Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2#include <linux/kernel.h>
3#include <linux/errno.h>
4#include <linux/err.h>
5#include <linux/spinlock.h>
6
7#include <linux/mm.h>
8#include <linux/memremap.h>
9#include <linux/pagemap.h>
10#include <linux/rmap.h>
11#include <linux/swap.h>
12#include <linux/swapops.h>
13#include <linux/secretmem.h>
14
15#include <linux/sched/signal.h>
16#include <linux/rwsem.h>
17#include <linux/hugetlb.h>
18#include <linux/migrate.h>
19#include <linux/mm_inline.h>
20#include <linux/sched/mm.h>
21#include <linux/shmem_fs.h>
22
23#include <asm/mmu_context.h>
24#include <asm/tlbflush.h>
25
26#include "internal.h"
27
28struct follow_page_context {
29 struct dev_pagemap *pgmap;
30 unsigned int page_mask;
31};
32
33static inline void sanity_check_pinned_pages(struct page **pages,
34 unsigned long npages)
35{
36 if (!IS_ENABLED(CONFIG_DEBUG_VM))
37 return;
38
39 /*
40 * We only pin anonymous pages if they are exclusive. Once pinned, we
41 * can no longer turn them possibly shared and PageAnonExclusive() will
42 * stick around until the page is freed.
43 *
44 * We'd like to verify that our pinned anonymous pages are still mapped
45 * exclusively. The issue with anon THP is that we don't know how
46 * they are/were mapped when pinning them. However, for anon
47 * THP we can assume that either the given page (PTE-mapped THP) or
48 * the head page (PMD-mapped THP) should be PageAnonExclusive(). If
49 * neither is the case, there is certainly something wrong.
50 */
51 for (; npages; npages--, pages++) {
52 struct page *page = *pages;
53 struct folio *folio = page_folio(page);
54
55 if (is_zero_page(page) ||
56 !folio_test_anon(folio))
57 continue;
58 if (!folio_test_large(folio) || folio_test_hugetlb(folio))
59 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page), page);
60 else
61 /* Either a PTE-mapped or a PMD-mapped THP. */
62 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page) &&
63 !PageAnonExclusive(page), page);
64 }
65}
66
67/*
68 * Return the folio with ref appropriately incremented,
69 * or NULL if that failed.
70 */
71static inline struct folio *try_get_folio(struct page *page, int refs)
72{
73 struct folio *folio;
74
75retry:
76 folio = page_folio(page);
77 if (WARN_ON_ONCE(folio_ref_count(folio) < 0))
78 return NULL;
79 if (unlikely(!folio_ref_try_add_rcu(folio, refs)))
80 return NULL;
81
82 /*
83 * At this point we have a stable reference to the folio; but it
84 * could be that between calling page_folio() and the refcount
85 * increment, the folio was split, in which case we'd end up
86 * holding a reference on a folio that has nothing to do with the page
87 * we were given anymore.
88 * So now that the folio is stable, recheck that the page still
89 * belongs to this folio.
90 */
91 if (unlikely(page_folio(page) != folio)) {
92 if (!put_devmap_managed_page_refs(&folio->page, refs))
93 folio_put_refs(folio, refs);
94 goto retry;
95 }
96
97 return folio;
98}
99
100/**
101 * try_grab_folio() - Attempt to get or pin a folio.
102 * @page: pointer to page to be grabbed
103 * @refs: the value to (effectively) add to the folio's refcount
104 * @flags: gup flags: these are the FOLL_* flag values.
105 *
106 * "grab" names in this file mean, "look at flags to decide whether to use
107 * FOLL_PIN or FOLL_GET behavior, when incrementing the folio's refcount.
108 *
109 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
110 * same time. (That's true throughout the get_user_pages*() and
111 * pin_user_pages*() APIs.) Cases:
112 *
113 * FOLL_GET: folio's refcount will be incremented by @refs.
114 *
115 * FOLL_PIN on large folios: folio's refcount will be incremented by
116 * @refs, and its pincount will be incremented by @refs.
117 *
118 * FOLL_PIN on single-page folios: folio's refcount will be incremented by
119 * @refs * GUP_PIN_COUNTING_BIAS.
120 *
121 * Return: The folio containing @page (with refcount appropriately
122 * incremented) for success, or NULL upon failure. If neither FOLL_GET
123 * nor FOLL_PIN was set, that's considered failure, and furthermore,
124 * a likely bug in the caller, so a warning is also emitted.
125 */
126struct folio *try_grab_folio(struct page *page, int refs, unsigned int flags)
127{
128 struct folio *folio;
129
130 if (WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == 0))
131 return NULL;
132
133 if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
134 return NULL;
135
136 if (flags & FOLL_GET)
137 return try_get_folio(page, refs);
138
139 /* FOLL_PIN is set */
140
141 /*
142 * Don't take a pin on the zero page - it's not going anywhere
143 * and it is used in a *lot* of places.
144 */
145 if (is_zero_page(page))
146 return page_folio(page);
147
148 folio = try_get_folio(page, refs);
149 if (!folio)
150 return NULL;
151
152 /*
153 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
154 * right zone, so fail and let the caller fall back to the slow
155 * path.
156 */
157 if (unlikely((flags & FOLL_LONGTERM) &&
158 !folio_is_longterm_pinnable(folio))) {
159 if (!put_devmap_managed_page_refs(&folio->page, refs))
160 folio_put_refs(folio, refs);
161 return NULL;
162 }
163
164 /*
165 * When pinning a large folio, use an exact count to track it.
166 *
167 * However, be sure to *also* increment the normal folio
168 * refcount field at least once, so that the folio really
169 * is pinned. That's why the refcount from the earlier
170 * try_get_folio() is left intact.
171 */
172 if (folio_test_large(folio))
173 atomic_add(refs, &folio->_pincount);
174 else
175 folio_ref_add(folio,
176 refs * (GUP_PIN_COUNTING_BIAS - 1));
177 /*
178 * Adjust the pincount before re-checking the PTE for changes.
179 * This is essentially a smp_mb() and is paired with a memory
180 * barrier in folio_try_share_anon_rmap_*().
181 */
182 smp_mb__after_atomic();
183
184 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs);
185
186 return folio;
187}
188
189static void gup_put_folio(struct folio *folio, int refs, unsigned int flags)
190{
191 if (flags & FOLL_PIN) {
192 if (is_zero_folio(folio))
193 return;
194 node_stat_mod_folio(folio, NR_FOLL_PIN_RELEASED, refs);
195 if (folio_test_large(folio))
196 atomic_sub(refs, &folio->_pincount);
197 else
198 refs *= GUP_PIN_COUNTING_BIAS;
199 }
200
201 if (!put_devmap_managed_page_refs(&folio->page, refs))
202 folio_put_refs(folio, refs);
203}
204
205/**
206 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
207 * @page: pointer to page to be grabbed
208 * @flags: gup flags: these are the FOLL_* flag values.
209 *
210 * This might not do anything at all, depending on the flags argument.
211 *
212 * "grab" names in this file mean, "look at flags to decide whether to use
213 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
214 *
215 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
216 * time. Cases: please see the try_grab_folio() documentation, with
217 * "refs=1".
218 *
219 * Return: 0 for success, or if no action was required (if neither FOLL_PIN
220 * nor FOLL_GET was set, nothing is done). A negative error code for failure:
221 *
222 * -ENOMEM FOLL_GET or FOLL_PIN was set, but the page could not
223 * be grabbed.
224 */
225int __must_check try_grab_page(struct page *page, unsigned int flags)
226{
227 struct folio *folio = page_folio(page);
228
229 if (WARN_ON_ONCE(folio_ref_count(folio) <= 0))
230 return -ENOMEM;
231
232 if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
233 return -EREMOTEIO;
234
235 if (flags & FOLL_GET)
236 folio_ref_inc(folio);
237 else if (flags & FOLL_PIN) {
238 /*
239 * Don't take a pin on the zero page - it's not going anywhere
240 * and it is used in a *lot* of places.
241 */
242 if (is_zero_page(page))
243 return 0;
244
245 /*
246 * Similar to try_grab_folio(): be sure to *also*
247 * increment the normal page refcount field at least once,
248 * so that the page really is pinned.
249 */
250 if (folio_test_large(folio)) {
251 folio_ref_add(folio, 1);
252 atomic_add(1, &folio->_pincount);
253 } else {
254 folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
255 }
256
257 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, 1);
258 }
259
260 return 0;
261}
262
263/**
264 * unpin_user_page() - release a dma-pinned page
265 * @page: pointer to page to be released
266 *
267 * Pages that were pinned via pin_user_pages*() must be released via either
268 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
269 * that such pages can be separately tracked and uniquely handled. In
270 * particular, interactions with RDMA and filesystems need special handling.
271 */
272void unpin_user_page(struct page *page)
273{
274 sanity_check_pinned_pages(&page, 1);
275 gup_put_folio(page_folio(page), 1, FOLL_PIN);
276}
277EXPORT_SYMBOL(unpin_user_page);
278
279/**
280 * folio_add_pin - Try to get an additional pin on a pinned folio
281 * @folio: The folio to be pinned
282 *
283 * Get an additional pin on a folio we already have a pin on. Makes no change
284 * if the folio is a zero_page.
285 */
286void folio_add_pin(struct folio *folio)
287{
288 if (is_zero_folio(folio))
289 return;
290
291 /*
292 * Similar to try_grab_folio(): be sure to *also* increment the normal
293 * page refcount field at least once, so that the page really is
294 * pinned.
295 */
296 if (folio_test_large(folio)) {
297 WARN_ON_ONCE(atomic_read(&folio->_pincount) < 1);
298 folio_ref_inc(folio);
299 atomic_inc(&folio->_pincount);
300 } else {
301 WARN_ON_ONCE(folio_ref_count(folio) < GUP_PIN_COUNTING_BIAS);
302 folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
303 }
304}
305
306static inline struct folio *gup_folio_range_next(struct page *start,
307 unsigned long npages, unsigned long i, unsigned int *ntails)
308{
309 struct page *next = nth_page(start, i);
310 struct folio *folio = page_folio(next);
311 unsigned int nr = 1;
312
313 if (folio_test_large(folio))
314 nr = min_t(unsigned int, npages - i,
315 folio_nr_pages(folio) - folio_page_idx(folio, next));
316
317 *ntails = nr;
318 return folio;
319}
320
321static inline struct folio *gup_folio_next(struct page **list,
322 unsigned long npages, unsigned long i, unsigned int *ntails)
323{
324 struct folio *folio = page_folio(list[i]);
325 unsigned int nr;
326
327 for (nr = i + 1; nr < npages; nr++) {
328 if (page_folio(list[nr]) != folio)
329 break;
330 }
331
332 *ntails = nr - i;
333 return folio;
334}
335
336/**
337 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
338 * @pages: array of pages to be maybe marked dirty, and definitely released.
339 * @npages: number of pages in the @pages array.
340 * @make_dirty: whether to mark the pages dirty
341 *
342 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
343 * variants called on that page.
344 *
345 * For each page in the @pages array, make that page (or its head page, if a
346 * compound page) dirty, if @make_dirty is true, and if the page was previously
347 * listed as clean. In any case, releases all pages using unpin_user_page(),
348 * possibly via unpin_user_pages(), for the non-dirty case.
349 *
350 * Please see the unpin_user_page() documentation for details.
351 *
352 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
353 * required, then the caller should a) verify that this is really correct,
354 * because _lock() is usually required, and b) hand code it:
355 * set_page_dirty_lock(), unpin_user_page().
356 *
357 */
358void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
359 bool make_dirty)
360{
361 unsigned long i;
362 struct folio *folio;
363 unsigned int nr;
364
365 if (!make_dirty) {
366 unpin_user_pages(pages, npages);
367 return;
368 }
369
370 sanity_check_pinned_pages(pages, npages);
371 for (i = 0; i < npages; i += nr) {
372 folio = gup_folio_next(pages, npages, i, &nr);
373 /*
374 * Checking PageDirty at this point may race with
375 * clear_page_dirty_for_io(), but that's OK. Two key
376 * cases:
377 *
378 * 1) This code sees the page as already dirty, so it
379 * skips the call to set_page_dirty(). That could happen
380 * because clear_page_dirty_for_io() called
381 * page_mkclean(), followed by set_page_dirty().
382 * However, now the page is going to get written back,
383 * which meets the original intention of setting it
384 * dirty, so all is well: clear_page_dirty_for_io() goes
385 * on to call TestClearPageDirty(), and write the page
386 * back.
387 *
388 * 2) This code sees the page as clean, so it calls
389 * set_page_dirty(). The page stays dirty, despite being
390 * written back, so it gets written back again in the
391 * next writeback cycle. This is harmless.
392 */
393 if (!folio_test_dirty(folio)) {
394 folio_lock(folio);
395 folio_mark_dirty(folio);
396 folio_unlock(folio);
397 }
398 gup_put_folio(folio, nr, FOLL_PIN);
399 }
400}
401EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
402
403/**
404 * unpin_user_page_range_dirty_lock() - release and optionally dirty
405 * gup-pinned page range
406 *
407 * @page: the starting page of a range maybe marked dirty, and definitely released.
408 * @npages: number of consecutive pages to release.
409 * @make_dirty: whether to mark the pages dirty
410 *
411 * "gup-pinned page range" refers to a range of pages that has had one of the
412 * pin_user_pages() variants called on that page.
413 *
414 * For the page ranges defined by [page .. page+npages], make that range (or
415 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
416 * page range was previously listed as clean.
417 *
418 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
419 * required, then the caller should a) verify that this is really correct,
420 * because _lock() is usually required, and b) hand code it:
421 * set_page_dirty_lock(), unpin_user_page().
422 *
423 */
424void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages,
425 bool make_dirty)
426{
427 unsigned long i;
428 struct folio *folio;
429 unsigned int nr;
430
431 for (i = 0; i < npages; i += nr) {
432 folio = gup_folio_range_next(page, npages, i, &nr);
433 if (make_dirty && !folio_test_dirty(folio)) {
434 folio_lock(folio);
435 folio_mark_dirty(folio);
436 folio_unlock(folio);
437 }
438 gup_put_folio(folio, nr, FOLL_PIN);
439 }
440}
441EXPORT_SYMBOL(unpin_user_page_range_dirty_lock);
442
443static void unpin_user_pages_lockless(struct page **pages, unsigned long npages)
444{
445 unsigned long i;
446 struct folio *folio;
447 unsigned int nr;
448
449 /*
450 * Don't perform any sanity checks because we might have raced with
451 * fork() and some anonymous pages might now actually be shared --
452 * which is why we're unpinning after all.
453 */
454 for (i = 0; i < npages; i += nr) {
455 folio = gup_folio_next(pages, npages, i, &nr);
456 gup_put_folio(folio, nr, FOLL_PIN);
457 }
458}
459
460/**
461 * unpin_user_pages() - release an array of gup-pinned pages.
462 * @pages: array of pages to be marked dirty and released.
463 * @npages: number of pages in the @pages array.
464 *
465 * For each page in the @pages array, release the page using unpin_user_page().
466 *
467 * Please see the unpin_user_page() documentation for details.
468 */
469void unpin_user_pages(struct page **pages, unsigned long npages)
470{
471 unsigned long i;
472 struct folio *folio;
473 unsigned int nr;
474
475 /*
476 * If this WARN_ON() fires, then the system *might* be leaking pages (by
477 * leaving them pinned), but probably not. More likely, gup/pup returned
478 * a hard -ERRNO error to the caller, who erroneously passed it here.
479 */
480 if (WARN_ON(IS_ERR_VALUE(npages)))
481 return;
482
483 sanity_check_pinned_pages(pages, npages);
484 for (i = 0; i < npages; i += nr) {
485 folio = gup_folio_next(pages, npages, i, &nr);
486 gup_put_folio(folio, nr, FOLL_PIN);
487 }
488}
489EXPORT_SYMBOL(unpin_user_pages);
490
491/*
492 * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's
493 * lifecycle. Avoid setting the bit unless necessary, or it might cause write
494 * cache bouncing on large SMP machines for concurrent pinned gups.
495 */
496static inline void mm_set_has_pinned_flag(unsigned long *mm_flags)
497{
498 if (!test_bit(MMF_HAS_PINNED, mm_flags))
499 set_bit(MMF_HAS_PINNED, mm_flags);
500}
501
502#ifdef CONFIG_MMU
503static struct page *no_page_table(struct vm_area_struct *vma,
504 unsigned int flags)
505{
506 /*
507 * When core dumping an enormous anonymous area that nobody
508 * has touched so far, we don't want to allocate unnecessary pages or
509 * page tables. Return error instead of NULL to skip handle_mm_fault,
510 * then get_dump_page() will return NULL to leave a hole in the dump.
511 * But we can only make this optimization where a hole would surely
512 * be zero-filled if handle_mm_fault() actually did handle it.
513 */
514 if ((flags & FOLL_DUMP) &&
515 (vma_is_anonymous(vma) || !vma->vm_ops->fault))
516 return ERR_PTR(-EFAULT);
517 return NULL;
518}
519
520static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
521 pte_t *pte, unsigned int flags)
522{
523 if (flags & FOLL_TOUCH) {
524 pte_t orig_entry = ptep_get(pte);
525 pte_t entry = orig_entry;
526
527 if (flags & FOLL_WRITE)
528 entry = pte_mkdirty(entry);
529 entry = pte_mkyoung(entry);
530
531 if (!pte_same(orig_entry, entry)) {
532 set_pte_at(vma->vm_mm, address, pte, entry);
533 update_mmu_cache(vma, address, pte);
534 }
535 }
536
537 /* Proper page table entry exists, but no corresponding struct page */
538 return -EEXIST;
539}
540
541/* FOLL_FORCE can write to even unwritable PTEs in COW mappings. */
542static inline bool can_follow_write_pte(pte_t pte, struct page *page,
543 struct vm_area_struct *vma,
544 unsigned int flags)
545{
546 /* If the pte is writable, we can write to the page. */
547 if (pte_write(pte))
548 return true;
549
550 /* Maybe FOLL_FORCE is set to override it? */
551 if (!(flags & FOLL_FORCE))
552 return false;
553
554 /* But FOLL_FORCE has no effect on shared mappings */
555 if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
556 return false;
557
558 /* ... or read-only private ones */
559 if (!(vma->vm_flags & VM_MAYWRITE))
560 return false;
561
562 /* ... or already writable ones that just need to take a write fault */
563 if (vma->vm_flags & VM_WRITE)
564 return false;
565
566 /*
567 * See can_change_pte_writable(): we broke COW and could map the page
568 * writable if we have an exclusive anonymous page ...
569 */
570 if (!page || !PageAnon(page) || !PageAnonExclusive(page))
571 return false;
572
573 /* ... and a write-fault isn't required for other reasons. */
574 if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte))
575 return false;
576 return !userfaultfd_pte_wp(vma, pte);
577}
578
579static struct page *follow_page_pte(struct vm_area_struct *vma,
580 unsigned long address, pmd_t *pmd, unsigned int flags,
581 struct dev_pagemap **pgmap)
582{
583 struct mm_struct *mm = vma->vm_mm;
584 struct page *page;
585 spinlock_t *ptl;
586 pte_t *ptep, pte;
587 int ret;
588
589 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
590 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
591 (FOLL_PIN | FOLL_GET)))
592 return ERR_PTR(-EINVAL);
593
594 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
595 if (!ptep)
596 return no_page_table(vma, flags);
597 pte = ptep_get(ptep);
598 if (!pte_present(pte))
599 goto no_page;
600 if (pte_protnone(pte) && !gup_can_follow_protnone(vma, flags))
601 goto no_page;
602
603 page = vm_normal_page(vma, address, pte);
604
605 /*
606 * We only care about anon pages in can_follow_write_pte() and don't
607 * have to worry about pte_devmap() because they are never anon.
608 */
609 if ((flags & FOLL_WRITE) &&
610 !can_follow_write_pte(pte, page, vma, flags)) {
611 page = NULL;
612 goto out;
613 }
614
615 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
616 /*
617 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
618 * case since they are only valid while holding the pgmap
619 * reference.
620 */
621 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
622 if (*pgmap)
623 page = pte_page(pte);
624 else
625 goto no_page;
626 } else if (unlikely(!page)) {
627 if (flags & FOLL_DUMP) {
628 /* Avoid special (like zero) pages in core dumps */
629 page = ERR_PTR(-EFAULT);
630 goto out;
631 }
632
633 if (is_zero_pfn(pte_pfn(pte))) {
634 page = pte_page(pte);
635 } else {
636 ret = follow_pfn_pte(vma, address, ptep, flags);
637 page = ERR_PTR(ret);
638 goto out;
639 }
640 }
641
642 if (!pte_write(pte) && gup_must_unshare(vma, flags, page)) {
643 page = ERR_PTR(-EMLINK);
644 goto out;
645 }
646
647 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
648 !PageAnonExclusive(page), page);
649
650 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
651 ret = try_grab_page(page, flags);
652 if (unlikely(ret)) {
653 page = ERR_PTR(ret);
654 goto out;
655 }
656
657 /*
658 * We need to make the page accessible if and only if we are going
659 * to access its content (the FOLL_PIN case). Please see
660 * Documentation/core-api/pin_user_pages.rst for details.
661 */
662 if (flags & FOLL_PIN) {
663 ret = arch_make_page_accessible(page);
664 if (ret) {
665 unpin_user_page(page);
666 page = ERR_PTR(ret);
667 goto out;
668 }
669 }
670 if (flags & FOLL_TOUCH) {
671 if ((flags & FOLL_WRITE) &&
672 !pte_dirty(pte) && !PageDirty(page))
673 set_page_dirty(page);
674 /*
675 * pte_mkyoung() would be more correct here, but atomic care
676 * is needed to avoid losing the dirty bit: it is easier to use
677 * mark_page_accessed().
678 */
679 mark_page_accessed(page);
680 }
681out:
682 pte_unmap_unlock(ptep, ptl);
683 return page;
684no_page:
685 pte_unmap_unlock(ptep, ptl);
686 if (!pte_none(pte))
687 return NULL;
688 return no_page_table(vma, flags);
689}
690
691static struct page *follow_pmd_mask(struct vm_area_struct *vma,
692 unsigned long address, pud_t *pudp,
693 unsigned int flags,
694 struct follow_page_context *ctx)
695{
696 pmd_t *pmd, pmdval;
697 spinlock_t *ptl;
698 struct page *page;
699 struct mm_struct *mm = vma->vm_mm;
700
701 pmd = pmd_offset(pudp, address);
702 pmdval = pmdp_get_lockless(pmd);
703 if (pmd_none(pmdval))
704 return no_page_table(vma, flags);
705 if (!pmd_present(pmdval))
706 return no_page_table(vma, flags);
707 if (pmd_devmap(pmdval)) {
708 ptl = pmd_lock(mm, pmd);
709 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
710 spin_unlock(ptl);
711 if (page)
712 return page;
713 return no_page_table(vma, flags);
714 }
715 if (likely(!pmd_trans_huge(pmdval)))
716 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
717
718 if (pmd_protnone(pmdval) && !gup_can_follow_protnone(vma, flags))
719 return no_page_table(vma, flags);
720
721 ptl = pmd_lock(mm, pmd);
722 if (unlikely(!pmd_present(*pmd))) {
723 spin_unlock(ptl);
724 return no_page_table(vma, flags);
725 }
726 if (unlikely(!pmd_trans_huge(*pmd))) {
727 spin_unlock(ptl);
728 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
729 }
730 if (flags & FOLL_SPLIT_PMD) {
731 spin_unlock(ptl);
732 split_huge_pmd(vma, pmd, address);
733 /* If pmd was left empty, stuff a page table in there quickly */
734 return pte_alloc(mm, pmd) ? ERR_PTR(-ENOMEM) :
735 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
736 }
737 page = follow_trans_huge_pmd(vma, address, pmd, flags);
738 spin_unlock(ptl);
739 ctx->page_mask = HPAGE_PMD_NR - 1;
740 return page;
741}
742
743static struct page *follow_pud_mask(struct vm_area_struct *vma,
744 unsigned long address, p4d_t *p4dp,
745 unsigned int flags,
746 struct follow_page_context *ctx)
747{
748 pud_t *pud;
749 spinlock_t *ptl;
750 struct page *page;
751 struct mm_struct *mm = vma->vm_mm;
752
753 pud = pud_offset(p4dp, address);
754 if (pud_none(*pud))
755 return no_page_table(vma, flags);
756 if (pud_devmap(*pud)) {
757 ptl = pud_lock(mm, pud);
758 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
759 spin_unlock(ptl);
760 if (page)
761 return page;
762 return no_page_table(vma, flags);
763 }
764 if (unlikely(pud_bad(*pud)))
765 return no_page_table(vma, flags);
766
767 return follow_pmd_mask(vma, address, pud, flags, ctx);
768}
769
770static struct page *follow_p4d_mask(struct vm_area_struct *vma,
771 unsigned long address, pgd_t *pgdp,
772 unsigned int flags,
773 struct follow_page_context *ctx)
774{
775 p4d_t *p4d;
776
777 p4d = p4d_offset(pgdp, address);
778 if (p4d_none(*p4d))
779 return no_page_table(vma, flags);
780 BUILD_BUG_ON(p4d_huge(*p4d));
781 if (unlikely(p4d_bad(*p4d)))
782 return no_page_table(vma, flags);
783
784 return follow_pud_mask(vma, address, p4d, flags, ctx);
785}
786
787/**
788 * follow_page_mask - look up a page descriptor from a user-virtual address
789 * @vma: vm_area_struct mapping @address
790 * @address: virtual address to look up
791 * @flags: flags modifying lookup behaviour
792 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
793 * pointer to output page_mask
794 *
795 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
796 *
797 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
798 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
799 *
800 * When getting an anonymous page and the caller has to trigger unsharing
801 * of a shared anonymous page first, -EMLINK is returned. The caller should
802 * trigger a fault with FAULT_FLAG_UNSHARE set. Note that unsharing is only
803 * relevant with FOLL_PIN and !FOLL_WRITE.
804 *
805 * On output, the @ctx->page_mask is set according to the size of the page.
806 *
807 * Return: the mapped (struct page *), %NULL if no mapping exists, or
808 * an error pointer if there is a mapping to something not represented
809 * by a page descriptor (see also vm_normal_page()).
810 */
811static struct page *follow_page_mask(struct vm_area_struct *vma,
812 unsigned long address, unsigned int flags,
813 struct follow_page_context *ctx)
814{
815 pgd_t *pgd;
816 struct mm_struct *mm = vma->vm_mm;
817
818 ctx->page_mask = 0;
819
820 /*
821 * Call hugetlb_follow_page_mask for hugetlb vmas as it will use
822 * special hugetlb page table walking code. This eliminates the
823 * need to check for hugetlb entries in the general walking code.
824 */
825 if (is_vm_hugetlb_page(vma))
826 return hugetlb_follow_page_mask(vma, address, flags,
827 &ctx->page_mask);
828
829 pgd = pgd_offset(mm, address);
830
831 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
832 return no_page_table(vma, flags);
833
834 return follow_p4d_mask(vma, address, pgd, flags, ctx);
835}
836
837struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
838 unsigned int foll_flags)
839{
840 struct follow_page_context ctx = { NULL };
841 struct page *page;
842
843 if (vma_is_secretmem(vma))
844 return NULL;
845
846 if (WARN_ON_ONCE(foll_flags & FOLL_PIN))
847 return NULL;
848
849 /*
850 * We never set FOLL_HONOR_NUMA_FAULT because callers don't expect
851 * to fail on PROT_NONE-mapped pages.
852 */
853 page = follow_page_mask(vma, address, foll_flags, &ctx);
854 if (ctx.pgmap)
855 put_dev_pagemap(ctx.pgmap);
856 return page;
857}
858
859static int get_gate_page(struct mm_struct *mm, unsigned long address,
860 unsigned int gup_flags, struct vm_area_struct **vma,
861 struct page **page)
862{
863 pgd_t *pgd;
864 p4d_t *p4d;
865 pud_t *pud;
866 pmd_t *pmd;
867 pte_t *pte;
868 pte_t entry;
869 int ret = -EFAULT;
870
871 /* user gate pages are read-only */
872 if (gup_flags & FOLL_WRITE)
873 return -EFAULT;
874 if (address > TASK_SIZE)
875 pgd = pgd_offset_k(address);
876 else
877 pgd = pgd_offset_gate(mm, address);
878 if (pgd_none(*pgd))
879 return -EFAULT;
880 p4d = p4d_offset(pgd, address);
881 if (p4d_none(*p4d))
882 return -EFAULT;
883 pud = pud_offset(p4d, address);
884 if (pud_none(*pud))
885 return -EFAULT;
886 pmd = pmd_offset(pud, address);
887 if (!pmd_present(*pmd))
888 return -EFAULT;
889 pte = pte_offset_map(pmd, address);
890 if (!pte)
891 return -EFAULT;
892 entry = ptep_get(pte);
893 if (pte_none(entry))
894 goto unmap;
895 *vma = get_gate_vma(mm);
896 if (!page)
897 goto out;
898 *page = vm_normal_page(*vma, address, entry);
899 if (!*page) {
900 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(entry)))
901 goto unmap;
902 *page = pte_page(entry);
903 }
904 ret = try_grab_page(*page, gup_flags);
905 if (unlikely(ret))
906 goto unmap;
907out:
908 ret = 0;
909unmap:
910 pte_unmap(pte);
911 return ret;
912}
913
914/*
915 * mmap_lock must be held on entry. If @flags has FOLL_UNLOCKABLE but not
916 * FOLL_NOWAIT, the mmap_lock may be released. If it is, *@locked will be set
917 * to 0 and -EBUSY returned.
918 */
919static int faultin_page(struct vm_area_struct *vma,
920 unsigned long address, unsigned int *flags, bool unshare,
921 int *locked)
922{
923 unsigned int fault_flags = 0;
924 vm_fault_t ret;
925
926 if (*flags & FOLL_NOFAULT)
927 return -EFAULT;
928 if (*flags & FOLL_WRITE)
929 fault_flags |= FAULT_FLAG_WRITE;
930 if (*flags & FOLL_REMOTE)
931 fault_flags |= FAULT_FLAG_REMOTE;
932 if (*flags & FOLL_UNLOCKABLE) {
933 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
934 /*
935 * FAULT_FLAG_INTERRUPTIBLE is opt-in. GUP callers must set
936 * FOLL_INTERRUPTIBLE to enable FAULT_FLAG_INTERRUPTIBLE.
937 * That's because some callers may not be prepared to
938 * handle early exits caused by non-fatal signals.
939 */
940 if (*flags & FOLL_INTERRUPTIBLE)
941 fault_flags |= FAULT_FLAG_INTERRUPTIBLE;
942 }
943 if (*flags & FOLL_NOWAIT)
944 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
945 if (*flags & FOLL_TRIED) {
946 /*
947 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
948 * can co-exist
949 */
950 fault_flags |= FAULT_FLAG_TRIED;
951 }
952 if (unshare) {
953 fault_flags |= FAULT_FLAG_UNSHARE;
954 /* FAULT_FLAG_WRITE and FAULT_FLAG_UNSHARE are incompatible */
955 VM_BUG_ON(fault_flags & FAULT_FLAG_WRITE);
956 }
957
958 ret = handle_mm_fault(vma, address, fault_flags, NULL);
959
960 if (ret & VM_FAULT_COMPLETED) {
961 /*
962 * With FAULT_FLAG_RETRY_NOWAIT we'll never release the
963 * mmap lock in the page fault handler. Sanity check this.
964 */
965 WARN_ON_ONCE(fault_flags & FAULT_FLAG_RETRY_NOWAIT);
966 *locked = 0;
967
968 /*
969 * We should do the same as VM_FAULT_RETRY, but let's not
970 * return -EBUSY since that's not reflecting the reality of
971 * what has happened - we've just fully completed a page
972 * fault, with the mmap lock released. Use -EAGAIN to show
973 * that we want to take the mmap lock _again_.
974 */
975 return -EAGAIN;
976 }
977
978 if (ret & VM_FAULT_ERROR) {
979 int err = vm_fault_to_errno(ret, *flags);
980
981 if (err)
982 return err;
983 BUG();
984 }
985
986 if (ret & VM_FAULT_RETRY) {
987 if (!(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
988 *locked = 0;
989 return -EBUSY;
990 }
991
992 return 0;
993}
994
995/*
996 * Writing to file-backed mappings which require folio dirty tracking using GUP
997 * is a fundamentally broken operation, as kernel write access to GUP mappings
998 * do not adhere to the semantics expected by a file system.
999 *
1000 * Consider the following scenario:-
1001 *
1002 * 1. A folio is written to via GUP which write-faults the memory, notifying
1003 * the file system and dirtying the folio.
1004 * 2. Later, writeback is triggered, resulting in the folio being cleaned and
1005 * the PTE being marked read-only.
1006 * 3. The GUP caller writes to the folio, as it is mapped read/write via the
1007 * direct mapping.
1008 * 4. The GUP caller, now done with the page, unpins it and sets it dirty
1009 * (though it does not have to).
1010 *
1011 * This results in both data being written to a folio without writenotify, and
1012 * the folio being dirtied unexpectedly (if the caller decides to do so).
1013 */
1014static bool writable_file_mapping_allowed(struct vm_area_struct *vma,
1015 unsigned long gup_flags)
1016{
1017 /*
1018 * If we aren't pinning then no problematic write can occur. A long term
1019 * pin is the most egregious case so this is the case we disallow.
1020 */
1021 if ((gup_flags & (FOLL_PIN | FOLL_LONGTERM)) !=
1022 (FOLL_PIN | FOLL_LONGTERM))
1023 return true;
1024
1025 /*
1026 * If the VMA does not require dirty tracking then no problematic write
1027 * can occur either.
1028 */
1029 return !vma_needs_dirty_tracking(vma);
1030}
1031
1032static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
1033{
1034 vm_flags_t vm_flags = vma->vm_flags;
1035 int write = (gup_flags & FOLL_WRITE);
1036 int foreign = (gup_flags & FOLL_REMOTE);
1037 bool vma_anon = vma_is_anonymous(vma);
1038
1039 if (vm_flags & (VM_IO | VM_PFNMAP))
1040 return -EFAULT;
1041
1042 if ((gup_flags & FOLL_ANON) && !vma_anon)
1043 return -EFAULT;
1044
1045 if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma))
1046 return -EOPNOTSUPP;
1047
1048 if (vma_is_secretmem(vma))
1049 return -EFAULT;
1050
1051 if (write) {
1052 if (!vma_anon &&
1053 !writable_file_mapping_allowed(vma, gup_flags))
1054 return -EFAULT;
1055
1056 if (!(vm_flags & VM_WRITE) || (vm_flags & VM_SHADOW_STACK)) {
1057 if (!(gup_flags & FOLL_FORCE))
1058 return -EFAULT;
1059 /* hugetlb does not support FOLL_FORCE|FOLL_WRITE. */
1060 if (is_vm_hugetlb_page(vma))
1061 return -EFAULT;
1062 /*
1063 * We used to let the write,force case do COW in a
1064 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
1065 * set a breakpoint in a read-only mapping of an
1066 * executable, without corrupting the file (yet only
1067 * when that file had been opened for writing!).
1068 * Anon pages in shared mappings are surprising: now
1069 * just reject it.
1070 */
1071 if (!is_cow_mapping(vm_flags))
1072 return -EFAULT;
1073 }
1074 } else if (!(vm_flags & VM_READ)) {
1075 if (!(gup_flags & FOLL_FORCE))
1076 return -EFAULT;
1077 /*
1078 * Is there actually any vma we can reach here which does not
1079 * have VM_MAYREAD set?
1080 */
1081 if (!(vm_flags & VM_MAYREAD))
1082 return -EFAULT;
1083 }
1084 /*
1085 * gups are always data accesses, not instruction
1086 * fetches, so execute=false here
1087 */
1088 if (!arch_vma_access_permitted(vma, write, false, foreign))
1089 return -EFAULT;
1090 return 0;
1091}
1092
1093/*
1094 * This is "vma_lookup()", but with a warning if we would have
1095 * historically expanded the stack in the GUP code.
1096 */
1097static struct vm_area_struct *gup_vma_lookup(struct mm_struct *mm,
1098 unsigned long addr)
1099{
1100#ifdef CONFIG_STACK_GROWSUP
1101 return vma_lookup(mm, addr);
1102#else
1103 static volatile unsigned long next_warn;
1104 struct vm_area_struct *vma;
1105 unsigned long now, next;
1106
1107 vma = find_vma(mm, addr);
1108 if (!vma || (addr >= vma->vm_start))
1109 return vma;
1110
1111 /* Only warn for half-way relevant accesses */
1112 if (!(vma->vm_flags & VM_GROWSDOWN))
1113 return NULL;
1114 if (vma->vm_start - addr > 65536)
1115 return NULL;
1116
1117 /* Let's not warn more than once an hour.. */
1118 now = jiffies; next = next_warn;
1119 if (next && time_before(now, next))
1120 return NULL;
1121 next_warn = now + 60*60*HZ;
1122
1123 /* Let people know things may have changed. */
1124 pr_warn("GUP no longer grows the stack in %s (%d): %lx-%lx (%lx)\n",
1125 current->comm, task_pid_nr(current),
1126 vma->vm_start, vma->vm_end, addr);
1127 dump_stack();
1128 return NULL;
1129#endif
1130}
1131
1132/**
1133 * __get_user_pages() - pin user pages in memory
1134 * @mm: mm_struct of target mm
1135 * @start: starting user address
1136 * @nr_pages: number of pages from start to pin
1137 * @gup_flags: flags modifying pin behaviour
1138 * @pages: array that receives pointers to the pages pinned.
1139 * Should be at least nr_pages long. Or NULL, if caller
1140 * only intends to ensure the pages are faulted in.
1141 * @locked: whether we're still with the mmap_lock held
1142 *
1143 * Returns either number of pages pinned (which may be less than the
1144 * number requested), or an error. Details about the return value:
1145 *
1146 * -- If nr_pages is 0, returns 0.
1147 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1148 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1149 * pages pinned. Again, this may be less than nr_pages.
1150 * -- 0 return value is possible when the fault would need to be retried.
1151 *
1152 * The caller is responsible for releasing returned @pages, via put_page().
1153 *
1154 * Must be called with mmap_lock held. It may be released. See below.
1155 *
1156 * __get_user_pages walks a process's page tables and takes a reference to
1157 * each struct page that each user address corresponds to at a given
1158 * instant. That is, it takes the page that would be accessed if a user
1159 * thread accesses the given user virtual address at that instant.
1160 *
1161 * This does not guarantee that the page exists in the user mappings when
1162 * __get_user_pages returns, and there may even be a completely different
1163 * page there in some cases (eg. if mmapped pagecache has been invalidated
1164 * and subsequently re-faulted). However it does guarantee that the page
1165 * won't be freed completely. And mostly callers simply care that the page
1166 * contains data that was valid *at some point in time*. Typically, an IO
1167 * or similar operation cannot guarantee anything stronger anyway because
1168 * locks can't be held over the syscall boundary.
1169 *
1170 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1171 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1172 * appropriate) must be called after the page is finished with, and
1173 * before put_page is called.
1174 *
1175 * If FOLL_UNLOCKABLE is set without FOLL_NOWAIT then the mmap_lock may
1176 * be released. If this happens *@locked will be set to 0 on return.
1177 *
1178 * A caller using such a combination of @gup_flags must therefore hold the
1179 * mmap_lock for reading only, and recognize when it's been released. Otherwise,
1180 * it must be held for either reading or writing and will not be released.
1181 *
1182 * In most cases, get_user_pages or get_user_pages_fast should be used
1183 * instead of __get_user_pages. __get_user_pages should be used only if
1184 * you need some special @gup_flags.
1185 */
1186static long __get_user_pages(struct mm_struct *mm,
1187 unsigned long start, unsigned long nr_pages,
1188 unsigned int gup_flags, struct page **pages,
1189 int *locked)
1190{
1191 long ret = 0, i = 0;
1192 struct vm_area_struct *vma = NULL;
1193 struct follow_page_context ctx = { NULL };
1194
1195 if (!nr_pages)
1196 return 0;
1197
1198 start = untagged_addr_remote(mm, start);
1199
1200 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1201
1202 do {
1203 struct page *page;
1204 unsigned int foll_flags = gup_flags;
1205 unsigned int page_increm;
1206
1207 /* first iteration or cross vma bound */
1208 if (!vma || start >= vma->vm_end) {
1209 /*
1210 * MADV_POPULATE_(READ|WRITE) wants to handle VMA
1211 * lookups+error reporting differently.
1212 */
1213 if (gup_flags & FOLL_MADV_POPULATE) {
1214 vma = vma_lookup(mm, start);
1215 if (!vma) {
1216 ret = -ENOMEM;
1217 goto out;
1218 }
1219 if (check_vma_flags(vma, gup_flags)) {
1220 ret = -EINVAL;
1221 goto out;
1222 }
1223 goto retry;
1224 }
1225 vma = gup_vma_lookup(mm, start);
1226 if (!vma && in_gate_area(mm, start)) {
1227 ret = get_gate_page(mm, start & PAGE_MASK,
1228 gup_flags, &vma,
1229 pages ? &page : NULL);
1230 if (ret)
1231 goto out;
1232 ctx.page_mask = 0;
1233 goto next_page;
1234 }
1235
1236 if (!vma) {
1237 ret = -EFAULT;
1238 goto out;
1239 }
1240 ret = check_vma_flags(vma, gup_flags);
1241 if (ret)
1242 goto out;
1243 }
1244retry:
1245 /*
1246 * If we have a pending SIGKILL, don't keep faulting pages and
1247 * potentially allocating memory.
1248 */
1249 if (fatal_signal_pending(current)) {
1250 ret = -EINTR;
1251 goto out;
1252 }
1253 cond_resched();
1254
1255 page = follow_page_mask(vma, start, foll_flags, &ctx);
1256 if (!page || PTR_ERR(page) == -EMLINK) {
1257 ret = faultin_page(vma, start, &foll_flags,
1258 PTR_ERR(page) == -EMLINK, locked);
1259 switch (ret) {
1260 case 0:
1261 goto retry;
1262 case -EBUSY:
1263 case -EAGAIN:
1264 ret = 0;
1265 fallthrough;
1266 case -EFAULT:
1267 case -ENOMEM:
1268 case -EHWPOISON:
1269 goto out;
1270 }
1271 BUG();
1272 } else if (PTR_ERR(page) == -EEXIST) {
1273 /*
1274 * Proper page table entry exists, but no corresponding
1275 * struct page. If the caller expects **pages to be
1276 * filled in, bail out now, because that can't be done
1277 * for this page.
1278 */
1279 if (pages) {
1280 ret = PTR_ERR(page);
1281 goto out;
1282 }
1283 } else if (IS_ERR(page)) {
1284 ret = PTR_ERR(page);
1285 goto out;
1286 }
1287next_page:
1288 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1289 if (page_increm > nr_pages)
1290 page_increm = nr_pages;
1291
1292 if (pages) {
1293 struct page *subpage;
1294 unsigned int j;
1295
1296 /*
1297 * This must be a large folio (and doesn't need to
1298 * be the whole folio; it can be part of it), do
1299 * the refcount work for all the subpages too.
1300 *
1301 * NOTE: here the page may not be the head page
1302 * e.g. when start addr is not thp-size aligned.
1303 * try_grab_folio() should have taken care of tail
1304 * pages.
1305 */
1306 if (page_increm > 1) {
1307 struct folio *folio;
1308
1309 /*
1310 * Since we already hold refcount on the
1311 * large folio, this should never fail.
1312 */
1313 folio = try_grab_folio(page, page_increm - 1,
1314 foll_flags);
1315 if (WARN_ON_ONCE(!folio)) {
1316 /*
1317 * Release the 1st page ref if the
1318 * folio is problematic, fail hard.
1319 */
1320 gup_put_folio(page_folio(page), 1,
1321 foll_flags);
1322 ret = -EFAULT;
1323 goto out;
1324 }
1325 }
1326
1327 for (j = 0; j < page_increm; j++) {
1328 subpage = nth_page(page, j);
1329 pages[i + j] = subpage;
1330 flush_anon_page(vma, subpage, start + j * PAGE_SIZE);
1331 flush_dcache_page(subpage);
1332 }
1333 }
1334
1335 i += page_increm;
1336 start += page_increm * PAGE_SIZE;
1337 nr_pages -= page_increm;
1338 } while (nr_pages);
1339out:
1340 if (ctx.pgmap)
1341 put_dev_pagemap(ctx.pgmap);
1342 return i ? i : ret;
1343}
1344
1345static bool vma_permits_fault(struct vm_area_struct *vma,
1346 unsigned int fault_flags)
1347{
1348 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1349 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1350 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1351
1352 if (!(vm_flags & vma->vm_flags))
1353 return false;
1354
1355 /*
1356 * The architecture might have a hardware protection
1357 * mechanism other than read/write that can deny access.
1358 *
1359 * gup always represents data access, not instruction
1360 * fetches, so execute=false here:
1361 */
1362 if (!arch_vma_access_permitted(vma, write, false, foreign))
1363 return false;
1364
1365 return true;
1366}
1367
1368/**
1369 * fixup_user_fault() - manually resolve a user page fault
1370 * @mm: mm_struct of target mm
1371 * @address: user address
1372 * @fault_flags:flags to pass down to handle_mm_fault()
1373 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1374 * does not allow retry. If NULL, the caller must guarantee
1375 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1376 *
1377 * This is meant to be called in the specific scenario where for locking reasons
1378 * we try to access user memory in atomic context (within a pagefault_disable()
1379 * section), this returns -EFAULT, and we want to resolve the user fault before
1380 * trying again.
1381 *
1382 * Typically this is meant to be used by the futex code.
1383 *
1384 * The main difference with get_user_pages() is that this function will
1385 * unconditionally call handle_mm_fault() which will in turn perform all the
1386 * necessary SW fixup of the dirty and young bits in the PTE, while
1387 * get_user_pages() only guarantees to update these in the struct page.
1388 *
1389 * This is important for some architectures where those bits also gate the
1390 * access permission to the page because they are maintained in software. On
1391 * such architectures, gup() will not be enough to make a subsequent access
1392 * succeed.
1393 *
1394 * This function will not return with an unlocked mmap_lock. So it has not the
1395 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1396 */
1397int fixup_user_fault(struct mm_struct *mm,
1398 unsigned long address, unsigned int fault_flags,
1399 bool *unlocked)
1400{
1401 struct vm_area_struct *vma;
1402 vm_fault_t ret;
1403
1404 address = untagged_addr_remote(mm, address);
1405
1406 if (unlocked)
1407 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1408
1409retry:
1410 vma = gup_vma_lookup(mm, address);
1411 if (!vma)
1412 return -EFAULT;
1413
1414 if (!vma_permits_fault(vma, fault_flags))
1415 return -EFAULT;
1416
1417 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1418 fatal_signal_pending(current))
1419 return -EINTR;
1420
1421 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1422
1423 if (ret & VM_FAULT_COMPLETED) {
1424 /*
1425 * NOTE: it's a pity that we need to retake the lock here
1426 * to pair with the unlock() in the callers. Ideally we
1427 * could tell the callers so they do not need to unlock.
1428 */
1429 mmap_read_lock(mm);
1430 *unlocked = true;
1431 return 0;
1432 }
1433
1434 if (ret & VM_FAULT_ERROR) {
1435 int err = vm_fault_to_errno(ret, 0);
1436
1437 if (err)
1438 return err;
1439 BUG();
1440 }
1441
1442 if (ret & VM_FAULT_RETRY) {
1443 mmap_read_lock(mm);
1444 *unlocked = true;
1445 fault_flags |= FAULT_FLAG_TRIED;
1446 goto retry;
1447 }
1448
1449 return 0;
1450}
1451EXPORT_SYMBOL_GPL(fixup_user_fault);
1452
1453/*
1454 * GUP always responds to fatal signals. When FOLL_INTERRUPTIBLE is
1455 * specified, it'll also respond to generic signals. The caller of GUP
1456 * that has FOLL_INTERRUPTIBLE should take care of the GUP interruption.
1457 */
1458static bool gup_signal_pending(unsigned int flags)
1459{
1460 if (fatal_signal_pending(current))
1461 return true;
1462
1463 if (!(flags & FOLL_INTERRUPTIBLE))
1464 return false;
1465
1466 return signal_pending(current);
1467}
1468
1469/*
1470 * Locking: (*locked == 1) means that the mmap_lock has already been acquired by
1471 * the caller. This function may drop the mmap_lock. If it does so, then it will
1472 * set (*locked = 0).
1473 *
1474 * (*locked == 0) means that the caller expects this function to acquire and
1475 * drop the mmap_lock. Therefore, the value of *locked will still be zero when
1476 * the function returns, even though it may have changed temporarily during
1477 * function execution.
1478 *
1479 * Please note that this function, unlike __get_user_pages(), will not return 0
1480 * for nr_pages > 0, unless FOLL_NOWAIT is used.
1481 */
1482static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1483 unsigned long start,
1484 unsigned long nr_pages,
1485 struct page **pages,
1486 int *locked,
1487 unsigned int flags)
1488{
1489 long ret, pages_done;
1490 bool must_unlock = false;
1491
1492 if (!nr_pages)
1493 return 0;
1494
1495 /*
1496 * The internal caller expects GUP to manage the lock internally and the
1497 * lock must be released when this returns.
1498 */
1499 if (!*locked) {
1500 if (mmap_read_lock_killable(mm))
1501 return -EAGAIN;
1502 must_unlock = true;
1503 *locked = 1;
1504 }
1505 else
1506 mmap_assert_locked(mm);
1507
1508 if (flags & FOLL_PIN)
1509 mm_set_has_pinned_flag(&mm->flags);
1510
1511 /*
1512 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1513 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1514 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1515 * for FOLL_GET, not for the newer FOLL_PIN.
1516 *
1517 * FOLL_PIN always expects pages to be non-null, but no need to assert
1518 * that here, as any failures will be obvious enough.
1519 */
1520 if (pages && !(flags & FOLL_PIN))
1521 flags |= FOLL_GET;
1522
1523 pages_done = 0;
1524 for (;;) {
1525 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1526 locked);
1527 if (!(flags & FOLL_UNLOCKABLE)) {
1528 /* VM_FAULT_RETRY couldn't trigger, bypass */
1529 pages_done = ret;
1530 break;
1531 }
1532
1533 /* VM_FAULT_RETRY or VM_FAULT_COMPLETED cannot return errors */
1534 if (!*locked) {
1535 BUG_ON(ret < 0);
1536 BUG_ON(ret >= nr_pages);
1537 }
1538
1539 if (ret > 0) {
1540 nr_pages -= ret;
1541 pages_done += ret;
1542 if (!nr_pages)
1543 break;
1544 }
1545 if (*locked) {
1546 /*
1547 * VM_FAULT_RETRY didn't trigger or it was a
1548 * FOLL_NOWAIT.
1549 */
1550 if (!pages_done)
1551 pages_done = ret;
1552 break;
1553 }
1554 /*
1555 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1556 * For the prefault case (!pages) we only update counts.
1557 */
1558 if (likely(pages))
1559 pages += ret;
1560 start += ret << PAGE_SHIFT;
1561
1562 /* The lock was temporarily dropped, so we must unlock later */
1563 must_unlock = true;
1564
1565retry:
1566 /*
1567 * Repeat on the address that fired VM_FAULT_RETRY
1568 * with both FAULT_FLAG_ALLOW_RETRY and
1569 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1570 * by fatal signals of even common signals, depending on
1571 * the caller's request. So we need to check it before we
1572 * start trying again otherwise it can loop forever.
1573 */
1574 if (gup_signal_pending(flags)) {
1575 if (!pages_done)
1576 pages_done = -EINTR;
1577 break;
1578 }
1579
1580 ret = mmap_read_lock_killable(mm);
1581 if (ret) {
1582 BUG_ON(ret > 0);
1583 if (!pages_done)
1584 pages_done = ret;
1585 break;
1586 }
1587
1588 *locked = 1;
1589 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1590 pages, locked);
1591 if (!*locked) {
1592 /* Continue to retry until we succeeded */
1593 BUG_ON(ret != 0);
1594 goto retry;
1595 }
1596 if (ret != 1) {
1597 BUG_ON(ret > 1);
1598 if (!pages_done)
1599 pages_done = ret;
1600 break;
1601 }
1602 nr_pages--;
1603 pages_done++;
1604 if (!nr_pages)
1605 break;
1606 if (likely(pages))
1607 pages++;
1608 start += PAGE_SIZE;
1609 }
1610 if (must_unlock && *locked) {
1611 /*
1612 * We either temporarily dropped the lock, or the caller
1613 * requested that we both acquire and drop the lock. Either way,
1614 * we must now unlock, and notify the caller of that state.
1615 */
1616 mmap_read_unlock(mm);
1617 *locked = 0;
1618 }
1619
1620 /*
1621 * Failing to pin anything implies something has gone wrong (except when
1622 * FOLL_NOWAIT is specified).
1623 */
1624 if (WARN_ON_ONCE(pages_done == 0 && !(flags & FOLL_NOWAIT)))
1625 return -EFAULT;
1626
1627 return pages_done;
1628}
1629
1630/**
1631 * populate_vma_page_range() - populate a range of pages in the vma.
1632 * @vma: target vma
1633 * @start: start address
1634 * @end: end address
1635 * @locked: whether the mmap_lock is still held
1636 *
1637 * This takes care of mlocking the pages too if VM_LOCKED is set.
1638 *
1639 * Return either number of pages pinned in the vma, or a negative error
1640 * code on error.
1641 *
1642 * vma->vm_mm->mmap_lock must be held.
1643 *
1644 * If @locked is NULL, it may be held for read or write and will
1645 * be unperturbed.
1646 *
1647 * If @locked is non-NULL, it must held for read only and may be
1648 * released. If it's released, *@locked will be set to 0.
1649 */
1650long populate_vma_page_range(struct vm_area_struct *vma,
1651 unsigned long start, unsigned long end, int *locked)
1652{
1653 struct mm_struct *mm = vma->vm_mm;
1654 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1655 int local_locked = 1;
1656 int gup_flags;
1657 long ret;
1658
1659 VM_BUG_ON(!PAGE_ALIGNED(start));
1660 VM_BUG_ON(!PAGE_ALIGNED(end));
1661 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1662 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1663 mmap_assert_locked(mm);
1664
1665 /*
1666 * Rightly or wrongly, the VM_LOCKONFAULT case has never used
1667 * faultin_page() to break COW, so it has no work to do here.
1668 */
1669 if (vma->vm_flags & VM_LOCKONFAULT)
1670 return nr_pages;
1671
1672 /* ... similarly, we've never faulted in PROT_NONE pages */
1673 if (!vma_is_accessible(vma))
1674 return -EFAULT;
1675
1676 gup_flags = FOLL_TOUCH;
1677 /*
1678 * We want to touch writable mappings with a write fault in order
1679 * to break COW, except for shared mappings because these don't COW
1680 * and we would not want to dirty them for nothing.
1681 *
1682 * Otherwise, do a read fault, and use FOLL_FORCE in case it's not
1683 * readable (ie write-only or executable).
1684 */
1685 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1686 gup_flags |= FOLL_WRITE;
1687 else
1688 gup_flags |= FOLL_FORCE;
1689
1690 if (locked)
1691 gup_flags |= FOLL_UNLOCKABLE;
1692
1693 /*
1694 * We made sure addr is within a VMA, so the following will
1695 * not result in a stack expansion that recurses back here.
1696 */
1697 ret = __get_user_pages(mm, start, nr_pages, gup_flags,
1698 NULL, locked ? locked : &local_locked);
1699 lru_add_drain();
1700 return ret;
1701}
1702
1703/*
1704 * faultin_page_range() - populate (prefault) page tables inside the
1705 * given range readable/writable
1706 *
1707 * This takes care of mlocking the pages, too, if VM_LOCKED is set.
1708 *
1709 * @mm: the mm to populate page tables in
1710 * @start: start address
1711 * @end: end address
1712 * @write: whether to prefault readable or writable
1713 * @locked: whether the mmap_lock is still held
1714 *
1715 * Returns either number of processed pages in the MM, or a negative error
1716 * code on error (see __get_user_pages()). Note that this function reports
1717 * errors related to VMAs, such as incompatible mappings, as expected by
1718 * MADV_POPULATE_(READ|WRITE).
1719 *
1720 * The range must be page-aligned.
1721 *
1722 * mm->mmap_lock must be held. If it's released, *@locked will be set to 0.
1723 */
1724long faultin_page_range(struct mm_struct *mm, unsigned long start,
1725 unsigned long end, bool write, int *locked)
1726{
1727 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1728 int gup_flags;
1729 long ret;
1730
1731 VM_BUG_ON(!PAGE_ALIGNED(start));
1732 VM_BUG_ON(!PAGE_ALIGNED(end));
1733 mmap_assert_locked(mm);
1734
1735 /*
1736 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark
1737 * the page dirty with FOLL_WRITE -- which doesn't make a
1738 * difference with !FOLL_FORCE, because the page is writable
1739 * in the page table.
1740 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit
1741 * a poisoned page.
1742 * !FOLL_FORCE: Require proper access permissions.
1743 */
1744 gup_flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_UNLOCKABLE |
1745 FOLL_MADV_POPULATE;
1746 if (write)
1747 gup_flags |= FOLL_WRITE;
1748
1749 ret = __get_user_pages_locked(mm, start, nr_pages, NULL, locked,
1750 gup_flags);
1751 lru_add_drain();
1752 return ret;
1753}
1754
1755/*
1756 * __mm_populate - populate and/or mlock pages within a range of address space.
1757 *
1758 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1759 * flags. VMAs must be already marked with the desired vm_flags, and
1760 * mmap_lock must not be held.
1761 */
1762int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1763{
1764 struct mm_struct *mm = current->mm;
1765 unsigned long end, nstart, nend;
1766 struct vm_area_struct *vma = NULL;
1767 int locked = 0;
1768 long ret = 0;
1769
1770 end = start + len;
1771
1772 for (nstart = start; nstart < end; nstart = nend) {
1773 /*
1774 * We want to fault in pages for [nstart; end) address range.
1775 * Find first corresponding VMA.
1776 */
1777 if (!locked) {
1778 locked = 1;
1779 mmap_read_lock(mm);
1780 vma = find_vma_intersection(mm, nstart, end);
1781 } else if (nstart >= vma->vm_end)
1782 vma = find_vma_intersection(mm, vma->vm_end, end);
1783
1784 if (!vma)
1785 break;
1786 /*
1787 * Set [nstart; nend) to intersection of desired address
1788 * range with the first VMA. Also, skip undesirable VMA types.
1789 */
1790 nend = min(end, vma->vm_end);
1791 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1792 continue;
1793 if (nstart < vma->vm_start)
1794 nstart = vma->vm_start;
1795 /*
1796 * Now fault in a range of pages. populate_vma_page_range()
1797 * double checks the vma flags, so that it won't mlock pages
1798 * if the vma was already munlocked.
1799 */
1800 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1801 if (ret < 0) {
1802 if (ignore_errors) {
1803 ret = 0;
1804 continue; /* continue at next VMA */
1805 }
1806 break;
1807 }
1808 nend = nstart + ret * PAGE_SIZE;
1809 ret = 0;
1810 }
1811 if (locked)
1812 mmap_read_unlock(mm);
1813 return ret; /* 0 or negative error code */
1814}
1815#else /* CONFIG_MMU */
1816static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1817 unsigned long nr_pages, struct page **pages,
1818 int *locked, unsigned int foll_flags)
1819{
1820 struct vm_area_struct *vma;
1821 bool must_unlock = false;
1822 unsigned long vm_flags;
1823 long i;
1824
1825 if (!nr_pages)
1826 return 0;
1827
1828 /*
1829 * The internal caller expects GUP to manage the lock internally and the
1830 * lock must be released when this returns.
1831 */
1832 if (!*locked) {
1833 if (mmap_read_lock_killable(mm))
1834 return -EAGAIN;
1835 must_unlock = true;
1836 *locked = 1;
1837 }
1838
1839 /* calculate required read or write permissions.
1840 * If FOLL_FORCE is set, we only require the "MAY" flags.
1841 */
1842 vm_flags = (foll_flags & FOLL_WRITE) ?
1843 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1844 vm_flags &= (foll_flags & FOLL_FORCE) ?
1845 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1846
1847 for (i = 0; i < nr_pages; i++) {
1848 vma = find_vma(mm, start);
1849 if (!vma)
1850 break;
1851
1852 /* protect what we can, including chardevs */
1853 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1854 !(vm_flags & vma->vm_flags))
1855 break;
1856
1857 if (pages) {
1858 pages[i] = virt_to_page((void *)start);
1859 if (pages[i])
1860 get_page(pages[i]);
1861 }
1862
1863 start = (start + PAGE_SIZE) & PAGE_MASK;
1864 }
1865
1866 if (must_unlock && *locked) {
1867 mmap_read_unlock(mm);
1868 *locked = 0;
1869 }
1870
1871 return i ? : -EFAULT;
1872}
1873#endif /* !CONFIG_MMU */
1874
1875/**
1876 * fault_in_writeable - fault in userspace address range for writing
1877 * @uaddr: start of address range
1878 * @size: size of address range
1879 *
1880 * Returns the number of bytes not faulted in (like copy_to_user() and
1881 * copy_from_user()).
1882 */
1883size_t fault_in_writeable(char __user *uaddr, size_t size)
1884{
1885 char __user *start = uaddr, *end;
1886
1887 if (unlikely(size == 0))
1888 return 0;
1889 if (!user_write_access_begin(uaddr, size))
1890 return size;
1891 if (!PAGE_ALIGNED(uaddr)) {
1892 unsafe_put_user(0, uaddr, out);
1893 uaddr = (char __user *)PAGE_ALIGN((unsigned long)uaddr);
1894 }
1895 end = (char __user *)PAGE_ALIGN((unsigned long)start + size);
1896 if (unlikely(end < start))
1897 end = NULL;
1898 while (uaddr != end) {
1899 unsafe_put_user(0, uaddr, out);
1900 uaddr += PAGE_SIZE;
1901 }
1902
1903out:
1904 user_write_access_end();
1905 if (size > uaddr - start)
1906 return size - (uaddr - start);
1907 return 0;
1908}
1909EXPORT_SYMBOL(fault_in_writeable);
1910
1911/**
1912 * fault_in_subpage_writeable - fault in an address range for writing
1913 * @uaddr: start of address range
1914 * @size: size of address range
1915 *
1916 * Fault in a user address range for writing while checking for permissions at
1917 * sub-page granularity (e.g. arm64 MTE). This function should be used when
1918 * the caller cannot guarantee forward progress of a copy_to_user() loop.
1919 *
1920 * Returns the number of bytes not faulted in (like copy_to_user() and
1921 * copy_from_user()).
1922 */
1923size_t fault_in_subpage_writeable(char __user *uaddr, size_t size)
1924{
1925 size_t faulted_in;
1926
1927 /*
1928 * Attempt faulting in at page granularity first for page table
1929 * permission checking. The arch-specific probe_subpage_writeable()
1930 * functions may not check for this.
1931 */
1932 faulted_in = size - fault_in_writeable(uaddr, size);
1933 if (faulted_in)
1934 faulted_in -= probe_subpage_writeable(uaddr, faulted_in);
1935
1936 return size - faulted_in;
1937}
1938EXPORT_SYMBOL(fault_in_subpage_writeable);
1939
1940/*
1941 * fault_in_safe_writeable - fault in an address range for writing
1942 * @uaddr: start of address range
1943 * @size: length of address range
1944 *
1945 * Faults in an address range for writing. This is primarily useful when we
1946 * already know that some or all of the pages in the address range aren't in
1947 * memory.
1948 *
1949 * Unlike fault_in_writeable(), this function is non-destructive.
1950 *
1951 * Note that we don't pin or otherwise hold the pages referenced that we fault
1952 * in. There's no guarantee that they'll stay in memory for any duration of
1953 * time.
1954 *
1955 * Returns the number of bytes not faulted in, like copy_to_user() and
1956 * copy_from_user().
1957 */
1958size_t fault_in_safe_writeable(const char __user *uaddr, size_t size)
1959{
1960 unsigned long start = (unsigned long)uaddr, end;
1961 struct mm_struct *mm = current->mm;
1962 bool unlocked = false;
1963
1964 if (unlikely(size == 0))
1965 return 0;
1966 end = PAGE_ALIGN(start + size);
1967 if (end < start)
1968 end = 0;
1969
1970 mmap_read_lock(mm);
1971 do {
1972 if (fixup_user_fault(mm, start, FAULT_FLAG_WRITE, &unlocked))
1973 break;
1974 start = (start + PAGE_SIZE) & PAGE_MASK;
1975 } while (start != end);
1976 mmap_read_unlock(mm);
1977
1978 if (size > (unsigned long)uaddr - start)
1979 return size - ((unsigned long)uaddr - start);
1980 return 0;
1981}
1982EXPORT_SYMBOL(fault_in_safe_writeable);
1983
1984/**
1985 * fault_in_readable - fault in userspace address range for reading
1986 * @uaddr: start of user address range
1987 * @size: size of user address range
1988 *
1989 * Returns the number of bytes not faulted in (like copy_to_user() and
1990 * copy_from_user()).
1991 */
1992size_t fault_in_readable(const char __user *uaddr, size_t size)
1993{
1994 const char __user *start = uaddr, *end;
1995 volatile char c;
1996
1997 if (unlikely(size == 0))
1998 return 0;
1999 if (!user_read_access_begin(uaddr, size))
2000 return size;
2001 if (!PAGE_ALIGNED(uaddr)) {
2002 unsafe_get_user(c, uaddr, out);
2003 uaddr = (const char __user *)PAGE_ALIGN((unsigned long)uaddr);
2004 }
2005 end = (const char __user *)PAGE_ALIGN((unsigned long)start + size);
2006 if (unlikely(end < start))
2007 end = NULL;
2008 while (uaddr != end) {
2009 unsafe_get_user(c, uaddr, out);
2010 uaddr += PAGE_SIZE;
2011 }
2012
2013out:
2014 user_read_access_end();
2015 (void)c;
2016 if (size > uaddr - start)
2017 return size - (uaddr - start);
2018 return 0;
2019}
2020EXPORT_SYMBOL(fault_in_readable);
2021
2022/**
2023 * get_dump_page() - pin user page in memory while writing it to core dump
2024 * @addr: user address
2025 *
2026 * Returns struct page pointer of user page pinned for dump,
2027 * to be freed afterwards by put_page().
2028 *
2029 * Returns NULL on any kind of failure - a hole must then be inserted into
2030 * the corefile, to preserve alignment with its headers; and also returns
2031 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2032 * allowing a hole to be left in the corefile to save disk space.
2033 *
2034 * Called without mmap_lock (takes and releases the mmap_lock by itself).
2035 */
2036#ifdef CONFIG_ELF_CORE
2037struct page *get_dump_page(unsigned long addr)
2038{
2039 struct page *page;
2040 int locked = 0;
2041 int ret;
2042
2043 ret = __get_user_pages_locked(current->mm, addr, 1, &page, &locked,
2044 FOLL_FORCE | FOLL_DUMP | FOLL_GET);
2045 return (ret == 1) ? page : NULL;
2046}
2047#endif /* CONFIG_ELF_CORE */
2048
2049#ifdef CONFIG_MIGRATION
2050/*
2051 * Returns the number of collected pages. Return value is always >= 0.
2052 */
2053static unsigned long collect_longterm_unpinnable_pages(
2054 struct list_head *movable_page_list,
2055 unsigned long nr_pages,
2056 struct page **pages)
2057{
2058 unsigned long i, collected = 0;
2059 struct folio *prev_folio = NULL;
2060 bool drain_allow = true;
2061
2062 for (i = 0; i < nr_pages; i++) {
2063 struct folio *folio = page_folio(pages[i]);
2064
2065 if (folio == prev_folio)
2066 continue;
2067 prev_folio = folio;
2068
2069 if (folio_is_longterm_pinnable(folio))
2070 continue;
2071
2072 collected++;
2073
2074 if (folio_is_device_coherent(folio))
2075 continue;
2076
2077 if (folio_test_hugetlb(folio)) {
2078 isolate_hugetlb(folio, movable_page_list);
2079 continue;
2080 }
2081
2082 if (!folio_test_lru(folio) && drain_allow) {
2083 lru_add_drain_all();
2084 drain_allow = false;
2085 }
2086
2087 if (!folio_isolate_lru(folio))
2088 continue;
2089
2090 list_add_tail(&folio->lru, movable_page_list);
2091 node_stat_mod_folio(folio,
2092 NR_ISOLATED_ANON + folio_is_file_lru(folio),
2093 folio_nr_pages(folio));
2094 }
2095
2096 return collected;
2097}
2098
2099/*
2100 * Unpins all pages and migrates device coherent pages and movable_page_list.
2101 * Returns -EAGAIN if all pages were successfully migrated or -errno for failure
2102 * (or partial success).
2103 */
2104static int migrate_longterm_unpinnable_pages(
2105 struct list_head *movable_page_list,
2106 unsigned long nr_pages,
2107 struct page **pages)
2108{
2109 int ret;
2110 unsigned long i;
2111
2112 for (i = 0; i < nr_pages; i++) {
2113 struct folio *folio = page_folio(pages[i]);
2114
2115 if (folio_is_device_coherent(folio)) {
2116 /*
2117 * Migration will fail if the page is pinned, so convert
2118 * the pin on the source page to a normal reference.
2119 */
2120 pages[i] = NULL;
2121 folio_get(folio);
2122 gup_put_folio(folio, 1, FOLL_PIN);
2123
2124 if (migrate_device_coherent_page(&folio->page)) {
2125 ret = -EBUSY;
2126 goto err;
2127 }
2128
2129 continue;
2130 }
2131
2132 /*
2133 * We can't migrate pages with unexpected references, so drop
2134 * the reference obtained by __get_user_pages_locked().
2135 * Migrating pages have been added to movable_page_list after
2136 * calling folio_isolate_lru() which takes a reference so the
2137 * page won't be freed if it's migrating.
2138 */
2139 unpin_user_page(pages[i]);
2140 pages[i] = NULL;
2141 }
2142
2143 if (!list_empty(movable_page_list)) {
2144 struct migration_target_control mtc = {
2145 .nid = NUMA_NO_NODE,
2146 .gfp_mask = GFP_USER | __GFP_NOWARN,
2147 };
2148
2149 if (migrate_pages(movable_page_list, alloc_migration_target,
2150 NULL, (unsigned long)&mtc, MIGRATE_SYNC,
2151 MR_LONGTERM_PIN, NULL)) {
2152 ret = -ENOMEM;
2153 goto err;
2154 }
2155 }
2156
2157 putback_movable_pages(movable_page_list);
2158
2159 return -EAGAIN;
2160
2161err:
2162 for (i = 0; i < nr_pages; i++)
2163 if (pages[i])
2164 unpin_user_page(pages[i]);
2165 putback_movable_pages(movable_page_list);
2166
2167 return ret;
2168}
2169
2170/*
2171 * Check whether all pages are *allowed* to be pinned. Rather confusingly, all
2172 * pages in the range are required to be pinned via FOLL_PIN, before calling
2173 * this routine.
2174 *
2175 * If any pages in the range are not allowed to be pinned, then this routine
2176 * will migrate those pages away, unpin all the pages in the range and return
2177 * -EAGAIN. The caller should re-pin the entire range with FOLL_PIN and then
2178 * call this routine again.
2179 *
2180 * If an error other than -EAGAIN occurs, this indicates a migration failure.
2181 * The caller should give up, and propagate the error back up the call stack.
2182 *
2183 * If everything is OK and all pages in the range are allowed to be pinned, then
2184 * this routine leaves all pages pinned and returns zero for success.
2185 */
2186static long check_and_migrate_movable_pages(unsigned long nr_pages,
2187 struct page **pages)
2188{
2189 unsigned long collected;
2190 LIST_HEAD(movable_page_list);
2191
2192 collected = collect_longterm_unpinnable_pages(&movable_page_list,
2193 nr_pages, pages);
2194 if (!collected)
2195 return 0;
2196
2197 return migrate_longterm_unpinnable_pages(&movable_page_list, nr_pages,
2198 pages);
2199}
2200#else
2201static long check_and_migrate_movable_pages(unsigned long nr_pages,
2202 struct page **pages)
2203{
2204 return 0;
2205}
2206#endif /* CONFIG_MIGRATION */
2207
2208/*
2209 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
2210 * allows us to process the FOLL_LONGTERM flag.
2211 */
2212static long __gup_longterm_locked(struct mm_struct *mm,
2213 unsigned long start,
2214 unsigned long nr_pages,
2215 struct page **pages,
2216 int *locked,
2217 unsigned int gup_flags)
2218{
2219 unsigned int flags;
2220 long rc, nr_pinned_pages;
2221
2222 if (!(gup_flags & FOLL_LONGTERM))
2223 return __get_user_pages_locked(mm, start, nr_pages, pages,
2224 locked, gup_flags);
2225
2226 flags = memalloc_pin_save();
2227 do {
2228 nr_pinned_pages = __get_user_pages_locked(mm, start, nr_pages,
2229 pages, locked,
2230 gup_flags);
2231 if (nr_pinned_pages <= 0) {
2232 rc = nr_pinned_pages;
2233 break;
2234 }
2235
2236 /* FOLL_LONGTERM implies FOLL_PIN */
2237 rc = check_and_migrate_movable_pages(nr_pinned_pages, pages);
2238 } while (rc == -EAGAIN);
2239 memalloc_pin_restore(flags);
2240 return rc ? rc : nr_pinned_pages;
2241}
2242
2243/*
2244 * Check that the given flags are valid for the exported gup/pup interface, and
2245 * update them with the required flags that the caller must have set.
2246 */
2247static bool is_valid_gup_args(struct page **pages, int *locked,
2248 unsigned int *gup_flags_p, unsigned int to_set)
2249{
2250 unsigned int gup_flags = *gup_flags_p;
2251
2252 /*
2253 * These flags not allowed to be specified externally to the gup
2254 * interfaces:
2255 * - FOLL_TOUCH/FOLL_PIN/FOLL_TRIED/FOLL_FAST_ONLY are internal only
2256 * - FOLL_REMOTE is internal only and used on follow_page()
2257 * - FOLL_UNLOCKABLE is internal only and used if locked is !NULL
2258 */
2259 if (WARN_ON_ONCE(gup_flags & INTERNAL_GUP_FLAGS))
2260 return false;
2261
2262 gup_flags |= to_set;
2263 if (locked) {
2264 /* At the external interface locked must be set */
2265 if (WARN_ON_ONCE(*locked != 1))
2266 return false;
2267
2268 gup_flags |= FOLL_UNLOCKABLE;
2269 }
2270
2271 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2272 if (WARN_ON_ONCE((gup_flags & (FOLL_PIN | FOLL_GET)) ==
2273 (FOLL_PIN | FOLL_GET)))
2274 return false;
2275
2276 /* LONGTERM can only be specified when pinning */
2277 if (WARN_ON_ONCE(!(gup_flags & FOLL_PIN) && (gup_flags & FOLL_LONGTERM)))
2278 return false;
2279
2280 /* Pages input must be given if using GET/PIN */
2281 if (WARN_ON_ONCE((gup_flags & (FOLL_GET | FOLL_PIN)) && !pages))
2282 return false;
2283
2284 /* We want to allow the pgmap to be hot-unplugged at all times */
2285 if (WARN_ON_ONCE((gup_flags & FOLL_LONGTERM) &&
2286 (gup_flags & FOLL_PCI_P2PDMA)))
2287 return false;
2288
2289 *gup_flags_p = gup_flags;
2290 return true;
2291}
2292
2293#ifdef CONFIG_MMU
2294/**
2295 * get_user_pages_remote() - pin user pages in memory
2296 * @mm: mm_struct of target mm
2297 * @start: starting user address
2298 * @nr_pages: number of pages from start to pin
2299 * @gup_flags: flags modifying lookup behaviour
2300 * @pages: array that receives pointers to the pages pinned.
2301 * Should be at least nr_pages long. Or NULL, if caller
2302 * only intends to ensure the pages are faulted in.
2303 * @locked: pointer to lock flag indicating whether lock is held and
2304 * subsequently whether VM_FAULT_RETRY functionality can be
2305 * utilised. Lock must initially be held.
2306 *
2307 * Returns either number of pages pinned (which may be less than the
2308 * number requested), or an error. Details about the return value:
2309 *
2310 * -- If nr_pages is 0, returns 0.
2311 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
2312 * -- If nr_pages is >0, and some pages were pinned, returns the number of
2313 * pages pinned. Again, this may be less than nr_pages.
2314 *
2315 * The caller is responsible for releasing returned @pages, via put_page().
2316 *
2317 * Must be called with mmap_lock held for read or write.
2318 *
2319 * get_user_pages_remote walks a process's page tables and takes a reference
2320 * to each struct page that each user address corresponds to at a given
2321 * instant. That is, it takes the page that would be accessed if a user
2322 * thread accesses the given user virtual address at that instant.
2323 *
2324 * This does not guarantee that the page exists in the user mappings when
2325 * get_user_pages_remote returns, and there may even be a completely different
2326 * page there in some cases (eg. if mmapped pagecache has been invalidated
2327 * and subsequently re-faulted). However it does guarantee that the page
2328 * won't be freed completely. And mostly callers simply care that the page
2329 * contains data that was valid *at some point in time*. Typically, an IO
2330 * or similar operation cannot guarantee anything stronger anyway because
2331 * locks can't be held over the syscall boundary.
2332 *
2333 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
2334 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
2335 * be called after the page is finished with, and before put_page is called.
2336 *
2337 * get_user_pages_remote is typically used for fewer-copy IO operations,
2338 * to get a handle on the memory by some means other than accesses
2339 * via the user virtual addresses. The pages may be submitted for
2340 * DMA to devices or accessed via their kernel linear mapping (via the
2341 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
2342 *
2343 * See also get_user_pages_fast, for performance critical applications.
2344 *
2345 * get_user_pages_remote should be phased out in favor of
2346 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
2347 * should use get_user_pages_remote because it cannot pass
2348 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
2349 */
2350long get_user_pages_remote(struct mm_struct *mm,
2351 unsigned long start, unsigned long nr_pages,
2352 unsigned int gup_flags, struct page **pages,
2353 int *locked)
2354{
2355 int local_locked = 1;
2356
2357 if (!is_valid_gup_args(pages, locked, &gup_flags,
2358 FOLL_TOUCH | FOLL_REMOTE))
2359 return -EINVAL;
2360
2361 return __get_user_pages_locked(mm, start, nr_pages, pages,
2362 locked ? locked : &local_locked,
2363 gup_flags);
2364}
2365EXPORT_SYMBOL(get_user_pages_remote);
2366
2367#else /* CONFIG_MMU */
2368long get_user_pages_remote(struct mm_struct *mm,
2369 unsigned long start, unsigned long nr_pages,
2370 unsigned int gup_flags, struct page **pages,
2371 int *locked)
2372{
2373 return 0;
2374}
2375#endif /* !CONFIG_MMU */
2376
2377/**
2378 * get_user_pages() - pin user pages in memory
2379 * @start: starting user address
2380 * @nr_pages: number of pages from start to pin
2381 * @gup_flags: flags modifying lookup behaviour
2382 * @pages: array that receives pointers to the pages pinned.
2383 * Should be at least nr_pages long. Or NULL, if caller
2384 * only intends to ensure the pages are faulted in.
2385 *
2386 * This is the same as get_user_pages_remote(), just with a less-flexible
2387 * calling convention where we assume that the mm being operated on belongs to
2388 * the current task, and doesn't allow passing of a locked parameter. We also
2389 * obviously don't pass FOLL_REMOTE in here.
2390 */
2391long get_user_pages(unsigned long start, unsigned long nr_pages,
2392 unsigned int gup_flags, struct page **pages)
2393{
2394 int locked = 1;
2395
2396 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_TOUCH))
2397 return -EINVAL;
2398
2399 return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2400 &locked, gup_flags);
2401}
2402EXPORT_SYMBOL(get_user_pages);
2403
2404/*
2405 * get_user_pages_unlocked() is suitable to replace the form:
2406 *
2407 * mmap_read_lock(mm);
2408 * get_user_pages(mm, ..., pages, NULL);
2409 * mmap_read_unlock(mm);
2410 *
2411 * with:
2412 *
2413 * get_user_pages_unlocked(mm, ..., pages);
2414 *
2415 * It is functionally equivalent to get_user_pages_fast so
2416 * get_user_pages_fast should be used instead if specific gup_flags
2417 * (e.g. FOLL_FORCE) are not required.
2418 */
2419long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2420 struct page **pages, unsigned int gup_flags)
2421{
2422 int locked = 0;
2423
2424 if (!is_valid_gup_args(pages, NULL, &gup_flags,
2425 FOLL_TOUCH | FOLL_UNLOCKABLE))
2426 return -EINVAL;
2427
2428 return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2429 &locked, gup_flags);
2430}
2431EXPORT_SYMBOL(get_user_pages_unlocked);
2432
2433/*
2434 * Fast GUP
2435 *
2436 * get_user_pages_fast attempts to pin user pages by walking the page
2437 * tables directly and avoids taking locks. Thus the walker needs to be
2438 * protected from page table pages being freed from under it, and should
2439 * block any THP splits.
2440 *
2441 * One way to achieve this is to have the walker disable interrupts, and
2442 * rely on IPIs from the TLB flushing code blocking before the page table
2443 * pages are freed. This is unsuitable for architectures that do not need
2444 * to broadcast an IPI when invalidating TLBs.
2445 *
2446 * Another way to achieve this is to batch up page table containing pages
2447 * belonging to more than one mm_user, then rcu_sched a callback to free those
2448 * pages. Disabling interrupts will allow the fast_gup walker to both block
2449 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2450 * (which is a relatively rare event). The code below adopts this strategy.
2451 *
2452 * Before activating this code, please be aware that the following assumptions
2453 * are currently made:
2454 *
2455 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2456 * free pages containing page tables or TLB flushing requires IPI broadcast.
2457 *
2458 * *) ptes can be read atomically by the architecture.
2459 *
2460 * *) access_ok is sufficient to validate userspace address ranges.
2461 *
2462 * The last two assumptions can be relaxed by the addition of helper functions.
2463 *
2464 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2465 */
2466#ifdef CONFIG_HAVE_FAST_GUP
2467
2468/*
2469 * Used in the GUP-fast path to determine whether a pin is permitted for a
2470 * specific folio.
2471 *
2472 * This call assumes the caller has pinned the folio, that the lowest page table
2473 * level still points to this folio, and that interrupts have been disabled.
2474 *
2475 * Writing to pinned file-backed dirty tracked folios is inherently problematic
2476 * (see comment describing the writable_file_mapping_allowed() function). We
2477 * therefore try to avoid the most egregious case of a long-term mapping doing
2478 * so.
2479 *
2480 * This function cannot be as thorough as that one as the VMA is not available
2481 * in the fast path, so instead we whitelist known good cases and if in doubt,
2482 * fall back to the slow path.
2483 */
2484static bool folio_fast_pin_allowed(struct folio *folio, unsigned int flags)
2485{
2486 struct address_space *mapping;
2487 unsigned long mapping_flags;
2488
2489 /*
2490 * If we aren't pinning then no problematic write can occur. A long term
2491 * pin is the most egregious case so this is the one we disallow.
2492 */
2493 if ((flags & (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE)) !=
2494 (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE))
2495 return true;
2496
2497 /* The folio is pinned, so we can safely access folio fields. */
2498
2499 if (WARN_ON_ONCE(folio_test_slab(folio)))
2500 return false;
2501
2502 /* hugetlb mappings do not require dirty-tracking. */
2503 if (folio_test_hugetlb(folio))
2504 return true;
2505
2506 /*
2507 * GUP-fast disables IRQs. When IRQS are disabled, RCU grace periods
2508 * cannot proceed, which means no actions performed under RCU can
2509 * proceed either.
2510 *
2511 * inodes and thus their mappings are freed under RCU, which means the
2512 * mapping cannot be freed beneath us and thus we can safely dereference
2513 * it.
2514 */
2515 lockdep_assert_irqs_disabled();
2516
2517 /*
2518 * However, there may be operations which _alter_ the mapping, so ensure
2519 * we read it once and only once.
2520 */
2521 mapping = READ_ONCE(folio->mapping);
2522
2523 /*
2524 * The mapping may have been truncated, in any case we cannot determine
2525 * if this mapping is safe - fall back to slow path to determine how to
2526 * proceed.
2527 */
2528 if (!mapping)
2529 return false;
2530
2531 /* Anonymous folios pose no problem. */
2532 mapping_flags = (unsigned long)mapping & PAGE_MAPPING_FLAGS;
2533 if (mapping_flags)
2534 return mapping_flags & PAGE_MAPPING_ANON;
2535
2536 /*
2537 * At this point, we know the mapping is non-null and points to an
2538 * address_space object. The only remaining whitelisted file system is
2539 * shmem.
2540 */
2541 return shmem_mapping(mapping);
2542}
2543
2544static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2545 unsigned int flags,
2546 struct page **pages)
2547{
2548 while ((*nr) - nr_start) {
2549 struct page *page = pages[--(*nr)];
2550
2551 ClearPageReferenced(page);
2552 if (flags & FOLL_PIN)
2553 unpin_user_page(page);
2554 else
2555 put_page(page);
2556 }
2557}
2558
2559#ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2560/*
2561 * Fast-gup relies on pte change detection to avoid concurrent pgtable
2562 * operations.
2563 *
2564 * To pin the page, fast-gup needs to do below in order:
2565 * (1) pin the page (by prefetching pte), then (2) check pte not changed.
2566 *
2567 * For the rest of pgtable operations where pgtable updates can be racy
2568 * with fast-gup, we need to do (1) clear pte, then (2) check whether page
2569 * is pinned.
2570 *
2571 * Above will work for all pte-level operations, including THP split.
2572 *
2573 * For THP collapse, it's a bit more complicated because fast-gup may be
2574 * walking a pgtable page that is being freed (pte is still valid but pmd
2575 * can be cleared already). To avoid race in such condition, we need to
2576 * also check pmd here to make sure pmd doesn't change (corresponds to
2577 * pmdp_collapse_flush() in the THP collapse code path).
2578 */
2579static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2580 unsigned long end, unsigned int flags,
2581 struct page **pages, int *nr)
2582{
2583 struct dev_pagemap *pgmap = NULL;
2584 int nr_start = *nr, ret = 0;
2585 pte_t *ptep, *ptem;
2586
2587 ptem = ptep = pte_offset_map(&pmd, addr);
2588 if (!ptep)
2589 return 0;
2590 do {
2591 pte_t pte = ptep_get_lockless(ptep);
2592 struct page *page;
2593 struct folio *folio;
2594
2595 /*
2596 * Always fallback to ordinary GUP on PROT_NONE-mapped pages:
2597 * pte_access_permitted() better should reject these pages
2598 * either way: otherwise, GUP-fast might succeed in
2599 * cases where ordinary GUP would fail due to VMA access
2600 * permissions.
2601 */
2602 if (pte_protnone(pte))
2603 goto pte_unmap;
2604
2605 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2606 goto pte_unmap;
2607
2608 if (pte_devmap(pte)) {
2609 if (unlikely(flags & FOLL_LONGTERM))
2610 goto pte_unmap;
2611
2612 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2613 if (unlikely(!pgmap)) {
2614 undo_dev_pagemap(nr, nr_start, flags, pages);
2615 goto pte_unmap;
2616 }
2617 } else if (pte_special(pte))
2618 goto pte_unmap;
2619
2620 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2621 page = pte_page(pte);
2622
2623 folio = try_grab_folio(page, 1, flags);
2624 if (!folio)
2625 goto pte_unmap;
2626
2627 if (unlikely(folio_is_secretmem(folio))) {
2628 gup_put_folio(folio, 1, flags);
2629 goto pte_unmap;
2630 }
2631
2632 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp)) ||
2633 unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) {
2634 gup_put_folio(folio, 1, flags);
2635 goto pte_unmap;
2636 }
2637
2638 if (!folio_fast_pin_allowed(folio, flags)) {
2639 gup_put_folio(folio, 1, flags);
2640 goto pte_unmap;
2641 }
2642
2643 if (!pte_write(pte) && gup_must_unshare(NULL, flags, page)) {
2644 gup_put_folio(folio, 1, flags);
2645 goto pte_unmap;
2646 }
2647
2648 /*
2649 * We need to make the page accessible if and only if we are
2650 * going to access its content (the FOLL_PIN case). Please
2651 * see Documentation/core-api/pin_user_pages.rst for
2652 * details.
2653 */
2654 if (flags & FOLL_PIN) {
2655 ret = arch_make_page_accessible(page);
2656 if (ret) {
2657 gup_put_folio(folio, 1, flags);
2658 goto pte_unmap;
2659 }
2660 }
2661 folio_set_referenced(folio);
2662 pages[*nr] = page;
2663 (*nr)++;
2664 } while (ptep++, addr += PAGE_SIZE, addr != end);
2665
2666 ret = 1;
2667
2668pte_unmap:
2669 if (pgmap)
2670 put_dev_pagemap(pgmap);
2671 pte_unmap(ptem);
2672 return ret;
2673}
2674#else
2675
2676/*
2677 * If we can't determine whether or not a pte is special, then fail immediately
2678 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2679 * to be special.
2680 *
2681 * For a futex to be placed on a THP tail page, get_futex_key requires a
2682 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2683 * useful to have gup_huge_pmd even if we can't operate on ptes.
2684 */
2685static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2686 unsigned long end, unsigned int flags,
2687 struct page **pages, int *nr)
2688{
2689 return 0;
2690}
2691#endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2692
2693#if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2694static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2695 unsigned long end, unsigned int flags,
2696 struct page **pages, int *nr)
2697{
2698 int nr_start = *nr;
2699 struct dev_pagemap *pgmap = NULL;
2700
2701 do {
2702 struct page *page = pfn_to_page(pfn);
2703
2704 pgmap = get_dev_pagemap(pfn, pgmap);
2705 if (unlikely(!pgmap)) {
2706 undo_dev_pagemap(nr, nr_start, flags, pages);
2707 break;
2708 }
2709
2710 if (!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)) {
2711 undo_dev_pagemap(nr, nr_start, flags, pages);
2712 break;
2713 }
2714
2715 SetPageReferenced(page);
2716 pages[*nr] = page;
2717 if (unlikely(try_grab_page(page, flags))) {
2718 undo_dev_pagemap(nr, nr_start, flags, pages);
2719 break;
2720 }
2721 (*nr)++;
2722 pfn++;
2723 } while (addr += PAGE_SIZE, addr != end);
2724
2725 put_dev_pagemap(pgmap);
2726 return addr == end;
2727}
2728
2729static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2730 unsigned long end, unsigned int flags,
2731 struct page **pages, int *nr)
2732{
2733 unsigned long fault_pfn;
2734 int nr_start = *nr;
2735
2736 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2737 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2738 return 0;
2739
2740 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2741 undo_dev_pagemap(nr, nr_start, flags, pages);
2742 return 0;
2743 }
2744 return 1;
2745}
2746
2747static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2748 unsigned long end, unsigned int flags,
2749 struct page **pages, int *nr)
2750{
2751 unsigned long fault_pfn;
2752 int nr_start = *nr;
2753
2754 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2755 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2756 return 0;
2757
2758 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2759 undo_dev_pagemap(nr, nr_start, flags, pages);
2760 return 0;
2761 }
2762 return 1;
2763}
2764#else
2765static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2766 unsigned long end, unsigned int flags,
2767 struct page **pages, int *nr)
2768{
2769 BUILD_BUG();
2770 return 0;
2771}
2772
2773static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2774 unsigned long end, unsigned int flags,
2775 struct page **pages, int *nr)
2776{
2777 BUILD_BUG();
2778 return 0;
2779}
2780#endif
2781
2782static int record_subpages(struct page *page, unsigned long addr,
2783 unsigned long end, struct page **pages)
2784{
2785 int nr;
2786
2787 for (nr = 0; addr != end; nr++, addr += PAGE_SIZE)
2788 pages[nr] = nth_page(page, nr);
2789
2790 return nr;
2791}
2792
2793#ifdef CONFIG_ARCH_HAS_HUGEPD
2794static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2795 unsigned long sz)
2796{
2797 unsigned long __boundary = (addr + sz) & ~(sz-1);
2798 return (__boundary - 1 < end - 1) ? __boundary : end;
2799}
2800
2801static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2802 unsigned long end, unsigned int flags,
2803 struct page **pages, int *nr)
2804{
2805 unsigned long pte_end;
2806 struct page *page;
2807 struct folio *folio;
2808 pte_t pte;
2809 int refs;
2810
2811 pte_end = (addr + sz) & ~(sz-1);
2812 if (pte_end < end)
2813 end = pte_end;
2814
2815 pte = huge_ptep_get(ptep);
2816
2817 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2818 return 0;
2819
2820 /* hugepages are never "special" */
2821 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2822
2823 page = nth_page(pte_page(pte), (addr & (sz - 1)) >> PAGE_SHIFT);
2824 refs = record_subpages(page, addr, end, pages + *nr);
2825
2826 folio = try_grab_folio(page, refs, flags);
2827 if (!folio)
2828 return 0;
2829
2830 if (unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) {
2831 gup_put_folio(folio, refs, flags);
2832 return 0;
2833 }
2834
2835 if (!folio_fast_pin_allowed(folio, flags)) {
2836 gup_put_folio(folio, refs, flags);
2837 return 0;
2838 }
2839
2840 if (!pte_write(pte) && gup_must_unshare(NULL, flags, &folio->page)) {
2841 gup_put_folio(folio, refs, flags);
2842 return 0;
2843 }
2844
2845 *nr += refs;
2846 folio_set_referenced(folio);
2847 return 1;
2848}
2849
2850static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2851 unsigned int pdshift, unsigned long end, unsigned int flags,
2852 struct page **pages, int *nr)
2853{
2854 pte_t *ptep;
2855 unsigned long sz = 1UL << hugepd_shift(hugepd);
2856 unsigned long next;
2857
2858 ptep = hugepte_offset(hugepd, addr, pdshift);
2859 do {
2860 next = hugepte_addr_end(addr, end, sz);
2861 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2862 return 0;
2863 } while (ptep++, addr = next, addr != end);
2864
2865 return 1;
2866}
2867#else
2868static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2869 unsigned int pdshift, unsigned long end, unsigned int flags,
2870 struct page **pages, int *nr)
2871{
2872 return 0;
2873}
2874#endif /* CONFIG_ARCH_HAS_HUGEPD */
2875
2876static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2877 unsigned long end, unsigned int flags,
2878 struct page **pages, int *nr)
2879{
2880 struct page *page;
2881 struct folio *folio;
2882 int refs;
2883
2884 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2885 return 0;
2886
2887 if (pmd_devmap(orig)) {
2888 if (unlikely(flags & FOLL_LONGTERM))
2889 return 0;
2890 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2891 pages, nr);
2892 }
2893
2894 page = nth_page(pmd_page(orig), (addr & ~PMD_MASK) >> PAGE_SHIFT);
2895 refs = record_subpages(page, addr, end, pages + *nr);
2896
2897 folio = try_grab_folio(page, refs, flags);
2898 if (!folio)
2899 return 0;
2900
2901 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2902 gup_put_folio(folio, refs, flags);
2903 return 0;
2904 }
2905
2906 if (!folio_fast_pin_allowed(folio, flags)) {
2907 gup_put_folio(folio, refs, flags);
2908 return 0;
2909 }
2910 if (!pmd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2911 gup_put_folio(folio, refs, flags);
2912 return 0;
2913 }
2914
2915 *nr += refs;
2916 folio_set_referenced(folio);
2917 return 1;
2918}
2919
2920static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2921 unsigned long end, unsigned int flags,
2922 struct page **pages, int *nr)
2923{
2924 struct page *page;
2925 struct folio *folio;
2926 int refs;
2927
2928 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2929 return 0;
2930
2931 if (pud_devmap(orig)) {
2932 if (unlikely(flags & FOLL_LONGTERM))
2933 return 0;
2934 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2935 pages, nr);
2936 }
2937
2938 page = nth_page(pud_page(orig), (addr & ~PUD_MASK) >> PAGE_SHIFT);
2939 refs = record_subpages(page, addr, end, pages + *nr);
2940
2941 folio = try_grab_folio(page, refs, flags);
2942 if (!folio)
2943 return 0;
2944
2945 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2946 gup_put_folio(folio, refs, flags);
2947 return 0;
2948 }
2949
2950 if (!folio_fast_pin_allowed(folio, flags)) {
2951 gup_put_folio(folio, refs, flags);
2952 return 0;
2953 }
2954
2955 if (!pud_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2956 gup_put_folio(folio, refs, flags);
2957 return 0;
2958 }
2959
2960 *nr += refs;
2961 folio_set_referenced(folio);
2962 return 1;
2963}
2964
2965static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2966 unsigned long end, unsigned int flags,
2967 struct page **pages, int *nr)
2968{
2969 int refs;
2970 struct page *page;
2971 struct folio *folio;
2972
2973 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2974 return 0;
2975
2976 BUILD_BUG_ON(pgd_devmap(orig));
2977
2978 page = nth_page(pgd_page(orig), (addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2979 refs = record_subpages(page, addr, end, pages + *nr);
2980
2981 folio = try_grab_folio(page, refs, flags);
2982 if (!folio)
2983 return 0;
2984
2985 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2986 gup_put_folio(folio, refs, flags);
2987 return 0;
2988 }
2989
2990 if (!pgd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2991 gup_put_folio(folio, refs, flags);
2992 return 0;
2993 }
2994
2995 if (!folio_fast_pin_allowed(folio, flags)) {
2996 gup_put_folio(folio, refs, flags);
2997 return 0;
2998 }
2999
3000 *nr += refs;
3001 folio_set_referenced(folio);
3002 return 1;
3003}
3004
3005static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
3006 unsigned int flags, struct page **pages, int *nr)
3007{
3008 unsigned long next;
3009 pmd_t *pmdp;
3010
3011 pmdp = pmd_offset_lockless(pudp, pud, addr);
3012 do {
3013 pmd_t pmd = pmdp_get_lockless(pmdp);
3014
3015 next = pmd_addr_end(addr, end);
3016 if (!pmd_present(pmd))
3017 return 0;
3018
3019 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
3020 pmd_devmap(pmd))) {
3021 /* See gup_pte_range() */
3022 if (pmd_protnone(pmd))
3023 return 0;
3024
3025 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
3026 pages, nr))
3027 return 0;
3028
3029 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
3030 /*
3031 * architecture have different format for hugetlbfs
3032 * pmd format and THP pmd format
3033 */
3034 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
3035 PMD_SHIFT, next, flags, pages, nr))
3036 return 0;
3037 } else if (!gup_pte_range(pmd, pmdp, addr, next, flags, pages, nr))
3038 return 0;
3039 } while (pmdp++, addr = next, addr != end);
3040
3041 return 1;
3042}
3043
3044static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
3045 unsigned int flags, struct page **pages, int *nr)
3046{
3047 unsigned long next;
3048 pud_t *pudp;
3049
3050 pudp = pud_offset_lockless(p4dp, p4d, addr);
3051 do {
3052 pud_t pud = READ_ONCE(*pudp);
3053
3054 next = pud_addr_end(addr, end);
3055 if (unlikely(!pud_present(pud)))
3056 return 0;
3057 if (unlikely(pud_huge(pud) || pud_devmap(pud))) {
3058 if (!gup_huge_pud(pud, pudp, addr, next, flags,
3059 pages, nr))
3060 return 0;
3061 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
3062 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
3063 PUD_SHIFT, next, flags, pages, nr))
3064 return 0;
3065 } else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
3066 return 0;
3067 } while (pudp++, addr = next, addr != end);
3068
3069 return 1;
3070}
3071
3072static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
3073 unsigned int flags, struct page **pages, int *nr)
3074{
3075 unsigned long next;
3076 p4d_t *p4dp;
3077
3078 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
3079 do {
3080 p4d_t p4d = READ_ONCE(*p4dp);
3081
3082 next = p4d_addr_end(addr, end);
3083 if (p4d_none(p4d))
3084 return 0;
3085 BUILD_BUG_ON(p4d_huge(p4d));
3086 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
3087 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
3088 P4D_SHIFT, next, flags, pages, nr))
3089 return 0;
3090 } else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
3091 return 0;
3092 } while (p4dp++, addr = next, addr != end);
3093
3094 return 1;
3095}
3096
3097static void gup_pgd_range(unsigned long addr, unsigned long end,
3098 unsigned int flags, struct page **pages, int *nr)
3099{
3100 unsigned long next;
3101 pgd_t *pgdp;
3102
3103 pgdp = pgd_offset(current->mm, addr);
3104 do {
3105 pgd_t pgd = READ_ONCE(*pgdp);
3106
3107 next = pgd_addr_end(addr, end);
3108 if (pgd_none(pgd))
3109 return;
3110 if (unlikely(pgd_huge(pgd))) {
3111 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
3112 pages, nr))
3113 return;
3114 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
3115 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
3116 PGDIR_SHIFT, next, flags, pages, nr))
3117 return;
3118 } else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
3119 return;
3120 } while (pgdp++, addr = next, addr != end);
3121}
3122#else
3123static inline void gup_pgd_range(unsigned long addr, unsigned long end,
3124 unsigned int flags, struct page **pages, int *nr)
3125{
3126}
3127#endif /* CONFIG_HAVE_FAST_GUP */
3128
3129#ifndef gup_fast_permitted
3130/*
3131 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
3132 * we need to fall back to the slow version:
3133 */
3134static bool gup_fast_permitted(unsigned long start, unsigned long end)
3135{
3136 return true;
3137}
3138#endif
3139
3140static unsigned long lockless_pages_from_mm(unsigned long start,
3141 unsigned long end,
3142 unsigned int gup_flags,
3143 struct page **pages)
3144{
3145 unsigned long flags;
3146 int nr_pinned = 0;
3147 unsigned seq;
3148
3149 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP) ||
3150 !gup_fast_permitted(start, end))
3151 return 0;
3152
3153 if (gup_flags & FOLL_PIN) {
3154 seq = raw_read_seqcount(¤t->mm->write_protect_seq);
3155 if (seq & 1)
3156 return 0;
3157 }
3158
3159 /*
3160 * Disable interrupts. The nested form is used, in order to allow full,
3161 * general purpose use of this routine.
3162 *
3163 * With interrupts disabled, we block page table pages from being freed
3164 * from under us. See struct mmu_table_batch comments in
3165 * include/asm-generic/tlb.h for more details.
3166 *
3167 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
3168 * that come from THPs splitting.
3169 */
3170 local_irq_save(flags);
3171 gup_pgd_range(start, end, gup_flags, pages, &nr_pinned);
3172 local_irq_restore(flags);
3173
3174 /*
3175 * When pinning pages for DMA there could be a concurrent write protect
3176 * from fork() via copy_page_range(), in this case always fail fast GUP.
3177 */
3178 if (gup_flags & FOLL_PIN) {
3179 if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) {
3180 unpin_user_pages_lockless(pages, nr_pinned);
3181 return 0;
3182 } else {
3183 sanity_check_pinned_pages(pages, nr_pinned);
3184 }
3185 }
3186 return nr_pinned;
3187}
3188
3189static int internal_get_user_pages_fast(unsigned long start,
3190 unsigned long nr_pages,
3191 unsigned int gup_flags,
3192 struct page **pages)
3193{
3194 unsigned long len, end;
3195 unsigned long nr_pinned;
3196 int locked = 0;
3197 int ret;
3198
3199 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
3200 FOLL_FORCE | FOLL_PIN | FOLL_GET |
3201 FOLL_FAST_ONLY | FOLL_NOFAULT |
3202 FOLL_PCI_P2PDMA | FOLL_HONOR_NUMA_FAULT)))
3203 return -EINVAL;
3204
3205 if (gup_flags & FOLL_PIN)
3206 mm_set_has_pinned_flag(¤t->mm->flags);
3207
3208 if (!(gup_flags & FOLL_FAST_ONLY))
3209 might_lock_read(¤t->mm->mmap_lock);
3210
3211 start = untagged_addr(start) & PAGE_MASK;
3212 len = nr_pages << PAGE_SHIFT;
3213 if (check_add_overflow(start, len, &end))
3214 return -EOVERFLOW;
3215 if (end > TASK_SIZE_MAX)
3216 return -EFAULT;
3217 if (unlikely(!access_ok((void __user *)start, len)))
3218 return -EFAULT;
3219
3220 nr_pinned = lockless_pages_from_mm(start, end, gup_flags, pages);
3221 if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
3222 return nr_pinned;
3223
3224 /* Slow path: try to get the remaining pages with get_user_pages */
3225 start += nr_pinned << PAGE_SHIFT;
3226 pages += nr_pinned;
3227 ret = __gup_longterm_locked(current->mm, start, nr_pages - nr_pinned,
3228 pages, &locked,
3229 gup_flags | FOLL_TOUCH | FOLL_UNLOCKABLE);
3230 if (ret < 0) {
3231 /*
3232 * The caller has to unpin the pages we already pinned so
3233 * returning -errno is not an option
3234 */
3235 if (nr_pinned)
3236 return nr_pinned;
3237 return ret;
3238 }
3239 return ret + nr_pinned;
3240}
3241
3242/**
3243 * get_user_pages_fast_only() - pin user pages in memory
3244 * @start: starting user address
3245 * @nr_pages: number of pages from start to pin
3246 * @gup_flags: flags modifying pin behaviour
3247 * @pages: array that receives pointers to the pages pinned.
3248 * Should be at least nr_pages long.
3249 *
3250 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
3251 * the regular GUP.
3252 *
3253 * If the architecture does not support this function, simply return with no
3254 * pages pinned.
3255 *
3256 * Careful, careful! COW breaking can go either way, so a non-write
3257 * access can get ambiguous page results. If you call this function without
3258 * 'write' set, you'd better be sure that you're ok with that ambiguity.
3259 */
3260int get_user_pages_fast_only(unsigned long start, int nr_pages,
3261 unsigned int gup_flags, struct page **pages)
3262{
3263 /*
3264 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
3265 * because gup fast is always a "pin with a +1 page refcount" request.
3266 *
3267 * FOLL_FAST_ONLY is required in order to match the API description of
3268 * this routine: no fall back to regular ("slow") GUP.
3269 */
3270 if (!is_valid_gup_args(pages, NULL, &gup_flags,
3271 FOLL_GET | FOLL_FAST_ONLY))
3272 return -EINVAL;
3273
3274 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3275}
3276EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
3277
3278/**
3279 * get_user_pages_fast() - pin user pages in memory
3280 * @start: starting user address
3281 * @nr_pages: number of pages from start to pin
3282 * @gup_flags: flags modifying pin behaviour
3283 * @pages: array that receives pointers to the pages pinned.
3284 * Should be at least nr_pages long.
3285 *
3286 * Attempt to pin user pages in memory without taking mm->mmap_lock.
3287 * If not successful, it will fall back to taking the lock and
3288 * calling get_user_pages().
3289 *
3290 * Returns number of pages pinned. This may be fewer than the number requested.
3291 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
3292 * -errno.
3293 */
3294int get_user_pages_fast(unsigned long start, int nr_pages,
3295 unsigned int gup_flags, struct page **pages)
3296{
3297 /*
3298 * The caller may or may not have explicitly set FOLL_GET; either way is
3299 * OK. However, internally (within mm/gup.c), gup fast variants must set
3300 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
3301 * request.
3302 */
3303 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_GET))
3304 return -EINVAL;
3305 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3306}
3307EXPORT_SYMBOL_GPL(get_user_pages_fast);
3308
3309/**
3310 * pin_user_pages_fast() - pin user pages in memory without taking locks
3311 *
3312 * @start: starting user address
3313 * @nr_pages: number of pages from start to pin
3314 * @gup_flags: flags modifying pin behaviour
3315 * @pages: array that receives pointers to the pages pinned.
3316 * Should be at least nr_pages long.
3317 *
3318 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
3319 * get_user_pages_fast() for documentation on the function arguments, because
3320 * the arguments here are identical.
3321 *
3322 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3323 * see Documentation/core-api/pin_user_pages.rst for further details.
3324 *
3325 * Note that if a zero_page is amongst the returned pages, it will not have
3326 * pins in it and unpin_user_page() will not remove pins from it.
3327 */
3328int pin_user_pages_fast(unsigned long start, int nr_pages,
3329 unsigned int gup_flags, struct page **pages)
3330{
3331 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN))
3332 return -EINVAL;
3333 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3334}
3335EXPORT_SYMBOL_GPL(pin_user_pages_fast);
3336
3337/**
3338 * pin_user_pages_remote() - pin pages of a remote process
3339 *
3340 * @mm: mm_struct of target mm
3341 * @start: starting user address
3342 * @nr_pages: number of pages from start to pin
3343 * @gup_flags: flags modifying lookup behaviour
3344 * @pages: array that receives pointers to the pages pinned.
3345 * Should be at least nr_pages long.
3346 * @locked: pointer to lock flag indicating whether lock is held and
3347 * subsequently whether VM_FAULT_RETRY functionality can be
3348 * utilised. Lock must initially be held.
3349 *
3350 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
3351 * get_user_pages_remote() for documentation on the function arguments, because
3352 * the arguments here are identical.
3353 *
3354 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3355 * see Documentation/core-api/pin_user_pages.rst for details.
3356 *
3357 * Note that if a zero_page is amongst the returned pages, it will not have
3358 * pins in it and unpin_user_page*() will not remove pins from it.
3359 */
3360long pin_user_pages_remote(struct mm_struct *mm,
3361 unsigned long start, unsigned long nr_pages,
3362 unsigned int gup_flags, struct page **pages,
3363 int *locked)
3364{
3365 int local_locked = 1;
3366
3367 if (!is_valid_gup_args(pages, locked, &gup_flags,
3368 FOLL_PIN | FOLL_TOUCH | FOLL_REMOTE))
3369 return 0;
3370 return __gup_longterm_locked(mm, start, nr_pages, pages,
3371 locked ? locked : &local_locked,
3372 gup_flags);
3373}
3374EXPORT_SYMBOL(pin_user_pages_remote);
3375
3376/**
3377 * pin_user_pages() - pin user pages in memory for use by other devices
3378 *
3379 * @start: starting user address
3380 * @nr_pages: number of pages from start to pin
3381 * @gup_flags: flags modifying lookup behaviour
3382 * @pages: array that receives pointers to the pages pinned.
3383 * Should be at least nr_pages long.
3384 *
3385 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
3386 * FOLL_PIN is set.
3387 *
3388 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3389 * see Documentation/core-api/pin_user_pages.rst for details.
3390 *
3391 * Note that if a zero_page is amongst the returned pages, it will not have
3392 * pins in it and unpin_user_page*() will not remove pins from it.
3393 */
3394long pin_user_pages(unsigned long start, unsigned long nr_pages,
3395 unsigned int gup_flags, struct page **pages)
3396{
3397 int locked = 1;
3398
3399 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN))
3400 return 0;
3401 return __gup_longterm_locked(current->mm, start, nr_pages,
3402 pages, &locked, gup_flags);
3403}
3404EXPORT_SYMBOL(pin_user_pages);
3405
3406/*
3407 * pin_user_pages_unlocked() is the FOLL_PIN variant of
3408 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
3409 * FOLL_PIN and rejects FOLL_GET.
3410 *
3411 * Note that if a zero_page is amongst the returned pages, it will not have
3412 * pins in it and unpin_user_page*() will not remove pins from it.
3413 */
3414long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
3415 struct page **pages, unsigned int gup_flags)
3416{
3417 int locked = 0;
3418
3419 if (!is_valid_gup_args(pages, NULL, &gup_flags,
3420 FOLL_PIN | FOLL_TOUCH | FOLL_UNLOCKABLE))
3421 return 0;
3422
3423 return __gup_longterm_locked(current->mm, start, nr_pages, pages,
3424 &locked, gup_flags);
3425}
3426EXPORT_SYMBOL(pin_user_pages_unlocked);
1// SPDX-License-Identifier: GPL-2.0-only
2#include <linux/kernel.h>
3#include <linux/errno.h>
4#include <linux/err.h>
5#include <linux/spinlock.h>
6
7#include <linux/mm.h>
8#include <linux/memremap.h>
9#include <linux/pagemap.h>
10#include <linux/rmap.h>
11#include <linux/swap.h>
12#include <linux/swapops.h>
13#include <linux/secretmem.h>
14
15#include <linux/sched/signal.h>
16#include <linux/rwsem.h>
17#include <linux/hugetlb.h>
18#include <linux/migrate.h>
19#include <linux/mm_inline.h>
20#include <linux/sched/mm.h>
21
22#include <asm/mmu_context.h>
23#include <asm/tlbflush.h>
24
25#include "internal.h"
26
27struct follow_page_context {
28 struct dev_pagemap *pgmap;
29 unsigned int page_mask;
30};
31
32static inline void sanity_check_pinned_pages(struct page **pages,
33 unsigned long npages)
34{
35 if (!IS_ENABLED(CONFIG_DEBUG_VM))
36 return;
37
38 /*
39 * We only pin anonymous pages if they are exclusive. Once pinned, we
40 * can no longer turn them possibly shared and PageAnonExclusive() will
41 * stick around until the page is freed.
42 *
43 * We'd like to verify that our pinned anonymous pages are still mapped
44 * exclusively. The issue with anon THP is that we don't know how
45 * they are/were mapped when pinning them. However, for anon
46 * THP we can assume that either the given page (PTE-mapped THP) or
47 * the head page (PMD-mapped THP) should be PageAnonExclusive(). If
48 * neither is the case, there is certainly something wrong.
49 */
50 for (; npages; npages--, pages++) {
51 struct page *page = *pages;
52 struct folio *folio = page_folio(page);
53
54 if (!folio_test_anon(folio))
55 continue;
56 if (!folio_test_large(folio) || folio_test_hugetlb(folio))
57 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page), page);
58 else
59 /* Either a PTE-mapped or a PMD-mapped THP. */
60 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page) &&
61 !PageAnonExclusive(page), page);
62 }
63}
64
65/*
66 * Return the folio with ref appropriately incremented,
67 * or NULL if that failed.
68 */
69static inline struct folio *try_get_folio(struct page *page, int refs)
70{
71 struct folio *folio;
72
73retry:
74 folio = page_folio(page);
75 if (WARN_ON_ONCE(folio_ref_count(folio) < 0))
76 return NULL;
77 if (unlikely(!folio_ref_try_add_rcu(folio, refs)))
78 return NULL;
79
80 /*
81 * At this point we have a stable reference to the folio; but it
82 * could be that between calling page_folio() and the refcount
83 * increment, the folio was split, in which case we'd end up
84 * holding a reference on a folio that has nothing to do with the page
85 * we were given anymore.
86 * So now that the folio is stable, recheck that the page still
87 * belongs to this folio.
88 */
89 if (unlikely(page_folio(page) != folio)) {
90 if (!put_devmap_managed_page_refs(&folio->page, refs))
91 folio_put_refs(folio, refs);
92 goto retry;
93 }
94
95 return folio;
96}
97
98/**
99 * try_grab_folio() - Attempt to get or pin a folio.
100 * @page: pointer to page to be grabbed
101 * @refs: the value to (effectively) add to the folio's refcount
102 * @flags: gup flags: these are the FOLL_* flag values.
103 *
104 * "grab" names in this file mean, "look at flags to decide whether to use
105 * FOLL_PIN or FOLL_GET behavior, when incrementing the folio's refcount.
106 *
107 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
108 * same time. (That's true throughout the get_user_pages*() and
109 * pin_user_pages*() APIs.) Cases:
110 *
111 * FOLL_GET: folio's refcount will be incremented by @refs.
112 *
113 * FOLL_PIN on large folios: folio's refcount will be incremented by
114 * @refs, and its compound_pincount will be incremented by @refs.
115 *
116 * FOLL_PIN on single-page folios: folio's refcount will be incremented by
117 * @refs * GUP_PIN_COUNTING_BIAS.
118 *
119 * Return: The folio containing @page (with refcount appropriately
120 * incremented) for success, or NULL upon failure. If neither FOLL_GET
121 * nor FOLL_PIN was set, that's considered failure, and furthermore,
122 * a likely bug in the caller, so a warning is also emitted.
123 */
124struct folio *try_grab_folio(struct page *page, int refs, unsigned int flags)
125{
126 if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
127 return NULL;
128
129 if (flags & FOLL_GET)
130 return try_get_folio(page, refs);
131 else if (flags & FOLL_PIN) {
132 struct folio *folio;
133
134 /*
135 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
136 * right zone, so fail and let the caller fall back to the slow
137 * path.
138 */
139 if (unlikely((flags & FOLL_LONGTERM) &&
140 !is_longterm_pinnable_page(page)))
141 return NULL;
142
143 /*
144 * CAUTION: Don't use compound_head() on the page before this
145 * point, the result won't be stable.
146 */
147 folio = try_get_folio(page, refs);
148 if (!folio)
149 return NULL;
150
151 /*
152 * When pinning a large folio, use an exact count to track it.
153 *
154 * However, be sure to *also* increment the normal folio
155 * refcount field at least once, so that the folio really
156 * is pinned. That's why the refcount from the earlier
157 * try_get_folio() is left intact.
158 */
159 if (folio_test_large(folio))
160 atomic_add(refs, folio_pincount_ptr(folio));
161 else
162 folio_ref_add(folio,
163 refs * (GUP_PIN_COUNTING_BIAS - 1));
164 /*
165 * Adjust the pincount before re-checking the PTE for changes.
166 * This is essentially a smp_mb() and is paired with a memory
167 * barrier in page_try_share_anon_rmap().
168 */
169 smp_mb__after_atomic();
170
171 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs);
172
173 return folio;
174 }
175
176 WARN_ON_ONCE(1);
177 return NULL;
178}
179
180static void gup_put_folio(struct folio *folio, int refs, unsigned int flags)
181{
182 if (flags & FOLL_PIN) {
183 node_stat_mod_folio(folio, NR_FOLL_PIN_RELEASED, refs);
184 if (folio_test_large(folio))
185 atomic_sub(refs, folio_pincount_ptr(folio));
186 else
187 refs *= GUP_PIN_COUNTING_BIAS;
188 }
189
190 if (!put_devmap_managed_page_refs(&folio->page, refs))
191 folio_put_refs(folio, refs);
192}
193
194/**
195 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
196 * @page: pointer to page to be grabbed
197 * @flags: gup flags: these are the FOLL_* flag values.
198 *
199 * This might not do anything at all, depending on the flags argument.
200 *
201 * "grab" names in this file mean, "look at flags to decide whether to use
202 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
203 *
204 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
205 * time. Cases: please see the try_grab_folio() documentation, with
206 * "refs=1".
207 *
208 * Return: 0 for success, or if no action was required (if neither FOLL_PIN
209 * nor FOLL_GET was set, nothing is done). A negative error code for failure:
210 *
211 * -ENOMEM FOLL_GET or FOLL_PIN was set, but the page could not
212 * be grabbed.
213 */
214int __must_check try_grab_page(struct page *page, unsigned int flags)
215{
216 struct folio *folio = page_folio(page);
217
218 WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == (FOLL_GET | FOLL_PIN));
219 if (WARN_ON_ONCE(folio_ref_count(folio) <= 0))
220 return -ENOMEM;
221
222 if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
223 return -EREMOTEIO;
224
225 if (flags & FOLL_GET)
226 folio_ref_inc(folio);
227 else if (flags & FOLL_PIN) {
228 /*
229 * Similar to try_grab_folio(): be sure to *also*
230 * increment the normal page refcount field at least once,
231 * so that the page really is pinned.
232 */
233 if (folio_test_large(folio)) {
234 folio_ref_add(folio, 1);
235 atomic_add(1, folio_pincount_ptr(folio));
236 } else {
237 folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
238 }
239
240 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, 1);
241 }
242
243 return 0;
244}
245
246/**
247 * unpin_user_page() - release a dma-pinned page
248 * @page: pointer to page to be released
249 *
250 * Pages that were pinned via pin_user_pages*() must be released via either
251 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
252 * that such pages can be separately tracked and uniquely handled. In
253 * particular, interactions with RDMA and filesystems need special handling.
254 */
255void unpin_user_page(struct page *page)
256{
257 sanity_check_pinned_pages(&page, 1);
258 gup_put_folio(page_folio(page), 1, FOLL_PIN);
259}
260EXPORT_SYMBOL(unpin_user_page);
261
262static inline struct folio *gup_folio_range_next(struct page *start,
263 unsigned long npages, unsigned long i, unsigned int *ntails)
264{
265 struct page *next = nth_page(start, i);
266 struct folio *folio = page_folio(next);
267 unsigned int nr = 1;
268
269 if (folio_test_large(folio))
270 nr = min_t(unsigned int, npages - i,
271 folio_nr_pages(folio) - folio_page_idx(folio, next));
272
273 *ntails = nr;
274 return folio;
275}
276
277static inline struct folio *gup_folio_next(struct page **list,
278 unsigned long npages, unsigned long i, unsigned int *ntails)
279{
280 struct folio *folio = page_folio(list[i]);
281 unsigned int nr;
282
283 for (nr = i + 1; nr < npages; nr++) {
284 if (page_folio(list[nr]) != folio)
285 break;
286 }
287
288 *ntails = nr - i;
289 return folio;
290}
291
292/**
293 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
294 * @pages: array of pages to be maybe marked dirty, and definitely released.
295 * @npages: number of pages in the @pages array.
296 * @make_dirty: whether to mark the pages dirty
297 *
298 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
299 * variants called on that page.
300 *
301 * For each page in the @pages array, make that page (or its head page, if a
302 * compound page) dirty, if @make_dirty is true, and if the page was previously
303 * listed as clean. In any case, releases all pages using unpin_user_page(),
304 * possibly via unpin_user_pages(), for the non-dirty case.
305 *
306 * Please see the unpin_user_page() documentation for details.
307 *
308 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
309 * required, then the caller should a) verify that this is really correct,
310 * because _lock() is usually required, and b) hand code it:
311 * set_page_dirty_lock(), unpin_user_page().
312 *
313 */
314void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
315 bool make_dirty)
316{
317 unsigned long i;
318 struct folio *folio;
319 unsigned int nr;
320
321 if (!make_dirty) {
322 unpin_user_pages(pages, npages);
323 return;
324 }
325
326 sanity_check_pinned_pages(pages, npages);
327 for (i = 0; i < npages; i += nr) {
328 folio = gup_folio_next(pages, npages, i, &nr);
329 /*
330 * Checking PageDirty at this point may race with
331 * clear_page_dirty_for_io(), but that's OK. Two key
332 * cases:
333 *
334 * 1) This code sees the page as already dirty, so it
335 * skips the call to set_page_dirty(). That could happen
336 * because clear_page_dirty_for_io() called
337 * page_mkclean(), followed by set_page_dirty().
338 * However, now the page is going to get written back,
339 * which meets the original intention of setting it
340 * dirty, so all is well: clear_page_dirty_for_io() goes
341 * on to call TestClearPageDirty(), and write the page
342 * back.
343 *
344 * 2) This code sees the page as clean, so it calls
345 * set_page_dirty(). The page stays dirty, despite being
346 * written back, so it gets written back again in the
347 * next writeback cycle. This is harmless.
348 */
349 if (!folio_test_dirty(folio)) {
350 folio_lock(folio);
351 folio_mark_dirty(folio);
352 folio_unlock(folio);
353 }
354 gup_put_folio(folio, nr, FOLL_PIN);
355 }
356}
357EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
358
359/**
360 * unpin_user_page_range_dirty_lock() - release and optionally dirty
361 * gup-pinned page range
362 *
363 * @page: the starting page of a range maybe marked dirty, and definitely released.
364 * @npages: number of consecutive pages to release.
365 * @make_dirty: whether to mark the pages dirty
366 *
367 * "gup-pinned page range" refers to a range of pages that has had one of the
368 * pin_user_pages() variants called on that page.
369 *
370 * For the page ranges defined by [page .. page+npages], make that range (or
371 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
372 * page range was previously listed as clean.
373 *
374 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
375 * required, then the caller should a) verify that this is really correct,
376 * because _lock() is usually required, and b) hand code it:
377 * set_page_dirty_lock(), unpin_user_page().
378 *
379 */
380void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages,
381 bool make_dirty)
382{
383 unsigned long i;
384 struct folio *folio;
385 unsigned int nr;
386
387 for (i = 0; i < npages; i += nr) {
388 folio = gup_folio_range_next(page, npages, i, &nr);
389 if (make_dirty && !folio_test_dirty(folio)) {
390 folio_lock(folio);
391 folio_mark_dirty(folio);
392 folio_unlock(folio);
393 }
394 gup_put_folio(folio, nr, FOLL_PIN);
395 }
396}
397EXPORT_SYMBOL(unpin_user_page_range_dirty_lock);
398
399static void unpin_user_pages_lockless(struct page **pages, unsigned long npages)
400{
401 unsigned long i;
402 struct folio *folio;
403 unsigned int nr;
404
405 /*
406 * Don't perform any sanity checks because we might have raced with
407 * fork() and some anonymous pages might now actually be shared --
408 * which is why we're unpinning after all.
409 */
410 for (i = 0; i < npages; i += nr) {
411 folio = gup_folio_next(pages, npages, i, &nr);
412 gup_put_folio(folio, nr, FOLL_PIN);
413 }
414}
415
416/**
417 * unpin_user_pages() - release an array of gup-pinned pages.
418 * @pages: array of pages to be marked dirty and released.
419 * @npages: number of pages in the @pages array.
420 *
421 * For each page in the @pages array, release the page using unpin_user_page().
422 *
423 * Please see the unpin_user_page() documentation for details.
424 */
425void unpin_user_pages(struct page **pages, unsigned long npages)
426{
427 unsigned long i;
428 struct folio *folio;
429 unsigned int nr;
430
431 /*
432 * If this WARN_ON() fires, then the system *might* be leaking pages (by
433 * leaving them pinned), but probably not. More likely, gup/pup returned
434 * a hard -ERRNO error to the caller, who erroneously passed it here.
435 */
436 if (WARN_ON(IS_ERR_VALUE(npages)))
437 return;
438
439 sanity_check_pinned_pages(pages, npages);
440 for (i = 0; i < npages; i += nr) {
441 folio = gup_folio_next(pages, npages, i, &nr);
442 gup_put_folio(folio, nr, FOLL_PIN);
443 }
444}
445EXPORT_SYMBOL(unpin_user_pages);
446
447/*
448 * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's
449 * lifecycle. Avoid setting the bit unless necessary, or it might cause write
450 * cache bouncing on large SMP machines for concurrent pinned gups.
451 */
452static inline void mm_set_has_pinned_flag(unsigned long *mm_flags)
453{
454 if (!test_bit(MMF_HAS_PINNED, mm_flags))
455 set_bit(MMF_HAS_PINNED, mm_flags);
456}
457
458#ifdef CONFIG_MMU
459static struct page *no_page_table(struct vm_area_struct *vma,
460 unsigned int flags)
461{
462 /*
463 * When core dumping an enormous anonymous area that nobody
464 * has touched so far, we don't want to allocate unnecessary pages or
465 * page tables. Return error instead of NULL to skip handle_mm_fault,
466 * then get_dump_page() will return NULL to leave a hole in the dump.
467 * But we can only make this optimization where a hole would surely
468 * be zero-filled if handle_mm_fault() actually did handle it.
469 */
470 if ((flags & FOLL_DUMP) &&
471 (vma_is_anonymous(vma) || !vma->vm_ops->fault))
472 return ERR_PTR(-EFAULT);
473 return NULL;
474}
475
476static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
477 pte_t *pte, unsigned int flags)
478{
479 if (flags & FOLL_TOUCH) {
480 pte_t entry = *pte;
481
482 if (flags & FOLL_WRITE)
483 entry = pte_mkdirty(entry);
484 entry = pte_mkyoung(entry);
485
486 if (!pte_same(*pte, entry)) {
487 set_pte_at(vma->vm_mm, address, pte, entry);
488 update_mmu_cache(vma, address, pte);
489 }
490 }
491
492 /* Proper page table entry exists, but no corresponding struct page */
493 return -EEXIST;
494}
495
496/* FOLL_FORCE can write to even unwritable PTEs in COW mappings. */
497static inline bool can_follow_write_pte(pte_t pte, struct page *page,
498 struct vm_area_struct *vma,
499 unsigned int flags)
500{
501 /* If the pte is writable, we can write to the page. */
502 if (pte_write(pte))
503 return true;
504
505 /* Maybe FOLL_FORCE is set to override it? */
506 if (!(flags & FOLL_FORCE))
507 return false;
508
509 /* But FOLL_FORCE has no effect on shared mappings */
510 if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
511 return false;
512
513 /* ... or read-only private ones */
514 if (!(vma->vm_flags & VM_MAYWRITE))
515 return false;
516
517 /* ... or already writable ones that just need to take a write fault */
518 if (vma->vm_flags & VM_WRITE)
519 return false;
520
521 /*
522 * See can_change_pte_writable(): we broke COW and could map the page
523 * writable if we have an exclusive anonymous page ...
524 */
525 if (!page || !PageAnon(page) || !PageAnonExclusive(page))
526 return false;
527
528 /* ... and a write-fault isn't required for other reasons. */
529 if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte))
530 return false;
531 return !userfaultfd_pte_wp(vma, pte);
532}
533
534static struct page *follow_page_pte(struct vm_area_struct *vma,
535 unsigned long address, pmd_t *pmd, unsigned int flags,
536 struct dev_pagemap **pgmap)
537{
538 struct mm_struct *mm = vma->vm_mm;
539 struct page *page;
540 spinlock_t *ptl;
541 pte_t *ptep, pte;
542 int ret;
543
544 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
545 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
546 (FOLL_PIN | FOLL_GET)))
547 return ERR_PTR(-EINVAL);
548 if (unlikely(pmd_bad(*pmd)))
549 return no_page_table(vma, flags);
550
551 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
552 pte = *ptep;
553 if (!pte_present(pte))
554 goto no_page;
555 if (pte_protnone(pte) && !gup_can_follow_protnone(flags))
556 goto no_page;
557
558 page = vm_normal_page(vma, address, pte);
559
560 /*
561 * We only care about anon pages in can_follow_write_pte() and don't
562 * have to worry about pte_devmap() because they are never anon.
563 */
564 if ((flags & FOLL_WRITE) &&
565 !can_follow_write_pte(pte, page, vma, flags)) {
566 page = NULL;
567 goto out;
568 }
569
570 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
571 /*
572 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
573 * case since they are only valid while holding the pgmap
574 * reference.
575 */
576 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
577 if (*pgmap)
578 page = pte_page(pte);
579 else
580 goto no_page;
581 } else if (unlikely(!page)) {
582 if (flags & FOLL_DUMP) {
583 /* Avoid special (like zero) pages in core dumps */
584 page = ERR_PTR(-EFAULT);
585 goto out;
586 }
587
588 if (is_zero_pfn(pte_pfn(pte))) {
589 page = pte_page(pte);
590 } else {
591 ret = follow_pfn_pte(vma, address, ptep, flags);
592 page = ERR_PTR(ret);
593 goto out;
594 }
595 }
596
597 if (!pte_write(pte) && gup_must_unshare(vma, flags, page)) {
598 page = ERR_PTR(-EMLINK);
599 goto out;
600 }
601
602 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
603 !PageAnonExclusive(page), page);
604
605 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
606 ret = try_grab_page(page, flags);
607 if (unlikely(ret)) {
608 page = ERR_PTR(ret);
609 goto out;
610 }
611
612 /*
613 * We need to make the page accessible if and only if we are going
614 * to access its content (the FOLL_PIN case). Please see
615 * Documentation/core-api/pin_user_pages.rst for details.
616 */
617 if (flags & FOLL_PIN) {
618 ret = arch_make_page_accessible(page);
619 if (ret) {
620 unpin_user_page(page);
621 page = ERR_PTR(ret);
622 goto out;
623 }
624 }
625 if (flags & FOLL_TOUCH) {
626 if ((flags & FOLL_WRITE) &&
627 !pte_dirty(pte) && !PageDirty(page))
628 set_page_dirty(page);
629 /*
630 * pte_mkyoung() would be more correct here, but atomic care
631 * is needed to avoid losing the dirty bit: it is easier to use
632 * mark_page_accessed().
633 */
634 mark_page_accessed(page);
635 }
636out:
637 pte_unmap_unlock(ptep, ptl);
638 return page;
639no_page:
640 pte_unmap_unlock(ptep, ptl);
641 if (!pte_none(pte))
642 return NULL;
643 return no_page_table(vma, flags);
644}
645
646static struct page *follow_pmd_mask(struct vm_area_struct *vma,
647 unsigned long address, pud_t *pudp,
648 unsigned int flags,
649 struct follow_page_context *ctx)
650{
651 pmd_t *pmd, pmdval;
652 spinlock_t *ptl;
653 struct page *page;
654 struct mm_struct *mm = vma->vm_mm;
655
656 pmd = pmd_offset(pudp, address);
657 /*
658 * The READ_ONCE() will stabilize the pmdval in a register or
659 * on the stack so that it will stop changing under the code.
660 */
661 pmdval = READ_ONCE(*pmd);
662 if (pmd_none(pmdval))
663 return no_page_table(vma, flags);
664 if (!pmd_present(pmdval))
665 return no_page_table(vma, flags);
666 if (pmd_devmap(pmdval)) {
667 ptl = pmd_lock(mm, pmd);
668 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
669 spin_unlock(ptl);
670 if (page)
671 return page;
672 }
673 if (likely(!pmd_trans_huge(pmdval)))
674 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
675
676 if (pmd_protnone(pmdval) && !gup_can_follow_protnone(flags))
677 return no_page_table(vma, flags);
678
679 ptl = pmd_lock(mm, pmd);
680 if (unlikely(!pmd_present(*pmd))) {
681 spin_unlock(ptl);
682 return no_page_table(vma, flags);
683 }
684 if (unlikely(!pmd_trans_huge(*pmd))) {
685 spin_unlock(ptl);
686 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
687 }
688 if (flags & FOLL_SPLIT_PMD) {
689 int ret;
690 page = pmd_page(*pmd);
691 if (is_huge_zero_page(page)) {
692 spin_unlock(ptl);
693 ret = 0;
694 split_huge_pmd(vma, pmd, address);
695 if (pmd_trans_unstable(pmd))
696 ret = -EBUSY;
697 } else {
698 spin_unlock(ptl);
699 split_huge_pmd(vma, pmd, address);
700 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
701 }
702
703 return ret ? ERR_PTR(ret) :
704 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
705 }
706 page = follow_trans_huge_pmd(vma, address, pmd, flags);
707 spin_unlock(ptl);
708 ctx->page_mask = HPAGE_PMD_NR - 1;
709 return page;
710}
711
712static struct page *follow_pud_mask(struct vm_area_struct *vma,
713 unsigned long address, p4d_t *p4dp,
714 unsigned int flags,
715 struct follow_page_context *ctx)
716{
717 pud_t *pud;
718 spinlock_t *ptl;
719 struct page *page;
720 struct mm_struct *mm = vma->vm_mm;
721
722 pud = pud_offset(p4dp, address);
723 if (pud_none(*pud))
724 return no_page_table(vma, flags);
725 if (pud_devmap(*pud)) {
726 ptl = pud_lock(mm, pud);
727 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
728 spin_unlock(ptl);
729 if (page)
730 return page;
731 }
732 if (unlikely(pud_bad(*pud)))
733 return no_page_table(vma, flags);
734
735 return follow_pmd_mask(vma, address, pud, flags, ctx);
736}
737
738static struct page *follow_p4d_mask(struct vm_area_struct *vma,
739 unsigned long address, pgd_t *pgdp,
740 unsigned int flags,
741 struct follow_page_context *ctx)
742{
743 p4d_t *p4d;
744
745 p4d = p4d_offset(pgdp, address);
746 if (p4d_none(*p4d))
747 return no_page_table(vma, flags);
748 BUILD_BUG_ON(p4d_huge(*p4d));
749 if (unlikely(p4d_bad(*p4d)))
750 return no_page_table(vma, flags);
751
752 return follow_pud_mask(vma, address, p4d, flags, ctx);
753}
754
755/**
756 * follow_page_mask - look up a page descriptor from a user-virtual address
757 * @vma: vm_area_struct mapping @address
758 * @address: virtual address to look up
759 * @flags: flags modifying lookup behaviour
760 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
761 * pointer to output page_mask
762 *
763 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
764 *
765 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
766 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
767 *
768 * When getting an anonymous page and the caller has to trigger unsharing
769 * of a shared anonymous page first, -EMLINK is returned. The caller should
770 * trigger a fault with FAULT_FLAG_UNSHARE set. Note that unsharing is only
771 * relevant with FOLL_PIN and !FOLL_WRITE.
772 *
773 * On output, the @ctx->page_mask is set according to the size of the page.
774 *
775 * Return: the mapped (struct page *), %NULL if no mapping exists, or
776 * an error pointer if there is a mapping to something not represented
777 * by a page descriptor (see also vm_normal_page()).
778 */
779static struct page *follow_page_mask(struct vm_area_struct *vma,
780 unsigned long address, unsigned int flags,
781 struct follow_page_context *ctx)
782{
783 pgd_t *pgd;
784 struct page *page;
785 struct mm_struct *mm = vma->vm_mm;
786
787 ctx->page_mask = 0;
788
789 /*
790 * Call hugetlb_follow_page_mask for hugetlb vmas as it will use
791 * special hugetlb page table walking code. This eliminates the
792 * need to check for hugetlb entries in the general walking code.
793 *
794 * hugetlb_follow_page_mask is only for follow_page() handling here.
795 * Ordinary GUP uses follow_hugetlb_page for hugetlb processing.
796 */
797 if (is_vm_hugetlb_page(vma)) {
798 page = hugetlb_follow_page_mask(vma, address, flags);
799 if (!page)
800 page = no_page_table(vma, flags);
801 return page;
802 }
803
804 pgd = pgd_offset(mm, address);
805
806 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
807 return no_page_table(vma, flags);
808
809 return follow_p4d_mask(vma, address, pgd, flags, ctx);
810}
811
812struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
813 unsigned int foll_flags)
814{
815 struct follow_page_context ctx = { NULL };
816 struct page *page;
817
818 if (vma_is_secretmem(vma))
819 return NULL;
820
821 if (foll_flags & FOLL_PIN)
822 return NULL;
823
824 page = follow_page_mask(vma, address, foll_flags, &ctx);
825 if (ctx.pgmap)
826 put_dev_pagemap(ctx.pgmap);
827 return page;
828}
829
830static int get_gate_page(struct mm_struct *mm, unsigned long address,
831 unsigned int gup_flags, struct vm_area_struct **vma,
832 struct page **page)
833{
834 pgd_t *pgd;
835 p4d_t *p4d;
836 pud_t *pud;
837 pmd_t *pmd;
838 pte_t *pte;
839 int ret = -EFAULT;
840
841 /* user gate pages are read-only */
842 if (gup_flags & FOLL_WRITE)
843 return -EFAULT;
844 if (address > TASK_SIZE)
845 pgd = pgd_offset_k(address);
846 else
847 pgd = pgd_offset_gate(mm, address);
848 if (pgd_none(*pgd))
849 return -EFAULT;
850 p4d = p4d_offset(pgd, address);
851 if (p4d_none(*p4d))
852 return -EFAULT;
853 pud = pud_offset(p4d, address);
854 if (pud_none(*pud))
855 return -EFAULT;
856 pmd = pmd_offset(pud, address);
857 if (!pmd_present(*pmd))
858 return -EFAULT;
859 VM_BUG_ON(pmd_trans_huge(*pmd));
860 pte = pte_offset_map(pmd, address);
861 if (pte_none(*pte))
862 goto unmap;
863 *vma = get_gate_vma(mm);
864 if (!page)
865 goto out;
866 *page = vm_normal_page(*vma, address, *pte);
867 if (!*page) {
868 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
869 goto unmap;
870 *page = pte_page(*pte);
871 }
872 ret = try_grab_page(*page, gup_flags);
873 if (unlikely(ret))
874 goto unmap;
875out:
876 ret = 0;
877unmap:
878 pte_unmap(pte);
879 return ret;
880}
881
882/*
883 * mmap_lock must be held on entry. If @locked != NULL and *@flags
884 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
885 * is, *@locked will be set to 0 and -EBUSY returned.
886 */
887static int faultin_page(struct vm_area_struct *vma,
888 unsigned long address, unsigned int *flags, bool unshare,
889 int *locked)
890{
891 unsigned int fault_flags = 0;
892 vm_fault_t ret;
893
894 if (*flags & FOLL_NOFAULT)
895 return -EFAULT;
896 if (*flags & FOLL_WRITE)
897 fault_flags |= FAULT_FLAG_WRITE;
898 if (*flags & FOLL_REMOTE)
899 fault_flags |= FAULT_FLAG_REMOTE;
900 if (locked) {
901 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
902 /*
903 * FAULT_FLAG_INTERRUPTIBLE is opt-in. GUP callers must set
904 * FOLL_INTERRUPTIBLE to enable FAULT_FLAG_INTERRUPTIBLE.
905 * That's because some callers may not be prepared to
906 * handle early exits caused by non-fatal signals.
907 */
908 if (*flags & FOLL_INTERRUPTIBLE)
909 fault_flags |= FAULT_FLAG_INTERRUPTIBLE;
910 }
911 if (*flags & FOLL_NOWAIT)
912 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
913 if (*flags & FOLL_TRIED) {
914 /*
915 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
916 * can co-exist
917 */
918 fault_flags |= FAULT_FLAG_TRIED;
919 }
920 if (unshare) {
921 fault_flags |= FAULT_FLAG_UNSHARE;
922 /* FAULT_FLAG_WRITE and FAULT_FLAG_UNSHARE are incompatible */
923 VM_BUG_ON(fault_flags & FAULT_FLAG_WRITE);
924 }
925
926 ret = handle_mm_fault(vma, address, fault_flags, NULL);
927
928 if (ret & VM_FAULT_COMPLETED) {
929 /*
930 * With FAULT_FLAG_RETRY_NOWAIT we'll never release the
931 * mmap lock in the page fault handler. Sanity check this.
932 */
933 WARN_ON_ONCE(fault_flags & FAULT_FLAG_RETRY_NOWAIT);
934 if (locked)
935 *locked = 0;
936 /*
937 * We should do the same as VM_FAULT_RETRY, but let's not
938 * return -EBUSY since that's not reflecting the reality of
939 * what has happened - we've just fully completed a page
940 * fault, with the mmap lock released. Use -EAGAIN to show
941 * that we want to take the mmap lock _again_.
942 */
943 return -EAGAIN;
944 }
945
946 if (ret & VM_FAULT_ERROR) {
947 int err = vm_fault_to_errno(ret, *flags);
948
949 if (err)
950 return err;
951 BUG();
952 }
953
954 if (ret & VM_FAULT_RETRY) {
955 if (locked && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
956 *locked = 0;
957 return -EBUSY;
958 }
959
960 return 0;
961}
962
963static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
964{
965 vm_flags_t vm_flags = vma->vm_flags;
966 int write = (gup_flags & FOLL_WRITE);
967 int foreign = (gup_flags & FOLL_REMOTE);
968
969 if (vm_flags & (VM_IO | VM_PFNMAP))
970 return -EFAULT;
971
972 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
973 return -EFAULT;
974
975 if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma))
976 return -EOPNOTSUPP;
977
978 if ((gup_flags & FOLL_LONGTERM) && (gup_flags & FOLL_PCI_P2PDMA))
979 return -EOPNOTSUPP;
980
981 if (vma_is_secretmem(vma))
982 return -EFAULT;
983
984 if (write) {
985 if (!(vm_flags & VM_WRITE)) {
986 if (!(gup_flags & FOLL_FORCE))
987 return -EFAULT;
988 /* hugetlb does not support FOLL_FORCE|FOLL_WRITE. */
989 if (is_vm_hugetlb_page(vma))
990 return -EFAULT;
991 /*
992 * We used to let the write,force case do COW in a
993 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
994 * set a breakpoint in a read-only mapping of an
995 * executable, without corrupting the file (yet only
996 * when that file had been opened for writing!).
997 * Anon pages in shared mappings are surprising: now
998 * just reject it.
999 */
1000 if (!is_cow_mapping(vm_flags))
1001 return -EFAULT;
1002 }
1003 } else if (!(vm_flags & VM_READ)) {
1004 if (!(gup_flags & FOLL_FORCE))
1005 return -EFAULT;
1006 /*
1007 * Is there actually any vma we can reach here which does not
1008 * have VM_MAYREAD set?
1009 */
1010 if (!(vm_flags & VM_MAYREAD))
1011 return -EFAULT;
1012 }
1013 /*
1014 * gups are always data accesses, not instruction
1015 * fetches, so execute=false here
1016 */
1017 if (!arch_vma_access_permitted(vma, write, false, foreign))
1018 return -EFAULT;
1019 return 0;
1020}
1021
1022/**
1023 * __get_user_pages() - pin user pages in memory
1024 * @mm: mm_struct of target mm
1025 * @start: starting user address
1026 * @nr_pages: number of pages from start to pin
1027 * @gup_flags: flags modifying pin behaviour
1028 * @pages: array that receives pointers to the pages pinned.
1029 * Should be at least nr_pages long. Or NULL, if caller
1030 * only intends to ensure the pages are faulted in.
1031 * @vmas: array of pointers to vmas corresponding to each page.
1032 * Or NULL if the caller does not require them.
1033 * @locked: whether we're still with the mmap_lock held
1034 *
1035 * Returns either number of pages pinned (which may be less than the
1036 * number requested), or an error. Details about the return value:
1037 *
1038 * -- If nr_pages is 0, returns 0.
1039 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1040 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1041 * pages pinned. Again, this may be less than nr_pages.
1042 * -- 0 return value is possible when the fault would need to be retried.
1043 *
1044 * The caller is responsible for releasing returned @pages, via put_page().
1045 *
1046 * @vmas are valid only as long as mmap_lock is held.
1047 *
1048 * Must be called with mmap_lock held. It may be released. See below.
1049 *
1050 * __get_user_pages walks a process's page tables and takes a reference to
1051 * each struct page that each user address corresponds to at a given
1052 * instant. That is, it takes the page that would be accessed if a user
1053 * thread accesses the given user virtual address at that instant.
1054 *
1055 * This does not guarantee that the page exists in the user mappings when
1056 * __get_user_pages returns, and there may even be a completely different
1057 * page there in some cases (eg. if mmapped pagecache has been invalidated
1058 * and subsequently re faulted). However it does guarantee that the page
1059 * won't be freed completely. And mostly callers simply care that the page
1060 * contains data that was valid *at some point in time*. Typically, an IO
1061 * or similar operation cannot guarantee anything stronger anyway because
1062 * locks can't be held over the syscall boundary.
1063 *
1064 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1065 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1066 * appropriate) must be called after the page is finished with, and
1067 * before put_page is called.
1068 *
1069 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
1070 * released by an up_read(). That can happen if @gup_flags does not
1071 * have FOLL_NOWAIT.
1072 *
1073 * A caller using such a combination of @locked and @gup_flags
1074 * must therefore hold the mmap_lock for reading only, and recognize
1075 * when it's been released. Otherwise, it must be held for either
1076 * reading or writing and will not be released.
1077 *
1078 * In most cases, get_user_pages or get_user_pages_fast should be used
1079 * instead of __get_user_pages. __get_user_pages should be used only if
1080 * you need some special @gup_flags.
1081 */
1082static long __get_user_pages(struct mm_struct *mm,
1083 unsigned long start, unsigned long nr_pages,
1084 unsigned int gup_flags, struct page **pages,
1085 struct vm_area_struct **vmas, int *locked)
1086{
1087 long ret = 0, i = 0;
1088 struct vm_area_struct *vma = NULL;
1089 struct follow_page_context ctx = { NULL };
1090
1091 if (!nr_pages)
1092 return 0;
1093
1094 start = untagged_addr(start);
1095
1096 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1097
1098 do {
1099 struct page *page;
1100 unsigned int foll_flags = gup_flags;
1101 unsigned int page_increm;
1102
1103 /* first iteration or cross vma bound */
1104 if (!vma || start >= vma->vm_end) {
1105 vma = find_extend_vma(mm, start);
1106 if (!vma && in_gate_area(mm, start)) {
1107 ret = get_gate_page(mm, start & PAGE_MASK,
1108 gup_flags, &vma,
1109 pages ? &pages[i] : NULL);
1110 if (ret)
1111 goto out;
1112 ctx.page_mask = 0;
1113 goto next_page;
1114 }
1115
1116 if (!vma) {
1117 ret = -EFAULT;
1118 goto out;
1119 }
1120 ret = check_vma_flags(vma, gup_flags);
1121 if (ret)
1122 goto out;
1123
1124 if (is_vm_hugetlb_page(vma)) {
1125 i = follow_hugetlb_page(mm, vma, pages, vmas,
1126 &start, &nr_pages, i,
1127 gup_flags, locked);
1128 if (locked && *locked == 0) {
1129 /*
1130 * We've got a VM_FAULT_RETRY
1131 * and we've lost mmap_lock.
1132 * We must stop here.
1133 */
1134 BUG_ON(gup_flags & FOLL_NOWAIT);
1135 goto out;
1136 }
1137 continue;
1138 }
1139 }
1140retry:
1141 /*
1142 * If we have a pending SIGKILL, don't keep faulting pages and
1143 * potentially allocating memory.
1144 */
1145 if (fatal_signal_pending(current)) {
1146 ret = -EINTR;
1147 goto out;
1148 }
1149 cond_resched();
1150
1151 page = follow_page_mask(vma, start, foll_flags, &ctx);
1152 if (!page || PTR_ERR(page) == -EMLINK) {
1153 ret = faultin_page(vma, start, &foll_flags,
1154 PTR_ERR(page) == -EMLINK, locked);
1155 switch (ret) {
1156 case 0:
1157 goto retry;
1158 case -EBUSY:
1159 case -EAGAIN:
1160 ret = 0;
1161 fallthrough;
1162 case -EFAULT:
1163 case -ENOMEM:
1164 case -EHWPOISON:
1165 goto out;
1166 }
1167 BUG();
1168 } else if (PTR_ERR(page) == -EEXIST) {
1169 /*
1170 * Proper page table entry exists, but no corresponding
1171 * struct page. If the caller expects **pages to be
1172 * filled in, bail out now, because that can't be done
1173 * for this page.
1174 */
1175 if (pages) {
1176 ret = PTR_ERR(page);
1177 goto out;
1178 }
1179
1180 goto next_page;
1181 } else if (IS_ERR(page)) {
1182 ret = PTR_ERR(page);
1183 goto out;
1184 }
1185 if (pages) {
1186 pages[i] = page;
1187 flush_anon_page(vma, page, start);
1188 flush_dcache_page(page);
1189 ctx.page_mask = 0;
1190 }
1191next_page:
1192 if (vmas) {
1193 vmas[i] = vma;
1194 ctx.page_mask = 0;
1195 }
1196 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1197 if (page_increm > nr_pages)
1198 page_increm = nr_pages;
1199 i += page_increm;
1200 start += page_increm * PAGE_SIZE;
1201 nr_pages -= page_increm;
1202 } while (nr_pages);
1203out:
1204 if (ctx.pgmap)
1205 put_dev_pagemap(ctx.pgmap);
1206 return i ? i : ret;
1207}
1208
1209static bool vma_permits_fault(struct vm_area_struct *vma,
1210 unsigned int fault_flags)
1211{
1212 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1213 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1214 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1215
1216 if (!(vm_flags & vma->vm_flags))
1217 return false;
1218
1219 /*
1220 * The architecture might have a hardware protection
1221 * mechanism other than read/write that can deny access.
1222 *
1223 * gup always represents data access, not instruction
1224 * fetches, so execute=false here:
1225 */
1226 if (!arch_vma_access_permitted(vma, write, false, foreign))
1227 return false;
1228
1229 return true;
1230}
1231
1232/**
1233 * fixup_user_fault() - manually resolve a user page fault
1234 * @mm: mm_struct of target mm
1235 * @address: user address
1236 * @fault_flags:flags to pass down to handle_mm_fault()
1237 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1238 * does not allow retry. If NULL, the caller must guarantee
1239 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1240 *
1241 * This is meant to be called in the specific scenario where for locking reasons
1242 * we try to access user memory in atomic context (within a pagefault_disable()
1243 * section), this returns -EFAULT, and we want to resolve the user fault before
1244 * trying again.
1245 *
1246 * Typically this is meant to be used by the futex code.
1247 *
1248 * The main difference with get_user_pages() is that this function will
1249 * unconditionally call handle_mm_fault() which will in turn perform all the
1250 * necessary SW fixup of the dirty and young bits in the PTE, while
1251 * get_user_pages() only guarantees to update these in the struct page.
1252 *
1253 * This is important for some architectures where those bits also gate the
1254 * access permission to the page because they are maintained in software. On
1255 * such architectures, gup() will not be enough to make a subsequent access
1256 * succeed.
1257 *
1258 * This function will not return with an unlocked mmap_lock. So it has not the
1259 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1260 */
1261int fixup_user_fault(struct mm_struct *mm,
1262 unsigned long address, unsigned int fault_flags,
1263 bool *unlocked)
1264{
1265 struct vm_area_struct *vma;
1266 vm_fault_t ret;
1267
1268 address = untagged_addr(address);
1269
1270 if (unlocked)
1271 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1272
1273retry:
1274 vma = find_extend_vma(mm, address);
1275 if (!vma || address < vma->vm_start)
1276 return -EFAULT;
1277
1278 if (!vma_permits_fault(vma, fault_flags))
1279 return -EFAULT;
1280
1281 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1282 fatal_signal_pending(current))
1283 return -EINTR;
1284
1285 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1286
1287 if (ret & VM_FAULT_COMPLETED) {
1288 /*
1289 * NOTE: it's a pity that we need to retake the lock here
1290 * to pair with the unlock() in the callers. Ideally we
1291 * could tell the callers so they do not need to unlock.
1292 */
1293 mmap_read_lock(mm);
1294 *unlocked = true;
1295 return 0;
1296 }
1297
1298 if (ret & VM_FAULT_ERROR) {
1299 int err = vm_fault_to_errno(ret, 0);
1300
1301 if (err)
1302 return err;
1303 BUG();
1304 }
1305
1306 if (ret & VM_FAULT_RETRY) {
1307 mmap_read_lock(mm);
1308 *unlocked = true;
1309 fault_flags |= FAULT_FLAG_TRIED;
1310 goto retry;
1311 }
1312
1313 return 0;
1314}
1315EXPORT_SYMBOL_GPL(fixup_user_fault);
1316
1317/*
1318 * GUP always responds to fatal signals. When FOLL_INTERRUPTIBLE is
1319 * specified, it'll also respond to generic signals. The caller of GUP
1320 * that has FOLL_INTERRUPTIBLE should take care of the GUP interruption.
1321 */
1322static bool gup_signal_pending(unsigned int flags)
1323{
1324 if (fatal_signal_pending(current))
1325 return true;
1326
1327 if (!(flags & FOLL_INTERRUPTIBLE))
1328 return false;
1329
1330 return signal_pending(current);
1331}
1332
1333/*
1334 * Please note that this function, unlike __get_user_pages will not
1335 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1336 */
1337static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1338 unsigned long start,
1339 unsigned long nr_pages,
1340 struct page **pages,
1341 struct vm_area_struct **vmas,
1342 int *locked,
1343 unsigned int flags)
1344{
1345 long ret, pages_done;
1346 bool lock_dropped;
1347
1348 if (locked) {
1349 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1350 BUG_ON(vmas);
1351 /* check caller initialized locked */
1352 BUG_ON(*locked != 1);
1353 }
1354
1355 if (flags & FOLL_PIN)
1356 mm_set_has_pinned_flag(&mm->flags);
1357
1358 /*
1359 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1360 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1361 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1362 * for FOLL_GET, not for the newer FOLL_PIN.
1363 *
1364 * FOLL_PIN always expects pages to be non-null, but no need to assert
1365 * that here, as any failures will be obvious enough.
1366 */
1367 if (pages && !(flags & FOLL_PIN))
1368 flags |= FOLL_GET;
1369
1370 pages_done = 0;
1371 lock_dropped = false;
1372 for (;;) {
1373 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1374 vmas, locked);
1375 if (!locked)
1376 /* VM_FAULT_RETRY couldn't trigger, bypass */
1377 return ret;
1378
1379 /* VM_FAULT_RETRY or VM_FAULT_COMPLETED cannot return errors */
1380 if (!*locked) {
1381 BUG_ON(ret < 0);
1382 BUG_ON(ret >= nr_pages);
1383 }
1384
1385 if (ret > 0) {
1386 nr_pages -= ret;
1387 pages_done += ret;
1388 if (!nr_pages)
1389 break;
1390 }
1391 if (*locked) {
1392 /*
1393 * VM_FAULT_RETRY didn't trigger or it was a
1394 * FOLL_NOWAIT.
1395 */
1396 if (!pages_done)
1397 pages_done = ret;
1398 break;
1399 }
1400 /*
1401 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1402 * For the prefault case (!pages) we only update counts.
1403 */
1404 if (likely(pages))
1405 pages += ret;
1406 start += ret << PAGE_SHIFT;
1407 lock_dropped = true;
1408
1409retry:
1410 /*
1411 * Repeat on the address that fired VM_FAULT_RETRY
1412 * with both FAULT_FLAG_ALLOW_RETRY and
1413 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1414 * by fatal signals of even common signals, depending on
1415 * the caller's request. So we need to check it before we
1416 * start trying again otherwise it can loop forever.
1417 */
1418 if (gup_signal_pending(flags)) {
1419 if (!pages_done)
1420 pages_done = -EINTR;
1421 break;
1422 }
1423
1424 ret = mmap_read_lock_killable(mm);
1425 if (ret) {
1426 BUG_ON(ret > 0);
1427 if (!pages_done)
1428 pages_done = ret;
1429 break;
1430 }
1431
1432 *locked = 1;
1433 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1434 pages, NULL, locked);
1435 if (!*locked) {
1436 /* Continue to retry until we succeeded */
1437 BUG_ON(ret != 0);
1438 goto retry;
1439 }
1440 if (ret != 1) {
1441 BUG_ON(ret > 1);
1442 if (!pages_done)
1443 pages_done = ret;
1444 break;
1445 }
1446 nr_pages--;
1447 pages_done++;
1448 if (!nr_pages)
1449 break;
1450 if (likely(pages))
1451 pages++;
1452 start += PAGE_SIZE;
1453 }
1454 if (lock_dropped && *locked) {
1455 /*
1456 * We must let the caller know we temporarily dropped the lock
1457 * and so the critical section protected by it was lost.
1458 */
1459 mmap_read_unlock(mm);
1460 *locked = 0;
1461 }
1462 return pages_done;
1463}
1464
1465/**
1466 * populate_vma_page_range() - populate a range of pages in the vma.
1467 * @vma: target vma
1468 * @start: start address
1469 * @end: end address
1470 * @locked: whether the mmap_lock is still held
1471 *
1472 * This takes care of mlocking the pages too if VM_LOCKED is set.
1473 *
1474 * Return either number of pages pinned in the vma, or a negative error
1475 * code on error.
1476 *
1477 * vma->vm_mm->mmap_lock must be held.
1478 *
1479 * If @locked is NULL, it may be held for read or write and will
1480 * be unperturbed.
1481 *
1482 * If @locked is non-NULL, it must held for read only and may be
1483 * released. If it's released, *@locked will be set to 0.
1484 */
1485long populate_vma_page_range(struct vm_area_struct *vma,
1486 unsigned long start, unsigned long end, int *locked)
1487{
1488 struct mm_struct *mm = vma->vm_mm;
1489 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1490 int gup_flags;
1491 long ret;
1492
1493 VM_BUG_ON(!PAGE_ALIGNED(start));
1494 VM_BUG_ON(!PAGE_ALIGNED(end));
1495 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1496 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1497 mmap_assert_locked(mm);
1498
1499 /*
1500 * Rightly or wrongly, the VM_LOCKONFAULT case has never used
1501 * faultin_page() to break COW, so it has no work to do here.
1502 */
1503 if (vma->vm_flags & VM_LOCKONFAULT)
1504 return nr_pages;
1505
1506 gup_flags = FOLL_TOUCH;
1507 /*
1508 * We want to touch writable mappings with a write fault in order
1509 * to break COW, except for shared mappings because these don't COW
1510 * and we would not want to dirty them for nothing.
1511 */
1512 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1513 gup_flags |= FOLL_WRITE;
1514
1515 /*
1516 * We want mlock to succeed for regions that have any permissions
1517 * other than PROT_NONE.
1518 */
1519 if (vma_is_accessible(vma))
1520 gup_flags |= FOLL_FORCE;
1521
1522 /*
1523 * We made sure addr is within a VMA, so the following will
1524 * not result in a stack expansion that recurses back here.
1525 */
1526 ret = __get_user_pages(mm, start, nr_pages, gup_flags,
1527 NULL, NULL, locked);
1528 lru_add_drain();
1529 return ret;
1530}
1531
1532/*
1533 * faultin_vma_page_range() - populate (prefault) page tables inside the
1534 * given VMA range readable/writable
1535 *
1536 * This takes care of mlocking the pages, too, if VM_LOCKED is set.
1537 *
1538 * @vma: target vma
1539 * @start: start address
1540 * @end: end address
1541 * @write: whether to prefault readable or writable
1542 * @locked: whether the mmap_lock is still held
1543 *
1544 * Returns either number of processed pages in the vma, or a negative error
1545 * code on error (see __get_user_pages()).
1546 *
1547 * vma->vm_mm->mmap_lock must be held. The range must be page-aligned and
1548 * covered by the VMA.
1549 *
1550 * If @locked is NULL, it may be held for read or write and will be unperturbed.
1551 *
1552 * If @locked is non-NULL, it must held for read only and may be released. If
1553 * it's released, *@locked will be set to 0.
1554 */
1555long faultin_vma_page_range(struct vm_area_struct *vma, unsigned long start,
1556 unsigned long end, bool write, int *locked)
1557{
1558 struct mm_struct *mm = vma->vm_mm;
1559 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1560 int gup_flags;
1561 long ret;
1562
1563 VM_BUG_ON(!PAGE_ALIGNED(start));
1564 VM_BUG_ON(!PAGE_ALIGNED(end));
1565 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1566 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1567 mmap_assert_locked(mm);
1568
1569 /*
1570 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark
1571 * the page dirty with FOLL_WRITE -- which doesn't make a
1572 * difference with !FOLL_FORCE, because the page is writable
1573 * in the page table.
1574 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit
1575 * a poisoned page.
1576 * !FOLL_FORCE: Require proper access permissions.
1577 */
1578 gup_flags = FOLL_TOUCH | FOLL_HWPOISON;
1579 if (write)
1580 gup_flags |= FOLL_WRITE;
1581
1582 /*
1583 * We want to report -EINVAL instead of -EFAULT for any permission
1584 * problems or incompatible mappings.
1585 */
1586 if (check_vma_flags(vma, gup_flags))
1587 return -EINVAL;
1588
1589 ret = __get_user_pages(mm, start, nr_pages, gup_flags,
1590 NULL, NULL, locked);
1591 lru_add_drain();
1592 return ret;
1593}
1594
1595/*
1596 * __mm_populate - populate and/or mlock pages within a range of address space.
1597 *
1598 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1599 * flags. VMAs must be already marked with the desired vm_flags, and
1600 * mmap_lock must not be held.
1601 */
1602int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1603{
1604 struct mm_struct *mm = current->mm;
1605 unsigned long end, nstart, nend;
1606 struct vm_area_struct *vma = NULL;
1607 int locked = 0;
1608 long ret = 0;
1609
1610 end = start + len;
1611
1612 for (nstart = start; nstart < end; nstart = nend) {
1613 /*
1614 * We want to fault in pages for [nstart; end) address range.
1615 * Find first corresponding VMA.
1616 */
1617 if (!locked) {
1618 locked = 1;
1619 mmap_read_lock(mm);
1620 vma = find_vma_intersection(mm, nstart, end);
1621 } else if (nstart >= vma->vm_end)
1622 vma = find_vma_intersection(mm, vma->vm_end, end);
1623
1624 if (!vma)
1625 break;
1626 /*
1627 * Set [nstart; nend) to intersection of desired address
1628 * range with the first VMA. Also, skip undesirable VMA types.
1629 */
1630 nend = min(end, vma->vm_end);
1631 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1632 continue;
1633 if (nstart < vma->vm_start)
1634 nstart = vma->vm_start;
1635 /*
1636 * Now fault in a range of pages. populate_vma_page_range()
1637 * double checks the vma flags, so that it won't mlock pages
1638 * if the vma was already munlocked.
1639 */
1640 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1641 if (ret < 0) {
1642 if (ignore_errors) {
1643 ret = 0;
1644 continue; /* continue at next VMA */
1645 }
1646 break;
1647 }
1648 nend = nstart + ret * PAGE_SIZE;
1649 ret = 0;
1650 }
1651 if (locked)
1652 mmap_read_unlock(mm);
1653 return ret; /* 0 or negative error code */
1654}
1655#else /* CONFIG_MMU */
1656static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1657 unsigned long nr_pages, struct page **pages,
1658 struct vm_area_struct **vmas, int *locked,
1659 unsigned int foll_flags)
1660{
1661 struct vm_area_struct *vma;
1662 unsigned long vm_flags;
1663 long i;
1664
1665 /* calculate required read or write permissions.
1666 * If FOLL_FORCE is set, we only require the "MAY" flags.
1667 */
1668 vm_flags = (foll_flags & FOLL_WRITE) ?
1669 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1670 vm_flags &= (foll_flags & FOLL_FORCE) ?
1671 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1672
1673 for (i = 0; i < nr_pages; i++) {
1674 vma = find_vma(mm, start);
1675 if (!vma)
1676 goto finish_or_fault;
1677
1678 /* protect what we can, including chardevs */
1679 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1680 !(vm_flags & vma->vm_flags))
1681 goto finish_or_fault;
1682
1683 if (pages) {
1684 pages[i] = virt_to_page((void *)start);
1685 if (pages[i])
1686 get_page(pages[i]);
1687 }
1688 if (vmas)
1689 vmas[i] = vma;
1690 start = (start + PAGE_SIZE) & PAGE_MASK;
1691 }
1692
1693 return i;
1694
1695finish_or_fault:
1696 return i ? : -EFAULT;
1697}
1698#endif /* !CONFIG_MMU */
1699
1700/**
1701 * fault_in_writeable - fault in userspace address range for writing
1702 * @uaddr: start of address range
1703 * @size: size of address range
1704 *
1705 * Returns the number of bytes not faulted in (like copy_to_user() and
1706 * copy_from_user()).
1707 */
1708size_t fault_in_writeable(char __user *uaddr, size_t size)
1709{
1710 char __user *start = uaddr, *end;
1711
1712 if (unlikely(size == 0))
1713 return 0;
1714 if (!user_write_access_begin(uaddr, size))
1715 return size;
1716 if (!PAGE_ALIGNED(uaddr)) {
1717 unsafe_put_user(0, uaddr, out);
1718 uaddr = (char __user *)PAGE_ALIGN((unsigned long)uaddr);
1719 }
1720 end = (char __user *)PAGE_ALIGN((unsigned long)start + size);
1721 if (unlikely(end < start))
1722 end = NULL;
1723 while (uaddr != end) {
1724 unsafe_put_user(0, uaddr, out);
1725 uaddr += PAGE_SIZE;
1726 }
1727
1728out:
1729 user_write_access_end();
1730 if (size > uaddr - start)
1731 return size - (uaddr - start);
1732 return 0;
1733}
1734EXPORT_SYMBOL(fault_in_writeable);
1735
1736/**
1737 * fault_in_subpage_writeable - fault in an address range for writing
1738 * @uaddr: start of address range
1739 * @size: size of address range
1740 *
1741 * Fault in a user address range for writing while checking for permissions at
1742 * sub-page granularity (e.g. arm64 MTE). This function should be used when
1743 * the caller cannot guarantee forward progress of a copy_to_user() loop.
1744 *
1745 * Returns the number of bytes not faulted in (like copy_to_user() and
1746 * copy_from_user()).
1747 */
1748size_t fault_in_subpage_writeable(char __user *uaddr, size_t size)
1749{
1750 size_t faulted_in;
1751
1752 /*
1753 * Attempt faulting in at page granularity first for page table
1754 * permission checking. The arch-specific probe_subpage_writeable()
1755 * functions may not check for this.
1756 */
1757 faulted_in = size - fault_in_writeable(uaddr, size);
1758 if (faulted_in)
1759 faulted_in -= probe_subpage_writeable(uaddr, faulted_in);
1760
1761 return size - faulted_in;
1762}
1763EXPORT_SYMBOL(fault_in_subpage_writeable);
1764
1765/*
1766 * fault_in_safe_writeable - fault in an address range for writing
1767 * @uaddr: start of address range
1768 * @size: length of address range
1769 *
1770 * Faults in an address range for writing. This is primarily useful when we
1771 * already know that some or all of the pages in the address range aren't in
1772 * memory.
1773 *
1774 * Unlike fault_in_writeable(), this function is non-destructive.
1775 *
1776 * Note that we don't pin or otherwise hold the pages referenced that we fault
1777 * in. There's no guarantee that they'll stay in memory for any duration of
1778 * time.
1779 *
1780 * Returns the number of bytes not faulted in, like copy_to_user() and
1781 * copy_from_user().
1782 */
1783size_t fault_in_safe_writeable(const char __user *uaddr, size_t size)
1784{
1785 unsigned long start = (unsigned long)uaddr, end;
1786 struct mm_struct *mm = current->mm;
1787 bool unlocked = false;
1788
1789 if (unlikely(size == 0))
1790 return 0;
1791 end = PAGE_ALIGN(start + size);
1792 if (end < start)
1793 end = 0;
1794
1795 mmap_read_lock(mm);
1796 do {
1797 if (fixup_user_fault(mm, start, FAULT_FLAG_WRITE, &unlocked))
1798 break;
1799 start = (start + PAGE_SIZE) & PAGE_MASK;
1800 } while (start != end);
1801 mmap_read_unlock(mm);
1802
1803 if (size > (unsigned long)uaddr - start)
1804 return size - ((unsigned long)uaddr - start);
1805 return 0;
1806}
1807EXPORT_SYMBOL(fault_in_safe_writeable);
1808
1809/**
1810 * fault_in_readable - fault in userspace address range for reading
1811 * @uaddr: start of user address range
1812 * @size: size of user address range
1813 *
1814 * Returns the number of bytes not faulted in (like copy_to_user() and
1815 * copy_from_user()).
1816 */
1817size_t fault_in_readable(const char __user *uaddr, size_t size)
1818{
1819 const char __user *start = uaddr, *end;
1820 volatile char c;
1821
1822 if (unlikely(size == 0))
1823 return 0;
1824 if (!user_read_access_begin(uaddr, size))
1825 return size;
1826 if (!PAGE_ALIGNED(uaddr)) {
1827 unsafe_get_user(c, uaddr, out);
1828 uaddr = (const char __user *)PAGE_ALIGN((unsigned long)uaddr);
1829 }
1830 end = (const char __user *)PAGE_ALIGN((unsigned long)start + size);
1831 if (unlikely(end < start))
1832 end = NULL;
1833 while (uaddr != end) {
1834 unsafe_get_user(c, uaddr, out);
1835 uaddr += PAGE_SIZE;
1836 }
1837
1838out:
1839 user_read_access_end();
1840 (void)c;
1841 if (size > uaddr - start)
1842 return size - (uaddr - start);
1843 return 0;
1844}
1845EXPORT_SYMBOL(fault_in_readable);
1846
1847/**
1848 * get_dump_page() - pin user page in memory while writing it to core dump
1849 * @addr: user address
1850 *
1851 * Returns struct page pointer of user page pinned for dump,
1852 * to be freed afterwards by put_page().
1853 *
1854 * Returns NULL on any kind of failure - a hole must then be inserted into
1855 * the corefile, to preserve alignment with its headers; and also returns
1856 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1857 * allowing a hole to be left in the corefile to save disk space.
1858 *
1859 * Called without mmap_lock (takes and releases the mmap_lock by itself).
1860 */
1861#ifdef CONFIG_ELF_CORE
1862struct page *get_dump_page(unsigned long addr)
1863{
1864 struct mm_struct *mm = current->mm;
1865 struct page *page;
1866 int locked = 1;
1867 int ret;
1868
1869 if (mmap_read_lock_killable(mm))
1870 return NULL;
1871 ret = __get_user_pages_locked(mm, addr, 1, &page, NULL, &locked,
1872 FOLL_FORCE | FOLL_DUMP | FOLL_GET);
1873 if (locked)
1874 mmap_read_unlock(mm);
1875 return (ret == 1) ? page : NULL;
1876}
1877#endif /* CONFIG_ELF_CORE */
1878
1879#ifdef CONFIG_MIGRATION
1880/*
1881 * Returns the number of collected pages. Return value is always >= 0.
1882 */
1883static unsigned long collect_longterm_unpinnable_pages(
1884 struct list_head *movable_page_list,
1885 unsigned long nr_pages,
1886 struct page **pages)
1887{
1888 unsigned long i, collected = 0;
1889 struct folio *prev_folio = NULL;
1890 bool drain_allow = true;
1891
1892 for (i = 0; i < nr_pages; i++) {
1893 struct folio *folio = page_folio(pages[i]);
1894
1895 if (folio == prev_folio)
1896 continue;
1897 prev_folio = folio;
1898
1899 if (folio_is_longterm_pinnable(folio))
1900 continue;
1901
1902 collected++;
1903
1904 if (folio_is_device_coherent(folio))
1905 continue;
1906
1907 if (folio_test_hugetlb(folio)) {
1908 isolate_hugetlb(&folio->page, movable_page_list);
1909 continue;
1910 }
1911
1912 if (!folio_test_lru(folio) && drain_allow) {
1913 lru_add_drain_all();
1914 drain_allow = false;
1915 }
1916
1917 if (folio_isolate_lru(folio))
1918 continue;
1919
1920 list_add_tail(&folio->lru, movable_page_list);
1921 node_stat_mod_folio(folio,
1922 NR_ISOLATED_ANON + folio_is_file_lru(folio),
1923 folio_nr_pages(folio));
1924 }
1925
1926 return collected;
1927}
1928
1929/*
1930 * Unpins all pages and migrates device coherent pages and movable_page_list.
1931 * Returns -EAGAIN if all pages were successfully migrated or -errno for failure
1932 * (or partial success).
1933 */
1934static int migrate_longterm_unpinnable_pages(
1935 struct list_head *movable_page_list,
1936 unsigned long nr_pages,
1937 struct page **pages)
1938{
1939 int ret;
1940 unsigned long i;
1941
1942 for (i = 0; i < nr_pages; i++) {
1943 struct folio *folio = page_folio(pages[i]);
1944
1945 if (folio_is_device_coherent(folio)) {
1946 /*
1947 * Migration will fail if the page is pinned, so convert
1948 * the pin on the source page to a normal reference.
1949 */
1950 pages[i] = NULL;
1951 folio_get(folio);
1952 gup_put_folio(folio, 1, FOLL_PIN);
1953
1954 if (migrate_device_coherent_page(&folio->page)) {
1955 ret = -EBUSY;
1956 goto err;
1957 }
1958
1959 continue;
1960 }
1961
1962 /*
1963 * We can't migrate pages with unexpected references, so drop
1964 * the reference obtained by __get_user_pages_locked().
1965 * Migrating pages have been added to movable_page_list after
1966 * calling folio_isolate_lru() which takes a reference so the
1967 * page won't be freed if it's migrating.
1968 */
1969 unpin_user_page(pages[i]);
1970 pages[i] = NULL;
1971 }
1972
1973 if (!list_empty(movable_page_list)) {
1974 struct migration_target_control mtc = {
1975 .nid = NUMA_NO_NODE,
1976 .gfp_mask = GFP_USER | __GFP_NOWARN,
1977 };
1978
1979 if (migrate_pages(movable_page_list, alloc_migration_target,
1980 NULL, (unsigned long)&mtc, MIGRATE_SYNC,
1981 MR_LONGTERM_PIN, NULL)) {
1982 ret = -ENOMEM;
1983 goto err;
1984 }
1985 }
1986
1987 putback_movable_pages(movable_page_list);
1988
1989 return -EAGAIN;
1990
1991err:
1992 for (i = 0; i < nr_pages; i++)
1993 if (pages[i])
1994 unpin_user_page(pages[i]);
1995 putback_movable_pages(movable_page_list);
1996
1997 return ret;
1998}
1999
2000/*
2001 * Check whether all pages are *allowed* to be pinned. Rather confusingly, all
2002 * pages in the range are required to be pinned via FOLL_PIN, before calling
2003 * this routine.
2004 *
2005 * If any pages in the range are not allowed to be pinned, then this routine
2006 * will migrate those pages away, unpin all the pages in the range and return
2007 * -EAGAIN. The caller should re-pin the entire range with FOLL_PIN and then
2008 * call this routine again.
2009 *
2010 * If an error other than -EAGAIN occurs, this indicates a migration failure.
2011 * The caller should give up, and propagate the error back up the call stack.
2012 *
2013 * If everything is OK and all pages in the range are allowed to be pinned, then
2014 * this routine leaves all pages pinned and returns zero for success.
2015 */
2016static long check_and_migrate_movable_pages(unsigned long nr_pages,
2017 struct page **pages)
2018{
2019 unsigned long collected;
2020 LIST_HEAD(movable_page_list);
2021
2022 collected = collect_longterm_unpinnable_pages(&movable_page_list,
2023 nr_pages, pages);
2024 if (!collected)
2025 return 0;
2026
2027 return migrate_longterm_unpinnable_pages(&movable_page_list, nr_pages,
2028 pages);
2029}
2030#else
2031static long check_and_migrate_movable_pages(unsigned long nr_pages,
2032 struct page **pages)
2033{
2034 return 0;
2035}
2036#endif /* CONFIG_MIGRATION */
2037
2038/*
2039 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
2040 * allows us to process the FOLL_LONGTERM flag.
2041 */
2042static long __gup_longterm_locked(struct mm_struct *mm,
2043 unsigned long start,
2044 unsigned long nr_pages,
2045 struct page **pages,
2046 struct vm_area_struct **vmas,
2047 int *locked,
2048 unsigned int gup_flags)
2049{
2050 bool must_unlock = false;
2051 unsigned int flags;
2052 long rc, nr_pinned_pages;
2053
2054 if (locked && WARN_ON_ONCE(!*locked))
2055 return -EINVAL;
2056
2057 if (!(gup_flags & FOLL_LONGTERM))
2058 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
2059 locked, gup_flags);
2060
2061 /*
2062 * If we get to this point then FOLL_LONGTERM is set, and FOLL_LONGTERM
2063 * implies FOLL_PIN (although the reverse is not true). Therefore it is
2064 * correct to unconditionally call check_and_migrate_movable_pages()
2065 * which assumes pages have been pinned via FOLL_PIN.
2066 *
2067 * Enforce the above reasoning by asserting that FOLL_PIN is set.
2068 */
2069 if (WARN_ON(!(gup_flags & FOLL_PIN)))
2070 return -EINVAL;
2071 flags = memalloc_pin_save();
2072 do {
2073 if (locked && !*locked) {
2074 mmap_read_lock(mm);
2075 must_unlock = true;
2076 *locked = 1;
2077 }
2078 nr_pinned_pages = __get_user_pages_locked(mm, start, nr_pages,
2079 pages, vmas, locked,
2080 gup_flags);
2081 if (nr_pinned_pages <= 0) {
2082 rc = nr_pinned_pages;
2083 break;
2084 }
2085 rc = check_and_migrate_movable_pages(nr_pinned_pages, pages);
2086 } while (rc == -EAGAIN);
2087 memalloc_pin_restore(flags);
2088
2089 if (locked && *locked && must_unlock) {
2090 mmap_read_unlock(mm);
2091 *locked = 0;
2092 }
2093 return rc ? rc : nr_pinned_pages;
2094}
2095
2096static bool is_valid_gup_flags(unsigned int gup_flags)
2097{
2098 /*
2099 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
2100 * never directly by the caller, so enforce that with an assertion:
2101 */
2102 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
2103 return false;
2104 /*
2105 * FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying
2106 * that is, FOLL_LONGTERM is a specific case, more restrictive case of
2107 * FOLL_PIN.
2108 */
2109 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2110 return false;
2111
2112 return true;
2113}
2114
2115#ifdef CONFIG_MMU
2116/**
2117 * get_user_pages_remote() - pin user pages in memory
2118 * @mm: mm_struct of target mm
2119 * @start: starting user address
2120 * @nr_pages: number of pages from start to pin
2121 * @gup_flags: flags modifying lookup behaviour
2122 * @pages: array that receives pointers to the pages pinned.
2123 * Should be at least nr_pages long. Or NULL, if caller
2124 * only intends to ensure the pages are faulted in.
2125 * @vmas: array of pointers to vmas corresponding to each page.
2126 * Or NULL if the caller does not require them.
2127 * @locked: pointer to lock flag indicating whether lock is held and
2128 * subsequently whether VM_FAULT_RETRY functionality can be
2129 * utilised. Lock must initially be held.
2130 *
2131 * Returns either number of pages pinned (which may be less than the
2132 * number requested), or an error. Details about the return value:
2133 *
2134 * -- If nr_pages is 0, returns 0.
2135 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
2136 * -- If nr_pages is >0, and some pages were pinned, returns the number of
2137 * pages pinned. Again, this may be less than nr_pages.
2138 *
2139 * The caller is responsible for releasing returned @pages, via put_page().
2140 *
2141 * @vmas are valid only as long as mmap_lock is held.
2142 *
2143 * Must be called with mmap_lock held for read or write.
2144 *
2145 * get_user_pages_remote walks a process's page tables and takes a reference
2146 * to each struct page that each user address corresponds to at a given
2147 * instant. That is, it takes the page that would be accessed if a user
2148 * thread accesses the given user virtual address at that instant.
2149 *
2150 * This does not guarantee that the page exists in the user mappings when
2151 * get_user_pages_remote returns, and there may even be a completely different
2152 * page there in some cases (eg. if mmapped pagecache has been invalidated
2153 * and subsequently re faulted). However it does guarantee that the page
2154 * won't be freed completely. And mostly callers simply care that the page
2155 * contains data that was valid *at some point in time*. Typically, an IO
2156 * or similar operation cannot guarantee anything stronger anyway because
2157 * locks can't be held over the syscall boundary.
2158 *
2159 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
2160 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
2161 * be called after the page is finished with, and before put_page is called.
2162 *
2163 * get_user_pages_remote is typically used for fewer-copy IO operations,
2164 * to get a handle on the memory by some means other than accesses
2165 * via the user virtual addresses. The pages may be submitted for
2166 * DMA to devices or accessed via their kernel linear mapping (via the
2167 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
2168 *
2169 * See also get_user_pages_fast, for performance critical applications.
2170 *
2171 * get_user_pages_remote should be phased out in favor of
2172 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
2173 * should use get_user_pages_remote because it cannot pass
2174 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
2175 */
2176long get_user_pages_remote(struct mm_struct *mm,
2177 unsigned long start, unsigned long nr_pages,
2178 unsigned int gup_flags, struct page **pages,
2179 struct vm_area_struct **vmas, int *locked)
2180{
2181 if (!is_valid_gup_flags(gup_flags))
2182 return -EINVAL;
2183
2184 return __gup_longterm_locked(mm, start, nr_pages, pages, vmas, locked,
2185 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
2186}
2187EXPORT_SYMBOL(get_user_pages_remote);
2188
2189#else /* CONFIG_MMU */
2190long get_user_pages_remote(struct mm_struct *mm,
2191 unsigned long start, unsigned long nr_pages,
2192 unsigned int gup_flags, struct page **pages,
2193 struct vm_area_struct **vmas, int *locked)
2194{
2195 return 0;
2196}
2197#endif /* !CONFIG_MMU */
2198
2199/**
2200 * get_user_pages() - pin user pages in memory
2201 * @start: starting user address
2202 * @nr_pages: number of pages from start to pin
2203 * @gup_flags: flags modifying lookup behaviour
2204 * @pages: array that receives pointers to the pages pinned.
2205 * Should be at least nr_pages long. Or NULL, if caller
2206 * only intends to ensure the pages are faulted in.
2207 * @vmas: array of pointers to vmas corresponding to each page.
2208 * Or NULL if the caller does not require them.
2209 *
2210 * This is the same as get_user_pages_remote(), just with a less-flexible
2211 * calling convention where we assume that the mm being operated on belongs to
2212 * the current task, and doesn't allow passing of a locked parameter. We also
2213 * obviously don't pass FOLL_REMOTE in here.
2214 */
2215long get_user_pages(unsigned long start, unsigned long nr_pages,
2216 unsigned int gup_flags, struct page **pages,
2217 struct vm_area_struct **vmas)
2218{
2219 if (!is_valid_gup_flags(gup_flags))
2220 return -EINVAL;
2221
2222 return __gup_longterm_locked(current->mm, start, nr_pages,
2223 pages, vmas, NULL, gup_flags | FOLL_TOUCH);
2224}
2225EXPORT_SYMBOL(get_user_pages);
2226
2227/*
2228 * get_user_pages_unlocked() is suitable to replace the form:
2229 *
2230 * mmap_read_lock(mm);
2231 * get_user_pages(mm, ..., pages, NULL);
2232 * mmap_read_unlock(mm);
2233 *
2234 * with:
2235 *
2236 * get_user_pages_unlocked(mm, ..., pages);
2237 *
2238 * It is functionally equivalent to get_user_pages_fast so
2239 * get_user_pages_fast should be used instead if specific gup_flags
2240 * (e.g. FOLL_FORCE) are not required.
2241 */
2242long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2243 struct page **pages, unsigned int gup_flags)
2244{
2245 struct mm_struct *mm = current->mm;
2246 int locked = 1;
2247 long ret;
2248
2249 mmap_read_lock(mm);
2250 ret = __gup_longterm_locked(mm, start, nr_pages, pages, NULL, &locked,
2251 gup_flags | FOLL_TOUCH);
2252 if (locked)
2253 mmap_read_unlock(mm);
2254 return ret;
2255}
2256EXPORT_SYMBOL(get_user_pages_unlocked);
2257
2258/*
2259 * Fast GUP
2260 *
2261 * get_user_pages_fast attempts to pin user pages by walking the page
2262 * tables directly and avoids taking locks. Thus the walker needs to be
2263 * protected from page table pages being freed from under it, and should
2264 * block any THP splits.
2265 *
2266 * One way to achieve this is to have the walker disable interrupts, and
2267 * rely on IPIs from the TLB flushing code blocking before the page table
2268 * pages are freed. This is unsuitable for architectures that do not need
2269 * to broadcast an IPI when invalidating TLBs.
2270 *
2271 * Another way to achieve this is to batch up page table containing pages
2272 * belonging to more than one mm_user, then rcu_sched a callback to free those
2273 * pages. Disabling interrupts will allow the fast_gup walker to both block
2274 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2275 * (which is a relatively rare event). The code below adopts this strategy.
2276 *
2277 * Before activating this code, please be aware that the following assumptions
2278 * are currently made:
2279 *
2280 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2281 * free pages containing page tables or TLB flushing requires IPI broadcast.
2282 *
2283 * *) ptes can be read atomically by the architecture.
2284 *
2285 * *) access_ok is sufficient to validate userspace address ranges.
2286 *
2287 * The last two assumptions can be relaxed by the addition of helper functions.
2288 *
2289 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2290 */
2291#ifdef CONFIG_HAVE_FAST_GUP
2292
2293static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2294 unsigned int flags,
2295 struct page **pages)
2296{
2297 while ((*nr) - nr_start) {
2298 struct page *page = pages[--(*nr)];
2299
2300 ClearPageReferenced(page);
2301 if (flags & FOLL_PIN)
2302 unpin_user_page(page);
2303 else
2304 put_page(page);
2305 }
2306}
2307
2308#ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2309/*
2310 * Fast-gup relies on pte change detection to avoid concurrent pgtable
2311 * operations.
2312 *
2313 * To pin the page, fast-gup needs to do below in order:
2314 * (1) pin the page (by prefetching pte), then (2) check pte not changed.
2315 *
2316 * For the rest of pgtable operations where pgtable updates can be racy
2317 * with fast-gup, we need to do (1) clear pte, then (2) check whether page
2318 * is pinned.
2319 *
2320 * Above will work for all pte-level operations, including THP split.
2321 *
2322 * For THP collapse, it's a bit more complicated because fast-gup may be
2323 * walking a pgtable page that is being freed (pte is still valid but pmd
2324 * can be cleared already). To avoid race in such condition, we need to
2325 * also check pmd here to make sure pmd doesn't change (corresponds to
2326 * pmdp_collapse_flush() in the THP collapse code path).
2327 */
2328static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2329 unsigned long end, unsigned int flags,
2330 struct page **pages, int *nr)
2331{
2332 struct dev_pagemap *pgmap = NULL;
2333 int nr_start = *nr, ret = 0;
2334 pte_t *ptep, *ptem;
2335
2336 ptem = ptep = pte_offset_map(&pmd, addr);
2337 do {
2338 pte_t pte = ptep_get_lockless(ptep);
2339 struct page *page;
2340 struct folio *folio;
2341
2342 if (pte_protnone(pte) && !gup_can_follow_protnone(flags))
2343 goto pte_unmap;
2344
2345 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2346 goto pte_unmap;
2347
2348 if (pte_devmap(pte)) {
2349 if (unlikely(flags & FOLL_LONGTERM))
2350 goto pte_unmap;
2351
2352 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2353 if (unlikely(!pgmap)) {
2354 undo_dev_pagemap(nr, nr_start, flags, pages);
2355 goto pte_unmap;
2356 }
2357 } else if (pte_special(pte))
2358 goto pte_unmap;
2359
2360 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2361 page = pte_page(pte);
2362
2363 folio = try_grab_folio(page, 1, flags);
2364 if (!folio)
2365 goto pte_unmap;
2366
2367 if (unlikely(page_is_secretmem(page))) {
2368 gup_put_folio(folio, 1, flags);
2369 goto pte_unmap;
2370 }
2371
2372 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp)) ||
2373 unlikely(pte_val(pte) != pte_val(*ptep))) {
2374 gup_put_folio(folio, 1, flags);
2375 goto pte_unmap;
2376 }
2377
2378 if (!pte_write(pte) && gup_must_unshare(NULL, flags, page)) {
2379 gup_put_folio(folio, 1, flags);
2380 goto pte_unmap;
2381 }
2382
2383 /*
2384 * We need to make the page accessible if and only if we are
2385 * going to access its content (the FOLL_PIN case). Please
2386 * see Documentation/core-api/pin_user_pages.rst for
2387 * details.
2388 */
2389 if (flags & FOLL_PIN) {
2390 ret = arch_make_page_accessible(page);
2391 if (ret) {
2392 gup_put_folio(folio, 1, flags);
2393 goto pte_unmap;
2394 }
2395 }
2396 folio_set_referenced(folio);
2397 pages[*nr] = page;
2398 (*nr)++;
2399 } while (ptep++, addr += PAGE_SIZE, addr != end);
2400
2401 ret = 1;
2402
2403pte_unmap:
2404 if (pgmap)
2405 put_dev_pagemap(pgmap);
2406 pte_unmap(ptem);
2407 return ret;
2408}
2409#else
2410
2411/*
2412 * If we can't determine whether or not a pte is special, then fail immediately
2413 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2414 * to be special.
2415 *
2416 * For a futex to be placed on a THP tail page, get_futex_key requires a
2417 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2418 * useful to have gup_huge_pmd even if we can't operate on ptes.
2419 */
2420static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2421 unsigned long end, unsigned int flags,
2422 struct page **pages, int *nr)
2423{
2424 return 0;
2425}
2426#endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2427
2428#if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2429static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2430 unsigned long end, unsigned int flags,
2431 struct page **pages, int *nr)
2432{
2433 int nr_start = *nr;
2434 struct dev_pagemap *pgmap = NULL;
2435
2436 do {
2437 struct page *page = pfn_to_page(pfn);
2438
2439 pgmap = get_dev_pagemap(pfn, pgmap);
2440 if (unlikely(!pgmap)) {
2441 undo_dev_pagemap(nr, nr_start, flags, pages);
2442 break;
2443 }
2444
2445 if (!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)) {
2446 undo_dev_pagemap(nr, nr_start, flags, pages);
2447 break;
2448 }
2449
2450 SetPageReferenced(page);
2451 pages[*nr] = page;
2452 if (unlikely(try_grab_page(page, flags))) {
2453 undo_dev_pagemap(nr, nr_start, flags, pages);
2454 break;
2455 }
2456 (*nr)++;
2457 pfn++;
2458 } while (addr += PAGE_SIZE, addr != end);
2459
2460 put_dev_pagemap(pgmap);
2461 return addr == end;
2462}
2463
2464static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2465 unsigned long end, unsigned int flags,
2466 struct page **pages, int *nr)
2467{
2468 unsigned long fault_pfn;
2469 int nr_start = *nr;
2470
2471 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2472 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2473 return 0;
2474
2475 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2476 undo_dev_pagemap(nr, nr_start, flags, pages);
2477 return 0;
2478 }
2479 return 1;
2480}
2481
2482static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2483 unsigned long end, unsigned int flags,
2484 struct page **pages, int *nr)
2485{
2486 unsigned long fault_pfn;
2487 int nr_start = *nr;
2488
2489 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2490 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2491 return 0;
2492
2493 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2494 undo_dev_pagemap(nr, nr_start, flags, pages);
2495 return 0;
2496 }
2497 return 1;
2498}
2499#else
2500static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2501 unsigned long end, unsigned int flags,
2502 struct page **pages, int *nr)
2503{
2504 BUILD_BUG();
2505 return 0;
2506}
2507
2508static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2509 unsigned long end, unsigned int flags,
2510 struct page **pages, int *nr)
2511{
2512 BUILD_BUG();
2513 return 0;
2514}
2515#endif
2516
2517static int record_subpages(struct page *page, unsigned long addr,
2518 unsigned long end, struct page **pages)
2519{
2520 int nr;
2521
2522 for (nr = 0; addr != end; nr++, addr += PAGE_SIZE)
2523 pages[nr] = nth_page(page, nr);
2524
2525 return nr;
2526}
2527
2528#ifdef CONFIG_ARCH_HAS_HUGEPD
2529static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2530 unsigned long sz)
2531{
2532 unsigned long __boundary = (addr + sz) & ~(sz-1);
2533 return (__boundary - 1 < end - 1) ? __boundary : end;
2534}
2535
2536static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2537 unsigned long end, unsigned int flags,
2538 struct page **pages, int *nr)
2539{
2540 unsigned long pte_end;
2541 struct page *page;
2542 struct folio *folio;
2543 pte_t pte;
2544 int refs;
2545
2546 pte_end = (addr + sz) & ~(sz-1);
2547 if (pte_end < end)
2548 end = pte_end;
2549
2550 pte = huge_ptep_get(ptep);
2551
2552 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2553 return 0;
2554
2555 /* hugepages are never "special" */
2556 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2557
2558 page = nth_page(pte_page(pte), (addr & (sz - 1)) >> PAGE_SHIFT);
2559 refs = record_subpages(page, addr, end, pages + *nr);
2560
2561 folio = try_grab_folio(page, refs, flags);
2562 if (!folio)
2563 return 0;
2564
2565 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2566 gup_put_folio(folio, refs, flags);
2567 return 0;
2568 }
2569
2570 if (!pte_write(pte) && gup_must_unshare(NULL, flags, &folio->page)) {
2571 gup_put_folio(folio, refs, flags);
2572 return 0;
2573 }
2574
2575 *nr += refs;
2576 folio_set_referenced(folio);
2577 return 1;
2578}
2579
2580static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2581 unsigned int pdshift, unsigned long end, unsigned int flags,
2582 struct page **pages, int *nr)
2583{
2584 pte_t *ptep;
2585 unsigned long sz = 1UL << hugepd_shift(hugepd);
2586 unsigned long next;
2587
2588 ptep = hugepte_offset(hugepd, addr, pdshift);
2589 do {
2590 next = hugepte_addr_end(addr, end, sz);
2591 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2592 return 0;
2593 } while (ptep++, addr = next, addr != end);
2594
2595 return 1;
2596}
2597#else
2598static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2599 unsigned int pdshift, unsigned long end, unsigned int flags,
2600 struct page **pages, int *nr)
2601{
2602 return 0;
2603}
2604#endif /* CONFIG_ARCH_HAS_HUGEPD */
2605
2606static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2607 unsigned long end, unsigned int flags,
2608 struct page **pages, int *nr)
2609{
2610 struct page *page;
2611 struct folio *folio;
2612 int refs;
2613
2614 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2615 return 0;
2616
2617 if (pmd_devmap(orig)) {
2618 if (unlikely(flags & FOLL_LONGTERM))
2619 return 0;
2620 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2621 pages, nr);
2622 }
2623
2624 page = nth_page(pmd_page(orig), (addr & ~PMD_MASK) >> PAGE_SHIFT);
2625 refs = record_subpages(page, addr, end, pages + *nr);
2626
2627 folio = try_grab_folio(page, refs, flags);
2628 if (!folio)
2629 return 0;
2630
2631 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2632 gup_put_folio(folio, refs, flags);
2633 return 0;
2634 }
2635
2636 if (!pmd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2637 gup_put_folio(folio, refs, flags);
2638 return 0;
2639 }
2640
2641 *nr += refs;
2642 folio_set_referenced(folio);
2643 return 1;
2644}
2645
2646static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2647 unsigned long end, unsigned int flags,
2648 struct page **pages, int *nr)
2649{
2650 struct page *page;
2651 struct folio *folio;
2652 int refs;
2653
2654 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2655 return 0;
2656
2657 if (pud_devmap(orig)) {
2658 if (unlikely(flags & FOLL_LONGTERM))
2659 return 0;
2660 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2661 pages, nr);
2662 }
2663
2664 page = nth_page(pud_page(orig), (addr & ~PUD_MASK) >> PAGE_SHIFT);
2665 refs = record_subpages(page, addr, end, pages + *nr);
2666
2667 folio = try_grab_folio(page, refs, flags);
2668 if (!folio)
2669 return 0;
2670
2671 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2672 gup_put_folio(folio, refs, flags);
2673 return 0;
2674 }
2675
2676 if (!pud_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2677 gup_put_folio(folio, refs, flags);
2678 return 0;
2679 }
2680
2681 *nr += refs;
2682 folio_set_referenced(folio);
2683 return 1;
2684}
2685
2686static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2687 unsigned long end, unsigned int flags,
2688 struct page **pages, int *nr)
2689{
2690 int refs;
2691 struct page *page;
2692 struct folio *folio;
2693
2694 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2695 return 0;
2696
2697 BUILD_BUG_ON(pgd_devmap(orig));
2698
2699 page = nth_page(pgd_page(orig), (addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2700 refs = record_subpages(page, addr, end, pages + *nr);
2701
2702 folio = try_grab_folio(page, refs, flags);
2703 if (!folio)
2704 return 0;
2705
2706 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2707 gup_put_folio(folio, refs, flags);
2708 return 0;
2709 }
2710
2711 *nr += refs;
2712 folio_set_referenced(folio);
2713 return 1;
2714}
2715
2716static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
2717 unsigned int flags, struct page **pages, int *nr)
2718{
2719 unsigned long next;
2720 pmd_t *pmdp;
2721
2722 pmdp = pmd_offset_lockless(pudp, pud, addr);
2723 do {
2724 pmd_t pmd = pmdp_get_lockless(pmdp);
2725
2726 next = pmd_addr_end(addr, end);
2727 if (!pmd_present(pmd))
2728 return 0;
2729
2730 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2731 pmd_devmap(pmd))) {
2732 if (pmd_protnone(pmd) &&
2733 !gup_can_follow_protnone(flags))
2734 return 0;
2735
2736 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2737 pages, nr))
2738 return 0;
2739
2740 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2741 /*
2742 * architecture have different format for hugetlbfs
2743 * pmd format and THP pmd format
2744 */
2745 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2746 PMD_SHIFT, next, flags, pages, nr))
2747 return 0;
2748 } else if (!gup_pte_range(pmd, pmdp, addr, next, flags, pages, nr))
2749 return 0;
2750 } while (pmdp++, addr = next, addr != end);
2751
2752 return 1;
2753}
2754
2755static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
2756 unsigned int flags, struct page **pages, int *nr)
2757{
2758 unsigned long next;
2759 pud_t *pudp;
2760
2761 pudp = pud_offset_lockless(p4dp, p4d, addr);
2762 do {
2763 pud_t pud = READ_ONCE(*pudp);
2764
2765 next = pud_addr_end(addr, end);
2766 if (unlikely(!pud_present(pud)))
2767 return 0;
2768 if (unlikely(pud_huge(pud) || pud_devmap(pud))) {
2769 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2770 pages, nr))
2771 return 0;
2772 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2773 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2774 PUD_SHIFT, next, flags, pages, nr))
2775 return 0;
2776 } else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
2777 return 0;
2778 } while (pudp++, addr = next, addr != end);
2779
2780 return 1;
2781}
2782
2783static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
2784 unsigned int flags, struct page **pages, int *nr)
2785{
2786 unsigned long next;
2787 p4d_t *p4dp;
2788
2789 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
2790 do {
2791 p4d_t p4d = READ_ONCE(*p4dp);
2792
2793 next = p4d_addr_end(addr, end);
2794 if (p4d_none(p4d))
2795 return 0;
2796 BUILD_BUG_ON(p4d_huge(p4d));
2797 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2798 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2799 P4D_SHIFT, next, flags, pages, nr))
2800 return 0;
2801 } else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
2802 return 0;
2803 } while (p4dp++, addr = next, addr != end);
2804
2805 return 1;
2806}
2807
2808static void gup_pgd_range(unsigned long addr, unsigned long end,
2809 unsigned int flags, struct page **pages, int *nr)
2810{
2811 unsigned long next;
2812 pgd_t *pgdp;
2813
2814 pgdp = pgd_offset(current->mm, addr);
2815 do {
2816 pgd_t pgd = READ_ONCE(*pgdp);
2817
2818 next = pgd_addr_end(addr, end);
2819 if (pgd_none(pgd))
2820 return;
2821 if (unlikely(pgd_huge(pgd))) {
2822 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2823 pages, nr))
2824 return;
2825 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2826 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2827 PGDIR_SHIFT, next, flags, pages, nr))
2828 return;
2829 } else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
2830 return;
2831 } while (pgdp++, addr = next, addr != end);
2832}
2833#else
2834static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2835 unsigned int flags, struct page **pages, int *nr)
2836{
2837}
2838#endif /* CONFIG_HAVE_FAST_GUP */
2839
2840#ifndef gup_fast_permitted
2841/*
2842 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2843 * we need to fall back to the slow version:
2844 */
2845static bool gup_fast_permitted(unsigned long start, unsigned long end)
2846{
2847 return true;
2848}
2849#endif
2850
2851static unsigned long lockless_pages_from_mm(unsigned long start,
2852 unsigned long end,
2853 unsigned int gup_flags,
2854 struct page **pages)
2855{
2856 unsigned long flags;
2857 int nr_pinned = 0;
2858 unsigned seq;
2859
2860 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP) ||
2861 !gup_fast_permitted(start, end))
2862 return 0;
2863
2864 if (gup_flags & FOLL_PIN) {
2865 seq = raw_read_seqcount(¤t->mm->write_protect_seq);
2866 if (seq & 1)
2867 return 0;
2868 }
2869
2870 /*
2871 * Disable interrupts. The nested form is used, in order to allow full,
2872 * general purpose use of this routine.
2873 *
2874 * With interrupts disabled, we block page table pages from being freed
2875 * from under us. See struct mmu_table_batch comments in
2876 * include/asm-generic/tlb.h for more details.
2877 *
2878 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
2879 * that come from THPs splitting.
2880 */
2881 local_irq_save(flags);
2882 gup_pgd_range(start, end, gup_flags, pages, &nr_pinned);
2883 local_irq_restore(flags);
2884
2885 /*
2886 * When pinning pages for DMA there could be a concurrent write protect
2887 * from fork() via copy_page_range(), in this case always fail fast GUP.
2888 */
2889 if (gup_flags & FOLL_PIN) {
2890 if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) {
2891 unpin_user_pages_lockless(pages, nr_pinned);
2892 return 0;
2893 } else {
2894 sanity_check_pinned_pages(pages, nr_pinned);
2895 }
2896 }
2897 return nr_pinned;
2898}
2899
2900static int internal_get_user_pages_fast(unsigned long start,
2901 unsigned long nr_pages,
2902 unsigned int gup_flags,
2903 struct page **pages)
2904{
2905 unsigned long len, end;
2906 unsigned long nr_pinned;
2907 int ret;
2908
2909 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
2910 FOLL_FORCE | FOLL_PIN | FOLL_GET |
2911 FOLL_FAST_ONLY | FOLL_NOFAULT |
2912 FOLL_PCI_P2PDMA)))
2913 return -EINVAL;
2914
2915 if (gup_flags & FOLL_PIN)
2916 mm_set_has_pinned_flag(¤t->mm->flags);
2917
2918 if (!(gup_flags & FOLL_FAST_ONLY))
2919 might_lock_read(¤t->mm->mmap_lock);
2920
2921 start = untagged_addr(start) & PAGE_MASK;
2922 len = nr_pages << PAGE_SHIFT;
2923 if (check_add_overflow(start, len, &end))
2924 return 0;
2925 if (unlikely(!access_ok((void __user *)start, len)))
2926 return -EFAULT;
2927
2928 nr_pinned = lockless_pages_from_mm(start, end, gup_flags, pages);
2929 if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
2930 return nr_pinned;
2931
2932 /* Slow path: try to get the remaining pages with get_user_pages */
2933 start += nr_pinned << PAGE_SHIFT;
2934 pages += nr_pinned;
2935 ret = get_user_pages_unlocked(start, nr_pages - nr_pinned, pages,
2936 gup_flags);
2937 if (ret < 0) {
2938 /*
2939 * The caller has to unpin the pages we already pinned so
2940 * returning -errno is not an option
2941 */
2942 if (nr_pinned)
2943 return nr_pinned;
2944 return ret;
2945 }
2946 return ret + nr_pinned;
2947}
2948
2949/**
2950 * get_user_pages_fast_only() - pin user pages in memory
2951 * @start: starting user address
2952 * @nr_pages: number of pages from start to pin
2953 * @gup_flags: flags modifying pin behaviour
2954 * @pages: array that receives pointers to the pages pinned.
2955 * Should be at least nr_pages long.
2956 *
2957 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2958 * the regular GUP.
2959 * Note a difference with get_user_pages_fast: this always returns the
2960 * number of pages pinned, 0 if no pages were pinned.
2961 *
2962 * If the architecture does not support this function, simply return with no
2963 * pages pinned.
2964 *
2965 * Careful, careful! COW breaking can go either way, so a non-write
2966 * access can get ambiguous page results. If you call this function without
2967 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2968 */
2969int get_user_pages_fast_only(unsigned long start, int nr_pages,
2970 unsigned int gup_flags, struct page **pages)
2971{
2972 int nr_pinned;
2973 /*
2974 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2975 * because gup fast is always a "pin with a +1 page refcount" request.
2976 *
2977 * FOLL_FAST_ONLY is required in order to match the API description of
2978 * this routine: no fall back to regular ("slow") GUP.
2979 */
2980 gup_flags |= FOLL_GET | FOLL_FAST_ONLY;
2981
2982 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2983 pages);
2984
2985 /*
2986 * As specified in the API description above, this routine is not
2987 * allowed to return negative values. However, the common core
2988 * routine internal_get_user_pages_fast() *can* return -errno.
2989 * Therefore, correct for that here:
2990 */
2991 if (nr_pinned < 0)
2992 nr_pinned = 0;
2993
2994 return nr_pinned;
2995}
2996EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
2997
2998/**
2999 * get_user_pages_fast() - pin user pages in memory
3000 * @start: starting user address
3001 * @nr_pages: number of pages from start to pin
3002 * @gup_flags: flags modifying pin behaviour
3003 * @pages: array that receives pointers to the pages pinned.
3004 * Should be at least nr_pages long.
3005 *
3006 * Attempt to pin user pages in memory without taking mm->mmap_lock.
3007 * If not successful, it will fall back to taking the lock and
3008 * calling get_user_pages().
3009 *
3010 * Returns number of pages pinned. This may be fewer than the number requested.
3011 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
3012 * -errno.
3013 */
3014int get_user_pages_fast(unsigned long start, int nr_pages,
3015 unsigned int gup_flags, struct page **pages)
3016{
3017 if (!is_valid_gup_flags(gup_flags))
3018 return -EINVAL;
3019
3020 /*
3021 * The caller may or may not have explicitly set FOLL_GET; either way is
3022 * OK. However, internally (within mm/gup.c), gup fast variants must set
3023 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
3024 * request.
3025 */
3026 gup_flags |= FOLL_GET;
3027 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3028}
3029EXPORT_SYMBOL_GPL(get_user_pages_fast);
3030
3031/**
3032 * pin_user_pages_fast() - pin user pages in memory without taking locks
3033 *
3034 * @start: starting user address
3035 * @nr_pages: number of pages from start to pin
3036 * @gup_flags: flags modifying pin behaviour
3037 * @pages: array that receives pointers to the pages pinned.
3038 * Should be at least nr_pages long.
3039 *
3040 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
3041 * get_user_pages_fast() for documentation on the function arguments, because
3042 * the arguments here are identical.
3043 *
3044 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3045 * see Documentation/core-api/pin_user_pages.rst for further details.
3046 */
3047int pin_user_pages_fast(unsigned long start, int nr_pages,
3048 unsigned int gup_flags, struct page **pages)
3049{
3050 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3051 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
3052 return -EINVAL;
3053
3054 if (WARN_ON_ONCE(!pages))
3055 return -EINVAL;
3056
3057 gup_flags |= FOLL_PIN;
3058 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3059}
3060EXPORT_SYMBOL_GPL(pin_user_pages_fast);
3061
3062/*
3063 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
3064 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
3065 *
3066 * The API rules are the same, too: no negative values may be returned.
3067 */
3068int pin_user_pages_fast_only(unsigned long start, int nr_pages,
3069 unsigned int gup_flags, struct page **pages)
3070{
3071 int nr_pinned;
3072
3073 /*
3074 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
3075 * rules require returning 0, rather than -errno:
3076 */
3077 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
3078 return 0;
3079
3080 if (WARN_ON_ONCE(!pages))
3081 return 0;
3082 /*
3083 * FOLL_FAST_ONLY is required in order to match the API description of
3084 * this routine: no fall back to regular ("slow") GUP.
3085 */
3086 gup_flags |= (FOLL_PIN | FOLL_FAST_ONLY);
3087 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
3088 pages);
3089 /*
3090 * This routine is not allowed to return negative values. However,
3091 * internal_get_user_pages_fast() *can* return -errno. Therefore,
3092 * correct for that here:
3093 */
3094 if (nr_pinned < 0)
3095 nr_pinned = 0;
3096
3097 return nr_pinned;
3098}
3099EXPORT_SYMBOL_GPL(pin_user_pages_fast_only);
3100
3101/**
3102 * pin_user_pages_remote() - pin pages of a remote process
3103 *
3104 * @mm: mm_struct of target mm
3105 * @start: starting user address
3106 * @nr_pages: number of pages from start to pin
3107 * @gup_flags: flags modifying lookup behaviour
3108 * @pages: array that receives pointers to the pages pinned.
3109 * Should be at least nr_pages long.
3110 * @vmas: array of pointers to vmas corresponding to each page.
3111 * Or NULL if the caller does not require them.
3112 * @locked: pointer to lock flag indicating whether lock is held and
3113 * subsequently whether VM_FAULT_RETRY functionality can be
3114 * utilised. Lock must initially be held.
3115 *
3116 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
3117 * get_user_pages_remote() for documentation on the function arguments, because
3118 * the arguments here are identical.
3119 *
3120 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3121 * see Documentation/core-api/pin_user_pages.rst for details.
3122 */
3123long pin_user_pages_remote(struct mm_struct *mm,
3124 unsigned long start, unsigned long nr_pages,
3125 unsigned int gup_flags, struct page **pages,
3126 struct vm_area_struct **vmas, int *locked)
3127{
3128 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3129 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
3130 return -EINVAL;
3131
3132 if (WARN_ON_ONCE(!pages))
3133 return -EINVAL;
3134
3135 return __gup_longterm_locked(mm, start, nr_pages, pages, vmas, locked,
3136 gup_flags | FOLL_PIN | FOLL_TOUCH |
3137 FOLL_REMOTE);
3138}
3139EXPORT_SYMBOL(pin_user_pages_remote);
3140
3141/**
3142 * pin_user_pages() - pin user pages in memory for use by other devices
3143 *
3144 * @start: starting user address
3145 * @nr_pages: number of pages from start to pin
3146 * @gup_flags: flags modifying lookup behaviour
3147 * @pages: array that receives pointers to the pages pinned.
3148 * Should be at least nr_pages long.
3149 * @vmas: array of pointers to vmas corresponding to each page.
3150 * Or NULL if the caller does not require them.
3151 *
3152 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
3153 * FOLL_PIN is set.
3154 *
3155 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3156 * see Documentation/core-api/pin_user_pages.rst for details.
3157 */
3158long pin_user_pages(unsigned long start, unsigned long nr_pages,
3159 unsigned int gup_flags, struct page **pages,
3160 struct vm_area_struct **vmas)
3161{
3162 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3163 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
3164 return -EINVAL;
3165
3166 if (WARN_ON_ONCE(!pages))
3167 return -EINVAL;
3168
3169 gup_flags |= FOLL_PIN;
3170 return __gup_longterm_locked(current->mm, start, nr_pages,
3171 pages, vmas, NULL, gup_flags);
3172}
3173EXPORT_SYMBOL(pin_user_pages);
3174
3175/*
3176 * pin_user_pages_unlocked() is the FOLL_PIN variant of
3177 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
3178 * FOLL_PIN and rejects FOLL_GET.
3179 */
3180long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
3181 struct page **pages, unsigned int gup_flags)
3182{
3183 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3184 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
3185 return -EINVAL;
3186
3187 if (WARN_ON_ONCE(!pages))
3188 return -EINVAL;
3189
3190 gup_flags |= FOLL_PIN;
3191 return get_user_pages_unlocked(start, nr_pages, pages, gup_flags);
3192}
3193EXPORT_SYMBOL(pin_user_pages_unlocked);