]>
git.proxmox.com Git - mirror_ubuntu-bionic-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>
17 #include <asm/mmu_context.h>
18 #include <asm/pgtable.h>
19 #include <asm/tlbflush.h>
23 static struct page
*no_page_table(struct vm_area_struct
*vma
,
27 * When core dumping an enormous anonymous area that nobody
28 * has touched so far, we don't want to allocate unnecessary pages or
29 * page tables. Return error instead of NULL to skip handle_mm_fault,
30 * then get_dump_page() will return NULL to leave a hole in the dump.
31 * But we can only make this optimization where a hole would surely
32 * be zero-filled if handle_mm_fault() actually did handle it.
34 if ((flags
& FOLL_DUMP
) && (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
35 return ERR_PTR(-EFAULT
);
39 static int follow_pfn_pte(struct vm_area_struct
*vma
, unsigned long address
,
40 pte_t
*pte
, unsigned int flags
)
42 /* No page to get reference */
46 if (flags
& FOLL_TOUCH
) {
49 if (flags
& FOLL_WRITE
)
50 entry
= pte_mkdirty(entry
);
51 entry
= pte_mkyoung(entry
);
53 if (!pte_same(*pte
, entry
)) {
54 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
55 update_mmu_cache(vma
, address
, pte
);
59 /* Proper page table entry exists, but no corresponding struct page */
64 * FOLL_FORCE can write to even unwritable pte's, but only
65 * after we've gone through a COW cycle and they are dirty.
67 static inline bool can_follow_write_pte(pte_t pte
, unsigned int flags
)
69 return pte_write(pte
) ||
70 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pte_dirty(pte
));
73 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
74 unsigned long address
, pmd_t
*pmd
, unsigned int flags
)
76 struct mm_struct
*mm
= vma
->vm_mm
;
77 struct dev_pagemap
*pgmap
= NULL
;
83 if (unlikely(pmd_bad(*pmd
)))
84 return no_page_table(vma
, flags
);
86 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
88 if (!pte_present(pte
)) {
91 * KSM's break_ksm() relies upon recognizing a ksm page
92 * even while it is being migrated, so for that case we
93 * need migration_entry_wait().
95 if (likely(!(flags
& FOLL_MIGRATION
)))
99 entry
= pte_to_swp_entry(pte
);
100 if (!is_migration_entry(entry
))
102 pte_unmap_unlock(ptep
, ptl
);
103 migration_entry_wait(mm
, pmd
, address
);
106 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
108 if ((flags
& FOLL_WRITE
) && !can_follow_write_pte(pte
, flags
)) {
109 pte_unmap_unlock(ptep
, ptl
);
113 page
= vm_normal_page(vma
, address
, pte
);
114 if (!page
&& pte_devmap(pte
) && (flags
& FOLL_GET
)) {
116 * Only return device mapping pages in the FOLL_GET case since
117 * they are only valid while holding the pgmap reference.
119 pgmap
= get_dev_pagemap(pte_pfn(pte
), NULL
);
121 page
= pte_page(pte
);
124 } else if (unlikely(!page
)) {
125 if (flags
& FOLL_DUMP
) {
126 /* Avoid special (like zero) pages in core dumps */
127 page
= ERR_PTR(-EFAULT
);
131 if (is_zero_pfn(pte_pfn(pte
))) {
132 page
= pte_page(pte
);
136 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
142 if (flags
& FOLL_SPLIT
&& PageTransCompound(page
)) {
145 pte_unmap_unlock(ptep
, ptl
);
147 ret
= split_huge_page(page
);
155 if (flags
& FOLL_GET
) {
158 /* drop the pgmap reference now that we hold the page */
160 put_dev_pagemap(pgmap
);
164 if (flags
& FOLL_TOUCH
) {
165 if ((flags
& FOLL_WRITE
) &&
166 !pte_dirty(pte
) && !PageDirty(page
))
167 set_page_dirty(page
);
169 * pte_mkyoung() would be more correct here, but atomic care
170 * is needed to avoid losing the dirty bit: it is easier to use
171 * mark_page_accessed().
173 mark_page_accessed(page
);
175 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
176 /* Do not mlock pte-mapped THP */
177 if (PageTransCompound(page
))
181 * The preliminary mapping check is mainly to avoid the
182 * pointless overhead of lock_page on the ZERO_PAGE
183 * which might bounce very badly if there is contention.
185 * If the page is already locked, we don't need to
186 * handle it now - vmscan will handle it later if and
187 * when it attempts to reclaim the page.
189 if (page
->mapping
&& trylock_page(page
)) {
190 lru_add_drain(); /* push cached pages to LRU */
192 * Because we lock page here, and migration is
193 * blocked by the pte's page reference, and we
194 * know the page is still mapped, we don't even
195 * need to check for file-cache page truncation.
197 mlock_vma_page(page
);
202 pte_unmap_unlock(ptep
, ptl
);
205 pte_unmap_unlock(ptep
, ptl
);
208 return no_page_table(vma
, flags
);
211 static struct page
*follow_pmd_mask(struct vm_area_struct
*vma
,
212 unsigned long address
, pud_t
*pudp
,
213 unsigned int flags
, unsigned int *page_mask
)
218 struct mm_struct
*mm
= vma
->vm_mm
;
220 pmd
= pmd_offset(pudp
, address
);
222 return no_page_table(vma
, flags
);
223 if (pmd_huge(*pmd
) && vma
->vm_flags
& VM_HUGETLB
) {
224 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
227 return no_page_table(vma
, flags
);
229 if (is_hugepd(__hugepd(pmd_val(*pmd
)))) {
230 page
= follow_huge_pd(vma
, address
,
231 __hugepd(pmd_val(*pmd
)), flags
,
235 return no_page_table(vma
, flags
);
238 if (!pmd_present(*pmd
)) {
239 if (likely(!(flags
& FOLL_MIGRATION
)))
240 return no_page_table(vma
, flags
);
241 VM_BUG_ON(thp_migration_supported() &&
242 !is_pmd_migration_entry(*pmd
));
243 if (is_pmd_migration_entry(*pmd
))
244 pmd_migration_entry_wait(mm
, pmd
);
247 if (pmd_devmap(*pmd
)) {
248 ptl
= pmd_lock(mm
, pmd
);
249 page
= follow_devmap_pmd(vma
, address
, pmd
, flags
);
254 if (likely(!pmd_trans_huge(*pmd
)))
255 return follow_page_pte(vma
, address
, pmd
, flags
);
257 if ((flags
& FOLL_NUMA
) && pmd_protnone(*pmd
))
258 return no_page_table(vma
, flags
);
261 ptl
= pmd_lock(mm
, pmd
);
262 if (unlikely(!pmd_present(*pmd
))) {
264 if (likely(!(flags
& FOLL_MIGRATION
)))
265 return no_page_table(vma
, flags
);
266 pmd_migration_entry_wait(mm
, pmd
);
269 if (unlikely(!pmd_trans_huge(*pmd
))) {
271 return follow_page_pte(vma
, address
, pmd
, flags
);
273 if (flags
& FOLL_SPLIT
) {
275 page
= pmd_page(*pmd
);
276 if (is_huge_zero_page(page
)) {
279 split_huge_pmd(vma
, pmd
, address
);
280 if (pmd_trans_unstable(pmd
))
286 ret
= split_huge_page(page
);
290 return no_page_table(vma
, flags
);
293 return ret
? ERR_PTR(ret
) :
294 follow_page_pte(vma
, address
, pmd
, flags
);
296 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
298 *page_mask
= HPAGE_PMD_NR
- 1;
303 static struct page
*follow_pud_mask(struct vm_area_struct
*vma
,
304 unsigned long address
, p4d_t
*p4dp
,
305 unsigned int flags
, unsigned int *page_mask
)
310 struct mm_struct
*mm
= vma
->vm_mm
;
312 pud
= pud_offset(p4dp
, address
);
314 return no_page_table(vma
, flags
);
315 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
316 page
= follow_huge_pud(mm
, address
, pud
, flags
);
319 return no_page_table(vma
, flags
);
321 if (is_hugepd(__hugepd(pud_val(*pud
)))) {
322 page
= follow_huge_pd(vma
, address
,
323 __hugepd(pud_val(*pud
)), flags
,
327 return no_page_table(vma
, flags
);
329 if (pud_devmap(*pud
)) {
330 ptl
= pud_lock(mm
, pud
);
331 page
= follow_devmap_pud(vma
, address
, pud
, flags
);
336 if (unlikely(pud_bad(*pud
)))
337 return no_page_table(vma
, flags
);
339 return follow_pmd_mask(vma
, address
, pud
, flags
, page_mask
);
343 static struct page
*follow_p4d_mask(struct vm_area_struct
*vma
,
344 unsigned long address
, pgd_t
*pgdp
,
345 unsigned int flags
, unsigned int *page_mask
)
350 p4d
= p4d_offset(pgdp
, address
);
352 return no_page_table(vma
, flags
);
353 BUILD_BUG_ON(p4d_huge(*p4d
));
354 if (unlikely(p4d_bad(*p4d
)))
355 return no_page_table(vma
, flags
);
357 if (is_hugepd(__hugepd(p4d_val(*p4d
)))) {
358 page
= follow_huge_pd(vma
, address
,
359 __hugepd(p4d_val(*p4d
)), flags
,
363 return no_page_table(vma
, flags
);
365 return follow_pud_mask(vma
, address
, p4d
, flags
, page_mask
);
369 * follow_page_mask - look up a page descriptor from a user-virtual address
370 * @vma: vm_area_struct mapping @address
371 * @address: virtual address to look up
372 * @flags: flags modifying lookup behaviour
373 * @page_mask: on output, *page_mask is set according to the size of the page
375 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
377 * Returns the mapped (struct page *), %NULL if no mapping exists, or
378 * an error pointer if there is a mapping to something not represented
379 * by a page descriptor (see also vm_normal_page()).
381 struct page
*follow_page_mask(struct vm_area_struct
*vma
,
382 unsigned long address
, unsigned int flags
,
383 unsigned int *page_mask
)
387 struct mm_struct
*mm
= vma
->vm_mm
;
391 /* make this handle hugepd */
392 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
394 BUG_ON(flags
& FOLL_GET
);
398 pgd
= pgd_offset(mm
, address
);
400 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
401 return no_page_table(vma
, flags
);
403 if (pgd_huge(*pgd
)) {
404 page
= follow_huge_pgd(mm
, address
, pgd
, flags
);
407 return no_page_table(vma
, flags
);
409 if (is_hugepd(__hugepd(pgd_val(*pgd
)))) {
410 page
= follow_huge_pd(vma
, address
,
411 __hugepd(pgd_val(*pgd
)), flags
,
415 return no_page_table(vma
, flags
);
418 return follow_p4d_mask(vma
, address
, pgd
, flags
, page_mask
);
421 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
422 unsigned int gup_flags
, struct vm_area_struct
**vma
,
432 /* user gate pages are read-only */
433 if (gup_flags
& FOLL_WRITE
)
435 if (address
> TASK_SIZE
)
436 pgd
= pgd_offset_k(address
);
438 pgd
= pgd_offset_gate(mm
, address
);
439 BUG_ON(pgd_none(*pgd
));
440 p4d
= p4d_offset(pgd
, address
);
441 BUG_ON(p4d_none(*p4d
));
442 pud
= pud_offset(p4d
, address
);
443 BUG_ON(pud_none(*pud
));
444 pmd
= pmd_offset(pud
, address
);
445 if (!pmd_present(*pmd
))
447 VM_BUG_ON(pmd_trans_huge(*pmd
));
448 pte
= pte_offset_map(pmd
, address
);
451 *vma
= get_gate_vma(mm
);
454 *page
= vm_normal_page(*vma
, address
, *pte
);
456 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
458 *page
= pte_page(*pte
);
461 * This should never happen (a device public page in the gate
464 if (is_device_public_page(*page
))
476 * mmap_sem must be held on entry. If @nonblocking != NULL and
477 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
478 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
480 static int faultin_page(struct task_struct
*tsk
, struct vm_area_struct
*vma
,
481 unsigned long address
, unsigned int *flags
, int *nonblocking
)
483 unsigned int fault_flags
= 0;
486 /* mlock all present pages, but do not fault in new pages */
487 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
489 if (*flags
& FOLL_WRITE
)
490 fault_flags
|= FAULT_FLAG_WRITE
;
491 if (*flags
& FOLL_REMOTE
)
492 fault_flags
|= FAULT_FLAG_REMOTE
;
494 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
495 if (*flags
& FOLL_NOWAIT
)
496 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
497 if (*flags
& FOLL_TRIED
) {
498 VM_WARN_ON_ONCE(fault_flags
& FAULT_FLAG_ALLOW_RETRY
);
499 fault_flags
|= FAULT_FLAG_TRIED
;
502 ret
= handle_mm_fault(vma
, address
, fault_flags
);
503 if (ret
& VM_FAULT_ERROR
) {
504 int err
= vm_fault_to_errno(ret
, *flags
);
512 if (ret
& VM_FAULT_MAJOR
)
518 if (ret
& VM_FAULT_RETRY
) {
525 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
526 * necessary, even if maybe_mkwrite decided not to set pte_write. We
527 * can thus safely do subsequent page lookups as if they were reads.
528 * But only do so when looping for pte_write is futile: in some cases
529 * userspace may also be wanting to write to the gotten user page,
530 * which a read fault here might prevent (a readonly page might get
531 * reCOWed by userspace write).
533 if ((ret
& VM_FAULT_WRITE
) && !(vma
->vm_flags
& VM_WRITE
))
538 static int check_vma_flags(struct vm_area_struct
*vma
, unsigned long gup_flags
)
540 vm_flags_t vm_flags
= vma
->vm_flags
;
541 int write
= (gup_flags
& FOLL_WRITE
);
542 int foreign
= (gup_flags
& FOLL_REMOTE
);
544 if (vm_flags
& (VM_IO
| VM_PFNMAP
))
547 if (gup_flags
& FOLL_ANON
&& !vma_is_anonymous(vma
))
551 if (!(vm_flags
& VM_WRITE
)) {
552 if (!(gup_flags
& FOLL_FORCE
))
555 * We used to let the write,force case do COW in a
556 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
557 * set a breakpoint in a read-only mapping of an
558 * executable, without corrupting the file (yet only
559 * when that file had been opened for writing!).
560 * Anon pages in shared mappings are surprising: now
563 if (!is_cow_mapping(vm_flags
))
566 } else if (!(vm_flags
& VM_READ
)) {
567 if (!(gup_flags
& FOLL_FORCE
))
570 * Is there actually any vma we can reach here which does not
571 * have VM_MAYREAD set?
573 if (!(vm_flags
& VM_MAYREAD
))
577 * gups are always data accesses, not instruction
578 * fetches, so execute=false here
580 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
586 * __get_user_pages() - pin user pages in memory
587 * @tsk: task_struct of target task
588 * @mm: mm_struct of target mm
589 * @start: starting user address
590 * @nr_pages: number of pages from start to pin
591 * @gup_flags: flags modifying pin behaviour
592 * @pages: array that receives pointers to the pages pinned.
593 * Should be at least nr_pages long. Or NULL, if caller
594 * only intends to ensure the pages are faulted in.
595 * @vmas: array of pointers to vmas corresponding to each page.
596 * Or NULL if the caller does not require them.
597 * @nonblocking: whether waiting for disk IO or mmap_sem contention
599 * Returns number of pages pinned. This may be fewer than the number
600 * requested. If nr_pages is 0 or negative, returns 0. If no pages
601 * were pinned, returns -errno. Each page returned must be released
602 * with a put_page() call when it is finished with. vmas will only
603 * remain valid while mmap_sem is held.
605 * Must be called with mmap_sem held. It may be released. See below.
607 * __get_user_pages walks a process's page tables and takes a reference to
608 * each struct page that each user address corresponds to at a given
609 * instant. That is, it takes the page that would be accessed if a user
610 * thread accesses the given user virtual address at that instant.
612 * This does not guarantee that the page exists in the user mappings when
613 * __get_user_pages returns, and there may even be a completely different
614 * page there in some cases (eg. if mmapped pagecache has been invalidated
615 * and subsequently re faulted). However it does guarantee that the page
616 * won't be freed completely. And mostly callers simply care that the page
617 * contains data that was valid *at some point in time*. Typically, an IO
618 * or similar operation cannot guarantee anything stronger anyway because
619 * locks can't be held over the syscall boundary.
621 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
622 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
623 * appropriate) must be called after the page is finished with, and
624 * before put_page is called.
626 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
627 * or mmap_sem contention, and if waiting is needed to pin all pages,
628 * *@nonblocking will be set to 0. Further, if @gup_flags does not
629 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
632 * A caller using such a combination of @nonblocking and @gup_flags
633 * must therefore hold the mmap_sem for reading only, and recognize
634 * when it's been released. Otherwise, it must be held for either
635 * reading or writing and will not be released.
637 * In most cases, get_user_pages or get_user_pages_fast should be used
638 * instead of __get_user_pages. __get_user_pages should be used only if
639 * you need some special @gup_flags.
641 static long __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
642 unsigned long start
, unsigned long nr_pages
,
643 unsigned int gup_flags
, struct page
**pages
,
644 struct vm_area_struct
**vmas
, int *nonblocking
)
647 unsigned int page_mask
;
648 struct vm_area_struct
*vma
= NULL
;
653 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
656 * If FOLL_FORCE is set then do not force a full fault as the hinting
657 * fault information is unrelated to the reference behaviour of a task
658 * using the address space
660 if (!(gup_flags
& FOLL_FORCE
))
661 gup_flags
|= FOLL_NUMA
;
665 unsigned int foll_flags
= gup_flags
;
666 unsigned int page_increm
;
668 /* first iteration or cross vma bound */
669 if (!vma
|| start
>= vma
->vm_end
) {
670 vma
= find_extend_vma(mm
, start
);
671 if (!vma
&& in_gate_area(mm
, start
)) {
673 ret
= get_gate_page(mm
, start
& PAGE_MASK
,
675 pages
? &pages
[i
] : NULL
);
682 if (!vma
|| check_vma_flags(vma
, gup_flags
))
683 return i
? : -EFAULT
;
684 if (is_vm_hugetlb_page(vma
)) {
685 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
686 &start
, &nr_pages
, i
,
687 gup_flags
, nonblocking
);
693 * If we have a pending SIGKILL, don't keep faulting pages and
694 * potentially allocating memory.
696 if (unlikely(fatal_signal_pending(current
)))
697 return i
? i
: -ERESTARTSYS
;
699 page
= follow_page_mask(vma
, start
, foll_flags
, &page_mask
);
702 ret
= faultin_page(tsk
, vma
, start
, &foll_flags
,
717 } else if (PTR_ERR(page
) == -EEXIST
) {
719 * Proper page table entry exists, but no corresponding
723 } else if (IS_ERR(page
)) {
724 return i
? i
: PTR_ERR(page
);
728 flush_anon_page(vma
, page
, start
);
729 flush_dcache_page(page
);
737 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & page_mask
);
738 if (page_increm
> nr_pages
)
739 page_increm
= nr_pages
;
741 start
+= page_increm
* PAGE_SIZE
;
742 nr_pages
-= page_increm
;
747 static bool vma_permits_fault(struct vm_area_struct
*vma
,
748 unsigned int fault_flags
)
750 bool write
= !!(fault_flags
& FAULT_FLAG_WRITE
);
751 bool foreign
= !!(fault_flags
& FAULT_FLAG_REMOTE
);
752 vm_flags_t vm_flags
= write
? VM_WRITE
: VM_READ
;
754 if (!(vm_flags
& vma
->vm_flags
))
758 * The architecture might have a hardware protection
759 * mechanism other than read/write that can deny access.
761 * gup always represents data access, not instruction
762 * fetches, so execute=false here:
764 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
771 * fixup_user_fault() - manually resolve a user page fault
772 * @tsk: the task_struct to use for page fault accounting, or
773 * NULL if faults are not to be recorded.
774 * @mm: mm_struct of target mm
775 * @address: user address
776 * @fault_flags:flags to pass down to handle_mm_fault()
777 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
778 * does not allow retry
780 * This is meant to be called in the specific scenario where for locking reasons
781 * we try to access user memory in atomic context (within a pagefault_disable()
782 * section), this returns -EFAULT, and we want to resolve the user fault before
785 * Typically this is meant to be used by the futex code.
787 * The main difference with get_user_pages() is that this function will
788 * unconditionally call handle_mm_fault() which will in turn perform all the
789 * necessary SW fixup of the dirty and young bits in the PTE, while
790 * get_user_pages() only guarantees to update these in the struct page.
792 * This is important for some architectures where those bits also gate the
793 * access permission to the page because they are maintained in software. On
794 * such architectures, gup() will not be enough to make a subsequent access
797 * This function will not return with an unlocked mmap_sem. So it has not the
798 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
800 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
801 unsigned long address
, unsigned int fault_flags
,
804 struct vm_area_struct
*vma
;
808 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
811 vma
= find_extend_vma(mm
, address
);
812 if (!vma
|| address
< vma
->vm_start
)
815 if (!vma_permits_fault(vma
, fault_flags
))
818 ret
= handle_mm_fault(vma
, address
, fault_flags
);
819 major
|= ret
& VM_FAULT_MAJOR
;
820 if (ret
& VM_FAULT_ERROR
) {
821 int err
= vm_fault_to_errno(ret
, 0);
828 if (ret
& VM_FAULT_RETRY
) {
829 down_read(&mm
->mmap_sem
);
830 if (!(fault_flags
& FAULT_FLAG_TRIED
)) {
832 fault_flags
&= ~FAULT_FLAG_ALLOW_RETRY
;
833 fault_flags
|= FAULT_FLAG_TRIED
;
846 EXPORT_SYMBOL_GPL(fixup_user_fault
);
848 static __always_inline
long __get_user_pages_locked(struct task_struct
*tsk
,
849 struct mm_struct
*mm
,
851 unsigned long nr_pages
,
853 struct vm_area_struct
**vmas
,
854 int *locked
, bool notify_drop
,
857 long ret
, pages_done
;
861 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
863 /* check caller initialized locked */
864 BUG_ON(*locked
!= 1);
871 lock_dropped
= false;
873 ret
= __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
,
876 /* VM_FAULT_RETRY couldn't trigger, bypass */
879 /* VM_FAULT_RETRY cannot return errors */
882 BUG_ON(ret
>= nr_pages
);
886 /* If it's a prefault don't insist harder */
896 /* VM_FAULT_RETRY didn't trigger */
901 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
903 start
+= ret
<< PAGE_SHIFT
;
906 * Repeat on the address that fired VM_FAULT_RETRY
907 * without FAULT_FLAG_ALLOW_RETRY but with
912 down_read(&mm
->mmap_sem
);
913 ret
= __get_user_pages(tsk
, mm
, start
, 1, flags
| FOLL_TRIED
,
928 if (notify_drop
&& lock_dropped
&& *locked
) {
930 * We must let the caller know we temporarily dropped the lock
931 * and so the critical section protected by it was lost.
933 up_read(&mm
->mmap_sem
);
940 * We can leverage the VM_FAULT_RETRY functionality in the page fault
941 * paths better by using either get_user_pages_locked() or
942 * get_user_pages_unlocked().
944 * get_user_pages_locked() is suitable to replace the form:
946 * down_read(&mm->mmap_sem);
948 * get_user_pages(tsk, mm, ..., pages, NULL);
949 * up_read(&mm->mmap_sem);
954 * down_read(&mm->mmap_sem);
956 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
958 * up_read(&mm->mmap_sem);
960 long get_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
961 unsigned int gup_flags
, struct page
**pages
,
964 return __get_user_pages_locked(current
, current
->mm
, start
, nr_pages
,
965 pages
, NULL
, locked
, true,
966 gup_flags
| FOLL_TOUCH
);
968 EXPORT_SYMBOL(get_user_pages_locked
);
971 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows for
972 * tsk, mm to be specified.
974 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
975 * caller if required (just like with __get_user_pages). "FOLL_GET"
976 * is set implicitly if "pages" is non-NULL.
978 static __always_inline
long __get_user_pages_unlocked(struct task_struct
*tsk
,
979 struct mm_struct
*mm
, unsigned long start
,
980 unsigned long nr_pages
, struct page
**pages
,
981 unsigned int gup_flags
)
986 down_read(&mm
->mmap_sem
);
987 ret
= __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
, NULL
,
988 &locked
, false, gup_flags
);
990 up_read(&mm
->mmap_sem
);
995 * get_user_pages_unlocked() is suitable to replace the form:
997 * down_read(&mm->mmap_sem);
998 * get_user_pages(tsk, mm, ..., pages, NULL);
999 * up_read(&mm->mmap_sem);
1003 * get_user_pages_unlocked(tsk, mm, ..., pages);
1005 * It is functionally equivalent to get_user_pages_fast so
1006 * get_user_pages_fast should be used instead if specific gup_flags
1007 * (e.g. FOLL_FORCE) are not required.
1009 long get_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
1010 struct page
**pages
, unsigned int gup_flags
)
1012 return __get_user_pages_unlocked(current
, current
->mm
, start
, nr_pages
,
1013 pages
, gup_flags
| FOLL_TOUCH
);
1015 EXPORT_SYMBOL(get_user_pages_unlocked
);
1018 * get_user_pages_remote() - pin user pages in memory
1019 * @tsk: the task_struct to use for page fault accounting, or
1020 * NULL if faults are not to be recorded.
1021 * @mm: mm_struct of target mm
1022 * @start: starting user address
1023 * @nr_pages: number of pages from start to pin
1024 * @gup_flags: flags modifying lookup behaviour
1025 * @pages: array that receives pointers to the pages pinned.
1026 * Should be at least nr_pages long. Or NULL, if caller
1027 * only intends to ensure the pages are faulted in.
1028 * @vmas: array of pointers to vmas corresponding to each page.
1029 * Or NULL if the caller does not require them.
1030 * @locked: pointer to lock flag indicating whether lock is held and
1031 * subsequently whether VM_FAULT_RETRY functionality can be
1032 * utilised. Lock must initially be held.
1034 * Returns number of pages pinned. This may be fewer than the number
1035 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1036 * were pinned, returns -errno. Each page returned must be released
1037 * with a put_page() call when it is finished with. vmas will only
1038 * remain valid while mmap_sem is held.
1040 * Must be called with mmap_sem held for read or write.
1042 * get_user_pages walks a process's page tables and takes a reference to
1043 * each struct page that each user address corresponds to at a given
1044 * instant. That is, it takes the page that would be accessed if a user
1045 * thread accesses the given user virtual address at that instant.
1047 * This does not guarantee that the page exists in the user mappings when
1048 * get_user_pages returns, and there may even be a completely different
1049 * page there in some cases (eg. if mmapped pagecache has been invalidated
1050 * and subsequently re faulted). However it does guarantee that the page
1051 * won't be freed completely. And mostly callers simply care that the page
1052 * contains data that was valid *at some point in time*. Typically, an IO
1053 * or similar operation cannot guarantee anything stronger anyway because
1054 * locks can't be held over the syscall boundary.
1056 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1057 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1058 * be called after the page is finished with, and before put_page is called.
1060 * get_user_pages is typically used for fewer-copy IO operations, to get a
1061 * handle on the memory by some means other than accesses via the user virtual
1062 * addresses. The pages may be submitted for DMA to devices or accessed via
1063 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1064 * use the correct cache flushing APIs.
1066 * See also get_user_pages_fast, for performance critical applications.
1068 * get_user_pages should be phased out in favor of
1069 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1070 * should use get_user_pages because it cannot pass
1071 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1073 long get_user_pages_remote(struct task_struct
*tsk
, struct mm_struct
*mm
,
1074 unsigned long start
, unsigned long nr_pages
,
1075 unsigned int gup_flags
, struct page
**pages
,
1076 struct vm_area_struct
**vmas
, int *locked
)
1078 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
, vmas
,
1080 gup_flags
| FOLL_TOUCH
| FOLL_REMOTE
);
1082 EXPORT_SYMBOL(get_user_pages_remote
);
1085 * This is the same as get_user_pages_remote(), just with a
1086 * less-flexible calling convention where we assume that the task
1087 * and mm being operated on are the current task's and don't allow
1088 * passing of a locked parameter. We also obviously don't pass
1089 * FOLL_REMOTE in here.
1091 long get_user_pages(unsigned long start
, unsigned long nr_pages
,
1092 unsigned int gup_flags
, struct page
**pages
,
1093 struct vm_area_struct
**vmas
)
1095 return __get_user_pages_locked(current
, current
->mm
, start
, nr_pages
,
1096 pages
, vmas
, NULL
, false,
1097 gup_flags
| FOLL_TOUCH
);
1099 EXPORT_SYMBOL(get_user_pages
);
1101 #ifdef CONFIG_FS_DAX
1103 * This is the same as get_user_pages() in that it assumes we are
1104 * operating on the current task's mm, but it goes further to validate
1105 * that the vmas associated with the address range are suitable for
1106 * longterm elevated page reference counts. For example, filesystem-dax
1107 * mappings are subject to the lifetime enforced by the filesystem and
1108 * we need guarantees that longterm users like RDMA and V4L2 only
1109 * establish mappings that have a kernel enforced revocation mechanism.
1111 * "longterm" == userspace controlled elevated page count lifetime.
1112 * Contrast this to iov_iter_get_pages() usages which are transient.
1114 long get_user_pages_longterm(unsigned long start
, unsigned long nr_pages
,
1115 unsigned int gup_flags
, struct page
**pages
,
1116 struct vm_area_struct
**vmas_arg
)
1118 struct vm_area_struct
**vmas
= vmas_arg
;
1119 struct vm_area_struct
*vma_prev
= NULL
;
1126 vmas
= kcalloc(nr_pages
, sizeof(struct vm_area_struct
*),
1132 rc
= get_user_pages(start
, nr_pages
, gup_flags
, pages
, vmas
);
1134 for (i
= 0; i
< rc
; i
++) {
1135 struct vm_area_struct
*vma
= vmas
[i
];
1137 if (vma
== vma_prev
)
1142 if (vma_is_fsdax(vma
))
1147 * Either get_user_pages() failed, or the vma validation
1148 * succeeded, in either case we don't need to put_page() before
1154 for (i
= 0; i
< rc
; i
++)
1158 if (vmas
!= vmas_arg
)
1162 EXPORT_SYMBOL(get_user_pages_longterm
);
1163 #endif /* CONFIG_FS_DAX */
1166 * populate_vma_page_range() - populate a range of pages in the vma.
1168 * @start: start address
1172 * This takes care of mlocking the pages too if VM_LOCKED is set.
1174 * return 0 on success, negative error code on error.
1176 * vma->vm_mm->mmap_sem must be held.
1178 * If @nonblocking is NULL, it may be held for read or write and will
1181 * If @nonblocking is non-NULL, it must held for read only and may be
1182 * released. If it's released, *@nonblocking will be set to 0.
1184 long populate_vma_page_range(struct vm_area_struct
*vma
,
1185 unsigned long start
, unsigned long end
, int *nonblocking
)
1187 struct mm_struct
*mm
= vma
->vm_mm
;
1188 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1191 VM_BUG_ON(start
& ~PAGE_MASK
);
1192 VM_BUG_ON(end
& ~PAGE_MASK
);
1193 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1194 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1195 VM_BUG_ON_MM(!rwsem_is_locked(&mm
->mmap_sem
), mm
);
1197 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
1198 if (vma
->vm_flags
& VM_LOCKONFAULT
)
1199 gup_flags
&= ~FOLL_POPULATE
;
1201 * We want to touch writable mappings with a write fault in order
1202 * to break COW, except for shared mappings because these don't COW
1203 * and we would not want to dirty them for nothing.
1205 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
1206 gup_flags
|= FOLL_WRITE
;
1209 * We want mlock to succeed for regions that have any permissions
1210 * other than PROT_NONE.
1212 if (vma
->vm_flags
& (VM_READ
| VM_WRITE
| VM_EXEC
))
1213 gup_flags
|= FOLL_FORCE
;
1216 * We made sure addr is within a VMA, so the following will
1217 * not result in a stack expansion that recurses back here.
1219 return __get_user_pages(current
, mm
, start
, nr_pages
, gup_flags
,
1220 NULL
, NULL
, nonblocking
);
1224 * __mm_populate - populate and/or mlock pages within a range of address space.
1226 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1227 * flags. VMAs must be already marked with the desired vm_flags, and
1228 * mmap_sem must not be held.
1230 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
1232 struct mm_struct
*mm
= current
->mm
;
1233 unsigned long end
, nstart
, nend
;
1234 struct vm_area_struct
*vma
= NULL
;
1240 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
1242 * We want to fault in pages for [nstart; end) address range.
1243 * Find first corresponding VMA.
1247 down_read(&mm
->mmap_sem
);
1248 vma
= find_vma(mm
, nstart
);
1249 } else if (nstart
>= vma
->vm_end
)
1251 if (!vma
|| vma
->vm_start
>= end
)
1254 * Set [nstart; nend) to intersection of desired address
1255 * range with the first VMA. Also, skip undesirable VMA types.
1257 nend
= min(end
, vma
->vm_end
);
1258 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1260 if (nstart
< vma
->vm_start
)
1261 nstart
= vma
->vm_start
;
1263 * Now fault in a range of pages. populate_vma_page_range()
1264 * double checks the vma flags, so that it won't mlock pages
1265 * if the vma was already munlocked.
1267 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1269 if (ignore_errors
) {
1271 continue; /* continue at next VMA */
1275 nend
= nstart
+ ret
* PAGE_SIZE
;
1279 up_read(&mm
->mmap_sem
);
1280 return ret
; /* 0 or negative error code */
1284 * get_dump_page() - pin user page in memory while writing it to core dump
1285 * @addr: user address
1287 * Returns struct page pointer of user page pinned for dump,
1288 * to be freed afterwards by put_page().
1290 * Returns NULL on any kind of failure - a hole must then be inserted into
1291 * the corefile, to preserve alignment with its headers; and also returns
1292 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1293 * allowing a hole to be left in the corefile to save diskspace.
1295 * Called without mmap_sem, but after all other threads have been killed.
1297 #ifdef CONFIG_ELF_CORE
1298 struct page
*get_dump_page(unsigned long addr
)
1300 struct vm_area_struct
*vma
;
1303 if (__get_user_pages(current
, current
->mm
, addr
, 1,
1304 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
1307 flush_cache_page(vma
, addr
, page_to_pfn(page
));
1310 #endif /* CONFIG_ELF_CORE */
1315 * get_user_pages_fast attempts to pin user pages by walking the page
1316 * tables directly and avoids taking locks. Thus the walker needs to be
1317 * protected from page table pages being freed from under it, and should
1318 * block any THP splits.
1320 * One way to achieve this is to have the walker disable interrupts, and
1321 * rely on IPIs from the TLB flushing code blocking before the page table
1322 * pages are freed. This is unsuitable for architectures that do not need
1323 * to broadcast an IPI when invalidating TLBs.
1325 * Another way to achieve this is to batch up page table containing pages
1326 * belonging to more than one mm_user, then rcu_sched a callback to free those
1327 * pages. Disabling interrupts will allow the fast_gup walker to both block
1328 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1329 * (which is a relatively rare event). The code below adopts this strategy.
1331 * Before activating this code, please be aware that the following assumptions
1332 * are currently made:
1334 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1335 * free pages containing page tables or TLB flushing requires IPI broadcast.
1337 * *) ptes can be read atomically by the architecture.
1339 * *) access_ok is sufficient to validate userspace address ranges.
1341 * The last two assumptions can be relaxed by the addition of helper functions.
1343 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1345 #ifdef CONFIG_HAVE_GENERIC_GUP
1349 * We assume that the PTE can be read atomically. If this is not the case for
1350 * your architecture, please provide the helper.
1352 static inline pte_t
gup_get_pte(pte_t
*ptep
)
1354 return READ_ONCE(*ptep
);
1358 static void undo_dev_pagemap(int *nr
, int nr_start
, struct page
**pages
)
1360 while ((*nr
) - nr_start
) {
1361 struct page
*page
= pages
[--(*nr
)];
1363 ClearPageReferenced(page
);
1368 #ifdef __HAVE_ARCH_PTE_SPECIAL
1369 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1370 int write
, struct page
**pages
, int *nr
)
1372 struct dev_pagemap
*pgmap
= NULL
;
1373 int nr_start
= *nr
, ret
= 0;
1376 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
1378 pte_t pte
= gup_get_pte(ptep
);
1379 struct page
*head
, *page
;
1382 * Similar to the PMD case below, NUMA hinting must take slow
1383 * path using the pte_protnone check.
1385 if (pte_protnone(pte
))
1388 if (!pte_access_permitted(pte
, write
))
1391 if (pte_devmap(pte
)) {
1392 pgmap
= get_dev_pagemap(pte_pfn(pte
), pgmap
);
1393 if (unlikely(!pgmap
)) {
1394 undo_dev_pagemap(nr
, nr_start
, pages
);
1397 } else if (pte_special(pte
))
1400 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
1401 page
= pte_page(pte
);
1402 head
= compound_head(page
);
1404 if (!page_cache_get_speculative(head
))
1407 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
1412 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1414 put_dev_pagemap(pgmap
);
1415 SetPageReferenced(page
);
1419 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
1430 * If we can't determine whether or not a pte is special, then fail immediately
1431 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1434 * For a futex to be placed on a THP tail page, get_futex_key requires a
1435 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1436 * useful to have gup_huge_pmd even if we can't operate on ptes.
1438 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1439 int write
, struct page
**pages
, int *nr
)
1443 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1445 #if defined(__HAVE_ARCH_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1446 static int __gup_device_huge(unsigned long pfn
, unsigned long addr
,
1447 unsigned long end
, struct page
**pages
, int *nr
)
1450 struct dev_pagemap
*pgmap
= NULL
;
1453 struct page
*page
= pfn_to_page(pfn
);
1455 pgmap
= get_dev_pagemap(pfn
, pgmap
);
1456 if (unlikely(!pgmap
)) {
1457 undo_dev_pagemap(nr
, nr_start
, pages
);
1460 SetPageReferenced(page
);
1463 put_dev_pagemap(pgmap
);
1466 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1470 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1471 unsigned long end
, struct page
**pages
, int *nr
)
1473 unsigned long fault_pfn
;
1476 fault_pfn
= pmd_pfn(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
1477 if (!__gup_device_huge(fault_pfn
, addr
, end
, pages
, nr
))
1480 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
1481 undo_dev_pagemap(nr
, nr_start
, pages
);
1487 static int __gup_device_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
1488 unsigned long end
, struct page
**pages
, int *nr
)
1490 unsigned long fault_pfn
;
1493 fault_pfn
= pud_pfn(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
1494 if (!__gup_device_huge(fault_pfn
, addr
, end
, pages
, nr
))
1497 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
1498 undo_dev_pagemap(nr
, nr_start
, pages
);
1504 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1505 unsigned long end
, struct page
**pages
, int *nr
)
1511 static int __gup_device_huge_pud(pud_t pud
, pud_t
*pudp
, unsigned long addr
,
1512 unsigned long end
, struct page
**pages
, int *nr
)
1519 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1520 unsigned long end
, int write
, struct page
**pages
, int *nr
)
1522 struct page
*head
, *page
;
1525 if (!pmd_access_permitted(orig
, write
))
1528 if (pmd_devmap(orig
))
1529 return __gup_device_huge_pmd(orig
, pmdp
, addr
, end
, pages
, nr
);
1532 page
= pmd_page(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
1538 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1540 head
= compound_head(pmd_page(orig
));
1541 if (!page_cache_add_speculative(head
, refs
)) {
1546 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
1553 SetPageReferenced(head
);
1557 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
1558 unsigned long end
, int write
, struct page
**pages
, int *nr
)
1560 struct page
*head
, *page
;
1563 if (!pud_access_permitted(orig
, write
))
1566 if (pud_devmap(orig
))
1567 return __gup_device_huge_pud(orig
, pudp
, addr
, end
, pages
, nr
);
1570 page
= pud_page(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
1576 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1578 head
= compound_head(pud_page(orig
));
1579 if (!page_cache_add_speculative(head
, refs
)) {
1584 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
1591 SetPageReferenced(head
);
1595 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
1596 unsigned long end
, int write
,
1597 struct page
**pages
, int *nr
)
1600 struct page
*head
, *page
;
1602 if (!pgd_access_permitted(orig
, write
))
1605 BUILD_BUG_ON(pgd_devmap(orig
));
1607 page
= pgd_page(orig
) + ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
1613 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1615 head
= compound_head(pgd_page(orig
));
1616 if (!page_cache_add_speculative(head
, refs
)) {
1621 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
1628 SetPageReferenced(head
);
1632 static int gup_pmd_range(pud_t pud
, unsigned long addr
, unsigned long end
,
1633 int write
, struct page
**pages
, int *nr
)
1638 pmdp
= pmd_offset(&pud
, addr
);
1640 pmd_t pmd
= READ_ONCE(*pmdp
);
1642 next
= pmd_addr_end(addr
, end
);
1643 if (!pmd_present(pmd
))
1646 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
))) {
1648 * NUMA hinting faults need to be handled in the GUP
1649 * slowpath for accounting purposes and so that they
1650 * can be serialised against THP migration.
1652 if (pmd_protnone(pmd
))
1655 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, write
,
1659 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
1661 * architecture have different format for hugetlbfs
1662 * pmd format and THP pmd format
1664 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
1665 PMD_SHIFT
, next
, write
, pages
, nr
))
1667 } else if (!gup_pte_range(pmd
, addr
, next
, write
, pages
, nr
))
1669 } while (pmdp
++, addr
= next
, addr
!= end
);
1674 static int gup_pud_range(p4d_t p4d
, unsigned long addr
, unsigned long end
,
1675 int write
, struct page
**pages
, int *nr
)
1680 pudp
= pud_offset(&p4d
, addr
);
1682 pud_t pud
= READ_ONCE(*pudp
);
1684 next
= pud_addr_end(addr
, end
);
1687 if (unlikely(pud_huge(pud
))) {
1688 if (!gup_huge_pud(pud
, pudp
, addr
, next
, write
,
1691 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
1692 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
1693 PUD_SHIFT
, next
, write
, pages
, nr
))
1695 } else if (!gup_pmd_range(pud
, addr
, next
, write
, pages
, nr
))
1697 } while (pudp
++, addr
= next
, addr
!= end
);
1702 static int gup_p4d_range(pgd_t pgd
, unsigned long addr
, unsigned long end
,
1703 int write
, struct page
**pages
, int *nr
)
1708 p4dp
= p4d_offset(&pgd
, addr
);
1710 p4d_t p4d
= READ_ONCE(*p4dp
);
1712 next
= p4d_addr_end(addr
, end
);
1715 BUILD_BUG_ON(p4d_huge(p4d
));
1716 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d
))))) {
1717 if (!gup_huge_pd(__hugepd(p4d_val(p4d
)), addr
,
1718 P4D_SHIFT
, next
, write
, pages
, nr
))
1720 } else if (!gup_pud_range(p4d
, addr
, next
, write
, pages
, nr
))
1722 } while (p4dp
++, addr
= next
, addr
!= end
);
1727 static void gup_pgd_range(unsigned long addr
, unsigned long end
,
1728 int write
, struct page
**pages
, int *nr
)
1733 pgdp
= pgd_offset(current
->mm
, addr
);
1735 pgd_t pgd
= READ_ONCE(*pgdp
);
1737 next
= pgd_addr_end(addr
, end
);
1740 if (unlikely(pgd_huge(pgd
))) {
1741 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, write
,
1744 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
1745 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
1746 PGDIR_SHIFT
, next
, write
, pages
, nr
))
1748 } else if (!gup_p4d_range(pgd
, addr
, next
, write
, pages
, nr
))
1750 } while (pgdp
++, addr
= next
, addr
!= end
);
1753 #ifndef gup_fast_permitted
1755 * Check if it's allowed to use __get_user_pages_fast() for the range, or
1756 * we need to fall back to the slow version:
1758 bool gup_fast_permitted(unsigned long start
, int nr_pages
, int write
)
1760 unsigned long len
, end
;
1762 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
1764 return end
>= start
;
1769 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1770 * the regular GUP. It will only return non-negative values.
1772 int __get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
1773 struct page
**pages
)
1775 unsigned long addr
, len
, end
;
1776 unsigned long flags
;
1781 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
1784 if (unlikely(!access_ok(write
? VERIFY_WRITE
: VERIFY_READ
,
1785 (void __user
*)start
, len
)))
1789 * Disable interrupts. We use the nested form as we can already have
1790 * interrupts disabled by get_futex_key.
1792 * With interrupts disabled, we block page table pages from being
1793 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1796 * We do not adopt an rcu_read_lock(.) here as we also want to
1797 * block IPIs that come from THPs splitting.
1800 if (gup_fast_permitted(start
, nr_pages
, write
)) {
1801 local_irq_save(flags
);
1802 gup_pgd_range(addr
, end
, write
, pages
, &nr
);
1803 local_irq_restore(flags
);
1810 * get_user_pages_fast() - pin user pages in memory
1811 * @start: starting user address
1812 * @nr_pages: number of pages from start to pin
1813 * @write: whether pages will be written to
1814 * @pages: array that receives pointers to the pages pinned.
1815 * Should be at least nr_pages long.
1817 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1818 * If not successful, it will fall back to taking the lock and
1819 * calling get_user_pages().
1821 * Returns number of pages pinned. This may be fewer than the number
1822 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1823 * were pinned, returns -errno.
1825 int get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
1826 struct page
**pages
)
1828 unsigned long addr
, len
, end
;
1829 int nr
= 0, ret
= 0;
1833 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
1839 if (unlikely(!access_ok(write
? VERIFY_WRITE
: VERIFY_READ
,
1840 (void __user
*)start
, len
)))
1843 if (gup_fast_permitted(start
, nr_pages
, write
)) {
1844 local_irq_disable();
1845 gup_pgd_range(addr
, end
, write
, pages
, &nr
);
1850 if (nr
< nr_pages
) {
1851 /* Try to get the remaining pages with get_user_pages */
1852 start
+= nr
<< PAGE_SHIFT
;
1855 ret
= get_user_pages_unlocked(start
, nr_pages
- nr
, pages
,
1856 write
? FOLL_WRITE
: 0);
1858 /* Have to be a bit careful with return values */
1870 #endif /* CONFIG_HAVE_GENERIC_GUP */