned int * ntails)253 static inline struct folio *gup_folio_next(struct page **list,
254 		unsigned long npages, unsigned long i, unsigned int *ntails)
255 {
256 	struct folio *folio = page_folio(list[i]);
257 	unsigned int nr;
258 
259 	for (nr = i + 1; nr < npages; nr++) {
260 		if (page_folio(list[nr]) != folio)
261 			break;
262 	}
263 
264 	*ntails = nr - i;
265 	return folio;
266 }
267 
268 /**
269  * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
270  * @pages:  array of pages to be maybe marked dirty, and definitely released.
271  * @npages: number of pages in the @pages array.
272  * @make_dirty: whether to mark the pages dirty
273  *
274  * "gup-pinned page" refers to a page that has had one of the get_user_pages()
275  * variants called on that page.
276  *
277  * For each page in the @pages array, make that page (or its head page, if a
278  * compound page) dirty, if @make_dirty is true, and if the page was previously
279  * listed as clean. In any case, releases all pages using unpin_user_page(),
280  * possibly via unpin_user_pages(), for the non-dirty case.
281  *
282  * Please see the unpin_user_page() documentation for details.
283  *
284  * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
285  * required, then the caller should a) verify that this is really correct,
286  * because _lock() is usually required, and b) hand code it:
287  * set_page_dirty_lock(), unpin_user_page().
288  *
289  */
unpin_user_pages_dirty_lock(struct page ** pages,unsigned long npages,bool make_dirty)290 void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
291 				 bool make_dirty)
292 {
293 	unsigned long i;
294 	struct folio *folio;
295 	unsigned int nr;
296 
297 	if (!make_dirty) {
298 		unpin_user_pages(pages, npages);
299 		return;
300 	}
301 
302 	sanity_check_pinned_pages(pages, npages);
303 	for (i = 0; i < npages; i += nr) {
304 		folio = gup_folio_next(pages, npages, i, &nr);
305 		/*
306 		 * Checking PageDirty at this point may race with
307 		 * clear_page_dirty_for_io(), but that's OK. Two key
308 		 * cases:
309 		 *
310 		 * 1) This code sees the page as already dirty, so it
311 		 * skips the call to set_page_dirty(). That could happen
312 		 * because clear_page_dirty_for_io() called
313 		 * folio_mkclean(), followed by set_page_dirty().
314 		 * However, now the page is going to get written back,
315 		 * which meets the original intention of setting it
316 		 * dirty, so all is well: clear_page_dirty_for_io() goes
317 		 * on to call TestClearPageDirty(), and write the page
318 		 * back.
319 		 *
320 		 * 2) This code sees the page as clean, so it calls
321 		 * set_page_dirty(). The page stays dirty, despite being
322 		 * written back, so it gets written back again in the
323 		 * next writeback cycle. This is harmless.
324 		 */
325 		if (!folio_test_dirty(folio)) {
326 			folio_lock(folio);
327 			folio_mark_dirty(folio);
328 			folio_unlock(folio);
329 		}
330 		gup_put_folio(folio, nr, FOLL_PIN);
331 	}
332 }
333 EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
334 
335 /**
336  * unpin_user_page_range_dirty_lock() - release and optionally dirty
337  * gup-pinned page range
338  *
339  * @page:  the starting page of a range maybe marked dirty, and definitely released.
340  * @npages: number of consecutive pages to release.
341  * @make_dirty: whether to mark the pages dirty
342  *
343  * "gup-pinned page range" refers to a range of pages that has had one of the
344  * pin_user_pages() variants called on that page.
345  *
346  * For the page ranges defined by [page .. page+npages], make that range (or
347  * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
348  * page range was previously listed as clean.
349  *
350  * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
351  * required, then the caller should a) verify that this is really correct,
352  * because _lock() is usually required, and b) hand code it:
353  * set_page_dirty_lock(), unpin_user_page().
354  *
355  */
unpin_user_page_range_dirty_lock(struct page * page,unsigned long npages,bool make_dirty)356 void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages,
357 				      bool make_dirty)
358 {
359 	unsigned long i;
360 	struct folio *folio;
361 	unsigned int nr;
362 
363 	for (i = 0; i < npages; i += nr) {
364 		folio = gup_folio_range_next(page, npages, i, &nr);
365 		if (make_dirty && !folio_test_dirty(folio)) {
366 			folio_lock(folio);
367 			folio_mark_dirty(folio);
368 			folio_unlock(folio);
369 		}
370 		gup_put_folio(folio, nr, FOLL_PIN);
371 	}
372 }
373 EXPORT_SYMBOL(unpin_user_page_range_dirty_lock);
374 
gup_fast_unpin_user_pages(struct page ** pages,unsigned long npages)375 static void gup_fast_unpin_user_pages(struct page **pages, unsigned long npages)
376 {
377 	unsigned long i;
378 	struct folio *folio;
379 	unsigned int nr;
380 
381 	/*
382 	 * Don't perform any sanity checks because we might have raced with
383 	 * fork() and some anonymous pages might now actually be shared --
384 	 * which is why we're unpinning after all.
385 	 */
386 	for (i = 0; i < npages; i += nr) {
387 		folio = gup_folio_next(pages, npages, i, &nr);
388 		gup_put_folio(folio, nr, FOLL_PIN);
389 	}
390 }
391 
392 /**
393  * unpin_user_pages() - release an array of gup-pinned pages.
394  * @pages:  array of pages to be marked dirty and released.
395  * @npages: number of pages in the @pages array.
396  *
397  * For each page in the @pages array, release the page using unpin_user_page().
398  *
399  * Please see the unpin_user_page() documentation for details.
400  */
unpin_user_pages(struct page ** pages,unsigned long npages)401 void unpin_user_pages(struct page **pages, unsigned long npages)
402 {
403 	unsigned long i;
404 	struct folio *folio;
405 	unsigned int nr;
406 
407 	/*
408 	 * If this WARN_ON() fires, then the system *might* be leaking pages (by
409 	 * leaving them pinned), but probably not. More likely, gup/pup returned
410 	 * a hard -ERRNO error to the caller, who erroneously passed it here.
411 	 */
412 	if (WARN_ON(IS_ERR_VALUE(npages)))
413 		return;
414 
415 	sanity_check_pinned_pages(pages, npages);
416 	for (i = 0; i < npages; i += nr) {
417 		if (!pages[i]) {
418 			nr = 1;
419 			continue;
420 		}
421 		folio = gup_folio_next(pages, npages, i, &nr);
422 		gup_put_folio(folio, nr, FOLL_PIN);
423 	}
424 }
425 EXPORT_SYMBOL(unpin_user_pages);
426 
427 /**
428  * unpin_user_folio() - release pages of a folio
429  * @folio:  pointer to folio to be released
430  * @npages: number of pages of same folio
431  *
432  * Release npages of the folio
433  */
unpin_user_folio(struct folio * folio,unsigned long npages)434 void unpin_user_folio(struct folio *folio, unsigned long npages)
435 {
436 	gup_put_folio(folio, npages, FOLL_PIN);
437 }
438 EXPORT_SYMBOL(unpin_user_folio);
439 
440 /**
441  * unpin_folios() - release an array of gup-pinned folios.
442  * @folios:  array of folios to be marked dirty and released.
443  * @nfolios: number of folios in the @folios array.
444  *
445  * For each folio in the @folios array, release the folio using gup_put_folio.
446  *
447  * Please see the unpin_folio() documentation for details.
448  */
unpin_folios(struct folio ** folios,unsigned long nfolios)449 void unpin_folios(struct folio **folios, unsigned long nfolios)
450 {
451 	unsigned long i = 0, j;
452 
453 	/*
454 	 * If this WARN_ON() fires, then the system *might* be leaking folios
455 	 * (by leaving them pinned), but probably not. More likely, gup/pup
456 	 * returned a hard -ERRNO error to the caller, who erroneously passed
457 	 * it here.
458 	 */
459 	if (WARN_ON(IS_ERR_VALUE(nfolios)))
460 		return;
461 
462 	while (i < nfolios) {
463 		for (j = i + 1; j < nfolios; j++)
464 			if (folios[i] != folios[j])
465 				break;
466 
467 		if (folios[i])
468 			gup_put_folio(folios[i], j - i, FOLL_PIN);
469 		i = j;
470 	}
471 }
472 EXPORT_SYMBOL_GPL(unpin_folios);
473 
474 /*
475  * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's
476  * lifecycle.  Avoid setting the bit unless necessary, or it might cause write
477  * cache bouncing on large SMP machines for concurrent pinned gups.
478  */
mm_set_has_pinned_flag(unsigned long * mm_flags)479 static inline void mm_set_has_pinned_flag(unsigned long *mm_flags)
480 {
481 	if (!test_bit(MMF_HAS_PINNED, mm_flags))
482 		set_bit(MMF_HAS_PINNED, mm_flags);
483 }
484 
485 #ifdef CONFIG_MMU
486 
487 #ifdef CONFIG_HAVE_GUP_FAST
record_subpages(struct page * page,unsigned long sz,unsigned long addr,unsigned long end,struct page ** pages)488 static int record_subpages(struct page *page, unsigned long sz,
489 			   unsigned long addr, unsigned long end,
490 			   struct page **pages)
491 {
492 	struct page *start_page;
493 	int nr;
494 
495 	start_page = nth_page(page, (addr & (sz - 1)) >> PAGE_SHIFT);
496 	for (nr = 0; addr != end; nr++, addr += PAGE_SIZE)
497 		pages[nr] = nth_page(start_page, nr);
498 
499 	return nr;
500 }
501 
502 /**
503  * try_grab_folio_fast() - Attempt to get or pin a folio in fast path.
504  * @page:  pointer to page to be grabbed
505  * @refs:  the value to (effectively) add to the folio's refcount
506  * @flags: gup flags: these are the FOLL_* flag values.
507  *
508  * "grab" names in this file mean, "look at flags to decide whether to use
509  * FOLL_PIN or FOLL_GET behavior, when incrementing the folio's refcount.
510  *
511  * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
512  * same time. (That's true throughout the get_user_pages*() and
513  * pin_user_pages*() APIs.) Cases:
514  *
515  *    FOLL_GET: folio's refcount will be incremented by @refs.
516  *
517  *    FOLL_PIN on large folios: folio's refcount will be incremented by
518  *    @refs, and its pincount will be incremented by @refs.
519  *
520  *    FOLL_PIN on single-page folios: folio's refcount will be incremented by
521  *    @refs * GUP_PIN_COUNTING_BIAS.
522  *
523  * Return: The folio containing @page (with refcount appropriately
524  * incremented) for success, or NULL upon failure. If neither FOLL_GET
525  * nor FOLL_PIN was set, that's considered failure, and furthermore,
526  * a likely bug in the caller, so a warning is also emitted.
527  *
528  * It uses add ref unless zero to elevate the folio refcount and must be called
529  * in fast path only.
530  */
try_grab_folio_fast(struct page * page,int refs,unsigned int flags)531 static struct folio *try_grab_folio_fast(struct page *page, int refs,
532 					 unsigned int flags)
533 {
534 	struct folio *folio;
535 
536 	/* Raise warn if it is not called in fast GUP */
537 	VM_WARN_ON_ONCE(!irqs_disabled());
538 
539 	if (WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == 0))
540 		return NULL;
541 
542 	if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
543 		return NULL;
544 
545 	if (flags & FOLL_GET)
546 		return try_get_folio(page, refs);
547 
548 	/* FOLL_PIN is set */
549 
550 	/*
551 	 * Don't take a pin on the zero page - it's not going anywhere
552 	 * and it is used in a *lot* of places.
553 	 */
554 	if (is_zero_page(page))
555 		return page_folio(page);
556 
557 	folio = try_get_folio(page, refs);
558 	if (!folio)
559 		return NULL;
560 
561 	/*
562 	 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
563 	 * right zone, so fail and let the caller fall back to the slow
564 	 * path.
565 	 */
566 	if (unlikely((flags & FOLL_LONGTERM) &&
567 		     !folio_is_longterm_pinnable(folio))) {
568 		if (!put_devmap_managed_folio_refs(folio, refs))
569 			folio_put_refs(folio, refs);
570 		return NULL;
571 	}
572 
573 	/*
574 	 * When pinning a large folio, use an exact count to track it.
575 	 *
576 	 * However, be sure to *also* increment the normal folio
577 	 * refcount field at least once, so that the folio really
578 	 * is pinned.  That's why the refcount from the earlier
579 	 * try_get_folio() is left intact.
580 	 */
581 	if (folio_test_large(folio))
582 		atomic_add(refs, &folio->_pincount);
583 	else
584 		folio_ref_add(folio,
585 				refs * (GUP_PIN_COUNTING_BIAS - 1));
586 	/*
587 	 * Adjust the pincount before re-checking the PTE for changes.
588 	 * This is essentially a smp_mb() and is paired with a memory
589 	 * barrier in folio_try_share_anon_rmap_*().
590 	 */
591 	smp_mb__after_atomic();
592 
593 	node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs);
594 
595 	return folio;
596 }
597 #endif	/* CONFIG_HAVE_GUP_FAST */
598 
599 /* Common code for can_follow_write_* */
can_follow_write_common(struct page * page,struct vm_area_struct * vma,unsigned int flags)600 static inline bool can_follow_write_common(struct page *page,
601 		struct vm_area_struct *vma, unsigned int flags)
602 {
603 	/* Maybe FOLL_FORCE is set to override it? */
604 	if (!(flags & FOLL_FORCE))
605 		return false;
606 
607 	/* But FOLL_FORCE has no effect on shared mappings */
608 	if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
609 		return false;
610 
611 	/* ... or read-only private ones */
612 	if (!(vma->vm_flags & VM_MAYWRITE))
613 		return false;
614 
615 	/* ... or already writable ones that just need to take a write fault */
616 	if (vma->vm_flags & VM_WRITE)
617 		return false;
618 
619 	/*
620 	 * See can_change_pte_writable(): we broke COW and could map the page
621 	 * writable if we have an exclusive anonymous page ...
622 	 */
623 	return page && PageAnon(page) && PageAnonExclusive(page);
624 }
625 
no_page_table(struct vm_area_struct * vma,unsigned int flags,unsigned long address)626 static struct page *no_page_table(struct vm_area_struct *vma,
627 				  unsigned int flags, unsigned long address)
628 {
629 	if (!(flags & FOLL_DUMP))
630 		return NULL;
631 
632 	/*
633 	 * When core dumping, we don't want to allocate unnecessary pages or
634 	 * page tables.  Return error instead of NULL to skip handle_mm_fault,
635 	 * then get_dump_page() will return NULL to leave a hole in the dump.
636 	 * But we can only make this optimization where a hole would surely
637 	 * be zero-filled if handle_mm_fault() actually did handle it.
638 	 */
639 	if (is_vm_hugetlb_page(vma)) {
640 		struct hstate *h = hstate_vma(vma);
641 
642 		if (!hugetlbfs_pagecache_present(h, vma, address))
643 			return ERR_PTR(-EFAULT);
644 	} else if ((vma_is_anonymous(vma) || !vma->vm_ops->fault)) {
645 		return ERR_PTR(-EFAULT);
646 	}
647 
648 	return NULL;
649 }
650 
651 #ifdef CONFIG_PGTABLE_HAS_HUGE_LEAVES
652 /* FOLL_FORCE can write to even unwritable PUDs in COW mappings. */
can_follow_write_pud(pud_t pud,struct page * page,struct vm_area_struct * vma,unsigned int flags)653 static inline bool can_follow_write_pud(pud_t pud, struct page *page,
654 					struct vm_area_struct *vma,
655 					unsigned int flags)
656 {
657 	/* If the pud is writable, we can write to the page. */
658 	if (pud_write(pud))
659 		return true;
660 
661 	return can_follow_write_common(page, vma, flags);
662 }
663 
follow_huge_pud(struct vm_area_struct * vma,unsigned long addr,pud_t * pudp,int flags,struct follow_page_context * ctx)664 static struct page *follow_huge_pud(struct vm_area_struct *vma,
665 				    unsigned long addr, pud_t *pudp,
666 				    int flags, struct follow_page_context *ctx)
667 {
668 	struct mm_struct *mm = vma->vm_mm;
669 	struct page *page;
670 	pud_t pud = *pudp;
671 	unsigned long pfn = pud_pfn(pud);
672 	int ret;
673 
674 	assert_spin_locked(pud_lockptr(mm, pudp));
675 
676 	if (!pud_present(pud))
677 		return NULL;
678 
679 	if ((flags & FOLL_WRITE) &&
680 	    !can_follow_write_pud(pud, pfn_to_page(pfn), vma, flags))
681 		return NULL;
682 
683 	pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
684 
685 	if (IS_ENABLED(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) &&
686 	    pud_devmap(pud)) {
687 		/*
688 		 * device mapped pages can only be returned if the caller
689 		 * will manage the page reference count.
690 		 *
691 		 * At least one of FOLL_GET | FOLL_PIN must be set, so
692 		 * assert that here:
693 		 */
694 		if (!(flags & (FOLL_GET | FOLL_PIN)))
695 			return ERR_PTR(-EEXIST);
696 
697 		if (flags & FOLL_TOUCH)
698 			touch_pud(vma, addr, pudp, flags & FOLL_WRITE);
699 
700 		ctx->pgmap = get_dev_pagemap(pfn, ctx->pgmap);
701 		if (!ctx->pgmap)
702 			return ERR_PTR(-EFAULT);
703 	}
704 
705 	page = pfn_to_page(pfn);
706 
707 	if (!pud_devmap(pud) && !pud_write(pud) &&
708 	    gup_must_unshare(vma, flags, page))
709 		return ERR_PTR(-EMLINK);
710 
711 	ret = try_grab_folio(page_folio(page), 1, flags);
712 	if (ret)
713 		page = ERR_PTR(ret);
714 	else
715 		ctx->page_mask = HPAGE_PUD_NR - 1;
716 
717 	return page;
718 }
719 
720 /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
can_follow_write_pmd(pmd_t pmd,struct page * page,struct vm_area_struct * vma,unsigned int flags)721 static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page,
722 					struct vm_area_struct *vma,
723 					unsigned int flags)
724 {
725 	/* If the pmd is writable, we can write to the page. */
726 	if (pmd_write(pmd))
727 		return true;
728 
729 	if (!can_follow_write_common(page, vma, flags))
730 		return false;
731 
732 	/* ... and a write-fault isn't required for other reasons. */
733 	if (pmd_needs_soft_dirty_wp(vma, pmd))
734 		return false;
735 	return !userfaultfd_huge_pmd_wp(vma, pmd);
736 }
737 
follow_huge_pmd(struct vm_area_struct * vma,unsigned long addr,pmd_t * pmd,unsigned int flags,struct follow_page_context * ctx)738 static struct page *follow_huge_pmd(struct vm_area_struct *vma,
739 				    unsigned long addr, pmd_t *pmd,
740 				    unsigned int flags,
741 				    struct follow_page_context *ctx)
742 {
743 	struct mm_struct *mm = vma->vm_mm;
744 	pmd_t pmdval = *pmd;
745 	struct page *page;
746 	int ret;
747 
748 	assert_spin_locked(pmd_lockptr(mm, pmd));
749 
750 	page = pmd_page(pmdval);
751 	if ((flags & FOLL_WRITE) &&
752 	    !can_follow_write_pmd(pmdval, page, vma, flags))
753 		return NULL;
754 
755 	/* Avoid dumping huge zero page */
756 	if ((flags & FOLL_DUMP) && is_huge_zero_pmd(pmdval))
757 		return ERR_PTR(-EFAULT);
758 
759 	if (pmd_protnone(*pmd) && !gup_can_follow_protnone(vma, flags))
760 		return NULL;
761 
762 	if (!pmd_write(pmdval) && gup_must_unshare(vma, flags, page))
763 		return ERR_PTR(-EMLINK);
764 
765 	VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
766 			!PageAnonExclusive(page), page);
767 
768 	ret = try_grab_folio(page_folio(page), 1, flags);
769 	if (ret)
770 		return ERR_PTR(ret);
771 
772 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
773 	if (pmd_trans_huge(pmdval) && (flags & FOLL_TOUCH))
774 		touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
775 #endif	/* CONFIG_TRANSPARENT_HUGEPAGE */
776 
777 	page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
778 	ctx->page_mask = HPAGE_PMD_NR - 1;
779 
780 	return page;
781 }
782 
783 #else  /* CONFIG_PGTABLE_HAS_HUGE_LEAVES */
follow_huge_pud(struct vm_area_struct * vma,unsigned long addr,pud_t * pudp,int flags,struct follow_page_context * ctx)784 static struct page *follow_huge_pud(struct vm_area_struct *vma,
785 				    unsigned long addr, pud_t *pudp,
786 				    int flags, struct follow_page_context *ctx)
787 {
788 	return NULL;
789 }
790 
follow_huge_pmd(struct vm_area_struct * vma,unsigned long addr,pmd_t * pmd,unsigned int flags,struct follow_page_context * ctx)791 static struct page *follow_huge_pmd(struct vm_area_struct *vma,
792 				    unsigned long addr, pmd_t *pmd,
793 				    unsigned int flags,
794 				    struct follow_page_context *ctx)
795 {
796 	return NULL;
797 }
798 #endif	/* CONFIG_PGTABLE_HAS_HUGE_LEAVES */
799 
follow_pfn_pte(struct vm_area_struct * vma,unsigned long address,pte_t * pte,unsigned int flags)800 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
801 		pte_t *pte, unsigned int flags)
802 {
803 	if (flags & FOLL_TOUCH) {
804 		pte_t orig_entry = ptep_get(pte);
805 		pte_t entry = orig_entry;
806 
807 		if (flags & FOLL_WRITE)
808 			entry = pte_mkdirty(entry);
809 		entry = pte_mkyoung(entry);
810 
811 		if (!pte_same(orig_entry, entry)) {
812 			set_pte_at(vma->vm_mm, address, pte, entry);
813 			update_mmu_cache(vma, address, pte);
814 		}
815 	}
816 
817 	/* Proper page table entry exists, but no corresponding struct page */
818 	return -EEXIST;
819 }
820 
821 /* 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)822 static inline bool can_follow_write_pte(pte_t pte, struct page *page,
823 					struct vm_area_struct *vma,
824 					unsigned int flags)
825 {
826 	/* If the pte is writable, we can write to the page. */
827 	if (pte_write(pte))
828 		return true;
829 
830 	if (!can_follow_write_common(page, vma, flags))
831 		return false;
832 
833 	/* ... and a write-fault isn't required for other reasons. */
834 	if (pte_needs_soft_dirty_wp(vma, pte))
835 		return false;
836 	return !userfaultfd_pte_wp(vma, pte);
837 }
838 
follow_page_pte(struct vm_area_struct * vma,unsigned long address,pmd_t * pmd,unsigned int flags,struct dev_pagemap ** pgmap)839 static struct page *follow_page_pte(struct vm_area_struct *vma,
840 		unsigned long address, pmd_t *pmd, unsigned int flags,
841 		struct dev_pagemap **pgmap)
842 {
843 	struct mm_struct *mm = vma->vm_mm;
844 	struct folio *folio;
845 	struct page *page;
846 	spinlock_t *ptl;
847 	pte_t *ptep, pte;
848 	int ret;
849 
850 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
851 	if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
852 			 (FOLL_PIN | FOLL_GET)))
853 		return ERR_PTR(-EINVAL);
854 
855 	ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
856 	if (!ptep)
857 		return no_page_table(vma, flags, address);
858 	pte = ptep_get(ptep);
859 	if (!pte_present(pte))
860 		goto no_page;
861 	if (pte_protnone(pte) && !gup_can_follow_protnone(vma, flags))
862 		goto no_page;
863 
864 	page = vm_normal_page(vma, address, pte);
865 
866 	/*
867 	 * We only care about anon pages in can_follow_write_pte() and don't
868 	 * have to worry about pte_devmap() because they are never anon.
869 	 */
870 	if ((flags & FOLL_WRITE) &&
871 	    !can_follow_write_pte(pte, page, vma, flags)) {
872 		page = NULL;
873 		goto out;
874 	}
875 
876 	if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
877 		/*
878 		 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
879 		 * case since they are only valid while holding the pgmap
880 		 * reference.
881 		 */
882 		*pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
883 		if (*pgmap)
884 			page = pte_page(pte);
885 		else
886 			goto no_page;
887 	} else if (unlikely(!page)) {
888 		if (flags & FOLL_DUMP) {
889 			/* Avoid special (like zero) pages in core dumps */
890 			page = ERR_PTR(-EFAULT);
891 			goto out;
892 		}
893 
894 		if (is_zero_pfn(pte_pfn(pte))) {
895 			page = pte_page(pte);
896 		} else {
897 			ret = follow_pfn_pte(vma, address, ptep, flags);
898 			page = ERR_PTR(ret);
899 			goto out;
900 		}
901 	}
902 	folio = page_folio(page);
903 
904 	if (!pte_write(pte) && gup_must_unshare(vma, flags, page)) {
905 		page = ERR_PTR(-EMLINK);
906 		goto out;
907 	}
908 
909 	VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
910 		       !PageAnonExclusive(page), page);
911 
912 	/* try_grab_folio() does nothing unless FOLL_GET or FOLL_PIN is set. */
913 	ret = try_grab_folio(folio, 1, flags);
914 	if (unlikely(ret)) {
915 		page = ERR_PTR(ret);
916 		goto out;
917 	}
918 
919 	/*
920 	 * We need to make the page accessible if and only if we are going
921 	 * to access its content (the FOLL_PIN case).  Please see
922 	 * Documentation/core-api/pin_user_pages.rst for details.
923 	 */
924 	if (flags & FOLL_PIN) {
925 		ret = arch_make_folio_accessible(folio);
926 		if (ret) {
927 			unpin_user_page(page);
928 			page = ERR_PTR(ret);
929 			goto out;
930 		}
931 	}
932 	if (flags & FOLL_TOUCH) {
933 		if ((flags & FOLL_WRITE) &&
934 		    !pte_dirty(pte) && !folio_test_dirty(folio))
935 			folio_mark_dirty(folio);
936 		/*
937 		 * pte_mkyoung() would be more correct here, but atomic care
938 		 * is needed to avoid losing the dirty bit: it is easier to use
939 		 * folio_mark_accessed().
940 		 */
941 		folio_mark_accessed(folio);
942 	}
943 out:
944 	pte_unmap_unlock(ptep, ptl);
945 	return page;
946 no_page:
947 	pte_unmap_unlock(ptep, ptl);
948 	if (!pte_none(pte))
949 		return NULL;
950 	return no_page_table(vma, flags, address);
951 }
952 
follow_pmd_mask(struct vm_area_struct * vma,unsigned long address,pud_t * pudp,unsigned int flags,struct follow_page_context * ctx)953 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
954 				    unsigned long address, pud_t *pudp,
955 				    unsigned int flags,
956 				    struct follow_page_context *ctx)
957 {
958 	pmd_t *pmd, pmdval;
959 	spinlock_t *ptl;
960 	struct page *page;
961 	struct mm_struct *mm = vma->vm_mm;
962 
963 	pmd = pmd_offset(pudp, address);
964 	pmdval = pmdp_get_lockless(pmd);
965 	if (pmd_none(pmdval))
966 		return no_page_table(vma, flags, address);
967 	if (!pmd_present(pmdval))
968 		return no_page_table(vma, flags, address);
969 	if (pmd_devmap(pmdval)) {
970 		ptl = pmd_lock(mm, pmd);
971 		page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
972 		spin_unlock(ptl);
973 		if (page)
974 			return page;
975 		return no_page_table(vma, flags, address);
976 	}
977 	if (likely(!pmd_leaf(pmdval)))
978 		return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
979 
980 	if (pmd_protnone(pmdval) && !gup_can_follow_protnone(vma, flags))
981 		return no_page_table(vma, flags, address);
982 
983 	ptl = pmd_lock(mm, pmd);
984 	pmdval = *pmd;
985 	if (unlikely(!pmd_present(pmdval))) {
986 		spin_unlock(ptl);
987 		return no_page_table(vma, flags, address);
988 	}
989 	if (unlikely(!pmd_leaf(pmdval))) {
990 		spin_unlock(ptl);
991 		return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
992 	}
993 	if (pmd_trans_huge(pmdval) && (flags & FOLL_SPLIT_PMD)) {
994 		spin_unlock(ptl);
995 		split_huge_pmd(vma, pmd, address);
996 		/* If pmd was left empty, stuff a page table in there quickly */
997 		return pte_alloc(mm, pmd) ? ERR_PTR(-ENOMEM) :
998 			follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
999 	}
1000 	page = follow_huge_pmd(vma, address, pmd, flags, ctx);
1001 	spin_unlock(ptl);
1002 	return page;
1003 }
1004 
follow_pud_mask(struct vm_area_struct * vma,unsigned long address,p4d_t * p4dp,unsigned int flags,struct follow_page_context * ctx)1005 static struct page *follow_pud_mask(struct vm_area_struct *vma,
1006 				    unsigned long address, p4d_t *p4dp,
1007 				    unsigned int flags,
1008 				    struct follow_page_context *ctx)
1009 {
1010 	pud_t *pudp, pud;
1011 	spinlock_t *ptl;
1012 	struct page *page;
1013 	struct mm_struct *mm = vma->vm_mm;
1014 
1015 	pudp = pud_offset(p4dp, address);
1016 	pud = READ_ONCE(*pudp);
1017 	if (!pud_present(pud))
1018 		return no_page_table(vma, flags, address);
1019 	if (pud_leaf(pud)) {
1020 		ptl = pud_lock(mm, pudp);
1021 		page = follow_huge_pud(vma, address, pudp, flags, ctx);
1022 		spin_unlock(ptl);
1023 		if (page)
1024 			return page;
1025 		return no_page_table(vma, flags, address);
1026 	}
1027 	if (unlikely(pud_bad(pud)))
1028 		return no_page_table(vma, flags, address);
1029 
1030 	return follow_pmd_mask(vma, address, pudp, flags, ctx);
1031 }
1032 
follow_p4d_mask(struct vm_area_struct * vma,unsigned long address,pgd_t * pgdp,unsigned int flags,struct follow_page_context * ctx)1033 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
1034 				    unsigned long address, pgd_t *pgdp,
1035 				    unsigned int flags,
1036 				    struct follow_page_context *ctx)
1037 {
1038 	p4d_t *p4dp, p4d;
1039 
1040 	p4dp = p4d_offset(pgdp, address);
1041 	p4d = READ_ONCE(*p4dp);
1042 	BUILD_BUG_ON(p4d_leaf(p4d));
1043 
1044 	if (!p4d_present(p4d) || p4d_bad(p4d))
1045 		return no_page_table(vma, flags, address);
1046 
1047 	return follow_pud_mask(vma, address, p4dp, flags, ctx);
1048 }
1049 
1050 /**
1051  * follow_page_mask - look up a page descriptor from a user-virtual address
1052  * @vma: vm_area_struct mapping @address
1053  * @address: virtual address to look up
1054  * @flags: flags modifying lookup behaviour
1055  * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
1056  *       pointer to output page_mask
1057  *
1058  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1059  *
1060  * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
1061  * the device's dev_pagemap metadata to avoid repeating expensive lookups.
1062  *
1063  * When getting an anonymous page and the caller has to trigger unsharing
1064  * of a shared anonymous page first, -EMLINK is returned. The caller should
1065  * trigger a fault with FAULT_FLAG_UNSHARE set. Note that unsharing is only
1066  * relevant with FOLL_PIN and !FOLL_WRITE.
1067  *
1068  * On output, the @ctx->page_mask is set according to the size of the page.
1069  *
1070  * Return: the mapped (struct page *), %NULL if no mapping exists, or
1071  * an error pointer if there is a mapping to something not represented
1072  * by a page descriptor (see also vm_normal_page()).
1073  */
follow_page_mask(struct vm_area_struct * vma,unsigned long address,unsigned int flags,struct follow_page_context * ctx)1074 static struct page *follow_page_mask(struct vm_area_struct *vma,
1075 			      unsigned long address, unsigned int flags,
1076 			      struct follow_page_context *ctx)
1077 {
1078 	pgd_t *pgd;
1079 	struct mm_struct *mm = vma->vm_mm;
1080 	struct page *page;
1081 
1082 	vma_pgtable_walk_begin(vma);
1083 
1084 	ctx->page_mask = 0;
1085 	pgd = pgd_offset(mm, address);
1086 
1087 	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1088 		page = no_page_table(vma, flags, address);
1089 	else
1090 		page = follow_p4d_mask(vma, address, pgd, flags, ctx);
1091 
1092 	vma_pgtable_walk_end(vma);
1093 
1094 	return page;
1095 }
1096 
get_gate_page(struct mm_struct * mm,unsigned long address,unsigned int gup_flags,struct vm_area_struct ** vma,struct page ** page)1097 static int get_gate_page(struct mm_struct *mm, unsigned long address,
1098 		unsigned int gup_flags, struct vm_area_struct **vma,
1099 		struct page **page)
1100 {
1101 	pgd_t *pgd;
1102 	p4d_t *p4d;
1103 	pud_t *pud;
1104 	pmd_t *pmd;
1105 	pte_t *pte;
1106 	pte_t entry;
1107 	int ret = -EFAULT;
1108 
1109 	/* user gate pages are read-only */
1110 	if (gup_flags & FOLL_WRITE)
1111 		return -EFAULT;
1112 	if (address > TASK_SIZE)
1113 		pgd = pgd_offset_k(address);
1114 	else
1115 		pgd = pgd_offset_gate(mm, address);
1116 	if (pgd_none(*pgd))
1117 		return -EFAULT;
1118 	p4d = p4d_offset(pgd, address);
1119 	if (p4d_none(*p4d))
1120 		return -EFAULT;
1121 	pud = pud_offset(p4d, address);
1122 	if (pud_none(*pud))
1123 		return -EFAULT;
1124 	pmd = pmd_offset(pud, address);
1125 	if (!pmd_present(*pmd))
1126 		return -EFAULT;
1127 	pte = pte_offset_map(pmd, address);
1128 	if (!pte)
1129 		return -EFAULT;
1130 	entry = ptep_get(pte);
1131 	if (pte_none(entry))
1132 		goto unmap;
1133 	*vma = get_gate_vma(mm);
1134 	if (!page)
1135 		goto out;
1136 	*page = vm_normal_page(*vma, address, entry);
1137 	if (!*page) {
1138 		if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(entry)))
1139 			goto unmap;
1140 		*page = pte_page(entry);
1141 	}
1142 	ret = try_grab_folio(page_folio(*page), 1, gup_flags);
1143 	if (unlikely(ret))
1144 		goto unmap;
1145 out:
1146 	ret = 0;
1147 unmap:
1148 	pte_unmap(pte);
1149 	return ret;
1150 }
1151 
1152 /*
1153  * mmap_lock must be held on entry.  If @flags has FOLL_UNLOCKABLE but not
1154  * FOLL_NOWAIT, the mmap_lock may be released.  If it is, *@locked will be set
1155  * to 0 and -EBUSY returned.
1156  */
faultin_page(struct vm_area_struct * vma,unsigned long address,unsigned int flags,bool unshare,int * locked)1157 static int faultin_page(struct vm_area_struct *vma,
1158 		unsigned long address, unsigned int flags, bool unshare,
1159 		int *locked)
1160 {
1161 	unsigned int fault_flags = 0;
1162 	vm_fault_t ret;
1163 
1164 	if (flags & FOLL_NOFAULT)
1165 		return -EFAULT;
1166 	if (flags & FOLL_WRITE)
1167 		fault_flags |= FAULT_FLAG_WRITE;
1168 	if (flags & FOLL_REMOTE)
1169 		fault_flags |= FAULT_FLAG_REMOTE;
1170 	if (flags & FOLL_UNLOCKABLE) {
1171 		fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1172 		/*
1173 		 * FAULT_FLAG_INTERRUPTIBLE is opt-in. GUP callers must set
1174 		 * FOLL_INTERRUPTIBLE to enable FAULT_FLAG_INTERRUPTIBLE.
1175 		 * That's because some callers may not be prepared to
1176 		 * handle early exits caused by non-fatal signals.
1177 		 */
1178 		if (flags & FOLL_INTERRUPTIBLE)
1179 			fault_flags |= FAULT_FLAG_INTERRUPTIBLE;
1180 	}
1181 	if (flags & FOLL_NOWAIT)
1182 		fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
1183 	if (flags & FOLL_TRIED) {
1184 		/*
1185 		 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
1186 		 * can co-exist
1187 		 */
1188 		fault_flags |= FAULT_FLAG_TRIED;
1189 	}
1190 	if (unshare) {
1191 		fault_flags |= FAULT_FLAG_UNSHARE;
1192 		/* FAULT_FLAG_WRITE and FAULT_FLAG_UNSHARE are incompatible */
1193 		VM_BUG_ON(fault_flags & FAULT_FLAG_WRITE);
1194 	}
1195 
1196 	ret = handle_mm_fault(vma, address, fault_flags, NULL);
1197 
1198 	if (ret & VM_FAULT_COMPLETED) {
1199 		/*
1200 		 * With FAULT_FLAG_RETRY_NOWAIT we'll never release the
1201 		 * mmap lock in the page fault handler. Sanity check this.
1202 		 */
1203 		WARN_ON_ONCE(fault_flags & FAULT_FLAG_RETRY_NOWAIT);
1204 		*locked = 0;
1205 
1206 		/*
1207 		 * We should do the same as VM_FAULT_RETRY, but let's not
1208 		 * return -EBUSY since that's not reflecting the reality of
1209 		 * what has happened - we've just fully completed a page
1210 		 * fault, with the mmap lock released.  Use -EAGAIN to show
1211 		 * that we want to take the mmap lock _again_.
1212 		 */
1213 		return -EAGAIN;
1214 	}
1215 
1216 	if (ret & VM_FAULT_ERROR) {
1217 		int err = vm_fault_to_errno(ret, flags);
1218 
1219 		if (err)
1220 			return err;
1221 		BUG();
1222 	}
1223 
1224 	if (ret & VM_FAULT_RETRY) {
1225 		if (!(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
1226 			*locked = 0;
1227 		return -EBUSY;
1228 	}
1229 
1230 	return 0;
1231 }
1232 
1233 /*
1234  * Writing to file-backed mappings which require folio dirty tracking using GUP
1235  * is a fundamentally broken operation, as kernel write access to GUP mappings
1236  * do not adhere to the semantics expected by a file system.
1237  *
1238  * Consider the following scenario:-
1239  *
1240  * 1. A folio is written to via GUP which write-faults the memory, notifying
1241  *    the file system and dirtying the folio.
1242  * 2. Later, writeback is triggered, resulting in the folio being cleaned and
1243  *    the PTE being marked read-only.
1244  * 3. The GUP caller writes to the folio, as it is mapped read/write via the
1245  *    direct mapping.
1246  * 4. The GUP caller, now done with the page, unpins it and sets it dirty
1247  *    (though it does not have to).
1248  *
1249  * This results in both data being written to a folio without writenotify, and
1250  * the folio being dirtied unexpectedly (if the caller decides to do so).
1251  */
writable_file_mapping_allowed(struct vm_area_struct * vma,unsigned long gup_flags)1252 static bool writable_file_mapping_allowed(struct vm_area_struct *vma,
1253 					  unsigned long gup_flags)
1254 {
1255 	/*
1256 	 * If we aren't pinning then no problematic write can occur. A long term
1257 	 * pin is the most egregious case so this is the case we disallow.
1258 	 */
1259 	if ((gup_flags & (FOLL_PIN | FOLL_LONGTERM)) !=
1260 	    (FOLL_PIN | FOLL_LONGTERM))
1261 		return true;
1262 
1263 	/*
1264 	 * If the VMA does not require dirty tracking then no problematic write
1265 	 * can occur either.
1266 	 */
1267 	return !vma_needs_dirty_tracking(vma);
1268 }
1269 
check_vma_flags(struct vm_area_struct * vma,unsigned long gup_flags)1270 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
1271 {
1272 	vm_flags_t vm_flags = vma->vm_flags;
1273 	int write = (gup_flags & FOLL_WRITE);
1274 	int foreign = (gup_flags & FOLL_REMOTE);
1275 	bool vma_anon = vma_is_anonymous(vma);
1276 
1277 	if (vm_flags & (VM_IO | VM_PFNMAP))
1278 		return -EFAULT;
1279 
1280 	if ((gup_flags & FOLL_ANON) && !vma_anon)
1281 		return -EFAULT;
1282 
1283 	if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma))
1284 		return -EOPNOTSUPP;
1285 
1286 	if (vma_is_secretmem(vma))
1287 		return -EFAULT;
1288 
1289 	if (write) {
1290 		if (!vma_anon &&
1291 		    !writable_file_mapping_allowed(vma, gup_flags))
1292 			return -EFAULT;
1293 
1294 		if (!(vm_flags & VM_WRITE) || (vm_flags & VM_SHADOW_STACK)) {
1295 			if (!(gup_flags & FOLL_FORCE))
1296 				return -EFAULT;
1297 			/*
1298 			 * We used to let the write,force case do COW in a
1299 			 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
1300 			 * set a breakpoint in a read-only mapping of an
1301 			 * executable, without corrupting the file (yet only
1302 			 * when that file had been opened for writing!).
1303 			 * Anon pages in shared mappings are surprising: now
1304 			 * just reject it.
1305 			 */
1306 			if (!is_cow_mapping(vm_flags))
1307 				return -EFAULT;
1308 		}
1309 	} else if (!(vm_flags & VM_READ)) {
1310 		if (!(gup_flags & FOLL_FORCE))
1311 			return -EFAULT;
1312 		/*
1313 		 * Is there actually any vma we can reach here which does not
1314 		 * have VM_MAYREAD set?
1315 		 */
1316 		if (!(vm_flags & VM_MAYREAD))
1317 			return -EFAULT;
1318 	}
1319 	/*
1320 	 * gups are always data accesses, not instruction
1321 	 * fetches, so execute=false here
1322 	 */
1323 	if (!arch_vma_access_permitted(vma, write, false, foreign))
1324 		return -EFAULT;
1325 	return 0;
1326 }
1327 
1328 /*
1329  * This is "vma_lookup()", but with a warning if we would have
1330  * historically expanded the stack in the GUP code.
1331  */
gup_vma_lookup(struct mm_struct * mm,unsigned long addr)1332 static struct vm_area_struct *gup_vma_lookup(struct mm_struct *mm,
1333 	 unsigned long addr)
1334 {
1335 #ifdef CONFIG_STACK_GROWSUP
1336 	return vma_lookup(mm, addr);
1337 #else
1338 	static volatile unsigned long next_warn;
1339 	struct vm_area_struct *vma;
1340 	unsigned long now, next;
1341 
1342 	vma = find_vma(mm, addr);
1343 	if (!vma || (addr >= vma->vm_start))
1344 		return vma;
1345 
1346 	/* Only warn for half-way relevant accesses */
1347 	if (!(vma->vm_flags & VM_GROWSDOWN))
1348 		return NULL;
1349 	if (vma->vm_start - addr > 65536)
1350 		return NULL;
1351 
1352 	/* Let's not warn more than once an hour.. */
1353 	now = jiffies; next = next_warn;
1354 	if (next && time_before(now, next))
1355 		return NULL;
1356 	next_warn = now + 60*60*HZ;
1357 
1358 	/* Let people know things may have changed. */
1359 	pr_warn("GUP no longer grows the stack in %s (%d): %lx-%lx (%lx)\n",
1360 		current->comm, task_pid_nr(current),
1361 		vma->vm_start, vma->vm_end, addr);
1362 	dump_stack();
1363 	return NULL;
1364 #endif
1365 }
1366 
1367 /**
1368  * __get_user_pages() - pin user pages in memory
1369  * @mm:		mm_struct of target mm
1370  * @start:	starting user address
1371  * @nr_pages:	number of pages from start to pin
1372  * @gup_flags:	flags modifying pin behaviour
1373  * @pages:	array that receives pointers to the pages pinned.
1374  *		Should be at least nr_pages long. Or NULL, if caller
1375  *		only intends to ensure the pages are faulted in.
1376  * @locked:     whether we're still with the mmap_lock held
1377  *
1378  * Returns either number of pages pinned (which may be less than the
1379  * number requested), or an error. Details about the return value:
1380  *
1381  * -- If nr_pages is 0, returns 0.
1382  * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1383  * -- If nr_pages is >0, and some pages were pinned, returns the number of
1384  *    pages pinned. Again, this may be less than nr_pages.
1385  * -- 0 return value is possible when the fault would need to be retried.
1386  *
1387  * The caller is responsible for releasing returned @pages, via put_page().
1388  *
1389  * Must be called with mmap_lock held.  It may be released.  See below.
1390  *
1391  * __get_user_pages walks a process's page tables and takes a reference to
1392  * each struct page that each user address corresponds to at a given
1393  * instant. That is, it takes the page that would be accessed if a user
1394  * thread accesses the given user virtual address at that instant.
1395  *
1396  * This does not guarantee that the page exists in the user mappings when
1397  * __get_user_pages returns, and there may even be a completely different
1398  * page there in some cases (eg. if mmapped pagecache has been invalidated
1399  * and subsequently re-faulted). However it does guarantee that the page
1400  * won't be freed completely. And mostly callers simply care that the page
1401  * contains data that was valid *at some point in time*. Typically, an IO
1402  * or similar operation cannot guarantee anything stronger anyway because
1403  * locks can't be held over the syscall boundary.
1404  *
1405  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1406  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1407  * appropriate) must be called after the page is finished with, and
1408  * before put_page is called.
1409  *
1410  * If FOLL_UNLOCKABLE is set without FOLL_NOWAIT then the mmap_lock may
1411  * be released. If this happens *@locked will be set to 0 on return.
1412  *
1413  * A caller using such a combination of @gup_flags must therefore hold the
1414  * mmap_lock for reading only, and recognize when it's been released. Otherwise,
1415  * it must be held for either reading or writing and will not be released.
1416  *
1417  * In most cases, get_user_pages or get_user_pages_fast should be used
1418  * instead of __get_user_pages. __get_user_pages should be used only if
1419  * you need some special @gup_flags.
1420  */
__get_user_pages(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)1421 static long __get_user_pages(struct mm_struct *mm,
1422 		unsigned long start, unsigned long nr_pages,
1423 		unsigned int gup_flags, struct page **pages,
1424 		int *locked)
1425 {
1426 	long ret = 0, i = 0;
1427 	struct vm_area_struct *vma = NULL;
1428 	struct follow_page_context ctx = { NULL };
1429 
1430 	if (!nr_pages)
1431 		return 0;
1432 
1433 	start = untagged_addr_remote(mm, start);
1434 
1435 	VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1436 
1437 	do {
1438 		struct page *page;
1439 		unsigned int page_increm;
1440 
1441 		/* first iteration or cross vma bound */
1442 		if (!vma || start >= vma->vm_end) {
1443 			/*
1444 			 * MADV_POPULATE_(READ|WRITE) wants to handle VMA
1445 			 * lookups+error reporting differently.
1446 			 */
1447 			if (gup_flags & FOLL_MADV_POPULATE) {
1448 				vma = vma_lookup(mm, start);
1449 				if (!vma) {
1450 					ret = -ENOMEM;
1451 					goto out;
1452 				}
1453 				if (check_vma_flags(vma, gup_flags)) {
1454 					ret = -EINVAL;
1455 					goto out;
1456 				}
1457 				goto retry;
1458 			}
1459 			vma = gup_vma_lookup(mm, start);
1460 			if (!vma && in_gate_area(mm, start)) {
1461 				ret = get_gate_page(mm, start & PAGE_MASK,
1462 						gup_flags, &vma,
1463 						pages ? &page : NULL);
1464 				if (ret)
1465 					goto out;
1466 				ctx.page_mask = 0;
1467 				goto next_page;
1468 			}
1469 
1470 			if (!vma) {
1471 				ret = -EFAULT;
1472 				goto out;
1473 			}
1474 			ret = check_vma_flags(vma, gup_flags);
1475 			if (ret)
1476 				goto out;
1477 		}
1478 retry:
1479 		/*
1480 		 * If we have a pending SIGKILL, don't keep faulting pages and
1481 		 * potentially allocating memory.
1482 		 */
1483 		if (fatal_signal_pending(current)) {
1484 			ret = -EINTR;
1485 			goto out;
1486 		}
1487 		cond_resched();
1488 
1489 		page = follow_page_mask(vma, start, gup_flags, &ctx);
1490 		if (!page || PTR_ERR(page) == -EMLINK) {
1491 			ret = faultin_page(vma, start, gup_flags,
1492 					   PTR_ERR(page) == -EMLINK, locked);
1493 			switch (ret) {
1494 			case 0:
1495 				goto retry;
1496 			case -EBUSY:
1497 			case -EAGAIN:
1498 				ret = 0;
1499 				fallthrough;
1500 			case -EFAULT:
1501 			case -ENOMEM:
1502 			case -EHWPOISON:
1503 				goto out;
1504 			}
1505 			BUG();
1506 		} else if (PTR_ERR(page) == -EEXIST) {
1507 			/*
1508 			 * Proper page table entry exists, but no corresponding
1509 			 * struct page. If the caller expects **pages to be
1510 			 * filled in, bail out now, because that can't be done
1511 			 * for this page.
1512 			 */
1513 			if (pages) {
1514 				ret = PTR_ERR(page);
1515 				goto out;
1516 			}
1517 		} else if (IS_ERR(page)) {
1518 			ret = PTR_ERR(page);
1519 			goto out;
1520 		}
1521 next_page:
1522 		page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1523 		if (page_increm > nr_pages)
1524 			page_increm = nr_pages;
1525 
1526 		if (pages) {
1527 			struct page *subpage;
1528 			unsigned int j;
1529 
1530 			/*
1531 			 * This must be a large folio (and doesn't need to
1532 			 * be the whole folio; it can be part of it), do
1533 			 * the refcount work for all the subpages too.
1534 			 *
1535 			 * NOTE: here the page may not be the head page
1536 			 * e.g. when start addr is not thp-size aligned.
1537 			 * try_grab_folio() should have taken care of tail
1538 			 * pages.
1539 			 */
1540 			if (page_increm > 1) {
1541 				struct folio *folio = page_folio(page);
1542 
1543 				/*
1544 				 * Since we already hold refcount on the
1545 				 * large folio, this should never fail.
1546 				 */
1547 				if (try_grab_folio(folio, page_increm - 1,
1548 						   gup_flags)) {
1549 					/*
1550 					 * Release the 1st page ref if the
1551 					 * folio is problematic, fail hard.
1552 					 */
1553 					gup_put_folio(folio, 1, gup_flags);
1554 					ret = -EFAULT;
1555 					goto out;
1556 				}
1557 			}
1558 
1559 			for (j = 0; j < page_increm; j++) {
1560 				subpage = nth_page(page, j);
1561 				pages[i + j] = subpage;
1562 				flush_anon_page(vma, subpage, start + j * PAGE_SIZE);
1563 				flush_dcache_page(subpage);
1564 			}
1565 		}
1566 
1567 		i += page_increm;
1568 		start += page_increm * PAGE_SIZE;
1569 		nr_pages -= page_increm;
1570 	} while (nr_pages);
1571 out:
1572 	if (ctx.pgmap)
1573 		put_dev_pagemap(ctx.pgmap);
1574 	return i ? i : ret;
1575 }
1576 
vma_permits_fault(struct vm_area_struct * vma,unsigned int fault_flags)1577 static bool vma_permits_fault(struct vm_area_struct *vma,
1578 			      unsigned int fault_flags)
1579 {
1580 	bool write   = !!(fault_flags & FAULT_FLAG_WRITE);
1581 	bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1582 	vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1583 
1584 	if (!(vm_flags & vma->vm_flags))
1585 		return false;
1586 
1587 	/*
1588 	 * The architecture might have a hardware protection
1589 	 * mechanism other than read/write that can deny access.
1590 	 *
1591 	 * gup always represents data access, not instruction
1592 	 * fetches, so execute=false here:
1593 	 */
1594 	if (!arch_vma_access_permitted(vma, write, false, foreign))
1595 		return false;
1596 
1597 	return true;
1598 }
1599 
1600 /**
1601  * fixup_user_fault() - manually resolve a user page fault
1602  * @mm:		mm_struct of target mm
1603  * @address:	user address
1604  * @fault_flags:flags to pass down to handle_mm_fault()
1605  * @unlocked:	did we unlock the mmap_lock while retrying, maybe NULL if caller
1606  *		does not allow retry. If NULL, the caller must guarantee
1607  *		that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1608  *
1609  * This is meant to be called in the specific scenario where for locking reasons
1610  * we try to access user memory in atomic context (within a pagefault_disable()
1611  * section), this returns -EFAULT, and we want to resolve the user fault before
1612  * trying again.
1613  *
1614  * Typically this is meant to be used by the futex code.
1615  *
1616  * The main difference with get_user_pages() is that this function will
1617  * unconditionally call handle_mm_fault() which will in turn perform all the
1618  * necessary SW fixup of the dirty and young bits in the PTE, while
1619  * get_user_pages() only guarantees to update these in the struct page.
1620  *
1621  * This is important for some architectures where those bits also gate the
1622  * access permission to the page because they are maintained in software.  On
1623  * such architectures, gup() will not be enough to make a subsequent access
1624  * succeed.
1625  *
1626  * This function will not return with an unlocked mmap_lock. So it has not the
1627  * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1628  */
fixup_user_fault(struct mm_struct * mm,unsigned long address,unsigned int fault_flags,bool * unlocked)1629 int fixup_user_fault(struct mm_struct *mm,
1630 		     unsigned long address, unsigned int fault_flags,
1631 		     bool *unlocked)
1632 {
1633 	struct vm_area_struct *vma;
1634 	vm_fault_t ret;
1635 
1636 	address = untagged_addr_remote(mm, address);
1637 
1638 	if (unlocked)
1639 		fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1640 
1641 retry:
1642 	vma = gup_vma_lookup(mm, address);
1643 	if (!vma)
1644 		return -EFAULT;
1645 
1646 	if (!vma_permits_fault(vma, fault_flags))
1647 		return -EFAULT;
1648 
1649 	if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1650 	    fatal_signal_pending(current))
1651 		return -EINTR;
1652 
1653 	ret = handle_mm_fault(vma, address, fault_flags, NULL);
1654 
1655 	if (ret & VM_FAULT_COMPLETED) {
1656 		/*
1657 		 * NOTE: it's a pity that we need to retake the lock here
1658 		 * to pair with the unlock() in the callers. Ideally we
1659 		 * could tell the callers so they do not need to unlock.
1660 		 */
1661 		mmap_read_lock(mm);
1662 		*unlocked = true;
1663 		return 0;
1664 	}
1665 
1666 	if (ret & VM_FAULT_ERROR) {
1667 		int err = vm_fault_to_errno(ret, 0);
1668 
1669 		if (err)
1670 			return err;
1671 		BUG();
1672 	}
1673 
1674 	if (ret & VM_FAULT_RETRY) {
1675 		mmap_read_lock(mm);
1676 		*unlocked = true;
1677 		fault_flags |= FAULT_FLAG_TRIED;
1678 		goto retry;
1679 	}
1680 
1681 	return 0;
1682 }
1683 EXPORT_SYMBOL_GPL(fixup_user_fault);
1684 
1685 /*
1686  * GUP always responds to fatal signals.  When FOLL_INTERRUPTIBLE is
1687  * specified, it'll also respond to generic signals.  The caller of GUP
1688  * that has FOLL_INTERRUPTIBLE should take care of the GUP interruption.
1689  */
gup_signal_pending(unsigned int flags)1690 static bool gup_signal_pending(unsigned int flags)
1691 {
1692 	if (fatal_signal_pending(current))
1693 		return true;
1694 
1695 	if (!(flags & FOLL_INTERRUPTIBLE))
1696 		return false;
1697 
1698 	return signal_pending(current);
1699 }
1700 
1701 /*
1702  * Locking: (*locked == 1) means that the mmap_lock has already been acquired by
1703  * the caller. This function may drop the mmap_lock. If it does so, then it will
1704  * set (*locked = 0).
1705  *
1706  * (*locked == 0) means that the caller expects this function to acquire and
1707  * drop the mmap_lock. Therefore, the value of *locked will still be zero when
1708  * the function returns, even though it may have changed temporarily during
1709  * function execution.
1710  *
1711  * Please note that this function, unlike __get_user_pages(), will not return 0
1712  * for nr_pages > 0, unless FOLL_NOWAIT is used.
1713  */
__get_user_pages_locked(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,int * locked,unsigned int flags)1714 static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1715 						unsigned long start,
1716 						unsigned long nr_pages,
1717 						struct page **pages,
1718 						int *locked,
1719 						unsigned int flags)
1720 {
1721 	long ret, pages_done;
1722 	bool must_unlock = false;
1723 
1724 	if (!nr_pages)
1725 		return 0;
1726 
1727 	/*
1728 	 * The internal caller expects GUP to manage the lock internally and the
1729 	 * lock must be released when this returns.
1730 	 */
1731 	if (!*locked) {
1732 		if (mmap_read_lock_killable(mm))
1733 			return -EAGAIN;
1734 		must_unlock = true;
1735 		*locked = 1;
1736 	}
1737 	else
1738 		mmap_assert_locked(mm);
1739 
1740 	if (flags & FOLL_PIN)
1741 		mm_set_has_pinned_flag(&mm->flags);
1742 
1743 	/*
1744 	 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1745 	 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1746 	 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1747 	 * for FOLL_GET, not for the newer FOLL_PIN.
1748 	 *
1749 	 * FOLL_PIN always expects pages to be non-null, but no need to assert
1750 	 * that here, as any failures will be obvious enough.
1751 	 */
1752 	if (pages && !(flags & FOLL_PIN))
1753 		flags |= FOLL_GET;
1754 
1755 	pages_done = 0;
1756 	for (;;) {
1757 		ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1758 				       locked);
1759 		if (!(flags & FOLL_UNLOCKABLE)) {
1760 			/* VM_FAULT_RETRY couldn't trigger, bypass */
1761 			pages_done = ret;
1762 			break;
1763 		}
1764 
1765 		/* VM_FAULT_RETRY or VM_FAULT_COMPLETED cannot return errors */
1766 		if (!*locked) {
1767 			BUG_ON(ret < 0);
1768 			BUG_ON(ret >= nr_pages);
1769 		}
1770 
1771 		if (ret > 0) {
1772 			nr_pages -= ret;
1773 			pages_done += ret;
1774 			if (!nr_pages)
1775 				break;
1776 		}
1777 		if (*locked) {
1778 			/*
1779 			 * VM_FAULT_RETRY didn't trigger or it was a
1780 			 * FOLL_NOWAIT.
1781 			 */
1782 			if (!pages_done)
1783 				pages_done = ret;
1784 			break;
1785 		}
1786 		/*
1787 		 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1788 		 * For the prefault case (!pages) we only update counts.
1789 		 */
1790 		if (likely(pages))
1791 			pages += ret;
1792 		start += ret << PAGE_SHIFT;
1793 
1794 		/* The lock was temporarily dropped, so we must unlock later */
1795 		must_unlock = true;
1796 
1797 retry:
1798 		/*
1799 		 * Repeat on the address that fired VM_FAULT_RETRY
1800 		 * with both FAULT_FLAG_ALLOW_RETRY and
1801 		 * FAULT_FLAG_TRIED.  Note that GUP can be interrupted
1802 		 * by fatal signals of even common signals, depending on
1803 		 * the caller's request. So we need to check it before we
1804 		 * start trying again otherwise it can loop forever.
1805 		 */
1806 		if (gup_signal_pending(flags)) {
1807 			if (!pages_done)
1808 				pages_done = -EINTR;
1809 			break;
1810 		}
1811 
1812 		ret = mmap_read_lock_killable(mm);
1813 		if (ret) {
1814 			BUG_ON(ret > 0);
1815 			if (!pages_done)
1816 				pages_done = ret;
1817 			break;
1818 		}
1819 
1820 		*locked = 1;
1821 		ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1822 				       pages, locked);
1823 		if (!*locked) {
1824 			/* Continue to retry until we succeeded */
1825 			BUG_ON(ret != 0);
1826 			goto retry;
1827 		}
1828 		if (ret != 1) {
1829 			BUG_ON(ret > 1);
1830 			if (!pages_done)
1831 				pages_done = ret;
1832 			break;
1833 		}
1834 		nr_pages--;
1835 		pages_done++;
1836 		if (!nr_pages)
1837 			break;
1838 		if (likely(pages))
1839 			pages++;
1840 		start += PAGE_SIZE;
1841 	}
1842 	if (must_unlock && *locked) {
1843 		/*
1844 		 * We either temporarily dropped the lock, or the caller
1845 		 * requested that we both acquire and drop the lock. Either way,
1846 		 * we must now unlock, and notify the caller of that state.
1847 		 */
1848 		mmap_read_unlock(mm);
1849 		*locked = 0;
1850 	}
1851 
1852 	/*
1853 	 * Failing to pin anything implies something has gone wrong (except when
1854 	 * FOLL_NOWAIT is specified).
1855 	 */
1856 	if (WARN_ON_ONCE(pages_done == 0 && !(flags & FOLL_NOWAIT)))
1857 		return -EFAULT;
1858 
1859 	return pages_done;
1860 }
1861 
1862 /**
1863  * populate_vma_page_range() -  populate a range of pages in the vma.
1864  * @vma:   target vma
1865  * @start: start address
1866  * @end:   end address
1867  * @locked: whether the mmap_lock is still held
1868  *
1869  * This takes care of mlocking the pages too if VM_LOCKED is set.
1870  *
1871  * Return either number of pages pinned in the vma, or a negative error
1872  * code on error.
1873  *
1874  * vma->vm_mm->mmap_lock must be held.
1875  *
1876  * If @locked is NULL, it may be held for read or write and will
1877  * be unperturbed.
1878  *
1879  * If @locked is non-NULL, it must held for read only and may be
1880  * released.  If it's released, *@locked will be set to 0.
1881  */
populate_vma_page_range(struct vm_area_struct * vma,unsigned long start,unsigned long end,int * locked)1882 long populate_vma_page_range(struct vm_area_struct *vma,
1883 		unsigned long start, unsigned long end, int *locked)
1884 {
1885 	struct mm_struct *mm = vma->vm_mm;
1886 	unsigned long nr_pages = (end - start) / PAGE_SIZE;
1887 	int local_locked = 1;
1888 	int gup_flags;
1889 	long ret;
1890 
1891 	VM_BUG_ON(!PAGE_ALIGNED(start));
1892 	VM_BUG_ON(!PAGE_ALIGNED(end));
1893 	VM_BUG_ON_VMA(start < vma->vm_start, vma);
1894 	VM_BUG_ON_VMA(end   > vma->vm_end, vma);
1895 	mmap_assert_locked(mm);
1896 
1897 	/*
1898 	 * Rightly or wrongly, the VM_LOCKONFAULT case has never used
1899 	 * faultin_page() to break COW, so it has no work to do here.
1900 	 */
1901 	if (vma->vm_flags & VM_LOCKONFAULT)
1902 		return nr_pages;
1903 
1904 	/* ... similarly, we've never faulted in PROT_NONE pages */
1905 	if (!vma_is_accessible(vma))
1906 		return -EFAULT;
1907 
1908 	gup_flags = FOLL_TOUCH;
1909 	/*
1910 	 * We want to touch writable mappings with a write fault in order
1911 	 * to break COW, except for shared mappings because these don't COW
1912 	 * and we would not want to dirty them for nothing.
1913 	 *
1914 	 * Otherwise, do a read fault, and use FOLL_FORCE in case it's not
1915 	 * readable (ie write-only or executable).
1916 	 */
1917 	if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1918 		gup_flags |= FOLL_WRITE;
1919 	else
1920 		gup_flags |= FOLL_FORCE;
1921 
1922 	if (locked)
1923 		gup_flags |= FOLL_UNLOCKABLE;
1924 
1925 	/*
1926 	 * We made sure addr is within a VMA, so the following will
1927 	 * not result in a stack expansion that recurses back here.
1928 	 */
1929 	ret = __get_user_pages(mm, start, nr_pages, gup_flags,
1930 			       NULL, locked ? locked : &local_locked);
1931 	lru_add_drain();
1932 	return ret;
1933 }
1934 
1935 /*
1936  * faultin_page_range() - populate (prefault) page tables inside the
1937  *			  given range readable/writable
1938  *
1939  * This takes care of mlocking the pages, too, if VM_LOCKED is set.
1940  *
1941  * @mm: the mm to populate page tables in
1942  * @start: start address
1943  * @end: end address
1944  * @write: whether to prefault readable or writable
1945  * @locked: whether the mmap_lock is still held
1946  *
1947  * Returns either number of processed pages in the MM, or a negative error
1948  * code on error (see __get_user_pages()). Note that this function reports
1949  * errors related to VMAs, such as incompatible mappings, as expected by
1950  * MADV_POPULATE_(READ|WRITE).
1951  *
1952  * The range must be page-aligned.
1953  *
1954  * mm->mmap_lock must be held. If it's released, *@locked will be set to 0.
1955  */
faultin_page_range(struct mm_struct * mm,unsigned long start,unsigned long end,bool write,int * locked)1956 long faultin_page_range(struct mm_struct *mm, unsigned long start,
1957 			unsigned long end, bool write, int *locked)
1958 {
1959 	unsigned long nr_pages = (end - start) / PAGE_SIZE;
1960 	int gup_flags;
1961 	long ret;
1962 
1963 	VM_BUG_ON(!PAGE_ALIGNED(start));
1964 	VM_BUG_ON(!PAGE_ALIGNED(end));
1965 	mmap_assert_locked(mm);
1966 
1967 	/*
1968 	 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark
1969 	 *	       the page dirty with FOLL_WRITE -- which doesn't make a
1970 	 *	       difference with !FOLL_FORCE, because the page is writable
1971 	 *	       in the page table.
1972 	 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit
1973 	 *		  a poisoned page.
1974 	 * !FOLL_FORCE: Require proper access permissions.
1975 	 */
1976 	gup_flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_UNLOCKABLE |
1977 		    FOLL_MADV_POPULATE;
1978 	if (write)
1979 		gup_flags |= FOLL_WRITE;
1980 
1981 	ret = __get_user_pages_locked(mm, start, nr_pages, NULL, locked,
1982 				      gup_flags);
1983 	lru_add_drain();
1984 	return ret;
1985 }
1986 
1987 /*
1988  * __mm_populate - populate and/or mlock pages within a range of address space.
1989  *
1990  * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1991  * flags. VMAs must be already marked with the desired vm_flags, and
1992  * mmap_lock must not be held.
1993  */
__mm_populate(unsigned long start,unsigned long len,int ignore_errors)1994 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1995 {
1996 	struct mm_struct *mm = current->mm;
1997 	unsigned long end, nstart, nend;
1998 	struct vm_area_struct *vma = NULL;
1999 	int locked = 0;
2000 	long ret = 0;
2001 
2002 	end = start + len;
2003 
2004 	for (nstart = start; nstart < end; nstart = nend) {
2005 		/*
2006 		 * We want to fault in pages for [nstart; end) address range.
2007 		 * Find first corresponding VMA.
2008 		 */
2009 		if (!locked) {
2010 			locked = 1;
2011 			mmap_read_lock(mm);
2012 			vma = find_vma_intersection(mm, nstart, end);
2013 		} else if (nstart >= vma->vm_end)
2014 			vma = find_vma_intersection(mm, vma->vm_end, end);
2015 
2016 		if (!vma)
2017 			break;
2018 		/*
2019 		 * Set [nstart; nend) to intersection of desired address
2020 		 * range with the first VMA. Also, skip undesirable VMA types.
2021 		 */
2022 		nend = min(end, vma->vm_end);
2023 		if (vma->vm_flags & (VM_IO | VM_PFNMAP))
2024 			continue;
2025 		if (nstart < vma->vm_start)
2026 			nstart = vma->vm_start;
2027 		/*
2028 		 * Now fault in a range of pages. populate_vma_page_range()
2029 		 * double checks the vma flags, so that it won't mlock pages
2030 		 * if the vma was already munlocked.
2031 		 */
2032 		ret = populate_vma_page_range(vma, nstart, nend, &locked);
2033 		if (ret < 0) {
2034 			if (ignore_errors) {
2035 				ret = 0;
2036 				continue;	/* continue at next VMA */
2037 			}
2038 			break;
2039 		}
2040 		nend = nstart + ret * PAGE_SIZE;
2041 		ret = 0;
2042 	}
2043 	if (locked)
2044 		mmap_read_unlock(mm);
2045 	return ret;	/* 0 or negative error code */
2046 }
2047 #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)2048 static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
2049 		unsigned long nr_pages, struct page **pages,
2050 		int *locked, unsigned int foll_flags)
2051 {
2052 	struct vm_area_struct *vma;
2053 	bool must_unlock = false;
2054 	unsigned long vm_flags;
2055 	long i;
2056 
2057 	if (!nr_pages)
2058 		return 0;
2059 
2060 	/*
2061 	 * The internal caller expects GUP to manage the lock internally and the
2062 	 * lock must be released when this returns.
2063 	 */
2064 	if (!*locked) {
2065 		if (mmap_read_lock_killable(mm))
2066 			return -EAGAIN;
2067 		must_unlock = true;
2068 		*locked = 1;
2069 	}
2070 
2071 	/* calculate required read or write permissions.
2072 	 * If FOLL_FORCE is set, we only require the "MAY" flags.
2073 	 */
2074 	vm_flags  = (foll_flags & FOLL_WRITE) ?
2075 			(VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
2076 	vm_flags &= (foll_flags & FOLL_FORCE) ?
2077 			(VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
2078 
2079 	for (i = 0; i < nr_pages; i++) {
2080 		vma = find_vma(mm, start);
2081 		if (!vma)
2082 			break;
2083 
2084 		/* protect what we can, including chardevs */
2085 		if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
2086 		    !(vm_flags & vma->vm_flags))
2087 			break;
2088 
2089 		if (pages) {
2090 			pages[i] = virt_to_page((void *)start);
2091 			if (pages[i])
2092 				get_page(pages[i]);
2093 		}
2094 
2095 		start = (start + PAGE_SIZE) & PAGE_MASK;
2096 	}
2097 
2098 	if (must_unlock && *locked) {
2099 		mmap_read_unlock(mm);
2100 		*locked = 0;
2101 	}
2102 
2103 	return i ? : -EFAULT;
2104 }
2105 #endif /* !CONFIG_MMU */
2106 
2107 /**
2108  * fault_in_writeable - fault in userspace address range for writing
2109  * @uaddr: start of address range
2110  * @size: size of address range
2111  *
2112  * Returns the number of bytes not faulted in (like copy_to_user() and
2113  * copy_from_user()).
2114  */
fault_in_writeable(char __user * uaddr,size_t size)2115 size_t fault_in_writeable(char __user *uaddr, size_t size)
2116 {
2117 	char __user *start = uaddr, *end;
2118 
2119 	if (unlikely(size == 0))
2120 		return 0;
2121 	if (!user_write_access_begin(uaddr, size))
2122 		return size;
2123 	if (!PAGE_ALIGNED(uaddr)) {
2124 		unsafe_put_user(0, uaddr, out);
2125 		uaddr = (char __user *)PAGE_ALIGN((unsigned long)uaddr);
2126 	}
2127 	end = (char __user *)PAGE_ALIGN((unsigned long)start + size);
2128 	if (unlikely(end < start))
2129 		end = NULL;
2130 	while (uaddr != end) {
2131 		unsafe_put_user(0, uaddr, out);
2132 		uaddr += PAGE_SIZE;
2133 	}
2134 
2135 out:
2136 	user_write_access_end();
2137 	if (size > uaddr - start)
2138 		return size - (uaddr - start);
2139 	return 0;
2140 }
2141 EXPORT_SYMBOL(fault_in_writeable);
2142 
2143 /**
2144  * fault_in_subpage_writeable - fault in an address range for writing
2145  * @uaddr: start of address range
2146  * @size: size of address range
2147  *
2148  * Fault in a user address range for writing while checking for permissions at
2149  * sub-page granularity (e.g. arm64 MTE). This function should be used when
2150  * the caller cannot guarantee forward progress of a copy_to_user() loop.
2151  *
2152  * Returns the number of bytes not faulted in (like copy_to_user() and
2153  * copy_from_user()).
2154  */
fault_in_subpage_writeable(char __user * uaddr,size_t size)2155 size_t fault_in_subpage_writeable(char __user *uaddr, size_t size)
2156 {
2157 	size_t faulted_in;
2158 
2159 	/*
2160 	 * Attempt faulting in at page granularity first for page table
2161 	 * permission checking. The arch-specific probe_subpage_writeable()
2162 	 * functions may not check for this.
2163 	 */
2164 	faulted_in = size - fault_in_writeable(uaddr, size);
2165 	if (faulted_in)
2166 		faulted_in -= probe_subpage_writeable(uaddr, faulted_in);
2167 
2168 	return size - faulted_in;
2169 }
2170 EXPORT_SYMBOL(fault_in_subpage_writeable);
2171 
2172 /*
2173  * fault_in_safe_writeable - fault in an address range for writing
2174  * @uaddr: start of address range
2175  * @size: length of address range
2176  *
2177  * Faults in an address range for writing.  This is primarily useful when we
2178  * already know that some or all of the pages in the address range aren't in
2179  * memory.
2180  *
2181  * Unlike fault_in_writeable(), this function is non-destructive.
2182  *
2183  * Note that we don't pin or otherwise hold the pages referenced that we fault
2184  * in.  There's no guarantee that they'll stay in memory for any duration of
2185  * time.
2186  *
2187  * Returns the number of bytes not faulted in, like copy_to_user() and
2188  * copy_from_user().
2189  */
fault_in_safe_writeable(const char __user * uaddr,size_t size)2190 size_t fault_in_safe_writeable(const char __user *uaddr, size_t size)
2191 {
2192 	unsigned long start = (unsigned long)uaddr, end;
2193 	struct mm_struct *mm = current->mm;
2194 	bool unlocked = false;
2195 
2196 	if (unlikely(size == 0))
2197 		return 0;
2198 	end = PAGE_ALIGN(start + size);
2199 	if (end < start)
2200 		end = 0;
2201 
2202 	mmap_read_lock(mm);
2203 	do {
2204 		if (fixup_user_fault(mm, start, FAULT_FLAG_WRITE, &unlocked))
2205 			break;
2206 		start = (start + PAGE_SIZE) & PAGE_MASK;
2207 	} while (start != end);
2208 	mmap_read_unlock(mm);
2209 
2210 	if (size > (unsigned long)uaddr - start)
2211 		return size - ((unsigned long)uaddr - start);
2212 	return 0;
2213 }
2214 EXPORT_SYMBOL(fault_in_safe_writeable);
2215 
2216 /**
2217  * fault_in_readable - fault in userspace address range for reading
2218  * @uaddr: start of user address range
2219  * @size: size of user address range
2220  *
2221  * Returns the number of bytes not faulted in (like copy_to_user() and
2222  * copy_from_user()).
2223  */
fault_in_readable(const char __user * uaddr,size_t size)2224 size_t fault_in_readable(const char __user *uaddr, size_t size)
2225 {
2226 	const char __user *start = uaddr, *end;
2227 	volatile char c;
2228 
2229 	if (unlikely(size == 0))
2230 		return 0;
2231 	if (!user_read_access_begin(uaddr, size))
2232 		return size;
2233 	if (!PAGE_ALIGNED(uaddr)) {
2234 		unsafe_get_user(c, uaddr, out);
2235 		uaddr = (const char __user *)PAGE_ALIGN((unsigned long)uaddr);
2236 	}
2237 	end = (const char __user *)PAGE_ALIGN((unsigned long)start + size);
2238 	if (unlikely(end < start))
2239 		end = NULL;
2240 	while (uaddr != end) {
2241 		unsafe_get_user(c, uaddr, out);
2242 		uaddr += PAGE_SIZE;
2243 	}
2244 
2245 out:
2246 	user_read_access_end();
2247 	(void)c;
2248 	if (size > uaddr - start)
2249 		return size - (uaddr - start);
2250 	return 0;
2251 }
2252 EXPORT_SYMBOL(fault_in_readable);
2253 
2254 /**
2255  * get_dump_page() - pin user page in memory while writing it to core dump
2256  * @addr: user address
2257  *
2258  * Returns struct page pointer of user page pinned for dump,
2259  * to be freed afterwards by put_page().
2260  *
2261  * Returns NULL on any kind of failure - a hole must then be inserted into
2262  * the corefile, to preserve alignment with its headers; and also returns
2263  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2264  * allowing a hole to be left in the corefile to save disk space.
2265  *
2266  * Called without mmap_lock (takes and releases the mmap_lock by itself).
2267  */
2268 #ifdef CONFIG_ELF_CORE
get_dump_page(unsigned long addr)2269 struct page *get_dump_page(unsigned long addr)
2270 {
2271 	struct page *page;
2272 	int locked = 0;
2273 	int ret;
2274 
2275 	ret = __get_user_pages_locked(current->mm, addr, 1, &page, &locked,
2276 				      FOLL_FORCE | FOLL_DUMP | FOLL_GET);
2277 	return (ret == 1) ? page : NULL;
2278 }
2279 #endif /* CONFIG_ELF_CORE */
2280 
2281 #ifdef CONFIG_MIGRATION
2282 
2283 /*
2284  * An array of either pages or folios ("pofs"). Although it may seem tempting to
2285  * avoid this complication, by simply interpreting a list of folios as a list of
2286  * pages, that approach won't work in the longer term, because eventually the
2287  * layouts of struct page and struct folio will become completely different.
2288  * Furthermore, this pof approach avoids excessive page_folio() calls.
2289  */
2290 struct pages_or_folios {
2291 	union {
2292 		struct page **pages;
2293 		struct folio **folios;
2294 		void **entries;
2295 	};
2296 	bool has_folios;
2297 	long nr_entries;
2298 };
2299 
pofs_get_folio(struct pages_or_folios * pofs,long i)2300 static struct folio *pofs_get_folio(struct pages_or_folios *pofs, long i)
2301 {
2302 	if (pofs->has_folios)
2303 		return pofs->folios[i];
2304 	return page_folio(pofs->pages[i]);
2305 }
2306 
pofs_clear_entry(struct pages_or_folios * pofs,long i)2307 static void pofs_clear_entry(struct pages_or_folios *pofs, long i)
2308 {
2309 	pofs->entries[i] = NULL;
2310 }
2311 
pofs_unpin(struct pages_or_folios * pofs)2312 static void pofs_unpin(struct pages_or_folios *pofs)
2313 {
2314 	if (pofs->has_folios)
2315 		unpin_folios(pofs->folios, pofs->nr_entries);
2316 	else
2317 		unpin_user_pages(pofs->pages, pofs->nr_entries);
2318 }
2319 
2320 /*
2321  * Returns the number of collected folios. Return value is always >= 0.
2322  */
collect_longterm_unpinnable_folios(struct list_head * movable_folio_list,struct pages_or_folios * pofs)2323 static void collect_longterm_unpinnable_folios(
2324 		struct list_head *movable_folio_list,
2325 		struct pages_or_folios *pofs)
2326 {
2327 	struct folio *prev_folio = NULL;
2328 	bool drain_allow = true;
2329 	unsigned long i;
2330 
2331 	for (i = 0; i < pofs->nr_entries; i++) {
2332 		struct folio *folio = pofs_get_folio(pofs, i);
2333 
2334 		if (folio == prev_folio)
2335 			continue;
2336 		prev_folio = folio;
2337 
2338 		if (folio_is_longterm_pinnable(folio))
2339 			continue;
2340 
2341 		if (folio_is_device_coherent(folio))
2342 			continue;
2343 
2344 		if (folio_test_hugetlb(folio)) {
2345 			folio_isolate_hugetlb(folio, movable_folio_list);
2346 			continue;
2347 		}
2348 
2349 		if (!folio_test_lru(folio) && drain_allow) {
2350 			lru_add_drain_all();
2351 			drain_allow = false;
2352 		}
2353 
2354 		if (!folio_isolate_lru(folio))
2355 			continue;
2356 
2357 		list_add_tail(&folio->lru, movable_folio_list);
2358 		node_stat_mod_folio(folio,
2359 				    NR_ISOLATED_ANON + folio_is_file_lru(folio),
2360 				    folio_nr_pages(folio));
2361 	}
2362 }
2363 
2364 /*
2365  * Unpins all folios and migrates device coherent folios and movable_folio_list.
2366  * Returns -EAGAIN if all folios were successfully migrated or -errno for
2367  * failure (or partial success).
2368  */
2369 static int
migrate_longterm_unpinnable_folios(struct list_head * movable_folio_list,struct pages_or_folios * pofs)2370 migrate_longterm_unpinnable_folios(struct list_head *movable_folio_list,
2371 				   struct pages_or_folios *pofs)
2372 {
2373 	int ret;
2374 	unsigned long i;
2375 
2376 	for (i = 0; i < pofs->nr_entries; i++) {
2377 		struct folio *folio = pofs_get_folio(pofs, i);
2378 
2379 		if (folio_is_device_coherent(folio)) {
2380 			/*
2381 			 * Migration will fail if the folio is pinned, so
2382 			 * convert the pin on the source folio to a normal
2383 			 * reference.
2384 			 */
2385 			pofs_clear_entry(pofs, i);
2386 			folio_get(folio);
2387 			gup_put_folio(folio, 1, FOLL_PIN);
2388 
2389 			if (migrate_device_coherent_folio(folio)) {
2390 				ret = -EBUSY;
2391 				goto err;
2392 			}
2393 
2394 			continue;
2395 		}
2396 
2397 		/*
2398 		 * We can't migrate folios with unexpected references, so drop
2399 		 * the reference obtained by __get_user_pages_locked().
2400 		 * Migrating folios have been added to movable_folio_list after
2401 		 * calling folio_isolate_lru() which takes a reference so the
2402 		 * folio won't be freed if it's migrating.
2403 		 */
2404 		unpin_folio(folio);
2405 		pofs_clear_entry(pofs, i);
2406 	}
2407 
2408 	if (!list_empty(movable_folio_list)) {
2409 		struct migration_target_control mtc = {
2410 			.nid = NUMA_NO_NODE,
2411 			.gfp_mask = GFP_USER | __GFP_NOWARN,
2412 			.reason = MR_LONGTERM_PIN,
2413 		};
2414 
2415 		if (migrate_pages(movable_folio_list, alloc_migration_target,
2416 				  NULL, (unsigned long)&mtc, MIGRATE_SYNC,
2417 				  MR_LONGTERM_PIN, NULL)) {
2418 			ret = -ENOMEM;
2419 			goto err;
2420 		}
2421 	}
2422 
2423 	putback_movable_pages(movable_folio_list);
2424 
2425 	return -EAGAIN;
2426 
2427 err:
2428 	pofs_unpin(pofs);
2429 	putback_movable_pages(movable_folio_list);
2430 
2431 	return ret;
2432 }
2433 
2434 static long
check_and_migrate_movable_pages_or_folios(struct pages_or_folios * pofs)2435 check_and_migrate_movable_pages_or_folios(struct pages_or_folios *pofs)
2436 {
2437 	LIST_HEAD(movable_folio_list);
2438 
2439 	collect_longterm_unpinnable_folios(&movable_folio_list, pofs);
2440 	if (list_empty(&movable_folio_list))
2441 		return 0;
2442 
2443 	return migrate_longterm_unpinnable_folios(&movable_folio_list, pofs);
2444 }
2445 
2446 /*
2447  * Check whether all folios are *allowed* to be pinned indefinitely (long term).
2448  * Rather confusingly, all folios in the range are required to be pinned via
2449  * FOLL_PIN, before calling this routine.
2450  *
2451  * Return values:
2452  *
2453  * 0: if everything is OK and all folios in the range are allowed to be pinned,
2454  * then this routine leaves all folios pinned and returns zero for success.
2455  *
2456  * -EAGAIN: if any folios in the range are not allowed to be pinned, then this
2457  * routine will migrate those folios away, unpin all the folios in the range. If
2458  * migration of the entire set of folios succeeds, then -EAGAIN is returned. The
2459  * caller should re-pin the entire range with FOLL_PIN and then call this
2460  * routine again.
2461  *
2462  * -ENOMEM, or any other -errno: if an error *other* than -EAGAIN occurs, this
2463  * indicates a migration failure. The caller should give up, and propagate the
2464  * error back up the call stack. The caller does not need to unpin any folios in
2465  * that case, because this routine will do the unpinning.
2466  */
check_and_migrate_movable_folios(unsigned long nr_folios,struct folio ** folios)2467 static long check_and_migrate_movable_folios(unsigned long nr_folios,
2468 					     struct folio **folios)
2469 {
2470 	struct pages_or_folios pofs = {
2471 		.folios = folios,
2472 		.has_folios = true,
2473 		.nr_entries = nr_folios,
2474 	};
2475 
2476 	return check_and_migrate_movable_pages_or_folios(&pofs);
2477 }
2478 
2479 /*
2480  * Return values and behavior are the same as those for
2481  * check_and_migrate_movable_folios().
2482  */
check_and_migrate_movable_pages(unsigned long nr_pages,struct page ** pages)2483 static long check_and_migrate_movable_pages(unsigned long nr_pages,
2484 					    struct page **pages)
2485 {
2486 	struct pages_or_folios pofs = {
2487 		.pages = pages,
2488 		.has_folios = false,
2489 		.nr_entries = nr_pages,
2490 	};
2491 
2492 	return check_and_migrate_movable_pages_or_folios(&pofs);
2493 }
2494 #else
check_and_migrate_movable_pages(unsigned long nr_pages,struct page ** pages)2495 static long check_and_migrate_movable_pages(unsigned long nr_pages,
2496 					    struct page **pages)
2497 {
2498 	return 0;
2499 }
2500 
check_and_migrate_movable_folios(unsigned long nr_folios,struct folio ** folios)2501 static long check_and_migrate_movable_folios(unsigned long nr_folios,
2502 					     struct folio **folios)
2503 {
2504 	return 0;
2505 }
2506 #endif /* CONFIG_MIGRATION */
2507 
2508 /*
2509  * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
2510  * allows us to process the FOLL_LONGTERM flag.
2511  */
__gup_longterm_locked(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,int * locked,unsigned int gup_flags)2512 static long __gup_longterm_locked(struct mm_struct *mm,
2513 				  unsigned long start,
2514 				  unsigned long nr_pages,
2515 				  struct page **pages,
2516 				  int *locked,
2517 				  unsigned int gup_flags)
2518 {
2519 	unsigned int flags;
2520 	long rc, nr_pinned_pages;
2521 
2522 	if (!(gup_flags & FOLL_LONGTERM))
2523 		return __get_user_pages_locked(mm, start, nr_pages, pages,
2524 					       locked, gup_flags);
2525 
2526 	flags = memalloc_pin_save();
2527 	do {
2528 		nr_pinned_pages = __get_user_pages_locked(mm, start, nr_pages,
2529 							  pages, locked,
2530 							  gup_flags);
2531 		if (nr_pinned_pages <= 0) {
2532 			rc = nr_pinned_pages;
2533 			break;
2534 		}
2535 
2536 		/* FOLL_LONGTERM implies FOLL_PIN */
2537 		rc = check_and_migrate_movable_pages(nr_pinned_pages, pages);
2538 	} while (rc == -EAGAIN);
2539 	memalloc_pin_restore(flags);
2540 	return rc ? rc : nr_pinned_pages;
2541 }
2542 
2543 /*
2544  * Check that the given flags are valid for the exported gup/pup interface, and
2545  * update them with the required flags that the caller must have set.
2546  */
is_valid_gup_args(struct page ** pages,int * locked,unsigned int * gup_flags_p,unsigned int to_set)2547 static bool is_valid_gup_args(struct page **pages, int *locked,
2548 			      unsigned int *gup_flags_p, unsigned int to_set)
2549 {
2550 	unsigned int gup_flags = *gup_flags_p;
2551 
2552 	/*
2553 	 * These flags not allowed to be specified externally to the gup
2554 	 * interfaces:
2555 	 * - FOLL_TOUCH/FOLL_PIN/FOLL_TRIED/FOLL_FAST_ONLY are internal only
2556 	 * - FOLL_REMOTE is internal only, set in (get|pin)_user_pages_remote()
2557 	 * - FOLL_UNLOCKABLE is internal only and used if locked is !NULL
2558 	 */
2559 	if (WARN_ON_ONCE(gup_flags & INTERNAL_GUP_FLAGS))
2560 		return false;
2561 
2562 	gup_flags |= to_set;
2563 	if (locked) {
2564 		/* At the external interface locked must be set */
2565 		if (WARN_ON_ONCE(*locked != 1))
2566 			return false;
2567 
2568 		gup_flags |= FOLL_UNLOCKABLE;
2569 	}
2570 
2571 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
2572 	if (WARN_ON_ONCE((gup_flags & (FOLL_PIN | FOLL_GET)) ==
2573 			 (FOLL_PIN | FOLL_GET)))
2574 		return false;
2575 
2576 	/* LONGTERM can only be specified when pinning */
2577 	if (WARN_ON_ONCE(!(gup_flags & FOLL_PIN) && (gup_flags & FOLL_LONGTERM)))
2578 		return false;
2579 
2580 	/* Pages input must be given if using GET/PIN */
2581 	if (WARN_ON_ONCE((gup_flags & (FOLL_GET | FOLL_PIN)) && !pages))
2582 		return false;
2583 
2584 	/* We want to allow the pgmap to be hot-unplugged at all times */
2585 	if (WARN_ON_ONCE((gup_flags & FOLL_LONGTERM) &&
2586 			 (gup_flags & FOLL_PCI_P2PDMA)))
2587 		return false;
2588 
2589 	*gup_flags_p = gup_flags;
2590 	return true;
2591 }
2592 
2593 #ifdef CONFIG_MMU
2594 /**
2595  * get_user_pages_remote() - pin user pages in memory
2596  * @mm:		mm_struct of target mm
2597  * @start:	starting user address
2598  * @nr_pages:	number of pages from start to pin
2599  * @gup_flags:	flags modifying lookup behaviour
2600  * @pages:	array that receives pointers to the pages pinned.
2601  *		Should be at least nr_pages long. Or NULL, if caller
2602  *		only intends to ensure the pages are faulted in.
2603  * @locked:	pointer to lock flag indicating whether lock is held and
2604  *		subsequently whether VM_FAULT_RETRY functionality can be
2605  *		utilised. Lock must initially be held.
2606  *
2607  * Returns either number of pages pinned (which may be less than the
2608  * number requested), or an error. Details about the return value:
2609  *
2610  * -- If nr_pages is 0, returns 0.
2611  * -- If nr_pages is >0, but no pages were pinned, returns -errno.
2612  * -- If nr_pages is >0, and some pages were pinned, returns the number of
2613  *    pages pinned. Again, this may be less than nr_pages.
2614  *
2615  * The caller is responsible for releasing returned @pages, via put_page().
2616  *
2617  * Must be called with mmap_lock held for read or write.
2618  *
2619  * get_user_pages_remote walks a process's page tables and takes a reference
2620  * to each struct page that each user address corresponds to at a given
2621  * instant. That is, it takes the page that would be accessed if a user
2622  * thread accesses the given user virtual address at that instant.
2623  *
2624  * This does not guarantee that the page exists in the user mappings when
2625  * get_user_pages_remote returns, and there may even be a completely different
2626  * page there in some cases (eg. if mmapped pagecache has been invalidated
2627  * and subsequently re-faulted). However it does guarantee that the page
2628  * won't be freed completely. And mostly callers simply care that the page
2629  * contains data that was valid *at some point in time*. Typically, an IO
2630  * or similar operation cannot guarantee anything stronger anyway because
2631  * locks can't be held over the syscall boundary.
2632  *
2633  * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
2634  * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
2635  * be called after the page is finished with, and before put_page is called.
2636  *
2637  * get_user_pages_remote is typically used for fewer-copy IO operations,
2638  * to get a handle on the memory by some means other than accesses
2639  * via the user virtual addresses. The pages may be submitted for
2640  * DMA to devices or accessed via their kernel linear mapping (via the
2641  * kmap APIs). Care should be taken to use the correct cache flushing APIs.
2642  *
2643  * See also get_user_pages_fast, for performance critical applications.
2644  *
2645  * get_user_pages_remote should be phased out in favor of
2646  * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
2647  * should use get_user_pages_remote because it cannot pass
2648  * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
2649  */
get_user_pages_remote(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)2650 long get_user_pages_remote(struct mm_struct *mm,
2651 		unsigned long start, unsigned long nr_pages,
2652 		unsigned int gup_flags, struct page **pages,
2653 		int *locked)
2654 {
2655 	int local_locked = 1;
2656 
2657 	if (!is_valid_gup_args(pages, locked, &gup_flags,
2658 			       FOLL_TOUCH | FOLL_REMOTE))
2659 		return -EINVAL;
2660 
2661 	return __get_user_pages_locked(mm, start, nr_pages, pages,
2662 				       locked ? locked : &local_locked,
2663 				       gup_flags);
2664 }
2665 EXPORT_SYMBOL(get_user_pages_remote);
2666 
2667 #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)2668 long get_user_pages_remote(struct mm_struct *mm,
2669 			   unsigned long start, unsigned long nr_pages,
2670 			   unsigned int gup_flags, struct page **pages,
2671 			   int *locked)
2672 {
2673 	return 0;
2674 }
2675 #endif /* !CONFIG_MMU */
2676 
2677 /**
2678  * get_user_pages() - pin user pages in memory
2679  * @start:      starting user address
2680  * @nr_pages:   number of pages from start to pin
2681  * @gup_flags:  flags modifying lookup behaviour
2682  * @pages:      array that receives pointers to the pages pinned.
2683  *              Should be at least nr_pages long. Or NULL, if caller
2684  *              only intends to ensure the pages are faulted in.
2685  *
2686  * This is the same as get_user_pages_remote(), just with a less-flexible
2687  * calling convention where we assume that the mm being operated on belongs to
2688  * the current task, and doesn't allow passing of a locked parameter.  We also
2689  * obviously don't pass FOLL_REMOTE in here.
2690  */
get_user_pages(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages)2691 long get_user_pages(unsigned long start, unsigned long nr_pages,
2692 		    unsigned int gup_flags, struct page **pages)
2693 {
2694 	int locked = 1;
2695 
2696 	if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_TOUCH))
2697 		return -EINVAL;
2698 
2699 	return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2700 				       &locked, gup_flags);
2701 }
2702 EXPORT_SYMBOL(get_user_pages);
2703 
2704 /*
2705  * get_user_pages_unlocked() is suitable to replace the form:
2706  *
2707  *      mmap_read_lock(mm);
2708  *      get_user_pages(mm, ..., pages, NULL);
2709  *      mmap_read_unlock(mm);
2710  *
2711  *  with:
2712  *
2713  *      get_user_pages_unlocked(mm, ..., pages);
2714  *
2715  * It is functionally equivalent to get_user_pages_fast so
2716  * get_user_pages_fast should be used instead if specific gup_flags
2717  * (e.g. FOLL_FORCE) are not required.
2718  */
get_user_pages_unlocked(unsigned long start,unsigned long nr_pages,struct page ** pages,unsigned int gup_flags)2719 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2720 			     struct page **pages, unsigned int gup_flags)
2721 {
2722 	int locked = 0;
2723 
2724 	if (!is_valid_gup_args(pages, NULL, &gup_flags,
2725 			       FOLL_TOUCH | FOLL_UNLOCKABLE))
2726 		return -EINVAL;
2727 
2728 	return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2729 				       &locked, gup_flags);
2730 }
2731 EXPORT_SYMBOL(get_user_pages_unlocked);
2732 
2733 /*
2734  * GUP-fast
2735  *
2736  * get_user_pages_fast attempts to pin user pages by walking the page
2737  * tables directly and avoids taking locks. Thus the walker needs to be
2738  * protected from page table pages being freed from under it, and should
2739  * block any THP splits.
2740  *
2741  * One way to achieve this is to have the walker disable interrupts, and
2742  * rely on IPIs from the TLB flushing code blocking before the page table
2743  * pages are freed. This is unsuitable for architectures that do not need
2744  * to broadcast an IPI when invalidating TLBs.
2745  *
2746  * Another way to achieve this is to batch up page table containing pages
2747  * belonging to more than one mm_user, then rcu_sched a callback to free those
2748  * pages. Disabling interrupts will allow the gup_fast() walker to both block
2749  * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2750  * (which is a relatively rare event). The code below adopts this strategy.
2751  *
2752  * Before activating this code, please be aware that the following assumptions
2753  * are currently made:
2754  *
2755  *  *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2756  *  free pages containing page tables or TLB flushing requires IPI broadcast.
2757  *
2758  *  *) ptes can be read atomically by the architecture.
2759  *
2760  *  *) access_ok is sufficient to validate userspace address ranges.
2761  *
2762  * The last two assumptions can be relaxed by the addition of helper functions.
2763  *
2764  * This code is based heavily on the PowerPC implementation by Nick Piggin.
2765  */
2766 #ifdef CONFIG_HAVE_GUP_FAST
2767 /*
2768  * Used in the GUP-fast path to determine whether GUP is permitted to work on
2769  * a specific folio.
2770  *
2771  * This call assumes the caller has pinned the folio, that the lowest page table
2772  * level still points to this folio, and that interrupts have been disabled.
2773  *
2774  * GUP-fast must reject all secretmem folios.
2775  *
2776  * Writing to pinned file-backed dirty tracked folios is inherently problematic
2777  * (see comment describing the writable_file_mapping_allowed() function). We
2778  * therefore try to avoid the most egregious case of a long-term mapping doing
2779  * so.
2780  *
2781  * This function cannot be as thorough as that one as the VMA is not available
2782  * in the fast path, so instead we whitelist known good cases and if in doubt,
2783  * fall back to the slow path.
2784  */
gup_fast_folio_allowed(struct folio * folio,unsigned int flags)2785 static bool gup_fast_folio_allowed(struct folio *folio, unsigned int flags)
2786 {
2787 	bool reject_file_backed = false;
2788 	struct address_space *mapping;
2789 	bool check_secretmem = false;
2790 	unsigned long mapping_flags;
2791 
2792 	/*
2793 	 * If we aren't pinning then no problematic write can occur. A long term
2794 	 * pin is the most egregious case so this is the one we disallow.
2795 	 */
2796 	if ((flags & (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE)) ==
2797 	    (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE))
2798 		reject_file_backed = true;
2799 
2800 	/* We hold a folio reference, so we can safely access folio fields. */
2801 
2802 	/* secretmem folios are always order-0 folios. */
2803 	if (IS_ENABLED(CONFIG_SECRETMEM) && !folio_test_large(folio))
2804 		check_secretmem = true;
2805 
2806 	if (!reject_file_backed && !check_secretmem)
2807 		return true;
2808 
2809 	if (WARN_ON_ONCE(folio_test_slab(folio)))
2810 		return false;
2811 
2812 	/* hugetlb neither requires dirty-tracking nor can be secretmem. */
2813 	if (folio_test_hugetlb(folio))
2814 		return true;
2815 
2816 	/*
2817 	 * GUP-fast disables IRQs. When IRQS are disabled, RCU grace periods
2818 	 * cannot proceed, which means no actions performed under RCU can
2819 	 * proceed either.
2820 	 *
2821 	 * inodes and thus their mappings are freed under RCU, which means the
2822 	 * mapping cannot be freed beneath us and thus we can safely dereference
2823 	 * it.
2824 	 */
2825 	lockdep_assert_irqs_disabled();
2826 
2827 	/*
2828 	 * However, there may be operations which _alter_ the mapping, so ensure
2829 	 * we read it once and only once.
2830 	 */
2831 	mapping = READ_ONCE(folio->mapping);
2832 
2833 	/*
2834 	 * The mapping may have been truncated, in any case we cannot determine
2835 	 * if this mapping is safe - fall back to slow path to determine how to
2836 	 * proceed.
2837 	 */
2838 	if (!mapping)
2839 		return false;
2840 
2841 	/* Anonymous folios pose no problem. */
2842 	mapping_flags = (unsigned long)mapping & PAGE_MAPPING_FLAGS;
2843 	if (mapping_flags)
2844 		return mapping_flags & PAGE_MAPPING_ANON;
2845 
2846 	/*
2847 	 * At this point, we know the mapping is non-null and points to an
2848 	 * address_space object.
2849 	 */
2850 	if (check_secretmem && secretmem_mapping(mapping))
2851 		return false;
2852 	/* The only remaining allowed file system is shmem. */
2853 	return !reject_file_backed || shmem_mapping(mapping);
2854 }
2855 
gup_fast_undo_dev_pagemap(int * nr,int nr_start,unsigned int flags,struct page ** pages)2856 static void __maybe_unused gup_fast_undo_dev_pagemap(int *nr, int nr_start,
2857 		unsigned int flags, struct page **pages)
2858 {
2859 	while ((*nr) - nr_start) {
2860 		struct folio *folio = page_folio(pages[--(*nr)]);
2861 
2862 		folio_clear_referenced(folio);
2863 		gup_put_folio(folio, 1, flags);
2864 	}
2865 }
2866 
2867 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2868 /*
2869  * GUP-fast relies on pte change detection to avoid concurrent pgtable
2870  * operations.
2871  *
2872  * To pin the page, GUP-fast needs to do below in order:
2873  * (1) pin the page (by prefetching pte), then (2) check pte not changed.
2874  *
2875  * For the rest of pgtable operations where pgtable updates can be racy
2876  * with GUP-fast, we need to do (1) clear pte, then (2) check whether page
2877  * is pinned.
2878  *
2879  * Above will work for all pte-level operations, including THP split.
2880  *
2881  * For THP collapse, it's a bit more complicated because GUP-fast may be
2882  * walking a pgtable page that is being freed (pte is still valid but pmd
2883  * can be cleared already).  To avoid race in such condition, we need to
2884  * also check pmd here to make sure pmd doesn't change (corresponds to
2885  * pmdp_collapse_flush() in the THP collapse code path).
2886  */
gup_fast_pte_range(pmd_t pmd,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2887 static int gup_fast_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2888 		unsigned long end, unsigned int flags, struct page **pages,
2889 		int *nr)
2890 {
2891 	struct dev_pagemap *pgmap = NULL;
2892 	int nr_start = *nr, ret = 0;
2893 	pte_t *ptep, *ptem;
2894 
2895 	ptem = ptep = pte_offset_map(&pmd, addr);
2896 	if (!ptep)
2897 		return 0;
2898 	do {
2899 		pte_t pte = ptep_get_lockless(ptep);
2900 		struct page *page;
2901 		struct folio *folio;
2902 
2903 		/*
2904 		 * Always fallback to ordinary GUP on PROT_NONE-mapped pages:
2905 		 * pte_access_permitted() better should reject these pages
2906 		 * either way: otherwise, GUP-fast might succeed in
2907 		 * cases where ordinary GUP would fail due to VMA access
2908 		 * permissions.
2909 		 */
2910 		if (pte_protnone(pte))
2911 			goto pte_unmap;
2912 
2913 		if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2914 			goto pte_unmap;
2915 
2916 		if (pte_devmap(pte)) {
2917 			if (unlikely(flags & FOLL_LONGTERM))
2918 				goto pte_unmap;
2919 
2920 			pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2921 			if (unlikely(!pgmap)) {
2922 				gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
2923 				goto pte_unmap;
2924 			}
2925 		} else if (pte_special(pte))
2926 			goto pte_unmap;
2927 
2928 		VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2929 		page = pte_page(pte);
2930 
2931 		folio = try_grab_folio_fast(page, 1, flags);
2932 		if (!folio)
2933 			goto pte_unmap;
2934 
2935 		if (unlikely(pmd_val(pmd) != pmd_val(*pmdp)) ||
2936 		    unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) {
2937 			gup_put_folio(folio, 1, flags);
2938 			goto pte_unmap;
2939 		}
2940 
2941 		if (!gup_fast_folio_allowed(folio, flags)) {
2942 			gup_put_folio(folio, 1, flags);
2943 			goto pte_unmap;
2944 		}
2945 
2946 		if (!pte_write(pte) && gup_must_unshare(NULL, flags, page)) {
2947 			gup_put_folio(folio, 1, flags);
2948 			goto pte_unmap;
2949 		}
2950 
2951 		/*
2952 		 * We need to make the page accessible if and only if we are
2953 		 * going to access its content (the FOLL_PIN case).  Please
2954 		 * see Documentation/core-api/pin_user_pages.rst for
2955 		 * details.
2956 		 */
2957 		if (flags & FOLL_PIN) {
2958 			ret = arch_make_folio_accessible(folio);
2959 			if (ret) {
2960 				gup_put_folio(folio, 1, flags);
2961 				goto pte_unmap;
2962 			}
2963 		}
2964 		folio_set_referenced(folio);
2965 		pages[*nr] = page;
2966 		(*nr)++;
2967 	} while (ptep++, addr += PAGE_SIZE, addr != end);
2968 
2969 	ret = 1;
2970 
2971 pte_unmap:
2972 	if (pgmap)
2973 		put_dev_pagemap(pgmap);
2974 	pte_unmap(ptem);
2975 	return ret;
2976 }
2977 #else
2978 
2979 /*
2980  * If we can't determine whether or not a pte is special, then fail immediately
2981  * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2982  * to be special.
2983  *
2984  * For a futex to be placed on a THP tail page, get_futex_key requires a
2985  * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2986  * useful to have gup_fast_pmd_leaf even if we can't operate on ptes.
2987  */
gup_fast_pte_range(pmd_t pmd,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2988 static int gup_fast_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2989 		unsigned long end, unsigned int flags, struct page **pages,
2990 		int *nr)
2991 {
2992 	return 0;
2993 }
2994 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2995 
2996 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
gup_fast_devmap_leaf(unsigned long pfn,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2997 static int gup_fast_devmap_leaf(unsigned long pfn, unsigned long addr,
2998 	unsigned long end, unsigned int flags, struct page **pages, int *nr)
2999 {
3000 	int nr_start = *nr;
3001 	struct dev_pagemap *pgmap = NULL;
3002 
3003 	do {
3004 		struct folio *folio;
3005 		struct page *page = pfn_to_page(pfn);
3006 
3007 		pgmap = get_dev_pagemap(pfn, pgmap);
3008 		if (unlikely(!pgmap)) {
3009 			gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
3010 			break;
3011 		}
3012 
3013 		if (!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)) {
3014 			gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
3015 			break;
3016 		}
3017 
3018 		folio = try_grab_folio_fast(page, 1, flags);
3019 		if (!folio) {
3020 			gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
3021 			break;
3022 		}
3023 		folio_set_referenced(folio);
3024 		pages[*nr] = page;
3025 		(*nr)++;
3026 		pfn++;
3027 	} while (addr += PAGE_SIZE, addr != end);
3028 
3029 	put_dev_pagemap(pgmap);
3030 	return addr == end;
3031 }
3032 
gup_fast_devmap_pmd_leaf(pmd_t orig,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3033 static int gup_fast_devmap_pmd_leaf(pmd_t orig, pmd_t *pmdp, unsigned long addr,
3034 		unsigned long end, unsigned int flags, struct page **pages,
3035 		int *nr)
3036 {
3037 	unsigned long fault_pfn;
3038 	int nr_start = *nr;
3039 
3040 	fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
3041 	if (!gup_fast_devmap_leaf(fault_pfn, addr, end, flags, pages, nr))
3042 		return 0;
3043 
3044 	if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
3045 		gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
3046 		return 0;
3047 	}
3048 	return 1;
3049 }
3050 
gup_fast_devmap_pud_leaf(pud_t orig,pud_t * pudp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3051 static int gup_fast_devmap_pud_leaf(pud_t orig, pud_t *pudp, unsigned long addr,
3052 		unsigned long end, unsigned int flags, struct page **pages,
3053 		int *nr)
3054 {
3055 	unsigned long fault_pfn;
3056 	int nr_start = *nr;
3057 
3058 	fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
3059 	if (!gup_fast_devmap_leaf(fault_pfn, addr, end, flags, pages, nr))
3060 		return 0;
3061 
3062 	if (unlikely(pud_val(orig) != pud_val(*pudp))) {
3063 		gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
3064 		return 0;
3065 	}
3066 	return 1;
3067 }
3068 #else
gup_fast_devmap_pmd_leaf(pmd_t orig,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3069 static int gup_fast_devmap_pmd_leaf(pmd_t orig, pmd_t *pmdp, unsigned long addr,
3070 		unsigned long end, unsigned int flags, struct page **pages,
3071 		int *nr)
3072 {
3073 	BUILD_BUG();
3074 	return 0;
3075 }
3076 
gup_fast_devmap_pud_leaf(pud_t pud,pud_t * pudp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3077 static int gup_fast_devmap_pud_leaf(pud_t pud, pud_t *pudp, unsigned long addr,
3078 		unsigned long end, unsigned int flags, struct page **pages,
3079 		int *nr)
3080 {
3081 	BUILD_BUG();
3082 	return 0;
3083 }
3084 #endif
3085 
gup_fast_pmd_leaf(pmd_t orig,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3086 static int gup_fast_pmd_leaf(pmd_t orig, pmd_t *pmdp, unsigned long addr,
3087 		unsigned long end, unsigned int flags, struct page **pages,
3088 		int *nr)
3089 {
3090 	struct page *page;
3091 	struct folio *folio;
3092 	int refs;
3093 
3094 	if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
3095 		return 0;
3096 
3097 	if (pmd_special(orig))
3098 		return 0;
3099 
3100 	if (pmd_devmap(orig)) {
3101 		if (unlikely(flags & FOLL_LONGTERM))
3102 			return 0;
3103 		return gup_fast_devmap_pmd_leaf(orig, pmdp, addr, end, flags,
3104 					        pages, nr);
3105 	}
3106 
3107 	page = pmd_page(orig);
3108 	refs = record_subpages(page, PMD_SIZE, addr, end, pages + *nr);
3109 
3110 	folio = try_grab_folio_fast(page, refs, flags);
3111 	if (!folio)
3112 		return 0;
3113 
3114 	if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
3115 		gup_put_folio(folio, refs, flags);
3116 		return 0;
3117 	}
3118 
3119 	if (!gup_fast_folio_allowed(folio, flags)) {
3120 		gup_put_folio(folio, refs, flags);
3121 		return 0;
3122 	}
3123 	if (!pmd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
3124 		gup_put_folio(folio, refs, flags);
3125 		return 0;
3126 	}
3127 
3128 	*nr += refs;
3129 	folio_set_referenced(folio);
3130 	return 1;
3131 }
3132 
gup_fast_pud_leaf(pud_t orig,pud_t * pudp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3133 static int gup_fast_pud_leaf(pud_t orig, pud_t *pudp, unsigned long addr,
3134 		unsigned long end, unsigned int flags, struct page **pages,
3135 		int *nr)
3136 {
3137 	struct page *page;
3138 	struct folio *folio;
3139 	int refs;
3140 
3141 	if (!pud_access_permitted(orig, flags & FOLL_WRITE))
3142 		return 0;
3143 
3144 	if (pud_special(orig))
3145 		return 0;
3146 
3147 	if (pud_devmap(orig)) {
3148 		if (unlikely(flags & FOLL_LONGTERM))
3149 			return 0;
3150 		return gup_fast_devmap_pud_leaf(orig, pudp, addr, end, flags,
3151 					        pages, nr);
3152 	}
3153 
3154 	page = pud_page(orig);
3155 	refs = record_subpages(page, PUD_SIZE, addr, end, pages + *nr);
3156 
3157 	folio = try_grab_folio_fast(page, refs, flags);
3158 	if (!folio)
3159 		return 0;
3160 
3161 	if (unlikely(pud_val(orig) != pud_val(*pudp))) {
3162 		gup_put_folio(folio, refs, flags);
3163 		return 0;
3164 	}
3165 
3166 	if (!gup_fast_folio_allowed(folio, flags)) {
3167 		gup_put_folio(folio, refs, flags);
3168 		return 0;
3169 	}
3170 
3171 	if (!pud_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
3172 		gup_put_folio(folio, refs, flags);
3173 		return 0;
3174 	}
3175 
3176 	*nr += refs;
3177 	folio_set_referenced(folio);
3178 	return 1;
3179 }
3180 
gup_fast_pgd_leaf(pgd_t orig,pgd_t * pgdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3181 static int gup_fast_pgd_leaf(pgd_t orig, pgd_t *pgdp, unsigned long addr,
3182 		unsigned long end, unsigned int flags, struct page **pages,
3183 		int *nr)
3184 {
3185 	int refs;
3186 	struct page *page;
3187 	struct folio *folio;
3188 
3189 	if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
3190 		return 0;
3191 
3192 	BUILD_BUG_ON(pgd_devmap(orig));
3193 
3194 	page = pgd_page(orig);
3195 	refs = record_subpages(page, PGDIR_SIZE, addr, end, pages + *nr);
3196 
3197 	folio = try_grab_folio_fast(page, refs, flags);
3198 	if (!folio)
3199 		return 0;
3200 
3201 	if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
3202 		gup_put_folio(folio, refs, flags);
3203 		return 0;
3204 	}
3205 
3206 	if (!pgd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
3207 		gup_put_folio(folio, refs, flags);
3208 		return 0;
3209 	}
3210 
3211 	if (!gup_fast_folio_allowed(folio, flags)) {
3212 		gup_put_folio(folio, refs, flags);
3213 		return 0;
3214 	}
3215 
3216 	*nr += refs;
3217 	folio_set_referenced(folio);
3218 	return 1;
3219 }
3220 
gup_fast_pmd_range(pud_t * pudp,pud_t pud,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3221 static int gup_fast_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr,
3222 		unsigned long end, unsigned int flags, struct page **pages,
3223 		int *nr)
3224 {
3225 	unsigned long next;
3226 	pmd_t *pmdp;
3227 
3228 	pmdp = pmd_offset_lockless(pudp, pud, addr);
3229 	do {
3230 		pmd_t pmd = pmdp_get_lockless(pmdp);
3231 
3232 		next = pmd_addr_end(addr, end);
3233 		if (!pmd_present(pmd))
3234 			return 0;
3235 
3236 		if (unlikely(pmd_leaf(pmd))) {
3237 			/* See gup_fast_pte_range() */
3238 			if (pmd_protnone(pmd))
3239 				return 0;
3240 
3241 			if (!gup_fast_pmd_leaf(pmd, pmdp, addr, next, flags,
3242 				pages, nr))
3243 				return 0;
3244 
3245 		} else if (!gup_fast_pte_range(pmd, pmdp, addr, next, flags,
3246 					       pages, nr))
3247 			return 0;
3248 	} while (pmdp++, addr = next, addr != end);
3249 
3250 	return 1;
3251 }
3252 
gup_fast_pud_range(p4d_t * p4dp,p4d_t p4d,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3253 static int gup_fast_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr,
3254 		unsigned long end, unsigned int flags, struct page **pages,
3255 		int *nr)
3256 {
3257 	unsigned long next;
3258 	pud_t *pudp;
3259 
3260 	pudp = pud_offset_lockless(p4dp, p4d, addr);
3261 	do {
3262 		pud_t pud = READ_ONCE(*pudp);
3263 
3264 		next = pud_addr_end(addr, end);
3265 		if (unlikely(!pud_present(pud)))
3266 			return 0;
3267 		if (unlikely(pud_leaf(pud))) {
3268 			if (!gup_fast_pud_leaf(pud, pudp, addr, next, flags,
3269 					       pages, nr))
3270 				return 0;
3271 		} else if (!gup_fast_pmd_range(pudp, pud, addr, next, flags,
3272 					       pages, nr))
3273 			return 0;
3274 	} while (pudp++, addr = next, addr != end);
3275 
3276 	return 1;
3277 }
3278 
gup_fast_p4d_range(pgd_t * pgdp,pgd_t pgd,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3279 static int gup_fast_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr,
3280 		unsigned long end, unsigned int flags, struct page **pages,
3281 		int *nr)
3282 {
3283 	unsigned long next;
3284 	p4d_t *p4dp;
3285 
3286 	p4dp = p4d_offset_lockless(pgdp, pgd, addr);
3287 	do {
3288 		p4d_t p4d = READ_ONCE(*p4dp);
3289 
3290 		next = p4d_addr_end(addr, end);
3291 		if (!p4d_present(p4d))
3292 			return 0;
3293 		BUILD_BUG_ON(p4d_leaf(p4d));
3294 		if (!gup_fast_pud_range(p4dp, p4d, addr, next, flags,
3295 					pages, nr))
3296 			return 0;
3297 	} while (p4dp++, addr = next, addr != end);
3298 
3299 	return 1;
3300 }
3301 
gup_fast_pgd_range(unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3302 static void gup_fast_pgd_range(unsigned long addr, unsigned long end,
3303 		unsigned int flags, struct page **pages, int *nr)
3304 {
3305 	unsigned long next;
3306 	pgd_t *pgdp;
3307 
3308 	pgdp = pgd_offset(current->mm, addr);
3309 	do {
3310 		pgd_t pgd = READ_ONCE(*pgdp);
3311 
3312 		next = pgd_addr_end(addr, end);
3313 		if (pgd_none(pgd))
3314 			return;
3315 		if (unlikely(pgd_leaf(pgd))) {
3316 			if (!gup_fast_pgd_leaf(pgd, pgdp, addr, next, flags,
3317 					       pages, nr))
3318 				return;
3319 		} else if (!gup_fast_p4d_range(pgdp, pgd, addr, next, flags,
3320 					       pages, nr))
3321 			return;
3322 	} while (pgdp++, addr = next, addr != end);
3323 }
3324 #else
gup_fast_pgd_range(unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3325 static inline void gup_fast_pgd_range(unsigned long addr, unsigned long end,
3326 		unsigned int flags, struct page **pages, int *nr)
3327 {
3328 }
3329 #endif /* CONFIG_HAVE_GUP_FAST */
3330 
3331 #ifndef gup_fast_permitted
3332 /*
3333  * Check if it's allowed to use get_user_pages_fast_only() for the range, or
3334  * we need to fall back to the slow version:
3335  */
gup_fast_permitted(unsigned long start,unsigned long end)3336 static bool gup_fast_permitted(unsigned long start, unsigned long end)
3337 {
3338 	return true;
3339 }
3340 #endif
3341 
gup_fast(unsigned long start,unsigned long end,unsigned int gup_flags,struct page ** pages)3342 static unsigned long gup_fast(unsigned long start, unsigned long end,
3343 		unsigned int gup_flags, struct page **pages)
3344 {
3345 	unsigned long flags;
3346 	int nr_pinned = 0;
3347 	unsigned seq;
3348 
3349 	if (!IS_ENABLED(CONFIG_HAVE_GUP_FAST) ||
3350 	    !gup_fast_permitted(start, end))
3351 		return 0;
3352 
3353 	if (gup_flags & FOLL_PIN) {
3354 		if (!raw_seqcount_try_begin(&current->mm->write_protect_seq, seq))
3355 			return 0;
3356 	}
3357 
3358 	/*
3359 	 * Disable interrupts. The nested form is used, in order to allow full,
3360 	 * general purpose use of this routine.
3361 	 *
3362 	 * With interrupts disabled, we block page table pages from being freed
3363 	 * from under us. See struct mmu_table_batch comments in
3364 	 * include/asm-generic/tlb.h for more details.
3365 	 *
3366 	 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
3367 	 * that come from THPs splitting.
3368 	 */
3369 	local_irq_save(flags);
3370 	gup_fast_pgd_range(start, end, gup_flags, pages, &nr_pinned);
3371 	local_irq_restore(flags);
3372 
3373 	/*
3374 	 * When pinning pages for DMA there could be a concurrent write protect
3375 	 * from fork() via copy_page_range(), in this case always fail GUP-fast.
3376 	 */
3377 	if (gup_flags & FOLL_PIN) {
3378 		if (read_seqcount_retry(&current->mm->write_protect_seq, seq)) {
3379 			gup_fast_unpin_user_pages(pages, nr_pinned);
3380 			return 0;
3381 		} else {
3382 			sanity_check_pinned_pages(pages, nr_pinned);
3383 		}
3384 	}
3385 	return nr_pinned;
3386 }
3387 
gup_fast_fallback(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages)3388 static int gup_fast_fallback(unsigned long start, unsigned long nr_pages,
3389 		unsigned int gup_flags, struct page **pages)
3390 {
3391 	unsigned long len, end;
3392 	unsigned long nr_pinned;
3393 	int locked = 0;
3394 	int ret;
3395 
3396 	if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
3397 				       FOLL_FORCE | FOLL_PIN | FOLL_GET |
3398 				       FOLL_FAST_ONLY | FOLL_NOFAULT |
3399 				       FOLL_PCI_P2PDMA | FOLL_HONOR_NUMA_FAULT)))
3400 		return -EINVAL;
3401 
3402 	if (gup_flags & FOLL_PIN)
3403 		mm_set_has_pinned_flag(&current->mm->flags);
3404 
3405 	if (!(gup_flags & FOLL_FAST_ONLY))
3406 		might_lock_read(&current->mm->mmap_lock);
3407 
3408 	start = untagged_addr(start) & PAGE_MASK;
3409 	len = nr_pages << PAGE_SHIFT;
3410 	if (check_add_overflow(start, len, &end))
3411 		return -EOVERFLOW;
3412 	if (end > TASK_SIZE_MAX)
3413 		return -EFAULT;
3414 	if (unlikely(!access_ok((void __user *)start, len)))
3415 		return -EFAULT;
3416 
3417 	nr_pinned = gup_fast(start, end, gup_flags, pages);
3418 	if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
3419 		return nr_pinned;
3420 
3421 	/* Slow path: try to get the remaining pages with get_user_pages */
3422 	start += nr_pinned << PAGE_SHIFT;
3423 	pages += nr_pinned;
3424 	ret = __gup_longterm_locked(current->mm, start, nr_pages - nr_pinned,
3425 				    pages, &locked,
3426 				    gup_flags | FOLL_TOUCH | FOLL_UNLOCKABLE);
3427 	if (ret < 0) {
3428 		/*
3429 		 * The caller has to unpin the pages we already pinned so
3430 		 * returning -errno is not an option
3431 		 */
3432 		if (nr_pinned)
3433 			return nr_pinned;
3434 		return ret;
3435 	}
3436 	return ret + nr_pinned;
3437 }
3438 
3439 /**
3440  * get_user_pages_fast_only() - pin user pages in memory
3441  * @start:      starting user address
3442  * @nr_pages:   number of pages from start to pin
3443  * @gup_flags:  flags modifying pin behaviour
3444  * @pages:      array that receives pointers to the pages pinned.
3445  *              Should be at least nr_pages long.
3446  *
3447  * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
3448  * the regular GUP.
3449  *
3450  * If the architecture does not support this function, simply return with no
3451  * pages pinned.
3452  *
3453  * Careful, careful! COW breaking can go either way, so a non-write
3454  * access can get ambiguous page results. If you call this function without
3455  * 'write' set, you'd better be sure that you're ok with that ambiguity.
3456  */
get_user_pages_fast_only(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)3457 int get_user_pages_fast_only(unsigned long start, int nr_pages,
3458 			     unsigned int gup_flags, struct page **pages)
3459 {
3460 	/*
3461 	 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
3462 	 * because gup fast is always a "pin with a +1 page refcount" request.
3463 	 *
3464 	 * FOLL_FAST_ONLY is required in order to match the API description of
3465 	 * this routine: no fall back to regular ("slow") GUP.
3466 	 */
3467 	if (!is_valid_gup_args(pages, NULL, &gup_flags,
3468 			       FOLL_GET | FOLL_FAST_ONLY))
3469 		return -EINVAL;
3470 
3471 	return gup_fast_fallback(start, nr_pages, gup_flags, pages);
3472 }
3473 EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
3474 
3475 /**
3476  * get_user_pages_fast() - pin user pages in memory
3477  * @start:      starting user address
3478  * @nr_pages:   number of pages from start to pin
3479  * @gup_flags:  flags modifying pin behaviour
3480  * @pages:      array that receives pointers to the pages pinned.
3481  *              Should be at least nr_pages long.
3482  *
3483  * Attempt to pin user pages in memory without taking mm->mmap_lock.
3484  * If not successful, it will fall back to taking the lock and
3485  * calling get_user_pages().
3486  *
3487  * Returns number of pages pinned. This may be fewer than the number requested.
3488  * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
3489  * -errno.
3490  */
get_user_pages_fast(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)3491 int get_user_pages_fast(unsigned long start, int nr_pages,
3492 			unsigned int gup_flags, struct page **pages)
3493 {
3494 	/*
3495 	 * The caller may or may not have explicitly set FOLL_GET; either way is
3496 	 * OK. However, internally (within mm/gup.c), gup fast variants must set
3497 	 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
3498 	 * request.
3499 	 */
3500 	if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_GET))
3501 		return -EINVAL;
3502 	return gup_fast_fallback(start, nr_pages, gup_flags, pages);
3503 }
3504 EXPORT_SYMBOL_GPL(get_user_pages_fast);
3505 
3506 /**
3507  * pin_user_pages_fast() - pin user pages in memory without taking locks
3508  *
3509  * @start:      starting user address
3510  * @nr_pages:   number of pages from start to pin
3511  * @gup_flags:  flags modifying pin behaviour
3512  * @pages:      array that receives pointers to the pages pinned.
3513  *              Should be at least nr_pages long.
3514  *
3515  * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
3516  * get_user_pages_fast() for documentation on the function arguments, because
3517  * the arguments here are identical.
3518  *
3519  * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3520  * see Documentation/core-api/pin_user_pages.rst for further details.
3521  *
3522  * Note that if a zero_page is amongst the returned pages, it will not have
3523  * pins in it and unpin_user_page() will not remove pins from it.
3524  */
pin_user_pages_fast(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)3525 int pin_user_pages_fast(unsigned long start, int nr_pages,
3526 			unsigned int gup_flags, struct page **pages)
3527 {
3528 	if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN))
3529 		return -EINVAL;
3530 	return gup_fast_fallback(start, nr_pages, gup_flags, pages);
3531 }
3532 EXPORT_SYMBOL_GPL(pin_user_pages_fast);
3533 
3534 /**
3535  * pin_user_pages_remote() - pin pages of a remote process
3536  *
3537  * @mm:		mm_struct of target mm
3538  * @start:	starting user address
3539  * @nr_pages:	number of pages from start to pin
3540  * @gup_flags:	flags modifying lookup behaviour
3541  * @pages:	array that receives pointers to the pages pinned.
3542  *		Should be at least nr_pages long.
3543  * @locked:	pointer to lock flag indicating whether lock is held and
3544  *		subsequently whether VM_FAULT_RETRY functionality can be
3545  *		utilised. Lock must initially be held.
3546  *
3547  * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
3548  * get_user_pages_remote() for documentation on the function arguments, because
3549  * the arguments here are identical.
3550  *
3551  * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3552  * see Documentation/core-api/pin_user_pages.rst for details.
3553  *
3554  * Note that if a zero_page is amongst the returned pages, it will not have
3555  * pins in it and unpin_user_page*() will not remove pins from it.
3556  */
pin_user_pages_remote(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)3557 long pin_user_pages_remote(struct mm_struct *mm,
3558 			   unsigned long start, unsigned long nr_pages,
3559 			   unsigned int gup_flags, struct page **pages,
3560 			   int *locked)
3561 {
3562 	int local_locked = 1;
3563 
3564 	if (!is_valid_gup_args(pages, locked, &gup_flags,
3565 			       FOLL_PIN | FOLL_TOUCH | FOLL_REMOTE))
3566 		return 0;
3567 	return __gup_longterm_locked(mm, start, nr_pages, pages,
3568 				     locked ? locked : &local_locked,
3569 				     gup_flags);
3570 }
3571 EXPORT_SYMBOL(pin_user_pages_remote);
3572 
3573 /**
3574  * pin_user_pages() - pin user pages in memory for use by other devices
3575  *
3576  * @start:	starting user address
3577  * @nr_pages:	number of pages from start to pin
3578  * @gup_flags:	flags modifying lookup behaviour
3579  * @pages:	array that receives pointers to the pages pinned.
3580  *		Should be at least nr_pages long.
3581  *
3582  * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
3583  * FOLL_PIN is set.
3584  *
3585  * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3586  * see Documentation/core-api/pin_user_pages.rst for details.
3587  *
3588  * Note that if a zero_page is amongst the returned pages, it will not have
3589  * pins in it and unpin_user_page*() will not remove pins from it.
3590  */
pin_user_pages(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages)3591 long pin_user_pages(unsigned long start, unsigned long nr_pages,
3592 		    unsigned int gup_flags, struct page **pages)
3593 {
3594 	int locked = 1;
3595 
3596 	if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN))
3597 		return 0;
3598 	return __gup_longterm_locked(current->mm, start, nr_pages,
3599 				     pages, &locked, gup_flags);
3600 }
3601 EXPORT_SYMBOL(pin_user_pages);
3602 
3603 /*
3604  * pin_user_pages_unlocked() is the FOLL_PIN variant of
3605  * get_user_pages_unlocked(). Behavior is the same, except that this one sets
3606  * FOLL_PIN and rejects FOLL_GET.
3607  *
3608  * Note that if a zero_page is amongst the returned pages, it will not have
3609  * pins in it and unpin_user_page*() will not remove pins from it.
3610  */
pin_user_pages_unlocked(unsigned long start,unsigned long nr_pages,struct page ** pages,unsigned int gup_flags)3611 long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
3612 			     struct page **pages, unsigned int gup_flags)
3613 {
3614 	int locked = 0;
3615 
3616 	if (!is_valid_gup_args(pages, NULL, &gup_flags,
3617 			       FOLL_PIN | FOLL_TOUCH | FOLL_UNLOCKABLE))
3618 		return 0;
3619 
3620 	return __gup_longterm_locked(current->mm, start, nr_pages, pages,
3621 				     &locked, gup_flags);
3622 }
3623 EXPORT_SYMBOL(pin_user_pages_unlocked);
3624 
3625 /**
3626  * memfd_pin_folios() - pin folios associated with a memfd
3627  * @memfd:      the memfd whose folios are to be pinned
3628  * @start:      the first memfd offset
3629  * @end:        the last memfd offset (inclusive)
3630  * @folios:     array that receives pointers to the folios pinned
3631  * @max_folios: maximum number of entries in @folios
3632  * @offset:     the offset into the first folio
3633  *
3634  * Attempt to pin folios associated with a memfd in the contiguous range
3635  * [start, end]. Given that a memfd is either backed by shmem or hugetlb,
3636  * the folios can either be found in the page cache or need to be allocated
3637  * if necessary. Once the folios are located, they are all pinned via
3638  * FOLL_PIN and @offset is populatedwith the offset into the first folio.
3639  * And, eventually, these pinned folios must be released either using
3640  * unpin_folios() or unpin_folio().
3641  *
3642  * It must be noted that the folios may be pinned for an indefinite amount
3643  * of time. And, in most cases, the duration of time they may stay pinned
3644  * would be controlled by the userspace. This behavior is effectively the
3645  * same as using FOLL_LONGTERM with other GUP APIs.
3646  *
3647  * Returns number of folios pinned, which could be less than @max_folios
3648  * as it depends on the folio sizes that cover the range [start, end].
3649  * If no folios were pinned, it returns -errno.
3650  */
memfd_pin_folios(struct file * memfd,loff_t start,loff_t end,struct folio ** folios,unsigned int max_folios,pgoff_t * offset)3651 long memfd_pin_folios(struct file *memfd, loff_t start, loff_t end,
3652 		      struct folio **folios, unsigned int max_folios,
3653 		      pgoff_t *offset)
3654 {
3655 	unsigned int flags, nr_folios, nr_found;
3656 	unsigned int i, pgshift = PAGE_SHIFT;
3657 	pgoff_t start_idx, end_idx, next_idx;
3658 	struct folio *folio = NULL;
3659 	struct folio_batch fbatch;
3660 	struct hstate *h;
3661 	long ret = -EINVAL;
3662 
3663 	if (start < 0 || start > end || !max_folios)
3664 		return -EINVAL;
3665 
3666 	if (!memfd)
3667 		return -EINVAL;
3668 
3669 	if (!shmem_file(memfd) && !is_file_hugepages(memfd))
3670 		return -EINVAL;
3671 
3672 	if (end >= i_size_read(file_inode(memfd)))
3673 		return -EINVAL;
3674 
3675 	if (is_file_hugepages(memfd)) {
3676 		h = hstate_file(memfd);
3677 		pgshift = huge_page_shift(h);
3678 	}
3679 
3680 	flags = memalloc_pin_save();
3681 	do {
3682 		nr_folios = 0;
3683 		start_idx = start >> pgshift;
3684 		end_idx = end >> pgshift;
3685 		if (is_file_hugepages(memfd)) {
3686 			start_idx <<= huge_page_order(h);
3687 			end_idx <<= huge_page_order(h);
3688 		}
3689 
3690 		folio_batch_init(&fbatch);
3691 		while (start_idx <= end_idx && nr_folios < max_folios) {
3692 			/*
3693 			 * In most cases, we should be able to find the folios
3694 			 * in the page cache. If we cannot find them for some
3695 			 * reason, we try to allocate them and add them to the
3696 			 * page cache.
3697 			 */
3698 			nr_found = filemap_get_folios_contig(memfd->f_mapping,
3699 							     &start_idx,
3700 							     end_idx,
3701 							     &fbatch);
3702 			if (folio) {
3703 				folio_put(folio);
3704 				folio = NULL;
3705 			}
3706 
3707 			next_idx = 0;
3708 			for (i = 0; i < nr_found; i++) {
3709 				/*
3710 				 * As there can be multiple entries for a
3711 				 * given folio in the batch returned by
3712 				 * filemap_get_folios_contig(), the below
3713 				 * check is to ensure that we pin and return a
3714 				 * unique set of folios between start and end.
3715 				 */
3716 				if (next_idx &&
3717 				    next_idx != folio_index(fbatch.folios[i]))
3718 					continue;
3719 
3720 				folio = page_folio(&fbatch.folios[i]->page);
3721 
3722 				if (try_grab_folio(folio, 1, FOLL_PIN)) {
3723 					folio_batch_release(&fbatch);
3724 					ret = -EINVAL;
3725 					goto err;
3726 				}
3727 
3728 				if (nr_folios == 0)
3729 					*offset = offset_in_folio(folio, start);
3730 
3731 				folios[nr_folios] = folio;
3732 				next_idx = folio_next_index(folio);
3733 				if (++nr_folios == max_folios)
3734 					break;
3735 			}
3736 
3737 			folio = NULL;
3738 			folio_batch_release(&fbatch);
3739 			if (!nr_found) {
3740 				folio = memfd_alloc_folio(memfd, start_idx);
3741 				if (IS_ERR(folio)) {
3742 					ret = PTR_ERR(folio);
3743 					if (ret != -EEXIST)
3744 						goto err;
3745 					folio = NULL;
3746 				}
3747 			}
3748 		}
3749 
3750 		ret = check_and_migrate_movable_folios(nr_folios, folios);
3751 	} while (ret == -EAGAIN);
3752 
3753 	memalloc_pin_restore(flags);
3754 	return ret ? ret : nr_folios;
3755 err:
3756 	memalloc_pin_restore(flags);
3757 	unpin_folios(folios, nr_folios);
3758 
3759 	return ret;
3760 }
3761 EXPORT_SYMBOL_GPL(memfd_pin_folios);
3762 
3763 /**
3764  * folio_add_pins() - add pins to an already-pinned folio
3765  * @folio: the folio to add more pins to
3766  * @pins: number of pins to add
3767  *
3768  * Try to add more pins to an already-pinned folio. The semantics
3769  * of the pin (e.g., FOLL_WRITE) follow any existing pin and cannot
3770  * be changed.
3771  *
3772  * This function is helpful when having obtained a pin on a large folio
3773  * using memfd_pin_folios(), but wanting to logically unpin parts
3774  * (e.g., individual pages) of the folio later, for example, using
3775  * unpin_user_page_range_dirty_lock().
3776  *
3777  * This is not the right interface to initially pin a folio.
3778  */
folio_add_pins(struct folio * folio,unsigned int pins)3779 int folio_add_pins(struct folio *folio, unsigned int pins)
3780 {
3781 	VM_WARN_ON_ONCE(!folio_maybe_dma_pinned(folio));
3782 
3783 	return try_grab_folio(folio, pins, FOLL_PIN);
3784 }
3785 EXPORT_SYMBOL_GPL(folio_add_pins);
3786