1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * demand-loading started 01.12.91 - seems it is high on the list of
10 * things wanted, and it should be easy to implement. - Linus
14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
15 * pages started 02.12.91, seems to work. - Linus.
17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
18 * would have taken more than the 6M I have free, but it worked well as
21 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
25 * Real VM (paging to/from disk) started 18.12.91. Much more work and
26 * thought has to go into this. Oh, well..
27 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
28 * Found it. Everything seems to work now.
29 * 20.12.91 - Ok, making the swap-device changeable like the root.
33 * 05.04.94 - Multi-page memory management added for v1.1.
34 * Idea by Alex Bligh (alex@cconcepts.co.uk)
36 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
37 * (Gerhard.Wichert@pdb.siemens.de)
39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
42 #include <linux/kernel_stat.h>
44 #include <linux/mm_inline.h>
45 #include <linux/sched/mm.h>
46 #include <linux/sched/numa_balancing.h>
47 #include <linux/sched/task.h>
48 #include <linux/hugetlb.h>
49 #include <linux/mman.h>
50 #include <linux/swap.h>
51 #include <linux/highmem.h>
52 #include <linux/pagemap.h>
53 #include <linux/memremap.h>
54 #include <linux/kmsan.h>
55 #include <linux/ksm.h>
56 #include <linux/rmap.h>
57 #include <linux/export.h>
58 #include <linux/delayacct.h>
59 #include <linux/init.h>
60 #include <linux/pfn_t.h>
61 #include <linux/writeback.h>
62 #include <linux/memcontrol.h>
63 #include <linux/mmu_notifier.h>
64 #include <linux/swapops.h>
65 #include <linux/elf.h>
66 #include <linux/gfp.h>
67 #include <linux/migrate.h>
68 #include <linux/string.h>
69 #include <linux/memory-tiers.h>
70 #include <linux/debugfs.h>
71 #include <linux/userfaultfd_k.h>
72 #include <linux/dax.h>
73 #include <linux/oom.h>
74 #include <linux/numa.h>
75 #include <linux/perf_event.h>
76 #include <linux/ptrace.h>
77 #include <linux/vmalloc.h>
78 #include <linux/sched/sysctl.h>
80 #include <trace/events/kmem.h>
83 #include <asm/mmu_context.h>
84 #include <asm/pgalloc.h>
85 #include <linux/uaccess.h>
87 #include <asm/tlbflush.h>
89 #include "pgalloc-track.h"
93 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
94 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
98 unsigned long max_mapnr
;
99 EXPORT_SYMBOL(max_mapnr
);
101 struct page
*mem_map
;
102 EXPORT_SYMBOL(mem_map
);
105 static vm_fault_t
do_fault(struct vm_fault
*vmf
);
106 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
);
107 static bool vmf_pte_changed(struct vm_fault
*vmf
);
110 * Return true if the original pte was a uffd-wp pte marker (so the pte was
113 static __always_inline
bool vmf_orig_pte_uffd_wp(struct vm_fault
*vmf
)
115 if (!userfaultfd_wp(vmf
->vma
))
117 if (!(vmf
->flags
& FAULT_FLAG_ORIG_PTE_VALID
))
120 return pte_marker_uffd_wp(vmf
->orig_pte
);
124 * A number of key systems in x86 including ioremap() rely on the assumption
125 * that high_memory defines the upper bound on direct map memory, then end
129 EXPORT_SYMBOL(high_memory
);
132 * Randomize the address space (stacks, mmaps, brk, etc.).
134 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
135 * as ancient (libc5 based) binaries can segfault. )
137 int randomize_va_space __read_mostly
=
138 #ifdef CONFIG_COMPAT_BRK
144 #ifndef arch_wants_old_prefaulted_pte
145 static inline bool arch_wants_old_prefaulted_pte(void)
148 * Transitioning a PTE from 'old' to 'young' can be expensive on
149 * some architectures, even if it's performed in hardware. By
150 * default, "false" means prefaulted entries will be 'young'.
156 static int __init
disable_randmaps(char *s
)
158 randomize_va_space
= 0;
161 __setup("norandmaps", disable_randmaps
);
163 unsigned long zero_pfn __read_mostly
;
164 EXPORT_SYMBOL(zero_pfn
);
166 unsigned long highest_memmap_pfn __read_mostly
;
169 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
171 static int __init
init_zero_pfn(void)
173 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
176 early_initcall(init_zero_pfn
);
178 void mm_trace_rss_stat(struct mm_struct
*mm
, int member
)
180 trace_rss_stat(mm
, member
);
184 * Note: this doesn't free the actual pages themselves. That
185 * has been handled earlier when unmapping all the memory regions.
187 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
190 pgtable_t token
= pmd_pgtable(*pmd
);
192 pte_free_tlb(tlb
, token
, addr
);
193 mm_dec_nr_ptes(tlb
->mm
);
196 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
197 unsigned long addr
, unsigned long end
,
198 unsigned long floor
, unsigned long ceiling
)
205 pmd
= pmd_offset(pud
, addr
);
207 next
= pmd_addr_end(addr
, end
);
208 if (pmd_none_or_clear_bad(pmd
))
210 free_pte_range(tlb
, pmd
, addr
);
211 } while (pmd
++, addr
= next
, addr
!= end
);
221 if (end
- 1 > ceiling
- 1)
224 pmd
= pmd_offset(pud
, start
);
226 pmd_free_tlb(tlb
, pmd
, start
);
227 mm_dec_nr_pmds(tlb
->mm
);
230 static inline void free_pud_range(struct mmu_gather
*tlb
, p4d_t
*p4d
,
231 unsigned long addr
, unsigned long end
,
232 unsigned long floor
, unsigned long ceiling
)
239 pud
= pud_offset(p4d
, addr
);
241 next
= pud_addr_end(addr
, end
);
242 if (pud_none_or_clear_bad(pud
))
244 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
245 } while (pud
++, addr
= next
, addr
!= end
);
255 if (end
- 1 > ceiling
- 1)
258 pud
= pud_offset(p4d
, start
);
260 pud_free_tlb(tlb
, pud
, start
);
261 mm_dec_nr_puds(tlb
->mm
);
264 static inline void free_p4d_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
265 unsigned long addr
, unsigned long end
,
266 unsigned long floor
, unsigned long ceiling
)
273 p4d
= p4d_offset(pgd
, addr
);
275 next
= p4d_addr_end(addr
, end
);
276 if (p4d_none_or_clear_bad(p4d
))
278 free_pud_range(tlb
, p4d
, addr
, next
, floor
, ceiling
);
279 } while (p4d
++, addr
= next
, addr
!= end
);
285 ceiling
&= PGDIR_MASK
;
289 if (end
- 1 > ceiling
- 1)
292 p4d
= p4d_offset(pgd
, start
);
294 p4d_free_tlb(tlb
, p4d
, start
);
298 * This function frees user-level page tables of a process.
300 void free_pgd_range(struct mmu_gather
*tlb
,
301 unsigned long addr
, unsigned long end
,
302 unsigned long floor
, unsigned long ceiling
)
308 * The next few lines have given us lots of grief...
310 * Why are we testing PMD* at this top level? Because often
311 * there will be no work to do at all, and we'd prefer not to
312 * go all the way down to the bottom just to discover that.
314 * Why all these "- 1"s? Because 0 represents both the bottom
315 * of the address space and the top of it (using -1 for the
316 * top wouldn't help much: the masks would do the wrong thing).
317 * The rule is that addr 0 and floor 0 refer to the bottom of
318 * the address space, but end 0 and ceiling 0 refer to the top
319 * Comparisons need to use "end - 1" and "ceiling - 1" (though
320 * that end 0 case should be mythical).
322 * Wherever addr is brought up or ceiling brought down, we must
323 * be careful to reject "the opposite 0" before it confuses the
324 * subsequent tests. But what about where end is brought down
325 * by PMD_SIZE below? no, end can't go down to 0 there.
327 * Whereas we round start (addr) and ceiling down, by different
328 * masks at different levels, in order to test whether a table
329 * now has no other vmas using it, so can be freed, we don't
330 * bother to round floor or end up - the tests don't need that.
344 if (end
- 1 > ceiling
- 1)
349 * We add page table cache pages with PAGE_SIZE,
350 * (see pte_free_tlb()), flush the tlb if we need
352 tlb_change_page_size(tlb
, PAGE_SIZE
);
353 pgd
= pgd_offset(tlb
->mm
, addr
);
355 next
= pgd_addr_end(addr
, end
);
356 if (pgd_none_or_clear_bad(pgd
))
358 free_p4d_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
359 } while (pgd
++, addr
= next
, addr
!= end
);
362 void free_pgtables(struct mmu_gather
*tlb
, struct ma_state
*mas
,
363 struct vm_area_struct
*vma
, unsigned long floor
,
364 unsigned long ceiling
, bool mm_wr_locked
)
366 struct unlink_vma_file_batch vb
;
369 unsigned long addr
= vma
->vm_start
;
370 struct vm_area_struct
*next
;
373 * Note: USER_PGTABLES_CEILING may be passed as ceiling and may
374 * be 0. This will underflow and is okay.
376 next
= mas_find(mas
, ceiling
- 1);
377 if (unlikely(xa_is_zero(next
)))
381 * Hide vma from rmap and truncate_pagecache before freeing
385 vma_start_write(vma
);
386 unlink_anon_vmas(vma
);
388 if (is_vm_hugetlb_page(vma
)) {
389 unlink_file_vma(vma
);
390 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
391 floor
, next
? next
->vm_start
: ceiling
);
393 unlink_file_vma_batch_init(&vb
);
394 unlink_file_vma_batch_add(&vb
, vma
);
397 * Optimization: gather nearby vmas into one call down
399 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
400 && !is_vm_hugetlb_page(next
)) {
402 next
= mas_find(mas
, ceiling
- 1);
403 if (unlikely(xa_is_zero(next
)))
406 vma_start_write(vma
);
407 unlink_anon_vmas(vma
);
408 unlink_file_vma_batch_add(&vb
, vma
);
410 unlink_file_vma_batch_final(&vb
);
411 free_pgd_range(tlb
, addr
, vma
->vm_end
,
412 floor
, next
? next
->vm_start
: ceiling
);
418 void pmd_install(struct mm_struct
*mm
, pmd_t
*pmd
, pgtable_t
*pte
)
420 spinlock_t
*ptl
= pmd_lock(mm
, pmd
);
422 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
425 * Ensure all pte setup (eg. pte page lock and page clearing) are
426 * visible before the pte is made visible to other CPUs by being
427 * put into page tables.
429 * The other side of the story is the pointer chasing in the page
430 * table walking code (when walking the page table without locking;
431 * ie. most of the time). Fortunately, these data accesses consist
432 * of a chain of data-dependent loads, meaning most CPUs (alpha
433 * being the notable exception) will already guarantee loads are
434 * seen in-order. See the alpha page table accessors for the
435 * smp_rmb() barriers in page table walking code.
437 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
438 pmd_populate(mm
, pmd
, *pte
);
444 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
)
446 pgtable_t
new = pte_alloc_one(mm
);
450 pmd_install(mm
, pmd
, &new);
456 int __pte_alloc_kernel(pmd_t
*pmd
)
458 pte_t
*new = pte_alloc_one_kernel(&init_mm
);
462 spin_lock(&init_mm
.page_table_lock
);
463 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
464 smp_wmb(); /* See comment in pmd_install() */
465 pmd_populate_kernel(&init_mm
, pmd
, new);
468 spin_unlock(&init_mm
.page_table_lock
);
470 pte_free_kernel(&init_mm
, new);
474 static inline void init_rss_vec(int *rss
)
476 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
479 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
483 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
485 add_mm_counter(mm
, i
, rss
[i
]);
489 * This function is called to print an error when a bad pte
490 * is found. For example, we might have a PFN-mapped pte in
491 * a region that doesn't allow it.
493 * The calling function must still handle the error.
495 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
496 pte_t pte
, struct page
*page
)
498 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
499 p4d_t
*p4d
= p4d_offset(pgd
, addr
);
500 pud_t
*pud
= pud_offset(p4d
, addr
);
501 pmd_t
*pmd
= pmd_offset(pud
, addr
);
502 struct address_space
*mapping
;
504 static unsigned long resume
;
505 static unsigned long nr_shown
;
506 static unsigned long nr_unshown
;
509 * Allow a burst of 60 reports, then keep quiet for that minute;
510 * or allow a steady drip of one report per second.
512 if (nr_shown
== 60) {
513 if (time_before(jiffies
, resume
)) {
518 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
525 resume
= jiffies
+ 60 * HZ
;
527 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
528 index
= linear_page_index(vma
, addr
);
530 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
532 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
534 dump_page(page
, "bad pte");
535 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
536 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
537 pr_alert("file:%pD fault:%ps mmap:%ps read_folio:%ps\n",
539 vma
->vm_ops
? vma
->vm_ops
->fault
: NULL
,
540 vma
->vm_file
? vma
->vm_file
->f_op
->mmap
: NULL
,
541 mapping
? mapping
->a_ops
->read_folio
: NULL
);
543 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
547 * vm_normal_page -- This function gets the "struct page" associated with a pte.
549 * "Special" mappings do not wish to be associated with a "struct page" (either
550 * it doesn't exist, or it exists but they don't want to touch it). In this
551 * case, NULL is returned here. "Normal" mappings do have a struct page.
553 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
554 * pte bit, in which case this function is trivial. Secondly, an architecture
555 * may not have a spare pte bit, which requires a more complicated scheme,
558 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
559 * special mapping (even if there are underlying and valid "struct pages").
560 * COWed pages of a VM_PFNMAP are always normal.
562 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
563 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
564 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
565 * mapping will always honor the rule
567 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
569 * And for normal mappings this is false.
571 * This restricts such mappings to be a linear translation from virtual address
572 * to pfn. To get around this restriction, we allow arbitrary mappings so long
573 * as the vma is not a COW mapping; in that case, we know that all ptes are
574 * special (because none can have been COWed).
577 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
579 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
580 * page" backing, however the difference is that _all_ pages with a struct
581 * page (that is, those where pfn_valid is true) are refcounted and considered
582 * normal pages by the VM. The only exception are zeropages, which are
583 * *never* refcounted.
585 * The disadvantage is that pages are refcounted (which can be slower and
586 * simply not an option for some PFNMAP users). The advantage is that we
587 * don't have to follow the strict linearity rule of PFNMAP mappings in
588 * order to support COWable mappings.
591 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
594 unsigned long pfn
= pte_pfn(pte
);
596 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
)) {
597 if (likely(!pte_special(pte
)))
599 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
600 return vma
->vm_ops
->find_special_page(vma
, addr
);
601 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
603 if (is_zero_pfn(pfn
))
607 * NOTE: New users of ZONE_DEVICE will not set pte_devmap()
608 * and will have refcounts incremented on their struct pages
609 * when they are inserted into PTEs, thus they are safe to
610 * return here. Legacy ZONE_DEVICE pages that set pte_devmap()
611 * do not have refcounts. Example of legacy ZONE_DEVICE is
612 * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers.
616 print_bad_pte(vma
, addr
, pte
, NULL
);
620 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
622 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
623 if (vma
->vm_flags
& VM_MIXEDMAP
) {
626 if (is_zero_pfn(pfn
))
631 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
632 if (pfn
== vma
->vm_pgoff
+ off
)
634 if (!is_cow_mapping(vma
->vm_flags
))
639 if (is_zero_pfn(pfn
))
643 if (unlikely(pfn
> highest_memmap_pfn
)) {
644 print_bad_pte(vma
, addr
, pte
, NULL
);
649 * NOTE! We still have PageReserved() pages in the page tables.
650 * eg. VDSO mappings can cause them to exist.
653 VM_WARN_ON_ONCE(is_zero_pfn(pfn
));
654 return pfn_to_page(pfn
);
657 struct folio
*vm_normal_folio(struct vm_area_struct
*vma
, unsigned long addr
,
660 struct page
*page
= vm_normal_page(vma
, addr
, pte
);
663 return page_folio(page
);
667 #ifdef CONFIG_PGTABLE_HAS_HUGE_LEAVES
668 struct page
*vm_normal_page_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
671 unsigned long pfn
= pmd_pfn(pmd
);
673 /* Currently it's only used for huge pfnmaps */
674 if (unlikely(pmd_special(pmd
)))
677 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
678 if (vma
->vm_flags
& VM_MIXEDMAP
) {
684 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
685 if (pfn
== vma
->vm_pgoff
+ off
)
687 if (!is_cow_mapping(vma
->vm_flags
))
694 if (is_huge_zero_pmd(pmd
))
696 if (unlikely(pfn
> highest_memmap_pfn
))
700 * NOTE! We still have PageReserved() pages in the page tables.
701 * eg. VDSO mappings can cause them to exist.
704 return pfn_to_page(pfn
);
707 struct folio
*vm_normal_folio_pmd(struct vm_area_struct
*vma
,
708 unsigned long addr
, pmd_t pmd
)
710 struct page
*page
= vm_normal_page_pmd(vma
, addr
, pmd
);
713 return page_folio(page
);
718 static void restore_exclusive_pte(struct vm_area_struct
*vma
,
719 struct page
*page
, unsigned long address
,
722 struct folio
*folio
= page_folio(page
);
727 orig_pte
= ptep_get(ptep
);
728 pte
= pte_mkold(mk_pte(page
, READ_ONCE(vma
->vm_page_prot
)));
729 if (pte_swp_soft_dirty(orig_pte
))
730 pte
= pte_mksoft_dirty(pte
);
732 entry
= pte_to_swp_entry(orig_pte
);
733 if (pte_swp_uffd_wp(orig_pte
))
734 pte
= pte_mkuffd_wp(pte
);
735 else if (is_writable_device_exclusive_entry(entry
))
736 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
738 VM_BUG_ON_FOLIO(pte_write(pte
) && (!folio_test_anon(folio
) &&
739 PageAnonExclusive(page
)), folio
);
742 * No need to take a page reference as one was already
743 * created when the swap entry was made.
745 if (folio_test_anon(folio
))
746 folio_add_anon_rmap_pte(folio
, page
, vma
, address
, RMAP_NONE
);
749 * Currently device exclusive access only supports anonymous
750 * memory so the entry shouldn't point to a filebacked page.
754 set_pte_at(vma
->vm_mm
, address
, ptep
, pte
);
757 * No need to invalidate - it was non-present before. However
758 * secondary CPUs may have mappings that need invalidating.
760 update_mmu_cache(vma
, address
, ptep
);
764 * Tries to restore an exclusive pte if the page lock can be acquired without
768 try_restore_exclusive_pte(pte_t
*src_pte
, struct vm_area_struct
*vma
,
771 swp_entry_t entry
= pte_to_swp_entry(ptep_get(src_pte
));
772 struct page
*page
= pfn_swap_entry_to_page(entry
);
774 if (trylock_page(page
)) {
775 restore_exclusive_pte(vma
, page
, addr
, src_pte
);
784 * copy one vm_area from one task to the other. Assumes the page tables
785 * already present in the new task to be cleared in the whole range
786 * covered by this vma.
790 copy_nonpresent_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
791 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*dst_vma
,
792 struct vm_area_struct
*src_vma
, unsigned long addr
, int *rss
)
794 unsigned long vm_flags
= dst_vma
->vm_flags
;
795 pte_t orig_pte
= ptep_get(src_pte
);
796 pte_t pte
= orig_pte
;
799 swp_entry_t entry
= pte_to_swp_entry(orig_pte
);
801 if (likely(!non_swap_entry(entry
))) {
802 if (swap_duplicate(entry
) < 0)
805 /* make sure dst_mm is on swapoff's mmlist. */
806 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
807 spin_lock(&mmlist_lock
);
808 if (list_empty(&dst_mm
->mmlist
))
809 list_add(&dst_mm
->mmlist
,
811 spin_unlock(&mmlist_lock
);
813 /* Mark the swap entry as shared. */
814 if (pte_swp_exclusive(orig_pte
)) {
815 pte
= pte_swp_clear_exclusive(orig_pte
);
816 set_pte_at(src_mm
, addr
, src_pte
, pte
);
819 } else if (is_migration_entry(entry
)) {
820 folio
= pfn_swap_entry_folio(entry
);
822 rss
[mm_counter(folio
)]++;
824 if (!is_readable_migration_entry(entry
) &&
825 is_cow_mapping(vm_flags
)) {
827 * COW mappings require pages in both parent and child
828 * to be set to read. A previously exclusive entry is
831 entry
= make_readable_migration_entry(
833 pte
= swp_entry_to_pte(entry
);
834 if (pte_swp_soft_dirty(orig_pte
))
835 pte
= pte_swp_mksoft_dirty(pte
);
836 if (pte_swp_uffd_wp(orig_pte
))
837 pte
= pte_swp_mkuffd_wp(pte
);
838 set_pte_at(src_mm
, addr
, src_pte
, pte
);
840 } else if (is_device_private_entry(entry
)) {
841 page
= pfn_swap_entry_to_page(entry
);
842 folio
= page_folio(page
);
845 * Update rss count even for unaddressable pages, as
846 * they should treated just like normal pages in this
849 * We will likely want to have some new rss counters
850 * for unaddressable pages, at some point. But for now
851 * keep things as they are.
854 rss
[mm_counter(folio
)]++;
855 /* Cannot fail as these pages cannot get pinned. */
856 folio_try_dup_anon_rmap_pte(folio
, page
, src_vma
);
859 * We do not preserve soft-dirty information, because so
860 * far, checkpoint/restore is the only feature that
861 * requires that. And checkpoint/restore does not work
862 * when a device driver is involved (you cannot easily
863 * save and restore device driver state).
865 if (is_writable_device_private_entry(entry
) &&
866 is_cow_mapping(vm_flags
)) {
867 entry
= make_readable_device_private_entry(
869 pte
= swp_entry_to_pte(entry
);
870 if (pte_swp_uffd_wp(orig_pte
))
871 pte
= pte_swp_mkuffd_wp(pte
);
872 set_pte_at(src_mm
, addr
, src_pte
, pte
);
874 } else if (is_device_exclusive_entry(entry
)) {
876 * Make device exclusive entries present by restoring the
877 * original entry then copying as for a present pte. Device
878 * exclusive entries currently only support private writable
879 * (ie. COW) mappings.
881 VM_BUG_ON(!is_cow_mapping(src_vma
->vm_flags
));
882 if (try_restore_exclusive_pte(src_pte
, src_vma
, addr
))
885 } else if (is_pte_marker_entry(entry
)) {
886 pte_marker marker
= copy_pte_marker(entry
, dst_vma
);
889 set_pte_at(dst_mm
, addr
, dst_pte
,
890 make_pte_marker(marker
));
893 if (!userfaultfd_wp(dst_vma
))
894 pte
= pte_swp_clear_uffd_wp(pte
);
895 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
900 * Copy a present and normal page.
902 * NOTE! The usual case is that this isn't required;
903 * instead, the caller can just increase the page refcount
904 * and re-use the pte the traditional way.
906 * And if we need a pre-allocated page but don't yet have
907 * one, return a negative error to let the preallocation
908 * code know so that it can do so outside the page table
912 copy_present_page(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
913 pte_t
*dst_pte
, pte_t
*src_pte
, unsigned long addr
, int *rss
,
914 struct folio
**prealloc
, struct page
*page
)
916 struct folio
*new_folio
;
919 new_folio
= *prealloc
;
924 * We have a prealloc page, all good! Take it
925 * over and copy the page & arm it.
928 if (copy_mc_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
, addr
);
1067 if (mem_cgroup_charge(new_folio
, src_mm
, GFP_KERNEL
)) {
1068 folio_put(new_folio
);
1071 folio_throttle_swaprate(new_folio
, GFP_KERNEL
);
1077 copy_pte_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1078 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
1081 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1082 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1083 pte_t
*orig_src_pte
, *orig_dst_pte
;
1084 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
);
1113 * We already hold the exclusive mmap_lock, the copy_pte_range() and
1114 * retract_page_tables() are using vma->anon_vma to be exclusive, so
1115 * the PTE page is stable, and there is no need to get pmdval and do
1118 src_pte
= pte_offset_map_rw_nolock(src_mm
, src_pmd
, addr
, &dummy_pmdval
,
1121 pte_unmap_unlock(dst_pte
, dst_ptl
);
1125 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1126 orig_src_pte
= src_pte
;
1127 orig_dst_pte
= dst_pte
;
1128 arch_enter_lazy_mmu_mode();
1134 * We are holding two locks at this point - either of them
1135 * could generate latencies in another task on another CPU.
1137 if (progress
>= 32) {
1139 if (need_resched() ||
1140 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
1143 ptent
= ptep_get(src_pte
);
1144 if (pte_none(ptent
)) {
1148 if (unlikely(!pte_present(ptent
))) {
1149 ret
= copy_nonpresent_pte(dst_mm
, src_mm
,
1154 entry
= pte_to_swp_entry(ptep_get(src_pte
));
1156 } else if (ret
== -EBUSY
) {
1162 ptent
= ptep_get(src_pte
);
1163 VM_WARN_ON_ONCE(!pte_present(ptent
));
1166 * Device exclusive entry restored, continue by copying
1167 * the now present pte.
1169 WARN_ON_ONCE(ret
!= -ENOENT
);
1171 /* copy_present_ptes() will clear `*prealloc' if consumed */
1172 max_nr
= (end
- addr
) / PAGE_SIZE
;
1173 ret
= copy_present_ptes(dst_vma
, src_vma
, dst_pte
, src_pte
,
1174 ptent
, addr
, max_nr
, rss
, &prealloc
);
1176 * If we need a pre-allocated page for this pte, drop the
1177 * locks, allocate, and try again.
1178 * If copy failed due to hwpoison in source page, break out.
1180 if (unlikely(ret
== -EAGAIN
|| ret
== -EHWPOISON
))
1182 if (unlikely(prealloc
)) {
1184 * pre-alloc page cannot be reused by next time so as
1185 * to strictly follow mempolicy (e.g., alloc_page_vma()
1186 * will allocate page according to address). This
1187 * could only happen if one pinned pte changed.
1189 folio_put(prealloc
);
1194 } while (dst_pte
+= nr
, src_pte
+= nr
, addr
+= PAGE_SIZE
* nr
,
1197 arch_leave_lazy_mmu_mode();
1198 pte_unmap_unlock(orig_src_pte
, src_ptl
);
1199 add_mm_rss_vec(dst_mm
, rss
);
1200 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
1204 VM_WARN_ON_ONCE(!entry
.val
);
1205 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0) {
1210 } else if (ret
== -EBUSY
|| unlikely(ret
== -EHWPOISON
)) {
1212 } else if (ret
== -EAGAIN
) {
1213 prealloc
= folio_prealloc(src_mm
, src_vma
, addr
, false);
1216 } else if (ret
< 0) {
1220 /* We've captured and resolved the error. Reset, try again. */
1226 if (unlikely(prealloc
))
1227 folio_put(prealloc
);
1232 copy_pmd_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1233 pud_t
*dst_pud
, pud_t
*src_pud
, unsigned long addr
,
1236 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1237 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1238 pmd_t
*src_pmd
, *dst_pmd
;
1241 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
1244 src_pmd
= pmd_offset(src_pud
, addr
);
1246 next
= pmd_addr_end(addr
, end
);
1247 if (is_swap_pmd(*src_pmd
) || pmd_trans_huge(*src_pmd
)
1248 || pmd_devmap(*src_pmd
)) {
1250 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PMD_SIZE
, src_vma
);
1251 err
= copy_huge_pmd(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
1252 addr
, dst_vma
, src_vma
);
1259 if (pmd_none_or_clear_bad(src_pmd
))
1261 if (copy_pte_range(dst_vma
, src_vma
, dst_pmd
, src_pmd
,
1264 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1269 copy_pud_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1270 p4d_t
*dst_p4d
, p4d_t
*src_p4d
, unsigned long addr
,
1273 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1274 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1275 pud_t
*src_pud
, *dst_pud
;
1278 dst_pud
= pud_alloc(dst_mm
, dst_p4d
, addr
);
1281 src_pud
= pud_offset(src_p4d
, addr
);
1283 next
= pud_addr_end(addr
, end
);
1284 if (pud_trans_huge(*src_pud
) || pud_devmap(*src_pud
)) {
1287 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PUD_SIZE
, src_vma
);
1288 err
= copy_huge_pud(dst_mm
, src_mm
,
1289 dst_pud
, src_pud
, addr
, src_vma
);
1296 if (pud_none_or_clear_bad(src_pud
))
1298 if (copy_pmd_range(dst_vma
, src_vma
, dst_pud
, src_pud
,
1301 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1306 copy_p4d_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1307 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, unsigned long addr
,
1310 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1311 p4d_t
*src_p4d
, *dst_p4d
;
1314 dst_p4d
= p4d_alloc(dst_mm
, dst_pgd
, addr
);
1317 src_p4d
= p4d_offset(src_pgd
, addr
);
1319 next
= p4d_addr_end(addr
, end
);
1320 if (p4d_none_or_clear_bad(src_p4d
))
1322 if (copy_pud_range(dst_vma
, src_vma
, dst_p4d
, src_p4d
,
1325 } while (dst_p4d
++, src_p4d
++, addr
= next
, addr
!= end
);
1330 * Return true if the vma needs to copy the pgtable during this fork(). Return
1331 * false when we can speed up fork() by allowing lazy page faults later until
1332 * when the child accesses the memory range.
1335 vma_needs_copy(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
)
1338 * Always copy pgtables when dst_vma has uffd-wp enabled even if it's
1339 * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable
1340 * contains uffd-wp protection information, that's something we can't
1341 * retrieve from page cache, and skip copying will lose those info.
1343 if (userfaultfd_wp(dst_vma
))
1346 if (src_vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
1349 if (src_vma
->anon_vma
)
1353 * Don't copy ptes where a page fault will fill them correctly. Fork
1354 * becomes much lighter when there are big shared or private readonly
1355 * mappings. The tradeoff is that copy_page_range is more efficient
1362 copy_page_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
)
1364 pgd_t
*src_pgd
, *dst_pgd
;
1366 unsigned long addr
= src_vma
->vm_start
;
1367 unsigned long end
= src_vma
->vm_end
;
1368 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1369 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1370 struct mmu_notifier_range range
;
1374 if (!vma_needs_copy(dst_vma
, src_vma
))
1377 if (is_vm_hugetlb_page(src_vma
))
1378 return copy_hugetlb_page_range(dst_mm
, src_mm
, dst_vma
, src_vma
);
1380 if (unlikely(src_vma
->vm_flags
& VM_PFNMAP
)) {
1382 * We do not free on error cases below as remove_vma
1383 * gets called on error from higher level routine
1385 ret
= track_pfn_copy(src_vma
);
1391 * We need to invalidate the secondary MMU mappings only when
1392 * there could be a permission downgrade on the ptes of the
1393 * parent mm. And a permission downgrade will only happen if
1394 * is_cow_mapping() returns true.
1396 is_cow
= is_cow_mapping(src_vma
->vm_flags
);
1399 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_PAGE
,
1400 0, src_mm
, addr
, end
);
1401 mmu_notifier_invalidate_range_start(&range
);
1403 * Disabling preemption is not needed for the write side, as
1404 * the read side doesn't spin, but goes to the mmap_lock.
1406 * Use the raw variant of the seqcount_t write API to avoid
1407 * lockdep complaining about preemptibility.
1409 vma_assert_write_locked(src_vma
);
1410 raw_write_seqcount_begin(&src_mm
->write_protect_seq
);
1414 dst_pgd
= pgd_offset(dst_mm
, addr
);
1415 src_pgd
= pgd_offset(src_mm
, addr
);
1417 next
= pgd_addr_end(addr
, end
);
1418 if (pgd_none_or_clear_bad(src_pgd
))
1420 if (unlikely(copy_p4d_range(dst_vma
, src_vma
, dst_pgd
, src_pgd
,
1422 untrack_pfn_clear(dst_vma
);
1426 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1429 raw_write_seqcount_end(&src_mm
->write_protect_seq
);
1430 mmu_notifier_invalidate_range_end(&range
);
1435 /* Whether we should zap all COWed (private) pages too */
1436 static inline bool should_zap_cows(struct zap_details
*details
)
1438 /* By default, zap all pages */
1442 /* Or, we zap COWed pages only if the caller wants to */
1443 return details
->even_cows
;
1446 /* Decides whether we should zap this folio with the folio pointer specified */
1447 static inline bool should_zap_folio(struct zap_details
*details
,
1448 struct folio
*folio
)
1450 /* If we can make a decision without *folio.. */
1451 if (should_zap_cows(details
))
1454 /* Otherwise we should only zap non-anon folios */
1455 return !folio_test_anon(folio
);
1458 static inline bool zap_drop_markers(struct zap_details
*details
)
1463 return details
->zap_flags
& ZAP_FLAG_DROP_MARKER
;
1467 * This function makes sure that we'll replace the none pte with an uffd-wp
1468 * swap special pte marker when necessary. Must be with the pgtable lock held.
1471 zap_install_uffd_wp_if_needed(struct vm_area_struct
*vma
,
1472 unsigned long addr
, pte_t
*pte
, int nr
,
1473 struct zap_details
*details
, pte_t pteval
)
1475 /* Zap on anonymous always means dropping everything */
1476 if (vma_is_anonymous(vma
))
1479 if (zap_drop_markers(details
))
1483 /* the PFN in the PTE is irrelevant. */
1484 pte_install_uffd_wp_if_needed(vma
, addr
, pte
, pteval
);
1492 static __always_inline
void zap_present_folio_ptes(struct mmu_gather
*tlb
,
1493 struct vm_area_struct
*vma
, struct folio
*folio
,
1494 struct page
*page
, pte_t
*pte
, pte_t ptent
, unsigned int nr
,
1495 unsigned long addr
, struct zap_details
*details
, int *rss
,
1496 bool *force_flush
, bool *force_break
)
1498 struct mm_struct
*mm
= tlb
->mm
;
1499 bool delay_rmap
= false;
1501 if (!folio_test_anon(folio
)) {
1502 ptent
= get_and_clear_full_ptes(mm
, addr
, pte
, nr
, tlb
->fullmm
);
1503 if (pte_dirty(ptent
)) {
1504 folio_mark_dirty(folio
);
1505 if (tlb_delay_rmap(tlb
)) {
1507 *force_flush
= true;
1510 if (pte_young(ptent
) && likely(vma_has_recency(vma
)))
1511 folio_mark_accessed(folio
);
1512 rss
[mm_counter(folio
)] -= nr
;
1514 /* We don't need up-to-date accessed/dirty bits. */
1515 clear_full_ptes(mm
, addr
, pte
, nr
, tlb
->fullmm
);
1516 rss
[MM_ANONPAGES
] -= nr
;
1518 /* Checking a single PTE in a batch is sufficient. */
1519 arch_check_zapped_pte(vma
, ptent
);
1520 tlb_remove_tlb_entries(tlb
, pte
, nr
, addr
);
1521 if (unlikely(userfaultfd_pte_wp(vma
, ptent
)))
1522 zap_install_uffd_wp_if_needed(vma
, addr
, pte
, nr
, details
,
1526 folio_remove_rmap_ptes(folio
, page
, nr
, vma
);
1528 if (unlikely(folio_mapcount(folio
) < 0))
1529 print_bad_pte(vma
, addr
, ptent
, page
);
1531 if (unlikely(__tlb_remove_folio_pages(tlb
, page
, nr
, delay_rmap
))) {
1532 *force_flush
= true;
1533 *force_break
= true;
1538 * Zap or skip at least one present PTE, trying to batch-process subsequent
1539 * PTEs that map consecutive pages of the same folio.
1541 * Returns the number of processed (skipped or zapped) PTEs (at least 1).
1543 static inline int zap_present_ptes(struct mmu_gather
*tlb
,
1544 struct vm_area_struct
*vma
, pte_t
*pte
, pte_t ptent
,
1545 unsigned int max_nr
, unsigned long addr
,
1546 struct zap_details
*details
, int *rss
, bool *force_flush
,
1549 const fpb_t fpb_flags
= FPB_IGNORE_DIRTY
| FPB_IGNORE_SOFT_DIRTY
;
1550 struct mm_struct
*mm
= tlb
->mm
;
1551 struct folio
*folio
;
1555 page
= vm_normal_page(vma
, addr
, ptent
);
1557 /* We don't need up-to-date accessed/dirty bits. */
1558 ptep_get_and_clear_full(mm
, addr
, pte
, tlb
->fullmm
);
1559 arch_check_zapped_pte(vma
, ptent
);
1560 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1561 if (userfaultfd_pte_wp(vma
, ptent
))
1562 zap_install_uffd_wp_if_needed(vma
, addr
, pte
, 1,
1564 ksm_might_unmap_zero_page(mm
, ptent
);
1568 folio
= page_folio(page
);
1569 if (unlikely(!should_zap_folio(details
, folio
)))
1573 * Make sure that the common "small folio" case is as fast as possible
1574 * by keeping the batching logic separate.
1576 if (unlikely(folio_test_large(folio
) && max_nr
!= 1)) {
1577 nr
= folio_pte_batch(folio
, addr
, pte
, ptent
, max_nr
, fpb_flags
,
1580 zap_present_folio_ptes(tlb
, vma
, folio
, page
, pte
, ptent
, nr
,
1581 addr
, details
, rss
, force_flush
,
1585 zap_present_folio_ptes(tlb
, vma
, folio
, page
, pte
, ptent
, 1, addr
,
1586 details
, rss
, force_flush
, force_break
);
1590 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1591 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1592 unsigned long addr
, unsigned long end
,
1593 struct zap_details
*details
)
1595 bool force_flush
= false, force_break
= false;
1596 struct mm_struct
*mm
= tlb
->mm
;
1597 int rss
[NR_MM_COUNTERS
];
1604 tlb_change_page_size(tlb
, PAGE_SIZE
);
1606 start_pte
= pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1610 flush_tlb_batched_pending(mm
);
1611 arch_enter_lazy_mmu_mode();
1613 pte_t ptent
= ptep_get(pte
);
1614 struct folio
*folio
;
1619 if (pte_none(ptent
))
1625 if (pte_present(ptent
)) {
1626 max_nr
= (end
- addr
) / PAGE_SIZE
;
1627 nr
= zap_present_ptes(tlb
, vma
, pte
, ptent
, max_nr
,
1628 addr
, details
, rss
, &force_flush
,
1630 if (unlikely(force_break
)) {
1631 addr
+= nr
* PAGE_SIZE
;
1637 entry
= pte_to_swp_entry(ptent
);
1638 if (is_device_private_entry(entry
) ||
1639 is_device_exclusive_entry(entry
)) {
1640 page
= pfn_swap_entry_to_page(entry
);
1641 folio
= page_folio(page
);
1642 if (unlikely(!should_zap_folio(details
, folio
)))
1645 * Both device private/exclusive mappings should only
1646 * work with anonymous page so far, so we don't need to
1647 * consider uffd-wp bit when zap. For more information,
1648 * see zap_install_uffd_wp_if_needed().
1650 WARN_ON_ONCE(!vma_is_anonymous(vma
));
1651 rss
[mm_counter(folio
)]--;
1652 if (is_device_private_entry(entry
))
1653 folio_remove_rmap_pte(folio
, page
, vma
);
1655 } else if (!non_swap_entry(entry
)) {
1656 max_nr
= (end
- addr
) / PAGE_SIZE
;
1657 nr
= swap_pte_batch(pte
, max_nr
, ptent
);
1658 /* Genuine swap entries, hence a private anon pages */
1659 if (!should_zap_cows(details
))
1661 rss
[MM_SWAPENTS
] -= nr
;
1662 free_swap_and_cache_nr(entry
, nr
);
1663 } else if (is_migration_entry(entry
)) {
1664 folio
= pfn_swap_entry_folio(entry
);
1665 if (!should_zap_folio(details
, folio
))
1667 rss
[mm_counter(folio
)]--;
1668 } else if (pte_marker_entry_uffd_wp(entry
)) {
1670 * For anon: always drop the marker; for file: only
1671 * drop the marker if explicitly requested.
1673 if (!vma_is_anonymous(vma
) &&
1674 !zap_drop_markers(details
))
1676 } else if (is_guard_swp_entry(entry
)) {
1678 * Ordinary zapping should not remove guard PTE
1679 * markers. Only do so if we should remove PTE markers
1682 if (!zap_drop_markers(details
))
1684 } else if (is_hwpoison_entry(entry
) ||
1685 is_poisoned_swp_entry(entry
)) {
1686 if (!should_zap_cows(details
))
1689 /* We should have covered all the swap entry types */
1690 pr_alert("unrecognized swap entry 0x%lx\n", entry
.val
);
1693 clear_not_present_full_ptes(mm
, addr
, pte
, nr
, tlb
->fullmm
);
1694 zap_install_uffd_wp_if_needed(vma
, addr
, pte
, nr
, details
, ptent
);
1695 } while (pte
+= nr
, addr
+= PAGE_SIZE
* nr
, addr
!= end
);
1697 add_mm_rss_vec(mm
, rss
);
1698 arch_leave_lazy_mmu_mode();
1700 /* Do the actual TLB flush before dropping ptl */
1702 tlb_flush_mmu_tlbonly(tlb
);
1703 tlb_flush_rmaps(tlb
, vma
);
1705 pte_unmap_unlock(start_pte
, ptl
);
1708 * If we forced a TLB flush (either due to running out of
1709 * batch buffers or because we needed to flush dirty TLB
1710 * entries before releasing the ptl), free the batched
1711 * memory too. Come back again if we didn't do everything.
1719 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1720 struct vm_area_struct
*vma
, pud_t
*pud
,
1721 unsigned long addr
, unsigned long end
,
1722 struct zap_details
*details
)
1727 pmd
= pmd_offset(pud
, addr
);
1729 next
= pmd_addr_end(addr
, end
);
1730 if (is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1731 if (next
- addr
!= HPAGE_PMD_SIZE
)
1732 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
1733 else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
)) {
1738 } else if (details
&& details
->single_folio
&&
1739 folio_test_pmd_mappable(details
->single_folio
) &&
1740 next
- addr
== HPAGE_PMD_SIZE
&& pmd_none(*pmd
)) {
1741 spinlock_t
*ptl
= pmd_lock(tlb
->mm
, pmd
);
1743 * Take and drop THP pmd lock so that we cannot return
1744 * prematurely, while zap_huge_pmd() has cleared *pmd,
1745 * but not yet decremented compound_mapcount().
1749 if (pmd_none(*pmd
)) {
1753 addr
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1756 } while (pmd
++, cond_resched(), addr
!= end
);
1761 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1762 struct vm_area_struct
*vma
, p4d_t
*p4d
,
1763 unsigned long addr
, unsigned long end
,
1764 struct zap_details
*details
)
1769 pud
= pud_offset(p4d
, addr
);
1771 next
= pud_addr_end(addr
, end
);
1772 if (pud_trans_huge(*pud
) || pud_devmap(*pud
)) {
1773 if (next
- addr
!= HPAGE_PUD_SIZE
) {
1774 mmap_assert_locked(tlb
->mm
);
1775 split_huge_pud(vma
, pud
, addr
);
1776 } else if (zap_huge_pud(tlb
, vma
, pud
, addr
))
1780 if (pud_none_or_clear_bad(pud
))
1782 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1785 } while (pud
++, addr
= next
, addr
!= end
);
1790 static inline unsigned long zap_p4d_range(struct mmu_gather
*tlb
,
1791 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1792 unsigned long addr
, unsigned long end
,
1793 struct zap_details
*details
)
1798 p4d
= p4d_offset(pgd
, addr
);
1800 next
= p4d_addr_end(addr
, end
);
1801 if (p4d_none_or_clear_bad(p4d
))
1803 next
= zap_pud_range(tlb
, vma
, p4d
, addr
, next
, details
);
1804 } while (p4d
++, addr
= next
, addr
!= end
);
1809 void unmap_page_range(struct mmu_gather
*tlb
,
1810 struct vm_area_struct
*vma
,
1811 unsigned long addr
, unsigned long end
,
1812 struct zap_details
*details
)
1817 BUG_ON(addr
>= end
);
1818 tlb_start_vma(tlb
, vma
);
1819 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1821 next
= pgd_addr_end(addr
, end
);
1822 if (pgd_none_or_clear_bad(pgd
))
1824 next
= zap_p4d_range(tlb
, vma
, pgd
, addr
, next
, details
);
1825 } while (pgd
++, addr
= next
, addr
!= end
);
1826 tlb_end_vma(tlb
, vma
);
1830 static void unmap_single_vma(struct mmu_gather
*tlb
,
1831 struct vm_area_struct
*vma
, unsigned long start_addr
,
1832 unsigned long end_addr
,
1833 struct zap_details
*details
, bool mm_wr_locked
)
1835 unsigned long start
= max(vma
->vm_start
, start_addr
);
1838 if (start
>= vma
->vm_end
)
1840 end
= min(vma
->vm_end
, end_addr
);
1841 if (end
<= vma
->vm_start
)
1845 uprobe_munmap(vma
, start
, end
);
1847 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1848 untrack_pfn(vma
, 0, 0, mm_wr_locked
);
1851 if (unlikely(is_vm_hugetlb_page(vma
))) {
1853 * It is undesirable to test vma->vm_file as it
1854 * should be non-null for valid hugetlb area.
1855 * However, vm_file will be NULL in the error
1856 * cleanup path of mmap_region. When
1857 * hugetlbfs ->mmap method fails,
1858 * mmap_region() nullifies vma->vm_file
1859 * before calling this function to clean up.
1860 * Since no pte has actually been setup, it is
1861 * safe to do nothing in this case.
1864 zap_flags_t zap_flags
= details
?
1865 details
->zap_flags
: 0;
1866 __unmap_hugepage_range(tlb
, vma
, start
, end
,
1870 unmap_page_range(tlb
, vma
, start
, end
, details
);
1875 * unmap_vmas - unmap a range of memory covered by a list of vma's
1876 * @tlb: address of the caller's struct mmu_gather
1877 * @mas: the maple state
1878 * @vma: the starting vma
1879 * @start_addr: virtual address at which to start unmapping
1880 * @end_addr: virtual address at which to end unmapping
1881 * @tree_end: The maximum index to check
1882 * @mm_wr_locked: lock flag
1884 * Unmap all pages in the vma list.
1886 * Only addresses between `start' and `end' will be unmapped.
1888 * The VMA list must be sorted in ascending virtual address order.
1890 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1891 * range after unmap_vmas() returns. So the only responsibility here is to
1892 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1893 * drops the lock and schedules.
1895 void unmap_vmas(struct mmu_gather
*tlb
, struct ma_state
*mas
,
1896 struct vm_area_struct
*vma
, unsigned long start_addr
,
1897 unsigned long end_addr
, unsigned long tree_end
,
1900 struct mmu_notifier_range range
;
1901 struct zap_details details
= {
1902 .zap_flags
= ZAP_FLAG_DROP_MARKER
| ZAP_FLAG_UNMAP
,
1903 /* Careful - we need to zap private pages too! */
1907 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
->vm_mm
,
1908 start_addr
, end_addr
);
1909 mmu_notifier_invalidate_range_start(&range
);
1911 unsigned long start
= start_addr
;
1912 unsigned long end
= end_addr
;
1913 hugetlb_zap_begin(vma
, &start
, &end
);
1914 unmap_single_vma(tlb
, vma
, start
, end
, &details
,
1916 hugetlb_zap_end(vma
, &details
);
1917 vma
= mas_find(mas
, tree_end
- 1);
1918 } while (vma
&& likely(!xa_is_zero(vma
)));
1919 mmu_notifier_invalidate_range_end(&range
);
1923 * zap_page_range_single - remove user pages in a given range
1924 * @vma: vm_area_struct holding the applicable pages
1925 * @address: starting address of pages to zap
1926 * @size: number of bytes to zap
1927 * @details: details of shared cache invalidation
1929 * The range must fit into one VMA.
1931 void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1932 unsigned long size
, struct zap_details
*details
)
1934 const unsigned long end
= address
+ size
;
1935 struct mmu_notifier_range range
;
1936 struct mmu_gather tlb
;
1939 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
->vm_mm
,
1941 hugetlb_zap_begin(vma
, &range
.start
, &range
.end
);
1942 tlb_gather_mmu(&tlb
, vma
->vm_mm
);
1943 update_hiwater_rss(vma
->vm_mm
);
1944 mmu_notifier_invalidate_range_start(&range
);
1946 * unmap 'address-end' not 'range.start-range.end' as range
1947 * could have been expanded for hugetlb pmd sharing.
1949 unmap_single_vma(&tlb
, vma
, address
, end
, details
, false);
1950 mmu_notifier_invalidate_range_end(&range
);
1951 tlb_finish_mmu(&tlb
);
1952 hugetlb_zap_end(vma
, details
);
1956 * zap_vma_ptes - remove ptes mapping the vma
1957 * @vma: vm_area_struct holding ptes to be zapped
1958 * @address: starting address of pages to zap
1959 * @size: number of bytes to zap
1961 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1963 * The entire address range must be fully contained within the vma.
1966 void zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1969 if (!range_in_vma(vma
, address
, address
+ size
) ||
1970 !(vma
->vm_flags
& VM_PFNMAP
))
1973 zap_page_range_single(vma
, address
, size
, NULL
);
1975 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1977 static pmd_t
*walk_to_pmd(struct mm_struct
*mm
, unsigned long addr
)
1984 pgd
= pgd_offset(mm
, addr
);
1985 p4d
= p4d_alloc(mm
, pgd
, addr
);
1988 pud
= pud_alloc(mm
, p4d
, addr
);
1991 pmd
= pmd_alloc(mm
, pud
, addr
);
1995 VM_BUG_ON(pmd_trans_huge(*pmd
));
1999 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
2002 pmd_t
*pmd
= walk_to_pmd(mm
, addr
);
2006 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
2009 static bool vm_mixed_zeropage_allowed(struct vm_area_struct
*vma
)
2011 VM_WARN_ON_ONCE(vma
->vm_flags
& VM_PFNMAP
);
2013 * Whoever wants to forbid the zeropage after some zeropages
2014 * might already have been mapped has to scan the page tables and
2015 * bail out on any zeropages. Zeropages in COW mappings can
2016 * be unshared using FAULT_FLAG_UNSHARE faults.
2018 if (mm_forbids_zeropage(vma
->vm_mm
))
2020 /* zeropages in COW mappings are common and unproblematic. */
2021 if (is_cow_mapping(vma
->vm_flags
))
2023 /* Mappings that do not allow for writable PTEs are unproblematic. */
2024 if (!(vma
->vm_flags
& (VM_WRITE
| VM_MAYWRITE
)))
2027 * Why not allow any VMA that has vm_ops->pfn_mkwrite? GUP could
2028 * find the shared zeropage and longterm-pin it, which would
2029 * be problematic as soon as the zeropage gets replaced by a different
2030 * page due to vma->vm_ops->pfn_mkwrite, because what's mapped would
2031 * now differ to what GUP looked up. FSDAX is incompatible to
2032 * FOLL_LONGTERM and VM_IO is incompatible to GUP completely (see
2035 return vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
&&
2036 (vma_is_fsdax(vma
) || vma
->vm_flags
& VM_IO
);
2039 static int validate_page_before_insert(struct vm_area_struct
*vma
,
2042 struct folio
*folio
= page_folio(page
);
2044 if (!folio_ref_count(folio
))
2046 if (unlikely(is_zero_folio(folio
))) {
2047 if (!vm_mixed_zeropage_allowed(vma
))
2051 if (folio_test_anon(folio
) || folio_test_slab(folio
) ||
2052 page_has_type(page
))
2054 flush_dcache_folio(folio
);
2058 static int insert_page_into_pte_locked(struct vm_area_struct
*vma
, pte_t
*pte
,
2059 unsigned long addr
, struct page
*page
, pgprot_t prot
)
2061 struct folio
*folio
= page_folio(page
);
2064 if (!pte_none(ptep_get(pte
)))
2066 /* Ok, finally just insert the thing.. */
2067 pteval
= mk_pte(page
, prot
);
2068 if (unlikely(is_zero_folio(folio
))) {
2069 pteval
= pte_mkspecial(pteval
);
2072 inc_mm_counter(vma
->vm_mm
, mm_counter_file(folio
));
2073 folio_add_file_rmap_pte(folio
, page
, vma
);
2075 set_pte_at(vma
->vm_mm
, addr
, pte
, pteval
);
2079 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2080 struct page
*page
, pgprot_t prot
)
2086 retval
= validate_page_before_insert(vma
, page
);
2090 pte
= get_locked_pte(vma
->vm_mm
, addr
, &ptl
);
2093 retval
= insert_page_into_pte_locked(vma
, pte
, addr
, page
, prot
);
2094 pte_unmap_unlock(pte
, ptl
);
2099 static int insert_page_in_batch_locked(struct vm_area_struct
*vma
, pte_t
*pte
,
2100 unsigned long addr
, struct page
*page
, pgprot_t prot
)
2104 err
= validate_page_before_insert(vma
, page
);
2107 return insert_page_into_pte_locked(vma
, pte
, addr
, page
, prot
);
2110 /* insert_pages() amortizes the cost of spinlock operations
2111 * when inserting pages in a loop.
2113 static int insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
2114 struct page
**pages
, unsigned long *num
, pgprot_t prot
)
2117 pte_t
*start_pte
, *pte
;
2118 spinlock_t
*pte_lock
;
2119 struct mm_struct
*const mm
= vma
->vm_mm
;
2120 unsigned long curr_page_idx
= 0;
2121 unsigned long remaining_pages_total
= *num
;
2122 unsigned long pages_to_write_in_pmd
;
2126 pmd
= walk_to_pmd(mm
, addr
);
2130 pages_to_write_in_pmd
= min_t(unsigned long,
2131 remaining_pages_total
, PTRS_PER_PTE
- pte_index(addr
));
2133 /* Allocate the PTE if necessary; takes PMD lock once only. */
2135 if (pte_alloc(mm
, pmd
))
2138 while (pages_to_write_in_pmd
) {
2140 const int batch_size
= min_t(int, pages_to_write_in_pmd
, 8);
2142 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &pte_lock
);
2147 for (pte
= start_pte
; pte_idx
< batch_size
; ++pte
, ++pte_idx
) {
2148 int err
= insert_page_in_batch_locked(vma
, pte
,
2149 addr
, pages
[curr_page_idx
], prot
);
2150 if (unlikely(err
)) {
2151 pte_unmap_unlock(start_pte
, pte_lock
);
2153 remaining_pages_total
-= pte_idx
;
2159 pte_unmap_unlock(start_pte
, pte_lock
);
2160 pages_to_write_in_pmd
-= batch_size
;
2161 remaining_pages_total
-= batch_size
;
2163 if (remaining_pages_total
)
2167 *num
= remaining_pages_total
;
2172 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
2173 * @vma: user vma to map to
2174 * @addr: target start user address of these pages
2175 * @pages: source kernel pages
2176 * @num: in: number of pages to map. out: number of pages that were *not*
2177 * mapped. (0 means all pages were successfully mapped).
2179 * Preferred over vm_insert_page() when inserting multiple pages.
2181 * In case of error, we may have mapped a subset of the provided
2182 * pages. It is the caller's responsibility to account for this case.
2184 * The same restrictions apply as in vm_insert_page().
2186 int vm_insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
2187 struct page
**pages
, unsigned long *num
)
2189 const unsigned long end_addr
= addr
+ (*num
* PAGE_SIZE
) - 1;
2191 if (addr
< vma
->vm_start
|| end_addr
>= vma
->vm_end
)
2193 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
2194 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
2195 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
2196 vm_flags_set(vma
, VM_MIXEDMAP
);
2198 /* Defer page refcount checking till we're about to map that page. */
2199 return insert_pages(vma
, addr
, pages
, num
, vma
->vm_page_prot
);
2201 EXPORT_SYMBOL(vm_insert_pages
);
2204 * vm_insert_page - insert single page into user vma
2205 * @vma: user vma to map to
2206 * @addr: target user address of this page
2207 * @page: source kernel page
2209 * This allows drivers to insert individual pages they've allocated
2210 * into a user vma. The zeropage is supported in some VMAs,
2211 * see vm_mixed_zeropage_allowed().
2213 * The page has to be a nice clean _individual_ kernel allocation.
2214 * If you allocate a compound page, you need to have marked it as
2215 * such (__GFP_COMP), or manually just split the page up yourself
2216 * (see split_page()).
2218 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2219 * took an arbitrary page protection parameter. This doesn't allow
2220 * that. Your vma protection will have to be set up correctly, which
2221 * means that if you want a shared writable mapping, you'd better
2222 * ask for a shared writable mapping!
2224 * The page does not need to be reserved.
2226 * Usually this function is called from f_op->mmap() handler
2227 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
2228 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2229 * function from other places, for example from page-fault handler.
2231 * Return: %0 on success, negative error code otherwise.
2233 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2236 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2238 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
2239 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
2240 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
2241 vm_flags_set(vma
, VM_MIXEDMAP
);
2243 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2245 EXPORT_SYMBOL(vm_insert_page
);
2248 * __vm_map_pages - maps range of kernel pages into user vma
2249 * @vma: user vma to map to
2250 * @pages: pointer to array of source kernel pages
2251 * @num: number of pages in page array
2252 * @offset: user's requested vm_pgoff
2254 * This allows drivers to map range of kernel pages into a user vma.
2255 * The zeropage is supported in some VMAs, see
2256 * vm_mixed_zeropage_allowed().
2258 * Return: 0 on success and error code otherwise.
2260 static int __vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
2261 unsigned long num
, unsigned long offset
)
2263 unsigned long count
= vma_pages(vma
);
2264 unsigned long uaddr
= vma
->vm_start
;
2267 /* Fail if the user requested offset is beyond the end of the object */
2271 /* Fail if the user requested size exceeds available object size */
2272 if (count
> num
- offset
)
2275 for (i
= 0; i
< count
; i
++) {
2276 ret
= vm_insert_page(vma
, uaddr
, pages
[offset
+ i
]);
2286 * vm_map_pages - maps range of kernel pages starts with non zero offset
2287 * @vma: user vma to map to
2288 * @pages: pointer to array of source kernel pages
2289 * @num: number of pages in page array
2291 * Maps an object consisting of @num pages, catering for the user's
2292 * requested vm_pgoff
2294 * If we fail to insert any page into the vma, the function will return
2295 * immediately leaving any previously inserted pages present. Callers
2296 * from the mmap handler may immediately return the error as their caller
2297 * will destroy the vma, removing any successfully inserted pages. Other
2298 * callers should make their own arrangements for calling unmap_region().
2300 * Context: Process context. Called by mmap handlers.
2301 * Return: 0 on success and error code otherwise.
2303 int vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
2306 return __vm_map_pages(vma
, pages
, num
, vma
->vm_pgoff
);
2308 EXPORT_SYMBOL(vm_map_pages
);
2311 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2312 * @vma: user vma to map to
2313 * @pages: pointer to array of source kernel pages
2314 * @num: number of pages in page array
2316 * Similar to vm_map_pages(), except that it explicitly sets the offset
2317 * to 0. This function is intended for the drivers that did not consider
2320 * Context: Process context. Called by mmap handlers.
2321 * Return: 0 on success and error code otherwise.
2323 int vm_map_pages_zero(struct vm_area_struct
*vma
, struct page
**pages
,
2326 return __vm_map_pages(vma
, pages
, num
, 0);
2328 EXPORT_SYMBOL(vm_map_pages_zero
);
2330 static vm_fault_t
insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2331 pfn_t pfn
, pgprot_t prot
, bool mkwrite
)
2333 struct mm_struct
*mm
= vma
->vm_mm
;
2337 pte
= get_locked_pte(mm
, addr
, &ptl
);
2339 return VM_FAULT_OOM
;
2340 entry
= ptep_get(pte
);
2341 if (!pte_none(entry
)) {
2344 * For read faults on private mappings the PFN passed
2345 * in may not match the PFN we have mapped if the
2346 * mapped PFN is a writeable COW page. In the mkwrite
2347 * case we are creating a writable PTE for a shared
2348 * mapping and we expect the PFNs to match. If they
2349 * don't match, we are likely racing with block
2350 * allocation and mapping invalidation so just skip the
2353 if (pte_pfn(entry
) != pfn_t_to_pfn(pfn
)) {
2354 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry
)));
2357 entry
= pte_mkyoung(entry
);
2358 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2359 if (ptep_set_access_flags(vma
, addr
, pte
, entry
, 1))
2360 update_mmu_cache(vma
, addr
, pte
);
2365 /* Ok, finally just insert the thing.. */
2366 if (pfn_t_devmap(pfn
))
2367 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
2369 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
2372 entry
= pte_mkyoung(entry
);
2373 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2376 set_pte_at(mm
, addr
, pte
, entry
);
2377 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
2380 pte_unmap_unlock(pte
, ptl
);
2381 return VM_FAULT_NOPAGE
;
2385 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2386 * @vma: user vma to map to
2387 * @addr: target user address of this page
2388 * @pfn: source kernel pfn
2389 * @pgprot: pgprot flags for the inserted page
2391 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2392 * to override pgprot on a per-page basis.
2394 * This only makes sense for IO mappings, and it makes no sense for
2395 * COW mappings. In general, using multiple vmas is preferable;
2396 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2399 * pgprot typically only differs from @vma->vm_page_prot when drivers set
2400 * caching- and encryption bits different than those of @vma->vm_page_prot,
2401 * because the caching- or encryption mode may not be known at mmap() time.
2403 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2404 * to set caching and encryption bits for those vmas (except for COW pages).
2405 * This is ensured by core vm only modifying these page table entries using
2406 * functions that don't touch caching- or encryption bits, using pte_modify()
2407 * if needed. (See for example mprotect()).
2409 * Also when new page-table entries are created, this is only done using the
2410 * fault() callback, and never using the value of vma->vm_page_prot,
2411 * except for page-table entries that point to anonymous pages as the result
2414 * Context: Process context. May allocate using %GFP_KERNEL.
2415 * Return: vm_fault_t value.
2417 vm_fault_t
vmf_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
2418 unsigned long pfn
, pgprot_t pgprot
)
2421 * Technically, architectures with pte_special can avoid all these
2422 * restrictions (same for remap_pfn_range). However we would like
2423 * consistency in testing and feature parity among all, so we should
2424 * try to keep these invariants in place for everybody.
2426 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
2427 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
2428 (VM_PFNMAP
|VM_MIXEDMAP
));
2429 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
2430 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
2432 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2433 return VM_FAULT_SIGBUS
;
2435 if (!pfn_modify_allowed(pfn
, pgprot
))
2436 return VM_FAULT_SIGBUS
;
2438 track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
));
2440 return insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
,
2443 EXPORT_SYMBOL(vmf_insert_pfn_prot
);
2446 * vmf_insert_pfn - insert single pfn into user vma
2447 * @vma: user vma to map to
2448 * @addr: target user address of this page
2449 * @pfn: source kernel pfn
2451 * Similar to vm_insert_page, this allows drivers to insert individual pages
2452 * they've allocated into a user vma. Same comments apply.
2454 * This function should only be called from a vm_ops->fault handler, and
2455 * in that case the handler should return the result of this function.
2457 * vma cannot be a COW mapping.
2459 * As this is called only for pages that do not currently exist, we
2460 * do not need to flush old virtual caches or the TLB.
2462 * Context: Process context. May allocate using %GFP_KERNEL.
2463 * Return: vm_fault_t value.
2465 vm_fault_t
vmf_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2468 return vmf_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
2470 EXPORT_SYMBOL(vmf_insert_pfn
);
2472 static bool vm_mixed_ok(struct vm_area_struct
*vma
, pfn_t pfn
, bool mkwrite
)
2474 if (unlikely(is_zero_pfn(pfn_t_to_pfn(pfn
))) &&
2475 (mkwrite
|| !vm_mixed_zeropage_allowed(vma
)))
2477 /* these checks mirror the abort conditions in vm_normal_page */
2478 if (vma
->vm_flags
& VM_MIXEDMAP
)
2480 if (pfn_t_devmap(pfn
))
2482 if (pfn_t_special(pfn
))
2484 if (is_zero_pfn(pfn_t_to_pfn(pfn
)))
2489 static vm_fault_t
__vm_insert_mixed(struct vm_area_struct
*vma
,
2490 unsigned long addr
, pfn_t pfn
, bool mkwrite
)
2492 pgprot_t pgprot
= vma
->vm_page_prot
;
2495 if (!vm_mixed_ok(vma
, pfn
, mkwrite
))
2496 return VM_FAULT_SIGBUS
;
2498 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2499 return VM_FAULT_SIGBUS
;
2501 track_pfn_insert(vma
, &pgprot
, pfn
);
2503 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn
), pgprot
))
2504 return VM_FAULT_SIGBUS
;
2507 * If we don't have pte special, then we have to use the pfn_valid()
2508 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2509 * refcount the page if pfn_valid is true (hence insert_page rather
2510 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2511 * without pte special, it would there be refcounted as a normal page.
2513 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
) &&
2514 !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
2518 * At this point we are committed to insert_page()
2519 * regardless of whether the caller specified flags that
2520 * result in pfn_t_has_page() == false.
2522 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
2523 err
= insert_page(vma
, addr
, page
, pgprot
);
2525 return insert_pfn(vma
, addr
, pfn
, pgprot
, mkwrite
);
2529 return VM_FAULT_OOM
;
2530 if (err
< 0 && err
!= -EBUSY
)
2531 return VM_FAULT_SIGBUS
;
2533 return VM_FAULT_NOPAGE
;
2536 vm_fault_t
vmf_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2539 return __vm_insert_mixed(vma
, addr
, pfn
, false);
2541 EXPORT_SYMBOL(vmf_insert_mixed
);
2544 * If the insertion of PTE failed because someone else already added a
2545 * different entry in the mean time, we treat that as success as we assume
2546 * the same entry was actually inserted.
2548 vm_fault_t
vmf_insert_mixed_mkwrite(struct vm_area_struct
*vma
,
2549 unsigned long addr
, pfn_t pfn
)
2551 return __vm_insert_mixed(vma
, addr
, pfn
, true);
2555 * maps a range of physical memory into the requested pages. the old
2556 * mappings are removed. any references to nonexistent pages results
2557 * in null mappings (currently treated as "copy-on-access")
2559 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2560 unsigned long addr
, unsigned long end
,
2561 unsigned long pfn
, pgprot_t prot
)
2563 pte_t
*pte
, *mapped_pte
;
2567 mapped_pte
= pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2570 arch_enter_lazy_mmu_mode();
2572 BUG_ON(!pte_none(ptep_get(pte
)));
2573 if (!pfn_modify_allowed(pfn
, prot
)) {
2577 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2579 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2580 arch_leave_lazy_mmu_mode();
2581 pte_unmap_unlock(mapped_pte
, ptl
);
2585 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2586 unsigned long addr
, unsigned long end
,
2587 unsigned long pfn
, pgprot_t prot
)
2593 pfn
-= addr
>> PAGE_SHIFT
;
2594 pmd
= pmd_alloc(mm
, pud
, addr
);
2597 VM_BUG_ON(pmd_trans_huge(*pmd
));
2599 next
= pmd_addr_end(addr
, end
);
2600 err
= remap_pte_range(mm
, pmd
, addr
, next
,
2601 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2604 } while (pmd
++, addr
= next
, addr
!= end
);
2608 static inline int remap_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2609 unsigned long addr
, unsigned long end
,
2610 unsigned long pfn
, pgprot_t prot
)
2616 pfn
-= addr
>> PAGE_SHIFT
;
2617 pud
= pud_alloc(mm
, p4d
, addr
);
2621 next
= pud_addr_end(addr
, end
);
2622 err
= remap_pmd_range(mm
, pud
, addr
, next
,
2623 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2626 } while (pud
++, addr
= next
, addr
!= end
);
2630 static inline int remap_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2631 unsigned long addr
, unsigned long end
,
2632 unsigned long pfn
, pgprot_t prot
)
2638 pfn
-= addr
>> PAGE_SHIFT
;
2639 p4d
= p4d_alloc(mm
, pgd
, addr
);
2643 next
= p4d_addr_end(addr
, end
);
2644 err
= remap_pud_range(mm
, p4d
, addr
, next
,
2645 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2648 } while (p4d
++, addr
= next
, addr
!= end
);
2652 static int remap_pfn_range_internal(struct vm_area_struct
*vma
, unsigned long addr
,
2653 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2657 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2658 struct mm_struct
*mm
= vma
->vm_mm
;
2661 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr
)))
2665 * Physically remapped pages are special. Tell the
2666 * rest of the world about it:
2667 * VM_IO tells people not to look at these pages
2668 * (accesses can have side effects).
2669 * VM_PFNMAP tells the core MM that the base pages are just
2670 * raw PFN mappings, and do not have a "struct page" associated
2673 * Disable vma merging and expanding with mremap().
2675 * Omit vma from core dump, even when VM_IO turned off.
2677 * There's a horrible special case to handle copy-on-write
2678 * behaviour that some programs depend on. We mark the "original"
2679 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2680 * See vm_normal_page() for details.
2682 if (is_cow_mapping(vma
->vm_flags
)) {
2683 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2685 vma
->vm_pgoff
= pfn
;
2688 vm_flags_set(vma
, VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
);
2690 BUG_ON(addr
>= end
);
2691 pfn
-= addr
>> PAGE_SHIFT
;
2692 pgd
= pgd_offset(mm
, addr
);
2693 flush_cache_range(vma
, addr
, end
);
2695 next
= pgd_addr_end(addr
, end
);
2696 err
= remap_p4d_range(mm
, pgd
, addr
, next
,
2697 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2700 } while (pgd
++, addr
= next
, addr
!= end
);
2706 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2707 * must have pre-validated the caching bits of the pgprot_t.
2709 int remap_pfn_range_notrack(struct vm_area_struct
*vma
, unsigned long addr
,
2710 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2712 int error
= remap_pfn_range_internal(vma
, addr
, pfn
, size
, prot
);
2718 * A partial pfn range mapping is dangerous: it does not
2719 * maintain page reference counts, and callers may free
2720 * pages due to the error. So zap it early.
2722 zap_page_range_single(vma
, addr
, size
, NULL
);
2727 * remap_pfn_range - remap kernel memory to userspace
2728 * @vma: user vma to map to
2729 * @addr: target page aligned user address to start at
2730 * @pfn: page frame number of kernel physical memory address
2731 * @size: size of mapping area
2732 * @prot: page protection flags for this mapping
2734 * Note: this is only safe if the mm semaphore is held when called.
2736 * Return: %0 on success, negative error code otherwise.
2738 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2739 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2743 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
2747 err
= remap_pfn_range_notrack(vma
, addr
, pfn
, size
, prot
);
2749 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
), true);
2752 EXPORT_SYMBOL(remap_pfn_range
);
2755 * vm_iomap_memory - remap memory to userspace
2756 * @vma: user vma to map to
2757 * @start: start of the physical memory to be mapped
2758 * @len: size of area
2760 * This is a simplified io_remap_pfn_range() for common driver use. The
2761 * driver just needs to give us the physical memory range to be mapped,
2762 * we'll figure out the rest from the vma information.
2764 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2765 * whatever write-combining details or similar.
2767 * Return: %0 on success, negative error code otherwise.
2769 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2771 unsigned long vm_len
, pfn
, pages
;
2773 /* Check that the physical memory area passed in looks valid */
2774 if (start
+ len
< start
)
2777 * You *really* shouldn't map things that aren't page-aligned,
2778 * but we've historically allowed it because IO memory might
2779 * just have smaller alignment.
2781 len
+= start
& ~PAGE_MASK
;
2782 pfn
= start
>> PAGE_SHIFT
;
2783 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2784 if (pfn
+ pages
< pfn
)
2787 /* We start the mapping 'vm_pgoff' pages into the area */
2788 if (vma
->vm_pgoff
> pages
)
2790 pfn
+= vma
->vm_pgoff
;
2791 pages
-= vma
->vm_pgoff
;
2793 /* Can we fit all of the mapping? */
2794 vm_len
= vma
->vm_end
- vma
->vm_start
;
2795 if (vm_len
>> PAGE_SHIFT
> pages
)
2798 /* Ok, let it rip */
2799 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2801 EXPORT_SYMBOL(vm_iomap_memory
);
2803 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2804 unsigned long addr
, unsigned long end
,
2805 pte_fn_t fn
, void *data
, bool create
,
2806 pgtbl_mod_mask
*mask
)
2808 pte_t
*pte
, *mapped_pte
;
2813 mapped_pte
= pte
= (mm
== &init_mm
) ?
2814 pte_alloc_kernel_track(pmd
, addr
, mask
) :
2815 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2819 mapped_pte
= pte
= (mm
== &init_mm
) ?
2820 pte_offset_kernel(pmd
, addr
) :
2821 pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
2826 arch_enter_lazy_mmu_mode();
2830 if (create
|| !pte_none(ptep_get(pte
))) {
2831 err
= fn(pte
++, addr
, data
);
2835 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2837 *mask
|= PGTBL_PTE_MODIFIED
;
2839 arch_leave_lazy_mmu_mode();
2842 pte_unmap_unlock(mapped_pte
, ptl
);
2846 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2847 unsigned long addr
, unsigned long end
,
2848 pte_fn_t fn
, void *data
, bool create
,
2849 pgtbl_mod_mask
*mask
)
2855 BUG_ON(pud_leaf(*pud
));
2858 pmd
= pmd_alloc_track(mm
, pud
, addr
, mask
);
2862 pmd
= pmd_offset(pud
, addr
);
2865 next
= pmd_addr_end(addr
, end
);
2866 if (pmd_none(*pmd
) && !create
)
2868 if (WARN_ON_ONCE(pmd_leaf(*pmd
)))
2870 if (!pmd_none(*pmd
) && WARN_ON_ONCE(pmd_bad(*pmd
))) {
2875 err
= apply_to_pte_range(mm
, pmd
, addr
, next
,
2876 fn
, data
, create
, mask
);
2879 } while (pmd
++, addr
= next
, addr
!= end
);
2884 static int apply_to_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2885 unsigned long addr
, unsigned long end
,
2886 pte_fn_t fn
, void *data
, bool create
,
2887 pgtbl_mod_mask
*mask
)
2894 pud
= pud_alloc_track(mm
, p4d
, addr
, mask
);
2898 pud
= pud_offset(p4d
, addr
);
2901 next
= pud_addr_end(addr
, end
);
2902 if (pud_none(*pud
) && !create
)
2904 if (WARN_ON_ONCE(pud_leaf(*pud
)))
2906 if (!pud_none(*pud
) && WARN_ON_ONCE(pud_bad(*pud
))) {
2911 err
= apply_to_pmd_range(mm
, pud
, addr
, next
,
2912 fn
, data
, create
, mask
);
2915 } while (pud
++, addr
= next
, addr
!= end
);
2920 static int apply_to_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2921 unsigned long addr
, unsigned long end
,
2922 pte_fn_t fn
, void *data
, bool create
,
2923 pgtbl_mod_mask
*mask
)
2930 p4d
= p4d_alloc_track(mm
, pgd
, addr
, mask
);
2934 p4d
= p4d_offset(pgd
, addr
);
2937 next
= p4d_addr_end(addr
, end
);
2938 if (p4d_none(*p4d
) && !create
)
2940 if (WARN_ON_ONCE(p4d_leaf(*p4d
)))
2942 if (!p4d_none(*p4d
) && WARN_ON_ONCE(p4d_bad(*p4d
))) {
2947 err
= apply_to_pud_range(mm
, p4d
, addr
, next
,
2948 fn
, data
, create
, mask
);
2951 } while (p4d
++, addr
= next
, addr
!= end
);
2956 static int __apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2957 unsigned long size
, pte_fn_t fn
,
2958 void *data
, bool create
)
2961 unsigned long start
= addr
, next
;
2962 unsigned long end
= addr
+ size
;
2963 pgtbl_mod_mask mask
= 0;
2966 if (WARN_ON(addr
>= end
))
2969 pgd
= pgd_offset(mm
, addr
);
2971 next
= pgd_addr_end(addr
, end
);
2972 if (pgd_none(*pgd
) && !create
)
2974 if (WARN_ON_ONCE(pgd_leaf(*pgd
)))
2976 if (!pgd_none(*pgd
) && WARN_ON_ONCE(pgd_bad(*pgd
))) {
2981 err
= apply_to_p4d_range(mm
, pgd
, addr
, next
,
2982 fn
, data
, create
, &mask
);
2985 } while (pgd
++, addr
= next
, addr
!= end
);
2987 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
2988 arch_sync_kernel_mappings(start
, start
+ size
);
2994 * Scan a region of virtual memory, filling in page tables as necessary
2995 * and calling a provided function on each leaf page table.
2997 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2998 unsigned long size
, pte_fn_t fn
, void *data
)
3000 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, true);
3002 EXPORT_SYMBOL_GPL(apply_to_page_range
);
3005 * Scan a region of virtual memory, calling a provided function on
3006 * each leaf page table where it exists.
3008 * Unlike apply_to_page_range, this does _not_ fill in page tables
3009 * where they are absent.
3011 int apply_to_existing_page_range(struct mm_struct
*mm
, unsigned long addr
,
3012 unsigned long size
, pte_fn_t fn
, void *data
)
3014 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, false);
3016 EXPORT_SYMBOL_GPL(apply_to_existing_page_range
);
3019 * handle_pte_fault chooses page fault handler according to an entry which was
3020 * read non-atomically. Before making any commitment, on those architectures
3021 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
3022 * parts, do_swap_page must check under lock before unmapping the pte and
3023 * proceeding (but do_wp_page is only called after already making such a check;
3024 * and do_anonymous_page can safely check later on).
3026 static inline int pte_unmap_same(struct vm_fault
*vmf
)
3029 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
3030 if (sizeof(pte_t
) > sizeof(unsigned long)) {
3031 spin_lock(vmf
->ptl
);
3032 same
= pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
);
3033 spin_unlock(vmf
->ptl
);
3036 pte_unmap(vmf
->pte
);
3043 * 0: copied succeeded
3044 * -EHWPOISON: copy failed due to hwpoison in source page
3045 * -EAGAIN: copied failed (some other reason)
3047 static inline int __wp_page_copy_user(struct page
*dst
, struct page
*src
,
3048 struct vm_fault
*vmf
)
3053 struct vm_area_struct
*vma
= vmf
->vma
;
3054 struct mm_struct
*mm
= vma
->vm_mm
;
3055 unsigned long addr
= vmf
->address
;
3058 if (copy_mc_user_highpage(dst
, src
, addr
, vma
))
3064 * If the source page was a PFN mapping, we don't have
3065 * a "struct page" for it. We do a best-effort copy by
3066 * just copying from the original user address. If that
3067 * fails, we just zero-fill it. Live with it.
3069 kaddr
= kmap_local_page(dst
);
3070 pagefault_disable();
3071 uaddr
= (void __user
*)(addr
& PAGE_MASK
);
3074 * On architectures with software "accessed" bits, we would
3075 * take a double page fault, so mark it accessed here.
3078 if (!arch_has_hw_pte_young() && !pte_young(vmf
->orig_pte
)) {
3081 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
3082 if (unlikely(!vmf
->pte
|| !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
3084 * Other thread has already handled the fault
3085 * and update local tlb only
3088 update_mmu_tlb(vma
, addr
, vmf
->pte
);
3093 entry
= pte_mkyoung(vmf
->orig_pte
);
3094 if (ptep_set_access_flags(vma
, addr
, vmf
->pte
, entry
, 0))
3095 update_mmu_cache_range(vmf
, vma
, addr
, vmf
->pte
, 1);
3099 * This really shouldn't fail, because the page is there
3100 * in the page tables. But it might just be unreadable,
3101 * in which case we just give up and fill the result with
3104 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
3108 /* Re-validate under PTL if the page is still mapped */
3109 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
3110 if (unlikely(!vmf
->pte
|| !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
3111 /* The PTE changed under us, update local tlb */
3113 update_mmu_tlb(vma
, addr
, vmf
->pte
);
3119 * The same page can be mapped back since last copy attempt.
3120 * Try to copy again under PTL.
3122 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
3124 * Give a warn in case there can be some obscure
3137 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3139 kunmap_local(kaddr
);
3140 flush_dcache_page(dst
);
3145 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
3147 struct file
*vm_file
= vma
->vm_file
;
3150 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
3153 * Special mappings (e.g. VDSO) do not have any file so fake
3154 * a default GFP_KERNEL for them.
3160 * Notify the address space that the page is about to become writable so that
3161 * it can prohibit this or wait for the page to get into an appropriate state.
3163 * We do this without the lock held, so that it can sleep if it needs to.
3165 static vm_fault_t
do_page_mkwrite(struct vm_fault
*vmf
, struct folio
*folio
)
3168 unsigned int old_flags
= vmf
->flags
;
3170 vmf
->flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
3172 if (vmf
->vma
->vm_file
&&
3173 IS_SWAPFILE(vmf
->vma
->vm_file
->f_mapping
->host
))
3174 return VM_FAULT_SIGBUS
;
3176 ret
= vmf
->vma
->vm_ops
->page_mkwrite(vmf
);
3177 /* Restore original flags so that caller is not surprised */
3178 vmf
->flags
= old_flags
;
3179 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
3181 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
3183 if (!folio
->mapping
) {
3184 folio_unlock(folio
);
3185 return 0; /* retry */
3187 ret
|= VM_FAULT_LOCKED
;
3189 VM_BUG_ON_FOLIO(!folio_test_locked(folio
), folio
);
3194 * Handle dirtying of a page in shared file mapping on a write fault.
3196 * The function expects the page to be locked and unlocks it.
3198 static vm_fault_t
fault_dirty_shared_page(struct vm_fault
*vmf
)
3200 struct vm_area_struct
*vma
= vmf
->vma
;
3201 struct address_space
*mapping
;
3202 struct folio
*folio
= page_folio(vmf
->page
);
3204 bool page_mkwrite
= vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
;
3206 dirtied
= folio_mark_dirty(folio
);
3207 VM_BUG_ON_FOLIO(folio_test_anon(folio
), folio
);
3209 * Take a local copy of the address_space - folio.mapping may be zeroed
3210 * by truncate after folio_unlock(). The address_space itself remains
3211 * pinned by vma->vm_file's reference. We rely on folio_unlock()'s
3212 * release semantics to prevent the compiler from undoing this copying.
3214 mapping
= folio_raw_mapping(folio
);
3215 folio_unlock(folio
);
3218 file_update_time(vma
->vm_file
);
3221 * Throttle page dirtying rate down to writeback speed.
3223 * mapping may be NULL here because some device drivers do not
3224 * set page.mapping but still dirty their pages
3226 * Drop the mmap_lock before waiting on IO, if we can. The file
3227 * is pinning the mapping, as per above.
3229 if ((dirtied
|| page_mkwrite
) && mapping
) {
3232 fpin
= maybe_unlock_mmap_for_io(vmf
, NULL
);
3233 balance_dirty_pages_ratelimited(mapping
);
3236 return VM_FAULT_COMPLETED
;
3244 * Handle write page faults for pages that can be reused in the current vma
3246 * This can happen either due to the mapping being with the VM_SHARED flag,
3247 * or due to us being the last reference standing to the page. In either
3248 * case, all we need to do here is to mark the page as writable and update
3249 * any related book-keeping.
3251 static inline void wp_page_reuse(struct vm_fault
*vmf
, struct folio
*folio
)
3252 __releases(vmf
->ptl
)
3254 struct vm_area_struct
*vma
= vmf
->vma
;
3257 VM_BUG_ON(!(vmf
->flags
& FAULT_FLAG_WRITE
));
3258 VM_WARN_ON(is_zero_pfn(pte_pfn(vmf
->orig_pte
)));
3261 VM_BUG_ON(folio_test_anon(folio
) &&
3262 !PageAnonExclusive(vmf
->page
));
3264 * Clear the folio's cpupid information as the existing
3265 * information potentially belongs to a now completely
3266 * unrelated process.
3268 folio_xchg_last_cpupid(folio
, (1 << LAST_CPUPID_SHIFT
) - 1);
3271 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
3272 entry
= pte_mkyoung(vmf
->orig_pte
);
3273 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3274 if (ptep_set_access_flags(vma
, vmf
->address
, vmf
->pte
, entry
, 1))
3275 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
3276 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3277 count_vm_event(PGREUSE
);
3281 * We could add a bitflag somewhere, but for now, we know that all
3282 * vm_ops that have a ->map_pages have been audited and don't need
3283 * the mmap_lock to be held.
3285 static inline vm_fault_t
vmf_can_call_fault(const struct vm_fault
*vmf
)
3287 struct vm_area_struct
*vma
= vmf
->vma
;
3289 if (vma
->vm_ops
->map_pages
|| !(vmf
->flags
& FAULT_FLAG_VMA_LOCK
))
3292 return VM_FAULT_RETRY
;
3296 * __vmf_anon_prepare - Prepare to handle an anonymous fault.
3297 * @vmf: The vm_fault descriptor passed from the fault handler.
3299 * When preparing to insert an anonymous page into a VMA from a
3300 * fault handler, call this function rather than anon_vma_prepare().
3301 * If this vma does not already have an associated anon_vma and we are
3302 * only protected by the per-VMA lock, the caller must retry with the
3303 * mmap_lock held. __anon_vma_prepare() will look at adjacent VMAs to
3304 * determine if this VMA can share its anon_vma, and that's not safe to
3305 * do with only the per-VMA lock held for this VMA.
3307 * Return: 0 if fault handling can proceed. Any other value should be
3308 * returned to the caller.
3310 vm_fault_t
__vmf_anon_prepare(struct vm_fault
*vmf
)
3312 struct vm_area_struct
*vma
= vmf
->vma
;
3315 if (likely(vma
->anon_vma
))
3317 if (vmf
->flags
& FAULT_FLAG_VMA_LOCK
) {
3318 if (!mmap_read_trylock(vma
->vm_mm
))
3319 return VM_FAULT_RETRY
;
3321 if (__anon_vma_prepare(vma
))
3323 if (vmf
->flags
& FAULT_FLAG_VMA_LOCK
)
3324 mmap_read_unlock(vma
->vm_mm
);
3329 * Handle the case of a page which we actually need to copy to a new page,
3330 * either due to COW or unsharing.
3332 * Called with mmap_lock locked and the old page referenced, but
3333 * without the ptl held.
3335 * High level logic flow:
3337 * - Allocate a page, copy the content of the old page to the new one.
3338 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
3339 * - Take the PTL. If the pte changed, bail out and release the allocated page
3340 * - If the pte is still the way we remember it, update the page table and all
3341 * relevant references. This includes dropping the reference the page-table
3342 * held to the old page, as well as updating the rmap.
3343 * - In any case, unlock the PTL and drop the reference we took to the old page.
3345 static vm_fault_t
wp_page_copy(struct vm_fault
*vmf
)
3347 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
3348 struct vm_area_struct
*vma
= vmf
->vma
;
3349 struct mm_struct
*mm
= vma
->vm_mm
;
3350 struct folio
*old_folio
= NULL
;
3351 struct folio
*new_folio
= NULL
;
3353 int page_copied
= 0;
3354 struct mmu_notifier_range range
;
3358 delayacct_wpcopy_start();
3361 old_folio
= page_folio(vmf
->page
);
3362 ret
= vmf_anon_prepare(vmf
);
3366 pfn_is_zero
= is_zero_pfn(pte_pfn(vmf
->orig_pte
));
3367 new_folio
= folio_prealloc(mm
, vma
, vmf
->address
, pfn_is_zero
);
3374 err
= __wp_page_copy_user(&new_folio
->page
, vmf
->page
, vmf
);
3377 * COW failed, if the fault was solved by other,
3378 * it's fine. If not, userspace would re-fault on
3379 * the same address and we will handle the fault
3380 * from the second attempt.
3381 * The -EHWPOISON case will not be retried.
3383 folio_put(new_folio
);
3385 folio_put(old_folio
);
3387 delayacct_wpcopy_end();
3388 return err
== -EHWPOISON
? VM_FAULT_HWPOISON
: 0;
3390 kmsan_copy_page_meta(&new_folio
->page
, vmf
->page
);
3393 __folio_mark_uptodate(new_folio
);
3395 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, mm
,
3396 vmf
->address
& PAGE_MASK
,
3397 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
);
3398 mmu_notifier_invalidate_range_start(&range
);
3401 * Re-check the pte - we dropped the lock
3403 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, vmf
->address
, &vmf
->ptl
);
3404 if (likely(vmf
->pte
&& pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
3406 if (!folio_test_anon(old_folio
)) {
3407 dec_mm_counter(mm
, mm_counter_file(old_folio
));
3408 inc_mm_counter(mm
, MM_ANONPAGES
);
3411 ksm_might_unmap_zero_page(mm
, vmf
->orig_pte
);
3412 inc_mm_counter(mm
, MM_ANONPAGES
);
3414 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
3415 entry
= mk_pte(&new_folio
->page
, vma
->vm_page_prot
);
3416 entry
= pte_sw_mkyoung(entry
);
3417 if (unlikely(unshare
)) {
3418 if (pte_soft_dirty(vmf
->orig_pte
))
3419 entry
= pte_mksoft_dirty(entry
);
3420 if (pte_uffd_wp(vmf
->orig_pte
))
3421 entry
= pte_mkuffd_wp(entry
);
3423 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3427 * Clear the pte entry and flush it first, before updating the
3428 * pte with the new entry, to keep TLBs on different CPUs in
3429 * sync. This code used to set the new PTE then flush TLBs, but
3430 * that left a window where the new PTE could be loaded into
3431 * some TLBs while the old PTE remains in others.
3433 ptep_clear_flush(vma
, vmf
->address
, vmf
->pte
);
3434 folio_add_new_anon_rmap(new_folio
, vma
, vmf
->address
, RMAP_EXCLUSIVE
);
3435 folio_add_lru_vma(new_folio
, vma
);
3436 BUG_ON(unshare
&& pte_write(entry
));
3437 set_pte_at(mm
, vmf
->address
, vmf
->pte
, entry
);
3438 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
3441 * Only after switching the pte to the new page may
3442 * we remove the mapcount here. Otherwise another
3443 * process may come and find the rmap count decremented
3444 * before the pte is switched to the new page, and
3445 * "reuse" the old page writing into it while our pte
3446 * here still points into it and can be read by other
3449 * The critical issue is to order this
3450 * folio_remove_rmap_pte() with the ptp_clear_flush
3451 * above. Those stores are ordered by (if nothing else,)
3452 * the barrier present in the atomic_add_negative
3453 * in folio_remove_rmap_pte();
3455 * Then the TLB flush in ptep_clear_flush ensures that
3456 * no process can access the old page before the
3457 * decremented mapcount is visible. And the old page
3458 * cannot be reused until after the decremented
3459 * mapcount is visible. So transitively, TLBs to
3460 * old page will be flushed before it can be reused.
3462 folio_remove_rmap_pte(old_folio
, vmf
->page
, vma
);
3465 /* Free the old page.. */
3466 new_folio
= old_folio
;
3468 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3469 } else if (vmf
->pte
) {
3470 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
3471 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3474 mmu_notifier_invalidate_range_end(&range
);
3477 folio_put(new_folio
);
3480 free_swap_cache(old_folio
);
3481 folio_put(old_folio
);
3484 delayacct_wpcopy_end();
3490 folio_put(old_folio
);
3492 delayacct_wpcopy_end();
3497 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3498 * writeable once the page is prepared
3500 * @vmf: structure describing the fault
3501 * @folio: the folio of vmf->page
3503 * This function handles all that is needed to finish a write page fault in a
3504 * shared mapping due to PTE being read-only once the mapped page is prepared.
3505 * It handles locking of PTE and modifying it.
3507 * The function expects the page to be locked or other protection against
3508 * concurrent faults / writeback (such as DAX radix tree locks).
3510 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3511 * we acquired PTE lock.
3513 static vm_fault_t
finish_mkwrite_fault(struct vm_fault
*vmf
, struct folio
*folio
)
3515 WARN_ON_ONCE(!(vmf
->vma
->vm_flags
& VM_SHARED
));
3516 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3519 return VM_FAULT_NOPAGE
;
3521 * We might have raced with another page fault while we released the
3522 * pte_offset_map_lock.
3524 if (!pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)) {
3525 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
3526 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3527 return VM_FAULT_NOPAGE
;
3529 wp_page_reuse(vmf
, folio
);
3534 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3537 static vm_fault_t
wp_pfn_shared(struct vm_fault
*vmf
)
3539 struct vm_area_struct
*vma
= vmf
->vma
;
3541 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
3544 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3545 ret
= vmf_can_call_fault(vmf
);
3549 vmf
->flags
|= FAULT_FLAG_MKWRITE
;
3550 ret
= vma
->vm_ops
->pfn_mkwrite(vmf
);
3551 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))
3553 return finish_mkwrite_fault(vmf
, NULL
);
3555 wp_page_reuse(vmf
, NULL
);
3559 static vm_fault_t
wp_page_shared(struct vm_fault
*vmf
, struct folio
*folio
)
3560 __releases(vmf
->ptl
)
3562 struct vm_area_struct
*vma
= vmf
->vma
;
3567 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
3570 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3571 tmp
= vmf_can_call_fault(vmf
);
3577 tmp
= do_page_mkwrite(vmf
, folio
);
3578 if (unlikely(!tmp
|| (tmp
&
3579 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3583 tmp
= finish_mkwrite_fault(vmf
, folio
);
3584 if (unlikely(tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3585 folio_unlock(folio
);
3590 wp_page_reuse(vmf
, folio
);
3593 ret
|= fault_dirty_shared_page(vmf
);
3599 static bool wp_can_reuse_anon_folio(struct folio
*folio
,
3600 struct vm_area_struct
*vma
)
3603 * We could currently only reuse a subpage of a large folio if no
3604 * other subpages of the large folios are still mapped. However,
3605 * let's just consistently not reuse subpages even if we could
3606 * reuse in that scenario, and give back a large folio a bit
3609 if (folio_test_large(folio
))
3613 * We have to verify under folio lock: these early checks are
3614 * just an optimization to avoid locking the folio and freeing
3615 * the swapcache if there is little hope that we can reuse.
3617 * KSM doesn't necessarily raise the folio refcount.
3619 if (folio_test_ksm(folio
) || folio_ref_count(folio
) > 3)
3621 if (!folio_test_lru(folio
))
3623 * We cannot easily detect+handle references from
3624 * remote LRU caches or references to LRU folios.
3627 if (folio_ref_count(folio
) > 1 + folio_test_swapcache(folio
))
3629 if (!folio_trylock(folio
))
3631 if (folio_test_swapcache(folio
))
3632 folio_free_swap(folio
);
3633 if (folio_test_ksm(folio
) || folio_ref_count(folio
) != 1) {
3634 folio_unlock(folio
);
3638 * Ok, we've got the only folio reference from our mapping
3639 * and the folio is locked, it's dark out, and we're wearing
3640 * sunglasses. Hit it.
3642 folio_move_anon_rmap(folio
, vma
);
3643 folio_unlock(folio
);
3648 * This routine handles present pages, when
3649 * * users try to write to a shared page (FAULT_FLAG_WRITE)
3650 * * GUP wants to take a R/O pin on a possibly shared anonymous page
3651 * (FAULT_FLAG_UNSHARE)
3653 * It is done by copying the page to a new address and decrementing the
3654 * shared-page counter for the old page.
3656 * Note that this routine assumes that the protection checks have been
3657 * done by the caller (the low-level page fault routine in most cases).
3658 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've
3659 * done any necessary COW.
3661 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even
3662 * though the page will change only once the write actually happens. This
3663 * avoids a few races, and potentially makes it more efficient.
3665 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3666 * but allow concurrent faults), with pte both mapped and locked.
3667 * We return with mmap_lock still held, but pte unmapped and unlocked.
3669 static vm_fault_t
do_wp_page(struct vm_fault
*vmf
)
3670 __releases(vmf
->ptl
)
3672 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
3673 struct vm_area_struct
*vma
= vmf
->vma
;
3674 struct folio
*folio
= NULL
;
3677 if (likely(!unshare
)) {
3678 if (userfaultfd_pte_wp(vma
, ptep_get(vmf
->pte
))) {
3679 if (!userfaultfd_wp_async(vma
)) {
3680 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3681 return handle_userfault(vmf
, VM_UFFD_WP
);
3685 * Nothing needed (cache flush, TLB invalidations,
3686 * etc.) because we're only removing the uffd-wp bit,
3687 * which is completely invisible to the user.
3689 pte
= pte_clear_uffd_wp(ptep_get(vmf
->pte
));
3691 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3693 * Update this to be prepared for following up CoW
3696 vmf
->orig_pte
= pte
;
3700 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3701 * is flushed in this case before copying.
3703 if (unlikely(userfaultfd_wp(vmf
->vma
) &&
3704 mm_tlb_flush_pending(vmf
->vma
->vm_mm
)))
3705 flush_tlb_page(vmf
->vma
, vmf
->address
);
3708 vmf
->page
= vm_normal_page(vma
, vmf
->address
, vmf
->orig_pte
);
3711 folio
= page_folio(vmf
->page
);
3714 * Shared mapping: we are guaranteed to have VM_WRITE and
3715 * FAULT_FLAG_WRITE set at this point.
3717 if (vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
)) {
3719 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3722 * We should not cow pages in a shared writeable mapping.
3723 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3726 return wp_pfn_shared(vmf
);
3727 return wp_page_shared(vmf
, folio
);
3731 * Private mapping: create an exclusive anonymous page copy if reuse
3732 * is impossible. We might miss VM_WRITE for FOLL_FORCE handling.
3734 * If we encounter a page that is marked exclusive, we must reuse
3735 * the page without further checks.
3737 if (folio
&& folio_test_anon(folio
) &&
3738 (PageAnonExclusive(vmf
->page
) || wp_can_reuse_anon_folio(folio
, vma
))) {
3739 if (!PageAnonExclusive(vmf
->page
))
3740 SetPageAnonExclusive(vmf
->page
);
3741 if (unlikely(unshare
)) {
3742 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3745 wp_page_reuse(vmf
, folio
);
3749 * Ok, we need to copy. Oh, well..
3754 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3756 if (folio
&& folio_test_ksm(folio
))
3757 count_vm_event(COW_KSM
);
3759 return wp_page_copy(vmf
);
3762 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
3763 unsigned long start_addr
, unsigned long end_addr
,
3764 struct zap_details
*details
)
3766 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
3769 static inline void unmap_mapping_range_tree(struct rb_root_cached
*root
,
3770 pgoff_t first_index
,
3772 struct zap_details
*details
)
3774 struct vm_area_struct
*vma
;
3775 pgoff_t vba
, vea
, zba
, zea
;
3777 vma_interval_tree_foreach(vma
, root
, first_index
, last_index
) {
3778 vba
= vma
->vm_pgoff
;
3779 vea
= vba
+ vma_pages(vma
) - 1;
3780 zba
= max(first_index
, vba
);
3781 zea
= min(last_index
, vea
);
3783 unmap_mapping_range_vma(vma
,
3784 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
3785 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
3791 * unmap_mapping_folio() - Unmap single folio from processes.
3792 * @folio: The locked folio to be unmapped.
3794 * Unmap this folio from any userspace process which still has it mmaped.
3795 * Typically, for efficiency, the range of nearby pages has already been
3796 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3797 * truncation or invalidation holds the lock on a folio, it may find that
3798 * the page has been remapped again: and then uses unmap_mapping_folio()
3799 * to unmap it finally.
3801 void unmap_mapping_folio(struct folio
*folio
)
3803 struct address_space
*mapping
= folio
->mapping
;
3804 struct zap_details details
= { };
3805 pgoff_t first_index
;
3808 VM_BUG_ON(!folio_test_locked(folio
));
3810 first_index
= folio
->index
;
3811 last_index
= folio_next_index(folio
) - 1;
3813 details
.even_cows
= false;
3814 details
.single_folio
= folio
;
3815 details
.zap_flags
= ZAP_FLAG_DROP_MARKER
;
3817 i_mmap_lock_read(mapping
);
3818 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3819 unmap_mapping_range_tree(&mapping
->i_mmap
, first_index
,
3820 last_index
, &details
);
3821 i_mmap_unlock_read(mapping
);
3825 * unmap_mapping_pages() - Unmap pages from processes.
3826 * @mapping: The address space containing pages to be unmapped.
3827 * @start: Index of first page to be unmapped.
3828 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3829 * @even_cows: Whether to unmap even private COWed pages.
3831 * Unmap the pages in this address space from any userspace process which
3832 * has them mmaped. Generally, you want to remove COWed pages as well when
3833 * a file is being truncated, but not when invalidating pages from the page
3836 void unmap_mapping_pages(struct address_space
*mapping
, pgoff_t start
,
3837 pgoff_t nr
, bool even_cows
)
3839 struct zap_details details
= { };
3840 pgoff_t first_index
= start
;
3841 pgoff_t last_index
= start
+ nr
- 1;
3843 details
.even_cows
= even_cows
;
3844 if (last_index
< first_index
)
3845 last_index
= ULONG_MAX
;
3847 i_mmap_lock_read(mapping
);
3848 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3849 unmap_mapping_range_tree(&mapping
->i_mmap
, first_index
,
3850 last_index
, &details
);
3851 i_mmap_unlock_read(mapping
);
3853 EXPORT_SYMBOL_GPL(unmap_mapping_pages
);
3856 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3857 * address_space corresponding to the specified byte range in the underlying
3860 * @mapping: the address space containing mmaps to be unmapped.
3861 * @holebegin: byte in first page to unmap, relative to the start of
3862 * the underlying file. This will be rounded down to a PAGE_SIZE
3863 * boundary. Note that this is different from truncate_pagecache(), which
3864 * must keep the partial page. In contrast, we must get rid of
3866 * @holelen: size of prospective hole in bytes. This will be rounded
3867 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3869 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3870 * but 0 when invalidating pagecache, don't throw away private data.
3872 void unmap_mapping_range(struct address_space
*mapping
,
3873 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
3875 pgoff_t hba
= (pgoff_t
)(holebegin
) >> PAGE_SHIFT
;
3876 pgoff_t hlen
= ((pgoff_t
)(holelen
) + PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3878 /* Check for overflow. */
3879 if (sizeof(holelen
) > sizeof(hlen
)) {
3881 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3882 if (holeend
& ~(long long)ULONG_MAX
)
3883 hlen
= ULONG_MAX
- hba
+ 1;
3886 unmap_mapping_pages(mapping
, hba
, hlen
, even_cows
);
3888 EXPORT_SYMBOL(unmap_mapping_range
);
3891 * Restore a potential device exclusive pte to a working pte entry
3893 static vm_fault_t
remove_device_exclusive_entry(struct vm_fault
*vmf
)
3895 struct folio
*folio
= page_folio(vmf
->page
);
3896 struct vm_area_struct
*vma
= vmf
->vma
;
3897 struct mmu_notifier_range range
;
3901 * We need a reference to lock the folio because we don't hold
3902 * the PTL so a racing thread can remove the device-exclusive
3903 * entry and unmap it. If the folio is free the entry must
3904 * have been removed already. If it happens to have already
3905 * been re-allocated after being freed all we do is lock and
3908 if (!folio_try_get(folio
))
3911 ret
= folio_lock_or_retry(folio
, vmf
);
3916 mmu_notifier_range_init_owner(&range
, MMU_NOTIFY_EXCLUSIVE
, 0,
3917 vma
->vm_mm
, vmf
->address
& PAGE_MASK
,
3918 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
, NULL
);
3919 mmu_notifier_invalidate_range_start(&range
);
3921 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3923 if (likely(vmf
->pte
&& pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)))
3924 restore_exclusive_pte(vma
, vmf
->page
, vmf
->address
, vmf
->pte
);
3927 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3928 folio_unlock(folio
);
3931 mmu_notifier_invalidate_range_end(&range
);
3935 static inline bool should_try_to_free_swap(struct folio
*folio
,
3936 struct vm_area_struct
*vma
,
3937 unsigned int fault_flags
)
3939 if (!folio_test_swapcache(folio
))
3941 if (mem_cgroup_swap_full(folio
) || (vma
->vm_flags
& VM_LOCKED
) ||
3942 folio_test_mlocked(folio
))
3945 * If we want to map a page that's in the swapcache writable, we
3946 * have to detect via the refcount if we're really the exclusive
3947 * user. Try freeing the swapcache to get rid of the swapcache
3948 * reference only in case it's likely that we'll be the exlusive user.
3950 return (fault_flags
& FAULT_FLAG_WRITE
) && !folio_test_ksm(folio
) &&
3951 folio_ref_count(folio
) == (1 + folio_nr_pages(folio
));
3954 static vm_fault_t
pte_marker_clear(struct vm_fault
*vmf
)
3956 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
,
3957 vmf
->address
, &vmf
->ptl
);
3961 * Be careful so that we will only recover a special uffd-wp pte into a
3962 * none pte. Otherwise it means the pte could have changed, so retry.
3964 * This should also cover the case where e.g. the pte changed
3965 * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_POISONED.
3966 * So is_pte_marker() check is not enough to safely drop the pte.
3968 if (pte_same(vmf
->orig_pte
, ptep_get(vmf
->pte
)))
3969 pte_clear(vmf
->vma
->vm_mm
, vmf
->address
, vmf
->pte
);
3970 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3974 static vm_fault_t
do_pte_missing(struct vm_fault
*vmf
)
3976 if (vma_is_anonymous(vmf
->vma
))
3977 return do_anonymous_page(vmf
);
3979 return do_fault(vmf
);
3983 * This is actually a page-missing access, but with uffd-wp special pte
3984 * installed. It means this pte was wr-protected before being unmapped.
3986 static vm_fault_t
pte_marker_handle_uffd_wp(struct vm_fault
*vmf
)
3989 * Just in case there're leftover special ptes even after the region
3990 * got unregistered - we can simply clear them.
3992 if (unlikely(!userfaultfd_wp(vmf
->vma
)))
3993 return pte_marker_clear(vmf
);
3995 return do_pte_missing(vmf
);
3998 static vm_fault_t
handle_pte_marker(struct vm_fault
*vmf
)
4000 swp_entry_t entry
= pte_to_swp_entry(vmf
->orig_pte
);
4001 unsigned long marker
= pte_marker_get(entry
);
4004 * PTE markers should never be empty. If anything weird happened,
4005 * the best thing to do is to kill the process along with its mm.
4007 if (WARN_ON_ONCE(!marker
))
4008 return VM_FAULT_SIGBUS
;
4010 /* Higher priority than uffd-wp when data corrupted */
4011 if (marker
& PTE_MARKER_POISONED
)
4012 return VM_FAULT_HWPOISON
;
4014 /* Hitting a guard page is always a fatal condition. */
4015 if (marker
& PTE_MARKER_GUARD
)
4016 return VM_FAULT_SIGSEGV
;
4018 if (pte_marker_entry_uffd_wp(entry
))
4019 return pte_marker_handle_uffd_wp(vmf
);
4021 /* This is an unknown pte marker */
4022 return VM_FAULT_SIGBUS
;
4025 static struct folio
*__alloc_swap_folio(struct vm_fault
*vmf
)
4027 struct vm_area_struct
*vma
= vmf
->vma
;
4028 struct folio
*folio
;
4031 folio
= vma_alloc_folio(GFP_HIGHUSER_MOVABLE
, 0, vma
, vmf
->address
);
4035 entry
= pte_to_swp_entry(vmf
->orig_pte
);
4036 if (mem_cgroup_swapin_charge_folio(folio
, vma
->vm_mm
,
4037 GFP_KERNEL
, entry
)) {
4045 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4046 static inline int non_swapcache_batch(swp_entry_t entry
, int max_nr
)
4048 struct swap_info_struct
*si
= swp_swap_info(entry
);
4049 pgoff_t offset
= swp_offset(entry
);
4053 * While allocating a large folio and doing swap_read_folio, which is
4054 * the case the being faulted pte doesn't have swapcache. We need to
4055 * ensure all PTEs have no cache as well, otherwise, we might go to
4056 * swap devices while the content is in swapcache.
4058 for (i
= 0; i
< max_nr
; i
++) {
4059 if ((si
->swap_map
[offset
+ i
] & SWAP_HAS_CACHE
))
4067 * Check if the PTEs within a range are contiguous swap entries
4068 * and have consistent swapcache, zeromap.
4070 static bool can_swapin_thp(struct vm_fault
*vmf
, pte_t
*ptep
, int nr_pages
)
4077 addr
= ALIGN_DOWN(vmf
->address
, nr_pages
* PAGE_SIZE
);
4078 idx
= (vmf
->address
- addr
) / PAGE_SIZE
;
4079 pte
= ptep_get(ptep
);
4081 if (!pte_same(pte
, pte_move_swp_offset(vmf
->orig_pte
, -idx
)))
4083 entry
= pte_to_swp_entry(pte
);
4084 if (swap_pte_batch(ptep
, nr_pages
, pte
) != nr_pages
)
4088 * swap_read_folio() can't handle the case a large folio is hybridly
4089 * from different backends. And they are likely corner cases. Similar
4090 * things might be added once zswap support large folios.
4092 if (unlikely(swap_zeromap_batch(entry
, nr_pages
, NULL
) != nr_pages
))
4094 if (unlikely(non_swapcache_batch(entry
, nr_pages
) != nr_pages
))
4100 static inline unsigned long thp_swap_suitable_orders(pgoff_t swp_offset
,
4102 unsigned long orders
)
4106 order
= highest_order(orders
);
4109 * To swap in a THP with nr pages, we require that its first swap_offset
4110 * is aligned with that number, as it was when the THP was swapped out.
4111 * This helps filter out most invalid entries.
4115 if ((addr
>> PAGE_SHIFT
) % nr
== swp_offset
% nr
)
4117 order
= next_order(&orders
, order
);
4123 static struct folio
*alloc_swap_folio(struct vm_fault
*vmf
)
4125 struct vm_area_struct
*vma
= vmf
->vma
;
4126 unsigned long orders
;
4127 struct folio
*folio
;
4136 * If uffd is active for the vma we need per-page fault fidelity to
4137 * maintain the uffd semantics.
4139 if (unlikely(userfaultfd_armed(vma
)))
4143 * A large swapped out folio could be partially or fully in zswap. We
4144 * lack handling for such cases, so fallback to swapping in order-0
4147 if (!zswap_never_enabled())
4150 entry
= pte_to_swp_entry(vmf
->orig_pte
);
4152 * Get a list of all the (large) orders below PMD_ORDER that are enabled
4153 * and suitable for swapping THP.
4155 orders
= thp_vma_allowable_orders(vma
, vma
->vm_flags
,
4156 TVA_IN_PF
| TVA_ENFORCE_SYSFS
, BIT(PMD_ORDER
) - 1);
4157 orders
= thp_vma_suitable_orders(vma
, vmf
->address
, orders
);
4158 orders
= thp_swap_suitable_orders(swp_offset(entry
),
4159 vmf
->address
, orders
);
4164 pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
,
4165 vmf
->address
& PMD_MASK
, &ptl
);
4170 * For do_swap_page, find the highest order where the aligned range is
4171 * completely swap entries with contiguous swap offsets.
4173 order
= highest_order(orders
);
4175 addr
= ALIGN_DOWN(vmf
->address
, PAGE_SIZE
<< order
);
4176 if (can_swapin_thp(vmf
, pte
+ pte_index(addr
), 1 << order
))
4178 order
= next_order(&orders
, order
);
4181 pte_unmap_unlock(pte
, ptl
);
4183 /* Try allocating the highest of the remaining orders. */
4184 gfp
= vma_thp_gfp_mask(vma
);
4186 addr
= ALIGN_DOWN(vmf
->address
, PAGE_SIZE
<< order
);
4187 folio
= vma_alloc_folio(gfp
, order
, vma
, addr
);
4189 if (!mem_cgroup_swapin_charge_folio(folio
, vma
->vm_mm
,
4194 order
= next_order(&orders
, order
);
4198 return __alloc_swap_folio(vmf
);
4200 #else /* !CONFIG_TRANSPARENT_HUGEPAGE */
4201 static struct folio
*alloc_swap_folio(struct vm_fault
*vmf
)
4203 return __alloc_swap_folio(vmf
);
4205 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4207 static DECLARE_WAIT_QUEUE_HEAD(swapcache_wq
);
4210 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4211 * but allow concurrent faults), and pte mapped but not yet locked.
4212 * We return with pte unmapped and unlocked.
4214 * We return with the mmap_lock locked or unlocked in the same cases
4215 * as does filemap_fault().
4217 vm_fault_t
do_swap_page(struct vm_fault
*vmf
)
4219 struct vm_area_struct
*vma
= vmf
->vma
;
4220 struct folio
*swapcache
, *folio
= NULL
;
4221 DECLARE_WAITQUEUE(wait
, current
);
4223 struct swap_info_struct
*si
= NULL
;
4224 rmap_t rmap_flags
= RMAP_NONE
;
4225 bool need_clear_cache
= false;
4226 bool exclusive
= false;
4230 void *shadow
= NULL
;
4232 unsigned long page_idx
;
4233 unsigned long address
;
4236 if (!pte_unmap_same(vmf
))
4239 entry
= pte_to_swp_entry(vmf
->orig_pte
);
4240 if (unlikely(non_swap_entry(entry
))) {
4241 if (is_migration_entry(entry
)) {
4242 migration_entry_wait(vma
->vm_mm
, vmf
->pmd
,
4244 } else if (is_device_exclusive_entry(entry
)) {
4245 vmf
->page
= pfn_swap_entry_to_page(entry
);
4246 ret
= remove_device_exclusive_entry(vmf
);
4247 } else if (is_device_private_entry(entry
)) {
4248 if (vmf
->flags
& FAULT_FLAG_VMA_LOCK
) {
4250 * migrate_to_ram is not yet ready to operate
4254 ret
= VM_FAULT_RETRY
;
4258 vmf
->page
= pfn_swap_entry_to_page(entry
);
4259 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4260 vmf
->address
, &vmf
->ptl
);
4261 if (unlikely(!vmf
->pte
||
4262 !pte_same(ptep_get(vmf
->pte
),
4267 * Get a page reference while we know the page can't be
4270 get_page(vmf
->page
);
4271 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4272 ret
= vmf
->page
->pgmap
->ops
->migrate_to_ram(vmf
);
4273 put_page(vmf
->page
);
4274 } else if (is_hwpoison_entry(entry
)) {
4275 ret
= VM_FAULT_HWPOISON
;
4276 } else if (is_pte_marker_entry(entry
)) {
4277 ret
= handle_pte_marker(vmf
);
4279 print_bad_pte(vma
, vmf
->address
, vmf
->orig_pte
, NULL
);
4280 ret
= VM_FAULT_SIGBUS
;
4285 /* Prevent swapoff from happening to us. */
4286 si
= get_swap_device(entry
);
4290 folio
= swap_cache_get_folio(entry
, vma
, vmf
->address
);
4292 page
= folio_file_page(folio
, swp_offset(entry
));
4296 if (data_race(si
->flags
& SWP_SYNCHRONOUS_IO
) &&
4297 __swap_count(entry
) == 1) {
4298 /* skip swapcache */
4299 folio
= alloc_swap_folio(vmf
);
4301 __folio_set_locked(folio
);
4302 __folio_set_swapbacked(folio
);
4304 nr_pages
= folio_nr_pages(folio
);
4305 if (folio_test_large(folio
))
4306 entry
.val
= ALIGN_DOWN(entry
.val
, nr_pages
);
4308 * Prevent parallel swapin from proceeding with
4309 * the cache flag. Otherwise, another thread
4310 * may finish swapin first, free the entry, and
4311 * swapout reusing the same entry. It's
4312 * undetectable as pte_same() returns true due
4315 if (swapcache_prepare(entry
, nr_pages
)) {
4317 * Relax a bit to prevent rapid
4318 * repeated page faults.
4320 add_wait_queue(&swapcache_wq
, &wait
);
4321 schedule_timeout_uninterruptible(1);
4322 remove_wait_queue(&swapcache_wq
, &wait
);
4325 need_clear_cache
= true;
4327 mem_cgroup_swapin_uncharge_swap(entry
, nr_pages
);
4329 shadow
= get_shadow_from_swap_cache(entry
);
4331 workingset_refault(folio
, shadow
);
4333 folio_add_lru(folio
);
4335 /* To provide entry to swap_read_folio() */
4336 folio
->swap
= entry
;
4337 swap_read_folio(folio
, NULL
);
4338 folio
->private = NULL
;
4341 folio
= swapin_readahead(entry
, GFP_HIGHUSER_MOVABLE
,
4348 * Back out if somebody else faulted in this pte
4349 * while we released the pte lock.
4351 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4352 vmf
->address
, &vmf
->ptl
);
4353 if (likely(vmf
->pte
&&
4354 pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)))
4359 /* Had to read the page from swap area: Major fault */
4360 ret
= VM_FAULT_MAJOR
;
4361 count_vm_event(PGMAJFAULT
);
4362 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
4363 page
= folio_file_page(folio
, swp_offset(entry
));
4364 } else if (PageHWPoison(page
)) {
4366 * hwpoisoned dirty swapcache pages are kept for killing
4367 * owner processes (which may be unknown at hwpoison time)
4369 ret
= VM_FAULT_HWPOISON
;
4373 ret
|= folio_lock_or_retry(folio
, vmf
);
4374 if (ret
& VM_FAULT_RETRY
)
4379 * Make sure folio_free_swap() or swapoff did not release the
4380 * swapcache from under us. The page pin, and pte_same test
4381 * below, are not enough to exclude that. Even if it is still
4382 * swapcache, we need to check that the page's swap has not
4385 if (unlikely(!folio_test_swapcache(folio
) ||
4386 page_swap_entry(page
).val
!= entry
.val
))
4390 * KSM sometimes has to copy on read faults, for example, if
4391 * page->index of !PageKSM() pages would be nonlinear inside the
4392 * anon VMA -- PageKSM() is lost on actual swapout.
4394 folio
= ksm_might_need_to_copy(folio
, vma
, vmf
->address
);
4395 if (unlikely(!folio
)) {
4399 } else if (unlikely(folio
== ERR_PTR(-EHWPOISON
))) {
4400 ret
= VM_FAULT_HWPOISON
;
4404 if (folio
!= swapcache
)
4405 page
= folio_page(folio
, 0);
4408 * If we want to map a page that's in the swapcache writable, we
4409 * have to detect via the refcount if we're really the exclusive
4410 * owner. Try removing the extra reference from the local LRU
4411 * caches if required.
4413 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && folio
== swapcache
&&
4414 !folio_test_ksm(folio
) && !folio_test_lru(folio
))
4418 folio_throttle_swaprate(folio
, GFP_KERNEL
);
4421 * Back out if somebody else already faulted in this pte.
4423 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
4425 if (unlikely(!vmf
->pte
|| !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)))
4428 if (unlikely(!folio_test_uptodate(folio
))) {
4429 ret
= VM_FAULT_SIGBUS
;
4433 /* allocated large folios for SWP_SYNCHRONOUS_IO */
4434 if (folio_test_large(folio
) && !folio_test_swapcache(folio
)) {
4435 unsigned long nr
= folio_nr_pages(folio
);
4436 unsigned long folio_start
= ALIGN_DOWN(vmf
->address
, nr
* PAGE_SIZE
);
4437 unsigned long idx
= (vmf
->address
- folio_start
) / PAGE_SIZE
;
4438 pte_t
*folio_ptep
= vmf
->pte
- idx
;
4439 pte_t folio_pte
= ptep_get(folio_ptep
);
4441 if (!pte_same(folio_pte
, pte_move_swp_offset(vmf
->orig_pte
, -idx
)) ||
4442 swap_pte_batch(folio_ptep
, nr
, folio_pte
) != nr
)
4446 address
= folio_start
;
4453 address
= vmf
->address
;
4455 if (folio_test_large(folio
) && folio_test_swapcache(folio
)) {
4456 int nr
= folio_nr_pages(folio
);
4457 unsigned long idx
= folio_page_idx(folio
, page
);
4458 unsigned long folio_start
= address
- idx
* PAGE_SIZE
;
4459 unsigned long folio_end
= folio_start
+ nr
* PAGE_SIZE
;
4463 if (unlikely(folio_start
< max(address
& PMD_MASK
, vma
->vm_start
)))
4465 if (unlikely(folio_end
> pmd_addr_end(address
, vma
->vm_end
)))
4468 folio_ptep
= vmf
->pte
- idx
;
4469 folio_pte
= ptep_get(folio_ptep
);
4470 if (!pte_same(folio_pte
, pte_move_swp_offset(vmf
->orig_pte
, -idx
)) ||
4471 swap_pte_batch(folio_ptep
, nr
, folio_pte
) != nr
)
4475 address
= folio_start
;
4478 entry
= folio
->swap
;
4479 page
= &folio
->page
;
4484 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte
4485 * must never point at an anonymous page in the swapcache that is
4486 * PG_anon_exclusive. Sanity check that this holds and especially, that
4487 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity
4488 * check after taking the PT lock and making sure that nobody
4489 * concurrently faulted in this page and set PG_anon_exclusive.
4491 BUG_ON(!folio_test_anon(folio
) && folio_test_mappedtodisk(folio
));
4492 BUG_ON(folio_test_anon(folio
) && PageAnonExclusive(page
));
4495 * Check under PT lock (to protect against concurrent fork() sharing
4496 * the swap entry concurrently) for certainly exclusive pages.
4498 if (!folio_test_ksm(folio
)) {
4499 exclusive
= pte_swp_exclusive(vmf
->orig_pte
);
4500 if (folio
!= swapcache
) {
4502 * We have a fresh page that is not exposed to the
4503 * swapcache -> certainly exclusive.
4506 } else if (exclusive
&& folio_test_writeback(folio
) &&
4507 data_race(si
->flags
& SWP_STABLE_WRITES
)) {
4509 * This is tricky: not all swap backends support
4510 * concurrent page modifications while under writeback.
4512 * So if we stumble over such a page in the swapcache
4513 * we must not set the page exclusive, otherwise we can
4514 * map it writable without further checks and modify it
4515 * while still under writeback.
4517 * For these problematic swap backends, simply drop the
4518 * exclusive marker: this is perfectly fine as we start
4519 * writeback only if we fully unmapped the page and
4520 * there are no unexpected references on the page after
4521 * unmapping succeeded. After fully unmapped, no
4522 * further GUP references (FOLL_GET and FOLL_PIN) can
4523 * appear, so dropping the exclusive marker and mapping
4524 * it only R/O is fine.
4531 * Some architectures may have to restore extra metadata to the page
4532 * when reading from swap. This metadata may be indexed by swap entry
4533 * so this must be called before swap_free().
4535 arch_swap_restore(folio_swap(entry
, folio
), folio
);
4538 * Remove the swap entry and conditionally try to free up the swapcache.
4539 * We're already holding a reference on the page but haven't mapped it
4542 swap_free_nr(entry
, nr_pages
);
4543 if (should_try_to_free_swap(folio
, vma
, vmf
->flags
))
4544 folio_free_swap(folio
);
4546 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, nr_pages
);
4547 add_mm_counter(vma
->vm_mm
, MM_SWAPENTS
, -nr_pages
);
4548 pte
= mk_pte(page
, vma
->vm_page_prot
);
4549 if (pte_swp_soft_dirty(vmf
->orig_pte
))
4550 pte
= pte_mksoft_dirty(pte
);
4551 if (pte_swp_uffd_wp(vmf
->orig_pte
))
4552 pte
= pte_mkuffd_wp(pte
);
4555 * Same logic as in do_wp_page(); however, optimize for pages that are
4556 * certainly not shared either because we just allocated them without
4557 * exposing them to the swapcache or because the swap entry indicates
4560 if (!folio_test_ksm(folio
) &&
4561 (exclusive
|| folio_ref_count(folio
) == 1)) {
4562 if ((vma
->vm_flags
& VM_WRITE
) && !userfaultfd_pte_wp(vma
, pte
) &&
4563 !pte_needs_soft_dirty_wp(vma
, pte
)) {
4564 pte
= pte_mkwrite(pte
, vma
);
4565 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4566 pte
= pte_mkdirty(pte
);
4567 vmf
->flags
&= ~FAULT_FLAG_WRITE
;
4570 rmap_flags
|= RMAP_EXCLUSIVE
;
4572 folio_ref_add(folio
, nr_pages
- 1);
4573 flush_icache_pages(vma
, page
, nr_pages
);
4574 vmf
->orig_pte
= pte_advance_pfn(pte
, page_idx
);
4576 /* ksm created a completely new copy */
4577 if (unlikely(folio
!= swapcache
&& swapcache
)) {
4578 folio_add_new_anon_rmap(folio
, vma
, address
, RMAP_EXCLUSIVE
);
4579 folio_add_lru_vma(folio
, vma
);
4580 } else if (!folio_test_anon(folio
)) {
4582 * We currently only expect small !anon folios which are either
4583 * fully exclusive or fully shared, or new allocated large
4584 * folios which are fully exclusive. If we ever get large
4585 * folios within swapcache here, we have to be careful.
4587 VM_WARN_ON_ONCE(folio_test_large(folio
) && folio_test_swapcache(folio
));
4588 VM_WARN_ON_FOLIO(!folio_test_locked(folio
), folio
);
4589 folio_add_new_anon_rmap(folio
, vma
, address
, rmap_flags
);
4591 folio_add_anon_rmap_ptes(folio
, page
, nr_pages
, vma
, address
,
4595 VM_BUG_ON(!folio_test_anon(folio
) ||
4596 (pte_write(pte
) && !PageAnonExclusive(page
)));
4597 set_ptes(vma
->vm_mm
, address
, ptep
, pte
, nr_pages
);
4598 arch_do_swap_page_nr(vma
->vm_mm
, vma
, address
,
4599 pte
, pte
, nr_pages
);
4601 folio_unlock(folio
);
4602 if (folio
!= swapcache
&& swapcache
) {
4604 * Hold the lock to avoid the swap entry to be reused
4605 * until we take the PT lock for the pte_same() check
4606 * (to avoid false positives from pte_same). For
4607 * further safety release the lock after the swap_free
4608 * so that the swap count won't change under a
4609 * parallel locked swapcache.
4611 folio_unlock(swapcache
);
4612 folio_put(swapcache
);
4615 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4616 ret
|= do_wp_page(vmf
);
4617 if (ret
& VM_FAULT_ERROR
)
4618 ret
&= VM_FAULT_ERROR
;
4622 /* No need to invalidate - it was non-present before */
4623 update_mmu_cache_range(vmf
, vma
, address
, ptep
, nr_pages
);
4626 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4628 /* Clear the swap cache pin for direct swapin after PTL unlock */
4629 if (need_clear_cache
) {
4630 swapcache_clear(si
, entry
, nr_pages
);
4631 if (waitqueue_active(&swapcache_wq
))
4632 wake_up(&swapcache_wq
);
4635 put_swap_device(si
);
4639 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4641 folio_unlock(folio
);
4644 if (folio
!= swapcache
&& swapcache
) {
4645 folio_unlock(swapcache
);
4646 folio_put(swapcache
);
4648 if (need_clear_cache
) {
4649 swapcache_clear(si
, entry
, nr_pages
);
4650 if (waitqueue_active(&swapcache_wq
))
4651 wake_up(&swapcache_wq
);
4654 put_swap_device(si
);
4658 static bool pte_range_none(pte_t
*pte
, int nr_pages
)
4662 for (i
= 0; i
< nr_pages
; i
++) {
4663 if (!pte_none(ptep_get_lockless(pte
+ i
)))
4670 static struct folio
*alloc_anon_folio(struct vm_fault
*vmf
)
4672 struct vm_area_struct
*vma
= vmf
->vma
;
4673 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4674 unsigned long orders
;
4675 struct folio
*folio
;
4682 * If uffd is active for the vma we need per-page fault fidelity to
4683 * maintain the uffd semantics.
4685 if (unlikely(userfaultfd_armed(vma
)))
4689 * Get a list of all the (large) orders below PMD_ORDER that are enabled
4690 * for this vma. Then filter out the orders that can't be allocated over
4691 * the faulting address and still be fully contained in the vma.
4693 orders
= thp_vma_allowable_orders(vma
, vma
->vm_flags
,
4694 TVA_IN_PF
| TVA_ENFORCE_SYSFS
, BIT(PMD_ORDER
) - 1);
4695 orders
= thp_vma_suitable_orders(vma
, vmf
->address
, orders
);
4700 pte
= pte_offset_map(vmf
->pmd
, vmf
->address
& PMD_MASK
);
4702 return ERR_PTR(-EAGAIN
);
4705 * Find the highest order where the aligned range is completely
4706 * pte_none(). Note that all remaining orders will be completely
4709 order
= highest_order(orders
);
4711 addr
= ALIGN_DOWN(vmf
->address
, PAGE_SIZE
<< order
);
4712 if (pte_range_none(pte
+ pte_index(addr
), 1 << order
))
4714 order
= next_order(&orders
, order
);
4722 /* Try allocating the highest of the remaining orders. */
4723 gfp
= vma_thp_gfp_mask(vma
);
4725 addr
= ALIGN_DOWN(vmf
->address
, PAGE_SIZE
<< order
);
4726 folio
= vma_alloc_folio(gfp
, order
, vma
, addr
);
4728 if (mem_cgroup_charge(folio
, vma
->vm_mm
, gfp
)) {
4729 count_mthp_stat(order
, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE
);
4733 folio_throttle_swaprate(folio
, gfp
);
4735 * When a folio is not zeroed during allocation
4736 * (__GFP_ZERO not used), folio_zero_user() is used
4737 * to make sure that the page corresponding to the
4738 * faulting address will be hot in the cache after
4741 if (!alloc_zeroed())
4742 folio_zero_user(folio
, vmf
->address
);
4746 count_mthp_stat(order
, MTHP_STAT_ANON_FAULT_FALLBACK
);
4747 order
= next_order(&orders
, order
);
4752 return folio_prealloc(vma
->vm_mm
, vma
, vmf
->address
, true);
4756 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4757 * but allow concurrent faults), and pte mapped but not yet locked.
4758 * We return with mmap_lock still held, but pte unmapped and unlocked.
4760 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
)
4762 struct vm_area_struct
*vma
= vmf
->vma
;
4763 unsigned long addr
= vmf
->address
;
4764 struct folio
*folio
;
4769 /* File mapping without ->vm_ops ? */
4770 if (vma
->vm_flags
& VM_SHARED
)
4771 return VM_FAULT_SIGBUS
;
4774 * Use pte_alloc() instead of pte_alloc_map(), so that OOM can
4775 * be distinguished from a transient failure of pte_offset_map().
4777 if (pte_alloc(vma
->vm_mm
, vmf
->pmd
))
4778 return VM_FAULT_OOM
;
4780 /* Use the zero-page for reads */
4781 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
4782 !mm_forbids_zeropage(vma
->vm_mm
)) {
4783 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(vmf
->address
),
4784 vma
->vm_page_prot
));
4785 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4786 vmf
->address
, &vmf
->ptl
);
4789 if (vmf_pte_changed(vmf
)) {
4790 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
4793 ret
= check_stable_address_space(vma
->vm_mm
);
4796 /* Deliver the page fault to userland, check inside PT lock */
4797 if (userfaultfd_missing(vma
)) {
4798 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4799 return handle_userfault(vmf
, VM_UFFD_MISSING
);
4804 /* Allocate our own private page. */
4805 ret
= vmf_anon_prepare(vmf
);
4808 /* Returns NULL on OOM or ERR_PTR(-EAGAIN) if we must retry the fault */
4809 folio
= alloc_anon_folio(vmf
);
4815 nr_pages
= folio_nr_pages(folio
);
4816 addr
= ALIGN_DOWN(vmf
->address
, nr_pages
* PAGE_SIZE
);
4819 * The memory barrier inside __folio_mark_uptodate makes sure that
4820 * preceding stores to the page contents become visible before
4821 * the set_pte_at() write.
4823 __folio_mark_uptodate(folio
);
4825 entry
= mk_pte(&folio
->page
, vma
->vm_page_prot
);
4826 entry
= pte_sw_mkyoung(entry
);
4827 if (vma
->vm_flags
& VM_WRITE
)
4828 entry
= pte_mkwrite(pte_mkdirty(entry
), vma
);
4830 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
4833 if (nr_pages
== 1 && vmf_pte_changed(vmf
)) {
4834 update_mmu_tlb(vma
, addr
, vmf
->pte
);
4836 } else if (nr_pages
> 1 && !pte_range_none(vmf
->pte
, nr_pages
)) {
4837 update_mmu_tlb_range(vma
, addr
, vmf
->pte
, nr_pages
);
4841 ret
= check_stable_address_space(vma
->vm_mm
);
4845 /* Deliver the page fault to userland, check inside PT lock */
4846 if (userfaultfd_missing(vma
)) {
4847 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4849 return handle_userfault(vmf
, VM_UFFD_MISSING
);
4852 folio_ref_add(folio
, nr_pages
- 1);
4853 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, nr_pages
);
4854 count_mthp_stat(folio_order(folio
), MTHP_STAT_ANON_FAULT_ALLOC
);
4855 folio_add_new_anon_rmap(folio
, vma
, addr
, RMAP_EXCLUSIVE
);
4856 folio_add_lru_vma(folio
, vma
);
4858 if (vmf_orig_pte_uffd_wp(vmf
))
4859 entry
= pte_mkuffd_wp(entry
);
4860 set_ptes(vma
->vm_mm
, addr
, vmf
->pte
, entry
, nr_pages
);
4862 /* No need to invalidate - it was non-present before */
4863 update_mmu_cache_range(vmf
, vma
, addr
, vmf
->pte
, nr_pages
);
4866 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4872 return VM_FAULT_OOM
;
4876 * The mmap_lock must have been held on entry, and may have been
4877 * released depending on flags and vma->vm_ops->fault() return value.
4878 * See filemap_fault() and __lock_page_retry().
4880 static vm_fault_t
__do_fault(struct vm_fault
*vmf
)
4882 struct vm_area_struct
*vma
= vmf
->vma
;
4883 struct folio
*folio
;
4887 * Preallocate pte before we take page_lock because this might lead to
4888 * deadlocks for memcg reclaim which waits for pages under writeback:
4890 * SetPageWriteback(A)
4896 * wait_on_page_writeback(A)
4897 * SetPageWriteback(B)
4899 * # flush A, B to clear the writeback
4901 if (pmd_none(*vmf
->pmd
) && !vmf
->prealloc_pte
) {
4902 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
4903 if (!vmf
->prealloc_pte
)
4904 return VM_FAULT_OOM
;
4907 ret
= vma
->vm_ops
->fault(vmf
);
4908 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
|
4909 VM_FAULT_DONE_COW
)))
4912 folio
= page_folio(vmf
->page
);
4913 if (unlikely(PageHWPoison(vmf
->page
))) {
4914 vm_fault_t poisonret
= VM_FAULT_HWPOISON
;
4915 if (ret
& VM_FAULT_LOCKED
) {
4916 if (page_mapped(vmf
->page
))
4917 unmap_mapping_folio(folio
);
4918 /* Retry if a clean folio was removed from the cache. */
4919 if (mapping_evict_folio(folio
->mapping
, folio
))
4920 poisonret
= VM_FAULT_NOPAGE
;
4921 folio_unlock(folio
);
4928 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
4931 VM_BUG_ON_PAGE(!folio_test_locked(folio
), vmf
->page
);
4936 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4937 static void deposit_prealloc_pte(struct vm_fault
*vmf
)
4939 struct vm_area_struct
*vma
= vmf
->vma
;
4941 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
4943 * We are going to consume the prealloc table,
4944 * count that as nr_ptes.
4946 mm_inc_nr_ptes(vma
->vm_mm
);
4947 vmf
->prealloc_pte
= NULL
;
4950 vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
4952 struct folio
*folio
= page_folio(page
);
4953 struct vm_area_struct
*vma
= vmf
->vma
;
4954 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
4955 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
4957 vm_fault_t ret
= VM_FAULT_FALLBACK
;
4960 * It is too late to allocate a small folio, we already have a large
4961 * folio in the pagecache: especially s390 KVM cannot tolerate any
4962 * PMD mappings, but PTE-mapped THP are fine. So let's simply refuse any
4963 * PMD mappings if THPs are disabled.
4965 if (thp_disabled_by_hw() || vma_thp_disabled(vma
, vma
->vm_flags
))
4968 if (!thp_vma_suitable_order(vma
, haddr
, PMD_ORDER
))
4971 if (folio_order(folio
) != HPAGE_PMD_ORDER
)
4973 page
= &folio
->page
;
4976 * Just backoff if any subpage of a THP is corrupted otherwise
4977 * the corrupted page may mapped by PMD silently to escape the
4978 * check. This kind of THP just can be PTE mapped. Access to
4979 * the corrupted subpage should trigger SIGBUS as expected.
4981 if (unlikely(folio_test_has_hwpoisoned(folio
)))
4985 * Archs like ppc64 need additional space to store information
4986 * related to pte entry. Use the preallocated table for that.
4988 if (arch_needs_pgtable_deposit() && !vmf
->prealloc_pte
) {
4989 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
4990 if (!vmf
->prealloc_pte
)
4991 return VM_FAULT_OOM
;
4994 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
4995 if (unlikely(!pmd_none(*vmf
->pmd
)))
4998 flush_icache_pages(vma
, page
, HPAGE_PMD_NR
);
5000 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
5002 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
5004 add_mm_counter(vma
->vm_mm
, mm_counter_file(folio
), HPAGE_PMD_NR
);
5005 folio_add_file_rmap_pmd(folio
, page
, vma
);
5008 * deposit and withdraw with pmd lock held
5010 if (arch_needs_pgtable_deposit())
5011 deposit_prealloc_pte(vmf
);
5013 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
5015 update_mmu_cache_pmd(vma
, haddr
, vmf
->pmd
);
5017 /* fault is handled */
5019 count_vm_event(THP_FILE_MAPPED
);
5021 spin_unlock(vmf
->ptl
);
5025 vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
5027 return VM_FAULT_FALLBACK
;
5032 * set_pte_range - Set a range of PTEs to point to pages in a folio.
5033 * @vmf: Fault decription.
5034 * @folio: The folio that contains @page.
5035 * @page: The first page to create a PTE for.
5036 * @nr: The number of PTEs to create.
5037 * @addr: The first address to create a PTE for.
5039 void set_pte_range(struct vm_fault
*vmf
, struct folio
*folio
,
5040 struct page
*page
, unsigned int nr
, unsigned long addr
)
5042 struct vm_area_struct
*vma
= vmf
->vma
;
5043 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
5044 bool prefault
= !in_range(vmf
->address
, addr
, nr
* PAGE_SIZE
);
5047 flush_icache_pages(vma
, page
, nr
);
5048 entry
= mk_pte(page
, vma
->vm_page_prot
);
5050 if (prefault
&& arch_wants_old_prefaulted_pte())
5051 entry
= pte_mkold(entry
);
5053 entry
= pte_sw_mkyoung(entry
);
5056 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
5057 if (unlikely(vmf_orig_pte_uffd_wp(vmf
)))
5058 entry
= pte_mkuffd_wp(entry
);
5059 /* copy-on-write page */
5060 if (write
&& !(vma
->vm_flags
& VM_SHARED
)) {
5061 VM_BUG_ON_FOLIO(nr
!= 1, folio
);
5062 folio_add_new_anon_rmap(folio
, vma
, addr
, RMAP_EXCLUSIVE
);
5063 folio_add_lru_vma(folio
, vma
);
5065 folio_add_file_rmap_ptes(folio
, page
, nr
, vma
);
5067 set_ptes(vma
->vm_mm
, addr
, vmf
->pte
, entry
, nr
);
5069 /* no need to invalidate: a not-present page won't be cached */
5070 update_mmu_cache_range(vmf
, vma
, addr
, vmf
->pte
, nr
);
5073 static bool vmf_pte_changed(struct vm_fault
*vmf
)
5075 if (vmf
->flags
& FAULT_FLAG_ORIG_PTE_VALID
)
5076 return !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
);
5078 return !pte_none(ptep_get(vmf
->pte
));
5082 * finish_fault - finish page fault once we have prepared the page to fault
5084 * @vmf: structure describing the fault
5086 * This function handles all that is needed to finish a page fault once the
5087 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
5088 * given page, adds reverse page mapping, handles memcg charges and LRU
5091 * The function expects the page to be locked and on success it consumes a
5092 * reference of a page being mapped (for the PTE which maps it).
5094 * Return: %0 on success, %VM_FAULT_ code in case of error.
5096 vm_fault_t
finish_fault(struct vm_fault
*vmf
)
5098 struct vm_area_struct
*vma
= vmf
->vma
;
5100 struct folio
*folio
;
5102 bool is_cow
= (vmf
->flags
& FAULT_FLAG_WRITE
) &&
5103 !(vma
->vm_flags
& VM_SHARED
);
5105 unsigned long addr
= vmf
->address
;
5107 /* Did we COW the page? */
5109 page
= vmf
->cow_page
;
5114 * check even for read faults because we might have lost our CoWed
5117 if (!(vma
->vm_flags
& VM_SHARED
)) {
5118 ret
= check_stable_address_space(vma
->vm_mm
);
5123 if (pmd_none(*vmf
->pmd
)) {
5124 if (PageTransCompound(page
)) {
5125 ret
= do_set_pmd(vmf
, page
);
5126 if (ret
!= VM_FAULT_FALLBACK
)
5130 if (vmf
->prealloc_pte
)
5131 pmd_install(vma
->vm_mm
, vmf
->pmd
, &vmf
->prealloc_pte
);
5132 else if (unlikely(pte_alloc(vma
->vm_mm
, vmf
->pmd
)))
5133 return VM_FAULT_OOM
;
5136 folio
= page_folio(page
);
5137 nr_pages
= folio_nr_pages(folio
);
5140 * Using per-page fault to maintain the uffd semantics, and same
5141 * approach also applies to non-anonymous-shmem faults to avoid
5142 * inflating the RSS of the process.
5144 if (!vma_is_anon_shmem(vma
) || unlikely(userfaultfd_armed(vma
))) {
5146 } else if (nr_pages
> 1) {
5147 pgoff_t idx
= folio_page_idx(folio
, page
);
5148 /* The page offset of vmf->address within the VMA. */
5149 pgoff_t vma_off
= vmf
->pgoff
- vmf
->vma
->vm_pgoff
;
5150 /* The index of the entry in the pagetable for fault page. */
5151 pgoff_t pte_off
= pte_index(vmf
->address
);
5154 * Fallback to per-page fault in case the folio size in page
5155 * cache beyond the VMA limits and PMD pagetable limits.
5157 if (unlikely(vma_off
< idx
||
5158 vma_off
+ (nr_pages
- idx
) > vma_pages(vma
) ||
5160 pte_off
+ (nr_pages
- idx
) > PTRS_PER_PTE
)) {
5163 /* Now we can set mappings for the whole large folio. */
5164 addr
= vmf
->address
- idx
* PAGE_SIZE
;
5165 page
= &folio
->page
;
5169 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
5172 return VM_FAULT_NOPAGE
;
5174 /* Re-check under ptl */
5175 if (nr_pages
== 1 && unlikely(vmf_pte_changed(vmf
))) {
5176 update_mmu_tlb(vma
, addr
, vmf
->pte
);
5177 ret
= VM_FAULT_NOPAGE
;
5179 } else if (nr_pages
> 1 && !pte_range_none(vmf
->pte
, nr_pages
)) {
5180 update_mmu_tlb_range(vma
, addr
, vmf
->pte
, nr_pages
);
5181 ret
= VM_FAULT_NOPAGE
;
5185 folio_ref_add(folio
, nr_pages
- 1);
5186 set_pte_range(vmf
, folio
, page
, nr_pages
, addr
);
5187 type
= is_cow
? MM_ANONPAGES
: mm_counter_file(folio
);
5188 add_mm_counter(vma
->vm_mm
, type
, nr_pages
);
5192 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5196 static unsigned long fault_around_pages __read_mostly
=
5197 65536 >> PAGE_SHIFT
;
5199 #ifdef CONFIG_DEBUG_FS
5200 static int fault_around_bytes_get(void *data
, u64
*val
)
5202 *val
= fault_around_pages
<< PAGE_SHIFT
;
5207 * fault_around_bytes must be rounded down to the nearest page order as it's
5208 * what do_fault_around() expects to see.
5210 static int fault_around_bytes_set(void *data
, u64 val
)
5212 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
5216 * The minimum value is 1 page, however this results in no fault-around
5217 * at all. See should_fault_around().
5219 val
= max(val
, PAGE_SIZE
);
5220 fault_around_pages
= rounddown_pow_of_two(val
) >> PAGE_SHIFT
;
5224 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops
,
5225 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
5227 static int __init
fault_around_debugfs(void)
5229 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL
, NULL
,
5230 &fault_around_bytes_fops
);
5233 late_initcall(fault_around_debugfs
);
5237 * do_fault_around() tries to map few pages around the fault address. The hope
5238 * is that the pages will be needed soon and this will lower the number of
5241 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
5242 * not ready to be mapped: not up-to-date, locked, etc.
5244 * This function doesn't cross VMA or page table boundaries, in order to call
5245 * map_pages() and acquire a PTE lock only once.
5247 * fault_around_pages defines how many pages we'll try to map.
5248 * do_fault_around() expects it to be set to a power of two less than or equal
5251 * The virtual address of the area that we map is naturally aligned to
5252 * fault_around_pages * PAGE_SIZE rounded down to the machine page size
5253 * (and therefore to page order). This way it's easier to guarantee
5254 * that we don't cross page table boundaries.
5256 static vm_fault_t
do_fault_around(struct vm_fault
*vmf
)
5258 pgoff_t nr_pages
= READ_ONCE(fault_around_pages
);
5259 pgoff_t pte_off
= pte_index(vmf
->address
);
5260 /* The page offset of vmf->address within the VMA. */
5261 pgoff_t vma_off
= vmf
->pgoff
- vmf
->vma
->vm_pgoff
;
5262 pgoff_t from_pte
, to_pte
;
5265 /* The PTE offset of the start address, clamped to the VMA. */
5266 from_pte
= max(ALIGN_DOWN(pte_off
, nr_pages
),
5267 pte_off
- min(pte_off
, vma_off
));
5269 /* The PTE offset of the end address, clamped to the VMA and PTE. */
5270 to_pte
= min3(from_pte
+ nr_pages
, (pgoff_t
)PTRS_PER_PTE
,
5271 pte_off
+ vma_pages(vmf
->vma
) - vma_off
) - 1;
5273 if (pmd_none(*vmf
->pmd
)) {
5274 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
);
5275 if (!vmf
->prealloc_pte
)
5276 return VM_FAULT_OOM
;
5280 ret
= vmf
->vma
->vm_ops
->map_pages(vmf
,
5281 vmf
->pgoff
+ from_pte
- pte_off
,
5282 vmf
->pgoff
+ to_pte
- pte_off
);
5288 /* Return true if we should do read fault-around, false otherwise */
5289 static inline bool should_fault_around(struct vm_fault
*vmf
)
5291 /* No ->map_pages? No way to fault around... */
5292 if (!vmf
->vma
->vm_ops
->map_pages
)
5295 if (uffd_disable_fault_around(vmf
->vma
))
5298 /* A single page implies no faulting 'around' at all. */
5299 return fault_around_pages
> 1;
5302 static vm_fault_t
do_read_fault(struct vm_fault
*vmf
)
5305 struct folio
*folio
;
5308 * Let's call ->map_pages() first and use ->fault() as fallback
5309 * if page by the offset is not ready to be mapped (cold cache or
5312 if (should_fault_around(vmf
)) {
5313 ret
= do_fault_around(vmf
);
5318 ret
= vmf_can_call_fault(vmf
);
5322 ret
= __do_fault(vmf
);
5323 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
5326 ret
|= finish_fault(vmf
);
5327 folio
= page_folio(vmf
->page
);
5328 folio_unlock(folio
);
5329 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
5334 static vm_fault_t
do_cow_fault(struct vm_fault
*vmf
)
5336 struct vm_area_struct
*vma
= vmf
->vma
;
5337 struct folio
*folio
;
5340 ret
= vmf_can_call_fault(vmf
);
5342 ret
= vmf_anon_prepare(vmf
);
5346 folio
= folio_prealloc(vma
->vm_mm
, vma
, vmf
->address
, false);
5348 return VM_FAULT_OOM
;
5350 vmf
->cow_page
= &folio
->page
;
5352 ret
= __do_fault(vmf
);
5353 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
5355 if (ret
& VM_FAULT_DONE_COW
)
5358 if (copy_mc_user_highpage(vmf
->cow_page
, vmf
->page
, vmf
->address
, vma
)) {
5359 ret
= VM_FAULT_HWPOISON
;
5362 __folio_mark_uptodate(folio
);
5364 ret
|= finish_fault(vmf
);
5366 unlock_page(vmf
->page
);
5367 put_page(vmf
->page
);
5368 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
5376 static vm_fault_t
do_shared_fault(struct vm_fault
*vmf
)
5378 struct vm_area_struct
*vma
= vmf
->vma
;
5379 vm_fault_t ret
, tmp
;
5380 struct folio
*folio
;
5382 ret
= vmf_can_call_fault(vmf
);
5386 ret
= __do_fault(vmf
);
5387 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
5390 folio
= page_folio(vmf
->page
);
5393 * Check if the backing address space wants to know that the page is
5394 * about to become writable
5396 if (vma
->vm_ops
->page_mkwrite
) {
5397 folio_unlock(folio
);
5398 tmp
= do_page_mkwrite(vmf
, folio
);
5399 if (unlikely(!tmp
||
5400 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
5406 ret
|= finish_fault(vmf
);
5407 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
5409 folio_unlock(folio
);
5414 ret
|= fault_dirty_shared_page(vmf
);
5419 * We enter with non-exclusive mmap_lock (to exclude vma changes,
5420 * but allow concurrent faults).
5421 * The mmap_lock may have been released depending on flags and our
5422 * return value. See filemap_fault() and __folio_lock_or_retry().
5423 * If mmap_lock is released, vma may become invalid (for example
5424 * by other thread calling munmap()).
5426 static vm_fault_t
do_fault(struct vm_fault
*vmf
)
5428 struct vm_area_struct
*vma
= vmf
->vma
;
5429 struct mm_struct
*vm_mm
= vma
->vm_mm
;
5433 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
5435 if (!vma
->vm_ops
->fault
) {
5436 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
,
5437 vmf
->address
, &vmf
->ptl
);
5438 if (unlikely(!vmf
->pte
))
5439 ret
= VM_FAULT_SIGBUS
;
5442 * Make sure this is not a temporary clearing of pte
5443 * by holding ptl and checking again. A R/M/W update
5444 * of pte involves: take ptl, clearing the pte so that
5445 * we don't have concurrent modification by hardware
5446 * followed by an update.
5448 if (unlikely(pte_none(ptep_get(vmf
->pte
))))
5449 ret
= VM_FAULT_SIGBUS
;
5451 ret
= VM_FAULT_NOPAGE
;
5453 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5455 } else if (!(vmf
->flags
& FAULT_FLAG_WRITE
))
5456 ret
= do_read_fault(vmf
);
5457 else if (!(vma
->vm_flags
& VM_SHARED
))
5458 ret
= do_cow_fault(vmf
);
5460 ret
= do_shared_fault(vmf
);
5462 /* preallocated pagetable is unused: free it */
5463 if (vmf
->prealloc_pte
) {
5464 pte_free(vm_mm
, vmf
->prealloc_pte
);
5465 vmf
->prealloc_pte
= NULL
;
5470 int numa_migrate_check(struct folio
*folio
, struct vm_fault
*vmf
,
5471 unsigned long addr
, int *flags
,
5472 bool writable
, int *last_cpupid
)
5474 struct vm_area_struct
*vma
= vmf
->vma
;
5477 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
5478 * much anyway since they can be in shared cache state. This misses
5479 * the case where a mapping is writable but the process never writes
5480 * to it but pte_write gets cleared during protection updates and
5481 * pte_dirty has unpredictable behaviour between PTE scan updates,
5482 * background writeback, dirty balancing and application behaviour.
5485 *flags
|= TNF_NO_GROUP
;
5488 * Flag if the folio is shared between multiple address spaces. This
5489 * is later used when determining whether to group tasks together
5491 if (folio_likely_mapped_shared(folio
) && (vma
->vm_flags
& VM_SHARED
))
5492 *flags
|= TNF_SHARED
;
5494 * For memory tiering mode, cpupid of slow memory page is used
5495 * to record page access time. So use default value.
5497 if (folio_use_access_time(folio
))
5498 *last_cpupid
= (-1 & LAST_CPUPID_MASK
);
5500 *last_cpupid
= folio_last_cpupid(folio
);
5502 /* Record the current PID acceesing VMA */
5503 vma_set_access_pid_bit(vma
);
5505 count_vm_numa_event(NUMA_HINT_FAULTS
);
5506 #ifdef CONFIG_NUMA_BALANCING
5507 count_memcg_folio_events(folio
, NUMA_HINT_FAULTS
, 1);
5509 if (folio_nid(folio
) == numa_node_id()) {
5510 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
5511 *flags
|= TNF_FAULT_LOCAL
;
5514 return mpol_misplaced(folio
, vmf
, addr
);
5517 static void numa_rebuild_single_mapping(struct vm_fault
*vmf
, struct vm_area_struct
*vma
,
5518 unsigned long fault_addr
, pte_t
*fault_pte
,
5523 old_pte
= ptep_modify_prot_start(vma
, fault_addr
, fault_pte
);
5524 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
5525 pte
= pte_mkyoung(pte
);
5527 pte
= pte_mkwrite(pte
, vma
);
5528 ptep_modify_prot_commit(vma
, fault_addr
, fault_pte
, old_pte
, pte
);
5529 update_mmu_cache_range(vmf
, vma
, fault_addr
, fault_pte
, 1);
5532 static void numa_rebuild_large_mapping(struct vm_fault
*vmf
, struct vm_area_struct
*vma
,
5533 struct folio
*folio
, pte_t fault_pte
,
5534 bool ignore_writable
, bool pte_write_upgrade
)
5536 int nr
= pte_pfn(fault_pte
) - folio_pfn(folio
);
5537 unsigned long start
, end
, addr
= vmf
->address
;
5538 unsigned long addr_start
= addr
- (nr
<< PAGE_SHIFT
);
5539 unsigned long pt_start
= ALIGN_DOWN(addr
, PMD_SIZE
);
5542 /* Stay within the VMA and within the page table. */
5543 start
= max3(addr_start
, pt_start
, vma
->vm_start
);
5544 end
= min3(addr_start
+ folio_size(folio
), pt_start
+ PMD_SIZE
,
5546 start_ptep
= vmf
->pte
- ((addr
- start
) >> PAGE_SHIFT
);
5548 /* Restore all PTEs' mapping of the large folio */
5549 for (addr
= start
; addr
!= end
; start_ptep
++, addr
+= PAGE_SIZE
) {
5550 pte_t ptent
= ptep_get(start_ptep
);
5551 bool writable
= false;
5553 if (!pte_present(ptent
) || !pte_protnone(ptent
))
5556 if (pfn_folio(pte_pfn(ptent
)) != folio
)
5559 if (!ignore_writable
) {
5560 ptent
= pte_modify(ptent
, vma
->vm_page_prot
);
5561 writable
= pte_write(ptent
);
5562 if (!writable
&& pte_write_upgrade
&&
5563 can_change_pte_writable(vma
, addr
, ptent
))
5567 numa_rebuild_single_mapping(vmf
, vma
, addr
, start_ptep
, writable
);
5571 static vm_fault_t
do_numa_page(struct vm_fault
*vmf
)
5573 struct vm_area_struct
*vma
= vmf
->vma
;
5574 struct folio
*folio
= NULL
;
5575 int nid
= NUMA_NO_NODE
;
5576 bool writable
= false, ignore_writable
= false;
5577 bool pte_write_upgrade
= vma_wants_manual_pte_write_upgrade(vma
);
5581 int flags
= 0, nr_pages
;
5584 * The pte cannot be used safely until we verify, while holding the page
5585 * table lock, that its contents have not changed during fault handling.
5587 spin_lock(vmf
->ptl
);
5588 /* Read the live PTE from the page tables: */
5589 old_pte
= ptep_get(vmf
->pte
);
5591 if (unlikely(!pte_same(old_pte
, vmf
->orig_pte
))) {
5592 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5596 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
5599 * Detect now whether the PTE could be writable; this information
5600 * is only valid while holding the PT lock.
5602 writable
= pte_write(pte
);
5603 if (!writable
&& pte_write_upgrade
&&
5604 can_change_pte_writable(vma
, vmf
->address
, pte
))
5607 folio
= vm_normal_folio(vma
, vmf
->address
, pte
);
5608 if (!folio
|| folio_is_zone_device(folio
))
5611 nid
= folio_nid(folio
);
5612 nr_pages
= folio_nr_pages(folio
);
5614 target_nid
= numa_migrate_check(folio
, vmf
, vmf
->address
, &flags
,
5615 writable
, &last_cpupid
);
5616 if (target_nid
== NUMA_NO_NODE
)
5618 if (migrate_misplaced_folio_prepare(folio
, vma
, target_nid
)) {
5619 flags
|= TNF_MIGRATE_FAIL
;
5622 /* The folio is isolated and isolation code holds a folio reference. */
5623 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5625 ignore_writable
= true;
5627 /* Migrate to the requested node */
5628 if (!migrate_misplaced_folio(folio
, vma
, target_nid
)) {
5630 flags
|= TNF_MIGRATED
;
5631 task_numa_fault(last_cpupid
, nid
, nr_pages
, flags
);
5635 flags
|= TNF_MIGRATE_FAIL
;
5636 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
5637 vmf
->address
, &vmf
->ptl
);
5638 if (unlikely(!vmf
->pte
))
5640 if (unlikely(!pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
5641 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5646 * Make it present again, depending on how arch implements
5647 * non-accessible ptes, some can allow access by kernel mode.
5649 if (folio
&& folio_test_large(folio
))
5650 numa_rebuild_large_mapping(vmf
, vma
, folio
, pte
, ignore_writable
,
5653 numa_rebuild_single_mapping(vmf
, vma
, vmf
->address
, vmf
->pte
,
5655 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5657 if (nid
!= NUMA_NO_NODE
)
5658 task_numa_fault(last_cpupid
, nid
, nr_pages
, flags
);
5662 static inline vm_fault_t
create_huge_pmd(struct vm_fault
*vmf
)
5664 struct vm_area_struct
*vma
= vmf
->vma
;
5665 if (vma_is_anonymous(vma
))
5666 return do_huge_pmd_anonymous_page(vmf
);
5667 if (vma
->vm_ops
->huge_fault
)
5668 return vma
->vm_ops
->huge_fault(vmf
, PMD_ORDER
);
5669 return VM_FAULT_FALLBACK
;
5672 /* `inline' is required to avoid gcc 4.1.2 build error */
5673 static inline vm_fault_t
wp_huge_pmd(struct vm_fault
*vmf
)
5675 struct vm_area_struct
*vma
= vmf
->vma
;
5676 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
5679 if (vma_is_anonymous(vma
)) {
5680 if (likely(!unshare
) &&
5681 userfaultfd_huge_pmd_wp(vma
, vmf
->orig_pmd
)) {
5682 if (userfaultfd_wp_async(vmf
->vma
))
5684 return handle_userfault(vmf
, VM_UFFD_WP
);
5686 return do_huge_pmd_wp_page(vmf
);
5689 if (vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
)) {
5690 if (vma
->vm_ops
->huge_fault
) {
5691 ret
= vma
->vm_ops
->huge_fault(vmf
, PMD_ORDER
);
5692 if (!(ret
& VM_FAULT_FALLBACK
))
5698 /* COW or write-notify handled on pte level: split pmd. */
5699 __split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
5701 return VM_FAULT_FALLBACK
;
5704 static vm_fault_t
create_huge_pud(struct vm_fault
*vmf
)
5706 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
5707 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5708 struct vm_area_struct
*vma
= vmf
->vma
;
5709 /* No support for anonymous transparent PUD pages yet */
5710 if (vma_is_anonymous(vma
))
5711 return VM_FAULT_FALLBACK
;
5712 if (vma
->vm_ops
->huge_fault
)
5713 return vma
->vm_ops
->huge_fault(vmf
, PUD_ORDER
);
5714 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
5715 return VM_FAULT_FALLBACK
;
5718 static vm_fault_t
wp_huge_pud(struct vm_fault
*vmf
, pud_t orig_pud
)
5720 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
5721 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5722 struct vm_area_struct
*vma
= vmf
->vma
;
5725 /* No support for anonymous transparent PUD pages yet */
5726 if (vma_is_anonymous(vma
))
5728 if (vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
)) {
5729 if (vma
->vm_ops
->huge_fault
) {
5730 ret
= vma
->vm_ops
->huge_fault(vmf
, PUD_ORDER
);
5731 if (!(ret
& VM_FAULT_FALLBACK
))
5736 /* COW or write-notify not handled on PUD level: split pud.*/
5737 __split_huge_pud(vma
, vmf
->pud
, vmf
->address
);
5738 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
5739 return VM_FAULT_FALLBACK
;
5743 * These routines also need to handle stuff like marking pages dirty
5744 * and/or accessed for architectures that don't do it in hardware (most
5745 * RISC architectures). The early dirtying is also good on the i386.
5747 * There is also a hook called "update_mmu_cache()" that architectures
5748 * with external mmu caches can use to update those (ie the Sparc or
5749 * PowerPC hashed page tables that act as extended TLBs).
5751 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
5752 * concurrent faults).
5754 * The mmap_lock may have been released depending on flags and our return value.
5755 * See filemap_fault() and __folio_lock_or_retry().
5757 static vm_fault_t
handle_pte_fault(struct vm_fault
*vmf
)
5761 if (unlikely(pmd_none(*vmf
->pmd
))) {
5763 * Leave __pte_alloc() until later: because vm_ops->fault may
5764 * want to allocate huge page, and if we expose page table
5765 * for an instant, it will be difficult to retract from
5766 * concurrent faults and from rmap lookups.
5769 vmf
->flags
&= ~FAULT_FLAG_ORIG_PTE_VALID
;
5774 * A regular pmd is established and it can't morph into a huge
5775 * pmd by anon khugepaged, since that takes mmap_lock in write
5776 * mode; but shmem or file collapse to THP could still morph
5777 * it into a huge pmd: just retry later if so.
5779 * Use the maywrite version to indicate that vmf->pte may be
5780 * modified, but since we will use pte_same() to detect the
5781 * change of the !pte_none() entry, there is no need to recheck
5782 * the pmdval. Here we chooes to pass a dummy variable instead
5783 * of NULL, which helps new user think about why this place is
5786 vmf
->pte
= pte_offset_map_rw_nolock(vmf
->vma
->vm_mm
, vmf
->pmd
,
5787 vmf
->address
, &dummy_pmdval
,
5789 if (unlikely(!vmf
->pte
))
5791 vmf
->orig_pte
= ptep_get_lockless(vmf
->pte
);
5792 vmf
->flags
|= FAULT_FLAG_ORIG_PTE_VALID
;
5794 if (pte_none(vmf
->orig_pte
)) {
5795 pte_unmap(vmf
->pte
);
5801 return do_pte_missing(vmf
);
5803 if (!pte_present(vmf
->orig_pte
))
5804 return do_swap_page(vmf
);
5806 if (pte_protnone(vmf
->orig_pte
) && vma_is_accessible(vmf
->vma
))
5807 return do_numa_page(vmf
);
5809 spin_lock(vmf
->ptl
);
5810 entry
= vmf
->orig_pte
;
5811 if (unlikely(!pte_same(ptep_get(vmf
->pte
), entry
))) {
5812 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
5815 if (vmf
->flags
& (FAULT_FLAG_WRITE
|FAULT_FLAG_UNSHARE
)) {
5816 if (!pte_write(entry
))
5817 return do_wp_page(vmf
);
5818 else if (likely(vmf
->flags
& FAULT_FLAG_WRITE
))
5819 entry
= pte_mkdirty(entry
);
5821 entry
= pte_mkyoung(entry
);
5822 if (ptep_set_access_flags(vmf
->vma
, vmf
->address
, vmf
->pte
, entry
,
5823 vmf
->flags
& FAULT_FLAG_WRITE
)) {
5824 update_mmu_cache_range(vmf
, vmf
->vma
, vmf
->address
,
5827 /* Skip spurious TLB flush for retried page fault */
5828 if (vmf
->flags
& FAULT_FLAG_TRIED
)
5831 * This is needed only for protection faults but the arch code
5832 * is not yet telling us if this is a protection fault or not.
5833 * This still avoids useless tlb flushes for .text page faults
5836 if (vmf
->flags
& FAULT_FLAG_WRITE
)
5837 flush_tlb_fix_spurious_fault(vmf
->vma
, vmf
->address
,
5841 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5846 * On entry, we hold either the VMA lock or the mmap_lock
5847 * (FAULT_FLAG_VMA_LOCK tells you which). If VM_FAULT_RETRY is set in
5848 * the result, the mmap_lock is not held on exit. See filemap_fault()
5849 * and __folio_lock_or_retry().
5851 static vm_fault_t
__handle_mm_fault(struct vm_area_struct
*vma
,
5852 unsigned long address
, unsigned int flags
)
5854 struct vm_fault vmf
= {
5856 .address
= address
& PAGE_MASK
,
5857 .real_address
= address
,
5859 .pgoff
= linear_page_index(vma
, address
),
5860 .gfp_mask
= __get_fault_gfp_mask(vma
),
5862 struct mm_struct
*mm
= vma
->vm_mm
;
5863 unsigned long vm_flags
= vma
->vm_flags
;
5868 pgd
= pgd_offset(mm
, address
);
5869 p4d
= p4d_alloc(mm
, pgd
, address
);
5871 return VM_FAULT_OOM
;
5873 vmf
.pud
= pud_alloc(mm
, p4d
, address
);
5875 return VM_FAULT_OOM
;
5877 if (pud_none(*vmf
.pud
) &&
5878 thp_vma_allowable_order(vma
, vm_flags
,
5879 TVA_IN_PF
| TVA_ENFORCE_SYSFS
, PUD_ORDER
)) {
5880 ret
= create_huge_pud(&vmf
);
5881 if (!(ret
& VM_FAULT_FALLBACK
))
5884 pud_t orig_pud
= *vmf
.pud
;
5887 if (pud_trans_huge(orig_pud
) || pud_devmap(orig_pud
)) {
5890 * TODO once we support anonymous PUDs: NUMA case and
5891 * FAULT_FLAG_UNSHARE handling.
5893 if ((flags
& FAULT_FLAG_WRITE
) && !pud_write(orig_pud
)) {
5894 ret
= wp_huge_pud(&vmf
, orig_pud
);
5895 if (!(ret
& VM_FAULT_FALLBACK
))
5898 huge_pud_set_accessed(&vmf
, orig_pud
);
5904 vmf
.pmd
= pmd_alloc(mm
, vmf
.pud
, address
);
5906 return VM_FAULT_OOM
;
5908 /* Huge pud page fault raced with pmd_alloc? */
5909 if (pud_trans_unstable(vmf
.pud
))
5912 if (pmd_none(*vmf
.pmd
) &&
5913 thp_vma_allowable_order(vma
, vm_flags
,
5914 TVA_IN_PF
| TVA_ENFORCE_SYSFS
, PMD_ORDER
)) {
5915 ret
= create_huge_pmd(&vmf
);
5916 if (!(ret
& VM_FAULT_FALLBACK
))
5919 vmf
.orig_pmd
= pmdp_get_lockless(vmf
.pmd
);
5921 if (unlikely(is_swap_pmd(vmf
.orig_pmd
))) {
5922 VM_BUG_ON(thp_migration_supported() &&
5923 !is_pmd_migration_entry(vmf
.orig_pmd
));
5924 if (is_pmd_migration_entry(vmf
.orig_pmd
))
5925 pmd_migration_entry_wait(mm
, vmf
.pmd
);
5928 if (pmd_trans_huge(vmf
.orig_pmd
) || pmd_devmap(vmf
.orig_pmd
)) {
5929 if (pmd_protnone(vmf
.orig_pmd
) && vma_is_accessible(vma
))
5930 return do_huge_pmd_numa_page(&vmf
);
5932 if ((flags
& (FAULT_FLAG_WRITE
|FAULT_FLAG_UNSHARE
)) &&
5933 !pmd_write(vmf
.orig_pmd
)) {
5934 ret
= wp_huge_pmd(&vmf
);
5935 if (!(ret
& VM_FAULT_FALLBACK
))
5938 huge_pmd_set_accessed(&vmf
);
5944 return handle_pte_fault(&vmf
);
5948 * mm_account_fault - Do page fault accounting
5949 * @mm: mm from which memcg should be extracted. It can be NULL.
5950 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
5951 * of perf event counters, but we'll still do the per-task accounting to
5952 * the task who triggered this page fault.
5953 * @address: the faulted address.
5954 * @flags: the fault flags.
5955 * @ret: the fault retcode.
5957 * This will take care of most of the page fault accounting. Meanwhile, it
5958 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
5959 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
5960 * still be in per-arch page fault handlers at the entry of page fault.
5962 static inline void mm_account_fault(struct mm_struct
*mm
, struct pt_regs
*regs
,
5963 unsigned long address
, unsigned int flags
,
5968 /* Incomplete faults will be accounted upon completion. */
5969 if (ret
& VM_FAULT_RETRY
)
5973 * To preserve the behavior of older kernels, PGFAULT counters record
5974 * both successful and failed faults, as opposed to perf counters,
5975 * which ignore failed cases.
5977 count_vm_event(PGFAULT
);
5978 count_memcg_event_mm(mm
, PGFAULT
);
5981 * Do not account for unsuccessful faults (e.g. when the address wasn't
5982 * valid). That includes arch_vma_access_permitted() failing before
5983 * reaching here. So this is not a "this many hardware page faults"
5984 * counter. We should use the hw profiling for that.
5986 if (ret
& VM_FAULT_ERROR
)
5990 * We define the fault as a major fault when the final successful fault
5991 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
5992 * handle it immediately previously).
5994 major
= (ret
& VM_FAULT_MAJOR
) || (flags
& FAULT_FLAG_TRIED
);
6002 * If the fault is done for GUP, regs will be NULL. We only do the
6003 * accounting for the per thread fault counters who triggered the
6004 * fault, and we skip the perf event updates.
6010 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ
, 1, regs
, address
);
6012 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN
, 1, regs
, address
);
6015 #ifdef CONFIG_LRU_GEN
6016 static void lru_gen_enter_fault(struct vm_area_struct
*vma
)
6018 /* the LRU algorithm only applies to accesses with recency */
6019 current
->in_lru_fault
= vma_has_recency(vma
);
6022 static void lru_gen_exit_fault(void)
6024 current
->in_lru_fault
= false;
6027 static void lru_gen_enter_fault(struct vm_area_struct
*vma
)
6031 static void lru_gen_exit_fault(void)
6034 #endif /* CONFIG_LRU_GEN */
6036 static vm_fault_t
sanitize_fault_flags(struct vm_area_struct
*vma
,
6037 unsigned int *flags
)
6039 if (unlikely(*flags
& FAULT_FLAG_UNSHARE
)) {
6040 if (WARN_ON_ONCE(*flags
& FAULT_FLAG_WRITE
))
6041 return VM_FAULT_SIGSEGV
;
6043 * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's
6044 * just treat it like an ordinary read-fault otherwise.
6046 if (!is_cow_mapping(vma
->vm_flags
))
6047 *flags
&= ~FAULT_FLAG_UNSHARE
;
6048 } else if (*flags
& FAULT_FLAG_WRITE
) {
6049 /* Write faults on read-only mappings are impossible ... */
6050 if (WARN_ON_ONCE(!(vma
->vm_flags
& VM_MAYWRITE
)))
6051 return VM_FAULT_SIGSEGV
;
6052 /* ... and FOLL_FORCE only applies to COW mappings. */
6053 if (WARN_ON_ONCE(!(vma
->vm_flags
& VM_WRITE
) &&
6054 !is_cow_mapping(vma
->vm_flags
)))
6055 return VM_FAULT_SIGSEGV
;
6057 #ifdef CONFIG_PER_VMA_LOCK
6059 * Per-VMA locks can't be used with FAULT_FLAG_RETRY_NOWAIT because of
6060 * the assumption that lock is dropped on VM_FAULT_RETRY.
6062 if (WARN_ON_ONCE((*flags
&
6063 (FAULT_FLAG_VMA_LOCK
| FAULT_FLAG_RETRY_NOWAIT
)) ==
6064 (FAULT_FLAG_VMA_LOCK
| FAULT_FLAG_RETRY_NOWAIT
)))
6065 return VM_FAULT_SIGSEGV
;
6072 * By the time we get here, we already hold the mm semaphore
6074 * The mmap_lock may have been released depending on flags and our
6075 * return value. See filemap_fault() and __folio_lock_or_retry().
6077 vm_fault_t
handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
6078 unsigned int flags
, struct pt_regs
*regs
)
6080 /* If the fault handler drops the mmap_lock, vma may be freed */
6081 struct mm_struct
*mm
= vma
->vm_mm
;
6085 __set_current_state(TASK_RUNNING
);
6087 ret
= sanitize_fault_flags(vma
, &flags
);
6091 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
6092 flags
& FAULT_FLAG_INSTRUCTION
,
6093 flags
& FAULT_FLAG_REMOTE
)) {
6094 ret
= VM_FAULT_SIGSEGV
;
6098 is_droppable
= !!(vma
->vm_flags
& VM_DROPPABLE
);
6101 * Enable the memcg OOM handling for faults triggered in user
6102 * space. Kernel faults are handled more gracefully.
6104 if (flags
& FAULT_FLAG_USER
)
6105 mem_cgroup_enter_user_fault();
6107 lru_gen_enter_fault(vma
);
6109 if (unlikely(is_vm_hugetlb_page(vma
)))
6110 ret
= hugetlb_fault(vma
->vm_mm
, vma
, address
, flags
);
6112 ret
= __handle_mm_fault(vma
, address
, flags
);
6115 * Warning: It is no longer safe to dereference vma-> after this point,
6116 * because mmap_lock might have been dropped by __handle_mm_fault(), so
6117 * vma might be destroyed from underneath us.
6120 lru_gen_exit_fault();
6122 /* If the mapping is droppable, then errors due to OOM aren't fatal. */
6124 ret
&= ~VM_FAULT_OOM
;
6126 if (flags
& FAULT_FLAG_USER
) {
6127 mem_cgroup_exit_user_fault();
6129 * The task may have entered a memcg OOM situation but
6130 * if the allocation error was handled gracefully (no
6131 * VM_FAULT_OOM), there is no need to kill anything.
6132 * Just clean up the OOM state peacefully.
6134 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
6135 mem_cgroup_oom_synchronize(false);
6138 mm_account_fault(mm
, regs
, address
, flags
, ret
);
6142 EXPORT_SYMBOL_GPL(handle_mm_fault
);
6144 #ifdef CONFIG_LOCK_MM_AND_FIND_VMA
6145 #include <linux/extable.h>
6147 static inline bool get_mmap_lock_carefully(struct mm_struct
*mm
, struct pt_regs
*regs
)
6149 if (likely(mmap_read_trylock(mm
)))
6152 if (regs
&& !user_mode(regs
)) {
6153 unsigned long ip
= exception_ip(regs
);
6154 if (!search_exception_tables(ip
))
6158 return !mmap_read_lock_killable(mm
);
6161 static inline bool mmap_upgrade_trylock(struct mm_struct
*mm
)
6164 * We don't have this operation yet.
6166 * It should be easy enough to do: it's basically a
6167 * atomic_long_try_cmpxchg_acquire()
6168 * from RWSEM_READER_BIAS -> RWSEM_WRITER_LOCKED, but
6169 * it also needs the proper lockdep magic etc.
6174 static inline bool upgrade_mmap_lock_carefully(struct mm_struct
*mm
, struct pt_regs
*regs
)
6176 mmap_read_unlock(mm
);
6177 if (regs
&& !user_mode(regs
)) {
6178 unsigned long ip
= exception_ip(regs
);
6179 if (!search_exception_tables(ip
))
6182 return !mmap_write_lock_killable(mm
);
6186 * Helper for page fault handling.
6188 * This is kind of equivalend to "mmap_read_lock()" followed
6189 * by "find_extend_vma()", except it's a lot more careful about
6190 * the locking (and will drop the lock on failure).
6192 * For example, if we have a kernel bug that causes a page
6193 * fault, we don't want to just use mmap_read_lock() to get
6194 * the mm lock, because that would deadlock if the bug were
6195 * to happen while we're holding the mm lock for writing.
6197 * So this checks the exception tables on kernel faults in
6198 * order to only do this all for instructions that are actually
6199 * expected to fault.
6201 * We can also actually take the mm lock for writing if we
6202 * need to extend the vma, which helps the VM layer a lot.
6204 struct vm_area_struct
*lock_mm_and_find_vma(struct mm_struct
*mm
,
6205 unsigned long addr
, struct pt_regs
*regs
)
6207 struct vm_area_struct
*vma
;
6209 if (!get_mmap_lock_carefully(mm
, regs
))
6212 vma
= find_vma(mm
, addr
);
6213 if (likely(vma
&& (vma
->vm_start
<= addr
)))
6217 * Well, dang. We might still be successful, but only
6218 * if we can extend a vma to do so.
6220 if (!vma
|| !(vma
->vm_flags
& VM_GROWSDOWN
)) {
6221 mmap_read_unlock(mm
);
6226 * We can try to upgrade the mmap lock atomically,
6227 * in which case we can continue to use the vma
6228 * we already looked up.
6230 * Otherwise we'll have to drop the mmap lock and
6231 * re-take it, and also look up the vma again,
6234 if (!mmap_upgrade_trylock(mm
)) {
6235 if (!upgrade_mmap_lock_carefully(mm
, regs
))
6238 vma
= find_vma(mm
, addr
);
6241 if (vma
->vm_start
<= addr
)
6243 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
6247 if (expand_stack_locked(vma
, addr
))
6251 mmap_write_downgrade(mm
);
6255 mmap_write_unlock(mm
);
6260 #ifdef CONFIG_PER_VMA_LOCK
6262 * Lookup and lock a VMA under RCU protection. Returned VMA is guaranteed to be
6263 * stable and not isolated. If the VMA is not found or is being modified the
6264 * function returns NULL.
6266 struct vm_area_struct
*lock_vma_under_rcu(struct mm_struct
*mm
,
6267 unsigned long address
)
6269 MA_STATE(mas
, &mm
->mm_mt
, address
, address
);
6270 struct vm_area_struct
*vma
;
6274 vma
= mas_walk(&mas
);
6278 if (!vma_start_read(vma
))
6281 /* Check if the VMA got isolated after we found it */
6282 if (vma
->detached
) {
6284 count_vm_vma_lock_event(VMA_LOCK_MISS
);
6285 /* The area was replaced with another one */
6289 * At this point, we have a stable reference to a VMA: The VMA is
6290 * locked and we know it hasn't already been isolated.
6291 * From here on, we can access the VMA without worrying about which
6292 * fields are accessible for RCU readers.
6295 /* Check since vm_start/vm_end might change before we lock the VMA */
6296 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
))
6297 goto inval_end_read
;
6306 count_vm_vma_lock_event(VMA_LOCK_ABORT
);
6309 #endif /* CONFIG_PER_VMA_LOCK */
6311 #ifndef __PAGETABLE_P4D_FOLDED
6313 * Allocate p4d page table.
6314 * We've already handled the fast-path in-line.
6316 int __p4d_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
6318 p4d_t
*new = p4d_alloc_one(mm
, address
);
6322 spin_lock(&mm
->page_table_lock
);
6323 if (pgd_present(*pgd
)) { /* Another has populated it */
6326 smp_wmb(); /* See comment in pmd_install() */
6327 pgd_populate(mm
, pgd
, new);
6329 spin_unlock(&mm
->page_table_lock
);
6332 #endif /* __PAGETABLE_P4D_FOLDED */
6334 #ifndef __PAGETABLE_PUD_FOLDED
6336 * Allocate page upper directory.
6337 * We've already handled the fast-path in-line.
6339 int __pud_alloc(struct mm_struct
*mm
, p4d_t
*p4d
, unsigned long address
)
6341 pud_t
*new = pud_alloc_one(mm
, address
);
6345 spin_lock(&mm
->page_table_lock
);
6346 if (!p4d_present(*p4d
)) {
6348 smp_wmb(); /* See comment in pmd_install() */
6349 p4d_populate(mm
, p4d
, new);
6350 } else /* Another has populated it */
6352 spin_unlock(&mm
->page_table_lock
);
6355 #endif /* __PAGETABLE_PUD_FOLDED */
6357 #ifndef __PAGETABLE_PMD_FOLDED
6359 * Allocate page middle directory.
6360 * We've already handled the fast-path in-line.
6362 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
6365 pmd_t
*new = pmd_alloc_one(mm
, address
);
6369 ptl
= pud_lock(mm
, pud
);
6370 if (!pud_present(*pud
)) {
6372 smp_wmb(); /* See comment in pmd_install() */
6373 pud_populate(mm
, pud
, new);
6374 } else { /* Another has populated it */
6380 #endif /* __PAGETABLE_PMD_FOLDED */
6382 static inline void pfnmap_args_setup(struct follow_pfnmap_args
*args
,
6383 spinlock_t
*lock
, pte_t
*ptep
,
6384 pgprot_t pgprot
, unsigned long pfn_base
,
6385 unsigned long addr_mask
, bool writable
,
6390 args
->pfn
= pfn_base
+ ((args
->address
& ~addr_mask
) >> PAGE_SHIFT
);
6391 args
->pgprot
= pgprot
;
6392 args
->writable
= writable
;
6393 args
->special
= special
;
6396 static inline void pfnmap_lockdep_assert(struct vm_area_struct
*vma
)
6398 #ifdef CONFIG_LOCKDEP
6399 struct file
*file
= vma
->vm_file
;
6400 struct address_space
*mapping
= file
? file
->f_mapping
: NULL
;
6403 lockdep_assert(lockdep_is_held(&mapping
->i_mmap_rwsem
) ||
6404 lockdep_is_held(&vma
->vm_mm
->mmap_lock
));
6406 lockdep_assert(lockdep_is_held(&vma
->vm_mm
->mmap_lock
));
6411 * follow_pfnmap_start() - Look up a pfn mapping at a user virtual address
6412 * @args: Pointer to struct @follow_pfnmap_args
6414 * The caller needs to setup args->vma and args->address to point to the
6415 * virtual address as the target of such lookup. On a successful return,
6416 * the results will be put into other output fields.
6418 * After the caller finished using the fields, the caller must invoke
6419 * another follow_pfnmap_end() to proper releases the locks and resources
6420 * of such look up request.
6422 * During the start() and end() calls, the results in @args will be valid
6423 * as proper locks will be held. After the end() is called, all the fields
6424 * in @follow_pfnmap_args will be invalid to be further accessed. Further
6425 * use of such information after end() may require proper synchronizations
6426 * by the caller with page table updates, otherwise it can create a
6429 * If the PTE maps a refcounted page, callers are responsible to protect
6430 * against invalidation with MMU notifiers; otherwise access to the PFN at
6431 * a later point in time can trigger use-after-free.
6433 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
6434 * should be taken for read, and the mmap semaphore cannot be released
6435 * before the end() is invoked.
6437 * This function must not be used to modify PTE content.
6439 * Return: zero on success, negative otherwise.
6441 int follow_pfnmap_start(struct follow_pfnmap_args
*args
)
6443 struct vm_area_struct
*vma
= args
->vma
;
6444 unsigned long address
= args
->address
;
6445 struct mm_struct
*mm
= vma
->vm_mm
;
6453 pfnmap_lockdep_assert(vma
);
6455 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
))
6458 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
6461 pgdp
= pgd_offset(mm
, address
);
6462 if (pgd_none(*pgdp
) || unlikely(pgd_bad(*pgdp
)))
6465 p4dp
= p4d_offset(pgdp
, address
);
6466 p4d
= READ_ONCE(*p4dp
);
6467 if (p4d_none(p4d
) || unlikely(p4d_bad(p4d
)))
6470 pudp
= pud_offset(p4dp
, address
);
6471 pud
= READ_ONCE(*pudp
);
6474 if (pud_leaf(pud
)) {
6475 lock
= pud_lock(mm
, pudp
);
6476 if (!unlikely(pud_leaf(pud
))) {
6480 pfnmap_args_setup(args
, lock
, NULL
, pud_pgprot(pud
),
6481 pud_pfn(pud
), PUD_MASK
, pud_write(pud
),
6486 pmdp
= pmd_offset(pudp
, address
);
6487 pmd
= pmdp_get_lockless(pmdp
);
6488 if (pmd_leaf(pmd
)) {
6489 lock
= pmd_lock(mm
, pmdp
);
6490 if (!unlikely(pmd_leaf(pmd
))) {
6494 pfnmap_args_setup(args
, lock
, NULL
, pmd_pgprot(pmd
),
6495 pmd_pfn(pmd
), PMD_MASK
, pmd_write(pmd
),
6500 ptep
= pte_offset_map_lock(mm
, pmdp
, address
, &lock
);
6503 pte
= ptep_get(ptep
);
6504 if (!pte_present(pte
))
6506 pfnmap_args_setup(args
, lock
, ptep
, pte_pgprot(pte
),
6507 pte_pfn(pte
), PAGE_MASK
, pte_write(pte
),
6511 pte_unmap_unlock(ptep
, lock
);
6515 EXPORT_SYMBOL_GPL(follow_pfnmap_start
);
6518 * follow_pfnmap_end(): End a follow_pfnmap_start() process
6519 * @args: Pointer to struct @follow_pfnmap_args
6521 * Must be used in pair of follow_pfnmap_start(). See the start() function
6522 * above for more information.
6524 void follow_pfnmap_end(struct follow_pfnmap_args
*args
)
6527 spin_unlock(args
->lock
);
6529 pte_unmap(args
->ptep
);
6531 EXPORT_SYMBOL_GPL(follow_pfnmap_end
);
6533 #ifdef CONFIG_HAVE_IOREMAP_PROT
6535 * generic_access_phys - generic implementation for iomem mmap access
6536 * @vma: the vma to access
6537 * @addr: userspace address, not relative offset within @vma
6538 * @buf: buffer to read/write
6539 * @len: length of transfer
6540 * @write: set to FOLL_WRITE when writing, otherwise reading
6542 * This is a generic implementation for &vm_operations_struct.access for an
6543 * iomem mapping. This callback is used by access_process_vm() when the @vma is
6546 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
6547 void *buf
, int len
, int write
)
6549 resource_size_t phys_addr
;
6550 unsigned long prot
= 0;
6551 void __iomem
*maddr
;
6552 int offset
= offset_in_page(addr
);
6555 struct follow_pfnmap_args args
= { .vma
= vma
, .address
= addr
};
6558 if (follow_pfnmap_start(&args
))
6560 prot
= pgprot_val(args
.pgprot
);
6561 phys_addr
= (resource_size_t
)args
.pfn
<< PAGE_SHIFT
;
6562 writable
= args
.writable
;
6563 follow_pfnmap_end(&args
);
6565 if ((write
& FOLL_WRITE
) && !writable
)
6568 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
6572 if (follow_pfnmap_start(&args
))
6575 if ((prot
!= pgprot_val(args
.pgprot
)) ||
6576 (phys_addr
!= (args
.pfn
<< PAGE_SHIFT
)) ||
6577 (writable
!= args
.writable
)) {
6578 follow_pfnmap_end(&args
);
6584 memcpy_toio(maddr
+ offset
, buf
, len
);
6586 memcpy_fromio(buf
, maddr
+ offset
, len
);
6588 follow_pfnmap_end(&args
);
6594 EXPORT_SYMBOL_GPL(generic_access_phys
);
6598 * Access another process' address space as given in mm.
6600 static int __access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
6601 void *buf
, int len
, unsigned int gup_flags
)
6603 void *old_buf
= buf
;
6604 int write
= gup_flags
& FOLL_WRITE
;
6606 if (mmap_read_lock_killable(mm
))
6609 /* Untag the address before looking up the VMA */
6610 addr
= untagged_addr_remote(mm
, addr
);
6612 /* Avoid triggering the temporary warning in __get_user_pages */
6613 if (!vma_lookup(mm
, addr
) && !expand_stack(mm
, addr
))
6616 /* ignore errors, just check how much was successfully transferred */
6620 struct vm_area_struct
*vma
= NULL
;
6621 struct page
*page
= get_user_page_vma_remote(mm
, addr
,
6625 /* We might need to expand the stack to access it */
6626 vma
= vma_lookup(mm
, addr
);
6628 vma
= expand_stack(mm
, addr
);
6630 /* mmap_lock was dropped on failure */
6632 return buf
- old_buf
;
6634 /* Try again if stack expansion worked */
6639 * Check if this is a VM_IO | VM_PFNMAP VMA, which
6640 * we can access using slightly different code.
6643 #ifdef CONFIG_HAVE_IOREMAP_PROT
6644 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
6645 bytes
= vma
->vm_ops
->access(vma
, addr
, buf
,
6652 offset
= addr
& (PAGE_SIZE
-1);
6653 if (bytes
> PAGE_SIZE
-offset
)
6654 bytes
= PAGE_SIZE
-offset
;
6656 maddr
= kmap_local_page(page
);
6658 copy_to_user_page(vma
, page
, addr
,
6659 maddr
+ offset
, buf
, bytes
);
6660 set_page_dirty_lock(page
);
6662 copy_from_user_page(vma
, page
, addr
,
6663 buf
, maddr
+ offset
, bytes
);
6665 unmap_and_put_page(page
, maddr
);
6671 mmap_read_unlock(mm
);
6673 return buf
- old_buf
;
6677 * access_remote_vm - access another process' address space
6678 * @mm: the mm_struct of the target address space
6679 * @addr: start address to access
6680 * @buf: source or destination buffer
6681 * @len: number of bytes to transfer
6682 * @gup_flags: flags modifying lookup behaviour
6684 * The caller must hold a reference on @mm.
6686 * Return: number of bytes copied from source to destination.
6688 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
6689 void *buf
, int len
, unsigned int gup_flags
)
6691 return __access_remote_vm(mm
, addr
, buf
, len
, gup_flags
);
6695 * Access another process' address space.
6696 * Source/target buffer must be kernel space,
6697 * Do not walk the page table directly, use get_user_pages
6699 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
6700 void *buf
, int len
, unsigned int gup_flags
)
6702 struct mm_struct
*mm
;
6705 mm
= get_task_mm(tsk
);
6709 ret
= __access_remote_vm(mm
, addr
, buf
, len
, gup_flags
);
6715 EXPORT_SYMBOL_GPL(access_process_vm
);
6718 * Print the name of a VMA.
6720 void print_vma_addr(char *prefix
, unsigned long ip
)
6722 struct mm_struct
*mm
= current
->mm
;
6723 struct vm_area_struct
*vma
;
6726 * we might be running from an atomic context so we cannot sleep
6728 if (!mmap_read_trylock(mm
))
6731 vma
= vma_lookup(mm
, ip
);
6732 if (vma
&& vma
->vm_file
) {
6733 struct file
*f
= vma
->vm_file
;
6734 ip
-= vma
->vm_start
;
6735 ip
+= vma
->vm_pgoff
<< PAGE_SHIFT
;
6736 printk("%s%pD[%lx,%lx+%lx]", prefix
, f
, ip
,
6738 vma
->vm_end
- vma
->vm_start
);
6740 mmap_read_unlock(mm
);
6743 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6744 void __might_fault(const char *file
, int line
)
6746 if (pagefault_disabled())
6748 __might_sleep(file
, line
);
6749 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6751 might_lock_read(¤t
->mm
->mmap_lock
);
6754 EXPORT_SYMBOL(__might_fault
);
6757 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
6759 * Process all subpages of the specified huge page with the specified
6760 * operation. The target subpage will be processed last to keep its
6763 static inline int process_huge_page(
6764 unsigned long addr_hint
, unsigned int nr_pages
,
6765 int (*process_subpage
)(unsigned long addr
, int idx
, void *arg
),
6768 int i
, n
, base
, l
, ret
;
6769 unsigned long addr
= addr_hint
&
6770 ~(((unsigned long)nr_pages
<< PAGE_SHIFT
) - 1);
6772 /* Process target subpage last to keep its cache lines hot */
6774 n
= (addr_hint
- addr
) / PAGE_SIZE
;
6775 if (2 * n
<= nr_pages
) {
6776 /* If target subpage in first half of huge page */
6779 /* Process subpages at the end of huge page */
6780 for (i
= nr_pages
- 1; i
>= 2 * n
; i
--) {
6782 ret
= process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
6787 /* If target subpage in second half of huge page */
6788 base
= nr_pages
- 2 * (nr_pages
- n
);
6790 /* Process subpages at the begin of huge page */
6791 for (i
= 0; i
< base
; i
++) {
6793 ret
= process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
6799 * Process remaining subpages in left-right-left-right pattern
6800 * towards the target subpage
6802 for (i
= 0; i
< l
; i
++) {
6803 int left_idx
= base
+ i
;
6804 int right_idx
= base
+ 2 * l
- 1 - i
;
6807 ret
= process_subpage(addr
+ left_idx
* PAGE_SIZE
, left_idx
, arg
);
6811 ret
= process_subpage(addr
+ right_idx
* PAGE_SIZE
, right_idx
, arg
);
6818 static void clear_gigantic_page(struct folio
*folio
, unsigned long addr
,
6819 unsigned int nr_pages
)
6824 for (i
= 0; i
< nr_pages
; i
++) {
6826 clear_user_highpage(folio_page(folio
, i
), addr
+ i
* PAGE_SIZE
);
6830 static int clear_subpage(unsigned long addr
, int idx
, void *arg
)
6832 struct folio
*folio
= arg
;
6834 clear_user_highpage(folio_page(folio
, idx
), addr
);
6839 * folio_zero_user - Zero a folio which will be mapped to userspace.
6840 * @folio: The folio to zero.
6841 * @addr_hint: The address will be accessed or the base address if uncelar.
6843 void folio_zero_user(struct folio
*folio
, unsigned long addr_hint
)
6845 unsigned int nr_pages
= folio_nr_pages(folio
);
6847 if (unlikely(nr_pages
> MAX_ORDER_NR_PAGES
))
6848 clear_gigantic_page(folio
, addr_hint
, nr_pages
);
6850 process_huge_page(addr_hint
, nr_pages
, clear_subpage
, folio
);
6853 static int copy_user_gigantic_page(struct folio
*dst
, struct folio
*src
,
6855 struct vm_area_struct
*vma
,
6856 unsigned int nr_pages
)
6859 struct page
*dst_page
;
6860 struct page
*src_page
;
6862 for (i
= 0; i
< nr_pages
; i
++) {
6863 dst_page
= folio_page(dst
, i
);
6864 src_page
= folio_page(src
, i
);
6867 if (copy_mc_user_highpage(dst_page
, src_page
,
6868 addr
+ i
*PAGE_SIZE
, vma
))
6874 struct copy_subpage_arg
{
6877 struct vm_area_struct
*vma
;
6880 static int copy_subpage(unsigned long addr
, int idx
, void *arg
)
6882 struct copy_subpage_arg
*copy_arg
= arg
;
6883 struct page
*dst
= folio_page(copy_arg
->dst
, idx
);
6884 struct page
*src
= folio_page(copy_arg
->src
, idx
);
6886 if (copy_mc_user_highpage(dst
, src
, addr
, copy_arg
->vma
))
6891 int copy_user_large_folio(struct folio
*dst
, struct folio
*src
,
6892 unsigned long addr_hint
, struct vm_area_struct
*vma
)
6894 unsigned int nr_pages
= folio_nr_pages(dst
);
6895 struct copy_subpage_arg arg
= {
6901 if (unlikely(nr_pages
> MAX_ORDER_NR_PAGES
))
6902 return copy_user_gigantic_page(dst
, src
, addr_hint
, vma
, nr_pages
);
6904 return process_huge_page(addr_hint
, nr_pages
, copy_subpage
, &arg
);
6907 long copy_folio_from_user(struct folio
*dst_folio
,
6908 const void __user
*usr_src
,
6909 bool allow_pagefault
)
6912 unsigned long i
, rc
= 0;
6913 unsigned int nr_pages
= folio_nr_pages(dst_folio
);
6914 unsigned long ret_val
= nr_pages
* PAGE_SIZE
;
6915 struct page
*subpage
;
6917 for (i
= 0; i
< nr_pages
; i
++) {
6918 subpage
= folio_page(dst_folio
, i
);
6919 kaddr
= kmap_local_page(subpage
);
6920 if (!allow_pagefault
)
6921 pagefault_disable();
6922 rc
= copy_from_user(kaddr
, usr_src
+ i
* PAGE_SIZE
, PAGE_SIZE
);
6923 if (!allow_pagefault
)
6925 kunmap_local(kaddr
);
6927 ret_val
-= (PAGE_SIZE
- rc
);
6931 flush_dcache_page(subpage
);
6937 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
6939 #if defined(CONFIG_SPLIT_PTE_PTLOCKS) && ALLOC_SPLIT_PTLOCKS
6941 static struct kmem_cache
*page_ptl_cachep
;
6943 void __init
ptlock_cache_init(void)
6945 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
6949 bool ptlock_alloc(struct ptdesc
*ptdesc
)
6953 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
6960 void ptlock_free(struct ptdesc
*ptdesc
)
6962 kmem_cache_free(page_ptl_cachep
, ptdesc
->ptl
);
6966 void vma_pgtable_walk_begin(struct vm_area_struct
*vma
)
6968 if (is_vm_hugetlb_page(vma
))
6969 hugetlb_vma_lock_read(vma
);
6972 void vma_pgtable_walk_end(struct vm_area_struct
*vma
)
6974 if (is_vm_hugetlb_page(vma
))
6975 hugetlb_vma_unlock_read(vma
);