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/coredump.h>
47 #include <linux/sched/numa_balancing.h>
48 #include <linux/sched/task.h>
49 #include <linux/hugetlb.h>
50 #include <linux/mman.h>
51 #include <linux/swap.h>
52 #include <linux/highmem.h>
53 #include <linux/pagemap.h>
54 #include <linux/memremap.h>
55 #include <linux/kmsan.h>
56 #include <linux/ksm.h>
57 #include <linux/rmap.h>
58 #include <linux/export.h>
59 #include <linux/delayacct.h>
60 #include <linux/init.h>
61 #include <linux/pfn_t.h>
62 #include <linux/writeback.h>
63 #include <linux/memcontrol.h>
64 #include <linux/mmu_notifier.h>
65 #include <linux/swapops.h>
66 #include <linux/elf.h>
67 #include <linux/gfp.h>
68 #include <linux/migrate.h>
69 #include <linux/string.h>
70 #include <linux/memory-tiers.h>
71 #include <linux/debugfs.h>
72 #include <linux/userfaultfd_k.h>
73 #include <linux/dax.h>
74 #include <linux/oom.h>
75 #include <linux/numa.h>
76 #include <linux/perf_event.h>
77 #include <linux/ptrace.h>
78 #include <linux/vmalloc.h>
79 #include <linux/sched/sysctl.h>
81 #include <trace/events/kmem.h>
84 #include <asm/mmu_context.h>
85 #include <asm/pgalloc.h>
86 #include <linux/uaccess.h>
88 #include <asm/tlbflush.h>
90 #include "pgalloc-track.h"
94 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
95 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
99 unsigned long max_mapnr
;
100 EXPORT_SYMBOL(max_mapnr
);
102 struct page
*mem_map
;
103 EXPORT_SYMBOL(mem_map
);
106 static vm_fault_t
do_fault(struct vm_fault
*vmf
);
107 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
);
108 static bool vmf_pte_changed(struct vm_fault
*vmf
);
111 * Return true if the original pte was a uffd-wp pte marker (so the pte was
114 static bool vmf_orig_pte_uffd_wp(struct vm_fault
*vmf
)
116 if (!(vmf
->flags
& FAULT_FLAG_ORIG_PTE_VALID
))
119 return pte_marker_uffd_wp(vmf
->orig_pte
);
123 * A number of key systems in x86 including ioremap() rely on the assumption
124 * that high_memory defines the upper bound on direct map memory, then end
125 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
126 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
130 EXPORT_SYMBOL(high_memory
);
133 * Randomize the address space (stacks, mmaps, brk, etc.).
135 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
136 * as ancient (libc5 based) binaries can segfault. )
138 int randomize_va_space __read_mostly
=
139 #ifdef CONFIG_COMPAT_BRK
145 #ifndef arch_wants_old_prefaulted_pte
146 static inline bool arch_wants_old_prefaulted_pte(void)
149 * Transitioning a PTE from 'old' to 'young' can be expensive on
150 * some architectures, even if it's performed in hardware. By
151 * default, "false" means prefaulted entries will be 'young'.
157 static int __init
disable_randmaps(char *s
)
159 randomize_va_space
= 0;
162 __setup("norandmaps", disable_randmaps
);
164 unsigned long zero_pfn __read_mostly
;
165 EXPORT_SYMBOL(zero_pfn
);
167 unsigned long highest_memmap_pfn __read_mostly
;
170 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
172 static int __init
init_zero_pfn(void)
174 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
177 early_initcall(init_zero_pfn
);
179 void mm_trace_rss_stat(struct mm_struct
*mm
, int member
)
181 trace_rss_stat(mm
, member
);
185 * Note: this doesn't free the actual pages themselves. That
186 * has been handled earlier when unmapping all the memory regions.
188 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
191 pgtable_t token
= pmd_pgtable(*pmd
);
193 pte_free_tlb(tlb
, token
, addr
);
194 mm_dec_nr_ptes(tlb
->mm
);
197 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
198 unsigned long addr
, unsigned long end
,
199 unsigned long floor
, unsigned long ceiling
)
206 pmd
= pmd_offset(pud
, addr
);
208 next
= pmd_addr_end(addr
, end
);
209 if (pmd_none_or_clear_bad(pmd
))
211 free_pte_range(tlb
, pmd
, addr
);
212 } while (pmd
++, addr
= next
, addr
!= end
);
222 if (end
- 1 > ceiling
- 1)
225 pmd
= pmd_offset(pud
, start
);
227 pmd_free_tlb(tlb
, pmd
, start
);
228 mm_dec_nr_pmds(tlb
->mm
);
231 static inline void free_pud_range(struct mmu_gather
*tlb
, p4d_t
*p4d
,
232 unsigned long addr
, unsigned long end
,
233 unsigned long floor
, unsigned long ceiling
)
240 pud
= pud_offset(p4d
, addr
);
242 next
= pud_addr_end(addr
, end
);
243 if (pud_none_or_clear_bad(pud
))
245 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
246 } while (pud
++, addr
= next
, addr
!= end
);
256 if (end
- 1 > ceiling
- 1)
259 pud
= pud_offset(p4d
, start
);
261 pud_free_tlb(tlb
, pud
, start
);
262 mm_dec_nr_puds(tlb
->mm
);
265 static inline void free_p4d_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
266 unsigned long addr
, unsigned long end
,
267 unsigned long floor
, unsigned long ceiling
)
274 p4d
= p4d_offset(pgd
, addr
);
276 next
= p4d_addr_end(addr
, end
);
277 if (p4d_none_or_clear_bad(p4d
))
279 free_pud_range(tlb
, p4d
, addr
, next
, floor
, ceiling
);
280 } while (p4d
++, addr
= next
, addr
!= end
);
286 ceiling
&= PGDIR_MASK
;
290 if (end
- 1 > ceiling
- 1)
293 p4d
= p4d_offset(pgd
, start
);
295 p4d_free_tlb(tlb
, p4d
, start
);
299 * This function frees user-level page tables of a process.
301 void free_pgd_range(struct mmu_gather
*tlb
,
302 unsigned long addr
, unsigned long end
,
303 unsigned long floor
, unsigned long ceiling
)
309 * The next few lines have given us lots of grief...
311 * Why are we testing PMD* at this top level? Because often
312 * there will be no work to do at all, and we'd prefer not to
313 * go all the way down to the bottom just to discover that.
315 * Why all these "- 1"s? Because 0 represents both the bottom
316 * of the address space and the top of it (using -1 for the
317 * top wouldn't help much: the masks would do the wrong thing).
318 * The rule is that addr 0 and floor 0 refer to the bottom of
319 * the address space, but end 0 and ceiling 0 refer to the top
320 * Comparisons need to use "end - 1" and "ceiling - 1" (though
321 * that end 0 case should be mythical).
323 * Wherever addr is brought up or ceiling brought down, we must
324 * be careful to reject "the opposite 0" before it confuses the
325 * subsequent tests. But what about where end is brought down
326 * by PMD_SIZE below? no, end can't go down to 0 there.
328 * Whereas we round start (addr) and ceiling down, by different
329 * masks at different levels, in order to test whether a table
330 * now has no other vmas using it, so can be freed, we don't
331 * bother to round floor or end up - the tests don't need that.
345 if (end
- 1 > ceiling
- 1)
350 * We add page table cache pages with PAGE_SIZE,
351 * (see pte_free_tlb()), flush the tlb if we need
353 tlb_change_page_size(tlb
, PAGE_SIZE
);
354 pgd
= pgd_offset(tlb
->mm
, addr
);
356 next
= pgd_addr_end(addr
, end
);
357 if (pgd_none_or_clear_bad(pgd
))
359 free_p4d_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
360 } while (pgd
++, addr
= next
, addr
!= end
);
363 void free_pgtables(struct mmu_gather
*tlb
, struct ma_state
*mas
,
364 struct vm_area_struct
*vma
, unsigned long floor
,
365 unsigned long ceiling
, bool mm_wr_locked
)
368 unsigned long addr
= vma
->vm_start
;
369 struct vm_area_struct
*next
;
372 * Note: USER_PGTABLES_CEILING may be passed as ceiling and may
373 * be 0. This will underflow and is okay.
375 next
= mas_find(mas
, ceiling
- 1);
378 * Hide vma from rmap and truncate_pagecache before freeing
382 vma_start_write(vma
);
383 unlink_anon_vmas(vma
);
384 unlink_file_vma(vma
);
386 if (is_vm_hugetlb_page(vma
)) {
387 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
388 floor
, next
? next
->vm_start
: ceiling
);
391 * Optimization: gather nearby vmas into one call down
393 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
394 && !is_vm_hugetlb_page(next
)) {
396 next
= mas_find(mas
, ceiling
- 1);
398 vma_start_write(vma
);
399 unlink_anon_vmas(vma
);
400 unlink_file_vma(vma
);
402 free_pgd_range(tlb
, addr
, vma
->vm_end
,
403 floor
, next
? next
->vm_start
: ceiling
);
409 void pmd_install(struct mm_struct
*mm
, pmd_t
*pmd
, pgtable_t
*pte
)
411 spinlock_t
*ptl
= pmd_lock(mm
, pmd
);
413 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
416 * Ensure all pte setup (eg. pte page lock and page clearing) are
417 * visible before the pte is made visible to other CPUs by being
418 * put into page tables.
420 * The other side of the story is the pointer chasing in the page
421 * table walking code (when walking the page table without locking;
422 * ie. most of the time). Fortunately, these data accesses consist
423 * of a chain of data-dependent loads, meaning most CPUs (alpha
424 * being the notable exception) will already guarantee loads are
425 * seen in-order. See the alpha page table accessors for the
426 * smp_rmb() barriers in page table walking code.
428 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
429 pmd_populate(mm
, pmd
, *pte
);
435 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
)
437 pgtable_t
new = pte_alloc_one(mm
);
441 pmd_install(mm
, pmd
, &new);
447 int __pte_alloc_kernel(pmd_t
*pmd
)
449 pte_t
*new = pte_alloc_one_kernel(&init_mm
);
453 spin_lock(&init_mm
.page_table_lock
);
454 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
455 smp_wmb(); /* See comment in pmd_install() */
456 pmd_populate_kernel(&init_mm
, pmd
, new);
459 spin_unlock(&init_mm
.page_table_lock
);
461 pte_free_kernel(&init_mm
, new);
465 static inline void init_rss_vec(int *rss
)
467 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
470 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
474 if (current
->mm
== mm
)
476 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
478 add_mm_counter(mm
, i
, rss
[i
]);
482 * This function is called to print an error when a bad pte
483 * is found. For example, we might have a PFN-mapped pte in
484 * a region that doesn't allow it.
486 * The calling function must still handle the error.
488 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
489 pte_t pte
, struct page
*page
)
491 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
492 p4d_t
*p4d
= p4d_offset(pgd
, addr
);
493 pud_t
*pud
= pud_offset(p4d
, addr
);
494 pmd_t
*pmd
= pmd_offset(pud
, addr
);
495 struct address_space
*mapping
;
497 static unsigned long resume
;
498 static unsigned long nr_shown
;
499 static unsigned long nr_unshown
;
502 * Allow a burst of 60 reports, then keep quiet for that minute;
503 * or allow a steady drip of one report per second.
505 if (nr_shown
== 60) {
506 if (time_before(jiffies
, resume
)) {
511 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
518 resume
= jiffies
+ 60 * HZ
;
520 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
521 index
= linear_page_index(vma
, addr
);
523 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
525 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
527 dump_page(page
, "bad pte");
528 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
529 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
530 pr_alert("file:%pD fault:%ps mmap:%ps read_folio:%ps\n",
532 vma
->vm_ops
? vma
->vm_ops
->fault
: NULL
,
533 vma
->vm_file
? vma
->vm_file
->f_op
->mmap
: NULL
,
534 mapping
? mapping
->a_ops
->read_folio
: NULL
);
536 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
540 * vm_normal_page -- This function gets the "struct page" associated with a pte.
542 * "Special" mappings do not wish to be associated with a "struct page" (either
543 * it doesn't exist, or it exists but they don't want to touch it). In this
544 * case, NULL is returned here. "Normal" mappings do have a struct page.
546 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
547 * pte bit, in which case this function is trivial. Secondly, an architecture
548 * may not have a spare pte bit, which requires a more complicated scheme,
551 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
552 * special mapping (even if there are underlying and valid "struct pages").
553 * COWed pages of a VM_PFNMAP are always normal.
555 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
556 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
557 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
558 * mapping will always honor the rule
560 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
562 * And for normal mappings this is false.
564 * This restricts such mappings to be a linear translation from virtual address
565 * to pfn. To get around this restriction, we allow arbitrary mappings so long
566 * as the vma is not a COW mapping; in that case, we know that all ptes are
567 * special (because none can have been COWed).
570 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
572 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
573 * page" backing, however the difference is that _all_ pages with a struct
574 * page (that is, those where pfn_valid is true) are refcounted and considered
575 * normal pages by the VM. The disadvantage is that pages are refcounted
576 * (which can be slower and simply not an option for some PFNMAP users). The
577 * advantage is that we don't have to follow the strict linearity rule of
578 * PFNMAP mappings in order to support COWable mappings.
581 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
584 unsigned long pfn
= pte_pfn(pte
);
586 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
)) {
587 if (likely(!pte_special(pte
)))
589 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
590 return vma
->vm_ops
->find_special_page(vma
, addr
);
591 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
593 if (is_zero_pfn(pfn
))
597 * NOTE: New users of ZONE_DEVICE will not set pte_devmap()
598 * and will have refcounts incremented on their struct pages
599 * when they are inserted into PTEs, thus they are safe to
600 * return here. Legacy ZONE_DEVICE pages that set pte_devmap()
601 * do not have refcounts. Example of legacy ZONE_DEVICE is
602 * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers.
606 print_bad_pte(vma
, addr
, pte
, NULL
);
610 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
612 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
613 if (vma
->vm_flags
& VM_MIXEDMAP
) {
619 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
620 if (pfn
== vma
->vm_pgoff
+ off
)
622 if (!is_cow_mapping(vma
->vm_flags
))
627 if (is_zero_pfn(pfn
))
631 if (unlikely(pfn
> highest_memmap_pfn
)) {
632 print_bad_pte(vma
, addr
, pte
, NULL
);
637 * NOTE! We still have PageReserved() pages in the page tables.
638 * eg. VDSO mappings can cause them to exist.
641 return pfn_to_page(pfn
);
644 struct folio
*vm_normal_folio(struct vm_area_struct
*vma
, unsigned long addr
,
647 struct page
*page
= vm_normal_page(vma
, addr
, pte
);
650 return page_folio(page
);
654 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
655 struct page
*vm_normal_page_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
658 unsigned long pfn
= pmd_pfn(pmd
);
661 * There is no pmd_special() but there may be special pmds, e.g.
662 * in a direct-access (dax) mapping, so let's just replicate the
663 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
665 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
666 if (vma
->vm_flags
& VM_MIXEDMAP
) {
672 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
673 if (pfn
== vma
->vm_pgoff
+ off
)
675 if (!is_cow_mapping(vma
->vm_flags
))
682 if (is_huge_zero_pmd(pmd
))
684 if (unlikely(pfn
> highest_memmap_pfn
))
688 * NOTE! We still have PageReserved() pages in the page tables.
689 * eg. VDSO mappings can cause them to exist.
692 return pfn_to_page(pfn
);
696 static void restore_exclusive_pte(struct vm_area_struct
*vma
,
697 struct page
*page
, unsigned long address
,
704 orig_pte
= ptep_get(ptep
);
705 pte
= pte_mkold(mk_pte(page
, READ_ONCE(vma
->vm_page_prot
)));
706 if (pte_swp_soft_dirty(orig_pte
))
707 pte
= pte_mksoft_dirty(pte
);
709 entry
= pte_to_swp_entry(orig_pte
);
710 if (pte_swp_uffd_wp(orig_pte
))
711 pte
= pte_mkuffd_wp(pte
);
712 else if (is_writable_device_exclusive_entry(entry
))
713 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
715 VM_BUG_ON(pte_write(pte
) && !(PageAnon(page
) && PageAnonExclusive(page
)));
718 * No need to take a page reference as one was already
719 * created when the swap entry was made.
722 page_add_anon_rmap(page
, vma
, address
, RMAP_NONE
);
725 * Currently device exclusive access only supports anonymous
726 * memory so the entry shouldn't point to a filebacked page.
730 set_pte_at(vma
->vm_mm
, address
, ptep
, pte
);
733 * No need to invalidate - it was non-present before. However
734 * secondary CPUs may have mappings that need invalidating.
736 update_mmu_cache(vma
, address
, ptep
);
740 * Tries to restore an exclusive pte if the page lock can be acquired without
744 try_restore_exclusive_pte(pte_t
*src_pte
, struct vm_area_struct
*vma
,
747 swp_entry_t entry
= pte_to_swp_entry(ptep_get(src_pte
));
748 struct page
*page
= pfn_swap_entry_to_page(entry
);
750 if (trylock_page(page
)) {
751 restore_exclusive_pte(vma
, page
, addr
, src_pte
);
760 * copy one vm_area from one task to the other. Assumes the page tables
761 * already present in the new task to be cleared in the whole range
762 * covered by this vma.
766 copy_nonpresent_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
767 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*dst_vma
,
768 struct vm_area_struct
*src_vma
, unsigned long addr
, int *rss
)
770 unsigned long vm_flags
= dst_vma
->vm_flags
;
771 pte_t orig_pte
= ptep_get(src_pte
);
772 pte_t pte
= orig_pte
;
774 swp_entry_t entry
= pte_to_swp_entry(orig_pte
);
776 if (likely(!non_swap_entry(entry
))) {
777 if (swap_duplicate(entry
) < 0)
780 /* make sure dst_mm is on swapoff's mmlist. */
781 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
782 spin_lock(&mmlist_lock
);
783 if (list_empty(&dst_mm
->mmlist
))
784 list_add(&dst_mm
->mmlist
,
786 spin_unlock(&mmlist_lock
);
788 /* Mark the swap entry as shared. */
789 if (pte_swp_exclusive(orig_pte
)) {
790 pte
= pte_swp_clear_exclusive(orig_pte
);
791 set_pte_at(src_mm
, addr
, src_pte
, pte
);
794 } else if (is_migration_entry(entry
)) {
795 page
= pfn_swap_entry_to_page(entry
);
797 rss
[mm_counter(page
)]++;
799 if (!is_readable_migration_entry(entry
) &&
800 is_cow_mapping(vm_flags
)) {
802 * COW mappings require pages in both parent and child
803 * to be set to read. A previously exclusive entry is
806 entry
= make_readable_migration_entry(
808 pte
= swp_entry_to_pte(entry
);
809 if (pte_swp_soft_dirty(orig_pte
))
810 pte
= pte_swp_mksoft_dirty(pte
);
811 if (pte_swp_uffd_wp(orig_pte
))
812 pte
= pte_swp_mkuffd_wp(pte
);
813 set_pte_at(src_mm
, addr
, src_pte
, pte
);
815 } else if (is_device_private_entry(entry
)) {
816 page
= pfn_swap_entry_to_page(entry
);
819 * Update rss count even for unaddressable pages, as
820 * they should treated just like normal pages in this
823 * We will likely want to have some new rss counters
824 * for unaddressable pages, at some point. But for now
825 * keep things as they are.
828 rss
[mm_counter(page
)]++;
829 /* Cannot fail as these pages cannot get pinned. */
830 BUG_ON(page_try_dup_anon_rmap(page
, false, src_vma
));
833 * We do not preserve soft-dirty information, because so
834 * far, checkpoint/restore is the only feature that
835 * requires that. And checkpoint/restore does not work
836 * when a device driver is involved (you cannot easily
837 * save and restore device driver state).
839 if (is_writable_device_private_entry(entry
) &&
840 is_cow_mapping(vm_flags
)) {
841 entry
= make_readable_device_private_entry(
843 pte
= swp_entry_to_pte(entry
);
844 if (pte_swp_uffd_wp(orig_pte
))
845 pte
= pte_swp_mkuffd_wp(pte
);
846 set_pte_at(src_mm
, addr
, src_pte
, pte
);
848 } else if (is_device_exclusive_entry(entry
)) {
850 * Make device exclusive entries present by restoring the
851 * original entry then copying as for a present pte. Device
852 * exclusive entries currently only support private writable
853 * (ie. COW) mappings.
855 VM_BUG_ON(!is_cow_mapping(src_vma
->vm_flags
));
856 if (try_restore_exclusive_pte(src_pte
, src_vma
, addr
))
859 } else if (is_pte_marker_entry(entry
)) {
860 pte_marker marker
= copy_pte_marker(entry
, dst_vma
);
863 set_pte_at(dst_mm
, addr
, dst_pte
,
864 make_pte_marker(marker
));
867 if (!userfaultfd_wp(dst_vma
))
868 pte
= pte_swp_clear_uffd_wp(pte
);
869 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
874 * Copy a present and normal page.
876 * NOTE! The usual case is that this isn't required;
877 * instead, the caller can just increase the page refcount
878 * and re-use the pte the traditional way.
880 * And if we need a pre-allocated page but don't yet have
881 * one, return a negative error to let the preallocation
882 * code know so that it can do so outside the page table
886 copy_present_page(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
887 pte_t
*dst_pte
, pte_t
*src_pte
, unsigned long addr
, int *rss
,
888 struct folio
**prealloc
, struct page
*page
)
890 struct folio
*new_folio
;
893 new_folio
= *prealloc
;
898 * We have a prealloc page, all good! Take it
899 * over and copy the page & arm it.
902 copy_user_highpage(&new_folio
->page
, page
, addr
, src_vma
);
903 __folio_mark_uptodate(new_folio
);
904 folio_add_new_anon_rmap(new_folio
, dst_vma
, addr
);
905 folio_add_lru_vma(new_folio
, dst_vma
);
908 /* All done, just insert the new page copy in the child */
909 pte
= mk_pte(&new_folio
->page
, dst_vma
->vm_page_prot
);
910 pte
= maybe_mkwrite(pte_mkdirty(pte
), dst_vma
);
911 if (userfaultfd_pte_wp(dst_vma
, ptep_get(src_pte
)))
912 /* Uffd-wp needs to be delivered to dest pte as well */
913 pte
= pte_mkuffd_wp(pte
);
914 set_pte_at(dst_vma
->vm_mm
, addr
, dst_pte
, pte
);
919 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
920 * is required to copy this pte.
923 copy_present_pte(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
924 pte_t
*dst_pte
, pte_t
*src_pte
, unsigned long addr
, int *rss
,
925 struct folio
**prealloc
)
927 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
928 unsigned long vm_flags
= src_vma
->vm_flags
;
929 pte_t pte
= ptep_get(src_pte
);
933 page
= vm_normal_page(src_vma
, addr
, pte
);
935 folio
= page_folio(page
);
936 if (page
&& folio_test_anon(folio
)) {
938 * If this page may have been pinned by the parent process,
939 * copy the page immediately for the child so that we'll always
940 * guarantee the pinned page won't be randomly replaced in the
944 if (unlikely(page_try_dup_anon_rmap(page
, false, src_vma
))) {
945 /* Page may be pinned, we have to copy. */
947 return copy_present_page(dst_vma
, src_vma
, dst_pte
, src_pte
,
948 addr
, rss
, prealloc
, page
);
953 page_dup_file_rmap(page
, false);
954 rss
[mm_counter_file(page
)]++;
958 * If it's a COW mapping, write protect it both
959 * in the parent and the child
961 if (is_cow_mapping(vm_flags
) && pte_write(pte
)) {
962 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
963 pte
= pte_wrprotect(pte
);
965 VM_BUG_ON(page
&& folio_test_anon(folio
) && PageAnonExclusive(page
));
968 * If it's a shared mapping, mark it clean in
971 if (vm_flags
& VM_SHARED
)
972 pte
= pte_mkclean(pte
);
973 pte
= pte_mkold(pte
);
975 if (!userfaultfd_wp(dst_vma
))
976 pte
= pte_clear_uffd_wp(pte
);
978 set_pte_at(dst_vma
->vm_mm
, addr
, dst_pte
, pte
);
982 static inline struct folio
*page_copy_prealloc(struct mm_struct
*src_mm
,
983 struct vm_area_struct
*vma
, unsigned long addr
)
985 struct folio
*new_folio
;
987 new_folio
= vma_alloc_folio(GFP_HIGHUSER_MOVABLE
, 0, vma
, addr
, false);
991 if (mem_cgroup_charge(new_folio
, src_mm
, GFP_KERNEL
)) {
992 folio_put(new_folio
);
995 folio_throttle_swaprate(new_folio
, GFP_KERNEL
);
1001 copy_pte_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1002 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
1005 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1006 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1007 pte_t
*orig_src_pte
, *orig_dst_pte
;
1008 pte_t
*src_pte
, *dst_pte
;
1010 spinlock_t
*src_ptl
, *dst_ptl
;
1011 int progress
, ret
= 0;
1012 int rss
[NR_MM_COUNTERS
];
1013 swp_entry_t entry
= (swp_entry_t
){0};
1014 struct folio
*prealloc
= NULL
;
1021 * copy_pmd_range()'s prior pmd_none_or_clear_bad(src_pmd), and the
1022 * error handling here, assume that exclusive mmap_lock on dst and src
1023 * protects anon from unexpected THP transitions; with shmem and file
1024 * protected by mmap_lock-less collapse skipping areas with anon_vma
1025 * (whereas vma_needs_copy() skips areas without anon_vma). A rework
1026 * can remove such assumptions later, but this is good enough for now.
1028 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
1033 src_pte
= pte_offset_map_nolock(src_mm
, src_pmd
, addr
, &src_ptl
);
1035 pte_unmap_unlock(dst_pte
, dst_ptl
);
1039 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1040 orig_src_pte
= src_pte
;
1041 orig_dst_pte
= dst_pte
;
1042 arch_enter_lazy_mmu_mode();
1046 * We are holding two locks at this point - either of them
1047 * could generate latencies in another task on another CPU.
1049 if (progress
>= 32) {
1051 if (need_resched() ||
1052 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
1055 ptent
= ptep_get(src_pte
);
1056 if (pte_none(ptent
)) {
1060 if (unlikely(!pte_present(ptent
))) {
1061 ret
= copy_nonpresent_pte(dst_mm
, src_mm
,
1066 entry
= pte_to_swp_entry(ptep_get(src_pte
));
1068 } else if (ret
== -EBUSY
) {
1076 * Device exclusive entry restored, continue by copying
1077 * the now present pte.
1079 WARN_ON_ONCE(ret
!= -ENOENT
);
1081 /* copy_present_pte() will clear `*prealloc' if consumed */
1082 ret
= copy_present_pte(dst_vma
, src_vma
, dst_pte
, src_pte
,
1083 addr
, rss
, &prealloc
);
1085 * If we need a pre-allocated page for this pte, drop the
1086 * locks, allocate, and try again.
1088 if (unlikely(ret
== -EAGAIN
))
1090 if (unlikely(prealloc
)) {
1092 * pre-alloc page cannot be reused by next time so as
1093 * to strictly follow mempolicy (e.g., alloc_page_vma()
1094 * will allocate page according to address). This
1095 * could only happen if one pinned pte changed.
1097 folio_put(prealloc
);
1101 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1103 arch_leave_lazy_mmu_mode();
1104 pte_unmap_unlock(orig_src_pte
, src_ptl
);
1105 add_mm_rss_vec(dst_mm
, rss
);
1106 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
1110 VM_WARN_ON_ONCE(!entry
.val
);
1111 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0) {
1116 } else if (ret
== -EBUSY
) {
1118 } else if (ret
== -EAGAIN
) {
1119 prealloc
= page_copy_prealloc(src_mm
, src_vma
, addr
);
1126 /* We've captured and resolved the error. Reset, try again. */
1132 if (unlikely(prealloc
))
1133 folio_put(prealloc
);
1138 copy_pmd_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1139 pud_t
*dst_pud
, pud_t
*src_pud
, unsigned long addr
,
1142 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1143 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1144 pmd_t
*src_pmd
, *dst_pmd
;
1147 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
1150 src_pmd
= pmd_offset(src_pud
, addr
);
1152 next
= pmd_addr_end(addr
, end
);
1153 if (is_swap_pmd(*src_pmd
) || pmd_trans_huge(*src_pmd
)
1154 || pmd_devmap(*src_pmd
)) {
1156 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PMD_SIZE
, src_vma
);
1157 err
= copy_huge_pmd(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
1158 addr
, dst_vma
, src_vma
);
1165 if (pmd_none_or_clear_bad(src_pmd
))
1167 if (copy_pte_range(dst_vma
, src_vma
, dst_pmd
, src_pmd
,
1170 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1175 copy_pud_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1176 p4d_t
*dst_p4d
, p4d_t
*src_p4d
, unsigned long addr
,
1179 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1180 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1181 pud_t
*src_pud
, *dst_pud
;
1184 dst_pud
= pud_alloc(dst_mm
, dst_p4d
, addr
);
1187 src_pud
= pud_offset(src_p4d
, addr
);
1189 next
= pud_addr_end(addr
, end
);
1190 if (pud_trans_huge(*src_pud
) || pud_devmap(*src_pud
)) {
1193 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PUD_SIZE
, src_vma
);
1194 err
= copy_huge_pud(dst_mm
, src_mm
,
1195 dst_pud
, src_pud
, addr
, src_vma
);
1202 if (pud_none_or_clear_bad(src_pud
))
1204 if (copy_pmd_range(dst_vma
, src_vma
, dst_pud
, src_pud
,
1207 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1212 copy_p4d_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1213 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, unsigned long addr
,
1216 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1217 p4d_t
*src_p4d
, *dst_p4d
;
1220 dst_p4d
= p4d_alloc(dst_mm
, dst_pgd
, addr
);
1223 src_p4d
= p4d_offset(src_pgd
, addr
);
1225 next
= p4d_addr_end(addr
, end
);
1226 if (p4d_none_or_clear_bad(src_p4d
))
1228 if (copy_pud_range(dst_vma
, src_vma
, dst_p4d
, src_p4d
,
1231 } while (dst_p4d
++, src_p4d
++, addr
= next
, addr
!= end
);
1236 * Return true if the vma needs to copy the pgtable during this fork(). Return
1237 * false when we can speed up fork() by allowing lazy page faults later until
1238 * when the child accesses the memory range.
1241 vma_needs_copy(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
)
1244 * Always copy pgtables when dst_vma has uffd-wp enabled even if it's
1245 * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable
1246 * contains uffd-wp protection information, that's something we can't
1247 * retrieve from page cache, and skip copying will lose those info.
1249 if (userfaultfd_wp(dst_vma
))
1252 if (src_vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
1255 if (src_vma
->anon_vma
)
1259 * Don't copy ptes where a page fault will fill them correctly. Fork
1260 * becomes much lighter when there are big shared or private readonly
1261 * mappings. The tradeoff is that copy_page_range is more efficient
1268 copy_page_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
)
1270 pgd_t
*src_pgd
, *dst_pgd
;
1272 unsigned long addr
= src_vma
->vm_start
;
1273 unsigned long end
= src_vma
->vm_end
;
1274 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1275 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1276 struct mmu_notifier_range range
;
1280 if (!vma_needs_copy(dst_vma
, src_vma
))
1283 if (is_vm_hugetlb_page(src_vma
))
1284 return copy_hugetlb_page_range(dst_mm
, src_mm
, dst_vma
, src_vma
);
1286 if (unlikely(src_vma
->vm_flags
& VM_PFNMAP
)) {
1288 * We do not free on error cases below as remove_vma
1289 * gets called on error from higher level routine
1291 ret
= track_pfn_copy(src_vma
);
1297 * We need to invalidate the secondary MMU mappings only when
1298 * there could be a permission downgrade on the ptes of the
1299 * parent mm. And a permission downgrade will only happen if
1300 * is_cow_mapping() returns true.
1302 is_cow
= is_cow_mapping(src_vma
->vm_flags
);
1305 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_PAGE
,
1306 0, src_mm
, addr
, end
);
1307 mmu_notifier_invalidate_range_start(&range
);
1309 * Disabling preemption is not needed for the write side, as
1310 * the read side doesn't spin, but goes to the mmap_lock.
1312 * Use the raw variant of the seqcount_t write API to avoid
1313 * lockdep complaining about preemptibility.
1315 vma_assert_write_locked(src_vma
);
1316 raw_write_seqcount_begin(&src_mm
->write_protect_seq
);
1320 dst_pgd
= pgd_offset(dst_mm
, addr
);
1321 src_pgd
= pgd_offset(src_mm
, addr
);
1323 next
= pgd_addr_end(addr
, end
);
1324 if (pgd_none_or_clear_bad(src_pgd
))
1326 if (unlikely(copy_p4d_range(dst_vma
, src_vma
, dst_pgd
, src_pgd
,
1328 untrack_pfn_clear(dst_vma
);
1332 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1335 raw_write_seqcount_end(&src_mm
->write_protect_seq
);
1336 mmu_notifier_invalidate_range_end(&range
);
1341 /* Whether we should zap all COWed (private) pages too */
1342 static inline bool should_zap_cows(struct zap_details
*details
)
1344 /* By default, zap all pages */
1348 /* Or, we zap COWed pages only if the caller wants to */
1349 return details
->even_cows
;
1352 /* Decides whether we should zap this page with the page pointer specified */
1353 static inline bool should_zap_page(struct zap_details
*details
, struct page
*page
)
1355 /* If we can make a decision without *page.. */
1356 if (should_zap_cows(details
))
1359 /* E.g. the caller passes NULL for the case of a zero page */
1363 /* Otherwise we should only zap non-anon pages */
1364 return !PageAnon(page
);
1367 static inline bool zap_drop_file_uffd_wp(struct zap_details
*details
)
1372 return details
->zap_flags
& ZAP_FLAG_DROP_MARKER
;
1376 * This function makes sure that we'll replace the none pte with an uffd-wp
1377 * swap special pte marker when necessary. Must be with the pgtable lock held.
1380 zap_install_uffd_wp_if_needed(struct vm_area_struct
*vma
,
1381 unsigned long addr
, pte_t
*pte
,
1382 struct zap_details
*details
, pte_t pteval
)
1384 /* Zap on anonymous always means dropping everything */
1385 if (vma_is_anonymous(vma
))
1388 if (zap_drop_file_uffd_wp(details
))
1391 pte_install_uffd_wp_if_needed(vma
, addr
, pte
, pteval
);
1394 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1395 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1396 unsigned long addr
, unsigned long end
,
1397 struct zap_details
*details
)
1399 struct mm_struct
*mm
= tlb
->mm
;
1400 int force_flush
= 0;
1401 int rss
[NR_MM_COUNTERS
];
1407 tlb_change_page_size(tlb
, PAGE_SIZE
);
1409 start_pte
= pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1413 flush_tlb_batched_pending(mm
);
1414 arch_enter_lazy_mmu_mode();
1416 pte_t ptent
= ptep_get(pte
);
1419 if (pte_none(ptent
))
1425 if (pte_present(ptent
)) {
1426 unsigned int delay_rmap
;
1428 page
= vm_normal_page(vma
, addr
, ptent
);
1429 if (unlikely(!should_zap_page(details
, page
)))
1431 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1433 arch_check_zapped_pte(vma
, ptent
);
1434 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1435 zap_install_uffd_wp_if_needed(vma
, addr
, pte
, details
,
1437 if (unlikely(!page
)) {
1438 ksm_might_unmap_zero_page(mm
, ptent
);
1443 if (!PageAnon(page
)) {
1444 if (pte_dirty(ptent
)) {
1445 set_page_dirty(page
);
1446 if (tlb_delay_rmap(tlb
)) {
1451 if (pte_young(ptent
) && likely(vma_has_recency(vma
)))
1452 mark_page_accessed(page
);
1454 rss
[mm_counter(page
)]--;
1456 page_remove_rmap(page
, vma
, false);
1457 if (unlikely(page_mapcount(page
) < 0))
1458 print_bad_pte(vma
, addr
, ptent
, page
);
1460 if (unlikely(__tlb_remove_page(tlb
, page
, delay_rmap
))) {
1468 entry
= pte_to_swp_entry(ptent
);
1469 if (is_device_private_entry(entry
) ||
1470 is_device_exclusive_entry(entry
)) {
1471 page
= pfn_swap_entry_to_page(entry
);
1472 if (unlikely(!should_zap_page(details
, page
)))
1475 * Both device private/exclusive mappings should only
1476 * work with anonymous page so far, so we don't need to
1477 * consider uffd-wp bit when zap. For more information,
1478 * see zap_install_uffd_wp_if_needed().
1480 WARN_ON_ONCE(!vma_is_anonymous(vma
));
1481 rss
[mm_counter(page
)]--;
1482 if (is_device_private_entry(entry
))
1483 page_remove_rmap(page
, vma
, false);
1485 } else if (!non_swap_entry(entry
)) {
1486 /* Genuine swap entry, hence a private anon page */
1487 if (!should_zap_cows(details
))
1490 if (unlikely(!free_swap_and_cache(entry
)))
1491 print_bad_pte(vma
, addr
, ptent
, NULL
);
1492 } else if (is_migration_entry(entry
)) {
1493 page
= pfn_swap_entry_to_page(entry
);
1494 if (!should_zap_page(details
, page
))
1496 rss
[mm_counter(page
)]--;
1497 } else if (pte_marker_entry_uffd_wp(entry
)) {
1499 * For anon: always drop the marker; for file: only
1500 * drop the marker if explicitly requested.
1502 if (!vma_is_anonymous(vma
) &&
1503 !zap_drop_file_uffd_wp(details
))
1505 } else if (is_hwpoison_entry(entry
) ||
1506 is_poisoned_swp_entry(entry
)) {
1507 if (!should_zap_cows(details
))
1510 /* We should have covered all the swap entry types */
1513 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1514 zap_install_uffd_wp_if_needed(vma
, addr
, pte
, details
, ptent
);
1515 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1517 add_mm_rss_vec(mm
, rss
);
1518 arch_leave_lazy_mmu_mode();
1520 /* Do the actual TLB flush before dropping ptl */
1522 tlb_flush_mmu_tlbonly(tlb
);
1523 tlb_flush_rmaps(tlb
, vma
);
1525 pte_unmap_unlock(start_pte
, ptl
);
1528 * If we forced a TLB flush (either due to running out of
1529 * batch buffers or because we needed to flush dirty TLB
1530 * entries before releasing the ptl), free the batched
1531 * memory too. Come back again if we didn't do everything.
1539 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1540 struct vm_area_struct
*vma
, pud_t
*pud
,
1541 unsigned long addr
, unsigned long end
,
1542 struct zap_details
*details
)
1547 pmd
= pmd_offset(pud
, addr
);
1549 next
= pmd_addr_end(addr
, end
);
1550 if (is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1551 if (next
- addr
!= HPAGE_PMD_SIZE
)
1552 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
1553 else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
)) {
1558 } else if (details
&& details
->single_folio
&&
1559 folio_test_pmd_mappable(details
->single_folio
) &&
1560 next
- addr
== HPAGE_PMD_SIZE
&& pmd_none(*pmd
)) {
1561 spinlock_t
*ptl
= pmd_lock(tlb
->mm
, pmd
);
1563 * Take and drop THP pmd lock so that we cannot return
1564 * prematurely, while zap_huge_pmd() has cleared *pmd,
1565 * but not yet decremented compound_mapcount().
1569 if (pmd_none(*pmd
)) {
1573 addr
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1576 } while (pmd
++, cond_resched(), addr
!= end
);
1581 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1582 struct vm_area_struct
*vma
, p4d_t
*p4d
,
1583 unsigned long addr
, unsigned long end
,
1584 struct zap_details
*details
)
1589 pud
= pud_offset(p4d
, addr
);
1591 next
= pud_addr_end(addr
, end
);
1592 if (pud_trans_huge(*pud
) || pud_devmap(*pud
)) {
1593 if (next
- addr
!= HPAGE_PUD_SIZE
) {
1594 mmap_assert_locked(tlb
->mm
);
1595 split_huge_pud(vma
, pud
, addr
);
1596 } else if (zap_huge_pud(tlb
, vma
, pud
, addr
))
1600 if (pud_none_or_clear_bad(pud
))
1602 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1605 } while (pud
++, addr
= next
, addr
!= end
);
1610 static inline unsigned long zap_p4d_range(struct mmu_gather
*tlb
,
1611 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1612 unsigned long addr
, unsigned long end
,
1613 struct zap_details
*details
)
1618 p4d
= p4d_offset(pgd
, addr
);
1620 next
= p4d_addr_end(addr
, end
);
1621 if (p4d_none_or_clear_bad(p4d
))
1623 next
= zap_pud_range(tlb
, vma
, p4d
, addr
, next
, details
);
1624 } while (p4d
++, addr
= next
, addr
!= end
);
1629 void unmap_page_range(struct mmu_gather
*tlb
,
1630 struct vm_area_struct
*vma
,
1631 unsigned long addr
, unsigned long end
,
1632 struct zap_details
*details
)
1637 BUG_ON(addr
>= end
);
1638 tlb_start_vma(tlb
, vma
);
1639 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1641 next
= pgd_addr_end(addr
, end
);
1642 if (pgd_none_or_clear_bad(pgd
))
1644 next
= zap_p4d_range(tlb
, vma
, pgd
, addr
, next
, details
);
1645 } while (pgd
++, addr
= next
, addr
!= end
);
1646 tlb_end_vma(tlb
, vma
);
1650 static void unmap_single_vma(struct mmu_gather
*tlb
,
1651 struct vm_area_struct
*vma
, unsigned long start_addr
,
1652 unsigned long end_addr
,
1653 struct zap_details
*details
, bool mm_wr_locked
)
1655 unsigned long start
= max(vma
->vm_start
, start_addr
);
1658 if (start
>= vma
->vm_end
)
1660 end
= min(vma
->vm_end
, end_addr
);
1661 if (end
<= vma
->vm_start
)
1665 uprobe_munmap(vma
, start
, end
);
1667 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1668 untrack_pfn(vma
, 0, 0, mm_wr_locked
);
1671 if (unlikely(is_vm_hugetlb_page(vma
))) {
1673 * It is undesirable to test vma->vm_file as it
1674 * should be non-null for valid hugetlb area.
1675 * However, vm_file will be NULL in the error
1676 * cleanup path of mmap_region. When
1677 * hugetlbfs ->mmap method fails,
1678 * mmap_region() nullifies vma->vm_file
1679 * before calling this function to clean up.
1680 * Since no pte has actually been setup, it is
1681 * safe to do nothing in this case.
1684 zap_flags_t zap_flags
= details
?
1685 details
->zap_flags
: 0;
1686 __unmap_hugepage_range_final(tlb
, vma
, start
, end
,
1690 unmap_page_range(tlb
, vma
, start
, end
, details
);
1695 * unmap_vmas - unmap a range of memory covered by a list of vma's
1696 * @tlb: address of the caller's struct mmu_gather
1697 * @mas: the maple state
1698 * @vma: the starting vma
1699 * @start_addr: virtual address at which to start unmapping
1700 * @end_addr: virtual address at which to end unmapping
1701 * @tree_end: The maximum index to check
1702 * @mm_wr_locked: lock flag
1704 * Unmap all pages in the vma list.
1706 * Only addresses between `start' and `end' will be unmapped.
1708 * The VMA list must be sorted in ascending virtual address order.
1710 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1711 * range after unmap_vmas() returns. So the only responsibility here is to
1712 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1713 * drops the lock and schedules.
1715 void unmap_vmas(struct mmu_gather
*tlb
, struct ma_state
*mas
,
1716 struct vm_area_struct
*vma
, unsigned long start_addr
,
1717 unsigned long end_addr
, unsigned long tree_end
,
1720 struct mmu_notifier_range range
;
1721 struct zap_details details
= {
1722 .zap_flags
= ZAP_FLAG_DROP_MARKER
| ZAP_FLAG_UNMAP
,
1723 /* Careful - we need to zap private pages too! */
1727 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
->vm_mm
,
1728 start_addr
, end_addr
);
1729 mmu_notifier_invalidate_range_start(&range
);
1731 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, &details
,
1733 } while ((vma
= mas_find(mas
, tree_end
- 1)) != NULL
);
1734 mmu_notifier_invalidate_range_end(&range
);
1738 * zap_page_range_single - remove user pages in a given range
1739 * @vma: vm_area_struct holding the applicable pages
1740 * @address: starting address of pages to zap
1741 * @size: number of bytes to zap
1742 * @details: details of shared cache invalidation
1744 * The range must fit into one VMA.
1746 void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1747 unsigned long size
, struct zap_details
*details
)
1749 const unsigned long end
= address
+ size
;
1750 struct mmu_notifier_range range
;
1751 struct mmu_gather tlb
;
1754 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
->vm_mm
,
1756 if (is_vm_hugetlb_page(vma
))
1757 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
,
1759 tlb_gather_mmu(&tlb
, vma
->vm_mm
);
1760 update_hiwater_rss(vma
->vm_mm
);
1761 mmu_notifier_invalidate_range_start(&range
);
1763 * unmap 'address-end' not 'range.start-range.end' as range
1764 * could have been expanded for hugetlb pmd sharing.
1766 unmap_single_vma(&tlb
, vma
, address
, end
, details
, false);
1767 mmu_notifier_invalidate_range_end(&range
);
1768 tlb_finish_mmu(&tlb
);
1772 * zap_vma_ptes - remove ptes mapping the vma
1773 * @vma: vm_area_struct holding ptes to be zapped
1774 * @address: starting address of pages to zap
1775 * @size: number of bytes to zap
1777 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1779 * The entire address range must be fully contained within the vma.
1782 void zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1785 if (!range_in_vma(vma
, address
, address
+ size
) ||
1786 !(vma
->vm_flags
& VM_PFNMAP
))
1789 zap_page_range_single(vma
, address
, size
, NULL
);
1791 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1793 static pmd_t
*walk_to_pmd(struct mm_struct
*mm
, unsigned long addr
)
1800 pgd
= pgd_offset(mm
, addr
);
1801 p4d
= p4d_alloc(mm
, pgd
, addr
);
1804 pud
= pud_alloc(mm
, p4d
, addr
);
1807 pmd
= pmd_alloc(mm
, pud
, addr
);
1811 VM_BUG_ON(pmd_trans_huge(*pmd
));
1815 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1818 pmd_t
*pmd
= walk_to_pmd(mm
, addr
);
1822 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1825 static int validate_page_before_insert(struct page
*page
)
1827 if (PageAnon(page
) || PageSlab(page
) || page_has_type(page
))
1829 flush_dcache_page(page
);
1833 static int insert_page_into_pte_locked(struct vm_area_struct
*vma
, pte_t
*pte
,
1834 unsigned long addr
, struct page
*page
, pgprot_t prot
)
1836 if (!pte_none(ptep_get(pte
)))
1838 /* Ok, finally just insert the thing.. */
1840 inc_mm_counter(vma
->vm_mm
, mm_counter_file(page
));
1841 page_add_file_rmap(page
, vma
, false);
1842 set_pte_at(vma
->vm_mm
, addr
, pte
, mk_pte(page
, prot
));
1847 * This is the old fallback for page remapping.
1849 * For historical reasons, it only allows reserved pages. Only
1850 * old drivers should use this, and they needed to mark their
1851 * pages reserved for the old functions anyway.
1853 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1854 struct page
*page
, pgprot_t prot
)
1860 retval
= validate_page_before_insert(page
);
1864 pte
= get_locked_pte(vma
->vm_mm
, addr
, &ptl
);
1867 retval
= insert_page_into_pte_locked(vma
, pte
, addr
, page
, prot
);
1868 pte_unmap_unlock(pte
, ptl
);
1873 static int insert_page_in_batch_locked(struct vm_area_struct
*vma
, pte_t
*pte
,
1874 unsigned long addr
, struct page
*page
, pgprot_t prot
)
1878 if (!page_count(page
))
1880 err
= validate_page_before_insert(page
);
1883 return insert_page_into_pte_locked(vma
, pte
, addr
, page
, prot
);
1886 /* insert_pages() amortizes the cost of spinlock operations
1887 * when inserting pages in a loop.
1889 static int insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
1890 struct page
**pages
, unsigned long *num
, pgprot_t prot
)
1893 pte_t
*start_pte
, *pte
;
1894 spinlock_t
*pte_lock
;
1895 struct mm_struct
*const mm
= vma
->vm_mm
;
1896 unsigned long curr_page_idx
= 0;
1897 unsigned long remaining_pages_total
= *num
;
1898 unsigned long pages_to_write_in_pmd
;
1902 pmd
= walk_to_pmd(mm
, addr
);
1906 pages_to_write_in_pmd
= min_t(unsigned long,
1907 remaining_pages_total
, PTRS_PER_PTE
- pte_index(addr
));
1909 /* Allocate the PTE if necessary; takes PMD lock once only. */
1911 if (pte_alloc(mm
, pmd
))
1914 while (pages_to_write_in_pmd
) {
1916 const int batch_size
= min_t(int, pages_to_write_in_pmd
, 8);
1918 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &pte_lock
);
1923 for (pte
= start_pte
; pte_idx
< batch_size
; ++pte
, ++pte_idx
) {
1924 int err
= insert_page_in_batch_locked(vma
, pte
,
1925 addr
, pages
[curr_page_idx
], prot
);
1926 if (unlikely(err
)) {
1927 pte_unmap_unlock(start_pte
, pte_lock
);
1929 remaining_pages_total
-= pte_idx
;
1935 pte_unmap_unlock(start_pte
, pte_lock
);
1936 pages_to_write_in_pmd
-= batch_size
;
1937 remaining_pages_total
-= batch_size
;
1939 if (remaining_pages_total
)
1943 *num
= remaining_pages_total
;
1948 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1949 * @vma: user vma to map to
1950 * @addr: target start user address of these pages
1951 * @pages: source kernel pages
1952 * @num: in: number of pages to map. out: number of pages that were *not*
1953 * mapped. (0 means all pages were successfully mapped).
1955 * Preferred over vm_insert_page() when inserting multiple pages.
1957 * In case of error, we may have mapped a subset of the provided
1958 * pages. It is the caller's responsibility to account for this case.
1960 * The same restrictions apply as in vm_insert_page().
1962 int vm_insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
1963 struct page
**pages
, unsigned long *num
)
1965 const unsigned long end_addr
= addr
+ (*num
* PAGE_SIZE
) - 1;
1967 if (addr
< vma
->vm_start
|| end_addr
>= vma
->vm_end
)
1969 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1970 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
1971 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1972 vm_flags_set(vma
, VM_MIXEDMAP
);
1974 /* Defer page refcount checking till we're about to map that page. */
1975 return insert_pages(vma
, addr
, pages
, num
, vma
->vm_page_prot
);
1977 EXPORT_SYMBOL(vm_insert_pages
);
1980 * vm_insert_page - insert single page into user vma
1981 * @vma: user vma to map to
1982 * @addr: target user address of this page
1983 * @page: source kernel page
1985 * This allows drivers to insert individual pages they've allocated
1988 * The page has to be a nice clean _individual_ kernel allocation.
1989 * If you allocate a compound page, you need to have marked it as
1990 * such (__GFP_COMP), or manually just split the page up yourself
1991 * (see split_page()).
1993 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1994 * took an arbitrary page protection parameter. This doesn't allow
1995 * that. Your vma protection will have to be set up correctly, which
1996 * means that if you want a shared writable mapping, you'd better
1997 * ask for a shared writable mapping!
1999 * The page does not need to be reserved.
2001 * Usually this function is called from f_op->mmap() handler
2002 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
2003 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2004 * function from other places, for example from page-fault handler.
2006 * Return: %0 on success, negative error code otherwise.
2008 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2011 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2013 if (!page_count(page
))
2015 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
2016 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
2017 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
2018 vm_flags_set(vma
, VM_MIXEDMAP
);
2020 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2022 EXPORT_SYMBOL(vm_insert_page
);
2025 * __vm_map_pages - maps range of kernel pages into user vma
2026 * @vma: user vma to map to
2027 * @pages: pointer to array of source kernel pages
2028 * @num: number of pages in page array
2029 * @offset: user's requested vm_pgoff
2031 * This allows drivers to map range of kernel pages into a user vma.
2033 * Return: 0 on success and error code otherwise.
2035 static int __vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
2036 unsigned long num
, unsigned long offset
)
2038 unsigned long count
= vma_pages(vma
);
2039 unsigned long uaddr
= vma
->vm_start
;
2042 /* Fail if the user requested offset is beyond the end of the object */
2046 /* Fail if the user requested size exceeds available object size */
2047 if (count
> num
- offset
)
2050 for (i
= 0; i
< count
; i
++) {
2051 ret
= vm_insert_page(vma
, uaddr
, pages
[offset
+ i
]);
2061 * vm_map_pages - maps range of kernel pages starts with non zero offset
2062 * @vma: user vma to map to
2063 * @pages: pointer to array of source kernel pages
2064 * @num: number of pages in page array
2066 * Maps an object consisting of @num pages, catering for the user's
2067 * requested vm_pgoff
2069 * If we fail to insert any page into the vma, the function will return
2070 * immediately leaving any previously inserted pages present. Callers
2071 * from the mmap handler may immediately return the error as their caller
2072 * will destroy the vma, removing any successfully inserted pages. Other
2073 * callers should make their own arrangements for calling unmap_region().
2075 * Context: Process context. Called by mmap handlers.
2076 * Return: 0 on success and error code otherwise.
2078 int vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
2081 return __vm_map_pages(vma
, pages
, num
, vma
->vm_pgoff
);
2083 EXPORT_SYMBOL(vm_map_pages
);
2086 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2087 * @vma: user vma to map to
2088 * @pages: pointer to array of source kernel pages
2089 * @num: number of pages in page array
2091 * Similar to vm_map_pages(), except that it explicitly sets the offset
2092 * to 0. This function is intended for the drivers that did not consider
2095 * Context: Process context. Called by mmap handlers.
2096 * Return: 0 on success and error code otherwise.
2098 int vm_map_pages_zero(struct vm_area_struct
*vma
, struct page
**pages
,
2101 return __vm_map_pages(vma
, pages
, num
, 0);
2103 EXPORT_SYMBOL(vm_map_pages_zero
);
2105 static vm_fault_t
insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2106 pfn_t pfn
, pgprot_t prot
, bool mkwrite
)
2108 struct mm_struct
*mm
= vma
->vm_mm
;
2112 pte
= get_locked_pte(mm
, addr
, &ptl
);
2114 return VM_FAULT_OOM
;
2115 entry
= ptep_get(pte
);
2116 if (!pte_none(entry
)) {
2119 * For read faults on private mappings the PFN passed
2120 * in may not match the PFN we have mapped if the
2121 * mapped PFN is a writeable COW page. In the mkwrite
2122 * case we are creating a writable PTE for a shared
2123 * mapping and we expect the PFNs to match. If they
2124 * don't match, we are likely racing with block
2125 * allocation and mapping invalidation so just skip the
2128 if (pte_pfn(entry
) != pfn_t_to_pfn(pfn
)) {
2129 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry
)));
2132 entry
= pte_mkyoung(entry
);
2133 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2134 if (ptep_set_access_flags(vma
, addr
, pte
, entry
, 1))
2135 update_mmu_cache(vma
, addr
, pte
);
2140 /* Ok, finally just insert the thing.. */
2141 if (pfn_t_devmap(pfn
))
2142 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
2144 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
2147 entry
= pte_mkyoung(entry
);
2148 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2151 set_pte_at(mm
, addr
, pte
, entry
);
2152 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
2155 pte_unmap_unlock(pte
, ptl
);
2156 return VM_FAULT_NOPAGE
;
2160 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2161 * @vma: user vma to map to
2162 * @addr: target user address of this page
2163 * @pfn: source kernel pfn
2164 * @pgprot: pgprot flags for the inserted page
2166 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2167 * to override pgprot on a per-page basis.
2169 * This only makes sense for IO mappings, and it makes no sense for
2170 * COW mappings. In general, using multiple vmas is preferable;
2171 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2174 * pgprot typically only differs from @vma->vm_page_prot when drivers set
2175 * caching- and encryption bits different than those of @vma->vm_page_prot,
2176 * because the caching- or encryption mode may not be known at mmap() time.
2178 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2179 * to set caching and encryption bits for those vmas (except for COW pages).
2180 * This is ensured by core vm only modifying these page table entries using
2181 * functions that don't touch caching- or encryption bits, using pte_modify()
2182 * if needed. (See for example mprotect()).
2184 * Also when new page-table entries are created, this is only done using the
2185 * fault() callback, and never using the value of vma->vm_page_prot,
2186 * except for page-table entries that point to anonymous pages as the result
2189 * Context: Process context. May allocate using %GFP_KERNEL.
2190 * Return: vm_fault_t value.
2192 vm_fault_t
vmf_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
2193 unsigned long pfn
, pgprot_t pgprot
)
2196 * Technically, architectures with pte_special can avoid all these
2197 * restrictions (same for remap_pfn_range). However we would like
2198 * consistency in testing and feature parity among all, so we should
2199 * try to keep these invariants in place for everybody.
2201 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
2202 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
2203 (VM_PFNMAP
|VM_MIXEDMAP
));
2204 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
2205 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
2207 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2208 return VM_FAULT_SIGBUS
;
2210 if (!pfn_modify_allowed(pfn
, pgprot
))
2211 return VM_FAULT_SIGBUS
;
2213 track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
));
2215 return insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
,
2218 EXPORT_SYMBOL(vmf_insert_pfn_prot
);
2221 * vmf_insert_pfn - insert single pfn into user vma
2222 * @vma: user vma to map to
2223 * @addr: target user address of this page
2224 * @pfn: source kernel pfn
2226 * Similar to vm_insert_page, this allows drivers to insert individual pages
2227 * they've allocated into a user vma. Same comments apply.
2229 * This function should only be called from a vm_ops->fault handler, and
2230 * in that case the handler should return the result of this function.
2232 * vma cannot be a COW mapping.
2234 * As this is called only for pages that do not currently exist, we
2235 * do not need to flush old virtual caches or the TLB.
2237 * Context: Process context. May allocate using %GFP_KERNEL.
2238 * Return: vm_fault_t value.
2240 vm_fault_t
vmf_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2243 return vmf_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
2245 EXPORT_SYMBOL(vmf_insert_pfn
);
2247 static bool vm_mixed_ok(struct vm_area_struct
*vma
, pfn_t pfn
)
2249 /* these checks mirror the abort conditions in vm_normal_page */
2250 if (vma
->vm_flags
& VM_MIXEDMAP
)
2252 if (pfn_t_devmap(pfn
))
2254 if (pfn_t_special(pfn
))
2256 if (is_zero_pfn(pfn_t_to_pfn(pfn
)))
2261 static vm_fault_t
__vm_insert_mixed(struct vm_area_struct
*vma
,
2262 unsigned long addr
, pfn_t pfn
, bool mkwrite
)
2264 pgprot_t pgprot
= vma
->vm_page_prot
;
2267 BUG_ON(!vm_mixed_ok(vma
, pfn
));
2269 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2270 return VM_FAULT_SIGBUS
;
2272 track_pfn_insert(vma
, &pgprot
, pfn
);
2274 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn
), pgprot
))
2275 return VM_FAULT_SIGBUS
;
2278 * If we don't have pte special, then we have to use the pfn_valid()
2279 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2280 * refcount the page if pfn_valid is true (hence insert_page rather
2281 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2282 * without pte special, it would there be refcounted as a normal page.
2284 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
) &&
2285 !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
2289 * At this point we are committed to insert_page()
2290 * regardless of whether the caller specified flags that
2291 * result in pfn_t_has_page() == false.
2293 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
2294 err
= insert_page(vma
, addr
, page
, pgprot
);
2296 return insert_pfn(vma
, addr
, pfn
, pgprot
, mkwrite
);
2300 return VM_FAULT_OOM
;
2301 if (err
< 0 && err
!= -EBUSY
)
2302 return VM_FAULT_SIGBUS
;
2304 return VM_FAULT_NOPAGE
;
2307 vm_fault_t
vmf_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2310 return __vm_insert_mixed(vma
, addr
, pfn
, false);
2312 EXPORT_SYMBOL(vmf_insert_mixed
);
2315 * If the insertion of PTE failed because someone else already added a
2316 * different entry in the mean time, we treat that as success as we assume
2317 * the same entry was actually inserted.
2319 vm_fault_t
vmf_insert_mixed_mkwrite(struct vm_area_struct
*vma
,
2320 unsigned long addr
, pfn_t pfn
)
2322 return __vm_insert_mixed(vma
, addr
, pfn
, true);
2324 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite
);
2327 * maps a range of physical memory into the requested pages. the old
2328 * mappings are removed. any references to nonexistent pages results
2329 * in null mappings (currently treated as "copy-on-access")
2331 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2332 unsigned long addr
, unsigned long end
,
2333 unsigned long pfn
, pgprot_t prot
)
2335 pte_t
*pte
, *mapped_pte
;
2339 mapped_pte
= pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2342 arch_enter_lazy_mmu_mode();
2344 BUG_ON(!pte_none(ptep_get(pte
)));
2345 if (!pfn_modify_allowed(pfn
, prot
)) {
2349 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2351 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2352 arch_leave_lazy_mmu_mode();
2353 pte_unmap_unlock(mapped_pte
, ptl
);
2357 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2358 unsigned long addr
, unsigned long end
,
2359 unsigned long pfn
, pgprot_t prot
)
2365 pfn
-= addr
>> PAGE_SHIFT
;
2366 pmd
= pmd_alloc(mm
, pud
, addr
);
2369 VM_BUG_ON(pmd_trans_huge(*pmd
));
2371 next
= pmd_addr_end(addr
, end
);
2372 err
= remap_pte_range(mm
, pmd
, addr
, next
,
2373 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2376 } while (pmd
++, addr
= next
, addr
!= end
);
2380 static inline int remap_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2381 unsigned long addr
, unsigned long end
,
2382 unsigned long pfn
, pgprot_t prot
)
2388 pfn
-= addr
>> PAGE_SHIFT
;
2389 pud
= pud_alloc(mm
, p4d
, addr
);
2393 next
= pud_addr_end(addr
, end
);
2394 err
= remap_pmd_range(mm
, pud
, addr
, next
,
2395 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2398 } while (pud
++, addr
= next
, addr
!= end
);
2402 static inline int remap_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2403 unsigned long addr
, unsigned long end
,
2404 unsigned long pfn
, pgprot_t prot
)
2410 pfn
-= addr
>> PAGE_SHIFT
;
2411 p4d
= p4d_alloc(mm
, pgd
, addr
);
2415 next
= p4d_addr_end(addr
, end
);
2416 err
= remap_pud_range(mm
, p4d
, addr
, next
,
2417 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2420 } while (p4d
++, addr
= next
, addr
!= end
);
2425 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2426 * must have pre-validated the caching bits of the pgprot_t.
2428 int remap_pfn_range_notrack(struct vm_area_struct
*vma
, unsigned long addr
,
2429 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2433 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2434 struct mm_struct
*mm
= vma
->vm_mm
;
2437 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr
)))
2441 * Physically remapped pages are special. Tell the
2442 * rest of the world about it:
2443 * VM_IO tells people not to look at these pages
2444 * (accesses can have side effects).
2445 * VM_PFNMAP tells the core MM that the base pages are just
2446 * raw PFN mappings, and do not have a "struct page" associated
2449 * Disable vma merging and expanding with mremap().
2451 * Omit vma from core dump, even when VM_IO turned off.
2453 * There's a horrible special case to handle copy-on-write
2454 * behaviour that some programs depend on. We mark the "original"
2455 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2456 * See vm_normal_page() for details.
2458 if (is_cow_mapping(vma
->vm_flags
)) {
2459 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2461 vma
->vm_pgoff
= pfn
;
2464 vm_flags_set(vma
, VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
);
2466 BUG_ON(addr
>= end
);
2467 pfn
-= addr
>> PAGE_SHIFT
;
2468 pgd
= pgd_offset(mm
, addr
);
2469 flush_cache_range(vma
, addr
, end
);
2471 next
= pgd_addr_end(addr
, end
);
2472 err
= remap_p4d_range(mm
, pgd
, addr
, next
,
2473 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2476 } while (pgd
++, addr
= next
, addr
!= end
);
2482 * remap_pfn_range - remap kernel memory to userspace
2483 * @vma: user vma to map to
2484 * @addr: target page aligned user address to start at
2485 * @pfn: page frame number of kernel physical memory address
2486 * @size: size of mapping area
2487 * @prot: page protection flags for this mapping
2489 * Note: this is only safe if the mm semaphore is held when called.
2491 * Return: %0 on success, negative error code otherwise.
2493 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2494 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2498 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
2502 err
= remap_pfn_range_notrack(vma
, addr
, pfn
, size
, prot
);
2504 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
), true);
2507 EXPORT_SYMBOL(remap_pfn_range
);
2510 * vm_iomap_memory - remap memory to userspace
2511 * @vma: user vma to map to
2512 * @start: start of the physical memory to be mapped
2513 * @len: size of area
2515 * This is a simplified io_remap_pfn_range() for common driver use. The
2516 * driver just needs to give us the physical memory range to be mapped,
2517 * we'll figure out the rest from the vma information.
2519 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2520 * whatever write-combining details or similar.
2522 * Return: %0 on success, negative error code otherwise.
2524 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2526 unsigned long vm_len
, pfn
, pages
;
2528 /* Check that the physical memory area passed in looks valid */
2529 if (start
+ len
< start
)
2532 * You *really* shouldn't map things that aren't page-aligned,
2533 * but we've historically allowed it because IO memory might
2534 * just have smaller alignment.
2536 len
+= start
& ~PAGE_MASK
;
2537 pfn
= start
>> PAGE_SHIFT
;
2538 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2539 if (pfn
+ pages
< pfn
)
2542 /* We start the mapping 'vm_pgoff' pages into the area */
2543 if (vma
->vm_pgoff
> pages
)
2545 pfn
+= vma
->vm_pgoff
;
2546 pages
-= vma
->vm_pgoff
;
2548 /* Can we fit all of the mapping? */
2549 vm_len
= vma
->vm_end
- vma
->vm_start
;
2550 if (vm_len
>> PAGE_SHIFT
> pages
)
2553 /* Ok, let it rip */
2554 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2556 EXPORT_SYMBOL(vm_iomap_memory
);
2558 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2559 unsigned long addr
, unsigned long end
,
2560 pte_fn_t fn
, void *data
, bool create
,
2561 pgtbl_mod_mask
*mask
)
2563 pte_t
*pte
, *mapped_pte
;
2568 mapped_pte
= pte
= (mm
== &init_mm
) ?
2569 pte_alloc_kernel_track(pmd
, addr
, mask
) :
2570 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2574 mapped_pte
= pte
= (mm
== &init_mm
) ?
2575 pte_offset_kernel(pmd
, addr
) :
2576 pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
2581 arch_enter_lazy_mmu_mode();
2585 if (create
|| !pte_none(ptep_get(pte
))) {
2586 err
= fn(pte
++, addr
, data
);
2590 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2592 *mask
|= PGTBL_PTE_MODIFIED
;
2594 arch_leave_lazy_mmu_mode();
2597 pte_unmap_unlock(mapped_pte
, ptl
);
2601 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2602 unsigned long addr
, unsigned long end
,
2603 pte_fn_t fn
, void *data
, bool create
,
2604 pgtbl_mod_mask
*mask
)
2610 BUG_ON(pud_huge(*pud
));
2613 pmd
= pmd_alloc_track(mm
, pud
, addr
, mask
);
2617 pmd
= pmd_offset(pud
, addr
);
2620 next
= pmd_addr_end(addr
, end
);
2621 if (pmd_none(*pmd
) && !create
)
2623 if (WARN_ON_ONCE(pmd_leaf(*pmd
)))
2625 if (!pmd_none(*pmd
) && WARN_ON_ONCE(pmd_bad(*pmd
))) {
2630 err
= apply_to_pte_range(mm
, pmd
, addr
, next
,
2631 fn
, data
, create
, mask
);
2634 } while (pmd
++, addr
= next
, addr
!= end
);
2639 static int apply_to_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2640 unsigned long addr
, unsigned long end
,
2641 pte_fn_t fn
, void *data
, bool create
,
2642 pgtbl_mod_mask
*mask
)
2649 pud
= pud_alloc_track(mm
, p4d
, addr
, mask
);
2653 pud
= pud_offset(p4d
, addr
);
2656 next
= pud_addr_end(addr
, end
);
2657 if (pud_none(*pud
) && !create
)
2659 if (WARN_ON_ONCE(pud_leaf(*pud
)))
2661 if (!pud_none(*pud
) && WARN_ON_ONCE(pud_bad(*pud
))) {
2666 err
= apply_to_pmd_range(mm
, pud
, addr
, next
,
2667 fn
, data
, create
, mask
);
2670 } while (pud
++, addr
= next
, addr
!= end
);
2675 static int apply_to_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2676 unsigned long addr
, unsigned long end
,
2677 pte_fn_t fn
, void *data
, bool create
,
2678 pgtbl_mod_mask
*mask
)
2685 p4d
= p4d_alloc_track(mm
, pgd
, addr
, mask
);
2689 p4d
= p4d_offset(pgd
, addr
);
2692 next
= p4d_addr_end(addr
, end
);
2693 if (p4d_none(*p4d
) && !create
)
2695 if (WARN_ON_ONCE(p4d_leaf(*p4d
)))
2697 if (!p4d_none(*p4d
) && WARN_ON_ONCE(p4d_bad(*p4d
))) {
2702 err
= apply_to_pud_range(mm
, p4d
, addr
, next
,
2703 fn
, data
, create
, mask
);
2706 } while (p4d
++, addr
= next
, addr
!= end
);
2711 static int __apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2712 unsigned long size
, pte_fn_t fn
,
2713 void *data
, bool create
)
2716 unsigned long start
= addr
, next
;
2717 unsigned long end
= addr
+ size
;
2718 pgtbl_mod_mask mask
= 0;
2721 if (WARN_ON(addr
>= end
))
2724 pgd
= pgd_offset(mm
, addr
);
2726 next
= pgd_addr_end(addr
, end
);
2727 if (pgd_none(*pgd
) && !create
)
2729 if (WARN_ON_ONCE(pgd_leaf(*pgd
)))
2731 if (!pgd_none(*pgd
) && WARN_ON_ONCE(pgd_bad(*pgd
))) {
2736 err
= apply_to_p4d_range(mm
, pgd
, addr
, next
,
2737 fn
, data
, create
, &mask
);
2740 } while (pgd
++, addr
= next
, addr
!= end
);
2742 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
2743 arch_sync_kernel_mappings(start
, start
+ size
);
2749 * Scan a region of virtual memory, filling in page tables as necessary
2750 * and calling a provided function on each leaf page table.
2752 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2753 unsigned long size
, pte_fn_t fn
, void *data
)
2755 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, true);
2757 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2760 * Scan a region of virtual memory, calling a provided function on
2761 * each leaf page table where it exists.
2763 * Unlike apply_to_page_range, this does _not_ fill in page tables
2764 * where they are absent.
2766 int apply_to_existing_page_range(struct mm_struct
*mm
, unsigned long addr
,
2767 unsigned long size
, pte_fn_t fn
, void *data
)
2769 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, false);
2771 EXPORT_SYMBOL_GPL(apply_to_existing_page_range
);
2774 * handle_pte_fault chooses page fault handler according to an entry which was
2775 * read non-atomically. Before making any commitment, on those architectures
2776 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2777 * parts, do_swap_page must check under lock before unmapping the pte and
2778 * proceeding (but do_wp_page is only called after already making such a check;
2779 * and do_anonymous_page can safely check later on).
2781 static inline int pte_unmap_same(struct vm_fault
*vmf
)
2784 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2785 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2786 spin_lock(vmf
->ptl
);
2787 same
= pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
);
2788 spin_unlock(vmf
->ptl
);
2791 pte_unmap(vmf
->pte
);
2798 * 0: copied succeeded
2799 * -EHWPOISON: copy failed due to hwpoison in source page
2800 * -EAGAIN: copied failed (some other reason)
2802 static inline int __wp_page_copy_user(struct page
*dst
, struct page
*src
,
2803 struct vm_fault
*vmf
)
2808 struct vm_area_struct
*vma
= vmf
->vma
;
2809 struct mm_struct
*mm
= vma
->vm_mm
;
2810 unsigned long addr
= vmf
->address
;
2813 if (copy_mc_user_highpage(dst
, src
, addr
, vma
)) {
2814 memory_failure_queue(page_to_pfn(src
), 0);
2821 * If the source page was a PFN mapping, we don't have
2822 * a "struct page" for it. We do a best-effort copy by
2823 * just copying from the original user address. If that
2824 * fails, we just zero-fill it. Live with it.
2826 kaddr
= kmap_atomic(dst
);
2827 uaddr
= (void __user
*)(addr
& PAGE_MASK
);
2830 * On architectures with software "accessed" bits, we would
2831 * take a double page fault, so mark it accessed here.
2834 if (!arch_has_hw_pte_young() && !pte_young(vmf
->orig_pte
)) {
2837 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2838 if (unlikely(!vmf
->pte
|| !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
2840 * Other thread has already handled the fault
2841 * and update local tlb only
2844 update_mmu_tlb(vma
, addr
, vmf
->pte
);
2849 entry
= pte_mkyoung(vmf
->orig_pte
);
2850 if (ptep_set_access_flags(vma
, addr
, vmf
->pte
, entry
, 0))
2851 update_mmu_cache_range(vmf
, vma
, addr
, vmf
->pte
, 1);
2855 * This really shouldn't fail, because the page is there
2856 * in the page tables. But it might just be unreadable,
2857 * in which case we just give up and fill the result with
2860 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
2864 /* Re-validate under PTL if the page is still mapped */
2865 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2866 if (unlikely(!vmf
->pte
|| !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
2867 /* The PTE changed under us, update local tlb */
2869 update_mmu_tlb(vma
, addr
, vmf
->pte
);
2875 * The same page can be mapped back since last copy attempt.
2876 * Try to copy again under PTL.
2878 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
2880 * Give a warn in case there can be some obscure
2893 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2894 kunmap_atomic(kaddr
);
2895 flush_dcache_page(dst
);
2900 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
2902 struct file
*vm_file
= vma
->vm_file
;
2905 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
2908 * Special mappings (e.g. VDSO) do not have any file so fake
2909 * a default GFP_KERNEL for them.
2915 * Notify the address space that the page is about to become writable so that
2916 * it can prohibit this or wait for the page to get into an appropriate state.
2918 * We do this without the lock held, so that it can sleep if it needs to.
2920 static vm_fault_t
do_page_mkwrite(struct vm_fault
*vmf
, struct folio
*folio
)
2923 unsigned int old_flags
= vmf
->flags
;
2925 vmf
->flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2927 if (vmf
->vma
->vm_file
&&
2928 IS_SWAPFILE(vmf
->vma
->vm_file
->f_mapping
->host
))
2929 return VM_FAULT_SIGBUS
;
2931 ret
= vmf
->vma
->vm_ops
->page_mkwrite(vmf
);
2932 /* Restore original flags so that caller is not surprised */
2933 vmf
->flags
= old_flags
;
2934 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2936 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
2938 if (!folio
->mapping
) {
2939 folio_unlock(folio
);
2940 return 0; /* retry */
2942 ret
|= VM_FAULT_LOCKED
;
2944 VM_BUG_ON_FOLIO(!folio_test_locked(folio
), folio
);
2949 * Handle dirtying of a page in shared file mapping on a write fault.
2951 * The function expects the page to be locked and unlocks it.
2953 static vm_fault_t
fault_dirty_shared_page(struct vm_fault
*vmf
)
2955 struct vm_area_struct
*vma
= vmf
->vma
;
2956 struct address_space
*mapping
;
2957 struct folio
*folio
= page_folio(vmf
->page
);
2959 bool page_mkwrite
= vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
;
2961 dirtied
= folio_mark_dirty(folio
);
2962 VM_BUG_ON_FOLIO(folio_test_anon(folio
), folio
);
2964 * Take a local copy of the address_space - folio.mapping may be zeroed
2965 * by truncate after folio_unlock(). The address_space itself remains
2966 * pinned by vma->vm_file's reference. We rely on folio_unlock()'s
2967 * release semantics to prevent the compiler from undoing this copying.
2969 mapping
= folio_raw_mapping(folio
);
2970 folio_unlock(folio
);
2973 file_update_time(vma
->vm_file
);
2976 * Throttle page dirtying rate down to writeback speed.
2978 * mapping may be NULL here because some device drivers do not
2979 * set page.mapping but still dirty their pages
2981 * Drop the mmap_lock before waiting on IO, if we can. The file
2982 * is pinning the mapping, as per above.
2984 if ((dirtied
|| page_mkwrite
) && mapping
) {
2987 fpin
= maybe_unlock_mmap_for_io(vmf
, NULL
);
2988 balance_dirty_pages_ratelimited(mapping
);
2991 return VM_FAULT_COMPLETED
;
2999 * Handle write page faults for pages that can be reused in the current vma
3001 * This can happen either due to the mapping being with the VM_SHARED flag,
3002 * or due to us being the last reference standing to the page. In either
3003 * case, all we need to do here is to mark the page as writable and update
3004 * any related book-keeping.
3006 static inline void wp_page_reuse(struct vm_fault
*vmf
)
3007 __releases(vmf
->ptl
)
3009 struct vm_area_struct
*vma
= vmf
->vma
;
3010 struct page
*page
= vmf
->page
;
3013 VM_BUG_ON(!(vmf
->flags
& FAULT_FLAG_WRITE
));
3014 VM_BUG_ON(page
&& PageAnon(page
) && !PageAnonExclusive(page
));
3017 * Clear the pages cpupid information as the existing
3018 * information potentially belongs to a now completely
3019 * unrelated process.
3022 page_cpupid_xchg_last(page
, (1 << LAST_CPUPID_SHIFT
) - 1);
3024 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
3025 entry
= pte_mkyoung(vmf
->orig_pte
);
3026 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3027 if (ptep_set_access_flags(vma
, vmf
->address
, vmf
->pte
, entry
, 1))
3028 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
3029 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3030 count_vm_event(PGREUSE
);
3034 * Handle the case of a page which we actually need to copy to a new page,
3035 * either due to COW or unsharing.
3037 * Called with mmap_lock locked and the old page referenced, but
3038 * without the ptl held.
3040 * High level logic flow:
3042 * - Allocate a page, copy the content of the old page to the new one.
3043 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
3044 * - Take the PTL. If the pte changed, bail out and release the allocated page
3045 * - If the pte is still the way we remember it, update the page table and all
3046 * relevant references. This includes dropping the reference the page-table
3047 * held to the old page, as well as updating the rmap.
3048 * - In any case, unlock the PTL and drop the reference we took to the old page.
3050 static vm_fault_t
wp_page_copy(struct vm_fault
*vmf
)
3052 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
3053 struct vm_area_struct
*vma
= vmf
->vma
;
3054 struct mm_struct
*mm
= vma
->vm_mm
;
3055 struct folio
*old_folio
= NULL
;
3056 struct folio
*new_folio
= NULL
;
3058 int page_copied
= 0;
3059 struct mmu_notifier_range range
;
3062 delayacct_wpcopy_start();
3065 old_folio
= page_folio(vmf
->page
);
3066 if (unlikely(anon_vma_prepare(vma
)))
3069 if (is_zero_pfn(pte_pfn(vmf
->orig_pte
))) {
3070 new_folio
= vma_alloc_zeroed_movable_folio(vma
, vmf
->address
);
3074 new_folio
= vma_alloc_folio(GFP_HIGHUSER_MOVABLE
, 0, vma
,
3075 vmf
->address
, false);
3079 ret
= __wp_page_copy_user(&new_folio
->page
, vmf
->page
, vmf
);
3082 * COW failed, if the fault was solved by other,
3083 * it's fine. If not, userspace would re-fault on
3084 * the same address and we will handle the fault
3085 * from the second attempt.
3086 * The -EHWPOISON case will not be retried.
3088 folio_put(new_folio
);
3090 folio_put(old_folio
);
3092 delayacct_wpcopy_end();
3093 return ret
== -EHWPOISON
? VM_FAULT_HWPOISON
: 0;
3095 kmsan_copy_page_meta(&new_folio
->page
, vmf
->page
);
3098 if (mem_cgroup_charge(new_folio
, mm
, GFP_KERNEL
))
3100 folio_throttle_swaprate(new_folio
, GFP_KERNEL
);
3102 __folio_mark_uptodate(new_folio
);
3104 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, mm
,
3105 vmf
->address
& PAGE_MASK
,
3106 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
);
3107 mmu_notifier_invalidate_range_start(&range
);
3110 * Re-check the pte - we dropped the lock
3112 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, vmf
->address
, &vmf
->ptl
);
3113 if (likely(vmf
->pte
&& pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
3115 if (!folio_test_anon(old_folio
)) {
3116 dec_mm_counter(mm
, mm_counter_file(&old_folio
->page
));
3117 inc_mm_counter(mm
, MM_ANONPAGES
);
3120 ksm_might_unmap_zero_page(mm
, vmf
->orig_pte
);
3121 inc_mm_counter(mm
, MM_ANONPAGES
);
3123 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
3124 entry
= mk_pte(&new_folio
->page
, vma
->vm_page_prot
);
3125 entry
= pte_sw_mkyoung(entry
);
3126 if (unlikely(unshare
)) {
3127 if (pte_soft_dirty(vmf
->orig_pte
))
3128 entry
= pte_mksoft_dirty(entry
);
3129 if (pte_uffd_wp(vmf
->orig_pte
))
3130 entry
= pte_mkuffd_wp(entry
);
3132 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3136 * Clear the pte entry and flush it first, before updating the
3137 * pte with the new entry, to keep TLBs on different CPUs in
3138 * sync. This code used to set the new PTE then flush TLBs, but
3139 * that left a window where the new PTE could be loaded into
3140 * some TLBs while the old PTE remains in others.
3142 ptep_clear_flush(vma
, vmf
->address
, vmf
->pte
);
3143 folio_add_new_anon_rmap(new_folio
, vma
, vmf
->address
);
3144 folio_add_lru_vma(new_folio
, vma
);
3146 * We call the notify macro here because, when using secondary
3147 * mmu page tables (such as kvm shadow page tables), we want the
3148 * new page to be mapped directly into the secondary page table.
3150 BUG_ON(unshare
&& pte_write(entry
));
3151 set_pte_at_notify(mm
, vmf
->address
, vmf
->pte
, entry
);
3152 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
3155 * Only after switching the pte to the new page may
3156 * we remove the mapcount here. Otherwise another
3157 * process may come and find the rmap count decremented
3158 * before the pte is switched to the new page, and
3159 * "reuse" the old page writing into it while our pte
3160 * here still points into it and can be read by other
3163 * The critical issue is to order this
3164 * page_remove_rmap with the ptp_clear_flush above.
3165 * Those stores are ordered by (if nothing else,)
3166 * the barrier present in the atomic_add_negative
3167 * in page_remove_rmap.
3169 * Then the TLB flush in ptep_clear_flush ensures that
3170 * no process can access the old page before the
3171 * decremented mapcount is visible. And the old page
3172 * cannot be reused until after the decremented
3173 * mapcount is visible. So transitively, TLBs to
3174 * old page will be flushed before it can be reused.
3176 page_remove_rmap(vmf
->page
, vma
, false);
3179 /* Free the old page.. */
3180 new_folio
= old_folio
;
3182 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3183 } else if (vmf
->pte
) {
3184 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
3185 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3188 mmu_notifier_invalidate_range_end(&range
);
3191 folio_put(new_folio
);
3194 free_swap_cache(&old_folio
->page
);
3195 folio_put(old_folio
);
3198 delayacct_wpcopy_end();
3201 folio_put(new_folio
);
3204 folio_put(old_folio
);
3206 delayacct_wpcopy_end();
3207 return VM_FAULT_OOM
;
3211 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3212 * writeable once the page is prepared
3214 * @vmf: structure describing the fault
3216 * This function handles all that is needed to finish a write page fault in a
3217 * shared mapping due to PTE being read-only once the mapped page is prepared.
3218 * It handles locking of PTE and modifying it.
3220 * The function expects the page to be locked or other protection against
3221 * concurrent faults / writeback (such as DAX radix tree locks).
3223 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3224 * we acquired PTE lock.
3226 vm_fault_t
finish_mkwrite_fault(struct vm_fault
*vmf
)
3228 WARN_ON_ONCE(!(vmf
->vma
->vm_flags
& VM_SHARED
));
3229 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3232 return VM_FAULT_NOPAGE
;
3234 * We might have raced with another page fault while we released the
3235 * pte_offset_map_lock.
3237 if (!pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)) {
3238 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
3239 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3240 return VM_FAULT_NOPAGE
;
3247 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3250 static vm_fault_t
wp_pfn_shared(struct vm_fault
*vmf
)
3252 struct vm_area_struct
*vma
= vmf
->vma
;
3254 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
3257 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3258 if (vmf
->flags
& FAULT_FLAG_VMA_LOCK
) {
3259 vma_end_read(vmf
->vma
);
3260 return VM_FAULT_RETRY
;
3263 vmf
->flags
|= FAULT_FLAG_MKWRITE
;
3264 ret
= vma
->vm_ops
->pfn_mkwrite(vmf
);
3265 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))
3267 return finish_mkwrite_fault(vmf
);
3273 static vm_fault_t
wp_page_shared(struct vm_fault
*vmf
, struct folio
*folio
)
3274 __releases(vmf
->ptl
)
3276 struct vm_area_struct
*vma
= vmf
->vma
;
3281 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
3284 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3285 if (vmf
->flags
& FAULT_FLAG_VMA_LOCK
) {
3287 vma_end_read(vmf
->vma
);
3288 return VM_FAULT_RETRY
;
3291 tmp
= do_page_mkwrite(vmf
, folio
);
3292 if (unlikely(!tmp
|| (tmp
&
3293 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3297 tmp
= finish_mkwrite_fault(vmf
);
3298 if (unlikely(tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3299 folio_unlock(folio
);
3307 ret
|= fault_dirty_shared_page(vmf
);
3314 * This routine handles present pages, when
3315 * * users try to write to a shared page (FAULT_FLAG_WRITE)
3316 * * GUP wants to take a R/O pin on a possibly shared anonymous page
3317 * (FAULT_FLAG_UNSHARE)
3319 * It is done by copying the page to a new address and decrementing the
3320 * shared-page counter for the old page.
3322 * Note that this routine assumes that the protection checks have been
3323 * done by the caller (the low-level page fault routine in most cases).
3324 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've
3325 * done any necessary COW.
3327 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even
3328 * though the page will change only once the write actually happens. This
3329 * avoids a few races, and potentially makes it more efficient.
3331 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3332 * but allow concurrent faults), with pte both mapped and locked.
3333 * We return with mmap_lock still held, but pte unmapped and unlocked.
3335 static vm_fault_t
do_wp_page(struct vm_fault
*vmf
)
3336 __releases(vmf
->ptl
)
3338 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
3339 struct vm_area_struct
*vma
= vmf
->vma
;
3340 struct folio
*folio
= NULL
;
3342 if (likely(!unshare
)) {
3343 if (userfaultfd_pte_wp(vma
, ptep_get(vmf
->pte
))) {
3344 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3345 return handle_userfault(vmf
, VM_UFFD_WP
);
3349 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3350 * is flushed in this case before copying.
3352 if (unlikely(userfaultfd_wp(vmf
->vma
) &&
3353 mm_tlb_flush_pending(vmf
->vma
->vm_mm
)))
3354 flush_tlb_page(vmf
->vma
, vmf
->address
);
3357 vmf
->page
= vm_normal_page(vma
, vmf
->address
, vmf
->orig_pte
);
3360 folio
= page_folio(vmf
->page
);
3363 * Shared mapping: we are guaranteed to have VM_WRITE and
3364 * FAULT_FLAG_WRITE set at this point.
3366 if (vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
)) {
3368 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3371 * We should not cow pages in a shared writeable mapping.
3372 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3375 return wp_pfn_shared(vmf
);
3376 return wp_page_shared(vmf
, folio
);
3380 * Private mapping: create an exclusive anonymous page copy if reuse
3381 * is impossible. We might miss VM_WRITE for FOLL_FORCE handling.
3383 if (folio
&& folio_test_anon(folio
)) {
3385 * If the page is exclusive to this process we must reuse the
3386 * page without further checks.
3388 if (PageAnonExclusive(vmf
->page
))
3392 * We have to verify under folio lock: these early checks are
3393 * just an optimization to avoid locking the folio and freeing
3394 * the swapcache if there is little hope that we can reuse.
3396 * KSM doesn't necessarily raise the folio refcount.
3398 if (folio_test_ksm(folio
) || folio_ref_count(folio
) > 3)
3400 if (!folio_test_lru(folio
))
3402 * We cannot easily detect+handle references from
3403 * remote LRU caches or references to LRU folios.
3406 if (folio_ref_count(folio
) > 1 + folio_test_swapcache(folio
))
3408 if (!folio_trylock(folio
))
3410 if (folio_test_swapcache(folio
))
3411 folio_free_swap(folio
);
3412 if (folio_test_ksm(folio
) || folio_ref_count(folio
) != 1) {
3413 folio_unlock(folio
);
3417 * Ok, we've got the only folio reference from our mapping
3418 * and the folio is locked, it's dark out, and we're wearing
3419 * sunglasses. Hit it.
3421 page_move_anon_rmap(vmf
->page
, vma
);
3422 folio_unlock(folio
);
3424 if (unlikely(unshare
)) {
3425 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3432 if ((vmf
->flags
& FAULT_FLAG_VMA_LOCK
) && !vma
->anon_vma
) {
3433 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3434 vma_end_read(vmf
->vma
);
3435 return VM_FAULT_RETRY
;
3439 * Ok, we need to copy. Oh, well..
3444 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3446 if (folio
&& folio_test_ksm(folio
))
3447 count_vm_event(COW_KSM
);
3449 return wp_page_copy(vmf
);
3452 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
3453 unsigned long start_addr
, unsigned long end_addr
,
3454 struct zap_details
*details
)
3456 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
3459 static inline void unmap_mapping_range_tree(struct rb_root_cached
*root
,
3460 pgoff_t first_index
,
3462 struct zap_details
*details
)
3464 struct vm_area_struct
*vma
;
3465 pgoff_t vba
, vea
, zba
, zea
;
3467 vma_interval_tree_foreach(vma
, root
, first_index
, last_index
) {
3468 vba
= vma
->vm_pgoff
;
3469 vea
= vba
+ vma_pages(vma
) - 1;
3470 zba
= max(first_index
, vba
);
3471 zea
= min(last_index
, vea
);
3473 unmap_mapping_range_vma(vma
,
3474 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
3475 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
3481 * unmap_mapping_folio() - Unmap single folio from processes.
3482 * @folio: The locked folio to be unmapped.
3484 * Unmap this folio from any userspace process which still has it mmaped.
3485 * Typically, for efficiency, the range of nearby pages has already been
3486 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3487 * truncation or invalidation holds the lock on a folio, it may find that
3488 * the page has been remapped again: and then uses unmap_mapping_folio()
3489 * to unmap it finally.
3491 void unmap_mapping_folio(struct folio
*folio
)
3493 struct address_space
*mapping
= folio
->mapping
;
3494 struct zap_details details
= { };
3495 pgoff_t first_index
;
3498 VM_BUG_ON(!folio_test_locked(folio
));
3500 first_index
= folio
->index
;
3501 last_index
= folio_next_index(folio
) - 1;
3503 details
.even_cows
= false;
3504 details
.single_folio
= folio
;
3505 details
.zap_flags
= ZAP_FLAG_DROP_MARKER
;
3507 i_mmap_lock_read(mapping
);
3508 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3509 unmap_mapping_range_tree(&mapping
->i_mmap
, first_index
,
3510 last_index
, &details
);
3511 i_mmap_unlock_read(mapping
);
3515 * unmap_mapping_pages() - Unmap pages from processes.
3516 * @mapping: The address space containing pages to be unmapped.
3517 * @start: Index of first page to be unmapped.
3518 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3519 * @even_cows: Whether to unmap even private COWed pages.
3521 * Unmap the pages in this address space from any userspace process which
3522 * has them mmaped. Generally, you want to remove COWed pages as well when
3523 * a file is being truncated, but not when invalidating pages from the page
3526 void unmap_mapping_pages(struct address_space
*mapping
, pgoff_t start
,
3527 pgoff_t nr
, bool even_cows
)
3529 struct zap_details details
= { };
3530 pgoff_t first_index
= start
;
3531 pgoff_t last_index
= start
+ nr
- 1;
3533 details
.even_cows
= even_cows
;
3534 if (last_index
< first_index
)
3535 last_index
= ULONG_MAX
;
3537 i_mmap_lock_read(mapping
);
3538 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3539 unmap_mapping_range_tree(&mapping
->i_mmap
, first_index
,
3540 last_index
, &details
);
3541 i_mmap_unlock_read(mapping
);
3543 EXPORT_SYMBOL_GPL(unmap_mapping_pages
);
3546 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3547 * address_space corresponding to the specified byte range in the underlying
3550 * @mapping: the address space containing mmaps to be unmapped.
3551 * @holebegin: byte in first page to unmap, relative to the start of
3552 * the underlying file. This will be rounded down to a PAGE_SIZE
3553 * boundary. Note that this is different from truncate_pagecache(), which
3554 * must keep the partial page. In contrast, we must get rid of
3556 * @holelen: size of prospective hole in bytes. This will be rounded
3557 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3559 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3560 * but 0 when invalidating pagecache, don't throw away private data.
3562 void unmap_mapping_range(struct address_space
*mapping
,
3563 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
3565 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
3566 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3568 /* Check for overflow. */
3569 if (sizeof(holelen
) > sizeof(hlen
)) {
3571 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3572 if (holeend
& ~(long long)ULONG_MAX
)
3573 hlen
= ULONG_MAX
- hba
+ 1;
3576 unmap_mapping_pages(mapping
, hba
, hlen
, even_cows
);
3578 EXPORT_SYMBOL(unmap_mapping_range
);
3581 * Restore a potential device exclusive pte to a working pte entry
3583 static vm_fault_t
remove_device_exclusive_entry(struct vm_fault
*vmf
)
3585 struct folio
*folio
= page_folio(vmf
->page
);
3586 struct vm_area_struct
*vma
= vmf
->vma
;
3587 struct mmu_notifier_range range
;
3591 * We need a reference to lock the folio because we don't hold
3592 * the PTL so a racing thread can remove the device-exclusive
3593 * entry and unmap it. If the folio is free the entry must
3594 * have been removed already. If it happens to have already
3595 * been re-allocated after being freed all we do is lock and
3598 if (!folio_try_get(folio
))
3601 ret
= folio_lock_or_retry(folio
, vmf
);
3606 mmu_notifier_range_init_owner(&range
, MMU_NOTIFY_EXCLUSIVE
, 0,
3607 vma
->vm_mm
, vmf
->address
& PAGE_MASK
,
3608 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
, NULL
);
3609 mmu_notifier_invalidate_range_start(&range
);
3611 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3613 if (likely(vmf
->pte
&& pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)))
3614 restore_exclusive_pte(vma
, vmf
->page
, vmf
->address
, vmf
->pte
);
3617 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3618 folio_unlock(folio
);
3621 mmu_notifier_invalidate_range_end(&range
);
3625 static inline bool should_try_to_free_swap(struct folio
*folio
,
3626 struct vm_area_struct
*vma
,
3627 unsigned int fault_flags
)
3629 if (!folio_test_swapcache(folio
))
3631 if (mem_cgroup_swap_full(folio
) || (vma
->vm_flags
& VM_LOCKED
) ||
3632 folio_test_mlocked(folio
))
3635 * If we want to map a page that's in the swapcache writable, we
3636 * have to detect via the refcount if we're really the exclusive
3637 * user. Try freeing the swapcache to get rid of the swapcache
3638 * reference only in case it's likely that we'll be the exlusive user.
3640 return (fault_flags
& FAULT_FLAG_WRITE
) && !folio_test_ksm(folio
) &&
3641 folio_ref_count(folio
) == 2;
3644 static vm_fault_t
pte_marker_clear(struct vm_fault
*vmf
)
3646 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
,
3647 vmf
->address
, &vmf
->ptl
);
3651 * Be careful so that we will only recover a special uffd-wp pte into a
3652 * none pte. Otherwise it means the pte could have changed, so retry.
3654 * This should also cover the case where e.g. the pte changed
3655 * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_POISONED.
3656 * So is_pte_marker() check is not enough to safely drop the pte.
3658 if (pte_same(vmf
->orig_pte
, ptep_get(vmf
->pte
)))
3659 pte_clear(vmf
->vma
->vm_mm
, vmf
->address
, vmf
->pte
);
3660 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3664 static vm_fault_t
do_pte_missing(struct vm_fault
*vmf
)
3666 if (vma_is_anonymous(vmf
->vma
))
3667 return do_anonymous_page(vmf
);
3669 return do_fault(vmf
);
3673 * This is actually a page-missing access, but with uffd-wp special pte
3674 * installed. It means this pte was wr-protected before being unmapped.
3676 static vm_fault_t
pte_marker_handle_uffd_wp(struct vm_fault
*vmf
)
3679 * Just in case there're leftover special ptes even after the region
3680 * got unregistered - we can simply clear them.
3682 if (unlikely(!userfaultfd_wp(vmf
->vma
)))
3683 return pte_marker_clear(vmf
);
3685 return do_pte_missing(vmf
);
3688 static vm_fault_t
handle_pte_marker(struct vm_fault
*vmf
)
3690 swp_entry_t entry
= pte_to_swp_entry(vmf
->orig_pte
);
3691 unsigned long marker
= pte_marker_get(entry
);
3694 * PTE markers should never be empty. If anything weird happened,
3695 * the best thing to do is to kill the process along with its mm.
3697 if (WARN_ON_ONCE(!marker
))
3698 return VM_FAULT_SIGBUS
;
3700 /* Higher priority than uffd-wp when data corrupted */
3701 if (marker
& PTE_MARKER_POISONED
)
3702 return VM_FAULT_HWPOISON
;
3704 if (pte_marker_entry_uffd_wp(entry
))
3705 return pte_marker_handle_uffd_wp(vmf
);
3707 /* This is an unknown pte marker */
3708 return VM_FAULT_SIGBUS
;
3712 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3713 * but allow concurrent faults), and pte mapped but not yet locked.
3714 * We return with pte unmapped and unlocked.
3716 * We return with the mmap_lock locked or unlocked in the same cases
3717 * as does filemap_fault().
3719 vm_fault_t
do_swap_page(struct vm_fault
*vmf
)
3721 struct vm_area_struct
*vma
= vmf
->vma
;
3722 struct folio
*swapcache
, *folio
= NULL
;
3724 struct swap_info_struct
*si
= NULL
;
3725 rmap_t rmap_flags
= RMAP_NONE
;
3726 bool exclusive
= false;
3730 void *shadow
= NULL
;
3732 if (!pte_unmap_same(vmf
))
3735 entry
= pte_to_swp_entry(vmf
->orig_pte
);
3736 if (unlikely(non_swap_entry(entry
))) {
3737 if (is_migration_entry(entry
)) {
3738 migration_entry_wait(vma
->vm_mm
, vmf
->pmd
,
3740 } else if (is_device_exclusive_entry(entry
)) {
3741 vmf
->page
= pfn_swap_entry_to_page(entry
);
3742 ret
= remove_device_exclusive_entry(vmf
);
3743 } else if (is_device_private_entry(entry
)) {
3744 if (vmf
->flags
& FAULT_FLAG_VMA_LOCK
) {
3746 * migrate_to_ram is not yet ready to operate
3750 ret
= VM_FAULT_RETRY
;
3754 vmf
->page
= pfn_swap_entry_to_page(entry
);
3755 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3756 vmf
->address
, &vmf
->ptl
);
3757 if (unlikely(!vmf
->pte
||
3758 !pte_same(ptep_get(vmf
->pte
),
3763 * Get a page reference while we know the page can't be
3766 get_page(vmf
->page
);
3767 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3768 ret
= vmf
->page
->pgmap
->ops
->migrate_to_ram(vmf
);
3769 put_page(vmf
->page
);
3770 } else if (is_hwpoison_entry(entry
)) {
3771 ret
= VM_FAULT_HWPOISON
;
3772 } else if (is_pte_marker_entry(entry
)) {
3773 ret
= handle_pte_marker(vmf
);
3775 print_bad_pte(vma
, vmf
->address
, vmf
->orig_pte
, NULL
);
3776 ret
= VM_FAULT_SIGBUS
;
3781 /* Prevent swapoff from happening to us. */
3782 si
= get_swap_device(entry
);
3786 folio
= swap_cache_get_folio(entry
, vma
, vmf
->address
);
3788 page
= folio_file_page(folio
, swp_offset(entry
));
3792 if (data_race(si
->flags
& SWP_SYNCHRONOUS_IO
) &&
3793 __swap_count(entry
) == 1) {
3794 /* skip swapcache */
3795 folio
= vma_alloc_folio(GFP_HIGHUSER_MOVABLE
, 0,
3796 vma
, vmf
->address
, false);
3797 page
= &folio
->page
;
3799 __folio_set_locked(folio
);
3800 __folio_set_swapbacked(folio
);
3802 if (mem_cgroup_swapin_charge_folio(folio
,
3803 vma
->vm_mm
, GFP_KERNEL
,
3808 mem_cgroup_swapin_uncharge_swap(entry
);
3810 shadow
= get_shadow_from_swap_cache(entry
);
3812 workingset_refault(folio
, shadow
);
3814 folio_add_lru(folio
);
3816 /* To provide entry to swap_readpage() */
3817 folio
->swap
= entry
;
3818 swap_readpage(page
, true, NULL
);
3819 folio
->private = NULL
;
3822 page
= swapin_readahead(entry
, GFP_HIGHUSER_MOVABLE
,
3825 folio
= page_folio(page
);
3831 * Back out if somebody else faulted in this pte
3832 * while we released the pte lock.
3834 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3835 vmf
->address
, &vmf
->ptl
);
3836 if (likely(vmf
->pte
&&
3837 pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)))
3842 /* Had to read the page from swap area: Major fault */
3843 ret
= VM_FAULT_MAJOR
;
3844 count_vm_event(PGMAJFAULT
);
3845 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
3846 } else if (PageHWPoison(page
)) {
3848 * hwpoisoned dirty swapcache pages are kept for killing
3849 * owner processes (which may be unknown at hwpoison time)
3851 ret
= VM_FAULT_HWPOISON
;
3855 ret
|= folio_lock_or_retry(folio
, vmf
);
3856 if (ret
& VM_FAULT_RETRY
)
3861 * Make sure folio_free_swap() or swapoff did not release the
3862 * swapcache from under us. The page pin, and pte_same test
3863 * below, are not enough to exclude that. Even if it is still
3864 * swapcache, we need to check that the page's swap has not
3867 if (unlikely(!folio_test_swapcache(folio
) ||
3868 page_swap_entry(page
).val
!= entry
.val
))
3872 * KSM sometimes has to copy on read faults, for example, if
3873 * page->index of !PageKSM() pages would be nonlinear inside the
3874 * anon VMA -- PageKSM() is lost on actual swapout.
3876 page
= ksm_might_need_to_copy(page
, vma
, vmf
->address
);
3877 if (unlikely(!page
)) {
3880 } else if (unlikely(PTR_ERR(page
) == -EHWPOISON
)) {
3881 ret
= VM_FAULT_HWPOISON
;
3884 folio
= page_folio(page
);
3887 * If we want to map a page that's in the swapcache writable, we
3888 * have to detect via the refcount if we're really the exclusive
3889 * owner. Try removing the extra reference from the local LRU
3890 * caches if required.
3892 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && folio
== swapcache
&&
3893 !folio_test_ksm(folio
) && !folio_test_lru(folio
))
3897 folio_throttle_swaprate(folio
, GFP_KERNEL
);
3900 * Back out if somebody else already faulted in this pte.
3902 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3904 if (unlikely(!vmf
->pte
|| !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)))
3907 if (unlikely(!folio_test_uptodate(folio
))) {
3908 ret
= VM_FAULT_SIGBUS
;
3913 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte
3914 * must never point at an anonymous page in the swapcache that is
3915 * PG_anon_exclusive. Sanity check that this holds and especially, that
3916 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity
3917 * check after taking the PT lock and making sure that nobody
3918 * concurrently faulted in this page and set PG_anon_exclusive.
3920 BUG_ON(!folio_test_anon(folio
) && folio_test_mappedtodisk(folio
));
3921 BUG_ON(folio_test_anon(folio
) && PageAnonExclusive(page
));
3924 * Check under PT lock (to protect against concurrent fork() sharing
3925 * the swap entry concurrently) for certainly exclusive pages.
3927 if (!folio_test_ksm(folio
)) {
3928 exclusive
= pte_swp_exclusive(vmf
->orig_pte
);
3929 if (folio
!= swapcache
) {
3931 * We have a fresh page that is not exposed to the
3932 * swapcache -> certainly exclusive.
3935 } else if (exclusive
&& folio_test_writeback(folio
) &&
3936 data_race(si
->flags
& SWP_STABLE_WRITES
)) {
3938 * This is tricky: not all swap backends support
3939 * concurrent page modifications while under writeback.
3941 * So if we stumble over such a page in the swapcache
3942 * we must not set the page exclusive, otherwise we can
3943 * map it writable without further checks and modify it
3944 * while still under writeback.
3946 * For these problematic swap backends, simply drop the
3947 * exclusive marker: this is perfectly fine as we start
3948 * writeback only if we fully unmapped the page and
3949 * there are no unexpected references on the page after
3950 * unmapping succeeded. After fully unmapped, no
3951 * further GUP references (FOLL_GET and FOLL_PIN) can
3952 * appear, so dropping the exclusive marker and mapping
3953 * it only R/O is fine.
3960 * Some architectures may have to restore extra metadata to the page
3961 * when reading from swap. This metadata may be indexed by swap entry
3962 * so this must be called before swap_free().
3964 arch_swap_restore(entry
, folio
);
3967 * Remove the swap entry and conditionally try to free up the swapcache.
3968 * We're already holding a reference on the page but haven't mapped it
3972 if (should_try_to_free_swap(folio
, vma
, vmf
->flags
))
3973 folio_free_swap(folio
);
3975 inc_mm_counter(vma
->vm_mm
, MM_ANONPAGES
);
3976 dec_mm_counter(vma
->vm_mm
, MM_SWAPENTS
);
3977 pte
= mk_pte(page
, vma
->vm_page_prot
);
3980 * Same logic as in do_wp_page(); however, optimize for pages that are
3981 * certainly not shared either because we just allocated them without
3982 * exposing them to the swapcache or because the swap entry indicates
3985 if (!folio_test_ksm(folio
) &&
3986 (exclusive
|| folio_ref_count(folio
) == 1)) {
3987 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
3988 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3989 vmf
->flags
&= ~FAULT_FLAG_WRITE
;
3991 rmap_flags
|= RMAP_EXCLUSIVE
;
3993 flush_icache_page(vma
, page
);
3994 if (pte_swp_soft_dirty(vmf
->orig_pte
))
3995 pte
= pte_mksoft_dirty(pte
);
3996 if (pte_swp_uffd_wp(vmf
->orig_pte
))
3997 pte
= pte_mkuffd_wp(pte
);
3998 vmf
->orig_pte
= pte
;
4000 /* ksm created a completely new copy */
4001 if (unlikely(folio
!= swapcache
&& swapcache
)) {
4002 page_add_new_anon_rmap(page
, vma
, vmf
->address
);
4003 folio_add_lru_vma(folio
, vma
);
4005 page_add_anon_rmap(page
, vma
, vmf
->address
, rmap_flags
);
4008 VM_BUG_ON(!folio_test_anon(folio
) ||
4009 (pte_write(pte
) && !PageAnonExclusive(page
)));
4010 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
4011 arch_do_swap_page(vma
->vm_mm
, vma
, vmf
->address
, pte
, vmf
->orig_pte
);
4013 folio_unlock(folio
);
4014 if (folio
!= swapcache
&& swapcache
) {
4016 * Hold the lock to avoid the swap entry to be reused
4017 * until we take the PT lock for the pte_same() check
4018 * (to avoid false positives from pte_same). For
4019 * further safety release the lock after the swap_free
4020 * so that the swap count won't change under a
4021 * parallel locked swapcache.
4023 folio_unlock(swapcache
);
4024 folio_put(swapcache
);
4027 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4028 ret
|= do_wp_page(vmf
);
4029 if (ret
& VM_FAULT_ERROR
)
4030 ret
&= VM_FAULT_ERROR
;
4034 /* No need to invalidate - it was non-present before */
4035 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
4038 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4041 put_swap_device(si
);
4045 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4047 folio_unlock(folio
);
4050 if (folio
!= swapcache
&& swapcache
) {
4051 folio_unlock(swapcache
);
4052 folio_put(swapcache
);
4055 put_swap_device(si
);
4060 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4061 * but allow concurrent faults), and pte mapped but not yet locked.
4062 * We return with mmap_lock still held, but pte unmapped and unlocked.
4064 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
)
4066 bool uffd_wp
= vmf_orig_pte_uffd_wp(vmf
);
4067 struct vm_area_struct
*vma
= vmf
->vma
;
4068 struct folio
*folio
;
4072 /* File mapping without ->vm_ops ? */
4073 if (vma
->vm_flags
& VM_SHARED
)
4074 return VM_FAULT_SIGBUS
;
4077 * Use pte_alloc() instead of pte_alloc_map(), so that OOM can
4078 * be distinguished from a transient failure of pte_offset_map().
4080 if (pte_alloc(vma
->vm_mm
, vmf
->pmd
))
4081 return VM_FAULT_OOM
;
4083 /* Use the zero-page for reads */
4084 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
4085 !mm_forbids_zeropage(vma
->vm_mm
)) {
4086 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(vmf
->address
),
4087 vma
->vm_page_prot
));
4088 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4089 vmf
->address
, &vmf
->ptl
);
4092 if (vmf_pte_changed(vmf
)) {
4093 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
4096 ret
= check_stable_address_space(vma
->vm_mm
);
4099 /* Deliver the page fault to userland, check inside PT lock */
4100 if (userfaultfd_missing(vma
)) {
4101 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4102 return handle_userfault(vmf
, VM_UFFD_MISSING
);
4107 /* Allocate our own private page. */
4108 if (unlikely(anon_vma_prepare(vma
)))
4110 folio
= vma_alloc_zeroed_movable_folio(vma
, vmf
->address
);
4114 if (mem_cgroup_charge(folio
, vma
->vm_mm
, GFP_KERNEL
))
4116 folio_throttle_swaprate(folio
, GFP_KERNEL
);
4119 * The memory barrier inside __folio_mark_uptodate makes sure that
4120 * preceding stores to the page contents become visible before
4121 * the set_pte_at() write.
4123 __folio_mark_uptodate(folio
);
4125 entry
= mk_pte(&folio
->page
, vma
->vm_page_prot
);
4126 entry
= pte_sw_mkyoung(entry
);
4127 if (vma
->vm_flags
& VM_WRITE
)
4128 entry
= pte_mkwrite(pte_mkdirty(entry
), vma
);
4130 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
4134 if (vmf_pte_changed(vmf
)) {
4135 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
4139 ret
= check_stable_address_space(vma
->vm_mm
);
4143 /* Deliver the page fault to userland, check inside PT lock */
4144 if (userfaultfd_missing(vma
)) {
4145 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4147 return handle_userfault(vmf
, VM_UFFD_MISSING
);
4150 inc_mm_counter(vma
->vm_mm
, MM_ANONPAGES
);
4151 folio_add_new_anon_rmap(folio
, vma
, vmf
->address
);
4152 folio_add_lru_vma(folio
, vma
);
4155 entry
= pte_mkuffd_wp(entry
);
4156 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
4158 /* No need to invalidate - it was non-present before */
4159 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
4162 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4170 return VM_FAULT_OOM
;
4174 * The mmap_lock must have been held on entry, and may have been
4175 * released depending on flags and vma->vm_ops->fault() return value.
4176 * See filemap_fault() and __lock_page_retry().
4178 static vm_fault_t
__do_fault(struct vm_fault
*vmf
)
4180 struct vm_area_struct
*vma
= vmf
->vma
;
4184 * Preallocate pte before we take page_lock because this might lead to
4185 * deadlocks for memcg reclaim which waits for pages under writeback:
4187 * SetPageWriteback(A)
4193 * wait_on_page_writeback(A)
4194 * SetPageWriteback(B)
4196 * # flush A, B to clear the writeback
4198 if (pmd_none(*vmf
->pmd
) && !vmf
->prealloc_pte
) {
4199 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
4200 if (!vmf
->prealloc_pte
)
4201 return VM_FAULT_OOM
;
4204 ret
= vma
->vm_ops
->fault(vmf
);
4205 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
|
4206 VM_FAULT_DONE_COW
)))
4209 if (unlikely(PageHWPoison(vmf
->page
))) {
4210 struct page
*page
= vmf
->page
;
4211 vm_fault_t poisonret
= VM_FAULT_HWPOISON
;
4212 if (ret
& VM_FAULT_LOCKED
) {
4213 if (page_mapped(page
))
4214 unmap_mapping_pages(page_mapping(page
),
4215 page
->index
, 1, false);
4216 /* Retry if a clean page was removed from the cache. */
4217 if (invalidate_inode_page(page
))
4218 poisonret
= VM_FAULT_NOPAGE
;
4226 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
4227 lock_page(vmf
->page
);
4229 VM_BUG_ON_PAGE(!PageLocked(vmf
->page
), vmf
->page
);
4234 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4235 static void deposit_prealloc_pte(struct vm_fault
*vmf
)
4237 struct vm_area_struct
*vma
= vmf
->vma
;
4239 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
4241 * We are going to consume the prealloc table,
4242 * count that as nr_ptes.
4244 mm_inc_nr_ptes(vma
->vm_mm
);
4245 vmf
->prealloc_pte
= NULL
;
4248 vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
4250 struct vm_area_struct
*vma
= vmf
->vma
;
4251 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
4252 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
4254 vm_fault_t ret
= VM_FAULT_FALLBACK
;
4256 if (!transhuge_vma_suitable(vma
, haddr
))
4259 page
= compound_head(page
);
4260 if (compound_order(page
) != HPAGE_PMD_ORDER
)
4264 * Just backoff if any subpage of a THP is corrupted otherwise
4265 * the corrupted page may mapped by PMD silently to escape the
4266 * check. This kind of THP just can be PTE mapped. Access to
4267 * the corrupted subpage should trigger SIGBUS as expected.
4269 if (unlikely(PageHasHWPoisoned(page
)))
4273 * Archs like ppc64 need additional space to store information
4274 * related to pte entry. Use the preallocated table for that.
4276 if (arch_needs_pgtable_deposit() && !vmf
->prealloc_pte
) {
4277 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
4278 if (!vmf
->prealloc_pte
)
4279 return VM_FAULT_OOM
;
4282 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
4283 if (unlikely(!pmd_none(*vmf
->pmd
)))
4286 flush_icache_pages(vma
, page
, HPAGE_PMD_NR
);
4288 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
4290 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
4292 add_mm_counter(vma
->vm_mm
, mm_counter_file(page
), HPAGE_PMD_NR
);
4293 page_add_file_rmap(page
, vma
, true);
4296 * deposit and withdraw with pmd lock held
4298 if (arch_needs_pgtable_deposit())
4299 deposit_prealloc_pte(vmf
);
4301 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
4303 update_mmu_cache_pmd(vma
, haddr
, vmf
->pmd
);
4305 /* fault is handled */
4307 count_vm_event(THP_FILE_MAPPED
);
4309 spin_unlock(vmf
->ptl
);
4313 vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
4315 return VM_FAULT_FALLBACK
;
4320 * set_pte_range - Set a range of PTEs to point to pages in a folio.
4321 * @vmf: Fault decription.
4322 * @folio: The folio that contains @page.
4323 * @page: The first page to create a PTE for.
4324 * @nr: The number of PTEs to create.
4325 * @addr: The first address to create a PTE for.
4327 void set_pte_range(struct vm_fault
*vmf
, struct folio
*folio
,
4328 struct page
*page
, unsigned int nr
, unsigned long addr
)
4330 struct vm_area_struct
*vma
= vmf
->vma
;
4331 bool uffd_wp
= vmf_orig_pte_uffd_wp(vmf
);
4332 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
4333 bool prefault
= in_range(vmf
->address
, addr
, nr
* PAGE_SIZE
);
4336 flush_icache_pages(vma
, page
, nr
);
4337 entry
= mk_pte(page
, vma
->vm_page_prot
);
4339 if (prefault
&& arch_wants_old_prefaulted_pte())
4340 entry
= pte_mkold(entry
);
4342 entry
= pte_sw_mkyoung(entry
);
4345 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
4346 if (unlikely(uffd_wp
))
4347 entry
= pte_mkuffd_wp(entry
);
4348 /* copy-on-write page */
4349 if (write
&& !(vma
->vm_flags
& VM_SHARED
)) {
4350 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, nr
);
4351 VM_BUG_ON_FOLIO(nr
!= 1, folio
);
4352 folio_add_new_anon_rmap(folio
, vma
, addr
);
4353 folio_add_lru_vma(folio
, vma
);
4355 add_mm_counter(vma
->vm_mm
, mm_counter_file(page
), nr
);
4356 folio_add_file_rmap_range(folio
, page
, nr
, vma
, false);
4358 set_ptes(vma
->vm_mm
, addr
, vmf
->pte
, entry
, nr
);
4360 /* no need to invalidate: a not-present page won't be cached */
4361 update_mmu_cache_range(vmf
, vma
, addr
, vmf
->pte
, nr
);
4364 static bool vmf_pte_changed(struct vm_fault
*vmf
)
4366 if (vmf
->flags
& FAULT_FLAG_ORIG_PTE_VALID
)
4367 return !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
);
4369 return !pte_none(ptep_get(vmf
->pte
));
4373 * finish_fault - finish page fault once we have prepared the page to fault
4375 * @vmf: structure describing the fault
4377 * This function handles all that is needed to finish a page fault once the
4378 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4379 * given page, adds reverse page mapping, handles memcg charges and LRU
4382 * The function expects the page to be locked and on success it consumes a
4383 * reference of a page being mapped (for the PTE which maps it).
4385 * Return: %0 on success, %VM_FAULT_ code in case of error.
4387 vm_fault_t
finish_fault(struct vm_fault
*vmf
)
4389 struct vm_area_struct
*vma
= vmf
->vma
;
4393 /* Did we COW the page? */
4394 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
4395 page
= vmf
->cow_page
;
4400 * check even for read faults because we might have lost our CoWed
4403 if (!(vma
->vm_flags
& VM_SHARED
)) {
4404 ret
= check_stable_address_space(vma
->vm_mm
);
4409 if (pmd_none(*vmf
->pmd
)) {
4410 if (PageTransCompound(page
)) {
4411 ret
= do_set_pmd(vmf
, page
);
4412 if (ret
!= VM_FAULT_FALLBACK
)
4416 if (vmf
->prealloc_pte
)
4417 pmd_install(vma
->vm_mm
, vmf
->pmd
, &vmf
->prealloc_pte
);
4418 else if (unlikely(pte_alloc(vma
->vm_mm
, vmf
->pmd
)))
4419 return VM_FAULT_OOM
;
4422 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4423 vmf
->address
, &vmf
->ptl
);
4425 return VM_FAULT_NOPAGE
;
4427 /* Re-check under ptl */
4428 if (likely(!vmf_pte_changed(vmf
))) {
4429 struct folio
*folio
= page_folio(page
);
4431 set_pte_range(vmf
, folio
, page
, 1, vmf
->address
);
4434 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
4435 ret
= VM_FAULT_NOPAGE
;
4438 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4442 static unsigned long fault_around_pages __read_mostly
=
4443 65536 >> PAGE_SHIFT
;
4445 #ifdef CONFIG_DEBUG_FS
4446 static int fault_around_bytes_get(void *data
, u64
*val
)
4448 *val
= fault_around_pages
<< PAGE_SHIFT
;
4453 * fault_around_bytes must be rounded down to the nearest page order as it's
4454 * what do_fault_around() expects to see.
4456 static int fault_around_bytes_set(void *data
, u64 val
)
4458 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
4462 * The minimum value is 1 page, however this results in no fault-around
4463 * at all. See should_fault_around().
4465 fault_around_pages
= max(rounddown_pow_of_two(val
) >> PAGE_SHIFT
, 1UL);
4469 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops
,
4470 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
4472 static int __init
fault_around_debugfs(void)
4474 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL
, NULL
,
4475 &fault_around_bytes_fops
);
4478 late_initcall(fault_around_debugfs
);
4482 * do_fault_around() tries to map few pages around the fault address. The hope
4483 * is that the pages will be needed soon and this will lower the number of
4486 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4487 * not ready to be mapped: not up-to-date, locked, etc.
4489 * This function doesn't cross VMA or page table boundaries, in order to call
4490 * map_pages() and acquire a PTE lock only once.
4492 * fault_around_pages defines how many pages we'll try to map.
4493 * do_fault_around() expects it to be set to a power of two less than or equal
4496 * The virtual address of the area that we map is naturally aligned to
4497 * fault_around_pages * PAGE_SIZE rounded down to the machine page size
4498 * (and therefore to page order). This way it's easier to guarantee
4499 * that we don't cross page table boundaries.
4501 static vm_fault_t
do_fault_around(struct vm_fault
*vmf
)
4503 pgoff_t nr_pages
= READ_ONCE(fault_around_pages
);
4504 pgoff_t pte_off
= pte_index(vmf
->address
);
4505 /* The page offset of vmf->address within the VMA. */
4506 pgoff_t vma_off
= vmf
->pgoff
- vmf
->vma
->vm_pgoff
;
4507 pgoff_t from_pte
, to_pte
;
4510 /* The PTE offset of the start address, clamped to the VMA. */
4511 from_pte
= max(ALIGN_DOWN(pte_off
, nr_pages
),
4512 pte_off
- min(pte_off
, vma_off
));
4514 /* The PTE offset of the end address, clamped to the VMA and PTE. */
4515 to_pte
= min3(from_pte
+ nr_pages
, (pgoff_t
)PTRS_PER_PTE
,
4516 pte_off
+ vma_pages(vmf
->vma
) - vma_off
) - 1;
4518 if (pmd_none(*vmf
->pmd
)) {
4519 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
);
4520 if (!vmf
->prealloc_pte
)
4521 return VM_FAULT_OOM
;
4525 ret
= vmf
->vma
->vm_ops
->map_pages(vmf
,
4526 vmf
->pgoff
+ from_pte
- pte_off
,
4527 vmf
->pgoff
+ to_pte
- pte_off
);
4533 /* Return true if we should do read fault-around, false otherwise */
4534 static inline bool should_fault_around(struct vm_fault
*vmf
)
4536 /* No ->map_pages? No way to fault around... */
4537 if (!vmf
->vma
->vm_ops
->map_pages
)
4540 if (uffd_disable_fault_around(vmf
->vma
))
4543 /* A single page implies no faulting 'around' at all. */
4544 return fault_around_pages
> 1;
4547 static vm_fault_t
do_read_fault(struct vm_fault
*vmf
)
4550 struct folio
*folio
;
4553 * Let's call ->map_pages() first and use ->fault() as fallback
4554 * if page by the offset is not ready to be mapped (cold cache or
4557 if (should_fault_around(vmf
)) {
4558 ret
= do_fault_around(vmf
);
4563 if (vmf
->flags
& FAULT_FLAG_VMA_LOCK
) {
4564 vma_end_read(vmf
->vma
);
4565 return VM_FAULT_RETRY
;
4568 ret
= __do_fault(vmf
);
4569 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4572 ret
|= finish_fault(vmf
);
4573 folio
= page_folio(vmf
->page
);
4574 folio_unlock(folio
);
4575 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4580 static vm_fault_t
do_cow_fault(struct vm_fault
*vmf
)
4582 struct vm_area_struct
*vma
= vmf
->vma
;
4585 if (vmf
->flags
& FAULT_FLAG_VMA_LOCK
) {
4587 return VM_FAULT_RETRY
;
4590 if (unlikely(anon_vma_prepare(vma
)))
4591 return VM_FAULT_OOM
;
4593 vmf
->cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, vmf
->address
);
4595 return VM_FAULT_OOM
;
4597 if (mem_cgroup_charge(page_folio(vmf
->cow_page
), vma
->vm_mm
,
4599 put_page(vmf
->cow_page
);
4600 return VM_FAULT_OOM
;
4602 folio_throttle_swaprate(page_folio(vmf
->cow_page
), GFP_KERNEL
);
4604 ret
= __do_fault(vmf
);
4605 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4607 if (ret
& VM_FAULT_DONE_COW
)
4610 copy_user_highpage(vmf
->cow_page
, vmf
->page
, vmf
->address
, vma
);
4611 __SetPageUptodate(vmf
->cow_page
);
4613 ret
|= finish_fault(vmf
);
4614 unlock_page(vmf
->page
);
4615 put_page(vmf
->page
);
4616 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4620 put_page(vmf
->cow_page
);
4624 static vm_fault_t
do_shared_fault(struct vm_fault
*vmf
)
4626 struct vm_area_struct
*vma
= vmf
->vma
;
4627 vm_fault_t ret
, tmp
;
4628 struct folio
*folio
;
4630 if (vmf
->flags
& FAULT_FLAG_VMA_LOCK
) {
4632 return VM_FAULT_RETRY
;
4635 ret
= __do_fault(vmf
);
4636 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4639 folio
= page_folio(vmf
->page
);
4642 * Check if the backing address space wants to know that the page is
4643 * about to become writable
4645 if (vma
->vm_ops
->page_mkwrite
) {
4646 folio_unlock(folio
);
4647 tmp
= do_page_mkwrite(vmf
, folio
);
4648 if (unlikely(!tmp
||
4649 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
4655 ret
|= finish_fault(vmf
);
4656 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
4658 folio_unlock(folio
);
4663 ret
|= fault_dirty_shared_page(vmf
);
4668 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4669 * but allow concurrent faults).
4670 * The mmap_lock may have been released depending on flags and our
4671 * return value. See filemap_fault() and __folio_lock_or_retry().
4672 * If mmap_lock is released, vma may become invalid (for example
4673 * by other thread calling munmap()).
4675 static vm_fault_t
do_fault(struct vm_fault
*vmf
)
4677 struct vm_area_struct
*vma
= vmf
->vma
;
4678 struct mm_struct
*vm_mm
= vma
->vm_mm
;
4682 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4684 if (!vma
->vm_ops
->fault
) {
4685 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
,
4686 vmf
->address
, &vmf
->ptl
);
4687 if (unlikely(!vmf
->pte
))
4688 ret
= VM_FAULT_SIGBUS
;
4691 * Make sure this is not a temporary clearing of pte
4692 * by holding ptl and checking again. A R/M/W update
4693 * of pte involves: take ptl, clearing the pte so that
4694 * we don't have concurrent modification by hardware
4695 * followed by an update.
4697 if (unlikely(pte_none(ptep_get(vmf
->pte
))))
4698 ret
= VM_FAULT_SIGBUS
;
4700 ret
= VM_FAULT_NOPAGE
;
4702 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4704 } else if (!(vmf
->flags
& FAULT_FLAG_WRITE
))
4705 ret
= do_read_fault(vmf
);
4706 else if (!(vma
->vm_flags
& VM_SHARED
))
4707 ret
= do_cow_fault(vmf
);
4709 ret
= do_shared_fault(vmf
);
4711 /* preallocated pagetable is unused: free it */
4712 if (vmf
->prealloc_pte
) {
4713 pte_free(vm_mm
, vmf
->prealloc_pte
);
4714 vmf
->prealloc_pte
= NULL
;
4719 int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
4720 unsigned long addr
, int page_nid
, int *flags
)
4724 /* Record the current PID acceesing VMA */
4725 vma_set_access_pid_bit(vma
);
4727 count_vm_numa_event(NUMA_HINT_FAULTS
);
4728 if (page_nid
== numa_node_id()) {
4729 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
4730 *flags
|= TNF_FAULT_LOCAL
;
4733 return mpol_misplaced(page
, vma
, addr
);
4736 static vm_fault_t
do_numa_page(struct vm_fault
*vmf
)
4738 struct vm_area_struct
*vma
= vmf
->vma
;
4739 struct page
*page
= NULL
;
4740 int page_nid
= NUMA_NO_NODE
;
4741 bool writable
= false;
4748 * The "pte" at this point cannot be used safely without
4749 * validation through pte_unmap_same(). It's of NUMA type but
4750 * the pfn may be screwed if the read is non atomic.
4752 spin_lock(vmf
->ptl
);
4753 if (unlikely(!pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
4754 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4758 /* Get the normal PTE */
4759 old_pte
= ptep_get(vmf
->pte
);
4760 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
4763 * Detect now whether the PTE could be writable; this information
4764 * is only valid while holding the PT lock.
4766 writable
= pte_write(pte
);
4767 if (!writable
&& vma_wants_manual_pte_write_upgrade(vma
) &&
4768 can_change_pte_writable(vma
, vmf
->address
, pte
))
4771 page
= vm_normal_page(vma
, vmf
->address
, pte
);
4772 if (!page
|| is_zone_device_page(page
))
4775 /* TODO: handle PTE-mapped THP */
4776 if (PageCompound(page
))
4780 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4781 * much anyway since they can be in shared cache state. This misses
4782 * the case where a mapping is writable but the process never writes
4783 * to it but pte_write gets cleared during protection updates and
4784 * pte_dirty has unpredictable behaviour between PTE scan updates,
4785 * background writeback, dirty balancing and application behaviour.
4788 flags
|= TNF_NO_GROUP
;
4791 * Flag if the page is shared between multiple address spaces. This
4792 * is later used when determining whether to group tasks together
4794 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
4795 flags
|= TNF_SHARED
;
4797 page_nid
= page_to_nid(page
);
4799 * For memory tiering mode, cpupid of slow memory page is used
4800 * to record page access time. So use default value.
4802 if ((sysctl_numa_balancing_mode
& NUMA_BALANCING_MEMORY_TIERING
) &&
4803 !node_is_toptier(page_nid
))
4804 last_cpupid
= (-1 & LAST_CPUPID_MASK
);
4806 last_cpupid
= page_cpupid_last(page
);
4807 target_nid
= numa_migrate_prep(page
, vma
, vmf
->address
, page_nid
,
4809 if (target_nid
== NUMA_NO_NODE
) {
4813 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4816 /* Migrate to the requested node */
4817 if (migrate_misplaced_page(page
, vma
, target_nid
)) {
4818 page_nid
= target_nid
;
4819 flags
|= TNF_MIGRATED
;
4821 flags
|= TNF_MIGRATE_FAIL
;
4822 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4823 vmf
->address
, &vmf
->ptl
);
4824 if (unlikely(!vmf
->pte
))
4826 if (unlikely(!pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
4827 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4834 if (page_nid
!= NUMA_NO_NODE
)
4835 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
4839 * Make it present again, depending on how arch implements
4840 * non-accessible ptes, some can allow access by kernel mode.
4842 old_pte
= ptep_modify_prot_start(vma
, vmf
->address
, vmf
->pte
);
4843 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
4844 pte
= pte_mkyoung(pte
);
4846 pte
= pte_mkwrite(pte
, vma
);
4847 ptep_modify_prot_commit(vma
, vmf
->address
, vmf
->pte
, old_pte
, pte
);
4848 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
4849 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4853 static inline vm_fault_t
create_huge_pmd(struct vm_fault
*vmf
)
4855 struct vm_area_struct
*vma
= vmf
->vma
;
4856 if (vma_is_anonymous(vma
))
4857 return do_huge_pmd_anonymous_page(vmf
);
4858 if (vma
->vm_ops
->huge_fault
)
4859 return vma
->vm_ops
->huge_fault(vmf
, PMD_ORDER
);
4860 return VM_FAULT_FALLBACK
;
4863 /* `inline' is required to avoid gcc 4.1.2 build error */
4864 static inline vm_fault_t
wp_huge_pmd(struct vm_fault
*vmf
)
4866 struct vm_area_struct
*vma
= vmf
->vma
;
4867 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
4870 if (vma_is_anonymous(vma
)) {
4871 if (likely(!unshare
) &&
4872 userfaultfd_huge_pmd_wp(vma
, vmf
->orig_pmd
))
4873 return handle_userfault(vmf
, VM_UFFD_WP
);
4874 return do_huge_pmd_wp_page(vmf
);
4877 if (vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
)) {
4878 if (vma
->vm_ops
->huge_fault
) {
4879 ret
= vma
->vm_ops
->huge_fault(vmf
, PMD_ORDER
);
4880 if (!(ret
& VM_FAULT_FALLBACK
))
4885 /* COW or write-notify handled on pte level: split pmd. */
4886 __split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
4888 return VM_FAULT_FALLBACK
;
4891 static vm_fault_t
create_huge_pud(struct vm_fault
*vmf
)
4893 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4894 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4895 struct vm_area_struct
*vma
= vmf
->vma
;
4896 /* No support for anonymous transparent PUD pages yet */
4897 if (vma_is_anonymous(vma
))
4898 return VM_FAULT_FALLBACK
;
4899 if (vma
->vm_ops
->huge_fault
)
4900 return vma
->vm_ops
->huge_fault(vmf
, PUD_ORDER
);
4901 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4902 return VM_FAULT_FALLBACK
;
4905 static vm_fault_t
wp_huge_pud(struct vm_fault
*vmf
, pud_t orig_pud
)
4907 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4908 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4909 struct vm_area_struct
*vma
= vmf
->vma
;
4912 /* No support for anonymous transparent PUD pages yet */
4913 if (vma_is_anonymous(vma
))
4915 if (vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
)) {
4916 if (vma
->vm_ops
->huge_fault
) {
4917 ret
= vma
->vm_ops
->huge_fault(vmf
, PUD_ORDER
);
4918 if (!(ret
& VM_FAULT_FALLBACK
))
4923 /* COW or write-notify not handled on PUD level: split pud.*/
4924 __split_huge_pud(vma
, vmf
->pud
, vmf
->address
);
4925 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
4926 return VM_FAULT_FALLBACK
;
4930 * These routines also need to handle stuff like marking pages dirty
4931 * and/or accessed for architectures that don't do it in hardware (most
4932 * RISC architectures). The early dirtying is also good on the i386.
4934 * There is also a hook called "update_mmu_cache()" that architectures
4935 * with external mmu caches can use to update those (ie the Sparc or
4936 * PowerPC hashed page tables that act as extended TLBs).
4938 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4939 * concurrent faults).
4941 * The mmap_lock may have been released depending on flags and our return value.
4942 * See filemap_fault() and __folio_lock_or_retry().
4944 static vm_fault_t
handle_pte_fault(struct vm_fault
*vmf
)
4948 if (unlikely(pmd_none(*vmf
->pmd
))) {
4950 * Leave __pte_alloc() until later: because vm_ops->fault may
4951 * want to allocate huge page, and if we expose page table
4952 * for an instant, it will be difficult to retract from
4953 * concurrent faults and from rmap lookups.
4956 vmf
->flags
&= ~FAULT_FLAG_ORIG_PTE_VALID
;
4959 * A regular pmd is established and it can't morph into a huge
4960 * pmd by anon khugepaged, since that takes mmap_lock in write
4961 * mode; but shmem or file collapse to THP could still morph
4962 * it into a huge pmd: just retry later if so.
4964 vmf
->pte
= pte_offset_map_nolock(vmf
->vma
->vm_mm
, vmf
->pmd
,
4965 vmf
->address
, &vmf
->ptl
);
4966 if (unlikely(!vmf
->pte
))
4968 vmf
->orig_pte
= ptep_get_lockless(vmf
->pte
);
4969 vmf
->flags
|= FAULT_FLAG_ORIG_PTE_VALID
;
4971 if (pte_none(vmf
->orig_pte
)) {
4972 pte_unmap(vmf
->pte
);
4978 return do_pte_missing(vmf
);
4980 if (!pte_present(vmf
->orig_pte
))
4981 return do_swap_page(vmf
);
4983 if (pte_protnone(vmf
->orig_pte
) && vma_is_accessible(vmf
->vma
))
4984 return do_numa_page(vmf
);
4986 spin_lock(vmf
->ptl
);
4987 entry
= vmf
->orig_pte
;
4988 if (unlikely(!pte_same(ptep_get(vmf
->pte
), entry
))) {
4989 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
4992 if (vmf
->flags
& (FAULT_FLAG_WRITE
|FAULT_FLAG_UNSHARE
)) {
4993 if (!pte_write(entry
))
4994 return do_wp_page(vmf
);
4995 else if (likely(vmf
->flags
& FAULT_FLAG_WRITE
))
4996 entry
= pte_mkdirty(entry
);
4998 entry
= pte_mkyoung(entry
);
4999 if (ptep_set_access_flags(vmf
->vma
, vmf
->address
, vmf
->pte
, entry
,
5000 vmf
->flags
& FAULT_FLAG_WRITE
)) {
5001 update_mmu_cache_range(vmf
, vmf
->vma
, vmf
->address
,
5004 /* Skip spurious TLB flush for retried page fault */
5005 if (vmf
->flags
& FAULT_FLAG_TRIED
)
5008 * This is needed only for protection faults but the arch code
5009 * is not yet telling us if this is a protection fault or not.
5010 * This still avoids useless tlb flushes for .text page faults
5013 if (vmf
->flags
& FAULT_FLAG_WRITE
)
5014 flush_tlb_fix_spurious_fault(vmf
->vma
, vmf
->address
,
5018 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5023 * On entry, we hold either the VMA lock or the mmap_lock
5024 * (FAULT_FLAG_VMA_LOCK tells you which). If VM_FAULT_RETRY is set in
5025 * the result, the mmap_lock is not held on exit. See filemap_fault()
5026 * and __folio_lock_or_retry().
5028 static vm_fault_t
__handle_mm_fault(struct vm_area_struct
*vma
,
5029 unsigned long address
, unsigned int flags
)
5031 struct vm_fault vmf
= {
5033 .address
= address
& PAGE_MASK
,
5034 .real_address
= address
,
5036 .pgoff
= linear_page_index(vma
, address
),
5037 .gfp_mask
= __get_fault_gfp_mask(vma
),
5039 struct mm_struct
*mm
= vma
->vm_mm
;
5040 unsigned long vm_flags
= vma
->vm_flags
;
5045 pgd
= pgd_offset(mm
, address
);
5046 p4d
= p4d_alloc(mm
, pgd
, address
);
5048 return VM_FAULT_OOM
;
5050 vmf
.pud
= pud_alloc(mm
, p4d
, address
);
5052 return VM_FAULT_OOM
;
5054 if (pud_none(*vmf
.pud
) &&
5055 hugepage_vma_check(vma
, vm_flags
, false, true, true)) {
5056 ret
= create_huge_pud(&vmf
);
5057 if (!(ret
& VM_FAULT_FALLBACK
))
5060 pud_t orig_pud
= *vmf
.pud
;
5063 if (pud_trans_huge(orig_pud
) || pud_devmap(orig_pud
)) {
5066 * TODO once we support anonymous PUDs: NUMA case and
5067 * FAULT_FLAG_UNSHARE handling.
5069 if ((flags
& FAULT_FLAG_WRITE
) && !pud_write(orig_pud
)) {
5070 ret
= wp_huge_pud(&vmf
, orig_pud
);
5071 if (!(ret
& VM_FAULT_FALLBACK
))
5074 huge_pud_set_accessed(&vmf
, orig_pud
);
5080 vmf
.pmd
= pmd_alloc(mm
, vmf
.pud
, address
);
5082 return VM_FAULT_OOM
;
5084 /* Huge pud page fault raced with pmd_alloc? */
5085 if (pud_trans_unstable(vmf
.pud
))
5088 if (pmd_none(*vmf
.pmd
) &&
5089 hugepage_vma_check(vma
, vm_flags
, false, true, true)) {
5090 ret
= create_huge_pmd(&vmf
);
5091 if (!(ret
& VM_FAULT_FALLBACK
))
5094 vmf
.orig_pmd
= pmdp_get_lockless(vmf
.pmd
);
5096 if (unlikely(is_swap_pmd(vmf
.orig_pmd
))) {
5097 VM_BUG_ON(thp_migration_supported() &&
5098 !is_pmd_migration_entry(vmf
.orig_pmd
));
5099 if (is_pmd_migration_entry(vmf
.orig_pmd
))
5100 pmd_migration_entry_wait(mm
, vmf
.pmd
);
5103 if (pmd_trans_huge(vmf
.orig_pmd
) || pmd_devmap(vmf
.orig_pmd
)) {
5104 if (pmd_protnone(vmf
.orig_pmd
) && vma_is_accessible(vma
))
5105 return do_huge_pmd_numa_page(&vmf
);
5107 if ((flags
& (FAULT_FLAG_WRITE
|FAULT_FLAG_UNSHARE
)) &&
5108 !pmd_write(vmf
.orig_pmd
)) {
5109 ret
= wp_huge_pmd(&vmf
);
5110 if (!(ret
& VM_FAULT_FALLBACK
))
5113 huge_pmd_set_accessed(&vmf
);
5119 return handle_pte_fault(&vmf
);
5123 * mm_account_fault - Do page fault accounting
5124 * @mm: mm from which memcg should be extracted. It can be NULL.
5125 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
5126 * of perf event counters, but we'll still do the per-task accounting to
5127 * the task who triggered this page fault.
5128 * @address: the faulted address.
5129 * @flags: the fault flags.
5130 * @ret: the fault retcode.
5132 * This will take care of most of the page fault accounting. Meanwhile, it
5133 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
5134 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
5135 * still be in per-arch page fault handlers at the entry of page fault.
5137 static inline void mm_account_fault(struct mm_struct
*mm
, struct pt_regs
*regs
,
5138 unsigned long address
, unsigned int flags
,
5143 /* Incomplete faults will be accounted upon completion. */
5144 if (ret
& VM_FAULT_RETRY
)
5148 * To preserve the behavior of older kernels, PGFAULT counters record
5149 * both successful and failed faults, as opposed to perf counters,
5150 * which ignore failed cases.
5152 count_vm_event(PGFAULT
);
5153 count_memcg_event_mm(mm
, PGFAULT
);
5156 * Do not account for unsuccessful faults (e.g. when the address wasn't
5157 * valid). That includes arch_vma_access_permitted() failing before
5158 * reaching here. So this is not a "this many hardware page faults"
5159 * counter. We should use the hw profiling for that.
5161 if (ret
& VM_FAULT_ERROR
)
5165 * We define the fault as a major fault when the final successful fault
5166 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
5167 * handle it immediately previously).
5169 major
= (ret
& VM_FAULT_MAJOR
) || (flags
& FAULT_FLAG_TRIED
);
5177 * If the fault is done for GUP, regs will be NULL. We only do the
5178 * accounting for the per thread fault counters who triggered the
5179 * fault, and we skip the perf event updates.
5185 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ
, 1, regs
, address
);
5187 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN
, 1, regs
, address
);
5190 #ifdef CONFIG_LRU_GEN
5191 static void lru_gen_enter_fault(struct vm_area_struct
*vma
)
5193 /* the LRU algorithm only applies to accesses with recency */
5194 current
->in_lru_fault
= vma_has_recency(vma
);
5197 static void lru_gen_exit_fault(void)
5199 current
->in_lru_fault
= false;
5202 static void lru_gen_enter_fault(struct vm_area_struct
*vma
)
5206 static void lru_gen_exit_fault(void)
5209 #endif /* CONFIG_LRU_GEN */
5211 static vm_fault_t
sanitize_fault_flags(struct vm_area_struct
*vma
,
5212 unsigned int *flags
)
5214 if (unlikely(*flags
& FAULT_FLAG_UNSHARE
)) {
5215 if (WARN_ON_ONCE(*flags
& FAULT_FLAG_WRITE
))
5216 return VM_FAULT_SIGSEGV
;
5218 * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's
5219 * just treat it like an ordinary read-fault otherwise.
5221 if (!is_cow_mapping(vma
->vm_flags
))
5222 *flags
&= ~FAULT_FLAG_UNSHARE
;
5223 } else if (*flags
& FAULT_FLAG_WRITE
) {
5224 /* Write faults on read-only mappings are impossible ... */
5225 if (WARN_ON_ONCE(!(vma
->vm_flags
& VM_MAYWRITE
)))
5226 return VM_FAULT_SIGSEGV
;
5227 /* ... and FOLL_FORCE only applies to COW mappings. */
5228 if (WARN_ON_ONCE(!(vma
->vm_flags
& VM_WRITE
) &&
5229 !is_cow_mapping(vma
->vm_flags
)))
5230 return VM_FAULT_SIGSEGV
;
5232 #ifdef CONFIG_PER_VMA_LOCK
5234 * Per-VMA locks can't be used with FAULT_FLAG_RETRY_NOWAIT because of
5235 * the assumption that lock is dropped on VM_FAULT_RETRY.
5237 if (WARN_ON_ONCE((*flags
&
5238 (FAULT_FLAG_VMA_LOCK
| FAULT_FLAG_RETRY_NOWAIT
)) ==
5239 (FAULT_FLAG_VMA_LOCK
| FAULT_FLAG_RETRY_NOWAIT
)))
5240 return VM_FAULT_SIGSEGV
;
5247 * By the time we get here, we already hold the mm semaphore
5249 * The mmap_lock may have been released depending on flags and our
5250 * return value. See filemap_fault() and __folio_lock_or_retry().
5252 vm_fault_t
handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
5253 unsigned int flags
, struct pt_regs
*regs
)
5255 /* If the fault handler drops the mmap_lock, vma may be freed */
5256 struct mm_struct
*mm
= vma
->vm_mm
;
5259 __set_current_state(TASK_RUNNING
);
5261 ret
= sanitize_fault_flags(vma
, &flags
);
5265 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
5266 flags
& FAULT_FLAG_INSTRUCTION
,
5267 flags
& FAULT_FLAG_REMOTE
)) {
5268 ret
= VM_FAULT_SIGSEGV
;
5273 * Enable the memcg OOM handling for faults triggered in user
5274 * space. Kernel faults are handled more gracefully.
5276 if (flags
& FAULT_FLAG_USER
)
5277 mem_cgroup_enter_user_fault();
5279 lru_gen_enter_fault(vma
);
5281 if (unlikely(is_vm_hugetlb_page(vma
)))
5282 ret
= hugetlb_fault(vma
->vm_mm
, vma
, address
, flags
);
5284 ret
= __handle_mm_fault(vma
, address
, flags
);
5286 lru_gen_exit_fault();
5288 if (flags
& FAULT_FLAG_USER
) {
5289 mem_cgroup_exit_user_fault();
5291 * The task may have entered a memcg OOM situation but
5292 * if the allocation error was handled gracefully (no
5293 * VM_FAULT_OOM), there is no need to kill anything.
5294 * Just clean up the OOM state peacefully.
5296 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
5297 mem_cgroup_oom_synchronize(false);
5300 mm_account_fault(mm
, regs
, address
, flags
, ret
);
5304 EXPORT_SYMBOL_GPL(handle_mm_fault
);
5306 #ifdef CONFIG_LOCK_MM_AND_FIND_VMA
5307 #include <linux/extable.h>
5309 static inline bool get_mmap_lock_carefully(struct mm_struct
*mm
, struct pt_regs
*regs
)
5311 if (likely(mmap_read_trylock(mm
)))
5314 if (regs
&& !user_mode(regs
)) {
5315 unsigned long ip
= instruction_pointer(regs
);
5316 if (!search_exception_tables(ip
))
5320 return !mmap_read_lock_killable(mm
);
5323 static inline bool mmap_upgrade_trylock(struct mm_struct
*mm
)
5326 * We don't have this operation yet.
5328 * It should be easy enough to do: it's basically a
5329 * atomic_long_try_cmpxchg_acquire()
5330 * from RWSEM_READER_BIAS -> RWSEM_WRITER_LOCKED, but
5331 * it also needs the proper lockdep magic etc.
5336 static inline bool upgrade_mmap_lock_carefully(struct mm_struct
*mm
, struct pt_regs
*regs
)
5338 mmap_read_unlock(mm
);
5339 if (regs
&& !user_mode(regs
)) {
5340 unsigned long ip
= instruction_pointer(regs
);
5341 if (!search_exception_tables(ip
))
5344 return !mmap_write_lock_killable(mm
);
5348 * Helper for page fault handling.
5350 * This is kind of equivalend to "mmap_read_lock()" followed
5351 * by "find_extend_vma()", except it's a lot more careful about
5352 * the locking (and will drop the lock on failure).
5354 * For example, if we have a kernel bug that causes a page
5355 * fault, we don't want to just use mmap_read_lock() to get
5356 * the mm lock, because that would deadlock if the bug were
5357 * to happen while we're holding the mm lock for writing.
5359 * So this checks the exception tables on kernel faults in
5360 * order to only do this all for instructions that are actually
5361 * expected to fault.
5363 * We can also actually take the mm lock for writing if we
5364 * need to extend the vma, which helps the VM layer a lot.
5366 struct vm_area_struct
*lock_mm_and_find_vma(struct mm_struct
*mm
,
5367 unsigned long addr
, struct pt_regs
*regs
)
5369 struct vm_area_struct
*vma
;
5371 if (!get_mmap_lock_carefully(mm
, regs
))
5374 vma
= find_vma(mm
, addr
);
5375 if (likely(vma
&& (vma
->vm_start
<= addr
)))
5379 * Well, dang. We might still be successful, but only
5380 * if we can extend a vma to do so.
5382 if (!vma
|| !(vma
->vm_flags
& VM_GROWSDOWN
)) {
5383 mmap_read_unlock(mm
);
5388 * We can try to upgrade the mmap lock atomically,
5389 * in which case we can continue to use the vma
5390 * we already looked up.
5392 * Otherwise we'll have to drop the mmap lock and
5393 * re-take it, and also look up the vma again,
5396 if (!mmap_upgrade_trylock(mm
)) {
5397 if (!upgrade_mmap_lock_carefully(mm
, regs
))
5400 vma
= find_vma(mm
, addr
);
5403 if (vma
->vm_start
<= addr
)
5405 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
5409 if (expand_stack_locked(vma
, addr
))
5413 mmap_write_downgrade(mm
);
5417 mmap_write_unlock(mm
);
5422 #ifdef CONFIG_PER_VMA_LOCK
5424 * Lookup and lock a VMA under RCU protection. Returned VMA is guaranteed to be
5425 * stable and not isolated. If the VMA is not found or is being modified the
5426 * function returns NULL.
5428 struct vm_area_struct
*lock_vma_under_rcu(struct mm_struct
*mm
,
5429 unsigned long address
)
5431 MA_STATE(mas
, &mm
->mm_mt
, address
, address
);
5432 struct vm_area_struct
*vma
;
5436 vma
= mas_walk(&mas
);
5440 if (!vma_start_read(vma
))
5444 * find_mergeable_anon_vma uses adjacent vmas which are not locked.
5445 * This check must happen after vma_start_read(); otherwise, a
5446 * concurrent mremap() with MREMAP_DONTUNMAP could dissociate the VMA
5447 * from its anon_vma.
5449 if (unlikely(vma_is_anonymous(vma
) && !vma
->anon_vma
))
5450 goto inval_end_read
;
5452 /* Check since vm_start/vm_end might change before we lock the VMA */
5453 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
))
5454 goto inval_end_read
;
5456 /* Check if the VMA got isolated after we found it */
5457 if (vma
->detached
) {
5459 count_vm_vma_lock_event(VMA_LOCK_MISS
);
5460 /* The area was replaced with another one */
5471 count_vm_vma_lock_event(VMA_LOCK_ABORT
);
5474 #endif /* CONFIG_PER_VMA_LOCK */
5476 #ifndef __PAGETABLE_P4D_FOLDED
5478 * Allocate p4d page table.
5479 * We've already handled the fast-path in-line.
5481 int __p4d_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
5483 p4d_t
*new = p4d_alloc_one(mm
, address
);
5487 spin_lock(&mm
->page_table_lock
);
5488 if (pgd_present(*pgd
)) { /* Another has populated it */
5491 smp_wmb(); /* See comment in pmd_install() */
5492 pgd_populate(mm
, pgd
, new);
5494 spin_unlock(&mm
->page_table_lock
);
5497 #endif /* __PAGETABLE_P4D_FOLDED */
5499 #ifndef __PAGETABLE_PUD_FOLDED
5501 * Allocate page upper directory.
5502 * We've already handled the fast-path in-line.
5504 int __pud_alloc(struct mm_struct
*mm
, p4d_t
*p4d
, unsigned long address
)
5506 pud_t
*new = pud_alloc_one(mm
, address
);
5510 spin_lock(&mm
->page_table_lock
);
5511 if (!p4d_present(*p4d
)) {
5513 smp_wmb(); /* See comment in pmd_install() */
5514 p4d_populate(mm
, p4d
, new);
5515 } else /* Another has populated it */
5517 spin_unlock(&mm
->page_table_lock
);
5520 #endif /* __PAGETABLE_PUD_FOLDED */
5522 #ifndef __PAGETABLE_PMD_FOLDED
5524 * Allocate page middle directory.
5525 * We've already handled the fast-path in-line.
5527 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
5530 pmd_t
*new = pmd_alloc_one(mm
, address
);
5534 ptl
= pud_lock(mm
, pud
);
5535 if (!pud_present(*pud
)) {
5537 smp_wmb(); /* See comment in pmd_install() */
5538 pud_populate(mm
, pud
, new);
5539 } else { /* Another has populated it */
5545 #endif /* __PAGETABLE_PMD_FOLDED */
5548 * follow_pte - look up PTE at a user virtual address
5549 * @mm: the mm_struct of the target address space
5550 * @address: user virtual address
5551 * @ptepp: location to store found PTE
5552 * @ptlp: location to store the lock for the PTE
5554 * On a successful return, the pointer to the PTE is stored in @ptepp;
5555 * the corresponding lock is taken and its location is stored in @ptlp.
5556 * The contents of the PTE are only stable until @ptlp is released;
5557 * any further use, if any, must be protected against invalidation
5558 * with MMU notifiers.
5560 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
5561 * should be taken for read.
5563 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
5564 * it is not a good general-purpose API.
5566 * Return: zero on success, -ve otherwise.
5568 int follow_pte(struct mm_struct
*mm
, unsigned long address
,
5569 pte_t
**ptepp
, spinlock_t
**ptlp
)
5577 pgd
= pgd_offset(mm
, address
);
5578 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
5581 p4d
= p4d_offset(pgd
, address
);
5582 if (p4d_none(*p4d
) || unlikely(p4d_bad(*p4d
)))
5585 pud
= pud_offset(p4d
, address
);
5586 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
5589 pmd
= pmd_offset(pud
, address
);
5590 VM_BUG_ON(pmd_trans_huge(*pmd
));
5592 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
5595 if (!pte_present(ptep_get(ptep
)))
5600 pte_unmap_unlock(ptep
, *ptlp
);
5604 EXPORT_SYMBOL_GPL(follow_pte
);
5607 * follow_pfn - look up PFN at a user virtual address
5608 * @vma: memory mapping
5609 * @address: user virtual address
5610 * @pfn: location to store found PFN
5612 * Only IO mappings and raw PFN mappings are allowed.
5614 * This function does not allow the caller to read the permissions
5615 * of the PTE. Do not use it.
5617 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5619 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
5626 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
5629 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
5632 *pfn
= pte_pfn(ptep_get(ptep
));
5633 pte_unmap_unlock(ptep
, ptl
);
5636 EXPORT_SYMBOL(follow_pfn
);
5638 #ifdef CONFIG_HAVE_IOREMAP_PROT
5639 int follow_phys(struct vm_area_struct
*vma
,
5640 unsigned long address
, unsigned int flags
,
5641 unsigned long *prot
, resource_size_t
*phys
)
5647 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
5650 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
5652 pte
= ptep_get(ptep
);
5654 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
5657 *prot
= pgprot_val(pte_pgprot(pte
));
5658 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
5662 pte_unmap_unlock(ptep
, ptl
);
5668 * generic_access_phys - generic implementation for iomem mmap access
5669 * @vma: the vma to access
5670 * @addr: userspace address, not relative offset within @vma
5671 * @buf: buffer to read/write
5672 * @len: length of transfer
5673 * @write: set to FOLL_WRITE when writing, otherwise reading
5675 * This is a generic implementation for &vm_operations_struct.access for an
5676 * iomem mapping. This callback is used by access_process_vm() when the @vma is
5679 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
5680 void *buf
, int len
, int write
)
5682 resource_size_t phys_addr
;
5683 unsigned long prot
= 0;
5684 void __iomem
*maddr
;
5687 int offset
= offset_in_page(addr
);
5690 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
5694 if (follow_pte(vma
->vm_mm
, addr
, &ptep
, &ptl
))
5696 pte
= ptep_get(ptep
);
5697 pte_unmap_unlock(ptep
, ptl
);
5699 prot
= pgprot_val(pte_pgprot(pte
));
5700 phys_addr
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
5702 if ((write
& FOLL_WRITE
) && !pte_write(pte
))
5705 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
5709 if (follow_pte(vma
->vm_mm
, addr
, &ptep
, &ptl
))
5712 if (!pte_same(pte
, ptep_get(ptep
))) {
5713 pte_unmap_unlock(ptep
, ptl
);
5720 memcpy_toio(maddr
+ offset
, buf
, len
);
5722 memcpy_fromio(buf
, maddr
+ offset
, len
);
5724 pte_unmap_unlock(ptep
, ptl
);
5730 EXPORT_SYMBOL_GPL(generic_access_phys
);
5734 * Access another process' address space as given in mm.
5736 int __access_remote_vm(struct mm_struct
*mm
, unsigned long addr
, void *buf
,
5737 int len
, unsigned int gup_flags
)
5739 void *old_buf
= buf
;
5740 int write
= gup_flags
& FOLL_WRITE
;
5742 if (mmap_read_lock_killable(mm
))
5745 /* Untag the address before looking up the VMA */
5746 addr
= untagged_addr_remote(mm
, addr
);
5748 /* Avoid triggering the temporary warning in __get_user_pages */
5749 if (!vma_lookup(mm
, addr
) && !expand_stack(mm
, addr
))
5752 /* ignore errors, just check how much was successfully transferred */
5756 struct vm_area_struct
*vma
= NULL
;
5757 struct page
*page
= get_user_page_vma_remote(mm
, addr
,
5760 if (IS_ERR_OR_NULL(page
)) {
5761 /* We might need to expand the stack to access it */
5762 vma
= vma_lookup(mm
, addr
);
5764 vma
= expand_stack(mm
, addr
);
5766 /* mmap_lock was dropped on failure */
5768 return buf
- old_buf
;
5770 /* Try again if stack expansion worked */
5776 * Check if this is a VM_IO | VM_PFNMAP VMA, which
5777 * we can access using slightly different code.
5780 #ifdef CONFIG_HAVE_IOREMAP_PROT
5781 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
5782 bytes
= vma
->vm_ops
->access(vma
, addr
, buf
,
5789 offset
= addr
& (PAGE_SIZE
-1);
5790 if (bytes
> PAGE_SIZE
-offset
)
5791 bytes
= PAGE_SIZE
-offset
;
5795 copy_to_user_page(vma
, page
, addr
,
5796 maddr
+ offset
, buf
, bytes
);
5797 set_page_dirty_lock(page
);
5799 copy_from_user_page(vma
, page
, addr
,
5800 buf
, maddr
+ offset
, bytes
);
5809 mmap_read_unlock(mm
);
5811 return buf
- old_buf
;
5815 * access_remote_vm - access another process' address space
5816 * @mm: the mm_struct of the target address space
5817 * @addr: start address to access
5818 * @buf: source or destination buffer
5819 * @len: number of bytes to transfer
5820 * @gup_flags: flags modifying lookup behaviour
5822 * The caller must hold a reference on @mm.
5824 * Return: number of bytes copied from source to destination.
5826 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
5827 void *buf
, int len
, unsigned int gup_flags
)
5829 return __access_remote_vm(mm
, addr
, buf
, len
, gup_flags
);
5833 * Access another process' address space.
5834 * Source/target buffer must be kernel space,
5835 * Do not walk the page table directly, use get_user_pages
5837 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
5838 void *buf
, int len
, unsigned int gup_flags
)
5840 struct mm_struct
*mm
;
5843 mm
= get_task_mm(tsk
);
5847 ret
= __access_remote_vm(mm
, addr
, buf
, len
, gup_flags
);
5853 EXPORT_SYMBOL_GPL(access_process_vm
);
5856 * Print the name of a VMA.
5858 void print_vma_addr(char *prefix
, unsigned long ip
)
5860 struct mm_struct
*mm
= current
->mm
;
5861 struct vm_area_struct
*vma
;
5864 * we might be running from an atomic context so we cannot sleep
5866 if (!mmap_read_trylock(mm
))
5869 vma
= find_vma(mm
, ip
);
5870 if (vma
&& vma
->vm_file
) {
5871 struct file
*f
= vma
->vm_file
;
5872 char *buf
= (char *)__get_free_page(GFP_NOWAIT
);
5876 p
= file_path(f
, buf
, PAGE_SIZE
);
5879 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
5881 vma
->vm_end
- vma
->vm_start
);
5882 free_page((unsigned long)buf
);
5885 mmap_read_unlock(mm
);
5888 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5889 void __might_fault(const char *file
, int line
)
5891 if (pagefault_disabled())
5893 __might_sleep(file
, line
);
5894 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5896 might_lock_read(¤t
->mm
->mmap_lock
);
5899 EXPORT_SYMBOL(__might_fault
);
5902 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5904 * Process all subpages of the specified huge page with the specified
5905 * operation. The target subpage will be processed last to keep its
5908 static inline int process_huge_page(
5909 unsigned long addr_hint
, unsigned int pages_per_huge_page
,
5910 int (*process_subpage
)(unsigned long addr
, int idx
, void *arg
),
5913 int i
, n
, base
, l
, ret
;
5914 unsigned long addr
= addr_hint
&
5915 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5917 /* Process target subpage last to keep its cache lines hot */
5919 n
= (addr_hint
- addr
) / PAGE_SIZE
;
5920 if (2 * n
<= pages_per_huge_page
) {
5921 /* If target subpage in first half of huge page */
5924 /* Process subpages at the end of huge page */
5925 for (i
= pages_per_huge_page
- 1; i
>= 2 * n
; i
--) {
5927 ret
= process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
5932 /* If target subpage in second half of huge page */
5933 base
= pages_per_huge_page
- 2 * (pages_per_huge_page
- n
);
5934 l
= pages_per_huge_page
- n
;
5935 /* Process subpages at the begin of huge page */
5936 for (i
= 0; i
< base
; i
++) {
5938 ret
= process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
5944 * Process remaining subpages in left-right-left-right pattern
5945 * towards the target subpage
5947 for (i
= 0; i
< l
; i
++) {
5948 int left_idx
= base
+ i
;
5949 int right_idx
= base
+ 2 * l
- 1 - i
;
5952 ret
= process_subpage(addr
+ left_idx
* PAGE_SIZE
, left_idx
, arg
);
5956 ret
= process_subpage(addr
+ right_idx
* PAGE_SIZE
, right_idx
, arg
);
5963 static void clear_gigantic_page(struct page
*page
,
5965 unsigned int pages_per_huge_page
)
5971 for (i
= 0; i
< pages_per_huge_page
; i
++) {
5972 p
= nth_page(page
, i
);
5974 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
5978 static int clear_subpage(unsigned long addr
, int idx
, void *arg
)
5980 struct page
*page
= arg
;
5982 clear_user_highpage(page
+ idx
, addr
);
5986 void clear_huge_page(struct page
*page
,
5987 unsigned long addr_hint
, unsigned int pages_per_huge_page
)
5989 unsigned long addr
= addr_hint
&
5990 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5992 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
5993 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
5997 process_huge_page(addr_hint
, pages_per_huge_page
, clear_subpage
, page
);
6000 static int copy_user_gigantic_page(struct folio
*dst
, struct folio
*src
,
6002 struct vm_area_struct
*vma
,
6003 unsigned int pages_per_huge_page
)
6006 struct page
*dst_page
;
6007 struct page
*src_page
;
6009 for (i
= 0; i
< pages_per_huge_page
; i
++) {
6010 dst_page
= folio_page(dst
, i
);
6011 src_page
= folio_page(src
, i
);
6014 if (copy_mc_user_highpage(dst_page
, src_page
,
6015 addr
+ i
*PAGE_SIZE
, vma
)) {
6016 memory_failure_queue(page_to_pfn(src_page
), 0);
6023 struct copy_subpage_arg
{
6026 struct vm_area_struct
*vma
;
6029 static int copy_subpage(unsigned long addr
, int idx
, void *arg
)
6031 struct copy_subpage_arg
*copy_arg
= arg
;
6033 if (copy_mc_user_highpage(copy_arg
->dst
+ idx
, copy_arg
->src
+ idx
,
6034 addr
, copy_arg
->vma
)) {
6035 memory_failure_queue(page_to_pfn(copy_arg
->src
+ idx
), 0);
6041 int copy_user_large_folio(struct folio
*dst
, struct folio
*src
,
6042 unsigned long addr_hint
, struct vm_area_struct
*vma
)
6044 unsigned int pages_per_huge_page
= folio_nr_pages(dst
);
6045 unsigned long addr
= addr_hint
&
6046 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
6047 struct copy_subpage_arg arg
= {
6053 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
))
6054 return copy_user_gigantic_page(dst
, src
, addr
, vma
,
6055 pages_per_huge_page
);
6057 return process_huge_page(addr_hint
, pages_per_huge_page
, copy_subpage
, &arg
);
6060 long copy_folio_from_user(struct folio
*dst_folio
,
6061 const void __user
*usr_src
,
6062 bool allow_pagefault
)
6065 unsigned long i
, rc
= 0;
6066 unsigned int nr_pages
= folio_nr_pages(dst_folio
);
6067 unsigned long ret_val
= nr_pages
* PAGE_SIZE
;
6068 struct page
*subpage
;
6070 for (i
= 0; i
< nr_pages
; i
++) {
6071 subpage
= folio_page(dst_folio
, i
);
6072 kaddr
= kmap_local_page(subpage
);
6073 if (!allow_pagefault
)
6074 pagefault_disable();
6075 rc
= copy_from_user(kaddr
, usr_src
+ i
* PAGE_SIZE
, PAGE_SIZE
);
6076 if (!allow_pagefault
)
6078 kunmap_local(kaddr
);
6080 ret_val
-= (PAGE_SIZE
- rc
);
6084 flush_dcache_page(subpage
);
6090 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
6092 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
6094 static struct kmem_cache
*page_ptl_cachep
;
6096 void __init
ptlock_cache_init(void)
6098 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
6102 bool ptlock_alloc(struct ptdesc
*ptdesc
)
6106 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
6113 void ptlock_free(struct ptdesc
*ptdesc
)
6115 kmem_cache_free(page_ptl_cachep
, ptdesc
->ptl
);