]> git.ipfire.org Git - thirdparty/linux.git/blob - mm/gup.c
Merge tag 'io_uring-5.7-2020-05-22' of git://git.kernel.dk/linux-block
[thirdparty/linux.git] / mm / gup.c
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/pgtable.h>
23 #include <asm/tlbflush.h>
24
25 #include "internal.h"
26
27 struct follow_page_context {
28 struct dev_pagemap *pgmap;
29 unsigned int page_mask;
30 };
31
32 static void hpage_pincount_add(struct page *page, int refs)
33 {
34 VM_BUG_ON_PAGE(!hpage_pincount_available(page), page);
35 VM_BUG_ON_PAGE(page != compound_head(page), page);
36
37 atomic_add(refs, compound_pincount_ptr(page));
38 }
39
40 static void hpage_pincount_sub(struct page *page, int refs)
41 {
42 VM_BUG_ON_PAGE(!hpage_pincount_available(page), page);
43 VM_BUG_ON_PAGE(page != compound_head(page), page);
44
45 atomic_sub(refs, compound_pincount_ptr(page));
46 }
47
48 /*
49 * Return the compound head page with ref appropriately incremented,
50 * or NULL if that failed.
51 */
52 static inline struct page *try_get_compound_head(struct page *page, int refs)
53 {
54 struct page *head = compound_head(page);
55
56 if (WARN_ON_ONCE(page_ref_count(head) < 0))
57 return NULL;
58 if (unlikely(!page_cache_add_speculative(head, refs)))
59 return NULL;
60 return head;
61 }
62
63 /*
64 * try_grab_compound_head() - attempt to elevate a page's refcount, by a
65 * flags-dependent amount.
66 *
67 * "grab" names in this file mean, "look at flags to decide whether to use
68 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
69 *
70 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
71 * same time. (That's true throughout the get_user_pages*() and
72 * pin_user_pages*() APIs.) Cases:
73 *
74 * FOLL_GET: page's refcount will be incremented by 1.
75 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
76 *
77 * Return: head page (with refcount appropriately incremented) for success, or
78 * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's
79 * considered failure, and furthermore, a likely bug in the caller, so a warning
80 * is also emitted.
81 */
82 static __maybe_unused struct page *try_grab_compound_head(struct page *page,
83 int refs,
84 unsigned int flags)
85 {
86 if (flags & FOLL_GET)
87 return try_get_compound_head(page, refs);
88 else if (flags & FOLL_PIN) {
89 int orig_refs = refs;
90
91 /*
92 * Can't do FOLL_LONGTERM + FOLL_PIN with CMA in the gup fast
93 * path, so fail and let the caller fall back to the slow path.
94 */
95 if (unlikely(flags & FOLL_LONGTERM) &&
96 is_migrate_cma_page(page))
97 return NULL;
98
99 /*
100 * When pinning a compound page of order > 1 (which is what
101 * hpage_pincount_available() checks for), use an exact count to
102 * track it, via hpage_pincount_add/_sub().
103 *
104 * However, be sure to *also* increment the normal page refcount
105 * field at least once, so that the page really is pinned.
106 */
107 if (!hpage_pincount_available(page))
108 refs *= GUP_PIN_COUNTING_BIAS;
109
110 page = try_get_compound_head(page, refs);
111 if (!page)
112 return NULL;
113
114 if (hpage_pincount_available(page))
115 hpage_pincount_add(page, refs);
116
117 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED,
118 orig_refs);
119
120 return page;
121 }
122
123 WARN_ON_ONCE(1);
124 return NULL;
125 }
126
127 /**
128 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
129 *
130 * This might not do anything at all, depending on the flags argument.
131 *
132 * "grab" names in this file mean, "look at flags to decide whether to use
133 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
134 *
135 * @page: pointer to page to be grabbed
136 * @flags: gup flags: these are the FOLL_* flag values.
137 *
138 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
139 * time. Cases:
140 *
141 * FOLL_GET: page's refcount will be incremented by 1.
142 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
143 *
144 * Return: true for success, or if no action was required (if neither FOLL_PIN
145 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
146 * FOLL_PIN was set, but the page could not be grabbed.
147 */
148 bool __must_check try_grab_page(struct page *page, unsigned int flags)
149 {
150 WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == (FOLL_GET | FOLL_PIN));
151
152 if (flags & FOLL_GET)
153 return try_get_page(page);
154 else if (flags & FOLL_PIN) {
155 int refs = 1;
156
157 page = compound_head(page);
158
159 if (WARN_ON_ONCE(page_ref_count(page) <= 0))
160 return false;
161
162 if (hpage_pincount_available(page))
163 hpage_pincount_add(page, 1);
164 else
165 refs = GUP_PIN_COUNTING_BIAS;
166
167 /*
168 * Similar to try_grab_compound_head(): even if using the
169 * hpage_pincount_add/_sub() routines, be sure to
170 * *also* increment the normal page refcount field at least
171 * once, so that the page really is pinned.
172 */
173 page_ref_add(page, refs);
174
175 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED, 1);
176 }
177
178 return true;
179 }
180
181 #ifdef CONFIG_DEV_PAGEMAP_OPS
182 static bool __unpin_devmap_managed_user_page(struct page *page)
183 {
184 int count, refs = 1;
185
186 if (!page_is_devmap_managed(page))
187 return false;
188
189 if (hpage_pincount_available(page))
190 hpage_pincount_sub(page, 1);
191 else
192 refs = GUP_PIN_COUNTING_BIAS;
193
194 count = page_ref_sub_return(page, refs);
195
196 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED, 1);
197 /*
198 * devmap page refcounts are 1-based, rather than 0-based: if
199 * refcount is 1, then the page is free and the refcount is
200 * stable because nobody holds a reference on the page.
201 */
202 if (count == 1)
203 free_devmap_managed_page(page);
204 else if (!count)
205 __put_page(page);
206
207 return true;
208 }
209 #else
210 static bool __unpin_devmap_managed_user_page(struct page *page)
211 {
212 return false;
213 }
214 #endif /* CONFIG_DEV_PAGEMAP_OPS */
215
216 /**
217 * unpin_user_page() - release a dma-pinned page
218 * @page: pointer to page to be released
219 *
220 * Pages that were pinned via pin_user_pages*() must be released via either
221 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
222 * that such pages can be separately tracked and uniquely handled. In
223 * particular, interactions with RDMA and filesystems need special handling.
224 */
225 void unpin_user_page(struct page *page)
226 {
227 int refs = 1;
228
229 page = compound_head(page);
230
231 /*
232 * For devmap managed pages we need to catch refcount transition from
233 * GUP_PIN_COUNTING_BIAS to 1, when refcount reach one it means the
234 * page is free and we need to inform the device driver through
235 * callback. See include/linux/memremap.h and HMM for details.
236 */
237 if (__unpin_devmap_managed_user_page(page))
238 return;
239
240 if (hpage_pincount_available(page))
241 hpage_pincount_sub(page, 1);
242 else
243 refs = GUP_PIN_COUNTING_BIAS;
244
245 if (page_ref_sub_and_test(page, refs))
246 __put_page(page);
247
248 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED, 1);
249 }
250 EXPORT_SYMBOL(unpin_user_page);
251
252 /**
253 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
254 * @pages: array of pages to be maybe marked dirty, and definitely released.
255 * @npages: number of pages in the @pages array.
256 * @make_dirty: whether to mark the pages dirty
257 *
258 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
259 * variants called on that page.
260 *
261 * For each page in the @pages array, make that page (or its head page, if a
262 * compound page) dirty, if @make_dirty is true, and if the page was previously
263 * listed as clean. In any case, releases all pages using unpin_user_page(),
264 * possibly via unpin_user_pages(), for the non-dirty case.
265 *
266 * Please see the unpin_user_page() documentation for details.
267 *
268 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
269 * required, then the caller should a) verify that this is really correct,
270 * because _lock() is usually required, and b) hand code it:
271 * set_page_dirty_lock(), unpin_user_page().
272 *
273 */
274 void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
275 bool make_dirty)
276 {
277 unsigned long index;
278
279 /*
280 * TODO: this can be optimized for huge pages: if a series of pages is
281 * physically contiguous and part of the same compound page, then a
282 * single operation to the head page should suffice.
283 */
284
285 if (!make_dirty) {
286 unpin_user_pages(pages, npages);
287 return;
288 }
289
290 for (index = 0; index < npages; index++) {
291 struct page *page = compound_head(pages[index]);
292 /*
293 * Checking PageDirty at this point may race with
294 * clear_page_dirty_for_io(), but that's OK. Two key
295 * cases:
296 *
297 * 1) This code sees the page as already dirty, so it
298 * skips the call to set_page_dirty(). That could happen
299 * because clear_page_dirty_for_io() called
300 * page_mkclean(), followed by set_page_dirty().
301 * However, now the page is going to get written back,
302 * which meets the original intention of setting it
303 * dirty, so all is well: clear_page_dirty_for_io() goes
304 * on to call TestClearPageDirty(), and write the page
305 * back.
306 *
307 * 2) This code sees the page as clean, so it calls
308 * set_page_dirty(). The page stays dirty, despite being
309 * written back, so it gets written back again in the
310 * next writeback cycle. This is harmless.
311 */
312 if (!PageDirty(page))
313 set_page_dirty_lock(page);
314 unpin_user_page(page);
315 }
316 }
317 EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
318
319 /**
320 * unpin_user_pages() - release an array of gup-pinned pages.
321 * @pages: array of pages to be marked dirty and released.
322 * @npages: number of pages in the @pages array.
323 *
324 * For each page in the @pages array, release the page using unpin_user_page().
325 *
326 * Please see the unpin_user_page() documentation for details.
327 */
328 void unpin_user_pages(struct page **pages, unsigned long npages)
329 {
330 unsigned long index;
331
332 /*
333 * TODO: this can be optimized for huge pages: if a series of pages is
334 * physically contiguous and part of the same compound page, then a
335 * single operation to the head page should suffice.
336 */
337 for (index = 0; index < npages; index++)
338 unpin_user_page(pages[index]);
339 }
340 EXPORT_SYMBOL(unpin_user_pages);
341
342 #ifdef CONFIG_MMU
343 static struct page *no_page_table(struct vm_area_struct *vma,
344 unsigned int flags)
345 {
346 /*
347 * When core dumping an enormous anonymous area that nobody
348 * has touched so far, we don't want to allocate unnecessary pages or
349 * page tables. Return error instead of NULL to skip handle_mm_fault,
350 * then get_dump_page() will return NULL to leave a hole in the dump.
351 * But we can only make this optimization where a hole would surely
352 * be zero-filled if handle_mm_fault() actually did handle it.
353 */
354 if ((flags & FOLL_DUMP) &&
355 (vma_is_anonymous(vma) || !vma->vm_ops->fault))
356 return ERR_PTR(-EFAULT);
357 return NULL;
358 }
359
360 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
361 pte_t *pte, unsigned int flags)
362 {
363 /* No page to get reference */
364 if (flags & FOLL_GET)
365 return -EFAULT;
366
367 if (flags & FOLL_TOUCH) {
368 pte_t entry = *pte;
369
370 if (flags & FOLL_WRITE)
371 entry = pte_mkdirty(entry);
372 entry = pte_mkyoung(entry);
373
374 if (!pte_same(*pte, entry)) {
375 set_pte_at(vma->vm_mm, address, pte, entry);
376 update_mmu_cache(vma, address, pte);
377 }
378 }
379
380 /* Proper page table entry exists, but no corresponding struct page */
381 return -EEXIST;
382 }
383
384 /*
385 * FOLL_FORCE can write to even unwritable pte's, but only
386 * after we've gone through a COW cycle and they are dirty.
387 */
388 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
389 {
390 return pte_write(pte) ||
391 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
392 }
393
394 static struct page *follow_page_pte(struct vm_area_struct *vma,
395 unsigned long address, pmd_t *pmd, unsigned int flags,
396 struct dev_pagemap **pgmap)
397 {
398 struct mm_struct *mm = vma->vm_mm;
399 struct page *page;
400 spinlock_t *ptl;
401 pte_t *ptep, pte;
402 int ret;
403
404 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
405 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
406 (FOLL_PIN | FOLL_GET)))
407 return ERR_PTR(-EINVAL);
408 retry:
409 if (unlikely(pmd_bad(*pmd)))
410 return no_page_table(vma, flags);
411
412 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
413 pte = *ptep;
414 if (!pte_present(pte)) {
415 swp_entry_t entry;
416 /*
417 * KSM's break_ksm() relies upon recognizing a ksm page
418 * even while it is being migrated, so for that case we
419 * need migration_entry_wait().
420 */
421 if (likely(!(flags & FOLL_MIGRATION)))
422 goto no_page;
423 if (pte_none(pte))
424 goto no_page;
425 entry = pte_to_swp_entry(pte);
426 if (!is_migration_entry(entry))
427 goto no_page;
428 pte_unmap_unlock(ptep, ptl);
429 migration_entry_wait(mm, pmd, address);
430 goto retry;
431 }
432 if ((flags & FOLL_NUMA) && pte_protnone(pte))
433 goto no_page;
434 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
435 pte_unmap_unlock(ptep, ptl);
436 return NULL;
437 }
438
439 page = vm_normal_page(vma, address, pte);
440 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
441 /*
442 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
443 * case since they are only valid while holding the pgmap
444 * reference.
445 */
446 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
447 if (*pgmap)
448 page = pte_page(pte);
449 else
450 goto no_page;
451 } else if (unlikely(!page)) {
452 if (flags & FOLL_DUMP) {
453 /* Avoid special (like zero) pages in core dumps */
454 page = ERR_PTR(-EFAULT);
455 goto out;
456 }
457
458 if (is_zero_pfn(pte_pfn(pte))) {
459 page = pte_page(pte);
460 } else {
461 ret = follow_pfn_pte(vma, address, ptep, flags);
462 page = ERR_PTR(ret);
463 goto out;
464 }
465 }
466
467 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
468 get_page(page);
469 pte_unmap_unlock(ptep, ptl);
470 lock_page(page);
471 ret = split_huge_page(page);
472 unlock_page(page);
473 put_page(page);
474 if (ret)
475 return ERR_PTR(ret);
476 goto retry;
477 }
478
479 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
480 if (unlikely(!try_grab_page(page, flags))) {
481 page = ERR_PTR(-ENOMEM);
482 goto out;
483 }
484 /*
485 * We need to make the page accessible if and only if we are going
486 * to access its content (the FOLL_PIN case). Please see
487 * Documentation/core-api/pin_user_pages.rst for details.
488 */
489 if (flags & FOLL_PIN) {
490 ret = arch_make_page_accessible(page);
491 if (ret) {
492 unpin_user_page(page);
493 page = ERR_PTR(ret);
494 goto out;
495 }
496 }
497 if (flags & FOLL_TOUCH) {
498 if ((flags & FOLL_WRITE) &&
499 !pte_dirty(pte) && !PageDirty(page))
500 set_page_dirty(page);
501 /*
502 * pte_mkyoung() would be more correct here, but atomic care
503 * is needed to avoid losing the dirty bit: it is easier to use
504 * mark_page_accessed().
505 */
506 mark_page_accessed(page);
507 }
508 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
509 /* Do not mlock pte-mapped THP */
510 if (PageTransCompound(page))
511 goto out;
512
513 /*
514 * The preliminary mapping check is mainly to avoid the
515 * pointless overhead of lock_page on the ZERO_PAGE
516 * which might bounce very badly if there is contention.
517 *
518 * If the page is already locked, we don't need to
519 * handle it now - vmscan will handle it later if and
520 * when it attempts to reclaim the page.
521 */
522 if (page->mapping && trylock_page(page)) {
523 lru_add_drain(); /* push cached pages to LRU */
524 /*
525 * Because we lock page here, and migration is
526 * blocked by the pte's page reference, and we
527 * know the page is still mapped, we don't even
528 * need to check for file-cache page truncation.
529 */
530 mlock_vma_page(page);
531 unlock_page(page);
532 }
533 }
534 out:
535 pte_unmap_unlock(ptep, ptl);
536 return page;
537 no_page:
538 pte_unmap_unlock(ptep, ptl);
539 if (!pte_none(pte))
540 return NULL;
541 return no_page_table(vma, flags);
542 }
543
544 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
545 unsigned long address, pud_t *pudp,
546 unsigned int flags,
547 struct follow_page_context *ctx)
548 {
549 pmd_t *pmd, pmdval;
550 spinlock_t *ptl;
551 struct page *page;
552 struct mm_struct *mm = vma->vm_mm;
553
554 pmd = pmd_offset(pudp, address);
555 /*
556 * The READ_ONCE() will stabilize the pmdval in a register or
557 * on the stack so that it will stop changing under the code.
558 */
559 pmdval = READ_ONCE(*pmd);
560 if (pmd_none(pmdval))
561 return no_page_table(vma, flags);
562 if (pmd_huge(pmdval) && is_vm_hugetlb_page(vma)) {
563 page = follow_huge_pmd(mm, address, pmd, flags);
564 if (page)
565 return page;
566 return no_page_table(vma, flags);
567 }
568 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
569 page = follow_huge_pd(vma, address,
570 __hugepd(pmd_val(pmdval)), flags,
571 PMD_SHIFT);
572 if (page)
573 return page;
574 return no_page_table(vma, flags);
575 }
576 retry:
577 if (!pmd_present(pmdval)) {
578 if (likely(!(flags & FOLL_MIGRATION)))
579 return no_page_table(vma, flags);
580 VM_BUG_ON(thp_migration_supported() &&
581 !is_pmd_migration_entry(pmdval));
582 if (is_pmd_migration_entry(pmdval))
583 pmd_migration_entry_wait(mm, pmd);
584 pmdval = READ_ONCE(*pmd);
585 /*
586 * MADV_DONTNEED may convert the pmd to null because
587 * mmap_sem is held in read mode
588 */
589 if (pmd_none(pmdval))
590 return no_page_table(vma, flags);
591 goto retry;
592 }
593 if (pmd_devmap(pmdval)) {
594 ptl = pmd_lock(mm, pmd);
595 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
596 spin_unlock(ptl);
597 if (page)
598 return page;
599 }
600 if (likely(!pmd_trans_huge(pmdval)))
601 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
602
603 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
604 return no_page_table(vma, flags);
605
606 retry_locked:
607 ptl = pmd_lock(mm, pmd);
608 if (unlikely(pmd_none(*pmd))) {
609 spin_unlock(ptl);
610 return no_page_table(vma, flags);
611 }
612 if (unlikely(!pmd_present(*pmd))) {
613 spin_unlock(ptl);
614 if (likely(!(flags & FOLL_MIGRATION)))
615 return no_page_table(vma, flags);
616 pmd_migration_entry_wait(mm, pmd);
617 goto retry_locked;
618 }
619 if (unlikely(!pmd_trans_huge(*pmd))) {
620 spin_unlock(ptl);
621 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
622 }
623 if (flags & (FOLL_SPLIT | FOLL_SPLIT_PMD)) {
624 int ret;
625 page = pmd_page(*pmd);
626 if (is_huge_zero_page(page)) {
627 spin_unlock(ptl);
628 ret = 0;
629 split_huge_pmd(vma, pmd, address);
630 if (pmd_trans_unstable(pmd))
631 ret = -EBUSY;
632 } else if (flags & FOLL_SPLIT) {
633 if (unlikely(!try_get_page(page))) {
634 spin_unlock(ptl);
635 return ERR_PTR(-ENOMEM);
636 }
637 spin_unlock(ptl);
638 lock_page(page);
639 ret = split_huge_page(page);
640 unlock_page(page);
641 put_page(page);
642 if (pmd_none(*pmd))
643 return no_page_table(vma, flags);
644 } else { /* flags & FOLL_SPLIT_PMD */
645 spin_unlock(ptl);
646 split_huge_pmd(vma, pmd, address);
647 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
648 }
649
650 return ret ? ERR_PTR(ret) :
651 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
652 }
653 page = follow_trans_huge_pmd(vma, address, pmd, flags);
654 spin_unlock(ptl);
655 ctx->page_mask = HPAGE_PMD_NR - 1;
656 return page;
657 }
658
659 static struct page *follow_pud_mask(struct vm_area_struct *vma,
660 unsigned long address, p4d_t *p4dp,
661 unsigned int flags,
662 struct follow_page_context *ctx)
663 {
664 pud_t *pud;
665 spinlock_t *ptl;
666 struct page *page;
667 struct mm_struct *mm = vma->vm_mm;
668
669 pud = pud_offset(p4dp, address);
670 if (pud_none(*pud))
671 return no_page_table(vma, flags);
672 if (pud_huge(*pud) && is_vm_hugetlb_page(vma)) {
673 page = follow_huge_pud(mm, address, pud, flags);
674 if (page)
675 return page;
676 return no_page_table(vma, flags);
677 }
678 if (is_hugepd(__hugepd(pud_val(*pud)))) {
679 page = follow_huge_pd(vma, address,
680 __hugepd(pud_val(*pud)), flags,
681 PUD_SHIFT);
682 if (page)
683 return page;
684 return no_page_table(vma, flags);
685 }
686 if (pud_devmap(*pud)) {
687 ptl = pud_lock(mm, pud);
688 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
689 spin_unlock(ptl);
690 if (page)
691 return page;
692 }
693 if (unlikely(pud_bad(*pud)))
694 return no_page_table(vma, flags);
695
696 return follow_pmd_mask(vma, address, pud, flags, ctx);
697 }
698
699 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
700 unsigned long address, pgd_t *pgdp,
701 unsigned int flags,
702 struct follow_page_context *ctx)
703 {
704 p4d_t *p4d;
705 struct page *page;
706
707 p4d = p4d_offset(pgdp, address);
708 if (p4d_none(*p4d))
709 return no_page_table(vma, flags);
710 BUILD_BUG_ON(p4d_huge(*p4d));
711 if (unlikely(p4d_bad(*p4d)))
712 return no_page_table(vma, flags);
713
714 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
715 page = follow_huge_pd(vma, address,
716 __hugepd(p4d_val(*p4d)), flags,
717 P4D_SHIFT);
718 if (page)
719 return page;
720 return no_page_table(vma, flags);
721 }
722 return follow_pud_mask(vma, address, p4d, flags, ctx);
723 }
724
725 /**
726 * follow_page_mask - look up a page descriptor from a user-virtual address
727 * @vma: vm_area_struct mapping @address
728 * @address: virtual address to look up
729 * @flags: flags modifying lookup behaviour
730 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
731 * pointer to output page_mask
732 *
733 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
734 *
735 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
736 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
737 *
738 * On output, the @ctx->page_mask is set according to the size of the page.
739 *
740 * Return: the mapped (struct page *), %NULL if no mapping exists, or
741 * an error pointer if there is a mapping to something not represented
742 * by a page descriptor (see also vm_normal_page()).
743 */
744 static struct page *follow_page_mask(struct vm_area_struct *vma,
745 unsigned long address, unsigned int flags,
746 struct follow_page_context *ctx)
747 {
748 pgd_t *pgd;
749 struct page *page;
750 struct mm_struct *mm = vma->vm_mm;
751
752 ctx->page_mask = 0;
753
754 /* make this handle hugepd */
755 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
756 if (!IS_ERR(page)) {
757 WARN_ON_ONCE(flags & (FOLL_GET | FOLL_PIN));
758 return page;
759 }
760
761 pgd = pgd_offset(mm, address);
762
763 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
764 return no_page_table(vma, flags);
765
766 if (pgd_huge(*pgd)) {
767 page = follow_huge_pgd(mm, address, pgd, flags);
768 if (page)
769 return page;
770 return no_page_table(vma, flags);
771 }
772 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
773 page = follow_huge_pd(vma, address,
774 __hugepd(pgd_val(*pgd)), flags,
775 PGDIR_SHIFT);
776 if (page)
777 return page;
778 return no_page_table(vma, flags);
779 }
780
781 return follow_p4d_mask(vma, address, pgd, flags, ctx);
782 }
783
784 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
785 unsigned int foll_flags)
786 {
787 struct follow_page_context ctx = { NULL };
788 struct page *page;
789
790 page = follow_page_mask(vma, address, foll_flags, &ctx);
791 if (ctx.pgmap)
792 put_dev_pagemap(ctx.pgmap);
793 return page;
794 }
795
796 static int get_gate_page(struct mm_struct *mm, unsigned long address,
797 unsigned int gup_flags, struct vm_area_struct **vma,
798 struct page **page)
799 {
800 pgd_t *pgd;
801 p4d_t *p4d;
802 pud_t *pud;
803 pmd_t *pmd;
804 pte_t *pte;
805 int ret = -EFAULT;
806
807 /* user gate pages are read-only */
808 if (gup_flags & FOLL_WRITE)
809 return -EFAULT;
810 if (address > TASK_SIZE)
811 pgd = pgd_offset_k(address);
812 else
813 pgd = pgd_offset_gate(mm, address);
814 if (pgd_none(*pgd))
815 return -EFAULT;
816 p4d = p4d_offset(pgd, address);
817 if (p4d_none(*p4d))
818 return -EFAULT;
819 pud = pud_offset(p4d, address);
820 if (pud_none(*pud))
821 return -EFAULT;
822 pmd = pmd_offset(pud, address);
823 if (!pmd_present(*pmd))
824 return -EFAULT;
825 VM_BUG_ON(pmd_trans_huge(*pmd));
826 pte = pte_offset_map(pmd, address);
827 if (pte_none(*pte))
828 goto unmap;
829 *vma = get_gate_vma(mm);
830 if (!page)
831 goto out;
832 *page = vm_normal_page(*vma, address, *pte);
833 if (!*page) {
834 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
835 goto unmap;
836 *page = pte_page(*pte);
837 }
838 if (unlikely(!try_get_page(*page))) {
839 ret = -ENOMEM;
840 goto unmap;
841 }
842 out:
843 ret = 0;
844 unmap:
845 pte_unmap(pte);
846 return ret;
847 }
848
849 /*
850 * mmap_sem must be held on entry. If @locked != NULL and *@flags
851 * does not include FOLL_NOWAIT, the mmap_sem may be released. If it
852 * is, *@locked will be set to 0 and -EBUSY returned.
853 */
854 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
855 unsigned long address, unsigned int *flags, int *locked)
856 {
857 unsigned int fault_flags = 0;
858 vm_fault_t ret;
859
860 /* mlock all present pages, but do not fault in new pages */
861 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
862 return -ENOENT;
863 if (*flags & FOLL_WRITE)
864 fault_flags |= FAULT_FLAG_WRITE;
865 if (*flags & FOLL_REMOTE)
866 fault_flags |= FAULT_FLAG_REMOTE;
867 if (locked)
868 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
869 if (*flags & FOLL_NOWAIT)
870 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
871 if (*flags & FOLL_TRIED) {
872 /*
873 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
874 * can co-exist
875 */
876 fault_flags |= FAULT_FLAG_TRIED;
877 }
878
879 ret = handle_mm_fault(vma, address, fault_flags);
880 if (ret & VM_FAULT_ERROR) {
881 int err = vm_fault_to_errno(ret, *flags);
882
883 if (err)
884 return err;
885 BUG();
886 }
887
888 if (tsk) {
889 if (ret & VM_FAULT_MAJOR)
890 tsk->maj_flt++;
891 else
892 tsk->min_flt++;
893 }
894
895 if (ret & VM_FAULT_RETRY) {
896 if (locked && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
897 *locked = 0;
898 return -EBUSY;
899 }
900
901 /*
902 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
903 * necessary, even if maybe_mkwrite decided not to set pte_write. We
904 * can thus safely do subsequent page lookups as if they were reads.
905 * But only do so when looping for pte_write is futile: in some cases
906 * userspace may also be wanting to write to the gotten user page,
907 * which a read fault here might prevent (a readonly page might get
908 * reCOWed by userspace write).
909 */
910 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
911 *flags |= FOLL_COW;
912 return 0;
913 }
914
915 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
916 {
917 vm_flags_t vm_flags = vma->vm_flags;
918 int write = (gup_flags & FOLL_WRITE);
919 int foreign = (gup_flags & FOLL_REMOTE);
920
921 if (vm_flags & (VM_IO | VM_PFNMAP))
922 return -EFAULT;
923
924 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
925 return -EFAULT;
926
927 if (write) {
928 if (!(vm_flags & VM_WRITE)) {
929 if (!(gup_flags & FOLL_FORCE))
930 return -EFAULT;
931 /*
932 * We used to let the write,force case do COW in a
933 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
934 * set a breakpoint in a read-only mapping of an
935 * executable, without corrupting the file (yet only
936 * when that file had been opened for writing!).
937 * Anon pages in shared mappings are surprising: now
938 * just reject it.
939 */
940 if (!is_cow_mapping(vm_flags))
941 return -EFAULT;
942 }
943 } else if (!(vm_flags & VM_READ)) {
944 if (!(gup_flags & FOLL_FORCE))
945 return -EFAULT;
946 /*
947 * Is there actually any vma we can reach here which does not
948 * have VM_MAYREAD set?
949 */
950 if (!(vm_flags & VM_MAYREAD))
951 return -EFAULT;
952 }
953 /*
954 * gups are always data accesses, not instruction
955 * fetches, so execute=false here
956 */
957 if (!arch_vma_access_permitted(vma, write, false, foreign))
958 return -EFAULT;
959 return 0;
960 }
961
962 /**
963 * __get_user_pages() - pin user pages in memory
964 * @tsk: task_struct of target task
965 * @mm: mm_struct of target mm
966 * @start: starting user address
967 * @nr_pages: number of pages from start to pin
968 * @gup_flags: flags modifying pin behaviour
969 * @pages: array that receives pointers to the pages pinned.
970 * Should be at least nr_pages long. Or NULL, if caller
971 * only intends to ensure the pages are faulted in.
972 * @vmas: array of pointers to vmas corresponding to each page.
973 * Or NULL if the caller does not require them.
974 * @locked: whether we're still with the mmap_sem held
975 *
976 * Returns either number of pages pinned (which may be less than the
977 * number requested), or an error. Details about the return value:
978 *
979 * -- If nr_pages is 0, returns 0.
980 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
981 * -- If nr_pages is >0, and some pages were pinned, returns the number of
982 * pages pinned. Again, this may be less than nr_pages.
983 *
984 * The caller is responsible for releasing returned @pages, via put_page().
985 *
986 * @vmas are valid only as long as mmap_sem is held.
987 *
988 * Must be called with mmap_sem held. It may be released. See below.
989 *
990 * __get_user_pages walks a process's page tables and takes a reference to
991 * each struct page that each user address corresponds to at a given
992 * instant. That is, it takes the page that would be accessed if a user
993 * thread accesses the given user virtual address at that instant.
994 *
995 * This does not guarantee that the page exists in the user mappings when
996 * __get_user_pages returns, and there may even be a completely different
997 * page there in some cases (eg. if mmapped pagecache has been invalidated
998 * and subsequently re faulted). However it does guarantee that the page
999 * won't be freed completely. And mostly callers simply care that the page
1000 * contains data that was valid *at some point in time*. Typically, an IO
1001 * or similar operation cannot guarantee anything stronger anyway because
1002 * locks can't be held over the syscall boundary.
1003 *
1004 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1005 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1006 * appropriate) must be called after the page is finished with, and
1007 * before put_page is called.
1008 *
1009 * If @locked != NULL, *@locked will be set to 0 when mmap_sem is
1010 * released by an up_read(). That can happen if @gup_flags does not
1011 * have FOLL_NOWAIT.
1012 *
1013 * A caller using such a combination of @locked and @gup_flags
1014 * must therefore hold the mmap_sem for reading only, and recognize
1015 * when it's been released. Otherwise, it must be held for either
1016 * reading or writing and will not be released.
1017 *
1018 * In most cases, get_user_pages or get_user_pages_fast should be used
1019 * instead of __get_user_pages. __get_user_pages should be used only if
1020 * you need some special @gup_flags.
1021 */
1022 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1023 unsigned long start, unsigned long nr_pages,
1024 unsigned int gup_flags, struct page **pages,
1025 struct vm_area_struct **vmas, int *locked)
1026 {
1027 long ret = 0, i = 0;
1028 struct vm_area_struct *vma = NULL;
1029 struct follow_page_context ctx = { NULL };
1030
1031 if (!nr_pages)
1032 return 0;
1033
1034 start = untagged_addr(start);
1035
1036 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1037
1038 /*
1039 * If FOLL_FORCE is set then do not force a full fault as the hinting
1040 * fault information is unrelated to the reference behaviour of a task
1041 * using the address space
1042 */
1043 if (!(gup_flags & FOLL_FORCE))
1044 gup_flags |= FOLL_NUMA;
1045
1046 do {
1047 struct page *page;
1048 unsigned int foll_flags = gup_flags;
1049 unsigned int page_increm;
1050
1051 /* first iteration or cross vma bound */
1052 if (!vma || start >= vma->vm_end) {
1053 vma = find_extend_vma(mm, start);
1054 if (!vma && in_gate_area(mm, start)) {
1055 ret = get_gate_page(mm, start & PAGE_MASK,
1056 gup_flags, &vma,
1057 pages ? &pages[i] : NULL);
1058 if (ret)
1059 goto out;
1060 ctx.page_mask = 0;
1061 goto next_page;
1062 }
1063
1064 if (!vma || check_vma_flags(vma, gup_flags)) {
1065 ret = -EFAULT;
1066 goto out;
1067 }
1068 if (is_vm_hugetlb_page(vma)) {
1069 i = follow_hugetlb_page(mm, vma, pages, vmas,
1070 &start, &nr_pages, i,
1071 gup_flags, locked);
1072 if (locked && *locked == 0) {
1073 /*
1074 * We've got a VM_FAULT_RETRY
1075 * and we've lost mmap_sem.
1076 * We must stop here.
1077 */
1078 BUG_ON(gup_flags & FOLL_NOWAIT);
1079 BUG_ON(ret != 0);
1080 goto out;
1081 }
1082 continue;
1083 }
1084 }
1085 retry:
1086 /*
1087 * If we have a pending SIGKILL, don't keep faulting pages and
1088 * potentially allocating memory.
1089 */
1090 if (fatal_signal_pending(current)) {
1091 ret = -EINTR;
1092 goto out;
1093 }
1094 cond_resched();
1095
1096 page = follow_page_mask(vma, start, foll_flags, &ctx);
1097 if (!page) {
1098 ret = faultin_page(tsk, vma, start, &foll_flags,
1099 locked);
1100 switch (ret) {
1101 case 0:
1102 goto retry;
1103 case -EBUSY:
1104 ret = 0;
1105 fallthrough;
1106 case -EFAULT:
1107 case -ENOMEM:
1108 case -EHWPOISON:
1109 goto out;
1110 case -ENOENT:
1111 goto next_page;
1112 }
1113 BUG();
1114 } else if (PTR_ERR(page) == -EEXIST) {
1115 /*
1116 * Proper page table entry exists, but no corresponding
1117 * struct page.
1118 */
1119 goto next_page;
1120 } else if (IS_ERR(page)) {
1121 ret = PTR_ERR(page);
1122 goto out;
1123 }
1124 if (pages) {
1125 pages[i] = page;
1126 flush_anon_page(vma, page, start);
1127 flush_dcache_page(page);
1128 ctx.page_mask = 0;
1129 }
1130 next_page:
1131 if (vmas) {
1132 vmas[i] = vma;
1133 ctx.page_mask = 0;
1134 }
1135 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1136 if (page_increm > nr_pages)
1137 page_increm = nr_pages;
1138 i += page_increm;
1139 start += page_increm * PAGE_SIZE;
1140 nr_pages -= page_increm;
1141 } while (nr_pages);
1142 out:
1143 if (ctx.pgmap)
1144 put_dev_pagemap(ctx.pgmap);
1145 return i ? i : ret;
1146 }
1147
1148 static bool vma_permits_fault(struct vm_area_struct *vma,
1149 unsigned int fault_flags)
1150 {
1151 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1152 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1153 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1154
1155 if (!(vm_flags & vma->vm_flags))
1156 return false;
1157
1158 /*
1159 * The architecture might have a hardware protection
1160 * mechanism other than read/write that can deny access.
1161 *
1162 * gup always represents data access, not instruction
1163 * fetches, so execute=false here:
1164 */
1165 if (!arch_vma_access_permitted(vma, write, false, foreign))
1166 return false;
1167
1168 return true;
1169 }
1170
1171 /*
1172 * fixup_user_fault() - manually resolve a user page fault
1173 * @tsk: the task_struct to use for page fault accounting, or
1174 * NULL if faults are not to be recorded.
1175 * @mm: mm_struct of target mm
1176 * @address: user address
1177 * @fault_flags:flags to pass down to handle_mm_fault()
1178 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
1179 * does not allow retry
1180 *
1181 * This is meant to be called in the specific scenario where for locking reasons
1182 * we try to access user memory in atomic context (within a pagefault_disable()
1183 * section), this returns -EFAULT, and we want to resolve the user fault before
1184 * trying again.
1185 *
1186 * Typically this is meant to be used by the futex code.
1187 *
1188 * The main difference with get_user_pages() is that this function will
1189 * unconditionally call handle_mm_fault() which will in turn perform all the
1190 * necessary SW fixup of the dirty and young bits in the PTE, while
1191 * get_user_pages() only guarantees to update these in the struct page.
1192 *
1193 * This is important for some architectures where those bits also gate the
1194 * access permission to the page because they are maintained in software. On
1195 * such architectures, gup() will not be enough to make a subsequent access
1196 * succeed.
1197 *
1198 * This function will not return with an unlocked mmap_sem. So it has not the
1199 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
1200 */
1201 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1202 unsigned long address, unsigned int fault_flags,
1203 bool *unlocked)
1204 {
1205 struct vm_area_struct *vma;
1206 vm_fault_t ret, major = 0;
1207
1208 address = untagged_addr(address);
1209
1210 if (unlocked)
1211 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1212
1213 retry:
1214 vma = find_extend_vma(mm, address);
1215 if (!vma || address < vma->vm_start)
1216 return -EFAULT;
1217
1218 if (!vma_permits_fault(vma, fault_flags))
1219 return -EFAULT;
1220
1221 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1222 fatal_signal_pending(current))
1223 return -EINTR;
1224
1225 ret = handle_mm_fault(vma, address, fault_flags);
1226 major |= ret & VM_FAULT_MAJOR;
1227 if (ret & VM_FAULT_ERROR) {
1228 int err = vm_fault_to_errno(ret, 0);
1229
1230 if (err)
1231 return err;
1232 BUG();
1233 }
1234
1235 if (ret & VM_FAULT_RETRY) {
1236 down_read(&mm->mmap_sem);
1237 *unlocked = true;
1238 fault_flags |= FAULT_FLAG_TRIED;
1239 goto retry;
1240 }
1241
1242 if (tsk) {
1243 if (major)
1244 tsk->maj_flt++;
1245 else
1246 tsk->min_flt++;
1247 }
1248 return 0;
1249 }
1250 EXPORT_SYMBOL_GPL(fixup_user_fault);
1251
1252 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
1253 struct mm_struct *mm,
1254 unsigned long start,
1255 unsigned long nr_pages,
1256 struct page **pages,
1257 struct vm_area_struct **vmas,
1258 int *locked,
1259 unsigned int flags)
1260 {
1261 long ret, pages_done;
1262 bool lock_dropped;
1263
1264 if (locked) {
1265 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1266 BUG_ON(vmas);
1267 /* check caller initialized locked */
1268 BUG_ON(*locked != 1);
1269 }
1270
1271 /*
1272 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1273 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1274 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1275 * for FOLL_GET, not for the newer FOLL_PIN.
1276 *
1277 * FOLL_PIN always expects pages to be non-null, but no need to assert
1278 * that here, as any failures will be obvious enough.
1279 */
1280 if (pages && !(flags & FOLL_PIN))
1281 flags |= FOLL_GET;
1282
1283 pages_done = 0;
1284 lock_dropped = false;
1285 for (;;) {
1286 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
1287 vmas, locked);
1288 if (!locked)
1289 /* VM_FAULT_RETRY couldn't trigger, bypass */
1290 return ret;
1291
1292 /* VM_FAULT_RETRY cannot return errors */
1293 if (!*locked) {
1294 BUG_ON(ret < 0);
1295 BUG_ON(ret >= nr_pages);
1296 }
1297
1298 if (ret > 0) {
1299 nr_pages -= ret;
1300 pages_done += ret;
1301 if (!nr_pages)
1302 break;
1303 }
1304 if (*locked) {
1305 /*
1306 * VM_FAULT_RETRY didn't trigger or it was a
1307 * FOLL_NOWAIT.
1308 */
1309 if (!pages_done)
1310 pages_done = ret;
1311 break;
1312 }
1313 /*
1314 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1315 * For the prefault case (!pages) we only update counts.
1316 */
1317 if (likely(pages))
1318 pages += ret;
1319 start += ret << PAGE_SHIFT;
1320 lock_dropped = true;
1321
1322 retry:
1323 /*
1324 * Repeat on the address that fired VM_FAULT_RETRY
1325 * with both FAULT_FLAG_ALLOW_RETRY and
1326 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1327 * by fatal signals, so we need to check it before we
1328 * start trying again otherwise it can loop forever.
1329 */
1330
1331 if (fatal_signal_pending(current)) {
1332 if (!pages_done)
1333 pages_done = -EINTR;
1334 break;
1335 }
1336
1337 ret = down_read_killable(&mm->mmap_sem);
1338 if (ret) {
1339 BUG_ON(ret > 0);
1340 if (!pages_done)
1341 pages_done = ret;
1342 break;
1343 }
1344
1345 *locked = 1;
1346 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
1347 pages, NULL, locked);
1348 if (!*locked) {
1349 /* Continue to retry until we succeeded */
1350 BUG_ON(ret != 0);
1351 goto retry;
1352 }
1353 if (ret != 1) {
1354 BUG_ON(ret > 1);
1355 if (!pages_done)
1356 pages_done = ret;
1357 break;
1358 }
1359 nr_pages--;
1360 pages_done++;
1361 if (!nr_pages)
1362 break;
1363 if (likely(pages))
1364 pages++;
1365 start += PAGE_SIZE;
1366 }
1367 if (lock_dropped && *locked) {
1368 /*
1369 * We must let the caller know we temporarily dropped the lock
1370 * and so the critical section protected by it was lost.
1371 */
1372 up_read(&mm->mmap_sem);
1373 *locked = 0;
1374 }
1375 return pages_done;
1376 }
1377
1378 /**
1379 * populate_vma_page_range() - populate a range of pages in the vma.
1380 * @vma: target vma
1381 * @start: start address
1382 * @end: end address
1383 * @locked: whether the mmap_sem is still held
1384 *
1385 * This takes care of mlocking the pages too if VM_LOCKED is set.
1386 *
1387 * return 0 on success, negative error code on error.
1388 *
1389 * vma->vm_mm->mmap_sem must be held.
1390 *
1391 * If @locked is NULL, it may be held for read or write and will
1392 * be unperturbed.
1393 *
1394 * If @locked is non-NULL, it must held for read only and may be
1395 * released. If it's released, *@locked will be set to 0.
1396 */
1397 long populate_vma_page_range(struct vm_area_struct *vma,
1398 unsigned long start, unsigned long end, int *locked)
1399 {
1400 struct mm_struct *mm = vma->vm_mm;
1401 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1402 int gup_flags;
1403
1404 VM_BUG_ON(start & ~PAGE_MASK);
1405 VM_BUG_ON(end & ~PAGE_MASK);
1406 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1407 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1408 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1409
1410 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1411 if (vma->vm_flags & VM_LOCKONFAULT)
1412 gup_flags &= ~FOLL_POPULATE;
1413 /*
1414 * We want to touch writable mappings with a write fault in order
1415 * to break COW, except for shared mappings because these don't COW
1416 * and we would not want to dirty them for nothing.
1417 */
1418 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1419 gup_flags |= FOLL_WRITE;
1420
1421 /*
1422 * We want mlock to succeed for regions that have any permissions
1423 * other than PROT_NONE.
1424 */
1425 if (vma_is_accessible(vma))
1426 gup_flags |= FOLL_FORCE;
1427
1428 /*
1429 * We made sure addr is within a VMA, so the following will
1430 * not result in a stack expansion that recurses back here.
1431 */
1432 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1433 NULL, NULL, locked);
1434 }
1435
1436 /*
1437 * __mm_populate - populate and/or mlock pages within a range of address space.
1438 *
1439 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1440 * flags. VMAs must be already marked with the desired vm_flags, and
1441 * mmap_sem must not be held.
1442 */
1443 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1444 {
1445 struct mm_struct *mm = current->mm;
1446 unsigned long end, nstart, nend;
1447 struct vm_area_struct *vma = NULL;
1448 int locked = 0;
1449 long ret = 0;
1450
1451 end = start + len;
1452
1453 for (nstart = start; nstart < end; nstart = nend) {
1454 /*
1455 * We want to fault in pages for [nstart; end) address range.
1456 * Find first corresponding VMA.
1457 */
1458 if (!locked) {
1459 locked = 1;
1460 down_read(&mm->mmap_sem);
1461 vma = find_vma(mm, nstart);
1462 } else if (nstart >= vma->vm_end)
1463 vma = vma->vm_next;
1464 if (!vma || vma->vm_start >= end)
1465 break;
1466 /*
1467 * Set [nstart; nend) to intersection of desired address
1468 * range with the first VMA. Also, skip undesirable VMA types.
1469 */
1470 nend = min(end, vma->vm_end);
1471 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1472 continue;
1473 if (nstart < vma->vm_start)
1474 nstart = vma->vm_start;
1475 /*
1476 * Now fault in a range of pages. populate_vma_page_range()
1477 * double checks the vma flags, so that it won't mlock pages
1478 * if the vma was already munlocked.
1479 */
1480 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1481 if (ret < 0) {
1482 if (ignore_errors) {
1483 ret = 0;
1484 continue; /* continue at next VMA */
1485 }
1486 break;
1487 }
1488 nend = nstart + ret * PAGE_SIZE;
1489 ret = 0;
1490 }
1491 if (locked)
1492 up_read(&mm->mmap_sem);
1493 return ret; /* 0 or negative error code */
1494 }
1495
1496 /**
1497 * get_dump_page() - pin user page in memory while writing it to core dump
1498 * @addr: user address
1499 *
1500 * Returns struct page pointer of user page pinned for dump,
1501 * to be freed afterwards by put_page().
1502 *
1503 * Returns NULL on any kind of failure - a hole must then be inserted into
1504 * the corefile, to preserve alignment with its headers; and also returns
1505 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1506 * allowing a hole to be left in the corefile to save diskspace.
1507 *
1508 * Called without mmap_sem, but after all other threads have been killed.
1509 */
1510 #ifdef CONFIG_ELF_CORE
1511 struct page *get_dump_page(unsigned long addr)
1512 {
1513 struct vm_area_struct *vma;
1514 struct page *page;
1515
1516 if (__get_user_pages(current, current->mm, addr, 1,
1517 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1518 NULL) < 1)
1519 return NULL;
1520 flush_cache_page(vma, addr, page_to_pfn(page));
1521 return page;
1522 }
1523 #endif /* CONFIG_ELF_CORE */
1524 #else /* CONFIG_MMU */
1525 static long __get_user_pages_locked(struct task_struct *tsk,
1526 struct mm_struct *mm, unsigned long start,
1527 unsigned long nr_pages, struct page **pages,
1528 struct vm_area_struct **vmas, int *locked,
1529 unsigned int foll_flags)
1530 {
1531 struct vm_area_struct *vma;
1532 unsigned long vm_flags;
1533 int i;
1534
1535 /* calculate required read or write permissions.
1536 * If FOLL_FORCE is set, we only require the "MAY" flags.
1537 */
1538 vm_flags = (foll_flags & FOLL_WRITE) ?
1539 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1540 vm_flags &= (foll_flags & FOLL_FORCE) ?
1541 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1542
1543 for (i = 0; i < nr_pages; i++) {
1544 vma = find_vma(mm, start);
1545 if (!vma)
1546 goto finish_or_fault;
1547
1548 /* protect what we can, including chardevs */
1549 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1550 !(vm_flags & vma->vm_flags))
1551 goto finish_or_fault;
1552
1553 if (pages) {
1554 pages[i] = virt_to_page(start);
1555 if (pages[i])
1556 get_page(pages[i]);
1557 }
1558 if (vmas)
1559 vmas[i] = vma;
1560 start = (start + PAGE_SIZE) & PAGE_MASK;
1561 }
1562
1563 return i;
1564
1565 finish_or_fault:
1566 return i ? : -EFAULT;
1567 }
1568 #endif /* !CONFIG_MMU */
1569
1570 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1571 static bool check_dax_vmas(struct vm_area_struct **vmas, long nr_pages)
1572 {
1573 long i;
1574 struct vm_area_struct *vma_prev = NULL;
1575
1576 for (i = 0; i < nr_pages; i++) {
1577 struct vm_area_struct *vma = vmas[i];
1578
1579 if (vma == vma_prev)
1580 continue;
1581
1582 vma_prev = vma;
1583
1584 if (vma_is_fsdax(vma))
1585 return true;
1586 }
1587 return false;
1588 }
1589
1590 #ifdef CONFIG_CMA
1591 static struct page *new_non_cma_page(struct page *page, unsigned long private)
1592 {
1593 /*
1594 * We want to make sure we allocate the new page from the same node
1595 * as the source page.
1596 */
1597 int nid = page_to_nid(page);
1598 /*
1599 * Trying to allocate a page for migration. Ignore allocation
1600 * failure warnings. We don't force __GFP_THISNODE here because
1601 * this node here is the node where we have CMA reservation and
1602 * in some case these nodes will have really less non movable
1603 * allocation memory.
1604 */
1605 gfp_t gfp_mask = GFP_USER | __GFP_NOWARN;
1606
1607 if (PageHighMem(page))
1608 gfp_mask |= __GFP_HIGHMEM;
1609
1610 #ifdef CONFIG_HUGETLB_PAGE
1611 if (PageHuge(page)) {
1612 struct hstate *h = page_hstate(page);
1613 /*
1614 * We don't want to dequeue from the pool because pool pages will
1615 * mostly be from the CMA region.
1616 */
1617 return alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1618 }
1619 #endif
1620 if (PageTransHuge(page)) {
1621 struct page *thp;
1622 /*
1623 * ignore allocation failure warnings
1624 */
1625 gfp_t thp_gfpmask = GFP_TRANSHUGE | __GFP_NOWARN;
1626
1627 /*
1628 * Remove the movable mask so that we don't allocate from
1629 * CMA area again.
1630 */
1631 thp_gfpmask &= ~__GFP_MOVABLE;
1632 thp = __alloc_pages_node(nid, thp_gfpmask, HPAGE_PMD_ORDER);
1633 if (!thp)
1634 return NULL;
1635 prep_transhuge_page(thp);
1636 return thp;
1637 }
1638
1639 return __alloc_pages_node(nid, gfp_mask, 0);
1640 }
1641
1642 static long check_and_migrate_cma_pages(struct task_struct *tsk,
1643 struct mm_struct *mm,
1644 unsigned long start,
1645 unsigned long nr_pages,
1646 struct page **pages,
1647 struct vm_area_struct **vmas,
1648 unsigned int gup_flags)
1649 {
1650 unsigned long i;
1651 unsigned long step;
1652 bool drain_allow = true;
1653 bool migrate_allow = true;
1654 LIST_HEAD(cma_page_list);
1655 long ret = nr_pages;
1656
1657 check_again:
1658 for (i = 0; i < nr_pages;) {
1659
1660 struct page *head = compound_head(pages[i]);
1661
1662 /*
1663 * gup may start from a tail page. Advance step by the left
1664 * part.
1665 */
1666 step = compound_nr(head) - (pages[i] - head);
1667 /*
1668 * If we get a page from the CMA zone, since we are going to
1669 * be pinning these entries, we might as well move them out
1670 * of the CMA zone if possible.
1671 */
1672 if (is_migrate_cma_page(head)) {
1673 if (PageHuge(head))
1674 isolate_huge_page(head, &cma_page_list);
1675 else {
1676 if (!PageLRU(head) && drain_allow) {
1677 lru_add_drain_all();
1678 drain_allow = false;
1679 }
1680
1681 if (!isolate_lru_page(head)) {
1682 list_add_tail(&head->lru, &cma_page_list);
1683 mod_node_page_state(page_pgdat(head),
1684 NR_ISOLATED_ANON +
1685 page_is_file_lru(head),
1686 hpage_nr_pages(head));
1687 }
1688 }
1689 }
1690
1691 i += step;
1692 }
1693
1694 if (!list_empty(&cma_page_list)) {
1695 /*
1696 * drop the above get_user_pages reference.
1697 */
1698 for (i = 0; i < nr_pages; i++)
1699 put_page(pages[i]);
1700
1701 if (migrate_pages(&cma_page_list, new_non_cma_page,
1702 NULL, 0, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
1703 /*
1704 * some of the pages failed migration. Do get_user_pages
1705 * without migration.
1706 */
1707 migrate_allow = false;
1708
1709 if (!list_empty(&cma_page_list))
1710 putback_movable_pages(&cma_page_list);
1711 }
1712 /*
1713 * We did migrate all the pages, Try to get the page references
1714 * again migrating any new CMA pages which we failed to isolate
1715 * earlier.
1716 */
1717 ret = __get_user_pages_locked(tsk, mm, start, nr_pages,
1718 pages, vmas, NULL,
1719 gup_flags);
1720
1721 if ((ret > 0) && migrate_allow) {
1722 nr_pages = ret;
1723 drain_allow = true;
1724 goto check_again;
1725 }
1726 }
1727
1728 return ret;
1729 }
1730 #else
1731 static long check_and_migrate_cma_pages(struct task_struct *tsk,
1732 struct mm_struct *mm,
1733 unsigned long start,
1734 unsigned long nr_pages,
1735 struct page **pages,
1736 struct vm_area_struct **vmas,
1737 unsigned int gup_flags)
1738 {
1739 return nr_pages;
1740 }
1741 #endif /* CONFIG_CMA */
1742
1743 /*
1744 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1745 * allows us to process the FOLL_LONGTERM flag.
1746 */
1747 static long __gup_longterm_locked(struct task_struct *tsk,
1748 struct mm_struct *mm,
1749 unsigned long start,
1750 unsigned long nr_pages,
1751 struct page **pages,
1752 struct vm_area_struct **vmas,
1753 unsigned int gup_flags)
1754 {
1755 struct vm_area_struct **vmas_tmp = vmas;
1756 unsigned long flags = 0;
1757 long rc, i;
1758
1759 if (gup_flags & FOLL_LONGTERM) {
1760 if (!pages)
1761 return -EINVAL;
1762
1763 if (!vmas_tmp) {
1764 vmas_tmp = kcalloc(nr_pages,
1765 sizeof(struct vm_area_struct *),
1766 GFP_KERNEL);
1767 if (!vmas_tmp)
1768 return -ENOMEM;
1769 }
1770 flags = memalloc_nocma_save();
1771 }
1772
1773 rc = __get_user_pages_locked(tsk, mm, start, nr_pages, pages,
1774 vmas_tmp, NULL, gup_flags);
1775
1776 if (gup_flags & FOLL_LONGTERM) {
1777 memalloc_nocma_restore(flags);
1778 if (rc < 0)
1779 goto out;
1780
1781 if (check_dax_vmas(vmas_tmp, rc)) {
1782 for (i = 0; i < rc; i++)
1783 put_page(pages[i]);
1784 rc = -EOPNOTSUPP;
1785 goto out;
1786 }
1787
1788 rc = check_and_migrate_cma_pages(tsk, mm, start, rc, pages,
1789 vmas_tmp, gup_flags);
1790 }
1791
1792 out:
1793 if (vmas_tmp != vmas)
1794 kfree(vmas_tmp);
1795 return rc;
1796 }
1797 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1798 static __always_inline long __gup_longterm_locked(struct task_struct *tsk,
1799 struct mm_struct *mm,
1800 unsigned long start,
1801 unsigned long nr_pages,
1802 struct page **pages,
1803 struct vm_area_struct **vmas,
1804 unsigned int flags)
1805 {
1806 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1807 NULL, flags);
1808 }
1809 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1810
1811 #ifdef CONFIG_MMU
1812 static long __get_user_pages_remote(struct task_struct *tsk,
1813 struct mm_struct *mm,
1814 unsigned long start, unsigned long nr_pages,
1815 unsigned int gup_flags, struct page **pages,
1816 struct vm_area_struct **vmas, int *locked)
1817 {
1818 /*
1819 * Parts of FOLL_LONGTERM behavior are incompatible with
1820 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1821 * vmas. However, this only comes up if locked is set, and there are
1822 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1823 * allow what we can.
1824 */
1825 if (gup_flags & FOLL_LONGTERM) {
1826 if (WARN_ON_ONCE(locked))
1827 return -EINVAL;
1828 /*
1829 * This will check the vmas (even if our vmas arg is NULL)
1830 * and return -ENOTSUPP if DAX isn't allowed in this case:
1831 */
1832 return __gup_longterm_locked(tsk, mm, start, nr_pages, pages,
1833 vmas, gup_flags | FOLL_TOUCH |
1834 FOLL_REMOTE);
1835 }
1836
1837 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1838 locked,
1839 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1840 }
1841
1842 /*
1843 * get_user_pages_remote() - pin user pages in memory
1844 * @tsk: the task_struct to use for page fault accounting, or
1845 * NULL if faults are not to be recorded.
1846 * @mm: mm_struct of target mm
1847 * @start: starting user address
1848 * @nr_pages: number of pages from start to pin
1849 * @gup_flags: flags modifying lookup behaviour
1850 * @pages: array that receives pointers to the pages pinned.
1851 * Should be at least nr_pages long. Or NULL, if caller
1852 * only intends to ensure the pages are faulted in.
1853 * @vmas: array of pointers to vmas corresponding to each page.
1854 * Or NULL if the caller does not require them.
1855 * @locked: pointer to lock flag indicating whether lock is held and
1856 * subsequently whether VM_FAULT_RETRY functionality can be
1857 * utilised. Lock must initially be held.
1858 *
1859 * Returns either number of pages pinned (which may be less than the
1860 * number requested), or an error. Details about the return value:
1861 *
1862 * -- If nr_pages is 0, returns 0.
1863 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1864 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1865 * pages pinned. Again, this may be less than nr_pages.
1866 *
1867 * The caller is responsible for releasing returned @pages, via put_page().
1868 *
1869 * @vmas are valid only as long as mmap_sem is held.
1870 *
1871 * Must be called with mmap_sem held for read or write.
1872 *
1873 * get_user_pages walks a process's page tables and takes a reference to
1874 * each struct page that each user address corresponds to at a given
1875 * instant. That is, it takes the page that would be accessed if a user
1876 * thread accesses the given user virtual address at that instant.
1877 *
1878 * This does not guarantee that the page exists in the user mappings when
1879 * get_user_pages returns, and there may even be a completely different
1880 * page there in some cases (eg. if mmapped pagecache has been invalidated
1881 * and subsequently re faulted). However it does guarantee that the page
1882 * won't be freed completely. And mostly callers simply care that the page
1883 * contains data that was valid *at some point in time*. Typically, an IO
1884 * or similar operation cannot guarantee anything stronger anyway because
1885 * locks can't be held over the syscall boundary.
1886 *
1887 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1888 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1889 * be called after the page is finished with, and before put_page is called.
1890 *
1891 * get_user_pages is typically used for fewer-copy IO operations, to get a
1892 * handle on the memory by some means other than accesses via the user virtual
1893 * addresses. The pages may be submitted for DMA to devices or accessed via
1894 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1895 * use the correct cache flushing APIs.
1896 *
1897 * See also get_user_pages_fast, for performance critical applications.
1898 *
1899 * get_user_pages should be phased out in favor of
1900 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1901 * should use get_user_pages because it cannot pass
1902 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1903 */
1904 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1905 unsigned long start, unsigned long nr_pages,
1906 unsigned int gup_flags, struct page **pages,
1907 struct vm_area_struct **vmas, int *locked)
1908 {
1909 /*
1910 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1911 * never directly by the caller, so enforce that with an assertion:
1912 */
1913 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1914 return -EINVAL;
1915
1916 return __get_user_pages_remote(tsk, mm, start, nr_pages, gup_flags,
1917 pages, vmas, locked);
1918 }
1919 EXPORT_SYMBOL(get_user_pages_remote);
1920
1921 #else /* CONFIG_MMU */
1922 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1923 unsigned long start, unsigned long nr_pages,
1924 unsigned int gup_flags, struct page **pages,
1925 struct vm_area_struct **vmas, int *locked)
1926 {
1927 return 0;
1928 }
1929
1930 static long __get_user_pages_remote(struct task_struct *tsk,
1931 struct mm_struct *mm,
1932 unsigned long start, unsigned long nr_pages,
1933 unsigned int gup_flags, struct page **pages,
1934 struct vm_area_struct **vmas, int *locked)
1935 {
1936 return 0;
1937 }
1938 #endif /* !CONFIG_MMU */
1939
1940 /*
1941 * This is the same as get_user_pages_remote(), just with a
1942 * less-flexible calling convention where we assume that the task
1943 * and mm being operated on are the current task's and don't allow
1944 * passing of a locked parameter. We also obviously don't pass
1945 * FOLL_REMOTE in here.
1946 */
1947 long get_user_pages(unsigned long start, unsigned long nr_pages,
1948 unsigned int gup_flags, struct page **pages,
1949 struct vm_area_struct **vmas)
1950 {
1951 /*
1952 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1953 * never directly by the caller, so enforce that with an assertion:
1954 */
1955 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1956 return -EINVAL;
1957
1958 return __gup_longterm_locked(current, current->mm, start, nr_pages,
1959 pages, vmas, gup_flags | FOLL_TOUCH);
1960 }
1961 EXPORT_SYMBOL(get_user_pages);
1962
1963 /*
1964 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1965 * paths better by using either get_user_pages_locked() or
1966 * get_user_pages_unlocked().
1967 *
1968 * get_user_pages_locked() is suitable to replace the form:
1969 *
1970 * down_read(&mm->mmap_sem);
1971 * do_something()
1972 * get_user_pages(tsk, mm, ..., pages, NULL);
1973 * up_read(&mm->mmap_sem);
1974 *
1975 * to:
1976 *
1977 * int locked = 1;
1978 * down_read(&mm->mmap_sem);
1979 * do_something()
1980 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
1981 * if (locked)
1982 * up_read(&mm->mmap_sem);
1983 */
1984 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1985 unsigned int gup_flags, struct page **pages,
1986 int *locked)
1987 {
1988 /*
1989 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1990 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1991 * vmas. As there are no users of this flag in this call we simply
1992 * disallow this option for now.
1993 */
1994 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1995 return -EINVAL;
1996
1997 return __get_user_pages_locked(current, current->mm, start, nr_pages,
1998 pages, NULL, locked,
1999 gup_flags | FOLL_TOUCH);
2000 }
2001 EXPORT_SYMBOL(get_user_pages_locked);
2002
2003 /*
2004 * get_user_pages_unlocked() is suitable to replace the form:
2005 *
2006 * down_read(&mm->mmap_sem);
2007 * get_user_pages(tsk, mm, ..., pages, NULL);
2008 * up_read(&mm->mmap_sem);
2009 *
2010 * with:
2011 *
2012 * get_user_pages_unlocked(tsk, mm, ..., pages);
2013 *
2014 * It is functionally equivalent to get_user_pages_fast so
2015 * get_user_pages_fast should be used instead if specific gup_flags
2016 * (e.g. FOLL_FORCE) are not required.
2017 */
2018 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2019 struct page **pages, unsigned int gup_flags)
2020 {
2021 struct mm_struct *mm = current->mm;
2022 int locked = 1;
2023 long ret;
2024
2025 /*
2026 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2027 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2028 * vmas. As there are no users of this flag in this call we simply
2029 * disallow this option for now.
2030 */
2031 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2032 return -EINVAL;
2033
2034 down_read(&mm->mmap_sem);
2035 ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL,
2036 &locked, gup_flags | FOLL_TOUCH);
2037 if (locked)
2038 up_read(&mm->mmap_sem);
2039 return ret;
2040 }
2041 EXPORT_SYMBOL(get_user_pages_unlocked);
2042
2043 /*
2044 * Fast GUP
2045 *
2046 * get_user_pages_fast attempts to pin user pages by walking the page
2047 * tables directly and avoids taking locks. Thus the walker needs to be
2048 * protected from page table pages being freed from under it, and should
2049 * block any THP splits.
2050 *
2051 * One way to achieve this is to have the walker disable interrupts, and
2052 * rely on IPIs from the TLB flushing code blocking before the page table
2053 * pages are freed. This is unsuitable for architectures that do not need
2054 * to broadcast an IPI when invalidating TLBs.
2055 *
2056 * Another way to achieve this is to batch up page table containing pages
2057 * belonging to more than one mm_user, then rcu_sched a callback to free those
2058 * pages. Disabling interrupts will allow the fast_gup walker to both block
2059 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2060 * (which is a relatively rare event). The code below adopts this strategy.
2061 *
2062 * Before activating this code, please be aware that the following assumptions
2063 * are currently made:
2064 *
2065 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2066 * free pages containing page tables or TLB flushing requires IPI broadcast.
2067 *
2068 * *) ptes can be read atomically by the architecture.
2069 *
2070 * *) access_ok is sufficient to validate userspace address ranges.
2071 *
2072 * The last two assumptions can be relaxed by the addition of helper functions.
2073 *
2074 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2075 */
2076 #ifdef CONFIG_HAVE_FAST_GUP
2077
2078 static void put_compound_head(struct page *page, int refs, unsigned int flags)
2079 {
2080 if (flags & FOLL_PIN) {
2081 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED,
2082 refs);
2083
2084 if (hpage_pincount_available(page))
2085 hpage_pincount_sub(page, refs);
2086 else
2087 refs *= GUP_PIN_COUNTING_BIAS;
2088 }
2089
2090 VM_BUG_ON_PAGE(page_ref_count(page) < refs, page);
2091 /*
2092 * Calling put_page() for each ref is unnecessarily slow. Only the last
2093 * ref needs a put_page().
2094 */
2095 if (refs > 1)
2096 page_ref_sub(page, refs - 1);
2097 put_page(page);
2098 }
2099
2100 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
2101
2102 /*
2103 * WARNING: only to be used in the get_user_pages_fast() implementation.
2104 *
2105 * With get_user_pages_fast(), we walk down the pagetables without taking any
2106 * locks. For this we would like to load the pointers atomically, but sometimes
2107 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
2108 * we do have is the guarantee that a PTE will only either go from not present
2109 * to present, or present to not present or both -- it will not switch to a
2110 * completely different present page without a TLB flush in between; something
2111 * that we are blocking by holding interrupts off.
2112 *
2113 * Setting ptes from not present to present goes:
2114 *
2115 * ptep->pte_high = h;
2116 * smp_wmb();
2117 * ptep->pte_low = l;
2118 *
2119 * And present to not present goes:
2120 *
2121 * ptep->pte_low = 0;
2122 * smp_wmb();
2123 * ptep->pte_high = 0;
2124 *
2125 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
2126 * We load pte_high *after* loading pte_low, which ensures we don't see an older
2127 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
2128 * picked up a changed pte high. We might have gotten rubbish values from
2129 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
2130 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
2131 * operates on present ptes we're safe.
2132 */
2133 static inline pte_t gup_get_pte(pte_t *ptep)
2134 {
2135 pte_t pte;
2136
2137 do {
2138 pte.pte_low = ptep->pte_low;
2139 smp_rmb();
2140 pte.pte_high = ptep->pte_high;
2141 smp_rmb();
2142 } while (unlikely(pte.pte_low != ptep->pte_low));
2143
2144 return pte;
2145 }
2146 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2147 /*
2148 * We require that the PTE can be read atomically.
2149 */
2150 static inline pte_t gup_get_pte(pte_t *ptep)
2151 {
2152 return READ_ONCE(*ptep);
2153 }
2154 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2155
2156 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2157 unsigned int flags,
2158 struct page **pages)
2159 {
2160 while ((*nr) - nr_start) {
2161 struct page *page = pages[--(*nr)];
2162
2163 ClearPageReferenced(page);
2164 if (flags & FOLL_PIN)
2165 unpin_user_page(page);
2166 else
2167 put_page(page);
2168 }
2169 }
2170
2171 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2172 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2173 unsigned int flags, struct page **pages, int *nr)
2174 {
2175 struct dev_pagemap *pgmap = NULL;
2176 int nr_start = *nr, ret = 0;
2177 pte_t *ptep, *ptem;
2178
2179 ptem = ptep = pte_offset_map(&pmd, addr);
2180 do {
2181 pte_t pte = gup_get_pte(ptep);
2182 struct page *head, *page;
2183
2184 /*
2185 * Similar to the PMD case below, NUMA hinting must take slow
2186 * path using the pte_protnone check.
2187 */
2188 if (pte_protnone(pte))
2189 goto pte_unmap;
2190
2191 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2192 goto pte_unmap;
2193
2194 if (pte_devmap(pte)) {
2195 if (unlikely(flags & FOLL_LONGTERM))
2196 goto pte_unmap;
2197
2198 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2199 if (unlikely(!pgmap)) {
2200 undo_dev_pagemap(nr, nr_start, flags, pages);
2201 goto pte_unmap;
2202 }
2203 } else if (pte_special(pte))
2204 goto pte_unmap;
2205
2206 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2207 page = pte_page(pte);
2208
2209 head = try_grab_compound_head(page, 1, flags);
2210 if (!head)
2211 goto pte_unmap;
2212
2213 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2214 put_compound_head(head, 1, flags);
2215 goto pte_unmap;
2216 }
2217
2218 VM_BUG_ON_PAGE(compound_head(page) != head, page);
2219
2220 /*
2221 * We need to make the page accessible if and only if we are
2222 * going to access its content (the FOLL_PIN case). Please
2223 * see Documentation/core-api/pin_user_pages.rst for
2224 * details.
2225 */
2226 if (flags & FOLL_PIN) {
2227 ret = arch_make_page_accessible(page);
2228 if (ret) {
2229 unpin_user_page(page);
2230 goto pte_unmap;
2231 }
2232 }
2233 SetPageReferenced(page);
2234 pages[*nr] = page;
2235 (*nr)++;
2236
2237 } while (ptep++, addr += PAGE_SIZE, addr != end);
2238
2239 ret = 1;
2240
2241 pte_unmap:
2242 if (pgmap)
2243 put_dev_pagemap(pgmap);
2244 pte_unmap(ptem);
2245 return ret;
2246 }
2247 #else
2248
2249 /*
2250 * If we can't determine whether or not a pte is special, then fail immediately
2251 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2252 * to be special.
2253 *
2254 * For a futex to be placed on a THP tail page, get_futex_key requires a
2255 * __get_user_pages_fast implementation that can pin pages. Thus it's still
2256 * useful to have gup_huge_pmd even if we can't operate on ptes.
2257 */
2258 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2259 unsigned int flags, struct page **pages, int *nr)
2260 {
2261 return 0;
2262 }
2263 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2264
2265 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2266 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2267 unsigned long end, unsigned int flags,
2268 struct page **pages, int *nr)
2269 {
2270 int nr_start = *nr;
2271 struct dev_pagemap *pgmap = NULL;
2272
2273 do {
2274 struct page *page = pfn_to_page(pfn);
2275
2276 pgmap = get_dev_pagemap(pfn, pgmap);
2277 if (unlikely(!pgmap)) {
2278 undo_dev_pagemap(nr, nr_start, flags, pages);
2279 return 0;
2280 }
2281 SetPageReferenced(page);
2282 pages[*nr] = page;
2283 if (unlikely(!try_grab_page(page, flags))) {
2284 undo_dev_pagemap(nr, nr_start, flags, pages);
2285 return 0;
2286 }
2287 (*nr)++;
2288 pfn++;
2289 } while (addr += PAGE_SIZE, addr != end);
2290
2291 if (pgmap)
2292 put_dev_pagemap(pgmap);
2293 return 1;
2294 }
2295
2296 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2297 unsigned long end, unsigned int flags,
2298 struct page **pages, int *nr)
2299 {
2300 unsigned long fault_pfn;
2301 int nr_start = *nr;
2302
2303 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2304 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2305 return 0;
2306
2307 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2308 undo_dev_pagemap(nr, nr_start, flags, pages);
2309 return 0;
2310 }
2311 return 1;
2312 }
2313
2314 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2315 unsigned long end, unsigned int flags,
2316 struct page **pages, int *nr)
2317 {
2318 unsigned long fault_pfn;
2319 int nr_start = *nr;
2320
2321 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2322 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2323 return 0;
2324
2325 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2326 undo_dev_pagemap(nr, nr_start, flags, pages);
2327 return 0;
2328 }
2329 return 1;
2330 }
2331 #else
2332 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2333 unsigned long end, unsigned int flags,
2334 struct page **pages, int *nr)
2335 {
2336 BUILD_BUG();
2337 return 0;
2338 }
2339
2340 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2341 unsigned long end, unsigned int flags,
2342 struct page **pages, int *nr)
2343 {
2344 BUILD_BUG();
2345 return 0;
2346 }
2347 #endif
2348
2349 static int record_subpages(struct page *page, unsigned long addr,
2350 unsigned long end, struct page **pages)
2351 {
2352 int nr;
2353
2354 for (nr = 0; addr != end; addr += PAGE_SIZE)
2355 pages[nr++] = page++;
2356
2357 return nr;
2358 }
2359
2360 #ifdef CONFIG_ARCH_HAS_HUGEPD
2361 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2362 unsigned long sz)
2363 {
2364 unsigned long __boundary = (addr + sz) & ~(sz-1);
2365 return (__boundary - 1 < end - 1) ? __boundary : end;
2366 }
2367
2368 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2369 unsigned long end, unsigned int flags,
2370 struct page **pages, int *nr)
2371 {
2372 unsigned long pte_end;
2373 struct page *head, *page;
2374 pte_t pte;
2375 int refs;
2376
2377 pte_end = (addr + sz) & ~(sz-1);
2378 if (pte_end < end)
2379 end = pte_end;
2380
2381 pte = READ_ONCE(*ptep);
2382
2383 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2384 return 0;
2385
2386 /* hugepages are never "special" */
2387 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2388
2389 head = pte_page(pte);
2390 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2391 refs = record_subpages(page, addr, end, pages + *nr);
2392
2393 head = try_grab_compound_head(head, refs, flags);
2394 if (!head)
2395 return 0;
2396
2397 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2398 put_compound_head(head, refs, flags);
2399 return 0;
2400 }
2401
2402 *nr += refs;
2403 SetPageReferenced(head);
2404 return 1;
2405 }
2406
2407 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2408 unsigned int pdshift, unsigned long end, unsigned int flags,
2409 struct page **pages, int *nr)
2410 {
2411 pte_t *ptep;
2412 unsigned long sz = 1UL << hugepd_shift(hugepd);
2413 unsigned long next;
2414
2415 ptep = hugepte_offset(hugepd, addr, pdshift);
2416 do {
2417 next = hugepte_addr_end(addr, end, sz);
2418 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2419 return 0;
2420 } while (ptep++, addr = next, addr != end);
2421
2422 return 1;
2423 }
2424 #else
2425 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2426 unsigned int pdshift, unsigned long end, unsigned int flags,
2427 struct page **pages, int *nr)
2428 {
2429 return 0;
2430 }
2431 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2432
2433 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2434 unsigned long end, unsigned int flags,
2435 struct page **pages, int *nr)
2436 {
2437 struct page *head, *page;
2438 int refs;
2439
2440 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2441 return 0;
2442
2443 if (pmd_devmap(orig)) {
2444 if (unlikely(flags & FOLL_LONGTERM))
2445 return 0;
2446 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2447 pages, nr);
2448 }
2449
2450 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2451 refs = record_subpages(page, addr, end, pages + *nr);
2452
2453 head = try_grab_compound_head(pmd_page(orig), refs, flags);
2454 if (!head)
2455 return 0;
2456
2457 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2458 put_compound_head(head, refs, flags);
2459 return 0;
2460 }
2461
2462 *nr += refs;
2463 SetPageReferenced(head);
2464 return 1;
2465 }
2466
2467 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2468 unsigned long end, unsigned int flags,
2469 struct page **pages, int *nr)
2470 {
2471 struct page *head, *page;
2472 int refs;
2473
2474 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2475 return 0;
2476
2477 if (pud_devmap(orig)) {
2478 if (unlikely(flags & FOLL_LONGTERM))
2479 return 0;
2480 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2481 pages, nr);
2482 }
2483
2484 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2485 refs = record_subpages(page, addr, end, pages + *nr);
2486
2487 head = try_grab_compound_head(pud_page(orig), refs, flags);
2488 if (!head)
2489 return 0;
2490
2491 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2492 put_compound_head(head, refs, flags);
2493 return 0;
2494 }
2495
2496 *nr += refs;
2497 SetPageReferenced(head);
2498 return 1;
2499 }
2500
2501 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2502 unsigned long end, unsigned int flags,
2503 struct page **pages, int *nr)
2504 {
2505 int refs;
2506 struct page *head, *page;
2507
2508 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2509 return 0;
2510
2511 BUILD_BUG_ON(pgd_devmap(orig));
2512
2513 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2514 refs = record_subpages(page, addr, end, pages + *nr);
2515
2516 head = try_grab_compound_head(pgd_page(orig), refs, flags);
2517 if (!head)
2518 return 0;
2519
2520 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2521 put_compound_head(head, refs, flags);
2522 return 0;
2523 }
2524
2525 *nr += refs;
2526 SetPageReferenced(head);
2527 return 1;
2528 }
2529
2530 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
2531 unsigned int flags, struct page **pages, int *nr)
2532 {
2533 unsigned long next;
2534 pmd_t *pmdp;
2535
2536 pmdp = pmd_offset(&pud, addr);
2537 do {
2538 pmd_t pmd = READ_ONCE(*pmdp);
2539
2540 next = pmd_addr_end(addr, end);
2541 if (!pmd_present(pmd))
2542 return 0;
2543
2544 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2545 pmd_devmap(pmd))) {
2546 /*
2547 * NUMA hinting faults need to be handled in the GUP
2548 * slowpath for accounting purposes and so that they
2549 * can be serialised against THP migration.
2550 */
2551 if (pmd_protnone(pmd))
2552 return 0;
2553
2554 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2555 pages, nr))
2556 return 0;
2557
2558 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2559 /*
2560 * architecture have different format for hugetlbfs
2561 * pmd format and THP pmd format
2562 */
2563 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2564 PMD_SHIFT, next, flags, pages, nr))
2565 return 0;
2566 } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2567 return 0;
2568 } while (pmdp++, addr = next, addr != end);
2569
2570 return 1;
2571 }
2572
2573 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
2574 unsigned int flags, struct page **pages, int *nr)
2575 {
2576 unsigned long next;
2577 pud_t *pudp;
2578
2579 pudp = pud_offset(&p4d, addr);
2580 do {
2581 pud_t pud = READ_ONCE(*pudp);
2582
2583 next = pud_addr_end(addr, end);
2584 if (unlikely(!pud_present(pud)))
2585 return 0;
2586 if (unlikely(pud_huge(pud))) {
2587 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2588 pages, nr))
2589 return 0;
2590 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2591 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2592 PUD_SHIFT, next, flags, pages, nr))
2593 return 0;
2594 } else if (!gup_pmd_range(pud, addr, next, flags, pages, nr))
2595 return 0;
2596 } while (pudp++, addr = next, addr != end);
2597
2598 return 1;
2599 }
2600
2601 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
2602 unsigned int flags, struct page **pages, int *nr)
2603 {
2604 unsigned long next;
2605 p4d_t *p4dp;
2606
2607 p4dp = p4d_offset(&pgd, addr);
2608 do {
2609 p4d_t p4d = READ_ONCE(*p4dp);
2610
2611 next = p4d_addr_end(addr, end);
2612 if (p4d_none(p4d))
2613 return 0;
2614 BUILD_BUG_ON(p4d_huge(p4d));
2615 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2616 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2617 P4D_SHIFT, next, flags, pages, nr))
2618 return 0;
2619 } else if (!gup_pud_range(p4d, addr, next, flags, pages, nr))
2620 return 0;
2621 } while (p4dp++, addr = next, addr != end);
2622
2623 return 1;
2624 }
2625
2626 static void gup_pgd_range(unsigned long addr, unsigned long end,
2627 unsigned int flags, struct page **pages, int *nr)
2628 {
2629 unsigned long next;
2630 pgd_t *pgdp;
2631
2632 pgdp = pgd_offset(current->mm, addr);
2633 do {
2634 pgd_t pgd = READ_ONCE(*pgdp);
2635
2636 next = pgd_addr_end(addr, end);
2637 if (pgd_none(pgd))
2638 return;
2639 if (unlikely(pgd_huge(pgd))) {
2640 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2641 pages, nr))
2642 return;
2643 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2644 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2645 PGDIR_SHIFT, next, flags, pages, nr))
2646 return;
2647 } else if (!gup_p4d_range(pgd, addr, next, flags, pages, nr))
2648 return;
2649 } while (pgdp++, addr = next, addr != end);
2650 }
2651 #else
2652 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2653 unsigned int flags, struct page **pages, int *nr)
2654 {
2655 }
2656 #endif /* CONFIG_HAVE_FAST_GUP */
2657
2658 #ifndef gup_fast_permitted
2659 /*
2660 * Check if it's allowed to use __get_user_pages_fast() for the range, or
2661 * we need to fall back to the slow version:
2662 */
2663 static bool gup_fast_permitted(unsigned long start, unsigned long end)
2664 {
2665 return true;
2666 }
2667 #endif
2668
2669 /*
2670 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2671 * the regular GUP.
2672 * Note a difference with get_user_pages_fast: this always returns the
2673 * number of pages pinned, 0 if no pages were pinned.
2674 *
2675 * If the architecture does not support this function, simply return with no
2676 * pages pinned.
2677 */
2678 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
2679 struct page **pages)
2680 {
2681 unsigned long len, end;
2682 unsigned long flags;
2683 int nr_pinned = 0;
2684 /*
2685 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2686 * because gup fast is always a "pin with a +1 page refcount" request.
2687 */
2688 unsigned int gup_flags = FOLL_GET;
2689
2690 if (write)
2691 gup_flags |= FOLL_WRITE;
2692
2693 start = untagged_addr(start) & PAGE_MASK;
2694 len = (unsigned long) nr_pages << PAGE_SHIFT;
2695 end = start + len;
2696
2697 if (end <= start)
2698 return 0;
2699 if (unlikely(!access_ok((void __user *)start, len)))
2700 return 0;
2701
2702 /*
2703 * Disable interrupts. We use the nested form as we can already have
2704 * interrupts disabled by get_futex_key.
2705 *
2706 * With interrupts disabled, we block page table pages from being
2707 * freed from under us. See struct mmu_table_batch comments in
2708 * include/asm-generic/tlb.h for more details.
2709 *
2710 * We do not adopt an rcu_read_lock(.) here as we also want to
2711 * block IPIs that come from THPs splitting.
2712 */
2713
2714 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
2715 gup_fast_permitted(start, end)) {
2716 local_irq_save(flags);
2717 gup_pgd_range(start, end, gup_flags, pages, &nr_pinned);
2718 local_irq_restore(flags);
2719 }
2720
2721 return nr_pinned;
2722 }
2723 EXPORT_SYMBOL_GPL(__get_user_pages_fast);
2724
2725 static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2726 unsigned int gup_flags, struct page **pages)
2727 {
2728 int ret;
2729
2730 /*
2731 * FIXME: FOLL_LONGTERM does not work with
2732 * get_user_pages_unlocked() (see comments in that function)
2733 */
2734 if (gup_flags & FOLL_LONGTERM) {
2735 down_read(&current->mm->mmap_sem);
2736 ret = __gup_longterm_locked(current, current->mm,
2737 start, nr_pages,
2738 pages, NULL, gup_flags);
2739 up_read(&current->mm->mmap_sem);
2740 } else {
2741 ret = get_user_pages_unlocked(start, nr_pages,
2742 pages, gup_flags);
2743 }
2744
2745 return ret;
2746 }
2747
2748 static int internal_get_user_pages_fast(unsigned long start, int nr_pages,
2749 unsigned int gup_flags,
2750 struct page **pages)
2751 {
2752 unsigned long addr, len, end;
2753 int nr_pinned = 0, ret = 0;
2754
2755 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
2756 FOLL_FORCE | FOLL_PIN | FOLL_GET)))
2757 return -EINVAL;
2758
2759 start = untagged_addr(start) & PAGE_MASK;
2760 addr = start;
2761 len = (unsigned long) nr_pages << PAGE_SHIFT;
2762 end = start + len;
2763
2764 if (end <= start)
2765 return 0;
2766 if (unlikely(!access_ok((void __user *)start, len)))
2767 return -EFAULT;
2768
2769 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
2770 gup_fast_permitted(start, end)) {
2771 local_irq_disable();
2772 gup_pgd_range(addr, end, gup_flags, pages, &nr_pinned);
2773 local_irq_enable();
2774 ret = nr_pinned;
2775 }
2776
2777 if (nr_pinned < nr_pages) {
2778 /* Try to get the remaining pages with get_user_pages */
2779 start += nr_pinned << PAGE_SHIFT;
2780 pages += nr_pinned;
2781
2782 ret = __gup_longterm_unlocked(start, nr_pages - nr_pinned,
2783 gup_flags, pages);
2784
2785 /* Have to be a bit careful with return values */
2786 if (nr_pinned > 0) {
2787 if (ret < 0)
2788 ret = nr_pinned;
2789 else
2790 ret += nr_pinned;
2791 }
2792 }
2793
2794 return ret;
2795 }
2796
2797 /**
2798 * get_user_pages_fast() - pin user pages in memory
2799 * @start: starting user address
2800 * @nr_pages: number of pages from start to pin
2801 * @gup_flags: flags modifying pin behaviour
2802 * @pages: array that receives pointers to the pages pinned.
2803 * Should be at least nr_pages long.
2804 *
2805 * Attempt to pin user pages in memory without taking mm->mmap_sem.
2806 * If not successful, it will fall back to taking the lock and
2807 * calling get_user_pages().
2808 *
2809 * Returns number of pages pinned. This may be fewer than the number requested.
2810 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2811 * -errno.
2812 */
2813 int get_user_pages_fast(unsigned long start, int nr_pages,
2814 unsigned int gup_flags, struct page **pages)
2815 {
2816 /*
2817 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
2818 * never directly by the caller, so enforce that:
2819 */
2820 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
2821 return -EINVAL;
2822
2823 /*
2824 * The caller may or may not have explicitly set FOLL_GET; either way is
2825 * OK. However, internally (within mm/gup.c), gup fast variants must set
2826 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2827 * request.
2828 */
2829 gup_flags |= FOLL_GET;
2830 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2831 }
2832 EXPORT_SYMBOL_GPL(get_user_pages_fast);
2833
2834 /**
2835 * pin_user_pages_fast() - pin user pages in memory without taking locks
2836 *
2837 * @start: starting user address
2838 * @nr_pages: number of pages from start to pin
2839 * @gup_flags: flags modifying pin behaviour
2840 * @pages: array that receives pointers to the pages pinned.
2841 * Should be at least nr_pages long.
2842 *
2843 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2844 * get_user_pages_fast() for documentation on the function arguments, because
2845 * the arguments here are identical.
2846 *
2847 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2848 * see Documentation/vm/pin_user_pages.rst for further details.
2849 *
2850 * This is intended for Case 1 (DIO) in Documentation/vm/pin_user_pages.rst. It
2851 * is NOT intended for Case 2 (RDMA: long-term pins).
2852 */
2853 int pin_user_pages_fast(unsigned long start, int nr_pages,
2854 unsigned int gup_flags, struct page **pages)
2855 {
2856 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2857 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2858 return -EINVAL;
2859
2860 gup_flags |= FOLL_PIN;
2861 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2862 }
2863 EXPORT_SYMBOL_GPL(pin_user_pages_fast);
2864
2865 /**
2866 * pin_user_pages_remote() - pin pages of a remote process (task != current)
2867 *
2868 * @tsk: the task_struct to use for page fault accounting, or
2869 * NULL if faults are not to be recorded.
2870 * @mm: mm_struct of target mm
2871 * @start: starting user address
2872 * @nr_pages: number of pages from start to pin
2873 * @gup_flags: flags modifying lookup behaviour
2874 * @pages: array that receives pointers to the pages pinned.
2875 * Should be at least nr_pages long. Or NULL, if caller
2876 * only intends to ensure the pages are faulted in.
2877 * @vmas: array of pointers to vmas corresponding to each page.
2878 * Or NULL if the caller does not require them.
2879 * @locked: pointer to lock flag indicating whether lock is held and
2880 * subsequently whether VM_FAULT_RETRY functionality can be
2881 * utilised. Lock must initially be held.
2882 *
2883 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2884 * get_user_pages_remote() for documentation on the function arguments, because
2885 * the arguments here are identical.
2886 *
2887 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2888 * see Documentation/vm/pin_user_pages.rst for details.
2889 *
2890 * This is intended for Case 1 (DIO) in Documentation/vm/pin_user_pages.rst. It
2891 * is NOT intended for Case 2 (RDMA: long-term pins).
2892 */
2893 long pin_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
2894 unsigned long start, unsigned long nr_pages,
2895 unsigned int gup_flags, struct page **pages,
2896 struct vm_area_struct **vmas, int *locked)
2897 {
2898 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2899 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2900 return -EINVAL;
2901
2902 gup_flags |= FOLL_PIN;
2903 return __get_user_pages_remote(tsk, mm, start, nr_pages, gup_flags,
2904 pages, vmas, locked);
2905 }
2906 EXPORT_SYMBOL(pin_user_pages_remote);
2907
2908 /**
2909 * pin_user_pages() - pin user pages in memory for use by other devices
2910 *
2911 * @start: starting user address
2912 * @nr_pages: number of pages from start to pin
2913 * @gup_flags: flags modifying lookup behaviour
2914 * @pages: array that receives pointers to the pages pinned.
2915 * Should be at least nr_pages long. Or NULL, if caller
2916 * only intends to ensure the pages are faulted in.
2917 * @vmas: array of pointers to vmas corresponding to each page.
2918 * Or NULL if the caller does not require them.
2919 *
2920 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2921 * FOLL_PIN is set.
2922 *
2923 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2924 * see Documentation/vm/pin_user_pages.rst for details.
2925 *
2926 * This is intended for Case 1 (DIO) in Documentation/vm/pin_user_pages.rst. It
2927 * is NOT intended for Case 2 (RDMA: long-term pins).
2928 */
2929 long pin_user_pages(unsigned long start, unsigned long nr_pages,
2930 unsigned int gup_flags, struct page **pages,
2931 struct vm_area_struct **vmas)
2932 {
2933 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2934 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2935 return -EINVAL;
2936
2937 gup_flags |= FOLL_PIN;
2938 return __gup_longterm_locked(current, current->mm, start, nr_pages,
2939 pages, vmas, gup_flags);
2940 }
2941 EXPORT_SYMBOL(pin_user_pages);