]>
git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blob - mm/gup.c
1 #include <linux/kernel.h>
2 #include <linux/errno.h>
4 #include <linux/spinlock.h>
7 #include <linux/memremap.h>
8 #include <linux/pagemap.h>
9 #include <linux/rmap.h>
10 #include <linux/swap.h>
11 #include <linux/swapops.h>
13 #include <linux/sched/signal.h>
14 #include <linux/rwsem.h>
15 #include <linux/hugetlb.h>
16 #include <linux/migrate.h>
17 #include <linux/mm_inline.h>
18 #include <linux/sched/mm.h>
20 #include <asm/mmu_context.h>
21 #include <asm/pgtable.h>
22 #include <asm/tlbflush.h>
26 struct follow_page_context
{
27 struct dev_pagemap
*pgmap
;
28 unsigned int page_mask
;
31 static struct page
*no_page_table(struct vm_area_struct
*vma
,
35 * When core dumping an enormous anonymous area that nobody
36 * has touched so far, we don't want to allocate unnecessary pages or
37 * page tables. Return error instead of NULL to skip handle_mm_fault,
38 * then get_dump_page() will return NULL to leave a hole in the dump.
39 * But we can only make this optimization where a hole would surely
40 * be zero-filled if handle_mm_fault() actually did handle it.
42 if ((flags
& FOLL_DUMP
) && (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
43 return ERR_PTR(-EFAULT
);
47 static int follow_pfn_pte(struct vm_area_struct
*vma
, unsigned long address
,
48 pte_t
*pte
, unsigned int flags
)
50 /* No page to get reference */
54 if (flags
& FOLL_TOUCH
) {
57 if (flags
& FOLL_WRITE
)
58 entry
= pte_mkdirty(entry
);
59 entry
= pte_mkyoung(entry
);
61 if (!pte_same(*pte
, entry
)) {
62 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
63 update_mmu_cache(vma
, address
, pte
);
67 /* Proper page table entry exists, but no corresponding struct page */
72 * FOLL_FORCE can write to even unwritable pte's, but only
73 * after we've gone through a COW cycle and they are dirty.
75 static inline bool can_follow_write_pte(pte_t pte
, unsigned int flags
)
77 return pte_write(pte
) ||
78 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pte_dirty(pte
));
81 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
82 unsigned long address
, pmd_t
*pmd
, unsigned int flags
,
83 struct dev_pagemap
**pgmap
)
85 struct mm_struct
*mm
= vma
->vm_mm
;
91 if (unlikely(pmd_bad(*pmd
)))
92 return no_page_table(vma
, flags
);
94 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
96 if (!pte_present(pte
)) {
99 * KSM's break_ksm() relies upon recognizing a ksm page
100 * even while it is being migrated, so for that case we
101 * need migration_entry_wait().
103 if (likely(!(flags
& FOLL_MIGRATION
)))
107 entry
= pte_to_swp_entry(pte
);
108 if (!is_migration_entry(entry
))
110 pte_unmap_unlock(ptep
, ptl
);
111 migration_entry_wait(mm
, pmd
, address
);
114 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
116 if ((flags
& FOLL_WRITE
) && !can_follow_write_pte(pte
, flags
)) {
117 pte_unmap_unlock(ptep
, ptl
);
121 page
= vm_normal_page(vma
, address
, pte
);
122 if (!page
&& pte_devmap(pte
) && (flags
& FOLL_GET
)) {
124 * Only return device mapping pages in the FOLL_GET case since
125 * they are only valid while holding the pgmap reference.
127 *pgmap
= get_dev_pagemap(pte_pfn(pte
), *pgmap
);
129 page
= pte_page(pte
);
132 } else if (unlikely(!page
)) {
133 if (flags
& FOLL_DUMP
) {
134 /* Avoid special (like zero) pages in core dumps */
135 page
= ERR_PTR(-EFAULT
);
139 if (is_zero_pfn(pte_pfn(pte
))) {
140 page
= pte_page(pte
);
144 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
150 if (flags
& FOLL_SPLIT
&& PageTransCompound(page
)) {
153 pte_unmap_unlock(ptep
, ptl
);
155 ret
= split_huge_page(page
);
163 if (flags
& FOLL_GET
) {
164 if (unlikely(!try_get_page(page
))) {
165 page
= ERR_PTR(-ENOMEM
);
169 if (flags
& FOLL_TOUCH
) {
170 if ((flags
& FOLL_WRITE
) &&
171 !pte_dirty(pte
) && !PageDirty(page
))
172 set_page_dirty(page
);
174 * pte_mkyoung() would be more correct here, but atomic care
175 * is needed to avoid losing the dirty bit: it is easier to use
176 * mark_page_accessed().
178 mark_page_accessed(page
);
180 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
181 /* Do not mlock pte-mapped THP */
182 if (PageTransCompound(page
))
186 * The preliminary mapping check is mainly to avoid the
187 * pointless overhead of lock_page on the ZERO_PAGE
188 * which might bounce very badly if there is contention.
190 * If the page is already locked, we don't need to
191 * handle it now - vmscan will handle it later if and
192 * when it attempts to reclaim the page.
194 if (page
->mapping
&& trylock_page(page
)) {
195 lru_add_drain(); /* push cached pages to LRU */
197 * Because we lock page here, and migration is
198 * blocked by the pte's page reference, and we
199 * know the page is still mapped, we don't even
200 * need to check for file-cache page truncation.
202 mlock_vma_page(page
);
207 pte_unmap_unlock(ptep
, ptl
);
210 pte_unmap_unlock(ptep
, ptl
);
213 return no_page_table(vma
, flags
);
216 static struct page
*follow_pmd_mask(struct vm_area_struct
*vma
,
217 unsigned long address
, pud_t
*pudp
,
219 struct follow_page_context
*ctx
)
224 struct mm_struct
*mm
= vma
->vm_mm
;
226 pmd
= pmd_offset(pudp
, address
);
228 * The READ_ONCE() will stabilize the pmdval in a register or
229 * on the stack so that it will stop changing under the code.
231 pmdval
= READ_ONCE(*pmd
);
232 if (pmd_none(pmdval
))
233 return no_page_table(vma
, flags
);
234 if (pmd_huge(pmdval
) && vma
->vm_flags
& VM_HUGETLB
) {
235 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
238 return no_page_table(vma
, flags
);
240 if (is_hugepd(__hugepd(pmd_val(pmdval
)))) {
241 page
= follow_huge_pd(vma
, address
,
242 __hugepd(pmd_val(pmdval
)), flags
,
246 return no_page_table(vma
, flags
);
249 if (!pmd_present(pmdval
)) {
250 if (likely(!(flags
& FOLL_MIGRATION
)))
251 return no_page_table(vma
, flags
);
252 VM_BUG_ON(thp_migration_supported() &&
253 !is_pmd_migration_entry(pmdval
));
254 if (is_pmd_migration_entry(pmdval
))
255 pmd_migration_entry_wait(mm
, pmd
);
256 pmdval
= READ_ONCE(*pmd
);
258 * MADV_DONTNEED may convert the pmd to null because
259 * mmap_sem is held in read mode
261 if (pmd_none(pmdval
))
262 return no_page_table(vma
, flags
);
265 if (pmd_devmap(pmdval
)) {
266 ptl
= pmd_lock(mm
, pmd
);
267 page
= follow_devmap_pmd(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
272 if (likely(!pmd_trans_huge(pmdval
)))
273 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
275 if ((flags
& FOLL_NUMA
) && pmd_protnone(pmdval
))
276 return no_page_table(vma
, flags
);
279 ptl
= pmd_lock(mm
, pmd
);
280 if (unlikely(pmd_none(*pmd
))) {
282 return no_page_table(vma
, flags
);
284 if (unlikely(!pmd_present(*pmd
))) {
286 if (likely(!(flags
& FOLL_MIGRATION
)))
287 return no_page_table(vma
, flags
);
288 pmd_migration_entry_wait(mm
, pmd
);
291 if (unlikely(!pmd_trans_huge(*pmd
))) {
293 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
295 if (flags
& FOLL_SPLIT
) {
297 page
= pmd_page(*pmd
);
298 if (is_huge_zero_page(page
)) {
301 split_huge_pmd(vma
, pmd
, address
);
302 if (pmd_trans_unstable(pmd
))
305 if (unlikely(!try_get_page(page
))) {
307 return ERR_PTR(-ENOMEM
);
311 ret
= split_huge_page(page
);
315 return no_page_table(vma
, flags
);
318 return ret
? ERR_PTR(ret
) :
319 follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
321 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
323 ctx
->page_mask
= HPAGE_PMD_NR
- 1;
327 static struct page
*follow_pud_mask(struct vm_area_struct
*vma
,
328 unsigned long address
, p4d_t
*p4dp
,
330 struct follow_page_context
*ctx
)
335 struct mm_struct
*mm
= vma
->vm_mm
;
337 pud
= pud_offset(p4dp
, address
);
339 return no_page_table(vma
, flags
);
340 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
341 page
= follow_huge_pud(mm
, address
, pud
, flags
);
344 return no_page_table(vma
, flags
);
346 if (is_hugepd(__hugepd(pud_val(*pud
)))) {
347 page
= follow_huge_pd(vma
, address
,
348 __hugepd(pud_val(*pud
)), flags
,
352 return no_page_table(vma
, flags
);
354 if (pud_devmap(*pud
)) {
355 ptl
= pud_lock(mm
, pud
);
356 page
= follow_devmap_pud(vma
, address
, pud
, flags
, &ctx
->pgmap
);
361 if (unlikely(pud_bad(*pud
)))
362 return no_page_table(vma
, flags
);
364 return follow_pmd_mask(vma
, address
, pud
, flags
, ctx
);
367 static struct page
*follow_p4d_mask(struct vm_area_struct
*vma
,
368 unsigned long address
, pgd_t
*pgdp
,
370 struct follow_page_context
*ctx
)
375 p4d
= p4d_offset(pgdp
, address
);
377 return no_page_table(vma
, flags
);
378 BUILD_BUG_ON(p4d_huge(*p4d
));
379 if (unlikely(p4d_bad(*p4d
)))
380 return no_page_table(vma
, flags
);
382 if (is_hugepd(__hugepd(p4d_val(*p4d
)))) {
383 page
= follow_huge_pd(vma
, address
,
384 __hugepd(p4d_val(*p4d
)), flags
,
388 return no_page_table(vma
, flags
);
390 return follow_pud_mask(vma
, address
, p4d
, flags
, ctx
);
394 * follow_page_mask - look up a page descriptor from a user-virtual address
395 * @vma: vm_area_struct mapping @address
396 * @address: virtual address to look up
397 * @flags: flags modifying lookup behaviour
398 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
399 * pointer to output page_mask
401 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
403 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
404 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
406 * On output, the @ctx->page_mask is set according to the size of the page.
408 * Return: the mapped (struct page *), %NULL if no mapping exists, or
409 * an error pointer if there is a mapping to something not represented
410 * by a page descriptor (see also vm_normal_page()).
412 struct page
*follow_page_mask(struct vm_area_struct
*vma
,
413 unsigned long address
, unsigned int flags
,
414 struct follow_page_context
*ctx
)
418 struct mm_struct
*mm
= vma
->vm_mm
;
422 /* make this handle hugepd */
423 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
425 BUG_ON(flags
& FOLL_GET
);
429 pgd
= pgd_offset(mm
, address
);
431 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
432 return no_page_table(vma
, flags
);
434 if (pgd_huge(*pgd
)) {
435 page
= follow_huge_pgd(mm
, address
, pgd
, flags
);
438 return no_page_table(vma
, flags
);
440 if (is_hugepd(__hugepd(pgd_val(*pgd
)))) {
441 page
= follow_huge_pd(vma
, address
,
442 __hugepd(pgd_val(*pgd
)), flags
,
446 return no_page_table(vma
, flags
);
449 return follow_p4d_mask(vma
, address
, pgd
, flags
, ctx
);
452 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
453 unsigned int foll_flags
)
455 struct follow_page_context ctx
= { NULL
};
458 page
= follow_page_mask(vma
, address
, foll_flags
, &ctx
);
460 put_dev_pagemap(ctx
.pgmap
);
464 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
465 unsigned int gup_flags
, struct vm_area_struct
**vma
,
475 /* user gate pages are read-only */
476 if (gup_flags
& FOLL_WRITE
)
478 if (address
> TASK_SIZE
)
479 pgd
= pgd_offset_k(address
);
481 pgd
= pgd_offset_gate(mm
, address
);
482 BUG_ON(pgd_none(*pgd
));
483 p4d
= p4d_offset(pgd
, address
);
484 BUG_ON(p4d_none(*p4d
));
485 pud
= pud_offset(p4d
, address
);
486 BUG_ON(pud_none(*pud
));
487 pmd
= pmd_offset(pud
, address
);
488 if (!pmd_present(*pmd
))
490 VM_BUG_ON(pmd_trans_huge(*pmd
));
491 pte
= pte_offset_map(pmd
, address
);
494 *vma
= get_gate_vma(mm
);
497 *page
= vm_normal_page(*vma
, address
, *pte
);
499 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
501 *page
= pte_page(*pte
);
504 * This should never happen (a device public page in the gate
507 if (is_device_public_page(*page
))
510 if (unlikely(!try_get_page(*page
))) {
522 * mmap_sem must be held on entry. If @nonblocking != NULL and
523 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
524 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
526 static int faultin_page(struct task_struct
*tsk
, struct vm_area_struct
*vma
,
527 unsigned long address
, unsigned int *flags
, int *nonblocking
)
529 unsigned int fault_flags
= 0;
532 /* mlock all present pages, but do not fault in new pages */
533 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
535 if (*flags
& FOLL_WRITE
)
536 fault_flags
|= FAULT_FLAG_WRITE
;
537 if (*flags
& FOLL_REMOTE
)
538 fault_flags
|= FAULT_FLAG_REMOTE
;
540 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
541 if (*flags
& FOLL_NOWAIT
)
542 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
543 if (*flags
& FOLL_TRIED
) {
544 VM_WARN_ON_ONCE(fault_flags
& FAULT_FLAG_ALLOW_RETRY
);
545 fault_flags
|= FAULT_FLAG_TRIED
;
548 ret
= handle_mm_fault(vma
, address
, fault_flags
);
549 if (ret
& VM_FAULT_ERROR
) {
550 int err
= vm_fault_to_errno(ret
, *flags
);
558 if (ret
& VM_FAULT_MAJOR
)
564 if (ret
& VM_FAULT_RETRY
) {
565 if (nonblocking
&& !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
571 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
572 * necessary, even if maybe_mkwrite decided not to set pte_write. We
573 * can thus safely do subsequent page lookups as if they were reads.
574 * But only do so when looping for pte_write is futile: in some cases
575 * userspace may also be wanting to write to the gotten user page,
576 * which a read fault here might prevent (a readonly page might get
577 * reCOWed by userspace write).
579 if ((ret
& VM_FAULT_WRITE
) && !(vma
->vm_flags
& VM_WRITE
))
584 static int check_vma_flags(struct vm_area_struct
*vma
, unsigned long gup_flags
)
586 vm_flags_t vm_flags
= vma
->vm_flags
;
587 int write
= (gup_flags
& FOLL_WRITE
);
588 int foreign
= (gup_flags
& FOLL_REMOTE
);
590 if (vm_flags
& (VM_IO
| VM_PFNMAP
))
593 if (gup_flags
& FOLL_ANON
&& !vma_is_anonymous(vma
))
597 if (!(vm_flags
& VM_WRITE
)) {
598 if (!(gup_flags
& FOLL_FORCE
))
601 * We used to let the write,force case do COW in a
602 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
603 * set a breakpoint in a read-only mapping of an
604 * executable, without corrupting the file (yet only
605 * when that file had been opened for writing!).
606 * Anon pages in shared mappings are surprising: now
609 if (!is_cow_mapping(vm_flags
))
612 } else if (!(vm_flags
& VM_READ
)) {
613 if (!(gup_flags
& FOLL_FORCE
))
616 * Is there actually any vma we can reach here which does not
617 * have VM_MAYREAD set?
619 if (!(vm_flags
& VM_MAYREAD
))
623 * gups are always data accesses, not instruction
624 * fetches, so execute=false here
626 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
632 * __get_user_pages() - pin user pages in memory
633 * @tsk: task_struct of target task
634 * @mm: mm_struct of target mm
635 * @start: starting user address
636 * @nr_pages: number of pages from start to pin
637 * @gup_flags: flags modifying pin behaviour
638 * @pages: array that receives pointers to the pages pinned.
639 * Should be at least nr_pages long. Or NULL, if caller
640 * only intends to ensure the pages are faulted in.
641 * @vmas: array of pointers to vmas corresponding to each page.
642 * Or NULL if the caller does not require them.
643 * @nonblocking: whether waiting for disk IO or mmap_sem contention
645 * Returns number of pages pinned. This may be fewer than the number
646 * requested. If nr_pages is 0 or negative, returns 0. If no pages
647 * were pinned, returns -errno. Each page returned must be released
648 * with a put_page() call when it is finished with. vmas will only
649 * remain valid while mmap_sem is held.
651 * Must be called with mmap_sem held. It may be released. See below.
653 * __get_user_pages walks a process's page tables and takes a reference to
654 * each struct page that each user address corresponds to at a given
655 * instant. That is, it takes the page that would be accessed if a user
656 * thread accesses the given user virtual address at that instant.
658 * This does not guarantee that the page exists in the user mappings when
659 * __get_user_pages returns, and there may even be a completely different
660 * page there in some cases (eg. if mmapped pagecache has been invalidated
661 * and subsequently re faulted). However it does guarantee that the page
662 * won't be freed completely. And mostly callers simply care that the page
663 * contains data that was valid *at some point in time*. Typically, an IO
664 * or similar operation cannot guarantee anything stronger anyway because
665 * locks can't be held over the syscall boundary.
667 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
668 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
669 * appropriate) must be called after the page is finished with, and
670 * before put_page is called.
672 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
673 * or mmap_sem contention, and if waiting is needed to pin all pages,
674 * *@nonblocking will be set to 0. Further, if @gup_flags does not
675 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
678 * A caller using such a combination of @nonblocking and @gup_flags
679 * must therefore hold the mmap_sem for reading only, and recognize
680 * when it's been released. Otherwise, it must be held for either
681 * reading or writing and will not be released.
683 * In most cases, get_user_pages or get_user_pages_fast should be used
684 * instead of __get_user_pages. __get_user_pages should be used only if
685 * you need some special @gup_flags.
687 static long __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
688 unsigned long start
, unsigned long nr_pages
,
689 unsigned int gup_flags
, struct page
**pages
,
690 struct vm_area_struct
**vmas
, int *nonblocking
)
693 struct vm_area_struct
*vma
= NULL
;
694 struct follow_page_context ctx
= { NULL
};
699 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
702 * If FOLL_FORCE is set then do not force a full fault as the hinting
703 * fault information is unrelated to the reference behaviour of a task
704 * using the address space
706 if (!(gup_flags
& FOLL_FORCE
))
707 gup_flags
|= FOLL_NUMA
;
711 unsigned int foll_flags
= gup_flags
;
712 unsigned int page_increm
;
714 /* first iteration or cross vma bound */
715 if (!vma
|| start
>= vma
->vm_end
) {
716 vma
= find_extend_vma(mm
, start
);
717 if (!vma
&& in_gate_area(mm
, start
)) {
718 ret
= get_gate_page(mm
, start
& PAGE_MASK
,
720 pages
? &pages
[i
] : NULL
);
727 if (!vma
|| check_vma_flags(vma
, gup_flags
)) {
731 if (is_vm_hugetlb_page(vma
)) {
732 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
733 &start
, &nr_pages
, i
,
734 gup_flags
, nonblocking
);
740 * If we have a pending SIGKILL, don't keep faulting pages and
741 * potentially allocating memory.
743 if (fatal_signal_pending(current
)) {
749 page
= follow_page_mask(vma
, start
, foll_flags
, &ctx
);
751 ret
= faultin_page(tsk
, vma
, start
, &foll_flags
,
767 } else if (PTR_ERR(page
) == -EEXIST
) {
769 * Proper page table entry exists, but no corresponding
773 } else if (IS_ERR(page
)) {
779 flush_anon_page(vma
, page
, start
);
780 flush_dcache_page(page
);
788 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & ctx
.page_mask
);
789 if (page_increm
> nr_pages
)
790 page_increm
= nr_pages
;
792 start
+= page_increm
* PAGE_SIZE
;
793 nr_pages
-= page_increm
;
797 put_dev_pagemap(ctx
.pgmap
);
801 static bool vma_permits_fault(struct vm_area_struct
*vma
,
802 unsigned int fault_flags
)
804 bool write
= !!(fault_flags
& FAULT_FLAG_WRITE
);
805 bool foreign
= !!(fault_flags
& FAULT_FLAG_REMOTE
);
806 vm_flags_t vm_flags
= write
? VM_WRITE
: VM_READ
;
808 if (!(vm_flags
& vma
->vm_flags
))
812 * The architecture might have a hardware protection
813 * mechanism other than read/write that can deny access.
815 * gup always represents data access, not instruction
816 * fetches, so execute=false here:
818 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
825 * fixup_user_fault() - manually resolve a user page fault
826 * @tsk: the task_struct to use for page fault accounting, or
827 * NULL if faults are not to be recorded.
828 * @mm: mm_struct of target mm
829 * @address: user address
830 * @fault_flags:flags to pass down to handle_mm_fault()
831 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
832 * does not allow retry
834 * This is meant to be called in the specific scenario where for locking reasons
835 * we try to access user memory in atomic context (within a pagefault_disable()
836 * section), this returns -EFAULT, and we want to resolve the user fault before
839 * Typically this is meant to be used by the futex code.
841 * The main difference with get_user_pages() is that this function will
842 * unconditionally call handle_mm_fault() which will in turn perform all the
843 * necessary SW fixup of the dirty and young bits in the PTE, while
844 * get_user_pages() only guarantees to update these in the struct page.
846 * This is important for some architectures where those bits also gate the
847 * access permission to the page because they are maintained in software. On
848 * such architectures, gup() will not be enough to make a subsequent access
851 * This function will not return with an unlocked mmap_sem. So it has not the
852 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
854 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
855 unsigned long address
, unsigned int fault_flags
,
858 struct vm_area_struct
*vma
;
859 vm_fault_t ret
, major
= 0;
862 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
865 vma
= find_extend_vma(mm
, address
);
866 if (!vma
|| address
< vma
->vm_start
)
869 if (!vma_permits_fault(vma
, fault_flags
))
872 ret
= handle_mm_fault(vma
, address
, fault_flags
);
873 major
|= ret
& VM_FAULT_MAJOR
;
874 if (ret
& VM_FAULT_ERROR
) {
875 int err
= vm_fault_to_errno(ret
, 0);
882 if (ret
& VM_FAULT_RETRY
) {
883 down_read(&mm
->mmap_sem
);
884 if (!(fault_flags
& FAULT_FLAG_TRIED
)) {
886 fault_flags
&= ~FAULT_FLAG_ALLOW_RETRY
;
887 fault_flags
|= FAULT_FLAG_TRIED
;
900 EXPORT_SYMBOL_GPL(fixup_user_fault
);
902 static __always_inline
long __get_user_pages_locked(struct task_struct
*tsk
,
903 struct mm_struct
*mm
,
905 unsigned long nr_pages
,
907 struct vm_area_struct
**vmas
,
911 long ret
, pages_done
;
915 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
917 /* check caller initialized locked */
918 BUG_ON(*locked
!= 1);
925 lock_dropped
= false;
927 ret
= __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
,
930 /* VM_FAULT_RETRY couldn't trigger, bypass */
933 /* VM_FAULT_RETRY cannot return errors */
936 BUG_ON(ret
>= nr_pages
);
940 /* If it's a prefault don't insist harder */
951 * VM_FAULT_RETRY didn't trigger or it was a
958 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
960 start
+= ret
<< PAGE_SHIFT
;
963 * Repeat on the address that fired VM_FAULT_RETRY
964 * without FAULT_FLAG_ALLOW_RETRY but with
969 down_read(&mm
->mmap_sem
);
970 ret
= __get_user_pages(tsk
, mm
, start
, 1, flags
| FOLL_TRIED
,
985 if (lock_dropped
&& *locked
) {
987 * We must let the caller know we temporarily dropped the lock
988 * and so the critical section protected by it was lost.
990 up_read(&mm
->mmap_sem
);
997 * We can leverage the VM_FAULT_RETRY functionality in the page fault
998 * paths better by using either get_user_pages_locked() or
999 * get_user_pages_unlocked().
1001 * get_user_pages_locked() is suitable to replace the form:
1003 * down_read(&mm->mmap_sem);
1005 * get_user_pages(tsk, mm, ..., pages, NULL);
1006 * up_read(&mm->mmap_sem);
1011 * down_read(&mm->mmap_sem);
1013 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
1015 * up_read(&mm->mmap_sem);
1017 long get_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
1018 unsigned int gup_flags
, struct page
**pages
,
1021 return __get_user_pages_locked(current
, current
->mm
, start
, nr_pages
,
1022 pages
, NULL
, locked
,
1023 gup_flags
| FOLL_TOUCH
);
1025 EXPORT_SYMBOL(get_user_pages_locked
);
1028 * get_user_pages_unlocked() is suitable to replace the form:
1030 * down_read(&mm->mmap_sem);
1031 * get_user_pages(tsk, mm, ..., pages, NULL);
1032 * up_read(&mm->mmap_sem);
1036 * get_user_pages_unlocked(tsk, mm, ..., pages);
1038 * It is functionally equivalent to get_user_pages_fast so
1039 * get_user_pages_fast should be used instead if specific gup_flags
1040 * (e.g. FOLL_FORCE) are not required.
1042 long get_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
1043 struct page
**pages
, unsigned int gup_flags
)
1045 struct mm_struct
*mm
= current
->mm
;
1049 down_read(&mm
->mmap_sem
);
1050 ret
= __get_user_pages_locked(current
, mm
, start
, nr_pages
, pages
, NULL
,
1051 &locked
, gup_flags
| FOLL_TOUCH
);
1053 up_read(&mm
->mmap_sem
);
1056 EXPORT_SYMBOL(get_user_pages_unlocked
);
1059 * get_user_pages_remote() - pin user pages in memory
1060 * @tsk: the task_struct to use for page fault accounting, or
1061 * NULL if faults are not to be recorded.
1062 * @mm: mm_struct of target mm
1063 * @start: starting user address
1064 * @nr_pages: number of pages from start to pin
1065 * @gup_flags: flags modifying lookup behaviour
1066 * @pages: array that receives pointers to the pages pinned.
1067 * Should be at least nr_pages long. Or NULL, if caller
1068 * only intends to ensure the pages are faulted in.
1069 * @vmas: array of pointers to vmas corresponding to each page.
1070 * Or NULL if the caller does not require them.
1071 * @locked: pointer to lock flag indicating whether lock is held and
1072 * subsequently whether VM_FAULT_RETRY functionality can be
1073 * utilised. Lock must initially be held.
1075 * Returns number of pages pinned. This may be fewer than the number
1076 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1077 * were pinned, returns -errno. Each page returned must be released
1078 * with a put_page() call when it is finished with. vmas will only
1079 * remain valid while mmap_sem is held.
1081 * Must be called with mmap_sem held for read or write.
1083 * get_user_pages walks a process's page tables and takes a reference to
1084 * each struct page that each user address corresponds to at a given
1085 * instant. That is, it takes the page that would be accessed if a user
1086 * thread accesses the given user virtual address at that instant.
1088 * This does not guarantee that the page exists in the user mappings when
1089 * get_user_pages returns, and there may even be a completely different
1090 * page there in some cases (eg. if mmapped pagecache has been invalidated
1091 * and subsequently re faulted). However it does guarantee that the page
1092 * won't be freed completely. And mostly callers simply care that the page
1093 * contains data that was valid *at some point in time*. Typically, an IO
1094 * or similar operation cannot guarantee anything stronger anyway because
1095 * locks can't be held over the syscall boundary.
1097 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1098 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1099 * be called after the page is finished with, and before put_page is called.
1101 * get_user_pages is typically used for fewer-copy IO operations, to get a
1102 * handle on the memory by some means other than accesses via the user virtual
1103 * addresses. The pages may be submitted for DMA to devices or accessed via
1104 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1105 * use the correct cache flushing APIs.
1107 * See also get_user_pages_fast, for performance critical applications.
1109 * get_user_pages should be phased out in favor of
1110 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1111 * should use get_user_pages because it cannot pass
1112 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1114 long get_user_pages_remote(struct task_struct
*tsk
, struct mm_struct
*mm
,
1115 unsigned long start
, unsigned long nr_pages
,
1116 unsigned int gup_flags
, struct page
**pages
,
1117 struct vm_area_struct
**vmas
, int *locked
)
1119 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
, vmas
,
1121 gup_flags
| FOLL_TOUCH
| FOLL_REMOTE
);
1123 EXPORT_SYMBOL(get_user_pages_remote
);
1126 * This is the same as get_user_pages_remote(), just with a
1127 * less-flexible calling convention where we assume that the task
1128 * and mm being operated on are the current task's and don't allow
1129 * passing of a locked parameter. We also obviously don't pass
1130 * FOLL_REMOTE in here.
1132 long get_user_pages(unsigned long start
, unsigned long nr_pages
,
1133 unsigned int gup_flags
, struct page
**pages
,
1134 struct vm_area_struct
**vmas
)
1136 return __get_user_pages_locked(current
, current
->mm
, start
, nr_pages
,
1138 gup_flags
| FOLL_TOUCH
);
1140 EXPORT_SYMBOL(get_user_pages
);
1142 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1144 #ifdef CONFIG_FS_DAX
1145 static bool check_dax_vmas(struct vm_area_struct
**vmas
, long nr_pages
)
1148 struct vm_area_struct
*vma_prev
= NULL
;
1150 for (i
= 0; i
< nr_pages
; i
++) {
1151 struct vm_area_struct
*vma
= vmas
[i
];
1153 if (vma
== vma_prev
)
1158 if (vma_is_fsdax(vma
))
1164 static inline bool check_dax_vmas(struct vm_area_struct
**vmas
, long nr_pages
)
1171 static struct page
*new_non_cma_page(struct page
*page
, unsigned long private)
1174 * We want to make sure we allocate the new page from the same node
1175 * as the source page.
1177 int nid
= page_to_nid(page
);
1179 * Trying to allocate a page for migration. Ignore allocation
1180 * failure warnings. We don't force __GFP_THISNODE here because
1181 * this node here is the node where we have CMA reservation and
1182 * in some case these nodes will have really less non movable
1183 * allocation memory.
1185 gfp_t gfp_mask
= GFP_USER
| __GFP_NOWARN
;
1187 if (PageHighMem(page
))
1188 gfp_mask
|= __GFP_HIGHMEM
;
1190 #ifdef CONFIG_HUGETLB_PAGE
1191 if (PageHuge(page
)) {
1192 struct hstate
*h
= page_hstate(page
);
1194 * We don't want to dequeue from the pool because pool pages will
1195 * mostly be from the CMA region.
1197 return alloc_migrate_huge_page(h
, gfp_mask
, nid
, NULL
);
1200 if (PageTransHuge(page
)) {
1203 * ignore allocation failure warnings
1205 gfp_t thp_gfpmask
= GFP_TRANSHUGE
| __GFP_NOWARN
;
1208 * Remove the movable mask so that we don't allocate from
1211 thp_gfpmask
&= ~__GFP_MOVABLE
;
1212 thp
= __alloc_pages_node(nid
, thp_gfpmask
, HPAGE_PMD_ORDER
);
1215 prep_transhuge_page(thp
);
1219 return __alloc_pages_node(nid
, gfp_mask
, 0);
1222 static long check_and_migrate_cma_pages(unsigned long start
, long nr_pages
,
1223 unsigned int gup_flags
,
1224 struct page
**pages
,
1225 struct vm_area_struct
**vmas
)
1228 bool drain_allow
= true;
1229 bool migrate_allow
= true;
1230 LIST_HEAD(cma_page_list
);
1233 for (i
= 0; i
< nr_pages
; i
++) {
1235 * If we get a page from the CMA zone, since we are going to
1236 * be pinning these entries, we might as well move them out
1237 * of the CMA zone if possible.
1239 if (is_migrate_cma_page(pages
[i
])) {
1241 struct page
*head
= compound_head(pages
[i
]);
1243 if (PageHuge(head
)) {
1244 isolate_huge_page(head
, &cma_page_list
);
1246 if (!PageLRU(head
) && drain_allow
) {
1247 lru_add_drain_all();
1248 drain_allow
= false;
1251 if (!isolate_lru_page(head
)) {
1252 list_add_tail(&head
->lru
, &cma_page_list
);
1253 mod_node_page_state(page_pgdat(head
),
1255 page_is_file_cache(head
),
1256 hpage_nr_pages(head
));
1262 if (!list_empty(&cma_page_list
)) {
1264 * drop the above get_user_pages reference.
1266 for (i
= 0; i
< nr_pages
; i
++)
1269 if (migrate_pages(&cma_page_list
, new_non_cma_page
,
1270 NULL
, 0, MIGRATE_SYNC
, MR_CONTIG_RANGE
)) {
1272 * some of the pages failed migration. Do get_user_pages
1273 * without migration.
1275 migrate_allow
= false;
1277 if (!list_empty(&cma_page_list
))
1278 putback_movable_pages(&cma_page_list
);
1281 * We did migrate all the pages, Try to get the page references again
1282 * migrating any new CMA pages which we failed to isolate earlier.
1284 nr_pages
= get_user_pages(start
, nr_pages
, gup_flags
, pages
, vmas
);
1285 if ((nr_pages
> 0) && migrate_allow
) {
1294 static inline long check_and_migrate_cma_pages(unsigned long start
, long nr_pages
,
1295 unsigned int gup_flags
,
1296 struct page
**pages
,
1297 struct vm_area_struct
**vmas
)
1304 * This is the same as get_user_pages() in that it assumes we are
1305 * operating on the current task's mm, but it goes further to validate
1306 * that the vmas associated with the address range are suitable for
1307 * longterm elevated page reference counts. For example, filesystem-dax
1308 * mappings are subject to the lifetime enforced by the filesystem and
1309 * we need guarantees that longterm users like RDMA and V4L2 only
1310 * establish mappings that have a kernel enforced revocation mechanism.
1312 * "longterm" == userspace controlled elevated page count lifetime.
1313 * Contrast this to iov_iter_get_pages() usages which are transient.
1315 long get_user_pages_longterm(unsigned long start
, unsigned long nr_pages
,
1316 unsigned int gup_flags
, struct page
**pages
,
1317 struct vm_area_struct
**vmas_arg
)
1319 struct vm_area_struct
**vmas
= vmas_arg
;
1320 unsigned long flags
;
1327 vmas
= kcalloc(nr_pages
, sizeof(struct vm_area_struct
*),
1333 flags
= memalloc_nocma_save();
1334 rc
= get_user_pages(start
, nr_pages
, gup_flags
, pages
, vmas
);
1335 memalloc_nocma_restore(flags
);
1339 if (check_dax_vmas(vmas
, rc
)) {
1340 for (i
= 0; i
< rc
; i
++)
1346 rc
= check_and_migrate_cma_pages(start
, rc
, gup_flags
, pages
, vmas
);
1348 if (vmas
!= vmas_arg
)
1352 EXPORT_SYMBOL(get_user_pages_longterm
);
1353 #endif /* CONFIG_FS_DAX */
1356 * populate_vma_page_range() - populate a range of pages in the vma.
1358 * @start: start address
1362 * This takes care of mlocking the pages too if VM_LOCKED is set.
1364 * return 0 on success, negative error code on error.
1366 * vma->vm_mm->mmap_sem must be held.
1368 * If @nonblocking is NULL, it may be held for read or write and will
1371 * If @nonblocking is non-NULL, it must held for read only and may be
1372 * released. If it's released, *@nonblocking will be set to 0.
1374 long populate_vma_page_range(struct vm_area_struct
*vma
,
1375 unsigned long start
, unsigned long end
, int *nonblocking
)
1377 struct mm_struct
*mm
= vma
->vm_mm
;
1378 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1381 VM_BUG_ON(start
& ~PAGE_MASK
);
1382 VM_BUG_ON(end
& ~PAGE_MASK
);
1383 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1384 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1385 VM_BUG_ON_MM(!rwsem_is_locked(&mm
->mmap_sem
), mm
);
1387 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
1388 if (vma
->vm_flags
& VM_LOCKONFAULT
)
1389 gup_flags
&= ~FOLL_POPULATE
;
1391 * We want to touch writable mappings with a write fault in order
1392 * to break COW, except for shared mappings because these don't COW
1393 * and we would not want to dirty them for nothing.
1395 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
1396 gup_flags
|= FOLL_WRITE
;
1399 * We want mlock to succeed for regions that have any permissions
1400 * other than PROT_NONE.
1402 if (vma
->vm_flags
& (VM_READ
| VM_WRITE
| VM_EXEC
))
1403 gup_flags
|= FOLL_FORCE
;
1406 * We made sure addr is within a VMA, so the following will
1407 * not result in a stack expansion that recurses back here.
1409 return __get_user_pages(current
, mm
, start
, nr_pages
, gup_flags
,
1410 NULL
, NULL
, nonblocking
);
1414 * __mm_populate - populate and/or mlock pages within a range of address space.
1416 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1417 * flags. VMAs must be already marked with the desired vm_flags, and
1418 * mmap_sem must not be held.
1420 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
1422 struct mm_struct
*mm
= current
->mm
;
1423 unsigned long end
, nstart
, nend
;
1424 struct vm_area_struct
*vma
= NULL
;
1430 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
1432 * We want to fault in pages for [nstart; end) address range.
1433 * Find first corresponding VMA.
1437 down_read(&mm
->mmap_sem
);
1438 vma
= find_vma(mm
, nstart
);
1439 } else if (nstart
>= vma
->vm_end
)
1441 if (!vma
|| vma
->vm_start
>= end
)
1444 * Set [nstart; nend) to intersection of desired address
1445 * range with the first VMA. Also, skip undesirable VMA types.
1447 nend
= min(end
, vma
->vm_end
);
1448 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1450 if (nstart
< vma
->vm_start
)
1451 nstart
= vma
->vm_start
;
1453 * Now fault in a range of pages. populate_vma_page_range()
1454 * double checks the vma flags, so that it won't mlock pages
1455 * if the vma was already munlocked.
1457 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1459 if (ignore_errors
) {
1461 continue; /* continue at next VMA */
1465 nend
= nstart
+ ret
* PAGE_SIZE
;
1469 up_read(&mm
->mmap_sem
);
1470 return ret
; /* 0 or negative error code */
1474 * get_dump_page() - pin user page in memory while writing it to core dump
1475 * @addr: user address
1477 * Returns struct page pointer of user page pinned for dump,
1478 * to be freed afterwards by put_page().
1480 * Returns NULL on any kind of failure - a hole must then be inserted into
1481 * the corefile, to preserve alignment with its headers; and also returns
1482 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1483 * allowing a hole to be left in the corefile to save diskspace.
1485 * Called without mmap_sem, but after all other threads have been killed.
1487 #ifdef CONFIG_ELF_CORE
1488 struct page
*get_dump_page(unsigned long addr
)
1490 struct vm_area_struct
*vma
;
1493 if (__get_user_pages(current
, current
->mm
, addr
, 1,
1494 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
1497 flush_cache_page(vma
, addr
, page_to_pfn(page
));
1500 #endif /* CONFIG_ELF_CORE */
1505 * get_user_pages_fast attempts to pin user pages by walking the page
1506 * tables directly and avoids taking locks. Thus the walker needs to be
1507 * protected from page table pages being freed from under it, and should
1508 * block any THP splits.
1510 * One way to achieve this is to have the walker disable interrupts, and
1511 * rely on IPIs from the TLB flushing code blocking before the page table
1512 * pages are freed. This is unsuitable for architectures that do not need
1513 * to broadcast an IPI when invalidating TLBs.
1515 * Another way to achieve this is to batch up page table containing pages
1516 * belonging to more than one mm_user, then rcu_sched a callback to free those
1517 * pages. Disabling interrupts will allow the fast_gup walker to both block
1518 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1519 * (which is a relatively rare event). The code below adopts this strategy.
1521 * Before activating this code, please be aware that the following assumptions
1522 * are currently made:
1524 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1525 * free pages containing page tables or TLB flushing requires IPI broadcast.
1527 * *) ptes can be read atomically by the architecture.
1529 * *) access_ok is sufficient to validate userspace address ranges.
1531 * The last two assumptions can be relaxed by the addition of helper functions.
1533 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1535 #ifdef CONFIG_HAVE_GENERIC_GUP
1539 * We assume that the PTE can be read atomically. If this is not the case for
1540 * your architecture, please provide the helper.
1542 static inline pte_t
gup_get_pte(pte_t
*ptep
)
1544 return READ_ONCE(*ptep
);
1548 static void undo_dev_pagemap(int *nr
, int nr_start
, struct page
**pages
)
1550 while ((*nr
) - nr_start
) {
1551 struct page
*page
= pages
[--(*nr
)];
1553 ClearPageReferenced(page
);
1559 * Return the compund head page with ref appropriately incremented,
1560 * or NULL if that failed.
1562 static inline struct page
*try_get_compound_head(struct page
*page
, int refs
)
1564 struct page
*head
= compound_head(page
);
1565 if (WARN_ON_ONCE(page_ref_count(head
) < 0))
1567 if (unlikely(!page_cache_add_speculative(head
, refs
)))
1572 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
1573 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1574 int write
, struct page
**pages
, int *nr
)
1576 struct dev_pagemap
*pgmap
= NULL
;
1577 int nr_start
= *nr
, ret
= 0;
1580 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
1582 pte_t pte
= gup_get_pte(ptep
);
1583 struct page
*head
, *page
;
1586 * Similar to the PMD case below, NUMA hinting must take slow
1587 * path using the pte_protnone check.
1589 if (pte_protnone(pte
))
1592 if (!pte_access_permitted(pte
, write
))
1595 if (pte_devmap(pte
)) {
1596 pgmap
= get_dev_pagemap(pte_pfn(pte
), pgmap
);
1597 if (unlikely(!pgmap
)) {
1598 undo_dev_pagemap(nr
, nr_start
, pages
);
1601 } else if (pte_special(pte
))
1604 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
1605 page
= pte_page(pte
);
1607 head
= try_get_compound_head(page
, 1);
1611 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
1616 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1618 SetPageReferenced(page
);
1622 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
1628 put_dev_pagemap(pgmap
);
1635 * If we can't determine whether or not a pte is special, then fail immediately
1636 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1639 * For a futex to be placed on a THP tail page, get_futex_key requires a
1640 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1641 * useful to have gup_huge_pmd even if we can't operate on ptes.
1643 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1644 int write
, struct page
**pages
, int *nr
)
1648 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
1650 #if defined(__HAVE_ARCH_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1651 static int __gup_device_huge(unsigned long pfn
, unsigned long addr
,
1652 unsigned long end
, struct page
**pages
, int *nr
)
1655 struct dev_pagemap
*pgmap
= NULL
;
1658 struct page
*page
= pfn_to_page(pfn
);
1660 pgmap
= get_dev_pagemap(pfn
, pgmap
);
1661 if (unlikely(!pgmap
)) {
1662 undo_dev_pagemap(nr
, nr_start
, pages
);
1665 SetPageReferenced(page
);
1670 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1673 put_dev_pagemap(pgmap
);
1677 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1678 unsigned long end
, struct page
**pages
, int *nr
)
1680 unsigned long fault_pfn
;
1683 fault_pfn
= pmd_pfn(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
1684 if (!__gup_device_huge(fault_pfn
, addr
, end
, pages
, nr
))
1687 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
1688 undo_dev_pagemap(nr
, nr_start
, pages
);
1694 static int __gup_device_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
1695 unsigned long end
, struct page
**pages
, int *nr
)
1697 unsigned long fault_pfn
;
1700 fault_pfn
= pud_pfn(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
1701 if (!__gup_device_huge(fault_pfn
, addr
, end
, pages
, nr
))
1704 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
1705 undo_dev_pagemap(nr
, nr_start
, pages
);
1711 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1712 unsigned long end
, struct page
**pages
, int *nr
)
1718 static int __gup_device_huge_pud(pud_t pud
, pud_t
*pudp
, unsigned long addr
,
1719 unsigned long end
, struct page
**pages
, int *nr
)
1726 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1727 unsigned long end
, int write
, struct page
**pages
, int *nr
)
1729 struct page
*head
, *page
;
1732 if (!pmd_access_permitted(orig
, write
))
1735 if (pmd_devmap(orig
))
1736 return __gup_device_huge_pmd(orig
, pmdp
, addr
, end
, pages
, nr
);
1739 page
= pmd_page(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
1745 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1747 head
= try_get_compound_head(pmd_page(orig
), refs
);
1753 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
1760 SetPageReferenced(head
);
1764 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
1765 unsigned long end
, int write
, struct page
**pages
, int *nr
)
1767 struct page
*head
, *page
;
1770 if (!pud_access_permitted(orig
, write
))
1773 if (pud_devmap(orig
))
1774 return __gup_device_huge_pud(orig
, pudp
, addr
, end
, pages
, nr
);
1777 page
= pud_page(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
1783 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1785 head
= try_get_compound_head(pud_page(orig
), refs
);
1791 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
1798 SetPageReferenced(head
);
1802 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
1803 unsigned long end
, int write
,
1804 struct page
**pages
, int *nr
)
1807 struct page
*head
, *page
;
1809 if (!pgd_access_permitted(orig
, write
))
1812 BUILD_BUG_ON(pgd_devmap(orig
));
1814 page
= pgd_page(orig
) + ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
1820 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1822 head
= try_get_compound_head(pgd_page(orig
), refs
);
1828 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
1835 SetPageReferenced(head
);
1839 static int gup_pmd_range(pud_t pud
, unsigned long addr
, unsigned long end
,
1840 int write
, struct page
**pages
, int *nr
)
1845 pmdp
= pmd_offset(&pud
, addr
);
1847 pmd_t pmd
= READ_ONCE(*pmdp
);
1849 next
= pmd_addr_end(addr
, end
);
1850 if (!pmd_present(pmd
))
1853 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
) ||
1856 * NUMA hinting faults need to be handled in the GUP
1857 * slowpath for accounting purposes and so that they
1858 * can be serialised against THP migration.
1860 if (pmd_protnone(pmd
))
1863 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, write
,
1867 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
1869 * architecture have different format for hugetlbfs
1870 * pmd format and THP pmd format
1872 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
1873 PMD_SHIFT
, next
, write
, pages
, nr
))
1875 } else if (!gup_pte_range(pmd
, addr
, next
, write
, pages
, nr
))
1877 } while (pmdp
++, addr
= next
, addr
!= end
);
1882 static int gup_pud_range(p4d_t p4d
, unsigned long addr
, unsigned long end
,
1883 int write
, struct page
**pages
, int *nr
)
1888 pudp
= pud_offset(&p4d
, addr
);
1890 pud_t pud
= READ_ONCE(*pudp
);
1892 next
= pud_addr_end(addr
, end
);
1895 if (unlikely(pud_huge(pud
))) {
1896 if (!gup_huge_pud(pud
, pudp
, addr
, next
, write
,
1899 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
1900 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
1901 PUD_SHIFT
, next
, write
, pages
, nr
))
1903 } else if (!gup_pmd_range(pud
, addr
, next
, write
, pages
, nr
))
1905 } while (pudp
++, addr
= next
, addr
!= end
);
1910 static int gup_p4d_range(pgd_t pgd
, unsigned long addr
, unsigned long end
,
1911 int write
, struct page
**pages
, int *nr
)
1916 p4dp
= p4d_offset(&pgd
, addr
);
1918 p4d_t p4d
= READ_ONCE(*p4dp
);
1920 next
= p4d_addr_end(addr
, end
);
1923 BUILD_BUG_ON(p4d_huge(p4d
));
1924 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d
))))) {
1925 if (!gup_huge_pd(__hugepd(p4d_val(p4d
)), addr
,
1926 P4D_SHIFT
, next
, write
, pages
, nr
))
1928 } else if (!gup_pud_range(p4d
, addr
, next
, write
, pages
, nr
))
1930 } while (p4dp
++, addr
= next
, addr
!= end
);
1935 static void gup_pgd_range(unsigned long addr
, unsigned long end
,
1936 int write
, struct page
**pages
, int *nr
)
1941 pgdp
= pgd_offset(current
->mm
, addr
);
1943 pgd_t pgd
= READ_ONCE(*pgdp
);
1945 next
= pgd_addr_end(addr
, end
);
1948 if (unlikely(pgd_huge(pgd
))) {
1949 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, write
,
1952 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
1953 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
1954 PGDIR_SHIFT
, next
, write
, pages
, nr
))
1956 } else if (!gup_p4d_range(pgd
, addr
, next
, write
, pages
, nr
))
1958 } while (pgdp
++, addr
= next
, addr
!= end
);
1961 #ifndef gup_fast_permitted
1963 * Check if it's allowed to use __get_user_pages_fast() for the range, or
1964 * we need to fall back to the slow version:
1966 bool gup_fast_permitted(unsigned long start
, int nr_pages
)
1968 unsigned long len
, end
;
1970 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
1972 return end
>= start
;
1977 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1979 * Note a difference with get_user_pages_fast: this always returns the
1980 * number of pages pinned, 0 if no pages were pinned.
1982 int __get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
1983 struct page
**pages
)
1985 unsigned long len
, end
;
1986 unsigned long flags
;
1990 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
1993 if (unlikely(!access_ok((void __user
*)start
, len
)))
1997 * Disable interrupts. We use the nested form as we can already have
1998 * interrupts disabled by get_futex_key.
2000 * With interrupts disabled, we block page table pages from being
2001 * freed from under us. See struct mmu_table_batch comments in
2002 * include/asm-generic/tlb.h for more details.
2004 * We do not adopt an rcu_read_lock(.) here as we also want to
2005 * block IPIs that come from THPs splitting.
2008 if (gup_fast_permitted(start
, nr_pages
)) {
2009 local_irq_save(flags
);
2010 gup_pgd_range(start
, end
, write
, pages
, &nr
);
2011 local_irq_restore(flags
);
2018 * get_user_pages_fast() - pin user pages in memory
2019 * @start: starting user address
2020 * @nr_pages: number of pages from start to pin
2021 * @write: whether pages will be written to
2022 * @pages: array that receives pointers to the pages pinned.
2023 * Should be at least nr_pages long.
2025 * Attempt to pin user pages in memory without taking mm->mmap_sem.
2026 * If not successful, it will fall back to taking the lock and
2027 * calling get_user_pages().
2029 * Returns number of pages pinned. This may be fewer than the number
2030 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2031 * were pinned, returns -errno.
2033 int get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
2034 struct page
**pages
)
2036 unsigned long addr
, len
, end
;
2037 int nr
= 0, ret
= 0;
2041 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
2047 if (unlikely(!access_ok((void __user
*)start
, len
)))
2050 if (gup_fast_permitted(start
, nr_pages
)) {
2051 local_irq_disable();
2052 gup_pgd_range(addr
, end
, write
, pages
, &nr
);
2057 if (nr
< nr_pages
) {
2058 /* Try to get the remaining pages with get_user_pages */
2059 start
+= nr
<< PAGE_SHIFT
;
2062 ret
= get_user_pages_unlocked(start
, nr_pages
- nr
, pages
,
2063 write
? FOLL_WRITE
: 0);
2065 /* Have to be a bit careful with return values */
2077 #endif /* CONFIG_HAVE_GENERIC_GUP */