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/memfd.h>
9#include <linux/memremap.h>
10#include <linux/pagemap.h>
11#include <linux/rmap.h>
12#include <linux/swap.h>
13#include <linux/swapops.h>
14#include <linux/secretmem.h>
15
16#include <linux/sched/signal.h>
17#include <linux/rwsem.h>
18#include <linux/hugetlb.h>
19#include <linux/migrate.h>
20#include <linux/mm_inline.h>
21#include <linux/pagevec.h>
22#include <linux/sched/mm.h>
23#include <linux/shmem_fs.h>
24
25#include <asm/mmu_context.h>
26#include <asm/tlbflush.h>
27
28#include "internal.h"
29
30struct follow_page_context {
31 struct dev_pagemap *pgmap;
32 unsigned int page_mask;
33};
34
35static inline void sanity_check_pinned_pages(struct page **pages,
36 unsigned long npages)
37{
38 if (!IS_ENABLED(CONFIG_DEBUG_VM))
39 return;
40
41 /*
42 * We only pin anonymous pages if they are exclusive. Once pinned, we
43 * can no longer turn them possibly shared and PageAnonExclusive() will
44 * stick around until the page is freed.
45 *
46 * We'd like to verify that our pinned anonymous pages are still mapped
47 * exclusively. The issue with anon THP is that we don't know how
48 * they are/were mapped when pinning them. However, for anon
49 * THP we can assume that either the given page (PTE-mapped THP) or
50 * the head page (PMD-mapped THP) should be PageAnonExclusive(). If
51 * neither is the case, there is certainly something wrong.
52 */
53 for (; npages; npages--, pages++) {
54 struct page *page = *pages;
55 struct folio *folio;
56
57 if (!page)
58 continue;
59
60 folio = page_folio(page);
61
62 if (is_zero_page(page) ||
63 !folio_test_anon(folio))
64 continue;
65 if (!folio_test_large(folio) || folio_test_hugetlb(folio))
66 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page), page);
67 else
68 /* Either a PTE-mapped or a PMD-mapped THP. */
69 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page) &&
70 !PageAnonExclusive(page), page);
71 }
72}
73
74/*
75 * Return the folio with ref appropriately incremented,
76 * or NULL if that failed.
77 */
78static inline struct folio *try_get_folio(struct page *page, int refs)
79{
80 struct folio *folio;
81
82retry:
83 folio = page_folio(page);
84 if (WARN_ON_ONCE(folio_ref_count(folio) < 0))
85 return NULL;
86 if (unlikely(!folio_ref_try_add(folio, refs)))
87 return NULL;
88
89 /*
90 * At this point we have a stable reference to the folio; but it
91 * could be that between calling page_folio() and the refcount
92 * increment, the folio was split, in which case we'd end up
93 * holding a reference on a folio that has nothing to do with the page
94 * we were given anymore.
95 * So now that the folio is stable, recheck that the page still
96 * belongs to this folio.
97 */
98 if (unlikely(page_folio(page) != folio)) {
99 if (!put_devmap_managed_folio_refs(folio, refs))
100 folio_put_refs(folio, refs);
101 goto retry;
102 }
103
104 return folio;
105}
106
107static void gup_put_folio(struct folio *folio, int refs, unsigned int flags)
108{
109 if (flags & FOLL_PIN) {
110 if (is_zero_folio(folio))
111 return;
112 node_stat_mod_folio(folio, NR_FOLL_PIN_RELEASED, refs);
113 if (folio_test_large(folio))
114 atomic_sub(refs, &folio->_pincount);
115 else
116 refs *= GUP_PIN_COUNTING_BIAS;
117 }
118
119 if (!put_devmap_managed_folio_refs(folio, refs))
120 folio_put_refs(folio, refs);
121}
122
123/**
124 * try_grab_folio() - add a folio's refcount by a flag-dependent amount
125 * @folio: pointer to folio to be grabbed
126 * @refs: the value to (effectively) add to the folio's refcount
127 * @flags: gup flags: these are the FOLL_* flag values
128 *
129 * This might not do anything at all, depending on the flags argument.
130 *
131 * "grab" names in this file mean, "look at flags to decide whether to use
132 * FOLL_PIN or FOLL_GET behavior, when incrementing the folio's refcount.
133 *
134 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
135 * time.
136 *
137 * Return: 0 for success, or if no action was required (if neither FOLL_PIN
138 * nor FOLL_GET was set, nothing is done). A negative error code for failure:
139 *
140 * -ENOMEM FOLL_GET or FOLL_PIN was set, but the folio could not
141 * be grabbed.
142 *
143 * It is called when we have a stable reference for the folio, typically in
144 * GUP slow path.
145 */
146int __must_check try_grab_folio(struct folio *folio, int refs,
147 unsigned int flags)
148{
149 if (WARN_ON_ONCE(folio_ref_count(folio) <= 0))
150 return -ENOMEM;
151
152 if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(&folio->page)))
153 return -EREMOTEIO;
154
155 if (flags & FOLL_GET)
156 folio_ref_add(folio, refs);
157 else if (flags & FOLL_PIN) {
158 /*
159 * Don't take a pin on the zero page - it's not going anywhere
160 * and it is used in a *lot* of places.
161 */
162 if (is_zero_folio(folio))
163 return 0;
164
165 /*
166 * Increment the normal page refcount field at least once,
167 * so that the page really is pinned.
168 */
169 if (folio_test_large(folio)) {
170 folio_ref_add(folio, refs);
171 atomic_add(refs, &folio->_pincount);
172 } else {
173 folio_ref_add(folio, refs * GUP_PIN_COUNTING_BIAS);
174 }
175
176 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs);
177 }
178
179 return 0;
180}
181
182/**
183 * unpin_user_page() - release a dma-pinned page
184 * @page: pointer to page to be released
185 *
186 * Pages that were pinned via pin_user_pages*() must be released via either
187 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
188 * that such pages can be separately tracked and uniquely handled. In
189 * particular, interactions with RDMA and filesystems need special handling.
190 */
191void unpin_user_page(struct page *page)
192{
193 sanity_check_pinned_pages(&page, 1);
194 gup_put_folio(page_folio(page), 1, FOLL_PIN);
195}
196EXPORT_SYMBOL(unpin_user_page);
197
198/**
199 * unpin_folio() - release a dma-pinned folio
200 * @folio: pointer to folio to be released
201 *
202 * Folios that were pinned via memfd_pin_folios() or other similar routines
203 * must be released either using unpin_folio() or unpin_folios().
204 */
205void unpin_folio(struct folio *folio)
206{
207 gup_put_folio(folio, 1, FOLL_PIN);
208}
209EXPORT_SYMBOL_GPL(unpin_folio);
210
211/**
212 * folio_add_pin - Try to get an additional pin on a pinned folio
213 * @folio: The folio to be pinned
214 *
215 * Get an additional pin on a folio we already have a pin on. Makes no change
216 * if the folio is a zero_page.
217 */
218void folio_add_pin(struct folio *folio)
219{
220 if (is_zero_folio(folio))
221 return;
222
223 /*
224 * Similar to try_grab_folio(): be sure to *also* increment the normal
225 * page refcount field at least once, so that the page really is
226 * pinned.
227 */
228 if (folio_test_large(folio)) {
229 WARN_ON_ONCE(atomic_read(&folio->_pincount) < 1);
230 folio_ref_inc(folio);
231 atomic_inc(&folio->_pincount);
232 } else {
233 WARN_ON_ONCE(folio_ref_count(folio) < GUP_PIN_COUNTING_BIAS);
234 folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
235 }
236}
237
238static inline struct folio *gup_folio_range_next(struct page *start,
239 unsigned long npages, unsigned long i, unsigned int *ntails)
240{
241 struct page *next = nth_page(start, i);
242 struct folio *folio = page_folio(next);
243 unsigned int nr = 1;
244
245 if (folio_test_large(folio))
246 nr = min_t(unsigned int, npages - i,
247 folio_nr_pages(folio) - folio_page_idx(folio, next));
248
249 *ntails = nr;
250 return folio;
251}
252
253static inline struct folio *gup_folio_next(struct page **list,
254 unsigned long npages, unsigned long i, unsigned int *ntails)
255{
256 struct folio *folio = page_folio(list[i]);
257 unsigned int nr;
258
259 for (nr = i + 1; nr < npages; nr++) {
260 if (page_folio(list[nr]) != folio)
261 break;
262 }
263
264 *ntails = nr - i;
265 return folio;
266}
267
268/**
269 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
270 * @pages: array of pages to be maybe marked dirty, and definitely released.
271 * @npages: number of pages in the @pages array.
272 * @make_dirty: whether to mark the pages dirty
273 *
274 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
275 * variants called on that page.
276 *
277 * For each page in the @pages array, make that page (or its head page, if a
278 * compound page) dirty, if @make_dirty is true, and if the page was previously
279 * listed as clean. In any case, releases all pages using unpin_user_page(),
280 * possibly via unpin_user_pages(), for the non-dirty case.
281 *
282 * Please see the unpin_user_page() documentation for details.
283 *
284 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
285 * required, then the caller should a) verify that this is really correct,
286 * because _lock() is usually required, and b) hand code it:
287 * set_page_dirty_lock(), unpin_user_page().
288 *
289 */
290void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
291 bool make_dirty)
292{
293 unsigned long i;
294 struct folio *folio;
295 unsigned int nr;
296
297 if (!make_dirty) {
298 unpin_user_pages(pages, npages);
299 return;
300 }
301
302 sanity_check_pinned_pages(pages, npages);
303 for (i = 0; i < npages; i += nr) {
304 folio = gup_folio_next(pages, npages, i, &nr);
305 /*
306 * Checking PageDirty at this point may race with
307 * clear_page_dirty_for_io(), but that's OK. Two key
308 * cases:
309 *
310 * 1) This code sees the page as already dirty, so it
311 * skips the call to set_page_dirty(). That could happen
312 * because clear_page_dirty_for_io() called
313 * folio_mkclean(), followed by set_page_dirty().
314 * However, now the page is going to get written back,
315 * which meets the original intention of setting it
316 * dirty, so all is well: clear_page_dirty_for_io() goes
317 * on to call TestClearPageDirty(), and write the page
318 * back.
319 *
320 * 2) This code sees the page as clean, so it calls
321 * set_page_dirty(). The page stays dirty, despite being
322 * written back, so it gets written back again in the
323 * next writeback cycle. This is harmless.
324 */
325 if (!folio_test_dirty(folio)) {
326 folio_lock(folio);
327 folio_mark_dirty(folio);
328 folio_unlock(folio);
329 }
330 gup_put_folio(folio, nr, FOLL_PIN);
331 }
332}
333EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
334
335/**
336 * unpin_user_page_range_dirty_lock() - release and optionally dirty
337 * gup-pinned page range
338 *
339 * @page: the starting page of a range maybe marked dirty, and definitely released.
340 * @npages: number of consecutive pages to release.
341 * @make_dirty: whether to mark the pages dirty
342 *
343 * "gup-pinned page range" refers to a range of pages that has had one of the
344 * pin_user_pages() variants called on that page.
345 *
346 * For the page ranges defined by [page .. page+npages], make that range (or
347 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
348 * page range was previously listed as clean.
349 *
350 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
351 * required, then the caller should a) verify that this is really correct,
352 * because _lock() is usually required, and b) hand code it:
353 * set_page_dirty_lock(), unpin_user_page().
354 *
355 */
356void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages,
357 bool make_dirty)
358{
359 unsigned long i;
360 struct folio *folio;
361 unsigned int nr;
362
363 for (i = 0; i < npages; i += nr) {
364 folio = gup_folio_range_next(page, npages, i, &nr);
365 if (make_dirty && !folio_test_dirty(folio)) {
366 folio_lock(folio);
367 folio_mark_dirty(folio);
368 folio_unlock(folio);
369 }
370 gup_put_folio(folio, nr, FOLL_PIN);
371 }
372}
373EXPORT_SYMBOL(unpin_user_page_range_dirty_lock);
374
375static void gup_fast_unpin_user_pages(struct page **pages, unsigned long npages)
376{
377 unsigned long i;
378 struct folio *folio;
379 unsigned int nr;
380
381 /*
382 * Don't perform any sanity checks because we might have raced with
383 * fork() and some anonymous pages might now actually be shared --
384 * which is why we're unpinning after all.
385 */
386 for (i = 0; i < npages; i += nr) {
387 folio = gup_folio_next(pages, npages, i, &nr);
388 gup_put_folio(folio, nr, FOLL_PIN);
389 }
390}
391
392/**
393 * unpin_user_pages() - release an array of gup-pinned pages.
394 * @pages: array of pages to be marked dirty and released.
395 * @npages: number of pages in the @pages array.
396 *
397 * For each page in the @pages array, release the page using unpin_user_page().
398 *
399 * Please see the unpin_user_page() documentation for details.
400 */
401void unpin_user_pages(struct page **pages, unsigned long npages)
402{
403 unsigned long i;
404 struct folio *folio;
405 unsigned int nr;
406
407 /*
408 * If this WARN_ON() fires, then the system *might* be leaking pages (by
409 * leaving them pinned), but probably not. More likely, gup/pup returned
410 * a hard -ERRNO error to the caller, who erroneously passed it here.
411 */
412 if (WARN_ON(IS_ERR_VALUE(npages)))
413 return;
414
415 sanity_check_pinned_pages(pages, npages);
416 for (i = 0; i < npages; i += nr) {
417 if (!pages[i]) {
418 nr = 1;
419 continue;
420 }
421 folio = gup_folio_next(pages, npages, i, &nr);
422 gup_put_folio(folio, nr, FOLL_PIN);
423 }
424}
425EXPORT_SYMBOL(unpin_user_pages);
426
427/**
428 * unpin_user_folio() - release pages of a folio
429 * @folio: pointer to folio to be released
430 * @npages: number of pages of same folio
431 *
432 * Release npages of the folio
433 */
434void unpin_user_folio(struct folio *folio, unsigned long npages)
435{
436 gup_put_folio(folio, npages, FOLL_PIN);
437}
438EXPORT_SYMBOL(unpin_user_folio);
439
440/**
441 * unpin_folios() - release an array of gup-pinned folios.
442 * @folios: array of folios to be marked dirty and released.
443 * @nfolios: number of folios in the @folios array.
444 *
445 * For each folio in the @folios array, release the folio using gup_put_folio.
446 *
447 * Please see the unpin_folio() documentation for details.
448 */
449void unpin_folios(struct folio **folios, unsigned long nfolios)
450{
451 unsigned long i = 0, j;
452
453 /*
454 * If this WARN_ON() fires, then the system *might* be leaking folios
455 * (by leaving them pinned), but probably not. More likely, gup/pup
456 * returned a hard -ERRNO error to the caller, who erroneously passed
457 * it here.
458 */
459 if (WARN_ON(IS_ERR_VALUE(nfolios)))
460 return;
461
462 while (i < nfolios) {
463 for (j = i + 1; j < nfolios; j++)
464 if (folios[i] != folios[j])
465 break;
466
467 if (folios[i])
468 gup_put_folio(folios[i], j - i, FOLL_PIN);
469 i = j;
470 }
471}
472EXPORT_SYMBOL_GPL(unpin_folios);
473
474/*
475 * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's
476 * lifecycle. Avoid setting the bit unless necessary, or it might cause write
477 * cache bouncing on large SMP machines for concurrent pinned gups.
478 */
479static inline void mm_set_has_pinned_flag(unsigned long *mm_flags)
480{
481 if (!test_bit(MMF_HAS_PINNED, mm_flags))
482 set_bit(MMF_HAS_PINNED, mm_flags);
483}
484
485#ifdef CONFIG_MMU
486
487#ifdef CONFIG_HAVE_GUP_FAST
488static int record_subpages(struct page *page, unsigned long sz,
489 unsigned long addr, unsigned long end,
490 struct page **pages)
491{
492 struct page *start_page;
493 int nr;
494
495 start_page = nth_page(page, (addr & (sz - 1)) >> PAGE_SHIFT);
496 for (nr = 0; addr != end; nr++, addr += PAGE_SIZE)
497 pages[nr] = nth_page(start_page, nr);
498
499 return nr;
500}
501
502/**
503 * try_grab_folio_fast() - Attempt to get or pin a folio in fast path.
504 * @page: pointer to page to be grabbed
505 * @refs: the value to (effectively) add to the folio's refcount
506 * @flags: gup flags: these are the FOLL_* flag values.
507 *
508 * "grab" names in this file mean, "look at flags to decide whether to use
509 * FOLL_PIN or FOLL_GET behavior, when incrementing the folio's refcount.
510 *
511 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
512 * same time. (That's true throughout the get_user_pages*() and
513 * pin_user_pages*() APIs.) Cases:
514 *
515 * FOLL_GET: folio's refcount will be incremented by @refs.
516 *
517 * FOLL_PIN on large folios: folio's refcount will be incremented by
518 * @refs, and its pincount will be incremented by @refs.
519 *
520 * FOLL_PIN on single-page folios: folio's refcount will be incremented by
521 * @refs * GUP_PIN_COUNTING_BIAS.
522 *
523 * Return: The folio containing @page (with refcount appropriately
524 * incremented) for success, or NULL upon failure. If neither FOLL_GET
525 * nor FOLL_PIN was set, that's considered failure, and furthermore,
526 * a likely bug in the caller, so a warning is also emitted.
527 *
528 * It uses add ref unless zero to elevate the folio refcount and must be called
529 * in fast path only.
530 */
531static struct folio *try_grab_folio_fast(struct page *page, int refs,
532 unsigned int flags)
533{
534 struct folio *folio;
535
536 /* Raise warn if it is not called in fast GUP */
537 VM_WARN_ON_ONCE(!irqs_disabled());
538
539 if (WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == 0))
540 return NULL;
541
542 if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
543 return NULL;
544
545 if (flags & FOLL_GET)
546 return try_get_folio(page, refs);
547
548 /* FOLL_PIN is set */
549
550 /*
551 * Don't take a pin on the zero page - it's not going anywhere
552 * and it is used in a *lot* of places.
553 */
554 if (is_zero_page(page))
555 return page_folio(page);
556
557 folio = try_get_folio(page, refs);
558 if (!folio)
559 return NULL;
560
561 /*
562 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
563 * right zone, so fail and let the caller fall back to the slow
564 * path.
565 */
566 if (unlikely((flags & FOLL_LONGTERM) &&
567 !folio_is_longterm_pinnable(folio))) {
568 if (!put_devmap_managed_folio_refs(folio, refs))
569 folio_put_refs(folio, refs);
570 return NULL;
571 }
572
573 /*
574 * When pinning a large folio, use an exact count to track it.
575 *
576 * However, be sure to *also* increment the normal folio
577 * refcount field at least once, so that the folio really
578 * is pinned. That's why the refcount from the earlier
579 * try_get_folio() is left intact.
580 */
581 if (folio_test_large(folio))
582 atomic_add(refs, &folio->_pincount);
583 else
584 folio_ref_add(folio,
585 refs * (GUP_PIN_COUNTING_BIAS - 1));
586 /*
587 * Adjust the pincount before re-checking the PTE for changes.
588 * This is essentially a smp_mb() and is paired with a memory
589 * barrier in folio_try_share_anon_rmap_*().
590 */
591 smp_mb__after_atomic();
592
593 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs);
594
595 return folio;
596}
597#endif /* CONFIG_HAVE_GUP_FAST */
598
599static struct page *no_page_table(struct vm_area_struct *vma,
600 unsigned int flags, unsigned long address)
601{
602 if (!(flags & FOLL_DUMP))
603 return NULL;
604
605 /*
606 * When core dumping, we don't want to allocate unnecessary pages or
607 * page tables. Return error instead of NULL to skip handle_mm_fault,
608 * then get_dump_page() will return NULL to leave a hole in the dump.
609 * But we can only make this optimization where a hole would surely
610 * be zero-filled if handle_mm_fault() actually did handle it.
611 */
612 if (is_vm_hugetlb_page(vma)) {
613 struct hstate *h = hstate_vma(vma);
614
615 if (!hugetlbfs_pagecache_present(h, vma, address))
616 return ERR_PTR(-EFAULT);
617 } else if ((vma_is_anonymous(vma) || !vma->vm_ops->fault)) {
618 return ERR_PTR(-EFAULT);
619 }
620
621 return NULL;
622}
623
624#ifdef CONFIG_PGTABLE_HAS_HUGE_LEAVES
625static struct page *follow_huge_pud(struct vm_area_struct *vma,
626 unsigned long addr, pud_t *pudp,
627 int flags, struct follow_page_context *ctx)
628{
629 struct mm_struct *mm = vma->vm_mm;
630 struct page *page;
631 pud_t pud = *pudp;
632 unsigned long pfn = pud_pfn(pud);
633 int ret;
634
635 assert_spin_locked(pud_lockptr(mm, pudp));
636
637 if ((flags & FOLL_WRITE) && !pud_write(pud))
638 return NULL;
639
640 if (!pud_present(pud))
641 return NULL;
642
643 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
644
645 if (IS_ENABLED(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) &&
646 pud_devmap(pud)) {
647 /*
648 * device mapped pages can only be returned if the caller
649 * will manage the page reference count.
650 *
651 * At least one of FOLL_GET | FOLL_PIN must be set, so
652 * assert that here:
653 */
654 if (!(flags & (FOLL_GET | FOLL_PIN)))
655 return ERR_PTR(-EEXIST);
656
657 if (flags & FOLL_TOUCH)
658 touch_pud(vma, addr, pudp, flags & FOLL_WRITE);
659
660 ctx->pgmap = get_dev_pagemap(pfn, ctx->pgmap);
661 if (!ctx->pgmap)
662 return ERR_PTR(-EFAULT);
663 }
664
665 page = pfn_to_page(pfn);
666
667 if (!pud_devmap(pud) && !pud_write(pud) &&
668 gup_must_unshare(vma, flags, page))
669 return ERR_PTR(-EMLINK);
670
671 ret = try_grab_folio(page_folio(page), 1, flags);
672 if (ret)
673 page = ERR_PTR(ret);
674 else
675 ctx->page_mask = HPAGE_PUD_NR - 1;
676
677 return page;
678}
679
680/* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
681static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page,
682 struct vm_area_struct *vma,
683 unsigned int flags)
684{
685 /* If the pmd is writable, we can write to the page. */
686 if (pmd_write(pmd))
687 return true;
688
689 /* Maybe FOLL_FORCE is set to override it? */
690 if (!(flags & FOLL_FORCE))
691 return false;
692
693 /* But FOLL_FORCE has no effect on shared mappings */
694 if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
695 return false;
696
697 /* ... or read-only private ones */
698 if (!(vma->vm_flags & VM_MAYWRITE))
699 return false;
700
701 /* ... or already writable ones that just need to take a write fault */
702 if (vma->vm_flags & VM_WRITE)
703 return false;
704
705 /*
706 * See can_change_pte_writable(): we broke COW and could map the page
707 * writable if we have an exclusive anonymous page ...
708 */
709 if (!page || !PageAnon(page) || !PageAnonExclusive(page))
710 return false;
711
712 /* ... and a write-fault isn't required for other reasons. */
713 if (pmd_needs_soft_dirty_wp(vma, pmd))
714 return false;
715 return !userfaultfd_huge_pmd_wp(vma, pmd);
716}
717
718static struct page *follow_huge_pmd(struct vm_area_struct *vma,
719 unsigned long addr, pmd_t *pmd,
720 unsigned int flags,
721 struct follow_page_context *ctx)
722{
723 struct mm_struct *mm = vma->vm_mm;
724 pmd_t pmdval = *pmd;
725 struct page *page;
726 int ret;
727
728 assert_spin_locked(pmd_lockptr(mm, pmd));
729
730 page = pmd_page(pmdval);
731 if ((flags & FOLL_WRITE) &&
732 !can_follow_write_pmd(pmdval, page, vma, flags))
733 return NULL;
734
735 /* Avoid dumping huge zero page */
736 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(pmdval))
737 return ERR_PTR(-EFAULT);
738
739 if (pmd_protnone(*pmd) && !gup_can_follow_protnone(vma, flags))
740 return NULL;
741
742 if (!pmd_write(pmdval) && gup_must_unshare(vma, flags, page))
743 return ERR_PTR(-EMLINK);
744
745 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
746 !PageAnonExclusive(page), page);
747
748 ret = try_grab_folio(page_folio(page), 1, flags);
749 if (ret)
750 return ERR_PTR(ret);
751
752#ifdef CONFIG_TRANSPARENT_HUGEPAGE
753 if (pmd_trans_huge(pmdval) && (flags & FOLL_TOUCH))
754 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
755#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
756
757 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
758 ctx->page_mask = HPAGE_PMD_NR - 1;
759
760 return page;
761}
762
763#else /* CONFIG_PGTABLE_HAS_HUGE_LEAVES */
764static struct page *follow_huge_pud(struct vm_area_struct *vma,
765 unsigned long addr, pud_t *pudp,
766 int flags, struct follow_page_context *ctx)
767{
768 return NULL;
769}
770
771static struct page *follow_huge_pmd(struct vm_area_struct *vma,
772 unsigned long addr, pmd_t *pmd,
773 unsigned int flags,
774 struct follow_page_context *ctx)
775{
776 return NULL;
777}
778#endif /* CONFIG_PGTABLE_HAS_HUGE_LEAVES */
779
780static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
781 pte_t *pte, unsigned int flags)
782{
783 if (flags & FOLL_TOUCH) {
784 pte_t orig_entry = ptep_get(pte);
785 pte_t entry = orig_entry;
786
787 if (flags & FOLL_WRITE)
788 entry = pte_mkdirty(entry);
789 entry = pte_mkyoung(entry);
790
791 if (!pte_same(orig_entry, entry)) {
792 set_pte_at(vma->vm_mm, address, pte, entry);
793 update_mmu_cache(vma, address, pte);
794 }
795 }
796
797 /* Proper page table entry exists, but no corresponding struct page */
798 return -EEXIST;
799}
800
801/* FOLL_FORCE can write to even unwritable PTEs in COW mappings. */
802static inline bool can_follow_write_pte(pte_t pte, struct page *page,
803 struct vm_area_struct *vma,
804 unsigned int flags)
805{
806 /* If the pte is writable, we can write to the page. */
807 if (pte_write(pte))
808 return true;
809
810 /* Maybe FOLL_FORCE is set to override it? */
811 if (!(flags & FOLL_FORCE))
812 return false;
813
814 /* But FOLL_FORCE has no effect on shared mappings */
815 if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
816 return false;
817
818 /* ... or read-only private ones */
819 if (!(vma->vm_flags & VM_MAYWRITE))
820 return false;
821
822 /* ... or already writable ones that just need to take a write fault */
823 if (vma->vm_flags & VM_WRITE)
824 return false;
825
826 /*
827 * See can_change_pte_writable(): we broke COW and could map the page
828 * writable if we have an exclusive anonymous page ...
829 */
830 if (!page || !PageAnon(page) || !PageAnonExclusive(page))
831 return false;
832
833 /* ... and a write-fault isn't required for other reasons. */
834 if (pte_needs_soft_dirty_wp(vma, pte))
835 return false;
836 return !userfaultfd_pte_wp(vma, pte);
837}
838
839static struct page *follow_page_pte(struct vm_area_struct *vma,
840 unsigned long address, pmd_t *pmd, unsigned int flags,
841 struct dev_pagemap **pgmap)
842{
843 struct mm_struct *mm = vma->vm_mm;
844 struct folio *folio;
845 struct page *page;
846 spinlock_t *ptl;
847 pte_t *ptep, pte;
848 int ret;
849
850 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
851 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
852 (FOLL_PIN | FOLL_GET)))
853 return ERR_PTR(-EINVAL);
854
855 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
856 if (!ptep)
857 return no_page_table(vma, flags, address);
858 pte = ptep_get(ptep);
859 if (!pte_present(pte))
860 goto no_page;
861 if (pte_protnone(pte) && !gup_can_follow_protnone(vma, flags))
862 goto no_page;
863
864 page = vm_normal_page(vma, address, pte);
865
866 /*
867 * We only care about anon pages in can_follow_write_pte() and don't
868 * have to worry about pte_devmap() because they are never anon.
869 */
870 if ((flags & FOLL_WRITE) &&
871 !can_follow_write_pte(pte, page, vma, flags)) {
872 page = NULL;
873 goto out;
874 }
875
876 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
877 /*
878 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
879 * case since they are only valid while holding the pgmap
880 * reference.
881 */
882 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
883 if (*pgmap)
884 page = pte_page(pte);
885 else
886 goto no_page;
887 } else if (unlikely(!page)) {
888 if (flags & FOLL_DUMP) {
889 /* Avoid special (like zero) pages in core dumps */
890 page = ERR_PTR(-EFAULT);
891 goto out;
892 }
893
894 if (is_zero_pfn(pte_pfn(pte))) {
895 page = pte_page(pte);
896 } else {
897 ret = follow_pfn_pte(vma, address, ptep, flags);
898 page = ERR_PTR(ret);
899 goto out;
900 }
901 }
902 folio = page_folio(page);
903
904 if (!pte_write(pte) && gup_must_unshare(vma, flags, page)) {
905 page = ERR_PTR(-EMLINK);
906 goto out;
907 }
908
909 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
910 !PageAnonExclusive(page), page);
911
912 /* try_grab_folio() does nothing unless FOLL_GET or FOLL_PIN is set. */
913 ret = try_grab_folio(folio, 1, flags);
914 if (unlikely(ret)) {
915 page = ERR_PTR(ret);
916 goto out;
917 }
918
919 /*
920 * We need to make the page accessible if and only if we are going
921 * to access its content (the FOLL_PIN case). Please see
922 * Documentation/core-api/pin_user_pages.rst for details.
923 */
924 if (flags & FOLL_PIN) {
925 ret = arch_make_folio_accessible(folio);
926 if (ret) {
927 unpin_user_page(page);
928 page = ERR_PTR(ret);
929 goto out;
930 }
931 }
932 if (flags & FOLL_TOUCH) {
933 if ((flags & FOLL_WRITE) &&
934 !pte_dirty(pte) && !folio_test_dirty(folio))
935 folio_mark_dirty(folio);
936 /*
937 * pte_mkyoung() would be more correct here, but atomic care
938 * is needed to avoid losing the dirty bit: it is easier to use
939 * folio_mark_accessed().
940 */
941 folio_mark_accessed(folio);
942 }
943out:
944 pte_unmap_unlock(ptep, ptl);
945 return page;
946no_page:
947 pte_unmap_unlock(ptep, ptl);
948 if (!pte_none(pte))
949 return NULL;
950 return no_page_table(vma, flags, address);
951}
952
953static struct page *follow_pmd_mask(struct vm_area_struct *vma,
954 unsigned long address, pud_t *pudp,
955 unsigned int flags,
956 struct follow_page_context *ctx)
957{
958 pmd_t *pmd, pmdval;
959 spinlock_t *ptl;
960 struct page *page;
961 struct mm_struct *mm = vma->vm_mm;
962
963 pmd = pmd_offset(pudp, address);
964 pmdval = pmdp_get_lockless(pmd);
965 if (pmd_none(pmdval))
966 return no_page_table(vma, flags, address);
967 if (!pmd_present(pmdval))
968 return no_page_table(vma, flags, address);
969 if (pmd_devmap(pmdval)) {
970 ptl = pmd_lock(mm, pmd);
971 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
972 spin_unlock(ptl);
973 if (page)
974 return page;
975 return no_page_table(vma, flags, address);
976 }
977 if (likely(!pmd_leaf(pmdval)))
978 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
979
980 if (pmd_protnone(pmdval) && !gup_can_follow_protnone(vma, flags))
981 return no_page_table(vma, flags, address);
982
983 ptl = pmd_lock(mm, pmd);
984 pmdval = *pmd;
985 if (unlikely(!pmd_present(pmdval))) {
986 spin_unlock(ptl);
987 return no_page_table(vma, flags, address);
988 }
989 if (unlikely(!pmd_leaf(pmdval))) {
990 spin_unlock(ptl);
991 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
992 }
993 if (pmd_trans_huge(pmdval) && (flags & FOLL_SPLIT_PMD)) {
994 spin_unlock(ptl);
995 split_huge_pmd(vma, pmd, address);
996 /* If pmd was left empty, stuff a page table in there quickly */
997 return pte_alloc(mm, pmd) ? ERR_PTR(-ENOMEM) :
998 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
999 }
1000 page = follow_huge_pmd(vma, address, pmd, flags, ctx);
1001 spin_unlock(ptl);
1002 return page;
1003}
1004
1005static struct page *follow_pud_mask(struct vm_area_struct *vma,
1006 unsigned long address, p4d_t *p4dp,
1007 unsigned int flags,
1008 struct follow_page_context *ctx)
1009{
1010 pud_t *pudp, pud;
1011 spinlock_t *ptl;
1012 struct page *page;
1013 struct mm_struct *mm = vma->vm_mm;
1014
1015 pudp = pud_offset(p4dp, address);
1016 pud = READ_ONCE(*pudp);
1017 if (!pud_present(pud))
1018 return no_page_table(vma, flags, address);
1019 if (pud_leaf(pud)) {
1020 ptl = pud_lock(mm, pudp);
1021 page = follow_huge_pud(vma, address, pudp, flags, ctx);
1022 spin_unlock(ptl);
1023 if (page)
1024 return page;
1025 return no_page_table(vma, flags, address);
1026 }
1027 if (unlikely(pud_bad(pud)))
1028 return no_page_table(vma, flags, address);
1029
1030 return follow_pmd_mask(vma, address, pudp, flags, ctx);
1031}
1032
1033static struct page *follow_p4d_mask(struct vm_area_struct *vma,
1034 unsigned long address, pgd_t *pgdp,
1035 unsigned int flags,
1036 struct follow_page_context *ctx)
1037{
1038 p4d_t *p4dp, p4d;
1039
1040 p4dp = p4d_offset(pgdp, address);
1041 p4d = READ_ONCE(*p4dp);
1042 BUILD_BUG_ON(p4d_leaf(p4d));
1043
1044 if (!p4d_present(p4d) || p4d_bad(p4d))
1045 return no_page_table(vma, flags, address);
1046
1047 return follow_pud_mask(vma, address, p4dp, flags, ctx);
1048}
1049
1050/**
1051 * follow_page_mask - look up a page descriptor from a user-virtual address
1052 * @vma: vm_area_struct mapping @address
1053 * @address: virtual address to look up
1054 * @flags: flags modifying lookup behaviour
1055 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
1056 * pointer to output page_mask
1057 *
1058 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1059 *
1060 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
1061 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
1062 *
1063 * When getting an anonymous page and the caller has to trigger unsharing
1064 * of a shared anonymous page first, -EMLINK is returned. The caller should
1065 * trigger a fault with FAULT_FLAG_UNSHARE set. Note that unsharing is only
1066 * relevant with FOLL_PIN and !FOLL_WRITE.
1067 *
1068 * On output, the @ctx->page_mask is set according to the size of the page.
1069 *
1070 * Return: the mapped (struct page *), %NULL if no mapping exists, or
1071 * an error pointer if there is a mapping to something not represented
1072 * by a page descriptor (see also vm_normal_page()).
1073 */
1074static struct page *follow_page_mask(struct vm_area_struct *vma,
1075 unsigned long address, unsigned int flags,
1076 struct follow_page_context *ctx)
1077{
1078 pgd_t *pgd;
1079 struct mm_struct *mm = vma->vm_mm;
1080 struct page *page;
1081
1082 vma_pgtable_walk_begin(vma);
1083
1084 ctx->page_mask = 0;
1085 pgd = pgd_offset(mm, address);
1086
1087 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1088 page = no_page_table(vma, flags, address);
1089 else
1090 page = follow_p4d_mask(vma, address, pgd, flags, ctx);
1091
1092 vma_pgtable_walk_end(vma);
1093
1094 return page;
1095}
1096
1097static int get_gate_page(struct mm_struct *mm, unsigned long address,
1098 unsigned int gup_flags, struct vm_area_struct **vma,
1099 struct page **page)
1100{
1101 pgd_t *pgd;
1102 p4d_t *p4d;
1103 pud_t *pud;
1104 pmd_t *pmd;
1105 pte_t *pte;
1106 pte_t entry;
1107 int ret = -EFAULT;
1108
1109 /* user gate pages are read-only */
1110 if (gup_flags & FOLL_WRITE)
1111 return -EFAULT;
1112 if (address > TASK_SIZE)
1113 pgd = pgd_offset_k(address);
1114 else
1115 pgd = pgd_offset_gate(mm, address);
1116 if (pgd_none(*pgd))
1117 return -EFAULT;
1118 p4d = p4d_offset(pgd, address);
1119 if (p4d_none(*p4d))
1120 return -EFAULT;
1121 pud = pud_offset(p4d, address);
1122 if (pud_none(*pud))
1123 return -EFAULT;
1124 pmd = pmd_offset(pud, address);
1125 if (!pmd_present(*pmd))
1126 return -EFAULT;
1127 pte = pte_offset_map(pmd, address);
1128 if (!pte)
1129 return -EFAULT;
1130 entry = ptep_get(pte);
1131 if (pte_none(entry))
1132 goto unmap;
1133 *vma = get_gate_vma(mm);
1134 if (!page)
1135 goto out;
1136 *page = vm_normal_page(*vma, address, entry);
1137 if (!*page) {
1138 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(entry)))
1139 goto unmap;
1140 *page = pte_page(entry);
1141 }
1142 ret = try_grab_folio(page_folio(*page), 1, gup_flags);
1143 if (unlikely(ret))
1144 goto unmap;
1145out:
1146 ret = 0;
1147unmap:
1148 pte_unmap(pte);
1149 return ret;
1150}
1151
1152/*
1153 * mmap_lock must be held on entry. If @flags has FOLL_UNLOCKABLE but not
1154 * FOLL_NOWAIT, the mmap_lock may be released. If it is, *@locked will be set
1155 * to 0 and -EBUSY returned.
1156 */
1157static int faultin_page(struct vm_area_struct *vma,
1158 unsigned long address, unsigned int flags, bool unshare,
1159 int *locked)
1160{
1161 unsigned int fault_flags = 0;
1162 vm_fault_t ret;
1163
1164 if (flags & FOLL_NOFAULT)
1165 return -EFAULT;
1166 if (flags & FOLL_WRITE)
1167 fault_flags |= FAULT_FLAG_WRITE;
1168 if (flags & FOLL_REMOTE)
1169 fault_flags |= FAULT_FLAG_REMOTE;
1170 if (flags & FOLL_UNLOCKABLE) {
1171 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1172 /*
1173 * FAULT_FLAG_INTERRUPTIBLE is opt-in. GUP callers must set
1174 * FOLL_INTERRUPTIBLE to enable FAULT_FLAG_INTERRUPTIBLE.
1175 * That's because some callers may not be prepared to
1176 * handle early exits caused by non-fatal signals.
1177 */
1178 if (flags & FOLL_INTERRUPTIBLE)
1179 fault_flags |= FAULT_FLAG_INTERRUPTIBLE;
1180 }
1181 if (flags & FOLL_NOWAIT)
1182 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
1183 if (flags & FOLL_TRIED) {
1184 /*
1185 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
1186 * can co-exist
1187 */
1188 fault_flags |= FAULT_FLAG_TRIED;
1189 }
1190 if (unshare) {
1191 fault_flags |= FAULT_FLAG_UNSHARE;
1192 /* FAULT_FLAG_WRITE and FAULT_FLAG_UNSHARE are incompatible */
1193 VM_BUG_ON(fault_flags & FAULT_FLAG_WRITE);
1194 }
1195
1196 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1197
1198 if (ret & VM_FAULT_COMPLETED) {
1199 /*
1200 * With FAULT_FLAG_RETRY_NOWAIT we'll never release the
1201 * mmap lock in the page fault handler. Sanity check this.
1202 */
1203 WARN_ON_ONCE(fault_flags & FAULT_FLAG_RETRY_NOWAIT);
1204 *locked = 0;
1205
1206 /*
1207 * We should do the same as VM_FAULT_RETRY, but let's not
1208 * return -EBUSY since that's not reflecting the reality of
1209 * what has happened - we've just fully completed a page
1210 * fault, with the mmap lock released. Use -EAGAIN to show
1211 * that we want to take the mmap lock _again_.
1212 */
1213 return -EAGAIN;
1214 }
1215
1216 if (ret & VM_FAULT_ERROR) {
1217 int err = vm_fault_to_errno(ret, flags);
1218
1219 if (err)
1220 return err;
1221 BUG();
1222 }
1223
1224 if (ret & VM_FAULT_RETRY) {
1225 if (!(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
1226 *locked = 0;
1227 return -EBUSY;
1228 }
1229
1230 return 0;
1231}
1232
1233/*
1234 * Writing to file-backed mappings which require folio dirty tracking using GUP
1235 * is a fundamentally broken operation, as kernel write access to GUP mappings
1236 * do not adhere to the semantics expected by a file system.
1237 *
1238 * Consider the following scenario:-
1239 *
1240 * 1. A folio is written to via GUP which write-faults the memory, notifying
1241 * the file system and dirtying the folio.
1242 * 2. Later, writeback is triggered, resulting in the folio being cleaned and
1243 * the PTE being marked read-only.
1244 * 3. The GUP caller writes to the folio, as it is mapped read/write via the
1245 * direct mapping.
1246 * 4. The GUP caller, now done with the page, unpins it and sets it dirty
1247 * (though it does not have to).
1248 *
1249 * This results in both data being written to a folio without writenotify, and
1250 * the folio being dirtied unexpectedly (if the caller decides to do so).
1251 */
1252static bool writable_file_mapping_allowed(struct vm_area_struct *vma,
1253 unsigned long gup_flags)
1254{
1255 /*
1256 * If we aren't pinning then no problematic write can occur. A long term
1257 * pin is the most egregious case so this is the case we disallow.
1258 */
1259 if ((gup_flags & (FOLL_PIN | FOLL_LONGTERM)) !=
1260 (FOLL_PIN | FOLL_LONGTERM))
1261 return true;
1262
1263 /*
1264 * If the VMA does not require dirty tracking then no problematic write
1265 * can occur either.
1266 */
1267 return !vma_needs_dirty_tracking(vma);
1268}
1269
1270static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
1271{
1272 vm_flags_t vm_flags = vma->vm_flags;
1273 int write = (gup_flags & FOLL_WRITE);
1274 int foreign = (gup_flags & FOLL_REMOTE);
1275 bool vma_anon = vma_is_anonymous(vma);
1276
1277 if (vm_flags & (VM_IO | VM_PFNMAP))
1278 return -EFAULT;
1279
1280 if ((gup_flags & FOLL_ANON) && !vma_anon)
1281 return -EFAULT;
1282
1283 if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma))
1284 return -EOPNOTSUPP;
1285
1286 if (vma_is_secretmem(vma))
1287 return -EFAULT;
1288
1289 if (write) {
1290 if (!vma_anon &&
1291 !writable_file_mapping_allowed(vma, gup_flags))
1292 return -EFAULT;
1293
1294 if (!(vm_flags & VM_WRITE) || (vm_flags & VM_SHADOW_STACK)) {
1295 if (!(gup_flags & FOLL_FORCE))
1296 return -EFAULT;
1297 /* hugetlb does not support FOLL_FORCE|FOLL_WRITE. */
1298 if (is_vm_hugetlb_page(vma))
1299 return -EFAULT;
1300 /*
1301 * We used to let the write,force case do COW in a
1302 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
1303 * set a breakpoint in a read-only mapping of an
1304 * executable, without corrupting the file (yet only
1305 * when that file had been opened for writing!).
1306 * Anon pages in shared mappings are surprising: now
1307 * just reject it.
1308 */
1309 if (!is_cow_mapping(vm_flags))
1310 return -EFAULT;
1311 }
1312 } else if (!(vm_flags & VM_READ)) {
1313 if (!(gup_flags & FOLL_FORCE))
1314 return -EFAULT;
1315 /*
1316 * Is there actually any vma we can reach here which does not
1317 * have VM_MAYREAD set?
1318 */
1319 if (!(vm_flags & VM_MAYREAD))
1320 return -EFAULT;
1321 }
1322 /*
1323 * gups are always data accesses, not instruction
1324 * fetches, so execute=false here
1325 */
1326 if (!arch_vma_access_permitted(vma, write, false, foreign))
1327 return -EFAULT;
1328 return 0;
1329}
1330
1331/*
1332 * This is "vma_lookup()", but with a warning if we would have
1333 * historically expanded the stack in the GUP code.
1334 */
1335static struct vm_area_struct *gup_vma_lookup(struct mm_struct *mm,
1336 unsigned long addr)
1337{
1338#ifdef CONFIG_STACK_GROWSUP
1339 return vma_lookup(mm, addr);
1340#else
1341 static volatile unsigned long next_warn;
1342 struct vm_area_struct *vma;
1343 unsigned long now, next;
1344
1345 vma = find_vma(mm, addr);
1346 if (!vma || (addr >= vma->vm_start))
1347 return vma;
1348
1349 /* Only warn for half-way relevant accesses */
1350 if (!(vma->vm_flags & VM_GROWSDOWN))
1351 return NULL;
1352 if (vma->vm_start - addr > 65536)
1353 return NULL;
1354
1355 /* Let's not warn more than once an hour.. */
1356 now = jiffies; next = next_warn;
1357 if (next && time_before(now, next))
1358 return NULL;
1359 next_warn = now + 60*60*HZ;
1360
1361 /* Let people know things may have changed. */
1362 pr_warn("GUP no longer grows the stack in %s (%d): %lx-%lx (%lx)\n",
1363 current->comm, task_pid_nr(current),
1364 vma->vm_start, vma->vm_end, addr);
1365 dump_stack();
1366 return NULL;
1367#endif
1368}
1369
1370/**
1371 * __get_user_pages() - pin user pages in memory
1372 * @mm: mm_struct of target mm
1373 * @start: starting user address
1374 * @nr_pages: number of pages from start to pin
1375 * @gup_flags: flags modifying pin behaviour
1376 * @pages: array that receives pointers to the pages pinned.
1377 * Should be at least nr_pages long. Or NULL, if caller
1378 * only intends to ensure the pages are faulted in.
1379 * @locked: whether we're still with the mmap_lock held
1380 *
1381 * Returns either number of pages pinned (which may be less than the
1382 * number requested), or an error. Details about the return value:
1383 *
1384 * -- If nr_pages is 0, returns 0.
1385 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1386 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1387 * pages pinned. Again, this may be less than nr_pages.
1388 * -- 0 return value is possible when the fault would need to be retried.
1389 *
1390 * The caller is responsible for releasing returned @pages, via put_page().
1391 *
1392 * Must be called with mmap_lock held. It may be released. See below.
1393 *
1394 * __get_user_pages walks a process's page tables and takes a reference to
1395 * each struct page that each user address corresponds to at a given
1396 * instant. That is, it takes the page that would be accessed if a user
1397 * thread accesses the given user virtual address at that instant.
1398 *
1399 * This does not guarantee that the page exists in the user mappings when
1400 * __get_user_pages returns, and there may even be a completely different
1401 * page there in some cases (eg. if mmapped pagecache has been invalidated
1402 * and subsequently re-faulted). However it does guarantee that the page
1403 * won't be freed completely. And mostly callers simply care that the page
1404 * contains data that was valid *at some point in time*. Typically, an IO
1405 * or similar operation cannot guarantee anything stronger anyway because
1406 * locks can't be held over the syscall boundary.
1407 *
1408 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1409 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1410 * appropriate) must be called after the page is finished with, and
1411 * before put_page is called.
1412 *
1413 * If FOLL_UNLOCKABLE is set without FOLL_NOWAIT then the mmap_lock may
1414 * be released. If this happens *@locked will be set to 0 on return.
1415 *
1416 * A caller using such a combination of @gup_flags must therefore hold the
1417 * mmap_lock for reading only, and recognize when it's been released. Otherwise,
1418 * it must be held for either reading or writing and will not be released.
1419 *
1420 * In most cases, get_user_pages or get_user_pages_fast should be used
1421 * instead of __get_user_pages. __get_user_pages should be used only if
1422 * you need some special @gup_flags.
1423 */
1424static long __get_user_pages(struct mm_struct *mm,
1425 unsigned long start, unsigned long nr_pages,
1426 unsigned int gup_flags, struct page **pages,
1427 int *locked)
1428{
1429 long ret = 0, i = 0;
1430 struct vm_area_struct *vma = NULL;
1431 struct follow_page_context ctx = { NULL };
1432
1433 if (!nr_pages)
1434 return 0;
1435
1436 start = untagged_addr_remote(mm, start);
1437
1438 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1439
1440 do {
1441 struct page *page;
1442 unsigned int page_increm;
1443
1444 /* first iteration or cross vma bound */
1445 if (!vma || start >= vma->vm_end) {
1446 /*
1447 * MADV_POPULATE_(READ|WRITE) wants to handle VMA
1448 * lookups+error reporting differently.
1449 */
1450 if (gup_flags & FOLL_MADV_POPULATE) {
1451 vma = vma_lookup(mm, start);
1452 if (!vma) {
1453 ret = -ENOMEM;
1454 goto out;
1455 }
1456 if (check_vma_flags(vma, gup_flags)) {
1457 ret = -EINVAL;
1458 goto out;
1459 }
1460 goto retry;
1461 }
1462 vma = gup_vma_lookup(mm, start);
1463 if (!vma && in_gate_area(mm, start)) {
1464 ret = get_gate_page(mm, start & PAGE_MASK,
1465 gup_flags, &vma,
1466 pages ? &page : NULL);
1467 if (ret)
1468 goto out;
1469 ctx.page_mask = 0;
1470 goto next_page;
1471 }
1472
1473 if (!vma) {
1474 ret = -EFAULT;
1475 goto out;
1476 }
1477 ret = check_vma_flags(vma, gup_flags);
1478 if (ret)
1479 goto out;
1480 }
1481retry:
1482 /*
1483 * If we have a pending SIGKILL, don't keep faulting pages and
1484 * potentially allocating memory.
1485 */
1486 if (fatal_signal_pending(current)) {
1487 ret = -EINTR;
1488 goto out;
1489 }
1490 cond_resched();
1491
1492 page = follow_page_mask(vma, start, gup_flags, &ctx);
1493 if (!page || PTR_ERR(page) == -EMLINK) {
1494 ret = faultin_page(vma, start, gup_flags,
1495 PTR_ERR(page) == -EMLINK, locked);
1496 switch (ret) {
1497 case 0:
1498 goto retry;
1499 case -EBUSY:
1500 case -EAGAIN:
1501 ret = 0;
1502 fallthrough;
1503 case -EFAULT:
1504 case -ENOMEM:
1505 case -EHWPOISON:
1506 goto out;
1507 }
1508 BUG();
1509 } else if (PTR_ERR(page) == -EEXIST) {
1510 /*
1511 * Proper page table entry exists, but no corresponding
1512 * struct page. If the caller expects **pages to be
1513 * filled in, bail out now, because that can't be done
1514 * for this page.
1515 */
1516 if (pages) {
1517 ret = PTR_ERR(page);
1518 goto out;
1519 }
1520 } else if (IS_ERR(page)) {
1521 ret = PTR_ERR(page);
1522 goto out;
1523 }
1524next_page:
1525 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1526 if (page_increm > nr_pages)
1527 page_increm = nr_pages;
1528
1529 if (pages) {
1530 struct page *subpage;
1531 unsigned int j;
1532
1533 /*
1534 * This must be a large folio (and doesn't need to
1535 * be the whole folio; it can be part of it), do
1536 * the refcount work for all the subpages too.
1537 *
1538 * NOTE: here the page may not be the head page
1539 * e.g. when start addr is not thp-size aligned.
1540 * try_grab_folio() should have taken care of tail
1541 * pages.
1542 */
1543 if (page_increm > 1) {
1544 struct folio *folio = page_folio(page);
1545
1546 /*
1547 * Since we already hold refcount on the
1548 * large folio, this should never fail.
1549 */
1550 if (try_grab_folio(folio, page_increm - 1,
1551 gup_flags)) {
1552 /*
1553 * Release the 1st page ref if the
1554 * folio is problematic, fail hard.
1555 */
1556 gup_put_folio(folio, 1, gup_flags);
1557 ret = -EFAULT;
1558 goto out;
1559 }
1560 }
1561
1562 for (j = 0; j < page_increm; j++) {
1563 subpage = nth_page(page, j);
1564 pages[i + j] = subpage;
1565 flush_anon_page(vma, subpage, start + j * PAGE_SIZE);
1566 flush_dcache_page(subpage);
1567 }
1568 }
1569
1570 i += page_increm;
1571 start += page_increm * PAGE_SIZE;
1572 nr_pages -= page_increm;
1573 } while (nr_pages);
1574out:
1575 if (ctx.pgmap)
1576 put_dev_pagemap(ctx.pgmap);
1577 return i ? i : ret;
1578}
1579
1580static bool vma_permits_fault(struct vm_area_struct *vma,
1581 unsigned int fault_flags)
1582{
1583 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1584 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1585 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1586
1587 if (!(vm_flags & vma->vm_flags))
1588 return false;
1589
1590 /*
1591 * The architecture might have a hardware protection
1592 * mechanism other than read/write that can deny access.
1593 *
1594 * gup always represents data access, not instruction
1595 * fetches, so execute=false here:
1596 */
1597 if (!arch_vma_access_permitted(vma, write, false, foreign))
1598 return false;
1599
1600 return true;
1601}
1602
1603/**
1604 * fixup_user_fault() - manually resolve a user page fault
1605 * @mm: mm_struct of target mm
1606 * @address: user address
1607 * @fault_flags:flags to pass down to handle_mm_fault()
1608 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1609 * does not allow retry. If NULL, the caller must guarantee
1610 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1611 *
1612 * This is meant to be called in the specific scenario where for locking reasons
1613 * we try to access user memory in atomic context (within a pagefault_disable()
1614 * section), this returns -EFAULT, and we want to resolve the user fault before
1615 * trying again.
1616 *
1617 * Typically this is meant to be used by the futex code.
1618 *
1619 * The main difference with get_user_pages() is that this function will
1620 * unconditionally call handle_mm_fault() which will in turn perform all the
1621 * necessary SW fixup of the dirty and young bits in the PTE, while
1622 * get_user_pages() only guarantees to update these in the struct page.
1623 *
1624 * This is important for some architectures where those bits also gate the
1625 * access permission to the page because they are maintained in software. On
1626 * such architectures, gup() will not be enough to make a subsequent access
1627 * succeed.
1628 *
1629 * This function will not return with an unlocked mmap_lock. So it has not the
1630 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1631 */
1632int fixup_user_fault(struct mm_struct *mm,
1633 unsigned long address, unsigned int fault_flags,
1634 bool *unlocked)
1635{
1636 struct vm_area_struct *vma;
1637 vm_fault_t ret;
1638
1639 address = untagged_addr_remote(mm, address);
1640
1641 if (unlocked)
1642 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1643
1644retry:
1645 vma = gup_vma_lookup(mm, address);
1646 if (!vma)
1647 return -EFAULT;
1648
1649 if (!vma_permits_fault(vma, fault_flags))
1650 return -EFAULT;
1651
1652 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1653 fatal_signal_pending(current))
1654 return -EINTR;
1655
1656 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1657
1658 if (ret & VM_FAULT_COMPLETED) {
1659 /*
1660 * NOTE: it's a pity that we need to retake the lock here
1661 * to pair with the unlock() in the callers. Ideally we
1662 * could tell the callers so they do not need to unlock.
1663 */
1664 mmap_read_lock(mm);
1665 *unlocked = true;
1666 return 0;
1667 }
1668
1669 if (ret & VM_FAULT_ERROR) {
1670 int err = vm_fault_to_errno(ret, 0);
1671
1672 if (err)
1673 return err;
1674 BUG();
1675 }
1676
1677 if (ret & VM_FAULT_RETRY) {
1678 mmap_read_lock(mm);
1679 *unlocked = true;
1680 fault_flags |= FAULT_FLAG_TRIED;
1681 goto retry;
1682 }
1683
1684 return 0;
1685}
1686EXPORT_SYMBOL_GPL(fixup_user_fault);
1687
1688/*
1689 * GUP always responds to fatal signals. When FOLL_INTERRUPTIBLE is
1690 * specified, it'll also respond to generic signals. The caller of GUP
1691 * that has FOLL_INTERRUPTIBLE should take care of the GUP interruption.
1692 */
1693static bool gup_signal_pending(unsigned int flags)
1694{
1695 if (fatal_signal_pending(current))
1696 return true;
1697
1698 if (!(flags & FOLL_INTERRUPTIBLE))
1699 return false;
1700
1701 return signal_pending(current);
1702}
1703
1704/*
1705 * Locking: (*locked == 1) means that the mmap_lock has already been acquired by
1706 * the caller. This function may drop the mmap_lock. If it does so, then it will
1707 * set (*locked = 0).
1708 *
1709 * (*locked == 0) means that the caller expects this function to acquire and
1710 * drop the mmap_lock. Therefore, the value of *locked will still be zero when
1711 * the function returns, even though it may have changed temporarily during
1712 * function execution.
1713 *
1714 * Please note that this function, unlike __get_user_pages(), will not return 0
1715 * for nr_pages > 0, unless FOLL_NOWAIT is used.
1716 */
1717static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1718 unsigned long start,
1719 unsigned long nr_pages,
1720 struct page **pages,
1721 int *locked,
1722 unsigned int flags)
1723{
1724 long ret, pages_done;
1725 bool must_unlock = false;
1726
1727 if (!nr_pages)
1728 return 0;
1729
1730 /*
1731 * The internal caller expects GUP to manage the lock internally and the
1732 * lock must be released when this returns.
1733 */
1734 if (!*locked) {
1735 if (mmap_read_lock_killable(mm))
1736 return -EAGAIN;
1737 must_unlock = true;
1738 *locked = 1;
1739 }
1740 else
1741 mmap_assert_locked(mm);
1742
1743 if (flags & FOLL_PIN)
1744 mm_set_has_pinned_flag(&mm->flags);
1745
1746 /*
1747 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1748 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1749 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1750 * for FOLL_GET, not for the newer FOLL_PIN.
1751 *
1752 * FOLL_PIN always expects pages to be non-null, but no need to assert
1753 * that here, as any failures will be obvious enough.
1754 */
1755 if (pages && !(flags & FOLL_PIN))
1756 flags |= FOLL_GET;
1757
1758 pages_done = 0;
1759 for (;;) {
1760 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1761 locked);
1762 if (!(flags & FOLL_UNLOCKABLE)) {
1763 /* VM_FAULT_RETRY couldn't trigger, bypass */
1764 pages_done = ret;
1765 break;
1766 }
1767
1768 /* VM_FAULT_RETRY or VM_FAULT_COMPLETED cannot return errors */
1769 if (!*locked) {
1770 BUG_ON(ret < 0);
1771 BUG_ON(ret >= nr_pages);
1772 }
1773
1774 if (ret > 0) {
1775 nr_pages -= ret;
1776 pages_done += ret;
1777 if (!nr_pages)
1778 break;
1779 }
1780 if (*locked) {
1781 /*
1782 * VM_FAULT_RETRY didn't trigger or it was a
1783 * FOLL_NOWAIT.
1784 */
1785 if (!pages_done)
1786 pages_done = ret;
1787 break;
1788 }
1789 /*
1790 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1791 * For the prefault case (!pages) we only update counts.
1792 */
1793 if (likely(pages))
1794 pages += ret;
1795 start += ret << PAGE_SHIFT;
1796
1797 /* The lock was temporarily dropped, so we must unlock later */
1798 must_unlock = true;
1799
1800retry:
1801 /*
1802 * Repeat on the address that fired VM_FAULT_RETRY
1803 * with both FAULT_FLAG_ALLOW_RETRY and
1804 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1805 * by fatal signals of even common signals, depending on
1806 * the caller's request. So we need to check it before we
1807 * start trying again otherwise it can loop forever.
1808 */
1809 if (gup_signal_pending(flags)) {
1810 if (!pages_done)
1811 pages_done = -EINTR;
1812 break;
1813 }
1814
1815 ret = mmap_read_lock_killable(mm);
1816 if (ret) {
1817 BUG_ON(ret > 0);
1818 if (!pages_done)
1819 pages_done = ret;
1820 break;
1821 }
1822
1823 *locked = 1;
1824 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1825 pages, locked);
1826 if (!*locked) {
1827 /* Continue to retry until we succeeded */
1828 BUG_ON(ret != 0);
1829 goto retry;
1830 }
1831 if (ret != 1) {
1832 BUG_ON(ret > 1);
1833 if (!pages_done)
1834 pages_done = ret;
1835 break;
1836 }
1837 nr_pages--;
1838 pages_done++;
1839 if (!nr_pages)
1840 break;
1841 if (likely(pages))
1842 pages++;
1843 start += PAGE_SIZE;
1844 }
1845 if (must_unlock && *locked) {
1846 /*
1847 * We either temporarily dropped the lock, or the caller
1848 * requested that we both acquire and drop the lock. Either way,
1849 * we must now unlock, and notify the caller of that state.
1850 */
1851 mmap_read_unlock(mm);
1852 *locked = 0;
1853 }
1854
1855 /*
1856 * Failing to pin anything implies something has gone wrong (except when
1857 * FOLL_NOWAIT is specified).
1858 */
1859 if (WARN_ON_ONCE(pages_done == 0 && !(flags & FOLL_NOWAIT)))
1860 return -EFAULT;
1861
1862 return pages_done;
1863}
1864
1865/**
1866 * populate_vma_page_range() - populate a range of pages in the vma.
1867 * @vma: target vma
1868 * @start: start address
1869 * @end: end address
1870 * @locked: whether the mmap_lock is still held
1871 *
1872 * This takes care of mlocking the pages too if VM_LOCKED is set.
1873 *
1874 * Return either number of pages pinned in the vma, or a negative error
1875 * code on error.
1876 *
1877 * vma->vm_mm->mmap_lock must be held.
1878 *
1879 * If @locked is NULL, it may be held for read or write and will
1880 * be unperturbed.
1881 *
1882 * If @locked is non-NULL, it must held for read only and may be
1883 * released. If it's released, *@locked will be set to 0.
1884 */
1885long populate_vma_page_range(struct vm_area_struct *vma,
1886 unsigned long start, unsigned long end, int *locked)
1887{
1888 struct mm_struct *mm = vma->vm_mm;
1889 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1890 int local_locked = 1;
1891 int gup_flags;
1892 long ret;
1893
1894 VM_BUG_ON(!PAGE_ALIGNED(start));
1895 VM_BUG_ON(!PAGE_ALIGNED(end));
1896 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1897 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1898 mmap_assert_locked(mm);
1899
1900 /*
1901 * Rightly or wrongly, the VM_LOCKONFAULT case has never used
1902 * faultin_page() to break COW, so it has no work to do here.
1903 */
1904 if (vma->vm_flags & VM_LOCKONFAULT)
1905 return nr_pages;
1906
1907 /* ... similarly, we've never faulted in PROT_NONE pages */
1908 if (!vma_is_accessible(vma))
1909 return -EFAULT;
1910
1911 gup_flags = FOLL_TOUCH;
1912 /*
1913 * We want to touch writable mappings with a write fault in order
1914 * to break COW, except for shared mappings because these don't COW
1915 * and we would not want to dirty them for nothing.
1916 *
1917 * Otherwise, do a read fault, and use FOLL_FORCE in case it's not
1918 * readable (ie write-only or executable).
1919 */
1920 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1921 gup_flags |= FOLL_WRITE;
1922 else
1923 gup_flags |= FOLL_FORCE;
1924
1925 if (locked)
1926 gup_flags |= FOLL_UNLOCKABLE;
1927
1928 /*
1929 * We made sure addr is within a VMA, so the following will
1930 * not result in a stack expansion that recurses back here.
1931 */
1932 ret = __get_user_pages(mm, start, nr_pages, gup_flags,
1933 NULL, locked ? locked : &local_locked);
1934 lru_add_drain();
1935 return ret;
1936}
1937
1938/*
1939 * faultin_page_range() - populate (prefault) page tables inside the
1940 * given range readable/writable
1941 *
1942 * This takes care of mlocking the pages, too, if VM_LOCKED is set.
1943 *
1944 * @mm: the mm to populate page tables in
1945 * @start: start address
1946 * @end: end address
1947 * @write: whether to prefault readable or writable
1948 * @locked: whether the mmap_lock is still held
1949 *
1950 * Returns either number of processed pages in the MM, or a negative error
1951 * code on error (see __get_user_pages()). Note that this function reports
1952 * errors related to VMAs, such as incompatible mappings, as expected by
1953 * MADV_POPULATE_(READ|WRITE).
1954 *
1955 * The range must be page-aligned.
1956 *
1957 * mm->mmap_lock must be held. If it's released, *@locked will be set to 0.
1958 */
1959long faultin_page_range(struct mm_struct *mm, unsigned long start,
1960 unsigned long end, bool write, int *locked)
1961{
1962 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1963 int gup_flags;
1964 long ret;
1965
1966 VM_BUG_ON(!PAGE_ALIGNED(start));
1967 VM_BUG_ON(!PAGE_ALIGNED(end));
1968 mmap_assert_locked(mm);
1969
1970 /*
1971 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark
1972 * the page dirty with FOLL_WRITE -- which doesn't make a
1973 * difference with !FOLL_FORCE, because the page is writable
1974 * in the page table.
1975 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit
1976 * a poisoned page.
1977 * !FOLL_FORCE: Require proper access permissions.
1978 */
1979 gup_flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_UNLOCKABLE |
1980 FOLL_MADV_POPULATE;
1981 if (write)
1982 gup_flags |= FOLL_WRITE;
1983
1984 ret = __get_user_pages_locked(mm, start, nr_pages, NULL, locked,
1985 gup_flags);
1986 lru_add_drain();
1987 return ret;
1988}
1989
1990/*
1991 * __mm_populate - populate and/or mlock pages within a range of address space.
1992 *
1993 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1994 * flags. VMAs must be already marked with the desired vm_flags, and
1995 * mmap_lock must not be held.
1996 */
1997int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1998{
1999 struct mm_struct *mm = current->mm;
2000 unsigned long end, nstart, nend;
2001 struct vm_area_struct *vma = NULL;
2002 int locked = 0;
2003 long ret = 0;
2004
2005 end = start + len;
2006
2007 for (nstart = start; nstart < end; nstart = nend) {
2008 /*
2009 * We want to fault in pages for [nstart; end) address range.
2010 * Find first corresponding VMA.
2011 */
2012 if (!locked) {
2013 locked = 1;
2014 mmap_read_lock(mm);
2015 vma = find_vma_intersection(mm, nstart, end);
2016 } else if (nstart >= vma->vm_end)
2017 vma = find_vma_intersection(mm, vma->vm_end, end);
2018
2019 if (!vma)
2020 break;
2021 /*
2022 * Set [nstart; nend) to intersection of desired address
2023 * range with the first VMA. Also, skip undesirable VMA types.
2024 */
2025 nend = min(end, vma->vm_end);
2026 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
2027 continue;
2028 if (nstart < vma->vm_start)
2029 nstart = vma->vm_start;
2030 /*
2031 * Now fault in a range of pages. populate_vma_page_range()
2032 * double checks the vma flags, so that it won't mlock pages
2033 * if the vma was already munlocked.
2034 */
2035 ret = populate_vma_page_range(vma, nstart, nend, &locked);
2036 if (ret < 0) {
2037 if (ignore_errors) {
2038 ret = 0;
2039 continue; /* continue at next VMA */
2040 }
2041 break;
2042 }
2043 nend = nstart + ret * PAGE_SIZE;
2044 ret = 0;
2045 }
2046 if (locked)
2047 mmap_read_unlock(mm);
2048 return ret; /* 0 or negative error code */
2049}
2050#else /* CONFIG_MMU */
2051static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
2052 unsigned long nr_pages, struct page **pages,
2053 int *locked, unsigned int foll_flags)
2054{
2055 struct vm_area_struct *vma;
2056 bool must_unlock = false;
2057 unsigned long vm_flags;
2058 long i;
2059
2060 if (!nr_pages)
2061 return 0;
2062
2063 /*
2064 * The internal caller expects GUP to manage the lock internally and the
2065 * lock must be released when this returns.
2066 */
2067 if (!*locked) {
2068 if (mmap_read_lock_killable(mm))
2069 return -EAGAIN;
2070 must_unlock = true;
2071 *locked = 1;
2072 }
2073
2074 /* calculate required read or write permissions.
2075 * If FOLL_FORCE is set, we only require the "MAY" flags.
2076 */
2077 vm_flags = (foll_flags & FOLL_WRITE) ?
2078 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
2079 vm_flags &= (foll_flags & FOLL_FORCE) ?
2080 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
2081
2082 for (i = 0; i < nr_pages; i++) {
2083 vma = find_vma(mm, start);
2084 if (!vma)
2085 break;
2086
2087 /* protect what we can, including chardevs */
2088 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
2089 !(vm_flags & vma->vm_flags))
2090 break;
2091
2092 if (pages) {
2093 pages[i] = virt_to_page((void *)start);
2094 if (pages[i])
2095 get_page(pages[i]);
2096 }
2097
2098 start = (start + PAGE_SIZE) & PAGE_MASK;
2099 }
2100
2101 if (must_unlock && *locked) {
2102 mmap_read_unlock(mm);
2103 *locked = 0;
2104 }
2105
2106 return i ? : -EFAULT;
2107}
2108#endif /* !CONFIG_MMU */
2109
2110/**
2111 * fault_in_writeable - fault in userspace address range for writing
2112 * @uaddr: start of address range
2113 * @size: size of address range
2114 *
2115 * Returns the number of bytes not faulted in (like copy_to_user() and
2116 * copy_from_user()).
2117 */
2118size_t fault_in_writeable(char __user *uaddr, size_t size)
2119{
2120 char __user *start = uaddr, *end;
2121
2122 if (unlikely(size == 0))
2123 return 0;
2124 if (!user_write_access_begin(uaddr, size))
2125 return size;
2126 if (!PAGE_ALIGNED(uaddr)) {
2127 unsafe_put_user(0, uaddr, out);
2128 uaddr = (char __user *)PAGE_ALIGN((unsigned long)uaddr);
2129 }
2130 end = (char __user *)PAGE_ALIGN((unsigned long)start + size);
2131 if (unlikely(end < start))
2132 end = NULL;
2133 while (uaddr != end) {
2134 unsafe_put_user(0, uaddr, out);
2135 uaddr += PAGE_SIZE;
2136 }
2137
2138out:
2139 user_write_access_end();
2140 if (size > uaddr - start)
2141 return size - (uaddr - start);
2142 return 0;
2143}
2144EXPORT_SYMBOL(fault_in_writeable);
2145
2146/**
2147 * fault_in_subpage_writeable - fault in an address range for writing
2148 * @uaddr: start of address range
2149 * @size: size of address range
2150 *
2151 * Fault in a user address range for writing while checking for permissions at
2152 * sub-page granularity (e.g. arm64 MTE). This function should be used when
2153 * the caller cannot guarantee forward progress of a copy_to_user() loop.
2154 *
2155 * Returns the number of bytes not faulted in (like copy_to_user() and
2156 * copy_from_user()).
2157 */
2158size_t fault_in_subpage_writeable(char __user *uaddr, size_t size)
2159{
2160 size_t faulted_in;
2161
2162 /*
2163 * Attempt faulting in at page granularity first for page table
2164 * permission checking. The arch-specific probe_subpage_writeable()
2165 * functions may not check for this.
2166 */
2167 faulted_in = size - fault_in_writeable(uaddr, size);
2168 if (faulted_in)
2169 faulted_in -= probe_subpage_writeable(uaddr, faulted_in);
2170
2171 return size - faulted_in;
2172}
2173EXPORT_SYMBOL(fault_in_subpage_writeable);
2174
2175/*
2176 * fault_in_safe_writeable - fault in an address range for writing
2177 * @uaddr: start of address range
2178 * @size: length of address range
2179 *
2180 * Faults in an address range for writing. This is primarily useful when we
2181 * already know that some or all of the pages in the address range aren't in
2182 * memory.
2183 *
2184 * Unlike fault_in_writeable(), this function is non-destructive.
2185 *
2186 * Note that we don't pin or otherwise hold the pages referenced that we fault
2187 * in. There's no guarantee that they'll stay in memory for any duration of
2188 * time.
2189 *
2190 * Returns the number of bytes not faulted in, like copy_to_user() and
2191 * copy_from_user().
2192 */
2193size_t fault_in_safe_writeable(const char __user *uaddr, size_t size)
2194{
2195 unsigned long start = (unsigned long)uaddr, end;
2196 struct mm_struct *mm = current->mm;
2197 bool unlocked = false;
2198
2199 if (unlikely(size == 0))
2200 return 0;
2201 end = PAGE_ALIGN(start + size);
2202 if (end < start)
2203 end = 0;
2204
2205 mmap_read_lock(mm);
2206 do {
2207 if (fixup_user_fault(mm, start, FAULT_FLAG_WRITE, &unlocked))
2208 break;
2209 start = (start + PAGE_SIZE) & PAGE_MASK;
2210 } while (start != end);
2211 mmap_read_unlock(mm);
2212
2213 if (size > (unsigned long)uaddr - start)
2214 return size - ((unsigned long)uaddr - start);
2215 return 0;
2216}
2217EXPORT_SYMBOL(fault_in_safe_writeable);
2218
2219/**
2220 * fault_in_readable - fault in userspace address range for reading
2221 * @uaddr: start of user address range
2222 * @size: size of user address range
2223 *
2224 * Returns the number of bytes not faulted in (like copy_to_user() and
2225 * copy_from_user()).
2226 */
2227size_t fault_in_readable(const char __user *uaddr, size_t size)
2228{
2229 const char __user *start = uaddr, *end;
2230 volatile char c;
2231
2232 if (unlikely(size == 0))
2233 return 0;
2234 if (!user_read_access_begin(uaddr, size))
2235 return size;
2236 if (!PAGE_ALIGNED(uaddr)) {
2237 unsafe_get_user(c, uaddr, out);
2238 uaddr = (const char __user *)PAGE_ALIGN((unsigned long)uaddr);
2239 }
2240 end = (const char __user *)PAGE_ALIGN((unsigned long)start + size);
2241 if (unlikely(end < start))
2242 end = NULL;
2243 while (uaddr != end) {
2244 unsafe_get_user(c, uaddr, out);
2245 uaddr += PAGE_SIZE;
2246 }
2247
2248out:
2249 user_read_access_end();
2250 (void)c;
2251 if (size > uaddr - start)
2252 return size - (uaddr - start);
2253 return 0;
2254}
2255EXPORT_SYMBOL(fault_in_readable);
2256
2257/**
2258 * get_dump_page() - pin user page in memory while writing it to core dump
2259 * @addr: user address
2260 *
2261 * Returns struct page pointer of user page pinned for dump,
2262 * to be freed afterwards by put_page().
2263 *
2264 * Returns NULL on any kind of failure - a hole must then be inserted into
2265 * the corefile, to preserve alignment with its headers; and also returns
2266 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2267 * allowing a hole to be left in the corefile to save disk space.
2268 *
2269 * Called without mmap_lock (takes and releases the mmap_lock by itself).
2270 */
2271#ifdef CONFIG_ELF_CORE
2272struct page *get_dump_page(unsigned long addr)
2273{
2274 struct page *page;
2275 int locked = 0;
2276 int ret;
2277
2278 ret = __get_user_pages_locked(current->mm, addr, 1, &page, &locked,
2279 FOLL_FORCE | FOLL_DUMP | FOLL_GET);
2280 return (ret == 1) ? page : NULL;
2281}
2282#endif /* CONFIG_ELF_CORE */
2283
2284#ifdef CONFIG_MIGRATION
2285
2286/*
2287 * An array of either pages or folios ("pofs"). Although it may seem tempting to
2288 * avoid this complication, by simply interpreting a list of folios as a list of
2289 * pages, that approach won't work in the longer term, because eventually the
2290 * layouts of struct page and struct folio will become completely different.
2291 * Furthermore, this pof approach avoids excessive page_folio() calls.
2292 */
2293struct pages_or_folios {
2294 union {
2295 struct page **pages;
2296 struct folio **folios;
2297 void **entries;
2298 };
2299 bool has_folios;
2300 long nr_entries;
2301};
2302
2303static struct folio *pofs_get_folio(struct pages_or_folios *pofs, long i)
2304{
2305 if (pofs->has_folios)
2306 return pofs->folios[i];
2307 return page_folio(pofs->pages[i]);
2308}
2309
2310static void pofs_clear_entry(struct pages_or_folios *pofs, long i)
2311{
2312 pofs->entries[i] = NULL;
2313}
2314
2315static void pofs_unpin(struct pages_or_folios *pofs)
2316{
2317 if (pofs->has_folios)
2318 unpin_folios(pofs->folios, pofs->nr_entries);
2319 else
2320 unpin_user_pages(pofs->pages, pofs->nr_entries);
2321}
2322
2323/*
2324 * Returns the number of collected folios. Return value is always >= 0.
2325 */
2326static void collect_longterm_unpinnable_folios(
2327 struct list_head *movable_folio_list,
2328 struct pages_or_folios *pofs)
2329{
2330 struct folio *prev_folio = NULL;
2331 bool drain_allow = true;
2332 unsigned long i;
2333
2334 for (i = 0; i < pofs->nr_entries; i++) {
2335 struct folio *folio = pofs_get_folio(pofs, i);
2336
2337 if (folio == prev_folio)
2338 continue;
2339 prev_folio = folio;
2340
2341 if (folio_is_longterm_pinnable(folio))
2342 continue;
2343
2344 if (folio_is_device_coherent(folio))
2345 continue;
2346
2347 if (folio_test_hugetlb(folio)) {
2348 isolate_hugetlb(folio, movable_folio_list);
2349 continue;
2350 }
2351
2352 if (!folio_test_lru(folio) && drain_allow) {
2353 lru_add_drain_all();
2354 drain_allow = false;
2355 }
2356
2357 if (!folio_isolate_lru(folio))
2358 continue;
2359
2360 list_add_tail(&folio->lru, movable_folio_list);
2361 node_stat_mod_folio(folio,
2362 NR_ISOLATED_ANON + folio_is_file_lru(folio),
2363 folio_nr_pages(folio));
2364 }
2365}
2366
2367/*
2368 * Unpins all folios and migrates device coherent folios and movable_folio_list.
2369 * Returns -EAGAIN if all folios were successfully migrated or -errno for
2370 * failure (or partial success).
2371 */
2372static int
2373migrate_longterm_unpinnable_folios(struct list_head *movable_folio_list,
2374 struct pages_or_folios *pofs)
2375{
2376 int ret;
2377 unsigned long i;
2378
2379 for (i = 0; i < pofs->nr_entries; i++) {
2380 struct folio *folio = pofs_get_folio(pofs, i);
2381
2382 if (folio_is_device_coherent(folio)) {
2383 /*
2384 * Migration will fail if the folio is pinned, so
2385 * convert the pin on the source folio to a normal
2386 * reference.
2387 */
2388 pofs_clear_entry(pofs, i);
2389 folio_get(folio);
2390 gup_put_folio(folio, 1, FOLL_PIN);
2391
2392 if (migrate_device_coherent_folio(folio)) {
2393 ret = -EBUSY;
2394 goto err;
2395 }
2396
2397 continue;
2398 }
2399
2400 /*
2401 * We can't migrate folios with unexpected references, so drop
2402 * the reference obtained by __get_user_pages_locked().
2403 * Migrating folios have been added to movable_folio_list after
2404 * calling folio_isolate_lru() which takes a reference so the
2405 * folio won't be freed if it's migrating.
2406 */
2407 unpin_folio(folio);
2408 pofs_clear_entry(pofs, i);
2409 }
2410
2411 if (!list_empty(movable_folio_list)) {
2412 struct migration_target_control mtc = {
2413 .nid = NUMA_NO_NODE,
2414 .gfp_mask = GFP_USER | __GFP_NOWARN,
2415 .reason = MR_LONGTERM_PIN,
2416 };
2417
2418 if (migrate_pages(movable_folio_list, alloc_migration_target,
2419 NULL, (unsigned long)&mtc, MIGRATE_SYNC,
2420 MR_LONGTERM_PIN, NULL)) {
2421 ret = -ENOMEM;
2422 goto err;
2423 }
2424 }
2425
2426 putback_movable_pages(movable_folio_list);
2427
2428 return -EAGAIN;
2429
2430err:
2431 pofs_unpin(pofs);
2432 putback_movable_pages(movable_folio_list);
2433
2434 return ret;
2435}
2436
2437static long
2438check_and_migrate_movable_pages_or_folios(struct pages_or_folios *pofs)
2439{
2440 LIST_HEAD(movable_folio_list);
2441
2442 collect_longterm_unpinnable_folios(&movable_folio_list, pofs);
2443 if (list_empty(&movable_folio_list))
2444 return 0;
2445
2446 return migrate_longterm_unpinnable_folios(&movable_folio_list, pofs);
2447}
2448
2449/*
2450 * Check whether all folios are *allowed* to be pinned indefinitely (long term).
2451 * Rather confusingly, all folios in the range are required to be pinned via
2452 * FOLL_PIN, before calling this routine.
2453 *
2454 * Return values:
2455 *
2456 * 0: if everything is OK and all folios in the range are allowed to be pinned,
2457 * then this routine leaves all folios pinned and returns zero for success.
2458 *
2459 * -EAGAIN: if any folios in the range are not allowed to be pinned, then this
2460 * routine will migrate those folios away, unpin all the folios in the range. If
2461 * migration of the entire set of folios succeeds, then -EAGAIN is returned. The
2462 * caller should re-pin the entire range with FOLL_PIN and then call this
2463 * routine again.
2464 *
2465 * -ENOMEM, or any other -errno: if an error *other* than -EAGAIN occurs, this
2466 * indicates a migration failure. The caller should give up, and propagate the
2467 * error back up the call stack. The caller does not need to unpin any folios in
2468 * that case, because this routine will do the unpinning.
2469 */
2470static long check_and_migrate_movable_folios(unsigned long nr_folios,
2471 struct folio **folios)
2472{
2473 struct pages_or_folios pofs = {
2474 .folios = folios,
2475 .has_folios = true,
2476 .nr_entries = nr_folios,
2477 };
2478
2479 return check_and_migrate_movable_pages_or_folios(&pofs);
2480}
2481
2482/*
2483 * Return values and behavior are the same as those for
2484 * check_and_migrate_movable_folios().
2485 */
2486static long check_and_migrate_movable_pages(unsigned long nr_pages,
2487 struct page **pages)
2488{
2489 struct pages_or_folios pofs = {
2490 .pages = pages,
2491 .has_folios = false,
2492 .nr_entries = nr_pages,
2493 };
2494
2495 return check_and_migrate_movable_pages_or_folios(&pofs);
2496}
2497#else
2498static long check_and_migrate_movable_pages(unsigned long nr_pages,
2499 struct page **pages)
2500{
2501 return 0;
2502}
2503
2504static long check_and_migrate_movable_folios(unsigned long nr_folios,
2505 struct folio **folios)
2506{
2507 return 0;
2508}
2509#endif /* CONFIG_MIGRATION */
2510
2511/*
2512 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
2513 * allows us to process the FOLL_LONGTERM flag.
2514 */
2515static long __gup_longterm_locked(struct mm_struct *mm,
2516 unsigned long start,
2517 unsigned long nr_pages,
2518 struct page **pages,
2519 int *locked,
2520 unsigned int gup_flags)
2521{
2522 unsigned int flags;
2523 long rc, nr_pinned_pages;
2524
2525 if (!(gup_flags & FOLL_LONGTERM))
2526 return __get_user_pages_locked(mm, start, nr_pages, pages,
2527 locked, gup_flags);
2528
2529 flags = memalloc_pin_save();
2530 do {
2531 nr_pinned_pages = __get_user_pages_locked(mm, start, nr_pages,
2532 pages, locked,
2533 gup_flags);
2534 if (nr_pinned_pages <= 0) {
2535 rc = nr_pinned_pages;
2536 break;
2537 }
2538
2539 /* FOLL_LONGTERM implies FOLL_PIN */
2540 rc = check_and_migrate_movable_pages(nr_pinned_pages, pages);
2541 } while (rc == -EAGAIN);
2542 memalloc_pin_restore(flags);
2543 return rc ? rc : nr_pinned_pages;
2544}
2545
2546/*
2547 * Check that the given flags are valid for the exported gup/pup interface, and
2548 * update them with the required flags that the caller must have set.
2549 */
2550static bool is_valid_gup_args(struct page **pages, int *locked,
2551 unsigned int *gup_flags_p, unsigned int to_set)
2552{
2553 unsigned int gup_flags = *gup_flags_p;
2554
2555 /*
2556 * These flags not allowed to be specified externally to the gup
2557 * interfaces:
2558 * - FOLL_TOUCH/FOLL_PIN/FOLL_TRIED/FOLL_FAST_ONLY are internal only
2559 * - FOLL_REMOTE is internal only, set in (get|pin)_user_pages_remote()
2560 * - FOLL_UNLOCKABLE is internal only and used if locked is !NULL
2561 */
2562 if (WARN_ON_ONCE(gup_flags & INTERNAL_GUP_FLAGS))
2563 return false;
2564
2565 gup_flags |= to_set;
2566 if (locked) {
2567 /* At the external interface locked must be set */
2568 if (WARN_ON_ONCE(*locked != 1))
2569 return false;
2570
2571 gup_flags |= FOLL_UNLOCKABLE;
2572 }
2573
2574 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2575 if (WARN_ON_ONCE((gup_flags & (FOLL_PIN | FOLL_GET)) ==
2576 (FOLL_PIN | FOLL_GET)))
2577 return false;
2578
2579 /* LONGTERM can only be specified when pinning */
2580 if (WARN_ON_ONCE(!(gup_flags & FOLL_PIN) && (gup_flags & FOLL_LONGTERM)))
2581 return false;
2582
2583 /* Pages input must be given if using GET/PIN */
2584 if (WARN_ON_ONCE((gup_flags & (FOLL_GET | FOLL_PIN)) && !pages))
2585 return false;
2586
2587 /* We want to allow the pgmap to be hot-unplugged at all times */
2588 if (WARN_ON_ONCE((gup_flags & FOLL_LONGTERM) &&
2589 (gup_flags & FOLL_PCI_P2PDMA)))
2590 return false;
2591
2592 *gup_flags_p = gup_flags;
2593 return true;
2594}
2595
2596#ifdef CONFIG_MMU
2597/**
2598 * get_user_pages_remote() - pin user pages in memory
2599 * @mm: mm_struct of target mm
2600 * @start: starting user address
2601 * @nr_pages: number of pages from start to pin
2602 * @gup_flags: flags modifying lookup behaviour
2603 * @pages: array that receives pointers to the pages pinned.
2604 * Should be at least nr_pages long. Or NULL, if caller
2605 * only intends to ensure the pages are faulted in.
2606 * @locked: pointer to lock flag indicating whether lock is held and
2607 * subsequently whether VM_FAULT_RETRY functionality can be
2608 * utilised. Lock must initially be held.
2609 *
2610 * Returns either number of pages pinned (which may be less than the
2611 * number requested), or an error. Details about the return value:
2612 *
2613 * -- If nr_pages is 0, returns 0.
2614 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
2615 * -- If nr_pages is >0, and some pages were pinned, returns the number of
2616 * pages pinned. Again, this may be less than nr_pages.
2617 *
2618 * The caller is responsible for releasing returned @pages, via put_page().
2619 *
2620 * Must be called with mmap_lock held for read or write.
2621 *
2622 * get_user_pages_remote walks a process's page tables and takes a reference
2623 * to each struct page that each user address corresponds to at a given
2624 * instant. That is, it takes the page that would be accessed if a user
2625 * thread accesses the given user virtual address at that instant.
2626 *
2627 * This does not guarantee that the page exists in the user mappings when
2628 * get_user_pages_remote returns, and there may even be a completely different
2629 * page there in some cases (eg. if mmapped pagecache has been invalidated
2630 * and subsequently re-faulted). However it does guarantee that the page
2631 * won't be freed completely. And mostly callers simply care that the page
2632 * contains data that was valid *at some point in time*. Typically, an IO
2633 * or similar operation cannot guarantee anything stronger anyway because
2634 * locks can't be held over the syscall boundary.
2635 *
2636 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
2637 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
2638 * be called after the page is finished with, and before put_page is called.
2639 *
2640 * get_user_pages_remote is typically used for fewer-copy IO operations,
2641 * to get a handle on the memory by some means other than accesses
2642 * via the user virtual addresses. The pages may be submitted for
2643 * DMA to devices or accessed via their kernel linear mapping (via the
2644 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
2645 *
2646 * See also get_user_pages_fast, for performance critical applications.
2647 *
2648 * get_user_pages_remote should be phased out in favor of
2649 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
2650 * should use get_user_pages_remote because it cannot pass
2651 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
2652 */
2653long get_user_pages_remote(struct mm_struct *mm,
2654 unsigned long start, unsigned long nr_pages,
2655 unsigned int gup_flags, struct page **pages,
2656 int *locked)
2657{
2658 int local_locked = 1;
2659
2660 if (!is_valid_gup_args(pages, locked, &gup_flags,
2661 FOLL_TOUCH | FOLL_REMOTE))
2662 return -EINVAL;
2663
2664 return __get_user_pages_locked(mm, start, nr_pages, pages,
2665 locked ? locked : &local_locked,
2666 gup_flags);
2667}
2668EXPORT_SYMBOL(get_user_pages_remote);
2669
2670#else /* CONFIG_MMU */
2671long get_user_pages_remote(struct mm_struct *mm,
2672 unsigned long start, unsigned long nr_pages,
2673 unsigned int gup_flags, struct page **pages,
2674 int *locked)
2675{
2676 return 0;
2677}
2678#endif /* !CONFIG_MMU */
2679
2680/**
2681 * get_user_pages() - pin user pages in memory
2682 * @start: starting user address
2683 * @nr_pages: number of pages from start to pin
2684 * @gup_flags: flags modifying lookup behaviour
2685 * @pages: array that receives pointers to the pages pinned.
2686 * Should be at least nr_pages long. Or NULL, if caller
2687 * only intends to ensure the pages are faulted in.
2688 *
2689 * This is the same as get_user_pages_remote(), just with a less-flexible
2690 * calling convention where we assume that the mm being operated on belongs to
2691 * the current task, and doesn't allow passing of a locked parameter. We also
2692 * obviously don't pass FOLL_REMOTE in here.
2693 */
2694long get_user_pages(unsigned long start, unsigned long nr_pages,
2695 unsigned int gup_flags, struct page **pages)
2696{
2697 int locked = 1;
2698
2699 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_TOUCH))
2700 return -EINVAL;
2701
2702 return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2703 &locked, gup_flags);
2704}
2705EXPORT_SYMBOL(get_user_pages);
2706
2707/*
2708 * get_user_pages_unlocked() is suitable to replace the form:
2709 *
2710 * mmap_read_lock(mm);
2711 * get_user_pages(mm, ..., pages, NULL);
2712 * mmap_read_unlock(mm);
2713 *
2714 * with:
2715 *
2716 * get_user_pages_unlocked(mm, ..., pages);
2717 *
2718 * It is functionally equivalent to get_user_pages_fast so
2719 * get_user_pages_fast should be used instead if specific gup_flags
2720 * (e.g. FOLL_FORCE) are not required.
2721 */
2722long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2723 struct page **pages, unsigned int gup_flags)
2724{
2725 int locked = 0;
2726
2727 if (!is_valid_gup_args(pages, NULL, &gup_flags,
2728 FOLL_TOUCH | FOLL_UNLOCKABLE))
2729 return -EINVAL;
2730
2731 return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2732 &locked, gup_flags);
2733}
2734EXPORT_SYMBOL(get_user_pages_unlocked);
2735
2736/*
2737 * GUP-fast
2738 *
2739 * get_user_pages_fast attempts to pin user pages by walking the page
2740 * tables directly and avoids taking locks. Thus the walker needs to be
2741 * protected from page table pages being freed from under it, and should
2742 * block any THP splits.
2743 *
2744 * One way to achieve this is to have the walker disable interrupts, and
2745 * rely on IPIs from the TLB flushing code blocking before the page table
2746 * pages are freed. This is unsuitable for architectures that do not need
2747 * to broadcast an IPI when invalidating TLBs.
2748 *
2749 * Another way to achieve this is to batch up page table containing pages
2750 * belonging to more than one mm_user, then rcu_sched a callback to free those
2751 * pages. Disabling interrupts will allow the gup_fast() walker to both block
2752 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2753 * (which is a relatively rare event). The code below adopts this strategy.
2754 *
2755 * Before activating this code, please be aware that the following assumptions
2756 * are currently made:
2757 *
2758 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2759 * free pages containing page tables or TLB flushing requires IPI broadcast.
2760 *
2761 * *) ptes can be read atomically by the architecture.
2762 *
2763 * *) access_ok is sufficient to validate userspace address ranges.
2764 *
2765 * The last two assumptions can be relaxed by the addition of helper functions.
2766 *
2767 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2768 */
2769#ifdef CONFIG_HAVE_GUP_FAST
2770/*
2771 * Used in the GUP-fast path to determine whether GUP is permitted to work on
2772 * a specific folio.
2773 *
2774 * This call assumes the caller has pinned the folio, that the lowest page table
2775 * level still points to this folio, and that interrupts have been disabled.
2776 *
2777 * GUP-fast must reject all secretmem folios.
2778 *
2779 * Writing to pinned file-backed dirty tracked folios is inherently problematic
2780 * (see comment describing the writable_file_mapping_allowed() function). We
2781 * therefore try to avoid the most egregious case of a long-term mapping doing
2782 * so.
2783 *
2784 * This function cannot be as thorough as that one as the VMA is not available
2785 * in the fast path, so instead we whitelist known good cases and if in doubt,
2786 * fall back to the slow path.
2787 */
2788static bool gup_fast_folio_allowed(struct folio *folio, unsigned int flags)
2789{
2790 bool reject_file_backed = false;
2791 struct address_space *mapping;
2792 bool check_secretmem = false;
2793 unsigned long mapping_flags;
2794
2795 /*
2796 * If we aren't pinning then no problematic write can occur. A long term
2797 * pin is the most egregious case so this is the one we disallow.
2798 */
2799 if ((flags & (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE)) ==
2800 (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE))
2801 reject_file_backed = true;
2802
2803 /* We hold a folio reference, so we can safely access folio fields. */
2804
2805 /* secretmem folios are always order-0 folios. */
2806 if (IS_ENABLED(CONFIG_SECRETMEM) && !folio_test_large(folio))
2807 check_secretmem = true;
2808
2809 if (!reject_file_backed && !check_secretmem)
2810 return true;
2811
2812 if (WARN_ON_ONCE(folio_test_slab(folio)))
2813 return false;
2814
2815 /* hugetlb neither requires dirty-tracking nor can be secretmem. */
2816 if (folio_test_hugetlb(folio))
2817 return true;
2818
2819 /*
2820 * GUP-fast disables IRQs. When IRQS are disabled, RCU grace periods
2821 * cannot proceed, which means no actions performed under RCU can
2822 * proceed either.
2823 *
2824 * inodes and thus their mappings are freed under RCU, which means the
2825 * mapping cannot be freed beneath us and thus we can safely dereference
2826 * it.
2827 */
2828 lockdep_assert_irqs_disabled();
2829
2830 /*
2831 * However, there may be operations which _alter_ the mapping, so ensure
2832 * we read it once and only once.
2833 */
2834 mapping = READ_ONCE(folio->mapping);
2835
2836 /*
2837 * The mapping may have been truncated, in any case we cannot determine
2838 * if this mapping is safe - fall back to slow path to determine how to
2839 * proceed.
2840 */
2841 if (!mapping)
2842 return false;
2843
2844 /* Anonymous folios pose no problem. */
2845 mapping_flags = (unsigned long)mapping & PAGE_MAPPING_FLAGS;
2846 if (mapping_flags)
2847 return mapping_flags & PAGE_MAPPING_ANON;
2848
2849 /*
2850 * At this point, we know the mapping is non-null and points to an
2851 * address_space object.
2852 */
2853 if (check_secretmem && secretmem_mapping(mapping))
2854 return false;
2855 /* The only remaining allowed file system is shmem. */
2856 return !reject_file_backed || shmem_mapping(mapping);
2857}
2858
2859static void __maybe_unused gup_fast_undo_dev_pagemap(int *nr, int nr_start,
2860 unsigned int flags, struct page **pages)
2861{
2862 while ((*nr) - nr_start) {
2863 struct folio *folio = page_folio(pages[--(*nr)]);
2864
2865 folio_clear_referenced(folio);
2866 gup_put_folio(folio, 1, flags);
2867 }
2868}
2869
2870#ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2871/*
2872 * GUP-fast relies on pte change detection to avoid concurrent pgtable
2873 * operations.
2874 *
2875 * To pin the page, GUP-fast needs to do below in order:
2876 * (1) pin the page (by prefetching pte), then (2) check pte not changed.
2877 *
2878 * For the rest of pgtable operations where pgtable updates can be racy
2879 * with GUP-fast, we need to do (1) clear pte, then (2) check whether page
2880 * is pinned.
2881 *
2882 * Above will work for all pte-level operations, including THP split.
2883 *
2884 * For THP collapse, it's a bit more complicated because GUP-fast may be
2885 * walking a pgtable page that is being freed (pte is still valid but pmd
2886 * can be cleared already). To avoid race in such condition, we need to
2887 * also check pmd here to make sure pmd doesn't change (corresponds to
2888 * pmdp_collapse_flush() in the THP collapse code path).
2889 */
2890static int gup_fast_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2891 unsigned long end, unsigned int flags, struct page **pages,
2892 int *nr)
2893{
2894 struct dev_pagemap *pgmap = NULL;
2895 int nr_start = *nr, ret = 0;
2896 pte_t *ptep, *ptem;
2897
2898 ptem = ptep = pte_offset_map(&pmd, addr);
2899 if (!ptep)
2900 return 0;
2901 do {
2902 pte_t pte = ptep_get_lockless(ptep);
2903 struct page *page;
2904 struct folio *folio;
2905
2906 /*
2907 * Always fallback to ordinary GUP on PROT_NONE-mapped pages:
2908 * pte_access_permitted() better should reject these pages
2909 * either way: otherwise, GUP-fast might succeed in
2910 * cases where ordinary GUP would fail due to VMA access
2911 * permissions.
2912 */
2913 if (pte_protnone(pte))
2914 goto pte_unmap;
2915
2916 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2917 goto pte_unmap;
2918
2919 if (pte_devmap(pte)) {
2920 if (unlikely(flags & FOLL_LONGTERM))
2921 goto pte_unmap;
2922
2923 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2924 if (unlikely(!pgmap)) {
2925 gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
2926 goto pte_unmap;
2927 }
2928 } else if (pte_special(pte))
2929 goto pte_unmap;
2930
2931 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2932 page = pte_page(pte);
2933
2934 folio = try_grab_folio_fast(page, 1, flags);
2935 if (!folio)
2936 goto pte_unmap;
2937
2938 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp)) ||
2939 unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) {
2940 gup_put_folio(folio, 1, flags);
2941 goto pte_unmap;
2942 }
2943
2944 if (!gup_fast_folio_allowed(folio, flags)) {
2945 gup_put_folio(folio, 1, flags);
2946 goto pte_unmap;
2947 }
2948
2949 if (!pte_write(pte) && gup_must_unshare(NULL, flags, page)) {
2950 gup_put_folio(folio, 1, flags);
2951 goto pte_unmap;
2952 }
2953
2954 /*
2955 * We need to make the page accessible if and only if we are
2956 * going to access its content (the FOLL_PIN case). Please
2957 * see Documentation/core-api/pin_user_pages.rst for
2958 * details.
2959 */
2960 if (flags & FOLL_PIN) {
2961 ret = arch_make_folio_accessible(folio);
2962 if (ret) {
2963 gup_put_folio(folio, 1, flags);
2964 goto pte_unmap;
2965 }
2966 }
2967 folio_set_referenced(folio);
2968 pages[*nr] = page;
2969 (*nr)++;
2970 } while (ptep++, addr += PAGE_SIZE, addr != end);
2971
2972 ret = 1;
2973
2974pte_unmap:
2975 if (pgmap)
2976 put_dev_pagemap(pgmap);
2977 pte_unmap(ptem);
2978 return ret;
2979}
2980#else
2981
2982/*
2983 * If we can't determine whether or not a pte is special, then fail immediately
2984 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2985 * to be special.
2986 *
2987 * For a futex to be placed on a THP tail page, get_futex_key requires a
2988 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2989 * useful to have gup_fast_pmd_leaf even if we can't operate on ptes.
2990 */
2991static int gup_fast_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2992 unsigned long end, unsigned int flags, struct page **pages,
2993 int *nr)
2994{
2995 return 0;
2996}
2997#endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2998
2999#if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
3000static int gup_fast_devmap_leaf(unsigned long pfn, unsigned long addr,
3001 unsigned long end, unsigned int flags, struct page **pages, int *nr)
3002{
3003 int nr_start = *nr;
3004 struct dev_pagemap *pgmap = NULL;
3005
3006 do {
3007 struct folio *folio;
3008 struct page *page = pfn_to_page(pfn);
3009
3010 pgmap = get_dev_pagemap(pfn, pgmap);
3011 if (unlikely(!pgmap)) {
3012 gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
3013 break;
3014 }
3015
3016 if (!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)) {
3017 gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
3018 break;
3019 }
3020
3021 folio = try_grab_folio_fast(page, 1, flags);
3022 if (!folio) {
3023 gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
3024 break;
3025 }
3026 folio_set_referenced(folio);
3027 pages[*nr] = page;
3028 (*nr)++;
3029 pfn++;
3030 } while (addr += PAGE_SIZE, addr != end);
3031
3032 put_dev_pagemap(pgmap);
3033 return addr == end;
3034}
3035
3036static int gup_fast_devmap_pmd_leaf(pmd_t orig, pmd_t *pmdp, unsigned long addr,
3037 unsigned long end, unsigned int flags, struct page **pages,
3038 int *nr)
3039{
3040 unsigned long fault_pfn;
3041 int nr_start = *nr;
3042
3043 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
3044 if (!gup_fast_devmap_leaf(fault_pfn, addr, end, flags, pages, nr))
3045 return 0;
3046
3047 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
3048 gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
3049 return 0;
3050 }
3051 return 1;
3052}
3053
3054static int gup_fast_devmap_pud_leaf(pud_t orig, pud_t *pudp, unsigned long addr,
3055 unsigned long end, unsigned int flags, struct page **pages,
3056 int *nr)
3057{
3058 unsigned long fault_pfn;
3059 int nr_start = *nr;
3060
3061 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
3062 if (!gup_fast_devmap_leaf(fault_pfn, addr, end, flags, pages, nr))
3063 return 0;
3064
3065 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
3066 gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
3067 return 0;
3068 }
3069 return 1;
3070}
3071#else
3072static int gup_fast_devmap_pmd_leaf(pmd_t orig, pmd_t *pmdp, unsigned long addr,
3073 unsigned long end, unsigned int flags, struct page **pages,
3074 int *nr)
3075{
3076 BUILD_BUG();
3077 return 0;
3078}
3079
3080static int gup_fast_devmap_pud_leaf(pud_t pud, pud_t *pudp, unsigned long addr,
3081 unsigned long end, unsigned int flags, struct page **pages,
3082 int *nr)
3083{
3084 BUILD_BUG();
3085 return 0;
3086}
3087#endif
3088
3089static int gup_fast_pmd_leaf(pmd_t orig, pmd_t *pmdp, unsigned long addr,
3090 unsigned long end, unsigned int flags, struct page **pages,
3091 int *nr)
3092{
3093 struct page *page;
3094 struct folio *folio;
3095 int refs;
3096
3097 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
3098 return 0;
3099
3100 if (pmd_special(orig))
3101 return 0;
3102
3103 if (pmd_devmap(orig)) {
3104 if (unlikely(flags & FOLL_LONGTERM))
3105 return 0;
3106 return gup_fast_devmap_pmd_leaf(orig, pmdp, addr, end, flags,
3107 pages, nr);
3108 }
3109
3110 page = pmd_page(orig);
3111 refs = record_subpages(page, PMD_SIZE, addr, end, pages + *nr);
3112
3113 folio = try_grab_folio_fast(page, refs, flags);
3114 if (!folio)
3115 return 0;
3116
3117 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
3118 gup_put_folio(folio, refs, flags);
3119 return 0;
3120 }
3121
3122 if (!gup_fast_folio_allowed(folio, flags)) {
3123 gup_put_folio(folio, refs, flags);
3124 return 0;
3125 }
3126 if (!pmd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
3127 gup_put_folio(folio, refs, flags);
3128 return 0;
3129 }
3130
3131 *nr += refs;
3132 folio_set_referenced(folio);
3133 return 1;
3134}
3135
3136static int gup_fast_pud_leaf(pud_t orig, pud_t *pudp, unsigned long addr,
3137 unsigned long end, unsigned int flags, struct page **pages,
3138 int *nr)
3139{
3140 struct page *page;
3141 struct folio *folio;
3142 int refs;
3143
3144 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
3145 return 0;
3146
3147 if (pud_special(orig))
3148 return 0;
3149
3150 if (pud_devmap(orig)) {
3151 if (unlikely(flags & FOLL_LONGTERM))
3152 return 0;
3153 return gup_fast_devmap_pud_leaf(orig, pudp, addr, end, flags,
3154 pages, nr);
3155 }
3156
3157 page = pud_page(orig);
3158 refs = record_subpages(page, PUD_SIZE, addr, end, pages + *nr);
3159
3160 folio = try_grab_folio_fast(page, refs, flags);
3161 if (!folio)
3162 return 0;
3163
3164 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
3165 gup_put_folio(folio, refs, flags);
3166 return 0;
3167 }
3168
3169 if (!gup_fast_folio_allowed(folio, flags)) {
3170 gup_put_folio(folio, refs, flags);
3171 return 0;
3172 }
3173
3174 if (!pud_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
3175 gup_put_folio(folio, refs, flags);
3176 return 0;
3177 }
3178
3179 *nr += refs;
3180 folio_set_referenced(folio);
3181 return 1;
3182}
3183
3184static int gup_fast_pgd_leaf(pgd_t orig, pgd_t *pgdp, unsigned long addr,
3185 unsigned long end, unsigned int flags, struct page **pages,
3186 int *nr)
3187{
3188 int refs;
3189 struct page *page;
3190 struct folio *folio;
3191
3192 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
3193 return 0;
3194
3195 BUILD_BUG_ON(pgd_devmap(orig));
3196
3197 page = pgd_page(orig);
3198 refs = record_subpages(page, PGDIR_SIZE, addr, end, pages + *nr);
3199
3200 folio = try_grab_folio_fast(page, refs, flags);
3201 if (!folio)
3202 return 0;
3203
3204 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
3205 gup_put_folio(folio, refs, flags);
3206 return 0;
3207 }
3208
3209 if (!pgd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
3210 gup_put_folio(folio, refs, flags);
3211 return 0;
3212 }
3213
3214 if (!gup_fast_folio_allowed(folio, flags)) {
3215 gup_put_folio(folio, refs, flags);
3216 return 0;
3217 }
3218
3219 *nr += refs;
3220 folio_set_referenced(folio);
3221 return 1;
3222}
3223
3224static int gup_fast_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr,
3225 unsigned long end, unsigned int flags, struct page **pages,
3226 int *nr)
3227{
3228 unsigned long next;
3229 pmd_t *pmdp;
3230
3231 pmdp = pmd_offset_lockless(pudp, pud, addr);
3232 do {
3233 pmd_t pmd = pmdp_get_lockless(pmdp);
3234
3235 next = pmd_addr_end(addr, end);
3236 if (!pmd_present(pmd))
3237 return 0;
3238
3239 if (unlikely(pmd_leaf(pmd))) {
3240 /* See gup_fast_pte_range() */
3241 if (pmd_protnone(pmd))
3242 return 0;
3243
3244 if (!gup_fast_pmd_leaf(pmd, pmdp, addr, next, flags,
3245 pages, nr))
3246 return 0;
3247
3248 } else if (!gup_fast_pte_range(pmd, pmdp, addr, next, flags,
3249 pages, nr))
3250 return 0;
3251 } while (pmdp++, addr = next, addr != end);
3252
3253 return 1;
3254}
3255
3256static int gup_fast_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr,
3257 unsigned long end, unsigned int flags, struct page **pages,
3258 int *nr)
3259{
3260 unsigned long next;
3261 pud_t *pudp;
3262
3263 pudp = pud_offset_lockless(p4dp, p4d, addr);
3264 do {
3265 pud_t pud = READ_ONCE(*pudp);
3266
3267 next = pud_addr_end(addr, end);
3268 if (unlikely(!pud_present(pud)))
3269 return 0;
3270 if (unlikely(pud_leaf(pud))) {
3271 if (!gup_fast_pud_leaf(pud, pudp, addr, next, flags,
3272 pages, nr))
3273 return 0;
3274 } else if (!gup_fast_pmd_range(pudp, pud, addr, next, flags,
3275 pages, nr))
3276 return 0;
3277 } while (pudp++, addr = next, addr != end);
3278
3279 return 1;
3280}
3281
3282static int gup_fast_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr,
3283 unsigned long end, unsigned int flags, struct page **pages,
3284 int *nr)
3285{
3286 unsigned long next;
3287 p4d_t *p4dp;
3288
3289 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
3290 do {
3291 p4d_t p4d = READ_ONCE(*p4dp);
3292
3293 next = p4d_addr_end(addr, end);
3294 if (!p4d_present(p4d))
3295 return 0;
3296 BUILD_BUG_ON(p4d_leaf(p4d));
3297 if (!gup_fast_pud_range(p4dp, p4d, addr, next, flags,
3298 pages, nr))
3299 return 0;
3300 } while (p4dp++, addr = next, addr != end);
3301
3302 return 1;
3303}
3304
3305static void gup_fast_pgd_range(unsigned long addr, unsigned long end,
3306 unsigned int flags, struct page **pages, int *nr)
3307{
3308 unsigned long next;
3309 pgd_t *pgdp;
3310
3311 pgdp = pgd_offset(current->mm, addr);
3312 do {
3313 pgd_t pgd = READ_ONCE(*pgdp);
3314
3315 next = pgd_addr_end(addr, end);
3316 if (pgd_none(pgd))
3317 return;
3318 if (unlikely(pgd_leaf(pgd))) {
3319 if (!gup_fast_pgd_leaf(pgd, pgdp, addr, next, flags,
3320 pages, nr))
3321 return;
3322 } else if (!gup_fast_p4d_range(pgdp, pgd, addr, next, flags,
3323 pages, nr))
3324 return;
3325 } while (pgdp++, addr = next, addr != end);
3326}
3327#else
3328static inline void gup_fast_pgd_range(unsigned long addr, unsigned long end,
3329 unsigned int flags, struct page **pages, int *nr)
3330{
3331}
3332#endif /* CONFIG_HAVE_GUP_FAST */
3333
3334#ifndef gup_fast_permitted
3335/*
3336 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
3337 * we need to fall back to the slow version:
3338 */
3339static bool gup_fast_permitted(unsigned long start, unsigned long end)
3340{
3341 return true;
3342}
3343#endif
3344
3345static unsigned long gup_fast(unsigned long start, unsigned long end,
3346 unsigned int gup_flags, struct page **pages)
3347{
3348 unsigned long flags;
3349 int nr_pinned = 0;
3350 unsigned seq;
3351
3352 if (!IS_ENABLED(CONFIG_HAVE_GUP_FAST) ||
3353 !gup_fast_permitted(start, end))
3354 return 0;
3355
3356 if (gup_flags & FOLL_PIN) {
3357 seq = raw_read_seqcount(¤t->mm->write_protect_seq);
3358 if (seq & 1)
3359 return 0;
3360 }
3361
3362 /*
3363 * Disable interrupts. The nested form is used, in order to allow full,
3364 * general purpose use of this routine.
3365 *
3366 * With interrupts disabled, we block page table pages from being freed
3367 * from under us. See struct mmu_table_batch comments in
3368 * include/asm-generic/tlb.h for more details.
3369 *
3370 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
3371 * that come from THPs splitting.
3372 */
3373 local_irq_save(flags);
3374 gup_fast_pgd_range(start, end, gup_flags, pages, &nr_pinned);
3375 local_irq_restore(flags);
3376
3377 /*
3378 * When pinning pages for DMA there could be a concurrent write protect
3379 * from fork() via copy_page_range(), in this case always fail GUP-fast.
3380 */
3381 if (gup_flags & FOLL_PIN) {
3382 if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) {
3383 gup_fast_unpin_user_pages(pages, nr_pinned);
3384 return 0;
3385 } else {
3386 sanity_check_pinned_pages(pages, nr_pinned);
3387 }
3388 }
3389 return nr_pinned;
3390}
3391
3392static int gup_fast_fallback(unsigned long start, unsigned long nr_pages,
3393 unsigned int gup_flags, struct page **pages)
3394{
3395 unsigned long len, end;
3396 unsigned long nr_pinned;
3397 int locked = 0;
3398 int ret;
3399
3400 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
3401 FOLL_FORCE | FOLL_PIN | FOLL_GET |
3402 FOLL_FAST_ONLY | FOLL_NOFAULT |
3403 FOLL_PCI_P2PDMA | FOLL_HONOR_NUMA_FAULT)))
3404 return -EINVAL;
3405
3406 if (gup_flags & FOLL_PIN)
3407 mm_set_has_pinned_flag(¤t->mm->flags);
3408
3409 if (!(gup_flags & FOLL_FAST_ONLY))
3410 might_lock_read(¤t->mm->mmap_lock);
3411
3412 start = untagged_addr(start) & PAGE_MASK;
3413 len = nr_pages << PAGE_SHIFT;
3414 if (check_add_overflow(start, len, &end))
3415 return -EOVERFLOW;
3416 if (end > TASK_SIZE_MAX)
3417 return -EFAULT;
3418 if (unlikely(!access_ok((void __user *)start, len)))
3419 return -EFAULT;
3420
3421 nr_pinned = gup_fast(start, end, gup_flags, pages);
3422 if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
3423 return nr_pinned;
3424
3425 /* Slow path: try to get the remaining pages with get_user_pages */
3426 start += nr_pinned << PAGE_SHIFT;
3427 pages += nr_pinned;
3428 ret = __gup_longterm_locked(current->mm, start, nr_pages - nr_pinned,
3429 pages, &locked,
3430 gup_flags | FOLL_TOUCH | FOLL_UNLOCKABLE);
3431 if (ret < 0) {
3432 /*
3433 * The caller has to unpin the pages we already pinned so
3434 * returning -errno is not an option
3435 */
3436 if (nr_pinned)
3437 return nr_pinned;
3438 return ret;
3439 }
3440 return ret + nr_pinned;
3441}
3442
3443/**
3444 * get_user_pages_fast_only() - pin user pages in memory
3445 * @start: starting user address
3446 * @nr_pages: number of pages from start to pin
3447 * @gup_flags: flags modifying pin behaviour
3448 * @pages: array that receives pointers to the pages pinned.
3449 * Should be at least nr_pages long.
3450 *
3451 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
3452 * the regular GUP.
3453 *
3454 * If the architecture does not support this function, simply return with no
3455 * pages pinned.
3456 *
3457 * Careful, careful! COW breaking can go either way, so a non-write
3458 * access can get ambiguous page results. If you call this function without
3459 * 'write' set, you'd better be sure that you're ok with that ambiguity.
3460 */
3461int get_user_pages_fast_only(unsigned long start, int nr_pages,
3462 unsigned int gup_flags, struct page **pages)
3463{
3464 /*
3465 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
3466 * because gup fast is always a "pin with a +1 page refcount" request.
3467 *
3468 * FOLL_FAST_ONLY is required in order to match the API description of
3469 * this routine: no fall back to regular ("slow") GUP.
3470 */
3471 if (!is_valid_gup_args(pages, NULL, &gup_flags,
3472 FOLL_GET | FOLL_FAST_ONLY))
3473 return -EINVAL;
3474
3475 return gup_fast_fallback(start, nr_pages, gup_flags, pages);
3476}
3477EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
3478
3479/**
3480 * get_user_pages_fast() - pin user pages in memory
3481 * @start: starting user address
3482 * @nr_pages: number of pages from start to pin
3483 * @gup_flags: flags modifying pin behaviour
3484 * @pages: array that receives pointers to the pages pinned.
3485 * Should be at least nr_pages long.
3486 *
3487 * Attempt to pin user pages in memory without taking mm->mmap_lock.
3488 * If not successful, it will fall back to taking the lock and
3489 * calling get_user_pages().
3490 *
3491 * Returns number of pages pinned. This may be fewer than the number requested.
3492 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
3493 * -errno.
3494 */
3495int get_user_pages_fast(unsigned long start, int nr_pages,
3496 unsigned int gup_flags, struct page **pages)
3497{
3498 /*
3499 * The caller may or may not have explicitly set FOLL_GET; either way is
3500 * OK. However, internally (within mm/gup.c), gup fast variants must set
3501 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
3502 * request.
3503 */
3504 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_GET))
3505 return -EINVAL;
3506 return gup_fast_fallback(start, nr_pages, gup_flags, pages);
3507}
3508EXPORT_SYMBOL_GPL(get_user_pages_fast);
3509
3510/**
3511 * pin_user_pages_fast() - pin user pages in memory without taking locks
3512 *
3513 * @start: starting user address
3514 * @nr_pages: number of pages from start to pin
3515 * @gup_flags: flags modifying pin behaviour
3516 * @pages: array that receives pointers to the pages pinned.
3517 * Should be at least nr_pages long.
3518 *
3519 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
3520 * get_user_pages_fast() for documentation on the function arguments, because
3521 * the arguments here are identical.
3522 *
3523 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3524 * see Documentation/core-api/pin_user_pages.rst for further details.
3525 *
3526 * Note that if a zero_page is amongst the returned pages, it will not have
3527 * pins in it and unpin_user_page() will not remove pins from it.
3528 */
3529int pin_user_pages_fast(unsigned long start, int nr_pages,
3530 unsigned int gup_flags, struct page **pages)
3531{
3532 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN))
3533 return -EINVAL;
3534 return gup_fast_fallback(start, nr_pages, gup_flags, pages);
3535}
3536EXPORT_SYMBOL_GPL(pin_user_pages_fast);
3537
3538/**
3539 * pin_user_pages_remote() - pin pages of a remote process
3540 *
3541 * @mm: mm_struct of target mm
3542 * @start: starting user address
3543 * @nr_pages: number of pages from start to pin
3544 * @gup_flags: flags modifying lookup behaviour
3545 * @pages: array that receives pointers to the pages pinned.
3546 * Should be at least nr_pages long.
3547 * @locked: pointer to lock flag indicating whether lock is held and
3548 * subsequently whether VM_FAULT_RETRY functionality can be
3549 * utilised. Lock must initially be held.
3550 *
3551 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
3552 * get_user_pages_remote() for documentation on the function arguments, because
3553 * the arguments here are identical.
3554 *
3555 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3556 * see Documentation/core-api/pin_user_pages.rst for details.
3557 *
3558 * Note that if a zero_page is amongst the returned pages, it will not have
3559 * pins in it and unpin_user_page*() will not remove pins from it.
3560 */
3561long pin_user_pages_remote(struct mm_struct *mm,
3562 unsigned long start, unsigned long nr_pages,
3563 unsigned int gup_flags, struct page **pages,
3564 int *locked)
3565{
3566 int local_locked = 1;
3567
3568 if (!is_valid_gup_args(pages, locked, &gup_flags,
3569 FOLL_PIN | FOLL_TOUCH | FOLL_REMOTE))
3570 return 0;
3571 return __gup_longterm_locked(mm, start, nr_pages, pages,
3572 locked ? locked : &local_locked,
3573 gup_flags);
3574}
3575EXPORT_SYMBOL(pin_user_pages_remote);
3576
3577/**
3578 * pin_user_pages() - pin user pages in memory for use by other devices
3579 *
3580 * @start: starting user address
3581 * @nr_pages: number of pages from start to pin
3582 * @gup_flags: flags modifying lookup behaviour
3583 * @pages: array that receives pointers to the pages pinned.
3584 * Should be at least nr_pages long.
3585 *
3586 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
3587 * FOLL_PIN is set.
3588 *
3589 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3590 * see Documentation/core-api/pin_user_pages.rst for details.
3591 *
3592 * Note that if a zero_page is amongst the returned pages, it will not have
3593 * pins in it and unpin_user_page*() will not remove pins from it.
3594 */
3595long pin_user_pages(unsigned long start, unsigned long nr_pages,
3596 unsigned int gup_flags, struct page **pages)
3597{
3598 int locked = 1;
3599
3600 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN))
3601 return 0;
3602 return __gup_longterm_locked(current->mm, start, nr_pages,
3603 pages, &locked, gup_flags);
3604}
3605EXPORT_SYMBOL(pin_user_pages);
3606
3607/*
3608 * pin_user_pages_unlocked() is the FOLL_PIN variant of
3609 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
3610 * FOLL_PIN and rejects FOLL_GET.
3611 *
3612 * Note that if a zero_page is amongst the returned pages, it will not have
3613 * pins in it and unpin_user_page*() will not remove pins from it.
3614 */
3615long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
3616 struct page **pages, unsigned int gup_flags)
3617{
3618 int locked = 0;
3619
3620 if (!is_valid_gup_args(pages, NULL, &gup_flags,
3621 FOLL_PIN | FOLL_TOUCH | FOLL_UNLOCKABLE))
3622 return 0;
3623
3624 return __gup_longterm_locked(current->mm, start, nr_pages, pages,
3625 &locked, gup_flags);
3626}
3627EXPORT_SYMBOL(pin_user_pages_unlocked);
3628
3629/**
3630 * memfd_pin_folios() - pin folios associated with a memfd
3631 * @memfd: the memfd whose folios are to be pinned
3632 * @start: the first memfd offset
3633 * @end: the last memfd offset (inclusive)
3634 * @folios: array that receives pointers to the folios pinned
3635 * @max_folios: maximum number of entries in @folios
3636 * @offset: the offset into the first folio
3637 *
3638 * Attempt to pin folios associated with a memfd in the contiguous range
3639 * [start, end]. Given that a memfd is either backed by shmem or hugetlb,
3640 * the folios can either be found in the page cache or need to be allocated
3641 * if necessary. Once the folios are located, they are all pinned via
3642 * FOLL_PIN and @offset is populatedwith the offset into the first folio.
3643 * And, eventually, these pinned folios must be released either using
3644 * unpin_folios() or unpin_folio().
3645 *
3646 * It must be noted that the folios may be pinned for an indefinite amount
3647 * of time. And, in most cases, the duration of time they may stay pinned
3648 * would be controlled by the userspace. This behavior is effectively the
3649 * same as using FOLL_LONGTERM with other GUP APIs.
3650 *
3651 * Returns number of folios pinned, which could be less than @max_folios
3652 * as it depends on the folio sizes that cover the range [start, end].
3653 * If no folios were pinned, it returns -errno.
3654 */
3655long memfd_pin_folios(struct file *memfd, loff_t start, loff_t end,
3656 struct folio **folios, unsigned int max_folios,
3657 pgoff_t *offset)
3658{
3659 unsigned int flags, nr_folios, nr_found;
3660 unsigned int i, pgshift = PAGE_SHIFT;
3661 pgoff_t start_idx, end_idx, next_idx;
3662 struct folio *folio = NULL;
3663 struct folio_batch fbatch;
3664 struct hstate *h;
3665 long ret = -EINVAL;
3666
3667 if (start < 0 || start > end || !max_folios)
3668 return -EINVAL;
3669
3670 if (!memfd)
3671 return -EINVAL;
3672
3673 if (!shmem_file(memfd) && !is_file_hugepages(memfd))
3674 return -EINVAL;
3675
3676 if (end >= i_size_read(file_inode(memfd)))
3677 return -EINVAL;
3678
3679 if (is_file_hugepages(memfd)) {
3680 h = hstate_file(memfd);
3681 pgshift = huge_page_shift(h);
3682 }
3683
3684 flags = memalloc_pin_save();
3685 do {
3686 nr_folios = 0;
3687 start_idx = start >> pgshift;
3688 end_idx = end >> pgshift;
3689 if (is_file_hugepages(memfd)) {
3690 start_idx <<= huge_page_order(h);
3691 end_idx <<= huge_page_order(h);
3692 }
3693
3694 folio_batch_init(&fbatch);
3695 while (start_idx <= end_idx && nr_folios < max_folios) {
3696 /*
3697 * In most cases, we should be able to find the folios
3698 * in the page cache. If we cannot find them for some
3699 * reason, we try to allocate them and add them to the
3700 * page cache.
3701 */
3702 nr_found = filemap_get_folios_contig(memfd->f_mapping,
3703 &start_idx,
3704 end_idx,
3705 &fbatch);
3706 if (folio) {
3707 folio_put(folio);
3708 folio = NULL;
3709 }
3710
3711 next_idx = 0;
3712 for (i = 0; i < nr_found; i++) {
3713 /*
3714 * As there can be multiple entries for a
3715 * given folio in the batch returned by
3716 * filemap_get_folios_contig(), the below
3717 * check is to ensure that we pin and return a
3718 * unique set of folios between start and end.
3719 */
3720 if (next_idx &&
3721 next_idx != folio_index(fbatch.folios[i]))
3722 continue;
3723
3724 folio = page_folio(&fbatch.folios[i]->page);
3725
3726 if (try_grab_folio(folio, 1, FOLL_PIN)) {
3727 folio_batch_release(&fbatch);
3728 ret = -EINVAL;
3729 goto err;
3730 }
3731
3732 if (nr_folios == 0)
3733 *offset = offset_in_folio(folio, start);
3734
3735 folios[nr_folios] = folio;
3736 next_idx = folio_next_index(folio);
3737 if (++nr_folios == max_folios)
3738 break;
3739 }
3740
3741 folio = NULL;
3742 folio_batch_release(&fbatch);
3743 if (!nr_found) {
3744 folio = memfd_alloc_folio(memfd, start_idx);
3745 if (IS_ERR(folio)) {
3746 ret = PTR_ERR(folio);
3747 if (ret != -EEXIST)
3748 goto err;
3749 folio = NULL;
3750 }
3751 }
3752 }
3753
3754 ret = check_and_migrate_movable_folios(nr_folios, folios);
3755 } while (ret == -EAGAIN);
3756
3757 memalloc_pin_restore(flags);
3758 return ret ? ret : nr_folios;
3759err:
3760 memalloc_pin_restore(flags);
3761 unpin_folios(folios, nr_folios);
3762
3763 return ret;
3764}
3765EXPORT_SYMBOL_GPL(memfd_pin_folios);
3766
3767/**
3768 * folio_add_pins() - add pins to an already-pinned folio
3769 * @folio: the folio to add more pins to
3770 * @pins: number of pins to add
3771 *
3772 * Try to add more pins to an already-pinned folio. The semantics
3773 * of the pin (e.g., FOLL_WRITE) follow any existing pin and cannot
3774 * be changed.
3775 *
3776 * This function is helpful when having obtained a pin on a large folio
3777 * using memfd_pin_folios(), but wanting to logically unpin parts
3778 * (e.g., individual pages) of the folio later, for example, using
3779 * unpin_user_page_range_dirty_lock().
3780 *
3781 * This is not the right interface to initially pin a folio.
3782 */
3783int folio_add_pins(struct folio *folio, unsigned int pins)
3784{
3785 VM_WARN_ON_ONCE(!folio_maybe_dma_pinned(folio));
3786
3787 return try_grab_folio(folio, pins, FOLL_PIN);
3788}
3789EXPORT_SYMBOL_GPL(folio_add_pins);
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 vma = gup_vma_lookup(mm, start);
1210 if (!vma && in_gate_area(mm, start)) {
1211 ret = get_gate_page(mm, start & PAGE_MASK,
1212 gup_flags, &vma,
1213 pages ? &page : NULL);
1214 if (ret)
1215 goto out;
1216 ctx.page_mask = 0;
1217 goto next_page;
1218 }
1219
1220 if (!vma) {
1221 ret = -EFAULT;
1222 goto out;
1223 }
1224 ret = check_vma_flags(vma, gup_flags);
1225 if (ret)
1226 goto out;
1227 }
1228retry:
1229 /*
1230 * If we have a pending SIGKILL, don't keep faulting pages and
1231 * potentially allocating memory.
1232 */
1233 if (fatal_signal_pending(current)) {
1234 ret = -EINTR;
1235 goto out;
1236 }
1237 cond_resched();
1238
1239 page = follow_page_mask(vma, start, foll_flags, &ctx);
1240 if (!page || PTR_ERR(page) == -EMLINK) {
1241 ret = faultin_page(vma, start, &foll_flags,
1242 PTR_ERR(page) == -EMLINK, locked);
1243 switch (ret) {
1244 case 0:
1245 goto retry;
1246 case -EBUSY:
1247 case -EAGAIN:
1248 ret = 0;
1249 fallthrough;
1250 case -EFAULT:
1251 case -ENOMEM:
1252 case -EHWPOISON:
1253 goto out;
1254 }
1255 BUG();
1256 } else if (PTR_ERR(page) == -EEXIST) {
1257 /*
1258 * Proper page table entry exists, but no corresponding
1259 * struct page. If the caller expects **pages to be
1260 * filled in, bail out now, because that can't be done
1261 * for this page.
1262 */
1263 if (pages) {
1264 ret = PTR_ERR(page);
1265 goto out;
1266 }
1267 } else if (IS_ERR(page)) {
1268 ret = PTR_ERR(page);
1269 goto out;
1270 }
1271next_page:
1272 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1273 if (page_increm > nr_pages)
1274 page_increm = nr_pages;
1275
1276 if (pages) {
1277 struct page *subpage;
1278 unsigned int j;
1279
1280 /*
1281 * This must be a large folio (and doesn't need to
1282 * be the whole folio; it can be part of it), do
1283 * the refcount work for all the subpages too.
1284 *
1285 * NOTE: here the page may not be the head page
1286 * e.g. when start addr is not thp-size aligned.
1287 * try_grab_folio() should have taken care of tail
1288 * pages.
1289 */
1290 if (page_increm > 1) {
1291 struct folio *folio;
1292
1293 /*
1294 * Since we already hold refcount on the
1295 * large folio, this should never fail.
1296 */
1297 folio = try_grab_folio(page, page_increm - 1,
1298 foll_flags);
1299 if (WARN_ON_ONCE(!folio)) {
1300 /*
1301 * Release the 1st page ref if the
1302 * folio is problematic, fail hard.
1303 */
1304 gup_put_folio(page_folio(page), 1,
1305 foll_flags);
1306 ret = -EFAULT;
1307 goto out;
1308 }
1309 }
1310
1311 for (j = 0; j < page_increm; j++) {
1312 subpage = nth_page(page, j);
1313 pages[i + j] = subpage;
1314 flush_anon_page(vma, subpage, start + j * PAGE_SIZE);
1315 flush_dcache_page(subpage);
1316 }
1317 }
1318
1319 i += page_increm;
1320 start += page_increm * PAGE_SIZE;
1321 nr_pages -= page_increm;
1322 } while (nr_pages);
1323out:
1324 if (ctx.pgmap)
1325 put_dev_pagemap(ctx.pgmap);
1326 return i ? i : ret;
1327}
1328
1329static bool vma_permits_fault(struct vm_area_struct *vma,
1330 unsigned int fault_flags)
1331{
1332 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1333 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1334 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1335
1336 if (!(vm_flags & vma->vm_flags))
1337 return false;
1338
1339 /*
1340 * The architecture might have a hardware protection
1341 * mechanism other than read/write that can deny access.
1342 *
1343 * gup always represents data access, not instruction
1344 * fetches, so execute=false here:
1345 */
1346 if (!arch_vma_access_permitted(vma, write, false, foreign))
1347 return false;
1348
1349 return true;
1350}
1351
1352/**
1353 * fixup_user_fault() - manually resolve a user page fault
1354 * @mm: mm_struct of target mm
1355 * @address: user address
1356 * @fault_flags:flags to pass down to handle_mm_fault()
1357 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1358 * does not allow retry. If NULL, the caller must guarantee
1359 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1360 *
1361 * This is meant to be called in the specific scenario where for locking reasons
1362 * we try to access user memory in atomic context (within a pagefault_disable()
1363 * section), this returns -EFAULT, and we want to resolve the user fault before
1364 * trying again.
1365 *
1366 * Typically this is meant to be used by the futex code.
1367 *
1368 * The main difference with get_user_pages() is that this function will
1369 * unconditionally call handle_mm_fault() which will in turn perform all the
1370 * necessary SW fixup of the dirty and young bits in the PTE, while
1371 * get_user_pages() only guarantees to update these in the struct page.
1372 *
1373 * This is important for some architectures where those bits also gate the
1374 * access permission to the page because they are maintained in software. On
1375 * such architectures, gup() will not be enough to make a subsequent access
1376 * succeed.
1377 *
1378 * This function will not return with an unlocked mmap_lock. So it has not the
1379 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1380 */
1381int fixup_user_fault(struct mm_struct *mm,
1382 unsigned long address, unsigned int fault_flags,
1383 bool *unlocked)
1384{
1385 struct vm_area_struct *vma;
1386 vm_fault_t ret;
1387
1388 address = untagged_addr_remote(mm, address);
1389
1390 if (unlocked)
1391 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1392
1393retry:
1394 vma = gup_vma_lookup(mm, address);
1395 if (!vma)
1396 return -EFAULT;
1397
1398 if (!vma_permits_fault(vma, fault_flags))
1399 return -EFAULT;
1400
1401 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1402 fatal_signal_pending(current))
1403 return -EINTR;
1404
1405 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1406
1407 if (ret & VM_FAULT_COMPLETED) {
1408 /*
1409 * NOTE: it's a pity that we need to retake the lock here
1410 * to pair with the unlock() in the callers. Ideally we
1411 * could tell the callers so they do not need to unlock.
1412 */
1413 mmap_read_lock(mm);
1414 *unlocked = true;
1415 return 0;
1416 }
1417
1418 if (ret & VM_FAULT_ERROR) {
1419 int err = vm_fault_to_errno(ret, 0);
1420
1421 if (err)
1422 return err;
1423 BUG();
1424 }
1425
1426 if (ret & VM_FAULT_RETRY) {
1427 mmap_read_lock(mm);
1428 *unlocked = true;
1429 fault_flags |= FAULT_FLAG_TRIED;
1430 goto retry;
1431 }
1432
1433 return 0;
1434}
1435EXPORT_SYMBOL_GPL(fixup_user_fault);
1436
1437/*
1438 * GUP always responds to fatal signals. When FOLL_INTERRUPTIBLE is
1439 * specified, it'll also respond to generic signals. The caller of GUP
1440 * that has FOLL_INTERRUPTIBLE should take care of the GUP interruption.
1441 */
1442static bool gup_signal_pending(unsigned int flags)
1443{
1444 if (fatal_signal_pending(current))
1445 return true;
1446
1447 if (!(flags & FOLL_INTERRUPTIBLE))
1448 return false;
1449
1450 return signal_pending(current);
1451}
1452
1453/*
1454 * Locking: (*locked == 1) means that the mmap_lock has already been acquired by
1455 * the caller. This function may drop the mmap_lock. If it does so, then it will
1456 * set (*locked = 0).
1457 *
1458 * (*locked == 0) means that the caller expects this function to acquire and
1459 * drop the mmap_lock. Therefore, the value of *locked will still be zero when
1460 * the function returns, even though it may have changed temporarily during
1461 * function execution.
1462 *
1463 * Please note that this function, unlike __get_user_pages(), will not return 0
1464 * for nr_pages > 0, unless FOLL_NOWAIT is used.
1465 */
1466static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1467 unsigned long start,
1468 unsigned long nr_pages,
1469 struct page **pages,
1470 int *locked,
1471 unsigned int flags)
1472{
1473 long ret, pages_done;
1474 bool must_unlock = false;
1475
1476 if (!nr_pages)
1477 return 0;
1478
1479 /*
1480 * The internal caller expects GUP to manage the lock internally and the
1481 * lock must be released when this returns.
1482 */
1483 if (!*locked) {
1484 if (mmap_read_lock_killable(mm))
1485 return -EAGAIN;
1486 must_unlock = true;
1487 *locked = 1;
1488 }
1489 else
1490 mmap_assert_locked(mm);
1491
1492 if (flags & FOLL_PIN)
1493 mm_set_has_pinned_flag(&mm->flags);
1494
1495 /*
1496 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1497 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1498 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1499 * for FOLL_GET, not for the newer FOLL_PIN.
1500 *
1501 * FOLL_PIN always expects pages to be non-null, but no need to assert
1502 * that here, as any failures will be obvious enough.
1503 */
1504 if (pages && !(flags & FOLL_PIN))
1505 flags |= FOLL_GET;
1506
1507 pages_done = 0;
1508 for (;;) {
1509 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1510 locked);
1511 if (!(flags & FOLL_UNLOCKABLE)) {
1512 /* VM_FAULT_RETRY couldn't trigger, bypass */
1513 pages_done = ret;
1514 break;
1515 }
1516
1517 /* VM_FAULT_RETRY or VM_FAULT_COMPLETED cannot return errors */
1518 if (!*locked) {
1519 BUG_ON(ret < 0);
1520 BUG_ON(ret >= nr_pages);
1521 }
1522
1523 if (ret > 0) {
1524 nr_pages -= ret;
1525 pages_done += ret;
1526 if (!nr_pages)
1527 break;
1528 }
1529 if (*locked) {
1530 /*
1531 * VM_FAULT_RETRY didn't trigger or it was a
1532 * FOLL_NOWAIT.
1533 */
1534 if (!pages_done)
1535 pages_done = ret;
1536 break;
1537 }
1538 /*
1539 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1540 * For the prefault case (!pages) we only update counts.
1541 */
1542 if (likely(pages))
1543 pages += ret;
1544 start += ret << PAGE_SHIFT;
1545
1546 /* The lock was temporarily dropped, so we must unlock later */
1547 must_unlock = true;
1548
1549retry:
1550 /*
1551 * Repeat on the address that fired VM_FAULT_RETRY
1552 * with both FAULT_FLAG_ALLOW_RETRY and
1553 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1554 * by fatal signals of even common signals, depending on
1555 * the caller's request. So we need to check it before we
1556 * start trying again otherwise it can loop forever.
1557 */
1558 if (gup_signal_pending(flags)) {
1559 if (!pages_done)
1560 pages_done = -EINTR;
1561 break;
1562 }
1563
1564 ret = mmap_read_lock_killable(mm);
1565 if (ret) {
1566 BUG_ON(ret > 0);
1567 if (!pages_done)
1568 pages_done = ret;
1569 break;
1570 }
1571
1572 *locked = 1;
1573 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1574 pages, locked);
1575 if (!*locked) {
1576 /* Continue to retry until we succeeded */
1577 BUG_ON(ret != 0);
1578 goto retry;
1579 }
1580 if (ret != 1) {
1581 BUG_ON(ret > 1);
1582 if (!pages_done)
1583 pages_done = ret;
1584 break;
1585 }
1586 nr_pages--;
1587 pages_done++;
1588 if (!nr_pages)
1589 break;
1590 if (likely(pages))
1591 pages++;
1592 start += PAGE_SIZE;
1593 }
1594 if (must_unlock && *locked) {
1595 /*
1596 * We either temporarily dropped the lock, or the caller
1597 * requested that we both acquire and drop the lock. Either way,
1598 * we must now unlock, and notify the caller of that state.
1599 */
1600 mmap_read_unlock(mm);
1601 *locked = 0;
1602 }
1603
1604 /*
1605 * Failing to pin anything implies something has gone wrong (except when
1606 * FOLL_NOWAIT is specified).
1607 */
1608 if (WARN_ON_ONCE(pages_done == 0 && !(flags & FOLL_NOWAIT)))
1609 return -EFAULT;
1610
1611 return pages_done;
1612}
1613
1614/**
1615 * populate_vma_page_range() - populate a range of pages in the vma.
1616 * @vma: target vma
1617 * @start: start address
1618 * @end: end address
1619 * @locked: whether the mmap_lock is still held
1620 *
1621 * This takes care of mlocking the pages too if VM_LOCKED is set.
1622 *
1623 * Return either number of pages pinned in the vma, or a negative error
1624 * code on error.
1625 *
1626 * vma->vm_mm->mmap_lock must be held.
1627 *
1628 * If @locked is NULL, it may be held for read or write and will
1629 * be unperturbed.
1630 *
1631 * If @locked is non-NULL, it must held for read only and may be
1632 * released. If it's released, *@locked will be set to 0.
1633 */
1634long populate_vma_page_range(struct vm_area_struct *vma,
1635 unsigned long start, unsigned long end, int *locked)
1636{
1637 struct mm_struct *mm = vma->vm_mm;
1638 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1639 int local_locked = 1;
1640 int gup_flags;
1641 long ret;
1642
1643 VM_BUG_ON(!PAGE_ALIGNED(start));
1644 VM_BUG_ON(!PAGE_ALIGNED(end));
1645 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1646 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1647 mmap_assert_locked(mm);
1648
1649 /*
1650 * Rightly or wrongly, the VM_LOCKONFAULT case has never used
1651 * faultin_page() to break COW, so it has no work to do here.
1652 */
1653 if (vma->vm_flags & VM_LOCKONFAULT)
1654 return nr_pages;
1655
1656 gup_flags = FOLL_TOUCH;
1657 /*
1658 * We want to touch writable mappings with a write fault in order
1659 * to break COW, except for shared mappings because these don't COW
1660 * and we would not want to dirty them for nothing.
1661 */
1662 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1663 gup_flags |= FOLL_WRITE;
1664
1665 /*
1666 * We want mlock to succeed for regions that have any permissions
1667 * other than PROT_NONE.
1668 */
1669 if (vma_is_accessible(vma))
1670 gup_flags |= FOLL_FORCE;
1671
1672 if (locked)
1673 gup_flags |= FOLL_UNLOCKABLE;
1674
1675 /*
1676 * We made sure addr is within a VMA, so the following will
1677 * not result in a stack expansion that recurses back here.
1678 */
1679 ret = __get_user_pages(mm, start, nr_pages, gup_flags,
1680 NULL, locked ? locked : &local_locked);
1681 lru_add_drain();
1682 return ret;
1683}
1684
1685/*
1686 * faultin_vma_page_range() - populate (prefault) page tables inside the
1687 * given VMA range readable/writable
1688 *
1689 * This takes care of mlocking the pages, too, if VM_LOCKED is set.
1690 *
1691 * @vma: target vma
1692 * @start: start address
1693 * @end: end address
1694 * @write: whether to prefault readable or writable
1695 * @locked: whether the mmap_lock is still held
1696 *
1697 * Returns either number of processed pages in the vma, or a negative error
1698 * code on error (see __get_user_pages()).
1699 *
1700 * vma->vm_mm->mmap_lock must be held. The range must be page-aligned and
1701 * covered by the VMA. If it's released, *@locked will be set to 0.
1702 */
1703long faultin_vma_page_range(struct vm_area_struct *vma, unsigned long start,
1704 unsigned long end, bool write, int *locked)
1705{
1706 struct mm_struct *mm = vma->vm_mm;
1707 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1708 int gup_flags;
1709 long ret;
1710
1711 VM_BUG_ON(!PAGE_ALIGNED(start));
1712 VM_BUG_ON(!PAGE_ALIGNED(end));
1713 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1714 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1715 mmap_assert_locked(mm);
1716
1717 /*
1718 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark
1719 * the page dirty with FOLL_WRITE -- which doesn't make a
1720 * difference with !FOLL_FORCE, because the page is writable
1721 * in the page table.
1722 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit
1723 * a poisoned page.
1724 * !FOLL_FORCE: Require proper access permissions.
1725 */
1726 gup_flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_UNLOCKABLE;
1727 if (write)
1728 gup_flags |= FOLL_WRITE;
1729
1730 /*
1731 * We want to report -EINVAL instead of -EFAULT for any permission
1732 * problems or incompatible mappings.
1733 */
1734 if (check_vma_flags(vma, gup_flags))
1735 return -EINVAL;
1736
1737 ret = __get_user_pages(mm, start, nr_pages, gup_flags,
1738 NULL, locked);
1739 lru_add_drain();
1740 return ret;
1741}
1742
1743/*
1744 * __mm_populate - populate and/or mlock pages within a range of address space.
1745 *
1746 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1747 * flags. VMAs must be already marked with the desired vm_flags, and
1748 * mmap_lock must not be held.
1749 */
1750int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1751{
1752 struct mm_struct *mm = current->mm;
1753 unsigned long end, nstart, nend;
1754 struct vm_area_struct *vma = NULL;
1755 int locked = 0;
1756 long ret = 0;
1757
1758 end = start + len;
1759
1760 for (nstart = start; nstart < end; nstart = nend) {
1761 /*
1762 * We want to fault in pages for [nstart; end) address range.
1763 * Find first corresponding VMA.
1764 */
1765 if (!locked) {
1766 locked = 1;
1767 mmap_read_lock(mm);
1768 vma = find_vma_intersection(mm, nstart, end);
1769 } else if (nstart >= vma->vm_end)
1770 vma = find_vma_intersection(mm, vma->vm_end, end);
1771
1772 if (!vma)
1773 break;
1774 /*
1775 * Set [nstart; nend) to intersection of desired address
1776 * range with the first VMA. Also, skip undesirable VMA types.
1777 */
1778 nend = min(end, vma->vm_end);
1779 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1780 continue;
1781 if (nstart < vma->vm_start)
1782 nstart = vma->vm_start;
1783 /*
1784 * Now fault in a range of pages. populate_vma_page_range()
1785 * double checks the vma flags, so that it won't mlock pages
1786 * if the vma was already munlocked.
1787 */
1788 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1789 if (ret < 0) {
1790 if (ignore_errors) {
1791 ret = 0;
1792 continue; /* continue at next VMA */
1793 }
1794 break;
1795 }
1796 nend = nstart + ret * PAGE_SIZE;
1797 ret = 0;
1798 }
1799 if (locked)
1800 mmap_read_unlock(mm);
1801 return ret; /* 0 or negative error code */
1802}
1803#else /* CONFIG_MMU */
1804static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1805 unsigned long nr_pages, struct page **pages,
1806 int *locked, unsigned int foll_flags)
1807{
1808 struct vm_area_struct *vma;
1809 bool must_unlock = false;
1810 unsigned long vm_flags;
1811 long i;
1812
1813 if (!nr_pages)
1814 return 0;
1815
1816 /*
1817 * The internal caller expects GUP to manage the lock internally and the
1818 * lock must be released when this returns.
1819 */
1820 if (!*locked) {
1821 if (mmap_read_lock_killable(mm))
1822 return -EAGAIN;
1823 must_unlock = true;
1824 *locked = 1;
1825 }
1826
1827 /* calculate required read or write permissions.
1828 * If FOLL_FORCE is set, we only require the "MAY" flags.
1829 */
1830 vm_flags = (foll_flags & FOLL_WRITE) ?
1831 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1832 vm_flags &= (foll_flags & FOLL_FORCE) ?
1833 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1834
1835 for (i = 0; i < nr_pages; i++) {
1836 vma = find_vma(mm, start);
1837 if (!vma)
1838 break;
1839
1840 /* protect what we can, including chardevs */
1841 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1842 !(vm_flags & vma->vm_flags))
1843 break;
1844
1845 if (pages) {
1846 pages[i] = virt_to_page((void *)start);
1847 if (pages[i])
1848 get_page(pages[i]);
1849 }
1850
1851 start = (start + PAGE_SIZE) & PAGE_MASK;
1852 }
1853
1854 if (must_unlock && *locked) {
1855 mmap_read_unlock(mm);
1856 *locked = 0;
1857 }
1858
1859 return i ? : -EFAULT;
1860}
1861#endif /* !CONFIG_MMU */
1862
1863/**
1864 * fault_in_writeable - fault in userspace address range for writing
1865 * @uaddr: start of address range
1866 * @size: size of address range
1867 *
1868 * Returns the number of bytes not faulted in (like copy_to_user() and
1869 * copy_from_user()).
1870 */
1871size_t fault_in_writeable(char __user *uaddr, size_t size)
1872{
1873 char __user *start = uaddr, *end;
1874
1875 if (unlikely(size == 0))
1876 return 0;
1877 if (!user_write_access_begin(uaddr, size))
1878 return size;
1879 if (!PAGE_ALIGNED(uaddr)) {
1880 unsafe_put_user(0, uaddr, out);
1881 uaddr = (char __user *)PAGE_ALIGN((unsigned long)uaddr);
1882 }
1883 end = (char __user *)PAGE_ALIGN((unsigned long)start + size);
1884 if (unlikely(end < start))
1885 end = NULL;
1886 while (uaddr != end) {
1887 unsafe_put_user(0, uaddr, out);
1888 uaddr += PAGE_SIZE;
1889 }
1890
1891out:
1892 user_write_access_end();
1893 if (size > uaddr - start)
1894 return size - (uaddr - start);
1895 return 0;
1896}
1897EXPORT_SYMBOL(fault_in_writeable);
1898
1899/**
1900 * fault_in_subpage_writeable - fault in an address range for writing
1901 * @uaddr: start of address range
1902 * @size: size of address range
1903 *
1904 * Fault in a user address range for writing while checking for permissions at
1905 * sub-page granularity (e.g. arm64 MTE). This function should be used when
1906 * the caller cannot guarantee forward progress of a copy_to_user() loop.
1907 *
1908 * Returns the number of bytes not faulted in (like copy_to_user() and
1909 * copy_from_user()).
1910 */
1911size_t fault_in_subpage_writeable(char __user *uaddr, size_t size)
1912{
1913 size_t faulted_in;
1914
1915 /*
1916 * Attempt faulting in at page granularity first for page table
1917 * permission checking. The arch-specific probe_subpage_writeable()
1918 * functions may not check for this.
1919 */
1920 faulted_in = size - fault_in_writeable(uaddr, size);
1921 if (faulted_in)
1922 faulted_in -= probe_subpage_writeable(uaddr, faulted_in);
1923
1924 return size - faulted_in;
1925}
1926EXPORT_SYMBOL(fault_in_subpage_writeable);
1927
1928/*
1929 * fault_in_safe_writeable - fault in an address range for writing
1930 * @uaddr: start of address range
1931 * @size: length of address range
1932 *
1933 * Faults in an address range for writing. This is primarily useful when we
1934 * already know that some or all of the pages in the address range aren't in
1935 * memory.
1936 *
1937 * Unlike fault_in_writeable(), this function is non-destructive.
1938 *
1939 * Note that we don't pin or otherwise hold the pages referenced that we fault
1940 * in. There's no guarantee that they'll stay in memory for any duration of
1941 * time.
1942 *
1943 * Returns the number of bytes not faulted in, like copy_to_user() and
1944 * copy_from_user().
1945 */
1946size_t fault_in_safe_writeable(const char __user *uaddr, size_t size)
1947{
1948 unsigned long start = (unsigned long)uaddr, end;
1949 struct mm_struct *mm = current->mm;
1950 bool unlocked = false;
1951
1952 if (unlikely(size == 0))
1953 return 0;
1954 end = PAGE_ALIGN(start + size);
1955 if (end < start)
1956 end = 0;
1957
1958 mmap_read_lock(mm);
1959 do {
1960 if (fixup_user_fault(mm, start, FAULT_FLAG_WRITE, &unlocked))
1961 break;
1962 start = (start + PAGE_SIZE) & PAGE_MASK;
1963 } while (start != end);
1964 mmap_read_unlock(mm);
1965
1966 if (size > (unsigned long)uaddr - start)
1967 return size - ((unsigned long)uaddr - start);
1968 return 0;
1969}
1970EXPORT_SYMBOL(fault_in_safe_writeable);
1971
1972/**
1973 * fault_in_readable - fault in userspace address range for reading
1974 * @uaddr: start of user address range
1975 * @size: size of user address range
1976 *
1977 * Returns the number of bytes not faulted in (like copy_to_user() and
1978 * copy_from_user()).
1979 */
1980size_t fault_in_readable(const char __user *uaddr, size_t size)
1981{
1982 const char __user *start = uaddr, *end;
1983 volatile char c;
1984
1985 if (unlikely(size == 0))
1986 return 0;
1987 if (!user_read_access_begin(uaddr, size))
1988 return size;
1989 if (!PAGE_ALIGNED(uaddr)) {
1990 unsafe_get_user(c, uaddr, out);
1991 uaddr = (const char __user *)PAGE_ALIGN((unsigned long)uaddr);
1992 }
1993 end = (const char __user *)PAGE_ALIGN((unsigned long)start + size);
1994 if (unlikely(end < start))
1995 end = NULL;
1996 while (uaddr != end) {
1997 unsafe_get_user(c, uaddr, out);
1998 uaddr += PAGE_SIZE;
1999 }
2000
2001out:
2002 user_read_access_end();
2003 (void)c;
2004 if (size > uaddr - start)
2005 return size - (uaddr - start);
2006 return 0;
2007}
2008EXPORT_SYMBOL(fault_in_readable);
2009
2010/**
2011 * get_dump_page() - pin user page in memory while writing it to core dump
2012 * @addr: user address
2013 *
2014 * Returns struct page pointer of user page pinned for dump,
2015 * to be freed afterwards by put_page().
2016 *
2017 * Returns NULL on any kind of failure - a hole must then be inserted into
2018 * the corefile, to preserve alignment with its headers; and also returns
2019 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2020 * allowing a hole to be left in the corefile to save disk space.
2021 *
2022 * Called without mmap_lock (takes and releases the mmap_lock by itself).
2023 */
2024#ifdef CONFIG_ELF_CORE
2025struct page *get_dump_page(unsigned long addr)
2026{
2027 struct page *page;
2028 int locked = 0;
2029 int ret;
2030
2031 ret = __get_user_pages_locked(current->mm, addr, 1, &page, &locked,
2032 FOLL_FORCE | FOLL_DUMP | FOLL_GET);
2033 return (ret == 1) ? page : NULL;
2034}
2035#endif /* CONFIG_ELF_CORE */
2036
2037#ifdef CONFIG_MIGRATION
2038/*
2039 * Returns the number of collected pages. Return value is always >= 0.
2040 */
2041static unsigned long collect_longterm_unpinnable_pages(
2042 struct list_head *movable_page_list,
2043 unsigned long nr_pages,
2044 struct page **pages)
2045{
2046 unsigned long i, collected = 0;
2047 struct folio *prev_folio = NULL;
2048 bool drain_allow = true;
2049
2050 for (i = 0; i < nr_pages; i++) {
2051 struct folio *folio = page_folio(pages[i]);
2052
2053 if (folio == prev_folio)
2054 continue;
2055 prev_folio = folio;
2056
2057 if (folio_is_longterm_pinnable(folio))
2058 continue;
2059
2060 collected++;
2061
2062 if (folio_is_device_coherent(folio))
2063 continue;
2064
2065 if (folio_test_hugetlb(folio)) {
2066 isolate_hugetlb(folio, movable_page_list);
2067 continue;
2068 }
2069
2070 if (!folio_test_lru(folio) && drain_allow) {
2071 lru_add_drain_all();
2072 drain_allow = false;
2073 }
2074
2075 if (!folio_isolate_lru(folio))
2076 continue;
2077
2078 list_add_tail(&folio->lru, movable_page_list);
2079 node_stat_mod_folio(folio,
2080 NR_ISOLATED_ANON + folio_is_file_lru(folio),
2081 folio_nr_pages(folio));
2082 }
2083
2084 return collected;
2085}
2086
2087/*
2088 * Unpins all pages and migrates device coherent pages and movable_page_list.
2089 * Returns -EAGAIN if all pages were successfully migrated or -errno for failure
2090 * (or partial success).
2091 */
2092static int migrate_longterm_unpinnable_pages(
2093 struct list_head *movable_page_list,
2094 unsigned long nr_pages,
2095 struct page **pages)
2096{
2097 int ret;
2098 unsigned long i;
2099
2100 for (i = 0; i < nr_pages; i++) {
2101 struct folio *folio = page_folio(pages[i]);
2102
2103 if (folio_is_device_coherent(folio)) {
2104 /*
2105 * Migration will fail if the page is pinned, so convert
2106 * the pin on the source page to a normal reference.
2107 */
2108 pages[i] = NULL;
2109 folio_get(folio);
2110 gup_put_folio(folio, 1, FOLL_PIN);
2111
2112 if (migrate_device_coherent_page(&folio->page)) {
2113 ret = -EBUSY;
2114 goto err;
2115 }
2116
2117 continue;
2118 }
2119
2120 /*
2121 * We can't migrate pages with unexpected references, so drop
2122 * the reference obtained by __get_user_pages_locked().
2123 * Migrating pages have been added to movable_page_list after
2124 * calling folio_isolate_lru() which takes a reference so the
2125 * page won't be freed if it's migrating.
2126 */
2127 unpin_user_page(pages[i]);
2128 pages[i] = NULL;
2129 }
2130
2131 if (!list_empty(movable_page_list)) {
2132 struct migration_target_control mtc = {
2133 .nid = NUMA_NO_NODE,
2134 .gfp_mask = GFP_USER | __GFP_NOWARN,
2135 };
2136
2137 if (migrate_pages(movable_page_list, alloc_migration_target,
2138 NULL, (unsigned long)&mtc, MIGRATE_SYNC,
2139 MR_LONGTERM_PIN, NULL)) {
2140 ret = -ENOMEM;
2141 goto err;
2142 }
2143 }
2144
2145 putback_movable_pages(movable_page_list);
2146
2147 return -EAGAIN;
2148
2149err:
2150 for (i = 0; i < nr_pages; i++)
2151 if (pages[i])
2152 unpin_user_page(pages[i]);
2153 putback_movable_pages(movable_page_list);
2154
2155 return ret;
2156}
2157
2158/*
2159 * Check whether all pages are *allowed* to be pinned. Rather confusingly, all
2160 * pages in the range are required to be pinned via FOLL_PIN, before calling
2161 * this routine.
2162 *
2163 * If any pages in the range are not allowed to be pinned, then this routine
2164 * will migrate those pages away, unpin all the pages in the range and return
2165 * -EAGAIN. The caller should re-pin the entire range with FOLL_PIN and then
2166 * call this routine again.
2167 *
2168 * If an error other than -EAGAIN occurs, this indicates a migration failure.
2169 * The caller should give up, and propagate the error back up the call stack.
2170 *
2171 * If everything is OK and all pages in the range are allowed to be pinned, then
2172 * this routine leaves all pages pinned and returns zero for success.
2173 */
2174static long check_and_migrate_movable_pages(unsigned long nr_pages,
2175 struct page **pages)
2176{
2177 unsigned long collected;
2178 LIST_HEAD(movable_page_list);
2179
2180 collected = collect_longterm_unpinnable_pages(&movable_page_list,
2181 nr_pages, pages);
2182 if (!collected)
2183 return 0;
2184
2185 return migrate_longterm_unpinnable_pages(&movable_page_list, nr_pages,
2186 pages);
2187}
2188#else
2189static long check_and_migrate_movable_pages(unsigned long nr_pages,
2190 struct page **pages)
2191{
2192 return 0;
2193}
2194#endif /* CONFIG_MIGRATION */
2195
2196/*
2197 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
2198 * allows us to process the FOLL_LONGTERM flag.
2199 */
2200static long __gup_longterm_locked(struct mm_struct *mm,
2201 unsigned long start,
2202 unsigned long nr_pages,
2203 struct page **pages,
2204 int *locked,
2205 unsigned int gup_flags)
2206{
2207 unsigned int flags;
2208 long rc, nr_pinned_pages;
2209
2210 if (!(gup_flags & FOLL_LONGTERM))
2211 return __get_user_pages_locked(mm, start, nr_pages, pages,
2212 locked, gup_flags);
2213
2214 flags = memalloc_pin_save();
2215 do {
2216 nr_pinned_pages = __get_user_pages_locked(mm, start, nr_pages,
2217 pages, locked,
2218 gup_flags);
2219 if (nr_pinned_pages <= 0) {
2220 rc = nr_pinned_pages;
2221 break;
2222 }
2223
2224 /* FOLL_LONGTERM implies FOLL_PIN */
2225 rc = check_and_migrate_movable_pages(nr_pinned_pages, pages);
2226 } while (rc == -EAGAIN);
2227 memalloc_pin_restore(flags);
2228 return rc ? rc : nr_pinned_pages;
2229}
2230
2231/*
2232 * Check that the given flags are valid for the exported gup/pup interface, and
2233 * update them with the required flags that the caller must have set.
2234 */
2235static bool is_valid_gup_args(struct page **pages, int *locked,
2236 unsigned int *gup_flags_p, unsigned int to_set)
2237{
2238 unsigned int gup_flags = *gup_flags_p;
2239
2240 /*
2241 * These flags not allowed to be specified externally to the gup
2242 * interfaces:
2243 * - FOLL_TOUCH/FOLL_PIN/FOLL_TRIED/FOLL_FAST_ONLY are internal only
2244 * - FOLL_REMOTE is internal only and used on follow_page()
2245 * - FOLL_UNLOCKABLE is internal only and used if locked is !NULL
2246 */
2247 if (WARN_ON_ONCE(gup_flags & INTERNAL_GUP_FLAGS))
2248 return false;
2249
2250 gup_flags |= to_set;
2251 if (locked) {
2252 /* At the external interface locked must be set */
2253 if (WARN_ON_ONCE(*locked != 1))
2254 return false;
2255
2256 gup_flags |= FOLL_UNLOCKABLE;
2257 }
2258
2259 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2260 if (WARN_ON_ONCE((gup_flags & (FOLL_PIN | FOLL_GET)) ==
2261 (FOLL_PIN | FOLL_GET)))
2262 return false;
2263
2264 /* LONGTERM can only be specified when pinning */
2265 if (WARN_ON_ONCE(!(gup_flags & FOLL_PIN) && (gup_flags & FOLL_LONGTERM)))
2266 return false;
2267
2268 /* Pages input must be given if using GET/PIN */
2269 if (WARN_ON_ONCE((gup_flags & (FOLL_GET | FOLL_PIN)) && !pages))
2270 return false;
2271
2272 /* We want to allow the pgmap to be hot-unplugged at all times */
2273 if (WARN_ON_ONCE((gup_flags & FOLL_LONGTERM) &&
2274 (gup_flags & FOLL_PCI_P2PDMA)))
2275 return false;
2276
2277 *gup_flags_p = gup_flags;
2278 return true;
2279}
2280
2281#ifdef CONFIG_MMU
2282/**
2283 * get_user_pages_remote() - pin user pages in memory
2284 * @mm: mm_struct of target mm
2285 * @start: starting user address
2286 * @nr_pages: number of pages from start to pin
2287 * @gup_flags: flags modifying lookup behaviour
2288 * @pages: array that receives pointers to the pages pinned.
2289 * Should be at least nr_pages long. Or NULL, if caller
2290 * only intends to ensure the pages are faulted in.
2291 * @locked: pointer to lock flag indicating whether lock is held and
2292 * subsequently whether VM_FAULT_RETRY functionality can be
2293 * utilised. Lock must initially be held.
2294 *
2295 * Returns either number of pages pinned (which may be less than the
2296 * number requested), or an error. Details about the return value:
2297 *
2298 * -- If nr_pages is 0, returns 0.
2299 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
2300 * -- If nr_pages is >0, and some pages were pinned, returns the number of
2301 * pages pinned. Again, this may be less than nr_pages.
2302 *
2303 * The caller is responsible for releasing returned @pages, via put_page().
2304 *
2305 * Must be called with mmap_lock held for read or write.
2306 *
2307 * get_user_pages_remote walks a process's page tables and takes a reference
2308 * to each struct page that each user address corresponds to at a given
2309 * instant. That is, it takes the page that would be accessed if a user
2310 * thread accesses the given user virtual address at that instant.
2311 *
2312 * This does not guarantee that the page exists in the user mappings when
2313 * get_user_pages_remote returns, and there may even be a completely different
2314 * page there in some cases (eg. if mmapped pagecache has been invalidated
2315 * and subsequently re-faulted). However it does guarantee that the page
2316 * won't be freed completely. And mostly callers simply care that the page
2317 * contains data that was valid *at some point in time*. Typically, an IO
2318 * or similar operation cannot guarantee anything stronger anyway because
2319 * locks can't be held over the syscall boundary.
2320 *
2321 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
2322 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
2323 * be called after the page is finished with, and before put_page is called.
2324 *
2325 * get_user_pages_remote is typically used for fewer-copy IO operations,
2326 * to get a handle on the memory by some means other than accesses
2327 * via the user virtual addresses. The pages may be submitted for
2328 * DMA to devices or accessed via their kernel linear mapping (via the
2329 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
2330 *
2331 * See also get_user_pages_fast, for performance critical applications.
2332 *
2333 * get_user_pages_remote should be phased out in favor of
2334 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
2335 * should use get_user_pages_remote because it cannot pass
2336 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
2337 */
2338long get_user_pages_remote(struct mm_struct *mm,
2339 unsigned long start, unsigned long nr_pages,
2340 unsigned int gup_flags, struct page **pages,
2341 int *locked)
2342{
2343 int local_locked = 1;
2344
2345 if (!is_valid_gup_args(pages, locked, &gup_flags,
2346 FOLL_TOUCH | FOLL_REMOTE))
2347 return -EINVAL;
2348
2349 return __get_user_pages_locked(mm, start, nr_pages, pages,
2350 locked ? locked : &local_locked,
2351 gup_flags);
2352}
2353EXPORT_SYMBOL(get_user_pages_remote);
2354
2355#else /* CONFIG_MMU */
2356long get_user_pages_remote(struct mm_struct *mm,
2357 unsigned long start, unsigned long nr_pages,
2358 unsigned int gup_flags, struct page **pages,
2359 int *locked)
2360{
2361 return 0;
2362}
2363#endif /* !CONFIG_MMU */
2364
2365/**
2366 * get_user_pages() - pin user pages in memory
2367 * @start: starting user address
2368 * @nr_pages: number of pages from start to pin
2369 * @gup_flags: flags modifying lookup behaviour
2370 * @pages: array that receives pointers to the pages pinned.
2371 * Should be at least nr_pages long. Or NULL, if caller
2372 * only intends to ensure the pages are faulted in.
2373 *
2374 * This is the same as get_user_pages_remote(), just with a less-flexible
2375 * calling convention where we assume that the mm being operated on belongs to
2376 * the current task, and doesn't allow passing of a locked parameter. We also
2377 * obviously don't pass FOLL_REMOTE in here.
2378 */
2379long get_user_pages(unsigned long start, unsigned long nr_pages,
2380 unsigned int gup_flags, struct page **pages)
2381{
2382 int locked = 1;
2383
2384 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_TOUCH))
2385 return -EINVAL;
2386
2387 return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2388 &locked, gup_flags);
2389}
2390EXPORT_SYMBOL(get_user_pages);
2391
2392/*
2393 * get_user_pages_unlocked() is suitable to replace the form:
2394 *
2395 * mmap_read_lock(mm);
2396 * get_user_pages(mm, ..., pages, NULL);
2397 * mmap_read_unlock(mm);
2398 *
2399 * with:
2400 *
2401 * get_user_pages_unlocked(mm, ..., pages);
2402 *
2403 * It is functionally equivalent to get_user_pages_fast so
2404 * get_user_pages_fast should be used instead if specific gup_flags
2405 * (e.g. FOLL_FORCE) are not required.
2406 */
2407long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2408 struct page **pages, unsigned int gup_flags)
2409{
2410 int locked = 0;
2411
2412 if (!is_valid_gup_args(pages, NULL, &gup_flags,
2413 FOLL_TOUCH | FOLL_UNLOCKABLE))
2414 return -EINVAL;
2415
2416 return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2417 &locked, gup_flags);
2418}
2419EXPORT_SYMBOL(get_user_pages_unlocked);
2420
2421/*
2422 * Fast GUP
2423 *
2424 * get_user_pages_fast attempts to pin user pages by walking the page
2425 * tables directly and avoids taking locks. Thus the walker needs to be
2426 * protected from page table pages being freed from under it, and should
2427 * block any THP splits.
2428 *
2429 * One way to achieve this is to have the walker disable interrupts, and
2430 * rely on IPIs from the TLB flushing code blocking before the page table
2431 * pages are freed. This is unsuitable for architectures that do not need
2432 * to broadcast an IPI when invalidating TLBs.
2433 *
2434 * Another way to achieve this is to batch up page table containing pages
2435 * belonging to more than one mm_user, then rcu_sched a callback to free those
2436 * pages. Disabling interrupts will allow the fast_gup walker to both block
2437 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2438 * (which is a relatively rare event). The code below adopts this strategy.
2439 *
2440 * Before activating this code, please be aware that the following assumptions
2441 * are currently made:
2442 *
2443 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2444 * free pages containing page tables or TLB flushing requires IPI broadcast.
2445 *
2446 * *) ptes can be read atomically by the architecture.
2447 *
2448 * *) access_ok is sufficient to validate userspace address ranges.
2449 *
2450 * The last two assumptions can be relaxed by the addition of helper functions.
2451 *
2452 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2453 */
2454#ifdef CONFIG_HAVE_FAST_GUP
2455
2456/*
2457 * Used in the GUP-fast path to determine whether a pin is permitted for a
2458 * specific folio.
2459 *
2460 * This call assumes the caller has pinned the folio, that the lowest page table
2461 * level still points to this folio, and that interrupts have been disabled.
2462 *
2463 * Writing to pinned file-backed dirty tracked folios is inherently problematic
2464 * (see comment describing the writable_file_mapping_allowed() function). We
2465 * therefore try to avoid the most egregious case of a long-term mapping doing
2466 * so.
2467 *
2468 * This function cannot be as thorough as that one as the VMA is not available
2469 * in the fast path, so instead we whitelist known good cases and if in doubt,
2470 * fall back to the slow path.
2471 */
2472static bool folio_fast_pin_allowed(struct folio *folio, unsigned int flags)
2473{
2474 struct address_space *mapping;
2475 unsigned long mapping_flags;
2476
2477 /*
2478 * If we aren't pinning then no problematic write can occur. A long term
2479 * pin is the most egregious case so this is the one we disallow.
2480 */
2481 if ((flags & (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE)) !=
2482 (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE))
2483 return true;
2484
2485 /* The folio is pinned, so we can safely access folio fields. */
2486
2487 if (WARN_ON_ONCE(folio_test_slab(folio)))
2488 return false;
2489
2490 /* hugetlb mappings do not require dirty-tracking. */
2491 if (folio_test_hugetlb(folio))
2492 return true;
2493
2494 /*
2495 * GUP-fast disables IRQs. When IRQS are disabled, RCU grace periods
2496 * cannot proceed, which means no actions performed under RCU can
2497 * proceed either.
2498 *
2499 * inodes and thus their mappings are freed under RCU, which means the
2500 * mapping cannot be freed beneath us and thus we can safely dereference
2501 * it.
2502 */
2503 lockdep_assert_irqs_disabled();
2504
2505 /*
2506 * However, there may be operations which _alter_ the mapping, so ensure
2507 * we read it once and only once.
2508 */
2509 mapping = READ_ONCE(folio->mapping);
2510
2511 /*
2512 * The mapping may have been truncated, in any case we cannot determine
2513 * if this mapping is safe - fall back to slow path to determine how to
2514 * proceed.
2515 */
2516 if (!mapping)
2517 return false;
2518
2519 /* Anonymous folios pose no problem. */
2520 mapping_flags = (unsigned long)mapping & PAGE_MAPPING_FLAGS;
2521 if (mapping_flags)
2522 return mapping_flags & PAGE_MAPPING_ANON;
2523
2524 /*
2525 * At this point, we know the mapping is non-null and points to an
2526 * address_space object. The only remaining whitelisted file system is
2527 * shmem.
2528 */
2529 return shmem_mapping(mapping);
2530}
2531
2532static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2533 unsigned int flags,
2534 struct page **pages)
2535{
2536 while ((*nr) - nr_start) {
2537 struct page *page = pages[--(*nr)];
2538
2539 ClearPageReferenced(page);
2540 if (flags & FOLL_PIN)
2541 unpin_user_page(page);
2542 else
2543 put_page(page);
2544 }
2545}
2546
2547#ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2548/*
2549 * Fast-gup relies on pte change detection to avoid concurrent pgtable
2550 * operations.
2551 *
2552 * To pin the page, fast-gup needs to do below in order:
2553 * (1) pin the page (by prefetching pte), then (2) check pte not changed.
2554 *
2555 * For the rest of pgtable operations where pgtable updates can be racy
2556 * with fast-gup, we need to do (1) clear pte, then (2) check whether page
2557 * is pinned.
2558 *
2559 * Above will work for all pte-level operations, including THP split.
2560 *
2561 * For THP collapse, it's a bit more complicated because fast-gup may be
2562 * walking a pgtable page that is being freed (pte is still valid but pmd
2563 * can be cleared already). To avoid race in such condition, we need to
2564 * also check pmd here to make sure pmd doesn't change (corresponds to
2565 * pmdp_collapse_flush() in the THP collapse code path).
2566 */
2567static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2568 unsigned long end, unsigned int flags,
2569 struct page **pages, int *nr)
2570{
2571 struct dev_pagemap *pgmap = NULL;
2572 int nr_start = *nr, ret = 0;
2573 pte_t *ptep, *ptem;
2574
2575 ptem = ptep = pte_offset_map(&pmd, addr);
2576 if (!ptep)
2577 return 0;
2578 do {
2579 pte_t pte = ptep_get_lockless(ptep);
2580 struct page *page;
2581 struct folio *folio;
2582
2583 /*
2584 * Always fallback to ordinary GUP on PROT_NONE-mapped pages:
2585 * pte_access_permitted() better should reject these pages
2586 * either way: otherwise, GUP-fast might succeed in
2587 * cases where ordinary GUP would fail due to VMA access
2588 * permissions.
2589 */
2590 if (pte_protnone(pte))
2591 goto pte_unmap;
2592
2593 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2594 goto pte_unmap;
2595
2596 if (pte_devmap(pte)) {
2597 if (unlikely(flags & FOLL_LONGTERM))
2598 goto pte_unmap;
2599
2600 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2601 if (unlikely(!pgmap)) {
2602 undo_dev_pagemap(nr, nr_start, flags, pages);
2603 goto pte_unmap;
2604 }
2605 } else if (pte_special(pte))
2606 goto pte_unmap;
2607
2608 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2609 page = pte_page(pte);
2610
2611 folio = try_grab_folio(page, 1, flags);
2612 if (!folio)
2613 goto pte_unmap;
2614
2615 if (unlikely(folio_is_secretmem(folio))) {
2616 gup_put_folio(folio, 1, flags);
2617 goto pte_unmap;
2618 }
2619
2620 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp)) ||
2621 unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) {
2622 gup_put_folio(folio, 1, flags);
2623 goto pte_unmap;
2624 }
2625
2626 if (!folio_fast_pin_allowed(folio, flags)) {
2627 gup_put_folio(folio, 1, flags);
2628 goto pte_unmap;
2629 }
2630
2631 if (!pte_write(pte) && gup_must_unshare(NULL, flags, page)) {
2632 gup_put_folio(folio, 1, flags);
2633 goto pte_unmap;
2634 }
2635
2636 /*
2637 * We need to make the page accessible if and only if we are
2638 * going to access its content (the FOLL_PIN case). Please
2639 * see Documentation/core-api/pin_user_pages.rst for
2640 * details.
2641 */
2642 if (flags & FOLL_PIN) {
2643 ret = arch_make_page_accessible(page);
2644 if (ret) {
2645 gup_put_folio(folio, 1, flags);
2646 goto pte_unmap;
2647 }
2648 }
2649 folio_set_referenced(folio);
2650 pages[*nr] = page;
2651 (*nr)++;
2652 } while (ptep++, addr += PAGE_SIZE, addr != end);
2653
2654 ret = 1;
2655
2656pte_unmap:
2657 if (pgmap)
2658 put_dev_pagemap(pgmap);
2659 pte_unmap(ptem);
2660 return ret;
2661}
2662#else
2663
2664/*
2665 * If we can't determine whether or not a pte is special, then fail immediately
2666 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2667 * to be special.
2668 *
2669 * For a futex to be placed on a THP tail page, get_futex_key requires a
2670 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2671 * useful to have gup_huge_pmd even if we can't operate on ptes.
2672 */
2673static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2674 unsigned long end, unsigned int flags,
2675 struct page **pages, int *nr)
2676{
2677 return 0;
2678}
2679#endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2680
2681#if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2682static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2683 unsigned long end, unsigned int flags,
2684 struct page **pages, int *nr)
2685{
2686 int nr_start = *nr;
2687 struct dev_pagemap *pgmap = NULL;
2688
2689 do {
2690 struct page *page = pfn_to_page(pfn);
2691
2692 pgmap = get_dev_pagemap(pfn, pgmap);
2693 if (unlikely(!pgmap)) {
2694 undo_dev_pagemap(nr, nr_start, flags, pages);
2695 break;
2696 }
2697
2698 if (!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)) {
2699 undo_dev_pagemap(nr, nr_start, flags, pages);
2700 break;
2701 }
2702
2703 SetPageReferenced(page);
2704 pages[*nr] = page;
2705 if (unlikely(try_grab_page(page, flags))) {
2706 undo_dev_pagemap(nr, nr_start, flags, pages);
2707 break;
2708 }
2709 (*nr)++;
2710 pfn++;
2711 } while (addr += PAGE_SIZE, addr != end);
2712
2713 put_dev_pagemap(pgmap);
2714 return addr == end;
2715}
2716
2717static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2718 unsigned long end, unsigned int flags,
2719 struct page **pages, int *nr)
2720{
2721 unsigned long fault_pfn;
2722 int nr_start = *nr;
2723
2724 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2725 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2726 return 0;
2727
2728 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2729 undo_dev_pagemap(nr, nr_start, flags, pages);
2730 return 0;
2731 }
2732 return 1;
2733}
2734
2735static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2736 unsigned long end, unsigned int flags,
2737 struct page **pages, int *nr)
2738{
2739 unsigned long fault_pfn;
2740 int nr_start = *nr;
2741
2742 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2743 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2744 return 0;
2745
2746 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2747 undo_dev_pagemap(nr, nr_start, flags, pages);
2748 return 0;
2749 }
2750 return 1;
2751}
2752#else
2753static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2754 unsigned long end, unsigned int flags,
2755 struct page **pages, int *nr)
2756{
2757 BUILD_BUG();
2758 return 0;
2759}
2760
2761static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2762 unsigned long end, unsigned int flags,
2763 struct page **pages, int *nr)
2764{
2765 BUILD_BUG();
2766 return 0;
2767}
2768#endif
2769
2770static int record_subpages(struct page *page, unsigned long addr,
2771 unsigned long end, struct page **pages)
2772{
2773 int nr;
2774
2775 for (nr = 0; addr != end; nr++, addr += PAGE_SIZE)
2776 pages[nr] = nth_page(page, nr);
2777
2778 return nr;
2779}
2780
2781#ifdef CONFIG_ARCH_HAS_HUGEPD
2782static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2783 unsigned long sz)
2784{
2785 unsigned long __boundary = (addr + sz) & ~(sz-1);
2786 return (__boundary - 1 < end - 1) ? __boundary : end;
2787}
2788
2789static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2790 unsigned long end, unsigned int flags,
2791 struct page **pages, int *nr)
2792{
2793 unsigned long pte_end;
2794 struct page *page;
2795 struct folio *folio;
2796 pte_t pte;
2797 int refs;
2798
2799 pte_end = (addr + sz) & ~(sz-1);
2800 if (pte_end < end)
2801 end = pte_end;
2802
2803 pte = huge_ptep_get(ptep);
2804
2805 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2806 return 0;
2807
2808 /* hugepages are never "special" */
2809 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2810
2811 page = nth_page(pte_page(pte), (addr & (sz - 1)) >> PAGE_SHIFT);
2812 refs = record_subpages(page, addr, end, pages + *nr);
2813
2814 folio = try_grab_folio(page, refs, flags);
2815 if (!folio)
2816 return 0;
2817
2818 if (unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) {
2819 gup_put_folio(folio, refs, flags);
2820 return 0;
2821 }
2822
2823 if (!folio_fast_pin_allowed(folio, flags)) {
2824 gup_put_folio(folio, refs, flags);
2825 return 0;
2826 }
2827
2828 if (!pte_write(pte) && gup_must_unshare(NULL, flags, &folio->page)) {
2829 gup_put_folio(folio, refs, flags);
2830 return 0;
2831 }
2832
2833 *nr += refs;
2834 folio_set_referenced(folio);
2835 return 1;
2836}
2837
2838static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2839 unsigned int pdshift, unsigned long end, unsigned int flags,
2840 struct page **pages, int *nr)
2841{
2842 pte_t *ptep;
2843 unsigned long sz = 1UL << hugepd_shift(hugepd);
2844 unsigned long next;
2845
2846 ptep = hugepte_offset(hugepd, addr, pdshift);
2847 do {
2848 next = hugepte_addr_end(addr, end, sz);
2849 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2850 return 0;
2851 } while (ptep++, addr = next, addr != end);
2852
2853 return 1;
2854}
2855#else
2856static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2857 unsigned int pdshift, unsigned long end, unsigned int flags,
2858 struct page **pages, int *nr)
2859{
2860 return 0;
2861}
2862#endif /* CONFIG_ARCH_HAS_HUGEPD */
2863
2864static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2865 unsigned long end, unsigned int flags,
2866 struct page **pages, int *nr)
2867{
2868 struct page *page;
2869 struct folio *folio;
2870 int refs;
2871
2872 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2873 return 0;
2874
2875 if (pmd_devmap(orig)) {
2876 if (unlikely(flags & FOLL_LONGTERM))
2877 return 0;
2878 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2879 pages, nr);
2880 }
2881
2882 page = nth_page(pmd_page(orig), (addr & ~PMD_MASK) >> PAGE_SHIFT);
2883 refs = record_subpages(page, addr, end, pages + *nr);
2884
2885 folio = try_grab_folio(page, refs, flags);
2886 if (!folio)
2887 return 0;
2888
2889 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2890 gup_put_folio(folio, refs, flags);
2891 return 0;
2892 }
2893
2894 if (!folio_fast_pin_allowed(folio, flags)) {
2895 gup_put_folio(folio, refs, flags);
2896 return 0;
2897 }
2898 if (!pmd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2899 gup_put_folio(folio, refs, flags);
2900 return 0;
2901 }
2902
2903 *nr += refs;
2904 folio_set_referenced(folio);
2905 return 1;
2906}
2907
2908static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2909 unsigned long end, unsigned int flags,
2910 struct page **pages, int *nr)
2911{
2912 struct page *page;
2913 struct folio *folio;
2914 int refs;
2915
2916 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2917 return 0;
2918
2919 if (pud_devmap(orig)) {
2920 if (unlikely(flags & FOLL_LONGTERM))
2921 return 0;
2922 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2923 pages, nr);
2924 }
2925
2926 page = nth_page(pud_page(orig), (addr & ~PUD_MASK) >> PAGE_SHIFT);
2927 refs = record_subpages(page, addr, end, pages + *nr);
2928
2929 folio = try_grab_folio(page, refs, flags);
2930 if (!folio)
2931 return 0;
2932
2933 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2934 gup_put_folio(folio, refs, flags);
2935 return 0;
2936 }
2937
2938 if (!folio_fast_pin_allowed(folio, flags)) {
2939 gup_put_folio(folio, refs, flags);
2940 return 0;
2941 }
2942
2943 if (!pud_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2944 gup_put_folio(folio, refs, flags);
2945 return 0;
2946 }
2947
2948 *nr += refs;
2949 folio_set_referenced(folio);
2950 return 1;
2951}
2952
2953static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2954 unsigned long end, unsigned int flags,
2955 struct page **pages, int *nr)
2956{
2957 int refs;
2958 struct page *page;
2959 struct folio *folio;
2960
2961 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2962 return 0;
2963
2964 BUILD_BUG_ON(pgd_devmap(orig));
2965
2966 page = nth_page(pgd_page(orig), (addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2967 refs = record_subpages(page, addr, end, pages + *nr);
2968
2969 folio = try_grab_folio(page, refs, flags);
2970 if (!folio)
2971 return 0;
2972
2973 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2974 gup_put_folio(folio, refs, flags);
2975 return 0;
2976 }
2977
2978 if (!pgd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2979 gup_put_folio(folio, refs, flags);
2980 return 0;
2981 }
2982
2983 if (!folio_fast_pin_allowed(folio, flags)) {
2984 gup_put_folio(folio, refs, flags);
2985 return 0;
2986 }
2987
2988 *nr += refs;
2989 folio_set_referenced(folio);
2990 return 1;
2991}
2992
2993static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
2994 unsigned int flags, struct page **pages, int *nr)
2995{
2996 unsigned long next;
2997 pmd_t *pmdp;
2998
2999 pmdp = pmd_offset_lockless(pudp, pud, addr);
3000 do {
3001 pmd_t pmd = pmdp_get_lockless(pmdp);
3002
3003 next = pmd_addr_end(addr, end);
3004 if (!pmd_present(pmd))
3005 return 0;
3006
3007 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
3008 pmd_devmap(pmd))) {
3009 /* See gup_pte_range() */
3010 if (pmd_protnone(pmd))
3011 return 0;
3012
3013 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
3014 pages, nr))
3015 return 0;
3016
3017 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
3018 /*
3019 * architecture have different format for hugetlbfs
3020 * pmd format and THP pmd format
3021 */
3022 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
3023 PMD_SHIFT, next, flags, pages, nr))
3024 return 0;
3025 } else if (!gup_pte_range(pmd, pmdp, addr, next, flags, pages, nr))
3026 return 0;
3027 } while (pmdp++, addr = next, addr != end);
3028
3029 return 1;
3030}
3031
3032static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
3033 unsigned int flags, struct page **pages, int *nr)
3034{
3035 unsigned long next;
3036 pud_t *pudp;
3037
3038 pudp = pud_offset_lockless(p4dp, p4d, addr);
3039 do {
3040 pud_t pud = READ_ONCE(*pudp);
3041
3042 next = pud_addr_end(addr, end);
3043 if (unlikely(!pud_present(pud)))
3044 return 0;
3045 if (unlikely(pud_huge(pud) || pud_devmap(pud))) {
3046 if (!gup_huge_pud(pud, pudp, addr, next, flags,
3047 pages, nr))
3048 return 0;
3049 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
3050 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
3051 PUD_SHIFT, next, flags, pages, nr))
3052 return 0;
3053 } else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
3054 return 0;
3055 } while (pudp++, addr = next, addr != end);
3056
3057 return 1;
3058}
3059
3060static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
3061 unsigned int flags, struct page **pages, int *nr)
3062{
3063 unsigned long next;
3064 p4d_t *p4dp;
3065
3066 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
3067 do {
3068 p4d_t p4d = READ_ONCE(*p4dp);
3069
3070 next = p4d_addr_end(addr, end);
3071 if (p4d_none(p4d))
3072 return 0;
3073 BUILD_BUG_ON(p4d_huge(p4d));
3074 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
3075 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
3076 P4D_SHIFT, next, flags, pages, nr))
3077 return 0;
3078 } else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
3079 return 0;
3080 } while (p4dp++, addr = next, addr != end);
3081
3082 return 1;
3083}
3084
3085static void gup_pgd_range(unsigned long addr, unsigned long end,
3086 unsigned int flags, struct page **pages, int *nr)
3087{
3088 unsigned long next;
3089 pgd_t *pgdp;
3090
3091 pgdp = pgd_offset(current->mm, addr);
3092 do {
3093 pgd_t pgd = READ_ONCE(*pgdp);
3094
3095 next = pgd_addr_end(addr, end);
3096 if (pgd_none(pgd))
3097 return;
3098 if (unlikely(pgd_huge(pgd))) {
3099 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
3100 pages, nr))
3101 return;
3102 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
3103 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
3104 PGDIR_SHIFT, next, flags, pages, nr))
3105 return;
3106 } else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
3107 return;
3108 } while (pgdp++, addr = next, addr != end);
3109}
3110#else
3111static inline void gup_pgd_range(unsigned long addr, unsigned long end,
3112 unsigned int flags, struct page **pages, int *nr)
3113{
3114}
3115#endif /* CONFIG_HAVE_FAST_GUP */
3116
3117#ifndef gup_fast_permitted
3118/*
3119 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
3120 * we need to fall back to the slow version:
3121 */
3122static bool gup_fast_permitted(unsigned long start, unsigned long end)
3123{
3124 return true;
3125}
3126#endif
3127
3128static unsigned long lockless_pages_from_mm(unsigned long start,
3129 unsigned long end,
3130 unsigned int gup_flags,
3131 struct page **pages)
3132{
3133 unsigned long flags;
3134 int nr_pinned = 0;
3135 unsigned seq;
3136
3137 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP) ||
3138 !gup_fast_permitted(start, end))
3139 return 0;
3140
3141 if (gup_flags & FOLL_PIN) {
3142 seq = raw_read_seqcount(¤t->mm->write_protect_seq);
3143 if (seq & 1)
3144 return 0;
3145 }
3146
3147 /*
3148 * Disable interrupts. The nested form is used, in order to allow full,
3149 * general purpose use of this routine.
3150 *
3151 * With interrupts disabled, we block page table pages from being freed
3152 * from under us. See struct mmu_table_batch comments in
3153 * include/asm-generic/tlb.h for more details.
3154 *
3155 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
3156 * that come from THPs splitting.
3157 */
3158 local_irq_save(flags);
3159 gup_pgd_range(start, end, gup_flags, pages, &nr_pinned);
3160 local_irq_restore(flags);
3161
3162 /*
3163 * When pinning pages for DMA there could be a concurrent write protect
3164 * from fork() via copy_page_range(), in this case always fail fast GUP.
3165 */
3166 if (gup_flags & FOLL_PIN) {
3167 if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) {
3168 unpin_user_pages_lockless(pages, nr_pinned);
3169 return 0;
3170 } else {
3171 sanity_check_pinned_pages(pages, nr_pinned);
3172 }
3173 }
3174 return nr_pinned;
3175}
3176
3177static int internal_get_user_pages_fast(unsigned long start,
3178 unsigned long nr_pages,
3179 unsigned int gup_flags,
3180 struct page **pages)
3181{
3182 unsigned long len, end;
3183 unsigned long nr_pinned;
3184 int locked = 0;
3185 int ret;
3186
3187 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
3188 FOLL_FORCE | FOLL_PIN | FOLL_GET |
3189 FOLL_FAST_ONLY | FOLL_NOFAULT |
3190 FOLL_PCI_P2PDMA | FOLL_HONOR_NUMA_FAULT)))
3191 return -EINVAL;
3192
3193 if (gup_flags & FOLL_PIN)
3194 mm_set_has_pinned_flag(¤t->mm->flags);
3195
3196 if (!(gup_flags & FOLL_FAST_ONLY))
3197 might_lock_read(¤t->mm->mmap_lock);
3198
3199 start = untagged_addr(start) & PAGE_MASK;
3200 len = nr_pages << PAGE_SHIFT;
3201 if (check_add_overflow(start, len, &end))
3202 return -EOVERFLOW;
3203 if (end > TASK_SIZE_MAX)
3204 return -EFAULT;
3205 if (unlikely(!access_ok((void __user *)start, len)))
3206 return -EFAULT;
3207
3208 nr_pinned = lockless_pages_from_mm(start, end, gup_flags, pages);
3209 if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
3210 return nr_pinned;
3211
3212 /* Slow path: try to get the remaining pages with get_user_pages */
3213 start += nr_pinned << PAGE_SHIFT;
3214 pages += nr_pinned;
3215 ret = __gup_longterm_locked(current->mm, start, nr_pages - nr_pinned,
3216 pages, &locked,
3217 gup_flags | FOLL_TOUCH | FOLL_UNLOCKABLE);
3218 if (ret < 0) {
3219 /*
3220 * The caller has to unpin the pages we already pinned so
3221 * returning -errno is not an option
3222 */
3223 if (nr_pinned)
3224 return nr_pinned;
3225 return ret;
3226 }
3227 return ret + nr_pinned;
3228}
3229
3230/**
3231 * get_user_pages_fast_only() - pin user pages in memory
3232 * @start: starting user address
3233 * @nr_pages: number of pages from start to pin
3234 * @gup_flags: flags modifying pin behaviour
3235 * @pages: array that receives pointers to the pages pinned.
3236 * Should be at least nr_pages long.
3237 *
3238 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
3239 * the regular GUP.
3240 *
3241 * If the architecture does not support this function, simply return with no
3242 * pages pinned.
3243 *
3244 * Careful, careful! COW breaking can go either way, so a non-write
3245 * access can get ambiguous page results. If you call this function without
3246 * 'write' set, you'd better be sure that you're ok with that ambiguity.
3247 */
3248int get_user_pages_fast_only(unsigned long start, int nr_pages,
3249 unsigned int gup_flags, struct page **pages)
3250{
3251 /*
3252 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
3253 * because gup fast is always a "pin with a +1 page refcount" request.
3254 *
3255 * FOLL_FAST_ONLY is required in order to match the API description of
3256 * this routine: no fall back to regular ("slow") GUP.
3257 */
3258 if (!is_valid_gup_args(pages, NULL, &gup_flags,
3259 FOLL_GET | FOLL_FAST_ONLY))
3260 return -EINVAL;
3261
3262 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3263}
3264EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
3265
3266/**
3267 * get_user_pages_fast() - pin user pages in memory
3268 * @start: starting user address
3269 * @nr_pages: number of pages from start to pin
3270 * @gup_flags: flags modifying pin behaviour
3271 * @pages: array that receives pointers to the pages pinned.
3272 * Should be at least nr_pages long.
3273 *
3274 * Attempt to pin user pages in memory without taking mm->mmap_lock.
3275 * If not successful, it will fall back to taking the lock and
3276 * calling get_user_pages().
3277 *
3278 * Returns number of pages pinned. This may be fewer than the number requested.
3279 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
3280 * -errno.
3281 */
3282int get_user_pages_fast(unsigned long start, int nr_pages,
3283 unsigned int gup_flags, struct page **pages)
3284{
3285 /*
3286 * The caller may or may not have explicitly set FOLL_GET; either way is
3287 * OK. However, internally (within mm/gup.c), gup fast variants must set
3288 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
3289 * request.
3290 */
3291 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_GET))
3292 return -EINVAL;
3293 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3294}
3295EXPORT_SYMBOL_GPL(get_user_pages_fast);
3296
3297/**
3298 * pin_user_pages_fast() - pin user pages in memory without taking locks
3299 *
3300 * @start: starting user address
3301 * @nr_pages: number of pages from start to pin
3302 * @gup_flags: flags modifying pin behaviour
3303 * @pages: array that receives pointers to the pages pinned.
3304 * Should be at least nr_pages long.
3305 *
3306 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
3307 * get_user_pages_fast() for documentation on the function arguments, because
3308 * the arguments here are identical.
3309 *
3310 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3311 * see Documentation/core-api/pin_user_pages.rst for further details.
3312 *
3313 * Note that if a zero_page is amongst the returned pages, it will not have
3314 * pins in it and unpin_user_page() will not remove pins from it.
3315 */
3316int pin_user_pages_fast(unsigned long start, int nr_pages,
3317 unsigned int gup_flags, struct page **pages)
3318{
3319 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN))
3320 return -EINVAL;
3321 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3322}
3323EXPORT_SYMBOL_GPL(pin_user_pages_fast);
3324
3325/**
3326 * pin_user_pages_remote() - pin pages of a remote process
3327 *
3328 * @mm: mm_struct of target mm
3329 * @start: starting user address
3330 * @nr_pages: number of pages from start to pin
3331 * @gup_flags: flags modifying lookup behaviour
3332 * @pages: array that receives pointers to the pages pinned.
3333 * Should be at least nr_pages long.
3334 * @locked: pointer to lock flag indicating whether lock is held and
3335 * subsequently whether VM_FAULT_RETRY functionality can be
3336 * utilised. Lock must initially be held.
3337 *
3338 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
3339 * get_user_pages_remote() for documentation on the function arguments, because
3340 * the arguments here are identical.
3341 *
3342 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3343 * see Documentation/core-api/pin_user_pages.rst for details.
3344 *
3345 * Note that if a zero_page is amongst the returned pages, it will not have
3346 * pins in it and unpin_user_page*() will not remove pins from it.
3347 */
3348long pin_user_pages_remote(struct mm_struct *mm,
3349 unsigned long start, unsigned long nr_pages,
3350 unsigned int gup_flags, struct page **pages,
3351 int *locked)
3352{
3353 int local_locked = 1;
3354
3355 if (!is_valid_gup_args(pages, locked, &gup_flags,
3356 FOLL_PIN | FOLL_TOUCH | FOLL_REMOTE))
3357 return 0;
3358 return __gup_longterm_locked(mm, start, nr_pages, pages,
3359 locked ? locked : &local_locked,
3360 gup_flags);
3361}
3362EXPORT_SYMBOL(pin_user_pages_remote);
3363
3364/**
3365 * pin_user_pages() - pin user pages in memory for use by other devices
3366 *
3367 * @start: starting user address
3368 * @nr_pages: number of pages from start to pin
3369 * @gup_flags: flags modifying lookup behaviour
3370 * @pages: array that receives pointers to the pages pinned.
3371 * Should be at least nr_pages long.
3372 *
3373 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
3374 * FOLL_PIN is set.
3375 *
3376 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3377 * see Documentation/core-api/pin_user_pages.rst for details.
3378 *
3379 * Note that if a zero_page is amongst the returned pages, it will not have
3380 * pins in it and unpin_user_page*() will not remove pins from it.
3381 */
3382long pin_user_pages(unsigned long start, unsigned long nr_pages,
3383 unsigned int gup_flags, struct page **pages)
3384{
3385 int locked = 1;
3386
3387 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN))
3388 return 0;
3389 return __gup_longterm_locked(current->mm, start, nr_pages,
3390 pages, &locked, gup_flags);
3391}
3392EXPORT_SYMBOL(pin_user_pages);
3393
3394/*
3395 * pin_user_pages_unlocked() is the FOLL_PIN variant of
3396 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
3397 * FOLL_PIN and rejects FOLL_GET.
3398 *
3399 * Note that if a zero_page is amongst the returned pages, it will not have
3400 * pins in it and unpin_user_page*() will not remove pins from it.
3401 */
3402long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
3403 struct page **pages, unsigned int gup_flags)
3404{
3405 int locked = 0;
3406
3407 if (!is_valid_gup_args(pages, NULL, &gup_flags,
3408 FOLL_PIN | FOLL_TOUCH | FOLL_UNLOCKABLE))
3409 return 0;
3410
3411 return __gup_longterm_locked(current->mm, start, nr_pages, pages,
3412 &locked, gup_flags);
3413}
3414EXPORT_SYMBOL(pin_user_pages_unlocked);