2 // SPDX-License-Identifier: GPL-2.0-only
6 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
10 * demand-loading started 01.12.91 - seems it is high on the list of
11 * things wanted, and it should be easy to implement. - Linus
15 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
16 * pages started 02.12.91, seems to work. - Linus.
18 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
19 * would have taken more than the 6M I have free, but it worked well as
22 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
26 * Real VM (paging to/from disk) started 18.12.91. Much more work and
27 * thought has to go into this. Oh, well..
28 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
29 * Found it. Everything seems to work now.
30 * 20.12.91 - Ok, making the swap-device changeable like the root.
34 * 05.04.94 - Multi-page memory management added for v1.1.
35 * Idea by Alex Bligh (alex@cconcepts.co.uk)
37 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
38 * (Gerhard.Wichert@pdb.siemens.de)
40 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
43 #include <linux/kernel_stat.h>
45 #include <linux/mm_inline.h>
46 #include <linux/sched/mm.h>
47 #include <linux/sched/coredump.h>
48 #include <linux/sched/numa_balancing.h>
49 #include <linux/sched/task.h>
50 #include <linux/hugetlb.h>
51 #include <linux/mman.h>
52 #include <linux/swap.h>
53 #include <linux/highmem.h>
54 #include <linux/pagemap.h>
55 #include <linux/memremap.h>
56 #include <linux/kmsan.h>
57 #include <linux/ksm.h>
58 #include <linux/rmap.h>
59 #include <linux/export.h>
60 #include <linux/delayacct.h>
61 #include <linux/init.h>
62 #include <linux/pfn_t.h>
63 #include <linux/writeback.h>
64 #include <linux/memcontrol.h>
65 #include <linux/mmu_notifier.h>
66 #include <linux/swapops.h>
67 #include <linux/elf.h>
68 #include <linux/gfp.h>
69 #include <linux/migrate.h>
70 #include <linux/string.h>
71 #include <linux/memory-tiers.h>
72 #include <linux/debugfs.h>
73 #include <linux/userfaultfd_k.h>
74 #include <linux/dax.h>
75 #include <linux/oom.h>
76 #include <linux/numa.h>
77 #include <linux/perf_event.h>
78 #include <linux/ptrace.h>
79 #include <linux/vmalloc.h>
80 #include <linux/sched/sysctl.h>
82 #include <trace/events/kmem.h>
85 #include <asm/mmu_context.h>
86 #include <asm/pgalloc.h>
87 #include <linux/uaccess.h>
89 #include <asm/tlbflush.h>
91 #include "pgalloc-track.h"
95 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
96 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
100 unsigned long max_mapnr
;
101 EXPORT_SYMBOL(max_mapnr
);
103 struct page
*mem_map
;
104 EXPORT_SYMBOL(mem_map
);
107 static vm_fault_t
do_fault(struct vm_fault
*vmf
);
108 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
);
109 static bool vmf_pte_changed(struct vm_fault
*vmf
);
112 * Return true if the original pte was a uffd-wp pte marker (so the pte was
115 static __always_inline
bool vmf_orig_pte_uffd_wp(struct vm_fault
*vmf
)
117 if (!userfaultfd_wp(vmf
->vma
))
119 if (!(vmf
->flags
& FAULT_FLAG_ORIG_PTE_VALID
))
122 return pte_marker_uffd_wp(vmf
->orig_pte
);
126 * A number of key systems in x86 including ioremap() rely on the assumption
127 * that high_memory defines the upper bound on direct map memory, then end
131 EXPORT_SYMBOL(high_memory
);
134 * Randomize the address space (stacks, mmaps, brk, etc.).
136 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
137 * as ancient (libc5 based) binaries can segfault. )
139 int randomize_va_space __read_mostly
=
140 #ifdef CONFIG_COMPAT_BRK
146 #ifndef arch_wants_old_prefaulted_pte
147 static inline bool arch_wants_old_prefaulted_pte(void)
150 * Transitioning a PTE from 'old' to 'young' can be expensive on
151 * some architectures, even if it's performed in hardware. By
152 * default, "false" means prefaulted entries will be 'young'.
158 static int __init
disable_randmaps(char *s
)
160 randomize_va_space
= 0;
163 __setup("norandmaps", disable_randmaps
);
165 unsigned long zero_pfn __read_mostly
;
166 EXPORT_SYMBOL(zero_pfn
);
168 unsigned long highest_memmap_pfn __read_mostly
;
171 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
173 static int __init
init_zero_pfn(void)
175 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
178 early_initcall(init_zero_pfn
);
180 void mm_trace_rss_stat(struct mm_struct
*mm
, int member
)
182 trace_rss_stat(mm
, member
);
186 * Note: this doesn't free the actual pages themselves. That
187 * has been handled earlier when unmapping all the memory regions.
189 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
192 pgtable_t token
= pmd_pgtable(*pmd
);
194 pte_free_tlb(tlb
, token
, addr
);
195 mm_dec_nr_ptes(tlb
->mm
);
198 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
199 unsigned long addr
, unsigned long end
,
200 unsigned long floor
, unsigned long ceiling
)
207 pmd
= pmd_offset(pud
, addr
);
209 next
= pmd_addr_end(addr
, end
);
210 if (pmd_none_or_clear_bad(pmd
))
212 free_pte_range(tlb
, pmd
, addr
);
213 } while (pmd
++, addr
= next
, addr
!= end
);
223 if (end
- 1 > ceiling
- 1)
226 pmd
= pmd_offset(pud
, start
);
228 pmd_free_tlb(tlb
, pmd
, start
);
229 mm_dec_nr_pmds(tlb
->mm
);
232 static inline void free_pud_range(struct mmu_gather
*tlb
, p4d_t
*p4d
,
233 unsigned long addr
, unsigned long end
,
234 unsigned long floor
, unsigned long ceiling
)
241 pud
= pud_offset(p4d
, addr
);
243 next
= pud_addr_end(addr
, end
);
244 if (pud_none_or_clear_bad(pud
))
246 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
247 } while (pud
++, addr
= next
, addr
!= end
);
257 if (end
- 1 > ceiling
- 1)
260 pud
= pud_offset(p4d
, start
);
262 pud_free_tlb(tlb
, pud
, start
);
263 mm_dec_nr_puds(tlb
->mm
);
266 static inline void free_p4d_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
267 unsigned long addr
, unsigned long end
,
268 unsigned long floor
, unsigned long ceiling
)
275 p4d
= p4d_offset(pgd
, addr
);
277 next
= p4d_addr_end(addr
, end
);
278 if (p4d_none_or_clear_bad(p4d
))
280 free_pud_range(tlb
, p4d
, addr
, next
, floor
, ceiling
);
281 } while (p4d
++, addr
= next
, addr
!= end
);
287 ceiling
&= PGDIR_MASK
;
291 if (end
- 1 > ceiling
- 1)
294 p4d
= p4d_offset(pgd
, start
);
296 p4d_free_tlb(tlb
, p4d
, start
);
300 * This function frees user-level page tables of a process.
302 void free_pgd_range(struct mmu_gather
*tlb
,
303 unsigned long addr
, unsigned long end
,
304 unsigned long floor
, unsigned long ceiling
)
310 * The next few lines have given us lots of grief...
312 * Why are we testing PMD* at this top level? Because often
313 * there will be no work to do at all, and we'd prefer not to
314 * go all the way down to the bottom just to discover that.
316 * Why all these "- 1"s? Because 0 represents both the bottom
317 * of the address space and the top of it (using -1 for the
318 * top wouldn't help much: the masks would do the wrong thing).
319 * The rule is that addr 0 and floor 0 refer to the bottom of
320 * the address space, but end 0 and ceiling 0 refer to the top
321 * Comparisons need to use "end - 1" and "ceiling - 1" (though
322 * that end 0 case should be mythical).
324 * Wherever addr is brought up or ceiling brought down, we must
325 * be careful to reject "the opposite 0" before it confuses the
326 * subsequent tests. But what about where end is brought down
327 * by PMD_SIZE below? no, end can't go down to 0 there.
329 * Whereas we round start (addr) and ceiling down, by different
330 * masks at different levels, in order to test whether a table
331 * now has no other vmas using it, so can be freed, we don't
332 * bother to round floor or end up - the tests don't need that.
346 if (end
- 1 > ceiling
- 1)
351 * We add page table cache pages with PAGE_SIZE,
352 * (see pte_free_tlb()), flush the tlb if we need
354 tlb_change_page_size(tlb
, PAGE_SIZE
);
355 pgd
= pgd_offset(tlb
->mm
, addr
);
357 next
= pgd_addr_end(addr
, end
);
358 if (pgd_none_or_clear_bad(pgd
))
360 free_p4d_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
361 } while (pgd
++, addr
= next
, addr
!= end
);
364 void free_pgtables(struct mmu_gather
*tlb
, struct ma_state
*mas
,
365 struct vm_area_struct
*vma
, unsigned long floor
,
366 unsigned long ceiling
, bool mm_wr_locked
)
368 struct unlink_vma_file_batch vb
;
371 unsigned long addr
= vma
->vm_start
;
372 struct vm_area_struct
*next
;
375 * Note: USER_PGTABLES_CEILING may be passed as ceiling and may
376 * be 0. This will underflow and is okay.
378 next
= mas_find(mas
, ceiling
- 1);
379 if (unlikely(xa_is_zero(next
)))
383 * Hide vma from rmap and truncate_pagecache before freeing
387 vma_start_write(vma
);
388 unlink_anon_vmas(vma
);
390 if (is_vm_hugetlb_page(vma
)) {
391 unlink_file_vma(vma
);
392 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
393 floor
, next
? next
->vm_start
: ceiling
);
395 unlink_file_vma_batch_init(&vb
);
396 unlink_file_vma_batch_add(&vb
, vma
);
399 * Optimization: gather nearby vmas into one call down
401 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
402 && !is_vm_hugetlb_page(next
)) {
404 next
= mas_find(mas
, ceiling
- 1);
405 if (unlikely(xa_is_zero(next
)))
408 vma_start_write(vma
);
409 unlink_anon_vmas(vma
);
410 unlink_file_vma_batch_add(&vb
, vma
);
412 unlink_file_vma_batch_final(&vb
);
413 free_pgd_range(tlb
, addr
, vma
->vm_end
,
414 floor
, next
? next
->vm_start
: ceiling
);
420 void pmd_install(struct mm_struct
*mm
, pmd_t
*pmd
, pgtable_t
*pte
)
422 spinlock_t
*ptl
= pmd_lock(mm
, pmd
);
424 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
427 * Ensure all pte setup (eg. pte page lock and page clearing) are
428 * visible before the pte is made visible to other CPUs by being
429 * put into page tables.
431 * The other side of the story is the pointer chasing in the page
432 * table walking code (when walking the page table without locking;
433 * ie. most of the time). Fortunately, these data accesses consist
434 * of a chain of data-dependent loads, meaning most CPUs (alpha
435 * being the notable exception) will already guarantee loads are
436 * seen in-order. See the alpha page table accessors for the
437 * smp_rmb() barriers in page table walking code.
439 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
440 pmd_populate(mm
, pmd
, *pte
);
446 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
)
448 pgtable_t
new = pte_alloc_one(mm
);
452 pmd_install(mm
, pmd
, &new);
458 int __pte_alloc_kernel(pmd_t
*pmd
)
460 pte_t
*new = pte_alloc_one_kernel(&init_mm
);
464 spin_lock(&init_mm
.page_table_lock
);
465 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
466 smp_wmb(); /* See comment in pmd_install() */
467 pmd_populate_kernel(&init_mm
, pmd
, new);
470 spin_unlock(&init_mm
.page_table_lock
);
472 pte_free_kernel(&init_mm
, new);
476 static inline void init_rss_vec(int *rss
)
478 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
481 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
485 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
487 add_mm_counter(mm
, i
, rss
[i
]);
491 * This function is called to print an error when a bad pte
492 * is found. For example, we might have a PFN-mapped pte in
493 * a region that doesn't allow it.
495 * The calling function must still handle the error.
497 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
498 pte_t pte
, struct page
*page
)
500 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
501 p4d_t
*p4d
= p4d_offset(pgd
, addr
);
502 pud_t
*pud
= pud_offset(p4d
, addr
);
503 pmd_t
*pmd
= pmd_offset(pud
, addr
);
504 struct address_space
*mapping
;
506 static unsigned long resume
;
507 static unsigned long nr_shown
;
508 static unsigned long nr_unshown
;
511 * Allow a burst of 60 reports, then keep quiet for that minute;
512 * or allow a steady drip of one report per second.
514 if (nr_shown
== 60) {
515 if (time_before(jiffies
, resume
)) {
520 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
527 resume
= jiffies
+ 60 * HZ
;
529 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
530 index
= linear_page_index(vma
, addr
);
532 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
534 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
536 dump_page(page
, "bad pte");
537 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
538 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
539 pr_alert("file:%pD fault:%ps mmap:%ps read_folio:%ps\n",
541 vma
->vm_ops
? vma
->vm_ops
->fault
: NULL
,
542 vma
->vm_file
? vma
->vm_file
->f_op
->mmap
: NULL
,
543 mapping
? mapping
->a_ops
->read_folio
: NULL
);
545 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
549 * vm_normal_page -- This function gets the "struct page" associated with a pte.
551 * "Special" mappings do not wish to be associated with a "struct page" (either
552 * it doesn't exist, or it exists but they don't want to touch it). In this
553 * case, NULL is returned here. "Normal" mappings do have a struct page.
555 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
556 * pte bit, in which case this function is trivial. Secondly, an architecture
557 * may not have a spare pte bit, which requires a more complicated scheme,
560 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
561 * special mapping (even if there are underlying and valid "struct pages").
562 * COWed pages of a VM_PFNMAP are always normal.
564 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
565 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
566 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
567 * mapping will always honor the rule
569 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
571 * And for normal mappings this is false.
573 * This restricts such mappings to be a linear translation from virtual address
574 * to pfn. To get around this restriction, we allow arbitrary mappings so long
575 * as the vma is not a COW mapping; in that case, we know that all ptes are
576 * special (because none can have been COWed).
579 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
581 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
582 * page" backing, however the difference is that _all_ pages with a struct
583 * page (that is, those where pfn_valid is true) are refcounted and considered
584 * normal pages by the VM. The only exception are zeropages, which are
585 * *never* refcounted.
587 * The disadvantage is that pages are refcounted (which can be slower and
588 * simply not an option for some PFNMAP users). The advantage is that we
589 * don't have to follow the strict linearity rule of PFNMAP mappings in
590 * order to support COWable mappings.
593 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
596 unsigned long pfn
= pte_pfn(pte
);
598 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
)) {
599 if (likely(!pte_special(pte
)))
601 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
602 return vma
->vm_ops
->find_special_page(vma
, addr
);
603 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
605 if (is_zero_pfn(pfn
))
609 * NOTE: New users of ZONE_DEVICE will not set pte_devmap()
610 * and will have refcounts incremented on their struct pages
611 * when they are inserted into PTEs, thus they are safe to
612 * return here. Legacy ZONE_DEVICE pages that set pte_devmap()
613 * do not have refcounts. Example of legacy ZONE_DEVICE is
614 * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers.
618 print_bad_pte(vma
, addr
, pte
, NULL
);
622 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
624 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
625 if (vma
->vm_flags
& VM_MIXEDMAP
) {
628 if (is_zero_pfn(pfn
))
633 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
634 if (pfn
== vma
->vm_pgoff
+ off
)
636 if (!is_cow_mapping(vma
->vm_flags
))
641 if (is_zero_pfn(pfn
))
645 if (unlikely(pfn
> highest_memmap_pfn
)) {
646 print_bad_pte(vma
, addr
, pte
, NULL
);
651 * NOTE! We still have PageReserved() pages in the page tables.
652 * eg. VDSO mappings can cause them to exist.
655 VM_WARN_ON_ONCE(is_zero_pfn(pfn
));
656 return pfn_to_page(pfn
);
659 struct folio
*vm_normal_folio(struct vm_area_struct
*vma
, unsigned long addr
,
662 struct page
*page
= vm_normal_page(vma
, addr
, pte
);
665 return page_folio(page
);
669 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
670 struct page
*vm_normal_page_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
673 unsigned long pfn
= pmd_pfn(pmd
);
676 * There is no pmd_special() but there may be special pmds, e.g.
677 * in a direct-access (dax) mapping, so let's just replicate the
678 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
680 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
681 if (vma
->vm_flags
& VM_MIXEDMAP
) {
687 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
688 if (pfn
== vma
->vm_pgoff
+ off
)
690 if (!is_cow_mapping(vma
->vm_flags
))
697 if (is_huge_zero_pmd(pmd
))
699 if (unlikely(pfn
> highest_memmap_pfn
))
703 * NOTE! We still have PageReserved() pages in the page tables.
704 * eg. VDSO mappings can cause them to exist.
707 return pfn_to_page(pfn
);
710 struct folio
*vm_normal_folio_pmd(struct vm_area_struct
*vma
,
711 unsigned long addr
, pmd_t pmd
)
713 struct page
*page
= vm_normal_page_pmd(vma
, addr
, pmd
);
716 return page_folio(page
);
721 static void restore_exclusive_pte(struct vm_area_struct
*vma
,
722 struct page
*page
, unsigned long address
,
725 struct folio
*folio
= page_folio(page
);
730 orig_pte
= ptep_get(ptep
);
731 pte
= pte_mkold(mk_pte(page
, READ_ONCE(vma
->vm_page_prot
)));
732 if (pte_swp_soft_dirty(orig_pte
))
733 pte
= pte_mksoft_dirty(pte
);
735 entry
= pte_to_swp_entry(orig_pte
);
736 if (pte_swp_uffd_wp(orig_pte
))
737 pte
= pte_mkuffd_wp(pte
);
738 else if (is_writable_device_exclusive_entry(entry
))
739 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
741 VM_BUG_ON_FOLIO(pte_write(pte
) && (!folio_test_anon(folio
) &&
742 PageAnonExclusive(page
)), folio
);
745 * No need to take a page reference as one was already
746 * created when the swap entry was made.
748 if (folio_test_anon(folio
))
749 folio_add_anon_rmap_pte(folio
, page
, vma
, address
, RMAP_NONE
);
752 * Currently device exclusive access only supports anonymous
753 * memory so the entry shouldn't point to a filebacked page.
757 set_pte_at(vma
->vm_mm
, address
, ptep
, pte
);
760 * No need to invalidate - it was non-present before. However
761 * secondary CPUs may have mappings that need invalidating.
763 update_mmu_cache(vma
, address
, ptep
);
767 * Tries to restore an exclusive pte if the page lock can be acquired without
771 try_restore_exclusive_pte(pte_t
*src_pte
, struct vm_area_struct
*vma
,
774 swp_entry_t entry
= pte_to_swp_entry(ptep_get(src_pte
));
775 struct page
*page
= pfn_swap_entry_to_page(entry
);
777 if (trylock_page(page
)) {
778 restore_exclusive_pte(vma
, page
, addr
, src_pte
);
787 * copy one vm_area from one task to the other. Assumes the page tables
788 * already present in the new task to be cleared in the whole range
789 * covered by this vma.
793 copy_nonpresent_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
794 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*dst_vma
,
795 struct vm_area_struct
*src_vma
, unsigned long addr
, int *rss
)
797 unsigned long vm_flags
= dst_vma
->vm_flags
;
798 pte_t orig_pte
= ptep_get(src_pte
);
799 pte_t pte
= orig_pte
;
802 swp_entry_t entry
= pte_to_swp_entry(orig_pte
);
804 if (likely(!non_swap_entry(entry
))) {
805 if (swap_duplicate(entry
) < 0)
808 /* make sure dst_mm is on swapoff's mmlist. */
809 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
810 spin_lock(&mmlist_lock
);
811 if (list_empty(&dst_mm
->mmlist
))
812 list_add(&dst_mm
->mmlist
,
814 spin_unlock(&mmlist_lock
);
816 /* Mark the swap entry as shared. */
817 if (pte_swp_exclusive(orig_pte
)) {
818 pte
= pte_swp_clear_exclusive(orig_pte
);
819 set_pte_at(src_mm
, addr
, src_pte
, pte
);
822 } else if (is_migration_entry(entry
)) {
823 folio
= pfn_swap_entry_folio(entry
);
825 rss
[mm_counter(folio
)]++;
827 if (!is_readable_migration_entry(entry
) &&
828 is_cow_mapping(vm_flags
)) {
830 * COW mappings require pages in both parent and child
831 * to be set to read. A previously exclusive entry is
834 entry
= make_readable_migration_entry(
836 pte
= swp_entry_to_pte(entry
);
837 if (pte_swp_soft_dirty(orig_pte
))
838 pte
= pte_swp_mksoft_dirty(pte
);
839 if (pte_swp_uffd_wp(orig_pte
))
840 pte
= pte_swp_mkuffd_wp(pte
);
841 set_pte_at(src_mm
, addr
, src_pte
, pte
);
843 } else if (is_device_private_entry(entry
)) {
844 page
= pfn_swap_entry_to_page(entry
);
845 folio
= page_folio(page
);
848 * Update rss count even for unaddressable pages, as
849 * they should treated just like normal pages in this
852 * We will likely want to have some new rss counters
853 * for unaddressable pages, at some point. But for now
854 * keep things as they are.
857 rss
[mm_counter(folio
)]++;
858 /* Cannot fail as these pages cannot get pinned. */
859 folio_try_dup_anon_rmap_pte(folio
, page
, src_vma
);
862 * We do not preserve soft-dirty information, because so
863 * far, checkpoint/restore is the only feature that
864 * requires that. And checkpoint/restore does not work
865 * when a device driver is involved (you cannot easily
866 * save and restore device driver state).
868 if (is_writable_device_private_entry(entry
) &&
869 is_cow_mapping(vm_flags
)) {
870 entry
= make_readable_device_private_entry(
872 pte
= swp_entry_to_pte(entry
);
873 if (pte_swp_uffd_wp(orig_pte
))
874 pte
= pte_swp_mkuffd_wp(pte
);
875 set_pte_at(src_mm
, addr
, src_pte
, pte
);
877 } else if (is_device_exclusive_entry(entry
)) {
879 * Make device exclusive entries present by restoring the
880 * original entry then copying as for a present pte. Device
881 * exclusive entries currently only support private writable
882 * (ie. COW) mappings.
884 VM_BUG_ON(!is_cow_mapping(src_vma
->vm_flags
));
885 if (try_restore_exclusive_pte(src_pte
, src_vma
, addr
))
888 } else if (is_pte_marker_entry(entry
)) {
889 pte_marker marker
= copy_pte_marker(entry
, dst_vma
);
892 set_pte_at(dst_mm
, addr
, dst_pte
,
893 make_pte_marker(marker
));
896 if (!userfaultfd_wp(dst_vma
))
897 pte
= pte_swp_clear_uffd_wp(pte
);
898 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
903 * Copy a present and normal page.
905 * NOTE! The usual case is that this isn't required;
906 * instead, the caller can just increase the page refcount
907 * and re-use the pte the traditional way.
909 * And if we need a pre-allocated page but don't yet have
910 * one, return a negative error to let the preallocation
911 * code know so that it can do so outside the page table
915 copy_present_page(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
916 pte_t
*dst_pte
, pte_t
*src_pte
, unsigned long addr
, int *rss
,
917 struct folio
**prealloc
, struct page
*page
)
919 struct folio
*new_folio
;
922 new_folio
= *prealloc
;
927 * We have a prealloc page, all good! Take it
928 * over and copy the page & arm it.
931 copy_user_highpage(&new_folio
->page
, page
, addr
, src_vma
);
932 __folio_mark_uptodate(new_folio
);
933 folio_add_new_anon_rmap(new_folio
, dst_vma
, addr
, RMAP_EXCLUSIVE
);
934 folio_add_lru_vma(new_folio
, dst_vma
);
937 /* All done, just insert the new page copy in the child */
938 pte
= mk_pte(&new_folio
->page
, dst_vma
->vm_page_prot
);
939 pte
= maybe_mkwrite(pte_mkdirty(pte
), dst_vma
);
940 if (userfaultfd_pte_wp(dst_vma
, ptep_get(src_pte
)))
941 /* Uffd-wp needs to be delivered to dest pte as well */
942 pte
= pte_mkuffd_wp(pte
);
943 set_pte_at(dst_vma
->vm_mm
, addr
, dst_pte
, pte
);
947 static __always_inline
void __copy_present_ptes(struct vm_area_struct
*dst_vma
,
948 struct vm_area_struct
*src_vma
, pte_t
*dst_pte
, pte_t
*src_pte
,
949 pte_t pte
, unsigned long addr
, int nr
)
951 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
953 /* If it's a COW mapping, write protect it both processes. */
954 if (is_cow_mapping(src_vma
->vm_flags
) && pte_write(pte
)) {
955 wrprotect_ptes(src_mm
, addr
, src_pte
, nr
);
956 pte
= pte_wrprotect(pte
);
959 /* If it's a shared mapping, mark it clean in the child. */
960 if (src_vma
->vm_flags
& VM_SHARED
)
961 pte
= pte_mkclean(pte
);
962 pte
= pte_mkold(pte
);
964 if (!userfaultfd_wp(dst_vma
))
965 pte
= pte_clear_uffd_wp(pte
);
967 set_ptes(dst_vma
->vm_mm
, addr
, dst_pte
, pte
, nr
);
971 * Copy one present PTE, trying to batch-process subsequent PTEs that map
972 * consecutive pages of the same folio by copying them as well.
974 * Returns -EAGAIN if one preallocated page is required to copy the next PTE.
975 * Otherwise, returns the number of copied PTEs (at least 1).
978 copy_present_ptes(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
979 pte_t
*dst_pte
, pte_t
*src_pte
, pte_t pte
, unsigned long addr
,
980 int max_nr
, int *rss
, struct folio
**prealloc
)
988 page
= vm_normal_page(src_vma
, addr
, pte
);
992 folio
= page_folio(page
);
995 * If we likely have to copy, just don't bother with batching. Make
996 * sure that the common "small folio" case is as fast as possible
997 * by keeping the batching logic separate.
999 if (unlikely(!*prealloc
&& folio_test_large(folio
) && max_nr
!= 1)) {
1000 if (src_vma
->vm_flags
& VM_SHARED
)
1001 flags
|= FPB_IGNORE_DIRTY
;
1002 if (!vma_soft_dirty_enabled(src_vma
))
1003 flags
|= FPB_IGNORE_SOFT_DIRTY
;
1005 nr
= folio_pte_batch(folio
, addr
, src_pte
, pte
, max_nr
, flags
,
1006 &any_writable
, NULL
, NULL
);
1007 folio_ref_add(folio
, nr
);
1008 if (folio_test_anon(folio
)) {
1009 if (unlikely(folio_try_dup_anon_rmap_ptes(folio
, page
,
1011 folio_ref_sub(folio
, nr
);
1014 rss
[MM_ANONPAGES
] += nr
;
1015 VM_WARN_ON_FOLIO(PageAnonExclusive(page
), folio
);
1017 folio_dup_file_rmap_ptes(folio
, page
, nr
);
1018 rss
[mm_counter_file(folio
)] += nr
;
1021 pte
= pte_mkwrite(pte
, src_vma
);
1022 __copy_present_ptes(dst_vma
, src_vma
, dst_pte
, src_pte
, pte
,
1028 if (folio_test_anon(folio
)) {
1030 * If this page may have been pinned by the parent process,
1031 * copy the page immediately for the child so that we'll always
1032 * guarantee the pinned page won't be randomly replaced in the
1035 if (unlikely(folio_try_dup_anon_rmap_pte(folio
, page
, src_vma
))) {
1036 /* Page may be pinned, we have to copy. */
1038 err
= copy_present_page(dst_vma
, src_vma
, dst_pte
, src_pte
,
1039 addr
, rss
, prealloc
, page
);
1040 return err
? err
: 1;
1042 rss
[MM_ANONPAGES
]++;
1043 VM_WARN_ON_FOLIO(PageAnonExclusive(page
), folio
);
1045 folio_dup_file_rmap_pte(folio
, page
);
1046 rss
[mm_counter_file(folio
)]++;
1050 __copy_present_ptes(dst_vma
, src_vma
, dst_pte
, src_pte
, pte
, addr
, 1);
1054 static inline struct folio
*folio_prealloc(struct mm_struct
*src_mm
,
1055 struct vm_area_struct
*vma
, unsigned long addr
, bool need_zero
)
1057 struct folio
*new_folio
;
1060 new_folio
= vma_alloc_zeroed_movable_folio(vma
, addr
);
1062 new_folio
= vma_alloc_folio(GFP_HIGHUSER_MOVABLE
, 0, vma
,
1068 if (mem_cgroup_charge(new_folio
, src_mm
, GFP_KERNEL
)) {
1069 folio_put(new_folio
);
1072 folio_throttle_swaprate(new_folio
, GFP_KERNEL
);
1078 copy_pte_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1079 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
1082 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1083 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1084 pte_t
*orig_src_pte
, *orig_dst_pte
;
1085 pte_t
*src_pte
, *dst_pte
;
1087 spinlock_t
*src_ptl
, *dst_ptl
;
1088 int progress
, max_nr
, ret
= 0;
1089 int rss
[NR_MM_COUNTERS
];
1090 swp_entry_t entry
= (swp_entry_t
){0};
1091 struct folio
*prealloc
= NULL
;
1099 * copy_pmd_range()'s prior pmd_none_or_clear_bad(src_pmd), and the
1100 * error handling here, assume that exclusive mmap_lock on dst and src
1101 * protects anon from unexpected THP transitions; with shmem and file
1102 * protected by mmap_lock-less collapse skipping areas with anon_vma
1103 * (whereas vma_needs_copy() skips areas without anon_vma). A rework
1104 * can remove such assumptions later, but this is good enough for now.
1106 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
1111 src_pte
= pte_offset_map_nolock(src_mm
, src_pmd
, addr
, &src_ptl
);
1113 pte_unmap_unlock(dst_pte
, dst_ptl
);
1117 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1118 orig_src_pte
= src_pte
;
1119 orig_dst_pte
= dst_pte
;
1120 arch_enter_lazy_mmu_mode();
1126 * We are holding two locks at this point - either of them
1127 * could generate latencies in another task on another CPU.
1129 if (progress
>= 32) {
1131 if (need_resched() ||
1132 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
1135 ptent
= ptep_get(src_pte
);
1136 if (pte_none(ptent
)) {
1140 if (unlikely(!pte_present(ptent
))) {
1141 ret
= copy_nonpresent_pte(dst_mm
, src_mm
,
1146 entry
= pte_to_swp_entry(ptep_get(src_pte
));
1148 } else if (ret
== -EBUSY
) {
1154 ptent
= ptep_get(src_pte
);
1155 VM_WARN_ON_ONCE(!pte_present(ptent
));
1158 * Device exclusive entry restored, continue by copying
1159 * the now present pte.
1161 WARN_ON_ONCE(ret
!= -ENOENT
);
1163 /* copy_present_ptes() will clear `*prealloc' if consumed */
1164 max_nr
= (end
- addr
) / PAGE_SIZE
;
1165 ret
= copy_present_ptes(dst_vma
, src_vma
, dst_pte
, src_pte
,
1166 ptent
, addr
, max_nr
, rss
, &prealloc
);
1168 * If we need a pre-allocated page for this pte, drop the
1169 * locks, allocate, and try again.
1171 if (unlikely(ret
== -EAGAIN
))
1173 if (unlikely(prealloc
)) {
1175 * pre-alloc page cannot be reused by next time so as
1176 * to strictly follow mempolicy (e.g., alloc_page_vma()
1177 * will allocate page according to address). This
1178 * could only happen if one pinned pte changed.
1180 folio_put(prealloc
);
1185 } while (dst_pte
+= nr
, src_pte
+= nr
, addr
+= PAGE_SIZE
* nr
,
1188 arch_leave_lazy_mmu_mode();
1189 pte_unmap_unlock(orig_src_pte
, src_ptl
);
1190 add_mm_rss_vec(dst_mm
, rss
);
1191 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
1195 VM_WARN_ON_ONCE(!entry
.val
);
1196 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0) {
1201 } else if (ret
== -EBUSY
) {
1203 } else if (ret
== -EAGAIN
) {
1204 prealloc
= folio_prealloc(src_mm
, src_vma
, addr
, false);
1207 } else if (ret
< 0) {
1211 /* We've captured and resolved the error. Reset, try again. */
1217 if (unlikely(prealloc
))
1218 folio_put(prealloc
);
1223 copy_pmd_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1224 pud_t
*dst_pud
, pud_t
*src_pud
, unsigned long addr
,
1227 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1228 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1229 pmd_t
*src_pmd
, *dst_pmd
;
1232 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
1235 src_pmd
= pmd_offset(src_pud
, addr
);
1237 next
= pmd_addr_end(addr
, end
);
1238 if (is_swap_pmd(*src_pmd
) || pmd_trans_huge(*src_pmd
)
1239 || pmd_devmap(*src_pmd
)) {
1241 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PMD_SIZE
, src_vma
);
1242 err
= copy_huge_pmd(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
1243 addr
, dst_vma
, src_vma
);
1250 if (pmd_none_or_clear_bad(src_pmd
))
1252 if (copy_pte_range(dst_vma
, src_vma
, dst_pmd
, src_pmd
,
1255 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1260 copy_pud_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1261 p4d_t
*dst_p4d
, p4d_t
*src_p4d
, unsigned long addr
,
1264 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1265 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1266 pud_t
*src_pud
, *dst_pud
;
1269 dst_pud
= pud_alloc(dst_mm
, dst_p4d
, addr
);
1272 src_pud
= pud_offset(src_p4d
, addr
);
1274 next
= pud_addr_end(addr
, end
);
1275 if (pud_trans_huge(*src_pud
) || pud_devmap(*src_pud
)) {
1278 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PUD_SIZE
, src_vma
);
1279 err
= copy_huge_pud(dst_mm
, src_mm
,
1280 dst_pud
, src_pud
, addr
, src_vma
);
1287 if (pud_none_or_clear_bad(src_pud
))
1289 if (copy_pmd_range(dst_vma
, src_vma
, dst_pud
, src_pud
,
1292 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1297 copy_p4d_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1298 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, unsigned long addr
,
1301 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1302 p4d_t
*src_p4d
, *dst_p4d
;
1305 dst_p4d
= p4d_alloc(dst_mm
, dst_pgd
, addr
);
1308 src_p4d
= p4d_offset(src_pgd
, addr
);
1310 next
= p4d_addr_end(addr
, end
);
1311 if (p4d_none_or_clear_bad(src_p4d
))
1313 if (copy_pud_range(dst_vma
, src_vma
, dst_p4d
, src_p4d
,
1316 } while (dst_p4d
++, src_p4d
++, addr
= next
, addr
!= end
);
1321 * Return true if the vma needs to copy the pgtable during this fork(). Return
1322 * false when we can speed up fork() by allowing lazy page faults later until
1323 * when the child accesses the memory range.
1326 vma_needs_copy(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
)
1329 * Always copy pgtables when dst_vma has uffd-wp enabled even if it's
1330 * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable
1331 * contains uffd-wp protection information, that's something we can't
1332 * retrieve from page cache, and skip copying will lose those info.
1334 if (userfaultfd_wp(dst_vma
))
1337 if (src_vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
1340 if (src_vma
->anon_vma
)
1344 * Don't copy ptes where a page fault will fill them correctly. Fork
1345 * becomes much lighter when there are big shared or private readonly
1346 * mappings. The tradeoff is that copy_page_range is more efficient
1353 copy_page_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
)
1355 pgd_t
*src_pgd
, *dst_pgd
;
1357 unsigned long addr
= src_vma
->vm_start
;
1358 unsigned long end
= src_vma
->vm_end
;
1359 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1360 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1361 struct mmu_notifier_range range
;
1365 if (!vma_needs_copy(dst_vma
, src_vma
))
1368 if (is_vm_hugetlb_page(src_vma
))
1369 return copy_hugetlb_page_range(dst_mm
, src_mm
, dst_vma
, src_vma
);
1371 if (unlikely(src_vma
->vm_flags
& VM_PFNMAP
)) {
1373 * We do not free on error cases below as remove_vma
1374 * gets called on error from higher level routine
1376 ret
= track_pfn_copy(src_vma
);
1382 * We need to invalidate the secondary MMU mappings only when
1383 * there could be a permission downgrade on the ptes of the
1384 * parent mm. And a permission downgrade will only happen if
1385 * is_cow_mapping() returns true.
1387 is_cow
= is_cow_mapping(src_vma
->vm_flags
);
1390 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_PAGE
,
1391 0, src_mm
, addr
, end
);
1392 mmu_notifier_invalidate_range_start(&range
);
1394 * Disabling preemption is not needed for the write side, as
1395 * the read side doesn't spin, but goes to the mmap_lock.
1397 * Use the raw variant of the seqcount_t write API to avoid
1398 * lockdep complaining about preemptibility.
1400 vma_assert_write_locked(src_vma
);
1401 raw_write_seqcount_begin(&src_mm
->write_protect_seq
);
1405 dst_pgd
= pgd_offset(dst_mm
, addr
);
1406 src_pgd
= pgd_offset(src_mm
, addr
);
1408 next
= pgd_addr_end(addr
, end
);
1409 if (pgd_none_or_clear_bad(src_pgd
))
1411 if (unlikely(copy_p4d_range(dst_vma
, src_vma
, dst_pgd
, src_pgd
,
1413 untrack_pfn_clear(dst_vma
);
1417 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1420 raw_write_seqcount_end(&src_mm
->write_protect_seq
);
1421 mmu_notifier_invalidate_range_end(&range
);
1426 /* Whether we should zap all COWed (private) pages too */
1427 static inline bool should_zap_cows(struct zap_details
*details
)
1429 /* By default, zap all pages */
1433 /* Or, we zap COWed pages only if the caller wants to */
1434 return details
->even_cows
;
1437 /* Decides whether we should zap this folio with the folio pointer specified */
1438 static inline bool should_zap_folio(struct zap_details
*details
,
1439 struct folio
*folio
)
1441 /* If we can make a decision without *folio.. */
1442 if (should_zap_cows(details
))
1445 /* Otherwise we should only zap non-anon folios */
1446 return !folio_test_anon(folio
);
1449 static inline bool zap_drop_file_uffd_wp(struct zap_details
*details
)
1454 return details
->zap_flags
& ZAP_FLAG_DROP_MARKER
;
1458 * This function makes sure that we'll replace the none pte with an uffd-wp
1459 * swap special pte marker when necessary. Must be with the pgtable lock held.
1462 zap_install_uffd_wp_if_needed(struct vm_area_struct
*vma
,
1463 unsigned long addr
, pte_t
*pte
, int nr
,
1464 struct zap_details
*details
, pte_t pteval
)
1466 /* Zap on anonymous always means dropping everything */
1467 if (vma_is_anonymous(vma
))
1470 if (zap_drop_file_uffd_wp(details
))
1474 /* the PFN in the PTE is irrelevant. */
1475 pte_install_uffd_wp_if_needed(vma
, addr
, pte
, pteval
);
1483 static __always_inline
void zap_present_folio_ptes(struct mmu_gather
*tlb
,
1484 struct vm_area_struct
*vma
, struct folio
*folio
,
1485 struct page
*page
, pte_t
*pte
, pte_t ptent
, unsigned int nr
,
1486 unsigned long addr
, struct zap_details
*details
, int *rss
,
1487 bool *force_flush
, bool *force_break
)
1489 struct mm_struct
*mm
= tlb
->mm
;
1490 bool delay_rmap
= false;
1492 if (!folio_test_anon(folio
)) {
1493 ptent
= get_and_clear_full_ptes(mm
, addr
, pte
, nr
, tlb
->fullmm
);
1494 if (pte_dirty(ptent
)) {
1495 folio_mark_dirty(folio
);
1496 if (tlb_delay_rmap(tlb
)) {
1498 *force_flush
= true;
1501 if (pte_young(ptent
) && likely(vma_has_recency(vma
)))
1502 folio_mark_accessed(folio
);
1503 rss
[mm_counter(folio
)] -= nr
;
1505 /* We don't need up-to-date accessed/dirty bits. */
1506 clear_full_ptes(mm
, addr
, pte
, nr
, tlb
->fullmm
);
1507 rss
[MM_ANONPAGES
] -= nr
;
1509 /* Checking a single PTE in a batch is sufficient. */
1510 arch_check_zapped_pte(vma
, ptent
);
1511 tlb_remove_tlb_entries(tlb
, pte
, nr
, addr
);
1512 if (unlikely(userfaultfd_pte_wp(vma
, ptent
)))
1513 zap_install_uffd_wp_if_needed(vma
, addr
, pte
, nr
, details
,
1517 folio_remove_rmap_ptes(folio
, page
, nr
, vma
);
1519 if (unlikely(folio_mapcount(folio
) < 0))
1520 print_bad_pte(vma
, addr
, ptent
, page
);
1522 if (unlikely(__tlb_remove_folio_pages(tlb
, page
, nr
, delay_rmap
))) {
1523 *force_flush
= true;
1524 *force_break
= true;
1529 * Zap or skip at least one present PTE, trying to batch-process subsequent
1530 * PTEs that map consecutive pages of the same folio.
1532 * Returns the number of processed (skipped or zapped) PTEs (at least 1).
1534 static inline int zap_present_ptes(struct mmu_gather
*tlb
,
1535 struct vm_area_struct
*vma
, pte_t
*pte
, pte_t ptent
,
1536 unsigned int max_nr
, unsigned long addr
,
1537 struct zap_details
*details
, int *rss
, bool *force_flush
,
1540 const fpb_t fpb_flags
= FPB_IGNORE_DIRTY
| FPB_IGNORE_SOFT_DIRTY
;
1541 struct mm_struct
*mm
= tlb
->mm
;
1542 struct folio
*folio
;
1546 page
= vm_normal_page(vma
, addr
, ptent
);
1548 /* We don't need up-to-date accessed/dirty bits. */
1549 ptep_get_and_clear_full(mm
, addr
, pte
, tlb
->fullmm
);
1550 arch_check_zapped_pte(vma
, ptent
);
1551 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1552 if (userfaultfd_pte_wp(vma
, ptent
))
1553 zap_install_uffd_wp_if_needed(vma
, addr
, pte
, 1,
1555 ksm_might_unmap_zero_page(mm
, ptent
);
1559 folio
= page_folio(page
);
1560 if (unlikely(!should_zap_folio(details
, folio
)))
1564 * Make sure that the common "small folio" case is as fast as possible
1565 * by keeping the batching logic separate.
1567 if (unlikely(folio_test_large(folio
) && max_nr
!= 1)) {
1568 nr
= folio_pte_batch(folio
, addr
, pte
, ptent
, max_nr
, fpb_flags
,
1571 zap_present_folio_ptes(tlb
, vma
, folio
, page
, pte
, ptent
, nr
,
1572 addr
, details
, rss
, force_flush
,
1576 zap_present_folio_ptes(tlb
, vma
, folio
, page
, pte
, ptent
, 1, addr
,
1577 details
, rss
, force_flush
, force_break
);
1581 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1582 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1583 unsigned long addr
, unsigned long end
,
1584 struct zap_details
*details
)
1586 bool force_flush
= false, force_break
= false;
1587 struct mm_struct
*mm
= tlb
->mm
;
1588 int rss
[NR_MM_COUNTERS
];
1595 tlb_change_page_size(tlb
, PAGE_SIZE
);
1597 start_pte
= pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1601 flush_tlb_batched_pending(mm
);
1602 arch_enter_lazy_mmu_mode();
1604 pte_t ptent
= ptep_get(pte
);
1605 struct folio
*folio
;
1610 if (pte_none(ptent
))
1616 if (pte_present(ptent
)) {
1617 max_nr
= (end
- addr
) / PAGE_SIZE
;
1618 nr
= zap_present_ptes(tlb
, vma
, pte
, ptent
, max_nr
,
1619 addr
, details
, rss
, &force_flush
,
1621 if (unlikely(force_break
)) {
1622 addr
+= nr
* PAGE_SIZE
;
1628 entry
= pte_to_swp_entry(ptent
);
1629 if (is_device_private_entry(entry
) ||
1630 is_device_exclusive_entry(entry
)) {
1631 page
= pfn_swap_entry_to_page(entry
);
1632 folio
= page_folio(page
);
1633 if (unlikely(!should_zap_folio(details
, folio
)))
1636 * Both device private/exclusive mappings should only
1637 * work with anonymous page so far, so we don't need to
1638 * consider uffd-wp bit when zap. For more information,
1639 * see zap_install_uffd_wp_if_needed().
1641 WARN_ON_ONCE(!vma_is_anonymous(vma
));
1642 rss
[mm_counter(folio
)]--;
1643 if (is_device_private_entry(entry
))
1644 folio_remove_rmap_pte(folio
, page
, vma
);
1646 } else if (!non_swap_entry(entry
)) {
1647 max_nr
= (end
- addr
) / PAGE_SIZE
;
1648 nr
= swap_pte_batch(pte
, max_nr
, ptent
);
1649 /* Genuine swap entries, hence a private anon pages */
1650 if (!should_zap_cows(details
))
1652 rss
[MM_SWAPENTS
] -= nr
;
1653 free_swap_and_cache_nr(entry
, nr
);
1654 } else if (is_migration_entry(entry
)) {
1655 folio
= pfn_swap_entry_folio(entry
);
1656 if (!should_zap_folio(details
, folio
))
1658 rss
[mm_counter(folio
)]--;
1659 } else if (pte_marker_entry_uffd_wp(entry
)) {
1661 * For anon: always drop the marker; for file: only
1662 * drop the marker if explicitly requested.
1664 if (!vma_is_anonymous(vma
) &&
1665 !zap_drop_file_uffd_wp(details
))
1667 } else if (is_hwpoison_entry(entry
) ||
1668 is_poisoned_swp_entry(entry
)) {
1669 if (!should_zap_cows(details
))
1672 /* We should have covered all the swap entry types */
1673 pr_alert("unrecognized swap entry 0x%lx\n", entry
.val
);
1676 clear_not_present_full_ptes(mm
, addr
, pte
, nr
, tlb
->fullmm
);
1677 zap_install_uffd_wp_if_needed(vma
, addr
, pte
, nr
, details
, ptent
);
1678 } while (pte
+= nr
, addr
+= PAGE_SIZE
* nr
, addr
!= end
);
1680 add_mm_rss_vec(mm
, rss
);
1681 arch_leave_lazy_mmu_mode();
1683 /* Do the actual TLB flush before dropping ptl */
1685 tlb_flush_mmu_tlbonly(tlb
);
1686 tlb_flush_rmaps(tlb
, vma
);
1688 pte_unmap_unlock(start_pte
, ptl
);
1691 * If we forced a TLB flush (either due to running out of
1692 * batch buffers or because we needed to flush dirty TLB
1693 * entries before releasing the ptl), free the batched
1694 * memory too. Come back again if we didn't do everything.
1702 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1703 struct vm_area_struct
*vma
, pud_t
*pud
,
1704 unsigned long addr
, unsigned long end
,
1705 struct zap_details
*details
)
1710 pmd
= pmd_offset(pud
, addr
);
1712 next
= pmd_addr_end(addr
, end
);
1713 if (is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1714 if (next
- addr
!= HPAGE_PMD_SIZE
)
1715 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
1716 else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
)) {
1721 } else if (details
&& details
->single_folio
&&
1722 folio_test_pmd_mappable(details
->single_folio
) &&
1723 next
- addr
== HPAGE_PMD_SIZE
&& pmd_none(*pmd
)) {
1724 spinlock_t
*ptl
= pmd_lock(tlb
->mm
, pmd
);
1726 * Take and drop THP pmd lock so that we cannot return
1727 * prematurely, while zap_huge_pmd() has cleared *pmd,
1728 * but not yet decremented compound_mapcount().
1732 if (pmd_none(*pmd
)) {
1736 addr
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1739 } while (pmd
++, cond_resched(), addr
!= end
);
1744 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1745 struct vm_area_struct
*vma
, p4d_t
*p4d
,
1746 unsigned long addr
, unsigned long end
,
1747 struct zap_details
*details
)
1752 pud
= pud_offset(p4d
, addr
);
1754 next
= pud_addr_end(addr
, end
);
1755 if (pud_trans_huge(*pud
) || pud_devmap(*pud
)) {
1756 if (next
- addr
!= HPAGE_PUD_SIZE
) {
1757 mmap_assert_locked(tlb
->mm
);
1758 split_huge_pud(vma
, pud
, addr
);
1759 } else if (zap_huge_pud(tlb
, vma
, pud
, addr
))
1763 if (pud_none_or_clear_bad(pud
))
1765 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1768 } while (pud
++, addr
= next
, addr
!= end
);
1773 static inline unsigned long zap_p4d_range(struct mmu_gather
*tlb
,
1774 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1775 unsigned long addr
, unsigned long end
,
1776 struct zap_details
*details
)
1781 p4d
= p4d_offset(pgd
, addr
);
1783 next
= p4d_addr_end(addr
, end
);
1784 if (p4d_none_or_clear_bad(p4d
))
1786 next
= zap_pud_range(tlb
, vma
, p4d
, addr
, next
, details
);
1787 } while (p4d
++, addr
= next
, addr
!= end
);
1792 void unmap_page_range(struct mmu_gather
*tlb
,
1793 struct vm_area_struct
*vma
,
1794 unsigned long addr
, unsigned long end
,
1795 struct zap_details
*details
)
1800 BUG_ON(addr
>= end
);
1801 tlb_start_vma(tlb
, vma
);
1802 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1804 next
= pgd_addr_end(addr
, end
);
1805 if (pgd_none_or_clear_bad(pgd
))
1807 next
= zap_p4d_range(tlb
, vma
, pgd
, addr
, next
, details
);
1808 } while (pgd
++, addr
= next
, addr
!= end
);
1809 tlb_end_vma(tlb
, vma
);
1813 static void unmap_single_vma(struct mmu_gather
*tlb
,
1814 struct vm_area_struct
*vma
, unsigned long start_addr
,
1815 unsigned long end_addr
,
1816 struct zap_details
*details
, bool mm_wr_locked
)
1818 unsigned long start
= max(vma
->vm_start
, start_addr
);
1821 if (start
>= vma
->vm_end
)
1823 end
= min(vma
->vm_end
, end_addr
);
1824 if (end
<= vma
->vm_start
)
1828 uprobe_munmap(vma
, start
, end
);
1830 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1831 untrack_pfn(vma
, 0, 0, mm_wr_locked
);
1834 if (unlikely(is_vm_hugetlb_page(vma
))) {
1836 * It is undesirable to test vma->vm_file as it
1837 * should be non-null for valid hugetlb area.
1838 * However, vm_file will be NULL in the error
1839 * cleanup path of mmap_region. When
1840 * hugetlbfs ->mmap method fails,
1841 * mmap_region() nullifies vma->vm_file
1842 * before calling this function to clean up.
1843 * Since no pte has actually been setup, it is
1844 * safe to do nothing in this case.
1847 zap_flags_t zap_flags
= details
?
1848 details
->zap_flags
: 0;
1849 __unmap_hugepage_range(tlb
, vma
, start
, end
,
1853 unmap_page_range(tlb
, vma
, start
, end
, details
);
1858 * unmap_vmas - unmap a range of memory covered by a list of vma's
1859 * @tlb: address of the caller's struct mmu_gather
1860 * @mas: the maple state
1861 * @vma: the starting vma
1862 * @start_addr: virtual address at which to start unmapping
1863 * @end_addr: virtual address at which to end unmapping
1864 * @tree_end: The maximum index to check
1865 * @mm_wr_locked: lock flag
1867 * Unmap all pages in the vma list.
1869 * Only addresses between `start' and `end' will be unmapped.
1871 * The VMA list must be sorted in ascending virtual address order.
1873 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1874 * range after unmap_vmas() returns. So the only responsibility here is to
1875 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1876 * drops the lock and schedules.
1878 void unmap_vmas(struct mmu_gather
*tlb
, struct ma_state
*mas
,
1879 struct vm_area_struct
*vma
, unsigned long start_addr
,
1880 unsigned long end_addr
, unsigned long tree_end
,
1883 struct mmu_notifier_range range
;
1884 struct zap_details details
= {
1885 .zap_flags
= ZAP_FLAG_DROP_MARKER
| ZAP_FLAG_UNMAP
,
1886 /* Careful - we need to zap private pages too! */
1890 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
->vm_mm
,
1891 start_addr
, end_addr
);
1892 mmu_notifier_invalidate_range_start(&range
);
1894 unsigned long start
= start_addr
;
1895 unsigned long end
= end_addr
;
1896 hugetlb_zap_begin(vma
, &start
, &end
);
1897 unmap_single_vma(tlb
, vma
, start
, end
, &details
,
1899 hugetlb_zap_end(vma
, &details
);
1900 vma
= mas_find(mas
, tree_end
- 1);
1901 } while (vma
&& likely(!xa_is_zero(vma
)));
1902 mmu_notifier_invalidate_range_end(&range
);
1906 * zap_page_range_single - remove user pages in a given range
1907 * @vma: vm_area_struct holding the applicable pages
1908 * @address: starting address of pages to zap
1909 * @size: number of bytes to zap
1910 * @details: details of shared cache invalidation
1912 * The range must fit into one VMA.
1914 void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1915 unsigned long size
, struct zap_details
*details
)
1917 const unsigned long end
= address
+ size
;
1918 struct mmu_notifier_range range
;
1919 struct mmu_gather tlb
;
1922 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
->vm_mm
,
1924 hugetlb_zap_begin(vma
, &range
.start
, &range
.end
);
1925 tlb_gather_mmu(&tlb
, vma
->vm_mm
);
1926 update_hiwater_rss(vma
->vm_mm
);
1927 mmu_notifier_invalidate_range_start(&range
);
1929 * unmap 'address-end' not 'range.start-range.end' as range
1930 * could have been expanded for hugetlb pmd sharing.
1932 unmap_single_vma(&tlb
, vma
, address
, end
, details
, false);
1933 mmu_notifier_invalidate_range_end(&range
);
1934 tlb_finish_mmu(&tlb
);
1935 hugetlb_zap_end(vma
, details
);
1939 * zap_vma_ptes - remove ptes mapping the vma
1940 * @vma: vm_area_struct holding ptes to be zapped
1941 * @address: starting address of pages to zap
1942 * @size: number of bytes to zap
1944 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1946 * The entire address range must be fully contained within the vma.
1949 void zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1952 if (!range_in_vma(vma
, address
, address
+ size
) ||
1953 !(vma
->vm_flags
& VM_PFNMAP
))
1956 zap_page_range_single(vma
, address
, size
, NULL
);
1958 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1960 static pmd_t
*walk_to_pmd(struct mm_struct
*mm
, unsigned long addr
)
1967 pgd
= pgd_offset(mm
, addr
);
1968 p4d
= p4d_alloc(mm
, pgd
, addr
);
1971 pud
= pud_alloc(mm
, p4d
, addr
);
1974 pmd
= pmd_alloc(mm
, pud
, addr
);
1978 VM_BUG_ON(pmd_trans_huge(*pmd
));
1982 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1985 pmd_t
*pmd
= walk_to_pmd(mm
, addr
);
1989 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1992 static bool vm_mixed_zeropage_allowed(struct vm_area_struct
*vma
)
1994 VM_WARN_ON_ONCE(vma
->vm_flags
& VM_PFNMAP
);
1996 * Whoever wants to forbid the zeropage after some zeropages
1997 * might already have been mapped has to scan the page tables and
1998 * bail out on any zeropages. Zeropages in COW mappings can
1999 * be unshared using FAULT_FLAG_UNSHARE faults.
2001 if (mm_forbids_zeropage(vma
->vm_mm
))
2003 /* zeropages in COW mappings are common and unproblematic. */
2004 if (is_cow_mapping(vma
->vm_flags
))
2006 /* Mappings that do not allow for writable PTEs are unproblematic. */
2007 if (!(vma
->vm_flags
& (VM_WRITE
| VM_MAYWRITE
)))
2010 * Why not allow any VMA that has vm_ops->pfn_mkwrite? GUP could
2011 * find the shared zeropage and longterm-pin it, which would
2012 * be problematic as soon as the zeropage gets replaced by a different
2013 * page due to vma->vm_ops->pfn_mkwrite, because what's mapped would
2014 * now differ to what GUP looked up. FSDAX is incompatible to
2015 * FOLL_LONGTERM and VM_IO is incompatible to GUP completely (see
2018 return vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
&&
2019 (vma_is_fsdax(vma
) || vma
->vm_flags
& VM_IO
);
2022 static int validate_page_before_insert(struct vm_area_struct
*vma
,
2025 struct folio
*folio
= page_folio(page
);
2027 if (!folio_ref_count(folio
))
2029 if (unlikely(is_zero_folio(folio
))) {
2030 if (!vm_mixed_zeropage_allowed(vma
))
2034 if (folio_test_anon(folio
) || folio_test_slab(folio
) ||
2035 page_has_type(page
))
2037 flush_dcache_folio(folio
);
2041 static int insert_page_into_pte_locked(struct vm_area_struct
*vma
, pte_t
*pte
,
2042 unsigned long addr
, struct page
*page
, pgprot_t prot
)
2044 struct folio
*folio
= page_folio(page
);
2047 if (!pte_none(ptep_get(pte
)))
2049 /* Ok, finally just insert the thing.. */
2050 pteval
= mk_pte(page
, prot
);
2051 if (unlikely(is_zero_folio(folio
))) {
2052 pteval
= pte_mkspecial(pteval
);
2055 inc_mm_counter(vma
->vm_mm
, mm_counter_file(folio
));
2056 folio_add_file_rmap_pte(folio
, page
, vma
);
2058 set_pte_at(vma
->vm_mm
, addr
, pte
, pteval
);
2062 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2063 struct page
*page
, pgprot_t prot
)
2069 retval
= validate_page_before_insert(vma
, page
);
2073 pte
= get_locked_pte(vma
->vm_mm
, addr
, &ptl
);
2076 retval
= insert_page_into_pte_locked(vma
, pte
, addr
, page
, prot
);
2077 pte_unmap_unlock(pte
, ptl
);
2082 static int insert_page_in_batch_locked(struct vm_area_struct
*vma
, pte_t
*pte
,
2083 unsigned long addr
, struct page
*page
, pgprot_t prot
)
2087 err
= validate_page_before_insert(vma
, page
);
2090 return insert_page_into_pte_locked(vma
, pte
, addr
, page
, prot
);
2093 /* insert_pages() amortizes the cost of spinlock operations
2094 * when inserting pages in a loop.
2096 static int insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
2097 struct page
**pages
, unsigned long *num
, pgprot_t prot
)
2100 pte_t
*start_pte
, *pte
;
2101 spinlock_t
*pte_lock
;
2102 struct mm_struct
*const mm
= vma
->vm_mm
;
2103 unsigned long curr_page_idx
= 0;
2104 unsigned long remaining_pages_total
= *num
;
2105 unsigned long pages_to_write_in_pmd
;
2109 pmd
= walk_to_pmd(mm
, addr
);
2113 pages_to_write_in_pmd
= min_t(unsigned long,
2114 remaining_pages_total
, PTRS_PER_PTE
- pte_index(addr
));
2116 /* Allocate the PTE if necessary; takes PMD lock once only. */
2118 if (pte_alloc(mm
, pmd
))
2121 while (pages_to_write_in_pmd
) {
2123 const int batch_size
= min_t(int, pages_to_write_in_pmd
, 8);
2125 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &pte_lock
);
2130 for (pte
= start_pte
; pte_idx
< batch_size
; ++pte
, ++pte_idx
) {
2131 int err
= insert_page_in_batch_locked(vma
, pte
,
2132 addr
, pages
[curr_page_idx
], prot
);
2133 if (unlikely(err
)) {
2134 pte_unmap_unlock(start_pte
, pte_lock
);
2136 remaining_pages_total
-= pte_idx
;
2142 pte_unmap_unlock(start_pte
, pte_lock
);
2143 pages_to_write_in_pmd
-= batch_size
;
2144 remaining_pages_total
-= batch_size
;
2146 if (remaining_pages_total
)
2150 *num
= remaining_pages_total
;
2155 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
2156 * @vma: user vma to map to
2157 * @addr: target start user address of these pages
2158 * @pages: source kernel pages
2159 * @num: in: number of pages to map. out: number of pages that were *not*
2160 * mapped. (0 means all pages were successfully mapped).
2162 * Preferred over vm_insert_page() when inserting multiple pages.
2164 * In case of error, we may have mapped a subset of the provided
2165 * pages. It is the caller's responsibility to account for this case.
2167 * The same restrictions apply as in vm_insert_page().
2169 int vm_insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
2170 struct page
**pages
, unsigned long *num
)
2172 const unsigned long end_addr
= addr
+ (*num
* PAGE_SIZE
) - 1;
2174 if (addr
< vma
->vm_start
|| end_addr
>= vma
->vm_end
)
2176 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
2177 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
2178 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
2179 vm_flags_set(vma
, VM_MIXEDMAP
);
2181 /* Defer page refcount checking till we're about to map that page. */
2182 return insert_pages(vma
, addr
, pages
, num
, vma
->vm_page_prot
);
2184 EXPORT_SYMBOL(vm_insert_pages
);
2187 * vm_insert_page - insert single page into user vma
2188 * @vma: user vma to map to
2189 * @addr: target user address of this page
2190 * @page: source kernel page
2192 * This allows drivers to insert individual pages they've allocated
2193 * into a user vma. The zeropage is supported in some VMAs,
2194 * see vm_mixed_zeropage_allowed().
2196 * The page has to be a nice clean _individual_ kernel allocation.
2197 * If you allocate a compound page, you need to have marked it as
2198 * such (__GFP_COMP), or manually just split the page up yourself
2199 * (see split_page()).
2201 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2202 * took an arbitrary page protection parameter. This doesn't allow
2203 * that. Your vma protection will have to be set up correctly, which
2204 * means that if you want a shared writable mapping, you'd better
2205 * ask for a shared writable mapping!
2207 * The page does not need to be reserved.
2209 * Usually this function is called from f_op->mmap() handler
2210 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
2211 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2212 * function from other places, for example from page-fault handler.
2214 * Return: %0 on success, negative error code otherwise.
2216 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2219 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2221 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
2222 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
2223 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
2224 vm_flags_set(vma
, VM_MIXEDMAP
);
2226 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2228 EXPORT_SYMBOL(vm_insert_page
);
2231 * __vm_map_pages - maps range of kernel pages into user vma
2232 * @vma: user vma to map to
2233 * @pages: pointer to array of source kernel pages
2234 * @num: number of pages in page array
2235 * @offset: user's requested vm_pgoff
2237 * This allows drivers to map range of kernel pages into a user vma.
2238 * The zeropage is supported in some VMAs, see
2239 * vm_mixed_zeropage_allowed().
2241 * Return: 0 on success and error code otherwise.
2243 static int __vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
2244 unsigned long num
, unsigned long offset
)
2246 unsigned long count
= vma_pages(vma
);
2247 unsigned long uaddr
= vma
->vm_start
;
2250 /* Fail if the user requested offset is beyond the end of the object */
2254 /* Fail if the user requested size exceeds available object size */
2255 if (count
> num
- offset
)
2258 for (i
= 0; i
< count
; i
++) {
2259 ret
= vm_insert_page(vma
, uaddr
, pages
[offset
+ i
]);
2269 * vm_map_pages - maps range of kernel pages starts with non zero offset
2270 * @vma: user vma to map to
2271 * @pages: pointer to array of source kernel pages
2272 * @num: number of pages in page array
2274 * Maps an object consisting of @num pages, catering for the user's
2275 * requested vm_pgoff
2277 * If we fail to insert any page into the vma, the function will return
2278 * immediately leaving any previously inserted pages present. Callers
2279 * from the mmap handler may immediately return the error as their caller
2280 * will destroy the vma, removing any successfully inserted pages. Other
2281 * callers should make their own arrangements for calling unmap_region().
2283 * Context: Process context. Called by mmap handlers.
2284 * Return: 0 on success and error code otherwise.
2286 int vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
2289 return __vm_map_pages(vma
, pages
, num
, vma
->vm_pgoff
);
2291 EXPORT_SYMBOL(vm_map_pages
);
2294 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2295 * @vma: user vma to map to
2296 * @pages: pointer to array of source kernel pages
2297 * @num: number of pages in page array
2299 * Similar to vm_map_pages(), except that it explicitly sets the offset
2300 * to 0. This function is intended for the drivers that did not consider
2303 * Context: Process context. Called by mmap handlers.
2304 * Return: 0 on success and error code otherwise.
2306 int vm_map_pages_zero(struct vm_area_struct
*vma
, struct page
**pages
,
2309 return __vm_map_pages(vma
, pages
, num
, 0);
2311 EXPORT_SYMBOL(vm_map_pages_zero
);
2313 static vm_fault_t
insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2314 pfn_t pfn
, pgprot_t prot
, bool mkwrite
)
2316 struct mm_struct
*mm
= vma
->vm_mm
;
2320 pte
= get_locked_pte(mm
, addr
, &ptl
);
2322 return VM_FAULT_OOM
;
2323 entry
= ptep_get(pte
);
2324 if (!pte_none(entry
)) {
2327 * For read faults on private mappings the PFN passed
2328 * in may not match the PFN we have mapped if the
2329 * mapped PFN is a writeable COW page. In the mkwrite
2330 * case we are creating a writable PTE for a shared
2331 * mapping and we expect the PFNs to match. If they
2332 * don't match, we are likely racing with block
2333 * allocation and mapping invalidation so just skip the
2336 if (pte_pfn(entry
) != pfn_t_to_pfn(pfn
)) {
2337 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry
)));
2340 entry
= pte_mkyoung(entry
);
2341 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2342 if (ptep_set_access_flags(vma
, addr
, pte
, entry
, 1))
2343 update_mmu_cache(vma
, addr
, pte
);
2348 /* Ok, finally just insert the thing.. */
2349 if (pfn_t_devmap(pfn
))
2350 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
2352 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
2355 entry
= pte_mkyoung(entry
);
2356 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2359 set_pte_at(mm
, addr
, pte
, entry
);
2360 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
2363 pte_unmap_unlock(pte
, ptl
);
2364 return VM_FAULT_NOPAGE
;
2368 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2369 * @vma: user vma to map to
2370 * @addr: target user address of this page
2371 * @pfn: source kernel pfn
2372 * @pgprot: pgprot flags for the inserted page
2374 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2375 * to override pgprot on a per-page basis.
2377 * This only makes sense for IO mappings, and it makes no sense for
2378 * COW mappings. In general, using multiple vmas is preferable;
2379 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2382 * pgprot typically only differs from @vma->vm_page_prot when drivers set
2383 * caching- and encryption bits different than those of @vma->vm_page_prot,
2384 * because the caching- or encryption mode may not be known at mmap() time.
2386 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2387 * to set caching and encryption bits for those vmas (except for COW pages).
2388 * This is ensured by core vm only modifying these page table entries using
2389 * functions that don't touch caching- or encryption bits, using pte_modify()
2390 * if needed. (See for example mprotect()).
2392 * Also when new page-table entries are created, this is only done using the
2393 * fault() callback, and never using the value of vma->vm_page_prot,
2394 * except for page-table entries that point to anonymous pages as the result
2397 * Context: Process context. May allocate using %GFP_KERNEL.
2398 * Return: vm_fault_t value.
2400 vm_fault_t
vmf_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
2401 unsigned long pfn
, pgprot_t pgprot
)
2404 * Technically, architectures with pte_special can avoid all these
2405 * restrictions (same for remap_pfn_range). However we would like
2406 * consistency in testing and feature parity among all, so we should
2407 * try to keep these invariants in place for everybody.
2409 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
2410 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
2411 (VM_PFNMAP
|VM_MIXEDMAP
));
2412 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
2413 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
2415 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2416 return VM_FAULT_SIGBUS
;
2418 if (!pfn_modify_allowed(pfn
, pgprot
))
2419 return VM_FAULT_SIGBUS
;
2421 track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
));
2423 return insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
,
2426 EXPORT_SYMBOL(vmf_insert_pfn_prot
);
2429 * vmf_insert_pfn - insert single pfn into user vma
2430 * @vma: user vma to map to
2431 * @addr: target user address of this page
2432 * @pfn: source kernel pfn
2434 * Similar to vm_insert_page, this allows drivers to insert individual pages
2435 * they've allocated into a user vma. Same comments apply.
2437 * This function should only be called from a vm_ops->fault handler, and
2438 * in that case the handler should return the result of this function.
2440 * vma cannot be a COW mapping.
2442 * As this is called only for pages that do not currently exist, we
2443 * do not need to flush old virtual caches or the TLB.
2445 * Context: Process context. May allocate using %GFP_KERNEL.
2446 * Return: vm_fault_t value.
2448 vm_fault_t
vmf_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2451 return vmf_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
2453 EXPORT_SYMBOL(vmf_insert_pfn
);
2455 static bool vm_mixed_ok(struct vm_area_struct
*vma
, pfn_t pfn
, bool mkwrite
)
2457 if (unlikely(is_zero_pfn(pfn_t_to_pfn(pfn
))) &&
2458 (mkwrite
|| !vm_mixed_zeropage_allowed(vma
)))
2460 /* these checks mirror the abort conditions in vm_normal_page */
2461 if (vma
->vm_flags
& VM_MIXEDMAP
)
2463 if (pfn_t_devmap(pfn
))
2465 if (pfn_t_special(pfn
))
2467 if (is_zero_pfn(pfn_t_to_pfn(pfn
)))
2472 static vm_fault_t
__vm_insert_mixed(struct vm_area_struct
*vma
,
2473 unsigned long addr
, pfn_t pfn
, bool mkwrite
)
2475 pgprot_t pgprot
= vma
->vm_page_prot
;
2478 if (!vm_mixed_ok(vma
, pfn
, mkwrite
))
2479 return VM_FAULT_SIGBUS
;
2481 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2482 return VM_FAULT_SIGBUS
;
2484 track_pfn_insert(vma
, &pgprot
, pfn
);
2486 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn
), pgprot
))
2487 return VM_FAULT_SIGBUS
;
2490 * If we don't have pte special, then we have to use the pfn_valid()
2491 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2492 * refcount the page if pfn_valid is true (hence insert_page rather
2493 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2494 * without pte special, it would there be refcounted as a normal page.
2496 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
) &&
2497 !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
2501 * At this point we are committed to insert_page()
2502 * regardless of whether the caller specified flags that
2503 * result in pfn_t_has_page() == false.
2505 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
2506 err
= insert_page(vma
, addr
, page
, pgprot
);
2508 return insert_pfn(vma
, addr
, pfn
, pgprot
, mkwrite
);
2512 return VM_FAULT_OOM
;
2513 if (err
< 0 && err
!= -EBUSY
)
2514 return VM_FAULT_SIGBUS
;
2516 return VM_FAULT_NOPAGE
;
2519 vm_fault_t
vmf_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2522 return __vm_insert_mixed(vma
, addr
, pfn
, false);
2524 EXPORT_SYMBOL(vmf_insert_mixed
);
2527 * If the insertion of PTE failed because someone else already added a
2528 * different entry in the mean time, we treat that as success as we assume
2529 * the same entry was actually inserted.
2531 vm_fault_t
vmf_insert_mixed_mkwrite(struct vm_area_struct
*vma
,
2532 unsigned long addr
, pfn_t pfn
)
2534 return __vm_insert_mixed(vma
, addr
, pfn
, true);
2538 * maps a range of physical memory into the requested pages. the old
2539 * mappings are removed. any references to nonexistent pages results
2540 * in null mappings (currently treated as "copy-on-access")
2542 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2543 unsigned long addr
, unsigned long end
,
2544 unsigned long pfn
, pgprot_t prot
)
2546 pte_t
*pte
, *mapped_pte
;
2550 mapped_pte
= pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2553 arch_enter_lazy_mmu_mode();
2555 BUG_ON(!pte_none(ptep_get(pte
)));
2556 if (!pfn_modify_allowed(pfn
, prot
)) {
2560 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2562 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2563 arch_leave_lazy_mmu_mode();
2564 pte_unmap_unlock(mapped_pte
, ptl
);
2568 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2569 unsigned long addr
, unsigned long end
,
2570 unsigned long pfn
, pgprot_t prot
)
2576 pfn
-= addr
>> PAGE_SHIFT
;
2577 pmd
= pmd_alloc(mm
, pud
, addr
);
2580 VM_BUG_ON(pmd_trans_huge(*pmd
));
2582 next
= pmd_addr_end(addr
, end
);
2583 err
= remap_pte_range(mm
, pmd
, addr
, next
,
2584 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2587 } while (pmd
++, addr
= next
, addr
!= end
);
2591 static inline int remap_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2592 unsigned long addr
, unsigned long end
,
2593 unsigned long pfn
, pgprot_t prot
)
2599 pfn
-= addr
>> PAGE_SHIFT
;
2600 pud
= pud_alloc(mm
, p4d
, addr
);
2604 next
= pud_addr_end(addr
, end
);
2605 err
= remap_pmd_range(mm
, pud
, addr
, next
,
2606 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2609 } while (pud
++, addr
= next
, addr
!= end
);
2613 static inline int remap_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2614 unsigned long addr
, unsigned long end
,
2615 unsigned long pfn
, pgprot_t prot
)
2621 pfn
-= addr
>> PAGE_SHIFT
;
2622 p4d
= p4d_alloc(mm
, pgd
, addr
);
2626 next
= p4d_addr_end(addr
, end
);
2627 err
= remap_pud_range(mm
, p4d
, addr
, next
,
2628 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2631 } while (p4d
++, addr
= next
, addr
!= end
);
2636 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2637 * must have pre-validated the caching bits of the pgprot_t.
2639 int remap_pfn_range_notrack(struct vm_area_struct
*vma
, unsigned long addr
,
2640 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2644 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2645 struct mm_struct
*mm
= vma
->vm_mm
;
2648 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr
)))
2652 * Physically remapped pages are special. Tell the
2653 * rest of the world about it:
2654 * VM_IO tells people not to look at these pages
2655 * (accesses can have side effects).
2656 * VM_PFNMAP tells the core MM that the base pages are just
2657 * raw PFN mappings, and do not have a "struct page" associated
2660 * Disable vma merging and expanding with mremap().
2662 * Omit vma from core dump, even when VM_IO turned off.
2664 * There's a horrible special case to handle copy-on-write
2665 * behaviour that some programs depend on. We mark the "original"
2666 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2667 * See vm_normal_page() for details.
2669 if (is_cow_mapping(vma
->vm_flags
)) {
2670 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2672 vma
->vm_pgoff
= pfn
;
2675 vm_flags_set(vma
, VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
);
2677 BUG_ON(addr
>= end
);
2678 pfn
-= addr
>> PAGE_SHIFT
;
2679 pgd
= pgd_offset(mm
, addr
);
2680 flush_cache_range(vma
, addr
, end
);
2682 next
= pgd_addr_end(addr
, end
);
2683 err
= remap_p4d_range(mm
, pgd
, addr
, next
,
2684 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2687 } while (pgd
++, addr
= next
, addr
!= end
);
2693 * remap_pfn_range - remap kernel memory to userspace
2694 * @vma: user vma to map to
2695 * @addr: target page aligned user address to start at
2696 * @pfn: page frame number of kernel physical memory address
2697 * @size: size of mapping area
2698 * @prot: page protection flags for this mapping
2700 * Note: this is only safe if the mm semaphore is held when called.
2702 * Return: %0 on success, negative error code otherwise.
2704 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2705 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2709 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
2713 err
= remap_pfn_range_notrack(vma
, addr
, pfn
, size
, prot
);
2715 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
), true);
2718 EXPORT_SYMBOL(remap_pfn_range
);
2721 * vm_iomap_memory - remap memory to userspace
2722 * @vma: user vma to map to
2723 * @start: start of the physical memory to be mapped
2724 * @len: size of area
2726 * This is a simplified io_remap_pfn_range() for common driver use. The
2727 * driver just needs to give us the physical memory range to be mapped,
2728 * we'll figure out the rest from the vma information.
2730 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2731 * whatever write-combining details or similar.
2733 * Return: %0 on success, negative error code otherwise.
2735 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2737 unsigned long vm_len
, pfn
, pages
;
2739 /* Check that the physical memory area passed in looks valid */
2740 if (start
+ len
< start
)
2743 * You *really* shouldn't map things that aren't page-aligned,
2744 * but we've historically allowed it because IO memory might
2745 * just have smaller alignment.
2747 len
+= start
& ~PAGE_MASK
;
2748 pfn
= start
>> PAGE_SHIFT
;
2749 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2750 if (pfn
+ pages
< pfn
)
2753 /* We start the mapping 'vm_pgoff' pages into the area */
2754 if (vma
->vm_pgoff
> pages
)
2756 pfn
+= vma
->vm_pgoff
;
2757 pages
-= vma
->vm_pgoff
;
2759 /* Can we fit all of the mapping? */
2760 vm_len
= vma
->vm_end
- vma
->vm_start
;
2761 if (vm_len
>> PAGE_SHIFT
> pages
)
2764 /* Ok, let it rip */
2765 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2767 EXPORT_SYMBOL(vm_iomap_memory
);
2769 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2770 unsigned long addr
, unsigned long end
,
2771 pte_fn_t fn
, void *data
, bool create
,
2772 pgtbl_mod_mask
*mask
)
2774 pte_t
*pte
, *mapped_pte
;
2779 mapped_pte
= pte
= (mm
== &init_mm
) ?
2780 pte_alloc_kernel_track(pmd
, addr
, mask
) :
2781 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2785 mapped_pte
= pte
= (mm
== &init_mm
) ?
2786 pte_offset_kernel(pmd
, addr
) :
2787 pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
2792 arch_enter_lazy_mmu_mode();
2796 if (create
|| !pte_none(ptep_get(pte
))) {
2797 err
= fn(pte
++, addr
, data
);
2801 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2803 *mask
|= PGTBL_PTE_MODIFIED
;
2805 arch_leave_lazy_mmu_mode();
2808 pte_unmap_unlock(mapped_pte
, ptl
);
2812 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2813 unsigned long addr
, unsigned long end
,
2814 pte_fn_t fn
, void *data
, bool create
,
2815 pgtbl_mod_mask
*mask
)
2821 BUG_ON(pud_leaf(*pud
));
2824 pmd
= pmd_alloc_track(mm
, pud
, addr
, mask
);
2828 pmd
= pmd_offset(pud
, addr
);
2831 next
= pmd_addr_end(addr
, end
);
2832 if (pmd_none(*pmd
) && !create
)
2834 if (WARN_ON_ONCE(pmd_leaf(*pmd
)))
2836 if (!pmd_none(*pmd
) && WARN_ON_ONCE(pmd_bad(*pmd
))) {
2841 err
= apply_to_pte_range(mm
, pmd
, addr
, next
,
2842 fn
, data
, create
, mask
);
2845 } while (pmd
++, addr
= next
, addr
!= end
);
2850 static int apply_to_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2851 unsigned long addr
, unsigned long end
,
2852 pte_fn_t fn
, void *data
, bool create
,
2853 pgtbl_mod_mask
*mask
)
2860 pud
= pud_alloc_track(mm
, p4d
, addr
, mask
);
2864 pud
= pud_offset(p4d
, addr
);
2867 next
= pud_addr_end(addr
, end
);
2868 if (pud_none(*pud
) && !create
)
2870 if (WARN_ON_ONCE(pud_leaf(*pud
)))
2872 if (!pud_none(*pud
) && WARN_ON_ONCE(pud_bad(*pud
))) {
2877 err
= apply_to_pmd_range(mm
, pud
, addr
, next
,
2878 fn
, data
, create
, mask
);
2881 } while (pud
++, addr
= next
, addr
!= end
);
2886 static int apply_to_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2887 unsigned long addr
, unsigned long end
,
2888 pte_fn_t fn
, void *data
, bool create
,
2889 pgtbl_mod_mask
*mask
)
2896 p4d
= p4d_alloc_track(mm
, pgd
, addr
, mask
);
2900 p4d
= p4d_offset(pgd
, addr
);
2903 next
= p4d_addr_end(addr
, end
);
2904 if (p4d_none(*p4d
) && !create
)
2906 if (WARN_ON_ONCE(p4d_leaf(*p4d
)))
2908 if (!p4d_none(*p4d
) && WARN_ON_ONCE(p4d_bad(*p4d
))) {
2913 err
= apply_to_pud_range(mm
, p4d
, addr
, next
,
2914 fn
, data
, create
, mask
);
2917 } while (p4d
++, addr
= next
, addr
!= end
);
2922 static int __apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2923 unsigned long size
, pte_fn_t fn
,
2924 void *data
, bool create
)
2927 unsigned long start
= addr
, next
;
2928 unsigned long end
= addr
+ size
;
2929 pgtbl_mod_mask mask
= 0;
2932 if (WARN_ON(addr
>= end
))
2935 pgd
= pgd_offset(mm
, addr
);
2937 next
= pgd_addr_end(addr
, end
);
2938 if (pgd_none(*pgd
) && !create
)
2940 if (WARN_ON_ONCE(pgd_leaf(*pgd
)))
2942 if (!pgd_none(*pgd
) && WARN_ON_ONCE(pgd_bad(*pgd
))) {
2947 err
= apply_to_p4d_range(mm
, pgd
, addr
, next
,
2948 fn
, data
, create
, &mask
);
2951 } while (pgd
++, addr
= next
, addr
!= end
);
2953 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
2954 arch_sync_kernel_mappings(start
, start
+ size
);
2960 * Scan a region of virtual memory, filling in page tables as necessary
2961 * and calling a provided function on each leaf page table.
2963 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2964 unsigned long size
, pte_fn_t fn
, void *data
)
2966 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, true);
2968 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2971 * Scan a region of virtual memory, calling a provided function on
2972 * each leaf page table where it exists.
2974 * Unlike apply_to_page_range, this does _not_ fill in page tables
2975 * where they are absent.
2977 int apply_to_existing_page_range(struct mm_struct
*mm
, unsigned long addr
,
2978 unsigned long size
, pte_fn_t fn
, void *data
)
2980 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, false);
2982 EXPORT_SYMBOL_GPL(apply_to_existing_page_range
);
2985 * handle_pte_fault chooses page fault handler according to an entry which was
2986 * read non-atomically. Before making any commitment, on those architectures
2987 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2988 * parts, do_swap_page must check under lock before unmapping the pte and
2989 * proceeding (but do_wp_page is only called after already making such a check;
2990 * and do_anonymous_page can safely check later on).
2992 static inline int pte_unmap_same(struct vm_fault
*vmf
)
2995 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2996 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2997 spin_lock(vmf
->ptl
);
2998 same
= pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
);
2999 spin_unlock(vmf
->ptl
);
3002 pte_unmap(vmf
->pte
);
3009 * 0: copied succeeded
3010 * -EHWPOISON: copy failed due to hwpoison in source page
3011 * -EAGAIN: copied failed (some other reason)
3013 static inline int __wp_page_copy_user(struct page
*dst
, struct page
*src
,
3014 struct vm_fault
*vmf
)
3019 struct vm_area_struct
*vma
= vmf
->vma
;
3020 struct mm_struct
*mm
= vma
->vm_mm
;
3021 unsigned long addr
= vmf
->address
;
3024 if (copy_mc_user_highpage(dst
, src
, addr
, vma
))
3030 * If the source page was a PFN mapping, we don't have
3031 * a "struct page" for it. We do a best-effort copy by
3032 * just copying from the original user address. If that
3033 * fails, we just zero-fill it. Live with it.
3035 kaddr
= kmap_local_page(dst
);
3036 pagefault_disable();
3037 uaddr
= (void __user
*)(addr
& PAGE_MASK
);
3040 * On architectures with software "accessed" bits, we would
3041 * take a double page fault, so mark it accessed here.
3044 if (!arch_has_hw_pte_young() && !pte_young(vmf
->orig_pte
)) {
3047 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
3048 if (unlikely(!vmf
->pte
|| !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
3050 * Other thread has already handled the fault
3051 * and update local tlb only
3054 update_mmu_tlb(vma
, addr
, vmf
->pte
);
3059 entry
= pte_mkyoung(vmf
->orig_pte
);
3060 if (ptep_set_access_flags(vma
, addr
, vmf
->pte
, entry
, 0))
3061 update_mmu_cache_range(vmf
, vma
, addr
, vmf
->pte
, 1);
3065 * This really shouldn't fail, because the page is there
3066 * in the page tables. But it might just be unreadable,
3067 * in which case we just give up and fill the result with
3070 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
3074 /* Re-validate under PTL if the page is still mapped */
3075 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
3076 if (unlikely(!vmf
->pte
|| !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
3077 /* The PTE changed under us, update local tlb */
3079 update_mmu_tlb(vma
, addr
, vmf
->pte
);
3085 * The same page can be mapped back since last copy attempt.
3086 * Try to copy again under PTL.
3088 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
3090 * Give a warn in case there can be some obscure
3103 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3105 kunmap_local(kaddr
);
3106 flush_dcache_page(dst
);
3111 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
3113 struct file
*vm_file
= vma
->vm_file
;
3116 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
3119 * Special mappings (e.g. VDSO) do not have any file so fake
3120 * a default GFP_KERNEL for them.
3126 * Notify the address space that the page is about to become writable so that
3127 * it can prohibit this or wait for the page to get into an appropriate state.
3129 * We do this without the lock held, so that it can sleep if it needs to.
3131 static vm_fault_t
do_page_mkwrite(struct vm_fault
*vmf
, struct folio
*folio
)
3134 unsigned int old_flags
= vmf
->flags
;
3136 vmf
->flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
3138 if (vmf
->vma
->vm_file
&&
3139 IS_SWAPFILE(vmf
->vma
->vm_file
->f_mapping
->host
))
3140 return VM_FAULT_SIGBUS
;
3142 ret
= vmf
->vma
->vm_ops
->page_mkwrite(vmf
);
3143 /* Restore original flags so that caller is not surprised */
3144 vmf
->flags
= old_flags
;
3145 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
3147 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
3149 if (!folio
->mapping
) {
3150 folio_unlock(folio
);
3151 return 0; /* retry */
3153 ret
|= VM_FAULT_LOCKED
;
3155 VM_BUG_ON_FOLIO(!folio_test_locked(folio
), folio
);
3160 * Handle dirtying of a page in shared file mapping on a write fault.
3162 * The function expects the page to be locked and unlocks it.
3164 static vm_fault_t
fault_dirty_shared_page(struct vm_fault
*vmf
)
3166 struct vm_area_struct
*vma
= vmf
->vma
;
3167 struct address_space
*mapping
;
3168 struct folio
*folio
= page_folio(vmf
->page
);
3170 bool page_mkwrite
= vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
;
3172 dirtied
= folio_mark_dirty(folio
);
3173 VM_BUG_ON_FOLIO(folio_test_anon(folio
), folio
);
3175 * Take a local copy of the address_space - folio.mapping may be zeroed
3176 * by truncate after folio_unlock(). The address_space itself remains
3177 * pinned by vma->vm_file's reference. We rely on folio_unlock()'s
3178 * release semantics to prevent the compiler from undoing this copying.
3180 mapping
= folio_raw_mapping(folio
);
3181 folio_unlock(folio
);
3184 file_update_time(vma
->vm_file
);
3187 * Throttle page dirtying rate down to writeback speed.
3189 * mapping may be NULL here because some device drivers do not
3190 * set page.mapping but still dirty their pages
3192 * Drop the mmap_lock before waiting on IO, if we can. The file
3193 * is pinning the mapping, as per above.
3195 if ((dirtied
|| page_mkwrite
) && mapping
) {
3198 fpin
= maybe_unlock_mmap_for_io(vmf
, NULL
);
3199 balance_dirty_pages_ratelimited(mapping
);
3202 return VM_FAULT_COMPLETED
;
3210 * Handle write page faults for pages that can be reused in the current vma
3212 * This can happen either due to the mapping being with the VM_SHARED flag,
3213 * or due to us being the last reference standing to the page. In either
3214 * case, all we need to do here is to mark the page as writable and update
3215 * any related book-keeping.
3217 static inline void wp_page_reuse(struct vm_fault
*vmf
, struct folio
*folio
)
3218 __releases(vmf
->ptl
)
3220 struct vm_area_struct
*vma
= vmf
->vma
;
3223 VM_BUG_ON(!(vmf
->flags
& FAULT_FLAG_WRITE
));
3224 VM_WARN_ON(is_zero_pfn(pte_pfn(vmf
->orig_pte
)));
3227 VM_BUG_ON(folio_test_anon(folio
) &&
3228 !PageAnonExclusive(vmf
->page
));
3230 * Clear the folio's cpupid information as the existing
3231 * information potentially belongs to a now completely
3232 * unrelated process.
3234 folio_xchg_last_cpupid(folio
, (1 << LAST_CPUPID_SHIFT
) - 1);
3237 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
3238 entry
= pte_mkyoung(vmf
->orig_pte
);
3239 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3240 if (ptep_set_access_flags(vma
, vmf
->address
, vmf
->pte
, entry
, 1))
3241 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
3242 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3243 count_vm_event(PGREUSE
);
3247 * We could add a bitflag somewhere, but for now, we know that all
3248 * vm_ops that have a ->map_pages have been audited and don't need
3249 * the mmap_lock to be held.
3251 static inline vm_fault_t
vmf_can_call_fault(const struct vm_fault
*vmf
)
3253 struct vm_area_struct
*vma
= vmf
->vma
;
3255 if (vma
->vm_ops
->map_pages
|| !(vmf
->flags
& FAULT_FLAG_VMA_LOCK
))
3258 return VM_FAULT_RETRY
;
3262 * vmf_anon_prepare - Prepare to handle an anonymous fault.
3263 * @vmf: The vm_fault descriptor passed from the fault handler.
3265 * When preparing to insert an anonymous page into a VMA from a
3266 * fault handler, call this function rather than anon_vma_prepare().
3267 * If this vma does not already have an associated anon_vma and we are
3268 * only protected by the per-VMA lock, the caller must retry with the
3269 * mmap_lock held. __anon_vma_prepare() will look at adjacent VMAs to
3270 * determine if this VMA can share its anon_vma, and that's not safe to
3271 * do with only the per-VMA lock held for this VMA.
3273 * Return: 0 if fault handling can proceed. Any other value should be
3274 * returned to the caller.
3276 vm_fault_t
vmf_anon_prepare(struct vm_fault
*vmf
)
3278 struct vm_area_struct
*vma
= vmf
->vma
;
3281 if (likely(vma
->anon_vma
))
3283 if (vmf
->flags
& FAULT_FLAG_VMA_LOCK
) {
3284 if (!mmap_read_trylock(vma
->vm_mm
)) {
3286 return VM_FAULT_RETRY
;
3289 if (__anon_vma_prepare(vma
))
3291 if (vmf
->flags
& FAULT_FLAG_VMA_LOCK
)
3292 mmap_read_unlock(vma
->vm_mm
);
3297 * Handle the case of a page which we actually need to copy to a new page,
3298 * either due to COW or unsharing.
3300 * Called with mmap_lock locked and the old page referenced, but
3301 * without the ptl held.
3303 * High level logic flow:
3305 * - Allocate a page, copy the content of the old page to the new one.
3306 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
3307 * - Take the PTL. If the pte changed, bail out and release the allocated page
3308 * - If the pte is still the way we remember it, update the page table and all
3309 * relevant references. This includes dropping the reference the page-table
3310 * held to the old page, as well as updating the rmap.
3311 * - In any case, unlock the PTL and drop the reference we took to the old page.
3313 static vm_fault_t
wp_page_copy(struct vm_fault
*vmf
)
3315 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
3316 struct vm_area_struct
*vma
= vmf
->vma
;
3317 struct mm_struct
*mm
= vma
->vm_mm
;
3318 struct folio
*old_folio
= NULL
;
3319 struct folio
*new_folio
= NULL
;
3321 int page_copied
= 0;
3322 struct mmu_notifier_range range
;
3326 delayacct_wpcopy_start();
3329 old_folio
= page_folio(vmf
->page
);
3330 ret
= vmf_anon_prepare(vmf
);
3334 pfn_is_zero
= is_zero_pfn(pte_pfn(vmf
->orig_pte
));
3335 new_folio
= folio_prealloc(mm
, vma
, vmf
->address
, pfn_is_zero
);
3342 err
= __wp_page_copy_user(&new_folio
->page
, vmf
->page
, vmf
);
3345 * COW failed, if the fault was solved by other,
3346 * it's fine. If not, userspace would re-fault on
3347 * the same address and we will handle the fault
3348 * from the second attempt.
3349 * The -EHWPOISON case will not be retried.
3351 folio_put(new_folio
);
3353 folio_put(old_folio
);
3355 delayacct_wpcopy_end();
3356 return err
== -EHWPOISON
? VM_FAULT_HWPOISON
: 0;
3358 kmsan_copy_page_meta(&new_folio
->page
, vmf
->page
);
3361 __folio_mark_uptodate(new_folio
);
3363 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, mm
,
3364 vmf
->address
& PAGE_MASK
,
3365 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
);
3366 mmu_notifier_invalidate_range_start(&range
);
3369 * Re-check the pte - we dropped the lock
3371 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, vmf
->address
, &vmf
->ptl
);
3372 if (likely(vmf
->pte
&& pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
3374 if (!folio_test_anon(old_folio
)) {
3375 dec_mm_counter(mm
, mm_counter_file(old_folio
));
3376 inc_mm_counter(mm
, MM_ANONPAGES
);
3379 ksm_might_unmap_zero_page(mm
, vmf
->orig_pte
);
3380 inc_mm_counter(mm
, MM_ANONPAGES
);
3382 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
3383 entry
= mk_pte(&new_folio
->page
, vma
->vm_page_prot
);
3384 entry
= pte_sw_mkyoung(entry
);
3385 if (unlikely(unshare
)) {
3386 if (pte_soft_dirty(vmf
->orig_pte
))
3387 entry
= pte_mksoft_dirty(entry
);
3388 if (pte_uffd_wp(vmf
->orig_pte
))
3389 entry
= pte_mkuffd_wp(entry
);
3391 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3395 * Clear the pte entry and flush it first, before updating the
3396 * pte with the new entry, to keep TLBs on different CPUs in
3397 * sync. This code used to set the new PTE then flush TLBs, but
3398 * that left a window where the new PTE could be loaded into
3399 * some TLBs while the old PTE remains in others.
3401 ptep_clear_flush(vma
, vmf
->address
, vmf
->pte
);
3402 folio_add_new_anon_rmap(new_folio
, vma
, vmf
->address
, RMAP_EXCLUSIVE
);
3403 folio_add_lru_vma(new_folio
, vma
);
3404 BUG_ON(unshare
&& pte_write(entry
));
3405 set_pte_at(mm
, vmf
->address
, vmf
->pte
, entry
);
3406 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
3409 * Only after switching the pte to the new page may
3410 * we remove the mapcount here. Otherwise another
3411 * process may come and find the rmap count decremented
3412 * before the pte is switched to the new page, and
3413 * "reuse" the old page writing into it while our pte
3414 * here still points into it and can be read by other
3417 * The critical issue is to order this
3418 * folio_remove_rmap_pte() with the ptp_clear_flush
3419 * above. Those stores are ordered by (if nothing else,)
3420 * the barrier present in the atomic_add_negative
3421 * in folio_remove_rmap_pte();
3423 * Then the TLB flush in ptep_clear_flush ensures that
3424 * no process can access the old page before the
3425 * decremented mapcount is visible. And the old page
3426 * cannot be reused until after the decremented
3427 * mapcount is visible. So transitively, TLBs to
3428 * old page will be flushed before it can be reused.
3430 folio_remove_rmap_pte(old_folio
, vmf
->page
, vma
);
3433 /* Free the old page.. */
3434 new_folio
= old_folio
;
3436 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3437 } else if (vmf
->pte
) {
3438 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
3439 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3442 mmu_notifier_invalidate_range_end(&range
);
3445 folio_put(new_folio
);
3448 free_swap_cache(old_folio
);
3449 folio_put(old_folio
);
3452 delayacct_wpcopy_end();
3458 folio_put(old_folio
);
3460 delayacct_wpcopy_end();
3465 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3466 * writeable once the page is prepared
3468 * @vmf: structure describing the fault
3469 * @folio: the folio of vmf->page
3471 * This function handles all that is needed to finish a write page fault in a
3472 * shared mapping due to PTE being read-only once the mapped page is prepared.
3473 * It handles locking of PTE and modifying it.
3475 * The function expects the page to be locked or other protection against
3476 * concurrent faults / writeback (such as DAX radix tree locks).
3478 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3479 * we acquired PTE lock.
3481 static vm_fault_t
finish_mkwrite_fault(struct vm_fault
*vmf
, struct folio
*folio
)
3483 WARN_ON_ONCE(!(vmf
->vma
->vm_flags
& VM_SHARED
));
3484 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3487 return VM_FAULT_NOPAGE
;
3489 * We might have raced with another page fault while we released the
3490 * pte_offset_map_lock.
3492 if (!pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)) {
3493 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
3494 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3495 return VM_FAULT_NOPAGE
;
3497 wp_page_reuse(vmf
, folio
);
3502 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3505 static vm_fault_t
wp_pfn_shared(struct vm_fault
*vmf
)
3507 struct vm_area_struct
*vma
= vmf
->vma
;
3509 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
3512 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3513 ret
= vmf_can_call_fault(vmf
);
3517 vmf
->flags
|= FAULT_FLAG_MKWRITE
;
3518 ret
= vma
->vm_ops
->pfn_mkwrite(vmf
);
3519 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))
3521 return finish_mkwrite_fault(vmf
, NULL
);
3523 wp_page_reuse(vmf
, NULL
);
3527 static vm_fault_t
wp_page_shared(struct vm_fault
*vmf
, struct folio
*folio
)
3528 __releases(vmf
->ptl
)
3530 struct vm_area_struct
*vma
= vmf
->vma
;
3535 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
3538 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3539 tmp
= vmf_can_call_fault(vmf
);
3545 tmp
= do_page_mkwrite(vmf
, folio
);
3546 if (unlikely(!tmp
|| (tmp
&
3547 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3551 tmp
= finish_mkwrite_fault(vmf
, folio
);
3552 if (unlikely(tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3553 folio_unlock(folio
);
3558 wp_page_reuse(vmf
, folio
);
3561 ret
|= fault_dirty_shared_page(vmf
);
3567 static bool wp_can_reuse_anon_folio(struct folio
*folio
,
3568 struct vm_area_struct
*vma
)
3571 * We could currently only reuse a subpage of a large folio if no
3572 * other subpages of the large folios are still mapped. However,
3573 * let's just consistently not reuse subpages even if we could
3574 * reuse in that scenario, and give back a large folio a bit
3577 if (folio_test_large(folio
))
3581 * We have to verify under folio lock: these early checks are
3582 * just an optimization to avoid locking the folio and freeing
3583 * the swapcache if there is little hope that we can reuse.
3585 * KSM doesn't necessarily raise the folio refcount.
3587 if (folio_test_ksm(folio
) || folio_ref_count(folio
) > 3)
3589 if (!folio_test_lru(folio
))
3591 * We cannot easily detect+handle references from
3592 * remote LRU caches or references to LRU folios.
3595 if (folio_ref_count(folio
) > 1 + folio_test_swapcache(folio
))
3597 if (!folio_trylock(folio
))
3599 if (folio_test_swapcache(folio
))
3600 folio_free_swap(folio
);
3601 if (folio_test_ksm(folio
) || folio_ref_count(folio
) != 1) {
3602 folio_unlock(folio
);
3606 * Ok, we've got the only folio reference from our mapping
3607 * and the folio is locked, it's dark out, and we're wearing
3608 * sunglasses. Hit it.
3610 folio_move_anon_rmap(folio
, vma
);
3611 folio_unlock(folio
);
3616 * This routine handles present pages, when
3617 * * users try to write to a shared page (FAULT_FLAG_WRITE)
3618 * * GUP wants to take a R/O pin on a possibly shared anonymous page
3619 * (FAULT_FLAG_UNSHARE)
3621 * It is done by copying the page to a new address and decrementing the
3622 * shared-page counter for the old page.
3624 * Note that this routine assumes that the protection checks have been
3625 * done by the caller (the low-level page fault routine in most cases).
3626 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've
3627 * done any necessary COW.
3629 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even
3630 * though the page will change only once the write actually happens. This
3631 * avoids a few races, and potentially makes it more efficient.
3633 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3634 * but allow concurrent faults), with pte both mapped and locked.
3635 * We return with mmap_lock still held, but pte unmapped and unlocked.
3637 static vm_fault_t
do_wp_page(struct vm_fault
*vmf
)
3638 __releases(vmf
->ptl
)
3640 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
3641 struct vm_area_struct
*vma
= vmf
->vma
;
3642 struct folio
*folio
= NULL
;
3645 if (likely(!unshare
)) {
3646 if (userfaultfd_pte_wp(vma
, ptep_get(vmf
->pte
))) {
3647 if (!userfaultfd_wp_async(vma
)) {
3648 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3649 return handle_userfault(vmf
, VM_UFFD_WP
);
3653 * Nothing needed (cache flush, TLB invalidations,
3654 * etc.) because we're only removing the uffd-wp bit,
3655 * which is completely invisible to the user.
3657 pte
= pte_clear_uffd_wp(ptep_get(vmf
->pte
));
3659 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3661 * Update this to be prepared for following up CoW
3664 vmf
->orig_pte
= pte
;
3668 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3669 * is flushed in this case before copying.
3671 if (unlikely(userfaultfd_wp(vmf
->vma
) &&
3672 mm_tlb_flush_pending(vmf
->vma
->vm_mm
)))
3673 flush_tlb_page(vmf
->vma
, vmf
->address
);
3676 vmf
->page
= vm_normal_page(vma
, vmf
->address
, vmf
->orig_pte
);
3679 folio
= page_folio(vmf
->page
);
3682 * Shared mapping: we are guaranteed to have VM_WRITE and
3683 * FAULT_FLAG_WRITE set at this point.
3685 if (vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
)) {
3687 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3690 * We should not cow pages in a shared writeable mapping.
3691 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3694 return wp_pfn_shared(vmf
);
3695 return wp_page_shared(vmf
, folio
);
3699 * Private mapping: create an exclusive anonymous page copy if reuse
3700 * is impossible. We might miss VM_WRITE for FOLL_FORCE handling.
3702 * If we encounter a page that is marked exclusive, we must reuse
3703 * the page without further checks.
3705 if (folio
&& folio_test_anon(folio
) &&
3706 (PageAnonExclusive(vmf
->page
) || wp_can_reuse_anon_folio(folio
, vma
))) {
3707 if (!PageAnonExclusive(vmf
->page
))
3708 SetPageAnonExclusive(vmf
->page
);
3709 if (unlikely(unshare
)) {
3710 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3713 wp_page_reuse(vmf
, folio
);
3717 * Ok, we need to copy. Oh, well..
3722 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3724 if (folio
&& folio_test_ksm(folio
))
3725 count_vm_event(COW_KSM
);
3727 return wp_page_copy(vmf
);
3730 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
3731 unsigned long start_addr
, unsigned long end_addr
,
3732 struct zap_details
*details
)
3734 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
3737 static inline void unmap_mapping_range_tree(struct rb_root_cached
*root
,
3738 pgoff_t first_index
,
3740 struct zap_details
*details
)
3742 struct vm_area_struct
*vma
;
3743 pgoff_t vba
, vea
, zba
, zea
;
3745 vma_interval_tree_foreach(vma
, root
, first_index
, last_index
) {
3746 vba
= vma
->vm_pgoff
;
3747 vea
= vba
+ vma_pages(vma
) - 1;
3748 zba
= max(first_index
, vba
);
3749 zea
= min(last_index
, vea
);
3751 unmap_mapping_range_vma(vma
,
3752 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
3753 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
3759 * unmap_mapping_folio() - Unmap single folio from processes.
3760 * @folio: The locked folio to be unmapped.
3762 * Unmap this folio from any userspace process which still has it mmaped.
3763 * Typically, for efficiency, the range of nearby pages has already been
3764 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3765 * truncation or invalidation holds the lock on a folio, it may find that
3766 * the page has been remapped again: and then uses unmap_mapping_folio()
3767 * to unmap it finally.
3769 void unmap_mapping_folio(struct folio
*folio
)
3771 struct address_space
*mapping
= folio
->mapping
;
3772 struct zap_details details
= { };
3773 pgoff_t first_index
;
3776 VM_BUG_ON(!folio_test_locked(folio
));
3778 first_index
= folio
->index
;
3779 last_index
= folio_next_index(folio
) - 1;
3781 details
.even_cows
= false;
3782 details
.single_folio
= folio
;
3783 details
.zap_flags
= ZAP_FLAG_DROP_MARKER
;
3785 i_mmap_lock_read(mapping
);
3786 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3787 unmap_mapping_range_tree(&mapping
->i_mmap
, first_index
,
3788 last_index
, &details
);
3789 i_mmap_unlock_read(mapping
);
3793 * unmap_mapping_pages() - Unmap pages from processes.
3794 * @mapping: The address space containing pages to be unmapped.
3795 * @start: Index of first page to be unmapped.
3796 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3797 * @even_cows: Whether to unmap even private COWed pages.
3799 * Unmap the pages in this address space from any userspace process which
3800 * has them mmaped. Generally, you want to remove COWed pages as well when
3801 * a file is being truncated, but not when invalidating pages from the page
3804 void unmap_mapping_pages(struct address_space
*mapping
, pgoff_t start
,
3805 pgoff_t nr
, bool even_cows
)
3807 struct zap_details details
= { };
3808 pgoff_t first_index
= start
;
3809 pgoff_t last_index
= start
+ nr
- 1;
3811 details
.even_cows
= even_cows
;
3812 if (last_index
< first_index
)
3813 last_index
= ULONG_MAX
;
3815 i_mmap_lock_read(mapping
);
3816 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3817 unmap_mapping_range_tree(&mapping
->i_mmap
, first_index
,
3818 last_index
, &details
);
3819 i_mmap_unlock_read(mapping
);
3821 EXPORT_SYMBOL_GPL(unmap_mapping_pages
);
3824 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3825 * address_space corresponding to the specified byte range in the underlying
3828 * @mapping: the address space containing mmaps to be unmapped.
3829 * @holebegin: byte in first page to unmap, relative to the start of
3830 * the underlying file. This will be rounded down to a PAGE_SIZE
3831 * boundary. Note that this is different from truncate_pagecache(), which
3832 * must keep the partial page. In contrast, we must get rid of
3834 * @holelen: size of prospective hole in bytes. This will be rounded
3835 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3837 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3838 * but 0 when invalidating pagecache, don't throw away private data.
3840 void unmap_mapping_range(struct address_space
*mapping
,
3841 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
3843 pgoff_t hba
= (pgoff_t
)(holebegin
) >> PAGE_SHIFT
;
3844 pgoff_t hlen
= ((pgoff_t
)(holelen
) + PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3846 /* Check for overflow. */
3847 if (sizeof(holelen
) > sizeof(hlen
)) {
3849 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3850 if (holeend
& ~(long long)ULONG_MAX
)
3851 hlen
= ULONG_MAX
- hba
+ 1;
3854 unmap_mapping_pages(mapping
, hba
, hlen
, even_cows
);
3856 EXPORT_SYMBOL(unmap_mapping_range
);
3859 * Restore a potential device exclusive pte to a working pte entry
3861 static vm_fault_t
remove_device_exclusive_entry(struct vm_fault
*vmf
)
3863 struct folio
*folio
= page_folio(vmf
->page
);
3864 struct vm_area_struct
*vma
= vmf
->vma
;
3865 struct mmu_notifier_range range
;
3869 * We need a reference to lock the folio because we don't hold
3870 * the PTL so a racing thread can remove the device-exclusive
3871 * entry and unmap it. If the folio is free the entry must
3872 * have been removed already. If it happens to have already
3873 * been re-allocated after being freed all we do is lock and
3876 if (!folio_try_get(folio
))
3879 ret
= folio_lock_or_retry(folio
, vmf
);
3884 mmu_notifier_range_init_owner(&range
, MMU_NOTIFY_EXCLUSIVE
, 0,
3885 vma
->vm_mm
, vmf
->address
& PAGE_MASK
,
3886 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
, NULL
);
3887 mmu_notifier_invalidate_range_start(&range
);
3889 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3891 if (likely(vmf
->pte
&& pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)))
3892 restore_exclusive_pte(vma
, vmf
->page
, vmf
->address
, vmf
->pte
);
3895 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3896 folio_unlock(folio
);
3899 mmu_notifier_invalidate_range_end(&range
);
3903 static inline bool should_try_to_free_swap(struct folio
*folio
,
3904 struct vm_area_struct
*vma
,
3905 unsigned int fault_flags
)
3907 if (!folio_test_swapcache(folio
))
3909 if (mem_cgroup_swap_full(folio
) || (vma
->vm_flags
& VM_LOCKED
) ||
3910 folio_test_mlocked(folio
))
3913 * If we want to map a page that's in the swapcache writable, we
3914 * have to detect via the refcount if we're really the exclusive
3915 * user. Try freeing the swapcache to get rid of the swapcache
3916 * reference only in case it's likely that we'll be the exlusive user.
3918 return (fault_flags
& FAULT_FLAG_WRITE
) && !folio_test_ksm(folio
) &&
3919 folio_ref_count(folio
) == (1 + folio_nr_pages(folio
));
3922 static vm_fault_t
pte_marker_clear(struct vm_fault
*vmf
)
3924 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
,
3925 vmf
->address
, &vmf
->ptl
);
3929 * Be careful so that we will only recover a special uffd-wp pte into a
3930 * none pte. Otherwise it means the pte could have changed, so retry.
3932 * This should also cover the case where e.g. the pte changed
3933 * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_POISONED.
3934 * So is_pte_marker() check is not enough to safely drop the pte.
3936 if (pte_same(vmf
->orig_pte
, ptep_get(vmf
->pte
)))
3937 pte_clear(vmf
->vma
->vm_mm
, vmf
->address
, vmf
->pte
);
3938 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3942 static vm_fault_t
do_pte_missing(struct vm_fault
*vmf
)
3944 if (vma_is_anonymous(vmf
->vma
))
3945 return do_anonymous_page(vmf
);
3947 return do_fault(vmf
);
3951 * This is actually a page-missing access, but with uffd-wp special pte
3952 * installed. It means this pte was wr-protected before being unmapped.
3954 static vm_fault_t
pte_marker_handle_uffd_wp(struct vm_fault
*vmf
)
3957 * Just in case there're leftover special ptes even after the region
3958 * got unregistered - we can simply clear them.
3960 if (unlikely(!userfaultfd_wp(vmf
->vma
)))
3961 return pte_marker_clear(vmf
);
3963 return do_pte_missing(vmf
);
3966 static vm_fault_t
handle_pte_marker(struct vm_fault
*vmf
)
3968 swp_entry_t entry
= pte_to_swp_entry(vmf
->orig_pte
);
3969 unsigned long marker
= pte_marker_get(entry
);
3972 * PTE markers should never be empty. If anything weird happened,
3973 * the best thing to do is to kill the process along with its mm.
3975 if (WARN_ON_ONCE(!marker
))
3976 return VM_FAULT_SIGBUS
;
3978 /* Higher priority than uffd-wp when data corrupted */
3979 if (marker
& PTE_MARKER_POISONED
)
3980 return VM_FAULT_HWPOISON
;
3982 if (pte_marker_entry_uffd_wp(entry
))
3983 return pte_marker_handle_uffd_wp(vmf
);
3985 /* This is an unknown pte marker */
3986 return VM_FAULT_SIGBUS
;
3990 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3991 * but allow concurrent faults), and pte mapped but not yet locked.
3992 * We return with pte unmapped and unlocked.
3994 * We return with the mmap_lock locked or unlocked in the same cases
3995 * as does filemap_fault().
3997 vm_fault_t
do_swap_page(struct vm_fault
*vmf
)
3999 struct vm_area_struct
*vma
= vmf
->vma
;
4000 struct folio
*swapcache
, *folio
= NULL
;
4002 struct swap_info_struct
*si
= NULL
;
4003 rmap_t rmap_flags
= RMAP_NONE
;
4004 bool need_clear_cache
= false;
4005 bool exclusive
= false;
4009 void *shadow
= NULL
;
4011 unsigned long page_idx
;
4012 unsigned long address
;
4015 if (!pte_unmap_same(vmf
))
4018 entry
= pte_to_swp_entry(vmf
->orig_pte
);
4019 if (unlikely(non_swap_entry(entry
))) {
4020 if (is_migration_entry(entry
)) {
4021 migration_entry_wait(vma
->vm_mm
, vmf
->pmd
,
4023 } else if (is_device_exclusive_entry(entry
)) {
4024 vmf
->page
= pfn_swap_entry_to_page(entry
);
4025 ret
= remove_device_exclusive_entry(vmf
);
4026 } else if (is_device_private_entry(entry
)) {
4027 if (vmf
->flags
& FAULT_FLAG_VMA_LOCK
) {
4029 * migrate_to_ram is not yet ready to operate
4033 ret
= VM_FAULT_RETRY
;
4037 vmf
->page
= pfn_swap_entry_to_page(entry
);
4038 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4039 vmf
->address
, &vmf
->ptl
);
4040 if (unlikely(!vmf
->pte
||
4041 !pte_same(ptep_get(vmf
->pte
),
4046 * Get a page reference while we know the page can't be
4049 get_page(vmf
->page
);
4050 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4051 ret
= vmf
->page
->pgmap
->ops
->migrate_to_ram(vmf
);
4052 put_page(vmf
->page
);
4053 } else if (is_hwpoison_entry(entry
)) {
4054 ret
= VM_FAULT_HWPOISON
;
4055 } else if (is_pte_marker_entry(entry
)) {
4056 ret
= handle_pte_marker(vmf
);
4058 print_bad_pte(vma
, vmf
->address
, vmf
->orig_pte
, NULL
);
4059 ret
= VM_FAULT_SIGBUS
;
4064 /* Prevent swapoff from happening to us. */
4065 si
= get_swap_device(entry
);
4069 folio
= swap_cache_get_folio(entry
, vma
, vmf
->address
);
4071 page
= folio_file_page(folio
, swp_offset(entry
));
4075 if (data_race(si
->flags
& SWP_SYNCHRONOUS_IO
) &&
4076 __swap_count(entry
) == 1) {
4078 * Prevent parallel swapin from proceeding with
4079 * the cache flag. Otherwise, another thread may
4080 * finish swapin first, free the entry, and swapout
4081 * reusing the same entry. It's undetectable as
4082 * pte_same() returns true due to entry reuse.
4084 if (swapcache_prepare(entry
)) {
4085 /* Relax a bit to prevent rapid repeated page faults */
4086 schedule_timeout_uninterruptible(1);
4089 need_clear_cache
= true;
4091 /* skip swapcache */
4092 folio
= vma_alloc_folio(GFP_HIGHUSER_MOVABLE
, 0,
4093 vma
, vmf
->address
, false);
4094 page
= &folio
->page
;
4096 __folio_set_locked(folio
);
4097 __folio_set_swapbacked(folio
);
4099 if (mem_cgroup_swapin_charge_folio(folio
,
4100 vma
->vm_mm
, GFP_KERNEL
,
4105 mem_cgroup_swapin_uncharge_swap(entry
);
4107 shadow
= get_shadow_from_swap_cache(entry
);
4109 workingset_refault(folio
, shadow
);
4111 folio_add_lru(folio
);
4113 /* To provide entry to swap_read_folio() */
4114 folio
->swap
= entry
;
4115 swap_read_folio(folio
, NULL
);
4116 folio
->private = NULL
;
4119 page
= swapin_readahead(entry
, GFP_HIGHUSER_MOVABLE
,
4122 folio
= page_folio(page
);
4128 * Back out if somebody else faulted in this pte
4129 * while we released the pte lock.
4131 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4132 vmf
->address
, &vmf
->ptl
);
4133 if (likely(vmf
->pte
&&
4134 pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)))
4139 /* Had to read the page from swap area: Major fault */
4140 ret
= VM_FAULT_MAJOR
;
4141 count_vm_event(PGMAJFAULT
);
4142 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
4143 } else if (PageHWPoison(page
)) {
4145 * hwpoisoned dirty swapcache pages are kept for killing
4146 * owner processes (which may be unknown at hwpoison time)
4148 ret
= VM_FAULT_HWPOISON
;
4152 ret
|= folio_lock_or_retry(folio
, vmf
);
4153 if (ret
& VM_FAULT_RETRY
)
4158 * Make sure folio_free_swap() or swapoff did not release the
4159 * swapcache from under us. The page pin, and pte_same test
4160 * below, are not enough to exclude that. Even if it is still
4161 * swapcache, we need to check that the page's swap has not
4164 if (unlikely(!folio_test_swapcache(folio
) ||
4165 page_swap_entry(page
).val
!= entry
.val
))
4169 * KSM sometimes has to copy on read faults, for example, if
4170 * page->index of !PageKSM() pages would be nonlinear inside the
4171 * anon VMA -- PageKSM() is lost on actual swapout.
4173 folio
= ksm_might_need_to_copy(folio
, vma
, vmf
->address
);
4174 if (unlikely(!folio
)) {
4178 } else if (unlikely(folio
== ERR_PTR(-EHWPOISON
))) {
4179 ret
= VM_FAULT_HWPOISON
;
4183 if (folio
!= swapcache
)
4184 page
= folio_page(folio
, 0);
4187 * If we want to map a page that's in the swapcache writable, we
4188 * have to detect via the refcount if we're really the exclusive
4189 * owner. Try removing the extra reference from the local LRU
4190 * caches if required.
4192 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && folio
== swapcache
&&
4193 !folio_test_ksm(folio
) && !folio_test_lru(folio
))
4197 folio_throttle_swaprate(folio
, GFP_KERNEL
);
4200 * Back out if somebody else already faulted in this pte.
4202 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
4204 if (unlikely(!vmf
->pte
|| !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)))
4207 if (unlikely(!folio_test_uptodate(folio
))) {
4208 ret
= VM_FAULT_SIGBUS
;
4214 address
= vmf
->address
;
4216 if (folio_test_large(folio
) && folio_test_swapcache(folio
)) {
4217 int nr
= folio_nr_pages(folio
);
4218 unsigned long idx
= folio_page_idx(folio
, page
);
4219 unsigned long folio_start
= address
- idx
* PAGE_SIZE
;
4220 unsigned long folio_end
= folio_start
+ nr
* PAGE_SIZE
;
4224 if (unlikely(folio_start
< max(address
& PMD_MASK
, vma
->vm_start
)))
4226 if (unlikely(folio_end
> pmd_addr_end(address
, vma
->vm_end
)))
4229 folio_ptep
= vmf
->pte
- idx
;
4230 folio_pte
= ptep_get(folio_ptep
);
4231 if (!pte_same(folio_pte
, pte_move_swp_offset(vmf
->orig_pte
, -idx
)) ||
4232 swap_pte_batch(folio_ptep
, nr
, folio_pte
) != nr
)
4236 address
= folio_start
;
4239 entry
= folio
->swap
;
4240 page
= &folio
->page
;
4245 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte
4246 * must never point at an anonymous page in the swapcache that is
4247 * PG_anon_exclusive. Sanity check that this holds and especially, that
4248 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity
4249 * check after taking the PT lock and making sure that nobody
4250 * concurrently faulted in this page and set PG_anon_exclusive.
4252 BUG_ON(!folio_test_anon(folio
) && folio_test_mappedtodisk(folio
));
4253 BUG_ON(folio_test_anon(folio
) && PageAnonExclusive(page
));
4256 * Check under PT lock (to protect against concurrent fork() sharing
4257 * the swap entry concurrently) for certainly exclusive pages.
4259 if (!folio_test_ksm(folio
)) {
4260 exclusive
= pte_swp_exclusive(vmf
->orig_pte
);
4261 if (folio
!= swapcache
) {
4263 * We have a fresh page that is not exposed to the
4264 * swapcache -> certainly exclusive.
4267 } else if (exclusive
&& folio_test_writeback(folio
) &&
4268 data_race(si
->flags
& SWP_STABLE_WRITES
)) {
4270 * This is tricky: not all swap backends support
4271 * concurrent page modifications while under writeback.
4273 * So if we stumble over such a page in the swapcache
4274 * we must not set the page exclusive, otherwise we can
4275 * map it writable without further checks and modify it
4276 * while still under writeback.
4278 * For these problematic swap backends, simply drop the
4279 * exclusive marker: this is perfectly fine as we start
4280 * writeback only if we fully unmapped the page and
4281 * there are no unexpected references on the page after
4282 * unmapping succeeded. After fully unmapped, no
4283 * further GUP references (FOLL_GET and FOLL_PIN) can
4284 * appear, so dropping the exclusive marker and mapping
4285 * it only R/O is fine.
4292 * Some architectures may have to restore extra metadata to the page
4293 * when reading from swap. This metadata may be indexed by swap entry
4294 * so this must be called before swap_free().
4296 arch_swap_restore(folio_swap(entry
, folio
), folio
);
4299 * Remove the swap entry and conditionally try to free up the swapcache.
4300 * We're already holding a reference on the page but haven't mapped it
4303 swap_free_nr(entry
, nr_pages
);
4304 if (should_try_to_free_swap(folio
, vma
, vmf
->flags
))
4305 folio_free_swap(folio
);
4307 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, nr_pages
);
4308 add_mm_counter(vma
->vm_mm
, MM_SWAPENTS
, -nr_pages
);
4309 pte
= mk_pte(page
, vma
->vm_page_prot
);
4310 if (pte_swp_soft_dirty(vmf
->orig_pte
))
4311 pte
= pte_mksoft_dirty(pte
);
4312 if (pte_swp_uffd_wp(vmf
->orig_pte
))
4313 pte
= pte_mkuffd_wp(pte
);
4316 * Same logic as in do_wp_page(); however, optimize for pages that are
4317 * certainly not shared either because we just allocated them without
4318 * exposing them to the swapcache or because the swap entry indicates
4321 if (!folio_test_ksm(folio
) &&
4322 (exclusive
|| folio_ref_count(folio
) == 1)) {
4323 if ((vma
->vm_flags
& VM_WRITE
) && !userfaultfd_pte_wp(vma
, pte
) &&
4324 !pte_needs_soft_dirty_wp(vma
, pte
)) {
4325 pte
= pte_mkwrite(pte
, vma
);
4326 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4327 pte
= pte_mkdirty(pte
);
4328 vmf
->flags
&= ~FAULT_FLAG_WRITE
;
4331 rmap_flags
|= RMAP_EXCLUSIVE
;
4333 folio_ref_add(folio
, nr_pages
- 1);
4334 flush_icache_pages(vma
, page
, nr_pages
);
4335 vmf
->orig_pte
= pte_advance_pfn(pte
, page_idx
);
4337 /* ksm created a completely new copy */
4338 if (unlikely(folio
!= swapcache
&& swapcache
)) {
4339 folio_add_new_anon_rmap(folio
, vma
, address
, RMAP_EXCLUSIVE
);
4340 folio_add_lru_vma(folio
, vma
);
4341 } else if (!folio_test_anon(folio
)) {
4343 * We currently only expect small !anon folios, which are either
4344 * fully exclusive or fully shared. If we ever get large folios
4345 * here, we have to be careful.
4347 VM_WARN_ON_ONCE(folio_test_large(folio
));
4348 VM_WARN_ON_FOLIO(!folio_test_locked(folio
), folio
);
4349 folio_add_new_anon_rmap(folio
, vma
, address
, rmap_flags
);
4351 folio_add_anon_rmap_ptes(folio
, page
, nr_pages
, vma
, address
,
4355 VM_BUG_ON(!folio_test_anon(folio
) ||
4356 (pte_write(pte
) && !PageAnonExclusive(page
)));
4357 set_ptes(vma
->vm_mm
, address
, ptep
, pte
, nr_pages
);
4358 arch_do_swap_page_nr(vma
->vm_mm
, vma
, address
,
4359 pte
, pte
, nr_pages
);
4361 folio_unlock(folio
);
4362 if (folio
!= swapcache
&& swapcache
) {
4364 * Hold the lock to avoid the swap entry to be reused
4365 * until we take the PT lock for the pte_same() check
4366 * (to avoid false positives from pte_same). For
4367 * further safety release the lock after the swap_free
4368 * so that the swap count won't change under a
4369 * parallel locked swapcache.
4371 folio_unlock(swapcache
);
4372 folio_put(swapcache
);
4375 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4376 ret
|= do_wp_page(vmf
);
4377 if (ret
& VM_FAULT_ERROR
)
4378 ret
&= VM_FAULT_ERROR
;
4382 /* No need to invalidate - it was non-present before */
4383 update_mmu_cache_range(vmf
, vma
, address
, ptep
, nr_pages
);
4386 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4388 /* Clear the swap cache pin for direct swapin after PTL unlock */
4389 if (need_clear_cache
)
4390 swapcache_clear(si
, entry
);
4392 put_swap_device(si
);
4396 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4398 folio_unlock(folio
);
4401 if (folio
!= swapcache
&& swapcache
) {
4402 folio_unlock(swapcache
);
4403 folio_put(swapcache
);
4405 if (need_clear_cache
)
4406 swapcache_clear(si
, entry
);
4408 put_swap_device(si
);
4412 static bool pte_range_none(pte_t
*pte
, int nr_pages
)
4416 for (i
= 0; i
< nr_pages
; i
++) {
4417 if (!pte_none(ptep_get_lockless(pte
+ i
)))
4424 static struct folio
*alloc_anon_folio(struct vm_fault
*vmf
)
4426 struct vm_area_struct
*vma
= vmf
->vma
;
4427 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4428 unsigned long orders
;
4429 struct folio
*folio
;
4436 * If uffd is active for the vma we need per-page fault fidelity to
4437 * maintain the uffd semantics.
4439 if (unlikely(userfaultfd_armed(vma
)))
4443 * Get a list of all the (large) orders below PMD_ORDER that are enabled
4444 * for this vma. Then filter out the orders that can't be allocated over
4445 * the faulting address and still be fully contained in the vma.
4447 orders
= thp_vma_allowable_orders(vma
, vma
->vm_flags
,
4448 TVA_IN_PF
| TVA_ENFORCE_SYSFS
, BIT(PMD_ORDER
) - 1);
4449 orders
= thp_vma_suitable_orders(vma
, vmf
->address
, orders
);
4454 pte
= pte_offset_map(vmf
->pmd
, vmf
->address
& PMD_MASK
);
4456 return ERR_PTR(-EAGAIN
);
4459 * Find the highest order where the aligned range is completely
4460 * pte_none(). Note that all remaining orders will be completely
4463 order
= highest_order(orders
);
4465 addr
= ALIGN_DOWN(vmf
->address
, PAGE_SIZE
<< order
);
4466 if (pte_range_none(pte
+ pte_index(addr
), 1 << order
))
4468 order
= next_order(&orders
, order
);
4476 /* Try allocating the highest of the remaining orders. */
4477 gfp
= vma_thp_gfp_mask(vma
);
4479 addr
= ALIGN_DOWN(vmf
->address
, PAGE_SIZE
<< order
);
4480 folio
= vma_alloc_folio(gfp
, order
, vma
, addr
, true);
4482 if (mem_cgroup_charge(folio
, vma
->vm_mm
, gfp
)) {
4483 count_mthp_stat(order
, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE
);
4487 folio_throttle_swaprate(folio
, gfp
);
4488 folio_zero_user(folio
, vmf
->address
);
4492 count_mthp_stat(order
, MTHP_STAT_ANON_FAULT_FALLBACK
);
4493 order
= next_order(&orders
, order
);
4498 return folio_prealloc(vma
->vm_mm
, vma
, vmf
->address
, true);
4502 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4503 * but allow concurrent faults), and pte mapped but not yet locked.
4504 * We return with mmap_lock still held, but pte unmapped and unlocked.
4506 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
)
4508 struct vm_area_struct
*vma
= vmf
->vma
;
4509 unsigned long addr
= vmf
->address
;
4510 struct folio
*folio
;
4515 /* File mapping without ->vm_ops ? */
4516 if (vma
->vm_flags
& VM_SHARED
)
4517 return VM_FAULT_SIGBUS
;
4520 * Use pte_alloc() instead of pte_alloc_map(), so that OOM can
4521 * be distinguished from a transient failure of pte_offset_map().
4523 if (pte_alloc(vma
->vm_mm
, vmf
->pmd
))
4524 return VM_FAULT_OOM
;
4526 /* Use the zero-page for reads */
4527 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
4528 !mm_forbids_zeropage(vma
->vm_mm
)) {
4529 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(vmf
->address
),
4530 vma
->vm_page_prot
));
4531 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4532 vmf
->address
, &vmf
->ptl
);
4535 if (vmf_pte_changed(vmf
)) {
4536 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
4539 ret
= check_stable_address_space(vma
->vm_mm
);
4542 /* Deliver the page fault to userland, check inside PT lock */
4543 if (userfaultfd_missing(vma
)) {
4544 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4545 return handle_userfault(vmf
, VM_UFFD_MISSING
);
4550 /* Allocate our own private page. */
4551 ret
= vmf_anon_prepare(vmf
);
4554 /* Returns NULL on OOM or ERR_PTR(-EAGAIN) if we must retry the fault */
4555 folio
= alloc_anon_folio(vmf
);
4561 nr_pages
= folio_nr_pages(folio
);
4562 addr
= ALIGN_DOWN(vmf
->address
, nr_pages
* PAGE_SIZE
);
4565 * The memory barrier inside __folio_mark_uptodate makes sure that
4566 * preceding stores to the page contents become visible before
4567 * the set_pte_at() write.
4569 __folio_mark_uptodate(folio
);
4571 entry
= mk_pte(&folio
->page
, vma
->vm_page_prot
);
4572 entry
= pte_sw_mkyoung(entry
);
4573 if (vma
->vm_flags
& VM_WRITE
)
4574 entry
= pte_mkwrite(pte_mkdirty(entry
), vma
);
4576 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
4579 if (nr_pages
== 1 && vmf_pte_changed(vmf
)) {
4580 update_mmu_tlb(vma
, addr
, vmf
->pte
);
4582 } else if (nr_pages
> 1 && !pte_range_none(vmf
->pte
, nr_pages
)) {
4583 update_mmu_tlb_range(vma
, addr
, vmf
->pte
, nr_pages
);
4587 ret
= check_stable_address_space(vma
->vm_mm
);
4591 /* Deliver the page fault to userland, check inside PT lock */
4592 if (userfaultfd_missing(vma
)) {
4593 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4595 return handle_userfault(vmf
, VM_UFFD_MISSING
);
4598 folio_ref_add(folio
, nr_pages
- 1);
4599 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, nr_pages
);
4600 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4601 count_mthp_stat(folio_order(folio
), MTHP_STAT_ANON_FAULT_ALLOC
);
4603 folio_add_new_anon_rmap(folio
, vma
, addr
, RMAP_EXCLUSIVE
);
4604 folio_add_lru_vma(folio
, vma
);
4606 if (vmf_orig_pte_uffd_wp(vmf
))
4607 entry
= pte_mkuffd_wp(entry
);
4608 set_ptes(vma
->vm_mm
, addr
, vmf
->pte
, entry
, nr_pages
);
4610 /* No need to invalidate - it was non-present before */
4611 update_mmu_cache_range(vmf
, vma
, addr
, vmf
->pte
, nr_pages
);
4614 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4620 return VM_FAULT_OOM
;
4624 * The mmap_lock must have been held on entry, and may have been
4625 * released depending on flags and vma->vm_ops->fault() return value.
4626 * See filemap_fault() and __lock_page_retry().
4628 static vm_fault_t
__do_fault(struct vm_fault
*vmf
)
4630 struct vm_area_struct
*vma
= vmf
->vma
;
4631 struct folio
*folio
;
4635 * Preallocate pte before we take page_lock because this might lead to
4636 * deadlocks for memcg reclaim which waits for pages under writeback:
4638 * SetPageWriteback(A)
4644 * wait_on_page_writeback(A)
4645 * SetPageWriteback(B)
4647 * # flush A, B to clear the writeback
4649 if (pmd_none(*vmf
->pmd
) && !vmf
->prealloc_pte
) {
4650 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
4651 if (!vmf
->prealloc_pte
)
4652 return VM_FAULT_OOM
;
4655 ret
= vma
->vm_ops
->fault(vmf
);
4656 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
|
4657 VM_FAULT_DONE_COW
)))
4660 folio
= page_folio(vmf
->page
);
4661 if (unlikely(PageHWPoison(vmf
->page
))) {
4662 vm_fault_t poisonret
= VM_FAULT_HWPOISON
;
4663 if (ret
& VM_FAULT_LOCKED
) {
4664 if (page_mapped(vmf
->page
))
4665 unmap_mapping_folio(folio
);
4666 /* Retry if a clean folio was removed from the cache. */
4667 if (mapping_evict_folio(folio
->mapping
, folio
))
4668 poisonret
= VM_FAULT_NOPAGE
;
4669 folio_unlock(folio
);
4676 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
4679 VM_BUG_ON_PAGE(!folio_test_locked(folio
), vmf
->page
);
4684 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4685 static void deposit_prealloc_pte(struct vm_fault
*vmf
)
4687 struct vm_area_struct
*vma
= vmf
->vma
;
4689 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
4691 * We are going to consume the prealloc table,
4692 * count that as nr_ptes.
4694 mm_inc_nr_ptes(vma
->vm_mm
);
4695 vmf
->prealloc_pte
= NULL
;
4698 vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
4700 struct folio
*folio
= page_folio(page
);
4701 struct vm_area_struct
*vma
= vmf
->vma
;
4702 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
4703 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
4705 vm_fault_t ret
= VM_FAULT_FALLBACK
;
4707 if (!thp_vma_suitable_order(vma
, haddr
, PMD_ORDER
))
4710 if (folio_order(folio
) != HPAGE_PMD_ORDER
)
4712 page
= &folio
->page
;
4715 * Just backoff if any subpage of a THP is corrupted otherwise
4716 * the corrupted page may mapped by PMD silently to escape the
4717 * check. This kind of THP just can be PTE mapped. Access to
4718 * the corrupted subpage should trigger SIGBUS as expected.
4720 if (unlikely(folio_test_has_hwpoisoned(folio
)))
4724 * Archs like ppc64 need additional space to store information
4725 * related to pte entry. Use the preallocated table for that.
4727 if (arch_needs_pgtable_deposit() && !vmf
->prealloc_pte
) {
4728 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
4729 if (!vmf
->prealloc_pte
)
4730 return VM_FAULT_OOM
;
4733 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
4734 if (unlikely(!pmd_none(*vmf
->pmd
)))
4737 flush_icache_pages(vma
, page
, HPAGE_PMD_NR
);
4739 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
4741 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
4743 add_mm_counter(vma
->vm_mm
, mm_counter_file(folio
), HPAGE_PMD_NR
);
4744 folio_add_file_rmap_pmd(folio
, page
, vma
);
4747 * deposit and withdraw with pmd lock held
4749 if (arch_needs_pgtable_deposit())
4750 deposit_prealloc_pte(vmf
);
4752 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
4754 update_mmu_cache_pmd(vma
, haddr
, vmf
->pmd
);
4756 /* fault is handled */
4758 count_vm_event(THP_FILE_MAPPED
);
4760 spin_unlock(vmf
->ptl
);
4764 vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
4766 return VM_FAULT_FALLBACK
;
4771 * set_pte_range - Set a range of PTEs to point to pages in a folio.
4772 * @vmf: Fault decription.
4773 * @folio: The folio that contains @page.
4774 * @page: The first page to create a PTE for.
4775 * @nr: The number of PTEs to create.
4776 * @addr: The first address to create a PTE for.
4778 void set_pte_range(struct vm_fault
*vmf
, struct folio
*folio
,
4779 struct page
*page
, unsigned int nr
, unsigned long addr
)
4781 struct vm_area_struct
*vma
= vmf
->vma
;
4782 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
4783 bool prefault
= !in_range(vmf
->address
, addr
, nr
* PAGE_SIZE
);
4786 flush_icache_pages(vma
, page
, nr
);
4787 entry
= mk_pte(page
, vma
->vm_page_prot
);
4789 if (prefault
&& arch_wants_old_prefaulted_pte())
4790 entry
= pte_mkold(entry
);
4792 entry
= pte_sw_mkyoung(entry
);
4795 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
4796 if (unlikely(vmf_orig_pte_uffd_wp(vmf
)))
4797 entry
= pte_mkuffd_wp(entry
);
4798 /* copy-on-write page */
4799 if (write
&& !(vma
->vm_flags
& VM_SHARED
)) {
4800 VM_BUG_ON_FOLIO(nr
!= 1, folio
);
4801 folio_add_new_anon_rmap(folio
, vma
, addr
, RMAP_EXCLUSIVE
);
4802 folio_add_lru_vma(folio
, vma
);
4804 folio_add_file_rmap_ptes(folio
, page
, nr
, vma
);
4806 set_ptes(vma
->vm_mm
, addr
, vmf
->pte
, entry
, nr
);
4808 /* no need to invalidate: a not-present page won't be cached */
4809 update_mmu_cache_range(vmf
, vma
, addr
, vmf
->pte
, nr
);
4812 static bool vmf_pte_changed(struct vm_fault
*vmf
)
4814 if (vmf
->flags
& FAULT_FLAG_ORIG_PTE_VALID
)
4815 return !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
);
4817 return !pte_none(ptep_get(vmf
->pte
));
4821 * finish_fault - finish page fault once we have prepared the page to fault
4823 * @vmf: structure describing the fault
4825 * This function handles all that is needed to finish a page fault once the
4826 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4827 * given page, adds reverse page mapping, handles memcg charges and LRU
4830 * The function expects the page to be locked and on success it consumes a
4831 * reference of a page being mapped (for the PTE which maps it).
4833 * Return: %0 on success, %VM_FAULT_ code in case of error.
4835 vm_fault_t
finish_fault(struct vm_fault
*vmf
)
4837 struct vm_area_struct
*vma
= vmf
->vma
;
4839 struct folio
*folio
;
4841 bool is_cow
= (vmf
->flags
& FAULT_FLAG_WRITE
) &&
4842 !(vma
->vm_flags
& VM_SHARED
);
4844 unsigned long addr
= vmf
->address
;
4846 /* Did we COW the page? */
4848 page
= vmf
->cow_page
;
4853 * check even for read faults because we might have lost our CoWed
4856 if (!(vma
->vm_flags
& VM_SHARED
)) {
4857 ret
= check_stable_address_space(vma
->vm_mm
);
4862 if (pmd_none(*vmf
->pmd
)) {
4863 if (PageTransCompound(page
)) {
4864 ret
= do_set_pmd(vmf
, page
);
4865 if (ret
!= VM_FAULT_FALLBACK
)
4869 if (vmf
->prealloc_pte
)
4870 pmd_install(vma
->vm_mm
, vmf
->pmd
, &vmf
->prealloc_pte
);
4871 else if (unlikely(pte_alloc(vma
->vm_mm
, vmf
->pmd
)))
4872 return VM_FAULT_OOM
;
4875 folio
= page_folio(page
);
4876 nr_pages
= folio_nr_pages(folio
);
4879 * Using per-page fault to maintain the uffd semantics, and same
4880 * approach also applies to non-anonymous-shmem faults to avoid
4881 * inflating the RSS of the process.
4883 if (!vma_is_anon_shmem(vma
) || unlikely(userfaultfd_armed(vma
))) {
4885 } else if (nr_pages
> 1) {
4886 pgoff_t idx
= folio_page_idx(folio
, page
);
4887 /* The page offset of vmf->address within the VMA. */
4888 pgoff_t vma_off
= vmf
->pgoff
- vmf
->vma
->vm_pgoff
;
4889 /* The index of the entry in the pagetable for fault page. */
4890 pgoff_t pte_off
= pte_index(vmf
->address
);
4893 * Fallback to per-page fault in case the folio size in page
4894 * cache beyond the VMA limits and PMD pagetable limits.
4896 if (unlikely(vma_off
< idx
||
4897 vma_off
+ (nr_pages
- idx
) > vma_pages(vma
) ||
4899 pte_off
+ (nr_pages
- idx
) > PTRS_PER_PTE
)) {
4902 /* Now we can set mappings for the whole large folio. */
4903 addr
= vmf
->address
- idx
* PAGE_SIZE
;
4904 page
= &folio
->page
;
4908 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4911 return VM_FAULT_NOPAGE
;
4913 /* Re-check under ptl */
4914 if (nr_pages
== 1 && unlikely(vmf_pte_changed(vmf
))) {
4915 update_mmu_tlb(vma
, addr
, vmf
->pte
);
4916 ret
= VM_FAULT_NOPAGE
;
4918 } else if (nr_pages
> 1 && !pte_range_none(vmf
->pte
, nr_pages
)) {
4919 update_mmu_tlb_range(vma
, addr
, vmf
->pte
, nr_pages
);
4920 ret
= VM_FAULT_NOPAGE
;
4924 folio_ref_add(folio
, nr_pages
- 1);
4925 set_pte_range(vmf
, folio
, page
, nr_pages
, addr
);
4926 type
= is_cow
? MM_ANONPAGES
: mm_counter_file(folio
);
4927 add_mm_counter(vma
->vm_mm
, type
, nr_pages
);
4931 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4935 static unsigned long fault_around_pages __read_mostly
=
4936 65536 >> PAGE_SHIFT
;
4938 #ifdef CONFIG_DEBUG_FS
4939 static int fault_around_bytes_get(void *data
, u64
*val
)
4941 *val
= fault_around_pages
<< PAGE_SHIFT
;
4946 * fault_around_bytes must be rounded down to the nearest page order as it's
4947 * what do_fault_around() expects to see.
4949 static int fault_around_bytes_set(void *data
, u64 val
)
4951 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
4955 * The minimum value is 1 page, however this results in no fault-around
4956 * at all. See should_fault_around().
4958 val
= max(val
, PAGE_SIZE
);
4959 fault_around_pages
= rounddown_pow_of_two(val
) >> PAGE_SHIFT
;
4963 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops
,
4964 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
4966 static int __init
fault_around_debugfs(void)
4968 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL
, NULL
,
4969 &fault_around_bytes_fops
);
4972 late_initcall(fault_around_debugfs
);
4976 * do_fault_around() tries to map few pages around the fault address. The hope
4977 * is that the pages will be needed soon and this will lower the number of
4980 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4981 * not ready to be mapped: not up-to-date, locked, etc.
4983 * This function doesn't cross VMA or page table boundaries, in order to call
4984 * map_pages() and acquire a PTE lock only once.
4986 * fault_around_pages defines how many pages we'll try to map.
4987 * do_fault_around() expects it to be set to a power of two less than or equal
4990 * The virtual address of the area that we map is naturally aligned to
4991 * fault_around_pages * PAGE_SIZE rounded down to the machine page size
4992 * (and therefore to page order). This way it's easier to guarantee
4993 * that we don't cross page table boundaries.
4995 static vm_fault_t
do_fault_around(struct vm_fault
*vmf
)
4997 pgoff_t nr_pages
= READ_ONCE(fault_around_pages
);
4998 pgoff_t pte_off
= pte_index(vmf
->address
);
4999 /* The page offset of vmf->address within the VMA. */
5000 pgoff_t vma_off
= vmf
->pgoff
- vmf
->vma
->vm_pgoff
;
5001 pgoff_t from_pte
, to_pte
;
5004 /* The PTE offset of the start address, clamped to the VMA. */
5005 from_pte
= max(ALIGN_DOWN(pte_off
, nr_pages
),
5006 pte_off
- min(pte_off
, vma_off
));
5008 /* The PTE offset of the end address, clamped to the VMA and PTE. */
5009 to_pte
= min3(from_pte
+ nr_pages
, (pgoff_t
)PTRS_PER_PTE
,
5010 pte_off
+ vma_pages(vmf
->vma
) - vma_off
) - 1;
5012 if (pmd_none(*vmf
->pmd
)) {
5013 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
);
5014 if (!vmf
->prealloc_pte
)
5015 return VM_FAULT_OOM
;
5019 ret
= vmf
->vma
->vm_ops
->map_pages(vmf
,
5020 vmf
->pgoff
+ from_pte
- pte_off
,
5021 vmf
->pgoff
+ to_pte
- pte_off
);
5027 /* Return true if we should do read fault-around, false otherwise */
5028 static inline bool should_fault_around(struct vm_fault
*vmf
)
5030 /* No ->map_pages? No way to fault around... */
5031 if (!vmf
->vma
->vm_ops
->map_pages
)
5034 if (uffd_disable_fault_around(vmf
->vma
))
5037 /* A single page implies no faulting 'around' at all. */
5038 return fault_around_pages
> 1;
5041 static vm_fault_t
do_read_fault(struct vm_fault
*vmf
)
5044 struct folio
*folio
;
5047 * Let's call ->map_pages() first and use ->fault() as fallback
5048 * if page by the offset is not ready to be mapped (cold cache or
5051 if (should_fault_around(vmf
)) {
5052 ret
= do_fault_around(vmf
);
5057 ret
= vmf_can_call_fault(vmf
);
5061 ret
= __do_fault(vmf
);
5062 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
5065 ret
|= finish_fault(vmf
);
5066 folio
= page_folio(vmf
->page
);
5067 folio_unlock(folio
);
5068 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
5073 static vm_fault_t
do_cow_fault(struct vm_fault
*vmf
)
5075 struct vm_area_struct
*vma
= vmf
->vma
;
5076 struct folio
*folio
;
5079 ret
= vmf_can_call_fault(vmf
);
5081 ret
= vmf_anon_prepare(vmf
);
5085 folio
= folio_prealloc(vma
->vm_mm
, vma
, vmf
->address
, false);
5087 return VM_FAULT_OOM
;
5089 vmf
->cow_page
= &folio
->page
;
5091 ret
= __do_fault(vmf
);
5092 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
5094 if (ret
& VM_FAULT_DONE_COW
)
5097 copy_user_highpage(vmf
->cow_page
, vmf
->page
, vmf
->address
, vma
);
5098 __folio_mark_uptodate(folio
);
5100 ret
|= finish_fault(vmf
);
5101 unlock_page(vmf
->page
);
5102 put_page(vmf
->page
);
5103 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
5111 static vm_fault_t
do_shared_fault(struct vm_fault
*vmf
)
5113 struct vm_area_struct
*vma
= vmf
->vma
;
5114 vm_fault_t ret
, tmp
;
5115 struct folio
*folio
;
5117 ret
= vmf_can_call_fault(vmf
);
5121 ret
= __do_fault(vmf
);
5122 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
5125 folio
= page_folio(vmf
->page
);
5128 * Check if the backing address space wants to know that the page is
5129 * about to become writable
5131 if (vma
->vm_ops
->page_mkwrite
) {
5132 folio_unlock(folio
);
5133 tmp
= do_page_mkwrite(vmf
, folio
);
5134 if (unlikely(!tmp
||
5135 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
5141 ret
|= finish_fault(vmf
);
5142 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
5144 folio_unlock(folio
);
5149 ret
|= fault_dirty_shared_page(vmf
);
5154 * We enter with non-exclusive mmap_lock (to exclude vma changes,
5155 * but allow concurrent faults).
5156 * The mmap_lock may have been released depending on flags and our
5157 * return value. See filemap_fault() and __folio_lock_or_retry().
5158 * If mmap_lock is released, vma may become invalid (for example
5159 * by other thread calling munmap()).
5161 static vm_fault_t
do_fault(struct vm_fault
*vmf
)
5163 struct vm_area_struct
*vma
= vmf
->vma
;
5164 struct mm_struct
*vm_mm
= vma
->vm_mm
;
5168 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
5170 if (!vma
->vm_ops
->fault
) {
5171 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
,
5172 vmf
->address
, &vmf
->ptl
);
5173 if (unlikely(!vmf
->pte
))
5174 ret
= VM_FAULT_SIGBUS
;
5177 * Make sure this is not a temporary clearing of pte
5178 * by holding ptl and checking again. A R/M/W update
5179 * of pte involves: take ptl, clearing the pte so that
5180 * we don't have concurrent modification by hardware
5181 * followed by an update.
5183 if (unlikely(pte_none(ptep_get(vmf
->pte
))))
5184 ret
= VM_FAULT_SIGBUS
;
5186 ret
= VM_FAULT_NOPAGE
;
5188 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5190 } else if (!(vmf
->flags
& FAULT_FLAG_WRITE
))
5191 ret
= do_read_fault(vmf
);
5192 else if (!(vma
->vm_flags
& VM_SHARED
))
5193 ret
= do_cow_fault(vmf
);
5195 ret
= do_shared_fault(vmf
);
5197 /* preallocated pagetable is unused: free it */
5198 if (vmf
->prealloc_pte
) {
5199 pte_free(vm_mm
, vmf
->prealloc_pte
);
5200 vmf
->prealloc_pte
= NULL
;
5205 int numa_migrate_prep(struct folio
*folio
, struct vm_fault
*vmf
,
5206 unsigned long addr
, int page_nid
, int *flags
)
5208 struct vm_area_struct
*vma
= vmf
->vma
;
5210 /* Record the current PID acceesing VMA */
5211 vma_set_access_pid_bit(vma
);
5213 count_vm_numa_event(NUMA_HINT_FAULTS
);
5214 if (page_nid
== numa_node_id()) {
5215 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
5216 *flags
|= TNF_FAULT_LOCAL
;
5219 return mpol_misplaced(folio
, vmf
, addr
);
5222 static void numa_rebuild_single_mapping(struct vm_fault
*vmf
, struct vm_area_struct
*vma
,
5223 unsigned long fault_addr
, pte_t
*fault_pte
,
5228 old_pte
= ptep_modify_prot_start(vma
, fault_addr
, fault_pte
);
5229 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
5230 pte
= pte_mkyoung(pte
);
5232 pte
= pte_mkwrite(pte
, vma
);
5233 ptep_modify_prot_commit(vma
, fault_addr
, fault_pte
, old_pte
, pte
);
5234 update_mmu_cache_range(vmf
, vma
, fault_addr
, fault_pte
, 1);
5237 static void numa_rebuild_large_mapping(struct vm_fault
*vmf
, struct vm_area_struct
*vma
,
5238 struct folio
*folio
, pte_t fault_pte
,
5239 bool ignore_writable
, bool pte_write_upgrade
)
5241 int nr
= pte_pfn(fault_pte
) - folio_pfn(folio
);
5242 unsigned long start
, end
, addr
= vmf
->address
;
5243 unsigned long addr_start
= addr
- (nr
<< PAGE_SHIFT
);
5244 unsigned long pt_start
= ALIGN_DOWN(addr
, PMD_SIZE
);
5247 /* Stay within the VMA and within the page table. */
5248 start
= max3(addr_start
, pt_start
, vma
->vm_start
);
5249 end
= min3(addr_start
+ folio_size(folio
), pt_start
+ PMD_SIZE
,
5251 start_ptep
= vmf
->pte
- ((addr
- start
) >> PAGE_SHIFT
);
5253 /* Restore all PTEs' mapping of the large folio */
5254 for (addr
= start
; addr
!= end
; start_ptep
++, addr
+= PAGE_SIZE
) {
5255 pte_t ptent
= ptep_get(start_ptep
);
5256 bool writable
= false;
5258 if (!pte_present(ptent
) || !pte_protnone(ptent
))
5261 if (pfn_folio(pte_pfn(ptent
)) != folio
)
5264 if (!ignore_writable
) {
5265 ptent
= pte_modify(ptent
, vma
->vm_page_prot
);
5266 writable
= pte_write(ptent
);
5267 if (!writable
&& pte_write_upgrade
&&
5268 can_change_pte_writable(vma
, addr
, ptent
))
5272 numa_rebuild_single_mapping(vmf
, vma
, addr
, start_ptep
, writable
);
5276 static vm_fault_t
do_numa_page(struct vm_fault
*vmf
)
5278 struct vm_area_struct
*vma
= vmf
->vma
;
5279 struct folio
*folio
= NULL
;
5280 int nid
= NUMA_NO_NODE
;
5281 bool writable
= false, ignore_writable
= false;
5282 bool pte_write_upgrade
= vma_wants_manual_pte_write_upgrade(vma
);
5286 int flags
= 0, nr_pages
;
5289 * The pte cannot be used safely until we verify, while holding the page
5290 * table lock, that its contents have not changed during fault handling.
5292 spin_lock(vmf
->ptl
);
5293 /* Read the live PTE from the page tables: */
5294 old_pte
= ptep_get(vmf
->pte
);
5296 if (unlikely(!pte_same(old_pte
, vmf
->orig_pte
))) {
5297 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5301 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
5304 * Detect now whether the PTE could be writable; this information
5305 * is only valid while holding the PT lock.
5307 writable
= pte_write(pte
);
5308 if (!writable
&& pte_write_upgrade
&&
5309 can_change_pte_writable(vma
, vmf
->address
, pte
))
5312 folio
= vm_normal_folio(vma
, vmf
->address
, pte
);
5313 if (!folio
|| folio_is_zone_device(folio
))
5317 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
5318 * much anyway since they can be in shared cache state. This misses
5319 * the case where a mapping is writable but the process never writes
5320 * to it but pte_write gets cleared during protection updates and
5321 * pte_dirty has unpredictable behaviour between PTE scan updates,
5322 * background writeback, dirty balancing and application behaviour.
5325 flags
|= TNF_NO_GROUP
;
5328 * Flag if the folio is shared between multiple address spaces. This
5329 * is later used when determining whether to group tasks together
5331 if (folio_likely_mapped_shared(folio
) && (vma
->vm_flags
& VM_SHARED
))
5332 flags
|= TNF_SHARED
;
5334 nid
= folio_nid(folio
);
5335 nr_pages
= folio_nr_pages(folio
);
5337 * For memory tiering mode, cpupid of slow memory page is used
5338 * to record page access time. So use default value.
5340 if ((sysctl_numa_balancing_mode
& NUMA_BALANCING_MEMORY_TIERING
) &&
5341 !node_is_toptier(nid
))
5342 last_cpupid
= (-1 & LAST_CPUPID_MASK
);
5344 last_cpupid
= folio_last_cpupid(folio
);
5345 target_nid
= numa_migrate_prep(folio
, vmf
, vmf
->address
, nid
, &flags
);
5346 if (target_nid
== NUMA_NO_NODE
)
5348 if (migrate_misplaced_folio_prepare(folio
, vma
, target_nid
)) {
5349 flags
|= TNF_MIGRATE_FAIL
;
5352 /* The folio is isolated and isolation code holds a folio reference. */
5353 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5355 ignore_writable
= true;
5357 /* Migrate to the requested node */
5358 if (!migrate_misplaced_folio(folio
, vma
, target_nid
)) {
5360 flags
|= TNF_MIGRATED
;
5362 flags
|= TNF_MIGRATE_FAIL
;
5363 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
5364 vmf
->address
, &vmf
->ptl
);
5365 if (unlikely(!vmf
->pte
))
5367 if (unlikely(!pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
5368 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5375 if (nid
!= NUMA_NO_NODE
)
5376 task_numa_fault(last_cpupid
, nid
, nr_pages
, flags
);
5380 * Make it present again, depending on how arch implements
5381 * non-accessible ptes, some can allow access by kernel mode.
5383 if (folio
&& folio_test_large(folio
))
5384 numa_rebuild_large_mapping(vmf
, vma
, folio
, pte
, ignore_writable
,
5387 numa_rebuild_single_mapping(vmf
, vma
, vmf
->address
, vmf
->pte
,
5389 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5393 static inline vm_fault_t
create_huge_pmd(struct vm_fault
*vmf
)
5395 struct vm_area_struct
*vma
= vmf
->vma
;
5396 if (vma_is_anonymous(vma
))
5397 return do_huge_pmd_anonymous_page(vmf
);
5398 if (vma
->vm_ops
->huge_fault
)
5399 return vma
->vm_ops
->huge_fault(vmf
, PMD_ORDER
);
5400 return VM_FAULT_FALLBACK
;
5403 /* `inline' is required to avoid gcc 4.1.2 build error */
5404 static inline vm_fault_t
wp_huge_pmd(struct vm_fault
*vmf
)
5406 struct vm_area_struct
*vma
= vmf
->vma
;
5407 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
5410 if (vma_is_anonymous(vma
)) {
5411 if (likely(!unshare
) &&
5412 userfaultfd_huge_pmd_wp(vma
, vmf
->orig_pmd
)) {
5413 if (userfaultfd_wp_async(vmf
->vma
))
5415 return handle_userfault(vmf
, VM_UFFD_WP
);
5417 return do_huge_pmd_wp_page(vmf
);
5420 if (vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
)) {
5421 if (vma
->vm_ops
->huge_fault
) {
5422 ret
= vma
->vm_ops
->huge_fault(vmf
, PMD_ORDER
);
5423 if (!(ret
& VM_FAULT_FALLBACK
))
5429 /* COW or write-notify handled on pte level: split pmd. */
5430 __split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
5432 return VM_FAULT_FALLBACK
;
5435 static vm_fault_t
create_huge_pud(struct vm_fault
*vmf
)
5437 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
5438 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5439 struct vm_area_struct
*vma
= vmf
->vma
;
5440 /* No support for anonymous transparent PUD pages yet */
5441 if (vma_is_anonymous(vma
))
5442 return VM_FAULT_FALLBACK
;
5443 if (vma
->vm_ops
->huge_fault
)
5444 return vma
->vm_ops
->huge_fault(vmf
, PUD_ORDER
);
5445 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
5446 return VM_FAULT_FALLBACK
;
5449 static vm_fault_t
wp_huge_pud(struct vm_fault
*vmf
, pud_t orig_pud
)
5451 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
5452 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5453 struct vm_area_struct
*vma
= vmf
->vma
;
5456 /* No support for anonymous transparent PUD pages yet */
5457 if (vma_is_anonymous(vma
))
5459 if (vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
)) {
5460 if (vma
->vm_ops
->huge_fault
) {
5461 ret
= vma
->vm_ops
->huge_fault(vmf
, PUD_ORDER
);
5462 if (!(ret
& VM_FAULT_FALLBACK
))
5467 /* COW or write-notify not handled on PUD level: split pud.*/
5468 __split_huge_pud(vma
, vmf
->pud
, vmf
->address
);
5469 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
5470 return VM_FAULT_FALLBACK
;
5474 * These routines also need to handle stuff like marking pages dirty
5475 * and/or accessed for architectures that don't do it in hardware (most
5476 * RISC architectures). The early dirtying is also good on the i386.
5478 * There is also a hook called "update_mmu_cache()" that architectures
5479 * with external mmu caches can use to update those (ie the Sparc or
5480 * PowerPC hashed page tables that act as extended TLBs).
5482 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
5483 * concurrent faults).
5485 * The mmap_lock may have been released depending on flags and our return value.
5486 * See filemap_fault() and __folio_lock_or_retry().
5488 static vm_fault_t
handle_pte_fault(struct vm_fault
*vmf
)
5492 if (unlikely(pmd_none(*vmf
->pmd
))) {
5494 * Leave __pte_alloc() until later: because vm_ops->fault may
5495 * want to allocate huge page, and if we expose page table
5496 * for an instant, it will be difficult to retract from
5497 * concurrent faults and from rmap lookups.
5500 vmf
->flags
&= ~FAULT_FLAG_ORIG_PTE_VALID
;
5503 * A regular pmd is established and it can't morph into a huge
5504 * pmd by anon khugepaged, since that takes mmap_lock in write
5505 * mode; but shmem or file collapse to THP could still morph
5506 * it into a huge pmd: just retry later if so.
5508 vmf
->pte
= pte_offset_map_nolock(vmf
->vma
->vm_mm
, vmf
->pmd
,
5509 vmf
->address
, &vmf
->ptl
);
5510 if (unlikely(!vmf
->pte
))
5512 vmf
->orig_pte
= ptep_get_lockless(vmf
->pte
);
5513 vmf
->flags
|= FAULT_FLAG_ORIG_PTE_VALID
;
5515 if (pte_none(vmf
->orig_pte
)) {
5516 pte_unmap(vmf
->pte
);
5522 return do_pte_missing(vmf
);
5524 if (!pte_present(vmf
->orig_pte
))
5525 return do_swap_page(vmf
);
5527 if (pte_protnone(vmf
->orig_pte
) && vma_is_accessible(vmf
->vma
))
5528 return do_numa_page(vmf
);
5530 spin_lock(vmf
->ptl
);
5531 entry
= vmf
->orig_pte
;
5532 if (unlikely(!pte_same(ptep_get(vmf
->pte
), entry
))) {
5533 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
5536 if (vmf
->flags
& (FAULT_FLAG_WRITE
|FAULT_FLAG_UNSHARE
)) {
5537 if (!pte_write(entry
))
5538 return do_wp_page(vmf
);
5539 else if (likely(vmf
->flags
& FAULT_FLAG_WRITE
))
5540 entry
= pte_mkdirty(entry
);
5542 entry
= pte_mkyoung(entry
);
5543 if (ptep_set_access_flags(vmf
->vma
, vmf
->address
, vmf
->pte
, entry
,
5544 vmf
->flags
& FAULT_FLAG_WRITE
)) {
5545 update_mmu_cache_range(vmf
, vmf
->vma
, vmf
->address
,
5548 /* Skip spurious TLB flush for retried page fault */
5549 if (vmf
->flags
& FAULT_FLAG_TRIED
)
5552 * This is needed only for protection faults but the arch code
5553 * is not yet telling us if this is a protection fault or not.
5554 * This still avoids useless tlb flushes for .text page faults
5557 if (vmf
->flags
& FAULT_FLAG_WRITE
)
5558 flush_tlb_fix_spurious_fault(vmf
->vma
, vmf
->address
,
5562 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5567 * On entry, we hold either the VMA lock or the mmap_lock
5568 * (FAULT_FLAG_VMA_LOCK tells you which). If VM_FAULT_RETRY is set in
5569 * the result, the mmap_lock is not held on exit. See filemap_fault()
5570 * and __folio_lock_or_retry().
5572 static vm_fault_t
__handle_mm_fault(struct vm_area_struct
*vma
,
5573 unsigned long address
, unsigned int flags
)
5575 struct vm_fault vmf
= {
5577 .address
= address
& PAGE_MASK
,
5578 .real_address
= address
,
5580 .pgoff
= linear_page_index(vma
, address
),
5581 .gfp_mask
= __get_fault_gfp_mask(vma
),
5583 struct mm_struct
*mm
= vma
->vm_mm
;
5584 unsigned long vm_flags
= vma
->vm_flags
;
5589 pgd
= pgd_offset(mm
, address
);
5590 p4d
= p4d_alloc(mm
, pgd
, address
);
5592 return VM_FAULT_OOM
;
5594 vmf
.pud
= pud_alloc(mm
, p4d
, address
);
5596 return VM_FAULT_OOM
;
5598 if (pud_none(*vmf
.pud
) &&
5599 thp_vma_allowable_order(vma
, vm_flags
,
5600 TVA_IN_PF
| TVA_ENFORCE_SYSFS
, PUD_ORDER
)) {
5601 ret
= create_huge_pud(&vmf
);
5602 if (!(ret
& VM_FAULT_FALLBACK
))
5605 pud_t orig_pud
= *vmf
.pud
;
5608 if (pud_trans_huge(orig_pud
) || pud_devmap(orig_pud
)) {
5611 * TODO once we support anonymous PUDs: NUMA case and
5612 * FAULT_FLAG_UNSHARE handling.
5614 if ((flags
& FAULT_FLAG_WRITE
) && !pud_write(orig_pud
)) {
5615 ret
= wp_huge_pud(&vmf
, orig_pud
);
5616 if (!(ret
& VM_FAULT_FALLBACK
))
5619 huge_pud_set_accessed(&vmf
, orig_pud
);
5625 vmf
.pmd
= pmd_alloc(mm
, vmf
.pud
, address
);
5627 return VM_FAULT_OOM
;
5629 /* Huge pud page fault raced with pmd_alloc? */
5630 if (pud_trans_unstable(vmf
.pud
))
5633 if (pmd_none(*vmf
.pmd
) &&
5634 thp_vma_allowable_order(vma
, vm_flags
,
5635 TVA_IN_PF
| TVA_ENFORCE_SYSFS
, PMD_ORDER
)) {
5636 ret
= create_huge_pmd(&vmf
);
5637 if (!(ret
& VM_FAULT_FALLBACK
))
5640 vmf
.orig_pmd
= pmdp_get_lockless(vmf
.pmd
);
5642 if (unlikely(is_swap_pmd(vmf
.orig_pmd
))) {
5643 VM_BUG_ON(thp_migration_supported() &&
5644 !is_pmd_migration_entry(vmf
.orig_pmd
));
5645 if (is_pmd_migration_entry(vmf
.orig_pmd
))
5646 pmd_migration_entry_wait(mm
, vmf
.pmd
);
5649 if (pmd_trans_huge(vmf
.orig_pmd
) || pmd_devmap(vmf
.orig_pmd
)) {
5650 if (pmd_protnone(vmf
.orig_pmd
) && vma_is_accessible(vma
))
5651 return do_huge_pmd_numa_page(&vmf
);
5653 if ((flags
& (FAULT_FLAG_WRITE
|FAULT_FLAG_UNSHARE
)) &&
5654 !pmd_write(vmf
.orig_pmd
)) {
5655 ret
= wp_huge_pmd(&vmf
);
5656 if (!(ret
& VM_FAULT_FALLBACK
))
5659 huge_pmd_set_accessed(&vmf
);
5665 return handle_pte_fault(&vmf
);
5669 * mm_account_fault - Do page fault accounting
5670 * @mm: mm from which memcg should be extracted. It can be NULL.
5671 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
5672 * of perf event counters, but we'll still do the per-task accounting to
5673 * the task who triggered this page fault.
5674 * @address: the faulted address.
5675 * @flags: the fault flags.
5676 * @ret: the fault retcode.
5678 * This will take care of most of the page fault accounting. Meanwhile, it
5679 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
5680 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
5681 * still be in per-arch page fault handlers at the entry of page fault.
5683 static inline void mm_account_fault(struct mm_struct
*mm
, struct pt_regs
*regs
,
5684 unsigned long address
, unsigned int flags
,
5689 /* Incomplete faults will be accounted upon completion. */
5690 if (ret
& VM_FAULT_RETRY
)
5694 * To preserve the behavior of older kernels, PGFAULT counters record
5695 * both successful and failed faults, as opposed to perf counters,
5696 * which ignore failed cases.
5698 count_vm_event(PGFAULT
);
5699 count_memcg_event_mm(mm
, PGFAULT
);
5702 * Do not account for unsuccessful faults (e.g. when the address wasn't
5703 * valid). That includes arch_vma_access_permitted() failing before
5704 * reaching here. So this is not a "this many hardware page faults"
5705 * counter. We should use the hw profiling for that.
5707 if (ret
& VM_FAULT_ERROR
)
5711 * We define the fault as a major fault when the final successful fault
5712 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
5713 * handle it immediately previously).
5715 major
= (ret
& VM_FAULT_MAJOR
) || (flags
& FAULT_FLAG_TRIED
);
5723 * If the fault is done for GUP, regs will be NULL. We only do the
5724 * accounting for the per thread fault counters who triggered the
5725 * fault, and we skip the perf event updates.
5731 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ
, 1, regs
, address
);
5733 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN
, 1, regs
, address
);
5736 #ifdef CONFIG_LRU_GEN
5737 static void lru_gen_enter_fault(struct vm_area_struct
*vma
)
5739 /* the LRU algorithm only applies to accesses with recency */
5740 current
->in_lru_fault
= vma_has_recency(vma
);
5743 static void lru_gen_exit_fault(void)
5745 current
->in_lru_fault
= false;
5748 static void lru_gen_enter_fault(struct vm_area_struct
*vma
)
5752 static void lru_gen_exit_fault(void)
5755 #endif /* CONFIG_LRU_GEN */
5757 static vm_fault_t
sanitize_fault_flags(struct vm_area_struct
*vma
,
5758 unsigned int *flags
)
5760 if (unlikely(*flags
& FAULT_FLAG_UNSHARE
)) {
5761 if (WARN_ON_ONCE(*flags
& FAULT_FLAG_WRITE
))
5762 return VM_FAULT_SIGSEGV
;
5764 * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's
5765 * just treat it like an ordinary read-fault otherwise.
5767 if (!is_cow_mapping(vma
->vm_flags
))
5768 *flags
&= ~FAULT_FLAG_UNSHARE
;
5769 } else if (*flags
& FAULT_FLAG_WRITE
) {
5770 /* Write faults on read-only mappings are impossible ... */
5771 if (WARN_ON_ONCE(!(vma
->vm_flags
& VM_MAYWRITE
)))
5772 return VM_FAULT_SIGSEGV
;
5773 /* ... and FOLL_FORCE only applies to COW mappings. */
5774 if (WARN_ON_ONCE(!(vma
->vm_flags
& VM_WRITE
) &&
5775 !is_cow_mapping(vma
->vm_flags
)))
5776 return VM_FAULT_SIGSEGV
;
5778 #ifdef CONFIG_PER_VMA_LOCK
5780 * Per-VMA locks can't be used with FAULT_FLAG_RETRY_NOWAIT because of
5781 * the assumption that lock is dropped on VM_FAULT_RETRY.
5783 if (WARN_ON_ONCE((*flags
&
5784 (FAULT_FLAG_VMA_LOCK
| FAULT_FLAG_RETRY_NOWAIT
)) ==
5785 (FAULT_FLAG_VMA_LOCK
| FAULT_FLAG_RETRY_NOWAIT
)))
5786 return VM_FAULT_SIGSEGV
;
5793 * By the time we get here, we already hold the mm semaphore
5795 * The mmap_lock may have been released depending on flags and our
5796 * return value. See filemap_fault() and __folio_lock_or_retry().
5798 vm_fault_t
handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
5799 unsigned int flags
, struct pt_regs
*regs
)
5801 /* If the fault handler drops the mmap_lock, vma may be freed */
5802 struct mm_struct
*mm
= vma
->vm_mm
;
5806 __set_current_state(TASK_RUNNING
);
5808 ret
= sanitize_fault_flags(vma
, &flags
);
5812 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
5813 flags
& FAULT_FLAG_INSTRUCTION
,
5814 flags
& FAULT_FLAG_REMOTE
)) {
5815 ret
= VM_FAULT_SIGSEGV
;
5819 is_droppable
= !!(vma
->vm_flags
& VM_DROPPABLE
);
5822 * Enable the memcg OOM handling for faults triggered in user
5823 * space. Kernel faults are handled more gracefully.
5825 if (flags
& FAULT_FLAG_USER
)
5826 mem_cgroup_enter_user_fault();
5828 lru_gen_enter_fault(vma
);
5830 if (unlikely(is_vm_hugetlb_page(vma
)))
5831 ret
= hugetlb_fault(vma
->vm_mm
, vma
, address
, flags
);
5833 ret
= __handle_mm_fault(vma
, address
, flags
);
5836 * Warning: It is no longer safe to dereference vma-> after this point,
5837 * because mmap_lock might have been dropped by __handle_mm_fault(), so
5838 * vma might be destroyed from underneath us.
5841 lru_gen_exit_fault();
5843 /* If the mapping is droppable, then errors due to OOM aren't fatal. */
5845 ret
&= ~VM_FAULT_OOM
;
5847 if (flags
& FAULT_FLAG_USER
) {
5848 mem_cgroup_exit_user_fault();
5850 * The task may have entered a memcg OOM situation but
5851 * if the allocation error was handled gracefully (no
5852 * VM_FAULT_OOM), there is no need to kill anything.
5853 * Just clean up the OOM state peacefully.
5855 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
5856 mem_cgroup_oom_synchronize(false);
5859 mm_account_fault(mm
, regs
, address
, flags
, ret
);
5863 EXPORT_SYMBOL_GPL(handle_mm_fault
);
5865 #ifdef CONFIG_LOCK_MM_AND_FIND_VMA
5866 #include <linux/extable.h>
5868 static inline bool get_mmap_lock_carefully(struct mm_struct
*mm
, struct pt_regs
*regs
)
5870 if (likely(mmap_read_trylock(mm
)))
5873 if (regs
&& !user_mode(regs
)) {
5874 unsigned long ip
= exception_ip(regs
);
5875 if (!search_exception_tables(ip
))
5879 return !mmap_read_lock_killable(mm
);
5882 static inline bool mmap_upgrade_trylock(struct mm_struct
*mm
)
5885 * We don't have this operation yet.
5887 * It should be easy enough to do: it's basically a
5888 * atomic_long_try_cmpxchg_acquire()
5889 * from RWSEM_READER_BIAS -> RWSEM_WRITER_LOCKED, but
5890 * it also needs the proper lockdep magic etc.
5895 static inline bool upgrade_mmap_lock_carefully(struct mm_struct
*mm
, struct pt_regs
*regs
)
5897 mmap_read_unlock(mm
);
5898 if (regs
&& !user_mode(regs
)) {
5899 unsigned long ip
= exception_ip(regs
);
5900 if (!search_exception_tables(ip
))
5903 return !mmap_write_lock_killable(mm
);
5907 * Helper for page fault handling.
5909 * This is kind of equivalend to "mmap_read_lock()" followed
5910 * by "find_extend_vma()", except it's a lot more careful about
5911 * the locking (and will drop the lock on failure).
5913 * For example, if we have a kernel bug that causes a page
5914 * fault, we don't want to just use mmap_read_lock() to get
5915 * the mm lock, because that would deadlock if the bug were
5916 * to happen while we're holding the mm lock for writing.
5918 * So this checks the exception tables on kernel faults in
5919 * order to only do this all for instructions that are actually
5920 * expected to fault.
5922 * We can also actually take the mm lock for writing if we
5923 * need to extend the vma, which helps the VM layer a lot.
5925 struct vm_area_struct
*lock_mm_and_find_vma(struct mm_struct
*mm
,
5926 unsigned long addr
, struct pt_regs
*regs
)
5928 struct vm_area_struct
*vma
;
5930 if (!get_mmap_lock_carefully(mm
, regs
))
5933 vma
= find_vma(mm
, addr
);
5934 if (likely(vma
&& (vma
->vm_start
<= addr
)))
5938 * Well, dang. We might still be successful, but only
5939 * if we can extend a vma to do so.
5941 if (!vma
|| !(vma
->vm_flags
& VM_GROWSDOWN
)) {
5942 mmap_read_unlock(mm
);
5947 * We can try to upgrade the mmap lock atomically,
5948 * in which case we can continue to use the vma
5949 * we already looked up.
5951 * Otherwise we'll have to drop the mmap lock and
5952 * re-take it, and also look up the vma again,
5955 if (!mmap_upgrade_trylock(mm
)) {
5956 if (!upgrade_mmap_lock_carefully(mm
, regs
))
5959 vma
= find_vma(mm
, addr
);
5962 if (vma
->vm_start
<= addr
)
5964 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
5968 if (expand_stack_locked(vma
, addr
))
5972 mmap_write_downgrade(mm
);
5976 mmap_write_unlock(mm
);
5981 #ifdef CONFIG_PER_VMA_LOCK
5983 * Lookup and lock a VMA under RCU protection. Returned VMA is guaranteed to be
5984 * stable and not isolated. If the VMA is not found or is being modified the
5985 * function returns NULL.
5987 struct vm_area_struct
*lock_vma_under_rcu(struct mm_struct
*mm
,
5988 unsigned long address
)
5990 MA_STATE(mas
, &mm
->mm_mt
, address
, address
);
5991 struct vm_area_struct
*vma
;
5995 vma
= mas_walk(&mas
);
5999 if (!vma_start_read(vma
))
6002 /* Check since vm_start/vm_end might change before we lock the VMA */
6003 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
))
6004 goto inval_end_read
;
6006 /* Check if the VMA got isolated after we found it */
6007 if (vma
->detached
) {
6009 count_vm_vma_lock_event(VMA_LOCK_MISS
);
6010 /* The area was replaced with another one */
6021 count_vm_vma_lock_event(VMA_LOCK_ABORT
);
6024 #endif /* CONFIG_PER_VMA_LOCK */
6026 #ifndef __PAGETABLE_P4D_FOLDED
6028 * Allocate p4d page table.
6029 * We've already handled the fast-path in-line.
6031 int __p4d_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
6033 p4d_t
*new = p4d_alloc_one(mm
, address
);
6037 spin_lock(&mm
->page_table_lock
);
6038 if (pgd_present(*pgd
)) { /* Another has populated it */
6041 smp_wmb(); /* See comment in pmd_install() */
6042 pgd_populate(mm
, pgd
, new);
6044 spin_unlock(&mm
->page_table_lock
);
6047 #endif /* __PAGETABLE_P4D_FOLDED */
6049 #ifndef __PAGETABLE_PUD_FOLDED
6051 * Allocate page upper directory.
6052 * We've already handled the fast-path in-line.
6054 int __pud_alloc(struct mm_struct
*mm
, p4d_t
*p4d
, unsigned long address
)
6056 pud_t
*new = pud_alloc_one(mm
, address
);
6060 spin_lock(&mm
->page_table_lock
);
6061 if (!p4d_present(*p4d
)) {
6063 smp_wmb(); /* See comment in pmd_install() */
6064 p4d_populate(mm
, p4d
, new);
6065 } else /* Another has populated it */
6067 spin_unlock(&mm
->page_table_lock
);
6070 #endif /* __PAGETABLE_PUD_FOLDED */
6072 #ifndef __PAGETABLE_PMD_FOLDED
6074 * Allocate page middle directory.
6075 * We've already handled the fast-path in-line.
6077 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
6080 pmd_t
*new = pmd_alloc_one(mm
, address
);
6084 ptl
= pud_lock(mm
, pud
);
6085 if (!pud_present(*pud
)) {
6087 smp_wmb(); /* See comment in pmd_install() */
6088 pud_populate(mm
, pud
, new);
6089 } else { /* Another has populated it */
6095 #endif /* __PAGETABLE_PMD_FOLDED */
6098 * follow_pte - look up PTE at a user virtual address
6099 * @vma: the memory mapping
6100 * @address: user virtual address
6101 * @ptepp: location to store found PTE
6102 * @ptlp: location to store the lock for the PTE
6104 * On a successful return, the pointer to the PTE is stored in @ptepp;
6105 * the corresponding lock is taken and its location is stored in @ptlp.
6107 * The contents of the PTE are only stable until @ptlp is released using
6108 * pte_unmap_unlock(). This function will fail if the PTE is non-present.
6109 * Present PTEs may include PTEs that map refcounted pages, such as
6110 * anonymous folios in COW mappings.
6112 * Callers must be careful when relying on PTE content after
6113 * pte_unmap_unlock(). Especially if the PTE maps a refcounted page,
6114 * callers must protect against invalidation with MMU notifiers; otherwise
6115 * access to the PFN at a later point in time can trigger use-after-free.
6117 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
6118 * should be taken for read.
6120 * This function must not be used to modify PTE content.
6122 * Return: zero on success, -ve otherwise.
6124 int follow_pte(struct vm_area_struct
*vma
, unsigned long address
,
6125 pte_t
**ptepp
, spinlock_t
**ptlp
)
6127 struct mm_struct
*mm
= vma
->vm_mm
;
6134 mmap_assert_locked(mm
);
6135 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
))
6138 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
6141 pgd
= pgd_offset(mm
, address
);
6142 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
6145 p4d
= p4d_offset(pgd
, address
);
6146 if (p4d_none(*p4d
) || unlikely(p4d_bad(*p4d
)))
6149 pud
= pud_offset(p4d
, address
);
6150 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
6153 pmd
= pmd_offset(pud
, address
);
6154 VM_BUG_ON(pmd_trans_huge(*pmd
));
6156 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
6159 if (!pte_present(ptep_get(ptep
)))
6164 pte_unmap_unlock(ptep
, *ptlp
);
6168 EXPORT_SYMBOL_GPL(follow_pte
);
6170 #ifdef CONFIG_HAVE_IOREMAP_PROT
6172 * generic_access_phys - generic implementation for iomem mmap access
6173 * @vma: the vma to access
6174 * @addr: userspace address, not relative offset within @vma
6175 * @buf: buffer to read/write
6176 * @len: length of transfer
6177 * @write: set to FOLL_WRITE when writing, otherwise reading
6179 * This is a generic implementation for &vm_operations_struct.access for an
6180 * iomem mapping. This callback is used by access_process_vm() when the @vma is
6183 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
6184 void *buf
, int len
, int write
)
6186 resource_size_t phys_addr
;
6187 unsigned long prot
= 0;
6188 void __iomem
*maddr
;
6191 int offset
= offset_in_page(addr
);
6195 if (follow_pte(vma
, addr
, &ptep
, &ptl
))
6197 pte
= ptep_get(ptep
);
6198 pte_unmap_unlock(ptep
, ptl
);
6200 prot
= pgprot_val(pte_pgprot(pte
));
6201 phys_addr
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
6203 if ((write
& FOLL_WRITE
) && !pte_write(pte
))
6206 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
6210 if (follow_pte(vma
, addr
, &ptep
, &ptl
))
6213 if (!pte_same(pte
, ptep_get(ptep
))) {
6214 pte_unmap_unlock(ptep
, ptl
);
6221 memcpy_toio(maddr
+ offset
, buf
, len
);
6223 memcpy_fromio(buf
, maddr
+ offset
, len
);
6225 pte_unmap_unlock(ptep
, ptl
);
6231 EXPORT_SYMBOL_GPL(generic_access_phys
);
6235 * Access another process' address space as given in mm.
6237 static int __access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
6238 void *buf
, int len
, unsigned int gup_flags
)
6240 void *old_buf
= buf
;
6241 int write
= gup_flags
& FOLL_WRITE
;
6243 if (mmap_read_lock_killable(mm
))
6246 /* Untag the address before looking up the VMA */
6247 addr
= untagged_addr_remote(mm
, addr
);
6249 /* Avoid triggering the temporary warning in __get_user_pages */
6250 if (!vma_lookup(mm
, addr
) && !expand_stack(mm
, addr
))
6253 /* ignore errors, just check how much was successfully transferred */
6257 struct vm_area_struct
*vma
= NULL
;
6258 struct page
*page
= get_user_page_vma_remote(mm
, addr
,
6262 /* We might need to expand the stack to access it */
6263 vma
= vma_lookup(mm
, addr
);
6265 vma
= expand_stack(mm
, addr
);
6267 /* mmap_lock was dropped on failure */
6269 return buf
- old_buf
;
6271 /* Try again if stack expansion worked */
6276 * Check if this is a VM_IO | VM_PFNMAP VMA, which
6277 * we can access using slightly different code.
6280 #ifdef CONFIG_HAVE_IOREMAP_PROT
6281 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
6282 bytes
= vma
->vm_ops
->access(vma
, addr
, buf
,
6289 offset
= addr
& (PAGE_SIZE
-1);
6290 if (bytes
> PAGE_SIZE
-offset
)
6291 bytes
= PAGE_SIZE
-offset
;
6293 maddr
= kmap_local_page(page
);
6295 copy_to_user_page(vma
, page
, addr
,
6296 maddr
+ offset
, buf
, bytes
);
6297 set_page_dirty_lock(page
);
6299 copy_from_user_page(vma
, page
, addr
,
6300 buf
, maddr
+ offset
, bytes
);
6302 unmap_and_put_page(page
, maddr
);
6308 mmap_read_unlock(mm
);
6310 return buf
- old_buf
;
6314 * access_remote_vm - access another process' address space
6315 * @mm: the mm_struct of the target address space
6316 * @addr: start address to access
6317 * @buf: source or destination buffer
6318 * @len: number of bytes to transfer
6319 * @gup_flags: flags modifying lookup behaviour
6321 * The caller must hold a reference on @mm.
6323 * Return: number of bytes copied from source to destination.
6325 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
6326 void *buf
, int len
, unsigned int gup_flags
)
6328 return __access_remote_vm(mm
, addr
, buf
, len
, gup_flags
);
6332 * Access another process' address space.
6333 * Source/target buffer must be kernel space,
6334 * Do not walk the page table directly, use get_user_pages
6336 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
6337 void *buf
, int len
, unsigned int gup_flags
)
6339 struct mm_struct
*mm
;
6342 mm
= get_task_mm(tsk
);
6346 ret
= __access_remote_vm(mm
, addr
, buf
, len
, gup_flags
);
6352 EXPORT_SYMBOL_GPL(access_process_vm
);
6355 * Print the name of a VMA.
6357 void print_vma_addr(char *prefix
, unsigned long ip
)
6359 struct mm_struct
*mm
= current
->mm
;
6360 struct vm_area_struct
*vma
;
6363 * we might be running from an atomic context so we cannot sleep
6365 if (!mmap_read_trylock(mm
))
6368 vma
= vma_lookup(mm
, ip
);
6369 if (vma
&& vma
->vm_file
) {
6370 struct file
*f
= vma
->vm_file
;
6371 ip
-= vma
->vm_start
;
6372 ip
+= vma
->vm_pgoff
<< PAGE_SHIFT
;
6373 printk("%s%pD[%lx,%lx+%lx]", prefix
, f
, ip
,
6375 vma
->vm_end
- vma
->vm_start
);
6377 mmap_read_unlock(mm
);
6380 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6381 void __might_fault(const char *file
, int line
)
6383 if (pagefault_disabled())
6385 __might_sleep(file
, line
);
6386 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6388 might_lock_read(¤t
->mm
->mmap_lock
);
6391 EXPORT_SYMBOL(__might_fault
);
6394 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
6396 * Process all subpages of the specified huge page with the specified
6397 * operation. The target subpage will be processed last to keep its
6400 static inline int process_huge_page(
6401 unsigned long addr_hint
, unsigned int nr_pages
,
6402 int (*process_subpage
)(unsigned long addr
, int idx
, void *arg
),
6405 int i
, n
, base
, l
, ret
;
6406 unsigned long addr
= addr_hint
&
6407 ~(((unsigned long)nr_pages
<< PAGE_SHIFT
) - 1);
6409 /* Process target subpage last to keep its cache lines hot */
6411 n
= (addr_hint
- addr
) / PAGE_SIZE
;
6412 if (2 * n
<= nr_pages
) {
6413 /* If target subpage in first half of huge page */
6416 /* Process subpages at the end of huge page */
6417 for (i
= nr_pages
- 1; i
>= 2 * n
; i
--) {
6419 ret
= process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
6424 /* If target subpage in second half of huge page */
6425 base
= nr_pages
- 2 * (nr_pages
- n
);
6427 /* Process subpages at the begin of huge page */
6428 for (i
= 0; i
< base
; i
++) {
6430 ret
= process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
6436 * Process remaining subpages in left-right-left-right pattern
6437 * towards the target subpage
6439 for (i
= 0; i
< l
; i
++) {
6440 int left_idx
= base
+ i
;
6441 int right_idx
= base
+ 2 * l
- 1 - i
;
6444 ret
= process_subpage(addr
+ left_idx
* PAGE_SIZE
, left_idx
, arg
);
6448 ret
= process_subpage(addr
+ right_idx
* PAGE_SIZE
, right_idx
, arg
);
6455 static void clear_gigantic_page(struct folio
*folio
, unsigned long addr
,
6456 unsigned int nr_pages
)
6461 for (i
= 0; i
< nr_pages
; i
++) {
6463 clear_user_highpage(folio_page(folio
, i
), addr
+ i
* PAGE_SIZE
);
6467 static int clear_subpage(unsigned long addr
, int idx
, void *arg
)
6469 struct folio
*folio
= arg
;
6471 clear_user_highpage(folio_page(folio
, idx
), addr
);
6476 * folio_zero_user - Zero a folio which will be mapped to userspace.
6477 * @folio: The folio to zero.
6478 * @addr_hint: The address will be accessed or the base address if uncelar.
6480 void folio_zero_user(struct folio
*folio
, unsigned long addr_hint
)
6482 unsigned int nr_pages
= folio_nr_pages(folio
);
6484 if (unlikely(nr_pages
> MAX_ORDER_NR_PAGES
))
6485 clear_gigantic_page(folio
, addr_hint
, nr_pages
);
6487 process_huge_page(addr_hint
, nr_pages
, clear_subpage
, folio
);
6490 static int copy_user_gigantic_page(struct folio
*dst
, struct folio
*src
,
6492 struct vm_area_struct
*vma
,
6493 unsigned int nr_pages
)
6496 struct page
*dst_page
;
6497 struct page
*src_page
;
6499 for (i
= 0; i
< nr_pages
; i
++) {
6500 dst_page
= folio_page(dst
, i
);
6501 src_page
= folio_page(src
, i
);
6504 if (copy_mc_user_highpage(dst_page
, src_page
,
6505 addr
+ i
*PAGE_SIZE
, vma
))
6511 struct copy_subpage_arg
{
6514 struct vm_area_struct
*vma
;
6517 static int copy_subpage(unsigned long addr
, int idx
, void *arg
)
6519 struct copy_subpage_arg
*copy_arg
= arg
;
6520 struct page
*dst
= folio_page(copy_arg
->dst
, idx
);
6521 struct page
*src
= folio_page(copy_arg
->src
, idx
);
6523 if (copy_mc_user_highpage(dst
, src
, addr
, copy_arg
->vma
))
6528 int copy_user_large_folio(struct folio
*dst
, struct folio
*src
,
6529 unsigned long addr_hint
, struct vm_area_struct
*vma
)
6531 unsigned int nr_pages
= folio_nr_pages(dst
);
6532 struct copy_subpage_arg arg
= {
6538 if (unlikely(nr_pages
> MAX_ORDER_NR_PAGES
))
6539 return copy_user_gigantic_page(dst
, src
, addr_hint
, vma
, nr_pages
);
6541 return process_huge_page(addr_hint
, nr_pages
, copy_subpage
, &arg
);
6544 long copy_folio_from_user(struct folio
*dst_folio
,
6545 const void __user
*usr_src
,
6546 bool allow_pagefault
)
6549 unsigned long i
, rc
= 0;
6550 unsigned int nr_pages
= folio_nr_pages(dst_folio
);
6551 unsigned long ret_val
= nr_pages
* PAGE_SIZE
;
6552 struct page
*subpage
;
6554 for (i
= 0; i
< nr_pages
; i
++) {
6555 subpage
= folio_page(dst_folio
, i
);
6556 kaddr
= kmap_local_page(subpage
);
6557 if (!allow_pagefault
)
6558 pagefault_disable();
6559 rc
= copy_from_user(kaddr
, usr_src
+ i
* PAGE_SIZE
, PAGE_SIZE
);
6560 if (!allow_pagefault
)
6562 kunmap_local(kaddr
);
6564 ret_val
-= (PAGE_SIZE
- rc
);
6568 flush_dcache_page(subpage
);
6574 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
6576 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
6578 static struct kmem_cache
*page_ptl_cachep
;
6580 void __init
ptlock_cache_init(void)
6582 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
6586 bool ptlock_alloc(struct ptdesc
*ptdesc
)
6590 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
6597 void ptlock_free(struct ptdesc
*ptdesc
)
6599 kmem_cache_free(page_ptl_cachep
, ptdesc
->ptl
);
6603 void vma_pgtable_walk_begin(struct vm_area_struct
*vma
)
6605 if (is_vm_hugetlb_page(vma
))
6606 hugetlb_vma_lock_read(vma
);
6609 void vma_pgtable_walk_end(struct vm_area_struct
*vma
)
6611 if (is_vm_hugetlb_page(vma
))
6612 hugetlb_vma_unlock_read(vma
);