Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2#include <linux/kernel.h>
3#include <linux/errno.h>
4#include <linux/err.h>
5#include <linux/spinlock.h>
6
7#include <linux/mm.h>
8#include <linux/memremap.h>
9#include <linux/pagemap.h>
10#include <linux/rmap.h>
11#include <linux/swap.h>
12#include <linux/swapops.h>
13#include <linux/secretmem.h>
14
15#include <linux/sched/signal.h>
16#include <linux/rwsem.h>
17#include <linux/hugetlb.h>
18#include <linux/migrate.h>
19#include <linux/mm_inline.h>
20#include <linux/sched/mm.h>
21#include <linux/shmem_fs.h>
22
23#include <asm/mmu_context.h>
24#include <asm/tlbflush.h>
25
26#include "internal.h"
27
28struct follow_page_context {
29 struct dev_pagemap *pgmap;
30 unsigned int page_mask;
31};
32
33static inline void sanity_check_pinned_pages(struct page **pages,
34 unsigned long npages)
35{
36 if (!IS_ENABLED(CONFIG_DEBUG_VM))
37 return;
38
39 /*
40 * We only pin anonymous pages if they are exclusive. Once pinned, we
41 * can no longer turn them possibly shared and PageAnonExclusive() will
42 * stick around until the page is freed.
43 *
44 * We'd like to verify that our pinned anonymous pages are still mapped
45 * exclusively. The issue with anon THP is that we don't know how
46 * they are/were mapped when pinning them. However, for anon
47 * THP we can assume that either the given page (PTE-mapped THP) or
48 * the head page (PMD-mapped THP) should be PageAnonExclusive(). If
49 * neither is the case, there is certainly something wrong.
50 */
51 for (; npages; npages--, pages++) {
52 struct page *page = *pages;
53 struct folio *folio = page_folio(page);
54
55 if (is_zero_page(page) ||
56 !folio_test_anon(folio))
57 continue;
58 if (!folio_test_large(folio) || folio_test_hugetlb(folio))
59 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page), page);
60 else
61 /* Either a PTE-mapped or a PMD-mapped THP. */
62 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page) &&
63 !PageAnonExclusive(page), page);
64 }
65}
66
67/*
68 * Return the folio with ref appropriately incremented,
69 * or NULL if that failed.
70 */
71static inline struct folio *try_get_folio(struct page *page, int refs)
72{
73 struct folio *folio;
74
75retry:
76 folio = page_folio(page);
77 if (WARN_ON_ONCE(folio_ref_count(folio) < 0))
78 return NULL;
79 if (unlikely(!folio_ref_try_add_rcu(folio, refs)))
80 return NULL;
81
82 /*
83 * At this point we have a stable reference to the folio; but it
84 * could be that between calling page_folio() and the refcount
85 * increment, the folio was split, in which case we'd end up
86 * holding a reference on a folio that has nothing to do with the page
87 * we were given anymore.
88 * So now that the folio is stable, recheck that the page still
89 * belongs to this folio.
90 */
91 if (unlikely(page_folio(page) != folio)) {
92 if (!put_devmap_managed_page_refs(&folio->page, refs))
93 folio_put_refs(folio, refs);
94 goto retry;
95 }
96
97 return folio;
98}
99
100/**
101 * try_grab_folio() - Attempt to get or pin a folio.
102 * @page: pointer to page to be grabbed
103 * @refs: the value to (effectively) add to the folio's refcount
104 * @flags: gup flags: these are the FOLL_* flag values.
105 *
106 * "grab" names in this file mean, "look at flags to decide whether to use
107 * FOLL_PIN or FOLL_GET behavior, when incrementing the folio's refcount.
108 *
109 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
110 * same time. (That's true throughout the get_user_pages*() and
111 * pin_user_pages*() APIs.) Cases:
112 *
113 * FOLL_GET: folio's refcount will be incremented by @refs.
114 *
115 * FOLL_PIN on large folios: folio's refcount will be incremented by
116 * @refs, and its pincount will be incremented by @refs.
117 *
118 * FOLL_PIN on single-page folios: folio's refcount will be incremented by
119 * @refs * GUP_PIN_COUNTING_BIAS.
120 *
121 * Return: The folio containing @page (with refcount appropriately
122 * incremented) for success, or NULL upon failure. If neither FOLL_GET
123 * nor FOLL_PIN was set, that's considered failure, and furthermore,
124 * a likely bug in the caller, so a warning is also emitted.
125 */
126struct folio *try_grab_folio(struct page *page, int refs, unsigned int flags)
127{
128 struct folio *folio;
129
130 if (WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == 0))
131 return NULL;
132
133 if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
134 return NULL;
135
136 if (flags & FOLL_GET)
137 return try_get_folio(page, refs);
138
139 /* FOLL_PIN is set */
140
141 /*
142 * Don't take a pin on the zero page - it's not going anywhere
143 * and it is used in a *lot* of places.
144 */
145 if (is_zero_page(page))
146 return page_folio(page);
147
148 folio = try_get_folio(page, refs);
149 if (!folio)
150 return NULL;
151
152 /*
153 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
154 * right zone, so fail and let the caller fall back to the slow
155 * path.
156 */
157 if (unlikely((flags & FOLL_LONGTERM) &&
158 !folio_is_longterm_pinnable(folio))) {
159 if (!put_devmap_managed_page_refs(&folio->page, refs))
160 folio_put_refs(folio, refs);
161 return NULL;
162 }
163
164 /*
165 * When pinning a large folio, use an exact count to track it.
166 *
167 * However, be sure to *also* increment the normal folio
168 * refcount field at least once, so that the folio really
169 * is pinned. That's why the refcount from the earlier
170 * try_get_folio() is left intact.
171 */
172 if (folio_test_large(folio))
173 atomic_add(refs, &folio->_pincount);
174 else
175 folio_ref_add(folio,
176 refs * (GUP_PIN_COUNTING_BIAS - 1));
177 /*
178 * Adjust the pincount before re-checking the PTE for changes.
179 * This is essentially a smp_mb() and is paired with a memory
180 * barrier in folio_try_share_anon_rmap_*().
181 */
182 smp_mb__after_atomic();
183
184 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs);
185
186 return folio;
187}
188
189static void gup_put_folio(struct folio *folio, int refs, unsigned int flags)
190{
191 if (flags & FOLL_PIN) {
192 if (is_zero_folio(folio))
193 return;
194 node_stat_mod_folio(folio, NR_FOLL_PIN_RELEASED, refs);
195 if (folio_test_large(folio))
196 atomic_sub(refs, &folio->_pincount);
197 else
198 refs *= GUP_PIN_COUNTING_BIAS;
199 }
200
201 if (!put_devmap_managed_page_refs(&folio->page, refs))
202 folio_put_refs(folio, refs);
203}
204
205/**
206 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
207 * @page: pointer to page to be grabbed
208 * @flags: gup flags: these are the FOLL_* flag values.
209 *
210 * This might not do anything at all, depending on the flags argument.
211 *
212 * "grab" names in this file mean, "look at flags to decide whether to use
213 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
214 *
215 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
216 * time. Cases: please see the try_grab_folio() documentation, with
217 * "refs=1".
218 *
219 * Return: 0 for success, or if no action was required (if neither FOLL_PIN
220 * nor FOLL_GET was set, nothing is done). A negative error code for failure:
221 *
222 * -ENOMEM FOLL_GET or FOLL_PIN was set, but the page could not
223 * be grabbed.
224 */
225int __must_check try_grab_page(struct page *page, unsigned int flags)
226{
227 struct folio *folio = page_folio(page);
228
229 if (WARN_ON_ONCE(folio_ref_count(folio) <= 0))
230 return -ENOMEM;
231
232 if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
233 return -EREMOTEIO;
234
235 if (flags & FOLL_GET)
236 folio_ref_inc(folio);
237 else if (flags & FOLL_PIN) {
238 /*
239 * Don't take a pin on the zero page - it's not going anywhere
240 * and it is used in a *lot* of places.
241 */
242 if (is_zero_page(page))
243 return 0;
244
245 /*
246 * Similar to try_grab_folio(): be sure to *also*
247 * increment the normal page refcount field at least once,
248 * so that the page really is pinned.
249 */
250 if (folio_test_large(folio)) {
251 folio_ref_add(folio, 1);
252 atomic_add(1, &folio->_pincount);
253 } else {
254 folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
255 }
256
257 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, 1);
258 }
259
260 return 0;
261}
262
263/**
264 * unpin_user_page() - release a dma-pinned page
265 * @page: pointer to page to be released
266 *
267 * Pages that were pinned via pin_user_pages*() must be released via either
268 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
269 * that such pages can be separately tracked and uniquely handled. In
270 * particular, interactions with RDMA and filesystems need special handling.
271 */
272void unpin_user_page(struct page *page)
273{
274 sanity_check_pinned_pages(&page, 1);
275 gup_put_folio(page_folio(page), 1, FOLL_PIN);
276}
277EXPORT_SYMBOL(unpin_user_page);
278
279/**
280 * folio_add_pin - Try to get an additional pin on a pinned folio
281 * @folio: The folio to be pinned
282 *
283 * Get an additional pin on a folio we already have a pin on. Makes no change
284 * if the folio is a zero_page.
285 */
286void folio_add_pin(struct folio *folio)
287{
288 if (is_zero_folio(folio))
289 return;
290
291 /*
292 * Similar to try_grab_folio(): be sure to *also* increment the normal
293 * page refcount field at least once, so that the page really is
294 * pinned.
295 */
296 if (folio_test_large(folio)) {
297 WARN_ON_ONCE(atomic_read(&folio->_pincount) < 1);
298 folio_ref_inc(folio);
299 atomic_inc(&folio->_pincount);
300 } else {
301 WARN_ON_ONCE(folio_ref_count(folio) < GUP_PIN_COUNTING_BIAS);
302 folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
303 }
304}
305
306static inline struct folio *gup_folio_range_next(struct page *start,
307 unsigned long npages, unsigned long i, unsigned int *ntails)
308{
309 struct page *next = nth_page(start, i);
310 struct folio *folio = page_folio(next);
311 unsigned int nr = 1;
312
313 if (folio_test_large(folio))
314 nr = min_t(unsigned int, npages - i,
315 folio_nr_pages(folio) - folio_page_idx(folio, next));
316
317 *ntails = nr;
318 return folio;
319}
320
321static inline struct folio *gup_folio_next(struct page **list,
322 unsigned long npages, unsigned long i, unsigned int *ntails)
323{
324 struct folio *folio = page_folio(list[i]);
325 unsigned int nr;
326
327 for (nr = i + 1; nr < npages; nr++) {
328 if (page_folio(list[nr]) != folio)
329 break;
330 }
331
332 *ntails = nr - i;
333 return folio;
334}
335
336/**
337 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
338 * @pages: array of pages to be maybe marked dirty, and definitely released.
339 * @npages: number of pages in the @pages array.
340 * @make_dirty: whether to mark the pages dirty
341 *
342 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
343 * variants called on that page.
344 *
345 * For each page in the @pages array, make that page (or its head page, if a
346 * compound page) dirty, if @make_dirty is true, and if the page was previously
347 * listed as clean. In any case, releases all pages using unpin_user_page(),
348 * possibly via unpin_user_pages(), for the non-dirty case.
349 *
350 * Please see the unpin_user_page() documentation for details.
351 *
352 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
353 * required, then the caller should a) verify that this is really correct,
354 * because _lock() is usually required, and b) hand code it:
355 * set_page_dirty_lock(), unpin_user_page().
356 *
357 */
358void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
359 bool make_dirty)
360{
361 unsigned long i;
362 struct folio *folio;
363 unsigned int nr;
364
365 if (!make_dirty) {
366 unpin_user_pages(pages, npages);
367 return;
368 }
369
370 sanity_check_pinned_pages(pages, npages);
371 for (i = 0; i < npages; i += nr) {
372 folio = gup_folio_next(pages, npages, i, &nr);
373 /*
374 * Checking PageDirty at this point may race with
375 * clear_page_dirty_for_io(), but that's OK. Two key
376 * cases:
377 *
378 * 1) This code sees the page as already dirty, so it
379 * skips the call to set_page_dirty(). That could happen
380 * because clear_page_dirty_for_io() called
381 * page_mkclean(), followed by set_page_dirty().
382 * However, now the page is going to get written back,
383 * which meets the original intention of setting it
384 * dirty, so all is well: clear_page_dirty_for_io() goes
385 * on to call TestClearPageDirty(), and write the page
386 * back.
387 *
388 * 2) This code sees the page as clean, so it calls
389 * set_page_dirty(). The page stays dirty, despite being
390 * written back, so it gets written back again in the
391 * next writeback cycle. This is harmless.
392 */
393 if (!folio_test_dirty(folio)) {
394 folio_lock(folio);
395 folio_mark_dirty(folio);
396 folio_unlock(folio);
397 }
398 gup_put_folio(folio, nr, FOLL_PIN);
399 }
400}
401EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
402
403/**
404 * unpin_user_page_range_dirty_lock() - release and optionally dirty
405 * gup-pinned page range
406 *
407 * @page: the starting page of a range maybe marked dirty, and definitely released.
408 * @npages: number of consecutive pages to release.
409 * @make_dirty: whether to mark the pages dirty
410 *
411 * "gup-pinned page range" refers to a range of pages that has had one of the
412 * pin_user_pages() variants called on that page.
413 *
414 * For the page ranges defined by [page .. page+npages], make that range (or
415 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
416 * page range was previously listed as clean.
417 *
418 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
419 * required, then the caller should a) verify that this is really correct,
420 * because _lock() is usually required, and b) hand code it:
421 * set_page_dirty_lock(), unpin_user_page().
422 *
423 */
424void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages,
425 bool make_dirty)
426{
427 unsigned long i;
428 struct folio *folio;
429 unsigned int nr;
430
431 for (i = 0; i < npages; i += nr) {
432 folio = gup_folio_range_next(page, npages, i, &nr);
433 if (make_dirty && !folio_test_dirty(folio)) {
434 folio_lock(folio);
435 folio_mark_dirty(folio);
436 folio_unlock(folio);
437 }
438 gup_put_folio(folio, nr, FOLL_PIN);
439 }
440}
441EXPORT_SYMBOL(unpin_user_page_range_dirty_lock);
442
443static void unpin_user_pages_lockless(struct page **pages, unsigned long npages)
444{
445 unsigned long i;
446 struct folio *folio;
447 unsigned int nr;
448
449 /*
450 * Don't perform any sanity checks because we might have raced with
451 * fork() and some anonymous pages might now actually be shared --
452 * which is why we're unpinning after all.
453 */
454 for (i = 0; i < npages; i += nr) {
455 folio = gup_folio_next(pages, npages, i, &nr);
456 gup_put_folio(folio, nr, FOLL_PIN);
457 }
458}
459
460/**
461 * unpin_user_pages() - release an array of gup-pinned pages.
462 * @pages: array of pages to be marked dirty and released.
463 * @npages: number of pages in the @pages array.
464 *
465 * For each page in the @pages array, release the page using unpin_user_page().
466 *
467 * Please see the unpin_user_page() documentation for details.
468 */
469void unpin_user_pages(struct page **pages, unsigned long npages)
470{
471 unsigned long i;
472 struct folio *folio;
473 unsigned int nr;
474
475 /*
476 * If this WARN_ON() fires, then the system *might* be leaking pages (by
477 * leaving them pinned), but probably not. More likely, gup/pup returned
478 * a hard -ERRNO error to the caller, who erroneously passed it here.
479 */
480 if (WARN_ON(IS_ERR_VALUE(npages)))
481 return;
482
483 sanity_check_pinned_pages(pages, npages);
484 for (i = 0; i < npages; i += nr) {
485 folio = gup_folio_next(pages, npages, i, &nr);
486 gup_put_folio(folio, nr, FOLL_PIN);
487 }
488}
489EXPORT_SYMBOL(unpin_user_pages);
490
491/*
492 * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's
493 * lifecycle. Avoid setting the bit unless necessary, or it might cause write
494 * cache bouncing on large SMP machines for concurrent pinned gups.
495 */
496static inline void mm_set_has_pinned_flag(unsigned long *mm_flags)
497{
498 if (!test_bit(MMF_HAS_PINNED, mm_flags))
499 set_bit(MMF_HAS_PINNED, mm_flags);
500}
501
502#ifdef CONFIG_MMU
503static struct page *no_page_table(struct vm_area_struct *vma,
504 unsigned int flags)
505{
506 /*
507 * When core dumping an enormous anonymous area that nobody
508 * has touched so far, we don't want to allocate unnecessary pages or
509 * page tables. Return error instead of NULL to skip handle_mm_fault,
510 * then get_dump_page() will return NULL to leave a hole in the dump.
511 * But we can only make this optimization where a hole would surely
512 * be zero-filled if handle_mm_fault() actually did handle it.
513 */
514 if ((flags & FOLL_DUMP) &&
515 (vma_is_anonymous(vma) || !vma->vm_ops->fault))
516 return ERR_PTR(-EFAULT);
517 return NULL;
518}
519
520static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
521 pte_t *pte, unsigned int flags)
522{
523 if (flags & FOLL_TOUCH) {
524 pte_t orig_entry = ptep_get(pte);
525 pte_t entry = orig_entry;
526
527 if (flags & FOLL_WRITE)
528 entry = pte_mkdirty(entry);
529 entry = pte_mkyoung(entry);
530
531 if (!pte_same(orig_entry, entry)) {
532 set_pte_at(vma->vm_mm, address, pte, entry);
533 update_mmu_cache(vma, address, pte);
534 }
535 }
536
537 /* Proper page table entry exists, but no corresponding struct page */
538 return -EEXIST;
539}
540
541/* FOLL_FORCE can write to even unwritable PTEs in COW mappings. */
542static inline bool can_follow_write_pte(pte_t pte, struct page *page,
543 struct vm_area_struct *vma,
544 unsigned int flags)
545{
546 /* If the pte is writable, we can write to the page. */
547 if (pte_write(pte))
548 return true;
549
550 /* Maybe FOLL_FORCE is set to override it? */
551 if (!(flags & FOLL_FORCE))
552 return false;
553
554 /* But FOLL_FORCE has no effect on shared mappings */
555 if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
556 return false;
557
558 /* ... or read-only private ones */
559 if (!(vma->vm_flags & VM_MAYWRITE))
560 return false;
561
562 /* ... or already writable ones that just need to take a write fault */
563 if (vma->vm_flags & VM_WRITE)
564 return false;
565
566 /*
567 * See can_change_pte_writable(): we broke COW and could map the page
568 * writable if we have an exclusive anonymous page ...
569 */
570 if (!page || !PageAnon(page) || !PageAnonExclusive(page))
571 return false;
572
573 /* ... and a write-fault isn't required for other reasons. */
574 if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte))
575 return false;
576 return !userfaultfd_pte_wp(vma, pte);
577}
578
579static struct page *follow_page_pte(struct vm_area_struct *vma,
580 unsigned long address, pmd_t *pmd, unsigned int flags,
581 struct dev_pagemap **pgmap)
582{
583 struct mm_struct *mm = vma->vm_mm;
584 struct page *page;
585 spinlock_t *ptl;
586 pte_t *ptep, pte;
587 int ret;
588
589 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
590 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
591 (FOLL_PIN | FOLL_GET)))
592 return ERR_PTR(-EINVAL);
593
594 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
595 if (!ptep)
596 return no_page_table(vma, flags);
597 pte = ptep_get(ptep);
598 if (!pte_present(pte))
599 goto no_page;
600 if (pte_protnone(pte) && !gup_can_follow_protnone(vma, flags))
601 goto no_page;
602
603 page = vm_normal_page(vma, address, pte);
604
605 /*
606 * We only care about anon pages in can_follow_write_pte() and don't
607 * have to worry about pte_devmap() because they are never anon.
608 */
609 if ((flags & FOLL_WRITE) &&
610 !can_follow_write_pte(pte, page, vma, flags)) {
611 page = NULL;
612 goto out;
613 }
614
615 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
616 /*
617 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
618 * case since they are only valid while holding the pgmap
619 * reference.
620 */
621 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
622 if (*pgmap)
623 page = pte_page(pte);
624 else
625 goto no_page;
626 } else if (unlikely(!page)) {
627 if (flags & FOLL_DUMP) {
628 /* Avoid special (like zero) pages in core dumps */
629 page = ERR_PTR(-EFAULT);
630 goto out;
631 }
632
633 if (is_zero_pfn(pte_pfn(pte))) {
634 page = pte_page(pte);
635 } else {
636 ret = follow_pfn_pte(vma, address, ptep, flags);
637 page = ERR_PTR(ret);
638 goto out;
639 }
640 }
641
642 if (!pte_write(pte) && gup_must_unshare(vma, flags, page)) {
643 page = ERR_PTR(-EMLINK);
644 goto out;
645 }
646
647 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
648 !PageAnonExclusive(page), page);
649
650 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
651 ret = try_grab_page(page, flags);
652 if (unlikely(ret)) {
653 page = ERR_PTR(ret);
654 goto out;
655 }
656
657 /*
658 * We need to make the page accessible if and only if we are going
659 * to access its content (the FOLL_PIN case). Please see
660 * Documentation/core-api/pin_user_pages.rst for details.
661 */
662 if (flags & FOLL_PIN) {
663 ret = arch_make_page_accessible(page);
664 if (ret) {
665 unpin_user_page(page);
666 page = ERR_PTR(ret);
667 goto out;
668 }
669 }
670 if (flags & FOLL_TOUCH) {
671 if ((flags & FOLL_WRITE) &&
672 !pte_dirty(pte) && !PageDirty(page))
673 set_page_dirty(page);
674 /*
675 * pte_mkyoung() would be more correct here, but atomic care
676 * is needed to avoid losing the dirty bit: it is easier to use
677 * mark_page_accessed().
678 */
679 mark_page_accessed(page);
680 }
681out:
682 pte_unmap_unlock(ptep, ptl);
683 return page;
684no_page:
685 pte_unmap_unlock(ptep, ptl);
686 if (!pte_none(pte))
687 return NULL;
688 return no_page_table(vma, flags);
689}
690
691static struct page *follow_pmd_mask(struct vm_area_struct *vma,
692 unsigned long address, pud_t *pudp,
693 unsigned int flags,
694 struct follow_page_context *ctx)
695{
696 pmd_t *pmd, pmdval;
697 spinlock_t *ptl;
698 struct page *page;
699 struct mm_struct *mm = vma->vm_mm;
700
701 pmd = pmd_offset(pudp, address);
702 pmdval = pmdp_get_lockless(pmd);
703 if (pmd_none(pmdval))
704 return no_page_table(vma, flags);
705 if (!pmd_present(pmdval))
706 return no_page_table(vma, flags);
707 if (pmd_devmap(pmdval)) {
708 ptl = pmd_lock(mm, pmd);
709 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
710 spin_unlock(ptl);
711 if (page)
712 return page;
713 return no_page_table(vma, flags);
714 }
715 if (likely(!pmd_trans_huge(pmdval)))
716 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
717
718 if (pmd_protnone(pmdval) && !gup_can_follow_protnone(vma, flags))
719 return no_page_table(vma, flags);
720
721 ptl = pmd_lock(mm, pmd);
722 if (unlikely(!pmd_present(*pmd))) {
723 spin_unlock(ptl);
724 return no_page_table(vma, flags);
725 }
726 if (unlikely(!pmd_trans_huge(*pmd))) {
727 spin_unlock(ptl);
728 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
729 }
730 if (flags & FOLL_SPLIT_PMD) {
731 spin_unlock(ptl);
732 split_huge_pmd(vma, pmd, address);
733 /* If pmd was left empty, stuff a page table in there quickly */
734 return pte_alloc(mm, pmd) ? ERR_PTR(-ENOMEM) :
735 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
736 }
737 page = follow_trans_huge_pmd(vma, address, pmd, flags);
738 spin_unlock(ptl);
739 ctx->page_mask = HPAGE_PMD_NR - 1;
740 return page;
741}
742
743static struct page *follow_pud_mask(struct vm_area_struct *vma,
744 unsigned long address, p4d_t *p4dp,
745 unsigned int flags,
746 struct follow_page_context *ctx)
747{
748 pud_t *pud;
749 spinlock_t *ptl;
750 struct page *page;
751 struct mm_struct *mm = vma->vm_mm;
752
753 pud = pud_offset(p4dp, address);
754 if (pud_none(*pud))
755 return no_page_table(vma, flags);
756 if (pud_devmap(*pud)) {
757 ptl = pud_lock(mm, pud);
758 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
759 spin_unlock(ptl);
760 if (page)
761 return page;
762 return no_page_table(vma, flags);
763 }
764 if (unlikely(pud_bad(*pud)))
765 return no_page_table(vma, flags);
766
767 return follow_pmd_mask(vma, address, pud, flags, ctx);
768}
769
770static struct page *follow_p4d_mask(struct vm_area_struct *vma,
771 unsigned long address, pgd_t *pgdp,
772 unsigned int flags,
773 struct follow_page_context *ctx)
774{
775 p4d_t *p4d;
776
777 p4d = p4d_offset(pgdp, address);
778 if (p4d_none(*p4d))
779 return no_page_table(vma, flags);
780 BUILD_BUG_ON(p4d_huge(*p4d));
781 if (unlikely(p4d_bad(*p4d)))
782 return no_page_table(vma, flags);
783
784 return follow_pud_mask(vma, address, p4d, flags, ctx);
785}
786
787/**
788 * follow_page_mask - look up a page descriptor from a user-virtual address
789 * @vma: vm_area_struct mapping @address
790 * @address: virtual address to look up
791 * @flags: flags modifying lookup behaviour
792 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
793 * pointer to output page_mask
794 *
795 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
796 *
797 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
798 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
799 *
800 * When getting an anonymous page and the caller has to trigger unsharing
801 * of a shared anonymous page first, -EMLINK is returned. The caller should
802 * trigger a fault with FAULT_FLAG_UNSHARE set. Note that unsharing is only
803 * relevant with FOLL_PIN and !FOLL_WRITE.
804 *
805 * On output, the @ctx->page_mask is set according to the size of the page.
806 *
807 * Return: the mapped (struct page *), %NULL if no mapping exists, or
808 * an error pointer if there is a mapping to something not represented
809 * by a page descriptor (see also vm_normal_page()).
810 */
811static struct page *follow_page_mask(struct vm_area_struct *vma,
812 unsigned long address, unsigned int flags,
813 struct follow_page_context *ctx)
814{
815 pgd_t *pgd;
816 struct mm_struct *mm = vma->vm_mm;
817
818 ctx->page_mask = 0;
819
820 /*
821 * Call hugetlb_follow_page_mask for hugetlb vmas as it will use
822 * special hugetlb page table walking code. This eliminates the
823 * need to check for hugetlb entries in the general walking code.
824 */
825 if (is_vm_hugetlb_page(vma))
826 return hugetlb_follow_page_mask(vma, address, flags,
827 &ctx->page_mask);
828
829 pgd = pgd_offset(mm, address);
830
831 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
832 return no_page_table(vma, flags);
833
834 return follow_p4d_mask(vma, address, pgd, flags, ctx);
835}
836
837struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
838 unsigned int foll_flags)
839{
840 struct follow_page_context ctx = { NULL };
841 struct page *page;
842
843 if (vma_is_secretmem(vma))
844 return NULL;
845
846 if (WARN_ON_ONCE(foll_flags & FOLL_PIN))
847 return NULL;
848
849 /*
850 * We never set FOLL_HONOR_NUMA_FAULT because callers don't expect
851 * to fail on PROT_NONE-mapped pages.
852 */
853 page = follow_page_mask(vma, address, foll_flags, &ctx);
854 if (ctx.pgmap)
855 put_dev_pagemap(ctx.pgmap);
856 return page;
857}
858
859static int get_gate_page(struct mm_struct *mm, unsigned long address,
860 unsigned int gup_flags, struct vm_area_struct **vma,
861 struct page **page)
862{
863 pgd_t *pgd;
864 p4d_t *p4d;
865 pud_t *pud;
866 pmd_t *pmd;
867 pte_t *pte;
868 pte_t entry;
869 int ret = -EFAULT;
870
871 /* user gate pages are read-only */
872 if (gup_flags & FOLL_WRITE)
873 return -EFAULT;
874 if (address > TASK_SIZE)
875 pgd = pgd_offset_k(address);
876 else
877 pgd = pgd_offset_gate(mm, address);
878 if (pgd_none(*pgd))
879 return -EFAULT;
880 p4d = p4d_offset(pgd, address);
881 if (p4d_none(*p4d))
882 return -EFAULT;
883 pud = pud_offset(p4d, address);
884 if (pud_none(*pud))
885 return -EFAULT;
886 pmd = pmd_offset(pud, address);
887 if (!pmd_present(*pmd))
888 return -EFAULT;
889 pte = pte_offset_map(pmd, address);
890 if (!pte)
891 return -EFAULT;
892 entry = ptep_get(pte);
893 if (pte_none(entry))
894 goto unmap;
895 *vma = get_gate_vma(mm);
896 if (!page)
897 goto out;
898 *page = vm_normal_page(*vma, address, entry);
899 if (!*page) {
900 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(entry)))
901 goto unmap;
902 *page = pte_page(entry);
903 }
904 ret = try_grab_page(*page, gup_flags);
905 if (unlikely(ret))
906 goto unmap;
907out:
908 ret = 0;
909unmap:
910 pte_unmap(pte);
911 return ret;
912}
913
914/*
915 * mmap_lock must be held on entry. If @flags has FOLL_UNLOCKABLE but not
916 * FOLL_NOWAIT, the mmap_lock may be released. If it is, *@locked will be set
917 * to 0 and -EBUSY returned.
918 */
919static int faultin_page(struct vm_area_struct *vma,
920 unsigned long address, unsigned int *flags, bool unshare,
921 int *locked)
922{
923 unsigned int fault_flags = 0;
924 vm_fault_t ret;
925
926 if (*flags & FOLL_NOFAULT)
927 return -EFAULT;
928 if (*flags & FOLL_WRITE)
929 fault_flags |= FAULT_FLAG_WRITE;
930 if (*flags & FOLL_REMOTE)
931 fault_flags |= FAULT_FLAG_REMOTE;
932 if (*flags & FOLL_UNLOCKABLE) {
933 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
934 /*
935 * FAULT_FLAG_INTERRUPTIBLE is opt-in. GUP callers must set
936 * FOLL_INTERRUPTIBLE to enable FAULT_FLAG_INTERRUPTIBLE.
937 * That's because some callers may not be prepared to
938 * handle early exits caused by non-fatal signals.
939 */
940 if (*flags & FOLL_INTERRUPTIBLE)
941 fault_flags |= FAULT_FLAG_INTERRUPTIBLE;
942 }
943 if (*flags & FOLL_NOWAIT)
944 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
945 if (*flags & FOLL_TRIED) {
946 /*
947 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
948 * can co-exist
949 */
950 fault_flags |= FAULT_FLAG_TRIED;
951 }
952 if (unshare) {
953 fault_flags |= FAULT_FLAG_UNSHARE;
954 /* FAULT_FLAG_WRITE and FAULT_FLAG_UNSHARE are incompatible */
955 VM_BUG_ON(fault_flags & FAULT_FLAG_WRITE);
956 }
957
958 ret = handle_mm_fault(vma, address, fault_flags, NULL);
959
960 if (ret & VM_FAULT_COMPLETED) {
961 /*
962 * With FAULT_FLAG_RETRY_NOWAIT we'll never release the
963 * mmap lock in the page fault handler. Sanity check this.
964 */
965 WARN_ON_ONCE(fault_flags & FAULT_FLAG_RETRY_NOWAIT);
966 *locked = 0;
967
968 /*
969 * We should do the same as VM_FAULT_RETRY, but let's not
970 * return -EBUSY since that's not reflecting the reality of
971 * what has happened - we've just fully completed a page
972 * fault, with the mmap lock released. Use -EAGAIN to show
973 * that we want to take the mmap lock _again_.
974 */
975 return -EAGAIN;
976 }
977
978 if (ret & VM_FAULT_ERROR) {
979 int err = vm_fault_to_errno(ret, *flags);
980
981 if (err)
982 return err;
983 BUG();
984 }
985
986 if (ret & VM_FAULT_RETRY) {
987 if (!(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
988 *locked = 0;
989 return -EBUSY;
990 }
991
992 return 0;
993}
994
995/*
996 * Writing to file-backed mappings which require folio dirty tracking using GUP
997 * is a fundamentally broken operation, as kernel write access to GUP mappings
998 * do not adhere to the semantics expected by a file system.
999 *
1000 * Consider the following scenario:-
1001 *
1002 * 1. A folio is written to via GUP which write-faults the memory, notifying
1003 * the file system and dirtying the folio.
1004 * 2. Later, writeback is triggered, resulting in the folio being cleaned and
1005 * the PTE being marked read-only.
1006 * 3. The GUP caller writes to the folio, as it is mapped read/write via the
1007 * direct mapping.
1008 * 4. The GUP caller, now done with the page, unpins it and sets it dirty
1009 * (though it does not have to).
1010 *
1011 * This results in both data being written to a folio without writenotify, and
1012 * the folio being dirtied unexpectedly (if the caller decides to do so).
1013 */
1014static bool writable_file_mapping_allowed(struct vm_area_struct *vma,
1015 unsigned long gup_flags)
1016{
1017 /*
1018 * If we aren't pinning then no problematic write can occur. A long term
1019 * pin is the most egregious case so this is the case we disallow.
1020 */
1021 if ((gup_flags & (FOLL_PIN | FOLL_LONGTERM)) !=
1022 (FOLL_PIN | FOLL_LONGTERM))
1023 return true;
1024
1025 /*
1026 * If the VMA does not require dirty tracking then no problematic write
1027 * can occur either.
1028 */
1029 return !vma_needs_dirty_tracking(vma);
1030}
1031
1032static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
1033{
1034 vm_flags_t vm_flags = vma->vm_flags;
1035 int write = (gup_flags & FOLL_WRITE);
1036 int foreign = (gup_flags & FOLL_REMOTE);
1037 bool vma_anon = vma_is_anonymous(vma);
1038
1039 if (vm_flags & (VM_IO | VM_PFNMAP))
1040 return -EFAULT;
1041
1042 if ((gup_flags & FOLL_ANON) && !vma_anon)
1043 return -EFAULT;
1044
1045 if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma))
1046 return -EOPNOTSUPP;
1047
1048 if (vma_is_secretmem(vma))
1049 return -EFAULT;
1050
1051 if (write) {
1052 if (!vma_anon &&
1053 !writable_file_mapping_allowed(vma, gup_flags))
1054 return -EFAULT;
1055
1056 if (!(vm_flags & VM_WRITE) || (vm_flags & VM_SHADOW_STACK)) {
1057 if (!(gup_flags & FOLL_FORCE))
1058 return -EFAULT;
1059 /* hugetlb does not support FOLL_FORCE|FOLL_WRITE. */
1060 if (is_vm_hugetlb_page(vma))
1061 return -EFAULT;
1062 /*
1063 * We used to let the write,force case do COW in a
1064 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
1065 * set a breakpoint in a read-only mapping of an
1066 * executable, without corrupting the file (yet only
1067 * when that file had been opened for writing!).
1068 * Anon pages in shared mappings are surprising: now
1069 * just reject it.
1070 */
1071 if (!is_cow_mapping(vm_flags))
1072 return -EFAULT;
1073 }
1074 } else if (!(vm_flags & VM_READ)) {
1075 if (!(gup_flags & FOLL_FORCE))
1076 return -EFAULT;
1077 /*
1078 * Is there actually any vma we can reach here which does not
1079 * have VM_MAYREAD set?
1080 */
1081 if (!(vm_flags & VM_MAYREAD))
1082 return -EFAULT;
1083 }
1084 /*
1085 * gups are always data accesses, not instruction
1086 * fetches, so execute=false here
1087 */
1088 if (!arch_vma_access_permitted(vma, write, false, foreign))
1089 return -EFAULT;
1090 return 0;
1091}
1092
1093/*
1094 * This is "vma_lookup()", but with a warning if we would have
1095 * historically expanded the stack in the GUP code.
1096 */
1097static struct vm_area_struct *gup_vma_lookup(struct mm_struct *mm,
1098 unsigned long addr)
1099{
1100#ifdef CONFIG_STACK_GROWSUP
1101 return vma_lookup(mm, addr);
1102#else
1103 static volatile unsigned long next_warn;
1104 struct vm_area_struct *vma;
1105 unsigned long now, next;
1106
1107 vma = find_vma(mm, addr);
1108 if (!vma || (addr >= vma->vm_start))
1109 return vma;
1110
1111 /* Only warn for half-way relevant accesses */
1112 if (!(vma->vm_flags & VM_GROWSDOWN))
1113 return NULL;
1114 if (vma->vm_start - addr > 65536)
1115 return NULL;
1116
1117 /* Let's not warn more than once an hour.. */
1118 now = jiffies; next = next_warn;
1119 if (next && time_before(now, next))
1120 return NULL;
1121 next_warn = now + 60*60*HZ;
1122
1123 /* Let people know things may have changed. */
1124 pr_warn("GUP no longer grows the stack in %s (%d): %lx-%lx (%lx)\n",
1125 current->comm, task_pid_nr(current),
1126 vma->vm_start, vma->vm_end, addr);
1127 dump_stack();
1128 return NULL;
1129#endif
1130}
1131
1132/**
1133 * __get_user_pages() - pin user pages in memory
1134 * @mm: mm_struct of target mm
1135 * @start: starting user address
1136 * @nr_pages: number of pages from start to pin
1137 * @gup_flags: flags modifying pin behaviour
1138 * @pages: array that receives pointers to the pages pinned.
1139 * Should be at least nr_pages long. Or NULL, if caller
1140 * only intends to ensure the pages are faulted in.
1141 * @locked: whether we're still with the mmap_lock held
1142 *
1143 * Returns either number of pages pinned (which may be less than the
1144 * number requested), or an error. Details about the return value:
1145 *
1146 * -- If nr_pages is 0, returns 0.
1147 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1148 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1149 * pages pinned. Again, this may be less than nr_pages.
1150 * -- 0 return value is possible when the fault would need to be retried.
1151 *
1152 * The caller is responsible for releasing returned @pages, via put_page().
1153 *
1154 * Must be called with mmap_lock held. It may be released. See below.
1155 *
1156 * __get_user_pages walks a process's page tables and takes a reference to
1157 * each struct page that each user address corresponds to at a given
1158 * instant. That is, it takes the page that would be accessed if a user
1159 * thread accesses the given user virtual address at that instant.
1160 *
1161 * This does not guarantee that the page exists in the user mappings when
1162 * __get_user_pages returns, and there may even be a completely different
1163 * page there in some cases (eg. if mmapped pagecache has been invalidated
1164 * and subsequently re-faulted). However it does guarantee that the page
1165 * won't be freed completely. And mostly callers simply care that the page
1166 * contains data that was valid *at some point in time*. Typically, an IO
1167 * or similar operation cannot guarantee anything stronger anyway because
1168 * locks can't be held over the syscall boundary.
1169 *
1170 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1171 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1172 * appropriate) must be called after the page is finished with, and
1173 * before put_page is called.
1174 *
1175 * If FOLL_UNLOCKABLE is set without FOLL_NOWAIT then the mmap_lock may
1176 * be released. If this happens *@locked will be set to 0 on return.
1177 *
1178 * A caller using such a combination of @gup_flags must therefore hold the
1179 * mmap_lock for reading only, and recognize when it's been released. Otherwise,
1180 * it must be held for either reading or writing and will not be released.
1181 *
1182 * In most cases, get_user_pages or get_user_pages_fast should be used
1183 * instead of __get_user_pages. __get_user_pages should be used only if
1184 * you need some special @gup_flags.
1185 */
1186static long __get_user_pages(struct mm_struct *mm,
1187 unsigned long start, unsigned long nr_pages,
1188 unsigned int gup_flags, struct page **pages,
1189 int *locked)
1190{
1191 long ret = 0, i = 0;
1192 struct vm_area_struct *vma = NULL;
1193 struct follow_page_context ctx = { NULL };
1194
1195 if (!nr_pages)
1196 return 0;
1197
1198 start = untagged_addr_remote(mm, start);
1199
1200 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1201
1202 do {
1203 struct page *page;
1204 unsigned int foll_flags = gup_flags;
1205 unsigned int page_increm;
1206
1207 /* first iteration or cross vma bound */
1208 if (!vma || start >= vma->vm_end) {
1209 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);
1#include <linux/kernel.h>
2#include <linux/errno.h>
3#include <linux/err.h>
4#include <linux/spinlock.h>
5
6#include <linux/mm.h>
7#include <linux/memremap.h>
8#include <linux/pagemap.h>
9#include <linux/rmap.h>
10#include <linux/swap.h>
11#include <linux/swapops.h>
12
13#include <linux/sched.h>
14#include <linux/rwsem.h>
15#include <linux/hugetlb.h>
16
17#include <asm/mmu_context.h>
18#include <asm/pgtable.h>
19#include <asm/tlbflush.h>
20
21#include "internal.h"
22
23static struct page *no_page_table(struct vm_area_struct *vma,
24 unsigned int flags)
25{
26 /*
27 * When core dumping an enormous anonymous area that nobody
28 * has touched so far, we don't want to allocate unnecessary pages or
29 * page tables. Return error instead of NULL to skip handle_mm_fault,
30 * then get_dump_page() will return NULL to leave a hole in the dump.
31 * But we can only make this optimization where a hole would surely
32 * be zero-filled if handle_mm_fault() actually did handle it.
33 */
34 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
35 return ERR_PTR(-EFAULT);
36 return NULL;
37}
38
39static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
40 pte_t *pte, unsigned int flags)
41{
42 /* No page to get reference */
43 if (flags & FOLL_GET)
44 return -EFAULT;
45
46 if (flags & FOLL_TOUCH) {
47 pte_t entry = *pte;
48
49 if (flags & FOLL_WRITE)
50 entry = pte_mkdirty(entry);
51 entry = pte_mkyoung(entry);
52
53 if (!pte_same(*pte, entry)) {
54 set_pte_at(vma->vm_mm, address, pte, entry);
55 update_mmu_cache(vma, address, pte);
56 }
57 }
58
59 /* Proper page table entry exists, but no corresponding struct page */
60 return -EEXIST;
61}
62
63static struct page *follow_page_pte(struct vm_area_struct *vma,
64 unsigned long address, pmd_t *pmd, unsigned int flags)
65{
66 struct mm_struct *mm = vma->vm_mm;
67 struct dev_pagemap *pgmap = NULL;
68 struct page *page;
69 spinlock_t *ptl;
70 pte_t *ptep, pte;
71
72retry:
73 if (unlikely(pmd_bad(*pmd)))
74 return no_page_table(vma, flags);
75
76 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
77 pte = *ptep;
78 if (!pte_present(pte)) {
79 swp_entry_t entry;
80 /*
81 * KSM's break_ksm() relies upon recognizing a ksm page
82 * even while it is being migrated, so for that case we
83 * need migration_entry_wait().
84 */
85 if (likely(!(flags & FOLL_MIGRATION)))
86 goto no_page;
87 if (pte_none(pte))
88 goto no_page;
89 entry = pte_to_swp_entry(pte);
90 if (!is_migration_entry(entry))
91 goto no_page;
92 pte_unmap_unlock(ptep, ptl);
93 migration_entry_wait(mm, pmd, address);
94 goto retry;
95 }
96 if ((flags & FOLL_NUMA) && pte_protnone(pte))
97 goto no_page;
98 if ((flags & FOLL_WRITE) && !pte_write(pte)) {
99 pte_unmap_unlock(ptep, ptl);
100 return NULL;
101 }
102
103 page = vm_normal_page(vma, address, pte);
104 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
105 /*
106 * Only return device mapping pages in the FOLL_GET case since
107 * they are only valid while holding the pgmap reference.
108 */
109 pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
110 if (pgmap)
111 page = pte_page(pte);
112 else
113 goto no_page;
114 } else if (unlikely(!page)) {
115 if (flags & FOLL_DUMP) {
116 /* Avoid special (like zero) pages in core dumps */
117 page = ERR_PTR(-EFAULT);
118 goto out;
119 }
120
121 if (is_zero_pfn(pte_pfn(pte))) {
122 page = pte_page(pte);
123 } else {
124 int ret;
125
126 ret = follow_pfn_pte(vma, address, ptep, flags);
127 page = ERR_PTR(ret);
128 goto out;
129 }
130 }
131
132 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
133 int ret;
134 get_page(page);
135 pte_unmap_unlock(ptep, ptl);
136 lock_page(page);
137 ret = split_huge_page(page);
138 unlock_page(page);
139 put_page(page);
140 if (ret)
141 return ERR_PTR(ret);
142 goto retry;
143 }
144
145 if (flags & FOLL_GET) {
146 get_page(page);
147
148 /* drop the pgmap reference now that we hold the page */
149 if (pgmap) {
150 put_dev_pagemap(pgmap);
151 pgmap = NULL;
152 }
153 }
154 if (flags & FOLL_TOUCH) {
155 if ((flags & FOLL_WRITE) &&
156 !pte_dirty(pte) && !PageDirty(page))
157 set_page_dirty(page);
158 /*
159 * pte_mkyoung() would be more correct here, but atomic care
160 * is needed to avoid losing the dirty bit: it is easier to use
161 * mark_page_accessed().
162 */
163 mark_page_accessed(page);
164 }
165 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
166 /* Do not mlock pte-mapped THP */
167 if (PageTransCompound(page))
168 goto out;
169
170 /*
171 * The preliminary mapping check is mainly to avoid the
172 * pointless overhead of lock_page on the ZERO_PAGE
173 * which might bounce very badly if there is contention.
174 *
175 * If the page is already locked, we don't need to
176 * handle it now - vmscan will handle it later if and
177 * when it attempts to reclaim the page.
178 */
179 if (page->mapping && trylock_page(page)) {
180 lru_add_drain(); /* push cached pages to LRU */
181 /*
182 * Because we lock page here, and migration is
183 * blocked by the pte's page reference, and we
184 * know the page is still mapped, we don't even
185 * need to check for file-cache page truncation.
186 */
187 mlock_vma_page(page);
188 unlock_page(page);
189 }
190 }
191out:
192 pte_unmap_unlock(ptep, ptl);
193 return page;
194no_page:
195 pte_unmap_unlock(ptep, ptl);
196 if (!pte_none(pte))
197 return NULL;
198 return no_page_table(vma, flags);
199}
200
201/**
202 * follow_page_mask - look up a page descriptor from a user-virtual address
203 * @vma: vm_area_struct mapping @address
204 * @address: virtual address to look up
205 * @flags: flags modifying lookup behaviour
206 * @page_mask: on output, *page_mask is set according to the size of the page
207 *
208 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
209 *
210 * Returns the mapped (struct page *), %NULL if no mapping exists, or
211 * an error pointer if there is a mapping to something not represented
212 * by a page descriptor (see also vm_normal_page()).
213 */
214struct page *follow_page_mask(struct vm_area_struct *vma,
215 unsigned long address, unsigned int flags,
216 unsigned int *page_mask)
217{
218 pgd_t *pgd;
219 pud_t *pud;
220 pmd_t *pmd;
221 spinlock_t *ptl;
222 struct page *page;
223 struct mm_struct *mm = vma->vm_mm;
224
225 *page_mask = 0;
226
227 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
228 if (!IS_ERR(page)) {
229 BUG_ON(flags & FOLL_GET);
230 return page;
231 }
232
233 pgd = pgd_offset(mm, address);
234 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
235 return no_page_table(vma, flags);
236
237 pud = pud_offset(pgd, address);
238 if (pud_none(*pud))
239 return no_page_table(vma, flags);
240 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
241 page = follow_huge_pud(mm, address, pud, flags);
242 if (page)
243 return page;
244 return no_page_table(vma, flags);
245 }
246 if (unlikely(pud_bad(*pud)))
247 return no_page_table(vma, flags);
248
249 pmd = pmd_offset(pud, address);
250 if (pmd_none(*pmd))
251 return no_page_table(vma, flags);
252 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
253 page = follow_huge_pmd(mm, address, pmd, flags);
254 if (page)
255 return page;
256 return no_page_table(vma, flags);
257 }
258 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
259 return no_page_table(vma, flags);
260 if (pmd_devmap(*pmd)) {
261 ptl = pmd_lock(mm, pmd);
262 page = follow_devmap_pmd(vma, address, pmd, flags);
263 spin_unlock(ptl);
264 if (page)
265 return page;
266 }
267 if (likely(!pmd_trans_huge(*pmd)))
268 return follow_page_pte(vma, address, pmd, flags);
269
270 ptl = pmd_lock(mm, pmd);
271 if (unlikely(!pmd_trans_huge(*pmd))) {
272 spin_unlock(ptl);
273 return follow_page_pte(vma, address, pmd, flags);
274 }
275 if (flags & FOLL_SPLIT) {
276 int ret;
277 page = pmd_page(*pmd);
278 if (is_huge_zero_page(page)) {
279 spin_unlock(ptl);
280 ret = 0;
281 split_huge_pmd(vma, pmd, address);
282 } else {
283 get_page(page);
284 spin_unlock(ptl);
285 lock_page(page);
286 ret = split_huge_page(page);
287 unlock_page(page);
288 put_page(page);
289 }
290
291 return ret ? ERR_PTR(ret) :
292 follow_page_pte(vma, address, pmd, flags);
293 }
294
295 page = follow_trans_huge_pmd(vma, address, pmd, flags);
296 spin_unlock(ptl);
297 *page_mask = HPAGE_PMD_NR - 1;
298 return page;
299}
300
301static int get_gate_page(struct mm_struct *mm, unsigned long address,
302 unsigned int gup_flags, struct vm_area_struct **vma,
303 struct page **page)
304{
305 pgd_t *pgd;
306 pud_t *pud;
307 pmd_t *pmd;
308 pte_t *pte;
309 int ret = -EFAULT;
310
311 /* user gate pages are read-only */
312 if (gup_flags & FOLL_WRITE)
313 return -EFAULT;
314 if (address > TASK_SIZE)
315 pgd = pgd_offset_k(address);
316 else
317 pgd = pgd_offset_gate(mm, address);
318 BUG_ON(pgd_none(*pgd));
319 pud = pud_offset(pgd, address);
320 BUG_ON(pud_none(*pud));
321 pmd = pmd_offset(pud, address);
322 if (pmd_none(*pmd))
323 return -EFAULT;
324 VM_BUG_ON(pmd_trans_huge(*pmd));
325 pte = pte_offset_map(pmd, address);
326 if (pte_none(*pte))
327 goto unmap;
328 *vma = get_gate_vma(mm);
329 if (!page)
330 goto out;
331 *page = vm_normal_page(*vma, address, *pte);
332 if (!*page) {
333 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
334 goto unmap;
335 *page = pte_page(*pte);
336 }
337 get_page(*page);
338out:
339 ret = 0;
340unmap:
341 pte_unmap(pte);
342 return ret;
343}
344
345/*
346 * mmap_sem must be held on entry. If @nonblocking != NULL and
347 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
348 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
349 */
350static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
351 unsigned long address, unsigned int *flags, int *nonblocking)
352{
353 struct mm_struct *mm = vma->vm_mm;
354 unsigned int fault_flags = 0;
355 int ret;
356
357 /* mlock all present pages, but do not fault in new pages */
358 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
359 return -ENOENT;
360 /* For mm_populate(), just skip the stack guard page. */
361 if ((*flags & FOLL_POPULATE) &&
362 (stack_guard_page_start(vma, address) ||
363 stack_guard_page_end(vma, address + PAGE_SIZE)))
364 return -ENOENT;
365 if (*flags & FOLL_WRITE)
366 fault_flags |= FAULT_FLAG_WRITE;
367 if (*flags & FOLL_REMOTE)
368 fault_flags |= FAULT_FLAG_REMOTE;
369 if (nonblocking)
370 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
371 if (*flags & FOLL_NOWAIT)
372 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
373 if (*flags & FOLL_TRIED) {
374 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
375 fault_flags |= FAULT_FLAG_TRIED;
376 }
377
378 ret = handle_mm_fault(mm, vma, address, fault_flags);
379 if (ret & VM_FAULT_ERROR) {
380 if (ret & VM_FAULT_OOM)
381 return -ENOMEM;
382 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
383 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
384 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
385 return -EFAULT;
386 BUG();
387 }
388
389 if (tsk) {
390 if (ret & VM_FAULT_MAJOR)
391 tsk->maj_flt++;
392 else
393 tsk->min_flt++;
394 }
395
396 if (ret & VM_FAULT_RETRY) {
397 if (nonblocking)
398 *nonblocking = 0;
399 return -EBUSY;
400 }
401
402 /*
403 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
404 * necessary, even if maybe_mkwrite decided not to set pte_write. We
405 * can thus safely do subsequent page lookups as if they were reads.
406 * But only do so when looping for pte_write is futile: in some cases
407 * userspace may also be wanting to write to the gotten user page,
408 * which a read fault here might prevent (a readonly page might get
409 * reCOWed by userspace write).
410 */
411 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
412 *flags &= ~FOLL_WRITE;
413 return 0;
414}
415
416static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
417{
418 vm_flags_t vm_flags = vma->vm_flags;
419 int write = (gup_flags & FOLL_WRITE);
420 int foreign = (gup_flags & FOLL_REMOTE);
421
422 if (vm_flags & (VM_IO | VM_PFNMAP))
423 return -EFAULT;
424
425 if (write) {
426 if (!(vm_flags & VM_WRITE)) {
427 if (!(gup_flags & FOLL_FORCE))
428 return -EFAULT;
429 /*
430 * We used to let the write,force case do COW in a
431 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
432 * set a breakpoint in a read-only mapping of an
433 * executable, without corrupting the file (yet only
434 * when that file had been opened for writing!).
435 * Anon pages in shared mappings are surprising: now
436 * just reject it.
437 */
438 if (!is_cow_mapping(vm_flags))
439 return -EFAULT;
440 }
441 } else if (!(vm_flags & VM_READ)) {
442 if (!(gup_flags & FOLL_FORCE))
443 return -EFAULT;
444 /*
445 * Is there actually any vma we can reach here which does not
446 * have VM_MAYREAD set?
447 */
448 if (!(vm_flags & VM_MAYREAD))
449 return -EFAULT;
450 }
451 /*
452 * gups are always data accesses, not instruction
453 * fetches, so execute=false here
454 */
455 if (!arch_vma_access_permitted(vma, write, false, foreign))
456 return -EFAULT;
457 return 0;
458}
459
460/**
461 * __get_user_pages() - pin user pages in memory
462 * @tsk: task_struct of target task
463 * @mm: mm_struct of target mm
464 * @start: starting user address
465 * @nr_pages: number of pages from start to pin
466 * @gup_flags: flags modifying pin behaviour
467 * @pages: array that receives pointers to the pages pinned.
468 * Should be at least nr_pages long. Or NULL, if caller
469 * only intends to ensure the pages are faulted in.
470 * @vmas: array of pointers to vmas corresponding to each page.
471 * Or NULL if the caller does not require them.
472 * @nonblocking: whether waiting for disk IO or mmap_sem contention
473 *
474 * Returns number of pages pinned. This may be fewer than the number
475 * requested. If nr_pages is 0 or negative, returns 0. If no pages
476 * were pinned, returns -errno. Each page returned must be released
477 * with a put_page() call when it is finished with. vmas will only
478 * remain valid while mmap_sem is held.
479 *
480 * Must be called with mmap_sem held. It may be released. See below.
481 *
482 * __get_user_pages walks a process's page tables and takes a reference to
483 * each struct page that each user address corresponds to at a given
484 * instant. That is, it takes the page that would be accessed if a user
485 * thread accesses the given user virtual address at that instant.
486 *
487 * This does not guarantee that the page exists in the user mappings when
488 * __get_user_pages returns, and there may even be a completely different
489 * page there in some cases (eg. if mmapped pagecache has been invalidated
490 * and subsequently re faulted). However it does guarantee that the page
491 * won't be freed completely. And mostly callers simply care that the page
492 * contains data that was valid *at some point in time*. Typically, an IO
493 * or similar operation cannot guarantee anything stronger anyway because
494 * locks can't be held over the syscall boundary.
495 *
496 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
497 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
498 * appropriate) must be called after the page is finished with, and
499 * before put_page is called.
500 *
501 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
502 * or mmap_sem contention, and if waiting is needed to pin all pages,
503 * *@nonblocking will be set to 0. Further, if @gup_flags does not
504 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
505 * this case.
506 *
507 * A caller using such a combination of @nonblocking and @gup_flags
508 * must therefore hold the mmap_sem for reading only, and recognize
509 * when it's been released. Otherwise, it must be held for either
510 * reading or writing and will not be released.
511 *
512 * In most cases, get_user_pages or get_user_pages_fast should be used
513 * instead of __get_user_pages. __get_user_pages should be used only if
514 * you need some special @gup_flags.
515 */
516long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
517 unsigned long start, unsigned long nr_pages,
518 unsigned int gup_flags, struct page **pages,
519 struct vm_area_struct **vmas, int *nonblocking)
520{
521 long i = 0;
522 unsigned int page_mask;
523 struct vm_area_struct *vma = NULL;
524
525 if (!nr_pages)
526 return 0;
527
528 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
529
530 /*
531 * If FOLL_FORCE is set then do not force a full fault as the hinting
532 * fault information is unrelated to the reference behaviour of a task
533 * using the address space
534 */
535 if (!(gup_flags & FOLL_FORCE))
536 gup_flags |= FOLL_NUMA;
537
538 do {
539 struct page *page;
540 unsigned int foll_flags = gup_flags;
541 unsigned int page_increm;
542
543 /* first iteration or cross vma bound */
544 if (!vma || start >= vma->vm_end) {
545 vma = find_extend_vma(mm, start);
546 if (!vma && in_gate_area(mm, start)) {
547 int ret;
548 ret = get_gate_page(mm, start & PAGE_MASK,
549 gup_flags, &vma,
550 pages ? &pages[i] : NULL);
551 if (ret)
552 return i ? : ret;
553 page_mask = 0;
554 goto next_page;
555 }
556
557 if (!vma || check_vma_flags(vma, gup_flags))
558 return i ? : -EFAULT;
559 if (is_vm_hugetlb_page(vma)) {
560 i = follow_hugetlb_page(mm, vma, pages, vmas,
561 &start, &nr_pages, i,
562 gup_flags);
563 continue;
564 }
565 }
566retry:
567 /*
568 * If we have a pending SIGKILL, don't keep faulting pages and
569 * potentially allocating memory.
570 */
571 if (unlikely(fatal_signal_pending(current)))
572 return i ? i : -ERESTARTSYS;
573 cond_resched();
574 page = follow_page_mask(vma, start, foll_flags, &page_mask);
575 if (!page) {
576 int ret;
577 ret = faultin_page(tsk, vma, start, &foll_flags,
578 nonblocking);
579 switch (ret) {
580 case 0:
581 goto retry;
582 case -EFAULT:
583 case -ENOMEM:
584 case -EHWPOISON:
585 return i ? i : ret;
586 case -EBUSY:
587 return i;
588 case -ENOENT:
589 goto next_page;
590 }
591 BUG();
592 } else if (PTR_ERR(page) == -EEXIST) {
593 /*
594 * Proper page table entry exists, but no corresponding
595 * struct page.
596 */
597 goto next_page;
598 } else if (IS_ERR(page)) {
599 return i ? i : PTR_ERR(page);
600 }
601 if (pages) {
602 pages[i] = page;
603 flush_anon_page(vma, page, start);
604 flush_dcache_page(page);
605 page_mask = 0;
606 }
607next_page:
608 if (vmas) {
609 vmas[i] = vma;
610 page_mask = 0;
611 }
612 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
613 if (page_increm > nr_pages)
614 page_increm = nr_pages;
615 i += page_increm;
616 start += page_increm * PAGE_SIZE;
617 nr_pages -= page_increm;
618 } while (nr_pages);
619 return i;
620}
621EXPORT_SYMBOL(__get_user_pages);
622
623bool vma_permits_fault(struct vm_area_struct *vma, unsigned int fault_flags)
624{
625 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
626 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
627 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
628
629 if (!(vm_flags & vma->vm_flags))
630 return false;
631
632 /*
633 * The architecture might have a hardware protection
634 * mechanism other than read/write that can deny access.
635 *
636 * gup always represents data access, not instruction
637 * fetches, so execute=false here:
638 */
639 if (!arch_vma_access_permitted(vma, write, false, foreign))
640 return false;
641
642 return true;
643}
644
645/*
646 * fixup_user_fault() - manually resolve a user page fault
647 * @tsk: the task_struct to use for page fault accounting, or
648 * NULL if faults are not to be recorded.
649 * @mm: mm_struct of target mm
650 * @address: user address
651 * @fault_flags:flags to pass down to handle_mm_fault()
652 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
653 * does not allow retry
654 *
655 * This is meant to be called in the specific scenario where for locking reasons
656 * we try to access user memory in atomic context (within a pagefault_disable()
657 * section), this returns -EFAULT, and we want to resolve the user fault before
658 * trying again.
659 *
660 * Typically this is meant to be used by the futex code.
661 *
662 * The main difference with get_user_pages() is that this function will
663 * unconditionally call handle_mm_fault() which will in turn perform all the
664 * necessary SW fixup of the dirty and young bits in the PTE, while
665 * get_user_pages() only guarantees to update these in the struct page.
666 *
667 * This is important for some architectures where those bits also gate the
668 * access permission to the page because they are maintained in software. On
669 * such architectures, gup() will not be enough to make a subsequent access
670 * succeed.
671 *
672 * This function will not return with an unlocked mmap_sem. So it has not the
673 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
674 */
675int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
676 unsigned long address, unsigned int fault_flags,
677 bool *unlocked)
678{
679 struct vm_area_struct *vma;
680 int ret, major = 0;
681
682 if (unlocked)
683 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
684
685retry:
686 vma = find_extend_vma(mm, address);
687 if (!vma || address < vma->vm_start)
688 return -EFAULT;
689
690 if (!vma_permits_fault(vma, fault_flags))
691 return -EFAULT;
692
693 ret = handle_mm_fault(mm, vma, address, fault_flags);
694 major |= ret & VM_FAULT_MAJOR;
695 if (ret & VM_FAULT_ERROR) {
696 if (ret & VM_FAULT_OOM)
697 return -ENOMEM;
698 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
699 return -EHWPOISON;
700 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
701 return -EFAULT;
702 BUG();
703 }
704
705 if (ret & VM_FAULT_RETRY) {
706 down_read(&mm->mmap_sem);
707 if (!(fault_flags & FAULT_FLAG_TRIED)) {
708 *unlocked = true;
709 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
710 fault_flags |= FAULT_FLAG_TRIED;
711 goto retry;
712 }
713 }
714
715 if (tsk) {
716 if (major)
717 tsk->maj_flt++;
718 else
719 tsk->min_flt++;
720 }
721 return 0;
722}
723
724static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
725 struct mm_struct *mm,
726 unsigned long start,
727 unsigned long nr_pages,
728 int write, int force,
729 struct page **pages,
730 struct vm_area_struct **vmas,
731 int *locked, bool notify_drop,
732 unsigned int flags)
733{
734 long ret, pages_done;
735 bool lock_dropped;
736
737 if (locked) {
738 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
739 BUG_ON(vmas);
740 /* check caller initialized locked */
741 BUG_ON(*locked != 1);
742 }
743
744 if (pages)
745 flags |= FOLL_GET;
746 if (write)
747 flags |= FOLL_WRITE;
748 if (force)
749 flags |= FOLL_FORCE;
750
751 pages_done = 0;
752 lock_dropped = false;
753 for (;;) {
754 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
755 vmas, locked);
756 if (!locked)
757 /* VM_FAULT_RETRY couldn't trigger, bypass */
758 return ret;
759
760 /* VM_FAULT_RETRY cannot return errors */
761 if (!*locked) {
762 BUG_ON(ret < 0);
763 BUG_ON(ret >= nr_pages);
764 }
765
766 if (!pages)
767 /* If it's a prefault don't insist harder */
768 return ret;
769
770 if (ret > 0) {
771 nr_pages -= ret;
772 pages_done += ret;
773 if (!nr_pages)
774 break;
775 }
776 if (*locked) {
777 /* VM_FAULT_RETRY didn't trigger */
778 if (!pages_done)
779 pages_done = ret;
780 break;
781 }
782 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
783 pages += ret;
784 start += ret << PAGE_SHIFT;
785
786 /*
787 * Repeat on the address that fired VM_FAULT_RETRY
788 * without FAULT_FLAG_ALLOW_RETRY but with
789 * FAULT_FLAG_TRIED.
790 */
791 *locked = 1;
792 lock_dropped = true;
793 down_read(&mm->mmap_sem);
794 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
795 pages, NULL, NULL);
796 if (ret != 1) {
797 BUG_ON(ret > 1);
798 if (!pages_done)
799 pages_done = ret;
800 break;
801 }
802 nr_pages--;
803 pages_done++;
804 if (!nr_pages)
805 break;
806 pages++;
807 start += PAGE_SIZE;
808 }
809 if (notify_drop && lock_dropped && *locked) {
810 /*
811 * We must let the caller know we temporarily dropped the lock
812 * and so the critical section protected by it was lost.
813 */
814 up_read(&mm->mmap_sem);
815 *locked = 0;
816 }
817 return pages_done;
818}
819
820/*
821 * We can leverage the VM_FAULT_RETRY functionality in the page fault
822 * paths better by using either get_user_pages_locked() or
823 * get_user_pages_unlocked().
824 *
825 * get_user_pages_locked() is suitable to replace the form:
826 *
827 * down_read(&mm->mmap_sem);
828 * do_something()
829 * get_user_pages(tsk, mm, ..., pages, NULL);
830 * up_read(&mm->mmap_sem);
831 *
832 * to:
833 *
834 * int locked = 1;
835 * down_read(&mm->mmap_sem);
836 * do_something()
837 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
838 * if (locked)
839 * up_read(&mm->mmap_sem);
840 */
841long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
842 int write, int force, struct page **pages,
843 int *locked)
844{
845 return __get_user_pages_locked(current, current->mm, start, nr_pages,
846 write, force, pages, NULL, locked, true,
847 FOLL_TOUCH);
848}
849EXPORT_SYMBOL(get_user_pages_locked);
850
851/*
852 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
853 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
854 *
855 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
856 * caller if required (just like with __get_user_pages). "FOLL_GET",
857 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
858 * according to the parameters "pages", "write", "force"
859 * respectively.
860 */
861__always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
862 unsigned long start, unsigned long nr_pages,
863 int write, int force, struct page **pages,
864 unsigned int gup_flags)
865{
866 long ret;
867 int locked = 1;
868 down_read(&mm->mmap_sem);
869 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
870 pages, NULL, &locked, false, gup_flags);
871 if (locked)
872 up_read(&mm->mmap_sem);
873 return ret;
874}
875EXPORT_SYMBOL(__get_user_pages_unlocked);
876
877/*
878 * get_user_pages_unlocked() is suitable to replace the form:
879 *
880 * down_read(&mm->mmap_sem);
881 * get_user_pages(tsk, mm, ..., pages, NULL);
882 * up_read(&mm->mmap_sem);
883 *
884 * with:
885 *
886 * get_user_pages_unlocked(tsk, mm, ..., pages);
887 *
888 * It is functionally equivalent to get_user_pages_fast so
889 * get_user_pages_fast should be used instead, if the two parameters
890 * "tsk" and "mm" are respectively equal to current and current->mm,
891 * or if "force" shall be set to 1 (get_user_pages_fast misses the
892 * "force" parameter).
893 */
894long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
895 int write, int force, struct page **pages)
896{
897 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
898 write, force, pages, FOLL_TOUCH);
899}
900EXPORT_SYMBOL(get_user_pages_unlocked);
901
902/*
903 * get_user_pages_remote() - pin user pages in memory
904 * @tsk: the task_struct to use for page fault accounting, or
905 * NULL if faults are not to be recorded.
906 * @mm: mm_struct of target mm
907 * @start: starting user address
908 * @nr_pages: number of pages from start to pin
909 * @write: whether pages will be written to by the caller
910 * @force: whether to force access even when user mapping is currently
911 * protected (but never forces write access to shared mapping).
912 * @pages: array that receives pointers to the pages pinned.
913 * Should be at least nr_pages long. Or NULL, if caller
914 * only intends to ensure the pages are faulted in.
915 * @vmas: array of pointers to vmas corresponding to each page.
916 * Or NULL if the caller does not require them.
917 *
918 * Returns number of pages pinned. This may be fewer than the number
919 * requested. If nr_pages is 0 or negative, returns 0. If no pages
920 * were pinned, returns -errno. Each page returned must be released
921 * with a put_page() call when it is finished with. vmas will only
922 * remain valid while mmap_sem is held.
923 *
924 * Must be called with mmap_sem held for read or write.
925 *
926 * get_user_pages walks a process's page tables and takes a reference to
927 * each struct page that each user address corresponds to at a given
928 * instant. That is, it takes the page that would be accessed if a user
929 * thread accesses the given user virtual address at that instant.
930 *
931 * This does not guarantee that the page exists in the user mappings when
932 * get_user_pages returns, and there may even be a completely different
933 * page there in some cases (eg. if mmapped pagecache has been invalidated
934 * and subsequently re faulted). However it does guarantee that the page
935 * won't be freed completely. And mostly callers simply care that the page
936 * contains data that was valid *at some point in time*. Typically, an IO
937 * or similar operation cannot guarantee anything stronger anyway because
938 * locks can't be held over the syscall boundary.
939 *
940 * If write=0, the page must not be written to. If the page is written to,
941 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
942 * after the page is finished with, and before put_page is called.
943 *
944 * get_user_pages is typically used for fewer-copy IO operations, to get a
945 * handle on the memory by some means other than accesses via the user virtual
946 * addresses. The pages may be submitted for DMA to devices or accessed via
947 * their kernel linear mapping (via the kmap APIs). Care should be taken to
948 * use the correct cache flushing APIs.
949 *
950 * See also get_user_pages_fast, for performance critical applications.
951 *
952 * get_user_pages should be phased out in favor of
953 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
954 * should use get_user_pages because it cannot pass
955 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
956 */
957long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
958 unsigned long start, unsigned long nr_pages,
959 int write, int force, struct page **pages,
960 struct vm_area_struct **vmas)
961{
962 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
963 pages, vmas, NULL, false,
964 FOLL_TOUCH | FOLL_REMOTE);
965}
966EXPORT_SYMBOL(get_user_pages_remote);
967
968/*
969 * This is the same as get_user_pages_remote(), just with a
970 * less-flexible calling convention where we assume that the task
971 * and mm being operated on are the current task's. We also
972 * obviously don't pass FOLL_REMOTE in here.
973 */
974long get_user_pages(unsigned long start, unsigned long nr_pages,
975 int write, int force, struct page **pages,
976 struct vm_area_struct **vmas)
977{
978 return __get_user_pages_locked(current, current->mm, start, nr_pages,
979 write, force, pages, vmas, NULL, false,
980 FOLL_TOUCH);
981}
982EXPORT_SYMBOL(get_user_pages);
983
984/**
985 * populate_vma_page_range() - populate a range of pages in the vma.
986 * @vma: target vma
987 * @start: start address
988 * @end: end address
989 * @nonblocking:
990 *
991 * This takes care of mlocking the pages too if VM_LOCKED is set.
992 *
993 * return 0 on success, negative error code on error.
994 *
995 * vma->vm_mm->mmap_sem must be held.
996 *
997 * If @nonblocking is NULL, it may be held for read or write and will
998 * be unperturbed.
999 *
1000 * If @nonblocking is non-NULL, it must held for read only and may be
1001 * released. If it's released, *@nonblocking will be set to 0.
1002 */
1003long populate_vma_page_range(struct vm_area_struct *vma,
1004 unsigned long start, unsigned long end, int *nonblocking)
1005{
1006 struct mm_struct *mm = vma->vm_mm;
1007 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1008 int gup_flags;
1009
1010 VM_BUG_ON(start & ~PAGE_MASK);
1011 VM_BUG_ON(end & ~PAGE_MASK);
1012 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1013 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1014 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1015
1016 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1017 if (vma->vm_flags & VM_LOCKONFAULT)
1018 gup_flags &= ~FOLL_POPULATE;
1019 /*
1020 * We want to touch writable mappings with a write fault in order
1021 * to break COW, except for shared mappings because these don't COW
1022 * and we would not want to dirty them for nothing.
1023 */
1024 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1025 gup_flags |= FOLL_WRITE;
1026
1027 /*
1028 * We want mlock to succeed for regions that have any permissions
1029 * other than PROT_NONE.
1030 */
1031 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1032 gup_flags |= FOLL_FORCE;
1033
1034 /*
1035 * We made sure addr is within a VMA, so the following will
1036 * not result in a stack expansion that recurses back here.
1037 */
1038 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1039 NULL, NULL, nonblocking);
1040}
1041
1042/*
1043 * __mm_populate - populate and/or mlock pages within a range of address space.
1044 *
1045 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1046 * flags. VMAs must be already marked with the desired vm_flags, and
1047 * mmap_sem must not be held.
1048 */
1049int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1050{
1051 struct mm_struct *mm = current->mm;
1052 unsigned long end, nstart, nend;
1053 struct vm_area_struct *vma = NULL;
1054 int locked = 0;
1055 long ret = 0;
1056
1057 VM_BUG_ON(start & ~PAGE_MASK);
1058 VM_BUG_ON(len != PAGE_ALIGN(len));
1059 end = start + len;
1060
1061 for (nstart = start; nstart < end; nstart = nend) {
1062 /*
1063 * We want to fault in pages for [nstart; end) address range.
1064 * Find first corresponding VMA.
1065 */
1066 if (!locked) {
1067 locked = 1;
1068 down_read(&mm->mmap_sem);
1069 vma = find_vma(mm, nstart);
1070 } else if (nstart >= vma->vm_end)
1071 vma = vma->vm_next;
1072 if (!vma || vma->vm_start >= end)
1073 break;
1074 /*
1075 * Set [nstart; nend) to intersection of desired address
1076 * range with the first VMA. Also, skip undesirable VMA types.
1077 */
1078 nend = min(end, vma->vm_end);
1079 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1080 continue;
1081 if (nstart < vma->vm_start)
1082 nstart = vma->vm_start;
1083 /*
1084 * Now fault in a range of pages. populate_vma_page_range()
1085 * double checks the vma flags, so that it won't mlock pages
1086 * if the vma was already munlocked.
1087 */
1088 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1089 if (ret < 0) {
1090 if (ignore_errors) {
1091 ret = 0;
1092 continue; /* continue at next VMA */
1093 }
1094 break;
1095 }
1096 nend = nstart + ret * PAGE_SIZE;
1097 ret = 0;
1098 }
1099 if (locked)
1100 up_read(&mm->mmap_sem);
1101 return ret; /* 0 or negative error code */
1102}
1103
1104/**
1105 * get_dump_page() - pin user page in memory while writing it to core dump
1106 * @addr: user address
1107 *
1108 * Returns struct page pointer of user page pinned for dump,
1109 * to be freed afterwards by put_page().
1110 *
1111 * Returns NULL on any kind of failure - a hole must then be inserted into
1112 * the corefile, to preserve alignment with its headers; and also returns
1113 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1114 * allowing a hole to be left in the corefile to save diskspace.
1115 *
1116 * Called without mmap_sem, but after all other threads have been killed.
1117 */
1118#ifdef CONFIG_ELF_CORE
1119struct page *get_dump_page(unsigned long addr)
1120{
1121 struct vm_area_struct *vma;
1122 struct page *page;
1123
1124 if (__get_user_pages(current, current->mm, addr, 1,
1125 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1126 NULL) < 1)
1127 return NULL;
1128 flush_cache_page(vma, addr, page_to_pfn(page));
1129 return page;
1130}
1131#endif /* CONFIG_ELF_CORE */
1132
1133/*
1134 * Generic RCU Fast GUP
1135 *
1136 * get_user_pages_fast attempts to pin user pages by walking the page
1137 * tables directly and avoids taking locks. Thus the walker needs to be
1138 * protected from page table pages being freed from under it, and should
1139 * block any THP splits.
1140 *
1141 * One way to achieve this is to have the walker disable interrupts, and
1142 * rely on IPIs from the TLB flushing code blocking before the page table
1143 * pages are freed. This is unsuitable for architectures that do not need
1144 * to broadcast an IPI when invalidating TLBs.
1145 *
1146 * Another way to achieve this is to batch up page table containing pages
1147 * belonging to more than one mm_user, then rcu_sched a callback to free those
1148 * pages. Disabling interrupts will allow the fast_gup walker to both block
1149 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1150 * (which is a relatively rare event). The code below adopts this strategy.
1151 *
1152 * Before activating this code, please be aware that the following assumptions
1153 * are currently made:
1154 *
1155 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1156 * pages containing page tables.
1157 *
1158 * *) ptes can be read atomically by the architecture.
1159 *
1160 * *) access_ok is sufficient to validate userspace address ranges.
1161 *
1162 * The last two assumptions can be relaxed by the addition of helper functions.
1163 *
1164 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1165 */
1166#ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1167
1168#ifdef __HAVE_ARCH_PTE_SPECIAL
1169static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1170 int write, struct page **pages, int *nr)
1171{
1172 pte_t *ptep, *ptem;
1173 int ret = 0;
1174
1175 ptem = ptep = pte_offset_map(&pmd, addr);
1176 do {
1177 /*
1178 * In the line below we are assuming that the pte can be read
1179 * atomically. If this is not the case for your architecture,
1180 * please wrap this in a helper function!
1181 *
1182 * for an example see gup_get_pte in arch/x86/mm/gup.c
1183 */
1184 pte_t pte = READ_ONCE(*ptep);
1185 struct page *head, *page;
1186
1187 /*
1188 * Similar to the PMD case below, NUMA hinting must take slow
1189 * path using the pte_protnone check.
1190 */
1191 if (!pte_present(pte) || pte_special(pte) ||
1192 pte_protnone(pte) || (write && !pte_write(pte)))
1193 goto pte_unmap;
1194
1195 if (!arch_pte_access_permitted(pte, write))
1196 goto pte_unmap;
1197
1198 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1199 page = pte_page(pte);
1200 head = compound_head(page);
1201
1202 if (!page_cache_get_speculative(head))
1203 goto pte_unmap;
1204
1205 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1206 put_page(head);
1207 goto pte_unmap;
1208 }
1209
1210 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1211 pages[*nr] = page;
1212 (*nr)++;
1213
1214 } while (ptep++, addr += PAGE_SIZE, addr != end);
1215
1216 ret = 1;
1217
1218pte_unmap:
1219 pte_unmap(ptem);
1220 return ret;
1221}
1222#else
1223
1224/*
1225 * If we can't determine whether or not a pte is special, then fail immediately
1226 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1227 * to be special.
1228 *
1229 * For a futex to be placed on a THP tail page, get_futex_key requires a
1230 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1231 * useful to have gup_huge_pmd even if we can't operate on ptes.
1232 */
1233static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1234 int write, struct page **pages, int *nr)
1235{
1236 return 0;
1237}
1238#endif /* __HAVE_ARCH_PTE_SPECIAL */
1239
1240static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1241 unsigned long end, int write, struct page **pages, int *nr)
1242{
1243 struct page *head, *page;
1244 int refs;
1245
1246 if (write && !pmd_write(orig))
1247 return 0;
1248
1249 refs = 0;
1250 head = pmd_page(orig);
1251 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1252 do {
1253 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1254 pages[*nr] = page;
1255 (*nr)++;
1256 page++;
1257 refs++;
1258 } while (addr += PAGE_SIZE, addr != end);
1259
1260 if (!page_cache_add_speculative(head, refs)) {
1261 *nr -= refs;
1262 return 0;
1263 }
1264
1265 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1266 *nr -= refs;
1267 while (refs--)
1268 put_page(head);
1269 return 0;
1270 }
1271
1272 return 1;
1273}
1274
1275static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1276 unsigned long end, int write, struct page **pages, int *nr)
1277{
1278 struct page *head, *page;
1279 int refs;
1280
1281 if (write && !pud_write(orig))
1282 return 0;
1283
1284 refs = 0;
1285 head = pud_page(orig);
1286 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1287 do {
1288 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1289 pages[*nr] = page;
1290 (*nr)++;
1291 page++;
1292 refs++;
1293 } while (addr += PAGE_SIZE, addr != end);
1294
1295 if (!page_cache_add_speculative(head, refs)) {
1296 *nr -= refs;
1297 return 0;
1298 }
1299
1300 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1301 *nr -= refs;
1302 while (refs--)
1303 put_page(head);
1304 return 0;
1305 }
1306
1307 return 1;
1308}
1309
1310static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1311 unsigned long end, int write,
1312 struct page **pages, int *nr)
1313{
1314 int refs;
1315 struct page *head, *page;
1316
1317 if (write && !pgd_write(orig))
1318 return 0;
1319
1320 refs = 0;
1321 head = pgd_page(orig);
1322 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1323 do {
1324 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1325 pages[*nr] = page;
1326 (*nr)++;
1327 page++;
1328 refs++;
1329 } while (addr += PAGE_SIZE, addr != end);
1330
1331 if (!page_cache_add_speculative(head, refs)) {
1332 *nr -= refs;
1333 return 0;
1334 }
1335
1336 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1337 *nr -= refs;
1338 while (refs--)
1339 put_page(head);
1340 return 0;
1341 }
1342
1343 return 1;
1344}
1345
1346static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1347 int write, struct page **pages, int *nr)
1348{
1349 unsigned long next;
1350 pmd_t *pmdp;
1351
1352 pmdp = pmd_offset(&pud, addr);
1353 do {
1354 pmd_t pmd = READ_ONCE(*pmdp);
1355
1356 next = pmd_addr_end(addr, end);
1357 if (pmd_none(pmd))
1358 return 0;
1359
1360 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1361 /*
1362 * NUMA hinting faults need to be handled in the GUP
1363 * slowpath for accounting purposes and so that they
1364 * can be serialised against THP migration.
1365 */
1366 if (pmd_protnone(pmd))
1367 return 0;
1368
1369 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1370 pages, nr))
1371 return 0;
1372
1373 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1374 /*
1375 * architecture have different format for hugetlbfs
1376 * pmd format and THP pmd format
1377 */
1378 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1379 PMD_SHIFT, next, write, pages, nr))
1380 return 0;
1381 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1382 return 0;
1383 } while (pmdp++, addr = next, addr != end);
1384
1385 return 1;
1386}
1387
1388static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1389 int write, struct page **pages, int *nr)
1390{
1391 unsigned long next;
1392 pud_t *pudp;
1393
1394 pudp = pud_offset(&pgd, addr);
1395 do {
1396 pud_t pud = READ_ONCE(*pudp);
1397
1398 next = pud_addr_end(addr, end);
1399 if (pud_none(pud))
1400 return 0;
1401 if (unlikely(pud_huge(pud))) {
1402 if (!gup_huge_pud(pud, pudp, addr, next, write,
1403 pages, nr))
1404 return 0;
1405 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1406 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1407 PUD_SHIFT, next, write, pages, nr))
1408 return 0;
1409 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1410 return 0;
1411 } while (pudp++, addr = next, addr != end);
1412
1413 return 1;
1414}
1415
1416/*
1417 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1418 * the regular GUP. It will only return non-negative values.
1419 */
1420int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1421 struct page **pages)
1422{
1423 struct mm_struct *mm = current->mm;
1424 unsigned long addr, len, end;
1425 unsigned long next, flags;
1426 pgd_t *pgdp;
1427 int nr = 0;
1428
1429 start &= PAGE_MASK;
1430 addr = start;
1431 len = (unsigned long) nr_pages << PAGE_SHIFT;
1432 end = start + len;
1433
1434 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1435 start, len)))
1436 return 0;
1437
1438 /*
1439 * Disable interrupts. We use the nested form as we can already have
1440 * interrupts disabled by get_futex_key.
1441 *
1442 * With interrupts disabled, we block page table pages from being
1443 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1444 * for more details.
1445 *
1446 * We do not adopt an rcu_read_lock(.) here as we also want to
1447 * block IPIs that come from THPs splitting.
1448 */
1449
1450 local_irq_save(flags);
1451 pgdp = pgd_offset(mm, addr);
1452 do {
1453 pgd_t pgd = READ_ONCE(*pgdp);
1454
1455 next = pgd_addr_end(addr, end);
1456 if (pgd_none(pgd))
1457 break;
1458 if (unlikely(pgd_huge(pgd))) {
1459 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1460 pages, &nr))
1461 break;
1462 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1463 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1464 PGDIR_SHIFT, next, write, pages, &nr))
1465 break;
1466 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
1467 break;
1468 } while (pgdp++, addr = next, addr != end);
1469 local_irq_restore(flags);
1470
1471 return nr;
1472}
1473
1474/**
1475 * get_user_pages_fast() - pin user pages in memory
1476 * @start: starting user address
1477 * @nr_pages: number of pages from start to pin
1478 * @write: whether pages will be written to
1479 * @pages: array that receives pointers to the pages pinned.
1480 * Should be at least nr_pages long.
1481 *
1482 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1483 * If not successful, it will fall back to taking the lock and
1484 * calling get_user_pages().
1485 *
1486 * Returns number of pages pinned. This may be fewer than the number
1487 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1488 * were pinned, returns -errno.
1489 */
1490int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1491 struct page **pages)
1492{
1493 int nr, ret;
1494
1495 start &= PAGE_MASK;
1496 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1497 ret = nr;
1498
1499 if (nr < nr_pages) {
1500 /* Try to get the remaining pages with get_user_pages */
1501 start += nr << PAGE_SHIFT;
1502 pages += nr;
1503
1504 ret = get_user_pages_unlocked(start, nr_pages - nr, write, 0, pages);
1505
1506 /* Have to be a bit careful with return values */
1507 if (nr > 0) {
1508 if (ret < 0)
1509 ret = nr;
1510 else
1511 ret += nr;
1512 }
1513 }
1514
1515 return ret;
1516}
1517
1518#endif /* CONFIG_HAVE_GENERIC_RCU_GUP */