Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2#include <linux/kernel.h>
3#include <linux/errno.h>
4#include <linux/err.h>
5#include <linux/spinlock.h>
6
7#include <linux/mm.h>
8#include <linux/memremap.h>
9#include <linux/pagemap.h>
10#include <linux/rmap.h>
11#include <linux/swap.h>
12#include <linux/swapops.h>
13
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
27struct follow_page_context {
28 struct dev_pagemap *pgmap;
29 unsigned int page_mask;
30};
31
32/**
33 * put_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
34 * @pages: array of pages to be maybe marked dirty, and definitely released.
35 * @npages: number of pages in the @pages array.
36 * @make_dirty: whether to mark the pages dirty
37 *
38 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
39 * variants called on that page.
40 *
41 * For each page in the @pages array, make that page (or its head page, if a
42 * compound page) dirty, if @make_dirty is true, and if the page was previously
43 * listed as clean. In any case, releases all pages using put_user_page(),
44 * possibly via put_user_pages(), for the non-dirty case.
45 *
46 * Please see the put_user_page() documentation for details.
47 *
48 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
49 * required, then the caller should a) verify that this is really correct,
50 * because _lock() is usually required, and b) hand code it:
51 * set_page_dirty_lock(), put_user_page().
52 *
53 */
54void put_user_pages_dirty_lock(struct page **pages, unsigned long npages,
55 bool make_dirty)
56{
57 unsigned long index;
58
59 /*
60 * TODO: this can be optimized for huge pages: if a series of pages is
61 * physically contiguous and part of the same compound page, then a
62 * single operation to the head page should suffice.
63 */
64
65 if (!make_dirty) {
66 put_user_pages(pages, npages);
67 return;
68 }
69
70 for (index = 0; index < npages; index++) {
71 struct page *page = compound_head(pages[index]);
72 /*
73 * Checking PageDirty at this point may race with
74 * clear_page_dirty_for_io(), but that's OK. Two key
75 * cases:
76 *
77 * 1) This code sees the page as already dirty, so it
78 * skips the call to set_page_dirty(). That could happen
79 * because clear_page_dirty_for_io() called
80 * page_mkclean(), followed by set_page_dirty().
81 * However, now the page is going to get written back,
82 * which meets the original intention of setting it
83 * dirty, so all is well: clear_page_dirty_for_io() goes
84 * on to call TestClearPageDirty(), and write the page
85 * back.
86 *
87 * 2) This code sees the page as clean, so it calls
88 * set_page_dirty(). The page stays dirty, despite being
89 * written back, so it gets written back again in the
90 * next writeback cycle. This is harmless.
91 */
92 if (!PageDirty(page))
93 set_page_dirty_lock(page);
94 put_user_page(page);
95 }
96}
97EXPORT_SYMBOL(put_user_pages_dirty_lock);
98
99/**
100 * put_user_pages() - release an array of gup-pinned pages.
101 * @pages: array of pages to be marked dirty and released.
102 * @npages: number of pages in the @pages array.
103 *
104 * For each page in the @pages array, release the page using put_user_page().
105 *
106 * Please see the put_user_page() documentation for details.
107 */
108void put_user_pages(struct page **pages, unsigned long npages)
109{
110 unsigned long index;
111
112 /*
113 * TODO: this can be optimized for huge pages: if a series of pages is
114 * physically contiguous and part of the same compound page, then a
115 * single operation to the head page should suffice.
116 */
117 for (index = 0; index < npages; index++)
118 put_user_page(pages[index]);
119}
120EXPORT_SYMBOL(put_user_pages);
121
122#ifdef CONFIG_MMU
123static struct page *no_page_table(struct vm_area_struct *vma,
124 unsigned int flags)
125{
126 /*
127 * When core dumping an enormous anonymous area that nobody
128 * has touched so far, we don't want to allocate unnecessary pages or
129 * page tables. Return error instead of NULL to skip handle_mm_fault,
130 * then get_dump_page() will return NULL to leave a hole in the dump.
131 * But we can only make this optimization where a hole would surely
132 * be zero-filled if handle_mm_fault() actually did handle it.
133 */
134 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
135 return ERR_PTR(-EFAULT);
136 return NULL;
137}
138
139static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
140 pte_t *pte, unsigned int flags)
141{
142 /* No page to get reference */
143 if (flags & FOLL_GET)
144 return -EFAULT;
145
146 if (flags & FOLL_TOUCH) {
147 pte_t entry = *pte;
148
149 if (flags & FOLL_WRITE)
150 entry = pte_mkdirty(entry);
151 entry = pte_mkyoung(entry);
152
153 if (!pte_same(*pte, entry)) {
154 set_pte_at(vma->vm_mm, address, pte, entry);
155 update_mmu_cache(vma, address, pte);
156 }
157 }
158
159 /* Proper page table entry exists, but no corresponding struct page */
160 return -EEXIST;
161}
162
163/*
164 * FOLL_FORCE can write to even unwritable pte's, but only
165 * after we've gone through a COW cycle and they are dirty.
166 */
167static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
168{
169 return pte_write(pte) ||
170 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
171}
172
173static struct page *follow_page_pte(struct vm_area_struct *vma,
174 unsigned long address, pmd_t *pmd, unsigned int flags,
175 struct dev_pagemap **pgmap)
176{
177 struct mm_struct *mm = vma->vm_mm;
178 struct page *page;
179 spinlock_t *ptl;
180 pte_t *ptep, pte;
181
182retry:
183 if (unlikely(pmd_bad(*pmd)))
184 return no_page_table(vma, flags);
185
186 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
187 pte = *ptep;
188 if (!pte_present(pte)) {
189 swp_entry_t entry;
190 /*
191 * KSM's break_ksm() relies upon recognizing a ksm page
192 * even while it is being migrated, so for that case we
193 * need migration_entry_wait().
194 */
195 if (likely(!(flags & FOLL_MIGRATION)))
196 goto no_page;
197 if (pte_none(pte))
198 goto no_page;
199 entry = pte_to_swp_entry(pte);
200 if (!is_migration_entry(entry))
201 goto no_page;
202 pte_unmap_unlock(ptep, ptl);
203 migration_entry_wait(mm, pmd, address);
204 goto retry;
205 }
206 if ((flags & FOLL_NUMA) && pte_protnone(pte))
207 goto no_page;
208 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
209 pte_unmap_unlock(ptep, ptl);
210 return NULL;
211 }
212
213 page = vm_normal_page(vma, address, pte);
214 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
215 /*
216 * Only return device mapping pages in the FOLL_GET case since
217 * they are only valid while holding the pgmap reference.
218 */
219 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
220 if (*pgmap)
221 page = pte_page(pte);
222 else
223 goto no_page;
224 } else if (unlikely(!page)) {
225 if (flags & FOLL_DUMP) {
226 /* Avoid special (like zero) pages in core dumps */
227 page = ERR_PTR(-EFAULT);
228 goto out;
229 }
230
231 if (is_zero_pfn(pte_pfn(pte))) {
232 page = pte_page(pte);
233 } else {
234 int ret;
235
236 ret = follow_pfn_pte(vma, address, ptep, flags);
237 page = ERR_PTR(ret);
238 goto out;
239 }
240 }
241
242 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
243 int ret;
244 get_page(page);
245 pte_unmap_unlock(ptep, ptl);
246 lock_page(page);
247 ret = split_huge_page(page);
248 unlock_page(page);
249 put_page(page);
250 if (ret)
251 return ERR_PTR(ret);
252 goto retry;
253 }
254
255 if (flags & FOLL_GET) {
256 if (unlikely(!try_get_page(page))) {
257 page = ERR_PTR(-ENOMEM);
258 goto out;
259 }
260 }
261 if (flags & FOLL_TOUCH) {
262 if ((flags & FOLL_WRITE) &&
263 !pte_dirty(pte) && !PageDirty(page))
264 set_page_dirty(page);
265 /*
266 * pte_mkyoung() would be more correct here, but atomic care
267 * is needed to avoid losing the dirty bit: it is easier to use
268 * mark_page_accessed().
269 */
270 mark_page_accessed(page);
271 }
272 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
273 /* Do not mlock pte-mapped THP */
274 if (PageTransCompound(page))
275 goto out;
276
277 /*
278 * The preliminary mapping check is mainly to avoid the
279 * pointless overhead of lock_page on the ZERO_PAGE
280 * which might bounce very badly if there is contention.
281 *
282 * If the page is already locked, we don't need to
283 * handle it now - vmscan will handle it later if and
284 * when it attempts to reclaim the page.
285 */
286 if (page->mapping && trylock_page(page)) {
287 lru_add_drain(); /* push cached pages to LRU */
288 /*
289 * Because we lock page here, and migration is
290 * blocked by the pte's page reference, and we
291 * know the page is still mapped, we don't even
292 * need to check for file-cache page truncation.
293 */
294 mlock_vma_page(page);
295 unlock_page(page);
296 }
297 }
298out:
299 pte_unmap_unlock(ptep, ptl);
300 return page;
301no_page:
302 pte_unmap_unlock(ptep, ptl);
303 if (!pte_none(pte))
304 return NULL;
305 return no_page_table(vma, flags);
306}
307
308static struct page *follow_pmd_mask(struct vm_area_struct *vma,
309 unsigned long address, pud_t *pudp,
310 unsigned int flags,
311 struct follow_page_context *ctx)
312{
313 pmd_t *pmd, pmdval;
314 spinlock_t *ptl;
315 struct page *page;
316 struct mm_struct *mm = vma->vm_mm;
317
318 pmd = pmd_offset(pudp, address);
319 /*
320 * The READ_ONCE() will stabilize the pmdval in a register or
321 * on the stack so that it will stop changing under the code.
322 */
323 pmdval = READ_ONCE(*pmd);
324 if (pmd_none(pmdval))
325 return no_page_table(vma, flags);
326 if (pmd_huge(pmdval) && vma->vm_flags & VM_HUGETLB) {
327 page = follow_huge_pmd(mm, address, pmd, flags);
328 if (page)
329 return page;
330 return no_page_table(vma, flags);
331 }
332 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
333 page = follow_huge_pd(vma, address,
334 __hugepd(pmd_val(pmdval)), flags,
335 PMD_SHIFT);
336 if (page)
337 return page;
338 return no_page_table(vma, flags);
339 }
340retry:
341 if (!pmd_present(pmdval)) {
342 if (likely(!(flags & FOLL_MIGRATION)))
343 return no_page_table(vma, flags);
344 VM_BUG_ON(thp_migration_supported() &&
345 !is_pmd_migration_entry(pmdval));
346 if (is_pmd_migration_entry(pmdval))
347 pmd_migration_entry_wait(mm, pmd);
348 pmdval = READ_ONCE(*pmd);
349 /*
350 * MADV_DONTNEED may convert the pmd to null because
351 * mmap_sem is held in read mode
352 */
353 if (pmd_none(pmdval))
354 return no_page_table(vma, flags);
355 goto retry;
356 }
357 if (pmd_devmap(pmdval)) {
358 ptl = pmd_lock(mm, pmd);
359 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
360 spin_unlock(ptl);
361 if (page)
362 return page;
363 }
364 if (likely(!pmd_trans_huge(pmdval)))
365 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
366
367 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
368 return no_page_table(vma, flags);
369
370retry_locked:
371 ptl = pmd_lock(mm, pmd);
372 if (unlikely(pmd_none(*pmd))) {
373 spin_unlock(ptl);
374 return no_page_table(vma, flags);
375 }
376 if (unlikely(!pmd_present(*pmd))) {
377 spin_unlock(ptl);
378 if (likely(!(flags & FOLL_MIGRATION)))
379 return no_page_table(vma, flags);
380 pmd_migration_entry_wait(mm, pmd);
381 goto retry_locked;
382 }
383 if (unlikely(!pmd_trans_huge(*pmd))) {
384 spin_unlock(ptl);
385 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
386 }
387 if (flags & (FOLL_SPLIT | FOLL_SPLIT_PMD)) {
388 int ret;
389 page = pmd_page(*pmd);
390 if (is_huge_zero_page(page)) {
391 spin_unlock(ptl);
392 ret = 0;
393 split_huge_pmd(vma, pmd, address);
394 if (pmd_trans_unstable(pmd))
395 ret = -EBUSY;
396 } else if (flags & FOLL_SPLIT) {
397 if (unlikely(!try_get_page(page))) {
398 spin_unlock(ptl);
399 return ERR_PTR(-ENOMEM);
400 }
401 spin_unlock(ptl);
402 lock_page(page);
403 ret = split_huge_page(page);
404 unlock_page(page);
405 put_page(page);
406 if (pmd_none(*pmd))
407 return no_page_table(vma, flags);
408 } else { /* flags & FOLL_SPLIT_PMD */
409 spin_unlock(ptl);
410 split_huge_pmd(vma, pmd, address);
411 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
412 }
413
414 return ret ? ERR_PTR(ret) :
415 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
416 }
417 page = follow_trans_huge_pmd(vma, address, pmd, flags);
418 spin_unlock(ptl);
419 ctx->page_mask = HPAGE_PMD_NR - 1;
420 return page;
421}
422
423static struct page *follow_pud_mask(struct vm_area_struct *vma,
424 unsigned long address, p4d_t *p4dp,
425 unsigned int flags,
426 struct follow_page_context *ctx)
427{
428 pud_t *pud;
429 spinlock_t *ptl;
430 struct page *page;
431 struct mm_struct *mm = vma->vm_mm;
432
433 pud = pud_offset(p4dp, address);
434 if (pud_none(*pud))
435 return no_page_table(vma, flags);
436 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
437 page = follow_huge_pud(mm, address, pud, flags);
438 if (page)
439 return page;
440 return no_page_table(vma, flags);
441 }
442 if (is_hugepd(__hugepd(pud_val(*pud)))) {
443 page = follow_huge_pd(vma, address,
444 __hugepd(pud_val(*pud)), flags,
445 PUD_SHIFT);
446 if (page)
447 return page;
448 return no_page_table(vma, flags);
449 }
450 if (pud_devmap(*pud)) {
451 ptl = pud_lock(mm, pud);
452 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
453 spin_unlock(ptl);
454 if (page)
455 return page;
456 }
457 if (unlikely(pud_bad(*pud)))
458 return no_page_table(vma, flags);
459
460 return follow_pmd_mask(vma, address, pud, flags, ctx);
461}
462
463static struct page *follow_p4d_mask(struct vm_area_struct *vma,
464 unsigned long address, pgd_t *pgdp,
465 unsigned int flags,
466 struct follow_page_context *ctx)
467{
468 p4d_t *p4d;
469 struct page *page;
470
471 p4d = p4d_offset(pgdp, address);
472 if (p4d_none(*p4d))
473 return no_page_table(vma, flags);
474 BUILD_BUG_ON(p4d_huge(*p4d));
475 if (unlikely(p4d_bad(*p4d)))
476 return no_page_table(vma, flags);
477
478 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
479 page = follow_huge_pd(vma, address,
480 __hugepd(p4d_val(*p4d)), flags,
481 P4D_SHIFT);
482 if (page)
483 return page;
484 return no_page_table(vma, flags);
485 }
486 return follow_pud_mask(vma, address, p4d, flags, ctx);
487}
488
489/**
490 * follow_page_mask - look up a page descriptor from a user-virtual address
491 * @vma: vm_area_struct mapping @address
492 * @address: virtual address to look up
493 * @flags: flags modifying lookup behaviour
494 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
495 * pointer to output page_mask
496 *
497 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
498 *
499 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
500 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
501 *
502 * On output, the @ctx->page_mask is set according to the size of the page.
503 *
504 * Return: the mapped (struct page *), %NULL if no mapping exists, or
505 * an error pointer if there is a mapping to something not represented
506 * by a page descriptor (see also vm_normal_page()).
507 */
508static struct page *follow_page_mask(struct vm_area_struct *vma,
509 unsigned long address, unsigned int flags,
510 struct follow_page_context *ctx)
511{
512 pgd_t *pgd;
513 struct page *page;
514 struct mm_struct *mm = vma->vm_mm;
515
516 ctx->page_mask = 0;
517
518 /* make this handle hugepd */
519 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
520 if (!IS_ERR(page)) {
521 BUG_ON(flags & FOLL_GET);
522 return page;
523 }
524
525 pgd = pgd_offset(mm, address);
526
527 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
528 return no_page_table(vma, flags);
529
530 if (pgd_huge(*pgd)) {
531 page = follow_huge_pgd(mm, address, pgd, flags);
532 if (page)
533 return page;
534 return no_page_table(vma, flags);
535 }
536 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
537 page = follow_huge_pd(vma, address,
538 __hugepd(pgd_val(*pgd)), flags,
539 PGDIR_SHIFT);
540 if (page)
541 return page;
542 return no_page_table(vma, flags);
543 }
544
545 return follow_p4d_mask(vma, address, pgd, flags, ctx);
546}
547
548struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
549 unsigned int foll_flags)
550{
551 struct follow_page_context ctx = { NULL };
552 struct page *page;
553
554 page = follow_page_mask(vma, address, foll_flags, &ctx);
555 if (ctx.pgmap)
556 put_dev_pagemap(ctx.pgmap);
557 return page;
558}
559
560static int get_gate_page(struct mm_struct *mm, unsigned long address,
561 unsigned int gup_flags, struct vm_area_struct **vma,
562 struct page **page)
563{
564 pgd_t *pgd;
565 p4d_t *p4d;
566 pud_t *pud;
567 pmd_t *pmd;
568 pte_t *pte;
569 int ret = -EFAULT;
570
571 /* user gate pages are read-only */
572 if (gup_flags & FOLL_WRITE)
573 return -EFAULT;
574 if (address > TASK_SIZE)
575 pgd = pgd_offset_k(address);
576 else
577 pgd = pgd_offset_gate(mm, address);
578 if (pgd_none(*pgd))
579 return -EFAULT;
580 p4d = p4d_offset(pgd, address);
581 if (p4d_none(*p4d))
582 return -EFAULT;
583 pud = pud_offset(p4d, address);
584 if (pud_none(*pud))
585 return -EFAULT;
586 pmd = pmd_offset(pud, address);
587 if (!pmd_present(*pmd))
588 return -EFAULT;
589 VM_BUG_ON(pmd_trans_huge(*pmd));
590 pte = pte_offset_map(pmd, address);
591 if (pte_none(*pte))
592 goto unmap;
593 *vma = get_gate_vma(mm);
594 if (!page)
595 goto out;
596 *page = vm_normal_page(*vma, address, *pte);
597 if (!*page) {
598 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
599 goto unmap;
600 *page = pte_page(*pte);
601 }
602 if (unlikely(!try_get_page(*page))) {
603 ret = -ENOMEM;
604 goto unmap;
605 }
606out:
607 ret = 0;
608unmap:
609 pte_unmap(pte);
610 return ret;
611}
612
613/*
614 * mmap_sem must be held on entry. If @nonblocking != NULL and
615 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
616 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
617 */
618static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
619 unsigned long address, unsigned int *flags, int *nonblocking)
620{
621 unsigned int fault_flags = 0;
622 vm_fault_t ret;
623
624 /* mlock all present pages, but do not fault in new pages */
625 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
626 return -ENOENT;
627 if (*flags & FOLL_WRITE)
628 fault_flags |= FAULT_FLAG_WRITE;
629 if (*flags & FOLL_REMOTE)
630 fault_flags |= FAULT_FLAG_REMOTE;
631 if (nonblocking)
632 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
633 if (*flags & FOLL_NOWAIT)
634 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
635 if (*flags & FOLL_TRIED) {
636 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
637 fault_flags |= FAULT_FLAG_TRIED;
638 }
639
640 ret = handle_mm_fault(vma, address, fault_flags);
641 if (ret & VM_FAULT_ERROR) {
642 int err = vm_fault_to_errno(ret, *flags);
643
644 if (err)
645 return err;
646 BUG();
647 }
648
649 if (tsk) {
650 if (ret & VM_FAULT_MAJOR)
651 tsk->maj_flt++;
652 else
653 tsk->min_flt++;
654 }
655
656 if (ret & VM_FAULT_RETRY) {
657 if (nonblocking && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
658 *nonblocking = 0;
659 return -EBUSY;
660 }
661
662 /*
663 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
664 * necessary, even if maybe_mkwrite decided not to set pte_write. We
665 * can thus safely do subsequent page lookups as if they were reads.
666 * But only do so when looping for pte_write is futile: in some cases
667 * userspace may also be wanting to write to the gotten user page,
668 * which a read fault here might prevent (a readonly page might get
669 * reCOWed by userspace write).
670 */
671 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
672 *flags |= FOLL_COW;
673 return 0;
674}
675
676static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
677{
678 vm_flags_t vm_flags = vma->vm_flags;
679 int write = (gup_flags & FOLL_WRITE);
680 int foreign = (gup_flags & FOLL_REMOTE);
681
682 if (vm_flags & (VM_IO | VM_PFNMAP))
683 return -EFAULT;
684
685 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
686 return -EFAULT;
687
688 if (write) {
689 if (!(vm_flags & VM_WRITE)) {
690 if (!(gup_flags & FOLL_FORCE))
691 return -EFAULT;
692 /*
693 * We used to let the write,force case do COW in a
694 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
695 * set a breakpoint in a read-only mapping of an
696 * executable, without corrupting the file (yet only
697 * when that file had been opened for writing!).
698 * Anon pages in shared mappings are surprising: now
699 * just reject it.
700 */
701 if (!is_cow_mapping(vm_flags))
702 return -EFAULT;
703 }
704 } else if (!(vm_flags & VM_READ)) {
705 if (!(gup_flags & FOLL_FORCE))
706 return -EFAULT;
707 /*
708 * Is there actually any vma we can reach here which does not
709 * have VM_MAYREAD set?
710 */
711 if (!(vm_flags & VM_MAYREAD))
712 return -EFAULT;
713 }
714 /*
715 * gups are always data accesses, not instruction
716 * fetches, so execute=false here
717 */
718 if (!arch_vma_access_permitted(vma, write, false, foreign))
719 return -EFAULT;
720 return 0;
721}
722
723/**
724 * __get_user_pages() - pin user pages in memory
725 * @tsk: task_struct of target task
726 * @mm: mm_struct of target mm
727 * @start: starting user address
728 * @nr_pages: number of pages from start to pin
729 * @gup_flags: flags modifying pin behaviour
730 * @pages: array that receives pointers to the pages pinned.
731 * Should be at least nr_pages long. Or NULL, if caller
732 * only intends to ensure the pages are faulted in.
733 * @vmas: array of pointers to vmas corresponding to each page.
734 * Or NULL if the caller does not require them.
735 * @nonblocking: whether waiting for disk IO or mmap_sem contention
736 *
737 * Returns number of pages pinned. This may be fewer than the number
738 * requested. If nr_pages is 0 or negative, returns 0. If no pages
739 * were pinned, returns -errno. Each page returned must be released
740 * with a put_page() call when it is finished with. vmas will only
741 * remain valid while mmap_sem is held.
742 *
743 * Must be called with mmap_sem held. It may be released. See below.
744 *
745 * __get_user_pages walks a process's page tables and takes a reference to
746 * each struct page that each user address corresponds to at a given
747 * instant. That is, it takes the page that would be accessed if a user
748 * thread accesses the given user virtual address at that instant.
749 *
750 * This does not guarantee that the page exists in the user mappings when
751 * __get_user_pages returns, and there may even be a completely different
752 * page there in some cases (eg. if mmapped pagecache has been invalidated
753 * and subsequently re faulted). However it does guarantee that the page
754 * won't be freed completely. And mostly callers simply care that the page
755 * contains data that was valid *at some point in time*. Typically, an IO
756 * or similar operation cannot guarantee anything stronger anyway because
757 * locks can't be held over the syscall boundary.
758 *
759 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
760 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
761 * appropriate) must be called after the page is finished with, and
762 * before put_page is called.
763 *
764 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
765 * or mmap_sem contention, and if waiting is needed to pin all pages,
766 * *@nonblocking will be set to 0. Further, if @gup_flags does not
767 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
768 * this case.
769 *
770 * A caller using such a combination of @nonblocking and @gup_flags
771 * must therefore hold the mmap_sem for reading only, and recognize
772 * when it's been released. Otherwise, it must be held for either
773 * reading or writing and will not be released.
774 *
775 * In most cases, get_user_pages or get_user_pages_fast should be used
776 * instead of __get_user_pages. __get_user_pages should be used only if
777 * you need some special @gup_flags.
778 */
779static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
780 unsigned long start, unsigned long nr_pages,
781 unsigned int gup_flags, struct page **pages,
782 struct vm_area_struct **vmas, int *nonblocking)
783{
784 long ret = 0, i = 0;
785 struct vm_area_struct *vma = NULL;
786 struct follow_page_context ctx = { NULL };
787
788 if (!nr_pages)
789 return 0;
790
791 start = untagged_addr(start);
792
793 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
794
795 /*
796 * If FOLL_FORCE is set then do not force a full fault as the hinting
797 * fault information is unrelated to the reference behaviour of a task
798 * using the address space
799 */
800 if (!(gup_flags & FOLL_FORCE))
801 gup_flags |= FOLL_NUMA;
802
803 do {
804 struct page *page;
805 unsigned int foll_flags = gup_flags;
806 unsigned int page_increm;
807
808 /* first iteration or cross vma bound */
809 if (!vma || start >= vma->vm_end) {
810 vma = find_extend_vma(mm, start);
811 if (!vma && in_gate_area(mm, start)) {
812 ret = get_gate_page(mm, start & PAGE_MASK,
813 gup_flags, &vma,
814 pages ? &pages[i] : NULL);
815 if (ret)
816 goto out;
817 ctx.page_mask = 0;
818 goto next_page;
819 }
820
821 if (!vma || check_vma_flags(vma, gup_flags)) {
822 ret = -EFAULT;
823 goto out;
824 }
825 if (is_vm_hugetlb_page(vma)) {
826 i = follow_hugetlb_page(mm, vma, pages, vmas,
827 &start, &nr_pages, i,
828 gup_flags, nonblocking);
829 continue;
830 }
831 }
832retry:
833 /*
834 * If we have a pending SIGKILL, don't keep faulting pages and
835 * potentially allocating memory.
836 */
837 if (fatal_signal_pending(current)) {
838 ret = -ERESTARTSYS;
839 goto out;
840 }
841 cond_resched();
842
843 page = follow_page_mask(vma, start, foll_flags, &ctx);
844 if (!page) {
845 ret = faultin_page(tsk, vma, start, &foll_flags,
846 nonblocking);
847 switch (ret) {
848 case 0:
849 goto retry;
850 case -EBUSY:
851 ret = 0;
852 /* FALLTHRU */
853 case -EFAULT:
854 case -ENOMEM:
855 case -EHWPOISON:
856 goto out;
857 case -ENOENT:
858 goto next_page;
859 }
860 BUG();
861 } else if (PTR_ERR(page) == -EEXIST) {
862 /*
863 * Proper page table entry exists, but no corresponding
864 * struct page.
865 */
866 goto next_page;
867 } else if (IS_ERR(page)) {
868 ret = PTR_ERR(page);
869 goto out;
870 }
871 if (pages) {
872 pages[i] = page;
873 flush_anon_page(vma, page, start);
874 flush_dcache_page(page);
875 ctx.page_mask = 0;
876 }
877next_page:
878 if (vmas) {
879 vmas[i] = vma;
880 ctx.page_mask = 0;
881 }
882 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
883 if (page_increm > nr_pages)
884 page_increm = nr_pages;
885 i += page_increm;
886 start += page_increm * PAGE_SIZE;
887 nr_pages -= page_increm;
888 } while (nr_pages);
889out:
890 if (ctx.pgmap)
891 put_dev_pagemap(ctx.pgmap);
892 return i ? i : ret;
893}
894
895static bool vma_permits_fault(struct vm_area_struct *vma,
896 unsigned int fault_flags)
897{
898 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
899 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
900 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
901
902 if (!(vm_flags & vma->vm_flags))
903 return false;
904
905 /*
906 * The architecture might have a hardware protection
907 * mechanism other than read/write that can deny access.
908 *
909 * gup always represents data access, not instruction
910 * fetches, so execute=false here:
911 */
912 if (!arch_vma_access_permitted(vma, write, false, foreign))
913 return false;
914
915 return true;
916}
917
918/*
919 * fixup_user_fault() - manually resolve a user page fault
920 * @tsk: the task_struct to use for page fault accounting, or
921 * NULL if faults are not to be recorded.
922 * @mm: mm_struct of target mm
923 * @address: user address
924 * @fault_flags:flags to pass down to handle_mm_fault()
925 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
926 * does not allow retry
927 *
928 * This is meant to be called in the specific scenario where for locking reasons
929 * we try to access user memory in atomic context (within a pagefault_disable()
930 * section), this returns -EFAULT, and we want to resolve the user fault before
931 * trying again.
932 *
933 * Typically this is meant to be used by the futex code.
934 *
935 * The main difference with get_user_pages() is that this function will
936 * unconditionally call handle_mm_fault() which will in turn perform all the
937 * necessary SW fixup of the dirty and young bits in the PTE, while
938 * get_user_pages() only guarantees to update these in the struct page.
939 *
940 * This is important for some architectures where those bits also gate the
941 * access permission to the page because they are maintained in software. On
942 * such architectures, gup() will not be enough to make a subsequent access
943 * succeed.
944 *
945 * This function will not return with an unlocked mmap_sem. So it has not the
946 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
947 */
948int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
949 unsigned long address, unsigned int fault_flags,
950 bool *unlocked)
951{
952 struct vm_area_struct *vma;
953 vm_fault_t ret, major = 0;
954
955 address = untagged_addr(address);
956
957 if (unlocked)
958 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
959
960retry:
961 vma = find_extend_vma(mm, address);
962 if (!vma || address < vma->vm_start)
963 return -EFAULT;
964
965 if (!vma_permits_fault(vma, fault_flags))
966 return -EFAULT;
967
968 ret = handle_mm_fault(vma, address, fault_flags);
969 major |= ret & VM_FAULT_MAJOR;
970 if (ret & VM_FAULT_ERROR) {
971 int err = vm_fault_to_errno(ret, 0);
972
973 if (err)
974 return err;
975 BUG();
976 }
977
978 if (ret & VM_FAULT_RETRY) {
979 down_read(&mm->mmap_sem);
980 if (!(fault_flags & FAULT_FLAG_TRIED)) {
981 *unlocked = true;
982 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
983 fault_flags |= FAULT_FLAG_TRIED;
984 goto retry;
985 }
986 }
987
988 if (tsk) {
989 if (major)
990 tsk->maj_flt++;
991 else
992 tsk->min_flt++;
993 }
994 return 0;
995}
996EXPORT_SYMBOL_GPL(fixup_user_fault);
997
998static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
999 struct mm_struct *mm,
1000 unsigned long start,
1001 unsigned long nr_pages,
1002 struct page **pages,
1003 struct vm_area_struct **vmas,
1004 int *locked,
1005 unsigned int flags)
1006{
1007 long ret, pages_done;
1008 bool lock_dropped;
1009
1010 if (locked) {
1011 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1012 BUG_ON(vmas);
1013 /* check caller initialized locked */
1014 BUG_ON(*locked != 1);
1015 }
1016
1017 if (pages)
1018 flags |= FOLL_GET;
1019
1020 pages_done = 0;
1021 lock_dropped = false;
1022 for (;;) {
1023 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
1024 vmas, locked);
1025 if (!locked)
1026 /* VM_FAULT_RETRY couldn't trigger, bypass */
1027 return ret;
1028
1029 /* VM_FAULT_RETRY cannot return errors */
1030 if (!*locked) {
1031 BUG_ON(ret < 0);
1032 BUG_ON(ret >= nr_pages);
1033 }
1034
1035 if (ret > 0) {
1036 nr_pages -= ret;
1037 pages_done += ret;
1038 if (!nr_pages)
1039 break;
1040 }
1041 if (*locked) {
1042 /*
1043 * VM_FAULT_RETRY didn't trigger or it was a
1044 * FOLL_NOWAIT.
1045 */
1046 if (!pages_done)
1047 pages_done = ret;
1048 break;
1049 }
1050 /*
1051 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1052 * For the prefault case (!pages) we only update counts.
1053 */
1054 if (likely(pages))
1055 pages += ret;
1056 start += ret << PAGE_SHIFT;
1057
1058 /*
1059 * Repeat on the address that fired VM_FAULT_RETRY
1060 * without FAULT_FLAG_ALLOW_RETRY but with
1061 * FAULT_FLAG_TRIED.
1062 */
1063 *locked = 1;
1064 lock_dropped = true;
1065 down_read(&mm->mmap_sem);
1066 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
1067 pages, NULL, NULL);
1068 if (ret != 1) {
1069 BUG_ON(ret > 1);
1070 if (!pages_done)
1071 pages_done = ret;
1072 break;
1073 }
1074 nr_pages--;
1075 pages_done++;
1076 if (!nr_pages)
1077 break;
1078 if (likely(pages))
1079 pages++;
1080 start += PAGE_SIZE;
1081 }
1082 if (lock_dropped && *locked) {
1083 /*
1084 * We must let the caller know we temporarily dropped the lock
1085 * and so the critical section protected by it was lost.
1086 */
1087 up_read(&mm->mmap_sem);
1088 *locked = 0;
1089 }
1090 return pages_done;
1091}
1092
1093/*
1094 * get_user_pages_remote() - pin user pages in memory
1095 * @tsk: the task_struct to use for page fault accounting, or
1096 * NULL if faults are not to be recorded.
1097 * @mm: mm_struct of target mm
1098 * @start: starting user address
1099 * @nr_pages: number of pages from start to pin
1100 * @gup_flags: flags modifying lookup behaviour
1101 * @pages: array that receives pointers to the pages pinned.
1102 * Should be at least nr_pages long. Or NULL, if caller
1103 * only intends to ensure the pages are faulted in.
1104 * @vmas: array of pointers to vmas corresponding to each page.
1105 * Or NULL if the caller does not require them.
1106 * @locked: pointer to lock flag indicating whether lock is held and
1107 * subsequently whether VM_FAULT_RETRY functionality can be
1108 * utilised. Lock must initially be held.
1109 *
1110 * Returns number of pages pinned. This may be fewer than the number
1111 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1112 * were pinned, returns -errno. Each page returned must be released
1113 * with a put_page() call when it is finished with. vmas will only
1114 * remain valid while mmap_sem is held.
1115 *
1116 * Must be called with mmap_sem held for read or write.
1117 *
1118 * get_user_pages walks a process's page tables and takes a reference to
1119 * each struct page that each user address corresponds to at a given
1120 * instant. That is, it takes the page that would be accessed if a user
1121 * thread accesses the given user virtual address at that instant.
1122 *
1123 * This does not guarantee that the page exists in the user mappings when
1124 * get_user_pages returns, and there may even be a completely different
1125 * page there in some cases (eg. if mmapped pagecache has been invalidated
1126 * and subsequently re faulted). However it does guarantee that the page
1127 * won't be freed completely. And mostly callers simply care that the page
1128 * contains data that was valid *at some point in time*. Typically, an IO
1129 * or similar operation cannot guarantee anything stronger anyway because
1130 * locks can't be held over the syscall boundary.
1131 *
1132 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1133 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1134 * be called after the page is finished with, and before put_page is called.
1135 *
1136 * get_user_pages is typically used for fewer-copy IO operations, to get a
1137 * handle on the memory by some means other than accesses via the user virtual
1138 * addresses. The pages may be submitted for DMA to devices or accessed via
1139 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1140 * use the correct cache flushing APIs.
1141 *
1142 * See also get_user_pages_fast, for performance critical applications.
1143 *
1144 * get_user_pages should be phased out in favor of
1145 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1146 * should use get_user_pages because it cannot pass
1147 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1148 */
1149long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1150 unsigned long start, unsigned long nr_pages,
1151 unsigned int gup_flags, struct page **pages,
1152 struct vm_area_struct **vmas, int *locked)
1153{
1154 /*
1155 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1156 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1157 * vmas. As there are no users of this flag in this call we simply
1158 * disallow this option for now.
1159 */
1160 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1161 return -EINVAL;
1162
1163 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1164 locked,
1165 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1166}
1167EXPORT_SYMBOL(get_user_pages_remote);
1168
1169/**
1170 * populate_vma_page_range() - populate a range of pages in the vma.
1171 * @vma: target vma
1172 * @start: start address
1173 * @end: end address
1174 * @nonblocking:
1175 *
1176 * This takes care of mlocking the pages too if VM_LOCKED is set.
1177 *
1178 * return 0 on success, negative error code on error.
1179 *
1180 * vma->vm_mm->mmap_sem must be held.
1181 *
1182 * If @nonblocking is NULL, it may be held for read or write and will
1183 * be unperturbed.
1184 *
1185 * If @nonblocking is non-NULL, it must held for read only and may be
1186 * released. If it's released, *@nonblocking will be set to 0.
1187 */
1188long populate_vma_page_range(struct vm_area_struct *vma,
1189 unsigned long start, unsigned long end, int *nonblocking)
1190{
1191 struct mm_struct *mm = vma->vm_mm;
1192 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1193 int gup_flags;
1194
1195 VM_BUG_ON(start & ~PAGE_MASK);
1196 VM_BUG_ON(end & ~PAGE_MASK);
1197 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1198 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1199 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1200
1201 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1202 if (vma->vm_flags & VM_LOCKONFAULT)
1203 gup_flags &= ~FOLL_POPULATE;
1204 /*
1205 * We want to touch writable mappings with a write fault in order
1206 * to break COW, except for shared mappings because these don't COW
1207 * and we would not want to dirty them for nothing.
1208 */
1209 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1210 gup_flags |= FOLL_WRITE;
1211
1212 /*
1213 * We want mlock to succeed for regions that have any permissions
1214 * other than PROT_NONE.
1215 */
1216 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1217 gup_flags |= FOLL_FORCE;
1218
1219 /*
1220 * We made sure addr is within a VMA, so the following will
1221 * not result in a stack expansion that recurses back here.
1222 */
1223 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1224 NULL, NULL, nonblocking);
1225}
1226
1227/*
1228 * __mm_populate - populate and/or mlock pages within a range of address space.
1229 *
1230 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1231 * flags. VMAs must be already marked with the desired vm_flags, and
1232 * mmap_sem must not be held.
1233 */
1234int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1235{
1236 struct mm_struct *mm = current->mm;
1237 unsigned long end, nstart, nend;
1238 struct vm_area_struct *vma = NULL;
1239 int locked = 0;
1240 long ret = 0;
1241
1242 end = start + len;
1243
1244 for (nstart = start; nstart < end; nstart = nend) {
1245 /*
1246 * We want to fault in pages for [nstart; end) address range.
1247 * Find first corresponding VMA.
1248 */
1249 if (!locked) {
1250 locked = 1;
1251 down_read(&mm->mmap_sem);
1252 vma = find_vma(mm, nstart);
1253 } else if (nstart >= vma->vm_end)
1254 vma = vma->vm_next;
1255 if (!vma || vma->vm_start >= end)
1256 break;
1257 /*
1258 * Set [nstart; nend) to intersection of desired address
1259 * range with the first VMA. Also, skip undesirable VMA types.
1260 */
1261 nend = min(end, vma->vm_end);
1262 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1263 continue;
1264 if (nstart < vma->vm_start)
1265 nstart = vma->vm_start;
1266 /*
1267 * Now fault in a range of pages. populate_vma_page_range()
1268 * double checks the vma flags, so that it won't mlock pages
1269 * if the vma was already munlocked.
1270 */
1271 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1272 if (ret < 0) {
1273 if (ignore_errors) {
1274 ret = 0;
1275 continue; /* continue at next VMA */
1276 }
1277 break;
1278 }
1279 nend = nstart + ret * PAGE_SIZE;
1280 ret = 0;
1281 }
1282 if (locked)
1283 up_read(&mm->mmap_sem);
1284 return ret; /* 0 or negative error code */
1285}
1286
1287/**
1288 * get_dump_page() - pin user page in memory while writing it to core dump
1289 * @addr: user address
1290 *
1291 * Returns struct page pointer of user page pinned for dump,
1292 * to be freed afterwards by put_page().
1293 *
1294 * Returns NULL on any kind of failure - a hole must then be inserted into
1295 * the corefile, to preserve alignment with its headers; and also returns
1296 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1297 * allowing a hole to be left in the corefile to save diskspace.
1298 *
1299 * Called without mmap_sem, but after all other threads have been killed.
1300 */
1301#ifdef CONFIG_ELF_CORE
1302struct page *get_dump_page(unsigned long addr)
1303{
1304 struct vm_area_struct *vma;
1305 struct page *page;
1306
1307 if (__get_user_pages(current, current->mm, addr, 1,
1308 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1309 NULL) < 1)
1310 return NULL;
1311 flush_cache_page(vma, addr, page_to_pfn(page));
1312 return page;
1313}
1314#endif /* CONFIG_ELF_CORE */
1315#else /* CONFIG_MMU */
1316static long __get_user_pages_locked(struct task_struct *tsk,
1317 struct mm_struct *mm, unsigned long start,
1318 unsigned long nr_pages, struct page **pages,
1319 struct vm_area_struct **vmas, int *locked,
1320 unsigned int foll_flags)
1321{
1322 struct vm_area_struct *vma;
1323 unsigned long vm_flags;
1324 int i;
1325
1326 /* calculate required read or write permissions.
1327 * If FOLL_FORCE is set, we only require the "MAY" flags.
1328 */
1329 vm_flags = (foll_flags & FOLL_WRITE) ?
1330 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1331 vm_flags &= (foll_flags & FOLL_FORCE) ?
1332 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1333
1334 for (i = 0; i < nr_pages; i++) {
1335 vma = find_vma(mm, start);
1336 if (!vma)
1337 goto finish_or_fault;
1338
1339 /* protect what we can, including chardevs */
1340 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1341 !(vm_flags & vma->vm_flags))
1342 goto finish_or_fault;
1343
1344 if (pages) {
1345 pages[i] = virt_to_page(start);
1346 if (pages[i])
1347 get_page(pages[i]);
1348 }
1349 if (vmas)
1350 vmas[i] = vma;
1351 start = (start + PAGE_SIZE) & PAGE_MASK;
1352 }
1353
1354 return i;
1355
1356finish_or_fault:
1357 return i ? : -EFAULT;
1358}
1359#endif /* !CONFIG_MMU */
1360
1361#if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1362static bool check_dax_vmas(struct vm_area_struct **vmas, long nr_pages)
1363{
1364 long i;
1365 struct vm_area_struct *vma_prev = NULL;
1366
1367 for (i = 0; i < nr_pages; i++) {
1368 struct vm_area_struct *vma = vmas[i];
1369
1370 if (vma == vma_prev)
1371 continue;
1372
1373 vma_prev = vma;
1374
1375 if (vma_is_fsdax(vma))
1376 return true;
1377 }
1378 return false;
1379}
1380
1381#ifdef CONFIG_CMA
1382static struct page *new_non_cma_page(struct page *page, unsigned long private)
1383{
1384 /*
1385 * We want to make sure we allocate the new page from the same node
1386 * as the source page.
1387 */
1388 int nid = page_to_nid(page);
1389 /*
1390 * Trying to allocate a page for migration. Ignore allocation
1391 * failure warnings. We don't force __GFP_THISNODE here because
1392 * this node here is the node where we have CMA reservation and
1393 * in some case these nodes will have really less non movable
1394 * allocation memory.
1395 */
1396 gfp_t gfp_mask = GFP_USER | __GFP_NOWARN;
1397
1398 if (PageHighMem(page))
1399 gfp_mask |= __GFP_HIGHMEM;
1400
1401#ifdef CONFIG_HUGETLB_PAGE
1402 if (PageHuge(page)) {
1403 struct hstate *h = page_hstate(page);
1404 /*
1405 * We don't want to dequeue from the pool because pool pages will
1406 * mostly be from the CMA region.
1407 */
1408 return alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1409 }
1410#endif
1411 if (PageTransHuge(page)) {
1412 struct page *thp;
1413 /*
1414 * ignore allocation failure warnings
1415 */
1416 gfp_t thp_gfpmask = GFP_TRANSHUGE | __GFP_NOWARN;
1417
1418 /*
1419 * Remove the movable mask so that we don't allocate from
1420 * CMA area again.
1421 */
1422 thp_gfpmask &= ~__GFP_MOVABLE;
1423 thp = __alloc_pages_node(nid, thp_gfpmask, HPAGE_PMD_ORDER);
1424 if (!thp)
1425 return NULL;
1426 prep_transhuge_page(thp);
1427 return thp;
1428 }
1429
1430 return __alloc_pages_node(nid, gfp_mask, 0);
1431}
1432
1433static long check_and_migrate_cma_pages(struct task_struct *tsk,
1434 struct mm_struct *mm,
1435 unsigned long start,
1436 unsigned long nr_pages,
1437 struct page **pages,
1438 struct vm_area_struct **vmas,
1439 unsigned int gup_flags)
1440{
1441 unsigned long i;
1442 unsigned long step;
1443 bool drain_allow = true;
1444 bool migrate_allow = true;
1445 LIST_HEAD(cma_page_list);
1446
1447check_again:
1448 for (i = 0; i < nr_pages;) {
1449
1450 struct page *head = compound_head(pages[i]);
1451
1452 /*
1453 * gup may start from a tail page. Advance step by the left
1454 * part.
1455 */
1456 step = compound_nr(head) - (pages[i] - head);
1457 /*
1458 * If we get a page from the CMA zone, since we are going to
1459 * be pinning these entries, we might as well move them out
1460 * of the CMA zone if possible.
1461 */
1462 if (is_migrate_cma_page(head)) {
1463 if (PageHuge(head))
1464 isolate_huge_page(head, &cma_page_list);
1465 else {
1466 if (!PageLRU(head) && drain_allow) {
1467 lru_add_drain_all();
1468 drain_allow = false;
1469 }
1470
1471 if (!isolate_lru_page(head)) {
1472 list_add_tail(&head->lru, &cma_page_list);
1473 mod_node_page_state(page_pgdat(head),
1474 NR_ISOLATED_ANON +
1475 page_is_file_cache(head),
1476 hpage_nr_pages(head));
1477 }
1478 }
1479 }
1480
1481 i += step;
1482 }
1483
1484 if (!list_empty(&cma_page_list)) {
1485 /*
1486 * drop the above get_user_pages reference.
1487 */
1488 for (i = 0; i < nr_pages; i++)
1489 put_page(pages[i]);
1490
1491 if (migrate_pages(&cma_page_list, new_non_cma_page,
1492 NULL, 0, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
1493 /*
1494 * some of the pages failed migration. Do get_user_pages
1495 * without migration.
1496 */
1497 migrate_allow = false;
1498
1499 if (!list_empty(&cma_page_list))
1500 putback_movable_pages(&cma_page_list);
1501 }
1502 /*
1503 * We did migrate all the pages, Try to get the page references
1504 * again migrating any new CMA pages which we failed to isolate
1505 * earlier.
1506 */
1507 nr_pages = __get_user_pages_locked(tsk, mm, start, nr_pages,
1508 pages, vmas, NULL,
1509 gup_flags);
1510
1511 if ((nr_pages > 0) && migrate_allow) {
1512 drain_allow = true;
1513 goto check_again;
1514 }
1515 }
1516
1517 return nr_pages;
1518}
1519#else
1520static long check_and_migrate_cma_pages(struct task_struct *tsk,
1521 struct mm_struct *mm,
1522 unsigned long start,
1523 unsigned long nr_pages,
1524 struct page **pages,
1525 struct vm_area_struct **vmas,
1526 unsigned int gup_flags)
1527{
1528 return nr_pages;
1529}
1530#endif /* CONFIG_CMA */
1531
1532/*
1533 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1534 * allows us to process the FOLL_LONGTERM flag.
1535 */
1536static long __gup_longterm_locked(struct task_struct *tsk,
1537 struct mm_struct *mm,
1538 unsigned long start,
1539 unsigned long nr_pages,
1540 struct page **pages,
1541 struct vm_area_struct **vmas,
1542 unsigned int gup_flags)
1543{
1544 struct vm_area_struct **vmas_tmp = vmas;
1545 unsigned long flags = 0;
1546 long rc, i;
1547
1548 if (gup_flags & FOLL_LONGTERM) {
1549 if (!pages)
1550 return -EINVAL;
1551
1552 if (!vmas_tmp) {
1553 vmas_tmp = kcalloc(nr_pages,
1554 sizeof(struct vm_area_struct *),
1555 GFP_KERNEL);
1556 if (!vmas_tmp)
1557 return -ENOMEM;
1558 }
1559 flags = memalloc_nocma_save();
1560 }
1561
1562 rc = __get_user_pages_locked(tsk, mm, start, nr_pages, pages,
1563 vmas_tmp, NULL, gup_flags);
1564
1565 if (gup_flags & FOLL_LONGTERM) {
1566 memalloc_nocma_restore(flags);
1567 if (rc < 0)
1568 goto out;
1569
1570 if (check_dax_vmas(vmas_tmp, rc)) {
1571 for (i = 0; i < rc; i++)
1572 put_page(pages[i]);
1573 rc = -EOPNOTSUPP;
1574 goto out;
1575 }
1576
1577 rc = check_and_migrate_cma_pages(tsk, mm, start, rc, pages,
1578 vmas_tmp, gup_flags);
1579 }
1580
1581out:
1582 if (vmas_tmp != vmas)
1583 kfree(vmas_tmp);
1584 return rc;
1585}
1586#else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1587static __always_inline long __gup_longterm_locked(struct task_struct *tsk,
1588 struct mm_struct *mm,
1589 unsigned long start,
1590 unsigned long nr_pages,
1591 struct page **pages,
1592 struct vm_area_struct **vmas,
1593 unsigned int flags)
1594{
1595 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1596 NULL, flags);
1597}
1598#endif /* CONFIG_FS_DAX || CONFIG_CMA */
1599
1600/*
1601 * This is the same as get_user_pages_remote(), just with a
1602 * less-flexible calling convention where we assume that the task
1603 * and mm being operated on are the current task's and don't allow
1604 * passing of a locked parameter. We also obviously don't pass
1605 * FOLL_REMOTE in here.
1606 */
1607long get_user_pages(unsigned long start, unsigned long nr_pages,
1608 unsigned int gup_flags, struct page **pages,
1609 struct vm_area_struct **vmas)
1610{
1611 return __gup_longterm_locked(current, current->mm, start, nr_pages,
1612 pages, vmas, gup_flags | FOLL_TOUCH);
1613}
1614EXPORT_SYMBOL(get_user_pages);
1615
1616/*
1617 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1618 * paths better by using either get_user_pages_locked() or
1619 * get_user_pages_unlocked().
1620 *
1621 * get_user_pages_locked() is suitable to replace the form:
1622 *
1623 * down_read(&mm->mmap_sem);
1624 * do_something()
1625 * get_user_pages(tsk, mm, ..., pages, NULL);
1626 * up_read(&mm->mmap_sem);
1627 *
1628 * to:
1629 *
1630 * int locked = 1;
1631 * down_read(&mm->mmap_sem);
1632 * do_something()
1633 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
1634 * if (locked)
1635 * up_read(&mm->mmap_sem);
1636 */
1637long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1638 unsigned int gup_flags, struct page **pages,
1639 int *locked)
1640{
1641 /*
1642 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1643 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1644 * vmas. As there are no users of this flag in this call we simply
1645 * disallow this option for now.
1646 */
1647 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1648 return -EINVAL;
1649
1650 return __get_user_pages_locked(current, current->mm, start, nr_pages,
1651 pages, NULL, locked,
1652 gup_flags | FOLL_TOUCH);
1653}
1654EXPORT_SYMBOL(get_user_pages_locked);
1655
1656/*
1657 * get_user_pages_unlocked() is suitable to replace the form:
1658 *
1659 * down_read(&mm->mmap_sem);
1660 * get_user_pages(tsk, mm, ..., pages, NULL);
1661 * up_read(&mm->mmap_sem);
1662 *
1663 * with:
1664 *
1665 * get_user_pages_unlocked(tsk, mm, ..., pages);
1666 *
1667 * It is functionally equivalent to get_user_pages_fast so
1668 * get_user_pages_fast should be used instead if specific gup_flags
1669 * (e.g. FOLL_FORCE) are not required.
1670 */
1671long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1672 struct page **pages, unsigned int gup_flags)
1673{
1674 struct mm_struct *mm = current->mm;
1675 int locked = 1;
1676 long ret;
1677
1678 /*
1679 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1680 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1681 * vmas. As there are no users of this flag in this call we simply
1682 * disallow this option for now.
1683 */
1684 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1685 return -EINVAL;
1686
1687 down_read(&mm->mmap_sem);
1688 ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL,
1689 &locked, gup_flags | FOLL_TOUCH);
1690 if (locked)
1691 up_read(&mm->mmap_sem);
1692 return ret;
1693}
1694EXPORT_SYMBOL(get_user_pages_unlocked);
1695
1696/*
1697 * Fast GUP
1698 *
1699 * get_user_pages_fast attempts to pin user pages by walking the page
1700 * tables directly and avoids taking locks. Thus the walker needs to be
1701 * protected from page table pages being freed from under it, and should
1702 * block any THP splits.
1703 *
1704 * One way to achieve this is to have the walker disable interrupts, and
1705 * rely on IPIs from the TLB flushing code blocking before the page table
1706 * pages are freed. This is unsuitable for architectures that do not need
1707 * to broadcast an IPI when invalidating TLBs.
1708 *
1709 * Another way to achieve this is to batch up page table containing pages
1710 * belonging to more than one mm_user, then rcu_sched a callback to free those
1711 * pages. Disabling interrupts will allow the fast_gup walker to both block
1712 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1713 * (which is a relatively rare event). The code below adopts this strategy.
1714 *
1715 * Before activating this code, please be aware that the following assumptions
1716 * are currently made:
1717 *
1718 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1719 * free pages containing page tables or TLB flushing requires IPI broadcast.
1720 *
1721 * *) ptes can be read atomically by the architecture.
1722 *
1723 * *) access_ok is sufficient to validate userspace address ranges.
1724 *
1725 * The last two assumptions can be relaxed by the addition of helper functions.
1726 *
1727 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1728 */
1729#ifdef CONFIG_HAVE_FAST_GUP
1730#ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
1731/*
1732 * WARNING: only to be used in the get_user_pages_fast() implementation.
1733 *
1734 * With get_user_pages_fast(), we walk down the pagetables without taking any
1735 * locks. For this we would like to load the pointers atomically, but sometimes
1736 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
1737 * we do have is the guarantee that a PTE will only either go from not present
1738 * to present, or present to not present or both -- it will not switch to a
1739 * completely different present page without a TLB flush in between; something
1740 * that we are blocking by holding interrupts off.
1741 *
1742 * Setting ptes from not present to present goes:
1743 *
1744 * ptep->pte_high = h;
1745 * smp_wmb();
1746 * ptep->pte_low = l;
1747 *
1748 * And present to not present goes:
1749 *
1750 * ptep->pte_low = 0;
1751 * smp_wmb();
1752 * ptep->pte_high = 0;
1753 *
1754 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
1755 * We load pte_high *after* loading pte_low, which ensures we don't see an older
1756 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
1757 * picked up a changed pte high. We might have gotten rubbish values from
1758 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
1759 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
1760 * operates on present ptes we're safe.
1761 */
1762static inline pte_t gup_get_pte(pte_t *ptep)
1763{
1764 pte_t pte;
1765
1766 do {
1767 pte.pte_low = ptep->pte_low;
1768 smp_rmb();
1769 pte.pte_high = ptep->pte_high;
1770 smp_rmb();
1771 } while (unlikely(pte.pte_low != ptep->pte_low));
1772
1773 return pte;
1774}
1775#else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
1776/*
1777 * We require that the PTE can be read atomically.
1778 */
1779static inline pte_t gup_get_pte(pte_t *ptep)
1780{
1781 return READ_ONCE(*ptep);
1782}
1783#endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
1784
1785static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
1786 struct page **pages)
1787{
1788 while ((*nr) - nr_start) {
1789 struct page *page = pages[--(*nr)];
1790
1791 ClearPageReferenced(page);
1792 put_page(page);
1793 }
1794}
1795
1796/*
1797 * Return the compund head page with ref appropriately incremented,
1798 * or NULL if that failed.
1799 */
1800static inline struct page *try_get_compound_head(struct page *page, int refs)
1801{
1802 struct page *head = compound_head(page);
1803 if (WARN_ON_ONCE(page_ref_count(head) < 0))
1804 return NULL;
1805 if (unlikely(!page_cache_add_speculative(head, refs)))
1806 return NULL;
1807 return head;
1808}
1809
1810#ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
1811static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1812 unsigned int flags, struct page **pages, int *nr)
1813{
1814 struct dev_pagemap *pgmap = NULL;
1815 int nr_start = *nr, ret = 0;
1816 pte_t *ptep, *ptem;
1817
1818 ptem = ptep = pte_offset_map(&pmd, addr);
1819 do {
1820 pte_t pte = gup_get_pte(ptep);
1821 struct page *head, *page;
1822
1823 /*
1824 * Similar to the PMD case below, NUMA hinting must take slow
1825 * path using the pte_protnone check.
1826 */
1827 if (pte_protnone(pte))
1828 goto pte_unmap;
1829
1830 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
1831 goto pte_unmap;
1832
1833 if (pte_devmap(pte)) {
1834 if (unlikely(flags & FOLL_LONGTERM))
1835 goto pte_unmap;
1836
1837 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1838 if (unlikely(!pgmap)) {
1839 undo_dev_pagemap(nr, nr_start, pages);
1840 goto pte_unmap;
1841 }
1842 } else if (pte_special(pte))
1843 goto pte_unmap;
1844
1845 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1846 page = pte_page(pte);
1847
1848 head = try_get_compound_head(page, 1);
1849 if (!head)
1850 goto pte_unmap;
1851
1852 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1853 put_page(head);
1854 goto pte_unmap;
1855 }
1856
1857 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1858
1859 SetPageReferenced(page);
1860 pages[*nr] = page;
1861 (*nr)++;
1862
1863 } while (ptep++, addr += PAGE_SIZE, addr != end);
1864
1865 ret = 1;
1866
1867pte_unmap:
1868 if (pgmap)
1869 put_dev_pagemap(pgmap);
1870 pte_unmap(ptem);
1871 return ret;
1872}
1873#else
1874
1875/*
1876 * If we can't determine whether or not a pte is special, then fail immediately
1877 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1878 * to be special.
1879 *
1880 * For a futex to be placed on a THP tail page, get_futex_key requires a
1881 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1882 * useful to have gup_huge_pmd even if we can't operate on ptes.
1883 */
1884static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1885 unsigned int flags, struct page **pages, int *nr)
1886{
1887 return 0;
1888}
1889#endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
1890
1891#if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1892static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1893 unsigned long end, struct page **pages, int *nr)
1894{
1895 int nr_start = *nr;
1896 struct dev_pagemap *pgmap = NULL;
1897
1898 do {
1899 struct page *page = pfn_to_page(pfn);
1900
1901 pgmap = get_dev_pagemap(pfn, pgmap);
1902 if (unlikely(!pgmap)) {
1903 undo_dev_pagemap(nr, nr_start, pages);
1904 return 0;
1905 }
1906 SetPageReferenced(page);
1907 pages[*nr] = page;
1908 get_page(page);
1909 (*nr)++;
1910 pfn++;
1911 } while (addr += PAGE_SIZE, addr != end);
1912
1913 if (pgmap)
1914 put_dev_pagemap(pgmap);
1915 return 1;
1916}
1917
1918static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1919 unsigned long end, struct page **pages, int *nr)
1920{
1921 unsigned long fault_pfn;
1922 int nr_start = *nr;
1923
1924 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1925 if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1926 return 0;
1927
1928 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1929 undo_dev_pagemap(nr, nr_start, pages);
1930 return 0;
1931 }
1932 return 1;
1933}
1934
1935static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1936 unsigned long end, struct page **pages, int *nr)
1937{
1938 unsigned long fault_pfn;
1939 int nr_start = *nr;
1940
1941 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1942 if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1943 return 0;
1944
1945 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1946 undo_dev_pagemap(nr, nr_start, pages);
1947 return 0;
1948 }
1949 return 1;
1950}
1951#else
1952static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1953 unsigned long end, struct page **pages, int *nr)
1954{
1955 BUILD_BUG();
1956 return 0;
1957}
1958
1959static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
1960 unsigned long end, struct page **pages, int *nr)
1961{
1962 BUILD_BUG();
1963 return 0;
1964}
1965#endif
1966
1967#ifdef CONFIG_ARCH_HAS_HUGEPD
1968static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
1969 unsigned long sz)
1970{
1971 unsigned long __boundary = (addr + sz) & ~(sz-1);
1972 return (__boundary - 1 < end - 1) ? __boundary : end;
1973}
1974
1975static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
1976 unsigned long end, unsigned int flags,
1977 struct page **pages, int *nr)
1978{
1979 unsigned long pte_end;
1980 struct page *head, *page;
1981 pte_t pte;
1982 int refs;
1983
1984 pte_end = (addr + sz) & ~(sz-1);
1985 if (pte_end < end)
1986 end = pte_end;
1987
1988 pte = READ_ONCE(*ptep);
1989
1990 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
1991 return 0;
1992
1993 /* hugepages are never "special" */
1994 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1995
1996 refs = 0;
1997 head = pte_page(pte);
1998
1999 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2000 do {
2001 VM_BUG_ON(compound_head(page) != head);
2002 pages[*nr] = page;
2003 (*nr)++;
2004 page++;
2005 refs++;
2006 } while (addr += PAGE_SIZE, addr != end);
2007
2008 head = try_get_compound_head(head, refs);
2009 if (!head) {
2010 *nr -= refs;
2011 return 0;
2012 }
2013
2014 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2015 /* Could be optimized better */
2016 *nr -= refs;
2017 while (refs--)
2018 put_page(head);
2019 return 0;
2020 }
2021
2022 SetPageReferenced(head);
2023 return 1;
2024}
2025
2026static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2027 unsigned int pdshift, unsigned long end, unsigned int flags,
2028 struct page **pages, int *nr)
2029{
2030 pte_t *ptep;
2031 unsigned long sz = 1UL << hugepd_shift(hugepd);
2032 unsigned long next;
2033
2034 ptep = hugepte_offset(hugepd, addr, pdshift);
2035 do {
2036 next = hugepte_addr_end(addr, end, sz);
2037 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2038 return 0;
2039 } while (ptep++, addr = next, addr != end);
2040
2041 return 1;
2042}
2043#else
2044static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2045 unsigned int pdshift, unsigned long end, unsigned int flags,
2046 struct page **pages, int *nr)
2047{
2048 return 0;
2049}
2050#endif /* CONFIG_ARCH_HAS_HUGEPD */
2051
2052static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2053 unsigned long end, unsigned int flags,
2054 struct page **pages, int *nr)
2055{
2056 struct page *head, *page;
2057 int refs;
2058
2059 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2060 return 0;
2061
2062 if (pmd_devmap(orig)) {
2063 if (unlikely(flags & FOLL_LONGTERM))
2064 return 0;
2065 return __gup_device_huge_pmd(orig, pmdp, addr, end, pages, nr);
2066 }
2067
2068 refs = 0;
2069 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2070 do {
2071 pages[*nr] = page;
2072 (*nr)++;
2073 page++;
2074 refs++;
2075 } while (addr += PAGE_SIZE, addr != end);
2076
2077 head = try_get_compound_head(pmd_page(orig), refs);
2078 if (!head) {
2079 *nr -= refs;
2080 return 0;
2081 }
2082
2083 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2084 *nr -= refs;
2085 while (refs--)
2086 put_page(head);
2087 return 0;
2088 }
2089
2090 SetPageReferenced(head);
2091 return 1;
2092}
2093
2094static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2095 unsigned long end, unsigned int flags, struct page **pages, int *nr)
2096{
2097 struct page *head, *page;
2098 int refs;
2099
2100 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2101 return 0;
2102
2103 if (pud_devmap(orig)) {
2104 if (unlikely(flags & FOLL_LONGTERM))
2105 return 0;
2106 return __gup_device_huge_pud(orig, pudp, addr, end, pages, nr);
2107 }
2108
2109 refs = 0;
2110 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2111 do {
2112 pages[*nr] = page;
2113 (*nr)++;
2114 page++;
2115 refs++;
2116 } while (addr += PAGE_SIZE, addr != end);
2117
2118 head = try_get_compound_head(pud_page(orig), refs);
2119 if (!head) {
2120 *nr -= refs;
2121 return 0;
2122 }
2123
2124 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2125 *nr -= refs;
2126 while (refs--)
2127 put_page(head);
2128 return 0;
2129 }
2130
2131 SetPageReferenced(head);
2132 return 1;
2133}
2134
2135static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2136 unsigned long end, unsigned int flags,
2137 struct page **pages, int *nr)
2138{
2139 int refs;
2140 struct page *head, *page;
2141
2142 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2143 return 0;
2144
2145 BUILD_BUG_ON(pgd_devmap(orig));
2146 refs = 0;
2147 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2148 do {
2149 pages[*nr] = page;
2150 (*nr)++;
2151 page++;
2152 refs++;
2153 } while (addr += PAGE_SIZE, addr != end);
2154
2155 head = try_get_compound_head(pgd_page(orig), refs);
2156 if (!head) {
2157 *nr -= refs;
2158 return 0;
2159 }
2160
2161 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2162 *nr -= refs;
2163 while (refs--)
2164 put_page(head);
2165 return 0;
2166 }
2167
2168 SetPageReferenced(head);
2169 return 1;
2170}
2171
2172static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
2173 unsigned int flags, struct page **pages, int *nr)
2174{
2175 unsigned long next;
2176 pmd_t *pmdp;
2177
2178 pmdp = pmd_offset(&pud, addr);
2179 do {
2180 pmd_t pmd = READ_ONCE(*pmdp);
2181
2182 next = pmd_addr_end(addr, end);
2183 if (!pmd_present(pmd))
2184 return 0;
2185
2186 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2187 pmd_devmap(pmd))) {
2188 /*
2189 * NUMA hinting faults need to be handled in the GUP
2190 * slowpath for accounting purposes and so that they
2191 * can be serialised against THP migration.
2192 */
2193 if (pmd_protnone(pmd))
2194 return 0;
2195
2196 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2197 pages, nr))
2198 return 0;
2199
2200 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2201 /*
2202 * architecture have different format for hugetlbfs
2203 * pmd format and THP pmd format
2204 */
2205 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2206 PMD_SHIFT, next, flags, pages, nr))
2207 return 0;
2208 } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2209 return 0;
2210 } while (pmdp++, addr = next, addr != end);
2211
2212 return 1;
2213}
2214
2215static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
2216 unsigned int flags, struct page **pages, int *nr)
2217{
2218 unsigned long next;
2219 pud_t *pudp;
2220
2221 pudp = pud_offset(&p4d, addr);
2222 do {
2223 pud_t pud = READ_ONCE(*pudp);
2224
2225 next = pud_addr_end(addr, end);
2226 if (pud_none(pud))
2227 return 0;
2228 if (unlikely(pud_huge(pud))) {
2229 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2230 pages, nr))
2231 return 0;
2232 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2233 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2234 PUD_SHIFT, next, flags, pages, nr))
2235 return 0;
2236 } else if (!gup_pmd_range(pud, addr, next, flags, pages, nr))
2237 return 0;
2238 } while (pudp++, addr = next, addr != end);
2239
2240 return 1;
2241}
2242
2243static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
2244 unsigned int flags, struct page **pages, int *nr)
2245{
2246 unsigned long next;
2247 p4d_t *p4dp;
2248
2249 p4dp = p4d_offset(&pgd, addr);
2250 do {
2251 p4d_t p4d = READ_ONCE(*p4dp);
2252
2253 next = p4d_addr_end(addr, end);
2254 if (p4d_none(p4d))
2255 return 0;
2256 BUILD_BUG_ON(p4d_huge(p4d));
2257 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2258 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2259 P4D_SHIFT, next, flags, pages, nr))
2260 return 0;
2261 } else if (!gup_pud_range(p4d, addr, next, flags, pages, nr))
2262 return 0;
2263 } while (p4dp++, addr = next, addr != end);
2264
2265 return 1;
2266}
2267
2268static void gup_pgd_range(unsigned long addr, unsigned long end,
2269 unsigned int flags, struct page **pages, int *nr)
2270{
2271 unsigned long next;
2272 pgd_t *pgdp;
2273
2274 pgdp = pgd_offset(current->mm, addr);
2275 do {
2276 pgd_t pgd = READ_ONCE(*pgdp);
2277
2278 next = pgd_addr_end(addr, end);
2279 if (pgd_none(pgd))
2280 return;
2281 if (unlikely(pgd_huge(pgd))) {
2282 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2283 pages, nr))
2284 return;
2285 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2286 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2287 PGDIR_SHIFT, next, flags, pages, nr))
2288 return;
2289 } else if (!gup_p4d_range(pgd, addr, next, flags, pages, nr))
2290 return;
2291 } while (pgdp++, addr = next, addr != end);
2292}
2293#else
2294static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2295 unsigned int flags, struct page **pages, int *nr)
2296{
2297}
2298#endif /* CONFIG_HAVE_FAST_GUP */
2299
2300#ifndef gup_fast_permitted
2301/*
2302 * Check if it's allowed to use __get_user_pages_fast() for the range, or
2303 * we need to fall back to the slow version:
2304 */
2305static bool gup_fast_permitted(unsigned long start, unsigned long end)
2306{
2307 return true;
2308}
2309#endif
2310
2311/*
2312 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2313 * the regular GUP.
2314 * Note a difference with get_user_pages_fast: this always returns the
2315 * number of pages pinned, 0 if no pages were pinned.
2316 *
2317 * If the architecture does not support this function, simply return with no
2318 * pages pinned.
2319 */
2320int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
2321 struct page **pages)
2322{
2323 unsigned long len, end;
2324 unsigned long flags;
2325 int nr = 0;
2326
2327 start = untagged_addr(start) & PAGE_MASK;
2328 len = (unsigned long) nr_pages << PAGE_SHIFT;
2329 end = start + len;
2330
2331 if (end <= start)
2332 return 0;
2333 if (unlikely(!access_ok((void __user *)start, len)))
2334 return 0;
2335
2336 /*
2337 * Disable interrupts. We use the nested form as we can already have
2338 * interrupts disabled by get_futex_key.
2339 *
2340 * With interrupts disabled, we block page table pages from being
2341 * freed from under us. See struct mmu_table_batch comments in
2342 * include/asm-generic/tlb.h for more details.
2343 *
2344 * We do not adopt an rcu_read_lock(.) here as we also want to
2345 * block IPIs that come from THPs splitting.
2346 */
2347
2348 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
2349 gup_fast_permitted(start, end)) {
2350 local_irq_save(flags);
2351 gup_pgd_range(start, end, write ? FOLL_WRITE : 0, pages, &nr);
2352 local_irq_restore(flags);
2353 }
2354
2355 return nr;
2356}
2357EXPORT_SYMBOL_GPL(__get_user_pages_fast);
2358
2359static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2360 unsigned int gup_flags, struct page **pages)
2361{
2362 int ret;
2363
2364 /*
2365 * FIXME: FOLL_LONGTERM does not work with
2366 * get_user_pages_unlocked() (see comments in that function)
2367 */
2368 if (gup_flags & FOLL_LONGTERM) {
2369 down_read(¤t->mm->mmap_sem);
2370 ret = __gup_longterm_locked(current, current->mm,
2371 start, nr_pages,
2372 pages, NULL, gup_flags);
2373 up_read(¤t->mm->mmap_sem);
2374 } else {
2375 ret = get_user_pages_unlocked(start, nr_pages,
2376 pages, gup_flags);
2377 }
2378
2379 return ret;
2380}
2381
2382/**
2383 * get_user_pages_fast() - pin user pages in memory
2384 * @start: starting user address
2385 * @nr_pages: number of pages from start to pin
2386 * @gup_flags: flags modifying pin behaviour
2387 * @pages: array that receives pointers to the pages pinned.
2388 * Should be at least nr_pages long.
2389 *
2390 * Attempt to pin user pages in memory without taking mm->mmap_sem.
2391 * If not successful, it will fall back to taking the lock and
2392 * calling get_user_pages().
2393 *
2394 * Returns number of pages pinned. This may be fewer than the number
2395 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2396 * were pinned, returns -errno.
2397 */
2398int get_user_pages_fast(unsigned long start, int nr_pages,
2399 unsigned int gup_flags, struct page **pages)
2400{
2401 unsigned long addr, len, end;
2402 int nr = 0, ret = 0;
2403
2404 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM)))
2405 return -EINVAL;
2406
2407 start = untagged_addr(start) & PAGE_MASK;
2408 addr = start;
2409 len = (unsigned long) nr_pages << PAGE_SHIFT;
2410 end = start + len;
2411
2412 if (end <= start)
2413 return 0;
2414 if (unlikely(!access_ok((void __user *)start, len)))
2415 return -EFAULT;
2416
2417 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
2418 gup_fast_permitted(start, end)) {
2419 local_irq_disable();
2420 gup_pgd_range(addr, end, gup_flags, pages, &nr);
2421 local_irq_enable();
2422 ret = nr;
2423 }
2424
2425 if (nr < nr_pages) {
2426 /* Try to get the remaining pages with get_user_pages */
2427 start += nr << PAGE_SHIFT;
2428 pages += nr;
2429
2430 ret = __gup_longterm_unlocked(start, nr_pages - nr,
2431 gup_flags, pages);
2432
2433 /* Have to be a bit careful with return values */
2434 if (nr > 0) {
2435 if (ret < 0)
2436 ret = nr;
2437 else
2438 ret += nr;
2439 }
2440 }
2441
2442 return ret;
2443}
2444EXPORT_SYMBOL_GPL(get_user_pages_fast);
1// SPDX-License-Identifier: GPL-2.0-only
2#include <linux/kernel.h>
3#include <linux/errno.h>
4#include <linux/err.h>
5#include <linux/spinlock.h>
6
7#include <linux/mm.h>
8#include <linux/memremap.h>
9#include <linux/pagemap.h>
10#include <linux/rmap.h>
11#include <linux/swap.h>
12#include <linux/swapops.h>
13#include <linux/secretmem.h>
14
15#include <linux/sched/signal.h>
16#include <linux/rwsem.h>
17#include <linux/hugetlb.h>
18#include <linux/migrate.h>
19#include <linux/mm_inline.h>
20#include <linux/sched/mm.h>
21#include <linux/shmem_fs.h>
22
23#include <asm/mmu_context.h>
24#include <asm/tlbflush.h>
25
26#include "internal.h"
27
28struct follow_page_context {
29 struct dev_pagemap *pgmap;
30 unsigned int page_mask;
31};
32
33static inline void sanity_check_pinned_pages(struct page **pages,
34 unsigned long npages)
35{
36 if (!IS_ENABLED(CONFIG_DEBUG_VM))
37 return;
38
39 /*
40 * We only pin anonymous pages if they are exclusive. Once pinned, we
41 * can no longer turn them possibly shared and PageAnonExclusive() will
42 * stick around until the page is freed.
43 *
44 * We'd like to verify that our pinned anonymous pages are still mapped
45 * exclusively. The issue with anon THP is that we don't know how
46 * they are/were mapped when pinning them. However, for anon
47 * THP we can assume that either the given page (PTE-mapped THP) or
48 * the head page (PMD-mapped THP) should be PageAnonExclusive(). If
49 * neither is the case, there is certainly something wrong.
50 */
51 for (; npages; npages--, pages++) {
52 struct page *page = *pages;
53 struct folio *folio = page_folio(page);
54
55 if (is_zero_page(page) ||
56 !folio_test_anon(folio))
57 continue;
58 if (!folio_test_large(folio) || folio_test_hugetlb(folio))
59 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page), page);
60 else
61 /* Either a PTE-mapped or a PMD-mapped THP. */
62 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page) &&
63 !PageAnonExclusive(page), page);
64 }
65}
66
67/*
68 * Return the folio with ref appropriately incremented,
69 * or NULL if that failed.
70 */
71static inline struct folio *try_get_folio(struct page *page, int refs)
72{
73 struct folio *folio;
74
75retry:
76 folio = page_folio(page);
77 if (WARN_ON_ONCE(folio_ref_count(folio) < 0))
78 return NULL;
79 if (unlikely(!folio_ref_try_add_rcu(folio, refs)))
80 return NULL;
81
82 /*
83 * At this point we have a stable reference to the folio; but it
84 * could be that between calling page_folio() and the refcount
85 * increment, the folio was split, in which case we'd end up
86 * holding a reference on a folio that has nothing to do with the page
87 * we were given anymore.
88 * So now that the folio is stable, recheck that the page still
89 * belongs to this folio.
90 */
91 if (unlikely(page_folio(page) != folio)) {
92 if (!put_devmap_managed_page_refs(&folio->page, refs))
93 folio_put_refs(folio, refs);
94 goto retry;
95 }
96
97 return folio;
98}
99
100/**
101 * try_grab_folio() - Attempt to get or pin a folio.
102 * @page: pointer to page to be grabbed
103 * @refs: the value to (effectively) add to the folio's refcount
104 * @flags: gup flags: these are the FOLL_* flag values.
105 *
106 * "grab" names in this file mean, "look at flags to decide whether to use
107 * FOLL_PIN or FOLL_GET behavior, when incrementing the folio's refcount.
108 *
109 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
110 * same time. (That's true throughout the get_user_pages*() and
111 * pin_user_pages*() APIs.) Cases:
112 *
113 * FOLL_GET: folio's refcount will be incremented by @refs.
114 *
115 * FOLL_PIN on large folios: folio's refcount will be incremented by
116 * @refs, and its pincount will be incremented by @refs.
117 *
118 * FOLL_PIN on single-page folios: folio's refcount will be incremented by
119 * @refs * GUP_PIN_COUNTING_BIAS.
120 *
121 * Return: The folio containing @page (with refcount appropriately
122 * incremented) for success, or NULL upon failure. If neither FOLL_GET
123 * nor FOLL_PIN was set, that's considered failure, and furthermore,
124 * a likely bug in the caller, so a warning is also emitted.
125 */
126struct folio *try_grab_folio(struct page *page, int refs, unsigned int flags)
127{
128 struct folio *folio;
129
130 if (WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == 0))
131 return NULL;
132
133 if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
134 return NULL;
135
136 if (flags & FOLL_GET)
137 return try_get_folio(page, refs);
138
139 /* FOLL_PIN is set */
140
141 /*
142 * Don't take a pin on the zero page - it's not going anywhere
143 * and it is used in a *lot* of places.
144 */
145 if (is_zero_page(page))
146 return page_folio(page);
147
148 folio = try_get_folio(page, refs);
149 if (!folio)
150 return NULL;
151
152 /*
153 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
154 * right zone, so fail and let the caller fall back to the slow
155 * path.
156 */
157 if (unlikely((flags & FOLL_LONGTERM) &&
158 !folio_is_longterm_pinnable(folio))) {
159 if (!put_devmap_managed_page_refs(&folio->page, refs))
160 folio_put_refs(folio, refs);
161 return NULL;
162 }
163
164 /*
165 * When pinning a large folio, use an exact count to track it.
166 *
167 * However, be sure to *also* increment the normal folio
168 * refcount field at least once, so that the folio really
169 * is pinned. That's why the refcount from the earlier
170 * try_get_folio() is left intact.
171 */
172 if (folio_test_large(folio))
173 atomic_add(refs, &folio->_pincount);
174 else
175 folio_ref_add(folio,
176 refs * (GUP_PIN_COUNTING_BIAS - 1));
177 /*
178 * Adjust the pincount before re-checking the PTE for changes.
179 * This is essentially a smp_mb() and is paired with a memory
180 * barrier in folio_try_share_anon_rmap_*().
181 */
182 smp_mb__after_atomic();
183
184 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs);
185
186 return folio;
187}
188
189static void gup_put_folio(struct folio *folio, int refs, unsigned int flags)
190{
191 if (flags & FOLL_PIN) {
192 if (is_zero_folio(folio))
193 return;
194 node_stat_mod_folio(folio, NR_FOLL_PIN_RELEASED, refs);
195 if (folio_test_large(folio))
196 atomic_sub(refs, &folio->_pincount);
197 else
198 refs *= GUP_PIN_COUNTING_BIAS;
199 }
200
201 if (!put_devmap_managed_page_refs(&folio->page, refs))
202 folio_put_refs(folio, refs);
203}
204
205/**
206 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
207 * @page: pointer to page to be grabbed
208 * @flags: gup flags: these are the FOLL_* flag values.
209 *
210 * This might not do anything at all, depending on the flags argument.
211 *
212 * "grab" names in this file mean, "look at flags to decide whether to use
213 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
214 *
215 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
216 * time. Cases: please see the try_grab_folio() documentation, with
217 * "refs=1".
218 *
219 * Return: 0 for success, or if no action was required (if neither FOLL_PIN
220 * nor FOLL_GET was set, nothing is done). A negative error code for failure:
221 *
222 * -ENOMEM FOLL_GET or FOLL_PIN was set, but the page could not
223 * be grabbed.
224 */
225int __must_check try_grab_page(struct page *page, unsigned int flags)
226{
227 struct folio *folio = page_folio(page);
228
229 if (WARN_ON_ONCE(folio_ref_count(folio) <= 0))
230 return -ENOMEM;
231
232 if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
233 return -EREMOTEIO;
234
235 if (flags & FOLL_GET)
236 folio_ref_inc(folio);
237 else if (flags & FOLL_PIN) {
238 /*
239 * Don't take a pin on the zero page - it's not going anywhere
240 * and it is used in a *lot* of places.
241 */
242 if (is_zero_page(page))
243 return 0;
244
245 /*
246 * Similar to try_grab_folio(): be sure to *also*
247 * increment the normal page refcount field at least once,
248 * so that the page really is pinned.
249 */
250 if (folio_test_large(folio)) {
251 folio_ref_add(folio, 1);
252 atomic_add(1, &folio->_pincount);
253 } else {
254 folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
255 }
256
257 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, 1);
258 }
259
260 return 0;
261}
262
263/**
264 * unpin_user_page() - release a dma-pinned page
265 * @page: pointer to page to be released
266 *
267 * Pages that were pinned via pin_user_pages*() must be released via either
268 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
269 * that such pages can be separately tracked and uniquely handled. In
270 * particular, interactions with RDMA and filesystems need special handling.
271 */
272void unpin_user_page(struct page *page)
273{
274 sanity_check_pinned_pages(&page, 1);
275 gup_put_folio(page_folio(page), 1, FOLL_PIN);
276}
277EXPORT_SYMBOL(unpin_user_page);
278
279/**
280 * folio_add_pin - Try to get an additional pin on a pinned folio
281 * @folio: The folio to be pinned
282 *
283 * Get an additional pin on a folio we already have a pin on. Makes no change
284 * if the folio is a zero_page.
285 */
286void folio_add_pin(struct folio *folio)
287{
288 if (is_zero_folio(folio))
289 return;
290
291 /*
292 * Similar to try_grab_folio(): be sure to *also* increment the normal
293 * page refcount field at least once, so that the page really is
294 * pinned.
295 */
296 if (folio_test_large(folio)) {
297 WARN_ON_ONCE(atomic_read(&folio->_pincount) < 1);
298 folio_ref_inc(folio);
299 atomic_inc(&folio->_pincount);
300 } else {
301 WARN_ON_ONCE(folio_ref_count(folio) < GUP_PIN_COUNTING_BIAS);
302 folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
303 }
304}
305
306static inline struct folio *gup_folio_range_next(struct page *start,
307 unsigned long npages, unsigned long i, unsigned int *ntails)
308{
309 struct page *next = nth_page(start, i);
310 struct folio *folio = page_folio(next);
311 unsigned int nr = 1;
312
313 if (folio_test_large(folio))
314 nr = min_t(unsigned int, npages - i,
315 folio_nr_pages(folio) - folio_page_idx(folio, next));
316
317 *ntails = nr;
318 return folio;
319}
320
321static inline struct folio *gup_folio_next(struct page **list,
322 unsigned long npages, unsigned long i, unsigned int *ntails)
323{
324 struct folio *folio = page_folio(list[i]);
325 unsigned int nr;
326
327 for (nr = i + 1; nr < npages; nr++) {
328 if (page_folio(list[nr]) != folio)
329 break;
330 }
331
332 *ntails = nr - i;
333 return folio;
334}
335
336/**
337 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
338 * @pages: array of pages to be maybe marked dirty, and definitely released.
339 * @npages: number of pages in the @pages array.
340 * @make_dirty: whether to mark the pages dirty
341 *
342 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
343 * variants called on that page.
344 *
345 * For each page in the @pages array, make that page (or its head page, if a
346 * compound page) dirty, if @make_dirty is true, and if the page was previously
347 * listed as clean. In any case, releases all pages using unpin_user_page(),
348 * possibly via unpin_user_pages(), for the non-dirty case.
349 *
350 * Please see the unpin_user_page() documentation for details.
351 *
352 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
353 * required, then the caller should a) verify that this is really correct,
354 * because _lock() is usually required, and b) hand code it:
355 * set_page_dirty_lock(), unpin_user_page().
356 *
357 */
358void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
359 bool make_dirty)
360{
361 unsigned long i;
362 struct folio *folio;
363 unsigned int nr;
364
365 if (!make_dirty) {
366 unpin_user_pages(pages, npages);
367 return;
368 }
369
370 sanity_check_pinned_pages(pages, npages);
371 for (i = 0; i < npages; i += nr) {
372 folio = gup_folio_next(pages, npages, i, &nr);
373 /*
374 * Checking PageDirty at this point may race with
375 * clear_page_dirty_for_io(), but that's OK. Two key
376 * cases:
377 *
378 * 1) This code sees the page as already dirty, so it
379 * skips the call to set_page_dirty(). That could happen
380 * because clear_page_dirty_for_io() called
381 * page_mkclean(), followed by set_page_dirty().
382 * However, now the page is going to get written back,
383 * which meets the original intention of setting it
384 * dirty, so all is well: clear_page_dirty_for_io() goes
385 * on to call TestClearPageDirty(), and write the page
386 * back.
387 *
388 * 2) This code sees the page as clean, so it calls
389 * set_page_dirty(). The page stays dirty, despite being
390 * written back, so it gets written back again in the
391 * next writeback cycle. This is harmless.
392 */
393 if (!folio_test_dirty(folio)) {
394 folio_lock(folio);
395 folio_mark_dirty(folio);
396 folio_unlock(folio);
397 }
398 gup_put_folio(folio, nr, FOLL_PIN);
399 }
400}
401EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
402
403/**
404 * unpin_user_page_range_dirty_lock() - release and optionally dirty
405 * gup-pinned page range
406 *
407 * @page: the starting page of a range maybe marked dirty, and definitely released.
408 * @npages: number of consecutive pages to release.
409 * @make_dirty: whether to mark the pages dirty
410 *
411 * "gup-pinned page range" refers to a range of pages that has had one of the
412 * pin_user_pages() variants called on that page.
413 *
414 * For the page ranges defined by [page .. page+npages], make that range (or
415 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
416 * page range was previously listed as clean.
417 *
418 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
419 * required, then the caller should a) verify that this is really correct,
420 * because _lock() is usually required, and b) hand code it:
421 * set_page_dirty_lock(), unpin_user_page().
422 *
423 */
424void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages,
425 bool make_dirty)
426{
427 unsigned long i;
428 struct folio *folio;
429 unsigned int nr;
430
431 for (i = 0; i < npages; i += nr) {
432 folio = gup_folio_range_next(page, npages, i, &nr);
433 if (make_dirty && !folio_test_dirty(folio)) {
434 folio_lock(folio);
435 folio_mark_dirty(folio);
436 folio_unlock(folio);
437 }
438 gup_put_folio(folio, nr, FOLL_PIN);
439 }
440}
441EXPORT_SYMBOL(unpin_user_page_range_dirty_lock);
442
443static void unpin_user_pages_lockless(struct page **pages, unsigned long npages)
444{
445 unsigned long i;
446 struct folio *folio;
447 unsigned int nr;
448
449 /*
450 * Don't perform any sanity checks because we might have raced with
451 * fork() and some anonymous pages might now actually be shared --
452 * which is why we're unpinning after all.
453 */
454 for (i = 0; i < npages; i += nr) {
455 folio = gup_folio_next(pages, npages, i, &nr);
456 gup_put_folio(folio, nr, FOLL_PIN);
457 }
458}
459
460/**
461 * unpin_user_pages() - release an array of gup-pinned pages.
462 * @pages: array of pages to be marked dirty and released.
463 * @npages: number of pages in the @pages array.
464 *
465 * For each page in the @pages array, release the page using unpin_user_page().
466 *
467 * Please see the unpin_user_page() documentation for details.
468 */
469void unpin_user_pages(struct page **pages, unsigned long npages)
470{
471 unsigned long i;
472 struct folio *folio;
473 unsigned int nr;
474
475 /*
476 * If this WARN_ON() fires, then the system *might* be leaking pages (by
477 * leaving them pinned), but probably not. More likely, gup/pup returned
478 * a hard -ERRNO error to the caller, who erroneously passed it here.
479 */
480 if (WARN_ON(IS_ERR_VALUE(npages)))
481 return;
482
483 sanity_check_pinned_pages(pages, npages);
484 for (i = 0; i < npages; i += nr) {
485 folio = gup_folio_next(pages, npages, i, &nr);
486 gup_put_folio(folio, nr, FOLL_PIN);
487 }
488}
489EXPORT_SYMBOL(unpin_user_pages);
490
491/*
492 * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's
493 * lifecycle. Avoid setting the bit unless necessary, or it might cause write
494 * cache bouncing on large SMP machines for concurrent pinned gups.
495 */
496static inline void mm_set_has_pinned_flag(unsigned long *mm_flags)
497{
498 if (!test_bit(MMF_HAS_PINNED, mm_flags))
499 set_bit(MMF_HAS_PINNED, mm_flags);
500}
501
502#ifdef CONFIG_MMU
503static struct page *no_page_table(struct vm_area_struct *vma,
504 unsigned int flags)
505{
506 /*
507 * When core dumping an enormous anonymous area that nobody
508 * has touched so far, we don't want to allocate unnecessary pages or
509 * page tables. Return error instead of NULL to skip handle_mm_fault,
510 * then get_dump_page() will return NULL to leave a hole in the dump.
511 * But we can only make this optimization where a hole would surely
512 * be zero-filled if handle_mm_fault() actually did handle it.
513 */
514 if ((flags & FOLL_DUMP) &&
515 (vma_is_anonymous(vma) || !vma->vm_ops->fault))
516 return ERR_PTR(-EFAULT);
517 return NULL;
518}
519
520static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
521 pte_t *pte, unsigned int flags)
522{
523 if (flags & FOLL_TOUCH) {
524 pte_t orig_entry = ptep_get(pte);
525 pte_t entry = orig_entry;
526
527 if (flags & FOLL_WRITE)
528 entry = pte_mkdirty(entry);
529 entry = pte_mkyoung(entry);
530
531 if (!pte_same(orig_entry, entry)) {
532 set_pte_at(vma->vm_mm, address, pte, entry);
533 update_mmu_cache(vma, address, pte);
534 }
535 }
536
537 /* Proper page table entry exists, but no corresponding struct page */
538 return -EEXIST;
539}
540
541/* FOLL_FORCE can write to even unwritable PTEs in COW mappings. */
542static inline bool can_follow_write_pte(pte_t pte, struct page *page,
543 struct vm_area_struct *vma,
544 unsigned int flags)
545{
546 /* If the pte is writable, we can write to the page. */
547 if (pte_write(pte))
548 return true;
549
550 /* Maybe FOLL_FORCE is set to override it? */
551 if (!(flags & FOLL_FORCE))
552 return false;
553
554 /* But FOLL_FORCE has no effect on shared mappings */
555 if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
556 return false;
557
558 /* ... or read-only private ones */
559 if (!(vma->vm_flags & VM_MAYWRITE))
560 return false;
561
562 /* ... or already writable ones that just need to take a write fault */
563 if (vma->vm_flags & VM_WRITE)
564 return false;
565
566 /*
567 * See can_change_pte_writable(): we broke COW and could map the page
568 * writable if we have an exclusive anonymous page ...
569 */
570 if (!page || !PageAnon(page) || !PageAnonExclusive(page))
571 return false;
572
573 /* ... and a write-fault isn't required for other reasons. */
574 if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte))
575 return false;
576 return !userfaultfd_pte_wp(vma, pte);
577}
578
579static struct page *follow_page_pte(struct vm_area_struct *vma,
580 unsigned long address, pmd_t *pmd, unsigned int flags,
581 struct dev_pagemap **pgmap)
582{
583 struct mm_struct *mm = vma->vm_mm;
584 struct page *page;
585 spinlock_t *ptl;
586 pte_t *ptep, pte;
587 int ret;
588
589 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
590 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
591 (FOLL_PIN | FOLL_GET)))
592 return ERR_PTR(-EINVAL);
593
594 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
595 if (!ptep)
596 return no_page_table(vma, flags);
597 pte = ptep_get(ptep);
598 if (!pte_present(pte))
599 goto no_page;
600 if (pte_protnone(pte) && !gup_can_follow_protnone(vma, flags))
601 goto no_page;
602
603 page = vm_normal_page(vma, address, pte);
604
605 /*
606 * We only care about anon pages in can_follow_write_pte() and don't
607 * have to worry about pte_devmap() because they are never anon.
608 */
609 if ((flags & FOLL_WRITE) &&
610 !can_follow_write_pte(pte, page, vma, flags)) {
611 page = NULL;
612 goto out;
613 }
614
615 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
616 /*
617 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
618 * case since they are only valid while holding the pgmap
619 * reference.
620 */
621 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
622 if (*pgmap)
623 page = pte_page(pte);
624 else
625 goto no_page;
626 } else if (unlikely(!page)) {
627 if (flags & FOLL_DUMP) {
628 /* Avoid special (like zero) pages in core dumps */
629 page = ERR_PTR(-EFAULT);
630 goto out;
631 }
632
633 if (is_zero_pfn(pte_pfn(pte))) {
634 page = pte_page(pte);
635 } else {
636 ret = follow_pfn_pte(vma, address, ptep, flags);
637 page = ERR_PTR(ret);
638 goto out;
639 }
640 }
641
642 if (!pte_write(pte) && gup_must_unshare(vma, flags, page)) {
643 page = ERR_PTR(-EMLINK);
644 goto out;
645 }
646
647 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
648 !PageAnonExclusive(page), page);
649
650 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
651 ret = try_grab_page(page, flags);
652 if (unlikely(ret)) {
653 page = ERR_PTR(ret);
654 goto out;
655 }
656
657 /*
658 * We need to make the page accessible if and only if we are going
659 * to access its content (the FOLL_PIN case). Please see
660 * Documentation/core-api/pin_user_pages.rst for details.
661 */
662 if (flags & FOLL_PIN) {
663 ret = arch_make_page_accessible(page);
664 if (ret) {
665 unpin_user_page(page);
666 page = ERR_PTR(ret);
667 goto out;
668 }
669 }
670 if (flags & FOLL_TOUCH) {
671 if ((flags & FOLL_WRITE) &&
672 !pte_dirty(pte) && !PageDirty(page))
673 set_page_dirty(page);
674 /*
675 * pte_mkyoung() would be more correct here, but atomic care
676 * is needed to avoid losing the dirty bit: it is easier to use
677 * mark_page_accessed().
678 */
679 mark_page_accessed(page);
680 }
681out:
682 pte_unmap_unlock(ptep, ptl);
683 return page;
684no_page:
685 pte_unmap_unlock(ptep, ptl);
686 if (!pte_none(pte))
687 return NULL;
688 return no_page_table(vma, flags);
689}
690
691static struct page *follow_pmd_mask(struct vm_area_struct *vma,
692 unsigned long address, pud_t *pudp,
693 unsigned int flags,
694 struct follow_page_context *ctx)
695{
696 pmd_t *pmd, pmdval;
697 spinlock_t *ptl;
698 struct page *page;
699 struct mm_struct *mm = vma->vm_mm;
700
701 pmd = pmd_offset(pudp, address);
702 pmdval = pmdp_get_lockless(pmd);
703 if (pmd_none(pmdval))
704 return no_page_table(vma, flags);
705 if (!pmd_present(pmdval))
706 return no_page_table(vma, flags);
707 if (pmd_devmap(pmdval)) {
708 ptl = pmd_lock(mm, pmd);
709 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
710 spin_unlock(ptl);
711 if (page)
712 return page;
713 return no_page_table(vma, flags);
714 }
715 if (likely(!pmd_trans_huge(pmdval)))
716 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
717
718 if (pmd_protnone(pmdval) && !gup_can_follow_protnone(vma, flags))
719 return no_page_table(vma, flags);
720
721 ptl = pmd_lock(mm, pmd);
722 if (unlikely(!pmd_present(*pmd))) {
723 spin_unlock(ptl);
724 return no_page_table(vma, flags);
725 }
726 if (unlikely(!pmd_trans_huge(*pmd))) {
727 spin_unlock(ptl);
728 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
729 }
730 if (flags & FOLL_SPLIT_PMD) {
731 spin_unlock(ptl);
732 split_huge_pmd(vma, pmd, address);
733 /* If pmd was left empty, stuff a page table in there quickly */
734 return pte_alloc(mm, pmd) ? ERR_PTR(-ENOMEM) :
735 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
736 }
737 page = follow_trans_huge_pmd(vma, address, pmd, flags);
738 spin_unlock(ptl);
739 ctx->page_mask = HPAGE_PMD_NR - 1;
740 return page;
741}
742
743static struct page *follow_pud_mask(struct vm_area_struct *vma,
744 unsigned long address, p4d_t *p4dp,
745 unsigned int flags,
746 struct follow_page_context *ctx)
747{
748 pud_t *pud;
749 spinlock_t *ptl;
750 struct page *page;
751 struct mm_struct *mm = vma->vm_mm;
752
753 pud = pud_offset(p4dp, address);
754 if (pud_none(*pud))
755 return no_page_table(vma, flags);
756 if (pud_devmap(*pud)) {
757 ptl = pud_lock(mm, pud);
758 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
759 spin_unlock(ptl);
760 if (page)
761 return page;
762 return no_page_table(vma, flags);
763 }
764 if (unlikely(pud_bad(*pud)))
765 return no_page_table(vma, flags);
766
767 return follow_pmd_mask(vma, address, pud, flags, ctx);
768}
769
770static struct page *follow_p4d_mask(struct vm_area_struct *vma,
771 unsigned long address, pgd_t *pgdp,
772 unsigned int flags,
773 struct follow_page_context *ctx)
774{
775 p4d_t *p4d;
776
777 p4d = p4d_offset(pgdp, address);
778 if (p4d_none(*p4d))
779 return no_page_table(vma, flags);
780 BUILD_BUG_ON(p4d_huge(*p4d));
781 if (unlikely(p4d_bad(*p4d)))
782 return no_page_table(vma, flags);
783
784 return follow_pud_mask(vma, address, p4d, flags, ctx);
785}
786
787/**
788 * follow_page_mask - look up a page descriptor from a user-virtual address
789 * @vma: vm_area_struct mapping @address
790 * @address: virtual address to look up
791 * @flags: flags modifying lookup behaviour
792 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
793 * pointer to output page_mask
794 *
795 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
796 *
797 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
798 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
799 *
800 * When getting an anonymous page and the caller has to trigger unsharing
801 * of a shared anonymous page first, -EMLINK is returned. The caller should
802 * trigger a fault with FAULT_FLAG_UNSHARE set. Note that unsharing is only
803 * relevant with FOLL_PIN and !FOLL_WRITE.
804 *
805 * On output, the @ctx->page_mask is set according to the size of the page.
806 *
807 * Return: the mapped (struct page *), %NULL if no mapping exists, or
808 * an error pointer if there is a mapping to something not represented
809 * by a page descriptor (see also vm_normal_page()).
810 */
811static struct page *follow_page_mask(struct vm_area_struct *vma,
812 unsigned long address, unsigned int flags,
813 struct follow_page_context *ctx)
814{
815 pgd_t *pgd;
816 struct mm_struct *mm = vma->vm_mm;
817
818 ctx->page_mask = 0;
819
820 /*
821 * Call hugetlb_follow_page_mask for hugetlb vmas as it will use
822 * special hugetlb page table walking code. This eliminates the
823 * need to check for hugetlb entries in the general walking code.
824 */
825 if (is_vm_hugetlb_page(vma))
826 return hugetlb_follow_page_mask(vma, address, flags,
827 &ctx->page_mask);
828
829 pgd = pgd_offset(mm, address);
830
831 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
832 return no_page_table(vma, flags);
833
834 return follow_p4d_mask(vma, address, pgd, flags, ctx);
835}
836
837struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
838 unsigned int foll_flags)
839{
840 struct follow_page_context ctx = { NULL };
841 struct page *page;
842
843 if (vma_is_secretmem(vma))
844 return NULL;
845
846 if (WARN_ON_ONCE(foll_flags & FOLL_PIN))
847 return NULL;
848
849 /*
850 * We never set FOLL_HONOR_NUMA_FAULT because callers don't expect
851 * to fail on PROT_NONE-mapped pages.
852 */
853 page = follow_page_mask(vma, address, foll_flags, &ctx);
854 if (ctx.pgmap)
855 put_dev_pagemap(ctx.pgmap);
856 return page;
857}
858
859static int get_gate_page(struct mm_struct *mm, unsigned long address,
860 unsigned int gup_flags, struct vm_area_struct **vma,
861 struct page **page)
862{
863 pgd_t *pgd;
864 p4d_t *p4d;
865 pud_t *pud;
866 pmd_t *pmd;
867 pte_t *pte;
868 pte_t entry;
869 int ret = -EFAULT;
870
871 /* user gate pages are read-only */
872 if (gup_flags & FOLL_WRITE)
873 return -EFAULT;
874 if (address > TASK_SIZE)
875 pgd = pgd_offset_k(address);
876 else
877 pgd = pgd_offset_gate(mm, address);
878 if (pgd_none(*pgd))
879 return -EFAULT;
880 p4d = p4d_offset(pgd, address);
881 if (p4d_none(*p4d))
882 return -EFAULT;
883 pud = pud_offset(p4d, address);
884 if (pud_none(*pud))
885 return -EFAULT;
886 pmd = pmd_offset(pud, address);
887 if (!pmd_present(*pmd))
888 return -EFAULT;
889 pte = pte_offset_map(pmd, address);
890 if (!pte)
891 return -EFAULT;
892 entry = ptep_get(pte);
893 if (pte_none(entry))
894 goto unmap;
895 *vma = get_gate_vma(mm);
896 if (!page)
897 goto out;
898 *page = vm_normal_page(*vma, address, entry);
899 if (!*page) {
900 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(entry)))
901 goto unmap;
902 *page = pte_page(entry);
903 }
904 ret = try_grab_page(*page, gup_flags);
905 if (unlikely(ret))
906 goto unmap;
907out:
908 ret = 0;
909unmap:
910 pte_unmap(pte);
911 return ret;
912}
913
914/*
915 * mmap_lock must be held on entry. If @flags has FOLL_UNLOCKABLE but not
916 * FOLL_NOWAIT, the mmap_lock may be released. If it is, *@locked will be set
917 * to 0 and -EBUSY returned.
918 */
919static int faultin_page(struct vm_area_struct *vma,
920 unsigned long address, unsigned int *flags, bool unshare,
921 int *locked)
922{
923 unsigned int fault_flags = 0;
924 vm_fault_t ret;
925
926 if (*flags & FOLL_NOFAULT)
927 return -EFAULT;
928 if (*flags & FOLL_WRITE)
929 fault_flags |= FAULT_FLAG_WRITE;
930 if (*flags & FOLL_REMOTE)
931 fault_flags |= FAULT_FLAG_REMOTE;
932 if (*flags & FOLL_UNLOCKABLE) {
933 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
934 /*
935 * FAULT_FLAG_INTERRUPTIBLE is opt-in. GUP callers must set
936 * FOLL_INTERRUPTIBLE to enable FAULT_FLAG_INTERRUPTIBLE.
937 * That's because some callers may not be prepared to
938 * handle early exits caused by non-fatal signals.
939 */
940 if (*flags & FOLL_INTERRUPTIBLE)
941 fault_flags |= FAULT_FLAG_INTERRUPTIBLE;
942 }
943 if (*flags & FOLL_NOWAIT)
944 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
945 if (*flags & FOLL_TRIED) {
946 /*
947 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
948 * can co-exist
949 */
950 fault_flags |= FAULT_FLAG_TRIED;
951 }
952 if (unshare) {
953 fault_flags |= FAULT_FLAG_UNSHARE;
954 /* FAULT_FLAG_WRITE and FAULT_FLAG_UNSHARE are incompatible */
955 VM_BUG_ON(fault_flags & FAULT_FLAG_WRITE);
956 }
957
958 ret = handle_mm_fault(vma, address, fault_flags, NULL);
959
960 if (ret & VM_FAULT_COMPLETED) {
961 /*
962 * With FAULT_FLAG_RETRY_NOWAIT we'll never release the
963 * mmap lock in the page fault handler. Sanity check this.
964 */
965 WARN_ON_ONCE(fault_flags & FAULT_FLAG_RETRY_NOWAIT);
966 *locked = 0;
967
968 /*
969 * We should do the same as VM_FAULT_RETRY, but let's not
970 * return -EBUSY since that's not reflecting the reality of
971 * what has happened - we've just fully completed a page
972 * fault, with the mmap lock released. Use -EAGAIN to show
973 * that we want to take the mmap lock _again_.
974 */
975 return -EAGAIN;
976 }
977
978 if (ret & VM_FAULT_ERROR) {
979 int err = vm_fault_to_errno(ret, *flags);
980
981 if (err)
982 return err;
983 BUG();
984 }
985
986 if (ret & VM_FAULT_RETRY) {
987 if (!(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
988 *locked = 0;
989 return -EBUSY;
990 }
991
992 return 0;
993}
994
995/*
996 * Writing to file-backed mappings which require folio dirty tracking using GUP
997 * is a fundamentally broken operation, as kernel write access to GUP mappings
998 * do not adhere to the semantics expected by a file system.
999 *
1000 * Consider the following scenario:-
1001 *
1002 * 1. A folio is written to via GUP which write-faults the memory, notifying
1003 * the file system and dirtying the folio.
1004 * 2. Later, writeback is triggered, resulting in the folio being cleaned and
1005 * the PTE being marked read-only.
1006 * 3. The GUP caller writes to the folio, as it is mapped read/write via the
1007 * direct mapping.
1008 * 4. The GUP caller, now done with the page, unpins it and sets it dirty
1009 * (though it does not have to).
1010 *
1011 * This results in both data being written to a folio without writenotify, and
1012 * the folio being dirtied unexpectedly (if the caller decides to do so).
1013 */
1014static bool writable_file_mapping_allowed(struct vm_area_struct *vma,
1015 unsigned long gup_flags)
1016{
1017 /*
1018 * If we aren't pinning then no problematic write can occur. A long term
1019 * pin is the most egregious case so this is the case we disallow.
1020 */
1021 if ((gup_flags & (FOLL_PIN | FOLL_LONGTERM)) !=
1022 (FOLL_PIN | FOLL_LONGTERM))
1023 return true;
1024
1025 /*
1026 * If the VMA does not require dirty tracking then no problematic write
1027 * can occur either.
1028 */
1029 return !vma_needs_dirty_tracking(vma);
1030}
1031
1032static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
1033{
1034 vm_flags_t vm_flags = vma->vm_flags;
1035 int write = (gup_flags & FOLL_WRITE);
1036 int foreign = (gup_flags & FOLL_REMOTE);
1037 bool vma_anon = vma_is_anonymous(vma);
1038
1039 if (vm_flags & (VM_IO | VM_PFNMAP))
1040 return -EFAULT;
1041
1042 if ((gup_flags & FOLL_ANON) && !vma_anon)
1043 return -EFAULT;
1044
1045 if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma))
1046 return -EOPNOTSUPP;
1047
1048 if (vma_is_secretmem(vma))
1049 return -EFAULT;
1050
1051 if (write) {
1052 if (!vma_anon &&
1053 !writable_file_mapping_allowed(vma, gup_flags))
1054 return -EFAULT;
1055
1056 if (!(vm_flags & VM_WRITE) || (vm_flags & VM_SHADOW_STACK)) {
1057 if (!(gup_flags & FOLL_FORCE))
1058 return -EFAULT;
1059 /* hugetlb does not support FOLL_FORCE|FOLL_WRITE. */
1060 if (is_vm_hugetlb_page(vma))
1061 return -EFAULT;
1062 /*
1063 * We used to let the write,force case do COW in a
1064 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
1065 * set a breakpoint in a read-only mapping of an
1066 * executable, without corrupting the file (yet only
1067 * when that file had been opened for writing!).
1068 * Anon pages in shared mappings are surprising: now
1069 * just reject it.
1070 */
1071 if (!is_cow_mapping(vm_flags))
1072 return -EFAULT;
1073 }
1074 } else if (!(vm_flags & VM_READ)) {
1075 if (!(gup_flags & FOLL_FORCE))
1076 return -EFAULT;
1077 /*
1078 * Is there actually any vma we can reach here which does not
1079 * have VM_MAYREAD set?
1080 */
1081 if (!(vm_flags & VM_MAYREAD))
1082 return -EFAULT;
1083 }
1084 /*
1085 * gups are always data accesses, not instruction
1086 * fetches, so execute=false here
1087 */
1088 if (!arch_vma_access_permitted(vma, write, false, foreign))
1089 return -EFAULT;
1090 return 0;
1091}
1092
1093/*
1094 * This is "vma_lookup()", but with a warning if we would have
1095 * historically expanded the stack in the GUP code.
1096 */
1097static struct vm_area_struct *gup_vma_lookup(struct mm_struct *mm,
1098 unsigned long addr)
1099{
1100#ifdef CONFIG_STACK_GROWSUP
1101 return vma_lookup(mm, addr);
1102#else
1103 static volatile unsigned long next_warn;
1104 struct vm_area_struct *vma;
1105 unsigned long now, next;
1106
1107 vma = find_vma(mm, addr);
1108 if (!vma || (addr >= vma->vm_start))
1109 return vma;
1110
1111 /* Only warn for half-way relevant accesses */
1112 if (!(vma->vm_flags & VM_GROWSDOWN))
1113 return NULL;
1114 if (vma->vm_start - addr > 65536)
1115 return NULL;
1116
1117 /* Let's not warn more than once an hour.. */
1118 now = jiffies; next = next_warn;
1119 if (next && time_before(now, next))
1120 return NULL;
1121 next_warn = now + 60*60*HZ;
1122
1123 /* Let people know things may have changed. */
1124 pr_warn("GUP no longer grows the stack in %s (%d): %lx-%lx (%lx)\n",
1125 current->comm, task_pid_nr(current),
1126 vma->vm_start, vma->vm_end, addr);
1127 dump_stack();
1128 return NULL;
1129#endif
1130}
1131
1132/**
1133 * __get_user_pages() - pin user pages in memory
1134 * @mm: mm_struct of target mm
1135 * @start: starting user address
1136 * @nr_pages: number of pages from start to pin
1137 * @gup_flags: flags modifying pin behaviour
1138 * @pages: array that receives pointers to the pages pinned.
1139 * Should be at least nr_pages long. Or NULL, if caller
1140 * only intends to ensure the pages are faulted in.
1141 * @locked: whether we're still with the mmap_lock held
1142 *
1143 * Returns either number of pages pinned (which may be less than the
1144 * number requested), or an error. Details about the return value:
1145 *
1146 * -- If nr_pages is 0, returns 0.
1147 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1148 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1149 * pages pinned. Again, this may be less than nr_pages.
1150 * -- 0 return value is possible when the fault would need to be retried.
1151 *
1152 * The caller is responsible for releasing returned @pages, via put_page().
1153 *
1154 * Must be called with mmap_lock held. It may be released. See below.
1155 *
1156 * __get_user_pages walks a process's page tables and takes a reference to
1157 * each struct page that each user address corresponds to at a given
1158 * instant. That is, it takes the page that would be accessed if a user
1159 * thread accesses the given user virtual address at that instant.
1160 *
1161 * This does not guarantee that the page exists in the user mappings when
1162 * __get_user_pages returns, and there may even be a completely different
1163 * page there in some cases (eg. if mmapped pagecache has been invalidated
1164 * and subsequently re-faulted). However it does guarantee that the page
1165 * won't be freed completely. And mostly callers simply care that the page
1166 * contains data that was valid *at some point in time*. Typically, an IO
1167 * or similar operation cannot guarantee anything stronger anyway because
1168 * locks can't be held over the syscall boundary.
1169 *
1170 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1171 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1172 * appropriate) must be called after the page is finished with, and
1173 * before put_page is called.
1174 *
1175 * If FOLL_UNLOCKABLE is set without FOLL_NOWAIT then the mmap_lock may
1176 * be released. If this happens *@locked will be set to 0 on return.
1177 *
1178 * A caller using such a combination of @gup_flags must therefore hold the
1179 * mmap_lock for reading only, and recognize when it's been released. Otherwise,
1180 * it must be held for either reading or writing and will not be released.
1181 *
1182 * In most cases, get_user_pages or get_user_pages_fast should be used
1183 * instead of __get_user_pages. __get_user_pages should be used only if
1184 * you need some special @gup_flags.
1185 */
1186static long __get_user_pages(struct mm_struct *mm,
1187 unsigned long start, unsigned long nr_pages,
1188 unsigned int gup_flags, struct page **pages,
1189 int *locked)
1190{
1191 long ret = 0, i = 0;
1192 struct vm_area_struct *vma = NULL;
1193 struct follow_page_context ctx = { NULL };
1194
1195 if (!nr_pages)
1196 return 0;
1197
1198 start = untagged_addr_remote(mm, start);
1199
1200 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1201
1202 do {
1203 struct page *page;
1204 unsigned int foll_flags = gup_flags;
1205 unsigned int page_increm;
1206
1207 /* first iteration or cross vma bound */
1208 if (!vma || start >= vma->vm_end) {
1209 /*
1210 * MADV_POPULATE_(READ|WRITE) wants to handle VMA
1211 * lookups+error reporting differently.
1212 */
1213 if (gup_flags & FOLL_MADV_POPULATE) {
1214 vma = vma_lookup(mm, start);
1215 if (!vma) {
1216 ret = -ENOMEM;
1217 goto out;
1218 }
1219 if (check_vma_flags(vma, gup_flags)) {
1220 ret = -EINVAL;
1221 goto out;
1222 }
1223 goto retry;
1224 }
1225 vma = gup_vma_lookup(mm, start);
1226 if (!vma && in_gate_area(mm, start)) {
1227 ret = get_gate_page(mm, start & PAGE_MASK,
1228 gup_flags, &vma,
1229 pages ? &page : NULL);
1230 if (ret)
1231 goto out;
1232 ctx.page_mask = 0;
1233 goto next_page;
1234 }
1235
1236 if (!vma) {
1237 ret = -EFAULT;
1238 goto out;
1239 }
1240 ret = check_vma_flags(vma, gup_flags);
1241 if (ret)
1242 goto out;
1243 }
1244retry:
1245 /*
1246 * If we have a pending SIGKILL, don't keep faulting pages and
1247 * potentially allocating memory.
1248 */
1249 if (fatal_signal_pending(current)) {
1250 ret = -EINTR;
1251 goto out;
1252 }
1253 cond_resched();
1254
1255 page = follow_page_mask(vma, start, foll_flags, &ctx);
1256 if (!page || PTR_ERR(page) == -EMLINK) {
1257 ret = faultin_page(vma, start, &foll_flags,
1258 PTR_ERR(page) == -EMLINK, locked);
1259 switch (ret) {
1260 case 0:
1261 goto retry;
1262 case -EBUSY:
1263 case -EAGAIN:
1264 ret = 0;
1265 fallthrough;
1266 case -EFAULT:
1267 case -ENOMEM:
1268 case -EHWPOISON:
1269 goto out;
1270 }
1271 BUG();
1272 } else if (PTR_ERR(page) == -EEXIST) {
1273 /*
1274 * Proper page table entry exists, but no corresponding
1275 * struct page. If the caller expects **pages to be
1276 * filled in, bail out now, because that can't be done
1277 * for this page.
1278 */
1279 if (pages) {
1280 ret = PTR_ERR(page);
1281 goto out;
1282 }
1283 } else if (IS_ERR(page)) {
1284 ret = PTR_ERR(page);
1285 goto out;
1286 }
1287next_page:
1288 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1289 if (page_increm > nr_pages)
1290 page_increm = nr_pages;
1291
1292 if (pages) {
1293 struct page *subpage;
1294 unsigned int j;
1295
1296 /*
1297 * This must be a large folio (and doesn't need to
1298 * be the whole folio; it can be part of it), do
1299 * the refcount work for all the subpages too.
1300 *
1301 * NOTE: here the page may not be the head page
1302 * e.g. when start addr is not thp-size aligned.
1303 * try_grab_folio() should have taken care of tail
1304 * pages.
1305 */
1306 if (page_increm > 1) {
1307 struct folio *folio;
1308
1309 /*
1310 * Since we already hold refcount on the
1311 * large folio, this should never fail.
1312 */
1313 folio = try_grab_folio(page, page_increm - 1,
1314 foll_flags);
1315 if (WARN_ON_ONCE(!folio)) {
1316 /*
1317 * Release the 1st page ref if the
1318 * folio is problematic, fail hard.
1319 */
1320 gup_put_folio(page_folio(page), 1,
1321 foll_flags);
1322 ret = -EFAULT;
1323 goto out;
1324 }
1325 }
1326
1327 for (j = 0; j < page_increm; j++) {
1328 subpage = nth_page(page, j);
1329 pages[i + j] = subpage;
1330 flush_anon_page(vma, subpage, start + j * PAGE_SIZE);
1331 flush_dcache_page(subpage);
1332 }
1333 }
1334
1335 i += page_increm;
1336 start += page_increm * PAGE_SIZE;
1337 nr_pages -= page_increm;
1338 } while (nr_pages);
1339out:
1340 if (ctx.pgmap)
1341 put_dev_pagemap(ctx.pgmap);
1342 return i ? i : ret;
1343}
1344
1345static bool vma_permits_fault(struct vm_area_struct *vma,
1346 unsigned int fault_flags)
1347{
1348 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1349 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1350 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1351
1352 if (!(vm_flags & vma->vm_flags))
1353 return false;
1354
1355 /*
1356 * The architecture might have a hardware protection
1357 * mechanism other than read/write that can deny access.
1358 *
1359 * gup always represents data access, not instruction
1360 * fetches, so execute=false here:
1361 */
1362 if (!arch_vma_access_permitted(vma, write, false, foreign))
1363 return false;
1364
1365 return true;
1366}
1367
1368/**
1369 * fixup_user_fault() - manually resolve a user page fault
1370 * @mm: mm_struct of target mm
1371 * @address: user address
1372 * @fault_flags:flags to pass down to handle_mm_fault()
1373 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1374 * does not allow retry. If NULL, the caller must guarantee
1375 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1376 *
1377 * This is meant to be called in the specific scenario where for locking reasons
1378 * we try to access user memory in atomic context (within a pagefault_disable()
1379 * section), this returns -EFAULT, and we want to resolve the user fault before
1380 * trying again.
1381 *
1382 * Typically this is meant to be used by the futex code.
1383 *
1384 * The main difference with get_user_pages() is that this function will
1385 * unconditionally call handle_mm_fault() which will in turn perform all the
1386 * necessary SW fixup of the dirty and young bits in the PTE, while
1387 * get_user_pages() only guarantees to update these in the struct page.
1388 *
1389 * This is important for some architectures where those bits also gate the
1390 * access permission to the page because they are maintained in software. On
1391 * such architectures, gup() will not be enough to make a subsequent access
1392 * succeed.
1393 *
1394 * This function will not return with an unlocked mmap_lock. So it has not the
1395 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1396 */
1397int fixup_user_fault(struct mm_struct *mm,
1398 unsigned long address, unsigned int fault_flags,
1399 bool *unlocked)
1400{
1401 struct vm_area_struct *vma;
1402 vm_fault_t ret;
1403
1404 address = untagged_addr_remote(mm, address);
1405
1406 if (unlocked)
1407 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1408
1409retry:
1410 vma = gup_vma_lookup(mm, address);
1411 if (!vma)
1412 return -EFAULT;
1413
1414 if (!vma_permits_fault(vma, fault_flags))
1415 return -EFAULT;
1416
1417 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1418 fatal_signal_pending(current))
1419 return -EINTR;
1420
1421 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1422
1423 if (ret & VM_FAULT_COMPLETED) {
1424 /*
1425 * NOTE: it's a pity that we need to retake the lock here
1426 * to pair with the unlock() in the callers. Ideally we
1427 * could tell the callers so they do not need to unlock.
1428 */
1429 mmap_read_lock(mm);
1430 *unlocked = true;
1431 return 0;
1432 }
1433
1434 if (ret & VM_FAULT_ERROR) {
1435 int err = vm_fault_to_errno(ret, 0);
1436
1437 if (err)
1438 return err;
1439 BUG();
1440 }
1441
1442 if (ret & VM_FAULT_RETRY) {
1443 mmap_read_lock(mm);
1444 *unlocked = true;
1445 fault_flags |= FAULT_FLAG_TRIED;
1446 goto retry;
1447 }
1448
1449 return 0;
1450}
1451EXPORT_SYMBOL_GPL(fixup_user_fault);
1452
1453/*
1454 * GUP always responds to fatal signals. When FOLL_INTERRUPTIBLE is
1455 * specified, it'll also respond to generic signals. The caller of GUP
1456 * that has FOLL_INTERRUPTIBLE should take care of the GUP interruption.
1457 */
1458static bool gup_signal_pending(unsigned int flags)
1459{
1460 if (fatal_signal_pending(current))
1461 return true;
1462
1463 if (!(flags & FOLL_INTERRUPTIBLE))
1464 return false;
1465
1466 return signal_pending(current);
1467}
1468
1469/*
1470 * Locking: (*locked == 1) means that the mmap_lock has already been acquired by
1471 * the caller. This function may drop the mmap_lock. If it does so, then it will
1472 * set (*locked = 0).
1473 *
1474 * (*locked == 0) means that the caller expects this function to acquire and
1475 * drop the mmap_lock. Therefore, the value of *locked will still be zero when
1476 * the function returns, even though it may have changed temporarily during
1477 * function execution.
1478 *
1479 * Please note that this function, unlike __get_user_pages(), will not return 0
1480 * for nr_pages > 0, unless FOLL_NOWAIT is used.
1481 */
1482static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1483 unsigned long start,
1484 unsigned long nr_pages,
1485 struct page **pages,
1486 int *locked,
1487 unsigned int flags)
1488{
1489 long ret, pages_done;
1490 bool must_unlock = false;
1491
1492 if (!nr_pages)
1493 return 0;
1494
1495 /*
1496 * The internal caller expects GUP to manage the lock internally and the
1497 * lock must be released when this returns.
1498 */
1499 if (!*locked) {
1500 if (mmap_read_lock_killable(mm))
1501 return -EAGAIN;
1502 must_unlock = true;
1503 *locked = 1;
1504 }
1505 else
1506 mmap_assert_locked(mm);
1507
1508 if (flags & FOLL_PIN)
1509 mm_set_has_pinned_flag(&mm->flags);
1510
1511 /*
1512 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1513 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1514 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1515 * for FOLL_GET, not for the newer FOLL_PIN.
1516 *
1517 * FOLL_PIN always expects pages to be non-null, but no need to assert
1518 * that here, as any failures will be obvious enough.
1519 */
1520 if (pages && !(flags & FOLL_PIN))
1521 flags |= FOLL_GET;
1522
1523 pages_done = 0;
1524 for (;;) {
1525 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1526 locked);
1527 if (!(flags & FOLL_UNLOCKABLE)) {
1528 /* VM_FAULT_RETRY couldn't trigger, bypass */
1529 pages_done = ret;
1530 break;
1531 }
1532
1533 /* VM_FAULT_RETRY or VM_FAULT_COMPLETED cannot return errors */
1534 if (!*locked) {
1535 BUG_ON(ret < 0);
1536 BUG_ON(ret >= nr_pages);
1537 }
1538
1539 if (ret > 0) {
1540 nr_pages -= ret;
1541 pages_done += ret;
1542 if (!nr_pages)
1543 break;
1544 }
1545 if (*locked) {
1546 /*
1547 * VM_FAULT_RETRY didn't trigger or it was a
1548 * FOLL_NOWAIT.
1549 */
1550 if (!pages_done)
1551 pages_done = ret;
1552 break;
1553 }
1554 /*
1555 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1556 * For the prefault case (!pages) we only update counts.
1557 */
1558 if (likely(pages))
1559 pages += ret;
1560 start += ret << PAGE_SHIFT;
1561
1562 /* The lock was temporarily dropped, so we must unlock later */
1563 must_unlock = true;
1564
1565retry:
1566 /*
1567 * Repeat on the address that fired VM_FAULT_RETRY
1568 * with both FAULT_FLAG_ALLOW_RETRY and
1569 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1570 * by fatal signals of even common signals, depending on
1571 * the caller's request. So we need to check it before we
1572 * start trying again otherwise it can loop forever.
1573 */
1574 if (gup_signal_pending(flags)) {
1575 if (!pages_done)
1576 pages_done = -EINTR;
1577 break;
1578 }
1579
1580 ret = mmap_read_lock_killable(mm);
1581 if (ret) {
1582 BUG_ON(ret > 0);
1583 if (!pages_done)
1584 pages_done = ret;
1585 break;
1586 }
1587
1588 *locked = 1;
1589 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1590 pages, locked);
1591 if (!*locked) {
1592 /* Continue to retry until we succeeded */
1593 BUG_ON(ret != 0);
1594 goto retry;
1595 }
1596 if (ret != 1) {
1597 BUG_ON(ret > 1);
1598 if (!pages_done)
1599 pages_done = ret;
1600 break;
1601 }
1602 nr_pages--;
1603 pages_done++;
1604 if (!nr_pages)
1605 break;
1606 if (likely(pages))
1607 pages++;
1608 start += PAGE_SIZE;
1609 }
1610 if (must_unlock && *locked) {
1611 /*
1612 * We either temporarily dropped the lock, or the caller
1613 * requested that we both acquire and drop the lock. Either way,
1614 * we must now unlock, and notify the caller of that state.
1615 */
1616 mmap_read_unlock(mm);
1617 *locked = 0;
1618 }
1619
1620 /*
1621 * Failing to pin anything implies something has gone wrong (except when
1622 * FOLL_NOWAIT is specified).
1623 */
1624 if (WARN_ON_ONCE(pages_done == 0 && !(flags & FOLL_NOWAIT)))
1625 return -EFAULT;
1626
1627 return pages_done;
1628}
1629
1630/**
1631 * populate_vma_page_range() - populate a range of pages in the vma.
1632 * @vma: target vma
1633 * @start: start address
1634 * @end: end address
1635 * @locked: whether the mmap_lock is still held
1636 *
1637 * This takes care of mlocking the pages too if VM_LOCKED is set.
1638 *
1639 * Return either number of pages pinned in the vma, or a negative error
1640 * code on error.
1641 *
1642 * vma->vm_mm->mmap_lock must be held.
1643 *
1644 * If @locked is NULL, it may be held for read or write and will
1645 * be unperturbed.
1646 *
1647 * If @locked is non-NULL, it must held for read only and may be
1648 * released. If it's released, *@locked will be set to 0.
1649 */
1650long populate_vma_page_range(struct vm_area_struct *vma,
1651 unsigned long start, unsigned long end, int *locked)
1652{
1653 struct mm_struct *mm = vma->vm_mm;
1654 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1655 int local_locked = 1;
1656 int gup_flags;
1657 long ret;
1658
1659 VM_BUG_ON(!PAGE_ALIGNED(start));
1660 VM_BUG_ON(!PAGE_ALIGNED(end));
1661 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1662 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1663 mmap_assert_locked(mm);
1664
1665 /*
1666 * Rightly or wrongly, the VM_LOCKONFAULT case has never used
1667 * faultin_page() to break COW, so it has no work to do here.
1668 */
1669 if (vma->vm_flags & VM_LOCKONFAULT)
1670 return nr_pages;
1671
1672 /* ... similarly, we've never faulted in PROT_NONE pages */
1673 if (!vma_is_accessible(vma))
1674 return -EFAULT;
1675
1676 gup_flags = FOLL_TOUCH;
1677 /*
1678 * We want to touch writable mappings with a write fault in order
1679 * to break COW, except for shared mappings because these don't COW
1680 * and we would not want to dirty them for nothing.
1681 *
1682 * Otherwise, do a read fault, and use FOLL_FORCE in case it's not
1683 * readable (ie write-only or executable).
1684 */
1685 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1686 gup_flags |= FOLL_WRITE;
1687 else
1688 gup_flags |= FOLL_FORCE;
1689
1690 if (locked)
1691 gup_flags |= FOLL_UNLOCKABLE;
1692
1693 /*
1694 * We made sure addr is within a VMA, so the following will
1695 * not result in a stack expansion that recurses back here.
1696 */
1697 ret = __get_user_pages(mm, start, nr_pages, gup_flags,
1698 NULL, locked ? locked : &local_locked);
1699 lru_add_drain();
1700 return ret;
1701}
1702
1703/*
1704 * faultin_page_range() - populate (prefault) page tables inside the
1705 * given range readable/writable
1706 *
1707 * This takes care of mlocking the pages, too, if VM_LOCKED is set.
1708 *
1709 * @mm: the mm to populate page tables in
1710 * @start: start address
1711 * @end: end address
1712 * @write: whether to prefault readable or writable
1713 * @locked: whether the mmap_lock is still held
1714 *
1715 * Returns either number of processed pages in the MM, or a negative error
1716 * code on error (see __get_user_pages()). Note that this function reports
1717 * errors related to VMAs, such as incompatible mappings, as expected by
1718 * MADV_POPULATE_(READ|WRITE).
1719 *
1720 * The range must be page-aligned.
1721 *
1722 * mm->mmap_lock must be held. If it's released, *@locked will be set to 0.
1723 */
1724long faultin_page_range(struct mm_struct *mm, unsigned long start,
1725 unsigned long end, bool write, int *locked)
1726{
1727 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1728 int gup_flags;
1729 long ret;
1730
1731 VM_BUG_ON(!PAGE_ALIGNED(start));
1732 VM_BUG_ON(!PAGE_ALIGNED(end));
1733 mmap_assert_locked(mm);
1734
1735 /*
1736 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark
1737 * the page dirty with FOLL_WRITE -- which doesn't make a
1738 * difference with !FOLL_FORCE, because the page is writable
1739 * in the page table.
1740 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit
1741 * a poisoned page.
1742 * !FOLL_FORCE: Require proper access permissions.
1743 */
1744 gup_flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_UNLOCKABLE |
1745 FOLL_MADV_POPULATE;
1746 if (write)
1747 gup_flags |= FOLL_WRITE;
1748
1749 ret = __get_user_pages_locked(mm, start, nr_pages, NULL, locked,
1750 gup_flags);
1751 lru_add_drain();
1752 return ret;
1753}
1754
1755/*
1756 * __mm_populate - populate and/or mlock pages within a range of address space.
1757 *
1758 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1759 * flags. VMAs must be already marked with the desired vm_flags, and
1760 * mmap_lock must not be held.
1761 */
1762int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1763{
1764 struct mm_struct *mm = current->mm;
1765 unsigned long end, nstart, nend;
1766 struct vm_area_struct *vma = NULL;
1767 int locked = 0;
1768 long ret = 0;
1769
1770 end = start + len;
1771
1772 for (nstart = start; nstart < end; nstart = nend) {
1773 /*
1774 * We want to fault in pages for [nstart; end) address range.
1775 * Find first corresponding VMA.
1776 */
1777 if (!locked) {
1778 locked = 1;
1779 mmap_read_lock(mm);
1780 vma = find_vma_intersection(mm, nstart, end);
1781 } else if (nstart >= vma->vm_end)
1782 vma = find_vma_intersection(mm, vma->vm_end, end);
1783
1784 if (!vma)
1785 break;
1786 /*
1787 * Set [nstart; nend) to intersection of desired address
1788 * range with the first VMA. Also, skip undesirable VMA types.
1789 */
1790 nend = min(end, vma->vm_end);
1791 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1792 continue;
1793 if (nstart < vma->vm_start)
1794 nstart = vma->vm_start;
1795 /*
1796 * Now fault in a range of pages. populate_vma_page_range()
1797 * double checks the vma flags, so that it won't mlock pages
1798 * if the vma was already munlocked.
1799 */
1800 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1801 if (ret < 0) {
1802 if (ignore_errors) {
1803 ret = 0;
1804 continue; /* continue at next VMA */
1805 }
1806 break;
1807 }
1808 nend = nstart + ret * PAGE_SIZE;
1809 ret = 0;
1810 }
1811 if (locked)
1812 mmap_read_unlock(mm);
1813 return ret; /* 0 or negative error code */
1814}
1815#else /* CONFIG_MMU */
1816static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1817 unsigned long nr_pages, struct page **pages,
1818 int *locked, unsigned int foll_flags)
1819{
1820 struct vm_area_struct *vma;
1821 bool must_unlock = false;
1822 unsigned long vm_flags;
1823 long i;
1824
1825 if (!nr_pages)
1826 return 0;
1827
1828 /*
1829 * The internal caller expects GUP to manage the lock internally and the
1830 * lock must be released when this returns.
1831 */
1832 if (!*locked) {
1833 if (mmap_read_lock_killable(mm))
1834 return -EAGAIN;
1835 must_unlock = true;
1836 *locked = 1;
1837 }
1838
1839 /* calculate required read or write permissions.
1840 * If FOLL_FORCE is set, we only require the "MAY" flags.
1841 */
1842 vm_flags = (foll_flags & FOLL_WRITE) ?
1843 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1844 vm_flags &= (foll_flags & FOLL_FORCE) ?
1845 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1846
1847 for (i = 0; i < nr_pages; i++) {
1848 vma = find_vma(mm, start);
1849 if (!vma)
1850 break;
1851
1852 /* protect what we can, including chardevs */
1853 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1854 !(vm_flags & vma->vm_flags))
1855 break;
1856
1857 if (pages) {
1858 pages[i] = virt_to_page((void *)start);
1859 if (pages[i])
1860 get_page(pages[i]);
1861 }
1862
1863 start = (start + PAGE_SIZE) & PAGE_MASK;
1864 }
1865
1866 if (must_unlock && *locked) {
1867 mmap_read_unlock(mm);
1868 *locked = 0;
1869 }
1870
1871 return i ? : -EFAULT;
1872}
1873#endif /* !CONFIG_MMU */
1874
1875/**
1876 * fault_in_writeable - fault in userspace address range for writing
1877 * @uaddr: start of address range
1878 * @size: size of address range
1879 *
1880 * Returns the number of bytes not faulted in (like copy_to_user() and
1881 * copy_from_user()).
1882 */
1883size_t fault_in_writeable(char __user *uaddr, size_t size)
1884{
1885 char __user *start = uaddr, *end;
1886
1887 if (unlikely(size == 0))
1888 return 0;
1889 if (!user_write_access_begin(uaddr, size))
1890 return size;
1891 if (!PAGE_ALIGNED(uaddr)) {
1892 unsafe_put_user(0, uaddr, out);
1893 uaddr = (char __user *)PAGE_ALIGN((unsigned long)uaddr);
1894 }
1895 end = (char __user *)PAGE_ALIGN((unsigned long)start + size);
1896 if (unlikely(end < start))
1897 end = NULL;
1898 while (uaddr != end) {
1899 unsafe_put_user(0, uaddr, out);
1900 uaddr += PAGE_SIZE;
1901 }
1902
1903out:
1904 user_write_access_end();
1905 if (size > uaddr - start)
1906 return size - (uaddr - start);
1907 return 0;
1908}
1909EXPORT_SYMBOL(fault_in_writeable);
1910
1911/**
1912 * fault_in_subpage_writeable - fault in an address range for writing
1913 * @uaddr: start of address range
1914 * @size: size of address range
1915 *
1916 * Fault in a user address range for writing while checking for permissions at
1917 * sub-page granularity (e.g. arm64 MTE). This function should be used when
1918 * the caller cannot guarantee forward progress of a copy_to_user() loop.
1919 *
1920 * Returns the number of bytes not faulted in (like copy_to_user() and
1921 * copy_from_user()).
1922 */
1923size_t fault_in_subpage_writeable(char __user *uaddr, size_t size)
1924{
1925 size_t faulted_in;
1926
1927 /*
1928 * Attempt faulting in at page granularity first for page table
1929 * permission checking. The arch-specific probe_subpage_writeable()
1930 * functions may not check for this.
1931 */
1932 faulted_in = size - fault_in_writeable(uaddr, size);
1933 if (faulted_in)
1934 faulted_in -= probe_subpage_writeable(uaddr, faulted_in);
1935
1936 return size - faulted_in;
1937}
1938EXPORT_SYMBOL(fault_in_subpage_writeable);
1939
1940/*
1941 * fault_in_safe_writeable - fault in an address range for writing
1942 * @uaddr: start of address range
1943 * @size: length of address range
1944 *
1945 * Faults in an address range for writing. This is primarily useful when we
1946 * already know that some or all of the pages in the address range aren't in
1947 * memory.
1948 *
1949 * Unlike fault_in_writeable(), this function is non-destructive.
1950 *
1951 * Note that we don't pin or otherwise hold the pages referenced that we fault
1952 * in. There's no guarantee that they'll stay in memory for any duration of
1953 * time.
1954 *
1955 * Returns the number of bytes not faulted in, like copy_to_user() and
1956 * copy_from_user().
1957 */
1958size_t fault_in_safe_writeable(const char __user *uaddr, size_t size)
1959{
1960 unsigned long start = (unsigned long)uaddr, end;
1961 struct mm_struct *mm = current->mm;
1962 bool unlocked = false;
1963
1964 if (unlikely(size == 0))
1965 return 0;
1966 end = PAGE_ALIGN(start + size);
1967 if (end < start)
1968 end = 0;
1969
1970 mmap_read_lock(mm);
1971 do {
1972 if (fixup_user_fault(mm, start, FAULT_FLAG_WRITE, &unlocked))
1973 break;
1974 start = (start + PAGE_SIZE) & PAGE_MASK;
1975 } while (start != end);
1976 mmap_read_unlock(mm);
1977
1978 if (size > (unsigned long)uaddr - start)
1979 return size - ((unsigned long)uaddr - start);
1980 return 0;
1981}
1982EXPORT_SYMBOL(fault_in_safe_writeable);
1983
1984/**
1985 * fault_in_readable - fault in userspace address range for reading
1986 * @uaddr: start of user address range
1987 * @size: size of user address range
1988 *
1989 * Returns the number of bytes not faulted in (like copy_to_user() and
1990 * copy_from_user()).
1991 */
1992size_t fault_in_readable(const char __user *uaddr, size_t size)
1993{
1994 const char __user *start = uaddr, *end;
1995 volatile char c;
1996
1997 if (unlikely(size == 0))
1998 return 0;
1999 if (!user_read_access_begin(uaddr, size))
2000 return size;
2001 if (!PAGE_ALIGNED(uaddr)) {
2002 unsafe_get_user(c, uaddr, out);
2003 uaddr = (const char __user *)PAGE_ALIGN((unsigned long)uaddr);
2004 }
2005 end = (const char __user *)PAGE_ALIGN((unsigned long)start + size);
2006 if (unlikely(end < start))
2007 end = NULL;
2008 while (uaddr != end) {
2009 unsafe_get_user(c, uaddr, out);
2010 uaddr += PAGE_SIZE;
2011 }
2012
2013out:
2014 user_read_access_end();
2015 (void)c;
2016 if (size > uaddr - start)
2017 return size - (uaddr - start);
2018 return 0;
2019}
2020EXPORT_SYMBOL(fault_in_readable);
2021
2022/**
2023 * get_dump_page() - pin user page in memory while writing it to core dump
2024 * @addr: user address
2025 *
2026 * Returns struct page pointer of user page pinned for dump,
2027 * to be freed afterwards by put_page().
2028 *
2029 * Returns NULL on any kind of failure - a hole must then be inserted into
2030 * the corefile, to preserve alignment with its headers; and also returns
2031 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2032 * allowing a hole to be left in the corefile to save disk space.
2033 *
2034 * Called without mmap_lock (takes and releases the mmap_lock by itself).
2035 */
2036#ifdef CONFIG_ELF_CORE
2037struct page *get_dump_page(unsigned long addr)
2038{
2039 struct page *page;
2040 int locked = 0;
2041 int ret;
2042
2043 ret = __get_user_pages_locked(current->mm, addr, 1, &page, &locked,
2044 FOLL_FORCE | FOLL_DUMP | FOLL_GET);
2045 return (ret == 1) ? page : NULL;
2046}
2047#endif /* CONFIG_ELF_CORE */
2048
2049#ifdef CONFIG_MIGRATION
2050/*
2051 * Returns the number of collected pages. Return value is always >= 0.
2052 */
2053static unsigned long collect_longterm_unpinnable_pages(
2054 struct list_head *movable_page_list,
2055 unsigned long nr_pages,
2056 struct page **pages)
2057{
2058 unsigned long i, collected = 0;
2059 struct folio *prev_folio = NULL;
2060 bool drain_allow = true;
2061
2062 for (i = 0; i < nr_pages; i++) {
2063 struct folio *folio = page_folio(pages[i]);
2064
2065 if (folio == prev_folio)
2066 continue;
2067 prev_folio = folio;
2068
2069 if (folio_is_longterm_pinnable(folio))
2070 continue;
2071
2072 collected++;
2073
2074 if (folio_is_device_coherent(folio))
2075 continue;
2076
2077 if (folio_test_hugetlb(folio)) {
2078 isolate_hugetlb(folio, movable_page_list);
2079 continue;
2080 }
2081
2082 if (!folio_test_lru(folio) && drain_allow) {
2083 lru_add_drain_all();
2084 drain_allow = false;
2085 }
2086
2087 if (!folio_isolate_lru(folio))
2088 continue;
2089
2090 list_add_tail(&folio->lru, movable_page_list);
2091 node_stat_mod_folio(folio,
2092 NR_ISOLATED_ANON + folio_is_file_lru(folio),
2093 folio_nr_pages(folio));
2094 }
2095
2096 return collected;
2097}
2098
2099/*
2100 * Unpins all pages and migrates device coherent pages and movable_page_list.
2101 * Returns -EAGAIN if all pages were successfully migrated or -errno for failure
2102 * (or partial success).
2103 */
2104static int migrate_longterm_unpinnable_pages(
2105 struct list_head *movable_page_list,
2106 unsigned long nr_pages,
2107 struct page **pages)
2108{
2109 int ret;
2110 unsigned long i;
2111
2112 for (i = 0; i < nr_pages; i++) {
2113 struct folio *folio = page_folio(pages[i]);
2114
2115 if (folio_is_device_coherent(folio)) {
2116 /*
2117 * Migration will fail if the page is pinned, so convert
2118 * the pin on the source page to a normal reference.
2119 */
2120 pages[i] = NULL;
2121 folio_get(folio);
2122 gup_put_folio(folio, 1, FOLL_PIN);
2123
2124 if (migrate_device_coherent_page(&folio->page)) {
2125 ret = -EBUSY;
2126 goto err;
2127 }
2128
2129 continue;
2130 }
2131
2132 /*
2133 * We can't migrate pages with unexpected references, so drop
2134 * the reference obtained by __get_user_pages_locked().
2135 * Migrating pages have been added to movable_page_list after
2136 * calling folio_isolate_lru() which takes a reference so the
2137 * page won't be freed if it's migrating.
2138 */
2139 unpin_user_page(pages[i]);
2140 pages[i] = NULL;
2141 }
2142
2143 if (!list_empty(movable_page_list)) {
2144 struct migration_target_control mtc = {
2145 .nid = NUMA_NO_NODE,
2146 .gfp_mask = GFP_USER | __GFP_NOWARN,
2147 };
2148
2149 if (migrate_pages(movable_page_list, alloc_migration_target,
2150 NULL, (unsigned long)&mtc, MIGRATE_SYNC,
2151 MR_LONGTERM_PIN, NULL)) {
2152 ret = -ENOMEM;
2153 goto err;
2154 }
2155 }
2156
2157 putback_movable_pages(movable_page_list);
2158
2159 return -EAGAIN;
2160
2161err:
2162 for (i = 0; i < nr_pages; i++)
2163 if (pages[i])
2164 unpin_user_page(pages[i]);
2165 putback_movable_pages(movable_page_list);
2166
2167 return ret;
2168}
2169
2170/*
2171 * Check whether all pages are *allowed* to be pinned. Rather confusingly, all
2172 * pages in the range are required to be pinned via FOLL_PIN, before calling
2173 * this routine.
2174 *
2175 * If any pages in the range are not allowed to be pinned, then this routine
2176 * will migrate those pages away, unpin all the pages in the range and return
2177 * -EAGAIN. The caller should re-pin the entire range with FOLL_PIN and then
2178 * call this routine again.
2179 *
2180 * If an error other than -EAGAIN occurs, this indicates a migration failure.
2181 * The caller should give up, and propagate the error back up the call stack.
2182 *
2183 * If everything is OK and all pages in the range are allowed to be pinned, then
2184 * this routine leaves all pages pinned and returns zero for success.
2185 */
2186static long check_and_migrate_movable_pages(unsigned long nr_pages,
2187 struct page **pages)
2188{
2189 unsigned long collected;
2190 LIST_HEAD(movable_page_list);
2191
2192 collected = collect_longterm_unpinnable_pages(&movable_page_list,
2193 nr_pages, pages);
2194 if (!collected)
2195 return 0;
2196
2197 return migrate_longterm_unpinnable_pages(&movable_page_list, nr_pages,
2198 pages);
2199}
2200#else
2201static long check_and_migrate_movable_pages(unsigned long nr_pages,
2202 struct page **pages)
2203{
2204 return 0;
2205}
2206#endif /* CONFIG_MIGRATION */
2207
2208/*
2209 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
2210 * allows us to process the FOLL_LONGTERM flag.
2211 */
2212static long __gup_longterm_locked(struct mm_struct *mm,
2213 unsigned long start,
2214 unsigned long nr_pages,
2215 struct page **pages,
2216 int *locked,
2217 unsigned int gup_flags)
2218{
2219 unsigned int flags;
2220 long rc, nr_pinned_pages;
2221
2222 if (!(gup_flags & FOLL_LONGTERM))
2223 return __get_user_pages_locked(mm, start, nr_pages, pages,
2224 locked, gup_flags);
2225
2226 flags = memalloc_pin_save();
2227 do {
2228 nr_pinned_pages = __get_user_pages_locked(mm, start, nr_pages,
2229 pages, locked,
2230 gup_flags);
2231 if (nr_pinned_pages <= 0) {
2232 rc = nr_pinned_pages;
2233 break;
2234 }
2235
2236 /* FOLL_LONGTERM implies FOLL_PIN */
2237 rc = check_and_migrate_movable_pages(nr_pinned_pages, pages);
2238 } while (rc == -EAGAIN);
2239 memalloc_pin_restore(flags);
2240 return rc ? rc : nr_pinned_pages;
2241}
2242
2243/*
2244 * Check that the given flags are valid for the exported gup/pup interface, and
2245 * update them with the required flags that the caller must have set.
2246 */
2247static bool is_valid_gup_args(struct page **pages, int *locked,
2248 unsigned int *gup_flags_p, unsigned int to_set)
2249{
2250 unsigned int gup_flags = *gup_flags_p;
2251
2252 /*
2253 * These flags not allowed to be specified externally to the gup
2254 * interfaces:
2255 * - FOLL_TOUCH/FOLL_PIN/FOLL_TRIED/FOLL_FAST_ONLY are internal only
2256 * - FOLL_REMOTE is internal only and used on follow_page()
2257 * - FOLL_UNLOCKABLE is internal only and used if locked is !NULL
2258 */
2259 if (WARN_ON_ONCE(gup_flags & INTERNAL_GUP_FLAGS))
2260 return false;
2261
2262 gup_flags |= to_set;
2263 if (locked) {
2264 /* At the external interface locked must be set */
2265 if (WARN_ON_ONCE(*locked != 1))
2266 return false;
2267
2268 gup_flags |= FOLL_UNLOCKABLE;
2269 }
2270
2271 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2272 if (WARN_ON_ONCE((gup_flags & (FOLL_PIN | FOLL_GET)) ==
2273 (FOLL_PIN | FOLL_GET)))
2274 return false;
2275
2276 /* LONGTERM can only be specified when pinning */
2277 if (WARN_ON_ONCE(!(gup_flags & FOLL_PIN) && (gup_flags & FOLL_LONGTERM)))
2278 return false;
2279
2280 /* Pages input must be given if using GET/PIN */
2281 if (WARN_ON_ONCE((gup_flags & (FOLL_GET | FOLL_PIN)) && !pages))
2282 return false;
2283
2284 /* We want to allow the pgmap to be hot-unplugged at all times */
2285 if (WARN_ON_ONCE((gup_flags & FOLL_LONGTERM) &&
2286 (gup_flags & FOLL_PCI_P2PDMA)))
2287 return false;
2288
2289 *gup_flags_p = gup_flags;
2290 return true;
2291}
2292
2293#ifdef CONFIG_MMU
2294/**
2295 * get_user_pages_remote() - pin user pages in memory
2296 * @mm: mm_struct of target mm
2297 * @start: starting user address
2298 * @nr_pages: number of pages from start to pin
2299 * @gup_flags: flags modifying lookup behaviour
2300 * @pages: array that receives pointers to the pages pinned.
2301 * Should be at least nr_pages long. Or NULL, if caller
2302 * only intends to ensure the pages are faulted in.
2303 * @locked: pointer to lock flag indicating whether lock is held and
2304 * subsequently whether VM_FAULT_RETRY functionality can be
2305 * utilised. Lock must initially be held.
2306 *
2307 * Returns either number of pages pinned (which may be less than the
2308 * number requested), or an error. Details about the return value:
2309 *
2310 * -- If nr_pages is 0, returns 0.
2311 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
2312 * -- If nr_pages is >0, and some pages were pinned, returns the number of
2313 * pages pinned. Again, this may be less than nr_pages.
2314 *
2315 * The caller is responsible for releasing returned @pages, via put_page().
2316 *
2317 * Must be called with mmap_lock held for read or write.
2318 *
2319 * get_user_pages_remote walks a process's page tables and takes a reference
2320 * to each struct page that each user address corresponds to at a given
2321 * instant. That is, it takes the page that would be accessed if a user
2322 * thread accesses the given user virtual address at that instant.
2323 *
2324 * This does not guarantee that the page exists in the user mappings when
2325 * get_user_pages_remote returns, and there may even be a completely different
2326 * page there in some cases (eg. if mmapped pagecache has been invalidated
2327 * and subsequently re-faulted). However it does guarantee that the page
2328 * won't be freed completely. And mostly callers simply care that the page
2329 * contains data that was valid *at some point in time*. Typically, an IO
2330 * or similar operation cannot guarantee anything stronger anyway because
2331 * locks can't be held over the syscall boundary.
2332 *
2333 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
2334 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
2335 * be called after the page is finished with, and before put_page is called.
2336 *
2337 * get_user_pages_remote is typically used for fewer-copy IO operations,
2338 * to get a handle on the memory by some means other than accesses
2339 * via the user virtual addresses. The pages may be submitted for
2340 * DMA to devices or accessed via their kernel linear mapping (via the
2341 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
2342 *
2343 * See also get_user_pages_fast, for performance critical applications.
2344 *
2345 * get_user_pages_remote should be phased out in favor of
2346 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
2347 * should use get_user_pages_remote because it cannot pass
2348 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
2349 */
2350long get_user_pages_remote(struct mm_struct *mm,
2351 unsigned long start, unsigned long nr_pages,
2352 unsigned int gup_flags, struct page **pages,
2353 int *locked)
2354{
2355 int local_locked = 1;
2356
2357 if (!is_valid_gup_args(pages, locked, &gup_flags,
2358 FOLL_TOUCH | FOLL_REMOTE))
2359 return -EINVAL;
2360
2361 return __get_user_pages_locked(mm, start, nr_pages, pages,
2362 locked ? locked : &local_locked,
2363 gup_flags);
2364}
2365EXPORT_SYMBOL(get_user_pages_remote);
2366
2367#else /* CONFIG_MMU */
2368long get_user_pages_remote(struct mm_struct *mm,
2369 unsigned long start, unsigned long nr_pages,
2370 unsigned int gup_flags, struct page **pages,
2371 int *locked)
2372{
2373 return 0;
2374}
2375#endif /* !CONFIG_MMU */
2376
2377/**
2378 * get_user_pages() - pin user pages in memory
2379 * @start: starting user address
2380 * @nr_pages: number of pages from start to pin
2381 * @gup_flags: flags modifying lookup behaviour
2382 * @pages: array that receives pointers to the pages pinned.
2383 * Should be at least nr_pages long. Or NULL, if caller
2384 * only intends to ensure the pages are faulted in.
2385 *
2386 * This is the same as get_user_pages_remote(), just with a less-flexible
2387 * calling convention where we assume that the mm being operated on belongs to
2388 * the current task, and doesn't allow passing of a locked parameter. We also
2389 * obviously don't pass FOLL_REMOTE in here.
2390 */
2391long get_user_pages(unsigned long start, unsigned long nr_pages,
2392 unsigned int gup_flags, struct page **pages)
2393{
2394 int locked = 1;
2395
2396 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_TOUCH))
2397 return -EINVAL;
2398
2399 return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2400 &locked, gup_flags);
2401}
2402EXPORT_SYMBOL(get_user_pages);
2403
2404/*
2405 * get_user_pages_unlocked() is suitable to replace the form:
2406 *
2407 * mmap_read_lock(mm);
2408 * get_user_pages(mm, ..., pages, NULL);
2409 * mmap_read_unlock(mm);
2410 *
2411 * with:
2412 *
2413 * get_user_pages_unlocked(mm, ..., pages);
2414 *
2415 * It is functionally equivalent to get_user_pages_fast so
2416 * get_user_pages_fast should be used instead if specific gup_flags
2417 * (e.g. FOLL_FORCE) are not required.
2418 */
2419long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2420 struct page **pages, unsigned int gup_flags)
2421{
2422 int locked = 0;
2423
2424 if (!is_valid_gup_args(pages, NULL, &gup_flags,
2425 FOLL_TOUCH | FOLL_UNLOCKABLE))
2426 return -EINVAL;
2427
2428 return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2429 &locked, gup_flags);
2430}
2431EXPORT_SYMBOL(get_user_pages_unlocked);
2432
2433/*
2434 * Fast GUP
2435 *
2436 * get_user_pages_fast attempts to pin user pages by walking the page
2437 * tables directly and avoids taking locks. Thus the walker needs to be
2438 * protected from page table pages being freed from under it, and should
2439 * block any THP splits.
2440 *
2441 * One way to achieve this is to have the walker disable interrupts, and
2442 * rely on IPIs from the TLB flushing code blocking before the page table
2443 * pages are freed. This is unsuitable for architectures that do not need
2444 * to broadcast an IPI when invalidating TLBs.
2445 *
2446 * Another way to achieve this is to batch up page table containing pages
2447 * belonging to more than one mm_user, then rcu_sched a callback to free those
2448 * pages. Disabling interrupts will allow the fast_gup walker to both block
2449 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2450 * (which is a relatively rare event). The code below adopts this strategy.
2451 *
2452 * Before activating this code, please be aware that the following assumptions
2453 * are currently made:
2454 *
2455 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2456 * free pages containing page tables or TLB flushing requires IPI broadcast.
2457 *
2458 * *) ptes can be read atomically by the architecture.
2459 *
2460 * *) access_ok is sufficient to validate userspace address ranges.
2461 *
2462 * The last two assumptions can be relaxed by the addition of helper functions.
2463 *
2464 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2465 */
2466#ifdef CONFIG_HAVE_FAST_GUP
2467
2468/*
2469 * Used in the GUP-fast path to determine whether a pin is permitted for a
2470 * specific folio.
2471 *
2472 * This call assumes the caller has pinned the folio, that the lowest page table
2473 * level still points to this folio, and that interrupts have been disabled.
2474 *
2475 * Writing to pinned file-backed dirty tracked folios is inherently problematic
2476 * (see comment describing the writable_file_mapping_allowed() function). We
2477 * therefore try to avoid the most egregious case of a long-term mapping doing
2478 * so.
2479 *
2480 * This function cannot be as thorough as that one as the VMA is not available
2481 * in the fast path, so instead we whitelist known good cases and if in doubt,
2482 * fall back to the slow path.
2483 */
2484static bool folio_fast_pin_allowed(struct folio *folio, unsigned int flags)
2485{
2486 struct address_space *mapping;
2487 unsigned long mapping_flags;
2488
2489 /*
2490 * If we aren't pinning then no problematic write can occur. A long term
2491 * pin is the most egregious case so this is the one we disallow.
2492 */
2493 if ((flags & (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE)) !=
2494 (FOLL_PIN | FOLL_LONGTERM | FOLL_WRITE))
2495 return true;
2496
2497 /* The folio is pinned, so we can safely access folio fields. */
2498
2499 if (WARN_ON_ONCE(folio_test_slab(folio)))
2500 return false;
2501
2502 /* hugetlb mappings do not require dirty-tracking. */
2503 if (folio_test_hugetlb(folio))
2504 return true;
2505
2506 /*
2507 * GUP-fast disables IRQs. When IRQS are disabled, RCU grace periods
2508 * cannot proceed, which means no actions performed under RCU can
2509 * proceed either.
2510 *
2511 * inodes and thus their mappings are freed under RCU, which means the
2512 * mapping cannot be freed beneath us and thus we can safely dereference
2513 * it.
2514 */
2515 lockdep_assert_irqs_disabled();
2516
2517 /*
2518 * However, there may be operations which _alter_ the mapping, so ensure
2519 * we read it once and only once.
2520 */
2521 mapping = READ_ONCE(folio->mapping);
2522
2523 /*
2524 * The mapping may have been truncated, in any case we cannot determine
2525 * if this mapping is safe - fall back to slow path to determine how to
2526 * proceed.
2527 */
2528 if (!mapping)
2529 return false;
2530
2531 /* Anonymous folios pose no problem. */
2532 mapping_flags = (unsigned long)mapping & PAGE_MAPPING_FLAGS;
2533 if (mapping_flags)
2534 return mapping_flags & PAGE_MAPPING_ANON;
2535
2536 /*
2537 * At this point, we know the mapping is non-null and points to an
2538 * address_space object. The only remaining whitelisted file system is
2539 * shmem.
2540 */
2541 return shmem_mapping(mapping);
2542}
2543
2544static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2545 unsigned int flags,
2546 struct page **pages)
2547{
2548 while ((*nr) - nr_start) {
2549 struct page *page = pages[--(*nr)];
2550
2551 ClearPageReferenced(page);
2552 if (flags & FOLL_PIN)
2553 unpin_user_page(page);
2554 else
2555 put_page(page);
2556 }
2557}
2558
2559#ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2560/*
2561 * Fast-gup relies on pte change detection to avoid concurrent pgtable
2562 * operations.
2563 *
2564 * To pin the page, fast-gup needs to do below in order:
2565 * (1) pin the page (by prefetching pte), then (2) check pte not changed.
2566 *
2567 * For the rest of pgtable operations where pgtable updates can be racy
2568 * with fast-gup, we need to do (1) clear pte, then (2) check whether page
2569 * is pinned.
2570 *
2571 * Above will work for all pte-level operations, including THP split.
2572 *
2573 * For THP collapse, it's a bit more complicated because fast-gup may be
2574 * walking a pgtable page that is being freed (pte is still valid but pmd
2575 * can be cleared already). To avoid race in such condition, we need to
2576 * also check pmd here to make sure pmd doesn't change (corresponds to
2577 * pmdp_collapse_flush() in the THP collapse code path).
2578 */
2579static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2580 unsigned long end, unsigned int flags,
2581 struct page **pages, int *nr)
2582{
2583 struct dev_pagemap *pgmap = NULL;
2584 int nr_start = *nr, ret = 0;
2585 pte_t *ptep, *ptem;
2586
2587 ptem = ptep = pte_offset_map(&pmd, addr);
2588 if (!ptep)
2589 return 0;
2590 do {
2591 pte_t pte = ptep_get_lockless(ptep);
2592 struct page *page;
2593 struct folio *folio;
2594
2595 /*
2596 * Always fallback to ordinary GUP on PROT_NONE-mapped pages:
2597 * pte_access_permitted() better should reject these pages
2598 * either way: otherwise, GUP-fast might succeed in
2599 * cases where ordinary GUP would fail due to VMA access
2600 * permissions.
2601 */
2602 if (pte_protnone(pte))
2603 goto pte_unmap;
2604
2605 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2606 goto pte_unmap;
2607
2608 if (pte_devmap(pte)) {
2609 if (unlikely(flags & FOLL_LONGTERM))
2610 goto pte_unmap;
2611
2612 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2613 if (unlikely(!pgmap)) {
2614 undo_dev_pagemap(nr, nr_start, flags, pages);
2615 goto pte_unmap;
2616 }
2617 } else if (pte_special(pte))
2618 goto pte_unmap;
2619
2620 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2621 page = pte_page(pte);
2622
2623 folio = try_grab_folio(page, 1, flags);
2624 if (!folio)
2625 goto pte_unmap;
2626
2627 if (unlikely(folio_is_secretmem(folio))) {
2628 gup_put_folio(folio, 1, flags);
2629 goto pte_unmap;
2630 }
2631
2632 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp)) ||
2633 unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) {
2634 gup_put_folio(folio, 1, flags);
2635 goto pte_unmap;
2636 }
2637
2638 if (!folio_fast_pin_allowed(folio, flags)) {
2639 gup_put_folio(folio, 1, flags);
2640 goto pte_unmap;
2641 }
2642
2643 if (!pte_write(pte) && gup_must_unshare(NULL, flags, page)) {
2644 gup_put_folio(folio, 1, flags);
2645 goto pte_unmap;
2646 }
2647
2648 /*
2649 * We need to make the page accessible if and only if we are
2650 * going to access its content (the FOLL_PIN case). Please
2651 * see Documentation/core-api/pin_user_pages.rst for
2652 * details.
2653 */
2654 if (flags & FOLL_PIN) {
2655 ret = arch_make_page_accessible(page);
2656 if (ret) {
2657 gup_put_folio(folio, 1, flags);
2658 goto pte_unmap;
2659 }
2660 }
2661 folio_set_referenced(folio);
2662 pages[*nr] = page;
2663 (*nr)++;
2664 } while (ptep++, addr += PAGE_SIZE, addr != end);
2665
2666 ret = 1;
2667
2668pte_unmap:
2669 if (pgmap)
2670 put_dev_pagemap(pgmap);
2671 pte_unmap(ptem);
2672 return ret;
2673}
2674#else
2675
2676/*
2677 * If we can't determine whether or not a pte is special, then fail immediately
2678 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2679 * to be special.
2680 *
2681 * For a futex to be placed on a THP tail page, get_futex_key requires a
2682 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2683 * useful to have gup_huge_pmd even if we can't operate on ptes.
2684 */
2685static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2686 unsigned long end, unsigned int flags,
2687 struct page **pages, int *nr)
2688{
2689 return 0;
2690}
2691#endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2692
2693#if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2694static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2695 unsigned long end, unsigned int flags,
2696 struct page **pages, int *nr)
2697{
2698 int nr_start = *nr;
2699 struct dev_pagemap *pgmap = NULL;
2700
2701 do {
2702 struct page *page = pfn_to_page(pfn);
2703
2704 pgmap = get_dev_pagemap(pfn, pgmap);
2705 if (unlikely(!pgmap)) {
2706 undo_dev_pagemap(nr, nr_start, flags, pages);
2707 break;
2708 }
2709
2710 if (!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)) {
2711 undo_dev_pagemap(nr, nr_start, flags, pages);
2712 break;
2713 }
2714
2715 SetPageReferenced(page);
2716 pages[*nr] = page;
2717 if (unlikely(try_grab_page(page, flags))) {
2718 undo_dev_pagemap(nr, nr_start, flags, pages);
2719 break;
2720 }
2721 (*nr)++;
2722 pfn++;
2723 } while (addr += PAGE_SIZE, addr != end);
2724
2725 put_dev_pagemap(pgmap);
2726 return addr == end;
2727}
2728
2729static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2730 unsigned long end, unsigned int flags,
2731 struct page **pages, int *nr)
2732{
2733 unsigned long fault_pfn;
2734 int nr_start = *nr;
2735
2736 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2737 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2738 return 0;
2739
2740 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2741 undo_dev_pagemap(nr, nr_start, flags, pages);
2742 return 0;
2743 }
2744 return 1;
2745}
2746
2747static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2748 unsigned long end, unsigned int flags,
2749 struct page **pages, int *nr)
2750{
2751 unsigned long fault_pfn;
2752 int nr_start = *nr;
2753
2754 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2755 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2756 return 0;
2757
2758 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2759 undo_dev_pagemap(nr, nr_start, flags, pages);
2760 return 0;
2761 }
2762 return 1;
2763}
2764#else
2765static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2766 unsigned long end, unsigned int flags,
2767 struct page **pages, int *nr)
2768{
2769 BUILD_BUG();
2770 return 0;
2771}
2772
2773static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2774 unsigned long end, unsigned int flags,
2775 struct page **pages, int *nr)
2776{
2777 BUILD_BUG();
2778 return 0;
2779}
2780#endif
2781
2782static int record_subpages(struct page *page, unsigned long addr,
2783 unsigned long end, struct page **pages)
2784{
2785 int nr;
2786
2787 for (nr = 0; addr != end; nr++, addr += PAGE_SIZE)
2788 pages[nr] = nth_page(page, nr);
2789
2790 return nr;
2791}
2792
2793#ifdef CONFIG_ARCH_HAS_HUGEPD
2794static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2795 unsigned long sz)
2796{
2797 unsigned long __boundary = (addr + sz) & ~(sz-1);
2798 return (__boundary - 1 < end - 1) ? __boundary : end;
2799}
2800
2801static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2802 unsigned long end, unsigned int flags,
2803 struct page **pages, int *nr)
2804{
2805 unsigned long pte_end;
2806 struct page *page;
2807 struct folio *folio;
2808 pte_t pte;
2809 int refs;
2810
2811 pte_end = (addr + sz) & ~(sz-1);
2812 if (pte_end < end)
2813 end = pte_end;
2814
2815 pte = huge_ptep_get(ptep);
2816
2817 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2818 return 0;
2819
2820 /* hugepages are never "special" */
2821 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2822
2823 page = nth_page(pte_page(pte), (addr & (sz - 1)) >> PAGE_SHIFT);
2824 refs = record_subpages(page, addr, end, pages + *nr);
2825
2826 folio = try_grab_folio(page, refs, flags);
2827 if (!folio)
2828 return 0;
2829
2830 if (unlikely(pte_val(pte) != pte_val(ptep_get(ptep)))) {
2831 gup_put_folio(folio, refs, flags);
2832 return 0;
2833 }
2834
2835 if (!folio_fast_pin_allowed(folio, flags)) {
2836 gup_put_folio(folio, refs, flags);
2837 return 0;
2838 }
2839
2840 if (!pte_write(pte) && gup_must_unshare(NULL, flags, &folio->page)) {
2841 gup_put_folio(folio, refs, flags);
2842 return 0;
2843 }
2844
2845 *nr += refs;
2846 folio_set_referenced(folio);
2847 return 1;
2848}
2849
2850static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2851 unsigned int pdshift, unsigned long end, unsigned int flags,
2852 struct page **pages, int *nr)
2853{
2854 pte_t *ptep;
2855 unsigned long sz = 1UL << hugepd_shift(hugepd);
2856 unsigned long next;
2857
2858 ptep = hugepte_offset(hugepd, addr, pdshift);
2859 do {
2860 next = hugepte_addr_end(addr, end, sz);
2861 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2862 return 0;
2863 } while (ptep++, addr = next, addr != end);
2864
2865 return 1;
2866}
2867#else
2868static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2869 unsigned int pdshift, unsigned long end, unsigned int flags,
2870 struct page **pages, int *nr)
2871{
2872 return 0;
2873}
2874#endif /* CONFIG_ARCH_HAS_HUGEPD */
2875
2876static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2877 unsigned long end, unsigned int flags,
2878 struct page **pages, int *nr)
2879{
2880 struct page *page;
2881 struct folio *folio;
2882 int refs;
2883
2884 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2885 return 0;
2886
2887 if (pmd_devmap(orig)) {
2888 if (unlikely(flags & FOLL_LONGTERM))
2889 return 0;
2890 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2891 pages, nr);
2892 }
2893
2894 page = nth_page(pmd_page(orig), (addr & ~PMD_MASK) >> PAGE_SHIFT);
2895 refs = record_subpages(page, addr, end, pages + *nr);
2896
2897 folio = try_grab_folio(page, refs, flags);
2898 if (!folio)
2899 return 0;
2900
2901 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2902 gup_put_folio(folio, refs, flags);
2903 return 0;
2904 }
2905
2906 if (!folio_fast_pin_allowed(folio, flags)) {
2907 gup_put_folio(folio, refs, flags);
2908 return 0;
2909 }
2910 if (!pmd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2911 gup_put_folio(folio, refs, flags);
2912 return 0;
2913 }
2914
2915 *nr += refs;
2916 folio_set_referenced(folio);
2917 return 1;
2918}
2919
2920static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2921 unsigned long end, unsigned int flags,
2922 struct page **pages, int *nr)
2923{
2924 struct page *page;
2925 struct folio *folio;
2926 int refs;
2927
2928 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2929 return 0;
2930
2931 if (pud_devmap(orig)) {
2932 if (unlikely(flags & FOLL_LONGTERM))
2933 return 0;
2934 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2935 pages, nr);
2936 }
2937
2938 page = nth_page(pud_page(orig), (addr & ~PUD_MASK) >> PAGE_SHIFT);
2939 refs = record_subpages(page, addr, end, pages + *nr);
2940
2941 folio = try_grab_folio(page, refs, flags);
2942 if (!folio)
2943 return 0;
2944
2945 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2946 gup_put_folio(folio, refs, flags);
2947 return 0;
2948 }
2949
2950 if (!folio_fast_pin_allowed(folio, flags)) {
2951 gup_put_folio(folio, refs, flags);
2952 return 0;
2953 }
2954
2955 if (!pud_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2956 gup_put_folio(folio, refs, flags);
2957 return 0;
2958 }
2959
2960 *nr += refs;
2961 folio_set_referenced(folio);
2962 return 1;
2963}
2964
2965static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2966 unsigned long end, unsigned int flags,
2967 struct page **pages, int *nr)
2968{
2969 int refs;
2970 struct page *page;
2971 struct folio *folio;
2972
2973 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2974 return 0;
2975
2976 BUILD_BUG_ON(pgd_devmap(orig));
2977
2978 page = nth_page(pgd_page(orig), (addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2979 refs = record_subpages(page, addr, end, pages + *nr);
2980
2981 folio = try_grab_folio(page, refs, flags);
2982 if (!folio)
2983 return 0;
2984
2985 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2986 gup_put_folio(folio, refs, flags);
2987 return 0;
2988 }
2989
2990 if (!pgd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2991 gup_put_folio(folio, refs, flags);
2992 return 0;
2993 }
2994
2995 if (!folio_fast_pin_allowed(folio, flags)) {
2996 gup_put_folio(folio, refs, flags);
2997 return 0;
2998 }
2999
3000 *nr += refs;
3001 folio_set_referenced(folio);
3002 return 1;
3003}
3004
3005static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
3006 unsigned int flags, struct page **pages, int *nr)
3007{
3008 unsigned long next;
3009 pmd_t *pmdp;
3010
3011 pmdp = pmd_offset_lockless(pudp, pud, addr);
3012 do {
3013 pmd_t pmd = pmdp_get_lockless(pmdp);
3014
3015 next = pmd_addr_end(addr, end);
3016 if (!pmd_present(pmd))
3017 return 0;
3018
3019 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
3020 pmd_devmap(pmd))) {
3021 /* See gup_pte_range() */
3022 if (pmd_protnone(pmd))
3023 return 0;
3024
3025 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
3026 pages, nr))
3027 return 0;
3028
3029 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
3030 /*
3031 * architecture have different format for hugetlbfs
3032 * pmd format and THP pmd format
3033 */
3034 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
3035 PMD_SHIFT, next, flags, pages, nr))
3036 return 0;
3037 } else if (!gup_pte_range(pmd, pmdp, addr, next, flags, pages, nr))
3038 return 0;
3039 } while (pmdp++, addr = next, addr != end);
3040
3041 return 1;
3042}
3043
3044static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
3045 unsigned int flags, struct page **pages, int *nr)
3046{
3047 unsigned long next;
3048 pud_t *pudp;
3049
3050 pudp = pud_offset_lockless(p4dp, p4d, addr);
3051 do {
3052 pud_t pud = READ_ONCE(*pudp);
3053
3054 next = pud_addr_end(addr, end);
3055 if (unlikely(!pud_present(pud)))
3056 return 0;
3057 if (unlikely(pud_huge(pud) || pud_devmap(pud))) {
3058 if (!gup_huge_pud(pud, pudp, addr, next, flags,
3059 pages, nr))
3060 return 0;
3061 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
3062 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
3063 PUD_SHIFT, next, flags, pages, nr))
3064 return 0;
3065 } else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
3066 return 0;
3067 } while (pudp++, addr = next, addr != end);
3068
3069 return 1;
3070}
3071
3072static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
3073 unsigned int flags, struct page **pages, int *nr)
3074{
3075 unsigned long next;
3076 p4d_t *p4dp;
3077
3078 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
3079 do {
3080 p4d_t p4d = READ_ONCE(*p4dp);
3081
3082 next = p4d_addr_end(addr, end);
3083 if (p4d_none(p4d))
3084 return 0;
3085 BUILD_BUG_ON(p4d_huge(p4d));
3086 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
3087 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
3088 P4D_SHIFT, next, flags, pages, nr))
3089 return 0;
3090 } else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
3091 return 0;
3092 } while (p4dp++, addr = next, addr != end);
3093
3094 return 1;
3095}
3096
3097static void gup_pgd_range(unsigned long addr, unsigned long end,
3098 unsigned int flags, struct page **pages, int *nr)
3099{
3100 unsigned long next;
3101 pgd_t *pgdp;
3102
3103 pgdp = pgd_offset(current->mm, addr);
3104 do {
3105 pgd_t pgd = READ_ONCE(*pgdp);
3106
3107 next = pgd_addr_end(addr, end);
3108 if (pgd_none(pgd))
3109 return;
3110 if (unlikely(pgd_huge(pgd))) {
3111 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
3112 pages, nr))
3113 return;
3114 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
3115 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
3116 PGDIR_SHIFT, next, flags, pages, nr))
3117 return;
3118 } else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
3119 return;
3120 } while (pgdp++, addr = next, addr != end);
3121}
3122#else
3123static inline void gup_pgd_range(unsigned long addr, unsigned long end,
3124 unsigned int flags, struct page **pages, int *nr)
3125{
3126}
3127#endif /* CONFIG_HAVE_FAST_GUP */
3128
3129#ifndef gup_fast_permitted
3130/*
3131 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
3132 * we need to fall back to the slow version:
3133 */
3134static bool gup_fast_permitted(unsigned long start, unsigned long end)
3135{
3136 return true;
3137}
3138#endif
3139
3140static unsigned long lockless_pages_from_mm(unsigned long start,
3141 unsigned long end,
3142 unsigned int gup_flags,
3143 struct page **pages)
3144{
3145 unsigned long flags;
3146 int nr_pinned = 0;
3147 unsigned seq;
3148
3149 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP) ||
3150 !gup_fast_permitted(start, end))
3151 return 0;
3152
3153 if (gup_flags & FOLL_PIN) {
3154 seq = raw_read_seqcount(¤t->mm->write_protect_seq);
3155 if (seq & 1)
3156 return 0;
3157 }
3158
3159 /*
3160 * Disable interrupts. The nested form is used, in order to allow full,
3161 * general purpose use of this routine.
3162 *
3163 * With interrupts disabled, we block page table pages from being freed
3164 * from under us. See struct mmu_table_batch comments in
3165 * include/asm-generic/tlb.h for more details.
3166 *
3167 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
3168 * that come from THPs splitting.
3169 */
3170 local_irq_save(flags);
3171 gup_pgd_range(start, end, gup_flags, pages, &nr_pinned);
3172 local_irq_restore(flags);
3173
3174 /*
3175 * When pinning pages for DMA there could be a concurrent write protect
3176 * from fork() via copy_page_range(), in this case always fail fast GUP.
3177 */
3178 if (gup_flags & FOLL_PIN) {
3179 if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) {
3180 unpin_user_pages_lockless(pages, nr_pinned);
3181 return 0;
3182 } else {
3183 sanity_check_pinned_pages(pages, nr_pinned);
3184 }
3185 }
3186 return nr_pinned;
3187}
3188
3189static int internal_get_user_pages_fast(unsigned long start,
3190 unsigned long nr_pages,
3191 unsigned int gup_flags,
3192 struct page **pages)
3193{
3194 unsigned long len, end;
3195 unsigned long nr_pinned;
3196 int locked = 0;
3197 int ret;
3198
3199 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
3200 FOLL_FORCE | FOLL_PIN | FOLL_GET |
3201 FOLL_FAST_ONLY | FOLL_NOFAULT |
3202 FOLL_PCI_P2PDMA | FOLL_HONOR_NUMA_FAULT)))
3203 return -EINVAL;
3204
3205 if (gup_flags & FOLL_PIN)
3206 mm_set_has_pinned_flag(¤t->mm->flags);
3207
3208 if (!(gup_flags & FOLL_FAST_ONLY))
3209 might_lock_read(¤t->mm->mmap_lock);
3210
3211 start = untagged_addr(start) & PAGE_MASK;
3212 len = nr_pages << PAGE_SHIFT;
3213 if (check_add_overflow(start, len, &end))
3214 return -EOVERFLOW;
3215 if (end > TASK_SIZE_MAX)
3216 return -EFAULT;
3217 if (unlikely(!access_ok((void __user *)start, len)))
3218 return -EFAULT;
3219
3220 nr_pinned = lockless_pages_from_mm(start, end, gup_flags, pages);
3221 if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
3222 return nr_pinned;
3223
3224 /* Slow path: try to get the remaining pages with get_user_pages */
3225 start += nr_pinned << PAGE_SHIFT;
3226 pages += nr_pinned;
3227 ret = __gup_longterm_locked(current->mm, start, nr_pages - nr_pinned,
3228 pages, &locked,
3229 gup_flags | FOLL_TOUCH | FOLL_UNLOCKABLE);
3230 if (ret < 0) {
3231 /*
3232 * The caller has to unpin the pages we already pinned so
3233 * returning -errno is not an option
3234 */
3235 if (nr_pinned)
3236 return nr_pinned;
3237 return ret;
3238 }
3239 return ret + nr_pinned;
3240}
3241
3242/**
3243 * get_user_pages_fast_only() - pin user pages in memory
3244 * @start: starting user address
3245 * @nr_pages: number of pages from start to pin
3246 * @gup_flags: flags modifying pin behaviour
3247 * @pages: array that receives pointers to the pages pinned.
3248 * Should be at least nr_pages long.
3249 *
3250 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
3251 * the regular GUP.
3252 *
3253 * If the architecture does not support this function, simply return with no
3254 * pages pinned.
3255 *
3256 * Careful, careful! COW breaking can go either way, so a non-write
3257 * access can get ambiguous page results. If you call this function without
3258 * 'write' set, you'd better be sure that you're ok with that ambiguity.
3259 */
3260int get_user_pages_fast_only(unsigned long start, int nr_pages,
3261 unsigned int gup_flags, struct page **pages)
3262{
3263 /*
3264 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
3265 * because gup fast is always a "pin with a +1 page refcount" request.
3266 *
3267 * FOLL_FAST_ONLY is required in order to match the API description of
3268 * this routine: no fall back to regular ("slow") GUP.
3269 */
3270 if (!is_valid_gup_args(pages, NULL, &gup_flags,
3271 FOLL_GET | FOLL_FAST_ONLY))
3272 return -EINVAL;
3273
3274 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3275}
3276EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
3277
3278/**
3279 * get_user_pages_fast() - pin user pages in memory
3280 * @start: starting user address
3281 * @nr_pages: number of pages from start to pin
3282 * @gup_flags: flags modifying pin behaviour
3283 * @pages: array that receives pointers to the pages pinned.
3284 * Should be at least nr_pages long.
3285 *
3286 * Attempt to pin user pages in memory without taking mm->mmap_lock.
3287 * If not successful, it will fall back to taking the lock and
3288 * calling get_user_pages().
3289 *
3290 * Returns number of pages pinned. This may be fewer than the number requested.
3291 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
3292 * -errno.
3293 */
3294int get_user_pages_fast(unsigned long start, int nr_pages,
3295 unsigned int gup_flags, struct page **pages)
3296{
3297 /*
3298 * The caller may or may not have explicitly set FOLL_GET; either way is
3299 * OK. However, internally (within mm/gup.c), gup fast variants must set
3300 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
3301 * request.
3302 */
3303 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_GET))
3304 return -EINVAL;
3305 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3306}
3307EXPORT_SYMBOL_GPL(get_user_pages_fast);
3308
3309/**
3310 * pin_user_pages_fast() - pin user pages in memory without taking locks
3311 *
3312 * @start: starting user address
3313 * @nr_pages: number of pages from start to pin
3314 * @gup_flags: flags modifying pin behaviour
3315 * @pages: array that receives pointers to the pages pinned.
3316 * Should be at least nr_pages long.
3317 *
3318 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
3319 * get_user_pages_fast() for documentation on the function arguments, because
3320 * the arguments here are identical.
3321 *
3322 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3323 * see Documentation/core-api/pin_user_pages.rst for further details.
3324 *
3325 * Note that if a zero_page is amongst the returned pages, it will not have
3326 * pins in it and unpin_user_page() will not remove pins from it.
3327 */
3328int pin_user_pages_fast(unsigned long start, int nr_pages,
3329 unsigned int gup_flags, struct page **pages)
3330{
3331 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN))
3332 return -EINVAL;
3333 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3334}
3335EXPORT_SYMBOL_GPL(pin_user_pages_fast);
3336
3337/**
3338 * pin_user_pages_remote() - pin pages of a remote process
3339 *
3340 * @mm: mm_struct of target mm
3341 * @start: starting user address
3342 * @nr_pages: number of pages from start to pin
3343 * @gup_flags: flags modifying lookup behaviour
3344 * @pages: array that receives pointers to the pages pinned.
3345 * Should be at least nr_pages long.
3346 * @locked: pointer to lock flag indicating whether lock is held and
3347 * subsequently whether VM_FAULT_RETRY functionality can be
3348 * utilised. Lock must initially be held.
3349 *
3350 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
3351 * get_user_pages_remote() for documentation on the function arguments, because
3352 * the arguments here are identical.
3353 *
3354 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3355 * see Documentation/core-api/pin_user_pages.rst for details.
3356 *
3357 * Note that if a zero_page is amongst the returned pages, it will not have
3358 * pins in it and unpin_user_page*() will not remove pins from it.
3359 */
3360long pin_user_pages_remote(struct mm_struct *mm,
3361 unsigned long start, unsigned long nr_pages,
3362 unsigned int gup_flags, struct page **pages,
3363 int *locked)
3364{
3365 int local_locked = 1;
3366
3367 if (!is_valid_gup_args(pages, locked, &gup_flags,
3368 FOLL_PIN | FOLL_TOUCH | FOLL_REMOTE))
3369 return 0;
3370 return __gup_longterm_locked(mm, start, nr_pages, pages,
3371 locked ? locked : &local_locked,
3372 gup_flags);
3373}
3374EXPORT_SYMBOL(pin_user_pages_remote);
3375
3376/**
3377 * pin_user_pages() - pin user pages in memory for use by other devices
3378 *
3379 * @start: starting user address
3380 * @nr_pages: number of pages from start to pin
3381 * @gup_flags: flags modifying lookup behaviour
3382 * @pages: array that receives pointers to the pages pinned.
3383 * Should be at least nr_pages long.
3384 *
3385 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
3386 * FOLL_PIN is set.
3387 *
3388 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3389 * see Documentation/core-api/pin_user_pages.rst for details.
3390 *
3391 * Note that if a zero_page is amongst the returned pages, it will not have
3392 * pins in it and unpin_user_page*() will not remove pins from it.
3393 */
3394long pin_user_pages(unsigned long start, unsigned long nr_pages,
3395 unsigned int gup_flags, struct page **pages)
3396{
3397 int locked = 1;
3398
3399 if (!is_valid_gup_args(pages, NULL, &gup_flags, FOLL_PIN))
3400 return 0;
3401 return __gup_longterm_locked(current->mm, start, nr_pages,
3402 pages, &locked, gup_flags);
3403}
3404EXPORT_SYMBOL(pin_user_pages);
3405
3406/*
3407 * pin_user_pages_unlocked() is the FOLL_PIN variant of
3408 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
3409 * FOLL_PIN and rejects FOLL_GET.
3410 *
3411 * Note that if a zero_page is amongst the returned pages, it will not have
3412 * pins in it and unpin_user_page*() will not remove pins from it.
3413 */
3414long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
3415 struct page **pages, unsigned int gup_flags)
3416{
3417 int locked = 0;
3418
3419 if (!is_valid_gup_args(pages, NULL, &gup_flags,
3420 FOLL_PIN | FOLL_TOUCH | FOLL_UNLOCKABLE))
3421 return 0;
3422
3423 return __gup_longterm_locked(current->mm, start, nr_pages, pages,
3424 &locked, gup_flags);
3425}
3426EXPORT_SYMBOL(pin_user_pages_unlocked);