Loading...
Note: File does not exist in v3.1.
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
14#include <linux/sched/signal.h>
15#include <linux/rwsem.h>
16#include <linux/hugetlb.h>
17#include <linux/migrate.h>
18#include <linux/mm_inline.h>
19#include <linux/sched/mm.h>
20
21#include <asm/mmu_context.h>
22#include <asm/tlbflush.h>
23
24#include "internal.h"
25
26struct follow_page_context {
27 struct dev_pagemap *pgmap;
28 unsigned int page_mask;
29};
30
31static void hpage_pincount_add(struct page *page, int refs)
32{
33 VM_BUG_ON_PAGE(!hpage_pincount_available(page), page);
34 VM_BUG_ON_PAGE(page != compound_head(page), page);
35
36 atomic_add(refs, compound_pincount_ptr(page));
37}
38
39static void hpage_pincount_sub(struct page *page, int refs)
40{
41 VM_BUG_ON_PAGE(!hpage_pincount_available(page), page);
42 VM_BUG_ON_PAGE(page != compound_head(page), page);
43
44 atomic_sub(refs, compound_pincount_ptr(page));
45}
46
47/*
48 * Return the compound head page with ref appropriately incremented,
49 * or NULL if that failed.
50 */
51static inline struct page *try_get_compound_head(struct page *page, int refs)
52{
53 struct page *head = compound_head(page);
54
55 if (WARN_ON_ONCE(page_ref_count(head) < 0))
56 return NULL;
57 if (unlikely(!page_cache_add_speculative(head, refs)))
58 return NULL;
59 return head;
60}
61
62/*
63 * try_grab_compound_head() - attempt to elevate a page's refcount, by a
64 * flags-dependent amount.
65 *
66 * "grab" names in this file mean, "look at flags to decide whether to use
67 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
68 *
69 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
70 * same time. (That's true throughout the get_user_pages*() and
71 * pin_user_pages*() APIs.) Cases:
72 *
73 * FOLL_GET: page's refcount will be incremented by 1.
74 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
75 *
76 * Return: head page (with refcount appropriately incremented) for success, or
77 * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's
78 * considered failure, and furthermore, a likely bug in the caller, so a warning
79 * is also emitted.
80 */
81static __maybe_unused struct page *try_grab_compound_head(struct page *page,
82 int refs,
83 unsigned int flags)
84{
85 if (flags & FOLL_GET)
86 return try_get_compound_head(page, refs);
87 else if (flags & FOLL_PIN) {
88 int orig_refs = refs;
89
90 /*
91 * Can't do FOLL_LONGTERM + FOLL_PIN with CMA in the gup fast
92 * path, so fail and let the caller fall back to the slow path.
93 */
94 if (unlikely(flags & FOLL_LONGTERM) &&
95 is_migrate_cma_page(page))
96 return NULL;
97
98 /*
99 * When pinning a compound page of order > 1 (which is what
100 * hpage_pincount_available() checks for), use an exact count to
101 * track it, via hpage_pincount_add/_sub().
102 *
103 * However, be sure to *also* increment the normal page refcount
104 * field at least once, so that the page really is pinned.
105 */
106 if (!hpage_pincount_available(page))
107 refs *= GUP_PIN_COUNTING_BIAS;
108
109 page = try_get_compound_head(page, refs);
110 if (!page)
111 return NULL;
112
113 if (hpage_pincount_available(page))
114 hpage_pincount_add(page, refs);
115
116 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED,
117 orig_refs);
118
119 return page;
120 }
121
122 WARN_ON_ONCE(1);
123 return NULL;
124}
125
126/**
127 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
128 *
129 * This might not do anything at all, depending on the flags argument.
130 *
131 * "grab" names in this file mean, "look at flags to decide whether to use
132 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
133 *
134 * @page: pointer to page to be grabbed
135 * @flags: gup flags: these are the FOLL_* flag values.
136 *
137 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
138 * time. Cases:
139 *
140 * FOLL_GET: page's refcount will be incremented by 1.
141 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
142 *
143 * Return: true for success, or if no action was required (if neither FOLL_PIN
144 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
145 * FOLL_PIN was set, but the page could not be grabbed.
146 */
147bool __must_check try_grab_page(struct page *page, unsigned int flags)
148{
149 WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == (FOLL_GET | FOLL_PIN));
150
151 if (flags & FOLL_GET)
152 return try_get_page(page);
153 else if (flags & FOLL_PIN) {
154 int refs = 1;
155
156 page = compound_head(page);
157
158 if (WARN_ON_ONCE(page_ref_count(page) <= 0))
159 return false;
160
161 if (hpage_pincount_available(page))
162 hpage_pincount_add(page, 1);
163 else
164 refs = GUP_PIN_COUNTING_BIAS;
165
166 /*
167 * Similar to try_grab_compound_head(): even if using the
168 * hpage_pincount_add/_sub() routines, be sure to
169 * *also* increment the normal page refcount field at least
170 * once, so that the page really is pinned.
171 */
172 page_ref_add(page, refs);
173
174 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED, 1);
175 }
176
177 return true;
178}
179
180#ifdef CONFIG_DEV_PAGEMAP_OPS
181static bool __unpin_devmap_managed_user_page(struct page *page)
182{
183 int count, refs = 1;
184
185 if (!page_is_devmap_managed(page))
186 return false;
187
188 if (hpage_pincount_available(page))
189 hpage_pincount_sub(page, 1);
190 else
191 refs = GUP_PIN_COUNTING_BIAS;
192
193 count = page_ref_sub_return(page, refs);
194
195 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED, 1);
196 /*
197 * devmap page refcounts are 1-based, rather than 0-based: if
198 * refcount is 1, then the page is free and the refcount is
199 * stable because nobody holds a reference on the page.
200 */
201 if (count == 1)
202 free_devmap_managed_page(page);
203 else if (!count)
204 __put_page(page);
205
206 return true;
207}
208#else
209static bool __unpin_devmap_managed_user_page(struct page *page)
210{
211 return false;
212}
213#endif /* CONFIG_DEV_PAGEMAP_OPS */
214
215/**
216 * unpin_user_page() - release a dma-pinned page
217 * @page: pointer to page to be released
218 *
219 * Pages that were pinned via pin_user_pages*() must be released via either
220 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
221 * that such pages can be separately tracked and uniquely handled. In
222 * particular, interactions with RDMA and filesystems need special handling.
223 */
224void unpin_user_page(struct page *page)
225{
226 int refs = 1;
227
228 page = compound_head(page);
229
230 /*
231 * For devmap managed pages we need to catch refcount transition from
232 * GUP_PIN_COUNTING_BIAS to 1, when refcount reach one it means the
233 * page is free and we need to inform the device driver through
234 * callback. See include/linux/memremap.h and HMM for details.
235 */
236 if (__unpin_devmap_managed_user_page(page))
237 return;
238
239 if (hpage_pincount_available(page))
240 hpage_pincount_sub(page, 1);
241 else
242 refs = GUP_PIN_COUNTING_BIAS;
243
244 if (page_ref_sub_and_test(page, refs))
245 __put_page(page);
246
247 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED, 1);
248}
249EXPORT_SYMBOL(unpin_user_page);
250
251/**
252 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
253 * @pages: array of pages to be maybe marked dirty, and definitely released.
254 * @npages: number of pages in the @pages array.
255 * @make_dirty: whether to mark the pages dirty
256 *
257 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
258 * variants called on that page.
259 *
260 * For each page in the @pages array, make that page (or its head page, if a
261 * compound page) dirty, if @make_dirty is true, and if the page was previously
262 * listed as clean. In any case, releases all pages using unpin_user_page(),
263 * possibly via unpin_user_pages(), for the non-dirty case.
264 *
265 * Please see the unpin_user_page() documentation for details.
266 *
267 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
268 * required, then the caller should a) verify that this is really correct,
269 * because _lock() is usually required, and b) hand code it:
270 * set_page_dirty_lock(), unpin_user_page().
271 *
272 */
273void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
274 bool make_dirty)
275{
276 unsigned long index;
277
278 /*
279 * TODO: this can be optimized for huge pages: if a series of pages is
280 * physically contiguous and part of the same compound page, then a
281 * single operation to the head page should suffice.
282 */
283
284 if (!make_dirty) {
285 unpin_user_pages(pages, npages);
286 return;
287 }
288
289 for (index = 0; index < npages; index++) {
290 struct page *page = compound_head(pages[index]);
291 /*
292 * Checking PageDirty at this point may race with
293 * clear_page_dirty_for_io(), but that's OK. Two key
294 * cases:
295 *
296 * 1) This code sees the page as already dirty, so it
297 * skips the call to set_page_dirty(). That could happen
298 * because clear_page_dirty_for_io() called
299 * page_mkclean(), followed by set_page_dirty().
300 * However, now the page is going to get written back,
301 * which meets the original intention of setting it
302 * dirty, so all is well: clear_page_dirty_for_io() goes
303 * on to call TestClearPageDirty(), and write the page
304 * back.
305 *
306 * 2) This code sees the page as clean, so it calls
307 * set_page_dirty(). The page stays dirty, despite being
308 * written back, so it gets written back again in the
309 * next writeback cycle. This is harmless.
310 */
311 if (!PageDirty(page))
312 set_page_dirty_lock(page);
313 unpin_user_page(page);
314 }
315}
316EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
317
318/**
319 * unpin_user_pages() - release an array of gup-pinned pages.
320 * @pages: array of pages to be marked dirty and released.
321 * @npages: number of pages in the @pages array.
322 *
323 * For each page in the @pages array, release the page using unpin_user_page().
324 *
325 * Please see the unpin_user_page() documentation for details.
326 */
327void unpin_user_pages(struct page **pages, unsigned long npages)
328{
329 unsigned long index;
330
331 /*
332 * TODO: this can be optimized for huge pages: if a series of pages is
333 * physically contiguous and part of the same compound page, then a
334 * single operation to the head page should suffice.
335 */
336 for (index = 0; index < npages; index++)
337 unpin_user_page(pages[index]);
338}
339EXPORT_SYMBOL(unpin_user_pages);
340
341#ifdef CONFIG_MMU
342static struct page *no_page_table(struct vm_area_struct *vma,
343 unsigned int flags)
344{
345 /*
346 * When core dumping an enormous anonymous area that nobody
347 * has touched so far, we don't want to allocate unnecessary pages or
348 * page tables. Return error instead of NULL to skip handle_mm_fault,
349 * then get_dump_page() will return NULL to leave a hole in the dump.
350 * But we can only make this optimization where a hole would surely
351 * be zero-filled if handle_mm_fault() actually did handle it.
352 */
353 if ((flags & FOLL_DUMP) &&
354 (vma_is_anonymous(vma) || !vma->vm_ops->fault))
355 return ERR_PTR(-EFAULT);
356 return NULL;
357}
358
359static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
360 pte_t *pte, unsigned int flags)
361{
362 /* No page to get reference */
363 if (flags & FOLL_GET)
364 return -EFAULT;
365
366 if (flags & FOLL_TOUCH) {
367 pte_t entry = *pte;
368
369 if (flags & FOLL_WRITE)
370 entry = pte_mkdirty(entry);
371 entry = pte_mkyoung(entry);
372
373 if (!pte_same(*pte, entry)) {
374 set_pte_at(vma->vm_mm, address, pte, entry);
375 update_mmu_cache(vma, address, pte);
376 }
377 }
378
379 /* Proper page table entry exists, but no corresponding struct page */
380 return -EEXIST;
381}
382
383/*
384 * FOLL_FORCE can write to even unwritable pte's, but only
385 * after we've gone through a COW cycle and they are dirty.
386 */
387static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
388{
389 return pte_write(pte) ||
390 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
391}
392
393static struct page *follow_page_pte(struct vm_area_struct *vma,
394 unsigned long address, pmd_t *pmd, unsigned int flags,
395 struct dev_pagemap **pgmap)
396{
397 struct mm_struct *mm = vma->vm_mm;
398 struct page *page;
399 spinlock_t *ptl;
400 pte_t *ptep, pte;
401 int ret;
402
403 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
404 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
405 (FOLL_PIN | FOLL_GET)))
406 return ERR_PTR(-EINVAL);
407retry:
408 if (unlikely(pmd_bad(*pmd)))
409 return no_page_table(vma, flags);
410
411 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
412 pte = *ptep;
413 if (!pte_present(pte)) {
414 swp_entry_t entry;
415 /*
416 * KSM's break_ksm() relies upon recognizing a ksm page
417 * even while it is being migrated, so for that case we
418 * need migration_entry_wait().
419 */
420 if (likely(!(flags & FOLL_MIGRATION)))
421 goto no_page;
422 if (pte_none(pte))
423 goto no_page;
424 entry = pte_to_swp_entry(pte);
425 if (!is_migration_entry(entry))
426 goto no_page;
427 pte_unmap_unlock(ptep, ptl);
428 migration_entry_wait(mm, pmd, address);
429 goto retry;
430 }
431 if ((flags & FOLL_NUMA) && pte_protnone(pte))
432 goto no_page;
433 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
434 pte_unmap_unlock(ptep, ptl);
435 return NULL;
436 }
437
438 page = vm_normal_page(vma, address, pte);
439 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
440 /*
441 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
442 * case since they are only valid while holding the pgmap
443 * reference.
444 */
445 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
446 if (*pgmap)
447 page = pte_page(pte);
448 else
449 goto no_page;
450 } else if (unlikely(!page)) {
451 if (flags & FOLL_DUMP) {
452 /* Avoid special (like zero) pages in core dumps */
453 page = ERR_PTR(-EFAULT);
454 goto out;
455 }
456
457 if (is_zero_pfn(pte_pfn(pte))) {
458 page = pte_page(pte);
459 } else {
460 ret = follow_pfn_pte(vma, address, ptep, flags);
461 page = ERR_PTR(ret);
462 goto out;
463 }
464 }
465
466 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
467 get_page(page);
468 pte_unmap_unlock(ptep, ptl);
469 lock_page(page);
470 ret = split_huge_page(page);
471 unlock_page(page);
472 put_page(page);
473 if (ret)
474 return ERR_PTR(ret);
475 goto retry;
476 }
477
478 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
479 if (unlikely(!try_grab_page(page, flags))) {
480 page = ERR_PTR(-ENOMEM);
481 goto out;
482 }
483 /*
484 * We need to make the page accessible if and only if we are going
485 * to access its content (the FOLL_PIN case). Please see
486 * Documentation/core-api/pin_user_pages.rst for details.
487 */
488 if (flags & FOLL_PIN) {
489 ret = arch_make_page_accessible(page);
490 if (ret) {
491 unpin_user_page(page);
492 page = ERR_PTR(ret);
493 goto out;
494 }
495 }
496 if (flags & FOLL_TOUCH) {
497 if ((flags & FOLL_WRITE) &&
498 !pte_dirty(pte) && !PageDirty(page))
499 set_page_dirty(page);
500 /*
501 * pte_mkyoung() would be more correct here, but atomic care
502 * is needed to avoid losing the dirty bit: it is easier to use
503 * mark_page_accessed().
504 */
505 mark_page_accessed(page);
506 }
507 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
508 /* Do not mlock pte-mapped THP */
509 if (PageTransCompound(page))
510 goto out;
511
512 /*
513 * The preliminary mapping check is mainly to avoid the
514 * pointless overhead of lock_page on the ZERO_PAGE
515 * which might bounce very badly if there is contention.
516 *
517 * If the page is already locked, we don't need to
518 * handle it now - vmscan will handle it later if and
519 * when it attempts to reclaim the page.
520 */
521 if (page->mapping && trylock_page(page)) {
522 lru_add_drain(); /* push cached pages to LRU */
523 /*
524 * Because we lock page here, and migration is
525 * blocked by the pte's page reference, and we
526 * know the page is still mapped, we don't even
527 * need to check for file-cache page truncation.
528 */
529 mlock_vma_page(page);
530 unlock_page(page);
531 }
532 }
533out:
534 pte_unmap_unlock(ptep, ptl);
535 return page;
536no_page:
537 pte_unmap_unlock(ptep, ptl);
538 if (!pte_none(pte))
539 return NULL;
540 return no_page_table(vma, flags);
541}
542
543static struct page *follow_pmd_mask(struct vm_area_struct *vma,
544 unsigned long address, pud_t *pudp,
545 unsigned int flags,
546 struct follow_page_context *ctx)
547{
548 pmd_t *pmd, pmdval;
549 spinlock_t *ptl;
550 struct page *page;
551 struct mm_struct *mm = vma->vm_mm;
552
553 pmd = pmd_offset(pudp, address);
554 /*
555 * The READ_ONCE() will stabilize the pmdval in a register or
556 * on the stack so that it will stop changing under the code.
557 */
558 pmdval = READ_ONCE(*pmd);
559 if (pmd_none(pmdval))
560 return no_page_table(vma, flags);
561 if (pmd_huge(pmdval) && is_vm_hugetlb_page(vma)) {
562 page = follow_huge_pmd(mm, address, pmd, flags);
563 if (page)
564 return page;
565 return no_page_table(vma, flags);
566 }
567 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
568 page = follow_huge_pd(vma, address,
569 __hugepd(pmd_val(pmdval)), flags,
570 PMD_SHIFT);
571 if (page)
572 return page;
573 return no_page_table(vma, flags);
574 }
575retry:
576 if (!pmd_present(pmdval)) {
577 if (likely(!(flags & FOLL_MIGRATION)))
578 return no_page_table(vma, flags);
579 VM_BUG_ON(thp_migration_supported() &&
580 !is_pmd_migration_entry(pmdval));
581 if (is_pmd_migration_entry(pmdval))
582 pmd_migration_entry_wait(mm, pmd);
583 pmdval = READ_ONCE(*pmd);
584 /*
585 * MADV_DONTNEED may convert the pmd to null because
586 * mmap_lock is held in read mode
587 */
588 if (pmd_none(pmdval))
589 return no_page_table(vma, flags);
590 goto retry;
591 }
592 if (pmd_devmap(pmdval)) {
593 ptl = pmd_lock(mm, pmd);
594 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
595 spin_unlock(ptl);
596 if (page)
597 return page;
598 }
599 if (likely(!pmd_trans_huge(pmdval)))
600 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
601
602 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
603 return no_page_table(vma, flags);
604
605retry_locked:
606 ptl = pmd_lock(mm, pmd);
607 if (unlikely(pmd_none(*pmd))) {
608 spin_unlock(ptl);
609 return no_page_table(vma, flags);
610 }
611 if (unlikely(!pmd_present(*pmd))) {
612 spin_unlock(ptl);
613 if (likely(!(flags & FOLL_MIGRATION)))
614 return no_page_table(vma, flags);
615 pmd_migration_entry_wait(mm, pmd);
616 goto retry_locked;
617 }
618 if (unlikely(!pmd_trans_huge(*pmd))) {
619 spin_unlock(ptl);
620 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
621 }
622 if (flags & (FOLL_SPLIT | FOLL_SPLIT_PMD)) {
623 int ret;
624 page = pmd_page(*pmd);
625 if (is_huge_zero_page(page)) {
626 spin_unlock(ptl);
627 ret = 0;
628 split_huge_pmd(vma, pmd, address);
629 if (pmd_trans_unstable(pmd))
630 ret = -EBUSY;
631 } else if (flags & FOLL_SPLIT) {
632 if (unlikely(!try_get_page(page))) {
633 spin_unlock(ptl);
634 return ERR_PTR(-ENOMEM);
635 }
636 spin_unlock(ptl);
637 lock_page(page);
638 ret = split_huge_page(page);
639 unlock_page(page);
640 put_page(page);
641 if (pmd_none(*pmd))
642 return no_page_table(vma, flags);
643 } else { /* flags & FOLL_SPLIT_PMD */
644 spin_unlock(ptl);
645 split_huge_pmd(vma, pmd, address);
646 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
647 }
648
649 return ret ? ERR_PTR(ret) :
650 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
651 }
652 page = follow_trans_huge_pmd(vma, address, pmd, flags);
653 spin_unlock(ptl);
654 ctx->page_mask = HPAGE_PMD_NR - 1;
655 return page;
656}
657
658static struct page *follow_pud_mask(struct vm_area_struct *vma,
659 unsigned long address, p4d_t *p4dp,
660 unsigned int flags,
661 struct follow_page_context *ctx)
662{
663 pud_t *pud;
664 spinlock_t *ptl;
665 struct page *page;
666 struct mm_struct *mm = vma->vm_mm;
667
668 pud = pud_offset(p4dp, address);
669 if (pud_none(*pud))
670 return no_page_table(vma, flags);
671 if (pud_huge(*pud) && is_vm_hugetlb_page(vma)) {
672 page = follow_huge_pud(mm, address, pud, flags);
673 if (page)
674 return page;
675 return no_page_table(vma, flags);
676 }
677 if (is_hugepd(__hugepd(pud_val(*pud)))) {
678 page = follow_huge_pd(vma, address,
679 __hugepd(pud_val(*pud)), flags,
680 PUD_SHIFT);
681 if (page)
682 return page;
683 return no_page_table(vma, flags);
684 }
685 if (pud_devmap(*pud)) {
686 ptl = pud_lock(mm, pud);
687 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
688 spin_unlock(ptl);
689 if (page)
690 return page;
691 }
692 if (unlikely(pud_bad(*pud)))
693 return no_page_table(vma, flags);
694
695 return follow_pmd_mask(vma, address, pud, flags, ctx);
696}
697
698static struct page *follow_p4d_mask(struct vm_area_struct *vma,
699 unsigned long address, pgd_t *pgdp,
700 unsigned int flags,
701 struct follow_page_context *ctx)
702{
703 p4d_t *p4d;
704 struct page *page;
705
706 p4d = p4d_offset(pgdp, address);
707 if (p4d_none(*p4d))
708 return no_page_table(vma, flags);
709 BUILD_BUG_ON(p4d_huge(*p4d));
710 if (unlikely(p4d_bad(*p4d)))
711 return no_page_table(vma, flags);
712
713 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
714 page = follow_huge_pd(vma, address,
715 __hugepd(p4d_val(*p4d)), flags,
716 P4D_SHIFT);
717 if (page)
718 return page;
719 return no_page_table(vma, flags);
720 }
721 return follow_pud_mask(vma, address, p4d, flags, ctx);
722}
723
724/**
725 * follow_page_mask - look up a page descriptor from a user-virtual address
726 * @vma: vm_area_struct mapping @address
727 * @address: virtual address to look up
728 * @flags: flags modifying lookup behaviour
729 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
730 * pointer to output page_mask
731 *
732 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
733 *
734 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
735 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
736 *
737 * On output, the @ctx->page_mask is set according to the size of the page.
738 *
739 * Return: the mapped (struct page *), %NULL if no mapping exists, or
740 * an error pointer if there is a mapping to something not represented
741 * by a page descriptor (see also vm_normal_page()).
742 */
743static struct page *follow_page_mask(struct vm_area_struct *vma,
744 unsigned long address, unsigned int flags,
745 struct follow_page_context *ctx)
746{
747 pgd_t *pgd;
748 struct page *page;
749 struct mm_struct *mm = vma->vm_mm;
750
751 ctx->page_mask = 0;
752
753 /* make this handle hugepd */
754 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
755 if (!IS_ERR(page)) {
756 WARN_ON_ONCE(flags & (FOLL_GET | FOLL_PIN));
757 return page;
758 }
759
760 pgd = pgd_offset(mm, address);
761
762 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
763 return no_page_table(vma, flags);
764
765 if (pgd_huge(*pgd)) {
766 page = follow_huge_pgd(mm, address, pgd, flags);
767 if (page)
768 return page;
769 return no_page_table(vma, flags);
770 }
771 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
772 page = follow_huge_pd(vma, address,
773 __hugepd(pgd_val(*pgd)), flags,
774 PGDIR_SHIFT);
775 if (page)
776 return page;
777 return no_page_table(vma, flags);
778 }
779
780 return follow_p4d_mask(vma, address, pgd, flags, ctx);
781}
782
783struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
784 unsigned int foll_flags)
785{
786 struct follow_page_context ctx = { NULL };
787 struct page *page;
788
789 page = follow_page_mask(vma, address, foll_flags, &ctx);
790 if (ctx.pgmap)
791 put_dev_pagemap(ctx.pgmap);
792 return page;
793}
794
795static int get_gate_page(struct mm_struct *mm, unsigned long address,
796 unsigned int gup_flags, struct vm_area_struct **vma,
797 struct page **page)
798{
799 pgd_t *pgd;
800 p4d_t *p4d;
801 pud_t *pud;
802 pmd_t *pmd;
803 pte_t *pte;
804 int ret = -EFAULT;
805
806 /* user gate pages are read-only */
807 if (gup_flags & FOLL_WRITE)
808 return -EFAULT;
809 if (address > TASK_SIZE)
810 pgd = pgd_offset_k(address);
811 else
812 pgd = pgd_offset_gate(mm, address);
813 if (pgd_none(*pgd))
814 return -EFAULT;
815 p4d = p4d_offset(pgd, address);
816 if (p4d_none(*p4d))
817 return -EFAULT;
818 pud = pud_offset(p4d, address);
819 if (pud_none(*pud))
820 return -EFAULT;
821 pmd = pmd_offset(pud, address);
822 if (!pmd_present(*pmd))
823 return -EFAULT;
824 VM_BUG_ON(pmd_trans_huge(*pmd));
825 pte = pte_offset_map(pmd, address);
826 if (pte_none(*pte))
827 goto unmap;
828 *vma = get_gate_vma(mm);
829 if (!page)
830 goto out;
831 *page = vm_normal_page(*vma, address, *pte);
832 if (!*page) {
833 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
834 goto unmap;
835 *page = pte_page(*pte);
836 }
837 if (unlikely(!try_grab_page(*page, gup_flags))) {
838 ret = -ENOMEM;
839 goto unmap;
840 }
841out:
842 ret = 0;
843unmap:
844 pte_unmap(pte);
845 return ret;
846}
847
848/*
849 * mmap_lock must be held on entry. If @locked != NULL and *@flags
850 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
851 * is, *@locked will be set to 0 and -EBUSY returned.
852 */
853static int faultin_page(struct vm_area_struct *vma,
854 unsigned long address, unsigned int *flags, int *locked)
855{
856 unsigned int fault_flags = 0;
857 vm_fault_t ret;
858
859 /* mlock all present pages, but do not fault in new pages */
860 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
861 return -ENOENT;
862 if (*flags & FOLL_WRITE)
863 fault_flags |= FAULT_FLAG_WRITE;
864 if (*flags & FOLL_REMOTE)
865 fault_flags |= FAULT_FLAG_REMOTE;
866 if (locked)
867 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
868 if (*flags & FOLL_NOWAIT)
869 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
870 if (*flags & FOLL_TRIED) {
871 /*
872 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
873 * can co-exist
874 */
875 fault_flags |= FAULT_FLAG_TRIED;
876 }
877
878 ret = handle_mm_fault(vma, address, fault_flags, NULL);
879 if (ret & VM_FAULT_ERROR) {
880 int err = vm_fault_to_errno(ret, *flags);
881
882 if (err)
883 return err;
884 BUG();
885 }
886
887 if (ret & VM_FAULT_RETRY) {
888 if (locked && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
889 *locked = 0;
890 return -EBUSY;
891 }
892
893 /*
894 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
895 * necessary, even if maybe_mkwrite decided not to set pte_write. We
896 * can thus safely do subsequent page lookups as if they were reads.
897 * But only do so when looping for pte_write is futile: in some cases
898 * userspace may also be wanting to write to the gotten user page,
899 * which a read fault here might prevent (a readonly page might get
900 * reCOWed by userspace write).
901 */
902 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
903 *flags |= FOLL_COW;
904 return 0;
905}
906
907static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
908{
909 vm_flags_t vm_flags = vma->vm_flags;
910 int write = (gup_flags & FOLL_WRITE);
911 int foreign = (gup_flags & FOLL_REMOTE);
912
913 if (vm_flags & (VM_IO | VM_PFNMAP))
914 return -EFAULT;
915
916 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
917 return -EFAULT;
918
919 if (write) {
920 if (!(vm_flags & VM_WRITE)) {
921 if (!(gup_flags & FOLL_FORCE))
922 return -EFAULT;
923 /*
924 * We used to let the write,force case do COW in a
925 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
926 * set a breakpoint in a read-only mapping of an
927 * executable, without corrupting the file (yet only
928 * when that file had been opened for writing!).
929 * Anon pages in shared mappings are surprising: now
930 * just reject it.
931 */
932 if (!is_cow_mapping(vm_flags))
933 return -EFAULT;
934 }
935 } else if (!(vm_flags & VM_READ)) {
936 if (!(gup_flags & FOLL_FORCE))
937 return -EFAULT;
938 /*
939 * Is there actually any vma we can reach here which does not
940 * have VM_MAYREAD set?
941 */
942 if (!(vm_flags & VM_MAYREAD))
943 return -EFAULT;
944 }
945 /*
946 * gups are always data accesses, not instruction
947 * fetches, so execute=false here
948 */
949 if (!arch_vma_access_permitted(vma, write, false, foreign))
950 return -EFAULT;
951 return 0;
952}
953
954/**
955 * __get_user_pages() - pin user pages in memory
956 * @mm: mm_struct of target mm
957 * @start: starting user address
958 * @nr_pages: number of pages from start to pin
959 * @gup_flags: flags modifying pin behaviour
960 * @pages: array that receives pointers to the pages pinned.
961 * Should be at least nr_pages long. Or NULL, if caller
962 * only intends to ensure the pages are faulted in.
963 * @vmas: array of pointers to vmas corresponding to each page.
964 * Or NULL if the caller does not require them.
965 * @locked: whether we're still with the mmap_lock held
966 *
967 * Returns either number of pages pinned (which may be less than the
968 * number requested), or an error. Details about the return value:
969 *
970 * -- If nr_pages is 0, returns 0.
971 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
972 * -- If nr_pages is >0, and some pages were pinned, returns the number of
973 * pages pinned. Again, this may be less than nr_pages.
974 * -- 0 return value is possible when the fault would need to be retried.
975 *
976 * The caller is responsible for releasing returned @pages, via put_page().
977 *
978 * @vmas are valid only as long as mmap_lock is held.
979 *
980 * Must be called with mmap_lock held. It may be released. See below.
981 *
982 * __get_user_pages walks a process's page tables and takes a reference to
983 * each struct page that each user address corresponds to at a given
984 * instant. That is, it takes the page that would be accessed if a user
985 * thread accesses the given user virtual address at that instant.
986 *
987 * This does not guarantee that the page exists in the user mappings when
988 * __get_user_pages returns, and there may even be a completely different
989 * page there in some cases (eg. if mmapped pagecache has been invalidated
990 * and subsequently re faulted). However it does guarantee that the page
991 * won't be freed completely. And mostly callers simply care that the page
992 * contains data that was valid *at some point in time*. Typically, an IO
993 * or similar operation cannot guarantee anything stronger anyway because
994 * locks can't be held over the syscall boundary.
995 *
996 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
997 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
998 * appropriate) must be called after the page is finished with, and
999 * before put_page is called.
1000 *
1001 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
1002 * released by an up_read(). That can happen if @gup_flags does not
1003 * have FOLL_NOWAIT.
1004 *
1005 * A caller using such a combination of @locked and @gup_flags
1006 * must therefore hold the mmap_lock for reading only, and recognize
1007 * when it's been released. Otherwise, it must be held for either
1008 * reading or writing and will not be released.
1009 *
1010 * In most cases, get_user_pages or get_user_pages_fast should be used
1011 * instead of __get_user_pages. __get_user_pages should be used only if
1012 * you need some special @gup_flags.
1013 */
1014static long __get_user_pages(struct mm_struct *mm,
1015 unsigned long start, unsigned long nr_pages,
1016 unsigned int gup_flags, struct page **pages,
1017 struct vm_area_struct **vmas, int *locked)
1018{
1019 long ret = 0, i = 0;
1020 struct vm_area_struct *vma = NULL;
1021 struct follow_page_context ctx = { NULL };
1022
1023 if (!nr_pages)
1024 return 0;
1025
1026 start = untagged_addr(start);
1027
1028 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1029
1030 /*
1031 * If FOLL_FORCE is set then do not force a full fault as the hinting
1032 * fault information is unrelated to the reference behaviour of a task
1033 * using the address space
1034 */
1035 if (!(gup_flags & FOLL_FORCE))
1036 gup_flags |= FOLL_NUMA;
1037
1038 do {
1039 struct page *page;
1040 unsigned int foll_flags = gup_flags;
1041 unsigned int page_increm;
1042
1043 /* first iteration or cross vma bound */
1044 if (!vma || start >= vma->vm_end) {
1045 vma = find_extend_vma(mm, start);
1046 if (!vma && in_gate_area(mm, start)) {
1047 ret = get_gate_page(mm, start & PAGE_MASK,
1048 gup_flags, &vma,
1049 pages ? &pages[i] : NULL);
1050 if (ret)
1051 goto out;
1052 ctx.page_mask = 0;
1053 goto next_page;
1054 }
1055
1056 if (!vma || check_vma_flags(vma, gup_flags)) {
1057 ret = -EFAULT;
1058 goto out;
1059 }
1060 if (is_vm_hugetlb_page(vma)) {
1061 i = follow_hugetlb_page(mm, vma, pages, vmas,
1062 &start, &nr_pages, i,
1063 gup_flags, locked);
1064 if (locked && *locked == 0) {
1065 /*
1066 * We've got a VM_FAULT_RETRY
1067 * and we've lost mmap_lock.
1068 * We must stop here.
1069 */
1070 BUG_ON(gup_flags & FOLL_NOWAIT);
1071 BUG_ON(ret != 0);
1072 goto out;
1073 }
1074 continue;
1075 }
1076 }
1077retry:
1078 /*
1079 * If we have a pending SIGKILL, don't keep faulting pages and
1080 * potentially allocating memory.
1081 */
1082 if (fatal_signal_pending(current)) {
1083 ret = -EINTR;
1084 goto out;
1085 }
1086 cond_resched();
1087
1088 page = follow_page_mask(vma, start, foll_flags, &ctx);
1089 if (!page) {
1090 ret = faultin_page(vma, start, &foll_flags, locked);
1091 switch (ret) {
1092 case 0:
1093 goto retry;
1094 case -EBUSY:
1095 ret = 0;
1096 fallthrough;
1097 case -EFAULT:
1098 case -ENOMEM:
1099 case -EHWPOISON:
1100 goto out;
1101 case -ENOENT:
1102 goto next_page;
1103 }
1104 BUG();
1105 } else if (PTR_ERR(page) == -EEXIST) {
1106 /*
1107 * Proper page table entry exists, but no corresponding
1108 * struct page.
1109 */
1110 goto next_page;
1111 } else if (IS_ERR(page)) {
1112 ret = PTR_ERR(page);
1113 goto out;
1114 }
1115 if (pages) {
1116 pages[i] = page;
1117 flush_anon_page(vma, page, start);
1118 flush_dcache_page(page);
1119 ctx.page_mask = 0;
1120 }
1121next_page:
1122 if (vmas) {
1123 vmas[i] = vma;
1124 ctx.page_mask = 0;
1125 }
1126 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1127 if (page_increm > nr_pages)
1128 page_increm = nr_pages;
1129 i += page_increm;
1130 start += page_increm * PAGE_SIZE;
1131 nr_pages -= page_increm;
1132 } while (nr_pages);
1133out:
1134 if (ctx.pgmap)
1135 put_dev_pagemap(ctx.pgmap);
1136 return i ? i : ret;
1137}
1138
1139static bool vma_permits_fault(struct vm_area_struct *vma,
1140 unsigned int fault_flags)
1141{
1142 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1143 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1144 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1145
1146 if (!(vm_flags & vma->vm_flags))
1147 return false;
1148
1149 /*
1150 * The architecture might have a hardware protection
1151 * mechanism other than read/write that can deny access.
1152 *
1153 * gup always represents data access, not instruction
1154 * fetches, so execute=false here:
1155 */
1156 if (!arch_vma_access_permitted(vma, write, false, foreign))
1157 return false;
1158
1159 return true;
1160}
1161
1162/**
1163 * fixup_user_fault() - manually resolve a user page fault
1164 * @mm: mm_struct of target mm
1165 * @address: user address
1166 * @fault_flags:flags to pass down to handle_mm_fault()
1167 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1168 * does not allow retry. If NULL, the caller must guarantee
1169 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1170 *
1171 * This is meant to be called in the specific scenario where for locking reasons
1172 * we try to access user memory in atomic context (within a pagefault_disable()
1173 * section), this returns -EFAULT, and we want to resolve the user fault before
1174 * trying again.
1175 *
1176 * Typically this is meant to be used by the futex code.
1177 *
1178 * The main difference with get_user_pages() is that this function will
1179 * unconditionally call handle_mm_fault() which will in turn perform all the
1180 * necessary SW fixup of the dirty and young bits in the PTE, while
1181 * get_user_pages() only guarantees to update these in the struct page.
1182 *
1183 * This is important for some architectures where those bits also gate the
1184 * access permission to the page because they are maintained in software. On
1185 * such architectures, gup() will not be enough to make a subsequent access
1186 * succeed.
1187 *
1188 * This function will not return with an unlocked mmap_lock. So it has not the
1189 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1190 */
1191int fixup_user_fault(struct mm_struct *mm,
1192 unsigned long address, unsigned int fault_flags,
1193 bool *unlocked)
1194{
1195 struct vm_area_struct *vma;
1196 vm_fault_t ret, major = 0;
1197
1198 address = untagged_addr(address);
1199
1200 if (unlocked)
1201 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1202
1203retry:
1204 vma = find_extend_vma(mm, address);
1205 if (!vma || address < vma->vm_start)
1206 return -EFAULT;
1207
1208 if (!vma_permits_fault(vma, fault_flags))
1209 return -EFAULT;
1210
1211 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1212 fatal_signal_pending(current))
1213 return -EINTR;
1214
1215 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1216 major |= ret & VM_FAULT_MAJOR;
1217 if (ret & VM_FAULT_ERROR) {
1218 int err = vm_fault_to_errno(ret, 0);
1219
1220 if (err)
1221 return err;
1222 BUG();
1223 }
1224
1225 if (ret & VM_FAULT_RETRY) {
1226 mmap_read_lock(mm);
1227 *unlocked = true;
1228 fault_flags |= FAULT_FLAG_TRIED;
1229 goto retry;
1230 }
1231
1232 return 0;
1233}
1234EXPORT_SYMBOL_GPL(fixup_user_fault);
1235
1236/*
1237 * Please note that this function, unlike __get_user_pages will not
1238 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1239 */
1240static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1241 unsigned long start,
1242 unsigned long nr_pages,
1243 struct page **pages,
1244 struct vm_area_struct **vmas,
1245 int *locked,
1246 unsigned int flags)
1247{
1248 long ret, pages_done;
1249 bool lock_dropped;
1250
1251 if (locked) {
1252 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1253 BUG_ON(vmas);
1254 /* check caller initialized locked */
1255 BUG_ON(*locked != 1);
1256 }
1257
1258 if (flags & FOLL_PIN)
1259 atomic_set(&mm->has_pinned, 1);
1260
1261 /*
1262 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1263 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1264 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1265 * for FOLL_GET, not for the newer FOLL_PIN.
1266 *
1267 * FOLL_PIN always expects pages to be non-null, but no need to assert
1268 * that here, as any failures will be obvious enough.
1269 */
1270 if (pages && !(flags & FOLL_PIN))
1271 flags |= FOLL_GET;
1272
1273 pages_done = 0;
1274 lock_dropped = false;
1275 for (;;) {
1276 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1277 vmas, locked);
1278 if (!locked)
1279 /* VM_FAULT_RETRY couldn't trigger, bypass */
1280 return ret;
1281
1282 /* VM_FAULT_RETRY cannot return errors */
1283 if (!*locked) {
1284 BUG_ON(ret < 0);
1285 BUG_ON(ret >= nr_pages);
1286 }
1287
1288 if (ret > 0) {
1289 nr_pages -= ret;
1290 pages_done += ret;
1291 if (!nr_pages)
1292 break;
1293 }
1294 if (*locked) {
1295 /*
1296 * VM_FAULT_RETRY didn't trigger or it was a
1297 * FOLL_NOWAIT.
1298 */
1299 if (!pages_done)
1300 pages_done = ret;
1301 break;
1302 }
1303 /*
1304 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1305 * For the prefault case (!pages) we only update counts.
1306 */
1307 if (likely(pages))
1308 pages += ret;
1309 start += ret << PAGE_SHIFT;
1310 lock_dropped = true;
1311
1312retry:
1313 /*
1314 * Repeat on the address that fired VM_FAULT_RETRY
1315 * with both FAULT_FLAG_ALLOW_RETRY and
1316 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1317 * by fatal signals, so we need to check it before we
1318 * start trying again otherwise it can loop forever.
1319 */
1320
1321 if (fatal_signal_pending(current)) {
1322 if (!pages_done)
1323 pages_done = -EINTR;
1324 break;
1325 }
1326
1327 ret = mmap_read_lock_killable(mm);
1328 if (ret) {
1329 BUG_ON(ret > 0);
1330 if (!pages_done)
1331 pages_done = ret;
1332 break;
1333 }
1334
1335 *locked = 1;
1336 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1337 pages, NULL, locked);
1338 if (!*locked) {
1339 /* Continue to retry until we succeeded */
1340 BUG_ON(ret != 0);
1341 goto retry;
1342 }
1343 if (ret != 1) {
1344 BUG_ON(ret > 1);
1345 if (!pages_done)
1346 pages_done = ret;
1347 break;
1348 }
1349 nr_pages--;
1350 pages_done++;
1351 if (!nr_pages)
1352 break;
1353 if (likely(pages))
1354 pages++;
1355 start += PAGE_SIZE;
1356 }
1357 if (lock_dropped && *locked) {
1358 /*
1359 * We must let the caller know we temporarily dropped the lock
1360 * and so the critical section protected by it was lost.
1361 */
1362 mmap_read_unlock(mm);
1363 *locked = 0;
1364 }
1365 return pages_done;
1366}
1367
1368/**
1369 * populate_vma_page_range() - populate a range of pages in the vma.
1370 * @vma: target vma
1371 * @start: start address
1372 * @end: end address
1373 * @locked: whether the mmap_lock is still held
1374 *
1375 * This takes care of mlocking the pages too if VM_LOCKED is set.
1376 *
1377 * Return either number of pages pinned in the vma, or a negative error
1378 * code on error.
1379 *
1380 * vma->vm_mm->mmap_lock must be held.
1381 *
1382 * If @locked is NULL, it may be held for read or write and will
1383 * be unperturbed.
1384 *
1385 * If @locked is non-NULL, it must held for read only and may be
1386 * released. If it's released, *@locked will be set to 0.
1387 */
1388long populate_vma_page_range(struct vm_area_struct *vma,
1389 unsigned long start, unsigned long end, int *locked)
1390{
1391 struct mm_struct *mm = vma->vm_mm;
1392 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1393 int gup_flags;
1394
1395 VM_BUG_ON(start & ~PAGE_MASK);
1396 VM_BUG_ON(end & ~PAGE_MASK);
1397 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1398 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1399 mmap_assert_locked(mm);
1400
1401 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1402 if (vma->vm_flags & VM_LOCKONFAULT)
1403 gup_flags &= ~FOLL_POPULATE;
1404 /*
1405 * We want to touch writable mappings with a write fault in order
1406 * to break COW, except for shared mappings because these don't COW
1407 * and we would not want to dirty them for nothing.
1408 */
1409 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1410 gup_flags |= FOLL_WRITE;
1411
1412 /*
1413 * We want mlock to succeed for regions that have any permissions
1414 * other than PROT_NONE.
1415 */
1416 if (vma_is_accessible(vma))
1417 gup_flags |= FOLL_FORCE;
1418
1419 /*
1420 * We made sure addr is within a VMA, so the following will
1421 * not result in a stack expansion that recurses back here.
1422 */
1423 return __get_user_pages(mm, start, nr_pages, gup_flags,
1424 NULL, NULL, locked);
1425}
1426
1427/*
1428 * __mm_populate - populate and/or mlock pages within a range of address space.
1429 *
1430 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1431 * flags. VMAs must be already marked with the desired vm_flags, and
1432 * mmap_lock must not be held.
1433 */
1434int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1435{
1436 struct mm_struct *mm = current->mm;
1437 unsigned long end, nstart, nend;
1438 struct vm_area_struct *vma = NULL;
1439 int locked = 0;
1440 long ret = 0;
1441
1442 end = start + len;
1443
1444 for (nstart = start; nstart < end; nstart = nend) {
1445 /*
1446 * We want to fault in pages for [nstart; end) address range.
1447 * Find first corresponding VMA.
1448 */
1449 if (!locked) {
1450 locked = 1;
1451 mmap_read_lock(mm);
1452 vma = find_vma(mm, nstart);
1453 } else if (nstart >= vma->vm_end)
1454 vma = vma->vm_next;
1455 if (!vma || vma->vm_start >= end)
1456 break;
1457 /*
1458 * Set [nstart; nend) to intersection of desired address
1459 * range with the first VMA. Also, skip undesirable VMA types.
1460 */
1461 nend = min(end, vma->vm_end);
1462 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1463 continue;
1464 if (nstart < vma->vm_start)
1465 nstart = vma->vm_start;
1466 /*
1467 * Now fault in a range of pages. populate_vma_page_range()
1468 * double checks the vma flags, so that it won't mlock pages
1469 * if the vma was already munlocked.
1470 */
1471 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1472 if (ret < 0) {
1473 if (ignore_errors) {
1474 ret = 0;
1475 continue; /* continue at next VMA */
1476 }
1477 break;
1478 }
1479 nend = nstart + ret * PAGE_SIZE;
1480 ret = 0;
1481 }
1482 if (locked)
1483 mmap_read_unlock(mm);
1484 return ret; /* 0 or negative error code */
1485}
1486
1487/**
1488 * get_dump_page() - pin user page in memory while writing it to core dump
1489 * @addr: user address
1490 *
1491 * Returns struct page pointer of user page pinned for dump,
1492 * to be freed afterwards by put_page().
1493 *
1494 * Returns NULL on any kind of failure - a hole must then be inserted into
1495 * the corefile, to preserve alignment with its headers; and also returns
1496 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1497 * allowing a hole to be left in the corefile to save diskspace.
1498 *
1499 * Called without mmap_lock, but after all other threads have been killed.
1500 */
1501#ifdef CONFIG_ELF_CORE
1502struct page *get_dump_page(unsigned long addr)
1503{
1504 struct vm_area_struct *vma;
1505 struct page *page;
1506
1507 if (__get_user_pages(current->mm, addr, 1,
1508 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1509 NULL) < 1)
1510 return NULL;
1511 flush_cache_page(vma, addr, page_to_pfn(page));
1512 return page;
1513}
1514#endif /* CONFIG_ELF_CORE */
1515#else /* CONFIG_MMU */
1516static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1517 unsigned long nr_pages, struct page **pages,
1518 struct vm_area_struct **vmas, int *locked,
1519 unsigned int foll_flags)
1520{
1521 struct vm_area_struct *vma;
1522 unsigned long vm_flags;
1523 int i;
1524
1525 /* calculate required read or write permissions.
1526 * If FOLL_FORCE is set, we only require the "MAY" flags.
1527 */
1528 vm_flags = (foll_flags & FOLL_WRITE) ?
1529 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1530 vm_flags &= (foll_flags & FOLL_FORCE) ?
1531 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1532
1533 for (i = 0; i < nr_pages; i++) {
1534 vma = find_vma(mm, start);
1535 if (!vma)
1536 goto finish_or_fault;
1537
1538 /* protect what we can, including chardevs */
1539 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1540 !(vm_flags & vma->vm_flags))
1541 goto finish_or_fault;
1542
1543 if (pages) {
1544 pages[i] = virt_to_page(start);
1545 if (pages[i])
1546 get_page(pages[i]);
1547 }
1548 if (vmas)
1549 vmas[i] = vma;
1550 start = (start + PAGE_SIZE) & PAGE_MASK;
1551 }
1552
1553 return i;
1554
1555finish_or_fault:
1556 return i ? : -EFAULT;
1557}
1558#endif /* !CONFIG_MMU */
1559
1560#if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1561static bool check_dax_vmas(struct vm_area_struct **vmas, long nr_pages)
1562{
1563 long i;
1564 struct vm_area_struct *vma_prev = NULL;
1565
1566 for (i = 0; i < nr_pages; i++) {
1567 struct vm_area_struct *vma = vmas[i];
1568
1569 if (vma == vma_prev)
1570 continue;
1571
1572 vma_prev = vma;
1573
1574 if (vma_is_fsdax(vma))
1575 return true;
1576 }
1577 return false;
1578}
1579
1580#ifdef CONFIG_CMA
1581static long check_and_migrate_cma_pages(struct mm_struct *mm,
1582 unsigned long start,
1583 unsigned long nr_pages,
1584 struct page **pages,
1585 struct vm_area_struct **vmas,
1586 unsigned int gup_flags)
1587{
1588 unsigned long i;
1589 unsigned long step;
1590 bool drain_allow = true;
1591 bool migrate_allow = true;
1592 LIST_HEAD(cma_page_list);
1593 long ret = nr_pages;
1594 struct migration_target_control mtc = {
1595 .nid = NUMA_NO_NODE,
1596 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_NOWARN,
1597 };
1598
1599check_again:
1600 for (i = 0; i < nr_pages;) {
1601
1602 struct page *head = compound_head(pages[i]);
1603
1604 /*
1605 * gup may start from a tail page. Advance step by the left
1606 * part.
1607 */
1608 step = compound_nr(head) - (pages[i] - head);
1609 /*
1610 * If we get a page from the CMA zone, since we are going to
1611 * be pinning these entries, we might as well move them out
1612 * of the CMA zone if possible.
1613 */
1614 if (is_migrate_cma_page(head)) {
1615 if (PageHuge(head))
1616 isolate_huge_page(head, &cma_page_list);
1617 else {
1618 if (!PageLRU(head) && drain_allow) {
1619 lru_add_drain_all();
1620 drain_allow = false;
1621 }
1622
1623 if (!isolate_lru_page(head)) {
1624 list_add_tail(&head->lru, &cma_page_list);
1625 mod_node_page_state(page_pgdat(head),
1626 NR_ISOLATED_ANON +
1627 page_is_file_lru(head),
1628 thp_nr_pages(head));
1629 }
1630 }
1631 }
1632
1633 i += step;
1634 }
1635
1636 if (!list_empty(&cma_page_list)) {
1637 /*
1638 * drop the above get_user_pages reference.
1639 */
1640 for (i = 0; i < nr_pages; i++)
1641 put_page(pages[i]);
1642
1643 if (migrate_pages(&cma_page_list, alloc_migration_target, NULL,
1644 (unsigned long)&mtc, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
1645 /*
1646 * some of the pages failed migration. Do get_user_pages
1647 * without migration.
1648 */
1649 migrate_allow = false;
1650
1651 if (!list_empty(&cma_page_list))
1652 putback_movable_pages(&cma_page_list);
1653 }
1654 /*
1655 * We did migrate all the pages, Try to get the page references
1656 * again migrating any new CMA pages which we failed to isolate
1657 * earlier.
1658 */
1659 ret = __get_user_pages_locked(mm, start, nr_pages,
1660 pages, vmas, NULL,
1661 gup_flags);
1662
1663 if ((ret > 0) && migrate_allow) {
1664 nr_pages = ret;
1665 drain_allow = true;
1666 goto check_again;
1667 }
1668 }
1669
1670 return ret;
1671}
1672#else
1673static long check_and_migrate_cma_pages(struct mm_struct *mm,
1674 unsigned long start,
1675 unsigned long nr_pages,
1676 struct page **pages,
1677 struct vm_area_struct **vmas,
1678 unsigned int gup_flags)
1679{
1680 return nr_pages;
1681}
1682#endif /* CONFIG_CMA */
1683
1684/*
1685 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1686 * allows us to process the FOLL_LONGTERM flag.
1687 */
1688static long __gup_longterm_locked(struct mm_struct *mm,
1689 unsigned long start,
1690 unsigned long nr_pages,
1691 struct page **pages,
1692 struct vm_area_struct **vmas,
1693 unsigned int gup_flags)
1694{
1695 struct vm_area_struct **vmas_tmp = vmas;
1696 unsigned long flags = 0;
1697 long rc, i;
1698
1699 if (gup_flags & FOLL_LONGTERM) {
1700 if (!pages)
1701 return -EINVAL;
1702
1703 if (!vmas_tmp) {
1704 vmas_tmp = kcalloc(nr_pages,
1705 sizeof(struct vm_area_struct *),
1706 GFP_KERNEL);
1707 if (!vmas_tmp)
1708 return -ENOMEM;
1709 }
1710 flags = memalloc_nocma_save();
1711 }
1712
1713 rc = __get_user_pages_locked(mm, start, nr_pages, pages,
1714 vmas_tmp, NULL, gup_flags);
1715
1716 if (gup_flags & FOLL_LONGTERM) {
1717 if (rc < 0)
1718 goto out;
1719
1720 if (check_dax_vmas(vmas_tmp, rc)) {
1721 for (i = 0; i < rc; i++)
1722 put_page(pages[i]);
1723 rc = -EOPNOTSUPP;
1724 goto out;
1725 }
1726
1727 rc = check_and_migrate_cma_pages(mm, start, rc, pages,
1728 vmas_tmp, gup_flags);
1729out:
1730 memalloc_nocma_restore(flags);
1731 }
1732
1733 if (vmas_tmp != vmas)
1734 kfree(vmas_tmp);
1735 return rc;
1736}
1737#else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1738static __always_inline long __gup_longterm_locked(struct mm_struct *mm,
1739 unsigned long start,
1740 unsigned long nr_pages,
1741 struct page **pages,
1742 struct vm_area_struct **vmas,
1743 unsigned int flags)
1744{
1745 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1746 NULL, flags);
1747}
1748#endif /* CONFIG_FS_DAX || CONFIG_CMA */
1749
1750#ifdef CONFIG_MMU
1751static long __get_user_pages_remote(struct mm_struct *mm,
1752 unsigned long start, unsigned long nr_pages,
1753 unsigned int gup_flags, struct page **pages,
1754 struct vm_area_struct **vmas, int *locked)
1755{
1756 /*
1757 * Parts of FOLL_LONGTERM behavior are incompatible with
1758 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1759 * vmas. However, this only comes up if locked is set, and there are
1760 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1761 * allow what we can.
1762 */
1763 if (gup_flags & FOLL_LONGTERM) {
1764 if (WARN_ON_ONCE(locked))
1765 return -EINVAL;
1766 /*
1767 * This will check the vmas (even if our vmas arg is NULL)
1768 * and return -ENOTSUPP if DAX isn't allowed in this case:
1769 */
1770 return __gup_longterm_locked(mm, start, nr_pages, pages,
1771 vmas, gup_flags | FOLL_TOUCH |
1772 FOLL_REMOTE);
1773 }
1774
1775 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1776 locked,
1777 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1778}
1779
1780/**
1781 * get_user_pages_remote() - pin user pages in memory
1782 * @mm: mm_struct of target mm
1783 * @start: starting user address
1784 * @nr_pages: number of pages from start to pin
1785 * @gup_flags: flags modifying lookup behaviour
1786 * @pages: array that receives pointers to the pages pinned.
1787 * Should be at least nr_pages long. Or NULL, if caller
1788 * only intends to ensure the pages are faulted in.
1789 * @vmas: array of pointers to vmas corresponding to each page.
1790 * Or NULL if the caller does not require them.
1791 * @locked: pointer to lock flag indicating whether lock is held and
1792 * subsequently whether VM_FAULT_RETRY functionality can be
1793 * utilised. Lock must initially be held.
1794 *
1795 * Returns either number of pages pinned (which may be less than the
1796 * number requested), or an error. Details about the return value:
1797 *
1798 * -- If nr_pages is 0, returns 0.
1799 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1800 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1801 * pages pinned. Again, this may be less than nr_pages.
1802 *
1803 * The caller is responsible for releasing returned @pages, via put_page().
1804 *
1805 * @vmas are valid only as long as mmap_lock is held.
1806 *
1807 * Must be called with mmap_lock held for read or write.
1808 *
1809 * get_user_pages_remote walks a process's page tables and takes a reference
1810 * to each struct page that each user address corresponds to at a given
1811 * instant. That is, it takes the page that would be accessed if a user
1812 * thread accesses the given user virtual address at that instant.
1813 *
1814 * This does not guarantee that the page exists in the user mappings when
1815 * get_user_pages_remote returns, and there may even be a completely different
1816 * page there in some cases (eg. if mmapped pagecache has been invalidated
1817 * and subsequently re faulted). However it does guarantee that the page
1818 * won't be freed completely. And mostly callers simply care that the page
1819 * contains data that was valid *at some point in time*. Typically, an IO
1820 * or similar operation cannot guarantee anything stronger anyway because
1821 * locks can't be held over the syscall boundary.
1822 *
1823 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1824 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1825 * be called after the page is finished with, and before put_page is called.
1826 *
1827 * get_user_pages_remote is typically used for fewer-copy IO operations,
1828 * to get a handle on the memory by some means other than accesses
1829 * via the user virtual addresses. The pages may be submitted for
1830 * DMA to devices or accessed via their kernel linear mapping (via the
1831 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1832 *
1833 * See also get_user_pages_fast, for performance critical applications.
1834 *
1835 * get_user_pages_remote should be phased out in favor of
1836 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1837 * should use get_user_pages_remote because it cannot pass
1838 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1839 */
1840long get_user_pages_remote(struct mm_struct *mm,
1841 unsigned long start, unsigned long nr_pages,
1842 unsigned int gup_flags, struct page **pages,
1843 struct vm_area_struct **vmas, int *locked)
1844{
1845 /*
1846 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1847 * never directly by the caller, so enforce that with an assertion:
1848 */
1849 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1850 return -EINVAL;
1851
1852 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
1853 pages, vmas, locked);
1854}
1855EXPORT_SYMBOL(get_user_pages_remote);
1856
1857#else /* CONFIG_MMU */
1858long get_user_pages_remote(struct mm_struct *mm,
1859 unsigned long start, unsigned long nr_pages,
1860 unsigned int gup_flags, struct page **pages,
1861 struct vm_area_struct **vmas, int *locked)
1862{
1863 return 0;
1864}
1865
1866static long __get_user_pages_remote(struct mm_struct *mm,
1867 unsigned long start, unsigned long nr_pages,
1868 unsigned int gup_flags, struct page **pages,
1869 struct vm_area_struct **vmas, int *locked)
1870{
1871 return 0;
1872}
1873#endif /* !CONFIG_MMU */
1874
1875/**
1876 * get_user_pages() - pin user pages in memory
1877 * @start: starting user address
1878 * @nr_pages: number of pages from start to pin
1879 * @gup_flags: flags modifying lookup behaviour
1880 * @pages: array that receives pointers to the pages pinned.
1881 * Should be at least nr_pages long. Or NULL, if caller
1882 * only intends to ensure the pages are faulted in.
1883 * @vmas: array of pointers to vmas corresponding to each page.
1884 * Or NULL if the caller does not require them.
1885 *
1886 * This is the same as get_user_pages_remote(), just with a less-flexible
1887 * calling convention where we assume that the mm being operated on belongs to
1888 * the current task, and doesn't allow passing of a locked parameter. We also
1889 * obviously don't pass FOLL_REMOTE in here.
1890 */
1891long get_user_pages(unsigned long start, unsigned long nr_pages,
1892 unsigned int gup_flags, struct page **pages,
1893 struct vm_area_struct **vmas)
1894{
1895 /*
1896 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1897 * never directly by the caller, so enforce that with an assertion:
1898 */
1899 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1900 return -EINVAL;
1901
1902 return __gup_longterm_locked(current->mm, start, nr_pages,
1903 pages, vmas, gup_flags | FOLL_TOUCH);
1904}
1905EXPORT_SYMBOL(get_user_pages);
1906
1907/**
1908 * get_user_pages_locked() is suitable to replace the form:
1909 *
1910 * mmap_read_lock(mm);
1911 * do_something()
1912 * get_user_pages(mm, ..., pages, NULL);
1913 * mmap_read_unlock(mm);
1914 *
1915 * to:
1916 *
1917 * int locked = 1;
1918 * mmap_read_lock(mm);
1919 * do_something()
1920 * get_user_pages_locked(mm, ..., pages, &locked);
1921 * if (locked)
1922 * mmap_read_unlock(mm);
1923 *
1924 * @start: starting user address
1925 * @nr_pages: number of pages from start to pin
1926 * @gup_flags: flags modifying lookup behaviour
1927 * @pages: array that receives pointers to the pages pinned.
1928 * Should be at least nr_pages long. Or NULL, if caller
1929 * only intends to ensure the pages are faulted in.
1930 * @locked: pointer to lock flag indicating whether lock is held and
1931 * subsequently whether VM_FAULT_RETRY functionality can be
1932 * utilised. Lock must initially be held.
1933 *
1934 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1935 * paths better by using either get_user_pages_locked() or
1936 * get_user_pages_unlocked().
1937 *
1938 */
1939long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1940 unsigned int gup_flags, struct page **pages,
1941 int *locked)
1942{
1943 /*
1944 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1945 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1946 * vmas. As there are no users of this flag in this call we simply
1947 * disallow this option for now.
1948 */
1949 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1950 return -EINVAL;
1951 /*
1952 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1953 * never directly by the caller, so enforce that:
1954 */
1955 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1956 return -EINVAL;
1957
1958 return __get_user_pages_locked(current->mm, start, nr_pages,
1959 pages, NULL, locked,
1960 gup_flags | FOLL_TOUCH);
1961}
1962EXPORT_SYMBOL(get_user_pages_locked);
1963
1964/*
1965 * get_user_pages_unlocked() is suitable to replace the form:
1966 *
1967 * mmap_read_lock(mm);
1968 * get_user_pages(mm, ..., pages, NULL);
1969 * mmap_read_unlock(mm);
1970 *
1971 * with:
1972 *
1973 * get_user_pages_unlocked(mm, ..., pages);
1974 *
1975 * It is functionally equivalent to get_user_pages_fast so
1976 * get_user_pages_fast should be used instead if specific gup_flags
1977 * (e.g. FOLL_FORCE) are not required.
1978 */
1979long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1980 struct page **pages, unsigned int gup_flags)
1981{
1982 struct mm_struct *mm = current->mm;
1983 int locked = 1;
1984 long ret;
1985
1986 /*
1987 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1988 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1989 * vmas. As there are no users of this flag in this call we simply
1990 * disallow this option for now.
1991 */
1992 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1993 return -EINVAL;
1994
1995 mmap_read_lock(mm);
1996 ret = __get_user_pages_locked(mm, start, nr_pages, pages, NULL,
1997 &locked, gup_flags | FOLL_TOUCH);
1998 if (locked)
1999 mmap_read_unlock(mm);
2000 return ret;
2001}
2002EXPORT_SYMBOL(get_user_pages_unlocked);
2003
2004/*
2005 * Fast GUP
2006 *
2007 * get_user_pages_fast attempts to pin user pages by walking the page
2008 * tables directly and avoids taking locks. Thus the walker needs to be
2009 * protected from page table pages being freed from under it, and should
2010 * block any THP splits.
2011 *
2012 * One way to achieve this is to have the walker disable interrupts, and
2013 * rely on IPIs from the TLB flushing code blocking before the page table
2014 * pages are freed. This is unsuitable for architectures that do not need
2015 * to broadcast an IPI when invalidating TLBs.
2016 *
2017 * Another way to achieve this is to batch up page table containing pages
2018 * belonging to more than one mm_user, then rcu_sched a callback to free those
2019 * pages. Disabling interrupts will allow the fast_gup walker to both block
2020 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2021 * (which is a relatively rare event). The code below adopts this strategy.
2022 *
2023 * Before activating this code, please be aware that the following assumptions
2024 * are currently made:
2025 *
2026 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2027 * free pages containing page tables or TLB flushing requires IPI broadcast.
2028 *
2029 * *) ptes can be read atomically by the architecture.
2030 *
2031 * *) access_ok is sufficient to validate userspace address ranges.
2032 *
2033 * The last two assumptions can be relaxed by the addition of helper functions.
2034 *
2035 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2036 */
2037#ifdef CONFIG_HAVE_FAST_GUP
2038
2039static void put_compound_head(struct page *page, int refs, unsigned int flags)
2040{
2041 if (flags & FOLL_PIN) {
2042 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED,
2043 refs);
2044
2045 if (hpage_pincount_available(page))
2046 hpage_pincount_sub(page, refs);
2047 else
2048 refs *= GUP_PIN_COUNTING_BIAS;
2049 }
2050
2051 VM_BUG_ON_PAGE(page_ref_count(page) < refs, page);
2052 /*
2053 * Calling put_page() for each ref is unnecessarily slow. Only the last
2054 * ref needs a put_page().
2055 */
2056 if (refs > 1)
2057 page_ref_sub(page, refs - 1);
2058 put_page(page);
2059}
2060
2061#ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
2062
2063/*
2064 * WARNING: only to be used in the get_user_pages_fast() implementation.
2065 *
2066 * With get_user_pages_fast(), we walk down the pagetables without taking any
2067 * locks. For this we would like to load the pointers atomically, but sometimes
2068 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
2069 * we do have is the guarantee that a PTE will only either go from not present
2070 * to present, or present to not present or both -- it will not switch to a
2071 * completely different present page without a TLB flush in between; something
2072 * that we are blocking by holding interrupts off.
2073 *
2074 * Setting ptes from not present to present goes:
2075 *
2076 * ptep->pte_high = h;
2077 * smp_wmb();
2078 * ptep->pte_low = l;
2079 *
2080 * And present to not present goes:
2081 *
2082 * ptep->pte_low = 0;
2083 * smp_wmb();
2084 * ptep->pte_high = 0;
2085 *
2086 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
2087 * We load pte_high *after* loading pte_low, which ensures we don't see an older
2088 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
2089 * picked up a changed pte high. We might have gotten rubbish values from
2090 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
2091 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
2092 * operates on present ptes we're safe.
2093 */
2094static inline pte_t gup_get_pte(pte_t *ptep)
2095{
2096 pte_t pte;
2097
2098 do {
2099 pte.pte_low = ptep->pte_low;
2100 smp_rmb();
2101 pte.pte_high = ptep->pte_high;
2102 smp_rmb();
2103 } while (unlikely(pte.pte_low != ptep->pte_low));
2104
2105 return pte;
2106}
2107#else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2108/*
2109 * We require that the PTE can be read atomically.
2110 */
2111static inline pte_t gup_get_pte(pte_t *ptep)
2112{
2113 return ptep_get(ptep);
2114}
2115#endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2116
2117static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2118 unsigned int flags,
2119 struct page **pages)
2120{
2121 while ((*nr) - nr_start) {
2122 struct page *page = pages[--(*nr)];
2123
2124 ClearPageReferenced(page);
2125 if (flags & FOLL_PIN)
2126 unpin_user_page(page);
2127 else
2128 put_page(page);
2129 }
2130}
2131
2132#ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2133static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2134 unsigned int flags, struct page **pages, int *nr)
2135{
2136 struct dev_pagemap *pgmap = NULL;
2137 int nr_start = *nr, ret = 0;
2138 pte_t *ptep, *ptem;
2139
2140 ptem = ptep = pte_offset_map(&pmd, addr);
2141 do {
2142 pte_t pte = gup_get_pte(ptep);
2143 struct page *head, *page;
2144
2145 /*
2146 * Similar to the PMD case below, NUMA hinting must take slow
2147 * path using the pte_protnone check.
2148 */
2149 if (pte_protnone(pte))
2150 goto pte_unmap;
2151
2152 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2153 goto pte_unmap;
2154
2155 if (pte_devmap(pte)) {
2156 if (unlikely(flags & FOLL_LONGTERM))
2157 goto pte_unmap;
2158
2159 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2160 if (unlikely(!pgmap)) {
2161 undo_dev_pagemap(nr, nr_start, flags, pages);
2162 goto pte_unmap;
2163 }
2164 } else if (pte_special(pte))
2165 goto pte_unmap;
2166
2167 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2168 page = pte_page(pte);
2169
2170 head = try_grab_compound_head(page, 1, flags);
2171 if (!head)
2172 goto pte_unmap;
2173
2174 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2175 put_compound_head(head, 1, flags);
2176 goto pte_unmap;
2177 }
2178
2179 VM_BUG_ON_PAGE(compound_head(page) != head, page);
2180
2181 /*
2182 * We need to make the page accessible if and only if we are
2183 * going to access its content (the FOLL_PIN case). Please
2184 * see Documentation/core-api/pin_user_pages.rst for
2185 * details.
2186 */
2187 if (flags & FOLL_PIN) {
2188 ret = arch_make_page_accessible(page);
2189 if (ret) {
2190 unpin_user_page(page);
2191 goto pte_unmap;
2192 }
2193 }
2194 SetPageReferenced(page);
2195 pages[*nr] = page;
2196 (*nr)++;
2197
2198 } while (ptep++, addr += PAGE_SIZE, addr != end);
2199
2200 ret = 1;
2201
2202pte_unmap:
2203 if (pgmap)
2204 put_dev_pagemap(pgmap);
2205 pte_unmap(ptem);
2206 return ret;
2207}
2208#else
2209
2210/*
2211 * If we can't determine whether or not a pte is special, then fail immediately
2212 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2213 * to be special.
2214 *
2215 * For a futex to be placed on a THP tail page, get_futex_key requires a
2216 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2217 * useful to have gup_huge_pmd even if we can't operate on ptes.
2218 */
2219static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2220 unsigned int flags, struct page **pages, int *nr)
2221{
2222 return 0;
2223}
2224#endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2225
2226#if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2227static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2228 unsigned long end, unsigned int flags,
2229 struct page **pages, int *nr)
2230{
2231 int nr_start = *nr;
2232 struct dev_pagemap *pgmap = NULL;
2233
2234 do {
2235 struct page *page = pfn_to_page(pfn);
2236
2237 pgmap = get_dev_pagemap(pfn, pgmap);
2238 if (unlikely(!pgmap)) {
2239 undo_dev_pagemap(nr, nr_start, flags, pages);
2240 return 0;
2241 }
2242 SetPageReferenced(page);
2243 pages[*nr] = page;
2244 if (unlikely(!try_grab_page(page, flags))) {
2245 undo_dev_pagemap(nr, nr_start, flags, pages);
2246 return 0;
2247 }
2248 (*nr)++;
2249 pfn++;
2250 } while (addr += PAGE_SIZE, addr != end);
2251
2252 if (pgmap)
2253 put_dev_pagemap(pgmap);
2254 return 1;
2255}
2256
2257static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2258 unsigned long end, unsigned int flags,
2259 struct page **pages, int *nr)
2260{
2261 unsigned long fault_pfn;
2262 int nr_start = *nr;
2263
2264 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2265 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2266 return 0;
2267
2268 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2269 undo_dev_pagemap(nr, nr_start, flags, pages);
2270 return 0;
2271 }
2272 return 1;
2273}
2274
2275static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2276 unsigned long end, unsigned int flags,
2277 struct page **pages, int *nr)
2278{
2279 unsigned long fault_pfn;
2280 int nr_start = *nr;
2281
2282 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2283 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2284 return 0;
2285
2286 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2287 undo_dev_pagemap(nr, nr_start, flags, pages);
2288 return 0;
2289 }
2290 return 1;
2291}
2292#else
2293static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2294 unsigned long end, unsigned int flags,
2295 struct page **pages, int *nr)
2296{
2297 BUILD_BUG();
2298 return 0;
2299}
2300
2301static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2302 unsigned long end, unsigned int flags,
2303 struct page **pages, int *nr)
2304{
2305 BUILD_BUG();
2306 return 0;
2307}
2308#endif
2309
2310static int record_subpages(struct page *page, unsigned long addr,
2311 unsigned long end, struct page **pages)
2312{
2313 int nr;
2314
2315 for (nr = 0; addr != end; addr += PAGE_SIZE)
2316 pages[nr++] = page++;
2317
2318 return nr;
2319}
2320
2321#ifdef CONFIG_ARCH_HAS_HUGEPD
2322static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2323 unsigned long sz)
2324{
2325 unsigned long __boundary = (addr + sz) & ~(sz-1);
2326 return (__boundary - 1 < end - 1) ? __boundary : end;
2327}
2328
2329static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2330 unsigned long end, unsigned int flags,
2331 struct page **pages, int *nr)
2332{
2333 unsigned long pte_end;
2334 struct page *head, *page;
2335 pte_t pte;
2336 int refs;
2337
2338 pte_end = (addr + sz) & ~(sz-1);
2339 if (pte_end < end)
2340 end = pte_end;
2341
2342 pte = huge_ptep_get(ptep);
2343
2344 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2345 return 0;
2346
2347 /* hugepages are never "special" */
2348 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2349
2350 head = pte_page(pte);
2351 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2352 refs = record_subpages(page, addr, end, pages + *nr);
2353
2354 head = try_grab_compound_head(head, refs, flags);
2355 if (!head)
2356 return 0;
2357
2358 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2359 put_compound_head(head, refs, flags);
2360 return 0;
2361 }
2362
2363 *nr += refs;
2364 SetPageReferenced(head);
2365 return 1;
2366}
2367
2368static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2369 unsigned int pdshift, unsigned long end, unsigned int flags,
2370 struct page **pages, int *nr)
2371{
2372 pte_t *ptep;
2373 unsigned long sz = 1UL << hugepd_shift(hugepd);
2374 unsigned long next;
2375
2376 ptep = hugepte_offset(hugepd, addr, pdshift);
2377 do {
2378 next = hugepte_addr_end(addr, end, sz);
2379 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2380 return 0;
2381 } while (ptep++, addr = next, addr != end);
2382
2383 return 1;
2384}
2385#else
2386static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2387 unsigned int pdshift, unsigned long end, unsigned int flags,
2388 struct page **pages, int *nr)
2389{
2390 return 0;
2391}
2392#endif /* CONFIG_ARCH_HAS_HUGEPD */
2393
2394static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2395 unsigned long end, unsigned int flags,
2396 struct page **pages, int *nr)
2397{
2398 struct page *head, *page;
2399 int refs;
2400
2401 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2402 return 0;
2403
2404 if (pmd_devmap(orig)) {
2405 if (unlikely(flags & FOLL_LONGTERM))
2406 return 0;
2407 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2408 pages, nr);
2409 }
2410
2411 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2412 refs = record_subpages(page, addr, end, pages + *nr);
2413
2414 head = try_grab_compound_head(pmd_page(orig), refs, flags);
2415 if (!head)
2416 return 0;
2417
2418 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2419 put_compound_head(head, refs, flags);
2420 return 0;
2421 }
2422
2423 *nr += refs;
2424 SetPageReferenced(head);
2425 return 1;
2426}
2427
2428static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2429 unsigned long end, unsigned int flags,
2430 struct page **pages, int *nr)
2431{
2432 struct page *head, *page;
2433 int refs;
2434
2435 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2436 return 0;
2437
2438 if (pud_devmap(orig)) {
2439 if (unlikely(flags & FOLL_LONGTERM))
2440 return 0;
2441 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2442 pages, nr);
2443 }
2444
2445 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2446 refs = record_subpages(page, addr, end, pages + *nr);
2447
2448 head = try_grab_compound_head(pud_page(orig), refs, flags);
2449 if (!head)
2450 return 0;
2451
2452 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2453 put_compound_head(head, refs, flags);
2454 return 0;
2455 }
2456
2457 *nr += refs;
2458 SetPageReferenced(head);
2459 return 1;
2460}
2461
2462static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2463 unsigned long end, unsigned int flags,
2464 struct page **pages, int *nr)
2465{
2466 int refs;
2467 struct page *head, *page;
2468
2469 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2470 return 0;
2471
2472 BUILD_BUG_ON(pgd_devmap(orig));
2473
2474 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2475 refs = record_subpages(page, addr, end, pages + *nr);
2476
2477 head = try_grab_compound_head(pgd_page(orig), refs, flags);
2478 if (!head)
2479 return 0;
2480
2481 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2482 put_compound_head(head, refs, flags);
2483 return 0;
2484 }
2485
2486 *nr += refs;
2487 SetPageReferenced(head);
2488 return 1;
2489}
2490
2491static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
2492 unsigned int flags, struct page **pages, int *nr)
2493{
2494 unsigned long next;
2495 pmd_t *pmdp;
2496
2497 pmdp = pmd_offset_lockless(pudp, pud, addr);
2498 do {
2499 pmd_t pmd = READ_ONCE(*pmdp);
2500
2501 next = pmd_addr_end(addr, end);
2502 if (!pmd_present(pmd))
2503 return 0;
2504
2505 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2506 pmd_devmap(pmd))) {
2507 /*
2508 * NUMA hinting faults need to be handled in the GUP
2509 * slowpath for accounting purposes and so that they
2510 * can be serialised against THP migration.
2511 */
2512 if (pmd_protnone(pmd))
2513 return 0;
2514
2515 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2516 pages, nr))
2517 return 0;
2518
2519 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2520 /*
2521 * architecture have different format for hugetlbfs
2522 * pmd format and THP pmd format
2523 */
2524 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2525 PMD_SHIFT, next, flags, pages, nr))
2526 return 0;
2527 } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2528 return 0;
2529 } while (pmdp++, addr = next, addr != end);
2530
2531 return 1;
2532}
2533
2534static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
2535 unsigned int flags, struct page **pages, int *nr)
2536{
2537 unsigned long next;
2538 pud_t *pudp;
2539
2540 pudp = pud_offset_lockless(p4dp, p4d, addr);
2541 do {
2542 pud_t pud = READ_ONCE(*pudp);
2543
2544 next = pud_addr_end(addr, end);
2545 if (unlikely(!pud_present(pud)))
2546 return 0;
2547 if (unlikely(pud_huge(pud))) {
2548 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2549 pages, nr))
2550 return 0;
2551 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2552 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2553 PUD_SHIFT, next, flags, pages, nr))
2554 return 0;
2555 } else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
2556 return 0;
2557 } while (pudp++, addr = next, addr != end);
2558
2559 return 1;
2560}
2561
2562static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
2563 unsigned int flags, struct page **pages, int *nr)
2564{
2565 unsigned long next;
2566 p4d_t *p4dp;
2567
2568 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
2569 do {
2570 p4d_t p4d = READ_ONCE(*p4dp);
2571
2572 next = p4d_addr_end(addr, end);
2573 if (p4d_none(p4d))
2574 return 0;
2575 BUILD_BUG_ON(p4d_huge(p4d));
2576 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2577 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2578 P4D_SHIFT, next, flags, pages, nr))
2579 return 0;
2580 } else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
2581 return 0;
2582 } while (p4dp++, addr = next, addr != end);
2583
2584 return 1;
2585}
2586
2587static void gup_pgd_range(unsigned long addr, unsigned long end,
2588 unsigned int flags, struct page **pages, int *nr)
2589{
2590 unsigned long next;
2591 pgd_t *pgdp;
2592
2593 pgdp = pgd_offset(current->mm, addr);
2594 do {
2595 pgd_t pgd = READ_ONCE(*pgdp);
2596
2597 next = pgd_addr_end(addr, end);
2598 if (pgd_none(pgd))
2599 return;
2600 if (unlikely(pgd_huge(pgd))) {
2601 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2602 pages, nr))
2603 return;
2604 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2605 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2606 PGDIR_SHIFT, next, flags, pages, nr))
2607 return;
2608 } else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
2609 return;
2610 } while (pgdp++, addr = next, addr != end);
2611}
2612#else
2613static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2614 unsigned int flags, struct page **pages, int *nr)
2615{
2616}
2617#endif /* CONFIG_HAVE_FAST_GUP */
2618
2619#ifndef gup_fast_permitted
2620/*
2621 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2622 * we need to fall back to the slow version:
2623 */
2624static bool gup_fast_permitted(unsigned long start, unsigned long end)
2625{
2626 return true;
2627}
2628#endif
2629
2630static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2631 unsigned int gup_flags, struct page **pages)
2632{
2633 int ret;
2634
2635 /*
2636 * FIXME: FOLL_LONGTERM does not work with
2637 * get_user_pages_unlocked() (see comments in that function)
2638 */
2639 if (gup_flags & FOLL_LONGTERM) {
2640 mmap_read_lock(current->mm);
2641 ret = __gup_longterm_locked(current->mm,
2642 start, nr_pages,
2643 pages, NULL, gup_flags);
2644 mmap_read_unlock(current->mm);
2645 } else {
2646 ret = get_user_pages_unlocked(start, nr_pages,
2647 pages, gup_flags);
2648 }
2649
2650 return ret;
2651}
2652
2653static int internal_get_user_pages_fast(unsigned long start, int nr_pages,
2654 unsigned int gup_flags,
2655 struct page **pages)
2656{
2657 unsigned long addr, len, end;
2658 unsigned long flags;
2659 int nr_pinned = 0, ret = 0;
2660
2661 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
2662 FOLL_FORCE | FOLL_PIN | FOLL_GET |
2663 FOLL_FAST_ONLY)))
2664 return -EINVAL;
2665
2666 if (gup_flags & FOLL_PIN)
2667 atomic_set(¤t->mm->has_pinned, 1);
2668
2669 if (!(gup_flags & FOLL_FAST_ONLY))
2670 might_lock_read(¤t->mm->mmap_lock);
2671
2672 start = untagged_addr(start) & PAGE_MASK;
2673 addr = start;
2674 len = (unsigned long) nr_pages << PAGE_SHIFT;
2675 end = start + len;
2676
2677 if (end <= start)
2678 return 0;
2679 if (unlikely(!access_ok((void __user *)start, len)))
2680 return -EFAULT;
2681
2682 /*
2683 * Disable interrupts. The nested form is used, in order to allow
2684 * full, general purpose use of this routine.
2685 *
2686 * With interrupts disabled, we block page table pages from being
2687 * freed from under us. See struct mmu_table_batch comments in
2688 * include/asm-generic/tlb.h for more details.
2689 *
2690 * We do not adopt an rcu_read_lock(.) here as we also want to
2691 * block IPIs that come from THPs splitting.
2692 */
2693 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) && gup_fast_permitted(start, end)) {
2694 unsigned long fast_flags = gup_flags;
2695
2696 local_irq_save(flags);
2697 gup_pgd_range(addr, end, fast_flags, pages, &nr_pinned);
2698 local_irq_restore(flags);
2699 ret = nr_pinned;
2700 }
2701
2702 if (nr_pinned < nr_pages && !(gup_flags & FOLL_FAST_ONLY)) {
2703 /* Try to get the remaining pages with get_user_pages */
2704 start += nr_pinned << PAGE_SHIFT;
2705 pages += nr_pinned;
2706
2707 ret = __gup_longterm_unlocked(start, nr_pages - nr_pinned,
2708 gup_flags, pages);
2709
2710 /* Have to be a bit careful with return values */
2711 if (nr_pinned > 0) {
2712 if (ret < 0)
2713 ret = nr_pinned;
2714 else
2715 ret += nr_pinned;
2716 }
2717 }
2718
2719 return ret;
2720}
2721/**
2722 * get_user_pages_fast_only() - pin user pages in memory
2723 * @start: starting user address
2724 * @nr_pages: number of pages from start to pin
2725 * @gup_flags: flags modifying pin behaviour
2726 * @pages: array that receives pointers to the pages pinned.
2727 * Should be at least nr_pages long.
2728 *
2729 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2730 * the regular GUP.
2731 * Note a difference with get_user_pages_fast: this always returns the
2732 * number of pages pinned, 0 if no pages were pinned.
2733 *
2734 * If the architecture does not support this function, simply return with no
2735 * pages pinned.
2736 *
2737 * Careful, careful! COW breaking can go either way, so a non-write
2738 * access can get ambiguous page results. If you call this function without
2739 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2740 */
2741int get_user_pages_fast_only(unsigned long start, int nr_pages,
2742 unsigned int gup_flags, struct page **pages)
2743{
2744 int nr_pinned;
2745 /*
2746 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2747 * because gup fast is always a "pin with a +1 page refcount" request.
2748 *
2749 * FOLL_FAST_ONLY is required in order to match the API description of
2750 * this routine: no fall back to regular ("slow") GUP.
2751 */
2752 gup_flags |= FOLL_GET | FOLL_FAST_ONLY;
2753
2754 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2755 pages);
2756
2757 /*
2758 * As specified in the API description above, this routine is not
2759 * allowed to return negative values. However, the common core
2760 * routine internal_get_user_pages_fast() *can* return -errno.
2761 * Therefore, correct for that here:
2762 */
2763 if (nr_pinned < 0)
2764 nr_pinned = 0;
2765
2766 return nr_pinned;
2767}
2768EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
2769
2770/**
2771 * get_user_pages_fast() - pin user pages in memory
2772 * @start: starting user address
2773 * @nr_pages: number of pages from start to pin
2774 * @gup_flags: flags modifying pin behaviour
2775 * @pages: array that receives pointers to the pages pinned.
2776 * Should be at least nr_pages long.
2777 *
2778 * Attempt to pin user pages in memory without taking mm->mmap_lock.
2779 * If not successful, it will fall back to taking the lock and
2780 * calling get_user_pages().
2781 *
2782 * Returns number of pages pinned. This may be fewer than the number requested.
2783 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2784 * -errno.
2785 */
2786int get_user_pages_fast(unsigned long start, int nr_pages,
2787 unsigned int gup_flags, struct page **pages)
2788{
2789 /*
2790 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
2791 * never directly by the caller, so enforce that:
2792 */
2793 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
2794 return -EINVAL;
2795
2796 /*
2797 * The caller may or may not have explicitly set FOLL_GET; either way is
2798 * OK. However, internally (within mm/gup.c), gup fast variants must set
2799 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2800 * request.
2801 */
2802 gup_flags |= FOLL_GET;
2803 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2804}
2805EXPORT_SYMBOL_GPL(get_user_pages_fast);
2806
2807/**
2808 * pin_user_pages_fast() - pin user pages in memory without taking locks
2809 *
2810 * @start: starting user address
2811 * @nr_pages: number of pages from start to pin
2812 * @gup_flags: flags modifying pin behaviour
2813 * @pages: array that receives pointers to the pages pinned.
2814 * Should be at least nr_pages long.
2815 *
2816 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2817 * get_user_pages_fast() for documentation on the function arguments, because
2818 * the arguments here are identical.
2819 *
2820 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2821 * see Documentation/core-api/pin_user_pages.rst for further details.
2822 */
2823int pin_user_pages_fast(unsigned long start, int nr_pages,
2824 unsigned int gup_flags, struct page **pages)
2825{
2826 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2827 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2828 return -EINVAL;
2829
2830 gup_flags |= FOLL_PIN;
2831 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2832}
2833EXPORT_SYMBOL_GPL(pin_user_pages_fast);
2834
2835/*
2836 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
2837 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
2838 *
2839 * The API rules are the same, too: no negative values may be returned.
2840 */
2841int pin_user_pages_fast_only(unsigned long start, int nr_pages,
2842 unsigned int gup_flags, struct page **pages)
2843{
2844 int nr_pinned;
2845
2846 /*
2847 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
2848 * rules require returning 0, rather than -errno:
2849 */
2850 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2851 return 0;
2852 /*
2853 * FOLL_FAST_ONLY is required in order to match the API description of
2854 * this routine: no fall back to regular ("slow") GUP.
2855 */
2856 gup_flags |= (FOLL_PIN | FOLL_FAST_ONLY);
2857 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2858 pages);
2859 /*
2860 * This routine is not allowed to return negative values. However,
2861 * internal_get_user_pages_fast() *can* return -errno. Therefore,
2862 * correct for that here:
2863 */
2864 if (nr_pinned < 0)
2865 nr_pinned = 0;
2866
2867 return nr_pinned;
2868}
2869EXPORT_SYMBOL_GPL(pin_user_pages_fast_only);
2870
2871/**
2872 * pin_user_pages_remote() - pin pages of a remote process
2873 *
2874 * @mm: mm_struct of target mm
2875 * @start: starting user address
2876 * @nr_pages: number of pages from start to pin
2877 * @gup_flags: flags modifying lookup behaviour
2878 * @pages: array that receives pointers to the pages pinned.
2879 * Should be at least nr_pages long. Or NULL, if caller
2880 * only intends to ensure the pages are faulted in.
2881 * @vmas: array of pointers to vmas corresponding to each page.
2882 * Or NULL if the caller does not require them.
2883 * @locked: pointer to lock flag indicating whether lock is held and
2884 * subsequently whether VM_FAULT_RETRY functionality can be
2885 * utilised. Lock must initially be held.
2886 *
2887 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2888 * get_user_pages_remote() for documentation on the function arguments, because
2889 * the arguments here are identical.
2890 *
2891 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2892 * see Documentation/core-api/pin_user_pages.rst for details.
2893 */
2894long pin_user_pages_remote(struct mm_struct *mm,
2895 unsigned long start, unsigned long nr_pages,
2896 unsigned int gup_flags, struct page **pages,
2897 struct vm_area_struct **vmas, int *locked)
2898{
2899 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2900 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2901 return -EINVAL;
2902
2903 gup_flags |= FOLL_PIN;
2904 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
2905 pages, vmas, locked);
2906}
2907EXPORT_SYMBOL(pin_user_pages_remote);
2908
2909/**
2910 * pin_user_pages() - pin user pages in memory for use by other devices
2911 *
2912 * @start: starting user address
2913 * @nr_pages: number of pages from start to pin
2914 * @gup_flags: flags modifying lookup behaviour
2915 * @pages: array that receives pointers to the pages pinned.
2916 * Should be at least nr_pages long. Or NULL, if caller
2917 * only intends to ensure the pages are faulted in.
2918 * @vmas: array of pointers to vmas corresponding to each page.
2919 * Or NULL if the caller does not require them.
2920 *
2921 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2922 * FOLL_PIN is set.
2923 *
2924 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2925 * see Documentation/core-api/pin_user_pages.rst for details.
2926 */
2927long pin_user_pages(unsigned long start, unsigned long nr_pages,
2928 unsigned int gup_flags, struct page **pages,
2929 struct vm_area_struct **vmas)
2930{
2931 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2932 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2933 return -EINVAL;
2934
2935 gup_flags |= FOLL_PIN;
2936 return __gup_longterm_locked(current->mm, start, nr_pages,
2937 pages, vmas, gup_flags);
2938}
2939EXPORT_SYMBOL(pin_user_pages);
2940
2941/*
2942 * pin_user_pages_unlocked() is the FOLL_PIN variant of
2943 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
2944 * FOLL_PIN and rejects FOLL_GET.
2945 */
2946long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2947 struct page **pages, unsigned int gup_flags)
2948{
2949 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2950 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2951 return -EINVAL;
2952
2953 gup_flags |= FOLL_PIN;
2954 return get_user_pages_unlocked(start, nr_pages, pages, gup_flags);
2955}
2956EXPORT_SYMBOL(pin_user_pages_unlocked);
2957
2958/*
2959 * pin_user_pages_locked() is the FOLL_PIN variant of get_user_pages_locked().
2960 * Behavior is the same, except that this one sets FOLL_PIN and rejects
2961 * FOLL_GET.
2962 */
2963long pin_user_pages_locked(unsigned long start, unsigned long nr_pages,
2964 unsigned int gup_flags, struct page **pages,
2965 int *locked)
2966{
2967 /*
2968 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2969 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2970 * vmas. As there are no users of this flag in this call we simply
2971 * disallow this option for now.
2972 */
2973 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2974 return -EINVAL;
2975
2976 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2977 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2978 return -EINVAL;
2979
2980 gup_flags |= FOLL_PIN;
2981 return __get_user_pages_locked(current->mm, start, nr_pages,
2982 pages, NULL, locked,
2983 gup_flags | FOLL_TOUCH);
2984}
2985EXPORT_SYMBOL(pin_user_pages_locked);