1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
5 #include <linux/spinlock.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>
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>
21 #include <asm/mmu_context.h>
22 #include <asm/pgtable.h>
23 #include <asm/tlbflush.h>
27 struct follow_page_context
{
28 struct dev_pagemap
*pgmap
;
29 unsigned int page_mask
;
32 typedef int (*set_dirty_func_t
)(struct page
*page
);
34 static void __put_user_pages_dirty(struct page
**pages
,
40 for (index
= 0; index
< npages
; index
++) {
41 struct page
*page
= compound_head(pages
[index
]);
44 * Checking PageDirty at this point may race with
45 * clear_page_dirty_for_io(), but that's OK. Two key cases:
47 * 1) This code sees the page as already dirty, so it skips
48 * the call to sdf(). That could happen because
49 * clear_page_dirty_for_io() called page_mkclean(),
50 * followed by set_page_dirty(). However, now the page is
51 * going to get written back, which meets the original
52 * intention of setting it dirty, so all is well:
53 * clear_page_dirty_for_io() goes on to call
54 * TestClearPageDirty(), and write the page back.
56 * 2) This code sees the page as clean, so it calls sdf().
57 * The page stays dirty, despite being written back, so it
58 * gets written back again in the next writeback cycle.
69 * put_user_pages_dirty() - release and dirty an array of gup-pinned pages
70 * @pages: array of pages to be marked dirty and released.
71 * @npages: number of pages in the @pages array.
73 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
74 * variants called on that page.
76 * For each page in the @pages array, make that page (or its head page, if a
77 * compound page) dirty, if it was previously listed as clean. Then, release
78 * the page using put_user_page().
80 * Please see the put_user_page() documentation for details.
82 * set_page_dirty(), which does not lock the page, is used here.
83 * Therefore, it is the caller's responsibility to ensure that this is
84 * safe. If not, then put_user_pages_dirty_lock() should be called instead.
87 void put_user_pages_dirty(struct page
**pages
, unsigned long npages
)
89 __put_user_pages_dirty(pages
, npages
, set_page_dirty
);
91 EXPORT_SYMBOL(put_user_pages_dirty
);
94 * put_user_pages_dirty_lock() - release and dirty an array of gup-pinned pages
95 * @pages: array of pages to be marked dirty and released.
96 * @npages: number of pages in the @pages array.
98 * For each page in the @pages array, make that page (or its head page, if a
99 * compound page) dirty, if it was previously listed as clean. Then, release
100 * the page using put_user_page().
102 * Please see the put_user_page() documentation for details.
104 * This is just like put_user_pages_dirty(), except that it invokes
105 * set_page_dirty_lock(), instead of set_page_dirty().
108 void put_user_pages_dirty_lock(struct page
**pages
, unsigned long npages
)
110 __put_user_pages_dirty(pages
, npages
, set_page_dirty_lock
);
112 EXPORT_SYMBOL(put_user_pages_dirty_lock
);
115 * put_user_pages() - release an array of gup-pinned pages.
116 * @pages: array of pages to be marked dirty and released.
117 * @npages: number of pages in the @pages array.
119 * For each page in the @pages array, release the page using put_user_page().
121 * Please see the put_user_page() documentation for details.
123 void put_user_pages(struct page
**pages
, unsigned long npages
)
128 * TODO: this can be optimized for huge pages: if a series of pages is
129 * physically contiguous and part of the same compound page, then a
130 * single operation to the head page should suffice.
132 for (index
= 0; index
< npages
; index
++)
133 put_user_page(pages
[index
]);
135 EXPORT_SYMBOL(put_user_pages
);
137 static struct page
*no_page_table(struct vm_area_struct
*vma
,
141 * When core dumping an enormous anonymous area that nobody
142 * has touched so far, we don't want to allocate unnecessary pages or
143 * page tables. Return error instead of NULL to skip handle_mm_fault,
144 * then get_dump_page() will return NULL to leave a hole in the dump.
145 * But we can only make this optimization where a hole would surely
146 * be zero-filled if handle_mm_fault() actually did handle it.
148 if ((flags
& FOLL_DUMP
) && (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
149 return ERR_PTR(-EFAULT
);
153 static int follow_pfn_pte(struct vm_area_struct
*vma
, unsigned long address
,
154 pte_t
*pte
, unsigned int flags
)
156 /* No page to get reference */
157 if (flags
& FOLL_GET
)
160 if (flags
& FOLL_TOUCH
) {
163 if (flags
& FOLL_WRITE
)
164 entry
= pte_mkdirty(entry
);
165 entry
= pte_mkyoung(entry
);
167 if (!pte_same(*pte
, entry
)) {
168 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
169 update_mmu_cache(vma
, address
, pte
);
173 /* Proper page table entry exists, but no corresponding struct page */
178 * FOLL_FORCE can write to even unwritable pte's, but only
179 * after we've gone through a COW cycle and they are dirty.
181 static inline bool can_follow_write_pte(pte_t pte
, unsigned int flags
)
183 return pte_write(pte
) ||
184 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pte_dirty(pte
));
187 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
188 unsigned long address
, pmd_t
*pmd
, unsigned int flags
,
189 struct dev_pagemap
**pgmap
)
191 struct mm_struct
*mm
= vma
->vm_mm
;
197 if (unlikely(pmd_bad(*pmd
)))
198 return no_page_table(vma
, flags
);
200 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
202 if (!pte_present(pte
)) {
205 * KSM's break_ksm() relies upon recognizing a ksm page
206 * even while it is being migrated, so for that case we
207 * need migration_entry_wait().
209 if (likely(!(flags
& FOLL_MIGRATION
)))
213 entry
= pte_to_swp_entry(pte
);
214 if (!is_migration_entry(entry
))
216 pte_unmap_unlock(ptep
, ptl
);
217 migration_entry_wait(mm
, pmd
, address
);
220 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
222 if ((flags
& FOLL_WRITE
) && !can_follow_write_pte(pte
, flags
)) {
223 pte_unmap_unlock(ptep
, ptl
);
227 page
= vm_normal_page(vma
, address
, pte
);
228 if (!page
&& pte_devmap(pte
) && (flags
& FOLL_GET
)) {
230 * Only return device mapping pages in the FOLL_GET case since
231 * they are only valid while holding the pgmap reference.
233 *pgmap
= get_dev_pagemap(pte_pfn(pte
), *pgmap
);
235 page
= pte_page(pte
);
238 } else if (unlikely(!page
)) {
239 if (flags
& FOLL_DUMP
) {
240 /* Avoid special (like zero) pages in core dumps */
241 page
= ERR_PTR(-EFAULT
);
245 if (is_zero_pfn(pte_pfn(pte
))) {
246 page
= pte_page(pte
);
250 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
256 if (flags
& FOLL_SPLIT
&& PageTransCompound(page
)) {
259 pte_unmap_unlock(ptep
, ptl
);
261 ret
= split_huge_page(page
);
269 if (flags
& FOLL_GET
) {
270 if (unlikely(!try_get_page(page
))) {
271 page
= ERR_PTR(-ENOMEM
);
275 if (flags
& FOLL_TOUCH
) {
276 if ((flags
& FOLL_WRITE
) &&
277 !pte_dirty(pte
) && !PageDirty(page
))
278 set_page_dirty(page
);
280 * pte_mkyoung() would be more correct here, but atomic care
281 * is needed to avoid losing the dirty bit: it is easier to use
282 * mark_page_accessed().
284 mark_page_accessed(page
);
286 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
287 /* Do not mlock pte-mapped THP */
288 if (PageTransCompound(page
))
292 * The preliminary mapping check is mainly to avoid the
293 * pointless overhead of lock_page on the ZERO_PAGE
294 * which might bounce very badly if there is contention.
296 * If the page is already locked, we don't need to
297 * handle it now - vmscan will handle it later if and
298 * when it attempts to reclaim the page.
300 if (page
->mapping
&& trylock_page(page
)) {
301 lru_add_drain(); /* push cached pages to LRU */
303 * Because we lock page here, and migration is
304 * blocked by the pte's page reference, and we
305 * know the page is still mapped, we don't even
306 * need to check for file-cache page truncation.
308 mlock_vma_page(page
);
313 pte_unmap_unlock(ptep
, ptl
);
316 pte_unmap_unlock(ptep
, ptl
);
319 return no_page_table(vma
, flags
);
322 static struct page
*follow_pmd_mask(struct vm_area_struct
*vma
,
323 unsigned long address
, pud_t
*pudp
,
325 struct follow_page_context
*ctx
)
330 struct mm_struct
*mm
= vma
->vm_mm
;
332 pmd
= pmd_offset(pudp
, address
);
334 * The READ_ONCE() will stabilize the pmdval in a register or
335 * on the stack so that it will stop changing under the code.
337 pmdval
= READ_ONCE(*pmd
);
338 if (pmd_none(pmdval
))
339 return no_page_table(vma
, flags
);
340 if (pmd_huge(pmdval
) && vma
->vm_flags
& VM_HUGETLB
) {
341 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
344 return no_page_table(vma
, flags
);
346 if (is_hugepd(__hugepd(pmd_val(pmdval
)))) {
347 page
= follow_huge_pd(vma
, address
,
348 __hugepd(pmd_val(pmdval
)), flags
,
352 return no_page_table(vma
, flags
);
355 if (!pmd_present(pmdval
)) {
356 if (likely(!(flags
& FOLL_MIGRATION
)))
357 return no_page_table(vma
, flags
);
358 VM_BUG_ON(thp_migration_supported() &&
359 !is_pmd_migration_entry(pmdval
));
360 if (is_pmd_migration_entry(pmdval
))
361 pmd_migration_entry_wait(mm
, pmd
);
362 pmdval
= READ_ONCE(*pmd
);
364 * MADV_DONTNEED may convert the pmd to null because
365 * mmap_sem is held in read mode
367 if (pmd_none(pmdval
))
368 return no_page_table(vma
, flags
);
371 if (pmd_devmap(pmdval
)) {
372 ptl
= pmd_lock(mm
, pmd
);
373 page
= follow_devmap_pmd(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
378 if (likely(!pmd_trans_huge(pmdval
)))
379 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
381 if ((flags
& FOLL_NUMA
) && pmd_protnone(pmdval
))
382 return no_page_table(vma
, flags
);
385 ptl
= pmd_lock(mm
, pmd
);
386 if (unlikely(pmd_none(*pmd
))) {
388 return no_page_table(vma
, flags
);
390 if (unlikely(!pmd_present(*pmd
))) {
392 if (likely(!(flags
& FOLL_MIGRATION
)))
393 return no_page_table(vma
, flags
);
394 pmd_migration_entry_wait(mm
, pmd
);
397 if (unlikely(!pmd_trans_huge(*pmd
))) {
399 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
401 if (flags
& FOLL_SPLIT
) {
403 page
= pmd_page(*pmd
);
404 if (is_huge_zero_page(page
)) {
407 split_huge_pmd(vma
, pmd
, address
);
408 if (pmd_trans_unstable(pmd
))
411 if (unlikely(!try_get_page(page
))) {
413 return ERR_PTR(-ENOMEM
);
417 ret
= split_huge_page(page
);
421 return no_page_table(vma
, flags
);
424 return ret
? ERR_PTR(ret
) :
425 follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
427 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
429 ctx
->page_mask
= HPAGE_PMD_NR
- 1;
433 static struct page
*follow_pud_mask(struct vm_area_struct
*vma
,
434 unsigned long address
, p4d_t
*p4dp
,
436 struct follow_page_context
*ctx
)
441 struct mm_struct
*mm
= vma
->vm_mm
;
443 pud
= pud_offset(p4dp
, address
);
445 return no_page_table(vma
, flags
);
446 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
447 page
= follow_huge_pud(mm
, address
, pud
, flags
);
450 return no_page_table(vma
, flags
);
452 if (is_hugepd(__hugepd(pud_val(*pud
)))) {
453 page
= follow_huge_pd(vma
, address
,
454 __hugepd(pud_val(*pud
)), flags
,
458 return no_page_table(vma
, flags
);
460 if (pud_devmap(*pud
)) {
461 ptl
= pud_lock(mm
, pud
);
462 page
= follow_devmap_pud(vma
, address
, pud
, flags
, &ctx
->pgmap
);
467 if (unlikely(pud_bad(*pud
)))
468 return no_page_table(vma
, flags
);
470 return follow_pmd_mask(vma
, address
, pud
, flags
, ctx
);
473 static struct page
*follow_p4d_mask(struct vm_area_struct
*vma
,
474 unsigned long address
, pgd_t
*pgdp
,
476 struct follow_page_context
*ctx
)
481 p4d
= p4d_offset(pgdp
, address
);
483 return no_page_table(vma
, flags
);
484 BUILD_BUG_ON(p4d_huge(*p4d
));
485 if (unlikely(p4d_bad(*p4d
)))
486 return no_page_table(vma
, flags
);
488 if (is_hugepd(__hugepd(p4d_val(*p4d
)))) {
489 page
= follow_huge_pd(vma
, address
,
490 __hugepd(p4d_val(*p4d
)), flags
,
494 return no_page_table(vma
, flags
);
496 return follow_pud_mask(vma
, address
, p4d
, flags
, ctx
);
500 * follow_page_mask - look up a page descriptor from a user-virtual address
501 * @vma: vm_area_struct mapping @address
502 * @address: virtual address to look up
503 * @flags: flags modifying lookup behaviour
504 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
505 * pointer to output page_mask
507 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
509 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
510 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
512 * On output, the @ctx->page_mask is set according to the size of the page.
514 * Return: the mapped (struct page *), %NULL if no mapping exists, or
515 * an error pointer if there is a mapping to something not represented
516 * by a page descriptor (see also vm_normal_page()).
518 struct page
*follow_page_mask(struct vm_area_struct
*vma
,
519 unsigned long address
, unsigned int flags
,
520 struct follow_page_context
*ctx
)
524 struct mm_struct
*mm
= vma
->vm_mm
;
528 /* make this handle hugepd */
529 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
531 BUG_ON(flags
& FOLL_GET
);
535 pgd
= pgd_offset(mm
, address
);
537 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
538 return no_page_table(vma
, flags
);
540 if (pgd_huge(*pgd
)) {
541 page
= follow_huge_pgd(mm
, address
, pgd
, flags
);
544 return no_page_table(vma
, flags
);
546 if (is_hugepd(__hugepd(pgd_val(*pgd
)))) {
547 page
= follow_huge_pd(vma
, address
,
548 __hugepd(pgd_val(*pgd
)), flags
,
552 return no_page_table(vma
, flags
);
555 return follow_p4d_mask(vma
, address
, pgd
, flags
, ctx
);
558 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
559 unsigned int foll_flags
)
561 struct follow_page_context ctx
= { NULL
};
564 page
= follow_page_mask(vma
, address
, foll_flags
, &ctx
);
566 put_dev_pagemap(ctx
.pgmap
);
570 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
571 unsigned int gup_flags
, struct vm_area_struct
**vma
,
581 /* user gate pages are read-only */
582 if (gup_flags
& FOLL_WRITE
)
584 if (address
> TASK_SIZE
)
585 pgd
= pgd_offset_k(address
);
587 pgd
= pgd_offset_gate(mm
, address
);
588 BUG_ON(pgd_none(*pgd
));
589 p4d
= p4d_offset(pgd
, address
);
590 BUG_ON(p4d_none(*p4d
));
591 pud
= pud_offset(p4d
, address
);
592 BUG_ON(pud_none(*pud
));
593 pmd
= pmd_offset(pud
, address
);
594 if (!pmd_present(*pmd
))
596 VM_BUG_ON(pmd_trans_huge(*pmd
));
597 pte
= pte_offset_map(pmd
, address
);
600 *vma
= get_gate_vma(mm
);
603 *page
= vm_normal_page(*vma
, address
, *pte
);
605 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
607 *page
= pte_page(*pte
);
610 * This should never happen (a device public page in the gate
613 if (is_device_public_page(*page
))
616 if (unlikely(!try_get_page(*page
))) {
628 * mmap_sem must be held on entry. If @nonblocking != NULL and
629 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
630 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
632 static int faultin_page(struct task_struct
*tsk
, struct vm_area_struct
*vma
,
633 unsigned long address
, unsigned int *flags
, int *nonblocking
)
635 unsigned int fault_flags
= 0;
638 /* mlock all present pages, but do not fault in new pages */
639 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
641 if (*flags
& FOLL_WRITE
)
642 fault_flags
|= FAULT_FLAG_WRITE
;
643 if (*flags
& FOLL_REMOTE
)
644 fault_flags
|= FAULT_FLAG_REMOTE
;
646 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
647 if (*flags
& FOLL_NOWAIT
)
648 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
649 if (*flags
& FOLL_TRIED
) {
650 VM_WARN_ON_ONCE(fault_flags
& FAULT_FLAG_ALLOW_RETRY
);
651 fault_flags
|= FAULT_FLAG_TRIED
;
654 ret
= handle_mm_fault(vma
, address
, fault_flags
);
655 if (ret
& VM_FAULT_ERROR
) {
656 int err
= vm_fault_to_errno(ret
, *flags
);
664 if (ret
& VM_FAULT_MAJOR
)
670 if (ret
& VM_FAULT_RETRY
) {
671 if (nonblocking
&& !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
677 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
678 * necessary, even if maybe_mkwrite decided not to set pte_write. We
679 * can thus safely do subsequent page lookups as if they were reads.
680 * But only do so when looping for pte_write is futile: in some cases
681 * userspace may also be wanting to write to the gotten user page,
682 * which a read fault here might prevent (a readonly page might get
683 * reCOWed by userspace write).
685 if ((ret
& VM_FAULT_WRITE
) && !(vma
->vm_flags
& VM_WRITE
))
690 static int check_vma_flags(struct vm_area_struct
*vma
, unsigned long gup_flags
)
692 vm_flags_t vm_flags
= vma
->vm_flags
;
693 int write
= (gup_flags
& FOLL_WRITE
);
694 int foreign
= (gup_flags
& FOLL_REMOTE
);
696 if (vm_flags
& (VM_IO
| VM_PFNMAP
))
699 if (gup_flags
& FOLL_ANON
&& !vma_is_anonymous(vma
))
703 if (!(vm_flags
& VM_WRITE
)) {
704 if (!(gup_flags
& FOLL_FORCE
))
707 * We used to let the write,force case do COW in a
708 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
709 * set a breakpoint in a read-only mapping of an
710 * executable, without corrupting the file (yet only
711 * when that file had been opened for writing!).
712 * Anon pages in shared mappings are surprising: now
715 if (!is_cow_mapping(vm_flags
))
718 } else if (!(vm_flags
& VM_READ
)) {
719 if (!(gup_flags
& FOLL_FORCE
))
722 * Is there actually any vma we can reach here which does not
723 * have VM_MAYREAD set?
725 if (!(vm_flags
& VM_MAYREAD
))
729 * gups are always data accesses, not instruction
730 * fetches, so execute=false here
732 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
738 * __get_user_pages() - pin user pages in memory
739 * @tsk: task_struct of target task
740 * @mm: mm_struct of target mm
741 * @start: starting user address
742 * @nr_pages: number of pages from start to pin
743 * @gup_flags: flags modifying pin behaviour
744 * @pages: array that receives pointers to the pages pinned.
745 * Should be at least nr_pages long. Or NULL, if caller
746 * only intends to ensure the pages are faulted in.
747 * @vmas: array of pointers to vmas corresponding to each page.
748 * Or NULL if the caller does not require them.
749 * @nonblocking: whether waiting for disk IO or mmap_sem contention
751 * Returns number of pages pinned. This may be fewer than the number
752 * requested. If nr_pages is 0 or negative, returns 0. If no pages
753 * were pinned, returns -errno. Each page returned must be released
754 * with a put_page() call when it is finished with. vmas will only
755 * remain valid while mmap_sem is held.
757 * Must be called with mmap_sem held. It may be released. See below.
759 * __get_user_pages walks a process's page tables and takes a reference to
760 * each struct page that each user address corresponds to at a given
761 * instant. That is, it takes the page that would be accessed if a user
762 * thread accesses the given user virtual address at that instant.
764 * This does not guarantee that the page exists in the user mappings when
765 * __get_user_pages returns, and there may even be a completely different
766 * page there in some cases (eg. if mmapped pagecache has been invalidated
767 * and subsequently re faulted). However it does guarantee that the page
768 * won't be freed completely. And mostly callers simply care that the page
769 * contains data that was valid *at some point in time*. Typically, an IO
770 * or similar operation cannot guarantee anything stronger anyway because
771 * locks can't be held over the syscall boundary.
773 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
774 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
775 * appropriate) must be called after the page is finished with, and
776 * before put_page is called.
778 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
779 * or mmap_sem contention, and if waiting is needed to pin all pages,
780 * *@nonblocking will be set to 0. Further, if @gup_flags does not
781 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
784 * A caller using such a combination of @nonblocking and @gup_flags
785 * must therefore hold the mmap_sem for reading only, and recognize
786 * when it's been released. Otherwise, it must be held for either
787 * reading or writing and will not be released.
789 * In most cases, get_user_pages or get_user_pages_fast should be used
790 * instead of __get_user_pages. __get_user_pages should be used only if
791 * you need some special @gup_flags.
793 static long __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
794 unsigned long start
, unsigned long nr_pages
,
795 unsigned int gup_flags
, struct page
**pages
,
796 struct vm_area_struct
**vmas
, int *nonblocking
)
799 struct vm_area_struct
*vma
= NULL
;
800 struct follow_page_context ctx
= { NULL
};
805 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
808 * If FOLL_FORCE is set then do not force a full fault as the hinting
809 * fault information is unrelated to the reference behaviour of a task
810 * using the address space
812 if (!(gup_flags
& FOLL_FORCE
))
813 gup_flags
|= FOLL_NUMA
;
817 unsigned int foll_flags
= gup_flags
;
818 unsigned int page_increm
;
820 /* first iteration or cross vma bound */
821 if (!vma
|| start
>= vma
->vm_end
) {
822 vma
= find_extend_vma(mm
, start
);
823 if (!vma
&& in_gate_area(mm
, start
)) {
824 ret
= get_gate_page(mm
, start
& PAGE_MASK
,
826 pages
? &pages
[i
] : NULL
);
833 if (!vma
|| check_vma_flags(vma
, gup_flags
)) {
837 if (is_vm_hugetlb_page(vma
)) {
838 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
839 &start
, &nr_pages
, i
,
840 gup_flags
, nonblocking
);
846 * If we have a pending SIGKILL, don't keep faulting pages and
847 * potentially allocating memory.
849 if (fatal_signal_pending(current
)) {
855 page
= follow_page_mask(vma
, start
, foll_flags
, &ctx
);
857 ret
= faultin_page(tsk
, vma
, start
, &foll_flags
,
873 } else if (PTR_ERR(page
) == -EEXIST
) {
875 * Proper page table entry exists, but no corresponding
879 } else if (IS_ERR(page
)) {
885 flush_anon_page(vma
, page
, start
);
886 flush_dcache_page(page
);
894 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & ctx
.page_mask
);
895 if (page_increm
> nr_pages
)
896 page_increm
= nr_pages
;
898 start
+= page_increm
* PAGE_SIZE
;
899 nr_pages
-= page_increm
;
903 put_dev_pagemap(ctx
.pgmap
);
907 static bool vma_permits_fault(struct vm_area_struct
*vma
,
908 unsigned int fault_flags
)
910 bool write
= !!(fault_flags
& FAULT_FLAG_WRITE
);
911 bool foreign
= !!(fault_flags
& FAULT_FLAG_REMOTE
);
912 vm_flags_t vm_flags
= write
? VM_WRITE
: VM_READ
;
914 if (!(vm_flags
& vma
->vm_flags
))
918 * The architecture might have a hardware protection
919 * mechanism other than read/write that can deny access.
921 * gup always represents data access, not instruction
922 * fetches, so execute=false here:
924 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
931 * fixup_user_fault() - manually resolve a user page fault
932 * @tsk: the task_struct to use for page fault accounting, or
933 * NULL if faults are not to be recorded.
934 * @mm: mm_struct of target mm
935 * @address: user address
936 * @fault_flags:flags to pass down to handle_mm_fault()
937 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
938 * does not allow retry
940 * This is meant to be called in the specific scenario where for locking reasons
941 * we try to access user memory in atomic context (within a pagefault_disable()
942 * section), this returns -EFAULT, and we want to resolve the user fault before
945 * Typically this is meant to be used by the futex code.
947 * The main difference with get_user_pages() is that this function will
948 * unconditionally call handle_mm_fault() which will in turn perform all the
949 * necessary SW fixup of the dirty and young bits in the PTE, while
950 * get_user_pages() only guarantees to update these in the struct page.
952 * This is important for some architectures where those bits also gate the
953 * access permission to the page because they are maintained in software. On
954 * such architectures, gup() will not be enough to make a subsequent access
957 * This function will not return with an unlocked mmap_sem. So it has not the
958 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
960 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
961 unsigned long address
, unsigned int fault_flags
,
964 struct vm_area_struct
*vma
;
965 vm_fault_t ret
, major
= 0;
968 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
971 vma
= find_extend_vma(mm
, address
);
972 if (!vma
|| address
< vma
->vm_start
)
975 if (!vma_permits_fault(vma
, fault_flags
))
978 ret
= handle_mm_fault(vma
, address
, fault_flags
);
979 major
|= ret
& VM_FAULT_MAJOR
;
980 if (ret
& VM_FAULT_ERROR
) {
981 int err
= vm_fault_to_errno(ret
, 0);
988 if (ret
& VM_FAULT_RETRY
) {
989 down_read(&mm
->mmap_sem
);
990 if (!(fault_flags
& FAULT_FLAG_TRIED
)) {
992 fault_flags
&= ~FAULT_FLAG_ALLOW_RETRY
;
993 fault_flags
|= FAULT_FLAG_TRIED
;
1006 EXPORT_SYMBOL_GPL(fixup_user_fault
);
1008 static __always_inline
long __get_user_pages_locked(struct task_struct
*tsk
,
1009 struct mm_struct
*mm
,
1010 unsigned long start
,
1011 unsigned long nr_pages
,
1012 struct page
**pages
,
1013 struct vm_area_struct
**vmas
,
1017 long ret
, pages_done
;
1021 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1023 /* check caller initialized locked */
1024 BUG_ON(*locked
!= 1);
1031 lock_dropped
= false;
1033 ret
= __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
,
1036 /* VM_FAULT_RETRY couldn't trigger, bypass */
1039 /* VM_FAULT_RETRY cannot return errors */
1042 BUG_ON(ret
>= nr_pages
);
1046 /* If it's a prefault don't insist harder */
1057 * VM_FAULT_RETRY didn't trigger or it was a
1064 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
1066 start
+= ret
<< PAGE_SHIFT
;
1069 * Repeat on the address that fired VM_FAULT_RETRY
1070 * without FAULT_FLAG_ALLOW_RETRY but with
1074 lock_dropped
= true;
1075 down_read(&mm
->mmap_sem
);
1076 ret
= __get_user_pages(tsk
, mm
, start
, 1, flags
| FOLL_TRIED
,
1091 if (lock_dropped
&& *locked
) {
1093 * We must let the caller know we temporarily dropped the lock
1094 * and so the critical section protected by it was lost.
1096 up_read(&mm
->mmap_sem
);
1103 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1104 * paths better by using either get_user_pages_locked() or
1105 * get_user_pages_unlocked().
1107 * get_user_pages_locked() is suitable to replace the form:
1109 * down_read(&mm->mmap_sem);
1111 * get_user_pages(tsk, mm, ..., pages, NULL);
1112 * up_read(&mm->mmap_sem);
1117 * down_read(&mm->mmap_sem);
1119 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
1121 * up_read(&mm->mmap_sem);
1123 long get_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
1124 unsigned int gup_flags
, struct page
**pages
,
1128 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1129 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1130 * vmas. As there are no users of this flag in this call we simply
1131 * disallow this option for now.
1133 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1136 return __get_user_pages_locked(current
, current
->mm
, start
, nr_pages
,
1137 pages
, NULL
, locked
,
1138 gup_flags
| FOLL_TOUCH
);
1140 EXPORT_SYMBOL(get_user_pages_locked
);
1143 * get_user_pages_unlocked() is suitable to replace the form:
1145 * down_read(&mm->mmap_sem);
1146 * get_user_pages(tsk, mm, ..., pages, NULL);
1147 * up_read(&mm->mmap_sem);
1151 * get_user_pages_unlocked(tsk, mm, ..., pages);
1153 * It is functionally equivalent to get_user_pages_fast so
1154 * get_user_pages_fast should be used instead if specific gup_flags
1155 * (e.g. FOLL_FORCE) are not required.
1157 long get_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
1158 struct page
**pages
, unsigned int gup_flags
)
1160 struct mm_struct
*mm
= current
->mm
;
1165 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1166 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1167 * vmas. As there are no users of this flag in this call we simply
1168 * disallow this option for now.
1170 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1173 down_read(&mm
->mmap_sem
);
1174 ret
= __get_user_pages_locked(current
, mm
, start
, nr_pages
, pages
, NULL
,
1175 &locked
, gup_flags
| FOLL_TOUCH
);
1177 up_read(&mm
->mmap_sem
);
1180 EXPORT_SYMBOL(get_user_pages_unlocked
);
1183 * get_user_pages_remote() - pin user pages in memory
1184 * @tsk: the task_struct to use for page fault accounting, or
1185 * NULL if faults are not to be recorded.
1186 * @mm: mm_struct of target mm
1187 * @start: starting user address
1188 * @nr_pages: number of pages from start to pin
1189 * @gup_flags: flags modifying lookup behaviour
1190 * @pages: array that receives pointers to the pages pinned.
1191 * Should be at least nr_pages long. Or NULL, if caller
1192 * only intends to ensure the pages are faulted in.
1193 * @vmas: array of pointers to vmas corresponding to each page.
1194 * Or NULL if the caller does not require them.
1195 * @locked: pointer to lock flag indicating whether lock is held and
1196 * subsequently whether VM_FAULT_RETRY functionality can be
1197 * utilised. Lock must initially be held.
1199 * Returns number of pages pinned. This may be fewer than the number
1200 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1201 * were pinned, returns -errno. Each page returned must be released
1202 * with a put_page() call when it is finished with. vmas will only
1203 * remain valid while mmap_sem is held.
1205 * Must be called with mmap_sem held for read or write.
1207 * get_user_pages walks a process's page tables and takes a reference to
1208 * each struct page that each user address corresponds to at a given
1209 * instant. That is, it takes the page that would be accessed if a user
1210 * thread accesses the given user virtual address at that instant.
1212 * This does not guarantee that the page exists in the user mappings when
1213 * get_user_pages returns, and there may even be a completely different
1214 * page there in some cases (eg. if mmapped pagecache has been invalidated
1215 * and subsequently re faulted). However it does guarantee that the page
1216 * won't be freed completely. And mostly callers simply care that the page
1217 * contains data that was valid *at some point in time*. Typically, an IO
1218 * or similar operation cannot guarantee anything stronger anyway because
1219 * locks can't be held over the syscall boundary.
1221 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1222 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1223 * be called after the page is finished with, and before put_page is called.
1225 * get_user_pages is typically used for fewer-copy IO operations, to get a
1226 * handle on the memory by some means other than accesses via the user virtual
1227 * addresses. The pages may be submitted for DMA to devices or accessed via
1228 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1229 * use the correct cache flushing APIs.
1231 * See also get_user_pages_fast, for performance critical applications.
1233 * get_user_pages should be phased out in favor of
1234 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1235 * should use get_user_pages because it cannot pass
1236 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1238 long get_user_pages_remote(struct task_struct
*tsk
, struct mm_struct
*mm
,
1239 unsigned long start
, unsigned long nr_pages
,
1240 unsigned int gup_flags
, struct page
**pages
,
1241 struct vm_area_struct
**vmas
, int *locked
)
1244 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1245 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1246 * vmas. As there are no users of this flag in this call we simply
1247 * disallow this option for now.
1249 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1252 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
, vmas
,
1254 gup_flags
| FOLL_TOUCH
| FOLL_REMOTE
);
1256 EXPORT_SYMBOL(get_user_pages_remote
);
1258 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1259 static bool check_dax_vmas(struct vm_area_struct
**vmas
, long nr_pages
)
1262 struct vm_area_struct
*vma_prev
= NULL
;
1264 for (i
= 0; i
< nr_pages
; i
++) {
1265 struct vm_area_struct
*vma
= vmas
[i
];
1267 if (vma
== vma_prev
)
1272 if (vma_is_fsdax(vma
))
1279 static struct page
*new_non_cma_page(struct page
*page
, unsigned long private)
1282 * We want to make sure we allocate the new page from the same node
1283 * as the source page.
1285 int nid
= page_to_nid(page
);
1287 * Trying to allocate a page for migration. Ignore allocation
1288 * failure warnings. We don't force __GFP_THISNODE here because
1289 * this node here is the node where we have CMA reservation and
1290 * in some case these nodes will have really less non movable
1291 * allocation memory.
1293 gfp_t gfp_mask
= GFP_USER
| __GFP_NOWARN
;
1295 if (PageHighMem(page
))
1296 gfp_mask
|= __GFP_HIGHMEM
;
1298 #ifdef CONFIG_HUGETLB_PAGE
1299 if (PageHuge(page
)) {
1300 struct hstate
*h
= page_hstate(page
);
1302 * We don't want to dequeue from the pool because pool pages will
1303 * mostly be from the CMA region.
1305 return alloc_migrate_huge_page(h
, gfp_mask
, nid
, NULL
);
1308 if (PageTransHuge(page
)) {
1311 * ignore allocation failure warnings
1313 gfp_t thp_gfpmask
= GFP_TRANSHUGE
| __GFP_NOWARN
;
1316 * Remove the movable mask so that we don't allocate from
1319 thp_gfpmask
&= ~__GFP_MOVABLE
;
1320 thp
= __alloc_pages_node(nid
, thp_gfpmask
, HPAGE_PMD_ORDER
);
1323 prep_transhuge_page(thp
);
1327 return __alloc_pages_node(nid
, gfp_mask
, 0);
1330 static long check_and_migrate_cma_pages(struct task_struct
*tsk
,
1331 struct mm_struct
*mm
,
1332 unsigned long start
,
1333 unsigned long nr_pages
,
1334 struct page
**pages
,
1335 struct vm_area_struct
**vmas
,
1336 unsigned int gup_flags
)
1339 bool drain_allow
= true;
1340 bool migrate_allow
= true;
1341 LIST_HEAD(cma_page_list
);
1344 for (i
= 0; i
< nr_pages
; i
++) {
1346 * If we get a page from the CMA zone, since we are going to
1347 * be pinning these entries, we might as well move them out
1348 * of the CMA zone if possible.
1350 if (is_migrate_cma_page(pages
[i
])) {
1352 struct page
*head
= compound_head(pages
[i
]);
1354 if (PageHuge(head
)) {
1355 isolate_huge_page(head
, &cma_page_list
);
1357 if (!PageLRU(head
) && drain_allow
) {
1358 lru_add_drain_all();
1359 drain_allow
= false;
1362 if (!isolate_lru_page(head
)) {
1363 list_add_tail(&head
->lru
, &cma_page_list
);
1364 mod_node_page_state(page_pgdat(head
),
1366 page_is_file_cache(head
),
1367 hpage_nr_pages(head
));
1373 if (!list_empty(&cma_page_list
)) {
1375 * drop the above get_user_pages reference.
1377 for (i
= 0; i
< nr_pages
; i
++)
1380 if (migrate_pages(&cma_page_list
, new_non_cma_page
,
1381 NULL
, 0, MIGRATE_SYNC
, MR_CONTIG_RANGE
)) {
1383 * some of the pages failed migration. Do get_user_pages
1384 * without migration.
1386 migrate_allow
= false;
1388 if (!list_empty(&cma_page_list
))
1389 putback_movable_pages(&cma_page_list
);
1392 * We did migrate all the pages, Try to get the page references
1393 * again migrating any new CMA pages which we failed to isolate
1396 nr_pages
= __get_user_pages_locked(tsk
, mm
, start
, nr_pages
,
1400 if ((nr_pages
> 0) && migrate_allow
) {
1409 static long check_and_migrate_cma_pages(struct task_struct
*tsk
,
1410 struct mm_struct
*mm
,
1411 unsigned long start
,
1412 unsigned long nr_pages
,
1413 struct page
**pages
,
1414 struct vm_area_struct
**vmas
,
1415 unsigned int gup_flags
)
1422 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1423 * allows us to process the FOLL_LONGTERM flag.
1425 static long __gup_longterm_locked(struct task_struct
*tsk
,
1426 struct mm_struct
*mm
,
1427 unsigned long start
,
1428 unsigned long nr_pages
,
1429 struct page
**pages
,
1430 struct vm_area_struct
**vmas
,
1431 unsigned int gup_flags
)
1433 struct vm_area_struct
**vmas_tmp
= vmas
;
1434 unsigned long flags
= 0;
1437 if (gup_flags
& FOLL_LONGTERM
) {
1442 vmas_tmp
= kcalloc(nr_pages
,
1443 sizeof(struct vm_area_struct
*),
1448 flags
= memalloc_nocma_save();
1451 rc
= __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
,
1452 vmas_tmp
, NULL
, gup_flags
);
1454 if (gup_flags
& FOLL_LONGTERM
) {
1455 memalloc_nocma_restore(flags
);
1459 if (check_dax_vmas(vmas_tmp
, rc
)) {
1460 for (i
= 0; i
< rc
; i
++)
1466 rc
= check_and_migrate_cma_pages(tsk
, mm
, start
, rc
, pages
,
1467 vmas_tmp
, gup_flags
);
1471 if (vmas_tmp
!= vmas
)
1475 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1476 static __always_inline
long __gup_longterm_locked(struct task_struct
*tsk
,
1477 struct mm_struct
*mm
,
1478 unsigned long start
,
1479 unsigned long nr_pages
,
1480 struct page
**pages
,
1481 struct vm_area_struct
**vmas
,
1484 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
, vmas
,
1487 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1490 * This is the same as get_user_pages_remote(), just with a
1491 * less-flexible calling convention where we assume that the task
1492 * and mm being operated on are the current task's and don't allow
1493 * passing of a locked parameter. We also obviously don't pass
1494 * FOLL_REMOTE in here.
1496 long get_user_pages(unsigned long start
, unsigned long nr_pages
,
1497 unsigned int gup_flags
, struct page
**pages
,
1498 struct vm_area_struct
**vmas
)
1500 return __gup_longterm_locked(current
, current
->mm
, start
, nr_pages
,
1501 pages
, vmas
, gup_flags
| FOLL_TOUCH
);
1503 EXPORT_SYMBOL(get_user_pages
);
1506 * populate_vma_page_range() - populate a range of pages in the vma.
1508 * @start: start address
1512 * This takes care of mlocking the pages too if VM_LOCKED is set.
1514 * return 0 on success, negative error code on error.
1516 * vma->vm_mm->mmap_sem must be held.
1518 * If @nonblocking is NULL, it may be held for read or write and will
1521 * If @nonblocking is non-NULL, it must held for read only and may be
1522 * released. If it's released, *@nonblocking will be set to 0.
1524 long populate_vma_page_range(struct vm_area_struct
*vma
,
1525 unsigned long start
, unsigned long end
, int *nonblocking
)
1527 struct mm_struct
*mm
= vma
->vm_mm
;
1528 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1531 VM_BUG_ON(start
& ~PAGE_MASK
);
1532 VM_BUG_ON(end
& ~PAGE_MASK
);
1533 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1534 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1535 VM_BUG_ON_MM(!rwsem_is_locked(&mm
->mmap_sem
), mm
);
1537 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
1538 if (vma
->vm_flags
& VM_LOCKONFAULT
)
1539 gup_flags
&= ~FOLL_POPULATE
;
1541 * We want to touch writable mappings with a write fault in order
1542 * to break COW, except for shared mappings because these don't COW
1543 * and we would not want to dirty them for nothing.
1545 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
1546 gup_flags
|= FOLL_WRITE
;
1549 * We want mlock to succeed for regions that have any permissions
1550 * other than PROT_NONE.
1552 if (vma
->vm_flags
& (VM_READ
| VM_WRITE
| VM_EXEC
))
1553 gup_flags
|= FOLL_FORCE
;
1556 * We made sure addr is within a VMA, so the following will
1557 * not result in a stack expansion that recurses back here.
1559 return __get_user_pages(current
, mm
, start
, nr_pages
, gup_flags
,
1560 NULL
, NULL
, nonblocking
);
1564 * __mm_populate - populate and/or mlock pages within a range of address space.
1566 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1567 * flags. VMAs must be already marked with the desired vm_flags, and
1568 * mmap_sem must not be held.
1570 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
1572 struct mm_struct
*mm
= current
->mm
;
1573 unsigned long end
, nstart
, nend
;
1574 struct vm_area_struct
*vma
= NULL
;
1580 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
1582 * We want to fault in pages for [nstart; end) address range.
1583 * Find first corresponding VMA.
1587 down_read(&mm
->mmap_sem
);
1588 vma
= find_vma(mm
, nstart
);
1589 } else if (nstart
>= vma
->vm_end
)
1591 if (!vma
|| vma
->vm_start
>= end
)
1594 * Set [nstart; nend) to intersection of desired address
1595 * range with the first VMA. Also, skip undesirable VMA types.
1597 nend
= min(end
, vma
->vm_end
);
1598 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1600 if (nstart
< vma
->vm_start
)
1601 nstart
= vma
->vm_start
;
1603 * Now fault in a range of pages. populate_vma_page_range()
1604 * double checks the vma flags, so that it won't mlock pages
1605 * if the vma was already munlocked.
1607 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1609 if (ignore_errors
) {
1611 continue; /* continue at next VMA */
1615 nend
= nstart
+ ret
* PAGE_SIZE
;
1619 up_read(&mm
->mmap_sem
);
1620 return ret
; /* 0 or negative error code */
1624 * get_dump_page() - pin user page in memory while writing it to core dump
1625 * @addr: user address
1627 * Returns struct page pointer of user page pinned for dump,
1628 * to be freed afterwards by put_page().
1630 * Returns NULL on any kind of failure - a hole must then be inserted into
1631 * the corefile, to preserve alignment with its headers; and also returns
1632 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1633 * allowing a hole to be left in the corefile to save diskspace.
1635 * Called without mmap_sem, but after all other threads have been killed.
1637 #ifdef CONFIG_ELF_CORE
1638 struct page
*get_dump_page(unsigned long addr
)
1640 struct vm_area_struct
*vma
;
1643 if (__get_user_pages(current
, current
->mm
, addr
, 1,
1644 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
1647 flush_cache_page(vma
, addr
, page_to_pfn(page
));
1650 #endif /* CONFIG_ELF_CORE */
1655 * get_user_pages_fast attempts to pin user pages by walking the page
1656 * tables directly and avoids taking locks. Thus the walker needs to be
1657 * protected from page table pages being freed from under it, and should
1658 * block any THP splits.
1660 * One way to achieve this is to have the walker disable interrupts, and
1661 * rely on IPIs from the TLB flushing code blocking before the page table
1662 * pages are freed. This is unsuitable for architectures that do not need
1663 * to broadcast an IPI when invalidating TLBs.
1665 * Another way to achieve this is to batch up page table containing pages
1666 * belonging to more than one mm_user, then rcu_sched a callback to free those
1667 * pages. Disabling interrupts will allow the fast_gup walker to both block
1668 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1669 * (which is a relatively rare event). The code below adopts this strategy.
1671 * Before activating this code, please be aware that the following assumptions
1672 * are currently made:
1674 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1675 * free pages containing page tables or TLB flushing requires IPI broadcast.
1677 * *) ptes can be read atomically by the architecture.
1679 * *) access_ok is sufficient to validate userspace address ranges.
1681 * The last two assumptions can be relaxed by the addition of helper functions.
1683 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1685 #ifdef CONFIG_HAVE_GENERIC_GUP
1689 * We assume that the PTE can be read atomically. If this is not the case for
1690 * your architecture, please provide the helper.
1692 static inline pte_t
gup_get_pte(pte_t
*ptep
)
1694 return READ_ONCE(*ptep
);
1698 static void undo_dev_pagemap(int *nr
, int nr_start
, struct page
**pages
)
1700 while ((*nr
) - nr_start
) {
1701 struct page
*page
= pages
[--(*nr
)];
1703 ClearPageReferenced(page
);
1709 * Return the compund head page with ref appropriately incremented,
1710 * or NULL if that failed.
1712 static inline struct page
*try_get_compound_head(struct page
*page
, int refs
)
1714 struct page
*head
= compound_head(page
);
1715 if (WARN_ON_ONCE(page_ref_count(head
) < 0))
1717 if (unlikely(!page_cache_add_speculative(head
, refs
)))
1722 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
1723 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1724 unsigned int flags
, struct page
**pages
, int *nr
)
1726 struct dev_pagemap
*pgmap
= NULL
;
1727 int nr_start
= *nr
, ret
= 0;
1730 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
1732 pte_t pte
= gup_get_pte(ptep
);
1733 struct page
*head
, *page
;
1736 * Similar to the PMD case below, NUMA hinting must take slow
1737 * path using the pte_protnone check.
1739 if (pte_protnone(pte
))
1742 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
1745 if (pte_devmap(pte
)) {
1746 if (unlikely(flags
& FOLL_LONGTERM
))
1749 pgmap
= get_dev_pagemap(pte_pfn(pte
), pgmap
);
1750 if (unlikely(!pgmap
)) {
1751 undo_dev_pagemap(nr
, nr_start
, pages
);
1754 } else if (pte_special(pte
))
1757 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
1758 page
= pte_page(pte
);
1760 head
= try_get_compound_head(page
, 1);
1764 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
1769 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1771 SetPageReferenced(page
);
1775 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
1781 put_dev_pagemap(pgmap
);
1788 * If we can't determine whether or not a pte is special, then fail immediately
1789 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1792 * For a futex to be placed on a THP tail page, get_futex_key requires a
1793 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1794 * useful to have gup_huge_pmd even if we can't operate on ptes.
1796 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1797 unsigned int flags
, struct page
**pages
, int *nr
)
1801 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
1803 #if defined(__HAVE_ARCH_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1804 static int __gup_device_huge(unsigned long pfn
, unsigned long addr
,
1805 unsigned long end
, struct page
**pages
, int *nr
)
1808 struct dev_pagemap
*pgmap
= NULL
;
1811 struct page
*page
= pfn_to_page(pfn
);
1813 pgmap
= get_dev_pagemap(pfn
, pgmap
);
1814 if (unlikely(!pgmap
)) {
1815 undo_dev_pagemap(nr
, nr_start
, pages
);
1818 SetPageReferenced(page
);
1823 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1826 put_dev_pagemap(pgmap
);
1830 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1831 unsigned long end
, struct page
**pages
, int *nr
)
1833 unsigned long fault_pfn
;
1836 fault_pfn
= pmd_pfn(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
1837 if (!__gup_device_huge(fault_pfn
, addr
, end
, pages
, nr
))
1840 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
1841 undo_dev_pagemap(nr
, nr_start
, pages
);
1847 static int __gup_device_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
1848 unsigned long end
, struct page
**pages
, int *nr
)
1850 unsigned long fault_pfn
;
1853 fault_pfn
= pud_pfn(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
1854 if (!__gup_device_huge(fault_pfn
, addr
, end
, pages
, nr
))
1857 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
1858 undo_dev_pagemap(nr
, nr_start
, pages
);
1864 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1865 unsigned long end
, struct page
**pages
, int *nr
)
1871 static int __gup_device_huge_pud(pud_t pud
, pud_t
*pudp
, unsigned long addr
,
1872 unsigned long end
, struct page
**pages
, int *nr
)
1879 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1880 unsigned long end
, unsigned int flags
, struct page
**pages
, int *nr
)
1882 struct page
*head
, *page
;
1885 if (!pmd_access_permitted(orig
, flags
& FOLL_WRITE
))
1888 if (pmd_devmap(orig
)) {
1889 if (unlikely(flags
& FOLL_LONGTERM
))
1891 return __gup_device_huge_pmd(orig
, pmdp
, addr
, end
, pages
, nr
);
1895 page
= pmd_page(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
1901 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1903 head
= try_get_compound_head(pmd_page(orig
), refs
);
1909 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
1916 SetPageReferenced(head
);
1920 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
1921 unsigned long end
, unsigned int flags
, struct page
**pages
, int *nr
)
1923 struct page
*head
, *page
;
1926 if (!pud_access_permitted(orig
, flags
& FOLL_WRITE
))
1929 if (pud_devmap(orig
)) {
1930 if (unlikely(flags
& FOLL_LONGTERM
))
1932 return __gup_device_huge_pud(orig
, pudp
, addr
, end
, pages
, nr
);
1936 page
= pud_page(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
1942 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1944 head
= try_get_compound_head(pud_page(orig
), refs
);
1950 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
1957 SetPageReferenced(head
);
1961 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
1962 unsigned long end
, unsigned int flags
,
1963 struct page
**pages
, int *nr
)
1966 struct page
*head
, *page
;
1968 if (!pgd_access_permitted(orig
, flags
& FOLL_WRITE
))
1971 BUILD_BUG_ON(pgd_devmap(orig
));
1973 page
= pgd_page(orig
) + ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
1979 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1981 head
= try_get_compound_head(pgd_page(orig
), refs
);
1987 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
1994 SetPageReferenced(head
);
1998 static int gup_pmd_range(pud_t pud
, unsigned long addr
, unsigned long end
,
1999 unsigned int flags
, struct page
**pages
, int *nr
)
2004 pmdp
= pmd_offset(&pud
, addr
);
2006 pmd_t pmd
= READ_ONCE(*pmdp
);
2008 next
= pmd_addr_end(addr
, end
);
2009 if (!pmd_present(pmd
))
2012 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
) ||
2015 * NUMA hinting faults need to be handled in the GUP
2016 * slowpath for accounting purposes and so that they
2017 * can be serialised against THP migration.
2019 if (pmd_protnone(pmd
))
2022 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, flags
,
2026 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
2028 * architecture have different format for hugetlbfs
2029 * pmd format and THP pmd format
2031 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
2032 PMD_SHIFT
, next
, flags
, pages
, nr
))
2034 } else if (!gup_pte_range(pmd
, addr
, next
, flags
, pages
, nr
))
2036 } while (pmdp
++, addr
= next
, addr
!= end
);
2041 static int gup_pud_range(p4d_t p4d
, unsigned long addr
, unsigned long end
,
2042 unsigned int flags
, struct page
**pages
, int *nr
)
2047 pudp
= pud_offset(&p4d
, addr
);
2049 pud_t pud
= READ_ONCE(*pudp
);
2051 next
= pud_addr_end(addr
, end
);
2054 if (unlikely(pud_huge(pud
))) {
2055 if (!gup_huge_pud(pud
, pudp
, addr
, next
, flags
,
2058 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
2059 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
2060 PUD_SHIFT
, next
, flags
, pages
, nr
))
2062 } else if (!gup_pmd_range(pud
, addr
, next
, flags
, pages
, nr
))
2064 } while (pudp
++, addr
= next
, addr
!= end
);
2069 static int gup_p4d_range(pgd_t pgd
, unsigned long addr
, unsigned long end
,
2070 unsigned int flags
, struct page
**pages
, int *nr
)
2075 p4dp
= p4d_offset(&pgd
, addr
);
2077 p4d_t p4d
= READ_ONCE(*p4dp
);
2079 next
= p4d_addr_end(addr
, end
);
2082 BUILD_BUG_ON(p4d_huge(p4d
));
2083 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d
))))) {
2084 if (!gup_huge_pd(__hugepd(p4d_val(p4d
)), addr
,
2085 P4D_SHIFT
, next
, flags
, pages
, nr
))
2087 } else if (!gup_pud_range(p4d
, addr
, next
, flags
, pages
, nr
))
2089 } while (p4dp
++, addr
= next
, addr
!= end
);
2094 static void gup_pgd_range(unsigned long addr
, unsigned long end
,
2095 unsigned int flags
, struct page
**pages
, int *nr
)
2100 pgdp
= pgd_offset(current
->mm
, addr
);
2102 pgd_t pgd
= READ_ONCE(*pgdp
);
2104 next
= pgd_addr_end(addr
, end
);
2107 if (unlikely(pgd_huge(pgd
))) {
2108 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, flags
,
2111 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
2112 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
2113 PGDIR_SHIFT
, next
, flags
, pages
, nr
))
2115 } else if (!gup_p4d_range(pgd
, addr
, next
, flags
, pages
, nr
))
2117 } while (pgdp
++, addr
= next
, addr
!= end
);
2120 #ifndef gup_fast_permitted
2122 * Check if it's allowed to use __get_user_pages_fast() for the range, or
2123 * we need to fall back to the slow version:
2125 bool gup_fast_permitted(unsigned long start
, int nr_pages
)
2127 unsigned long len
, end
;
2129 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
2131 return end
>= start
;
2136 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2138 * Note a difference with get_user_pages_fast: this always returns the
2139 * number of pages pinned, 0 if no pages were pinned.
2141 int __get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
2142 struct page
**pages
)
2144 unsigned long len
, end
;
2145 unsigned long flags
;
2149 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
2152 if (unlikely(!access_ok((void __user
*)start
, len
)))
2156 * Disable interrupts. We use the nested form as we can already have
2157 * interrupts disabled by get_futex_key.
2159 * With interrupts disabled, we block page table pages from being
2160 * freed from under us. See struct mmu_table_batch comments in
2161 * include/asm-generic/tlb.h for more details.
2163 * We do not adopt an rcu_read_lock(.) here as we also want to
2164 * block IPIs that come from THPs splitting.
2167 if (gup_fast_permitted(start
, nr_pages
)) {
2168 local_irq_save(flags
);
2169 gup_pgd_range(start
, end
, write
? FOLL_WRITE
: 0, pages
, &nr
);
2170 local_irq_restore(flags
);
2176 static int __gup_longterm_unlocked(unsigned long start
, int nr_pages
,
2177 unsigned int gup_flags
, struct page
**pages
)
2182 * FIXME: FOLL_LONGTERM does not work with
2183 * get_user_pages_unlocked() (see comments in that function)
2185 if (gup_flags
& FOLL_LONGTERM
) {
2186 down_read(¤t
->mm
->mmap_sem
);
2187 ret
= __gup_longterm_locked(current
, current
->mm
,
2189 pages
, NULL
, gup_flags
);
2190 up_read(¤t
->mm
->mmap_sem
);
2192 ret
= get_user_pages_unlocked(start
, nr_pages
,
2200 * get_user_pages_fast() - pin user pages in memory
2201 * @start: starting user address
2202 * @nr_pages: number of pages from start to pin
2203 * @gup_flags: flags modifying pin behaviour
2204 * @pages: array that receives pointers to the pages pinned.
2205 * Should be at least nr_pages long.
2207 * Attempt to pin user pages in memory without taking mm->mmap_sem.
2208 * If not successful, it will fall back to taking the lock and
2209 * calling get_user_pages().
2211 * Returns number of pages pinned. This may be fewer than the number
2212 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2213 * were pinned, returns -errno.
2215 int get_user_pages_fast(unsigned long start
, int nr_pages
,
2216 unsigned int gup_flags
, struct page
**pages
)
2218 unsigned long addr
, len
, end
;
2219 int nr
= 0, ret
= 0;
2223 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
2229 if (unlikely(!access_ok((void __user
*)start
, len
)))
2232 if (gup_fast_permitted(start
, nr_pages
)) {
2233 local_irq_disable();
2234 gup_pgd_range(addr
, end
, gup_flags
, pages
, &nr
);
2239 if (nr
< nr_pages
) {
2240 /* Try to get the remaining pages with get_user_pages */
2241 start
+= nr
<< PAGE_SHIFT
;
2244 ret
= __gup_longterm_unlocked(start
, nr_pages
- nr
,
2247 /* Have to be a bit careful with return values */
2259 #endif /* CONFIG_HAVE_GENERIC_GUP */