1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
4 #include <linux/err.h>
5 #include <linux/spinlock.h>
6
7 #include <linux/mm.h>
8 #include <linux/memremap.h>
9 #include <linux/pagemap.h>
10 #include <linux/rmap.h>
11 #include <linux/swap.h>
12 #include <linux/swapops.h>
13 #include <linux/secretmem.h>
14
15 #include <linux/sched/signal.h>
16 #include <linux/rwsem.h>
17 #include <linux/hugetlb.h>
18 #include <linux/migrate.h>
19 #include <linux/mm_inline.h>
20 #include <linux/sched/mm.h>
21 #include <linux/shmem_fs.h>
22
23 #include <asm/mmu_context.h>
24 #include <asm/tlbflush.h>
25
26 #include "internal.h"
27
28 struct follow_page_context {
29 struct dev_pagemap *pgmap;
30 unsigned int page_mask;
31 };
32
sanity_check_pinned_pages(struct page ** pages,unsigned long npages)33 static inline void sanity_check_pinned_pages(struct page **pages,
34 unsigned long npages)
35 {
36 if (!IS_ENABLED(CONFIG_DEBUG_VM))
37 return;
38
39 /*
40 * We only pin anonymous pages if they are exclusive. Once pinned, we
41 * can no longer turn them possibly shared and PageAnonExclusive() will
42 * stick around until the page is freed.
43 *
44 * We'd like to verify that our pinned anonymous pages are still mapped
45 * exclusively. The issue with anon THP is that we don't know how
46 * they are/were mapped when pinning them. However, for anon
47 * THP we can assume that either the given page (PTE-mapped THP) or
48 * the head page (PMD-mapped THP) should be PageAnonExclusive(). If
49 * neither is the case, there is certainly something wrong.
50 */
51 for (; npages; npages--, pages++) {
52 struct page *page = *pages;
53 struct folio *folio = page_folio(page);
54
55 if (is_zero_page(page) ||
56 !folio_test_anon(folio))
57 continue;
58 if (!folio_test_large(folio) || folio_test_hugetlb(folio))
59 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page), page);
60 else
61 /* Either a PTE-mapped or a PMD-mapped THP. */
62 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page) &&
63 !PageAnonExclusive(page), page);
64 }
65 }
66
67 /*
68 * Return the folio with ref appropriately incremented,
69 * or NULL if that failed.
70 */
try_get_folio(struct page * page,int refs)71 static inline struct folio *try_get_folio(struct page *page, int refs)
72 {
73 struct folio *folio;
74
75 retry:
76 folio = page_folio(page);
77 if (WARN_ON_ONCE(folio_ref_count(folio) < 0))
78 return NULL;
79 if (unlikely(!folio_ref_try_add_rcu(folio, refs)))
80 return NULL;
81
82 /*
83 * At this point we have a stable reference to the folio; but it
84 * could be that between calling page_folio() and the refcount
85 * increment, the folio was split, in which case we'd end up
86 * holding a reference on a folio that has nothing to do with the page
87 * we were given anymore.
88 * So now that the folio is stable, recheck that the page still
89 * belongs to this folio.
90 */
91 if (unlikely(page_folio(page) != folio)) {
92 if (!put_devmap_managed_folio_refs(folio, refs))
93 folio_put_refs(folio, refs);
94 goto retry;
95 }
96
97 return folio;
98 }
99
100 /**
101 * try_grab_folio() - Attempt to get or pin a folio.
102 * @page: pointer to page to be grabbed
103 * @refs: the value to (effectively) add to the folio's refcount
104 * @flags: gup flags: these are the FOLL_* flag values.
105 *
106 * "grab" names in this file mean, "look at flags to decide whether to use
107 * FOLL_PIN or FOLL_GET behavior, when incrementing the folio's refcount.
108 *
109 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
110 * same time. (That's true throughout the get_user_pages*() and
111 * pin_user_pages*() APIs.) Cases:
112 *
113 * FOLL_GET: folio's refcount will be incremented by @refs.
114 *
115 * FOLL_PIN on large folios: folio's refcount will be incremented by
116 * @refs, and its pincount will be incremented by @refs.
117 *
118 * FOLL_PIN on single-page folios: folio's refcount will be incremented by
119 * @refs * GUP_PIN_COUNTING_BIAS.
120 *
121 * Return: The folio containing @page (with refcount appropriately
122 * incremented) for success, or NULL upon failure. If neither FOLL_GET
123 * nor FOLL_PIN was set, that's considered failure, and furthermore,
124 * a likely bug in the caller, so a warning is also emitted.
125 */
try_grab_folio(struct page * page,int refs,unsigned int flags)126 struct folio *try_grab_folio(struct page *page, int refs, unsigned int flags)
127 {
128 struct folio *folio;
129
130 if (WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == 0))
131 return NULL;
132
133 if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
134 return NULL;
135
136 if (flags & FOLL_GET)
137 return try_get_folio(page, refs);
138
139 /* FOLL_PIN is set */
140
141 /*
142 * Don't take a pin on the zero page - it's not going anywhere
143 * and it is used in a *lot* of places.
144 */
145 if (is_zero_page(page))
146 return page_folio(page);
147
148 folio = try_get_folio(page, refs);
149 if (!folio)
150 return NULL;
151
152 /*
153 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
154 * right zone, so fail and let the caller fall back to the slow
155 * path.
156 */
157 if (unlikely((flags & FOLL_LONGTERM) &&
158 !folio_is_longterm_pinnable(folio))) {
159 if (!put_devmap_managed_folio_refs(folio, 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
gup_put_folio(struct folio * folio,int refs,unsigned int flags)189 static void gup_put_folio(struct folio *folio, int refs, unsigned int flags)
190 {
191 if (flags & FOLL_PIN) {
192 if (is_zero_folio(folio))
193 return;
194 node_stat_mod_folio(folio, NR_FOLL_PIN_RELEASED, refs);
195 if (folio_test_large(folio))
196 atomic_sub(refs, &folio->_pincount);
197 else
198 refs *= GUP_PIN_COUNTING_BIAS;
199 }
200
201 if (!put_devmap_managed_folio_refs(folio, refs))
202 folio_put_refs(folio, refs);
203 }
204
205 /**
206 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
207 * @page: pointer to page to be grabbed
208 * @flags: gup flags: these are the FOLL_* flag values.
209 *
210 * This might not do anything at all, depending on the flags argument.
211 *
212 * "grab" names in this file mean, "look at flags to decide whether to use
213 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
214 *
215 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
216 * time. Cases: please see the try_grab_folio() documentation, with
217 * "refs=1".
218 *
219 * Return: 0 for success, or if no action was required (if neither FOLL_PIN
220 * nor FOLL_GET was set, nothing is done). A negative error code for failure:
221 *
222 * -ENOMEM FOLL_GET or FOLL_PIN was set, but the page could not
223 * be grabbed.
224 */
try_grab_page(struct page * page,unsigned int flags)225 int __must_check try_grab_page(struct page *page, unsigned int flags)
226 {
227 struct folio *folio = page_folio(page);
228
229 if (WARN_ON_ONCE(folio_ref_count(folio) <= 0))
230 return -ENOMEM;
231
232 if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
233 return -EREMOTEIO;
234
235 if (flags & FOLL_GET)
236 folio_ref_inc(folio);
237 else if (flags & FOLL_PIN) {
238 /*
239 * Don't take a pin on the zero page - it's not going anywhere
240 * and it is used in a *lot* of places.
241 */
242 if (is_zero_page(page))
243 return 0;
244
245 /*
246 * Similar to try_grab_folio(): be sure to *also*
247 * increment the normal page refcount field at least once,
248 * so that the page really is pinned.
249 */
250 if (folio_test_large(folio)) {
251 folio_ref_add(folio, 1);
252 atomic_add(1, &folio->_pincount);
253 } else {
254 folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
255 }
256
257 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, 1);
258 }
259
260 return 0;
261 }
262
263 /**
264 * unpin_user_page() - release a dma-pinned page
265 * @page: pointer to page to be released
266 *
267 * Pages that were pinned via pin_user_pages*() must be released via either
268 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
269 * that such pages can be separately tracked and uniquely handled. In
270 * particular, interactions with RDMA and filesystems need special handling.
271 */
unpin_user_page(struct page * page)272 void unpin_user_page(struct page *page)
273 {
274 sanity_check_pinned_pages(&page, 1);
275 gup_put_folio(page_folio(page), 1, FOLL_PIN);
276 }
277 EXPORT_SYMBOL(unpin_user_page);
278
279 /**
280 * folio_add_pin - Try to get an additional pin on a pinned folio
281 * @folio: The folio to be pinned
282 *
283 * Get an additional pin on a folio we already have a pin on. Makes no change
284 * if the folio is a zero_page.
285 */
folio_add_pin(struct folio * folio)286 void folio_add_pin(struct folio *folio)
287 {
288 if (is_zero_folio(folio))
289 return;
290
291 /*
292 * Similar to try_grab_folio(): be sure to *also* increment the normal
293 * page refcount field at least once, so that the page really is
294 * pinned.
295 */
296 if (folio_test_large(folio)) {
297 WARN_ON_ONCE(atomic_read(&folio->_pincount) < 1);
298 folio_ref_inc(folio);
299 atomic_inc(&folio->_pincount);
300 } else {
301 WARN_ON_ONCE(folio_ref_count(folio) < GUP_PIN_COUNTING_BIAS);
302 folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
303 }
304 }
305
gup_folio_range_next(struct page * start,unsigned long npages,unsigned long i,unsigned int * ntails)306 static inline struct folio *gup_folio_range_next(struct page *start,
307 unsigned long npages, unsigned long i, unsigned int *ntails)
308 {
309 struct page *next = nth_page(start, i);
310 struct folio *folio = page_folio(next);
311 unsigned int nr = 1;
312
313 if (folio_test_large(folio))
314 nr = min_t(unsigned int, npages - i,
315 folio_nr_pages(folio) - folio_page_idx(folio, next));
316
317 *ntails = nr;
318 return folio;
319 }
320
gup_folio_next(struct page ** list,unsigned long npages,unsigned long i,unsigned int * ntails)321 static inline struct folio *gup_folio_next(struct page **list,
322 unsigned long npages, unsigned long i, unsigned int *ntails)
323 {
324 struct folio *folio = page_folio(list[i]);
325 unsigned int nr;
326
327 for (nr = i + 1; nr < npages; nr++) {
328 if (page_folio(list[nr]) != folio)
329 break;
330 }
331
332 *ntails = nr - i;
333 return folio;
334 }
335
336 /**
337 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
338 * @pages: array of pages to be maybe marked dirty, and definitely released.
339 * @npages: number of pages in the @pages array.
340 * @make_dirty: whether to mark the pages dirty
341 *
342 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
343 * variants called on that page.
344 *
345 * For each page in the @pages array, make that page (or its head page, if a
346 * compound page) dirty, if @make_dirty is true, and if the page was previously
347 * listed as clean. In any case, releases all pages using unpin_user_page(),
348 * possibly via unpin_user_pages(), for the non-dirty case.
349 *
350 * Please see the unpin_user_page() documentation for details.
351 *
352 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
353 * required, then the caller should a) verify that this is really correct,
354 * because _lock() is usually required, and b) hand code it:
355 * set_page_dirty_lock(), unpin_user_page().
356 *
357 */
unpin_user_pages_dirty_lock(struct page ** pages,unsigned long npages,bool make_dirty)358 void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
359 bool make_dirty)
360 {
361 unsigned long i;
362 struct folio *folio;
363 unsigned int nr;
364
365 if (!make_dirty) {
366 unpin_user_pages(pages, npages);
367 return;
368 }
369
370 sanity_check_pinned_pages(pages, npages);
371 for (i = 0; i < npages; i += nr) {
372 folio = gup_folio_next(pages, npages, i, &nr);
373 /*
374 * Checking PageDirty at this point may race with
375 * clear_page_dirty_for_io(), but that's OK. Two key
376 * cases:
377 *
378 * 1) This code sees the page as already dirty, so it
379 * skips the call to set_page_dirty(). That could happen
380 * because clear_page_dirty_for_io() called
381 * page_mkclean(), followed by set_page_dirty().
382 * However, now the page is going to get written back,
383 * which meets the original intention of setting it
384 * dirty, so all is well: clear_page_dirty_for_io() goes
385 * on to call TestClearPageDirty(), and write the page
386 * back.
387 *
388 * 2) This code sees the page as clean, so it calls
389 * set_page_dirty(). The page stays dirty, despite being
390 * written back, so it gets written back again in the
391 * next writeback cycle. This is harmless.
392 */
393 if (!folio_test_dirty(folio)) {
394 folio_lock(folio);
395 folio_mark_dirty(folio);
396 folio_unlock(folio);
397 }
398 gup_put_folio(folio, nr, FOLL_PIN);
399 }
400 }
401 EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
402
403 /**
404 * unpin_user_page_range_dirty_lock() - release and optionally dirty
405 * gup-pinned page range
406 *
407 * @page: the starting page of a range maybe marked dirty, and definitely released.
408 * @npages: number of consecutive pages to release.
409 * @make_dirty: whether to mark the pages dirty
410 *
411 * "gup-pinned page range" refers to a range of pages that has had one of the
412 * pin_user_pages() variants called on that page.
413 *
414 * For the page ranges defined by [page .. page+npages], make that range (or
415 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
416 * page range was previously listed as clean.
417 *
418 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
419 * required, then the caller should a) verify that this is really correct,
420 * because _lock() is usually required, and b) hand code it:
421 * set_page_dirty_lock(), unpin_user_page().
422 *
423 */
unpin_user_page_range_dirty_lock(struct page * page,unsigned long npages,bool make_dirty)424 void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages,
425 bool make_dirty)
426 {
427 unsigned long i;
428 struct folio *folio;
429 unsigned int nr;
430
431 for (i = 0; i < npages; i += nr) {
432 folio = gup_folio_range_next(page, npages, i, &nr);
433 if (make_dirty && !folio_test_dirty(folio)) {
434 folio_lock(folio);
435 folio_mark_dirty(folio);
436 folio_unlock(folio);
437 }
438 gup_put_folio(folio, nr, FOLL_PIN);
439 }
440 }
441 EXPORT_SYMBOL(unpin_user_page_range_dirty_lock);
442
gup_fast_unpin_user_pages(struct page ** pages,unsigned long npages)443 static void gup_fast_unpin_user_pages(struct page **pages, unsigned long npages)
444 {
445 unsigned long i;
446 struct folio *folio;
447 unsigned int nr;
448
449 /*
450 * Don't perform any sanity checks because we might have raced with
451 * fork() and some anonymous pages might now actually be shared --
452 * which is why we're unpinning after all.
453 */
454 for (i = 0; i < npages; i += nr) {
455 folio = gup_folio_next(pages, npages, i, &nr);
456 gup_put_folio(folio, nr, FOLL_PIN);
457 }
458 }
459
460 /**
461 * unpin_user_pages() - release an array of gup-pinned pages.
462 * @pages: array of pages to be marked dirty and released.
463 * @npages: number of pages in the @pages array.
464 *
465 * For each page in the @pages array, release the page using unpin_user_page().
466 *
467 * Please see the unpin_user_page() documentation for details.
468 */
unpin_user_pages(struct page ** pages,unsigned long npages)469 void unpin_user_pages(struct page **pages, unsigned long npages)
470 {
471 unsigned long i;
472 struct folio *folio;
473 unsigned int nr;
474
475 /*
476 * If this WARN_ON() fires, then the system *might* be leaking pages (by
477 * leaving them pinned), but probably not. More likely, gup/pup returned
478 * a hard -ERRNO error to the caller, who erroneously passed it here.
479 */
480 if (WARN_ON(IS_ERR_VALUE(npages)))
481 return;
482
483 sanity_check_pinned_pages(pages, npages);
484 for (i = 0; i < npages; i += nr) {
485 folio = gup_folio_next(pages, npages, i, &nr);
486 gup_put_folio(folio, nr, FOLL_PIN);
487 }
488 }
489 EXPORT_SYMBOL(unpin_user_pages);
490
491 /*
492 * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's
493 * lifecycle. Avoid setting the bit unless necessary, or it might cause write
494 * cache bouncing on large SMP machines for concurrent pinned gups.
495 */
mm_set_has_pinned_flag(unsigned long * mm_flags)496 static inline void mm_set_has_pinned_flag(unsigned long *mm_flags)
497 {
498 if (!test_bit(MMF_HAS_PINNED, mm_flags))
499 set_bit(MMF_HAS_PINNED, mm_flags);
500 }
501
502 #ifdef CONFIG_MMU
503
504 #if defined(CONFIG_ARCH_HAS_HUGEPD) || defined(CONFIG_HAVE_GUP_FAST)
record_subpages(struct page * page,unsigned long sz,unsigned long addr,unsigned long end,struct page ** pages)505 static int record_subpages(struct page *page, unsigned long sz,
506 unsigned long addr, unsigned long end,
507 struct page **pages)
508 {
509 struct page *start_page;
510 int nr;
511
512 start_page = nth_page(page, (addr & (sz - 1)) >> PAGE_SHIFT);
513 for (nr = 0; addr != end; nr++, addr += PAGE_SIZE)
514 pages[nr] = nth_page(start_page, nr);
515
516 return nr;
517 }
518 #endif /* CONFIG_ARCH_HAS_HUGEPD || CONFIG_HAVE_GUP_FAST */
519
520 #ifdef CONFIG_ARCH_HAS_HUGEPD
hugepte_addr_end(unsigned long addr,unsigned long end,unsigned long sz)521 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
522 unsigned long sz)
523 {
524 unsigned long __boundary = (addr + sz) & ~(sz-1);
525 return (__boundary - 1 < end - 1) ? __boundary : end;
526 }
527
528 /*
529 * Returns 1 if succeeded, 0 if failed, -EMLINK if unshare needed.
530 *
531 * NOTE: for the same entry, gup-fast and gup-slow can return different
532 * results (0 v.s. -EMLINK) depending on whether vma is available. This is
533 * the expected behavior, where we simply want gup-fast to fallback to
534 * gup-slow to take the vma reference first.
535 */
gup_hugepte(struct vm_area_struct * vma,pte_t * ptep,unsigned long sz,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)536 static int gup_hugepte(struct vm_area_struct *vma, pte_t *ptep, unsigned long sz,
537 unsigned long addr, unsigned long end, unsigned int flags,
538 struct page **pages, int *nr)
539 {
540 unsigned long pte_end;
541 struct page *page;
542 struct folio *folio;
543 pte_t pte;
544 int refs;
545
546 pte_end = (addr + sz) & ~(sz-1);
547 if (pte_end < end)
548 end = pte_end;
549
550 pte = huge_ptep_get(ptep);
551
552 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
553 return 0;
554
555 /* hugepages are never "special" */
556 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
557
558 page = pte_page(pte);
559 refs = record_subpages(page, sz, addr, end, pages + *nr);
560
561 folio = try_grab_folio(page, refs, flags);
562 if (!folio)
563 return 0;
564
565 if (unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) {
566 gup_put_folio(folio, refs, flags);
567 return 0;
568 }
569
570 if (!pte_write(pte) && gup_must_unshare(vma, flags, &folio->page)) {
571 gup_put_folio(folio, refs, flags);
572 return -EMLINK;
573 }
574
575 *nr += refs;
576 folio_set_referenced(folio);
577 return 1;
578 }
579
580 /*
581 * NOTE: currently GUP for a hugepd is only possible on hugetlbfs file
582 * systems on Power, which does not have issue with folio writeback against
583 * GUP updates. When hugepd will be extended to support non-hugetlbfs or
584 * even anonymous memory, we need to do extra check as what we do with most
585 * of the other folios. See writable_file_mapping_allowed() and
586 * gup_fast_folio_allowed() for more information.
587 */
gup_hugepd(struct vm_area_struct * vma,hugepd_t hugepd,unsigned long addr,unsigned int pdshift,unsigned long end,unsigned int flags,struct page ** pages,int * nr)588 static int gup_hugepd(struct vm_area_struct *vma, hugepd_t hugepd,
589 unsigned long addr, unsigned int pdshift,
590 unsigned long end, unsigned int flags,
591 struct page **pages, int *nr)
592 {
593 pte_t *ptep;
594 unsigned long sz = 1UL << hugepd_shift(hugepd);
595 unsigned long next;
596 int ret;
597
598 ptep = hugepte_offset(hugepd, addr, pdshift);
599 do {
600 next = hugepte_addr_end(addr, end, sz);
601 ret = gup_hugepte(vma, ptep, sz, addr, end, flags, pages, nr);
602 if (ret != 1)
603 return ret;
604 } while (ptep++, addr = next, addr != end);
605
606 return 1;
607 }
608
follow_hugepd(struct vm_area_struct * vma,hugepd_t hugepd,unsigned long addr,unsigned int pdshift,unsigned int flags,struct follow_page_context * ctx)609 static struct page *follow_hugepd(struct vm_area_struct *vma, hugepd_t hugepd,
610 unsigned long addr, unsigned int pdshift,
611 unsigned int flags,
612 struct follow_page_context *ctx)
613 {
614 struct page *page;
615 struct hstate *h;
616 spinlock_t *ptl;
617 int nr = 0, ret;
618 pte_t *ptep;
619
620 /* Only hugetlb supports hugepd */
621 if (WARN_ON_ONCE(!is_vm_hugetlb_page(vma)))
622 return ERR_PTR(-EFAULT);
623
624 h = hstate_vma(vma);
625 ptep = hugepte_offset(hugepd, addr, pdshift);
626 ptl = huge_pte_lock(h, vma->vm_mm, ptep);
627 ret = gup_hugepd(vma, hugepd, addr, pdshift, addr + PAGE_SIZE,
628 flags, &page, &nr);
629 spin_unlock(ptl);
630
631 if (ret == 1) {
632 /* GUP succeeded */
633 WARN_ON_ONCE(nr != 1);
634 ctx->page_mask = (1U << huge_page_order(h)) - 1;
635 return page;
636 }
637
638 /* ret can be either 0 (translates to NULL) or negative */
639 return ERR_PTR(ret);
640 }
641 #else /* CONFIG_ARCH_HAS_HUGEPD */
gup_hugepd(struct vm_area_struct * vma,hugepd_t hugepd,unsigned long addr,unsigned int pdshift,unsigned long end,unsigned int flags,struct page ** pages,int * nr)642 static inline int gup_hugepd(struct vm_area_struct *vma, hugepd_t hugepd,
643 unsigned long addr, unsigned int pdshift,
644 unsigned long end, unsigned int flags,
645 struct page **pages, int *nr)
646 {
647 return 0;
648 }
649
follow_hugepd(struct vm_area_struct * vma,hugepd_t hugepd,unsigned long addr,unsigned int pdshift,unsigned int flags,struct follow_page_context * ctx)650 static struct page *follow_hugepd(struct vm_area_struct *vma, hugepd_t hugepd,
651 unsigned long addr, unsigned int pdshift,
652 unsigned int flags,
653 struct follow_page_context *ctx)
654 {
655 return NULL;
656 }
657 #endif /* CONFIG_ARCH_HAS_HUGEPD */
658
659
no_page_table(struct vm_area_struct * vma,unsigned int flags,unsigned long address)660 static struct page *no_page_table(struct vm_area_struct *vma,
661 unsigned int flags, unsigned long address)
662 {
663 if (!(flags & FOLL_DUMP))
664 return NULL;
665
666 /*
667 * When core dumping, we don't want to allocate unnecessary pages or
668 * page tables. Return error instead of NULL to skip handle_mm_fault,
669 * then get_dump_page() will return NULL to leave a hole in the dump.
670 * But we can only make this optimization where a hole would surely
671 * be zero-filled if handle_mm_fault() actually did handle it.
672 */
673 if (is_vm_hugetlb_page(vma)) {
674 struct hstate *h = hstate_vma(vma);
675
676 if (!hugetlbfs_pagecache_present(h, vma, address))
677 return ERR_PTR(-EFAULT);
678 } else if ((vma_is_anonymous(vma) || !vma->vm_ops->fault)) {
679 return ERR_PTR(-EFAULT);
680 }
681
682 return NULL;
683 }
684
685 #ifdef CONFIG_PGTABLE_HAS_HUGE_LEAVES
follow_huge_pud(struct vm_area_struct * vma,unsigned long addr,pud_t * pudp,int flags,struct follow_page_context * ctx)686 static struct page *follow_huge_pud(struct vm_area_struct *vma,
687 unsigned long addr, pud_t *pudp,
688 int flags, struct follow_page_context *ctx)
689 {
690 struct mm_struct *mm = vma->vm_mm;
691 struct page *page;
692 pud_t pud = *pudp;
693 unsigned long pfn = pud_pfn(pud);
694 int ret;
695
696 assert_spin_locked(pud_lockptr(mm, pudp));
697
698 if ((flags & FOLL_WRITE) && !pud_write(pud))
699 return NULL;
700
701 if (!pud_present(pud))
702 return NULL;
703
704 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
705
706 if (IS_ENABLED(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) &&
707 pud_devmap(pud)) {
708 /*
709 * device mapped pages can only be returned if the caller
710 * will manage the page reference count.
711 *
712 * At least one of FOLL_GET | FOLL_PIN must be set, so
713 * assert that here:
714 */
715 if (!(flags & (FOLL_GET | FOLL_PIN)))
716 return ERR_PTR(-EEXIST);
717
718 if (flags & FOLL_TOUCH)
719 touch_pud(vma, addr, pudp, flags & FOLL_WRITE);
720
721 ctx->pgmap = get_dev_pagemap(pfn, ctx->pgmap);
722 if (!ctx->pgmap)
723 return ERR_PTR(-EFAULT);
724 }
725
726 page = pfn_to_page(pfn);
727
728 if (!pud_devmap(pud) && !pud_write(pud) &&
729 gup_must_unshare(vma, flags, page))
730 return ERR_PTR(-EMLINK);
731
732 ret = try_grab_page(page, flags);
733 if (ret)
734 page = ERR_PTR(ret);
735 else
736 ctx->page_mask = HPAGE_PUD_NR - 1;
737
738 return page;
739 }
740
741 /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
can_follow_write_pmd(pmd_t pmd,struct page * page,struct vm_area_struct * vma,unsigned int flags)742 static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page,
743 struct vm_area_struct *vma,
744 unsigned int flags)
745 {
746 /* If the pmd is writable, we can write to the page. */
747 if (pmd_write(pmd))
748 return true;
749
750 /* Maybe FOLL_FORCE is set to override it? */
751 if (!(flags & FOLL_FORCE))
752 return false;
753
754 /* But FOLL_FORCE has no effect on shared mappings */
755 if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
756 return false;
757
758 /* ... or read-only private ones */
759 if (!(vma->vm_flags & VM_MAYWRITE))
760 return false;
761
762 /* ... or already writable ones that just need to take a write fault */
763 if (vma->vm_flags & VM_WRITE)
764 return false;
765
766 /*
767 * See can_change_pte_writable(): we broke COW and could map the page
768 * writable if we have an exclusive anonymous page ...
769 */
770 if (!page || !PageAnon(page) || !PageAnonExclusive(page))
771 return false;
772
773 /* ... and a write-fault isn't required for other reasons. */
774 if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
775 return false;
776 return !userfaultfd_huge_pmd_wp(vma, pmd);
777 }
778
follow_huge_pmd(struct vm_area_struct * vma,unsigned long addr,pmd_t * pmd,unsigned int flags,struct follow_page_context * ctx)779 static struct page *follow_huge_pmd(struct vm_area_struct *vma,
780 unsigned long addr, pmd_t *pmd,
781 unsigned int flags,
782 struct follow_page_context *ctx)
783 {
784 struct mm_struct *mm = vma->vm_mm;
785 pmd_t pmdval = *pmd;
786 struct page *page;
787 int ret;
788
789 assert_spin_locked(pmd_lockptr(mm, pmd));
790
791 page = pmd_page(pmdval);
792 if ((flags & FOLL_WRITE) &&
793 !can_follow_write_pmd(pmdval, page, vma, flags))
794 return NULL;
795
796 /* Avoid dumping huge zero page */
797 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(pmdval))
798 return ERR_PTR(-EFAULT);
799
800 if (pmd_protnone(*pmd) && !gup_can_follow_protnone(vma, flags))
801 return NULL;
802
803 if (!pmd_write(pmdval) && gup_must_unshare(vma, flags, page))
804 return ERR_PTR(-EMLINK);
805
806 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
807 !PageAnonExclusive(page), page);
808
809 ret = try_grab_page(page, flags);
810 if (ret)
811 return ERR_PTR(ret);
812
813 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
814 if (pmd_trans_huge(pmdval) && (flags & FOLL_TOUCH))
815 touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
816 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
817
818 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
819 ctx->page_mask = HPAGE_PMD_NR - 1;
820
821 return page;
822 }
823
824 #else /* CONFIG_PGTABLE_HAS_HUGE_LEAVES */
follow_huge_pud(struct vm_area_struct * vma,unsigned long addr,pud_t * pudp,int flags,struct follow_page_context * ctx)825 static struct page *follow_huge_pud(struct vm_area_struct *vma,
826 unsigned long addr, pud_t *pudp,
827 int flags, struct follow_page_context *ctx)
828 {
829 return NULL;
830 }
831
follow_huge_pmd(struct vm_area_struct * vma,unsigned long addr,pmd_t * pmd,unsigned int flags,struct follow_page_context * ctx)832 static struct page *follow_huge_pmd(struct vm_area_struct *vma,
833 unsigned long addr, pmd_t *pmd,
834 unsigned int flags,
835 struct follow_page_context *ctx)
836 {
837 return NULL;
838 }
839 #endif /* CONFIG_PGTABLE_HAS_HUGE_LEAVES */
840
follow_pfn_pte(struct vm_area_struct * vma,unsigned long address,pte_t * pte,unsigned int flags)841 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
842 pte_t *pte, unsigned int flags)
843 {
844 if (flags & FOLL_TOUCH) {
845 pte_t orig_entry = ptep_get(pte);
846 pte_t entry = orig_entry;
847
848 if (flags & FOLL_WRITE)
849 entry = pte_mkdirty(entry);
850 entry = pte_mkyoung(entry);
851
852 if (!pte_same(orig_entry, entry)) {
853 set_pte_at(vma->vm_mm, address, pte, entry);
854 update_mmu_cache(vma, address, pte);
855 }
856 }
857
858 /* Proper page table entry exists, but no corresponding struct page */
859 return -EEXIST;
860 }
861
862 /* FOLL_FORCE can write to even unwritable PTEs in COW mappings. */
can_follow_write_pte(pte_t pte,struct page * page,struct vm_area_struct * vma,unsigned int flags)863 static inline bool can_follow_write_pte(pte_t pte, struct page *page,
864 struct vm_area_struct *vma,
865 unsigned int flags)
866 {
867 /* If the pte is writable, we can write to the page. */
868 if (pte_write(pte))
869 return true;
870
871 /* Maybe FOLL_FORCE is set to override it? */
872 if (!(flags & FOLL_FORCE))
873 return false;
874
875 /* But FOLL_FORCE has no effect on shared mappings */
876 if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
877 return false;
878
879 /* ... or read-only private ones */
880 if (!(vma->vm_flags & VM_MAYWRITE))
881 return false;
882
883 /* ... or already writable ones that just need to take a write fault */
884 if (vma->vm_flags & VM_WRITE)
885 return false;
886
887 /*
888 * See can_change_pte_writable(): we broke COW and could map the page
889 * writable if we have an exclusive anonymous page ...
890 */
891 if (!page || !PageAnon(page) || !PageAnonExclusive(page))
892 return false;
893
894 /* ... and a write-fault isn't required for other reasons. */
895 if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte))
896 return false;
897 return !userfaultfd_pte_wp(vma, pte);
898 }
899
follow_page_pte(struct vm_area_struct * vma,unsigned long address,pmd_t * pmd,unsigned int flags,struct dev_pagemap ** pgmap)900 static struct page *follow_page_pte(struct vm_area_struct *vma,
901 unsigned long address, pmd_t *pmd, unsigned int flags,
902 struct dev_pagemap **pgmap)
903 {
904 struct mm_struct *mm = vma->vm_mm;
905 struct page *page;
906 spinlock_t *ptl;
907 pte_t *ptep, pte;
908 int ret;
909
910 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
911 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
912 (FOLL_PIN | FOLL_GET)))
913 return ERR_PTR(-EINVAL);
914
915 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
916 if (!ptep)
917 return no_page_table(vma, flags, address);
918 pte = ptep_get(ptep);
919 if (!pte_present(pte))
920 goto no_page;
921 if (pte_protnone(pte) && !gup_can_follow_protnone(vma, flags))
922 goto no_page;
923
924 page = vm_normal_page(vma, address, pte);
925
926 /*
927 * We only care about anon pages in can_follow_write_pte() and don't
928 * have to worry about pte_devmap() because they are never anon.
929 */
930 if ((flags & FOLL_WRITE) &&
931 !can_follow_write_pte(pte, page, vma, flags)) {
932 page = NULL;
933 goto out;
934 }
935
936 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
937 /*
938 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
939 * case since they are only valid while holding the pgmap
940 * reference.
941 */
942 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
943 if (*pgmap)
944 page = pte_page(pte);
945 else
946 goto no_page;
947 } else if (unlikely(!page)) {
948 if (flags & FOLL_DUMP) {
949 /* Avoid special (like zero) pages in core dumps */
950 page = ERR_PTR(-EFAULT);
951 goto out;
952 }
953
954 if (is_zero_pfn(pte_pfn(pte))) {
955 page = pte_page(pte);
956 } else {
957 ret = follow_pfn_pte(vma, address, ptep, flags);
958 page = ERR_PTR(ret);
959 goto out;
960 }
961 }
962
963 if (!pte_write(pte) && gup_must_unshare(vma, flags, page)) {
964 page = ERR_PTR(-EMLINK);
965 goto out;
966 }
967
968 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
969 !PageAnonExclusive(page), page);
970
971 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
972 ret = try_grab_page(page, flags);
973 if (unlikely(ret)) {
974 page = ERR_PTR(ret);
975 goto out;
976 }
977
978 /*
979 * We need to make the page accessible if and only if we are going
980 * to access its content (the FOLL_PIN case). Please see
981 * Documentation/core-api/pin_user_pages.rst for details.
982 */
983 if (flags & FOLL_PIN) {
984 ret = arch_make_page_accessible(page);
985 if (ret) {
986 unpin_user_page(page);
987 page = ERR_PTR(ret);
988 goto out;
989 }
990 }
991 if (flags & FOLL_TOUCH) {
992 if ((flags & FOLL_WRITE) &&
993 !pte_dirty(pte) && !PageDirty(page))
994 set_page_dirty(page);
995 /*
996 * pte_mkyoung() would be more correct here, but atomic care
997 * is needed to avoid losing the dirty bit: it is easier to use
998 * mark_page_accessed().
999 */
1000 mark_page_accessed(page);
1001 }
1002 out:
1003 pte_unmap_unlock(ptep, ptl);
1004 return page;
1005 no_page:
1006 pte_unmap_unlock(ptep, ptl);
1007 if (!pte_none(pte))
1008 return NULL;
1009 return no_page_table(vma, flags, address);
1010 }
1011
follow_pmd_mask(struct vm_area_struct * vma,unsigned long address,pud_t * pudp,unsigned int flags,struct follow_page_context * ctx)1012 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
1013 unsigned long address, pud_t *pudp,
1014 unsigned int flags,
1015 struct follow_page_context *ctx)
1016 {
1017 pmd_t *pmd, pmdval;
1018 spinlock_t *ptl;
1019 struct page *page;
1020 struct mm_struct *mm = vma->vm_mm;
1021
1022 pmd = pmd_offset(pudp, address);
1023 pmdval = pmdp_get_lockless(pmd);
1024 if (pmd_none(pmdval))
1025 return no_page_table(vma, flags, address);
1026 if (!pmd_present(pmdval))
1027 return no_page_table(vma, flags, address);
1028 if (unlikely(is_hugepd(__hugepd(pmd_val(pmdval)))))
1029 return follow_hugepd(vma, __hugepd(pmd_val(pmdval)),
1030 address, PMD_SHIFT, flags, ctx);
1031 if (pmd_devmap(pmdval)) {
1032 ptl = pmd_lock(mm, pmd);
1033 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
1034 spin_unlock(ptl);
1035 if (page)
1036 return page;
1037 return no_page_table(vma, flags, address);
1038 }
1039 if (likely(!pmd_leaf(pmdval)))
1040 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
1041
1042 if (pmd_protnone(pmdval) && !gup_can_follow_protnone(vma, flags))
1043 return no_page_table(vma, flags, address);
1044
1045 ptl = pmd_lock(mm, pmd);
1046 pmdval = *pmd;
1047 if (unlikely(!pmd_present(pmdval))) {
1048 spin_unlock(ptl);
1049 return no_page_table(vma, flags, address);
1050 }
1051 if (unlikely(!pmd_leaf(pmdval))) {
1052 spin_unlock(ptl);
1053 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
1054 }
1055 if (pmd_trans_huge(pmdval) && (flags & FOLL_SPLIT_PMD)) {
1056 spin_unlock(ptl);
1057 split_huge_pmd(vma, pmd, address);
1058 /* If pmd was left empty, stuff a page table in there quickly */
1059 return pte_alloc(mm, pmd) ? ERR_PTR(-ENOMEM) :
1060 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
1061 }
1062 page = follow_huge_pmd(vma, address, pmd, flags, ctx);
1063 spin_unlock(ptl);
1064 return page;
1065 }
1066
follow_pud_mask(struct vm_area_struct * vma,unsigned long address,p4d_t * p4dp,unsigned int flags,struct follow_page_context * ctx)1067 static struct page *follow_pud_mask(struct vm_area_struct *vma,
1068 unsigned long address, p4d_t *p4dp,
1069 unsigned int flags,
1070 struct follow_page_context *ctx)
1071 {
1072 pud_t *pudp, pud;
1073 spinlock_t *ptl;
1074 struct page *page;
1075 struct mm_struct *mm = vma->vm_mm;
1076
1077 pudp = pud_offset(p4dp, address);
1078 pud = READ_ONCE(*pudp);
1079 if (!pud_present(pud))
1080 return no_page_table(vma, flags, address);
1081 if (unlikely(is_hugepd(__hugepd(pud_val(pud)))))
1082 return follow_hugepd(vma, __hugepd(pud_val(pud)),
1083 address, PUD_SHIFT, flags, ctx);
1084 if (pud_leaf(pud)) {
1085 ptl = pud_lock(mm, pudp);
1086 page = follow_huge_pud(vma, address, pudp, flags, ctx);
1087 spin_unlock(ptl);
1088 if (page)
1089 return page;
1090 return no_page_table(vma, flags, address);
1091 }
1092 if (unlikely(pud_bad(pud)))
1093 return no_page_table(vma, flags, address);
1094
1095 return follow_pmd_mask(vma, address, pudp, flags, ctx);
1096 }
1097
follow_p4d_mask(struct vm_area_struct * vma,unsigned long address,pgd_t * pgdp,unsigned int flags,struct follow_page_context * ctx)1098 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
1099 unsigned long address, pgd_t *pgdp,
1100 unsigned int flags,
1101 struct follow_page_context *ctx)
1102 {
1103 p4d_t *p4dp, p4d;
1104
1105 p4dp = p4d_offset(pgdp, address);
1106 p4d = READ_ONCE(*p4dp);
1107 BUILD_BUG_ON(p4d_leaf(p4d));
1108
1109 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d)))))
1110 return follow_hugepd(vma, __hugepd(p4d_val(p4d)),
1111 address, P4D_SHIFT, flags, ctx);
1112
1113 if (!p4d_present(p4d) || p4d_bad(p4d))
1114 return no_page_table(vma, flags, address);
1115
1116 return follow_pud_mask(vma, address, p4dp, flags, ctx);
1117 }
1118
1119 /**
1120 * follow_page_mask - look up a page descriptor from a user-virtual address
1121 * @vma: vm_area_struct mapping @address
1122 * @address: virtual address to look up
1123 * @flags: flags modifying lookup behaviour
1124 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
1125 * pointer to output page_mask
1126 *
1127 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1128 *
1129 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
1130 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
1131 *
1132 * When getting an anonymous page and the caller has to trigger unsharing
1133 * of a shared anonymous page first, -EMLINK is returned. The caller should
1134 * trigger a fault with FAULT_FLAG_UNSHARE set. Note that unsharing is only
1135 * relevant with FOLL_PIN and !FOLL_WRITE.
1136 *
1137 * On output, the @ctx->page_mask is set according to the size of the page.
1138 *
1139 * Return: the mapped (struct page *), %NULL if no mapping exists, or
1140 * an error pointer if there is a mapping to something not represented
1141 * by a page descriptor (see also vm_normal_page()).
1142 */
follow_page_mask(struct vm_area_struct * vma,unsigned long address,unsigned int flags,struct follow_page_context * ctx)1143 static struct page *follow_page_mask(struct vm_area_struct *vma,
1144 unsigned long address, unsigned int flags,
1145 struct follow_page_context *ctx)
1146 {
1147 pgd_t *pgd;
1148 struct mm_struct *mm = vma->vm_mm;
1149 struct page *page;
1150
1151 vma_pgtable_walk_begin(vma);
1152
1153 ctx->page_mask = 0;
1154 pgd = pgd_offset(mm, address);
1155
1156 if (unlikely(is_hugepd(__hugepd(pgd_val(*pgd)))))
1157 page = follow_hugepd(vma, __hugepd(pgd_val(*pgd)),
1158 address, PGDIR_SHIFT, flags, ctx);
1159 else if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1160 page = no_page_table(vma, flags, address);
1161 else
1162 page = follow_p4d_mask(vma, address, pgd, flags, ctx);
1163
1164 vma_pgtable_walk_end(vma);
1165
1166 return page;
1167 }
1168
follow_page(struct vm_area_struct * vma,unsigned long address,unsigned int foll_flags)1169 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1170 unsigned int foll_flags)
1171 {
1172 struct follow_page_context ctx = { NULL };
1173 struct page *page;
1174
1175 if (vma_is_secretmem(vma))
1176 return NULL;
1177
1178 if (WARN_ON_ONCE(foll_flags & FOLL_PIN))
1179 return NULL;
1180
1181 /*
1182 * We never set FOLL_HONOR_NUMA_FAULT because callers don't expect
1183 * to fail on PROT_NONE-mapped pages.
1184 */
1185 page = follow_page_mask(vma, address, foll_flags, &ctx);
1186 if (ctx.pgmap)
1187 put_dev_pagemap(ctx.pgmap);
1188 return page;
1189 }
1190
get_gate_page(struct mm_struct * mm,unsigned long address,unsigned int gup_flags,struct vm_area_struct ** vma,struct page ** page)1191 static int get_gate_page(struct mm_struct *mm, unsigned long address,
1192 unsigned int gup_flags, struct vm_area_struct **vma,
1193 struct page **page)
1194 {
1195 pgd_t *pgd;
1196 p4d_t *p4d;
1197 pud_t *pud;
1198 pmd_t *pmd;
1199 pte_t *pte;
1200 pte_t entry;
1201 int ret = -EFAULT;
1202
1203 /* user gate pages are read-only */
1204 if (gup_flags & FOLL_WRITE)
1205 return -EFAULT;
1206 if (address > TASK_SIZE)
1207 pgd = pgd_offset_k(address);
1208 else
1209 pgd = pgd_offset_gate(mm, address);
1210 if (pgd_none(*pgd))
1211 return -EFAULT;
1212 p4d = p4d_offset(pgd, address);
1213 if (p4d_none(*p4d))
1214 return -EFAULT;
1215 pud = pud_offset(p4d, address);
1216 if (pud_none(*pud))
1217 return -EFAULT;
1218 pmd = pmd_offset(pud, address);
1219 if (!pmd_present(*pmd))
1220 return -EFAULT;
1221 pte = pte_offset_map(pmd, address);
1222 if (!pte)
1223 return -EFAULT;
1224 entry = ptep_get(pte);
1225 if (pte_none(entry))
1226 goto unmap;
1227 *vma = get_gate_vma(mm);
1228 if (!page)
1229 goto out;
1230 *page = vm_normal_page(*vma, address, entry);
1231 if (!*page) {
1232 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(entry)))
1233 goto unmap;
1234 *page = pte_page(entry);
1235 }
1236 ret = try_grab_page(*page, gup_flags);
1237 if (unlikely(ret))
1238 goto unmap;
1239 out:
1240 ret = 0;
1241 unmap:
1242 pte_unmap(pte);
1243 return ret;
1244 }
1245
1246 /*
1247 * mmap_lock must be held on entry. If @flags has FOLL_UNLOCKABLE but not
1248 * FOLL_NOWAIT, the mmap_lock may be released. If it is, *@locked will be set
1249 * to 0 and -EBUSY returned.
1250 */
faultin_page(struct vm_area_struct * vma,unsigned long address,unsigned int * flags,bool unshare,int * locked)1251 static int faultin_page(struct vm_area_struct *vma,
1252 unsigned long address, unsigned int *flags, bool unshare,
1253 int *locked)
1254 {
1255 unsigned int fault_flags = 0;
1256 vm_fault_t ret;
1257
1258 if (*flags & FOLL_NOFAULT)
1259 return -EFAULT;
1260 if (*flags & FOLL_WRITE)
1261 fault_flags |= FAULT_FLAG_WRITE;
1262 if (*flags & FOLL_REMOTE)
1263 fault_flags |= FAULT_FLAG_REMOTE;
1264 if (*flags & FOLL_UNLOCKABLE) {
1265 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1266 /*
1267 * FAULT_FLAG_INTERRUPTIBLE is opt-in. GUP callers must set
1268 * FOLL_INTERRUPTIBLE to enable FAULT_FLAG_INTERRUPTIBLE.
1269 * That's because some callers may not be prepared to
1270 * handle early exits caused by non-fatal signals.
1271 */
1272 if (*flags & FOLL_INTERRUPTIBLE)
1273 fault_flags |= FAULT_FLAG_INTERRUPTIBLE;
1274 }
1275 if (*flags & FOLL_NOWAIT)
1276 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
1277 if (*flags & FOLL_TRIED) {
1278 /*
1279 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
1280 * can co-exist
1281 */
1282 fault_flags |= FAULT_FLAG_TRIED;
1283 }
1284 if (unshare) {
1285 fault_flags |= FAULT_FLAG_UNSHARE;
1286 /* FAULT_FLAG_WRITE and FAULT_FLAG_UNSHARE are incompatible */
1287 VM_BUG_ON(fault_flags & FAULT_FLAG_WRITE);
1288 }
1289
1290 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1291
1292 if (ret & VM_FAULT_COMPLETED) {
1293 /*
1294 * With FAULT_FLAG_RETRY_NOWAIT we'll never release the
1295 * mmap lock in the page fault handler. Sanity check this.
1296 */
1297 WARN_ON_ONCE(fault_flags & FAULT_FLAG_RETRY_NOWAIT);
1298 *locked = 0;
1299
1300 /*
1301 * We should do the same as VM_FAULT_RETRY, but let's not
1302 * return -EBUSY since that's not reflecting the reality of
1303 * what has happened - we've just fully completed a page
1304 * fault, with the mmap lock released. Use -EAGAIN to show
1305 * that we want to take the mmap lock _again_.
1306 */
1307 return -EAGAIN;
1308 }
1309
1310 if (ret & VM_FAULT_ERROR) {
1311 int err = vm_fault_to_errno(ret, *flags);
1312
1313 if (err)
1314 return err;
1315 BUG();
1316 }
1317
1318 if (ret & VM_FAULT_RETRY) {
1319 if (!(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
1320 *locked = 0;
1321 return -EBUSY;
1322 }
1323
1324 return 0;
1325 }
1326
1327 /*
1328 * Writing to file-backed mappings which require folio dirty tracking using GUP
1329 * is a fundamentally broken operation, as kernel write access to GUP mappings
1330 * do not adhere to the semantics expected by a file system.
1331 *
1332 * Consider the following scenario:-
1333 *
1334 * 1. A folio is written to via GUP which write-faults the memory, notifying
1335 * the file system and dirtying the folio.
1336 * 2. Later, writeback is triggered, resulting in the folio being cleaned and
1337 * the PTE being marked read-only.
1338 * 3. The GUP caller writes to the folio, as it is mapped read/write via the
1339 * direct mapping.
1340 * 4. The GUP caller, now done with the page, unpins it and sets it dirty
1341 * (though it does not have to).
1342 *
1343 * This results in both data being written to a folio without writenotify, and
1344 * the folio being dirtied unexpectedly (if the caller decides to do so).
1345 */
writable_file_mapping_allowed(struct vm_area_struct * vma,unsigned long gup_flags)1346 static bool writable_file_mapping_allowed(struct vm_area_struct *vma,
1347 unsigned long gup_flags)
1348 {
1349 /*
1350 * If we aren't pinning then no problematic write can occur. A long term
1351 * pin is the most egregious case so this is the case we disallow.
1352 */
1353 if ((gup_flags & (FOLL_PIN | FOLL_LONGTERM)) !=
1354 (FOLL_PIN | FOLL_LONGTERM))
1355 return true;
1356
1357 /*
1358 * If the VMA does not require dirty tracking then no problematic write
1359 * can occur either.
1360 */
1361 return !vma_needs_dirty_tracking(vma);
1362 }
1363
check_vma_flags(struct vm_area_struct * vma,unsigned long gup_flags)1364 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
1365 {
1366 vm_flags_t vm_flags = vma->vm_flags;
1367 int write = (gup_flags & FOLL_WRITE);
1368 int foreign = (gup_flags & FOLL_REMOTE);
1369 bool vma_anon = vma_is_anonymous(vma);
1370
1371 if (vm_flags & (VM_IO | VM_PFNMAP))
1372 return -EFAULT;
1373
1374 if ((gup_flags & FOLL_ANON) && !vma_anon)
1375 return -EFAULT;
1376
1377 if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma))
1378 return -EOPNOTSUPP;
1379
1380 if (vma_is_secretmem(vma))
1381 return -EFAULT;
1382
1383 if (write) {
1384 if (!vma_anon &&
1385 !writable_file_mapping_allowed(vma, gup_flags))
1386 return -EFAULT;
1387
1388 if (!(vm_flags & VM_WRITE) || (vm_flags & VM_SHADOW_STACK)) {
1389 if (!(gup_flags & FOLL_FORCE))
1390 return -EFAULT;
1391 /* hugetlb does not support FOLL_FORCE|FOLL_WRITE. */
1392 if (is_vm_hugetlb_page(vma))
1393 return -EFAULT;
1394 /*
1395 * We used to let the write,force case do COW in a
1396 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
1397 * set a breakpoint in a read-only mapping of an
1398 * executable, without corrupting the file (yet only
1399 * when that file had been opened for writing!).
1400 * Anon pages in shared mappings are surprising: now
1401 * just reject it.
1402 */
1403 if (!is_cow_mapping(vm_flags))
1404 return -EFAULT;
1405 }
1406 } else if (!(vm_flags & VM_READ)) {
1407 if (!(gup_flags & FOLL_FORCE))
1408 return -EFAULT;
1409 /*
1410 * Is there actually any vma we can reach here which does not
1411 * have VM_MAYREAD set?
1412 */
1413 if (!(vm_flags & VM_MAYREAD))
1414 return -EFAULT;
1415 }
1416 /*
1417 * gups are always data accesses, not instruction
1418 * fetches, so execute=false here
1419 */
1420 if (!arch_vma_access_permitted(vma, write, false, foreign))
1421 return -EFAULT;
1422 return 0;
1423 }
1424
1425 /*
1426 * This is "vma_lookup()", but with a warning if we would have
1427 * historically expanded the stack in the GUP code.
1428 */
gup_vma_lookup(struct mm_struct * mm,unsigned long addr)1429 static struct vm_area_struct *gup_vma_lookup(struct mm_struct *mm,
1430 unsigned long addr)
1431 {
1432 #ifdef CONFIG_STACK_GROWSUP
1433 return vma_lookup(mm, addr);
1434 #else
1435 static volatile unsigned long next_warn;
1436 struct vm_area_struct *vma;
1437 unsigned long now, next;
1438
1439 vma = find_vma(mm, addr);
1440 if (!vma || (addr >= vma->vm_start))
1441 return vma;
1442
1443 /* Only warn for half-way relevant accesses */
1444 if (!(vma->vm_flags & VM_GROWSDOWN))
1445 return NULL;
1446 if (vma->vm_start - addr > 65536)
1447 return NULL;
1448
1449 /* Let's not warn more than once an hour.. */
1450 now = jiffies; next = next_warn;
1451 if (next && time_before(now, next))
1452 return NULL;
1453 next_warn = now + 60*60*HZ;
1454
1455 /* Let people know things may have changed. */
1456 pr_warn("GUP no longer grows the stack in %s (%d): %lx-%lx (%lx)\n",
1457 current->comm, task_pid_nr(current),
1458 vma->vm_start, vma->vm_end, addr);
1459 dump_stack();
1460 return NULL;
1461 #endif
1462 }
1463
1464 /**
1465 * __get_user_pages() - pin user pages in memory
1466 * @mm: mm_struct of target mm
1467 * @start: starting user address
1468 * @nr_pages: number of pages from start to pin
1469 * @gup_flags: flags modifying pin behaviour
1470 * @pages: array that receives pointers to the pages pinned.
1471 * Should be at least nr_pages long. Or NULL, if caller
1472 * only intends to ensure the pages are faulted in.
1473 * @locked: whether we're still with the mmap_lock held
1474 *
1475 * Returns either number of pages pinned (which may be less than the
1476 * number requested), or an error. Details about the return value:
1477 *
1478 * -- If nr_pages is 0, returns 0.
1479 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1480 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1481 * pages pinned. Again, this may be less than nr_pages.
1482 * -- 0 return value is possible when the fault would need to be retried.
1483 *
1484 * The caller is responsible for releasing returned @pages, via put_page().
1485 *
1486 * Must be called with mmap_lock held. It may be released. See below.
1487 *
1488 * __get_user_pages walks a process's page tables and takes a reference to
1489 * each struct page that each user address corresponds to at a given
1490 * instant. That is, it takes the page that would be accessed if a user
1491 * thread accesses the given user virtual address at that instant.
1492 *
1493 * This does not guarantee that the page exists in the user mappings when
1494 * __get_user_pages returns, and there may even be a completely different
1495 * page there in some cases (eg. if mmapped pagecache has been invalidated
1496 * and subsequently re-faulted). However it does guarantee that the page
1497 * won't be freed completely. And mostly callers simply care that the page
1498 * contains data that was valid *at some point in time*. Typically, an IO
1499 * or similar operation cannot guarantee anything stronger anyway because
1500 * locks can't be held over the syscall boundary.
1501 *
1502 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1503 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1504 * appropriate) must be called after the page is finished with, and
1505 * before put_page is called.
1506 *
1507 * If FOLL_UNLOCKABLE is set without FOLL_NOWAIT then the mmap_lock may
1508 * be released. If this happens *@locked will be set to 0 on return.
1509 *
1510 * A caller using such a combination of @gup_flags must therefore hold the
1511 * mmap_lock for reading only, and recognize when it's been released. Otherwise,
1512 * it must be held for either reading or writing and will not be released.
1513 *
1514 * In most cases, get_user_pages or get_user_pages_fast should be used
1515 * instead of __get_user_pages. __get_user_pages should be used only if
1516 * you need some special @gup_flags.
1517 */
__get_user_pages(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)1518 static long __get_user_pages(struct mm_struct *mm,
1519 unsigned long start, unsigned long nr_pages,
1520 unsigned int gup_flags, struct page **pages,
1521 int *locked)
1522 {
1523 long ret = 0, i = 0;
1524 struct vm_area_struct *vma = NULL;
1525 struct follow_page_context ctx = { NULL };
1526
1527 if (!nr_pages)
1528 return 0;
1529
1530 start = untagged_addr_remote(mm, start);
1531
1532 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1533
1534 do {
1535 struct page *page;
1536 unsigned int foll_flags = gup_flags;
1537 unsigned int page_increm;
1538
1539 /* first iteration or cross vma bound */
1540 if (!vma || start >= vma->vm_end) {
1541 /*
1542 * MADV_POPULATE_(READ|WRITE) wants to handle VMA
1543 * lookups+error reporting differently.
1544 */
1545 if (gup_flags & FOLL_MADV_POPULATE) {
1546 vma = vma_lookup(mm, start);
1547 if (!vma) {
1548 ret = -ENOMEM;
1549 goto out;
1550 }
1551 if (check_vma_flags(vma, gup_flags)) {
1552 ret = -EINVAL;
1553 goto out;
1554 }
1555 goto retry;
1556 }
1557 vma = gup_vma_lookup(mm, start);
1558 if (!vma && in_gate_area(mm, start)) {
1559 ret = get_gate_page(mm, start & PAGE_MASK,
1560 gup_flags, &vma,
1561 pages ? &page : NULL);
1562 if (ret)
1563 goto out;
1564 ctx.page_mask = 0;
1565 goto next_page;
1566 }
1567
1568 if (!vma) {
1569 ret = -EFAULT;
1570 goto out;
1571 }
1572 ret = check_vma_flags(vma, gup_flags);
1573 if (ret)
1574 goto out;
1575 }
1576 retry:
1577 /*
1578 * If we have a pending SIGKILL, don't keep faulting pages and
1579 * potentially allocating memory.
1580 */
1581 if (fatal_signal_pending(current)) {
1582 ret = -EINTR;
1583 goto out;
1584 }
1585 cond_resched();
1586
1587 page = follow_page_mask(vma, start, foll_flags, &ctx);
1588 if (!page || PTR_ERR(page) == -EMLINK) {
1589 ret = faultin_page(vma, start, &foll_flags,
1590 PTR_ERR(page) == -EMLINK, locked);
1591 switch (ret) {
1592 case 0:
1593 goto retry;
1594 case -EBUSY:
1595 case -EAGAIN:
1596 ret = 0;
1597 fallthrough;
1598 case -EFAULT:
1599 case -ENOMEM:
1600 case -EHWPOISON:
1601 goto out;
1602 }
1603 BUG();
1604 } else if (PTR_ERR(page) == -EEXIST) {
1605 /*
1606 * Proper page table entry exists, but no corresponding
1607 * struct page. If the caller expects **pages to be
1608 * filled in, bail out now, because that can't be done
1609 * for this page.
1610 */
1611 if (pages) {
1612 ret = PTR_ERR(page);
1613 goto out;
1614 }
1615 } else if (IS_ERR(page)) {
1616 ret = PTR_ERR(page);
1617 goto out;
1618 }
1619 next_page:
1620 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1621 if (page_increm > nr_pages)
1622 page_increm = nr_pages;
1623
1624 if (pages) {
1625 struct page *subpage;
1626 unsigned int j;
1627
1628 /*
1629 * This must be a large folio (and doesn't need to
1630 * be the whole folio; it can be part of it), do
1631 * the refcount work for all the subpages too.
1632 *
1633 * NOTE: here the page may not be the head page
1634 * e.g. when start addr is not thp-size aligned.
1635 * try_grab_folio() should have taken care of tail
1636 * pages.
1637 */
1638 if (page_increm > 1) {
1639 struct folio *folio;
1640
1641 /*
1642 * Since we already hold refcount on the
1643 * large folio, this should never fail.
1644 */
1645 folio = try_grab_folio(page, page_increm - 1,
1646 foll_flags);
1647 if (WARN_ON_ONCE(!folio)) {
1648 /*
1649 * Release the 1st page ref if the
1650 * folio is problematic, fail hard.
1651 */
1652 gup_put_folio(page_folio(page), 1,
1653 foll_flags);
1654 ret = -EFAULT;
1655 goto out;
1656 }
1657 }
1658
1659 for (j = 0; j < page_increm; j++) {
1660 subpage = nth_page(page, j);
1661 pages[i + j] = subpage;
1662 flush_anon_page(vma, subpage, start + j * PAGE_SIZE);
1663 flush_dcache_page(subpage);
1664 }
1665 }
1666
1667 i += page_increm;
1668 start += page_increm * PAGE_SIZE;
1669 nr_pages -= page_increm;
1670 } while (nr_pages);
1671 out:
1672 if (ctx.pgmap)
1673 put_dev_pagemap(ctx.pgmap);
1674 return i ? i : ret;
1675 }
1676
vma_permits_fault(struct vm_area_struct * vma,unsigned int fault_flags)1677 static bool vma_permits_fault(struct vm_area_struct *vma,
1678 unsigned int fault_flags)
1679 {
1680 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1681 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1682 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1683
1684 if (!(vm_flags & vma->vm_flags))
1685 return false;
1686
1687 /*
1688 * The architecture might have a hardware protection
1689 * mechanism other than read/write that can deny access.
1690 *
1691 * gup always represents data access, not instruction
1692 * fetches, so execute=false here:
1693 */
1694 if (!arch_vma_access_permitted(vma, write, false, foreign))
1695 return false;
1696
1697 return true;
1698 }
1699
1700 /**
1701 * fixup_user_fault() - manually resolve a user page fault
1702 * @mm: mm_struct of target mm
1703 * @address: user address
1704 * @fault_flags:flags to pass down to handle_mm_fault()
1705 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1706 * does not allow retry. If NULL, the caller must guarantee
1707 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1708 *
1709 * This is meant to be called in the specific scenario where for locking reasons
1710 * we try to access user memory in atomic context (within a pagefault_disable()
1711 * section), this returns -EFAULT, and we want to resolve the user fault before
1712 * trying again.
1713 *
1714 * Typically this is meant to be used by the futex code.
1715 *
1716 * The main difference with get_user_pages() is that this function will
1717 * unconditionally call handle_mm_fault() which will in turn perform all the
1718 * necessary SW fixup of the dirty and young bits in the PTE, while
1719 * get_user_pages() only guarantees to update these in the struct page.
1720 *
1721 * This is important for some architectures where those bits also gate the
1722 * access permission to the page because they are maintained in software. On
1723 * such architectures, gup() will not be enough to make a subsequent access
1724 * succeed.
1725 *
1726 * This function will not return with an unlocked mmap_lock. So it has not the
1727 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1728 */
fixup_user_fault(struct mm_struct * mm,unsigned long address,unsigned int fault_flags,bool * unlocked)1729 int fixup_user_fault(struct mm_struct *mm,
1730 unsigned long address, unsigned int fault_flags,
1731 bool *unlocked)
1732 {
1733 struct vm_area_struct *vma;
1734 vm_fault_t ret;
1735
1736 address = untagged_addr_remote(mm, address);
1737
1738 if (unlocked)
1739 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1740
1741 retry:
1742 vma = gup_vma_lookup(mm, address);
1743 if (!vma)
1744 return -EFAULT;
1745
1746 if (!vma_permits_fault(vma, fault_flags))
1747 return -EFAULT;
1748
1749 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1750 fatal_signal_pending(current))
1751 return -EINTR;
1752
1753 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1754
1755 if (ret & VM_FAULT_COMPLETED) {
1756 /*
1757 * NOTE: it's a pity that we need to retake the lock here
1758 * to pair with the unlock() in the callers. Ideally we
1759 * could tell the callers so they do not need to unlock.
1760 */
1761 mmap_read_lock(mm);
1762 *unlocked = true;
1763 return 0;
1764 }
1765
1766 if (ret & VM_FAULT_ERROR) {
1767 int err = vm_fault_to_errno(ret, 0);
1768
1769 if (err)
1770 return err;
1771 BUG();
1772 }
1773
1774 if (ret & VM_FAULT_RETRY) {
1775 mmap_read_lock(mm);
1776 *unlocked = true;
1777 fault_flags |= FAULT_FLAG_TRIED;
1778 goto retry;
1779 }
1780
1781 return 0;
1782 }
1783 EXPORT_SYMBOL_GPL(fixup_user_fault);
1784
1785 /*
1786 * GUP always responds to fatal signals. When FOLL_INTERRUPTIBLE is
1787 * specified, it'll also respond to generic signals. The caller of GUP
1788 * that has FOLL_INTERRUPTIBLE should take care of the GUP interruption.
1789 */
gup_signal_pending(unsigned int flags)1790 static bool gup_signal_pending(unsigned int flags)
1791 {
1792 if (fatal_signal_pending(current))
1793 return true;
1794
1795 if (!(flags & FOLL_INTERRUPTIBLE))
1796 return false;
1797
1798 return signal_pending(current);
1799 }
1800
1801 /*
1802 * Locking: (*locked == 1) means that the mmap_lock has already been acquired by
1803 * the caller. This function may drop the mmap_lock. If it does so, then it will
1804 * set (*locked = 0).
1805 *
1806 * (*locked == 0) means that the caller expects this function to acquire and
1807 * drop the mmap_lock. Therefore, the value of *locked will still be zero when
1808 * the function returns, even though it may have changed temporarily during
1809 * function execution.
1810 *
1811 * Please note that this function, unlike __get_user_pages(), will not return 0
1812 * for nr_pages > 0, unless FOLL_NOWAIT is used.
1813 */
__get_user_pages_locked(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,int * locked,unsigned int flags)1814 static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1815 unsigned long start,
1816 unsigned long nr_pages,
1817 struct page **pages,
1818 int *locked,
1819 unsigned int flags)
1820 {
1821 long ret, pages_done;
1822 bool must_unlock = false;
1823
1824 if (!nr_pages)
1825 return 0;
1826
1827 /*
1828 * The internal caller expects GUP to manage the lock internally and the
1829 * lock must be released when this returns.
1830 */
1831 if (!*locked) {
1832 if (mmap_read_lock_killable(mm))
1833 return -EAGAIN;
1834 must_unlock = true;
1835 *locked = 1;
1836 }
1837 else
1838 mmap_assert_locked(mm);
1839
1840 if (flags & FOLL_PIN)
1841 mm_set_has_pinned_flag(&mm->flags);
1842
1843 /*
1844 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1845 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1846 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1847 * for FOLL_GET, not for the newer FOLL_PIN.
1848 *
1849 * FOLL_PIN always expects pages to be non-null, but no need to assert
1850 * that here, as any failures will be obvious enough.
1851 */
1852 if (pages && !(flags & FOLL_PIN))
1853 flags |= FOLL_GET;
1854
1855 pages_done = 0;
1856 for (;;) {
1857 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1858 locked);
1859 if (!(flags & FOLL_UNLOCKABLE)) {
1860 /* VM_FAULT_RETRY couldn't trigger, bypass */
1861 pages_done = ret;
1862 break;
1863 }
1864
1865 /* VM_FAULT_RETRY or VM_FAULT_COMPLETED cannot return errors */
1866 if (!*locked) {
1867 BUG_ON(ret < 0);
1868 BUG_ON(ret >= nr_pages);
1869 }
1870
1871 if (ret > 0) {
1872 nr_pages -= ret;
1873 pages_done += ret;
1874 if (!nr_pages)
1875 break;
1876 }
1877 if (*locked) {
1878 /*
1879 * VM_FAULT_RETRY didn't trigger or it was a
1880 * FOLL_NOWAIT.
1881 */
1882 if (!pages_done)
1883 pages_done = ret;
1884 break;
1885 }
1886 /*
1887 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1888 * For the prefault case (!pages) we only update counts.
1889 */
1890 if (likely(pages))
1891 pages += ret;
1892 start += ret << PAGE_SHIFT;
1893
1894 /* The lock was temporarily dropped, so we must unlock later */
1895 must_unlock = true;
1896
1897 retry:
1898 /*
1899 * Repeat on the address that fired VM_FAULT_RETRY
1900 * with both FAULT_FLAG_ALLOW_RETRY and
1901 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1902 * by fatal signals of even common signals, depending on
1903 * the caller's request. So we need to check it before we
1904 * start trying again otherwise it can loop forever.
1905 */
1906 if (gup_signal_pending(flags)) {
1907 if (!pages_done)
1908 pages_done = -EINTR;
1909 break;
1910 }
1911
1912 ret = mmap_read_lock_killable(mm);
1913 if (ret) {
1914 BUG_ON(ret > 0);
1915 if (!pages_done)
1916 pages_done = ret;
1917 break;
1918 }
1919
1920 *locked = 1;
1921 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1922 pages, locked);
1923 if (!*locked) {
1924 /* Continue to retry until we succeeded */
1925 BUG_ON(ret != 0);
1926 goto retry;
1927 }
1928 if (ret != 1) {
1929 BUG_ON(ret > 1);
1930 if (!pages_done)
1931 pages_done = ret;
1932 break;
1933 }
1934 nr_pages--;
1935 pages_done++;
1936 if (!nr_pages)
1937 break;
1938 if (likely(pages))
1939 pages++;
1940 start += PAGE_SIZE;
1941 }
1942 if (must_unlock && *locked) {
1943 /*
1944 * We either temporarily dropped the lock, or the caller
1945 * requested that we both acquire and drop the lock. Either way,
1946 * we must now unlock, and notify the caller of that state.
1947 */
1948 mmap_read_unlock(mm);
1949 *locked = 0;
1950 }
1951
1952 /*
1953 * Failing to pin anything implies something has gone wrong (except when
1954 * FOLL_NOWAIT is specified).
1955 */
1956 if (WARN_ON_ONCE(pages_done == 0 && !(flags & FOLL_NOWAIT)))
1957 return -EFAULT;
1958
1959 return pages_done;
1960 }
1961
1962 /**
1963 * populate_vma_page_range() - populate a range of pages in the vma.
1964 * @vma: target vma
1965 * @start: start address
1966 * @end: end address
1967 * @locked: whether the mmap_lock is still held
1968 *
1969 * This takes care of mlocking the pages too if VM_LOCKED is set.
1970 *
1971 * Return either number of pages pinned in the vma, or a negative error
1972 * code on error.
1973 *
1974 * vma->vm_mm->mmap_lock must be held.
1975 *
1976 * If @locked is NULL, it may be held for read or write and will
1977 * be unperturbed.
1978 *
1979 * If @locked is non-NULL, it must held for read only and may be
1980 * released. If it's released, *@locked will be set to 0.
1981 */
populate_vma_page_range(struct vm_area_struct * vma,unsigned long start,unsigned long end,int * locked)1982 long populate_vma_page_range(struct vm_area_struct *vma,
1983 unsigned long start, unsigned long end, int *locked)
1984 {
1985 struct mm_struct *mm = vma->vm_mm;
1986 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1987 int local_locked = 1;
1988 int gup_flags;
1989 long ret;
1990
1991 VM_BUG_ON(!PAGE_ALIGNED(start));
1992 VM_BUG_ON(!PAGE_ALIGNED(end));
1993 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1994 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1995 mmap_assert_locked(mm);
1996
1997 /*
1998 * Rightly or wrongly, the VM_LOCKONFAULT case has never used
1999 * faultin_page() to break COW, so it has no work to do here.
2000 */
2001 if (vma->vm_flags & VM_LOCKONFAULT)
2002 return nr_pages;
2003
2004 /* ... similarly, we've never faulted in PROT_NONE pages */
2005 if (!vma_is_accessible(vma))
2006 return -EFAULT;
2007
2008 gup_flags = FOLL_TOUCH;
2009 /*
2010 * We want to touch writable mappings with a write fault in order
2011 * to break COW, except for shared mappings because these don't COW
2012 * and we would not want to dirty them for nothing.
2013 *
2014 * Otherwise, do a read fault, and use FOLL_FORCE in case it's not
2015 * readable (ie write-only or executable).
2016 */
2017 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
2018 gup_flags |= FOLL_WRITE;
2019 else
2020 gup_flags |= FOLL_FORCE;
2021
2022 if (locked)
2023 gup_flags |= FOLL_UNLOCKABLE;
2024
2025 /*
2026 * We made sure addr is within a VMA, so the following will
2027 * not result in a stack expansion that recurses back here.
2028 */
2029 ret = __get_user_pages(mm, start, nr_pages, gup_flags,
2030 NULL, locked ? locked : &local_locked);
2031 lru_add_drain();
2032 return ret;
2033 }
2034
2035 /*
2036 * faultin_page_range() - populate (prefault) page tables inside the
2037 * given range readable/writable
2038 *
2039 * This takes care of mlocking the pages, too, if VM_LOCKED is set.
2040 *
2041 * @mm: the mm to populate page tables in
2042 * @start: start address
2043 * @end: end address
2044 * @write: whether to prefault readable or writable
2045 * @locked: whether the mmap_lock is still held
2046 *
2047 * Returns either number of processed pages in the MM, or a negative error
2048 * code on error (see __get_user_pages()). Note that this function reports
2049 * errors related to VMAs, such as incompatible mappings, as expected by
2050 * MADV_POPULATE_(READ|WRITE).
2051 *
2052 * The range must be page-aligned.
2053 *
2054 * mm->mmap_lock must be held. If it's released, *@locked will be set to 0.
2055 */
faultin_page_range(struct mm_struct * mm,unsigned long start,unsigned long end,bool write,int * locked)2056 long faultin_page_range(struct mm_struct *mm, unsigned long start,
2057 unsigned long end, bool write, int *locked)
2058 {
2059 unsigned long nr_pages = (end - start) / PAGE_SIZE;
2060 int gup_flags;
2061 long ret;
2062
2063 VM_BUG_ON(!PAGE_ALIGNED(start));
2064 VM_BUG_ON(!PAGE_ALIGNED(end));
2065 mmap_assert_locked(mm);
2066
2067 /*
2068 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark
2069 * the page dirty with FOLL_WRITE -- which doesn't make a
2070 * difference with !FOLL_FORCE, because the page is writable
2071 * in the page table.
2072 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit
2073 * a poisoned page.
2074 * !FOLL_FORCE: Require proper access permissions.
2075 */
2076 gup_flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_UNLOCKABLE |
2077 FOLL_MADV_POPULATE;
2078 if (write)
2079 gup_flags |= FOLL_WRITE;
2080
2081 ret = __get_user_pages_locked(mm, start, nr_pages, NULL, locked,
2082 gup_flags);
2083 lru_add_drain();
2084 return ret;
2085 }
2086
2087 /*
2088 * __mm_populate - populate and/or mlock pages within a range of address space.
2089 *
2090 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
2091 * flags. VMAs must be already marked with the desired vm_flags, and
2092 * mmap_lock must not be held.
2093 */
__mm_populate(unsigned long start,unsigned long len,int ignore_errors)2094 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
2095 {
2096 struct mm_struct *mm = current->mm;
2097 unsigned long end, nstart, nend;
2098 struct vm_area_struct *vma = NULL;
2099 int locked = 0;
2100 long ret = 0;
2101
2102 end = start + len;
2103
2104 for (nstart = start; nstart < end; nstart = nend) {
2105 /*
2106 * We want to fault in pages for [nstart; end) address range.
2107 * Find first corresponding VMA.
2108 */
2109 if (!locked) {
2110 locked = 1;
2111 mmap_read_lock(mm);
2112 vma = find_vma_intersection(mm, nstart, end);
2113 } else if (nstart >= vma->vm_end)
2114 vma = find_vma_intersection(mm, vma->vm_end, end);
2115
2116 if (!vma)
2117 break;
2118 /*
2119 * Set [nstart; nend) to intersection of desired address
2120 * range with the first VMA. Also, skip undesirable VMA types.
2121 */
2122 nend = min(end, vma->vm_end);
2123 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
2124 continue;
2125 if (nstart < vma->vm_start)
2126 nstart = vma->vm_start;
2127 /*
2128 * Now fault in a range of pages. populate_vma_page_range()
2129 * double checks the vma flags, so that it won't mlock pages
2130 * if the vma was already munlocked.
2131 */
2132 ret = populate_vma_page_range(vma, nstart, nend, &locked);
2133 if (ret < 0) {
2134 if (ignore_errors) {
2135 ret = 0;
2136 continue; /* continue at next VMA */
2137 }
2138 break;
2139 }
2140 nend = nstart + ret * PAGE_SIZE;
2141 ret = 0;
2142 }
2143 if (locked)
2144 mmap_read_unlock(mm);
2145 return ret; /* 0 or negative error code */
2146 }
2147 #else /* CONFIG_MMU */
__get_user_pages_locked(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,int * locked,unsigned int foll_flags)2148 static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
2149 unsigned long nr_pages, struct page **pages,
2150 int *locked, unsigned int foll_flags)
2151 {
2152 struct vm_area_struct *vma;
2153 bool must_unlock = false;
2154 unsigned long vm_flags;
2155 long i;
2156
2157 if (!nr_pages)
2158 return 0;
2159
2160 /*
2161 * The internal caller expects GUP to manage the lock internally and the
2162 * lock must be released when this returns.
2163 */
2164 if (!*locked) {
2165 if (mmap_read_lock_killable(mm))
2166 return -EAGAIN;
2167 must_unlock = true;
2168 *locked = 1;
2169 }
2170
2171 /* calculate required read or write permissions.
2172 * If FOLL_FORCE is set, we only require the "MAY" flags.
2173 */
2174 vm_flags = (foll_flags & FOLL_WRITE) ?
2175 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
2176 vm_flags &= (foll_flags & FOLL_FORCE) ?
2177 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
2178
2179 for (i = 0; i < nr_pages; i++) {
2180 vma = find_vma(mm, start);
2181 if (!vma)
2182 break;
2183
2184 /* protect what we can, including chardevs */
2185 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
2186 !(vm_flags & vma->vm_flags))
2187 break;
2188
2189 if (pages) {
2190 pages[i] = virt_to_page((void *)start);
2191 if (pages[i])
2192 get_page(pages[i]);
2193 }
2194
2195 start = (start + PAGE_SIZE) & PAGE_MASK;
2196 }
2197
2198 if (must_unlock && *locked) {
2199 mmap_read_unlock(mm);
2200 *locked = 0;
2201 }
2202
2203 return i ? : -EFAULT;
2204 }
2205 #endif /* !CONFIG_MMU */
2206
2207 /**
2208 * fault_in_writeable - fault in userspace address range for writing
2209 * @uaddr: start of address range
2210 * @size: size of address range
2211 *
2212 * Returns the number of bytes not faulted in (like copy_to_user() and
2213 * copy_from_user()).
2214 */
fault_in_writeable(char __user * uaddr,size_t size)2215 size_t fault_in_writeable(char __user *uaddr, size_t size)
2216 {
2217 char __user *start = uaddr, *end;
2218
2219 if (unlikely(size == 0))
2220 return 0;
2221 if (!user_write_access_begin(uaddr, size))
2222 return size;
2223 if (!PAGE_ALIGNED(uaddr)) {
2224 unsafe_put_user(0, uaddr, out);
2225 uaddr = (char __user *)PAGE_ALIGN((unsigned long)uaddr);
2226 }
2227 end = (char __user *)PAGE_ALIGN((unsigned long)start + size);
2228 if (unlikely(end < start))
2229 end = NULL;
2230 while (uaddr != end) {
2231 unsafe_put_user(0, uaddr, out);
2232 uaddr += PAGE_SIZE;
2233 }
2234
2235 out:
2236 user_write_access_end();
2237 if (size > uaddr - start)
2238 return size - (uaddr - start);
2239 return 0;
2240 }
2241 EXPORT_SYMBOL(fault_in_writeable);
2242
2243 /**
2244 * fault_in_subpage_writeable - fault in an address range for writing
2245 * @uaddr: start of address range
2246 * @size: size of address range
2247 *
2248 * Fault in a user address range for writing while checking for permissions at
2249 * sub-page granularity (e.g. arm64 MTE). This function should be used when
2250 * the caller cannot guarantee forward progress of a copy_to_user() loop.
2251 *
2252 * Returns the number of bytes not faulted in (like copy_to_user() and
2253 * copy_from_user()).
2254 */
fault_in_subpage_writeable(char __user * uaddr,size_t size)2255 size_t fault_in_subpage_writeable(char __user *uaddr, size_t size)
2256 {
2257 size_t faulted_in;
2258
2259 /*
2260 * Attempt faulting in at page granularity first for page table
2261 * permission checking. The arch-specific probe_subpage_writeable()
2262 * functions may not check for this.
2263 */
2264 faulted_in = size - fault_in_writeable(uaddr, size);
2265 if (faulted_in)
2266 faulted_in -= probe_subpage_writeable(uaddr, faulted_in);
2267
2268 return size - faulted_in;
2269 }
2270 EXPORT_SYMBOL(fault_in_subpage_writeable);
2271
2272 /*
2273 * fault_in_safe_writeable - fault in an address range for writing
2274 * @uaddr: start of address range
2275 * @size: length of address range
2276 *
2277 * Faults in an address range for writing. This is primarily useful when we
2278 * already know that some or all of the pages in the address range aren't in
2279 * memory.
2280 *
2281 * Unlike fault_in_writeable(), this function is non-destructive.
2282 *
2283 * Note that we don't pin or otherwise hold the pages referenced that we fault
2284 * in. There's no guarantee that they'll stay in memory for any duration of
2285 * time.
2286 *
2287 * Returns the number of bytes not faulted in, like copy_to_user() and
2288 * copy_from_user().
2289 */
fault_in_safe_writeable(const char __user * uaddr,size_t size)2290 size_t fault_in_safe_writeable(const char __user *uaddr, size_t size)
2291 {
2292 unsigned long start = (unsigned long)uaddr, end;
2293 struct mm_struct *mm = current->mm;
2294 bool unlocked = false;
2295
2296 if (unlikely(size == 0))
2297 return 0;
2298 end = PAGE_ALIGN(start + size);
2299 if (end < start)
2300 end = 0;
2301
2302 mmap_read_lock(mm);
2303 do {
2304 if (fixup_user_fault(mm, start, FAULT_FLAG_WRITE, &unlocked))
2305 break;
2306 start = (start + PAGE_SIZE) & PAGE_MASK;
2307 } while (start != end);
2308 mmap_read_unlock(mm);
2309
2310 if (size > (unsigned long)uaddr - start)
2311 return size - ((unsigned long)uaddr - start);
2312 return 0;
2313 }
2314 EXPORT_SYMBOL(fault_in_safe_writeable);
2315
2316 /**
2317 * fault_in_readable - fault in userspace address range for reading
2318 * @uaddr: start of user address range
2319 * @size: size of user address range
2320 *
2321 * Returns the number of bytes not faulted in (like copy_to_user() and
2322 * copy_from_user()).
2323 */
fault_in_readable(const char __user * uaddr,size_t size)2324 size_t fault_in_readable(const char __user *uaddr, size_t size)
2325 {
2326 const char __user *start = uaddr, *end;
2327 volatile char c;
2328
2329 if (unlikely(size == 0))
2330 return 0;
2331 if (!user_read_access_begin(uaddr, size))
2332 return size;
2333 if (!PAGE_ALIGNED(uaddr)) {
2334 unsafe_get_user(c, uaddr, out);
2335 uaddr = (const char __user *)PAGE_ALIGN((unsigned long)uaddr);
2336 }
2337 end = (const char __user *)PAGE_ALIGN((unsigned long)start + size);
2338 if (unlikely(end < start))
2339 end = NULL;
2340 while (uaddr != end) {
2341 unsafe_get_user(c, uaddr, out);
2342 uaddr += PAGE_SIZE;
2343 }
2344
2345 out:
2346 user_read_access_end();
2347 (void)c;
2348 if (size > uaddr - start)
2349 return size - (uaddr - start);
2350 return 0;
2351 }
2352 EXPORT_SYMBOL(fault_in_readable);
2353
2354 /**
2355 * get_dump_page() - pin user page in memory while writing it to core dump
2356 * @addr: user address
2357 *
2358 * Returns struct page pointer of user page pinned for dump,
2359 * to be freed afterwards by put_page().
2360 *
2361 * Returns NULL on any kind of failure - a hole must then be inserted into
2362 * the corefile, to preserve alignment with its headers; and also returns
2363 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2364 * allowing a hole to be left in the corefile to save disk space.
2365 *
2366 * Called without mmap_lock (takes and releases the mmap_lock by itself).
2367 */
2368 #ifdef CONFIG_ELF_CORE
get_dump_page(unsigned long addr)2369 struct page *get_dump_page(unsigned long addr)
2370 {
2371 struct page *page;
2372 int locked = 0;
2373 int ret;
2374
2375 ret = __get_user_pages_locked(current->mm, addr, 1, &page, &locked,
2376 FOLL_FORCE | FOLL_DUMP | FOLL_GET);
2377 return (ret == 1) ? page : NULL;
2378 }
2379 #endif /* CONFIG_ELF_CORE */
2380
2381 #ifdef CONFIG_MIGRATION
2382 /*
2383 * Returns the number of collected pages. Return value is always >= 0.
2384 */
collect_longterm_unpinnable_pages(struct list_head * movable_page_list,unsigned long nr_pages,struct page ** pages)2385 static unsigned long collect_longterm_unpinnable_pages(
2386 struct list_head *movable_page_list,
2387 unsigned long nr_pages,
2388 struct page **pages)
2389 {
2390 unsigned long i, collected = 0;
2391 struct folio *prev_folio = NULL;
2392 bool drain_allow = true;
2393
2394 for (i = 0; i < nr_pages; i++) {
2395 struct folio *folio = page_folio(pages[i]);
2396
2397 if (folio == prev_folio)
2398 continue;
2399 prev_folio = folio;
2400
2401 if (folio_is_longterm_pinnable(folio))
2402 continue;
2403
2404 collected++;
2405
2406 if (folio_is_device_coherent(folio))
2407 continue;
2408
2409 if (folio_test_hugetlb(folio)) {
2410 isolate_hugetlb(folio, movable_page_list);
2411 continue;
2412 }
2413
2414 if (!folio_test_lru(folio) && drain_allow) {
2415 lru_add_drain_all();
2416 drain_allow = false;
2417 }
2418
2419 if (!folio_isolate_lru(folio))
2420 continue;
2421
2422 list_add_tail(&folio->lru, movable_page_list);
2423 node_stat_mod_folio(folio,
2424 NR_ISOLATED_ANON + folio_is_file_lru(folio),
2425 folio_nr_pages(folio));
2426 }
2427
2428 return collected;
2429 }
2430
2431 /*
2432 * Unpins all pages and migrates device coherent pages and movable_page_list.
2433 * Returns -EAGAIN if all pages were successfully migrated or -errno for failure
2434 * (or partial success).
2435 */
migrate_longterm_unpinnable_pages(struct list_head * movable_page_list,unsigned long nr_pages,struct page ** pages)2436 static int migrate_longterm_unpinnable_pages(
2437 struct list_head *movable_page_list,
2438 unsigned long nr_pages,
2439 struct page **pages)
2440 {
2441 int ret;
2442 unsigned long i;
2443
2444 for (i = 0; i < nr_pages; i++) {
2445 struct folio *folio = page_folio(pages[i]);
2446
2447 if (folio_is_device_coherent(folio)) {
2448 /*
2449 * Migration will fail if the page is pinned, so convert
2450 * the pin on the source page to a normal reference.
2451 */
2452 pages[i] = NULL;
2453 folio_get(folio);
2454 gup_put_folio(folio, 1, FOLL_PIN);
2455
2456 if (migrate_device_coherent_page(&folio->page)) {
2457 ret = -EBUSY;
2458 goto err;
2459 }
2460
2461 continue;
2462 }
2463
2464 /*
2465 * We can't migrate pages with unexpected references, so drop
2466 * the reference obtained by __get_user_pages_locked().
2467 * Migrating pages have been added to movable_page_list after
2468 * calling folio_isolate_lru() which takes a reference so the
2469 * page won't be freed if it's migrating.
2470 */
2471 unpin_user_page(pages[i]);
2472 pages[i] = NULL;
2473 }
2474
2475 if (!list_empty(movable_page_list)) {
2476 struct migration_target_control mtc = {
2477 .nid = NUMA_NO_NODE,
2478 .gfp_mask = GFP_USER | __GFP_NOWARN,
2479 .reason = MR_LONGTERM_PIN,
2480 };
2481
2482 if (migrate_pages(movable_page_list, alloc_migration_target,
2483 NULL, (unsigned long)&mtc, MIGRATE_SYNC,
2484 MR_LONGTERM_PIN, NULL)) {
2485 ret = -ENOMEM;
2486 goto err;
2487 }
2488 }
2489
2490 putback_movable_pages(movable_page_list);
2491
2492 return -EAGAIN;
2493
2494 err:
2495 for (i = 0; i < nr_pages; i++)
2496 if (pages[i])
2497 unpin_user_page(pages[i]);
2498 putback_movable_pages(movable_page_list);
2499
2500 return ret;
2501 }
2502
2503 /*
2504 * Check whether all pages are *allowed* to be pinned. Rather confusingly, all
2505 * pages in the range are required to be pinned via FOLL_PIN, before calling
2506 * this routine.
2507 *
2508 * If any pages in the range are not allowed to be pinned, then this routine
2509 * will migrate those pages away, unpin all the pages in the range and return
2510 * -EAGAIN. The caller should re-pin the entire range with FOLL_PIN and then
2511 * call this routine again.
2512 *
2513 * If an error other than -EAGAIN occurs, this indicates a migration failure.
2514 * The caller should give up, and propagate the error back up the call stack.
2515 *
2516 * If everything is OK and all pages in the range are allowed to be pinned, then
2517 * this routine leaves all pages pinned and returns zero for success.
2518 */
check_and_migrate_movable_pages(unsigned long nr_pages,struct page ** pages)2519 static long check_and_migrate_movable_pages(unsigned long nr_pages,
2520 struct page **pages)
2521 {
2522 unsigned long collected;
2523 LIST_HEAD(movable_page_list);
2524
2525 collected = collect_longterm_unpinnable_pages(&movable_page_list,
2526 nr_pages, pages);
2527 if (!collected)
2528 return 0;
2529
2530 return migrate_longterm_unpinnable_pages(&movable_page_list, nr_pages,
2531 pages);
2532 }
2533 #else
check_and_migrate_movable_pages(unsigned long nr_pages,struct page ** pages)2534 static long check_and_migrate_movable_pages(unsigned long nr_pages,
2535 struct page **pages)
2536 {
2537 return 0;
2538 }
2539 #endif /* CONFIG_MIGRATION */
2540
2541 /*
2542 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
2543 * allows us to process the FOLL_LONGTERM flag.
2544 */
__gup_longterm_locked(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,int * locked,unsigned int gup_flags)2545 static long __gup_longterm_locked(struct mm_struct *mm,
2546 unsigned long start,
2547 unsigned long nr_pages,
2548 struct page **pages,
2549 int *locked,
2550 unsigned int gup_flags)
2551 {
2552 unsigned int flags;
2553 long rc, nr_pinned_pages;
2554
2555 if (!(gup_flags & FOLL_LONGTERM))
2556 return __get_user_pages_locked(mm, start, nr_pages, pages,
2557 locked, gup_flags);
2558
2559 flags = memalloc_pin_save();
2560 do {
2561 nr_pinned_pages = __get_user_pages_locked(mm, start, nr_pages,
2562 pages, locked,
2563 gup_flags);
2564 if (nr_pinned_pages <= 0) {
2565 rc = nr_pinned_pages;
2566 break;
2567 }
2568
2569 /* FOLL_LONGTERM implies FOLL_PIN */
2570 rc = check_and_migrate_movable_pages(nr_pinned_pages, pages);
2571 } while (rc == -EAGAIN);
2572 memalloc_pin_restore(flags);
2573 return rc ? rc : nr_pinned_pages;
2574 }
2575
2576 /*
2577 * Check that the given flags are valid for the exported gup/pup interface, and
2578 * update them with the required flags that the caller must have set.
2579 */
is_valid_gup_args(struct page ** pages,int * locked,unsigned int * gup_flags_p,unsigned int to_set)2580 static bool is_valid_gup_args(struct page **pages, int *locked,
2581 unsigned int *gup_flags_p, unsigned int to_set)
2582 {
2583 unsigned int gup_flags = *gup_flags_p;
2584
2585 /*
2586 * These flags not allowed to be specified externally to the gup
2587 * interfaces:
2588 * - FOLL_TOUCH/FOLL_PIN/FOLL_TRIED/FOLL_FAST_ONLY are internal only
2589 * - FOLL_REMOTE is internal only and used on follow_page()
2590 * - FOLL_UNLOCKABLE is internal only and used if locked is !NULL
2591 */
2592 if (WARN_ON_ONCE(gup_flags & INTERNAL_GUP_FLAGS))
2593 return false;
2594
2595 gup_flags |= to_set;
2596 if (locked) {
2597 /* At the external interface locked must be set */
2598 if (WARN_ON_ONCE(*locked != 1))
2599 return false;
2600
2601 gup_flags |= FOLL_UNLOCKABLE;
2602 }
2603
2604 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2605 if (WARN_ON_ONCE((gup_flags & (FOLL_PIN | FOLL_GET)) ==
2606 (FOLL_PIN | FOLL_GET)))
2607 return false;
2608
2609 /* LONGTERM can only be specified when pinning */
2610 if (WARN_ON_ONCE(!(gup_flags & FOLL_PIN) && (gup_flags & FOLL_LONGTERM)))
2611 return false;
2612
2613 /* Pages input must be given if using GET/PIN */
2614 if (WARN_ON_ONCE((gup_flags & (FOLL_GET | FOLL_PIN)) && !pages))
2615 return false;
2616
2617 /* We want to allow the pgmap to be hot-unplugged at all times */
2618 if (WARN_ON_ONCE((gup_flags & FOLL_LONGTERM) &&
2619 (gup_flags & FOLL_PCI_P2PDMA)))
2620 return false;
2621
2622 *gup_flags_p = gup_flags;
2623 return true;
2624 }
2625
2626 #ifdef CONFIG_MMU
2627 /**
2628 * get_user_pages_remote() - pin user pages in memory
2629 * @mm: mm_struct of target mm
2630 * @start: starting user address
2631 * @nr_pages: number of pages from start to pin
2632 * @gup_flags: flags modifying lookup behaviour
2633 * @pages: array that receives pointers to the pages pinned.
2634 * Should be at least nr_pages long. Or NULL, if caller
2635 * only intends to ensure the pages are faulted in.
2636 * @locked: pointer to lock flag indicating whether lock is held and
2637 * subsequently whether VM_FAULT_RETRY functionality can be
2638 * utilised. Lock must initially be held.
2639 *
2640 * Returns either number of pages pinned (which may be less than the
2641 * number requested), or an error. Details about the return value:
2642 *
2643 * -- If nr_pages is 0, returns 0.
2644 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
2645 * -- If nr_pages is >0, and some pages were pinned, returns the number of
2646 * pages pinned. Again, this may be less than nr_pages.
2647 *
2648 * The caller is responsible for releasing returned @pages, via put_page().
2649 *
2650 * Must be called with mmap_lock held for read or write.
2651 *
2652 * get_user_pages_remote walks a process's page tables and takes a reference
2653 * to each struct page that each user address corresponds to at a given
2654 * instant. That is, it takes the page that would be accessed if a user
2655 * thread accesses the given user virtual address at that instant.
2656 *
2657 * This does not guarantee that the page exists in the user mappings when
2658 * get_user_pages_remote returns, and there may even be a completely different
2659 * page there in some cases (eg. if mmapped pagecache has been invalidated
2660 * and subsequently re-faulted). However it does guarantee that the page
2661 * won't be freed completely. And mostly callers simply care that the page
2662 * contains data that was valid *at some point in time*. Typically, an IO
2663 * or similar operation cannot guarantee anything stronger anyway because
2664 * locks can't be held over the syscall boundary.
2665 *
2666 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
2667 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
2668 * be called after the page is finished with, and before put_page is called.
2669 *
2670 * get_user_pages_remote is typically used for fewer-copy IO operations,
2671 * to get a handle on the memory by some means other than accesses
2672 * via the user virtual addresses. The pages may be submitted for
2673 * DMA to devices or accessed via their kernel linear mapping (via the
2674 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
2675 *
2676 * See also get_user_pages_fast, for performance critical applications.
2677 *
2678 * get_user_pages_remote should be phased out in favor of
2679 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
2680 * should use get_user_pages_remote because it cannot pass
2681 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
2682 */
get_user_pages_remote(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)2683 long get_user_pages_remote(struct mm_struct *mm,
2684 unsigned long start, unsigned long nr_pages,
2685 unsigned int gup_flags, struct page **pages,
2686 int *locked)
2687 {
2688 int local_locked = 1;
2689
2690 if (!is_valid_gup_args(pages, locked, &gup_flags,
2691 FOLL_TOUCH | FOLL_REMOTE))
2692 return -EINVAL;
2693
2694 return __get_user_pages_locked(mm, start, nr_pages, pages,
2695 locked ? locked : &local_locked,
2696 gup_flags);
2697 }
2698 EXPORT_SYMBOL(get_user_pages_remote);
2699
2700 #else /* CONFIG_MMU */
get_user_pages_remote(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)2701 long get_user_pages_remote(struct mm_struct *mm,
2702 unsigned long start, unsigned long nr_pages,
2703 unsigned int gup_flags, struct page **pages,
2704 int *locked)
2705 {
2706 return 0;
2707 }
2708 #endif /* !CONFIG_MMU */
2709
2710 /**
2711 * get_user_pages() - pin user pages in memory
2712 * @start: starting user address
2713 * @nr_pages: number of pages from start to pin
2714 * @gup_flags: flags modifying lookup behaviour
2715 * @pages: array that receives pointers to the pages pinned.
2716 * Should be at least nr_pages long. Or NULL, if caller
2717 * only intends to ensure the pages are faulted in.
2718 *
2719 * This is the same as get_user_pages_remote(), just with a less-flexible
2720 * calling convention where we assume that the mm being operated on belongs to
2721 * the current task, and doesn't allow passing of a locked parameter. We also
2722 * obviously don't pass FOLL_REMOTE in here.
2723 */
get_user_pages(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages)2724 long get_user_pages(unsigned long start, unsigned long nr_pages,
2725 unsigned int gup_flags, struct page **pages)
2726 {
2727 int locked = 1;
2728
2729 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_TOUCH))
2730 return -EINVAL;
2731
2732 return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2733 &locked, gup_flags);
2734 }
2735 EXPORT_SYMBOL(get_user_pages);
2736
2737 /*
2738 * get_user_pages_unlocked() is suitable to replace the form:
2739 *
2740 * mmap_read_lock(mm);
2741 * get_user_pages(mm, ..., pages, NULL);
2742 * mmap_read_unlock(mm);
2743 *
2744 * with:
2745 *
2746 * get_user_pages_unlocked(mm, ..., pages);
2747 *
2748 * It is functionally equivalent to get_user_pages_fast so
2749 * get_user_pages_fast should be used instead if specific gup_flags
2750 * (e.g. FOLL_FORCE) are not required.
2751 */
get_user_pages_unlocked(unsigned long start,unsigned long nr_pages,struct page ** pages,unsigned int gup_flags)2752 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2753 struct page **pages, unsigned int gup_flags)
2754 {
2755 int locked = 0;
2756
2757 if (!is_valid_gup_args(pages, NULL, &gup_flags,
2758 FOLL_TOUCH | FOLL_UNLOCKABLE))
2759 return -EINVAL;
2760
2761 return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2762 &locked, gup_flags);
2763 }
2764 EXPORT_SYMBOL(get_user_pages_unlocked);
2765
2766 /*
2767 * GUP-fast
2768 *
2769 * get_user_pages_fast attempts to pin user pages by walking the page
2770 * tables directly and avoids taking locks. Thus the walker needs to be
2771 * protected from page table pages being freed from under it, and should
2772 * block any THP splits.
2773 *
2774 * One way to achieve this is to have the walker disable interrupts, and
2775 * rely on IPIs from the TLB flushing code blocking before the page table
2776 * pages are freed. This is unsuitable for architectures that do not need
2777 * to broadcast an IPI when invalidating TLBs.
2778 *
2779 * Another way to achieve this is to batch up page table containing pages
2780 * belonging to more than one mm_user, then rcu_sched a callback to free those
2781 * pages. Disabling interrupts will allow the gup_fast() walker to both block
2782 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2783 * (which is a relatively rare event). The code below adopts this strategy.
2784 *
2785 * Before activating this code, please be aware that the following assumptions
2786 * are currently made:
2787 *
2788 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2789 * free pages containing page tables or TLB flushing requires IPI broadcast.
2790 *
2791 * *) ptes can be read atomically by the architecture.
2792 *
2793 * *) access_ok is sufficient to validate userspace address ranges.
2794 *
2795 * The last two assumptions can be relaxed by the addition of helper functions.
2796 *
2797 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2798 */
2799 #ifdef CONFIG_HAVE_GUP_FAST
2800
2801 /*
2802 * Used in the GUP-fast path to determine whether GUP is permitted to work on
2803 * a specific folio.
2804 *
2805 * This call assumes the caller has pinned the folio, that the lowest page table
2806 * level still points to this folio, and that interrupts have been disabled.
2807 *
2808 * GUP-fast must reject all secretmem folios.
2809 *
2810 * Writing to pinned file-backed dirty tracked folios is inherently problematic
2811 * (see comment describing the writable_file_mapping_allowed() function). We
2812 * therefore try to avoid the most egregious case of a long-term mapping doing
2813 * so.
2814 *
2815 * This function cannot be as thorough as that one as the VMA is not available
2816 * in the fast path, so instead we whitelist known good cases and if in doubt,
2817 * fall back to the slow path.
2818 */
gup_fast_folio_allowed(struct folio * folio,unsigned int flags)2819 static bool gup_fast_folio_allowed(struct folio *folio, unsigned int flags)
2820 {
2821 bool reject_file_backed = false;
2822 struct address_space *mapping;
2823 bool check_secretmem = false;
2824 unsigned long mapping_flags;
2825
2826 /*
2827 * If we aren't pinning then no problematic write can occur. A long term
2828 * pin is the most egregious case so this is the one we disallow.
2829 */
2830 if ((flags & (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE)) ==
2831 (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE))
2832 reject_file_backed = true;
2833
2834 /* We hold a folio reference, so we can safely access folio fields. */
2835
2836 /* secretmem folios are always order-0 folios. */
2837 if (IS_ENABLED(CONFIG_SECRETMEM) && !folio_test_large(folio))
2838 check_secretmem = true;
2839
2840 if (!reject_file_backed && !check_secretmem)
2841 return true;
2842
2843 if (WARN_ON_ONCE(folio_test_slab(folio)))
2844 return false;
2845
2846 /* hugetlb neither requires dirty-tracking nor can be secretmem. */
2847 if (folio_test_hugetlb(folio))
2848 return true;
2849
2850 /*
2851 * GUP-fast disables IRQs. When IRQS are disabled, RCU grace periods
2852 * cannot proceed, which means no actions performed under RCU can
2853 * proceed either.
2854 *
2855 * inodes and thus their mappings are freed under RCU, which means the
2856 * mapping cannot be freed beneath us and thus we can safely dereference
2857 * it.
2858 */
2859 lockdep_assert_irqs_disabled();
2860
2861 /*
2862 * However, there may be operations which _alter_ the mapping, so ensure
2863 * we read it once and only once.
2864 */
2865 mapping = READ_ONCE(folio->mapping);
2866
2867 /*
2868 * The mapping may have been truncated, in any case we cannot determine
2869 * if this mapping is safe - fall back to slow path to determine how to
2870 * proceed.
2871 */
2872 if (!mapping)
2873 return false;
2874
2875 /* Anonymous folios pose no problem. */
2876 mapping_flags = (unsigned long)mapping & PAGE_MAPPING_FLAGS;
2877 if (mapping_flags)
2878 return mapping_flags & PAGE_MAPPING_ANON;
2879
2880 /*
2881 * At this point, we know the mapping is non-null and points to an
2882 * address_space object.
2883 */
2884 if (check_secretmem && secretmem_mapping(mapping))
2885 return false;
2886 /* The only remaining allowed file system is shmem. */
2887 return !reject_file_backed || shmem_mapping(mapping);
2888 }
2889
gup_fast_undo_dev_pagemap(int * nr,int nr_start,unsigned int flags,struct page ** pages)2890 static void __maybe_unused gup_fast_undo_dev_pagemap(int *nr, int nr_start,
2891 unsigned int flags, struct page **pages)
2892 {
2893 while ((*nr) - nr_start) {
2894 struct folio *folio = page_folio(pages[--(*nr)]);
2895
2896 folio_clear_referenced(folio);
2897 gup_put_folio(folio, 1, flags);
2898 }
2899 }
2900
2901 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2902 /*
2903 * GUP-fast relies on pte change detection to avoid concurrent pgtable
2904 * operations.
2905 *
2906 * To pin the page, GUP-fast needs to do below in order:
2907 * (1) pin the page (by prefetching pte), then (2) check pte not changed.
2908 *
2909 * For the rest of pgtable operations where pgtable updates can be racy
2910 * with GUP-fast, we need to do (1) clear pte, then (2) check whether page
2911 * is pinned.
2912 *
2913 * Above will work for all pte-level operations, including THP split.
2914 *
2915 * For THP collapse, it's a bit more complicated because GUP-fast may be
2916 * walking a pgtable page that is being freed (pte is still valid but pmd
2917 * can be cleared already). To avoid race in such condition, we need to
2918 * also check pmd here to make sure pmd doesn't change (corresponds to
2919 * pmdp_collapse_flush() in the THP collapse code path).
2920 */
gup_fast_pte_range(pmd_t pmd,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)2921 static int gup_fast_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2922 unsigned long end, unsigned int flags, struct page **pages,
2923 int *nr)
2924 {
2925 struct dev_pagemap *pgmap = NULL;
2926 int nr_start = *nr, ret = 0;
2927 pte_t *ptep, *ptem;
2928
2929 ptem = ptep = pte_offset_map(&pmd, addr);
2930 if (!ptep)
2931 return 0;
2932 do {
2933 pte_t pte = ptep_get_lockless(ptep);
2934 struct page *page;
2935 struct folio *folio;
2936
2937 /*
2938 * Always fallback to ordinary GUP on PROT_NONE-mapped pages:
2939 * pte_access_permitted() better should reject these pages
2940 * either way: otherwise, GUP-fast might succeed in
2941 * cases where ordinary GUP would fail due to VMA access
2942 * permissions.
2943 */
2944 if (pte_protnone(pte))
2945 goto pte_unmap;
2946
2947 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2948 goto pte_unmap;
2949
2950 if (pte_devmap(pte)) {
2951 if (unlikely(flags & FOLL_LONGTERM))
2952 goto pte_unmap;
2953
2954 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2955 if (unlikely(!pgmap)) {
2956 gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
2957 goto pte_unmap;
2958 }
2959 } else if (pte_special(pte))
2960 goto pte_unmap;
2961
2962 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2963 page = pte_page(pte);
2964
2965 folio = try_grab_folio(page, 1, flags);
2966 if (!folio)
2967 goto pte_unmap;
2968
2969 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp)) ||
2970 unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) {
2971 gup_put_folio(folio, 1, flags);
2972 goto pte_unmap;
2973 }
2974
2975 if (!gup_fast_folio_allowed(folio, flags)) {
2976 gup_put_folio(folio, 1, flags);
2977 goto pte_unmap;
2978 }
2979
2980 if (!pte_write(pte) && gup_must_unshare(NULL, flags, page)) {
2981 gup_put_folio(folio, 1, flags);
2982 goto pte_unmap;
2983 }
2984
2985 /*
2986 * We need to make the page accessible if and only if we are
2987 * going to access its content (the FOLL_PIN case). Please
2988 * see Documentation/core-api/pin_user_pages.rst for
2989 * details.
2990 */
2991 if (flags & FOLL_PIN) {
2992 ret = arch_make_page_accessible(page);
2993 if (ret) {
2994 gup_put_folio(folio, 1, flags);
2995 goto pte_unmap;
2996 }
2997 }
2998 folio_set_referenced(folio);
2999 pages[*nr] = page;
3000 (*nr)++;
3001 } while (ptep++, addr += PAGE_SIZE, addr != end);
3002
3003 ret = 1;
3004
3005 pte_unmap:
3006 if (pgmap)
3007 put_dev_pagemap(pgmap);
3008 pte_unmap(ptem);
3009 return ret;
3010 }
3011 #else
3012
3013 /*
3014 * If we can't determine whether or not a pte is special, then fail immediately
3015 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
3016 * to be special.
3017 *
3018 * For a futex to be placed on a THP tail page, get_futex_key requires a
3019 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
3020 * useful to have gup_fast_pmd_leaf even if we can't operate on ptes.
3021 */
gup_fast_pte_range(pmd_t pmd,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3022 static int gup_fast_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
3023 unsigned long end, unsigned int flags, struct page **pages,
3024 int *nr)
3025 {
3026 return 0;
3027 }
3028 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
3029
3030 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
gup_fast_devmap_leaf(unsigned long pfn,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3031 static int gup_fast_devmap_leaf(unsigned long pfn, unsigned long addr,
3032 unsigned long end, unsigned int flags, struct page **pages, int *nr)
3033 {
3034 int nr_start = *nr;
3035 struct dev_pagemap *pgmap = NULL;
3036
3037 do {
3038 struct folio *folio;
3039 struct page *page = pfn_to_page(pfn);
3040
3041 pgmap = get_dev_pagemap(pfn, pgmap);
3042 if (unlikely(!pgmap)) {
3043 gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
3044 break;
3045 }
3046
3047 if (!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)) {
3048 gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
3049 break;
3050 }
3051
3052 folio = try_grab_folio(page, 1, flags);
3053 if (!folio) {
3054 gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
3055 break;
3056 }
3057 folio_set_referenced(folio);
3058 pages[*nr] = page;
3059 (*nr)++;
3060 pfn++;
3061 } while (addr += PAGE_SIZE, addr != end);
3062
3063 put_dev_pagemap(pgmap);
3064 return addr == end;
3065 }
3066
gup_fast_devmap_pmd_leaf(pmd_t orig,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3067 static int gup_fast_devmap_pmd_leaf(pmd_t orig, pmd_t *pmdp, unsigned long addr,
3068 unsigned long end, unsigned int flags, struct page **pages,
3069 int *nr)
3070 {
3071 unsigned long fault_pfn;
3072 int nr_start = *nr;
3073
3074 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
3075 if (!gup_fast_devmap_leaf(fault_pfn, addr, end, flags, pages, nr))
3076 return 0;
3077
3078 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
3079 gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
3080 return 0;
3081 }
3082 return 1;
3083 }
3084
gup_fast_devmap_pud_leaf(pud_t orig,pud_t * pudp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3085 static int gup_fast_devmap_pud_leaf(pud_t orig, pud_t *pudp, unsigned long addr,
3086 unsigned long end, unsigned int flags, struct page **pages,
3087 int *nr)
3088 {
3089 unsigned long fault_pfn;
3090 int nr_start = *nr;
3091
3092 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
3093 if (!gup_fast_devmap_leaf(fault_pfn, addr, end, flags, pages, nr))
3094 return 0;
3095
3096 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
3097 gup_fast_undo_dev_pagemap(nr, nr_start, flags, pages);
3098 return 0;
3099 }
3100 return 1;
3101 }
3102 #else
gup_fast_devmap_pmd_leaf(pmd_t orig,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3103 static int gup_fast_devmap_pmd_leaf(pmd_t orig, pmd_t *pmdp, unsigned long addr,
3104 unsigned long end, unsigned int flags, struct page **pages,
3105 int *nr)
3106 {
3107 BUILD_BUG();
3108 return 0;
3109 }
3110
gup_fast_devmap_pud_leaf(pud_t pud,pud_t * pudp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3111 static int gup_fast_devmap_pud_leaf(pud_t pud, pud_t *pudp, unsigned long addr,
3112 unsigned long end, unsigned int flags, struct page **pages,
3113 int *nr)
3114 {
3115 BUILD_BUG();
3116 return 0;
3117 }
3118 #endif
3119
gup_fast_pmd_leaf(pmd_t orig,pmd_t * pmdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3120 static int gup_fast_pmd_leaf(pmd_t orig, pmd_t *pmdp, unsigned long addr,
3121 unsigned long end, unsigned int flags, struct page **pages,
3122 int *nr)
3123 {
3124 struct page *page;
3125 struct folio *folio;
3126 int refs;
3127
3128 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
3129 return 0;
3130
3131 if (pmd_devmap(orig)) {
3132 if (unlikely(flags & FOLL_LONGTERM))
3133 return 0;
3134 return gup_fast_devmap_pmd_leaf(orig, pmdp, addr, end, flags,
3135 pages, nr);
3136 }
3137
3138 page = pmd_page(orig);
3139 refs = record_subpages(page, PMD_SIZE, addr, end, pages + *nr);
3140
3141 folio = try_grab_folio(page, refs, flags);
3142 if (!folio)
3143 return 0;
3144
3145 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
3146 gup_put_folio(folio, refs, flags);
3147 return 0;
3148 }
3149
3150 if (!gup_fast_folio_allowed(folio, flags)) {
3151 gup_put_folio(folio, refs, flags);
3152 return 0;
3153 }
3154 if (!pmd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
3155 gup_put_folio(folio, refs, flags);
3156 return 0;
3157 }
3158
3159 *nr += refs;
3160 folio_set_referenced(folio);
3161 return 1;
3162 }
3163
gup_fast_pud_leaf(pud_t orig,pud_t * pudp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3164 static int gup_fast_pud_leaf(pud_t orig, pud_t *pudp, unsigned long addr,
3165 unsigned long end, unsigned int flags, struct page **pages,
3166 int *nr)
3167 {
3168 struct page *page;
3169 struct folio *folio;
3170 int refs;
3171
3172 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
3173 return 0;
3174
3175 if (pud_devmap(orig)) {
3176 if (unlikely(flags & FOLL_LONGTERM))
3177 return 0;
3178 return gup_fast_devmap_pud_leaf(orig, pudp, addr, end, flags,
3179 pages, nr);
3180 }
3181
3182 page = pud_page(orig);
3183 refs = record_subpages(page, PUD_SIZE, addr, end, pages + *nr);
3184
3185 folio = try_grab_folio(page, refs, flags);
3186 if (!folio)
3187 return 0;
3188
3189 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
3190 gup_put_folio(folio, refs, flags);
3191 return 0;
3192 }
3193
3194 if (!gup_fast_folio_allowed(folio, flags)) {
3195 gup_put_folio(folio, refs, flags);
3196 return 0;
3197 }
3198
3199 if (!pud_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
3200 gup_put_folio(folio, refs, flags);
3201 return 0;
3202 }
3203
3204 *nr += refs;
3205 folio_set_referenced(folio);
3206 return 1;
3207 }
3208
gup_fast_pgd_leaf(pgd_t orig,pgd_t * pgdp,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3209 static int gup_fast_pgd_leaf(pgd_t orig, pgd_t *pgdp, unsigned long addr,
3210 unsigned long end, unsigned int flags, struct page **pages,
3211 int *nr)
3212 {
3213 int refs;
3214 struct page *page;
3215 struct folio *folio;
3216
3217 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
3218 return 0;
3219
3220 BUILD_BUG_ON(pgd_devmap(orig));
3221
3222 page = pgd_page(orig);
3223 refs = record_subpages(page, PGDIR_SIZE, addr, end, pages + *nr);
3224
3225 folio = try_grab_folio(page, refs, flags);
3226 if (!folio)
3227 return 0;
3228
3229 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
3230 gup_put_folio(folio, refs, flags);
3231 return 0;
3232 }
3233
3234 if (!pgd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
3235 gup_put_folio(folio, refs, flags);
3236 return 0;
3237 }
3238
3239 if (!gup_fast_folio_allowed(folio, flags)) {
3240 gup_put_folio(folio, refs, flags);
3241 return 0;
3242 }
3243
3244 *nr += refs;
3245 folio_set_referenced(folio);
3246 return 1;
3247 }
3248
gup_fast_pmd_range(pud_t * pudp,pud_t pud,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3249 static int gup_fast_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr,
3250 unsigned long end, unsigned int flags, struct page **pages,
3251 int *nr)
3252 {
3253 unsigned long next;
3254 pmd_t *pmdp;
3255
3256 pmdp = pmd_offset_lockless(pudp, pud, addr);
3257 do {
3258 pmd_t pmd = pmdp_get_lockless(pmdp);
3259
3260 next = pmd_addr_end(addr, end);
3261 if (!pmd_present(pmd))
3262 return 0;
3263
3264 if (unlikely(pmd_leaf(pmd))) {
3265 /* See gup_fast_pte_range() */
3266 if (pmd_protnone(pmd))
3267 return 0;
3268
3269 if (!gup_fast_pmd_leaf(pmd, pmdp, addr, next, flags,
3270 pages, nr))
3271 return 0;
3272
3273 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
3274 /*
3275 * architecture have different format for hugetlbfs
3276 * pmd format and THP pmd format
3277 */
3278 if (gup_hugepd(NULL, __hugepd(pmd_val(pmd)), addr,
3279 PMD_SHIFT, next, flags, pages, nr) != 1)
3280 return 0;
3281 } else if (!gup_fast_pte_range(pmd, pmdp, addr, next, flags,
3282 pages, nr))
3283 return 0;
3284 } while (pmdp++, addr = next, addr != end);
3285
3286 return 1;
3287 }
3288
gup_fast_pud_range(p4d_t * p4dp,p4d_t p4d,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3289 static int gup_fast_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr,
3290 unsigned long end, unsigned int flags, struct page **pages,
3291 int *nr)
3292 {
3293 unsigned long next;
3294 pud_t *pudp;
3295
3296 pudp = pud_offset_lockless(p4dp, p4d, addr);
3297 do {
3298 pud_t pud = READ_ONCE(*pudp);
3299
3300 next = pud_addr_end(addr, end);
3301 if (unlikely(!pud_present(pud)))
3302 return 0;
3303 if (unlikely(pud_leaf(pud))) {
3304 if (!gup_fast_pud_leaf(pud, pudp, addr, next, flags,
3305 pages, nr))
3306 return 0;
3307 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
3308 if (gup_hugepd(NULL, __hugepd(pud_val(pud)), addr,
3309 PUD_SHIFT, next, flags, pages, nr) != 1)
3310 return 0;
3311 } else if (!gup_fast_pmd_range(pudp, pud, addr, next, flags,
3312 pages, nr))
3313 return 0;
3314 } while (pudp++, addr = next, addr != end);
3315
3316 return 1;
3317 }
3318
gup_fast_p4d_range(pgd_t * pgdp,pgd_t pgd,unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3319 static int gup_fast_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr,
3320 unsigned long end, unsigned int flags, struct page **pages,
3321 int *nr)
3322 {
3323 unsigned long next;
3324 p4d_t *p4dp;
3325
3326 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
3327 do {
3328 p4d_t p4d = READ_ONCE(*p4dp);
3329
3330 next = p4d_addr_end(addr, end);
3331 if (!p4d_present(p4d))
3332 return 0;
3333 BUILD_BUG_ON(p4d_leaf(p4d));
3334 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
3335 if (gup_hugepd(NULL, __hugepd(p4d_val(p4d)), addr,
3336 P4D_SHIFT, next, flags, pages, nr) != 1)
3337 return 0;
3338 } else if (!gup_fast_pud_range(p4dp, p4d, addr, next, flags,
3339 pages, nr))
3340 return 0;
3341 } while (p4dp++, addr = next, addr != end);
3342
3343 return 1;
3344 }
3345
gup_fast_pgd_range(unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3346 static void gup_fast_pgd_range(unsigned long addr, unsigned long end,
3347 unsigned int flags, struct page **pages, int *nr)
3348 {
3349 unsigned long next;
3350 pgd_t *pgdp;
3351
3352 pgdp = pgd_offset(current->mm, addr);
3353 do {
3354 pgd_t pgd = READ_ONCE(*pgdp);
3355
3356 next = pgd_addr_end(addr, end);
3357 if (pgd_none(pgd))
3358 return;
3359 if (unlikely(pgd_leaf(pgd))) {
3360 if (!gup_fast_pgd_leaf(pgd, pgdp, addr, next, flags,
3361 pages, nr))
3362 return;
3363 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
3364 if (gup_hugepd(NULL, __hugepd(pgd_val(pgd)), addr,
3365 PGDIR_SHIFT, next, flags, pages, nr) != 1)
3366 return;
3367 } else if (!gup_fast_p4d_range(pgdp, pgd, addr, next, flags,
3368 pages, nr))
3369 return;
3370 } while (pgdp++, addr = next, addr != end);
3371 }
3372 #else
gup_fast_pgd_range(unsigned long addr,unsigned long end,unsigned int flags,struct page ** pages,int * nr)3373 static inline void gup_fast_pgd_range(unsigned long addr, unsigned long end,
3374 unsigned int flags, struct page **pages, int *nr)
3375 {
3376 }
3377 #endif /* CONFIG_HAVE_GUP_FAST */
3378
3379 #ifndef gup_fast_permitted
3380 /*
3381 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
3382 * we need to fall back to the slow version:
3383 */
gup_fast_permitted(unsigned long start,unsigned long end)3384 static bool gup_fast_permitted(unsigned long start, unsigned long end)
3385 {
3386 return true;
3387 }
3388 #endif
3389
gup_fast(unsigned long start,unsigned long end,unsigned int gup_flags,struct page ** pages)3390 static unsigned long gup_fast(unsigned long start, unsigned long end,
3391 unsigned int gup_flags, struct page **pages)
3392 {
3393 unsigned long flags;
3394 int nr_pinned = 0;
3395 unsigned seq;
3396
3397 if (!IS_ENABLED(CONFIG_HAVE_GUP_FAST) ||
3398 !gup_fast_permitted(start, end))
3399 return 0;
3400
3401 if (gup_flags & FOLL_PIN) {
3402 seq = raw_read_seqcount(¤t->mm->write_protect_seq);
3403 if (seq & 1)
3404 return 0;
3405 }
3406
3407 /*
3408 * Disable interrupts. The nested form is used, in order to allow full,
3409 * general purpose use of this routine.
3410 *
3411 * With interrupts disabled, we block page table pages from being freed
3412 * from under us. See struct mmu_table_batch comments in
3413 * include/asm-generic/tlb.h for more details.
3414 *
3415 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
3416 * that come from THPs splitting.
3417 */
3418 local_irq_save(flags);
3419 gup_fast_pgd_range(start, end, gup_flags, pages, &nr_pinned);
3420 local_irq_restore(flags);
3421
3422 /*
3423 * When pinning pages for DMA there could be a concurrent write protect
3424 * from fork() via copy_page_range(), in this case always fail GUP-fast.
3425 */
3426 if (gup_flags & FOLL_PIN) {
3427 if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) {
3428 gup_fast_unpin_user_pages(pages, nr_pinned);
3429 return 0;
3430 } else {
3431 sanity_check_pinned_pages(pages, nr_pinned);
3432 }
3433 }
3434 return nr_pinned;
3435 }
3436
gup_fast_fallback(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages)3437 static int gup_fast_fallback(unsigned long start, unsigned long nr_pages,
3438 unsigned int gup_flags, struct page **pages)
3439 {
3440 unsigned long len, end;
3441 unsigned long nr_pinned;
3442 int locked = 0;
3443 int ret;
3444
3445 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
3446 FOLL_FORCE | FOLL_PIN | FOLL_GET |
3447 FOLL_FAST_ONLY | FOLL_NOFAULT |
3448 FOLL_PCI_P2PDMA | FOLL_HONOR_NUMA_FAULT)))
3449 return -EINVAL;
3450
3451 if (gup_flags & FOLL_PIN)
3452 mm_set_has_pinned_flag(¤t->mm->flags);
3453
3454 if (!(gup_flags & FOLL_FAST_ONLY))
3455 might_lock_read(¤t->mm->mmap_lock);
3456
3457 start = untagged_addr(start) & PAGE_MASK;
3458 len = nr_pages << PAGE_SHIFT;
3459 if (check_add_overflow(start, len, &end))
3460 return -EOVERFLOW;
3461 if (end > TASK_SIZE_MAX)
3462 return -EFAULT;
3463 if (unlikely(!access_ok((void __user *)start, len)))
3464 return -EFAULT;
3465
3466 nr_pinned = gup_fast(start, end, gup_flags, pages);
3467 if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
3468 return nr_pinned;
3469
3470 /* Slow path: try to get the remaining pages with get_user_pages */
3471 start += nr_pinned << PAGE_SHIFT;
3472 pages += nr_pinned;
3473 ret = __gup_longterm_locked(current->mm, start, nr_pages - nr_pinned,
3474 pages, &locked,
3475 gup_flags | FOLL_TOUCH | FOLL_UNLOCKABLE);
3476 if (ret < 0) {
3477 /*
3478 * The caller has to unpin the pages we already pinned so
3479 * returning -errno is not an option
3480 */
3481 if (nr_pinned)
3482 return nr_pinned;
3483 return ret;
3484 }
3485 return ret + nr_pinned;
3486 }
3487
3488 /**
3489 * get_user_pages_fast_only() - pin user pages in memory
3490 * @start: starting user address
3491 * @nr_pages: number of pages from start to pin
3492 * @gup_flags: flags modifying pin behaviour
3493 * @pages: array that receives pointers to the pages pinned.
3494 * Should be at least nr_pages long.
3495 *
3496 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
3497 * the regular GUP.
3498 *
3499 * If the architecture does not support this function, simply return with no
3500 * pages pinned.
3501 *
3502 * Careful, careful! COW breaking can go either way, so a non-write
3503 * access can get ambiguous page results. If you call this function without
3504 * 'write' set, you'd better be sure that you're ok with that ambiguity.
3505 */
get_user_pages_fast_only(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)3506 int get_user_pages_fast_only(unsigned long start, int nr_pages,
3507 unsigned int gup_flags, struct page **pages)
3508 {
3509 /*
3510 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
3511 * because gup fast is always a "pin with a +1 page refcount" request.
3512 *
3513 * FOLL_FAST_ONLY is required in order to match the API description of
3514 * this routine: no fall back to regular ("slow") GUP.
3515 */
3516 if (!is_valid_gup_args(pages, NULL, &gup_flags,
3517 FOLL_GET | FOLL_FAST_ONLY))
3518 return -EINVAL;
3519
3520 return gup_fast_fallback(start, nr_pages, gup_flags, pages);
3521 }
3522 EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
3523
3524 /**
3525 * get_user_pages_fast() - pin user pages in memory
3526 * @start: starting user address
3527 * @nr_pages: number of pages from start to pin
3528 * @gup_flags: flags modifying pin behaviour
3529 * @pages: array that receives pointers to the pages pinned.
3530 * Should be at least nr_pages long.
3531 *
3532 * Attempt to pin user pages in memory without taking mm->mmap_lock.
3533 * If not successful, it will fall back to taking the lock and
3534 * calling get_user_pages().
3535 *
3536 * Returns number of pages pinned. This may be fewer than the number requested.
3537 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
3538 * -errno.
3539 */
get_user_pages_fast(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)3540 int get_user_pages_fast(unsigned long start, int nr_pages,
3541 unsigned int gup_flags, struct page **pages)
3542 {
3543 /*
3544 * The caller may or may not have explicitly set FOLL_GET; either way is
3545 * OK. However, internally (within mm/gup.c), gup fast variants must set
3546 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
3547 * request.
3548 */
3549 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_GET))
3550 return -EINVAL;
3551 return gup_fast_fallback(start, nr_pages, gup_flags, pages);
3552 }
3553 EXPORT_SYMBOL_GPL(get_user_pages_fast);
3554
3555 /**
3556 * pin_user_pages_fast() - pin user pages in memory without taking locks
3557 *
3558 * @start: starting user address
3559 * @nr_pages: number of pages from start to pin
3560 * @gup_flags: flags modifying pin behaviour
3561 * @pages: array that receives pointers to the pages pinned.
3562 * Should be at least nr_pages long.
3563 *
3564 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
3565 * get_user_pages_fast() for documentation on the function arguments, because
3566 * the arguments here are identical.
3567 *
3568 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3569 * see Documentation/core-api/pin_user_pages.rst for further details.
3570 *
3571 * Note that if a zero_page is amongst the returned pages, it will not have
3572 * pins in it and unpin_user_page() will not remove pins from it.
3573 */
pin_user_pages_fast(unsigned long start,int nr_pages,unsigned int gup_flags,struct page ** pages)3574 int pin_user_pages_fast(unsigned long start, int nr_pages,
3575 unsigned int gup_flags, struct page **pages)
3576 {
3577 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN))
3578 return -EINVAL;
3579 return gup_fast_fallback(start, nr_pages, gup_flags, pages);
3580 }
3581 EXPORT_SYMBOL_GPL(pin_user_pages_fast);
3582
3583 /**
3584 * pin_user_pages_remote() - pin pages of a remote process
3585 *
3586 * @mm: mm_struct of target mm
3587 * @start: starting user address
3588 * @nr_pages: number of pages from start to pin
3589 * @gup_flags: flags modifying lookup behaviour
3590 * @pages: array that receives pointers to the pages pinned.
3591 * Should be at least nr_pages long.
3592 * @locked: pointer to lock flag indicating whether lock is held and
3593 * subsequently whether VM_FAULT_RETRY functionality can be
3594 * utilised. Lock must initially be held.
3595 *
3596 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
3597 * get_user_pages_remote() for documentation on the function arguments, because
3598 * the arguments here are identical.
3599 *
3600 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3601 * see Documentation/core-api/pin_user_pages.rst for details.
3602 *
3603 * Note that if a zero_page is amongst the returned pages, it will not have
3604 * pins in it and unpin_user_page*() will not remove pins from it.
3605 */
pin_user_pages_remote(struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)3606 long pin_user_pages_remote(struct mm_struct *mm,
3607 unsigned long start, unsigned long nr_pages,
3608 unsigned int gup_flags, struct page **pages,
3609 int *locked)
3610 {
3611 int local_locked = 1;
3612
3613 if (!is_valid_gup_args(pages, locked, &gup_flags,
3614 FOLL_PIN | FOLL_TOUCH | FOLL_REMOTE))
3615 return 0;
3616 return __gup_longterm_locked(mm, start, nr_pages, pages,
3617 locked ? locked : &local_locked,
3618 gup_flags);
3619 }
3620 EXPORT_SYMBOL(pin_user_pages_remote);
3621
3622 /**
3623 * pin_user_pages() - pin user pages in memory for use by other devices
3624 *
3625 * @start: starting user address
3626 * @nr_pages: number of pages from start to pin
3627 * @gup_flags: flags modifying lookup behaviour
3628 * @pages: array that receives pointers to the pages pinned.
3629 * Should be at least nr_pages long.
3630 *
3631 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
3632 * FOLL_PIN is set.
3633 *
3634 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3635 * see Documentation/core-api/pin_user_pages.rst for details.
3636 *
3637 * Note that if a zero_page is amongst the returned pages, it will not have
3638 * pins in it and unpin_user_page*() will not remove pins from it.
3639 */
pin_user_pages(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages)3640 long pin_user_pages(unsigned long start, unsigned long nr_pages,
3641 unsigned int gup_flags, struct page **pages)
3642 {
3643 int locked = 1;
3644
3645 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN))
3646 return 0;
3647 return __gup_longterm_locked(current->mm, start, nr_pages,
3648 pages, &locked, gup_flags);
3649 }
3650 EXPORT_SYMBOL(pin_user_pages);
3651
3652 /*
3653 * pin_user_pages_unlocked() is the FOLL_PIN variant of
3654 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
3655 * FOLL_PIN and rejects FOLL_GET.
3656 *
3657 * Note that if a zero_page is amongst the returned pages, it will not have
3658 * pins in it and unpin_user_page*() will not remove pins from it.
3659 */
pin_user_pages_unlocked(unsigned long start,unsigned long nr_pages,struct page ** pages,unsigned int gup_flags)3660 long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
3661 struct page **pages, unsigned int gup_flags)
3662 {
3663 int locked = 0;
3664
3665 if (!is_valid_gup_args(pages, NULL, &gup_flags,
3666 FOLL_PIN | FOLL_TOUCH | FOLL_UNLOCKABLE))
3667 return 0;
3668
3669 return __gup_longterm_locked(current->mm, start, nr_pages, pages,
3670 &locked, gup_flags);
3671 }
3672 EXPORT_SYMBOL(pin_user_pages_unlocked);
3673