2 // SPDX-License-Identifier: GPL-2.0-only
6 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
10 * demand-loading started 01.12.91 - seems it is high on the list of
11 * things wanted, and it should be easy to implement. - Linus
15 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
16 * pages started 02.12.91, seems to work. - Linus.
18 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
19 * would have taken more than the 6M I have free, but it worked well as
22 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
26 * Real VM (paging to/from disk) started 18.12.91. Much more work and
27 * thought has to go into this. Oh, well..
28 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
29 * Found it. Everything seems to work now.
30 * 20.12.91 - Ok, making the swap-device changeable like the root.
34 * 05.04.94 - Multi-page memory management added for v1.1.
35 * Idea by Alex Bligh (alex@cconcepts.co.uk)
37 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
38 * (Gerhard.Wichert@pdb.siemens.de)
40 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
43 #include <linux/kernel_stat.h>
45 #include <linux/mm_inline.h>
46 #include <linux/sched/mm.h>
47 #include <linux/sched/coredump.h>
48 #include <linux/sched/numa_balancing.h>
49 #include <linux/sched/task.h>
50 #include <linux/hugetlb.h>
51 #include <linux/mman.h>
52 #include <linux/swap.h>
53 #include <linux/highmem.h>
54 #include <linux/pagemap.h>
55 #include <linux/memremap.h>
56 #include <linux/kmsan.h>
57 #include <linux/ksm.h>
58 #include <linux/rmap.h>
59 #include <linux/export.h>
60 #include <linux/delayacct.h>
61 #include <linux/init.h>
62 #include <linux/pfn_t.h>
63 #include <linux/writeback.h>
64 #include <linux/memcontrol.h>
65 #include <linux/mmu_notifier.h>
66 #include <linux/swapops.h>
67 #include <linux/elf.h>
68 #include <linux/gfp.h>
69 #include <linux/migrate.h>
70 #include <linux/string.h>
71 #include <linux/memory-tiers.h>
72 #include <linux/debugfs.h>
73 #include <linux/userfaultfd_k.h>
74 #include <linux/dax.h>
75 #include <linux/oom.h>
76 #include <linux/numa.h>
77 #include <linux/perf_event.h>
78 #include <linux/ptrace.h>
79 #include <linux/vmalloc.h>
80 #include <linux/sched/sysctl.h>
82 #include <trace/events/kmem.h>
85 #include <asm/mmu_context.h>
86 #include <asm/pgalloc.h>
87 #include <linux/uaccess.h>
89 #include <asm/tlbflush.h>
91 #include "pgalloc-track.h"
95 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
96 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
100 unsigned long max_mapnr
;
101 EXPORT_SYMBOL(max_mapnr
);
103 struct page
*mem_map
;
104 EXPORT_SYMBOL(mem_map
);
107 static vm_fault_t
do_fault(struct vm_fault
*vmf
);
108 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
);
109 static bool vmf_pte_changed(struct vm_fault
*vmf
);
112 * Return true if the original pte was a uffd-wp pte marker (so the pte was
115 static bool vmf_orig_pte_uffd_wp(struct vm_fault
*vmf
)
117 if (!(vmf
->flags
& FAULT_FLAG_ORIG_PTE_VALID
))
120 return pte_marker_uffd_wp(vmf
->orig_pte
);
124 * A number of key systems in x86 including ioremap() rely on the assumption
125 * that high_memory defines the upper bound on direct map memory, then end
129 EXPORT_SYMBOL(high_memory
);
132 * Randomize the address space (stacks, mmaps, brk, etc.).
134 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
135 * as ancient (libc5 based) binaries can segfault. )
137 int randomize_va_space __read_mostly
=
138 #ifdef CONFIG_COMPAT_BRK
144 #ifndef arch_wants_old_prefaulted_pte
145 static inline bool arch_wants_old_prefaulted_pte(void)
148 * Transitioning a PTE from 'old' to 'young' can be expensive on
149 * some architectures, even if it's performed in hardware. By
150 * default, "false" means prefaulted entries will be 'young'.
156 static int __init
disable_randmaps(char *s
)
158 randomize_va_space
= 0;
161 __setup("norandmaps", disable_randmaps
);
163 unsigned long zero_pfn __read_mostly
;
164 EXPORT_SYMBOL(zero_pfn
);
166 unsigned long highest_memmap_pfn __read_mostly
;
169 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
171 static int __init
init_zero_pfn(void)
173 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
176 early_initcall(init_zero_pfn
);
178 void mm_trace_rss_stat(struct mm_struct
*mm
, int member
)
180 trace_rss_stat(mm
, member
);
184 * Note: this doesn't free the actual pages themselves. That
185 * has been handled earlier when unmapping all the memory regions.
187 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
190 pgtable_t token
= pmd_pgtable(*pmd
);
192 pte_free_tlb(tlb
, token
, addr
);
193 mm_dec_nr_ptes(tlb
->mm
);
196 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
197 unsigned long addr
, unsigned long end
,
198 unsigned long floor
, unsigned long ceiling
)
205 pmd
= pmd_offset(pud
, addr
);
207 next
= pmd_addr_end(addr
, end
);
208 if (pmd_none_or_clear_bad(pmd
))
210 free_pte_range(tlb
, pmd
, addr
);
211 } while (pmd
++, addr
= next
, addr
!= end
);
221 if (end
- 1 > ceiling
- 1)
224 pmd
= pmd_offset(pud
, start
);
226 pmd_free_tlb(tlb
, pmd
, start
);
227 mm_dec_nr_pmds(tlb
->mm
);
230 static inline void free_pud_range(struct mmu_gather
*tlb
, p4d_t
*p4d
,
231 unsigned long addr
, unsigned long end
,
232 unsigned long floor
, unsigned long ceiling
)
239 pud
= pud_offset(p4d
, addr
);
241 next
= pud_addr_end(addr
, end
);
242 if (pud_none_or_clear_bad(pud
))
244 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
245 } while (pud
++, addr
= next
, addr
!= end
);
255 if (end
- 1 > ceiling
- 1)
258 pud
= pud_offset(p4d
, start
);
260 pud_free_tlb(tlb
, pud
, start
);
261 mm_dec_nr_puds(tlb
->mm
);
264 static inline void free_p4d_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
265 unsigned long addr
, unsigned long end
,
266 unsigned long floor
, unsigned long ceiling
)
273 p4d
= p4d_offset(pgd
, addr
);
275 next
= p4d_addr_end(addr
, end
);
276 if (p4d_none_or_clear_bad(p4d
))
278 free_pud_range(tlb
, p4d
, addr
, next
, floor
, ceiling
);
279 } while (p4d
++, addr
= next
, addr
!= end
);
285 ceiling
&= PGDIR_MASK
;
289 if (end
- 1 > ceiling
- 1)
292 p4d
= p4d_offset(pgd
, start
);
294 p4d_free_tlb(tlb
, p4d
, start
);
298 * This function frees user-level page tables of a process.
300 void free_pgd_range(struct mmu_gather
*tlb
,
301 unsigned long addr
, unsigned long end
,
302 unsigned long floor
, unsigned long ceiling
)
308 * The next few lines have given us lots of grief...
310 * Why are we testing PMD* at this top level? Because often
311 * there will be no work to do at all, and we'd prefer not to
312 * go all the way down to the bottom just to discover that.
314 * Why all these "- 1"s? Because 0 represents both the bottom
315 * of the address space and the top of it (using -1 for the
316 * top wouldn't help much: the masks would do the wrong thing).
317 * The rule is that addr 0 and floor 0 refer to the bottom of
318 * the address space, but end 0 and ceiling 0 refer to the top
319 * Comparisons need to use "end - 1" and "ceiling - 1" (though
320 * that end 0 case should be mythical).
322 * Wherever addr is brought up or ceiling brought down, we must
323 * be careful to reject "the opposite 0" before it confuses the
324 * subsequent tests. But what about where end is brought down
325 * by PMD_SIZE below? no, end can't go down to 0 there.
327 * Whereas we round start (addr) and ceiling down, by different
328 * masks at different levels, in order to test whether a table
329 * now has no other vmas using it, so can be freed, we don't
330 * bother to round floor or end up - the tests don't need that.
344 if (end
- 1 > ceiling
- 1)
349 * We add page table cache pages with PAGE_SIZE,
350 * (see pte_free_tlb()), flush the tlb if we need
352 tlb_change_page_size(tlb
, PAGE_SIZE
);
353 pgd
= pgd_offset(tlb
->mm
, addr
);
355 next
= pgd_addr_end(addr
, end
);
356 if (pgd_none_or_clear_bad(pgd
))
358 free_p4d_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
359 } while (pgd
++, addr
= next
, addr
!= end
);
362 void free_pgtables(struct mmu_gather
*tlb
, struct ma_state
*mas
,
363 struct vm_area_struct
*vma
, unsigned long floor
,
364 unsigned long ceiling
, bool mm_wr_locked
)
367 unsigned long addr
= vma
->vm_start
;
368 struct vm_area_struct
*next
;
371 * Note: USER_PGTABLES_CEILING may be passed as ceiling and may
372 * be 0. This will underflow and is okay.
374 next
= mas_find(mas
, ceiling
- 1);
375 if (unlikely(xa_is_zero(next
)))
379 * Hide vma from rmap and truncate_pagecache before freeing
383 vma_start_write(vma
);
384 unlink_anon_vmas(vma
);
385 unlink_file_vma(vma
);
387 if (is_vm_hugetlb_page(vma
)) {
388 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
389 floor
, next
? next
->vm_start
: ceiling
);
392 * Optimization: gather nearby vmas into one call down
394 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
395 && !is_vm_hugetlb_page(next
)) {
397 next
= mas_find(mas
, ceiling
- 1);
398 if (unlikely(xa_is_zero(next
)))
401 vma_start_write(vma
);
402 unlink_anon_vmas(vma
);
403 unlink_file_vma(vma
);
405 free_pgd_range(tlb
, addr
, vma
->vm_end
,
406 floor
, next
? next
->vm_start
: ceiling
);
412 void pmd_install(struct mm_struct
*mm
, pmd_t
*pmd
, pgtable_t
*pte
)
414 spinlock_t
*ptl
= pmd_lock(mm
, pmd
);
416 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
419 * Ensure all pte setup (eg. pte page lock and page clearing) are
420 * visible before the pte is made visible to other CPUs by being
421 * put into page tables.
423 * The other side of the story is the pointer chasing in the page
424 * table walking code (when walking the page table without locking;
425 * ie. most of the time). Fortunately, these data accesses consist
426 * of a chain of data-dependent loads, meaning most CPUs (alpha
427 * being the notable exception) will already guarantee loads are
428 * seen in-order. See the alpha page table accessors for the
429 * smp_rmb() barriers in page table walking code.
431 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
432 pmd_populate(mm
, pmd
, *pte
);
438 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
)
440 pgtable_t
new = pte_alloc_one(mm
);
444 pmd_install(mm
, pmd
, &new);
450 int __pte_alloc_kernel(pmd_t
*pmd
)
452 pte_t
*new = pte_alloc_one_kernel(&init_mm
);
456 spin_lock(&init_mm
.page_table_lock
);
457 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
458 smp_wmb(); /* See comment in pmd_install() */
459 pmd_populate_kernel(&init_mm
, pmd
, new);
462 spin_unlock(&init_mm
.page_table_lock
);
464 pte_free_kernel(&init_mm
, new);
468 static inline void init_rss_vec(int *rss
)
470 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
473 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
477 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
479 add_mm_counter(mm
, i
, rss
[i
]);
483 * This function is called to print an error when a bad pte
484 * is found. For example, we might have a PFN-mapped pte in
485 * a region that doesn't allow it.
487 * The calling function must still handle the error.
489 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
490 pte_t pte
, struct page
*page
)
492 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
493 p4d_t
*p4d
= p4d_offset(pgd
, addr
);
494 pud_t
*pud
= pud_offset(p4d
, addr
);
495 pmd_t
*pmd
= pmd_offset(pud
, addr
);
496 struct address_space
*mapping
;
498 static unsigned long resume
;
499 static unsigned long nr_shown
;
500 static unsigned long nr_unshown
;
503 * Allow a burst of 60 reports, then keep quiet for that minute;
504 * or allow a steady drip of one report per second.
506 if (nr_shown
== 60) {
507 if (time_before(jiffies
, resume
)) {
512 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
519 resume
= jiffies
+ 60 * HZ
;
521 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
522 index
= linear_page_index(vma
, addr
);
524 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
526 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
528 dump_page(page
, "bad pte");
529 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
530 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
531 pr_alert("file:%pD fault:%ps mmap:%ps read_folio:%ps\n",
533 vma
->vm_ops
? vma
->vm_ops
->fault
: NULL
,
534 vma
->vm_file
? vma
->vm_file
->f_op
->mmap
: NULL
,
535 mapping
? mapping
->a_ops
->read_folio
: NULL
);
537 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
541 * vm_normal_page -- This function gets the "struct page" associated with a pte.
543 * "Special" mappings do not wish to be associated with a "struct page" (either
544 * it doesn't exist, or it exists but they don't want to touch it). In this
545 * case, NULL is returned here. "Normal" mappings do have a struct page.
547 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
548 * pte bit, in which case this function is trivial. Secondly, an architecture
549 * may not have a spare pte bit, which requires a more complicated scheme,
552 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
553 * special mapping (even if there are underlying and valid "struct pages").
554 * COWed pages of a VM_PFNMAP are always normal.
556 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
557 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
558 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
559 * mapping will always honor the rule
561 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
563 * And for normal mappings this is false.
565 * This restricts such mappings to be a linear translation from virtual address
566 * to pfn. To get around this restriction, we allow arbitrary mappings so long
567 * as the vma is not a COW mapping; in that case, we know that all ptes are
568 * special (because none can have been COWed).
571 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
573 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
574 * page" backing, however the difference is that _all_ pages with a struct
575 * page (that is, those where pfn_valid is true) are refcounted and considered
576 * normal pages by the VM. The disadvantage is that pages are refcounted
577 * (which can be slower and simply not an option for some PFNMAP users). The
578 * advantage is that we don't have to follow the strict linearity rule of
579 * PFNMAP mappings in order to support COWable mappings.
582 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
585 unsigned long pfn
= pte_pfn(pte
);
587 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
)) {
588 if (likely(!pte_special(pte
)))
590 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
591 return vma
->vm_ops
->find_special_page(vma
, addr
);
592 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
594 if (is_zero_pfn(pfn
))
598 * NOTE: New users of ZONE_DEVICE will not set pte_devmap()
599 * and will have refcounts incremented on their struct pages
600 * when they are inserted into PTEs, thus they are safe to
601 * return here. Legacy ZONE_DEVICE pages that set pte_devmap()
602 * do not have refcounts. Example of legacy ZONE_DEVICE is
603 * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers.
607 print_bad_pte(vma
, addr
, pte
, NULL
);
611 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
613 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
614 if (vma
->vm_flags
& VM_MIXEDMAP
) {
620 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
621 if (pfn
== vma
->vm_pgoff
+ off
)
623 if (!is_cow_mapping(vma
->vm_flags
))
628 if (is_zero_pfn(pfn
))
632 if (unlikely(pfn
> highest_memmap_pfn
)) {
633 print_bad_pte(vma
, addr
, pte
, NULL
);
638 * NOTE! We still have PageReserved() pages in the page tables.
639 * eg. VDSO mappings can cause them to exist.
642 return pfn_to_page(pfn
);
645 struct folio
*vm_normal_folio(struct vm_area_struct
*vma
, unsigned long addr
,
648 struct page
*page
= vm_normal_page(vma
, addr
, pte
);
651 return page_folio(page
);
655 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
656 struct page
*vm_normal_page_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
659 unsigned long pfn
= pmd_pfn(pmd
);
662 * There is no pmd_special() but there may be special pmds, e.g.
663 * in a direct-access (dax) mapping, so let's just replicate the
664 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
666 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
667 if (vma
->vm_flags
& VM_MIXEDMAP
) {
673 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
674 if (pfn
== vma
->vm_pgoff
+ off
)
676 if (!is_cow_mapping(vma
->vm_flags
))
683 if (is_huge_zero_pmd(pmd
))
685 if (unlikely(pfn
> highest_memmap_pfn
))
689 * NOTE! We still have PageReserved() pages in the page tables.
690 * eg. VDSO mappings can cause them to exist.
693 return pfn_to_page(pfn
);
696 struct folio
*vm_normal_folio_pmd(struct vm_area_struct
*vma
,
697 unsigned long addr
, pmd_t pmd
)
699 struct page
*page
= vm_normal_page_pmd(vma
, addr
, pmd
);
702 return page_folio(page
);
707 static void restore_exclusive_pte(struct vm_area_struct
*vma
,
708 struct page
*page
, unsigned long address
,
711 struct folio
*folio
= page_folio(page
);
716 orig_pte
= ptep_get(ptep
);
717 pte
= pte_mkold(mk_pte(page
, READ_ONCE(vma
->vm_page_prot
)));
718 if (pte_swp_soft_dirty(orig_pte
))
719 pte
= pte_mksoft_dirty(pte
);
721 entry
= pte_to_swp_entry(orig_pte
);
722 if (pte_swp_uffd_wp(orig_pte
))
723 pte
= pte_mkuffd_wp(pte
);
724 else if (is_writable_device_exclusive_entry(entry
))
725 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
727 VM_BUG_ON_FOLIO(pte_write(pte
) && (!folio_test_anon(folio
) &&
728 PageAnonExclusive(page
)), folio
);
731 * No need to take a page reference as one was already
732 * created when the swap entry was made.
734 if (folio_test_anon(folio
))
735 folio_add_anon_rmap_pte(folio
, page
, vma
, address
, RMAP_NONE
);
738 * Currently device exclusive access only supports anonymous
739 * memory so the entry shouldn't point to a filebacked page.
743 set_pte_at(vma
->vm_mm
, address
, ptep
, pte
);
746 * No need to invalidate - it was non-present before. However
747 * secondary CPUs may have mappings that need invalidating.
749 update_mmu_cache(vma
, address
, ptep
);
753 * Tries to restore an exclusive pte if the page lock can be acquired without
757 try_restore_exclusive_pte(pte_t
*src_pte
, struct vm_area_struct
*vma
,
760 swp_entry_t entry
= pte_to_swp_entry(ptep_get(src_pte
));
761 struct page
*page
= pfn_swap_entry_to_page(entry
);
763 if (trylock_page(page
)) {
764 restore_exclusive_pte(vma
, page
, addr
, src_pte
);
773 * copy one vm_area from one task to the other. Assumes the page tables
774 * already present in the new task to be cleared in the whole range
775 * covered by this vma.
779 copy_nonpresent_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
780 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*dst_vma
,
781 struct vm_area_struct
*src_vma
, unsigned long addr
, int *rss
)
783 unsigned long vm_flags
= dst_vma
->vm_flags
;
784 pte_t orig_pte
= ptep_get(src_pte
);
785 pte_t pte
= orig_pte
;
788 swp_entry_t entry
= pte_to_swp_entry(orig_pte
);
790 if (likely(!non_swap_entry(entry
))) {
791 if (swap_duplicate(entry
) < 0)
794 /* make sure dst_mm is on swapoff's mmlist. */
795 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
796 spin_lock(&mmlist_lock
);
797 if (list_empty(&dst_mm
->mmlist
))
798 list_add(&dst_mm
->mmlist
,
800 spin_unlock(&mmlist_lock
);
802 /* Mark the swap entry as shared. */
803 if (pte_swp_exclusive(orig_pte
)) {
804 pte
= pte_swp_clear_exclusive(orig_pte
);
805 set_pte_at(src_mm
, addr
, src_pte
, pte
);
808 } else if (is_migration_entry(entry
)) {
809 folio
= pfn_swap_entry_folio(entry
);
811 rss
[mm_counter(folio
)]++;
813 if (!is_readable_migration_entry(entry
) &&
814 is_cow_mapping(vm_flags
)) {
816 * COW mappings require pages in both parent and child
817 * to be set to read. A previously exclusive entry is
820 entry
= make_readable_migration_entry(
822 pte
= swp_entry_to_pte(entry
);
823 if (pte_swp_soft_dirty(orig_pte
))
824 pte
= pte_swp_mksoft_dirty(pte
);
825 if (pte_swp_uffd_wp(orig_pte
))
826 pte
= pte_swp_mkuffd_wp(pte
);
827 set_pte_at(src_mm
, addr
, src_pte
, pte
);
829 } else if (is_device_private_entry(entry
)) {
830 page
= pfn_swap_entry_to_page(entry
);
831 folio
= page_folio(page
);
834 * Update rss count even for unaddressable pages, as
835 * they should treated just like normal pages in this
838 * We will likely want to have some new rss counters
839 * for unaddressable pages, at some point. But for now
840 * keep things as they are.
843 rss
[mm_counter(folio
)]++;
844 /* Cannot fail as these pages cannot get pinned. */
845 folio_try_dup_anon_rmap_pte(folio
, page
, src_vma
);
848 * We do not preserve soft-dirty information, because so
849 * far, checkpoint/restore is the only feature that
850 * requires that. And checkpoint/restore does not work
851 * when a device driver is involved (you cannot easily
852 * save and restore device driver state).
854 if (is_writable_device_private_entry(entry
) &&
855 is_cow_mapping(vm_flags
)) {
856 entry
= make_readable_device_private_entry(
858 pte
= swp_entry_to_pte(entry
);
859 if (pte_swp_uffd_wp(orig_pte
))
860 pte
= pte_swp_mkuffd_wp(pte
);
861 set_pte_at(src_mm
, addr
, src_pte
, pte
);
863 } else if (is_device_exclusive_entry(entry
)) {
865 * Make device exclusive entries present by restoring the
866 * original entry then copying as for a present pte. Device
867 * exclusive entries currently only support private writable
868 * (ie. COW) mappings.
870 VM_BUG_ON(!is_cow_mapping(src_vma
->vm_flags
));
871 if (try_restore_exclusive_pte(src_pte
, src_vma
, addr
))
874 } else if (is_pte_marker_entry(entry
)) {
875 pte_marker marker
= copy_pte_marker(entry
, dst_vma
);
878 set_pte_at(dst_mm
, addr
, dst_pte
,
879 make_pte_marker(marker
));
882 if (!userfaultfd_wp(dst_vma
))
883 pte
= pte_swp_clear_uffd_wp(pte
);
884 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
889 * Copy a present and normal page.
891 * NOTE! The usual case is that this isn't required;
892 * instead, the caller can just increase the page refcount
893 * and re-use the pte the traditional way.
895 * And if we need a pre-allocated page but don't yet have
896 * one, return a negative error to let the preallocation
897 * code know so that it can do so outside the page table
901 copy_present_page(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
902 pte_t
*dst_pte
, pte_t
*src_pte
, unsigned long addr
, int *rss
,
903 struct folio
**prealloc
, struct page
*page
)
905 struct folio
*new_folio
;
908 new_folio
= *prealloc
;
913 * We have a prealloc page, all good! Take it
914 * over and copy the page & arm it.
917 copy_user_highpage(&new_folio
->page
, page
, addr
, src_vma
);
918 __folio_mark_uptodate(new_folio
);
919 folio_add_new_anon_rmap(new_folio
, dst_vma
, addr
);
920 folio_add_lru_vma(new_folio
, dst_vma
);
923 /* All done, just insert the new page copy in the child */
924 pte
= mk_pte(&new_folio
->page
, dst_vma
->vm_page_prot
);
925 pte
= maybe_mkwrite(pte_mkdirty(pte
), dst_vma
);
926 if (userfaultfd_pte_wp(dst_vma
, ptep_get(src_pte
)))
927 /* Uffd-wp needs to be delivered to dest pte as well */
928 pte
= pte_mkuffd_wp(pte
);
929 set_pte_at(dst_vma
->vm_mm
, addr
, dst_pte
, pte
);
933 static __always_inline
void __copy_present_ptes(struct vm_area_struct
*dst_vma
,
934 struct vm_area_struct
*src_vma
, pte_t
*dst_pte
, pte_t
*src_pte
,
935 pte_t pte
, unsigned long addr
, int nr
)
937 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
939 /* If it's a COW mapping, write protect it both processes. */
940 if (is_cow_mapping(src_vma
->vm_flags
) && pte_write(pte
)) {
941 wrprotect_ptes(src_mm
, addr
, src_pte
, nr
);
942 pte
= pte_wrprotect(pte
);
945 /* If it's a shared mapping, mark it clean in the child. */
946 if (src_vma
->vm_flags
& VM_SHARED
)
947 pte
= pte_mkclean(pte
);
948 pte
= pte_mkold(pte
);
950 if (!userfaultfd_wp(dst_vma
))
951 pte
= pte_clear_uffd_wp(pte
);
953 set_ptes(dst_vma
->vm_mm
, addr
, dst_pte
, pte
, nr
);
957 * Copy one present PTE, trying to batch-process subsequent PTEs that map
958 * consecutive pages of the same folio by copying them as well.
960 * Returns -EAGAIN if one preallocated page is required to copy the next PTE.
961 * Otherwise, returns the number of copied PTEs (at least 1).
964 copy_present_ptes(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
965 pte_t
*dst_pte
, pte_t
*src_pte
, pte_t pte
, unsigned long addr
,
966 int max_nr
, int *rss
, struct folio
**prealloc
)
974 page
= vm_normal_page(src_vma
, addr
, pte
);
978 folio
= page_folio(page
);
981 * If we likely have to copy, just don't bother with batching. Make
982 * sure that the common "small folio" case is as fast as possible
983 * by keeping the batching logic separate.
985 if (unlikely(!*prealloc
&& folio_test_large(folio
) && max_nr
!= 1)) {
986 if (src_vma
->vm_flags
& VM_SHARED
)
987 flags
|= FPB_IGNORE_DIRTY
;
988 if (!vma_soft_dirty_enabled(src_vma
))
989 flags
|= FPB_IGNORE_SOFT_DIRTY
;
991 nr
= folio_pte_batch(folio
, addr
, src_pte
, pte
, max_nr
, flags
,
993 folio_ref_add(folio
, nr
);
994 if (folio_test_anon(folio
)) {
995 if (unlikely(folio_try_dup_anon_rmap_ptes(folio
, page
,
997 folio_ref_sub(folio
, nr
);
1000 rss
[MM_ANONPAGES
] += nr
;
1001 VM_WARN_ON_FOLIO(PageAnonExclusive(page
), folio
);
1003 folio_dup_file_rmap_ptes(folio
, page
, nr
);
1004 rss
[mm_counter_file(folio
)] += nr
;
1007 pte
= pte_mkwrite(pte
, src_vma
);
1008 __copy_present_ptes(dst_vma
, src_vma
, dst_pte
, src_pte
, pte
,
1014 if (folio_test_anon(folio
)) {
1016 * If this page may have been pinned by the parent process,
1017 * copy the page immediately for the child so that we'll always
1018 * guarantee the pinned page won't be randomly replaced in the
1021 if (unlikely(folio_try_dup_anon_rmap_pte(folio
, page
, src_vma
))) {
1022 /* Page may be pinned, we have to copy. */
1024 err
= copy_present_page(dst_vma
, src_vma
, dst_pte
, src_pte
,
1025 addr
, rss
, prealloc
, page
);
1026 return err
? err
: 1;
1028 rss
[MM_ANONPAGES
]++;
1029 VM_WARN_ON_FOLIO(PageAnonExclusive(page
), folio
);
1031 folio_dup_file_rmap_pte(folio
, page
);
1032 rss
[mm_counter_file(folio
)]++;
1036 __copy_present_ptes(dst_vma
, src_vma
, dst_pte
, src_pte
, pte
, addr
, 1);
1040 static inline struct folio
*folio_prealloc(struct mm_struct
*src_mm
,
1041 struct vm_area_struct
*vma
, unsigned long addr
, bool need_zero
)
1043 struct folio
*new_folio
;
1046 new_folio
= vma_alloc_zeroed_movable_folio(vma
, addr
);
1048 new_folio
= vma_alloc_folio(GFP_HIGHUSER_MOVABLE
, 0, vma
,
1054 if (mem_cgroup_charge(new_folio
, src_mm
, GFP_KERNEL
)) {
1055 folio_put(new_folio
);
1058 folio_throttle_swaprate(new_folio
, GFP_KERNEL
);
1064 copy_pte_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1065 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
1068 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1069 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1070 pte_t
*orig_src_pte
, *orig_dst_pte
;
1071 pte_t
*src_pte
, *dst_pte
;
1073 spinlock_t
*src_ptl
, *dst_ptl
;
1074 int progress
, max_nr
, ret
= 0;
1075 int rss
[NR_MM_COUNTERS
];
1076 swp_entry_t entry
= (swp_entry_t
){0};
1077 struct folio
*prealloc
= NULL
;
1085 * copy_pmd_range()'s prior pmd_none_or_clear_bad(src_pmd), and the
1086 * error handling here, assume that exclusive mmap_lock on dst and src
1087 * protects anon from unexpected THP transitions; with shmem and file
1088 * protected by mmap_lock-less collapse skipping areas with anon_vma
1089 * (whereas vma_needs_copy() skips areas without anon_vma). A rework
1090 * can remove such assumptions later, but this is good enough for now.
1092 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
1097 src_pte
= pte_offset_map_nolock(src_mm
, src_pmd
, addr
, &src_ptl
);
1099 pte_unmap_unlock(dst_pte
, dst_ptl
);
1103 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1104 orig_src_pte
= src_pte
;
1105 orig_dst_pte
= dst_pte
;
1106 arch_enter_lazy_mmu_mode();
1112 * We are holding two locks at this point - either of them
1113 * could generate latencies in another task on another CPU.
1115 if (progress
>= 32) {
1117 if (need_resched() ||
1118 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
1121 ptent
= ptep_get(src_pte
);
1122 if (pte_none(ptent
)) {
1126 if (unlikely(!pte_present(ptent
))) {
1127 ret
= copy_nonpresent_pte(dst_mm
, src_mm
,
1132 entry
= pte_to_swp_entry(ptep_get(src_pte
));
1134 } else if (ret
== -EBUSY
) {
1140 ptent
= ptep_get(src_pte
);
1141 VM_WARN_ON_ONCE(!pte_present(ptent
));
1144 * Device exclusive entry restored, continue by copying
1145 * the now present pte.
1147 WARN_ON_ONCE(ret
!= -ENOENT
);
1149 /* copy_present_ptes() will clear `*prealloc' if consumed */
1150 max_nr
= (end
- addr
) / PAGE_SIZE
;
1151 ret
= copy_present_ptes(dst_vma
, src_vma
, dst_pte
, src_pte
,
1152 ptent
, addr
, max_nr
, rss
, &prealloc
);
1154 * If we need a pre-allocated page for this pte, drop the
1155 * locks, allocate, and try again.
1157 if (unlikely(ret
== -EAGAIN
))
1159 if (unlikely(prealloc
)) {
1161 * pre-alloc page cannot be reused by next time so as
1162 * to strictly follow mempolicy (e.g., alloc_page_vma()
1163 * will allocate page according to address). This
1164 * could only happen if one pinned pte changed.
1166 folio_put(prealloc
);
1171 } while (dst_pte
+= nr
, src_pte
+= nr
, addr
+= PAGE_SIZE
* nr
,
1174 arch_leave_lazy_mmu_mode();
1175 pte_unmap_unlock(orig_src_pte
, src_ptl
);
1176 add_mm_rss_vec(dst_mm
, rss
);
1177 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
1181 VM_WARN_ON_ONCE(!entry
.val
);
1182 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0) {
1187 } else if (ret
== -EBUSY
) {
1189 } else if (ret
== -EAGAIN
) {
1190 prealloc
= folio_prealloc(src_mm
, src_vma
, addr
, false);
1193 } else if (ret
< 0) {
1197 /* We've captured and resolved the error. Reset, try again. */
1203 if (unlikely(prealloc
))
1204 folio_put(prealloc
);
1209 copy_pmd_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1210 pud_t
*dst_pud
, pud_t
*src_pud
, unsigned long addr
,
1213 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1214 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1215 pmd_t
*src_pmd
, *dst_pmd
;
1218 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
1221 src_pmd
= pmd_offset(src_pud
, addr
);
1223 next
= pmd_addr_end(addr
, end
);
1224 if (is_swap_pmd(*src_pmd
) || pmd_trans_huge(*src_pmd
)
1225 || pmd_devmap(*src_pmd
)) {
1227 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PMD_SIZE
, src_vma
);
1228 err
= copy_huge_pmd(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
1229 addr
, dst_vma
, src_vma
);
1236 if (pmd_none_or_clear_bad(src_pmd
))
1238 if (copy_pte_range(dst_vma
, src_vma
, dst_pmd
, src_pmd
,
1241 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1246 copy_pud_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1247 p4d_t
*dst_p4d
, p4d_t
*src_p4d
, unsigned long addr
,
1250 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1251 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1252 pud_t
*src_pud
, *dst_pud
;
1255 dst_pud
= pud_alloc(dst_mm
, dst_p4d
, addr
);
1258 src_pud
= pud_offset(src_p4d
, addr
);
1260 next
= pud_addr_end(addr
, end
);
1261 if (pud_trans_huge(*src_pud
) || pud_devmap(*src_pud
)) {
1264 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PUD_SIZE
, src_vma
);
1265 err
= copy_huge_pud(dst_mm
, src_mm
,
1266 dst_pud
, src_pud
, addr
, src_vma
);
1273 if (pud_none_or_clear_bad(src_pud
))
1275 if (copy_pmd_range(dst_vma
, src_vma
, dst_pud
, src_pud
,
1278 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1283 copy_p4d_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1284 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, unsigned long addr
,
1287 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1288 p4d_t
*src_p4d
, *dst_p4d
;
1291 dst_p4d
= p4d_alloc(dst_mm
, dst_pgd
, addr
);
1294 src_p4d
= p4d_offset(src_pgd
, addr
);
1296 next
= p4d_addr_end(addr
, end
);
1297 if (p4d_none_or_clear_bad(src_p4d
))
1299 if (copy_pud_range(dst_vma
, src_vma
, dst_p4d
, src_p4d
,
1302 } while (dst_p4d
++, src_p4d
++, addr
= next
, addr
!= end
);
1307 * Return true if the vma needs to copy the pgtable during this fork(). Return
1308 * false when we can speed up fork() by allowing lazy page faults later until
1309 * when the child accesses the memory range.
1312 vma_needs_copy(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
)
1315 * Always copy pgtables when dst_vma has uffd-wp enabled even if it's
1316 * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable
1317 * contains uffd-wp protection information, that's something we can't
1318 * retrieve from page cache, and skip copying will lose those info.
1320 if (userfaultfd_wp(dst_vma
))
1323 if (src_vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
1326 if (src_vma
->anon_vma
)
1330 * Don't copy ptes where a page fault will fill them correctly. Fork
1331 * becomes much lighter when there are big shared or private readonly
1332 * mappings. The tradeoff is that copy_page_range is more efficient
1339 copy_page_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
)
1341 pgd_t
*src_pgd
, *dst_pgd
;
1343 unsigned long addr
= src_vma
->vm_start
;
1344 unsigned long end
= src_vma
->vm_end
;
1345 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1346 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1347 struct mmu_notifier_range range
;
1351 if (!vma_needs_copy(dst_vma
, src_vma
))
1354 if (is_vm_hugetlb_page(src_vma
))
1355 return copy_hugetlb_page_range(dst_mm
, src_mm
, dst_vma
, src_vma
);
1357 if (unlikely(src_vma
->vm_flags
& VM_PFNMAP
)) {
1359 * We do not free on error cases below as remove_vma
1360 * gets called on error from higher level routine
1362 ret
= track_pfn_copy(src_vma
);
1368 * We need to invalidate the secondary MMU mappings only when
1369 * there could be a permission downgrade on the ptes of the
1370 * parent mm. And a permission downgrade will only happen if
1371 * is_cow_mapping() returns true.
1373 is_cow
= is_cow_mapping(src_vma
->vm_flags
);
1376 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_PAGE
,
1377 0, src_mm
, addr
, end
);
1378 mmu_notifier_invalidate_range_start(&range
);
1380 * Disabling preemption is not needed for the write side, as
1381 * the read side doesn't spin, but goes to the mmap_lock.
1383 * Use the raw variant of the seqcount_t write API to avoid
1384 * lockdep complaining about preemptibility.
1386 vma_assert_write_locked(src_vma
);
1387 raw_write_seqcount_begin(&src_mm
->write_protect_seq
);
1391 dst_pgd
= pgd_offset(dst_mm
, addr
);
1392 src_pgd
= pgd_offset(src_mm
, addr
);
1394 next
= pgd_addr_end(addr
, end
);
1395 if (pgd_none_or_clear_bad(src_pgd
))
1397 if (unlikely(copy_p4d_range(dst_vma
, src_vma
, dst_pgd
, src_pgd
,
1399 untrack_pfn_clear(dst_vma
);
1403 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1406 raw_write_seqcount_end(&src_mm
->write_protect_seq
);
1407 mmu_notifier_invalidate_range_end(&range
);
1412 /* Whether we should zap all COWed (private) pages too */
1413 static inline bool should_zap_cows(struct zap_details
*details
)
1415 /* By default, zap all pages */
1419 /* Or, we zap COWed pages only if the caller wants to */
1420 return details
->even_cows
;
1423 /* Decides whether we should zap this folio with the folio pointer specified */
1424 static inline bool should_zap_folio(struct zap_details
*details
,
1425 struct folio
*folio
)
1427 /* If we can make a decision without *folio.. */
1428 if (should_zap_cows(details
))
1431 /* Otherwise we should only zap non-anon folios */
1432 return !folio_test_anon(folio
);
1435 static inline bool zap_drop_file_uffd_wp(struct zap_details
*details
)
1440 return details
->zap_flags
& ZAP_FLAG_DROP_MARKER
;
1444 * This function makes sure that we'll replace the none pte with an uffd-wp
1445 * swap special pte marker when necessary. Must be with the pgtable lock held.
1448 zap_install_uffd_wp_if_needed(struct vm_area_struct
*vma
,
1449 unsigned long addr
, pte_t
*pte
, int nr
,
1450 struct zap_details
*details
, pte_t pteval
)
1452 /* Zap on anonymous always means dropping everything */
1453 if (vma_is_anonymous(vma
))
1456 if (zap_drop_file_uffd_wp(details
))
1460 /* the PFN in the PTE is irrelevant. */
1461 pte_install_uffd_wp_if_needed(vma
, addr
, pte
, pteval
);
1469 static __always_inline
void zap_present_folio_ptes(struct mmu_gather
*tlb
,
1470 struct vm_area_struct
*vma
, struct folio
*folio
,
1471 struct page
*page
, pte_t
*pte
, pte_t ptent
, unsigned int nr
,
1472 unsigned long addr
, struct zap_details
*details
, int *rss
,
1473 bool *force_flush
, bool *force_break
)
1475 struct mm_struct
*mm
= tlb
->mm
;
1476 bool delay_rmap
= false;
1478 if (!folio_test_anon(folio
)) {
1479 ptent
= get_and_clear_full_ptes(mm
, addr
, pte
, nr
, tlb
->fullmm
);
1480 if (pte_dirty(ptent
)) {
1481 folio_mark_dirty(folio
);
1482 if (tlb_delay_rmap(tlb
)) {
1484 *force_flush
= true;
1487 if (pte_young(ptent
) && likely(vma_has_recency(vma
)))
1488 folio_mark_accessed(folio
);
1489 rss
[mm_counter(folio
)] -= nr
;
1491 /* We don't need up-to-date accessed/dirty bits. */
1492 clear_full_ptes(mm
, addr
, pte
, nr
, tlb
->fullmm
);
1493 rss
[MM_ANONPAGES
] -= nr
;
1495 /* Checking a single PTE in a batch is sufficient. */
1496 arch_check_zapped_pte(vma
, ptent
);
1497 tlb_remove_tlb_entries(tlb
, pte
, nr
, addr
);
1498 if (unlikely(userfaultfd_pte_wp(vma
, ptent
)))
1499 zap_install_uffd_wp_if_needed(vma
, addr
, pte
, nr
, details
,
1503 folio_remove_rmap_ptes(folio
, page
, nr
, vma
);
1505 /* Only sanity-check the first page in a batch. */
1506 if (unlikely(page_mapcount(page
) < 0))
1507 print_bad_pte(vma
, addr
, ptent
, page
);
1509 if (unlikely(__tlb_remove_folio_pages(tlb
, page
, nr
, delay_rmap
))) {
1510 *force_flush
= true;
1511 *force_break
= true;
1516 * Zap or skip at least one present PTE, trying to batch-process subsequent
1517 * PTEs that map consecutive pages of the same folio.
1519 * Returns the number of processed (skipped or zapped) PTEs (at least 1).
1521 static inline int zap_present_ptes(struct mmu_gather
*tlb
,
1522 struct vm_area_struct
*vma
, pte_t
*pte
, pte_t ptent
,
1523 unsigned int max_nr
, unsigned long addr
,
1524 struct zap_details
*details
, int *rss
, bool *force_flush
,
1527 const fpb_t fpb_flags
= FPB_IGNORE_DIRTY
| FPB_IGNORE_SOFT_DIRTY
;
1528 struct mm_struct
*mm
= tlb
->mm
;
1529 struct folio
*folio
;
1533 page
= vm_normal_page(vma
, addr
, ptent
);
1535 /* We don't need up-to-date accessed/dirty bits. */
1536 ptep_get_and_clear_full(mm
, addr
, pte
, tlb
->fullmm
);
1537 arch_check_zapped_pte(vma
, ptent
);
1538 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1539 if (userfaultfd_pte_wp(vma
, ptent
))
1540 zap_install_uffd_wp_if_needed(vma
, addr
, pte
, 1,
1542 ksm_might_unmap_zero_page(mm
, ptent
);
1546 folio
= page_folio(page
);
1547 if (unlikely(!should_zap_folio(details
, folio
)))
1551 * Make sure that the common "small folio" case is as fast as possible
1552 * by keeping the batching logic separate.
1554 if (unlikely(folio_test_large(folio
) && max_nr
!= 1)) {
1555 nr
= folio_pte_batch(folio
, addr
, pte
, ptent
, max_nr
, fpb_flags
,
1558 zap_present_folio_ptes(tlb
, vma
, folio
, page
, pte
, ptent
, nr
,
1559 addr
, details
, rss
, force_flush
,
1563 zap_present_folio_ptes(tlb
, vma
, folio
, page
, pte
, ptent
, 1, addr
,
1564 details
, rss
, force_flush
, force_break
);
1568 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1569 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1570 unsigned long addr
, unsigned long end
,
1571 struct zap_details
*details
)
1573 bool force_flush
= false, force_break
= false;
1574 struct mm_struct
*mm
= tlb
->mm
;
1575 int rss
[NR_MM_COUNTERS
];
1582 tlb_change_page_size(tlb
, PAGE_SIZE
);
1584 start_pte
= pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1588 flush_tlb_batched_pending(mm
);
1589 arch_enter_lazy_mmu_mode();
1591 pte_t ptent
= ptep_get(pte
);
1592 struct folio
*folio
;
1597 if (pte_none(ptent
))
1603 if (pte_present(ptent
)) {
1604 max_nr
= (end
- addr
) / PAGE_SIZE
;
1605 nr
= zap_present_ptes(tlb
, vma
, pte
, ptent
, max_nr
,
1606 addr
, details
, rss
, &force_flush
,
1608 if (unlikely(force_break
)) {
1609 addr
+= nr
* PAGE_SIZE
;
1615 entry
= pte_to_swp_entry(ptent
);
1616 if (is_device_private_entry(entry
) ||
1617 is_device_exclusive_entry(entry
)) {
1618 page
= pfn_swap_entry_to_page(entry
);
1619 folio
= page_folio(page
);
1620 if (unlikely(!should_zap_folio(details
, folio
)))
1623 * Both device private/exclusive mappings should only
1624 * work with anonymous page so far, so we don't need to
1625 * consider uffd-wp bit when zap. For more information,
1626 * see zap_install_uffd_wp_if_needed().
1628 WARN_ON_ONCE(!vma_is_anonymous(vma
));
1629 rss
[mm_counter(folio
)]--;
1630 if (is_device_private_entry(entry
))
1631 folio_remove_rmap_pte(folio
, page
, vma
);
1633 } else if (!non_swap_entry(entry
)) {
1634 /* Genuine swap entry, hence a private anon page */
1635 if (!should_zap_cows(details
))
1638 if (unlikely(!free_swap_and_cache(entry
)))
1639 print_bad_pte(vma
, addr
, ptent
, NULL
);
1640 } else if (is_migration_entry(entry
)) {
1641 folio
= pfn_swap_entry_folio(entry
);
1642 if (!should_zap_folio(details
, folio
))
1644 rss
[mm_counter(folio
)]--;
1645 } else if (pte_marker_entry_uffd_wp(entry
)) {
1647 * For anon: always drop the marker; for file: only
1648 * drop the marker if explicitly requested.
1650 if (!vma_is_anonymous(vma
) &&
1651 !zap_drop_file_uffd_wp(details
))
1653 } else if (is_hwpoison_entry(entry
) ||
1654 is_poisoned_swp_entry(entry
)) {
1655 if (!should_zap_cows(details
))
1658 /* We should have covered all the swap entry types */
1659 pr_alert("unrecognized swap entry 0x%lx\n", entry
.val
);
1662 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1663 zap_install_uffd_wp_if_needed(vma
, addr
, pte
, 1, details
, ptent
);
1664 } while (pte
+= nr
, addr
+= PAGE_SIZE
* nr
, addr
!= end
);
1666 add_mm_rss_vec(mm
, rss
);
1667 arch_leave_lazy_mmu_mode();
1669 /* Do the actual TLB flush before dropping ptl */
1671 tlb_flush_mmu_tlbonly(tlb
);
1672 tlb_flush_rmaps(tlb
, vma
);
1674 pte_unmap_unlock(start_pte
, ptl
);
1677 * If we forced a TLB flush (either due to running out of
1678 * batch buffers or because we needed to flush dirty TLB
1679 * entries before releasing the ptl), free the batched
1680 * memory too. Come back again if we didn't do everything.
1688 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1689 struct vm_area_struct
*vma
, pud_t
*pud
,
1690 unsigned long addr
, unsigned long end
,
1691 struct zap_details
*details
)
1696 pmd
= pmd_offset(pud
, addr
);
1698 next
= pmd_addr_end(addr
, end
);
1699 if (is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1700 if (next
- addr
!= HPAGE_PMD_SIZE
)
1701 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
1702 else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
)) {
1707 } else if (details
&& details
->single_folio
&&
1708 folio_test_pmd_mappable(details
->single_folio
) &&
1709 next
- addr
== HPAGE_PMD_SIZE
&& pmd_none(*pmd
)) {
1710 spinlock_t
*ptl
= pmd_lock(tlb
->mm
, pmd
);
1712 * Take and drop THP pmd lock so that we cannot return
1713 * prematurely, while zap_huge_pmd() has cleared *pmd,
1714 * but not yet decremented compound_mapcount().
1718 if (pmd_none(*pmd
)) {
1722 addr
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1725 } while (pmd
++, cond_resched(), addr
!= end
);
1730 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1731 struct vm_area_struct
*vma
, p4d_t
*p4d
,
1732 unsigned long addr
, unsigned long end
,
1733 struct zap_details
*details
)
1738 pud
= pud_offset(p4d
, addr
);
1740 next
= pud_addr_end(addr
, end
);
1741 if (pud_trans_huge(*pud
) || pud_devmap(*pud
)) {
1742 if (next
- addr
!= HPAGE_PUD_SIZE
) {
1743 mmap_assert_locked(tlb
->mm
);
1744 split_huge_pud(vma
, pud
, addr
);
1745 } else if (zap_huge_pud(tlb
, vma
, pud
, addr
))
1749 if (pud_none_or_clear_bad(pud
))
1751 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1754 } while (pud
++, addr
= next
, addr
!= end
);
1759 static inline unsigned long zap_p4d_range(struct mmu_gather
*tlb
,
1760 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1761 unsigned long addr
, unsigned long end
,
1762 struct zap_details
*details
)
1767 p4d
= p4d_offset(pgd
, addr
);
1769 next
= p4d_addr_end(addr
, end
);
1770 if (p4d_none_or_clear_bad(p4d
))
1772 next
= zap_pud_range(tlb
, vma
, p4d
, addr
, next
, details
);
1773 } while (p4d
++, addr
= next
, addr
!= end
);
1778 void unmap_page_range(struct mmu_gather
*tlb
,
1779 struct vm_area_struct
*vma
,
1780 unsigned long addr
, unsigned long end
,
1781 struct zap_details
*details
)
1786 BUG_ON(addr
>= end
);
1787 tlb_start_vma(tlb
, vma
);
1788 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1790 next
= pgd_addr_end(addr
, end
);
1791 if (pgd_none_or_clear_bad(pgd
))
1793 next
= zap_p4d_range(tlb
, vma
, pgd
, addr
, next
, details
);
1794 } while (pgd
++, addr
= next
, addr
!= end
);
1795 tlb_end_vma(tlb
, vma
);
1799 static void unmap_single_vma(struct mmu_gather
*tlb
,
1800 struct vm_area_struct
*vma
, unsigned long start_addr
,
1801 unsigned long end_addr
,
1802 struct zap_details
*details
, bool mm_wr_locked
)
1804 unsigned long start
= max(vma
->vm_start
, start_addr
);
1807 if (start
>= vma
->vm_end
)
1809 end
= min(vma
->vm_end
, end_addr
);
1810 if (end
<= vma
->vm_start
)
1814 uprobe_munmap(vma
, start
, end
);
1816 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1817 untrack_pfn(vma
, 0, 0, mm_wr_locked
);
1820 if (unlikely(is_vm_hugetlb_page(vma
))) {
1822 * It is undesirable to test vma->vm_file as it
1823 * should be non-null for valid hugetlb area.
1824 * However, vm_file will be NULL in the error
1825 * cleanup path of mmap_region. When
1826 * hugetlbfs ->mmap method fails,
1827 * mmap_region() nullifies vma->vm_file
1828 * before calling this function to clean up.
1829 * Since no pte has actually been setup, it is
1830 * safe to do nothing in this case.
1833 zap_flags_t zap_flags
= details
?
1834 details
->zap_flags
: 0;
1835 __unmap_hugepage_range(tlb
, vma
, start
, end
,
1839 unmap_page_range(tlb
, vma
, start
, end
, details
);
1844 * unmap_vmas - unmap a range of memory covered by a list of vma's
1845 * @tlb: address of the caller's struct mmu_gather
1846 * @mas: the maple state
1847 * @vma: the starting vma
1848 * @start_addr: virtual address at which to start unmapping
1849 * @end_addr: virtual address at which to end unmapping
1850 * @tree_end: The maximum index to check
1851 * @mm_wr_locked: lock flag
1853 * Unmap all pages in the vma list.
1855 * Only addresses between `start' and `end' will be unmapped.
1857 * The VMA list must be sorted in ascending virtual address order.
1859 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1860 * range after unmap_vmas() returns. So the only responsibility here is to
1861 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1862 * drops the lock and schedules.
1864 void unmap_vmas(struct mmu_gather
*tlb
, struct ma_state
*mas
,
1865 struct vm_area_struct
*vma
, unsigned long start_addr
,
1866 unsigned long end_addr
, unsigned long tree_end
,
1869 struct mmu_notifier_range range
;
1870 struct zap_details details
= {
1871 .zap_flags
= ZAP_FLAG_DROP_MARKER
| ZAP_FLAG_UNMAP
,
1872 /* Careful - we need to zap private pages too! */
1876 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
->vm_mm
,
1877 start_addr
, end_addr
);
1878 mmu_notifier_invalidate_range_start(&range
);
1880 unsigned long start
= start_addr
;
1881 unsigned long end
= end_addr
;
1882 hugetlb_zap_begin(vma
, &start
, &end
);
1883 unmap_single_vma(tlb
, vma
, start
, end
, &details
,
1885 hugetlb_zap_end(vma
, &details
);
1886 vma
= mas_find(mas
, tree_end
- 1);
1887 } while (vma
&& likely(!xa_is_zero(vma
)));
1888 mmu_notifier_invalidate_range_end(&range
);
1892 * zap_page_range_single - remove user pages in a given range
1893 * @vma: vm_area_struct holding the applicable pages
1894 * @address: starting address of pages to zap
1895 * @size: number of bytes to zap
1896 * @details: details of shared cache invalidation
1898 * The range must fit into one VMA.
1900 void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1901 unsigned long size
, struct zap_details
*details
)
1903 const unsigned long end
= address
+ size
;
1904 struct mmu_notifier_range range
;
1905 struct mmu_gather tlb
;
1908 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
->vm_mm
,
1910 hugetlb_zap_begin(vma
, &range
.start
, &range
.end
);
1911 tlb_gather_mmu(&tlb
, vma
->vm_mm
);
1912 update_hiwater_rss(vma
->vm_mm
);
1913 mmu_notifier_invalidate_range_start(&range
);
1915 * unmap 'address-end' not 'range.start-range.end' as range
1916 * could have been expanded for hugetlb pmd sharing.
1918 unmap_single_vma(&tlb
, vma
, address
, end
, details
, false);
1919 mmu_notifier_invalidate_range_end(&range
);
1920 tlb_finish_mmu(&tlb
);
1921 hugetlb_zap_end(vma
, details
);
1925 * zap_vma_ptes - remove ptes mapping the vma
1926 * @vma: vm_area_struct holding ptes to be zapped
1927 * @address: starting address of pages to zap
1928 * @size: number of bytes to zap
1930 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1932 * The entire address range must be fully contained within the vma.
1935 void zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1938 if (!range_in_vma(vma
, address
, address
+ size
) ||
1939 !(vma
->vm_flags
& VM_PFNMAP
))
1942 zap_page_range_single(vma
, address
, size
, NULL
);
1944 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1946 static pmd_t
*walk_to_pmd(struct mm_struct
*mm
, unsigned long addr
)
1953 pgd
= pgd_offset(mm
, addr
);
1954 p4d
= p4d_alloc(mm
, pgd
, addr
);
1957 pud
= pud_alloc(mm
, p4d
, addr
);
1960 pmd
= pmd_alloc(mm
, pud
, addr
);
1964 VM_BUG_ON(pmd_trans_huge(*pmd
));
1968 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1971 pmd_t
*pmd
= walk_to_pmd(mm
, addr
);
1975 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1978 static int validate_page_before_insert(struct page
*page
)
1980 struct folio
*folio
= page_folio(page
);
1982 if (folio_test_anon(folio
) || folio_test_slab(folio
) ||
1983 page_has_type(page
))
1985 flush_dcache_folio(folio
);
1989 static int insert_page_into_pte_locked(struct vm_area_struct
*vma
, pte_t
*pte
,
1990 unsigned long addr
, struct page
*page
, pgprot_t prot
)
1992 struct folio
*folio
= page_folio(page
);
1994 if (!pte_none(ptep_get(pte
)))
1996 /* Ok, finally just insert the thing.. */
1998 inc_mm_counter(vma
->vm_mm
, mm_counter_file(folio
));
1999 folio_add_file_rmap_pte(folio
, page
, vma
);
2000 set_pte_at(vma
->vm_mm
, addr
, pte
, mk_pte(page
, prot
));
2005 * This is the old fallback for page remapping.
2007 * For historical reasons, it only allows reserved pages. Only
2008 * old drivers should use this, and they needed to mark their
2009 * pages reserved for the old functions anyway.
2011 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2012 struct page
*page
, pgprot_t prot
)
2018 retval
= validate_page_before_insert(page
);
2022 pte
= get_locked_pte(vma
->vm_mm
, addr
, &ptl
);
2025 retval
= insert_page_into_pte_locked(vma
, pte
, addr
, page
, prot
);
2026 pte_unmap_unlock(pte
, ptl
);
2031 static int insert_page_in_batch_locked(struct vm_area_struct
*vma
, pte_t
*pte
,
2032 unsigned long addr
, struct page
*page
, pgprot_t prot
)
2036 if (!page_count(page
))
2038 err
= validate_page_before_insert(page
);
2041 return insert_page_into_pte_locked(vma
, pte
, addr
, page
, prot
);
2044 /* insert_pages() amortizes the cost of spinlock operations
2045 * when inserting pages in a loop.
2047 static int insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
2048 struct page
**pages
, unsigned long *num
, pgprot_t prot
)
2051 pte_t
*start_pte
, *pte
;
2052 spinlock_t
*pte_lock
;
2053 struct mm_struct
*const mm
= vma
->vm_mm
;
2054 unsigned long curr_page_idx
= 0;
2055 unsigned long remaining_pages_total
= *num
;
2056 unsigned long pages_to_write_in_pmd
;
2060 pmd
= walk_to_pmd(mm
, addr
);
2064 pages_to_write_in_pmd
= min_t(unsigned long,
2065 remaining_pages_total
, PTRS_PER_PTE
- pte_index(addr
));
2067 /* Allocate the PTE if necessary; takes PMD lock once only. */
2069 if (pte_alloc(mm
, pmd
))
2072 while (pages_to_write_in_pmd
) {
2074 const int batch_size
= min_t(int, pages_to_write_in_pmd
, 8);
2076 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &pte_lock
);
2081 for (pte
= start_pte
; pte_idx
< batch_size
; ++pte
, ++pte_idx
) {
2082 int err
= insert_page_in_batch_locked(vma
, pte
,
2083 addr
, pages
[curr_page_idx
], prot
);
2084 if (unlikely(err
)) {
2085 pte_unmap_unlock(start_pte
, pte_lock
);
2087 remaining_pages_total
-= pte_idx
;
2093 pte_unmap_unlock(start_pte
, pte_lock
);
2094 pages_to_write_in_pmd
-= batch_size
;
2095 remaining_pages_total
-= batch_size
;
2097 if (remaining_pages_total
)
2101 *num
= remaining_pages_total
;
2106 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
2107 * @vma: user vma to map to
2108 * @addr: target start user address of these pages
2109 * @pages: source kernel pages
2110 * @num: in: number of pages to map. out: number of pages that were *not*
2111 * mapped. (0 means all pages were successfully mapped).
2113 * Preferred over vm_insert_page() when inserting multiple pages.
2115 * In case of error, we may have mapped a subset of the provided
2116 * pages. It is the caller's responsibility to account for this case.
2118 * The same restrictions apply as in vm_insert_page().
2120 int vm_insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
2121 struct page
**pages
, unsigned long *num
)
2123 const unsigned long end_addr
= addr
+ (*num
* PAGE_SIZE
) - 1;
2125 if (addr
< vma
->vm_start
|| end_addr
>= vma
->vm_end
)
2127 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
2128 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
2129 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
2130 vm_flags_set(vma
, VM_MIXEDMAP
);
2132 /* Defer page refcount checking till we're about to map that page. */
2133 return insert_pages(vma
, addr
, pages
, num
, vma
->vm_page_prot
);
2135 EXPORT_SYMBOL(vm_insert_pages
);
2138 * vm_insert_page - insert single page into user vma
2139 * @vma: user vma to map to
2140 * @addr: target user address of this page
2141 * @page: source kernel page
2143 * This allows drivers to insert individual pages they've allocated
2146 * The page has to be a nice clean _individual_ kernel allocation.
2147 * If you allocate a compound page, you need to have marked it as
2148 * such (__GFP_COMP), or manually just split the page up yourself
2149 * (see split_page()).
2151 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2152 * took an arbitrary page protection parameter. This doesn't allow
2153 * that. Your vma protection will have to be set up correctly, which
2154 * means that if you want a shared writable mapping, you'd better
2155 * ask for a shared writable mapping!
2157 * The page does not need to be reserved.
2159 * Usually this function is called from f_op->mmap() handler
2160 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
2161 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2162 * function from other places, for example from page-fault handler.
2164 * Return: %0 on success, negative error code otherwise.
2166 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2169 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2171 if (!page_count(page
))
2173 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
2174 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
2175 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
2176 vm_flags_set(vma
, VM_MIXEDMAP
);
2178 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2180 EXPORT_SYMBOL(vm_insert_page
);
2183 * __vm_map_pages - maps range of kernel pages into user vma
2184 * @vma: user vma to map to
2185 * @pages: pointer to array of source kernel pages
2186 * @num: number of pages in page array
2187 * @offset: user's requested vm_pgoff
2189 * This allows drivers to map range of kernel pages into a user vma.
2191 * Return: 0 on success and error code otherwise.
2193 static int __vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
2194 unsigned long num
, unsigned long offset
)
2196 unsigned long count
= vma_pages(vma
);
2197 unsigned long uaddr
= vma
->vm_start
;
2200 /* Fail if the user requested offset is beyond the end of the object */
2204 /* Fail if the user requested size exceeds available object size */
2205 if (count
> num
- offset
)
2208 for (i
= 0; i
< count
; i
++) {
2209 ret
= vm_insert_page(vma
, uaddr
, pages
[offset
+ i
]);
2219 * vm_map_pages - maps range of kernel pages starts with non zero offset
2220 * @vma: user vma to map to
2221 * @pages: pointer to array of source kernel pages
2222 * @num: number of pages in page array
2224 * Maps an object consisting of @num pages, catering for the user's
2225 * requested vm_pgoff
2227 * If we fail to insert any page into the vma, the function will return
2228 * immediately leaving any previously inserted pages present. Callers
2229 * from the mmap handler may immediately return the error as their caller
2230 * will destroy the vma, removing any successfully inserted pages. Other
2231 * callers should make their own arrangements for calling unmap_region().
2233 * Context: Process context. Called by mmap handlers.
2234 * Return: 0 on success and error code otherwise.
2236 int vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
2239 return __vm_map_pages(vma
, pages
, num
, vma
->vm_pgoff
);
2241 EXPORT_SYMBOL(vm_map_pages
);
2244 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2245 * @vma: user vma to map to
2246 * @pages: pointer to array of source kernel pages
2247 * @num: number of pages in page array
2249 * Similar to vm_map_pages(), except that it explicitly sets the offset
2250 * to 0. This function is intended for the drivers that did not consider
2253 * Context: Process context. Called by mmap handlers.
2254 * Return: 0 on success and error code otherwise.
2256 int vm_map_pages_zero(struct vm_area_struct
*vma
, struct page
**pages
,
2259 return __vm_map_pages(vma
, pages
, num
, 0);
2261 EXPORT_SYMBOL(vm_map_pages_zero
);
2263 static vm_fault_t
insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2264 pfn_t pfn
, pgprot_t prot
, bool mkwrite
)
2266 struct mm_struct
*mm
= vma
->vm_mm
;
2270 pte
= get_locked_pte(mm
, addr
, &ptl
);
2272 return VM_FAULT_OOM
;
2273 entry
= ptep_get(pte
);
2274 if (!pte_none(entry
)) {
2277 * For read faults on private mappings the PFN passed
2278 * in may not match the PFN we have mapped if the
2279 * mapped PFN is a writeable COW page. In the mkwrite
2280 * case we are creating a writable PTE for a shared
2281 * mapping and we expect the PFNs to match. If they
2282 * don't match, we are likely racing with block
2283 * allocation and mapping invalidation so just skip the
2286 if (pte_pfn(entry
) != pfn_t_to_pfn(pfn
)) {
2287 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry
)));
2290 entry
= pte_mkyoung(entry
);
2291 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2292 if (ptep_set_access_flags(vma
, addr
, pte
, entry
, 1))
2293 update_mmu_cache(vma
, addr
, pte
);
2298 /* Ok, finally just insert the thing.. */
2299 if (pfn_t_devmap(pfn
))
2300 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
2302 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
2305 entry
= pte_mkyoung(entry
);
2306 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2309 set_pte_at(mm
, addr
, pte
, entry
);
2310 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
2313 pte_unmap_unlock(pte
, ptl
);
2314 return VM_FAULT_NOPAGE
;
2318 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2319 * @vma: user vma to map to
2320 * @addr: target user address of this page
2321 * @pfn: source kernel pfn
2322 * @pgprot: pgprot flags for the inserted page
2324 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2325 * to override pgprot on a per-page basis.
2327 * This only makes sense for IO mappings, and it makes no sense for
2328 * COW mappings. In general, using multiple vmas is preferable;
2329 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2332 * pgprot typically only differs from @vma->vm_page_prot when drivers set
2333 * caching- and encryption bits different than those of @vma->vm_page_prot,
2334 * because the caching- or encryption mode may not be known at mmap() time.
2336 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2337 * to set caching and encryption bits for those vmas (except for COW pages).
2338 * This is ensured by core vm only modifying these page table entries using
2339 * functions that don't touch caching- or encryption bits, using pte_modify()
2340 * if needed. (See for example mprotect()).
2342 * Also when new page-table entries are created, this is only done using the
2343 * fault() callback, and never using the value of vma->vm_page_prot,
2344 * except for page-table entries that point to anonymous pages as the result
2347 * Context: Process context. May allocate using %GFP_KERNEL.
2348 * Return: vm_fault_t value.
2350 vm_fault_t
vmf_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
2351 unsigned long pfn
, pgprot_t pgprot
)
2354 * Technically, architectures with pte_special can avoid all these
2355 * restrictions (same for remap_pfn_range). However we would like
2356 * consistency in testing and feature parity among all, so we should
2357 * try to keep these invariants in place for everybody.
2359 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
2360 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
2361 (VM_PFNMAP
|VM_MIXEDMAP
));
2362 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
2363 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
2365 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2366 return VM_FAULT_SIGBUS
;
2368 if (!pfn_modify_allowed(pfn
, pgprot
))
2369 return VM_FAULT_SIGBUS
;
2371 track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
));
2373 return insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
,
2376 EXPORT_SYMBOL(vmf_insert_pfn_prot
);
2379 * vmf_insert_pfn - insert single pfn into user vma
2380 * @vma: user vma to map to
2381 * @addr: target user address of this page
2382 * @pfn: source kernel pfn
2384 * Similar to vm_insert_page, this allows drivers to insert individual pages
2385 * they've allocated into a user vma. Same comments apply.
2387 * This function should only be called from a vm_ops->fault handler, and
2388 * in that case the handler should return the result of this function.
2390 * vma cannot be a COW mapping.
2392 * As this is called only for pages that do not currently exist, we
2393 * do not need to flush old virtual caches or the TLB.
2395 * Context: Process context. May allocate using %GFP_KERNEL.
2396 * Return: vm_fault_t value.
2398 vm_fault_t
vmf_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2401 return vmf_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
2403 EXPORT_SYMBOL(vmf_insert_pfn
);
2405 static bool vm_mixed_ok(struct vm_area_struct
*vma
, pfn_t pfn
)
2407 /* these checks mirror the abort conditions in vm_normal_page */
2408 if (vma
->vm_flags
& VM_MIXEDMAP
)
2410 if (pfn_t_devmap(pfn
))
2412 if (pfn_t_special(pfn
))
2414 if (is_zero_pfn(pfn_t_to_pfn(pfn
)))
2419 static vm_fault_t
__vm_insert_mixed(struct vm_area_struct
*vma
,
2420 unsigned long addr
, pfn_t pfn
, bool mkwrite
)
2422 pgprot_t pgprot
= vma
->vm_page_prot
;
2425 BUG_ON(!vm_mixed_ok(vma
, pfn
));
2427 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2428 return VM_FAULT_SIGBUS
;
2430 track_pfn_insert(vma
, &pgprot
, pfn
);
2432 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn
), pgprot
))
2433 return VM_FAULT_SIGBUS
;
2436 * If we don't have pte special, then we have to use the pfn_valid()
2437 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2438 * refcount the page if pfn_valid is true (hence insert_page rather
2439 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2440 * without pte special, it would there be refcounted as a normal page.
2442 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
) &&
2443 !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
2447 * At this point we are committed to insert_page()
2448 * regardless of whether the caller specified flags that
2449 * result in pfn_t_has_page() == false.
2451 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
2452 err
= insert_page(vma
, addr
, page
, pgprot
);
2454 return insert_pfn(vma
, addr
, pfn
, pgprot
, mkwrite
);
2458 return VM_FAULT_OOM
;
2459 if (err
< 0 && err
!= -EBUSY
)
2460 return VM_FAULT_SIGBUS
;
2462 return VM_FAULT_NOPAGE
;
2465 vm_fault_t
vmf_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2468 return __vm_insert_mixed(vma
, addr
, pfn
, false);
2470 EXPORT_SYMBOL(vmf_insert_mixed
);
2473 * If the insertion of PTE failed because someone else already added a
2474 * different entry in the mean time, we treat that as success as we assume
2475 * the same entry was actually inserted.
2477 vm_fault_t
vmf_insert_mixed_mkwrite(struct vm_area_struct
*vma
,
2478 unsigned long addr
, pfn_t pfn
)
2480 return __vm_insert_mixed(vma
, addr
, pfn
, true);
2482 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite
);
2485 * maps a range of physical memory into the requested pages. the old
2486 * mappings are removed. any references to nonexistent pages results
2487 * in null mappings (currently treated as "copy-on-access")
2489 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2490 unsigned long addr
, unsigned long end
,
2491 unsigned long pfn
, pgprot_t prot
)
2493 pte_t
*pte
, *mapped_pte
;
2497 mapped_pte
= pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2500 arch_enter_lazy_mmu_mode();
2502 BUG_ON(!pte_none(ptep_get(pte
)));
2503 if (!pfn_modify_allowed(pfn
, prot
)) {
2507 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2509 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2510 arch_leave_lazy_mmu_mode();
2511 pte_unmap_unlock(mapped_pte
, ptl
);
2515 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2516 unsigned long addr
, unsigned long end
,
2517 unsigned long pfn
, pgprot_t prot
)
2523 pfn
-= addr
>> PAGE_SHIFT
;
2524 pmd
= pmd_alloc(mm
, pud
, addr
);
2527 VM_BUG_ON(pmd_trans_huge(*pmd
));
2529 next
= pmd_addr_end(addr
, end
);
2530 err
= remap_pte_range(mm
, pmd
, addr
, next
,
2531 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2534 } while (pmd
++, addr
= next
, addr
!= end
);
2538 static inline int remap_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2539 unsigned long addr
, unsigned long end
,
2540 unsigned long pfn
, pgprot_t prot
)
2546 pfn
-= addr
>> PAGE_SHIFT
;
2547 pud
= pud_alloc(mm
, p4d
, addr
);
2551 next
= pud_addr_end(addr
, end
);
2552 err
= remap_pmd_range(mm
, pud
, addr
, next
,
2553 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2556 } while (pud
++, addr
= next
, addr
!= end
);
2560 static inline int remap_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2561 unsigned long addr
, unsigned long end
,
2562 unsigned long pfn
, pgprot_t prot
)
2568 pfn
-= addr
>> PAGE_SHIFT
;
2569 p4d
= p4d_alloc(mm
, pgd
, addr
);
2573 next
= p4d_addr_end(addr
, end
);
2574 err
= remap_pud_range(mm
, p4d
, addr
, next
,
2575 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2578 } while (p4d
++, addr
= next
, addr
!= end
);
2583 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2584 * must have pre-validated the caching bits of the pgprot_t.
2586 int remap_pfn_range_notrack(struct vm_area_struct
*vma
, unsigned long addr
,
2587 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2591 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2592 struct mm_struct
*mm
= vma
->vm_mm
;
2595 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr
)))
2599 * Physically remapped pages are special. Tell the
2600 * rest of the world about it:
2601 * VM_IO tells people not to look at these pages
2602 * (accesses can have side effects).
2603 * VM_PFNMAP tells the core MM that the base pages are just
2604 * raw PFN mappings, and do not have a "struct page" associated
2607 * Disable vma merging and expanding with mremap().
2609 * Omit vma from core dump, even when VM_IO turned off.
2611 * There's a horrible special case to handle copy-on-write
2612 * behaviour that some programs depend on. We mark the "original"
2613 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2614 * See vm_normal_page() for details.
2616 if (is_cow_mapping(vma
->vm_flags
)) {
2617 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2619 vma
->vm_pgoff
= pfn
;
2622 vm_flags_set(vma
, VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
);
2624 BUG_ON(addr
>= end
);
2625 pfn
-= addr
>> PAGE_SHIFT
;
2626 pgd
= pgd_offset(mm
, addr
);
2627 flush_cache_range(vma
, addr
, end
);
2629 next
= pgd_addr_end(addr
, end
);
2630 err
= remap_p4d_range(mm
, pgd
, addr
, next
,
2631 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2634 } while (pgd
++, addr
= next
, addr
!= end
);
2640 * remap_pfn_range - remap kernel memory to userspace
2641 * @vma: user vma to map to
2642 * @addr: target page aligned user address to start at
2643 * @pfn: page frame number of kernel physical memory address
2644 * @size: size of mapping area
2645 * @prot: page protection flags for this mapping
2647 * Note: this is only safe if the mm semaphore is held when called.
2649 * Return: %0 on success, negative error code otherwise.
2651 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2652 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2656 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
2660 err
= remap_pfn_range_notrack(vma
, addr
, pfn
, size
, prot
);
2662 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
), true);
2665 EXPORT_SYMBOL(remap_pfn_range
);
2668 * vm_iomap_memory - remap memory to userspace
2669 * @vma: user vma to map to
2670 * @start: start of the physical memory to be mapped
2671 * @len: size of area
2673 * This is a simplified io_remap_pfn_range() for common driver use. The
2674 * driver just needs to give us the physical memory range to be mapped,
2675 * we'll figure out the rest from the vma information.
2677 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2678 * whatever write-combining details or similar.
2680 * Return: %0 on success, negative error code otherwise.
2682 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2684 unsigned long vm_len
, pfn
, pages
;
2686 /* Check that the physical memory area passed in looks valid */
2687 if (start
+ len
< start
)
2690 * You *really* shouldn't map things that aren't page-aligned,
2691 * but we've historically allowed it because IO memory might
2692 * just have smaller alignment.
2694 len
+= start
& ~PAGE_MASK
;
2695 pfn
= start
>> PAGE_SHIFT
;
2696 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2697 if (pfn
+ pages
< pfn
)
2700 /* We start the mapping 'vm_pgoff' pages into the area */
2701 if (vma
->vm_pgoff
> pages
)
2703 pfn
+= vma
->vm_pgoff
;
2704 pages
-= vma
->vm_pgoff
;
2706 /* Can we fit all of the mapping? */
2707 vm_len
= vma
->vm_end
- vma
->vm_start
;
2708 if (vm_len
>> PAGE_SHIFT
> pages
)
2711 /* Ok, let it rip */
2712 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2714 EXPORT_SYMBOL(vm_iomap_memory
);
2716 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2717 unsigned long addr
, unsigned long end
,
2718 pte_fn_t fn
, void *data
, bool create
,
2719 pgtbl_mod_mask
*mask
)
2721 pte_t
*pte
, *mapped_pte
;
2726 mapped_pte
= pte
= (mm
== &init_mm
) ?
2727 pte_alloc_kernel_track(pmd
, addr
, mask
) :
2728 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2732 mapped_pte
= pte
= (mm
== &init_mm
) ?
2733 pte_offset_kernel(pmd
, addr
) :
2734 pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
2739 arch_enter_lazy_mmu_mode();
2743 if (create
|| !pte_none(ptep_get(pte
))) {
2744 err
= fn(pte
++, addr
, data
);
2748 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2750 *mask
|= PGTBL_PTE_MODIFIED
;
2752 arch_leave_lazy_mmu_mode();
2755 pte_unmap_unlock(mapped_pte
, ptl
);
2759 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2760 unsigned long addr
, unsigned long end
,
2761 pte_fn_t fn
, void *data
, bool create
,
2762 pgtbl_mod_mask
*mask
)
2768 BUG_ON(pud_huge(*pud
));
2771 pmd
= pmd_alloc_track(mm
, pud
, addr
, mask
);
2775 pmd
= pmd_offset(pud
, addr
);
2778 next
= pmd_addr_end(addr
, end
);
2779 if (pmd_none(*pmd
) && !create
)
2781 if (WARN_ON_ONCE(pmd_leaf(*pmd
)))
2783 if (!pmd_none(*pmd
) && WARN_ON_ONCE(pmd_bad(*pmd
))) {
2788 err
= apply_to_pte_range(mm
, pmd
, addr
, next
,
2789 fn
, data
, create
, mask
);
2792 } while (pmd
++, addr
= next
, addr
!= end
);
2797 static int apply_to_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2798 unsigned long addr
, unsigned long end
,
2799 pte_fn_t fn
, void *data
, bool create
,
2800 pgtbl_mod_mask
*mask
)
2807 pud
= pud_alloc_track(mm
, p4d
, addr
, mask
);
2811 pud
= pud_offset(p4d
, addr
);
2814 next
= pud_addr_end(addr
, end
);
2815 if (pud_none(*pud
) && !create
)
2817 if (WARN_ON_ONCE(pud_leaf(*pud
)))
2819 if (!pud_none(*pud
) && WARN_ON_ONCE(pud_bad(*pud
))) {
2824 err
= apply_to_pmd_range(mm
, pud
, addr
, next
,
2825 fn
, data
, create
, mask
);
2828 } while (pud
++, addr
= next
, addr
!= end
);
2833 static int apply_to_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2834 unsigned long addr
, unsigned long end
,
2835 pte_fn_t fn
, void *data
, bool create
,
2836 pgtbl_mod_mask
*mask
)
2843 p4d
= p4d_alloc_track(mm
, pgd
, addr
, mask
);
2847 p4d
= p4d_offset(pgd
, addr
);
2850 next
= p4d_addr_end(addr
, end
);
2851 if (p4d_none(*p4d
) && !create
)
2853 if (WARN_ON_ONCE(p4d_leaf(*p4d
)))
2855 if (!p4d_none(*p4d
) && WARN_ON_ONCE(p4d_bad(*p4d
))) {
2860 err
= apply_to_pud_range(mm
, p4d
, addr
, next
,
2861 fn
, data
, create
, mask
);
2864 } while (p4d
++, addr
= next
, addr
!= end
);
2869 static int __apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2870 unsigned long size
, pte_fn_t fn
,
2871 void *data
, bool create
)
2874 unsigned long start
= addr
, next
;
2875 unsigned long end
= addr
+ size
;
2876 pgtbl_mod_mask mask
= 0;
2879 if (WARN_ON(addr
>= end
))
2882 pgd
= pgd_offset(mm
, addr
);
2884 next
= pgd_addr_end(addr
, end
);
2885 if (pgd_none(*pgd
) && !create
)
2887 if (WARN_ON_ONCE(pgd_leaf(*pgd
)))
2889 if (!pgd_none(*pgd
) && WARN_ON_ONCE(pgd_bad(*pgd
))) {
2894 err
= apply_to_p4d_range(mm
, pgd
, addr
, next
,
2895 fn
, data
, create
, &mask
);
2898 } while (pgd
++, addr
= next
, addr
!= end
);
2900 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
2901 arch_sync_kernel_mappings(start
, start
+ size
);
2907 * Scan a region of virtual memory, filling in page tables as necessary
2908 * and calling a provided function on each leaf page table.
2910 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2911 unsigned long size
, pte_fn_t fn
, void *data
)
2913 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, true);
2915 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2918 * Scan a region of virtual memory, calling a provided function on
2919 * each leaf page table where it exists.
2921 * Unlike apply_to_page_range, this does _not_ fill in page tables
2922 * where they are absent.
2924 int apply_to_existing_page_range(struct mm_struct
*mm
, unsigned long addr
,
2925 unsigned long size
, pte_fn_t fn
, void *data
)
2927 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, false);
2929 EXPORT_SYMBOL_GPL(apply_to_existing_page_range
);
2932 * handle_pte_fault chooses page fault handler according to an entry which was
2933 * read non-atomically. Before making any commitment, on those architectures
2934 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2935 * parts, do_swap_page must check under lock before unmapping the pte and
2936 * proceeding (but do_wp_page is only called after already making such a check;
2937 * and do_anonymous_page can safely check later on).
2939 static inline int pte_unmap_same(struct vm_fault
*vmf
)
2942 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2943 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2944 spin_lock(vmf
->ptl
);
2945 same
= pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
);
2946 spin_unlock(vmf
->ptl
);
2949 pte_unmap(vmf
->pte
);
2956 * 0: copied succeeded
2957 * -EHWPOISON: copy failed due to hwpoison in source page
2958 * -EAGAIN: copied failed (some other reason)
2960 static inline int __wp_page_copy_user(struct page
*dst
, struct page
*src
,
2961 struct vm_fault
*vmf
)
2966 struct vm_area_struct
*vma
= vmf
->vma
;
2967 struct mm_struct
*mm
= vma
->vm_mm
;
2968 unsigned long addr
= vmf
->address
;
2971 if (copy_mc_user_highpage(dst
, src
, addr
, vma
)) {
2972 memory_failure_queue(page_to_pfn(src
), 0);
2979 * If the source page was a PFN mapping, we don't have
2980 * a "struct page" for it. We do a best-effort copy by
2981 * just copying from the original user address. If that
2982 * fails, we just zero-fill it. Live with it.
2984 kaddr
= kmap_local_page(dst
);
2985 pagefault_disable();
2986 uaddr
= (void __user
*)(addr
& PAGE_MASK
);
2989 * On architectures with software "accessed" bits, we would
2990 * take a double page fault, so mark it accessed here.
2993 if (!arch_has_hw_pte_young() && !pte_young(vmf
->orig_pte
)) {
2996 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2997 if (unlikely(!vmf
->pte
|| !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
2999 * Other thread has already handled the fault
3000 * and update local tlb only
3003 update_mmu_tlb(vma
, addr
, vmf
->pte
);
3008 entry
= pte_mkyoung(vmf
->orig_pte
);
3009 if (ptep_set_access_flags(vma
, addr
, vmf
->pte
, entry
, 0))
3010 update_mmu_cache_range(vmf
, vma
, addr
, vmf
->pte
, 1);
3014 * This really shouldn't fail, because the page is there
3015 * in the page tables. But it might just be unreadable,
3016 * in which case we just give up and fill the result with
3019 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
3023 /* Re-validate under PTL if the page is still mapped */
3024 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
3025 if (unlikely(!vmf
->pte
|| !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
3026 /* The PTE changed under us, update local tlb */
3028 update_mmu_tlb(vma
, addr
, vmf
->pte
);
3034 * The same page can be mapped back since last copy attempt.
3035 * Try to copy again under PTL.
3037 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
3039 * Give a warn in case there can be some obscure
3052 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3054 kunmap_local(kaddr
);
3055 flush_dcache_page(dst
);
3060 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
3062 struct file
*vm_file
= vma
->vm_file
;
3065 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
3068 * Special mappings (e.g. VDSO) do not have any file so fake
3069 * a default GFP_KERNEL for them.
3075 * Notify the address space that the page is about to become writable so that
3076 * it can prohibit this or wait for the page to get into an appropriate state.
3078 * We do this without the lock held, so that it can sleep if it needs to.
3080 static vm_fault_t
do_page_mkwrite(struct vm_fault
*vmf
, struct folio
*folio
)
3083 unsigned int old_flags
= vmf
->flags
;
3085 vmf
->flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
3087 if (vmf
->vma
->vm_file
&&
3088 IS_SWAPFILE(vmf
->vma
->vm_file
->f_mapping
->host
))
3089 return VM_FAULT_SIGBUS
;
3091 ret
= vmf
->vma
->vm_ops
->page_mkwrite(vmf
);
3092 /* Restore original flags so that caller is not surprised */
3093 vmf
->flags
= old_flags
;
3094 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
3096 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
3098 if (!folio
->mapping
) {
3099 folio_unlock(folio
);
3100 return 0; /* retry */
3102 ret
|= VM_FAULT_LOCKED
;
3104 VM_BUG_ON_FOLIO(!folio_test_locked(folio
), folio
);
3109 * Handle dirtying of a page in shared file mapping on a write fault.
3111 * The function expects the page to be locked and unlocks it.
3113 static vm_fault_t
fault_dirty_shared_page(struct vm_fault
*vmf
)
3115 struct vm_area_struct
*vma
= vmf
->vma
;
3116 struct address_space
*mapping
;
3117 struct folio
*folio
= page_folio(vmf
->page
);
3119 bool page_mkwrite
= vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
;
3121 dirtied
= folio_mark_dirty(folio
);
3122 VM_BUG_ON_FOLIO(folio_test_anon(folio
), folio
);
3124 * Take a local copy of the address_space - folio.mapping may be zeroed
3125 * by truncate after folio_unlock(). The address_space itself remains
3126 * pinned by vma->vm_file's reference. We rely on folio_unlock()'s
3127 * release semantics to prevent the compiler from undoing this copying.
3129 mapping
= folio_raw_mapping(folio
);
3130 folio_unlock(folio
);
3133 file_update_time(vma
->vm_file
);
3136 * Throttle page dirtying rate down to writeback speed.
3138 * mapping may be NULL here because some device drivers do not
3139 * set page.mapping but still dirty their pages
3141 * Drop the mmap_lock before waiting on IO, if we can. The file
3142 * is pinning the mapping, as per above.
3144 if ((dirtied
|| page_mkwrite
) && mapping
) {
3147 fpin
= maybe_unlock_mmap_for_io(vmf
, NULL
);
3148 balance_dirty_pages_ratelimited(mapping
);
3151 return VM_FAULT_COMPLETED
;
3159 * Handle write page faults for pages that can be reused in the current vma
3161 * This can happen either due to the mapping being with the VM_SHARED flag,
3162 * or due to us being the last reference standing to the page. In either
3163 * case, all we need to do here is to mark the page as writable and update
3164 * any related book-keeping.
3166 static inline void wp_page_reuse(struct vm_fault
*vmf
, struct folio
*folio
)
3167 __releases(vmf
->ptl
)
3169 struct vm_area_struct
*vma
= vmf
->vma
;
3172 VM_BUG_ON(!(vmf
->flags
& FAULT_FLAG_WRITE
));
3175 VM_BUG_ON(folio_test_anon(folio
) &&
3176 !PageAnonExclusive(vmf
->page
));
3178 * Clear the folio's cpupid information as the existing
3179 * information potentially belongs to a now completely
3180 * unrelated process.
3182 folio_xchg_last_cpupid(folio
, (1 << LAST_CPUPID_SHIFT
) - 1);
3185 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
3186 entry
= pte_mkyoung(vmf
->orig_pte
);
3187 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3188 if (ptep_set_access_flags(vma
, vmf
->address
, vmf
->pte
, entry
, 1))
3189 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
3190 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3191 count_vm_event(PGREUSE
);
3195 * We could add a bitflag somewhere, but for now, we know that all
3196 * vm_ops that have a ->map_pages have been audited and don't need
3197 * the mmap_lock to be held.
3199 static inline vm_fault_t
vmf_can_call_fault(const struct vm_fault
*vmf
)
3201 struct vm_area_struct
*vma
= vmf
->vma
;
3203 if (vma
->vm_ops
->map_pages
|| !(vmf
->flags
& FAULT_FLAG_VMA_LOCK
))
3206 return VM_FAULT_RETRY
;
3209 vm_fault_t
vmf_anon_prepare(struct vm_fault
*vmf
)
3211 struct vm_area_struct
*vma
= vmf
->vma
;
3213 if (likely(vma
->anon_vma
))
3215 if (vmf
->flags
& FAULT_FLAG_VMA_LOCK
) {
3217 return VM_FAULT_RETRY
;
3219 if (__anon_vma_prepare(vma
))
3220 return VM_FAULT_OOM
;
3225 * Handle the case of a page which we actually need to copy to a new page,
3226 * either due to COW or unsharing.
3228 * Called with mmap_lock locked and the old page referenced, but
3229 * without the ptl held.
3231 * High level logic flow:
3233 * - Allocate a page, copy the content of the old page to the new one.
3234 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
3235 * - Take the PTL. If the pte changed, bail out and release the allocated page
3236 * - If the pte is still the way we remember it, update the page table and all
3237 * relevant references. This includes dropping the reference the page-table
3238 * held to the old page, as well as updating the rmap.
3239 * - In any case, unlock the PTL and drop the reference we took to the old page.
3241 static vm_fault_t
wp_page_copy(struct vm_fault
*vmf
)
3243 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
3244 struct vm_area_struct
*vma
= vmf
->vma
;
3245 struct mm_struct
*mm
= vma
->vm_mm
;
3246 struct folio
*old_folio
= NULL
;
3247 struct folio
*new_folio
= NULL
;
3249 int page_copied
= 0;
3250 struct mmu_notifier_range range
;
3254 delayacct_wpcopy_start();
3257 old_folio
= page_folio(vmf
->page
);
3258 ret
= vmf_anon_prepare(vmf
);
3262 pfn_is_zero
= is_zero_pfn(pte_pfn(vmf
->orig_pte
));
3263 new_folio
= folio_prealloc(mm
, vma
, vmf
->address
, pfn_is_zero
);
3270 err
= __wp_page_copy_user(&new_folio
->page
, vmf
->page
, vmf
);
3273 * COW failed, if the fault was solved by other,
3274 * it's fine. If not, userspace would re-fault on
3275 * the same address and we will handle the fault
3276 * from the second attempt.
3277 * The -EHWPOISON case will not be retried.
3279 folio_put(new_folio
);
3281 folio_put(old_folio
);
3283 delayacct_wpcopy_end();
3284 return err
== -EHWPOISON
? VM_FAULT_HWPOISON
: 0;
3286 kmsan_copy_page_meta(&new_folio
->page
, vmf
->page
);
3289 __folio_mark_uptodate(new_folio
);
3291 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, mm
,
3292 vmf
->address
& PAGE_MASK
,
3293 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
);
3294 mmu_notifier_invalidate_range_start(&range
);
3297 * Re-check the pte - we dropped the lock
3299 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, vmf
->address
, &vmf
->ptl
);
3300 if (likely(vmf
->pte
&& pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
3302 if (!folio_test_anon(old_folio
)) {
3303 dec_mm_counter(mm
, mm_counter_file(old_folio
));
3304 inc_mm_counter(mm
, MM_ANONPAGES
);
3307 ksm_might_unmap_zero_page(mm
, vmf
->orig_pte
);
3308 inc_mm_counter(mm
, MM_ANONPAGES
);
3310 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
3311 entry
= mk_pte(&new_folio
->page
, vma
->vm_page_prot
);
3312 entry
= pte_sw_mkyoung(entry
);
3313 if (unlikely(unshare
)) {
3314 if (pte_soft_dirty(vmf
->orig_pte
))
3315 entry
= pte_mksoft_dirty(entry
);
3316 if (pte_uffd_wp(vmf
->orig_pte
))
3317 entry
= pte_mkuffd_wp(entry
);
3319 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3323 * Clear the pte entry and flush it first, before updating the
3324 * pte with the new entry, to keep TLBs on different CPUs in
3325 * sync. This code used to set the new PTE then flush TLBs, but
3326 * that left a window where the new PTE could be loaded into
3327 * some TLBs while the old PTE remains in others.
3329 ptep_clear_flush(vma
, vmf
->address
, vmf
->pte
);
3330 folio_add_new_anon_rmap(new_folio
, vma
, vmf
->address
);
3331 folio_add_lru_vma(new_folio
, vma
);
3333 * We call the notify macro here because, when using secondary
3334 * mmu page tables (such as kvm shadow page tables), we want the
3335 * new page to be mapped directly into the secondary page table.
3337 BUG_ON(unshare
&& pte_write(entry
));
3338 set_pte_at_notify(mm
, vmf
->address
, vmf
->pte
, entry
);
3339 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
3342 * Only after switching the pte to the new page may
3343 * we remove the mapcount here. Otherwise another
3344 * process may come and find the rmap count decremented
3345 * before the pte is switched to the new page, and
3346 * "reuse" the old page writing into it while our pte
3347 * here still points into it and can be read by other
3350 * The critical issue is to order this
3351 * folio_remove_rmap_pte() with the ptp_clear_flush
3352 * above. Those stores are ordered by (if nothing else,)
3353 * the barrier present in the atomic_add_negative
3354 * in folio_remove_rmap_pte();
3356 * Then the TLB flush in ptep_clear_flush ensures that
3357 * no process can access the old page before the
3358 * decremented mapcount is visible. And the old page
3359 * cannot be reused until after the decremented
3360 * mapcount is visible. So transitively, TLBs to
3361 * old page will be flushed before it can be reused.
3363 folio_remove_rmap_pte(old_folio
, vmf
->page
, vma
);
3366 /* Free the old page.. */
3367 new_folio
= old_folio
;
3369 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3370 } else if (vmf
->pte
) {
3371 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
3372 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3375 mmu_notifier_invalidate_range_end(&range
);
3378 folio_put(new_folio
);
3381 free_swap_cache(old_folio
);
3382 folio_put(old_folio
);
3385 delayacct_wpcopy_end();
3391 folio_put(old_folio
);
3393 delayacct_wpcopy_end();
3398 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3399 * writeable once the page is prepared
3401 * @vmf: structure describing the fault
3402 * @folio: the folio of vmf->page
3404 * This function handles all that is needed to finish a write page fault in a
3405 * shared mapping due to PTE being read-only once the mapped page is prepared.
3406 * It handles locking of PTE and modifying it.
3408 * The function expects the page to be locked or other protection against
3409 * concurrent faults / writeback (such as DAX radix tree locks).
3411 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3412 * we acquired PTE lock.
3414 static vm_fault_t
finish_mkwrite_fault(struct vm_fault
*vmf
, struct folio
*folio
)
3416 WARN_ON_ONCE(!(vmf
->vma
->vm_flags
& VM_SHARED
));
3417 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3420 return VM_FAULT_NOPAGE
;
3422 * We might have raced with another page fault while we released the
3423 * pte_offset_map_lock.
3425 if (!pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)) {
3426 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
3427 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3428 return VM_FAULT_NOPAGE
;
3430 wp_page_reuse(vmf
, folio
);
3435 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3438 static vm_fault_t
wp_pfn_shared(struct vm_fault
*vmf
)
3440 struct vm_area_struct
*vma
= vmf
->vma
;
3442 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
3445 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3446 ret
= vmf_can_call_fault(vmf
);
3450 vmf
->flags
|= FAULT_FLAG_MKWRITE
;
3451 ret
= vma
->vm_ops
->pfn_mkwrite(vmf
);
3452 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))
3454 return finish_mkwrite_fault(vmf
, NULL
);
3456 wp_page_reuse(vmf
, NULL
);
3460 static vm_fault_t
wp_page_shared(struct vm_fault
*vmf
, struct folio
*folio
)
3461 __releases(vmf
->ptl
)
3463 struct vm_area_struct
*vma
= vmf
->vma
;
3468 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
3471 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3472 tmp
= vmf_can_call_fault(vmf
);
3478 tmp
= do_page_mkwrite(vmf
, folio
);
3479 if (unlikely(!tmp
|| (tmp
&
3480 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3484 tmp
= finish_mkwrite_fault(vmf
, folio
);
3485 if (unlikely(tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3486 folio_unlock(folio
);
3491 wp_page_reuse(vmf
, folio
);
3494 ret
|= fault_dirty_shared_page(vmf
);
3500 static bool wp_can_reuse_anon_folio(struct folio
*folio
,
3501 struct vm_area_struct
*vma
)
3504 * We could currently only reuse a subpage of a large folio if no
3505 * other subpages of the large folios are still mapped. However,
3506 * let's just consistently not reuse subpages even if we could
3507 * reuse in that scenario, and give back a large folio a bit
3510 if (folio_test_large(folio
))
3514 * We have to verify under folio lock: these early checks are
3515 * just an optimization to avoid locking the folio and freeing
3516 * the swapcache if there is little hope that we can reuse.
3518 * KSM doesn't necessarily raise the folio refcount.
3520 if (folio_test_ksm(folio
) || folio_ref_count(folio
) > 3)
3522 if (!folio_test_lru(folio
))
3524 * We cannot easily detect+handle references from
3525 * remote LRU caches or references to LRU folios.
3528 if (folio_ref_count(folio
) > 1 + folio_test_swapcache(folio
))
3530 if (!folio_trylock(folio
))
3532 if (folio_test_swapcache(folio
))
3533 folio_free_swap(folio
);
3534 if (folio_test_ksm(folio
) || folio_ref_count(folio
) != 1) {
3535 folio_unlock(folio
);
3539 * Ok, we've got the only folio reference from our mapping
3540 * and the folio is locked, it's dark out, and we're wearing
3541 * sunglasses. Hit it.
3543 folio_move_anon_rmap(folio
, vma
);
3544 folio_unlock(folio
);
3549 * This routine handles present pages, when
3550 * * users try to write to a shared page (FAULT_FLAG_WRITE)
3551 * * GUP wants to take a R/O pin on a possibly shared anonymous page
3552 * (FAULT_FLAG_UNSHARE)
3554 * It is done by copying the page to a new address and decrementing the
3555 * shared-page counter for the old page.
3557 * Note that this routine assumes that the protection checks have been
3558 * done by the caller (the low-level page fault routine in most cases).
3559 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've
3560 * done any necessary COW.
3562 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even
3563 * though the page will change only once the write actually happens. This
3564 * avoids a few races, and potentially makes it more efficient.
3566 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3567 * but allow concurrent faults), with pte both mapped and locked.
3568 * We return with mmap_lock still held, but pte unmapped and unlocked.
3570 static vm_fault_t
do_wp_page(struct vm_fault
*vmf
)
3571 __releases(vmf
->ptl
)
3573 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
3574 struct vm_area_struct
*vma
= vmf
->vma
;
3575 struct folio
*folio
= NULL
;
3578 if (likely(!unshare
)) {
3579 if (userfaultfd_pte_wp(vma
, ptep_get(vmf
->pte
))) {
3580 if (!userfaultfd_wp_async(vma
)) {
3581 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3582 return handle_userfault(vmf
, VM_UFFD_WP
);
3586 * Nothing needed (cache flush, TLB invalidations,
3587 * etc.) because we're only removing the uffd-wp bit,
3588 * which is completely invisible to the user.
3590 pte
= pte_clear_uffd_wp(ptep_get(vmf
->pte
));
3592 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3594 * Update this to be prepared for following up CoW
3597 vmf
->orig_pte
= pte
;
3601 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3602 * is flushed in this case before copying.
3604 if (unlikely(userfaultfd_wp(vmf
->vma
) &&
3605 mm_tlb_flush_pending(vmf
->vma
->vm_mm
)))
3606 flush_tlb_page(vmf
->vma
, vmf
->address
);
3609 vmf
->page
= vm_normal_page(vma
, vmf
->address
, vmf
->orig_pte
);
3612 folio
= page_folio(vmf
->page
);
3615 * Shared mapping: we are guaranteed to have VM_WRITE and
3616 * FAULT_FLAG_WRITE set at this point.
3618 if (vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
)) {
3620 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3623 * We should not cow pages in a shared writeable mapping.
3624 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3627 return wp_pfn_shared(vmf
);
3628 return wp_page_shared(vmf
, folio
);
3632 * Private mapping: create an exclusive anonymous page copy if reuse
3633 * is impossible. We might miss VM_WRITE for FOLL_FORCE handling.
3635 * If we encounter a page that is marked exclusive, we must reuse
3636 * the page without further checks.
3638 if (folio
&& folio_test_anon(folio
) &&
3639 (PageAnonExclusive(vmf
->page
) || wp_can_reuse_anon_folio(folio
, vma
))) {
3640 if (!PageAnonExclusive(vmf
->page
))
3641 SetPageAnonExclusive(vmf
->page
);
3642 if (unlikely(unshare
)) {
3643 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3646 wp_page_reuse(vmf
, folio
);
3650 * Ok, we need to copy. Oh, well..
3655 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3657 if (folio
&& folio_test_ksm(folio
))
3658 count_vm_event(COW_KSM
);
3660 return wp_page_copy(vmf
);
3663 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
3664 unsigned long start_addr
, unsigned long end_addr
,
3665 struct zap_details
*details
)
3667 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
3670 static inline void unmap_mapping_range_tree(struct rb_root_cached
*root
,
3671 pgoff_t first_index
,
3673 struct zap_details
*details
)
3675 struct vm_area_struct
*vma
;
3676 pgoff_t vba
, vea
, zba
, zea
;
3678 vma_interval_tree_foreach(vma
, root
, first_index
, last_index
) {
3679 vba
= vma
->vm_pgoff
;
3680 vea
= vba
+ vma_pages(vma
) - 1;
3681 zba
= max(first_index
, vba
);
3682 zea
= min(last_index
, vea
);
3684 unmap_mapping_range_vma(vma
,
3685 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
3686 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
3692 * unmap_mapping_folio() - Unmap single folio from processes.
3693 * @folio: The locked folio to be unmapped.
3695 * Unmap this folio from any userspace process which still has it mmaped.
3696 * Typically, for efficiency, the range of nearby pages has already been
3697 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3698 * truncation or invalidation holds the lock on a folio, it may find that
3699 * the page has been remapped again: and then uses unmap_mapping_folio()
3700 * to unmap it finally.
3702 void unmap_mapping_folio(struct folio
*folio
)
3704 struct address_space
*mapping
= folio
->mapping
;
3705 struct zap_details details
= { };
3706 pgoff_t first_index
;
3709 VM_BUG_ON(!folio_test_locked(folio
));
3711 first_index
= folio
->index
;
3712 last_index
= folio_next_index(folio
) - 1;
3714 details
.even_cows
= false;
3715 details
.single_folio
= folio
;
3716 details
.zap_flags
= ZAP_FLAG_DROP_MARKER
;
3718 i_mmap_lock_read(mapping
);
3719 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3720 unmap_mapping_range_tree(&mapping
->i_mmap
, first_index
,
3721 last_index
, &details
);
3722 i_mmap_unlock_read(mapping
);
3726 * unmap_mapping_pages() - Unmap pages from processes.
3727 * @mapping: The address space containing pages to be unmapped.
3728 * @start: Index of first page to be unmapped.
3729 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3730 * @even_cows: Whether to unmap even private COWed pages.
3732 * Unmap the pages in this address space from any userspace process which
3733 * has them mmaped. Generally, you want to remove COWed pages as well when
3734 * a file is being truncated, but not when invalidating pages from the page
3737 void unmap_mapping_pages(struct address_space
*mapping
, pgoff_t start
,
3738 pgoff_t nr
, bool even_cows
)
3740 struct zap_details details
= { };
3741 pgoff_t first_index
= start
;
3742 pgoff_t last_index
= start
+ nr
- 1;
3744 details
.even_cows
= even_cows
;
3745 if (last_index
< first_index
)
3746 last_index
= ULONG_MAX
;
3748 i_mmap_lock_read(mapping
);
3749 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3750 unmap_mapping_range_tree(&mapping
->i_mmap
, first_index
,
3751 last_index
, &details
);
3752 i_mmap_unlock_read(mapping
);
3754 EXPORT_SYMBOL_GPL(unmap_mapping_pages
);
3757 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3758 * address_space corresponding to the specified byte range in the underlying
3761 * @mapping: the address space containing mmaps to be unmapped.
3762 * @holebegin: byte in first page to unmap, relative to the start of
3763 * the underlying file. This will be rounded down to a PAGE_SIZE
3764 * boundary. Note that this is different from truncate_pagecache(), which
3765 * must keep the partial page. In contrast, we must get rid of
3767 * @holelen: size of prospective hole in bytes. This will be rounded
3768 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3770 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3771 * but 0 when invalidating pagecache, don't throw away private data.
3773 void unmap_mapping_range(struct address_space
*mapping
,
3774 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
3776 pgoff_t hba
= (pgoff_t
)(holebegin
) >> PAGE_SHIFT
;
3777 pgoff_t hlen
= ((pgoff_t
)(holelen
) + PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3779 /* Check for overflow. */
3780 if (sizeof(holelen
) > sizeof(hlen
)) {
3782 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3783 if (holeend
& ~(long long)ULONG_MAX
)
3784 hlen
= ULONG_MAX
- hba
+ 1;
3787 unmap_mapping_pages(mapping
, hba
, hlen
, even_cows
);
3789 EXPORT_SYMBOL(unmap_mapping_range
);
3792 * Restore a potential device exclusive pte to a working pte entry
3794 static vm_fault_t
remove_device_exclusive_entry(struct vm_fault
*vmf
)
3796 struct folio
*folio
= page_folio(vmf
->page
);
3797 struct vm_area_struct
*vma
= vmf
->vma
;
3798 struct mmu_notifier_range range
;
3802 * We need a reference to lock the folio because we don't hold
3803 * the PTL so a racing thread can remove the device-exclusive
3804 * entry and unmap it. If the folio is free the entry must
3805 * have been removed already. If it happens to have already
3806 * been re-allocated after being freed all we do is lock and
3809 if (!folio_try_get(folio
))
3812 ret
= folio_lock_or_retry(folio
, vmf
);
3817 mmu_notifier_range_init_owner(&range
, MMU_NOTIFY_EXCLUSIVE
, 0,
3818 vma
->vm_mm
, vmf
->address
& PAGE_MASK
,
3819 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
, NULL
);
3820 mmu_notifier_invalidate_range_start(&range
);
3822 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3824 if (likely(vmf
->pte
&& pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)))
3825 restore_exclusive_pte(vma
, vmf
->page
, vmf
->address
, vmf
->pte
);
3828 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3829 folio_unlock(folio
);
3832 mmu_notifier_invalidate_range_end(&range
);
3836 static inline bool should_try_to_free_swap(struct folio
*folio
,
3837 struct vm_area_struct
*vma
,
3838 unsigned int fault_flags
)
3840 if (!folio_test_swapcache(folio
))
3842 if (mem_cgroup_swap_full(folio
) || (vma
->vm_flags
& VM_LOCKED
) ||
3843 folio_test_mlocked(folio
))
3846 * If we want to map a page that's in the swapcache writable, we
3847 * have to detect via the refcount if we're really the exclusive
3848 * user. Try freeing the swapcache to get rid of the swapcache
3849 * reference only in case it's likely that we'll be the exlusive user.
3851 return (fault_flags
& FAULT_FLAG_WRITE
) && !folio_test_ksm(folio
) &&
3852 folio_ref_count(folio
) == 2;
3855 static vm_fault_t
pte_marker_clear(struct vm_fault
*vmf
)
3857 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
,
3858 vmf
->address
, &vmf
->ptl
);
3862 * Be careful so that we will only recover a special uffd-wp pte into a
3863 * none pte. Otherwise it means the pte could have changed, so retry.
3865 * This should also cover the case where e.g. the pte changed
3866 * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_POISONED.
3867 * So is_pte_marker() check is not enough to safely drop the pte.
3869 if (pte_same(vmf
->orig_pte
, ptep_get(vmf
->pte
)))
3870 pte_clear(vmf
->vma
->vm_mm
, vmf
->address
, vmf
->pte
);
3871 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3875 static vm_fault_t
do_pte_missing(struct vm_fault
*vmf
)
3877 if (vma_is_anonymous(vmf
->vma
))
3878 return do_anonymous_page(vmf
);
3880 return do_fault(vmf
);
3884 * This is actually a page-missing access, but with uffd-wp special pte
3885 * installed. It means this pte was wr-protected before being unmapped.
3887 static vm_fault_t
pte_marker_handle_uffd_wp(struct vm_fault
*vmf
)
3890 * Just in case there're leftover special ptes even after the region
3891 * got unregistered - we can simply clear them.
3893 if (unlikely(!userfaultfd_wp(vmf
->vma
)))
3894 return pte_marker_clear(vmf
);
3896 return do_pte_missing(vmf
);
3899 static vm_fault_t
handle_pte_marker(struct vm_fault
*vmf
)
3901 swp_entry_t entry
= pte_to_swp_entry(vmf
->orig_pte
);
3902 unsigned long marker
= pte_marker_get(entry
);
3905 * PTE markers should never be empty. If anything weird happened,
3906 * the best thing to do is to kill the process along with its mm.
3908 if (WARN_ON_ONCE(!marker
))
3909 return VM_FAULT_SIGBUS
;
3911 /* Higher priority than uffd-wp when data corrupted */
3912 if (marker
& PTE_MARKER_POISONED
)
3913 return VM_FAULT_HWPOISON
;
3915 if (pte_marker_entry_uffd_wp(entry
))
3916 return pte_marker_handle_uffd_wp(vmf
);
3918 /* This is an unknown pte marker */
3919 return VM_FAULT_SIGBUS
;
3923 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3924 * but allow concurrent faults), and pte mapped but not yet locked.
3925 * We return with pte unmapped and unlocked.
3927 * We return with the mmap_lock locked or unlocked in the same cases
3928 * as does filemap_fault().
3930 vm_fault_t
do_swap_page(struct vm_fault
*vmf
)
3932 struct vm_area_struct
*vma
= vmf
->vma
;
3933 struct folio
*swapcache
, *folio
= NULL
;
3935 struct swap_info_struct
*si
= NULL
;
3936 rmap_t rmap_flags
= RMAP_NONE
;
3937 bool need_clear_cache
= false;
3938 bool exclusive
= false;
3942 void *shadow
= NULL
;
3944 if (!pte_unmap_same(vmf
))
3947 entry
= pte_to_swp_entry(vmf
->orig_pte
);
3948 if (unlikely(non_swap_entry(entry
))) {
3949 if (is_migration_entry(entry
)) {
3950 migration_entry_wait(vma
->vm_mm
, vmf
->pmd
,
3952 } else if (is_device_exclusive_entry(entry
)) {
3953 vmf
->page
= pfn_swap_entry_to_page(entry
);
3954 ret
= remove_device_exclusive_entry(vmf
);
3955 } else if (is_device_private_entry(entry
)) {
3956 if (vmf
->flags
& FAULT_FLAG_VMA_LOCK
) {
3958 * migrate_to_ram is not yet ready to operate
3962 ret
= VM_FAULT_RETRY
;
3966 vmf
->page
= pfn_swap_entry_to_page(entry
);
3967 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3968 vmf
->address
, &vmf
->ptl
);
3969 if (unlikely(!vmf
->pte
||
3970 !pte_same(ptep_get(vmf
->pte
),
3975 * Get a page reference while we know the page can't be
3978 get_page(vmf
->page
);
3979 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3980 ret
= vmf
->page
->pgmap
->ops
->migrate_to_ram(vmf
);
3981 put_page(vmf
->page
);
3982 } else if (is_hwpoison_entry(entry
)) {
3983 ret
= VM_FAULT_HWPOISON
;
3984 } else if (is_pte_marker_entry(entry
)) {
3985 ret
= handle_pte_marker(vmf
);
3987 print_bad_pte(vma
, vmf
->address
, vmf
->orig_pte
, NULL
);
3988 ret
= VM_FAULT_SIGBUS
;
3993 /* Prevent swapoff from happening to us. */
3994 si
= get_swap_device(entry
);
3998 folio
= swap_cache_get_folio(entry
, vma
, vmf
->address
);
4000 page
= folio_file_page(folio
, swp_offset(entry
));
4004 if (data_race(si
->flags
& SWP_SYNCHRONOUS_IO
) &&
4005 __swap_count(entry
) == 1) {
4007 * Prevent parallel swapin from proceeding with
4008 * the cache flag. Otherwise, another thread may
4009 * finish swapin first, free the entry, and swapout
4010 * reusing the same entry. It's undetectable as
4011 * pte_same() returns true due to entry reuse.
4013 if (swapcache_prepare(entry
)) {
4014 /* Relax a bit to prevent rapid repeated page faults */
4015 schedule_timeout_uninterruptible(1);
4018 need_clear_cache
= true;
4020 /* skip swapcache */
4021 folio
= vma_alloc_folio(GFP_HIGHUSER_MOVABLE
, 0,
4022 vma
, vmf
->address
, false);
4023 page
= &folio
->page
;
4025 __folio_set_locked(folio
);
4026 __folio_set_swapbacked(folio
);
4028 if (mem_cgroup_swapin_charge_folio(folio
,
4029 vma
->vm_mm
, GFP_KERNEL
,
4034 mem_cgroup_swapin_uncharge_swap(entry
);
4036 shadow
= get_shadow_from_swap_cache(entry
);
4038 workingset_refault(folio
, shadow
);
4040 folio_add_lru(folio
);
4042 /* To provide entry to swap_read_folio() */
4043 folio
->swap
= entry
;
4044 swap_read_folio(folio
, true, NULL
);
4045 folio
->private = NULL
;
4048 page
= swapin_readahead(entry
, GFP_HIGHUSER_MOVABLE
,
4051 folio
= page_folio(page
);
4057 * Back out if somebody else faulted in this pte
4058 * while we released the pte lock.
4060 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4061 vmf
->address
, &vmf
->ptl
);
4062 if (likely(vmf
->pte
&&
4063 pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)))
4068 /* Had to read the page from swap area: Major fault */
4069 ret
= VM_FAULT_MAJOR
;
4070 count_vm_event(PGMAJFAULT
);
4071 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
4072 } else if (PageHWPoison(page
)) {
4074 * hwpoisoned dirty swapcache pages are kept for killing
4075 * owner processes (which may be unknown at hwpoison time)
4077 ret
= VM_FAULT_HWPOISON
;
4081 ret
|= folio_lock_or_retry(folio
, vmf
);
4082 if (ret
& VM_FAULT_RETRY
)
4087 * Make sure folio_free_swap() or swapoff did not release the
4088 * swapcache from under us. The page pin, and pte_same test
4089 * below, are not enough to exclude that. Even if it is still
4090 * swapcache, we need to check that the page's swap has not
4093 if (unlikely(!folio_test_swapcache(folio
) ||
4094 page_swap_entry(page
).val
!= entry
.val
))
4098 * KSM sometimes has to copy on read faults, for example, if
4099 * page->index of !PageKSM() pages would be nonlinear inside the
4100 * anon VMA -- PageKSM() is lost on actual swapout.
4102 folio
= ksm_might_need_to_copy(folio
, vma
, vmf
->address
);
4103 if (unlikely(!folio
)) {
4107 } else if (unlikely(folio
== ERR_PTR(-EHWPOISON
))) {
4108 ret
= VM_FAULT_HWPOISON
;
4112 if (folio
!= swapcache
)
4113 page
= folio_page(folio
, 0);
4116 * If we want to map a page that's in the swapcache writable, we
4117 * have to detect via the refcount if we're really the exclusive
4118 * owner. Try removing the extra reference from the local LRU
4119 * caches if required.
4121 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && folio
== swapcache
&&
4122 !folio_test_ksm(folio
) && !folio_test_lru(folio
))
4126 folio_throttle_swaprate(folio
, GFP_KERNEL
);
4129 * Back out if somebody else already faulted in this pte.
4131 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
4133 if (unlikely(!vmf
->pte
|| !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)))
4136 if (unlikely(!folio_test_uptodate(folio
))) {
4137 ret
= VM_FAULT_SIGBUS
;
4142 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte
4143 * must never point at an anonymous page in the swapcache that is
4144 * PG_anon_exclusive. Sanity check that this holds and especially, that
4145 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity
4146 * check after taking the PT lock and making sure that nobody
4147 * concurrently faulted in this page and set PG_anon_exclusive.
4149 BUG_ON(!folio_test_anon(folio
) && folio_test_mappedtodisk(folio
));
4150 BUG_ON(folio_test_anon(folio
) && PageAnonExclusive(page
));
4153 * Check under PT lock (to protect against concurrent fork() sharing
4154 * the swap entry concurrently) for certainly exclusive pages.
4156 if (!folio_test_ksm(folio
)) {
4157 exclusive
= pte_swp_exclusive(vmf
->orig_pte
);
4158 if (folio
!= swapcache
) {
4160 * We have a fresh page that is not exposed to the
4161 * swapcache -> certainly exclusive.
4164 } else if (exclusive
&& folio_test_writeback(folio
) &&
4165 data_race(si
->flags
& SWP_STABLE_WRITES
)) {
4167 * This is tricky: not all swap backends support
4168 * concurrent page modifications while under writeback.
4170 * So if we stumble over such a page in the swapcache
4171 * we must not set the page exclusive, otherwise we can
4172 * map it writable without further checks and modify it
4173 * while still under writeback.
4175 * For these problematic swap backends, simply drop the
4176 * exclusive marker: this is perfectly fine as we start
4177 * writeback only if we fully unmapped the page and
4178 * there are no unexpected references on the page after
4179 * unmapping succeeded. After fully unmapped, no
4180 * further GUP references (FOLL_GET and FOLL_PIN) can
4181 * appear, so dropping the exclusive marker and mapping
4182 * it only R/O is fine.
4189 * Some architectures may have to restore extra metadata to the page
4190 * when reading from swap. This metadata may be indexed by swap entry
4191 * so this must be called before swap_free().
4193 arch_swap_restore(entry
, folio
);
4196 * Remove the swap entry and conditionally try to free up the swapcache.
4197 * We're already holding a reference on the page but haven't mapped it
4201 if (should_try_to_free_swap(folio
, vma
, vmf
->flags
))
4202 folio_free_swap(folio
);
4204 inc_mm_counter(vma
->vm_mm
, MM_ANONPAGES
);
4205 dec_mm_counter(vma
->vm_mm
, MM_SWAPENTS
);
4206 pte
= mk_pte(page
, vma
->vm_page_prot
);
4209 * Same logic as in do_wp_page(); however, optimize for pages that are
4210 * certainly not shared either because we just allocated them without
4211 * exposing them to the swapcache or because the swap entry indicates
4214 if (!folio_test_ksm(folio
) &&
4215 (exclusive
|| folio_ref_count(folio
) == 1)) {
4216 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4217 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
4218 vmf
->flags
&= ~FAULT_FLAG_WRITE
;
4220 rmap_flags
|= RMAP_EXCLUSIVE
;
4222 flush_icache_page(vma
, page
);
4223 if (pte_swp_soft_dirty(vmf
->orig_pte
))
4224 pte
= pte_mksoft_dirty(pte
);
4225 if (pte_swp_uffd_wp(vmf
->orig_pte
))
4226 pte
= pte_mkuffd_wp(pte
);
4227 vmf
->orig_pte
= pte
;
4229 /* ksm created a completely new copy */
4230 if (unlikely(folio
!= swapcache
&& swapcache
)) {
4231 folio_add_new_anon_rmap(folio
, vma
, vmf
->address
);
4232 folio_add_lru_vma(folio
, vma
);
4234 folio_add_anon_rmap_pte(folio
, page
, vma
, vmf
->address
,
4238 VM_BUG_ON(!folio_test_anon(folio
) ||
4239 (pte_write(pte
) && !PageAnonExclusive(page
)));
4240 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
4241 arch_do_swap_page(vma
->vm_mm
, vma
, vmf
->address
, pte
, vmf
->orig_pte
);
4243 folio_unlock(folio
);
4244 if (folio
!= swapcache
&& swapcache
) {
4246 * Hold the lock to avoid the swap entry to be reused
4247 * until we take the PT lock for the pte_same() check
4248 * (to avoid false positives from pte_same). For
4249 * further safety release the lock after the swap_free
4250 * so that the swap count won't change under a
4251 * parallel locked swapcache.
4253 folio_unlock(swapcache
);
4254 folio_put(swapcache
);
4257 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4258 ret
|= do_wp_page(vmf
);
4259 if (ret
& VM_FAULT_ERROR
)
4260 ret
&= VM_FAULT_ERROR
;
4264 /* No need to invalidate - it was non-present before */
4265 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
4268 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4270 /* Clear the swap cache pin for direct swapin after PTL unlock */
4271 if (need_clear_cache
)
4272 swapcache_clear(si
, entry
);
4274 put_swap_device(si
);
4278 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4280 folio_unlock(folio
);
4283 if (folio
!= swapcache
&& swapcache
) {
4284 folio_unlock(swapcache
);
4285 folio_put(swapcache
);
4287 if (need_clear_cache
)
4288 swapcache_clear(si
, entry
);
4290 put_swap_device(si
);
4294 static bool pte_range_none(pte_t
*pte
, int nr_pages
)
4298 for (i
= 0; i
< nr_pages
; i
++) {
4299 if (!pte_none(ptep_get_lockless(pte
+ i
)))
4306 static struct folio
*alloc_anon_folio(struct vm_fault
*vmf
)
4308 struct vm_area_struct
*vma
= vmf
->vma
;
4309 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4310 unsigned long orders
;
4311 struct folio
*folio
;
4318 * If uffd is active for the vma we need per-page fault fidelity to
4319 * maintain the uffd semantics.
4321 if (unlikely(userfaultfd_armed(vma
)))
4325 * Get a list of all the (large) orders below PMD_ORDER that are enabled
4326 * for this vma. Then filter out the orders that can't be allocated over
4327 * the faulting address and still be fully contained in the vma.
4329 orders
= thp_vma_allowable_orders(vma
, vma
->vm_flags
, false, true, true,
4330 BIT(PMD_ORDER
) - 1);
4331 orders
= thp_vma_suitable_orders(vma
, vmf
->address
, orders
);
4336 pte
= pte_offset_map(vmf
->pmd
, vmf
->address
& PMD_MASK
);
4338 return ERR_PTR(-EAGAIN
);
4341 * Find the highest order where the aligned range is completely
4342 * pte_none(). Note that all remaining orders will be completely
4345 order
= highest_order(orders
);
4347 addr
= ALIGN_DOWN(vmf
->address
, PAGE_SIZE
<< order
);
4348 if (pte_range_none(pte
+ pte_index(addr
), 1 << order
))
4350 order
= next_order(&orders
, order
);
4355 /* Try allocating the highest of the remaining orders. */
4356 gfp
= vma_thp_gfp_mask(vma
);
4358 addr
= ALIGN_DOWN(vmf
->address
, PAGE_SIZE
<< order
);
4359 folio
= vma_alloc_folio(gfp
, order
, vma
, addr
, true);
4361 if (mem_cgroup_charge(folio
, vma
->vm_mm
, gfp
)) {
4365 folio_throttle_swaprate(folio
, gfp
);
4366 clear_huge_page(&folio
->page
, vmf
->address
, 1 << order
);
4370 order
= next_order(&orders
, order
);
4375 return folio_prealloc(vma
->vm_mm
, vma
, vmf
->address
, true);
4379 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4380 * but allow concurrent faults), and pte mapped but not yet locked.
4381 * We return with mmap_lock still held, but pte unmapped and unlocked.
4383 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
)
4385 bool uffd_wp
= vmf_orig_pte_uffd_wp(vmf
);
4386 struct vm_area_struct
*vma
= vmf
->vma
;
4387 unsigned long addr
= vmf
->address
;
4388 struct folio
*folio
;
4394 /* File mapping without ->vm_ops ? */
4395 if (vma
->vm_flags
& VM_SHARED
)
4396 return VM_FAULT_SIGBUS
;
4399 * Use pte_alloc() instead of pte_alloc_map(), so that OOM can
4400 * be distinguished from a transient failure of pte_offset_map().
4402 if (pte_alloc(vma
->vm_mm
, vmf
->pmd
))
4403 return VM_FAULT_OOM
;
4405 /* Use the zero-page for reads */
4406 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
4407 !mm_forbids_zeropage(vma
->vm_mm
)) {
4408 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(vmf
->address
),
4409 vma
->vm_page_prot
));
4410 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4411 vmf
->address
, &vmf
->ptl
);
4414 if (vmf_pte_changed(vmf
)) {
4415 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
4418 ret
= check_stable_address_space(vma
->vm_mm
);
4421 /* Deliver the page fault to userland, check inside PT lock */
4422 if (userfaultfd_missing(vma
)) {
4423 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4424 return handle_userfault(vmf
, VM_UFFD_MISSING
);
4429 /* Allocate our own private page. */
4430 if (unlikely(anon_vma_prepare(vma
)))
4432 /* Returns NULL on OOM or ERR_PTR(-EAGAIN) if we must retry the fault */
4433 folio
= alloc_anon_folio(vmf
);
4439 nr_pages
= folio_nr_pages(folio
);
4440 addr
= ALIGN_DOWN(vmf
->address
, nr_pages
* PAGE_SIZE
);
4443 * The memory barrier inside __folio_mark_uptodate makes sure that
4444 * preceding stores to the page contents become visible before
4445 * the set_pte_at() write.
4447 __folio_mark_uptodate(folio
);
4449 entry
= mk_pte(&folio
->page
, vma
->vm_page_prot
);
4450 entry
= pte_sw_mkyoung(entry
);
4451 if (vma
->vm_flags
& VM_WRITE
)
4452 entry
= pte_mkwrite(pte_mkdirty(entry
), vma
);
4454 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
4457 if (nr_pages
== 1 && vmf_pte_changed(vmf
)) {
4458 update_mmu_tlb(vma
, addr
, vmf
->pte
);
4460 } else if (nr_pages
> 1 && !pte_range_none(vmf
->pte
, nr_pages
)) {
4461 for (i
= 0; i
< nr_pages
; i
++)
4462 update_mmu_tlb(vma
, addr
+ PAGE_SIZE
* i
, vmf
->pte
+ i
);
4466 ret
= check_stable_address_space(vma
->vm_mm
);
4470 /* Deliver the page fault to userland, check inside PT lock */
4471 if (userfaultfd_missing(vma
)) {
4472 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4474 return handle_userfault(vmf
, VM_UFFD_MISSING
);
4477 folio_ref_add(folio
, nr_pages
- 1);
4478 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, nr_pages
);
4479 folio_add_new_anon_rmap(folio
, vma
, addr
);
4480 folio_add_lru_vma(folio
, vma
);
4483 entry
= pte_mkuffd_wp(entry
);
4484 set_ptes(vma
->vm_mm
, addr
, vmf
->pte
, entry
, nr_pages
);
4486 /* No need to invalidate - it was non-present before */
4487 update_mmu_cache_range(vmf
, vma
, addr
, vmf
->pte
, nr_pages
);
4490 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4496 return VM_FAULT_OOM
;
4500 * The mmap_lock must have been held on entry, and may have been
4501 * released depending on flags and vma->vm_ops->fault() return value.
4502 * See filemap_fault() and __lock_page_retry().
4504 static vm_fault_t
__do_fault(struct vm_fault
*vmf
)
4506 struct vm_area_struct
*vma
= vmf
->vma
;
4507 struct folio
*folio
;
4511 * Preallocate pte before we take page_lock because this might lead to
4512 * deadlocks for memcg reclaim which waits for pages under writeback:
4514 * SetPageWriteback(A)
4520 * wait_on_page_writeback(A)
4521 * SetPageWriteback(B)
4523 * # flush A, B to clear the writeback
4525 if (pmd_none(*vmf
->pmd
) && !vmf
->prealloc_pte
) {
4526 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
4527 if (!vmf
->prealloc_pte
)
4528 return VM_FAULT_OOM
;
4531 ret
= vma
->vm_ops
->fault(vmf
);
4532 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
|
4533 VM_FAULT_DONE_COW
)))
4536 folio
= page_folio(vmf
->page
);
4537 if (unlikely(PageHWPoison(vmf
->page
))) {
4538 vm_fault_t poisonret
= VM_FAULT_HWPOISON
;
4539 if (ret
& VM_FAULT_LOCKED
) {
4540 if (page_mapped(vmf
->page
))
4541 unmap_mapping_folio(folio
);
4542 /* Retry if a clean folio was removed from the cache. */
4543 if (mapping_evict_folio(folio
->mapping
, folio
))
4544 poisonret
= VM_FAULT_NOPAGE
;
4545 folio_unlock(folio
);
4552 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
4555 VM_BUG_ON_PAGE(!folio_test_locked(folio
), vmf
->page
);
4560 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4561 static void deposit_prealloc_pte(struct vm_fault
*vmf
)
4563 struct vm_area_struct
*vma
= vmf
->vma
;
4565 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
4567 * We are going to consume the prealloc table,
4568 * count that as nr_ptes.
4570 mm_inc_nr_ptes(vma
->vm_mm
);
4571 vmf
->prealloc_pte
= NULL
;
4574 vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
4576 struct folio
*folio
= page_folio(page
);
4577 struct vm_area_struct
*vma
= vmf
->vma
;
4578 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
4579 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
4581 vm_fault_t ret
= VM_FAULT_FALLBACK
;
4583 if (!thp_vma_suitable_order(vma
, haddr
, PMD_ORDER
))
4586 if (page
!= &folio
->page
|| folio_order(folio
) != HPAGE_PMD_ORDER
)
4590 * Just backoff if any subpage of a THP is corrupted otherwise
4591 * the corrupted page may mapped by PMD silently to escape the
4592 * check. This kind of THP just can be PTE mapped. Access to
4593 * the corrupted subpage should trigger SIGBUS as expected.
4595 if (unlikely(folio_test_has_hwpoisoned(folio
)))
4599 * Archs like ppc64 need additional space to store information
4600 * related to pte entry. Use the preallocated table for that.
4602 if (arch_needs_pgtable_deposit() && !vmf
->prealloc_pte
) {
4603 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
4604 if (!vmf
->prealloc_pte
)
4605 return VM_FAULT_OOM
;
4608 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
4609 if (unlikely(!pmd_none(*vmf
->pmd
)))
4612 flush_icache_pages(vma
, page
, HPAGE_PMD_NR
);
4614 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
4616 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
4618 add_mm_counter(vma
->vm_mm
, mm_counter_file(folio
), HPAGE_PMD_NR
);
4619 folio_add_file_rmap_pmd(folio
, page
, vma
);
4622 * deposit and withdraw with pmd lock held
4624 if (arch_needs_pgtable_deposit())
4625 deposit_prealloc_pte(vmf
);
4627 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
4629 update_mmu_cache_pmd(vma
, haddr
, vmf
->pmd
);
4631 /* fault is handled */
4633 count_vm_event(THP_FILE_MAPPED
);
4635 spin_unlock(vmf
->ptl
);
4639 vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
4641 return VM_FAULT_FALLBACK
;
4646 * set_pte_range - Set a range of PTEs to point to pages in a folio.
4647 * @vmf: Fault decription.
4648 * @folio: The folio that contains @page.
4649 * @page: The first page to create a PTE for.
4650 * @nr: The number of PTEs to create.
4651 * @addr: The first address to create a PTE for.
4653 void set_pte_range(struct vm_fault
*vmf
, struct folio
*folio
,
4654 struct page
*page
, unsigned int nr
, unsigned long addr
)
4656 struct vm_area_struct
*vma
= vmf
->vma
;
4657 bool uffd_wp
= vmf_orig_pte_uffd_wp(vmf
);
4658 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
4659 bool prefault
= in_range(vmf
->address
, addr
, nr
* PAGE_SIZE
);
4662 flush_icache_pages(vma
, page
, nr
);
4663 entry
= mk_pte(page
, vma
->vm_page_prot
);
4665 if (prefault
&& arch_wants_old_prefaulted_pte())
4666 entry
= pte_mkold(entry
);
4668 entry
= pte_sw_mkyoung(entry
);
4671 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
4672 if (unlikely(uffd_wp
))
4673 entry
= pte_mkuffd_wp(entry
);
4674 /* copy-on-write page */
4675 if (write
&& !(vma
->vm_flags
& VM_SHARED
)) {
4676 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, nr
);
4677 VM_BUG_ON_FOLIO(nr
!= 1, folio
);
4678 folio_add_new_anon_rmap(folio
, vma
, addr
);
4679 folio_add_lru_vma(folio
, vma
);
4681 add_mm_counter(vma
->vm_mm
, mm_counter_file(folio
), nr
);
4682 folio_add_file_rmap_ptes(folio
, page
, nr
, vma
);
4684 set_ptes(vma
->vm_mm
, addr
, vmf
->pte
, entry
, nr
);
4686 /* no need to invalidate: a not-present page won't be cached */
4687 update_mmu_cache_range(vmf
, vma
, addr
, vmf
->pte
, nr
);
4690 static bool vmf_pte_changed(struct vm_fault
*vmf
)
4692 if (vmf
->flags
& FAULT_FLAG_ORIG_PTE_VALID
)
4693 return !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
);
4695 return !pte_none(ptep_get(vmf
->pte
));
4699 * finish_fault - finish page fault once we have prepared the page to fault
4701 * @vmf: structure describing the fault
4703 * This function handles all that is needed to finish a page fault once the
4704 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4705 * given page, adds reverse page mapping, handles memcg charges and LRU
4708 * The function expects the page to be locked and on success it consumes a
4709 * reference of a page being mapped (for the PTE which maps it).
4711 * Return: %0 on success, %VM_FAULT_ code in case of error.
4713 vm_fault_t
finish_fault(struct vm_fault
*vmf
)
4715 struct vm_area_struct
*vma
= vmf
->vma
;
4719 /* Did we COW the page? */
4720 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
4721 page
= vmf
->cow_page
;
4726 * check even for read faults because we might have lost our CoWed
4729 if (!(vma
->vm_flags
& VM_SHARED
)) {
4730 ret
= check_stable_address_space(vma
->vm_mm
);
4735 if (pmd_none(*vmf
->pmd
)) {
4736 if (PageTransCompound(page
)) {
4737 ret
= do_set_pmd(vmf
, page
);
4738 if (ret
!= VM_FAULT_FALLBACK
)
4742 if (vmf
->prealloc_pte
)
4743 pmd_install(vma
->vm_mm
, vmf
->pmd
, &vmf
->prealloc_pte
);
4744 else if (unlikely(pte_alloc(vma
->vm_mm
, vmf
->pmd
)))
4745 return VM_FAULT_OOM
;
4748 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4749 vmf
->address
, &vmf
->ptl
);
4751 return VM_FAULT_NOPAGE
;
4753 /* Re-check under ptl */
4754 if (likely(!vmf_pte_changed(vmf
))) {
4755 struct folio
*folio
= page_folio(page
);
4757 set_pte_range(vmf
, folio
, page
, 1, vmf
->address
);
4760 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
4761 ret
= VM_FAULT_NOPAGE
;
4764 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4768 static unsigned long fault_around_pages __read_mostly
=
4769 65536 >> PAGE_SHIFT
;
4771 #ifdef CONFIG_DEBUG_FS
4772 static int fault_around_bytes_get(void *data
, u64
*val
)
4774 *val
= fault_around_pages
<< PAGE_SHIFT
;
4779 * fault_around_bytes must be rounded down to the nearest page order as it's
4780 * what do_fault_around() expects to see.
4782 static int fault_around_bytes_set(void *data
, u64 val
)
4784 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
4788 * The minimum value is 1 page, however this results in no fault-around
4789 * at all. See should_fault_around().
4791 val
= max(val
, PAGE_SIZE
);
4792 fault_around_pages
= rounddown_pow_of_two(val
) >> PAGE_SHIFT
;
4796 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops
,
4797 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
4799 static int __init
fault_around_debugfs(void)
4801 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL
, NULL
,
4802 &fault_around_bytes_fops
);
4805 late_initcall(fault_around_debugfs
);
4809 * do_fault_around() tries to map few pages around the fault address. The hope
4810 * is that the pages will be needed soon and this will lower the number of
4813 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4814 * not ready to be mapped: not up-to-date, locked, etc.
4816 * This function doesn't cross VMA or page table boundaries, in order to call
4817 * map_pages() and acquire a PTE lock only once.
4819 * fault_around_pages defines how many pages we'll try to map.
4820 * do_fault_around() expects it to be set to a power of two less than or equal
4823 * The virtual address of the area that we map is naturally aligned to
4824 * fault_around_pages * PAGE_SIZE rounded down to the machine page size
4825 * (and therefore to page order). This way it's easier to guarantee
4826 * that we don't cross page table boundaries.
4828 static vm_fault_t
do_fault_around(struct vm_fault
*vmf
)
4830 pgoff_t nr_pages
= READ_ONCE(fault_around_pages
);
4831 pgoff_t pte_off
= pte_index(vmf
->address
);
4832 /* The page offset of vmf->address within the VMA. */
4833 pgoff_t vma_off
= vmf
->pgoff
- vmf
->vma
->vm_pgoff
;
4834 pgoff_t from_pte
, to_pte
;
4837 /* The PTE offset of the start address, clamped to the VMA. */
4838 from_pte
= max(ALIGN_DOWN(pte_off
, nr_pages
),
4839 pte_off
- min(pte_off
, vma_off
));
4841 /* The PTE offset of the end address, clamped to the VMA and PTE. */
4842 to_pte
= min3(from_pte
+ nr_pages
, (pgoff_t
)PTRS_PER_PTE
,
4843 pte_off
+ vma_pages(vmf
->vma
) - vma_off
) - 1;
4845 if (pmd_none(*vmf
->pmd
)) {
4846 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
);
4847 if (!vmf
->prealloc_pte
)
4848 return VM_FAULT_OOM
;
4852 ret
= vmf
->vma
->vm_ops
->map_pages(vmf
,
4853 vmf
->pgoff
+ from_pte
- pte_off
,
4854 vmf
->pgoff
+ to_pte
- pte_off
);
4860 /* Return true if we should do read fault-around, false otherwise */
4861 static inline bool should_fault_around(struct vm_fault
*vmf
)
4863 /* No ->map_pages? No way to fault around... */
4864 if (!vmf
->vma
->vm_ops
->map_pages
)
4867 if (uffd_disable_fault_around(vmf
->vma
))
4870 /* A single page implies no faulting 'around' at all. */
4871 return fault_around_pages
> 1;
4874 static vm_fault_t
do_read_fault(struct vm_fault
*vmf
)
4877 struct folio
*folio
;
4880 * Let's call ->map_pages() first and use ->fault() as fallback
4881 * if page by the offset is not ready to be mapped (cold cache or
4884 if (should_fault_around(vmf
)) {
4885 ret
= do_fault_around(vmf
);
4890 ret
= vmf_can_call_fault(vmf
);
4894 ret
= __do_fault(vmf
);
4895 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4898 ret
|= finish_fault(vmf
);
4899 folio
= page_folio(vmf
->page
);
4900 folio_unlock(folio
);
4901 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4906 static vm_fault_t
do_cow_fault(struct vm_fault
*vmf
)
4908 struct vm_area_struct
*vma
= vmf
->vma
;
4909 struct folio
*folio
;
4912 ret
= vmf_can_call_fault(vmf
);
4914 ret
= vmf_anon_prepare(vmf
);
4918 folio
= folio_prealloc(vma
->vm_mm
, vma
, vmf
->address
, false);
4920 return VM_FAULT_OOM
;
4922 vmf
->cow_page
= &folio
->page
;
4924 ret
= __do_fault(vmf
);
4925 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4927 if (ret
& VM_FAULT_DONE_COW
)
4930 copy_user_highpage(vmf
->cow_page
, vmf
->page
, vmf
->address
, vma
);
4931 __folio_mark_uptodate(folio
);
4933 ret
|= finish_fault(vmf
);
4934 unlock_page(vmf
->page
);
4935 put_page(vmf
->page
);
4936 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4944 static vm_fault_t
do_shared_fault(struct vm_fault
*vmf
)
4946 struct vm_area_struct
*vma
= vmf
->vma
;
4947 vm_fault_t ret
, tmp
;
4948 struct folio
*folio
;
4950 ret
= vmf_can_call_fault(vmf
);
4954 ret
= __do_fault(vmf
);
4955 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4958 folio
= page_folio(vmf
->page
);
4961 * Check if the backing address space wants to know that the page is
4962 * about to become writable
4964 if (vma
->vm_ops
->page_mkwrite
) {
4965 folio_unlock(folio
);
4966 tmp
= do_page_mkwrite(vmf
, folio
);
4967 if (unlikely(!tmp
||
4968 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
4974 ret
|= finish_fault(vmf
);
4975 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
4977 folio_unlock(folio
);
4982 ret
|= fault_dirty_shared_page(vmf
);
4987 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4988 * but allow concurrent faults).
4989 * The mmap_lock may have been released depending on flags and our
4990 * return value. See filemap_fault() and __folio_lock_or_retry().
4991 * If mmap_lock is released, vma may become invalid (for example
4992 * by other thread calling munmap()).
4994 static vm_fault_t
do_fault(struct vm_fault
*vmf
)
4996 struct vm_area_struct
*vma
= vmf
->vma
;
4997 struct mm_struct
*vm_mm
= vma
->vm_mm
;
5001 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
5003 if (!vma
->vm_ops
->fault
) {
5004 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
,
5005 vmf
->address
, &vmf
->ptl
);
5006 if (unlikely(!vmf
->pte
))
5007 ret
= VM_FAULT_SIGBUS
;
5010 * Make sure this is not a temporary clearing of pte
5011 * by holding ptl and checking again. A R/M/W update
5012 * of pte involves: take ptl, clearing the pte so that
5013 * we don't have concurrent modification by hardware
5014 * followed by an update.
5016 if (unlikely(pte_none(ptep_get(vmf
->pte
))))
5017 ret
= VM_FAULT_SIGBUS
;
5019 ret
= VM_FAULT_NOPAGE
;
5021 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5023 } else if (!(vmf
->flags
& FAULT_FLAG_WRITE
))
5024 ret
= do_read_fault(vmf
);
5025 else if (!(vma
->vm_flags
& VM_SHARED
))
5026 ret
= do_cow_fault(vmf
);
5028 ret
= do_shared_fault(vmf
);
5030 /* preallocated pagetable is unused: free it */
5031 if (vmf
->prealloc_pte
) {
5032 pte_free(vm_mm
, vmf
->prealloc_pte
);
5033 vmf
->prealloc_pte
= NULL
;
5038 int numa_migrate_prep(struct folio
*folio
, struct vm_area_struct
*vma
,
5039 unsigned long addr
, int page_nid
, int *flags
)
5043 /* Record the current PID acceesing VMA */
5044 vma_set_access_pid_bit(vma
);
5046 count_vm_numa_event(NUMA_HINT_FAULTS
);
5047 if (page_nid
== numa_node_id()) {
5048 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
5049 *flags
|= TNF_FAULT_LOCAL
;
5052 return mpol_misplaced(folio
, vma
, addr
);
5055 static vm_fault_t
do_numa_page(struct vm_fault
*vmf
)
5057 struct vm_area_struct
*vma
= vmf
->vma
;
5058 struct folio
*folio
= NULL
;
5059 int nid
= NUMA_NO_NODE
;
5060 bool writable
= false;
5067 * The pte cannot be used safely until we verify, while holding the page
5068 * table lock, that its contents have not changed during fault handling.
5070 spin_lock(vmf
->ptl
);
5071 /* Read the live PTE from the page tables: */
5072 old_pte
= ptep_get(vmf
->pte
);
5074 if (unlikely(!pte_same(old_pte
, vmf
->orig_pte
))) {
5075 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5079 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
5082 * Detect now whether the PTE could be writable; this information
5083 * is only valid while holding the PT lock.
5085 writable
= pte_write(pte
);
5086 if (!writable
&& vma_wants_manual_pte_write_upgrade(vma
) &&
5087 can_change_pte_writable(vma
, vmf
->address
, pte
))
5090 folio
= vm_normal_folio(vma
, vmf
->address
, pte
);
5091 if (!folio
|| folio_is_zone_device(folio
))
5094 /* TODO: handle PTE-mapped THP */
5095 if (folio_test_large(folio
))
5099 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
5100 * much anyway since they can be in shared cache state. This misses
5101 * the case where a mapping is writable but the process never writes
5102 * to it but pte_write gets cleared during protection updates and
5103 * pte_dirty has unpredictable behaviour between PTE scan updates,
5104 * background writeback, dirty balancing and application behaviour.
5107 flags
|= TNF_NO_GROUP
;
5110 * Flag if the folio is shared between multiple address spaces. This
5111 * is later used when determining whether to group tasks together
5113 if (folio_estimated_sharers(folio
) > 1 && (vma
->vm_flags
& VM_SHARED
))
5114 flags
|= TNF_SHARED
;
5116 nid
= folio_nid(folio
);
5118 * For memory tiering mode, cpupid of slow memory page is used
5119 * to record page access time. So use default value.
5121 if ((sysctl_numa_balancing_mode
& NUMA_BALANCING_MEMORY_TIERING
) &&
5122 !node_is_toptier(nid
))
5123 last_cpupid
= (-1 & LAST_CPUPID_MASK
);
5125 last_cpupid
= folio_last_cpupid(folio
);
5126 target_nid
= numa_migrate_prep(folio
, vma
, vmf
->address
, nid
, &flags
);
5127 if (target_nid
== NUMA_NO_NODE
) {
5131 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5134 /* Migrate to the requested node */
5135 if (migrate_misplaced_folio(folio
, vma
, target_nid
)) {
5137 flags
|= TNF_MIGRATED
;
5139 flags
|= TNF_MIGRATE_FAIL
;
5140 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
5141 vmf
->address
, &vmf
->ptl
);
5142 if (unlikely(!vmf
->pte
))
5144 if (unlikely(!pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
5145 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5152 if (nid
!= NUMA_NO_NODE
)
5153 task_numa_fault(last_cpupid
, nid
, 1, flags
);
5157 * Make it present again, depending on how arch implements
5158 * non-accessible ptes, some can allow access by kernel mode.
5160 old_pte
= ptep_modify_prot_start(vma
, vmf
->address
, vmf
->pte
);
5161 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
5162 pte
= pte_mkyoung(pte
);
5164 pte
= pte_mkwrite(pte
, vma
);
5165 ptep_modify_prot_commit(vma
, vmf
->address
, vmf
->pte
, old_pte
, pte
);
5166 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
5167 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5171 static inline vm_fault_t
create_huge_pmd(struct vm_fault
*vmf
)
5173 struct vm_area_struct
*vma
= vmf
->vma
;
5174 if (vma_is_anonymous(vma
))
5175 return do_huge_pmd_anonymous_page(vmf
);
5176 if (vma
->vm_ops
->huge_fault
)
5177 return vma
->vm_ops
->huge_fault(vmf
, PMD_ORDER
);
5178 return VM_FAULT_FALLBACK
;
5181 /* `inline' is required to avoid gcc 4.1.2 build error */
5182 static inline vm_fault_t
wp_huge_pmd(struct vm_fault
*vmf
)
5184 struct vm_area_struct
*vma
= vmf
->vma
;
5185 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
5188 if (vma_is_anonymous(vma
)) {
5189 if (likely(!unshare
) &&
5190 userfaultfd_huge_pmd_wp(vma
, vmf
->orig_pmd
)) {
5191 if (userfaultfd_wp_async(vmf
->vma
))
5193 return handle_userfault(vmf
, VM_UFFD_WP
);
5195 return do_huge_pmd_wp_page(vmf
);
5198 if (vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
)) {
5199 if (vma
->vm_ops
->huge_fault
) {
5200 ret
= vma
->vm_ops
->huge_fault(vmf
, PMD_ORDER
);
5201 if (!(ret
& VM_FAULT_FALLBACK
))
5207 /* COW or write-notify handled on pte level: split pmd. */
5208 __split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
5210 return VM_FAULT_FALLBACK
;
5213 static vm_fault_t
create_huge_pud(struct vm_fault
*vmf
)
5215 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
5216 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5217 struct vm_area_struct
*vma
= vmf
->vma
;
5218 /* No support for anonymous transparent PUD pages yet */
5219 if (vma_is_anonymous(vma
))
5220 return VM_FAULT_FALLBACK
;
5221 if (vma
->vm_ops
->huge_fault
)
5222 return vma
->vm_ops
->huge_fault(vmf
, PUD_ORDER
);
5223 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
5224 return VM_FAULT_FALLBACK
;
5227 static vm_fault_t
wp_huge_pud(struct vm_fault
*vmf
, pud_t orig_pud
)
5229 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
5230 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5231 struct vm_area_struct
*vma
= vmf
->vma
;
5234 /* No support for anonymous transparent PUD pages yet */
5235 if (vma_is_anonymous(vma
))
5237 if (vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
)) {
5238 if (vma
->vm_ops
->huge_fault
) {
5239 ret
= vma
->vm_ops
->huge_fault(vmf
, PUD_ORDER
);
5240 if (!(ret
& VM_FAULT_FALLBACK
))
5245 /* COW or write-notify not handled on PUD level: split pud.*/
5246 __split_huge_pud(vma
, vmf
->pud
, vmf
->address
);
5247 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
5248 return VM_FAULT_FALLBACK
;
5252 * These routines also need to handle stuff like marking pages dirty
5253 * and/or accessed for architectures that don't do it in hardware (most
5254 * RISC architectures). The early dirtying is also good on the i386.
5256 * There is also a hook called "update_mmu_cache()" that architectures
5257 * with external mmu caches can use to update those (ie the Sparc or
5258 * PowerPC hashed page tables that act as extended TLBs).
5260 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
5261 * concurrent faults).
5263 * The mmap_lock may have been released depending on flags and our return value.
5264 * See filemap_fault() and __folio_lock_or_retry().
5266 static vm_fault_t
handle_pte_fault(struct vm_fault
*vmf
)
5270 if (unlikely(pmd_none(*vmf
->pmd
))) {
5272 * Leave __pte_alloc() until later: because vm_ops->fault may
5273 * want to allocate huge page, and if we expose page table
5274 * for an instant, it will be difficult to retract from
5275 * concurrent faults and from rmap lookups.
5278 vmf
->flags
&= ~FAULT_FLAG_ORIG_PTE_VALID
;
5281 * A regular pmd is established and it can't morph into a huge
5282 * pmd by anon khugepaged, since that takes mmap_lock in write
5283 * mode; but shmem or file collapse to THP could still morph
5284 * it into a huge pmd: just retry later if so.
5286 vmf
->pte
= pte_offset_map_nolock(vmf
->vma
->vm_mm
, vmf
->pmd
,
5287 vmf
->address
, &vmf
->ptl
);
5288 if (unlikely(!vmf
->pte
))
5290 vmf
->orig_pte
= ptep_get_lockless(vmf
->pte
);
5291 vmf
->flags
|= FAULT_FLAG_ORIG_PTE_VALID
;
5293 if (pte_none(vmf
->orig_pte
)) {
5294 pte_unmap(vmf
->pte
);
5300 return do_pte_missing(vmf
);
5302 if (!pte_present(vmf
->orig_pte
))
5303 return do_swap_page(vmf
);
5305 if (pte_protnone(vmf
->orig_pte
) && vma_is_accessible(vmf
->vma
))
5306 return do_numa_page(vmf
);
5308 spin_lock(vmf
->ptl
);
5309 entry
= vmf
->orig_pte
;
5310 if (unlikely(!pte_same(ptep_get(vmf
->pte
), entry
))) {
5311 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
5314 if (vmf
->flags
& (FAULT_FLAG_WRITE
|FAULT_FLAG_UNSHARE
)) {
5315 if (!pte_write(entry
))
5316 return do_wp_page(vmf
);
5317 else if (likely(vmf
->flags
& FAULT_FLAG_WRITE
))
5318 entry
= pte_mkdirty(entry
);
5320 entry
= pte_mkyoung(entry
);
5321 if (ptep_set_access_flags(vmf
->vma
, vmf
->address
, vmf
->pte
, entry
,
5322 vmf
->flags
& FAULT_FLAG_WRITE
)) {
5323 update_mmu_cache_range(vmf
, vmf
->vma
, vmf
->address
,
5326 /* Skip spurious TLB flush for retried page fault */
5327 if (vmf
->flags
& FAULT_FLAG_TRIED
)
5330 * This is needed only for protection faults but the arch code
5331 * is not yet telling us if this is a protection fault or not.
5332 * This still avoids useless tlb flushes for .text page faults
5335 if (vmf
->flags
& FAULT_FLAG_WRITE
)
5336 flush_tlb_fix_spurious_fault(vmf
->vma
, vmf
->address
,
5340 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5345 * On entry, we hold either the VMA lock or the mmap_lock
5346 * (FAULT_FLAG_VMA_LOCK tells you which). If VM_FAULT_RETRY is set in
5347 * the result, the mmap_lock is not held on exit. See filemap_fault()
5348 * and __folio_lock_or_retry().
5350 static vm_fault_t
__handle_mm_fault(struct vm_area_struct
*vma
,
5351 unsigned long address
, unsigned int flags
)
5353 struct vm_fault vmf
= {
5355 .address
= address
& PAGE_MASK
,
5356 .real_address
= address
,
5358 .pgoff
= linear_page_index(vma
, address
),
5359 .gfp_mask
= __get_fault_gfp_mask(vma
),
5361 struct mm_struct
*mm
= vma
->vm_mm
;
5362 unsigned long vm_flags
= vma
->vm_flags
;
5367 pgd
= pgd_offset(mm
, address
);
5368 p4d
= p4d_alloc(mm
, pgd
, address
);
5370 return VM_FAULT_OOM
;
5372 vmf
.pud
= pud_alloc(mm
, p4d
, address
);
5374 return VM_FAULT_OOM
;
5376 if (pud_none(*vmf
.pud
) &&
5377 thp_vma_allowable_order(vma
, vm_flags
, false, true, true, PUD_ORDER
)) {
5378 ret
= create_huge_pud(&vmf
);
5379 if (!(ret
& VM_FAULT_FALLBACK
))
5382 pud_t orig_pud
= *vmf
.pud
;
5385 if (pud_trans_huge(orig_pud
) || pud_devmap(orig_pud
)) {
5388 * TODO once we support anonymous PUDs: NUMA case and
5389 * FAULT_FLAG_UNSHARE handling.
5391 if ((flags
& FAULT_FLAG_WRITE
) && !pud_write(orig_pud
)) {
5392 ret
= wp_huge_pud(&vmf
, orig_pud
);
5393 if (!(ret
& VM_FAULT_FALLBACK
))
5396 huge_pud_set_accessed(&vmf
, orig_pud
);
5402 vmf
.pmd
= pmd_alloc(mm
, vmf
.pud
, address
);
5404 return VM_FAULT_OOM
;
5406 /* Huge pud page fault raced with pmd_alloc? */
5407 if (pud_trans_unstable(vmf
.pud
))
5410 if (pmd_none(*vmf
.pmd
) &&
5411 thp_vma_allowable_order(vma
, vm_flags
, false, true, true, PMD_ORDER
)) {
5412 ret
= create_huge_pmd(&vmf
);
5413 if (!(ret
& VM_FAULT_FALLBACK
))
5416 vmf
.orig_pmd
= pmdp_get_lockless(vmf
.pmd
);
5418 if (unlikely(is_swap_pmd(vmf
.orig_pmd
))) {
5419 VM_BUG_ON(thp_migration_supported() &&
5420 !is_pmd_migration_entry(vmf
.orig_pmd
));
5421 if (is_pmd_migration_entry(vmf
.orig_pmd
))
5422 pmd_migration_entry_wait(mm
, vmf
.pmd
);
5425 if (pmd_trans_huge(vmf
.orig_pmd
) || pmd_devmap(vmf
.orig_pmd
)) {
5426 if (pmd_protnone(vmf
.orig_pmd
) && vma_is_accessible(vma
))
5427 return do_huge_pmd_numa_page(&vmf
);
5429 if ((flags
& (FAULT_FLAG_WRITE
|FAULT_FLAG_UNSHARE
)) &&
5430 !pmd_write(vmf
.orig_pmd
)) {
5431 ret
= wp_huge_pmd(&vmf
);
5432 if (!(ret
& VM_FAULT_FALLBACK
))
5435 huge_pmd_set_accessed(&vmf
);
5441 return handle_pte_fault(&vmf
);
5445 * mm_account_fault - Do page fault accounting
5446 * @mm: mm from which memcg should be extracted. It can be NULL.
5447 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
5448 * of perf event counters, but we'll still do the per-task accounting to
5449 * the task who triggered this page fault.
5450 * @address: the faulted address.
5451 * @flags: the fault flags.
5452 * @ret: the fault retcode.
5454 * This will take care of most of the page fault accounting. Meanwhile, it
5455 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
5456 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
5457 * still be in per-arch page fault handlers at the entry of page fault.
5459 static inline void mm_account_fault(struct mm_struct
*mm
, struct pt_regs
*regs
,
5460 unsigned long address
, unsigned int flags
,
5465 /* Incomplete faults will be accounted upon completion. */
5466 if (ret
& VM_FAULT_RETRY
)
5470 * To preserve the behavior of older kernels, PGFAULT counters record
5471 * both successful and failed faults, as opposed to perf counters,
5472 * which ignore failed cases.
5474 count_vm_event(PGFAULT
);
5475 count_memcg_event_mm(mm
, PGFAULT
);
5478 * Do not account for unsuccessful faults (e.g. when the address wasn't
5479 * valid). That includes arch_vma_access_permitted() failing before
5480 * reaching here. So this is not a "this many hardware page faults"
5481 * counter. We should use the hw profiling for that.
5483 if (ret
& VM_FAULT_ERROR
)
5487 * We define the fault as a major fault when the final successful fault
5488 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
5489 * handle it immediately previously).
5491 major
= (ret
& VM_FAULT_MAJOR
) || (flags
& FAULT_FLAG_TRIED
);
5499 * If the fault is done for GUP, regs will be NULL. We only do the
5500 * accounting for the per thread fault counters who triggered the
5501 * fault, and we skip the perf event updates.
5507 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ
, 1, regs
, address
);
5509 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN
, 1, regs
, address
);
5512 #ifdef CONFIG_LRU_GEN
5513 static void lru_gen_enter_fault(struct vm_area_struct
*vma
)
5515 /* the LRU algorithm only applies to accesses with recency */
5516 current
->in_lru_fault
= vma_has_recency(vma
);
5519 static void lru_gen_exit_fault(void)
5521 current
->in_lru_fault
= false;
5524 static void lru_gen_enter_fault(struct vm_area_struct
*vma
)
5528 static void lru_gen_exit_fault(void)
5531 #endif /* CONFIG_LRU_GEN */
5533 static vm_fault_t
sanitize_fault_flags(struct vm_area_struct
*vma
,
5534 unsigned int *flags
)
5536 if (unlikely(*flags
& FAULT_FLAG_UNSHARE
)) {
5537 if (WARN_ON_ONCE(*flags
& FAULT_FLAG_WRITE
))
5538 return VM_FAULT_SIGSEGV
;
5540 * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's
5541 * just treat it like an ordinary read-fault otherwise.
5543 if (!is_cow_mapping(vma
->vm_flags
))
5544 *flags
&= ~FAULT_FLAG_UNSHARE
;
5545 } else if (*flags
& FAULT_FLAG_WRITE
) {
5546 /* Write faults on read-only mappings are impossible ... */
5547 if (WARN_ON_ONCE(!(vma
->vm_flags
& VM_MAYWRITE
)))
5548 return VM_FAULT_SIGSEGV
;
5549 /* ... and FOLL_FORCE only applies to COW mappings. */
5550 if (WARN_ON_ONCE(!(vma
->vm_flags
& VM_WRITE
) &&
5551 !is_cow_mapping(vma
->vm_flags
)))
5552 return VM_FAULT_SIGSEGV
;
5554 #ifdef CONFIG_PER_VMA_LOCK
5556 * Per-VMA locks can't be used with FAULT_FLAG_RETRY_NOWAIT because of
5557 * the assumption that lock is dropped on VM_FAULT_RETRY.
5559 if (WARN_ON_ONCE((*flags
&
5560 (FAULT_FLAG_VMA_LOCK
| FAULT_FLAG_RETRY_NOWAIT
)) ==
5561 (FAULT_FLAG_VMA_LOCK
| FAULT_FLAG_RETRY_NOWAIT
)))
5562 return VM_FAULT_SIGSEGV
;
5569 * By the time we get here, we already hold the mm semaphore
5571 * The mmap_lock may have been released depending on flags and our
5572 * return value. See filemap_fault() and __folio_lock_or_retry().
5574 vm_fault_t
handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
5575 unsigned int flags
, struct pt_regs
*regs
)
5577 /* If the fault handler drops the mmap_lock, vma may be freed */
5578 struct mm_struct
*mm
= vma
->vm_mm
;
5581 __set_current_state(TASK_RUNNING
);
5583 ret
= sanitize_fault_flags(vma
, &flags
);
5587 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
5588 flags
& FAULT_FLAG_INSTRUCTION
,
5589 flags
& FAULT_FLAG_REMOTE
)) {
5590 ret
= VM_FAULT_SIGSEGV
;
5595 * Enable the memcg OOM handling for faults triggered in user
5596 * space. Kernel faults are handled more gracefully.
5598 if (flags
& FAULT_FLAG_USER
)
5599 mem_cgroup_enter_user_fault();
5601 lru_gen_enter_fault(vma
);
5603 if (unlikely(is_vm_hugetlb_page(vma
)))
5604 ret
= hugetlb_fault(vma
->vm_mm
, vma
, address
, flags
);
5606 ret
= __handle_mm_fault(vma
, address
, flags
);
5608 lru_gen_exit_fault();
5610 if (flags
& FAULT_FLAG_USER
) {
5611 mem_cgroup_exit_user_fault();
5613 * The task may have entered a memcg OOM situation but
5614 * if the allocation error was handled gracefully (no
5615 * VM_FAULT_OOM), there is no need to kill anything.
5616 * Just clean up the OOM state peacefully.
5618 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
5619 mem_cgroup_oom_synchronize(false);
5622 mm_account_fault(mm
, regs
, address
, flags
, ret
);
5626 EXPORT_SYMBOL_GPL(handle_mm_fault
);
5628 #ifdef CONFIG_LOCK_MM_AND_FIND_VMA
5629 #include <linux/extable.h>
5631 static inline bool get_mmap_lock_carefully(struct mm_struct
*mm
, struct pt_regs
*regs
)
5633 if (likely(mmap_read_trylock(mm
)))
5636 if (regs
&& !user_mode(regs
)) {
5637 unsigned long ip
= exception_ip(regs
);
5638 if (!search_exception_tables(ip
))
5642 return !mmap_read_lock_killable(mm
);
5645 static inline bool mmap_upgrade_trylock(struct mm_struct
*mm
)
5648 * We don't have this operation yet.
5650 * It should be easy enough to do: it's basically a
5651 * atomic_long_try_cmpxchg_acquire()
5652 * from RWSEM_READER_BIAS -> RWSEM_WRITER_LOCKED, but
5653 * it also needs the proper lockdep magic etc.
5658 static inline bool upgrade_mmap_lock_carefully(struct mm_struct
*mm
, struct pt_regs
*regs
)
5660 mmap_read_unlock(mm
);
5661 if (regs
&& !user_mode(regs
)) {
5662 unsigned long ip
= exception_ip(regs
);
5663 if (!search_exception_tables(ip
))
5666 return !mmap_write_lock_killable(mm
);
5670 * Helper for page fault handling.
5672 * This is kind of equivalend to "mmap_read_lock()" followed
5673 * by "find_extend_vma()", except it's a lot more careful about
5674 * the locking (and will drop the lock on failure).
5676 * For example, if we have a kernel bug that causes a page
5677 * fault, we don't want to just use mmap_read_lock() to get
5678 * the mm lock, because that would deadlock if the bug were
5679 * to happen while we're holding the mm lock for writing.
5681 * So this checks the exception tables on kernel faults in
5682 * order to only do this all for instructions that are actually
5683 * expected to fault.
5685 * We can also actually take the mm lock for writing if we
5686 * need to extend the vma, which helps the VM layer a lot.
5688 struct vm_area_struct
*lock_mm_and_find_vma(struct mm_struct
*mm
,
5689 unsigned long addr
, struct pt_regs
*regs
)
5691 struct vm_area_struct
*vma
;
5693 if (!get_mmap_lock_carefully(mm
, regs
))
5696 vma
= find_vma(mm
, addr
);
5697 if (likely(vma
&& (vma
->vm_start
<= addr
)))
5701 * Well, dang. We might still be successful, but only
5702 * if we can extend a vma to do so.
5704 if (!vma
|| !(vma
->vm_flags
& VM_GROWSDOWN
)) {
5705 mmap_read_unlock(mm
);
5710 * We can try to upgrade the mmap lock atomically,
5711 * in which case we can continue to use the vma
5712 * we already looked up.
5714 * Otherwise we'll have to drop the mmap lock and
5715 * re-take it, and also look up the vma again,
5718 if (!mmap_upgrade_trylock(mm
)) {
5719 if (!upgrade_mmap_lock_carefully(mm
, regs
))
5722 vma
= find_vma(mm
, addr
);
5725 if (vma
->vm_start
<= addr
)
5727 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
5731 if (expand_stack_locked(vma
, addr
))
5735 mmap_write_downgrade(mm
);
5739 mmap_write_unlock(mm
);
5744 #ifdef CONFIG_PER_VMA_LOCK
5746 * Lookup and lock a VMA under RCU protection. Returned VMA is guaranteed to be
5747 * stable and not isolated. If the VMA is not found or is being modified the
5748 * function returns NULL.
5750 struct vm_area_struct
*lock_vma_under_rcu(struct mm_struct
*mm
,
5751 unsigned long address
)
5753 MA_STATE(mas
, &mm
->mm_mt
, address
, address
);
5754 struct vm_area_struct
*vma
;
5758 vma
= mas_walk(&mas
);
5762 if (!vma_start_read(vma
))
5766 * find_mergeable_anon_vma uses adjacent vmas which are not locked.
5767 * This check must happen after vma_start_read(); otherwise, a
5768 * concurrent mremap() with MREMAP_DONTUNMAP could dissociate the VMA
5769 * from its anon_vma.
5771 if (unlikely(vma_is_anonymous(vma
) && !vma
->anon_vma
))
5772 goto inval_end_read
;
5774 /* Check since vm_start/vm_end might change before we lock the VMA */
5775 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
))
5776 goto inval_end_read
;
5778 /* Check if the VMA got isolated after we found it */
5779 if (vma
->detached
) {
5781 count_vm_vma_lock_event(VMA_LOCK_MISS
);
5782 /* The area was replaced with another one */
5793 count_vm_vma_lock_event(VMA_LOCK_ABORT
);
5796 #endif /* CONFIG_PER_VMA_LOCK */
5798 #ifndef __PAGETABLE_P4D_FOLDED
5800 * Allocate p4d page table.
5801 * We've already handled the fast-path in-line.
5803 int __p4d_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
5805 p4d_t
*new = p4d_alloc_one(mm
, address
);
5809 spin_lock(&mm
->page_table_lock
);
5810 if (pgd_present(*pgd
)) { /* Another has populated it */
5813 smp_wmb(); /* See comment in pmd_install() */
5814 pgd_populate(mm
, pgd
, new);
5816 spin_unlock(&mm
->page_table_lock
);
5819 #endif /* __PAGETABLE_P4D_FOLDED */
5821 #ifndef __PAGETABLE_PUD_FOLDED
5823 * Allocate page upper directory.
5824 * We've already handled the fast-path in-line.
5826 int __pud_alloc(struct mm_struct
*mm
, p4d_t
*p4d
, unsigned long address
)
5828 pud_t
*new = pud_alloc_one(mm
, address
);
5832 spin_lock(&mm
->page_table_lock
);
5833 if (!p4d_present(*p4d
)) {
5835 smp_wmb(); /* See comment in pmd_install() */
5836 p4d_populate(mm
, p4d
, new);
5837 } else /* Another has populated it */
5839 spin_unlock(&mm
->page_table_lock
);
5842 #endif /* __PAGETABLE_PUD_FOLDED */
5844 #ifndef __PAGETABLE_PMD_FOLDED
5846 * Allocate page middle directory.
5847 * We've already handled the fast-path in-line.
5849 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
5852 pmd_t
*new = pmd_alloc_one(mm
, address
);
5856 ptl
= pud_lock(mm
, pud
);
5857 if (!pud_present(*pud
)) {
5859 smp_wmb(); /* See comment in pmd_install() */
5860 pud_populate(mm
, pud
, new);
5861 } else { /* Another has populated it */
5867 #endif /* __PAGETABLE_PMD_FOLDED */
5870 * follow_pte - look up PTE at a user virtual address
5871 * @mm: the mm_struct of the target address space
5872 * @address: user virtual address
5873 * @ptepp: location to store found PTE
5874 * @ptlp: location to store the lock for the PTE
5876 * On a successful return, the pointer to the PTE is stored in @ptepp;
5877 * the corresponding lock is taken and its location is stored in @ptlp.
5878 * The contents of the PTE are only stable until @ptlp is released;
5879 * any further use, if any, must be protected against invalidation
5880 * with MMU notifiers.
5882 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
5883 * should be taken for read.
5885 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
5886 * it is not a good general-purpose API.
5888 * Return: zero on success, -ve otherwise.
5890 int follow_pte(struct mm_struct
*mm
, unsigned long address
,
5891 pte_t
**ptepp
, spinlock_t
**ptlp
)
5899 pgd
= pgd_offset(mm
, address
);
5900 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
5903 p4d
= p4d_offset(pgd
, address
);
5904 if (p4d_none(*p4d
) || unlikely(p4d_bad(*p4d
)))
5907 pud
= pud_offset(p4d
, address
);
5908 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
5911 pmd
= pmd_offset(pud
, address
);
5912 VM_BUG_ON(pmd_trans_huge(*pmd
));
5914 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
5917 if (!pte_present(ptep_get(ptep
)))
5922 pte_unmap_unlock(ptep
, *ptlp
);
5926 EXPORT_SYMBOL_GPL(follow_pte
);
5929 * follow_pfn - look up PFN at a user virtual address
5930 * @vma: memory mapping
5931 * @address: user virtual address
5932 * @pfn: location to store found PFN
5934 * Only IO mappings and raw PFN mappings are allowed.
5936 * This function does not allow the caller to read the permissions
5937 * of the PTE. Do not use it.
5939 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5941 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
5948 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
5951 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
5954 *pfn
= pte_pfn(ptep_get(ptep
));
5955 pte_unmap_unlock(ptep
, ptl
);
5958 EXPORT_SYMBOL(follow_pfn
);
5960 #ifdef CONFIG_HAVE_IOREMAP_PROT
5961 int follow_phys(struct vm_area_struct
*vma
,
5962 unsigned long address
, unsigned int flags
,
5963 unsigned long *prot
, resource_size_t
*phys
)
5969 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
5972 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
5974 pte
= ptep_get(ptep
);
5976 /* Never return PFNs of anon folios in COW mappings. */
5977 if (vm_normal_folio(vma
, address
, pte
))
5980 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
5983 *prot
= pgprot_val(pte_pgprot(pte
));
5984 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
5988 pte_unmap_unlock(ptep
, ptl
);
5994 * generic_access_phys - generic implementation for iomem mmap access
5995 * @vma: the vma to access
5996 * @addr: userspace address, not relative offset within @vma
5997 * @buf: buffer to read/write
5998 * @len: length of transfer
5999 * @write: set to FOLL_WRITE when writing, otherwise reading
6001 * This is a generic implementation for &vm_operations_struct.access for an
6002 * iomem mapping. This callback is used by access_process_vm() when the @vma is
6005 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
6006 void *buf
, int len
, int write
)
6008 resource_size_t phys_addr
;
6009 unsigned long prot
= 0;
6010 void __iomem
*maddr
;
6013 int offset
= offset_in_page(addr
);
6016 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
6020 if (follow_pte(vma
->vm_mm
, addr
, &ptep
, &ptl
))
6022 pte
= ptep_get(ptep
);
6023 pte_unmap_unlock(ptep
, ptl
);
6025 prot
= pgprot_val(pte_pgprot(pte
));
6026 phys_addr
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
6028 if ((write
& FOLL_WRITE
) && !pte_write(pte
))
6031 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
6035 if (follow_pte(vma
->vm_mm
, addr
, &ptep
, &ptl
))
6038 if (!pte_same(pte
, ptep_get(ptep
))) {
6039 pte_unmap_unlock(ptep
, ptl
);
6046 memcpy_toio(maddr
+ offset
, buf
, len
);
6048 memcpy_fromio(buf
, maddr
+ offset
, len
);
6050 pte_unmap_unlock(ptep
, ptl
);
6056 EXPORT_SYMBOL_GPL(generic_access_phys
);
6060 * Access another process' address space as given in mm.
6062 static int __access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
6063 void *buf
, int len
, unsigned int gup_flags
)
6065 void *old_buf
= buf
;
6066 int write
= gup_flags
& FOLL_WRITE
;
6068 if (mmap_read_lock_killable(mm
))
6071 /* Untag the address before looking up the VMA */
6072 addr
= untagged_addr_remote(mm
, addr
);
6074 /* Avoid triggering the temporary warning in __get_user_pages */
6075 if (!vma_lookup(mm
, addr
) && !expand_stack(mm
, addr
))
6078 /* ignore errors, just check how much was successfully transferred */
6082 struct vm_area_struct
*vma
= NULL
;
6083 struct page
*page
= get_user_page_vma_remote(mm
, addr
,
6087 /* We might need to expand the stack to access it */
6088 vma
= vma_lookup(mm
, addr
);
6090 vma
= expand_stack(mm
, addr
);
6092 /* mmap_lock was dropped on failure */
6094 return buf
- old_buf
;
6096 /* Try again if stack expansion worked */
6101 * Check if this is a VM_IO | VM_PFNMAP VMA, which
6102 * we can access using slightly different code.
6105 #ifdef CONFIG_HAVE_IOREMAP_PROT
6106 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
6107 bytes
= vma
->vm_ops
->access(vma
, addr
, buf
,
6114 offset
= addr
& (PAGE_SIZE
-1);
6115 if (bytes
> PAGE_SIZE
-offset
)
6116 bytes
= PAGE_SIZE
-offset
;
6118 maddr
= kmap_local_page(page
);
6120 copy_to_user_page(vma
, page
, addr
,
6121 maddr
+ offset
, buf
, bytes
);
6122 set_page_dirty_lock(page
);
6124 copy_from_user_page(vma
, page
, addr
,
6125 buf
, maddr
+ offset
, bytes
);
6127 unmap_and_put_page(page
, maddr
);
6133 mmap_read_unlock(mm
);
6135 return buf
- old_buf
;
6139 * access_remote_vm - access another process' address space
6140 * @mm: the mm_struct of the target address space
6141 * @addr: start address to access
6142 * @buf: source or destination buffer
6143 * @len: number of bytes to transfer
6144 * @gup_flags: flags modifying lookup behaviour
6146 * The caller must hold a reference on @mm.
6148 * Return: number of bytes copied from source to destination.
6150 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
6151 void *buf
, int len
, unsigned int gup_flags
)
6153 return __access_remote_vm(mm
, addr
, buf
, len
, gup_flags
);
6157 * Access another process' address space.
6158 * Source/target buffer must be kernel space,
6159 * Do not walk the page table directly, use get_user_pages
6161 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
6162 void *buf
, int len
, unsigned int gup_flags
)
6164 struct mm_struct
*mm
;
6167 mm
= get_task_mm(tsk
);
6171 ret
= __access_remote_vm(mm
, addr
, buf
, len
, gup_flags
);
6177 EXPORT_SYMBOL_GPL(access_process_vm
);
6180 * Print the name of a VMA.
6182 void print_vma_addr(char *prefix
, unsigned long ip
)
6184 struct mm_struct
*mm
= current
->mm
;
6185 struct vm_area_struct
*vma
;
6188 * we might be running from an atomic context so we cannot sleep
6190 if (!mmap_read_trylock(mm
))
6193 vma
= find_vma(mm
, ip
);
6194 if (vma
&& vma
->vm_file
) {
6195 struct file
*f
= vma
->vm_file
;
6196 char *buf
= (char *)__get_free_page(GFP_NOWAIT
);
6200 p
= file_path(f
, buf
, PAGE_SIZE
);
6203 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
6205 vma
->vm_end
- vma
->vm_start
);
6206 free_page((unsigned long)buf
);
6209 mmap_read_unlock(mm
);
6212 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6213 void __might_fault(const char *file
, int line
)
6215 if (pagefault_disabled())
6217 __might_sleep(file
, line
);
6218 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6220 might_lock_read(¤t
->mm
->mmap_lock
);
6223 EXPORT_SYMBOL(__might_fault
);
6226 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
6228 * Process all subpages of the specified huge page with the specified
6229 * operation. The target subpage will be processed last to keep its
6232 static inline int process_huge_page(
6233 unsigned long addr_hint
, unsigned int pages_per_huge_page
,
6234 int (*process_subpage
)(unsigned long addr
, int idx
, void *arg
),
6237 int i
, n
, base
, l
, ret
;
6238 unsigned long addr
= addr_hint
&
6239 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
6241 /* Process target subpage last to keep its cache lines hot */
6243 n
= (addr_hint
- addr
) / PAGE_SIZE
;
6244 if (2 * n
<= pages_per_huge_page
) {
6245 /* If target subpage in first half of huge page */
6248 /* Process subpages at the end of huge page */
6249 for (i
= pages_per_huge_page
- 1; i
>= 2 * n
; i
--) {
6251 ret
= process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
6256 /* If target subpage in second half of huge page */
6257 base
= pages_per_huge_page
- 2 * (pages_per_huge_page
- n
);
6258 l
= pages_per_huge_page
- n
;
6259 /* Process subpages at the begin of huge page */
6260 for (i
= 0; i
< base
; i
++) {
6262 ret
= process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
6268 * Process remaining subpages in left-right-left-right pattern
6269 * towards the target subpage
6271 for (i
= 0; i
< l
; i
++) {
6272 int left_idx
= base
+ i
;
6273 int right_idx
= base
+ 2 * l
- 1 - i
;
6276 ret
= process_subpage(addr
+ left_idx
* PAGE_SIZE
, left_idx
, arg
);
6280 ret
= process_subpage(addr
+ right_idx
* PAGE_SIZE
, right_idx
, arg
);
6287 static void clear_gigantic_page(struct page
*page
,
6289 unsigned int pages_per_huge_page
)
6295 for (i
= 0; i
< pages_per_huge_page
; i
++) {
6296 p
= nth_page(page
, i
);
6298 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
6302 static int clear_subpage(unsigned long addr
, int idx
, void *arg
)
6304 struct page
*page
= arg
;
6306 clear_user_highpage(nth_page(page
, idx
), addr
);
6310 void clear_huge_page(struct page
*page
,
6311 unsigned long addr_hint
, unsigned int pages_per_huge_page
)
6313 unsigned long addr
= addr_hint
&
6314 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
6316 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
6317 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
6321 process_huge_page(addr_hint
, pages_per_huge_page
, clear_subpage
, page
);
6324 static int copy_user_gigantic_page(struct folio
*dst
, struct folio
*src
,
6326 struct vm_area_struct
*vma
,
6327 unsigned int pages_per_huge_page
)
6330 struct page
*dst_page
;
6331 struct page
*src_page
;
6333 for (i
= 0; i
< pages_per_huge_page
; i
++) {
6334 dst_page
= folio_page(dst
, i
);
6335 src_page
= folio_page(src
, i
);
6338 if (copy_mc_user_highpage(dst_page
, src_page
,
6339 addr
+ i
*PAGE_SIZE
, vma
)) {
6340 memory_failure_queue(page_to_pfn(src_page
), 0);
6347 struct copy_subpage_arg
{
6350 struct vm_area_struct
*vma
;
6353 static int copy_subpage(unsigned long addr
, int idx
, void *arg
)
6355 struct copy_subpage_arg
*copy_arg
= arg
;
6356 struct page
*dst
= nth_page(copy_arg
->dst
, idx
);
6357 struct page
*src
= nth_page(copy_arg
->src
, idx
);
6359 if (copy_mc_user_highpage(dst
, src
, addr
, copy_arg
->vma
)) {
6360 memory_failure_queue(page_to_pfn(src
), 0);
6366 int copy_user_large_folio(struct folio
*dst
, struct folio
*src
,
6367 unsigned long addr_hint
, struct vm_area_struct
*vma
)
6369 unsigned int pages_per_huge_page
= folio_nr_pages(dst
);
6370 unsigned long addr
= addr_hint
&
6371 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
6372 struct copy_subpage_arg arg
= {
6378 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
))
6379 return copy_user_gigantic_page(dst
, src
, addr
, vma
,
6380 pages_per_huge_page
);
6382 return process_huge_page(addr_hint
, pages_per_huge_page
, copy_subpage
, &arg
);
6385 long copy_folio_from_user(struct folio
*dst_folio
,
6386 const void __user
*usr_src
,
6387 bool allow_pagefault
)
6390 unsigned long i
, rc
= 0;
6391 unsigned int nr_pages
= folio_nr_pages(dst_folio
);
6392 unsigned long ret_val
= nr_pages
* PAGE_SIZE
;
6393 struct page
*subpage
;
6395 for (i
= 0; i
< nr_pages
; i
++) {
6396 subpage
= folio_page(dst_folio
, i
);
6397 kaddr
= kmap_local_page(subpage
);
6398 if (!allow_pagefault
)
6399 pagefault_disable();
6400 rc
= copy_from_user(kaddr
, usr_src
+ i
* PAGE_SIZE
, PAGE_SIZE
);
6401 if (!allow_pagefault
)
6403 kunmap_local(kaddr
);
6405 ret_val
-= (PAGE_SIZE
- rc
);
6409 flush_dcache_page(subpage
);
6415 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
6417 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
6419 static struct kmem_cache
*page_ptl_cachep
;
6421 void __init
ptlock_cache_init(void)
6423 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
6427 bool ptlock_alloc(struct ptdesc
*ptdesc
)
6431 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
6438 void ptlock_free(struct ptdesc
*ptdesc
)
6440 kmem_cache_free(page_ptl_cachep
, ptdesc
->ptl
);