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 VM_WARN_ON_ONCE(userfaultfd_wp(vma
));
1540 ksm_might_unmap_zero_page(mm
, ptent
);
1544 folio
= page_folio(page
);
1545 if (unlikely(!should_zap_folio(details
, folio
)))
1549 * Make sure that the common "small folio" case is as fast as possible
1550 * by keeping the batching logic separate.
1552 if (unlikely(folio_test_large(folio
) && max_nr
!= 1)) {
1553 nr
= folio_pte_batch(folio
, addr
, pte
, ptent
, max_nr
, fpb_flags
,
1556 zap_present_folio_ptes(tlb
, vma
, folio
, page
, pte
, ptent
, nr
,
1557 addr
, details
, rss
, force_flush
,
1561 zap_present_folio_ptes(tlb
, vma
, folio
, page
, pte
, ptent
, 1, addr
,
1562 details
, rss
, force_flush
, force_break
);
1566 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1567 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1568 unsigned long addr
, unsigned long end
,
1569 struct zap_details
*details
)
1571 bool force_flush
= false, force_break
= false;
1572 struct mm_struct
*mm
= tlb
->mm
;
1573 int rss
[NR_MM_COUNTERS
];
1580 tlb_change_page_size(tlb
, PAGE_SIZE
);
1582 start_pte
= pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1586 flush_tlb_batched_pending(mm
);
1587 arch_enter_lazy_mmu_mode();
1589 pte_t ptent
= ptep_get(pte
);
1590 struct folio
*folio
;
1595 if (pte_none(ptent
))
1601 if (pte_present(ptent
)) {
1602 max_nr
= (end
- addr
) / PAGE_SIZE
;
1603 nr
= zap_present_ptes(tlb
, vma
, pte
, ptent
, max_nr
,
1604 addr
, details
, rss
, &force_flush
,
1606 if (unlikely(force_break
)) {
1607 addr
+= nr
* PAGE_SIZE
;
1613 entry
= pte_to_swp_entry(ptent
);
1614 if (is_device_private_entry(entry
) ||
1615 is_device_exclusive_entry(entry
)) {
1616 page
= pfn_swap_entry_to_page(entry
);
1617 folio
= page_folio(page
);
1618 if (unlikely(!should_zap_folio(details
, folio
)))
1621 * Both device private/exclusive mappings should only
1622 * work with anonymous page so far, so we don't need to
1623 * consider uffd-wp bit when zap. For more information,
1624 * see zap_install_uffd_wp_if_needed().
1626 WARN_ON_ONCE(!vma_is_anonymous(vma
));
1627 rss
[mm_counter(folio
)]--;
1628 if (is_device_private_entry(entry
))
1629 folio_remove_rmap_pte(folio
, page
, vma
);
1631 } else if (!non_swap_entry(entry
)) {
1632 /* Genuine swap entry, hence a private anon page */
1633 if (!should_zap_cows(details
))
1636 if (unlikely(!free_swap_and_cache(entry
)))
1637 print_bad_pte(vma
, addr
, ptent
, NULL
);
1638 } else if (is_migration_entry(entry
)) {
1639 folio
= pfn_swap_entry_folio(entry
);
1640 if (!should_zap_folio(details
, folio
))
1642 rss
[mm_counter(folio
)]--;
1643 } else if (pte_marker_entry_uffd_wp(entry
)) {
1645 * For anon: always drop the marker; for file: only
1646 * drop the marker if explicitly requested.
1648 if (!vma_is_anonymous(vma
) &&
1649 !zap_drop_file_uffd_wp(details
))
1651 } else if (is_hwpoison_entry(entry
) ||
1652 is_poisoned_swp_entry(entry
)) {
1653 if (!should_zap_cows(details
))
1656 /* We should have covered all the swap entry types */
1657 pr_alert("unrecognized swap entry 0x%lx\n", entry
.val
);
1660 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1661 zap_install_uffd_wp_if_needed(vma
, addr
, pte
, 1, details
, ptent
);
1662 } while (pte
+= nr
, addr
+= PAGE_SIZE
* nr
, addr
!= end
);
1664 add_mm_rss_vec(mm
, rss
);
1665 arch_leave_lazy_mmu_mode();
1667 /* Do the actual TLB flush before dropping ptl */
1669 tlb_flush_mmu_tlbonly(tlb
);
1670 tlb_flush_rmaps(tlb
, vma
);
1672 pte_unmap_unlock(start_pte
, ptl
);
1675 * If we forced a TLB flush (either due to running out of
1676 * batch buffers or because we needed to flush dirty TLB
1677 * entries before releasing the ptl), free the batched
1678 * memory too. Come back again if we didn't do everything.
1686 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1687 struct vm_area_struct
*vma
, pud_t
*pud
,
1688 unsigned long addr
, unsigned long end
,
1689 struct zap_details
*details
)
1694 pmd
= pmd_offset(pud
, addr
);
1696 next
= pmd_addr_end(addr
, end
);
1697 if (is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1698 if (next
- addr
!= HPAGE_PMD_SIZE
)
1699 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
1700 else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
)) {
1705 } else if (details
&& details
->single_folio
&&
1706 folio_test_pmd_mappable(details
->single_folio
) &&
1707 next
- addr
== HPAGE_PMD_SIZE
&& pmd_none(*pmd
)) {
1708 spinlock_t
*ptl
= pmd_lock(tlb
->mm
, pmd
);
1710 * Take and drop THP pmd lock so that we cannot return
1711 * prematurely, while zap_huge_pmd() has cleared *pmd,
1712 * but not yet decremented compound_mapcount().
1716 if (pmd_none(*pmd
)) {
1720 addr
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1723 } while (pmd
++, cond_resched(), addr
!= end
);
1728 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1729 struct vm_area_struct
*vma
, p4d_t
*p4d
,
1730 unsigned long addr
, unsigned long end
,
1731 struct zap_details
*details
)
1736 pud
= pud_offset(p4d
, addr
);
1738 next
= pud_addr_end(addr
, end
);
1739 if (pud_trans_huge(*pud
) || pud_devmap(*pud
)) {
1740 if (next
- addr
!= HPAGE_PUD_SIZE
) {
1741 mmap_assert_locked(tlb
->mm
);
1742 split_huge_pud(vma
, pud
, addr
);
1743 } else if (zap_huge_pud(tlb
, vma
, pud
, addr
))
1747 if (pud_none_or_clear_bad(pud
))
1749 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1752 } while (pud
++, addr
= next
, addr
!= end
);
1757 static inline unsigned long zap_p4d_range(struct mmu_gather
*tlb
,
1758 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1759 unsigned long addr
, unsigned long end
,
1760 struct zap_details
*details
)
1765 p4d
= p4d_offset(pgd
, addr
);
1767 next
= p4d_addr_end(addr
, end
);
1768 if (p4d_none_or_clear_bad(p4d
))
1770 next
= zap_pud_range(tlb
, vma
, p4d
, addr
, next
, details
);
1771 } while (p4d
++, addr
= next
, addr
!= end
);
1776 void unmap_page_range(struct mmu_gather
*tlb
,
1777 struct vm_area_struct
*vma
,
1778 unsigned long addr
, unsigned long end
,
1779 struct zap_details
*details
)
1784 BUG_ON(addr
>= end
);
1785 tlb_start_vma(tlb
, vma
);
1786 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1788 next
= pgd_addr_end(addr
, end
);
1789 if (pgd_none_or_clear_bad(pgd
))
1791 next
= zap_p4d_range(tlb
, vma
, pgd
, addr
, next
, details
);
1792 } while (pgd
++, addr
= next
, addr
!= end
);
1793 tlb_end_vma(tlb
, vma
);
1797 static void unmap_single_vma(struct mmu_gather
*tlb
,
1798 struct vm_area_struct
*vma
, unsigned long start_addr
,
1799 unsigned long end_addr
,
1800 struct zap_details
*details
, bool mm_wr_locked
)
1802 unsigned long start
= max(vma
->vm_start
, start_addr
);
1805 if (start
>= vma
->vm_end
)
1807 end
= min(vma
->vm_end
, end_addr
);
1808 if (end
<= vma
->vm_start
)
1812 uprobe_munmap(vma
, start
, end
);
1814 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1815 untrack_pfn(vma
, 0, 0, mm_wr_locked
);
1818 if (unlikely(is_vm_hugetlb_page(vma
))) {
1820 * It is undesirable to test vma->vm_file as it
1821 * should be non-null for valid hugetlb area.
1822 * However, vm_file will be NULL in the error
1823 * cleanup path of mmap_region. When
1824 * hugetlbfs ->mmap method fails,
1825 * mmap_region() nullifies vma->vm_file
1826 * before calling this function to clean up.
1827 * Since no pte has actually been setup, it is
1828 * safe to do nothing in this case.
1831 zap_flags_t zap_flags
= details
?
1832 details
->zap_flags
: 0;
1833 __unmap_hugepage_range(tlb
, vma
, start
, end
,
1837 unmap_page_range(tlb
, vma
, start
, end
, details
);
1842 * unmap_vmas - unmap a range of memory covered by a list of vma's
1843 * @tlb: address of the caller's struct mmu_gather
1844 * @mas: the maple state
1845 * @vma: the starting vma
1846 * @start_addr: virtual address at which to start unmapping
1847 * @end_addr: virtual address at which to end unmapping
1848 * @tree_end: The maximum index to check
1849 * @mm_wr_locked: lock flag
1851 * Unmap all pages in the vma list.
1853 * Only addresses between `start' and `end' will be unmapped.
1855 * The VMA list must be sorted in ascending virtual address order.
1857 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1858 * range after unmap_vmas() returns. So the only responsibility here is to
1859 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1860 * drops the lock and schedules.
1862 void unmap_vmas(struct mmu_gather
*tlb
, struct ma_state
*mas
,
1863 struct vm_area_struct
*vma
, unsigned long start_addr
,
1864 unsigned long end_addr
, unsigned long tree_end
,
1867 struct mmu_notifier_range range
;
1868 struct zap_details details
= {
1869 .zap_flags
= ZAP_FLAG_DROP_MARKER
| ZAP_FLAG_UNMAP
,
1870 /* Careful - we need to zap private pages too! */
1874 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
->vm_mm
,
1875 start_addr
, end_addr
);
1876 mmu_notifier_invalidate_range_start(&range
);
1878 unsigned long start
= start_addr
;
1879 unsigned long end
= end_addr
;
1880 hugetlb_zap_begin(vma
, &start
, &end
);
1881 unmap_single_vma(tlb
, vma
, start
, end
, &details
,
1883 hugetlb_zap_end(vma
, &details
);
1884 vma
= mas_find(mas
, tree_end
- 1);
1885 } while (vma
&& likely(!xa_is_zero(vma
)));
1886 mmu_notifier_invalidate_range_end(&range
);
1890 * zap_page_range_single - remove user pages in a given range
1891 * @vma: vm_area_struct holding the applicable pages
1892 * @address: starting address of pages to zap
1893 * @size: number of bytes to zap
1894 * @details: details of shared cache invalidation
1896 * The range must fit into one VMA.
1898 void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1899 unsigned long size
, struct zap_details
*details
)
1901 const unsigned long end
= address
+ size
;
1902 struct mmu_notifier_range range
;
1903 struct mmu_gather tlb
;
1906 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
->vm_mm
,
1908 hugetlb_zap_begin(vma
, &range
.start
, &range
.end
);
1909 tlb_gather_mmu(&tlb
, vma
->vm_mm
);
1910 update_hiwater_rss(vma
->vm_mm
);
1911 mmu_notifier_invalidate_range_start(&range
);
1913 * unmap 'address-end' not 'range.start-range.end' as range
1914 * could have been expanded for hugetlb pmd sharing.
1916 unmap_single_vma(&tlb
, vma
, address
, end
, details
, false);
1917 mmu_notifier_invalidate_range_end(&range
);
1918 tlb_finish_mmu(&tlb
);
1919 hugetlb_zap_end(vma
, details
);
1923 * zap_vma_ptes - remove ptes mapping the vma
1924 * @vma: vm_area_struct holding ptes to be zapped
1925 * @address: starting address of pages to zap
1926 * @size: number of bytes to zap
1928 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1930 * The entire address range must be fully contained within the vma.
1933 void zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1936 if (!range_in_vma(vma
, address
, address
+ size
) ||
1937 !(vma
->vm_flags
& VM_PFNMAP
))
1940 zap_page_range_single(vma
, address
, size
, NULL
);
1942 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1944 static pmd_t
*walk_to_pmd(struct mm_struct
*mm
, unsigned long addr
)
1951 pgd
= pgd_offset(mm
, addr
);
1952 p4d
= p4d_alloc(mm
, pgd
, addr
);
1955 pud
= pud_alloc(mm
, p4d
, addr
);
1958 pmd
= pmd_alloc(mm
, pud
, addr
);
1962 VM_BUG_ON(pmd_trans_huge(*pmd
));
1966 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1969 pmd_t
*pmd
= walk_to_pmd(mm
, addr
);
1973 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1976 static int validate_page_before_insert(struct page
*page
)
1978 struct folio
*folio
= page_folio(page
);
1980 if (folio_test_anon(folio
) || folio_test_slab(folio
) ||
1981 page_has_type(page
))
1983 flush_dcache_folio(folio
);
1987 static int insert_page_into_pte_locked(struct vm_area_struct
*vma
, pte_t
*pte
,
1988 unsigned long addr
, struct page
*page
, pgprot_t prot
)
1990 struct folio
*folio
= page_folio(page
);
1992 if (!pte_none(ptep_get(pte
)))
1994 /* Ok, finally just insert the thing.. */
1996 inc_mm_counter(vma
->vm_mm
, mm_counter_file(folio
));
1997 folio_add_file_rmap_pte(folio
, page
, vma
);
1998 set_pte_at(vma
->vm_mm
, addr
, pte
, mk_pte(page
, prot
));
2003 * This is the old fallback for page remapping.
2005 * For historical reasons, it only allows reserved pages. Only
2006 * old drivers should use this, and they needed to mark their
2007 * pages reserved for the old functions anyway.
2009 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2010 struct page
*page
, pgprot_t prot
)
2016 retval
= validate_page_before_insert(page
);
2020 pte
= get_locked_pte(vma
->vm_mm
, addr
, &ptl
);
2023 retval
= insert_page_into_pte_locked(vma
, pte
, addr
, page
, prot
);
2024 pte_unmap_unlock(pte
, ptl
);
2029 static int insert_page_in_batch_locked(struct vm_area_struct
*vma
, pte_t
*pte
,
2030 unsigned long addr
, struct page
*page
, pgprot_t prot
)
2034 if (!page_count(page
))
2036 err
= validate_page_before_insert(page
);
2039 return insert_page_into_pte_locked(vma
, pte
, addr
, page
, prot
);
2042 /* insert_pages() amortizes the cost of spinlock operations
2043 * when inserting pages in a loop.
2045 static int insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
2046 struct page
**pages
, unsigned long *num
, pgprot_t prot
)
2049 pte_t
*start_pte
, *pte
;
2050 spinlock_t
*pte_lock
;
2051 struct mm_struct
*const mm
= vma
->vm_mm
;
2052 unsigned long curr_page_idx
= 0;
2053 unsigned long remaining_pages_total
= *num
;
2054 unsigned long pages_to_write_in_pmd
;
2058 pmd
= walk_to_pmd(mm
, addr
);
2062 pages_to_write_in_pmd
= min_t(unsigned long,
2063 remaining_pages_total
, PTRS_PER_PTE
- pte_index(addr
));
2065 /* Allocate the PTE if necessary; takes PMD lock once only. */
2067 if (pte_alloc(mm
, pmd
))
2070 while (pages_to_write_in_pmd
) {
2072 const int batch_size
= min_t(int, pages_to_write_in_pmd
, 8);
2074 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &pte_lock
);
2079 for (pte
= start_pte
; pte_idx
< batch_size
; ++pte
, ++pte_idx
) {
2080 int err
= insert_page_in_batch_locked(vma
, pte
,
2081 addr
, pages
[curr_page_idx
], prot
);
2082 if (unlikely(err
)) {
2083 pte_unmap_unlock(start_pte
, pte_lock
);
2085 remaining_pages_total
-= pte_idx
;
2091 pte_unmap_unlock(start_pte
, pte_lock
);
2092 pages_to_write_in_pmd
-= batch_size
;
2093 remaining_pages_total
-= batch_size
;
2095 if (remaining_pages_total
)
2099 *num
= remaining_pages_total
;
2104 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
2105 * @vma: user vma to map to
2106 * @addr: target start user address of these pages
2107 * @pages: source kernel pages
2108 * @num: in: number of pages to map. out: number of pages that were *not*
2109 * mapped. (0 means all pages were successfully mapped).
2111 * Preferred over vm_insert_page() when inserting multiple pages.
2113 * In case of error, we may have mapped a subset of the provided
2114 * pages. It is the caller's responsibility to account for this case.
2116 * The same restrictions apply as in vm_insert_page().
2118 int vm_insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
2119 struct page
**pages
, unsigned long *num
)
2121 const unsigned long end_addr
= addr
+ (*num
* PAGE_SIZE
) - 1;
2123 if (addr
< vma
->vm_start
|| end_addr
>= vma
->vm_end
)
2125 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
2126 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
2127 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
2128 vm_flags_set(vma
, VM_MIXEDMAP
);
2130 /* Defer page refcount checking till we're about to map that page. */
2131 return insert_pages(vma
, addr
, pages
, num
, vma
->vm_page_prot
);
2133 EXPORT_SYMBOL(vm_insert_pages
);
2136 * vm_insert_page - insert single page into user vma
2137 * @vma: user vma to map to
2138 * @addr: target user address of this page
2139 * @page: source kernel page
2141 * This allows drivers to insert individual pages they've allocated
2144 * The page has to be a nice clean _individual_ kernel allocation.
2145 * If you allocate a compound page, you need to have marked it as
2146 * such (__GFP_COMP), or manually just split the page up yourself
2147 * (see split_page()).
2149 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2150 * took an arbitrary page protection parameter. This doesn't allow
2151 * that. Your vma protection will have to be set up correctly, which
2152 * means that if you want a shared writable mapping, you'd better
2153 * ask for a shared writable mapping!
2155 * The page does not need to be reserved.
2157 * Usually this function is called from f_op->mmap() handler
2158 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
2159 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2160 * function from other places, for example from page-fault handler.
2162 * Return: %0 on success, negative error code otherwise.
2164 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2167 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2169 if (!page_count(page
))
2171 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
2172 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
2173 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
2174 vm_flags_set(vma
, VM_MIXEDMAP
);
2176 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2178 EXPORT_SYMBOL(vm_insert_page
);
2181 * __vm_map_pages - maps range of kernel pages into user vma
2182 * @vma: user vma to map to
2183 * @pages: pointer to array of source kernel pages
2184 * @num: number of pages in page array
2185 * @offset: user's requested vm_pgoff
2187 * This allows drivers to map range of kernel pages into a user vma.
2189 * Return: 0 on success and error code otherwise.
2191 static int __vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
2192 unsigned long num
, unsigned long offset
)
2194 unsigned long count
= vma_pages(vma
);
2195 unsigned long uaddr
= vma
->vm_start
;
2198 /* Fail if the user requested offset is beyond the end of the object */
2202 /* Fail if the user requested size exceeds available object size */
2203 if (count
> num
- offset
)
2206 for (i
= 0; i
< count
; i
++) {
2207 ret
= vm_insert_page(vma
, uaddr
, pages
[offset
+ i
]);
2217 * vm_map_pages - maps range of kernel pages starts with non zero offset
2218 * @vma: user vma to map to
2219 * @pages: pointer to array of source kernel pages
2220 * @num: number of pages in page array
2222 * Maps an object consisting of @num pages, catering for the user's
2223 * requested vm_pgoff
2225 * If we fail to insert any page into the vma, the function will return
2226 * immediately leaving any previously inserted pages present. Callers
2227 * from the mmap handler may immediately return the error as their caller
2228 * will destroy the vma, removing any successfully inserted pages. Other
2229 * callers should make their own arrangements for calling unmap_region().
2231 * Context: Process context. Called by mmap handlers.
2232 * Return: 0 on success and error code otherwise.
2234 int vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
2237 return __vm_map_pages(vma
, pages
, num
, vma
->vm_pgoff
);
2239 EXPORT_SYMBOL(vm_map_pages
);
2242 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2243 * @vma: user vma to map to
2244 * @pages: pointer to array of source kernel pages
2245 * @num: number of pages in page array
2247 * Similar to vm_map_pages(), except that it explicitly sets the offset
2248 * to 0. This function is intended for the drivers that did not consider
2251 * Context: Process context. Called by mmap handlers.
2252 * Return: 0 on success and error code otherwise.
2254 int vm_map_pages_zero(struct vm_area_struct
*vma
, struct page
**pages
,
2257 return __vm_map_pages(vma
, pages
, num
, 0);
2259 EXPORT_SYMBOL(vm_map_pages_zero
);
2261 static vm_fault_t
insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2262 pfn_t pfn
, pgprot_t prot
, bool mkwrite
)
2264 struct mm_struct
*mm
= vma
->vm_mm
;
2268 pte
= get_locked_pte(mm
, addr
, &ptl
);
2270 return VM_FAULT_OOM
;
2271 entry
= ptep_get(pte
);
2272 if (!pte_none(entry
)) {
2275 * For read faults on private mappings the PFN passed
2276 * in may not match the PFN we have mapped if the
2277 * mapped PFN is a writeable COW page. In the mkwrite
2278 * case we are creating a writable PTE for a shared
2279 * mapping and we expect the PFNs to match. If they
2280 * don't match, we are likely racing with block
2281 * allocation and mapping invalidation so just skip the
2284 if (pte_pfn(entry
) != pfn_t_to_pfn(pfn
)) {
2285 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry
)));
2288 entry
= pte_mkyoung(entry
);
2289 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2290 if (ptep_set_access_flags(vma
, addr
, pte
, entry
, 1))
2291 update_mmu_cache(vma
, addr
, pte
);
2296 /* Ok, finally just insert the thing.. */
2297 if (pfn_t_devmap(pfn
))
2298 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
2300 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
2303 entry
= pte_mkyoung(entry
);
2304 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2307 set_pte_at(mm
, addr
, pte
, entry
);
2308 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
2311 pte_unmap_unlock(pte
, ptl
);
2312 return VM_FAULT_NOPAGE
;
2316 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2317 * @vma: user vma to map to
2318 * @addr: target user address of this page
2319 * @pfn: source kernel pfn
2320 * @pgprot: pgprot flags for the inserted page
2322 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2323 * to override pgprot on a per-page basis.
2325 * This only makes sense for IO mappings, and it makes no sense for
2326 * COW mappings. In general, using multiple vmas is preferable;
2327 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2330 * pgprot typically only differs from @vma->vm_page_prot when drivers set
2331 * caching- and encryption bits different than those of @vma->vm_page_prot,
2332 * because the caching- or encryption mode may not be known at mmap() time.
2334 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2335 * to set caching and encryption bits for those vmas (except for COW pages).
2336 * This is ensured by core vm only modifying these page table entries using
2337 * functions that don't touch caching- or encryption bits, using pte_modify()
2338 * if needed. (See for example mprotect()).
2340 * Also when new page-table entries are created, this is only done using the
2341 * fault() callback, and never using the value of vma->vm_page_prot,
2342 * except for page-table entries that point to anonymous pages as the result
2345 * Context: Process context. May allocate using %GFP_KERNEL.
2346 * Return: vm_fault_t value.
2348 vm_fault_t
vmf_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
2349 unsigned long pfn
, pgprot_t pgprot
)
2352 * Technically, architectures with pte_special can avoid all these
2353 * restrictions (same for remap_pfn_range). However we would like
2354 * consistency in testing and feature parity among all, so we should
2355 * try to keep these invariants in place for everybody.
2357 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
2358 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
2359 (VM_PFNMAP
|VM_MIXEDMAP
));
2360 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
2361 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
2363 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2364 return VM_FAULT_SIGBUS
;
2366 if (!pfn_modify_allowed(pfn
, pgprot
))
2367 return VM_FAULT_SIGBUS
;
2369 track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
));
2371 return insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
,
2374 EXPORT_SYMBOL(vmf_insert_pfn_prot
);
2377 * vmf_insert_pfn - insert single pfn into user vma
2378 * @vma: user vma to map to
2379 * @addr: target user address of this page
2380 * @pfn: source kernel pfn
2382 * Similar to vm_insert_page, this allows drivers to insert individual pages
2383 * they've allocated into a user vma. Same comments apply.
2385 * This function should only be called from a vm_ops->fault handler, and
2386 * in that case the handler should return the result of this function.
2388 * vma cannot be a COW mapping.
2390 * As this is called only for pages that do not currently exist, we
2391 * do not need to flush old virtual caches or the TLB.
2393 * Context: Process context. May allocate using %GFP_KERNEL.
2394 * Return: vm_fault_t value.
2396 vm_fault_t
vmf_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2399 return vmf_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
2401 EXPORT_SYMBOL(vmf_insert_pfn
);
2403 static bool vm_mixed_ok(struct vm_area_struct
*vma
, pfn_t pfn
)
2405 /* these checks mirror the abort conditions in vm_normal_page */
2406 if (vma
->vm_flags
& VM_MIXEDMAP
)
2408 if (pfn_t_devmap(pfn
))
2410 if (pfn_t_special(pfn
))
2412 if (is_zero_pfn(pfn_t_to_pfn(pfn
)))
2417 static vm_fault_t
__vm_insert_mixed(struct vm_area_struct
*vma
,
2418 unsigned long addr
, pfn_t pfn
, bool mkwrite
)
2420 pgprot_t pgprot
= vma
->vm_page_prot
;
2423 BUG_ON(!vm_mixed_ok(vma
, pfn
));
2425 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2426 return VM_FAULT_SIGBUS
;
2428 track_pfn_insert(vma
, &pgprot
, pfn
);
2430 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn
), pgprot
))
2431 return VM_FAULT_SIGBUS
;
2434 * If we don't have pte special, then we have to use the pfn_valid()
2435 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2436 * refcount the page if pfn_valid is true (hence insert_page rather
2437 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2438 * without pte special, it would there be refcounted as a normal page.
2440 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
) &&
2441 !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
2445 * At this point we are committed to insert_page()
2446 * regardless of whether the caller specified flags that
2447 * result in pfn_t_has_page() == false.
2449 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
2450 err
= insert_page(vma
, addr
, page
, pgprot
);
2452 return insert_pfn(vma
, addr
, pfn
, pgprot
, mkwrite
);
2456 return VM_FAULT_OOM
;
2457 if (err
< 0 && err
!= -EBUSY
)
2458 return VM_FAULT_SIGBUS
;
2460 return VM_FAULT_NOPAGE
;
2463 vm_fault_t
vmf_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2466 return __vm_insert_mixed(vma
, addr
, pfn
, false);
2468 EXPORT_SYMBOL(vmf_insert_mixed
);
2471 * If the insertion of PTE failed because someone else already added a
2472 * different entry in the mean time, we treat that as success as we assume
2473 * the same entry was actually inserted.
2475 vm_fault_t
vmf_insert_mixed_mkwrite(struct vm_area_struct
*vma
,
2476 unsigned long addr
, pfn_t pfn
)
2478 return __vm_insert_mixed(vma
, addr
, pfn
, true);
2480 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite
);
2483 * maps a range of physical memory into the requested pages. the old
2484 * mappings are removed. any references to nonexistent pages results
2485 * in null mappings (currently treated as "copy-on-access")
2487 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2488 unsigned long addr
, unsigned long end
,
2489 unsigned long pfn
, pgprot_t prot
)
2491 pte_t
*pte
, *mapped_pte
;
2495 mapped_pte
= pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2498 arch_enter_lazy_mmu_mode();
2500 BUG_ON(!pte_none(ptep_get(pte
)));
2501 if (!pfn_modify_allowed(pfn
, prot
)) {
2505 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2507 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2508 arch_leave_lazy_mmu_mode();
2509 pte_unmap_unlock(mapped_pte
, ptl
);
2513 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2514 unsigned long addr
, unsigned long end
,
2515 unsigned long pfn
, pgprot_t prot
)
2521 pfn
-= addr
>> PAGE_SHIFT
;
2522 pmd
= pmd_alloc(mm
, pud
, addr
);
2525 VM_BUG_ON(pmd_trans_huge(*pmd
));
2527 next
= pmd_addr_end(addr
, end
);
2528 err
= remap_pte_range(mm
, pmd
, addr
, next
,
2529 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2532 } while (pmd
++, addr
= next
, addr
!= end
);
2536 static inline int remap_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2537 unsigned long addr
, unsigned long end
,
2538 unsigned long pfn
, pgprot_t prot
)
2544 pfn
-= addr
>> PAGE_SHIFT
;
2545 pud
= pud_alloc(mm
, p4d
, addr
);
2549 next
= pud_addr_end(addr
, end
);
2550 err
= remap_pmd_range(mm
, pud
, addr
, next
,
2551 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2554 } while (pud
++, addr
= next
, addr
!= end
);
2558 static inline int remap_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2559 unsigned long addr
, unsigned long end
,
2560 unsigned long pfn
, pgprot_t prot
)
2566 pfn
-= addr
>> PAGE_SHIFT
;
2567 p4d
= p4d_alloc(mm
, pgd
, addr
);
2571 next
= p4d_addr_end(addr
, end
);
2572 err
= remap_pud_range(mm
, p4d
, addr
, next
,
2573 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2576 } while (p4d
++, addr
= next
, addr
!= end
);
2581 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2582 * must have pre-validated the caching bits of the pgprot_t.
2584 int remap_pfn_range_notrack(struct vm_area_struct
*vma
, unsigned long addr
,
2585 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2589 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2590 struct mm_struct
*mm
= vma
->vm_mm
;
2593 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr
)))
2597 * Physically remapped pages are special. Tell the
2598 * rest of the world about it:
2599 * VM_IO tells people not to look at these pages
2600 * (accesses can have side effects).
2601 * VM_PFNMAP tells the core MM that the base pages are just
2602 * raw PFN mappings, and do not have a "struct page" associated
2605 * Disable vma merging and expanding with mremap().
2607 * Omit vma from core dump, even when VM_IO turned off.
2609 * There's a horrible special case to handle copy-on-write
2610 * behaviour that some programs depend on. We mark the "original"
2611 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2612 * See vm_normal_page() for details.
2614 if (is_cow_mapping(vma
->vm_flags
)) {
2615 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2617 vma
->vm_pgoff
= pfn
;
2620 vm_flags_set(vma
, VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
);
2622 BUG_ON(addr
>= end
);
2623 pfn
-= addr
>> PAGE_SHIFT
;
2624 pgd
= pgd_offset(mm
, addr
);
2625 flush_cache_range(vma
, addr
, end
);
2627 next
= pgd_addr_end(addr
, end
);
2628 err
= remap_p4d_range(mm
, pgd
, addr
, next
,
2629 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2632 } while (pgd
++, addr
= next
, addr
!= end
);
2638 * remap_pfn_range - remap kernel memory to userspace
2639 * @vma: user vma to map to
2640 * @addr: target page aligned user address to start at
2641 * @pfn: page frame number of kernel physical memory address
2642 * @size: size of mapping area
2643 * @prot: page protection flags for this mapping
2645 * Note: this is only safe if the mm semaphore is held when called.
2647 * Return: %0 on success, negative error code otherwise.
2649 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2650 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2654 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
2658 err
= remap_pfn_range_notrack(vma
, addr
, pfn
, size
, prot
);
2660 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
), true);
2663 EXPORT_SYMBOL(remap_pfn_range
);
2666 * vm_iomap_memory - remap memory to userspace
2667 * @vma: user vma to map to
2668 * @start: start of the physical memory to be mapped
2669 * @len: size of area
2671 * This is a simplified io_remap_pfn_range() for common driver use. The
2672 * driver just needs to give us the physical memory range to be mapped,
2673 * we'll figure out the rest from the vma information.
2675 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2676 * whatever write-combining details or similar.
2678 * Return: %0 on success, negative error code otherwise.
2680 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2682 unsigned long vm_len
, pfn
, pages
;
2684 /* Check that the physical memory area passed in looks valid */
2685 if (start
+ len
< start
)
2688 * You *really* shouldn't map things that aren't page-aligned,
2689 * but we've historically allowed it because IO memory might
2690 * just have smaller alignment.
2692 len
+= start
& ~PAGE_MASK
;
2693 pfn
= start
>> PAGE_SHIFT
;
2694 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2695 if (pfn
+ pages
< pfn
)
2698 /* We start the mapping 'vm_pgoff' pages into the area */
2699 if (vma
->vm_pgoff
> pages
)
2701 pfn
+= vma
->vm_pgoff
;
2702 pages
-= vma
->vm_pgoff
;
2704 /* Can we fit all of the mapping? */
2705 vm_len
= vma
->vm_end
- vma
->vm_start
;
2706 if (vm_len
>> PAGE_SHIFT
> pages
)
2709 /* Ok, let it rip */
2710 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2712 EXPORT_SYMBOL(vm_iomap_memory
);
2714 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2715 unsigned long addr
, unsigned long end
,
2716 pte_fn_t fn
, void *data
, bool create
,
2717 pgtbl_mod_mask
*mask
)
2719 pte_t
*pte
, *mapped_pte
;
2724 mapped_pte
= pte
= (mm
== &init_mm
) ?
2725 pte_alloc_kernel_track(pmd
, addr
, mask
) :
2726 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2730 mapped_pte
= pte
= (mm
== &init_mm
) ?
2731 pte_offset_kernel(pmd
, addr
) :
2732 pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
2737 arch_enter_lazy_mmu_mode();
2741 if (create
|| !pte_none(ptep_get(pte
))) {
2742 err
= fn(pte
++, addr
, data
);
2746 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2748 *mask
|= PGTBL_PTE_MODIFIED
;
2750 arch_leave_lazy_mmu_mode();
2753 pte_unmap_unlock(mapped_pte
, ptl
);
2757 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2758 unsigned long addr
, unsigned long end
,
2759 pte_fn_t fn
, void *data
, bool create
,
2760 pgtbl_mod_mask
*mask
)
2766 BUG_ON(pud_huge(*pud
));
2769 pmd
= pmd_alloc_track(mm
, pud
, addr
, mask
);
2773 pmd
= pmd_offset(pud
, addr
);
2776 next
= pmd_addr_end(addr
, end
);
2777 if (pmd_none(*pmd
) && !create
)
2779 if (WARN_ON_ONCE(pmd_leaf(*pmd
)))
2781 if (!pmd_none(*pmd
) && WARN_ON_ONCE(pmd_bad(*pmd
))) {
2786 err
= apply_to_pte_range(mm
, pmd
, addr
, next
,
2787 fn
, data
, create
, mask
);
2790 } while (pmd
++, addr
= next
, addr
!= end
);
2795 static int apply_to_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2796 unsigned long addr
, unsigned long end
,
2797 pte_fn_t fn
, void *data
, bool create
,
2798 pgtbl_mod_mask
*mask
)
2805 pud
= pud_alloc_track(mm
, p4d
, addr
, mask
);
2809 pud
= pud_offset(p4d
, addr
);
2812 next
= pud_addr_end(addr
, end
);
2813 if (pud_none(*pud
) && !create
)
2815 if (WARN_ON_ONCE(pud_leaf(*pud
)))
2817 if (!pud_none(*pud
) && WARN_ON_ONCE(pud_bad(*pud
))) {
2822 err
= apply_to_pmd_range(mm
, pud
, addr
, next
,
2823 fn
, data
, create
, mask
);
2826 } while (pud
++, addr
= next
, addr
!= end
);
2831 static int apply_to_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2832 unsigned long addr
, unsigned long end
,
2833 pte_fn_t fn
, void *data
, bool create
,
2834 pgtbl_mod_mask
*mask
)
2841 p4d
= p4d_alloc_track(mm
, pgd
, addr
, mask
);
2845 p4d
= p4d_offset(pgd
, addr
);
2848 next
= p4d_addr_end(addr
, end
);
2849 if (p4d_none(*p4d
) && !create
)
2851 if (WARN_ON_ONCE(p4d_leaf(*p4d
)))
2853 if (!p4d_none(*p4d
) && WARN_ON_ONCE(p4d_bad(*p4d
))) {
2858 err
= apply_to_pud_range(mm
, p4d
, addr
, next
,
2859 fn
, data
, create
, mask
);
2862 } while (p4d
++, addr
= next
, addr
!= end
);
2867 static int __apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2868 unsigned long size
, pte_fn_t fn
,
2869 void *data
, bool create
)
2872 unsigned long start
= addr
, next
;
2873 unsigned long end
= addr
+ size
;
2874 pgtbl_mod_mask mask
= 0;
2877 if (WARN_ON(addr
>= end
))
2880 pgd
= pgd_offset(mm
, addr
);
2882 next
= pgd_addr_end(addr
, end
);
2883 if (pgd_none(*pgd
) && !create
)
2885 if (WARN_ON_ONCE(pgd_leaf(*pgd
)))
2887 if (!pgd_none(*pgd
) && WARN_ON_ONCE(pgd_bad(*pgd
))) {
2892 err
= apply_to_p4d_range(mm
, pgd
, addr
, next
,
2893 fn
, data
, create
, &mask
);
2896 } while (pgd
++, addr
= next
, addr
!= end
);
2898 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
2899 arch_sync_kernel_mappings(start
, start
+ size
);
2905 * Scan a region of virtual memory, filling in page tables as necessary
2906 * and calling a provided function on each leaf page table.
2908 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2909 unsigned long size
, pte_fn_t fn
, void *data
)
2911 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, true);
2913 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2916 * Scan a region of virtual memory, calling a provided function on
2917 * each leaf page table where it exists.
2919 * Unlike apply_to_page_range, this does _not_ fill in page tables
2920 * where they are absent.
2922 int apply_to_existing_page_range(struct mm_struct
*mm
, unsigned long addr
,
2923 unsigned long size
, pte_fn_t fn
, void *data
)
2925 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, false);
2927 EXPORT_SYMBOL_GPL(apply_to_existing_page_range
);
2930 * handle_pte_fault chooses page fault handler according to an entry which was
2931 * read non-atomically. Before making any commitment, on those architectures
2932 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2933 * parts, do_swap_page must check under lock before unmapping the pte and
2934 * proceeding (but do_wp_page is only called after already making such a check;
2935 * and do_anonymous_page can safely check later on).
2937 static inline int pte_unmap_same(struct vm_fault
*vmf
)
2940 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2941 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2942 spin_lock(vmf
->ptl
);
2943 same
= pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
);
2944 spin_unlock(vmf
->ptl
);
2947 pte_unmap(vmf
->pte
);
2954 * 0: copied succeeded
2955 * -EHWPOISON: copy failed due to hwpoison in source page
2956 * -EAGAIN: copied failed (some other reason)
2958 static inline int __wp_page_copy_user(struct page
*dst
, struct page
*src
,
2959 struct vm_fault
*vmf
)
2964 struct vm_area_struct
*vma
= vmf
->vma
;
2965 struct mm_struct
*mm
= vma
->vm_mm
;
2966 unsigned long addr
= vmf
->address
;
2969 if (copy_mc_user_highpage(dst
, src
, addr
, vma
)) {
2970 memory_failure_queue(page_to_pfn(src
), 0);
2977 * If the source page was a PFN mapping, we don't have
2978 * a "struct page" for it. We do a best-effort copy by
2979 * just copying from the original user address. If that
2980 * fails, we just zero-fill it. Live with it.
2982 kaddr
= kmap_local_page(dst
);
2983 pagefault_disable();
2984 uaddr
= (void __user
*)(addr
& PAGE_MASK
);
2987 * On architectures with software "accessed" bits, we would
2988 * take a double page fault, so mark it accessed here.
2991 if (!arch_has_hw_pte_young() && !pte_young(vmf
->orig_pte
)) {
2994 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2995 if (unlikely(!vmf
->pte
|| !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
2997 * Other thread has already handled the fault
2998 * and update local tlb only
3001 update_mmu_tlb(vma
, addr
, vmf
->pte
);
3006 entry
= pte_mkyoung(vmf
->orig_pte
);
3007 if (ptep_set_access_flags(vma
, addr
, vmf
->pte
, entry
, 0))
3008 update_mmu_cache_range(vmf
, vma
, addr
, vmf
->pte
, 1);
3012 * This really shouldn't fail, because the page is there
3013 * in the page tables. But it might just be unreadable,
3014 * in which case we just give up and fill the result with
3017 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
3021 /* Re-validate under PTL if the page is still mapped */
3022 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
3023 if (unlikely(!vmf
->pte
|| !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
3024 /* The PTE changed under us, update local tlb */
3026 update_mmu_tlb(vma
, addr
, vmf
->pte
);
3032 * The same page can be mapped back since last copy attempt.
3033 * Try to copy again under PTL.
3035 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
3037 * Give a warn in case there can be some obscure
3050 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3052 kunmap_local(kaddr
);
3053 flush_dcache_page(dst
);
3058 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
3060 struct file
*vm_file
= vma
->vm_file
;
3063 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
3066 * Special mappings (e.g. VDSO) do not have any file so fake
3067 * a default GFP_KERNEL for them.
3073 * Notify the address space that the page is about to become writable so that
3074 * it can prohibit this or wait for the page to get into an appropriate state.
3076 * We do this without the lock held, so that it can sleep if it needs to.
3078 static vm_fault_t
do_page_mkwrite(struct vm_fault
*vmf
, struct folio
*folio
)
3081 unsigned int old_flags
= vmf
->flags
;
3083 vmf
->flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
3085 if (vmf
->vma
->vm_file
&&
3086 IS_SWAPFILE(vmf
->vma
->vm_file
->f_mapping
->host
))
3087 return VM_FAULT_SIGBUS
;
3089 ret
= vmf
->vma
->vm_ops
->page_mkwrite(vmf
);
3090 /* Restore original flags so that caller is not surprised */
3091 vmf
->flags
= old_flags
;
3092 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
3094 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
3096 if (!folio
->mapping
) {
3097 folio_unlock(folio
);
3098 return 0; /* retry */
3100 ret
|= VM_FAULT_LOCKED
;
3102 VM_BUG_ON_FOLIO(!folio_test_locked(folio
), folio
);
3107 * Handle dirtying of a page in shared file mapping on a write fault.
3109 * The function expects the page to be locked and unlocks it.
3111 static vm_fault_t
fault_dirty_shared_page(struct vm_fault
*vmf
)
3113 struct vm_area_struct
*vma
= vmf
->vma
;
3114 struct address_space
*mapping
;
3115 struct folio
*folio
= page_folio(vmf
->page
);
3117 bool page_mkwrite
= vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
;
3119 dirtied
= folio_mark_dirty(folio
);
3120 VM_BUG_ON_FOLIO(folio_test_anon(folio
), folio
);
3122 * Take a local copy of the address_space - folio.mapping may be zeroed
3123 * by truncate after folio_unlock(). The address_space itself remains
3124 * pinned by vma->vm_file's reference. We rely on folio_unlock()'s
3125 * release semantics to prevent the compiler from undoing this copying.
3127 mapping
= folio_raw_mapping(folio
);
3128 folio_unlock(folio
);
3131 file_update_time(vma
->vm_file
);
3134 * Throttle page dirtying rate down to writeback speed.
3136 * mapping may be NULL here because some device drivers do not
3137 * set page.mapping but still dirty their pages
3139 * Drop the mmap_lock before waiting on IO, if we can. The file
3140 * is pinning the mapping, as per above.
3142 if ((dirtied
|| page_mkwrite
) && mapping
) {
3145 fpin
= maybe_unlock_mmap_for_io(vmf
, NULL
);
3146 balance_dirty_pages_ratelimited(mapping
);
3149 return VM_FAULT_COMPLETED
;
3157 * Handle write page faults for pages that can be reused in the current vma
3159 * This can happen either due to the mapping being with the VM_SHARED flag,
3160 * or due to us being the last reference standing to the page. In either
3161 * case, all we need to do here is to mark the page as writable and update
3162 * any related book-keeping.
3164 static inline void wp_page_reuse(struct vm_fault
*vmf
, struct folio
*folio
)
3165 __releases(vmf
->ptl
)
3167 struct vm_area_struct
*vma
= vmf
->vma
;
3170 VM_BUG_ON(!(vmf
->flags
& FAULT_FLAG_WRITE
));
3173 VM_BUG_ON(folio_test_anon(folio
) &&
3174 !PageAnonExclusive(vmf
->page
));
3176 * Clear the folio's cpupid information as the existing
3177 * information potentially belongs to a now completely
3178 * unrelated process.
3180 folio_xchg_last_cpupid(folio
, (1 << LAST_CPUPID_SHIFT
) - 1);
3183 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
3184 entry
= pte_mkyoung(vmf
->orig_pte
);
3185 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3186 if (ptep_set_access_flags(vma
, vmf
->address
, vmf
->pte
, entry
, 1))
3187 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
3188 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3189 count_vm_event(PGREUSE
);
3193 * We could add a bitflag somewhere, but for now, we know that all
3194 * vm_ops that have a ->map_pages have been audited and don't need
3195 * the mmap_lock to be held.
3197 static inline vm_fault_t
vmf_can_call_fault(const struct vm_fault
*vmf
)
3199 struct vm_area_struct
*vma
= vmf
->vma
;
3201 if (vma
->vm_ops
->map_pages
|| !(vmf
->flags
& FAULT_FLAG_VMA_LOCK
))
3204 return VM_FAULT_RETRY
;
3207 vm_fault_t
vmf_anon_prepare(struct vm_fault
*vmf
)
3209 struct vm_area_struct
*vma
= vmf
->vma
;
3211 if (likely(vma
->anon_vma
))
3213 if (vmf
->flags
& FAULT_FLAG_VMA_LOCK
) {
3215 return VM_FAULT_RETRY
;
3217 if (__anon_vma_prepare(vma
))
3218 return VM_FAULT_OOM
;
3223 * Handle the case of a page which we actually need to copy to a new page,
3224 * either due to COW or unsharing.
3226 * Called with mmap_lock locked and the old page referenced, but
3227 * without the ptl held.
3229 * High level logic flow:
3231 * - Allocate a page, copy the content of the old page to the new one.
3232 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
3233 * - Take the PTL. If the pte changed, bail out and release the allocated page
3234 * - If the pte is still the way we remember it, update the page table and all
3235 * relevant references. This includes dropping the reference the page-table
3236 * held to the old page, as well as updating the rmap.
3237 * - In any case, unlock the PTL and drop the reference we took to the old page.
3239 static vm_fault_t
wp_page_copy(struct vm_fault
*vmf
)
3241 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
3242 struct vm_area_struct
*vma
= vmf
->vma
;
3243 struct mm_struct
*mm
= vma
->vm_mm
;
3244 struct folio
*old_folio
= NULL
;
3245 struct folio
*new_folio
= NULL
;
3247 int page_copied
= 0;
3248 struct mmu_notifier_range range
;
3252 delayacct_wpcopy_start();
3255 old_folio
= page_folio(vmf
->page
);
3256 ret
= vmf_anon_prepare(vmf
);
3260 pfn_is_zero
= is_zero_pfn(pte_pfn(vmf
->orig_pte
));
3261 new_folio
= folio_prealloc(mm
, vma
, vmf
->address
, pfn_is_zero
);
3268 err
= __wp_page_copy_user(&new_folio
->page
, vmf
->page
, vmf
);
3271 * COW failed, if the fault was solved by other,
3272 * it's fine. If not, userspace would re-fault on
3273 * the same address and we will handle the fault
3274 * from the second attempt.
3275 * The -EHWPOISON case will not be retried.
3277 folio_put(new_folio
);
3279 folio_put(old_folio
);
3281 delayacct_wpcopy_end();
3282 return err
== -EHWPOISON
? VM_FAULT_HWPOISON
: 0;
3284 kmsan_copy_page_meta(&new_folio
->page
, vmf
->page
);
3287 __folio_mark_uptodate(new_folio
);
3289 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, mm
,
3290 vmf
->address
& PAGE_MASK
,
3291 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
);
3292 mmu_notifier_invalidate_range_start(&range
);
3295 * Re-check the pte - we dropped the lock
3297 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, vmf
->address
, &vmf
->ptl
);
3298 if (likely(vmf
->pte
&& pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
3300 if (!folio_test_anon(old_folio
)) {
3301 dec_mm_counter(mm
, mm_counter_file(old_folio
));
3302 inc_mm_counter(mm
, MM_ANONPAGES
);
3305 ksm_might_unmap_zero_page(mm
, vmf
->orig_pte
);
3306 inc_mm_counter(mm
, MM_ANONPAGES
);
3308 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
3309 entry
= mk_pte(&new_folio
->page
, vma
->vm_page_prot
);
3310 entry
= pte_sw_mkyoung(entry
);
3311 if (unlikely(unshare
)) {
3312 if (pte_soft_dirty(vmf
->orig_pte
))
3313 entry
= pte_mksoft_dirty(entry
);
3314 if (pte_uffd_wp(vmf
->orig_pte
))
3315 entry
= pte_mkuffd_wp(entry
);
3317 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3321 * Clear the pte entry and flush it first, before updating the
3322 * pte with the new entry, to keep TLBs on different CPUs in
3323 * sync. This code used to set the new PTE then flush TLBs, but
3324 * that left a window where the new PTE could be loaded into
3325 * some TLBs while the old PTE remains in others.
3327 ptep_clear_flush(vma
, vmf
->address
, vmf
->pte
);
3328 folio_add_new_anon_rmap(new_folio
, vma
, vmf
->address
);
3329 folio_add_lru_vma(new_folio
, vma
);
3331 * We call the notify macro here because, when using secondary
3332 * mmu page tables (such as kvm shadow page tables), we want the
3333 * new page to be mapped directly into the secondary page table.
3335 BUG_ON(unshare
&& pte_write(entry
));
3336 set_pte_at_notify(mm
, vmf
->address
, vmf
->pte
, entry
);
3337 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
3340 * Only after switching the pte to the new page may
3341 * we remove the mapcount here. Otherwise another
3342 * process may come and find the rmap count decremented
3343 * before the pte is switched to the new page, and
3344 * "reuse" the old page writing into it while our pte
3345 * here still points into it and can be read by other
3348 * The critical issue is to order this
3349 * folio_remove_rmap_pte() with the ptp_clear_flush
3350 * above. Those stores are ordered by (if nothing else,)
3351 * the barrier present in the atomic_add_negative
3352 * in folio_remove_rmap_pte();
3354 * Then the TLB flush in ptep_clear_flush ensures that
3355 * no process can access the old page before the
3356 * decremented mapcount is visible. And the old page
3357 * cannot be reused until after the decremented
3358 * mapcount is visible. So transitively, TLBs to
3359 * old page will be flushed before it can be reused.
3361 folio_remove_rmap_pte(old_folio
, vmf
->page
, vma
);
3364 /* Free the old page.. */
3365 new_folio
= old_folio
;
3367 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3368 } else if (vmf
->pte
) {
3369 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
3370 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3373 mmu_notifier_invalidate_range_end(&range
);
3376 folio_put(new_folio
);
3379 free_swap_cache(old_folio
);
3380 folio_put(old_folio
);
3383 delayacct_wpcopy_end();
3389 folio_put(old_folio
);
3391 delayacct_wpcopy_end();
3396 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3397 * writeable once the page is prepared
3399 * @vmf: structure describing the fault
3400 * @folio: the folio of vmf->page
3402 * This function handles all that is needed to finish a write page fault in a
3403 * shared mapping due to PTE being read-only once the mapped page is prepared.
3404 * It handles locking of PTE and modifying it.
3406 * The function expects the page to be locked or other protection against
3407 * concurrent faults / writeback (such as DAX radix tree locks).
3409 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3410 * we acquired PTE lock.
3412 static vm_fault_t
finish_mkwrite_fault(struct vm_fault
*vmf
, struct folio
*folio
)
3414 WARN_ON_ONCE(!(vmf
->vma
->vm_flags
& VM_SHARED
));
3415 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3418 return VM_FAULT_NOPAGE
;
3420 * We might have raced with another page fault while we released the
3421 * pte_offset_map_lock.
3423 if (!pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)) {
3424 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
3425 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3426 return VM_FAULT_NOPAGE
;
3428 wp_page_reuse(vmf
, folio
);
3433 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3436 static vm_fault_t
wp_pfn_shared(struct vm_fault
*vmf
)
3438 struct vm_area_struct
*vma
= vmf
->vma
;
3440 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
3443 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3444 ret
= vmf_can_call_fault(vmf
);
3448 vmf
->flags
|= FAULT_FLAG_MKWRITE
;
3449 ret
= vma
->vm_ops
->pfn_mkwrite(vmf
);
3450 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))
3452 return finish_mkwrite_fault(vmf
, NULL
);
3454 wp_page_reuse(vmf
, NULL
);
3458 static vm_fault_t
wp_page_shared(struct vm_fault
*vmf
, struct folio
*folio
)
3459 __releases(vmf
->ptl
)
3461 struct vm_area_struct
*vma
= vmf
->vma
;
3466 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
3469 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3470 tmp
= vmf_can_call_fault(vmf
);
3476 tmp
= do_page_mkwrite(vmf
, folio
);
3477 if (unlikely(!tmp
|| (tmp
&
3478 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3482 tmp
= finish_mkwrite_fault(vmf
, folio
);
3483 if (unlikely(tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3484 folio_unlock(folio
);
3489 wp_page_reuse(vmf
, folio
);
3492 ret
|= fault_dirty_shared_page(vmf
);
3498 static bool wp_can_reuse_anon_folio(struct folio
*folio
,
3499 struct vm_area_struct
*vma
)
3502 * We could currently only reuse a subpage of a large folio if no
3503 * other subpages of the large folios are still mapped. However,
3504 * let's just consistently not reuse subpages even if we could
3505 * reuse in that scenario, and give back a large folio a bit
3508 if (folio_test_large(folio
))
3512 * We have to verify under folio lock: these early checks are
3513 * just an optimization to avoid locking the folio and freeing
3514 * the swapcache if there is little hope that we can reuse.
3516 * KSM doesn't necessarily raise the folio refcount.
3518 if (folio_test_ksm(folio
) || folio_ref_count(folio
) > 3)
3520 if (!folio_test_lru(folio
))
3522 * We cannot easily detect+handle references from
3523 * remote LRU caches or references to LRU folios.
3526 if (folio_ref_count(folio
) > 1 + folio_test_swapcache(folio
))
3528 if (!folio_trylock(folio
))
3530 if (folio_test_swapcache(folio
))
3531 folio_free_swap(folio
);
3532 if (folio_test_ksm(folio
) || folio_ref_count(folio
) != 1) {
3533 folio_unlock(folio
);
3537 * Ok, we've got the only folio reference from our mapping
3538 * and the folio is locked, it's dark out, and we're wearing
3539 * sunglasses. Hit it.
3541 folio_move_anon_rmap(folio
, vma
);
3542 folio_unlock(folio
);
3547 * This routine handles present pages, when
3548 * * users try to write to a shared page (FAULT_FLAG_WRITE)
3549 * * GUP wants to take a R/O pin on a possibly shared anonymous page
3550 * (FAULT_FLAG_UNSHARE)
3552 * It is done by copying the page to a new address and decrementing the
3553 * shared-page counter for the old page.
3555 * Note that this routine assumes that the protection checks have been
3556 * done by the caller (the low-level page fault routine in most cases).
3557 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've
3558 * done any necessary COW.
3560 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even
3561 * though the page will change only once the write actually happens. This
3562 * avoids a few races, and potentially makes it more efficient.
3564 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3565 * but allow concurrent faults), with pte both mapped and locked.
3566 * We return with mmap_lock still held, but pte unmapped and unlocked.
3568 static vm_fault_t
do_wp_page(struct vm_fault
*vmf
)
3569 __releases(vmf
->ptl
)
3571 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
3572 struct vm_area_struct
*vma
= vmf
->vma
;
3573 struct folio
*folio
= NULL
;
3576 if (likely(!unshare
)) {
3577 if (userfaultfd_pte_wp(vma
, ptep_get(vmf
->pte
))) {
3578 if (!userfaultfd_wp_async(vma
)) {
3579 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3580 return handle_userfault(vmf
, VM_UFFD_WP
);
3584 * Nothing needed (cache flush, TLB invalidations,
3585 * etc.) because we're only removing the uffd-wp bit,
3586 * which is completely invisible to the user.
3588 pte
= pte_clear_uffd_wp(ptep_get(vmf
->pte
));
3590 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3592 * Update this to be prepared for following up CoW
3595 vmf
->orig_pte
= pte
;
3599 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3600 * is flushed in this case before copying.
3602 if (unlikely(userfaultfd_wp(vmf
->vma
) &&
3603 mm_tlb_flush_pending(vmf
->vma
->vm_mm
)))
3604 flush_tlb_page(vmf
->vma
, vmf
->address
);
3607 vmf
->page
= vm_normal_page(vma
, vmf
->address
, vmf
->orig_pte
);
3610 folio
= page_folio(vmf
->page
);
3613 * Shared mapping: we are guaranteed to have VM_WRITE and
3614 * FAULT_FLAG_WRITE set at this point.
3616 if (vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
)) {
3618 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3621 * We should not cow pages in a shared writeable mapping.
3622 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3625 return wp_pfn_shared(vmf
);
3626 return wp_page_shared(vmf
, folio
);
3630 * Private mapping: create an exclusive anonymous page copy if reuse
3631 * is impossible. We might miss VM_WRITE for FOLL_FORCE handling.
3633 * If we encounter a page that is marked exclusive, we must reuse
3634 * the page without further checks.
3636 if (folio
&& folio_test_anon(folio
) &&
3637 (PageAnonExclusive(vmf
->page
) || wp_can_reuse_anon_folio(folio
, vma
))) {
3638 if (!PageAnonExclusive(vmf
->page
))
3639 SetPageAnonExclusive(vmf
->page
);
3640 if (unlikely(unshare
)) {
3641 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3644 wp_page_reuse(vmf
, folio
);
3648 * Ok, we need to copy. Oh, well..
3653 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3655 if (folio
&& folio_test_ksm(folio
))
3656 count_vm_event(COW_KSM
);
3658 return wp_page_copy(vmf
);
3661 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
3662 unsigned long start_addr
, unsigned long end_addr
,
3663 struct zap_details
*details
)
3665 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
3668 static inline void unmap_mapping_range_tree(struct rb_root_cached
*root
,
3669 pgoff_t first_index
,
3671 struct zap_details
*details
)
3673 struct vm_area_struct
*vma
;
3674 pgoff_t vba
, vea
, zba
, zea
;
3676 vma_interval_tree_foreach(vma
, root
, first_index
, last_index
) {
3677 vba
= vma
->vm_pgoff
;
3678 vea
= vba
+ vma_pages(vma
) - 1;
3679 zba
= max(first_index
, vba
);
3680 zea
= min(last_index
, vea
);
3682 unmap_mapping_range_vma(vma
,
3683 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
3684 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
3690 * unmap_mapping_folio() - Unmap single folio from processes.
3691 * @folio: The locked folio to be unmapped.
3693 * Unmap this folio from any userspace process which still has it mmaped.
3694 * Typically, for efficiency, the range of nearby pages has already been
3695 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3696 * truncation or invalidation holds the lock on a folio, it may find that
3697 * the page has been remapped again: and then uses unmap_mapping_folio()
3698 * to unmap it finally.
3700 void unmap_mapping_folio(struct folio
*folio
)
3702 struct address_space
*mapping
= folio
->mapping
;
3703 struct zap_details details
= { };
3704 pgoff_t first_index
;
3707 VM_BUG_ON(!folio_test_locked(folio
));
3709 first_index
= folio
->index
;
3710 last_index
= folio_next_index(folio
) - 1;
3712 details
.even_cows
= false;
3713 details
.single_folio
= folio
;
3714 details
.zap_flags
= ZAP_FLAG_DROP_MARKER
;
3716 i_mmap_lock_read(mapping
);
3717 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3718 unmap_mapping_range_tree(&mapping
->i_mmap
, first_index
,
3719 last_index
, &details
);
3720 i_mmap_unlock_read(mapping
);
3724 * unmap_mapping_pages() - Unmap pages from processes.
3725 * @mapping: The address space containing pages to be unmapped.
3726 * @start: Index of first page to be unmapped.
3727 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3728 * @even_cows: Whether to unmap even private COWed pages.
3730 * Unmap the pages in this address space from any userspace process which
3731 * has them mmaped. Generally, you want to remove COWed pages as well when
3732 * a file is being truncated, but not when invalidating pages from the page
3735 void unmap_mapping_pages(struct address_space
*mapping
, pgoff_t start
,
3736 pgoff_t nr
, bool even_cows
)
3738 struct zap_details details
= { };
3739 pgoff_t first_index
= start
;
3740 pgoff_t last_index
= start
+ nr
- 1;
3742 details
.even_cows
= even_cows
;
3743 if (last_index
< first_index
)
3744 last_index
= ULONG_MAX
;
3746 i_mmap_lock_read(mapping
);
3747 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3748 unmap_mapping_range_tree(&mapping
->i_mmap
, first_index
,
3749 last_index
, &details
);
3750 i_mmap_unlock_read(mapping
);
3752 EXPORT_SYMBOL_GPL(unmap_mapping_pages
);
3755 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3756 * address_space corresponding to the specified byte range in the underlying
3759 * @mapping: the address space containing mmaps to be unmapped.
3760 * @holebegin: byte in first page to unmap, relative to the start of
3761 * the underlying file. This will be rounded down to a PAGE_SIZE
3762 * boundary. Note that this is different from truncate_pagecache(), which
3763 * must keep the partial page. In contrast, we must get rid of
3765 * @holelen: size of prospective hole in bytes. This will be rounded
3766 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3768 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3769 * but 0 when invalidating pagecache, don't throw away private data.
3771 void unmap_mapping_range(struct address_space
*mapping
,
3772 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
3774 pgoff_t hba
= (pgoff_t
)(holebegin
) >> PAGE_SHIFT
;
3775 pgoff_t hlen
= ((pgoff_t
)(holelen
) + PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3777 /* Check for overflow. */
3778 if (sizeof(holelen
) > sizeof(hlen
)) {
3780 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3781 if (holeend
& ~(long long)ULONG_MAX
)
3782 hlen
= ULONG_MAX
- hba
+ 1;
3785 unmap_mapping_pages(mapping
, hba
, hlen
, even_cows
);
3787 EXPORT_SYMBOL(unmap_mapping_range
);
3790 * Restore a potential device exclusive pte to a working pte entry
3792 static vm_fault_t
remove_device_exclusive_entry(struct vm_fault
*vmf
)
3794 struct folio
*folio
= page_folio(vmf
->page
);
3795 struct vm_area_struct
*vma
= vmf
->vma
;
3796 struct mmu_notifier_range range
;
3800 * We need a reference to lock the folio because we don't hold
3801 * the PTL so a racing thread can remove the device-exclusive
3802 * entry and unmap it. If the folio is free the entry must
3803 * have been removed already. If it happens to have already
3804 * been re-allocated after being freed all we do is lock and
3807 if (!folio_try_get(folio
))
3810 ret
= folio_lock_or_retry(folio
, vmf
);
3815 mmu_notifier_range_init_owner(&range
, MMU_NOTIFY_EXCLUSIVE
, 0,
3816 vma
->vm_mm
, vmf
->address
& PAGE_MASK
,
3817 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
, NULL
);
3818 mmu_notifier_invalidate_range_start(&range
);
3820 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3822 if (likely(vmf
->pte
&& pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)))
3823 restore_exclusive_pte(vma
, vmf
->page
, vmf
->address
, vmf
->pte
);
3826 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3827 folio_unlock(folio
);
3830 mmu_notifier_invalidate_range_end(&range
);
3834 static inline bool should_try_to_free_swap(struct folio
*folio
,
3835 struct vm_area_struct
*vma
,
3836 unsigned int fault_flags
)
3838 if (!folio_test_swapcache(folio
))
3840 if (mem_cgroup_swap_full(folio
) || (vma
->vm_flags
& VM_LOCKED
) ||
3841 folio_test_mlocked(folio
))
3844 * If we want to map a page that's in the swapcache writable, we
3845 * have to detect via the refcount if we're really the exclusive
3846 * user. Try freeing the swapcache to get rid of the swapcache
3847 * reference only in case it's likely that we'll be the exlusive user.
3849 return (fault_flags
& FAULT_FLAG_WRITE
) && !folio_test_ksm(folio
) &&
3850 folio_ref_count(folio
) == 2;
3853 static vm_fault_t
pte_marker_clear(struct vm_fault
*vmf
)
3855 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
,
3856 vmf
->address
, &vmf
->ptl
);
3860 * Be careful so that we will only recover a special uffd-wp pte into a
3861 * none pte. Otherwise it means the pte could have changed, so retry.
3863 * This should also cover the case where e.g. the pte changed
3864 * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_POISONED.
3865 * So is_pte_marker() check is not enough to safely drop the pte.
3867 if (pte_same(vmf
->orig_pte
, ptep_get(vmf
->pte
)))
3868 pte_clear(vmf
->vma
->vm_mm
, vmf
->address
, vmf
->pte
);
3869 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3873 static vm_fault_t
do_pte_missing(struct vm_fault
*vmf
)
3875 if (vma_is_anonymous(vmf
->vma
))
3876 return do_anonymous_page(vmf
);
3878 return do_fault(vmf
);
3882 * This is actually a page-missing access, but with uffd-wp special pte
3883 * installed. It means this pte was wr-protected before being unmapped.
3885 static vm_fault_t
pte_marker_handle_uffd_wp(struct vm_fault
*vmf
)
3888 * Just in case there're leftover special ptes even after the region
3889 * got unregistered - we can simply clear them.
3891 if (unlikely(!userfaultfd_wp(vmf
->vma
)))
3892 return pte_marker_clear(vmf
);
3894 return do_pte_missing(vmf
);
3897 static vm_fault_t
handle_pte_marker(struct vm_fault
*vmf
)
3899 swp_entry_t entry
= pte_to_swp_entry(vmf
->orig_pte
);
3900 unsigned long marker
= pte_marker_get(entry
);
3903 * PTE markers should never be empty. If anything weird happened,
3904 * the best thing to do is to kill the process along with its mm.
3906 if (WARN_ON_ONCE(!marker
))
3907 return VM_FAULT_SIGBUS
;
3909 /* Higher priority than uffd-wp when data corrupted */
3910 if (marker
& PTE_MARKER_POISONED
)
3911 return VM_FAULT_HWPOISON
;
3913 if (pte_marker_entry_uffd_wp(entry
))
3914 return pte_marker_handle_uffd_wp(vmf
);
3916 /* This is an unknown pte marker */
3917 return VM_FAULT_SIGBUS
;
3921 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3922 * but allow concurrent faults), and pte mapped but not yet locked.
3923 * We return with pte unmapped and unlocked.
3925 * We return with the mmap_lock locked or unlocked in the same cases
3926 * as does filemap_fault().
3928 vm_fault_t
do_swap_page(struct vm_fault
*vmf
)
3930 struct vm_area_struct
*vma
= vmf
->vma
;
3931 struct folio
*swapcache
, *folio
= NULL
;
3933 struct swap_info_struct
*si
= NULL
;
3934 rmap_t rmap_flags
= RMAP_NONE
;
3935 bool need_clear_cache
= false;
3936 bool exclusive
= false;
3940 void *shadow
= NULL
;
3942 if (!pte_unmap_same(vmf
))
3945 entry
= pte_to_swp_entry(vmf
->orig_pte
);
3946 if (unlikely(non_swap_entry(entry
))) {
3947 if (is_migration_entry(entry
)) {
3948 migration_entry_wait(vma
->vm_mm
, vmf
->pmd
,
3950 } else if (is_device_exclusive_entry(entry
)) {
3951 vmf
->page
= pfn_swap_entry_to_page(entry
);
3952 ret
= remove_device_exclusive_entry(vmf
);
3953 } else if (is_device_private_entry(entry
)) {
3954 if (vmf
->flags
& FAULT_FLAG_VMA_LOCK
) {
3956 * migrate_to_ram is not yet ready to operate
3960 ret
= VM_FAULT_RETRY
;
3964 vmf
->page
= pfn_swap_entry_to_page(entry
);
3965 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3966 vmf
->address
, &vmf
->ptl
);
3967 if (unlikely(!vmf
->pte
||
3968 !pte_same(ptep_get(vmf
->pte
),
3973 * Get a page reference while we know the page can't be
3976 get_page(vmf
->page
);
3977 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3978 ret
= vmf
->page
->pgmap
->ops
->migrate_to_ram(vmf
);
3979 put_page(vmf
->page
);
3980 } else if (is_hwpoison_entry(entry
)) {
3981 ret
= VM_FAULT_HWPOISON
;
3982 } else if (is_pte_marker_entry(entry
)) {
3983 ret
= handle_pte_marker(vmf
);
3985 print_bad_pte(vma
, vmf
->address
, vmf
->orig_pte
, NULL
);
3986 ret
= VM_FAULT_SIGBUS
;
3991 /* Prevent swapoff from happening to us. */
3992 si
= get_swap_device(entry
);
3996 folio
= swap_cache_get_folio(entry
, vma
, vmf
->address
);
3998 page
= folio_file_page(folio
, swp_offset(entry
));
4002 if (data_race(si
->flags
& SWP_SYNCHRONOUS_IO
) &&
4003 __swap_count(entry
) == 1) {
4005 * Prevent parallel swapin from proceeding with
4006 * the cache flag. Otherwise, another thread may
4007 * finish swapin first, free the entry, and swapout
4008 * reusing the same entry. It's undetectable as
4009 * pte_same() returns true due to entry reuse.
4011 if (swapcache_prepare(entry
)) {
4012 /* Relax a bit to prevent rapid repeated page faults */
4013 schedule_timeout_uninterruptible(1);
4016 need_clear_cache
= true;
4018 /* skip swapcache */
4019 folio
= vma_alloc_folio(GFP_HIGHUSER_MOVABLE
, 0,
4020 vma
, vmf
->address
, false);
4021 page
= &folio
->page
;
4023 __folio_set_locked(folio
);
4024 __folio_set_swapbacked(folio
);
4026 if (mem_cgroup_swapin_charge_folio(folio
,
4027 vma
->vm_mm
, GFP_KERNEL
,
4032 mem_cgroup_swapin_uncharge_swap(entry
);
4034 shadow
= get_shadow_from_swap_cache(entry
);
4036 workingset_refault(folio
, shadow
);
4038 folio_add_lru(folio
);
4040 /* To provide entry to swap_read_folio() */
4041 folio
->swap
= entry
;
4042 swap_read_folio(folio
, true, NULL
);
4043 folio
->private = NULL
;
4046 page
= swapin_readahead(entry
, GFP_HIGHUSER_MOVABLE
,
4049 folio
= page_folio(page
);
4055 * Back out if somebody else faulted in this pte
4056 * while we released the pte lock.
4058 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4059 vmf
->address
, &vmf
->ptl
);
4060 if (likely(vmf
->pte
&&
4061 pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)))
4066 /* Had to read the page from swap area: Major fault */
4067 ret
= VM_FAULT_MAJOR
;
4068 count_vm_event(PGMAJFAULT
);
4069 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
4070 } else if (PageHWPoison(page
)) {
4072 * hwpoisoned dirty swapcache pages are kept for killing
4073 * owner processes (which may be unknown at hwpoison time)
4075 ret
= VM_FAULT_HWPOISON
;
4079 ret
|= folio_lock_or_retry(folio
, vmf
);
4080 if (ret
& VM_FAULT_RETRY
)
4085 * Make sure folio_free_swap() or swapoff did not release the
4086 * swapcache from under us. The page pin, and pte_same test
4087 * below, are not enough to exclude that. Even if it is still
4088 * swapcache, we need to check that the page's swap has not
4091 if (unlikely(!folio_test_swapcache(folio
) ||
4092 page_swap_entry(page
).val
!= entry
.val
))
4096 * KSM sometimes has to copy on read faults, for example, if
4097 * page->index of !PageKSM() pages would be nonlinear inside the
4098 * anon VMA -- PageKSM() is lost on actual swapout.
4100 folio
= ksm_might_need_to_copy(folio
, vma
, vmf
->address
);
4101 if (unlikely(!folio
)) {
4105 } else if (unlikely(folio
== ERR_PTR(-EHWPOISON
))) {
4106 ret
= VM_FAULT_HWPOISON
;
4110 if (folio
!= swapcache
)
4111 page
= folio_page(folio
, 0);
4114 * If we want to map a page that's in the swapcache writable, we
4115 * have to detect via the refcount if we're really the exclusive
4116 * owner. Try removing the extra reference from the local LRU
4117 * caches if required.
4119 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && folio
== swapcache
&&
4120 !folio_test_ksm(folio
) && !folio_test_lru(folio
))
4124 folio_throttle_swaprate(folio
, GFP_KERNEL
);
4127 * Back out if somebody else already faulted in this pte.
4129 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
4131 if (unlikely(!vmf
->pte
|| !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
)))
4134 if (unlikely(!folio_test_uptodate(folio
))) {
4135 ret
= VM_FAULT_SIGBUS
;
4140 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte
4141 * must never point at an anonymous page in the swapcache that is
4142 * PG_anon_exclusive. Sanity check that this holds and especially, that
4143 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity
4144 * check after taking the PT lock and making sure that nobody
4145 * concurrently faulted in this page and set PG_anon_exclusive.
4147 BUG_ON(!folio_test_anon(folio
) && folio_test_mappedtodisk(folio
));
4148 BUG_ON(folio_test_anon(folio
) && PageAnonExclusive(page
));
4151 * Check under PT lock (to protect against concurrent fork() sharing
4152 * the swap entry concurrently) for certainly exclusive pages.
4154 if (!folio_test_ksm(folio
)) {
4155 exclusive
= pte_swp_exclusive(vmf
->orig_pte
);
4156 if (folio
!= swapcache
) {
4158 * We have a fresh page that is not exposed to the
4159 * swapcache -> certainly exclusive.
4162 } else if (exclusive
&& folio_test_writeback(folio
) &&
4163 data_race(si
->flags
& SWP_STABLE_WRITES
)) {
4165 * This is tricky: not all swap backends support
4166 * concurrent page modifications while under writeback.
4168 * So if we stumble over such a page in the swapcache
4169 * we must not set the page exclusive, otherwise we can
4170 * map it writable without further checks and modify it
4171 * while still under writeback.
4173 * For these problematic swap backends, simply drop the
4174 * exclusive marker: this is perfectly fine as we start
4175 * writeback only if we fully unmapped the page and
4176 * there are no unexpected references on the page after
4177 * unmapping succeeded. After fully unmapped, no
4178 * further GUP references (FOLL_GET and FOLL_PIN) can
4179 * appear, so dropping the exclusive marker and mapping
4180 * it only R/O is fine.
4187 * Some architectures may have to restore extra metadata to the page
4188 * when reading from swap. This metadata may be indexed by swap entry
4189 * so this must be called before swap_free().
4191 arch_swap_restore(entry
, folio
);
4194 * Remove the swap entry and conditionally try to free up the swapcache.
4195 * We're already holding a reference on the page but haven't mapped it
4199 if (should_try_to_free_swap(folio
, vma
, vmf
->flags
))
4200 folio_free_swap(folio
);
4202 inc_mm_counter(vma
->vm_mm
, MM_ANONPAGES
);
4203 dec_mm_counter(vma
->vm_mm
, MM_SWAPENTS
);
4204 pte
= mk_pte(page
, vma
->vm_page_prot
);
4207 * Same logic as in do_wp_page(); however, optimize for pages that are
4208 * certainly not shared either because we just allocated them without
4209 * exposing them to the swapcache or because the swap entry indicates
4212 if (!folio_test_ksm(folio
) &&
4213 (exclusive
|| folio_ref_count(folio
) == 1)) {
4214 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4215 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
4216 vmf
->flags
&= ~FAULT_FLAG_WRITE
;
4218 rmap_flags
|= RMAP_EXCLUSIVE
;
4220 flush_icache_page(vma
, page
);
4221 if (pte_swp_soft_dirty(vmf
->orig_pte
))
4222 pte
= pte_mksoft_dirty(pte
);
4223 if (pte_swp_uffd_wp(vmf
->orig_pte
))
4224 pte
= pte_mkuffd_wp(pte
);
4225 vmf
->orig_pte
= pte
;
4227 /* ksm created a completely new copy */
4228 if (unlikely(folio
!= swapcache
&& swapcache
)) {
4229 folio_add_new_anon_rmap(folio
, vma
, vmf
->address
);
4230 folio_add_lru_vma(folio
, vma
);
4232 folio_add_anon_rmap_pte(folio
, page
, vma
, vmf
->address
,
4236 VM_BUG_ON(!folio_test_anon(folio
) ||
4237 (pte_write(pte
) && !PageAnonExclusive(page
)));
4238 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
4239 arch_do_swap_page(vma
->vm_mm
, vma
, vmf
->address
, pte
, vmf
->orig_pte
);
4241 folio_unlock(folio
);
4242 if (folio
!= swapcache
&& swapcache
) {
4244 * Hold the lock to avoid the swap entry to be reused
4245 * until we take the PT lock for the pte_same() check
4246 * (to avoid false positives from pte_same). For
4247 * further safety release the lock after the swap_free
4248 * so that the swap count won't change under a
4249 * parallel locked swapcache.
4251 folio_unlock(swapcache
);
4252 folio_put(swapcache
);
4255 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4256 ret
|= do_wp_page(vmf
);
4257 if (ret
& VM_FAULT_ERROR
)
4258 ret
&= VM_FAULT_ERROR
;
4262 /* No need to invalidate - it was non-present before */
4263 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
4266 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4268 /* Clear the swap cache pin for direct swapin after PTL unlock */
4269 if (need_clear_cache
)
4270 swapcache_clear(si
, entry
);
4272 put_swap_device(si
);
4276 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4278 folio_unlock(folio
);
4281 if (folio
!= swapcache
&& swapcache
) {
4282 folio_unlock(swapcache
);
4283 folio_put(swapcache
);
4285 if (need_clear_cache
)
4286 swapcache_clear(si
, entry
);
4288 put_swap_device(si
);
4292 static bool pte_range_none(pte_t
*pte
, int nr_pages
)
4296 for (i
= 0; i
< nr_pages
; i
++) {
4297 if (!pte_none(ptep_get_lockless(pte
+ i
)))
4304 static struct folio
*alloc_anon_folio(struct vm_fault
*vmf
)
4306 struct vm_area_struct
*vma
= vmf
->vma
;
4307 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4308 unsigned long orders
;
4309 struct folio
*folio
;
4316 * If uffd is active for the vma we need per-page fault fidelity to
4317 * maintain the uffd semantics.
4319 if (unlikely(userfaultfd_armed(vma
)))
4323 * Get a list of all the (large) orders below PMD_ORDER that are enabled
4324 * for this vma. Then filter out the orders that can't be allocated over
4325 * the faulting address and still be fully contained in the vma.
4327 orders
= thp_vma_allowable_orders(vma
, vma
->vm_flags
, false, true, true,
4328 BIT(PMD_ORDER
) - 1);
4329 orders
= thp_vma_suitable_orders(vma
, vmf
->address
, orders
);
4334 pte
= pte_offset_map(vmf
->pmd
, vmf
->address
& PMD_MASK
);
4336 return ERR_PTR(-EAGAIN
);
4339 * Find the highest order where the aligned range is completely
4340 * pte_none(). Note that all remaining orders will be completely
4343 order
= highest_order(orders
);
4345 addr
= ALIGN_DOWN(vmf
->address
, PAGE_SIZE
<< order
);
4346 if (pte_range_none(pte
+ pte_index(addr
), 1 << order
))
4348 order
= next_order(&orders
, order
);
4353 /* Try allocating the highest of the remaining orders. */
4354 gfp
= vma_thp_gfp_mask(vma
);
4356 addr
= ALIGN_DOWN(vmf
->address
, PAGE_SIZE
<< order
);
4357 folio
= vma_alloc_folio(gfp
, order
, vma
, addr
, true);
4359 if (mem_cgroup_charge(folio
, vma
->vm_mm
, gfp
)) {
4363 folio_throttle_swaprate(folio
, gfp
);
4364 clear_huge_page(&folio
->page
, vmf
->address
, 1 << order
);
4368 order
= next_order(&orders
, order
);
4373 return folio_prealloc(vma
->vm_mm
, vma
, vmf
->address
, true);
4377 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4378 * but allow concurrent faults), and pte mapped but not yet locked.
4379 * We return with mmap_lock still held, but pte unmapped and unlocked.
4381 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
)
4383 bool uffd_wp
= vmf_orig_pte_uffd_wp(vmf
);
4384 struct vm_area_struct
*vma
= vmf
->vma
;
4385 unsigned long addr
= vmf
->address
;
4386 struct folio
*folio
;
4392 /* File mapping without ->vm_ops ? */
4393 if (vma
->vm_flags
& VM_SHARED
)
4394 return VM_FAULT_SIGBUS
;
4397 * Use pte_alloc() instead of pte_alloc_map(), so that OOM can
4398 * be distinguished from a transient failure of pte_offset_map().
4400 if (pte_alloc(vma
->vm_mm
, vmf
->pmd
))
4401 return VM_FAULT_OOM
;
4403 /* Use the zero-page for reads */
4404 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
4405 !mm_forbids_zeropage(vma
->vm_mm
)) {
4406 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(vmf
->address
),
4407 vma
->vm_page_prot
));
4408 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4409 vmf
->address
, &vmf
->ptl
);
4412 if (vmf_pte_changed(vmf
)) {
4413 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
4416 ret
= check_stable_address_space(vma
->vm_mm
);
4419 /* Deliver the page fault to userland, check inside PT lock */
4420 if (userfaultfd_missing(vma
)) {
4421 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4422 return handle_userfault(vmf
, VM_UFFD_MISSING
);
4427 /* Allocate our own private page. */
4428 if (unlikely(anon_vma_prepare(vma
)))
4430 /* Returns NULL on OOM or ERR_PTR(-EAGAIN) if we must retry the fault */
4431 folio
= alloc_anon_folio(vmf
);
4437 nr_pages
= folio_nr_pages(folio
);
4438 addr
= ALIGN_DOWN(vmf
->address
, nr_pages
* PAGE_SIZE
);
4441 * The memory barrier inside __folio_mark_uptodate makes sure that
4442 * preceding stores to the page contents become visible before
4443 * the set_pte_at() write.
4445 __folio_mark_uptodate(folio
);
4447 entry
= mk_pte(&folio
->page
, vma
->vm_page_prot
);
4448 entry
= pte_sw_mkyoung(entry
);
4449 if (vma
->vm_flags
& VM_WRITE
)
4450 entry
= pte_mkwrite(pte_mkdirty(entry
), vma
);
4452 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
4455 if (nr_pages
== 1 && vmf_pte_changed(vmf
)) {
4456 update_mmu_tlb(vma
, addr
, vmf
->pte
);
4458 } else if (nr_pages
> 1 && !pte_range_none(vmf
->pte
, nr_pages
)) {
4459 for (i
= 0; i
< nr_pages
; i
++)
4460 update_mmu_tlb(vma
, addr
+ PAGE_SIZE
* i
, vmf
->pte
+ i
);
4464 ret
= check_stable_address_space(vma
->vm_mm
);
4468 /* Deliver the page fault to userland, check inside PT lock */
4469 if (userfaultfd_missing(vma
)) {
4470 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4472 return handle_userfault(vmf
, VM_UFFD_MISSING
);
4475 folio_ref_add(folio
, nr_pages
- 1);
4476 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, nr_pages
);
4477 folio_add_new_anon_rmap(folio
, vma
, addr
);
4478 folio_add_lru_vma(folio
, vma
);
4481 entry
= pte_mkuffd_wp(entry
);
4482 set_ptes(vma
->vm_mm
, addr
, vmf
->pte
, entry
, nr_pages
);
4484 /* No need to invalidate - it was non-present before */
4485 update_mmu_cache_range(vmf
, vma
, addr
, vmf
->pte
, nr_pages
);
4488 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4494 return VM_FAULT_OOM
;
4498 * The mmap_lock must have been held on entry, and may have been
4499 * released depending on flags and vma->vm_ops->fault() return value.
4500 * See filemap_fault() and __lock_page_retry().
4502 static vm_fault_t
__do_fault(struct vm_fault
*vmf
)
4504 struct vm_area_struct
*vma
= vmf
->vma
;
4505 struct folio
*folio
;
4509 * Preallocate pte before we take page_lock because this might lead to
4510 * deadlocks for memcg reclaim which waits for pages under writeback:
4512 * SetPageWriteback(A)
4518 * wait_on_page_writeback(A)
4519 * SetPageWriteback(B)
4521 * # flush A, B to clear the writeback
4523 if (pmd_none(*vmf
->pmd
) && !vmf
->prealloc_pte
) {
4524 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
4525 if (!vmf
->prealloc_pte
)
4526 return VM_FAULT_OOM
;
4529 ret
= vma
->vm_ops
->fault(vmf
);
4530 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
|
4531 VM_FAULT_DONE_COW
)))
4534 folio
= page_folio(vmf
->page
);
4535 if (unlikely(PageHWPoison(vmf
->page
))) {
4536 vm_fault_t poisonret
= VM_FAULT_HWPOISON
;
4537 if (ret
& VM_FAULT_LOCKED
) {
4538 if (page_mapped(vmf
->page
))
4539 unmap_mapping_folio(folio
);
4540 /* Retry if a clean folio was removed from the cache. */
4541 if (mapping_evict_folio(folio
->mapping
, folio
))
4542 poisonret
= VM_FAULT_NOPAGE
;
4543 folio_unlock(folio
);
4550 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
4553 VM_BUG_ON_PAGE(!folio_test_locked(folio
), vmf
->page
);
4558 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4559 static void deposit_prealloc_pte(struct vm_fault
*vmf
)
4561 struct vm_area_struct
*vma
= vmf
->vma
;
4563 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
4565 * We are going to consume the prealloc table,
4566 * count that as nr_ptes.
4568 mm_inc_nr_ptes(vma
->vm_mm
);
4569 vmf
->prealloc_pte
= NULL
;
4572 vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
4574 struct folio
*folio
= page_folio(page
);
4575 struct vm_area_struct
*vma
= vmf
->vma
;
4576 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
4577 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
4579 vm_fault_t ret
= VM_FAULT_FALLBACK
;
4581 if (!thp_vma_suitable_order(vma
, haddr
, PMD_ORDER
))
4584 if (page
!= &folio
->page
|| folio_order(folio
) != HPAGE_PMD_ORDER
)
4588 * Just backoff if any subpage of a THP is corrupted otherwise
4589 * the corrupted page may mapped by PMD silently to escape the
4590 * check. This kind of THP just can be PTE mapped. Access to
4591 * the corrupted subpage should trigger SIGBUS as expected.
4593 if (unlikely(folio_test_has_hwpoisoned(folio
)))
4597 * Archs like ppc64 need additional space to store information
4598 * related to pte entry. Use the preallocated table for that.
4600 if (arch_needs_pgtable_deposit() && !vmf
->prealloc_pte
) {
4601 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
4602 if (!vmf
->prealloc_pte
)
4603 return VM_FAULT_OOM
;
4606 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
4607 if (unlikely(!pmd_none(*vmf
->pmd
)))
4610 flush_icache_pages(vma
, page
, HPAGE_PMD_NR
);
4612 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
4614 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
4616 add_mm_counter(vma
->vm_mm
, mm_counter_file(folio
), HPAGE_PMD_NR
);
4617 folio_add_file_rmap_pmd(folio
, page
, vma
);
4620 * deposit and withdraw with pmd lock held
4622 if (arch_needs_pgtable_deposit())
4623 deposit_prealloc_pte(vmf
);
4625 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
4627 update_mmu_cache_pmd(vma
, haddr
, vmf
->pmd
);
4629 /* fault is handled */
4631 count_vm_event(THP_FILE_MAPPED
);
4633 spin_unlock(vmf
->ptl
);
4637 vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
4639 return VM_FAULT_FALLBACK
;
4644 * set_pte_range - Set a range of PTEs to point to pages in a folio.
4645 * @vmf: Fault decription.
4646 * @folio: The folio that contains @page.
4647 * @page: The first page to create a PTE for.
4648 * @nr: The number of PTEs to create.
4649 * @addr: The first address to create a PTE for.
4651 void set_pte_range(struct vm_fault
*vmf
, struct folio
*folio
,
4652 struct page
*page
, unsigned int nr
, unsigned long addr
)
4654 struct vm_area_struct
*vma
= vmf
->vma
;
4655 bool uffd_wp
= vmf_orig_pte_uffd_wp(vmf
);
4656 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
4657 bool prefault
= in_range(vmf
->address
, addr
, nr
* PAGE_SIZE
);
4660 flush_icache_pages(vma
, page
, nr
);
4661 entry
= mk_pte(page
, vma
->vm_page_prot
);
4663 if (prefault
&& arch_wants_old_prefaulted_pte())
4664 entry
= pte_mkold(entry
);
4666 entry
= pte_sw_mkyoung(entry
);
4669 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
4670 if (unlikely(uffd_wp
))
4671 entry
= pte_mkuffd_wp(entry
);
4672 /* copy-on-write page */
4673 if (write
&& !(vma
->vm_flags
& VM_SHARED
)) {
4674 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, nr
);
4675 VM_BUG_ON_FOLIO(nr
!= 1, folio
);
4676 folio_add_new_anon_rmap(folio
, vma
, addr
);
4677 folio_add_lru_vma(folio
, vma
);
4679 add_mm_counter(vma
->vm_mm
, mm_counter_file(folio
), nr
);
4680 folio_add_file_rmap_ptes(folio
, page
, nr
, vma
);
4682 set_ptes(vma
->vm_mm
, addr
, vmf
->pte
, entry
, nr
);
4684 /* no need to invalidate: a not-present page won't be cached */
4685 update_mmu_cache_range(vmf
, vma
, addr
, vmf
->pte
, nr
);
4688 static bool vmf_pte_changed(struct vm_fault
*vmf
)
4690 if (vmf
->flags
& FAULT_FLAG_ORIG_PTE_VALID
)
4691 return !pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
);
4693 return !pte_none(ptep_get(vmf
->pte
));
4697 * finish_fault - finish page fault once we have prepared the page to fault
4699 * @vmf: structure describing the fault
4701 * This function handles all that is needed to finish a page fault once the
4702 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4703 * given page, adds reverse page mapping, handles memcg charges and LRU
4706 * The function expects the page to be locked and on success it consumes a
4707 * reference of a page being mapped (for the PTE which maps it).
4709 * Return: %0 on success, %VM_FAULT_ code in case of error.
4711 vm_fault_t
finish_fault(struct vm_fault
*vmf
)
4713 struct vm_area_struct
*vma
= vmf
->vma
;
4717 /* Did we COW the page? */
4718 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
4719 page
= vmf
->cow_page
;
4724 * check even for read faults because we might have lost our CoWed
4727 if (!(vma
->vm_flags
& VM_SHARED
)) {
4728 ret
= check_stable_address_space(vma
->vm_mm
);
4733 if (pmd_none(*vmf
->pmd
)) {
4734 if (PageTransCompound(page
)) {
4735 ret
= do_set_pmd(vmf
, page
);
4736 if (ret
!= VM_FAULT_FALLBACK
)
4740 if (vmf
->prealloc_pte
)
4741 pmd_install(vma
->vm_mm
, vmf
->pmd
, &vmf
->prealloc_pte
);
4742 else if (unlikely(pte_alloc(vma
->vm_mm
, vmf
->pmd
)))
4743 return VM_FAULT_OOM
;
4746 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4747 vmf
->address
, &vmf
->ptl
);
4749 return VM_FAULT_NOPAGE
;
4751 /* Re-check under ptl */
4752 if (likely(!vmf_pte_changed(vmf
))) {
4753 struct folio
*folio
= page_folio(page
);
4755 set_pte_range(vmf
, folio
, page
, 1, vmf
->address
);
4758 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
4759 ret
= VM_FAULT_NOPAGE
;
4762 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4766 static unsigned long fault_around_pages __read_mostly
=
4767 65536 >> PAGE_SHIFT
;
4769 #ifdef CONFIG_DEBUG_FS
4770 static int fault_around_bytes_get(void *data
, u64
*val
)
4772 *val
= fault_around_pages
<< PAGE_SHIFT
;
4777 * fault_around_bytes must be rounded down to the nearest page order as it's
4778 * what do_fault_around() expects to see.
4780 static int fault_around_bytes_set(void *data
, u64 val
)
4782 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
4786 * The minimum value is 1 page, however this results in no fault-around
4787 * at all. See should_fault_around().
4789 val
= max(val
, PAGE_SIZE
);
4790 fault_around_pages
= rounddown_pow_of_two(val
) >> PAGE_SHIFT
;
4794 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops
,
4795 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
4797 static int __init
fault_around_debugfs(void)
4799 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL
, NULL
,
4800 &fault_around_bytes_fops
);
4803 late_initcall(fault_around_debugfs
);
4807 * do_fault_around() tries to map few pages around the fault address. The hope
4808 * is that the pages will be needed soon and this will lower the number of
4811 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4812 * not ready to be mapped: not up-to-date, locked, etc.
4814 * This function doesn't cross VMA or page table boundaries, in order to call
4815 * map_pages() and acquire a PTE lock only once.
4817 * fault_around_pages defines how many pages we'll try to map.
4818 * do_fault_around() expects it to be set to a power of two less than or equal
4821 * The virtual address of the area that we map is naturally aligned to
4822 * fault_around_pages * PAGE_SIZE rounded down to the machine page size
4823 * (and therefore to page order). This way it's easier to guarantee
4824 * that we don't cross page table boundaries.
4826 static vm_fault_t
do_fault_around(struct vm_fault
*vmf
)
4828 pgoff_t nr_pages
= READ_ONCE(fault_around_pages
);
4829 pgoff_t pte_off
= pte_index(vmf
->address
);
4830 /* The page offset of vmf->address within the VMA. */
4831 pgoff_t vma_off
= vmf
->pgoff
- vmf
->vma
->vm_pgoff
;
4832 pgoff_t from_pte
, to_pte
;
4835 /* The PTE offset of the start address, clamped to the VMA. */
4836 from_pte
= max(ALIGN_DOWN(pte_off
, nr_pages
),
4837 pte_off
- min(pte_off
, vma_off
));
4839 /* The PTE offset of the end address, clamped to the VMA and PTE. */
4840 to_pte
= min3(from_pte
+ nr_pages
, (pgoff_t
)PTRS_PER_PTE
,
4841 pte_off
+ vma_pages(vmf
->vma
) - vma_off
) - 1;
4843 if (pmd_none(*vmf
->pmd
)) {
4844 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
);
4845 if (!vmf
->prealloc_pte
)
4846 return VM_FAULT_OOM
;
4850 ret
= vmf
->vma
->vm_ops
->map_pages(vmf
,
4851 vmf
->pgoff
+ from_pte
- pte_off
,
4852 vmf
->pgoff
+ to_pte
- pte_off
);
4858 /* Return true if we should do read fault-around, false otherwise */
4859 static inline bool should_fault_around(struct vm_fault
*vmf
)
4861 /* No ->map_pages? No way to fault around... */
4862 if (!vmf
->vma
->vm_ops
->map_pages
)
4865 if (uffd_disable_fault_around(vmf
->vma
))
4868 /* A single page implies no faulting 'around' at all. */
4869 return fault_around_pages
> 1;
4872 static vm_fault_t
do_read_fault(struct vm_fault
*vmf
)
4875 struct folio
*folio
;
4878 * Let's call ->map_pages() first and use ->fault() as fallback
4879 * if page by the offset is not ready to be mapped (cold cache or
4882 if (should_fault_around(vmf
)) {
4883 ret
= do_fault_around(vmf
);
4888 ret
= vmf_can_call_fault(vmf
);
4892 ret
= __do_fault(vmf
);
4893 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4896 ret
|= finish_fault(vmf
);
4897 folio
= page_folio(vmf
->page
);
4898 folio_unlock(folio
);
4899 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4904 static vm_fault_t
do_cow_fault(struct vm_fault
*vmf
)
4906 struct vm_area_struct
*vma
= vmf
->vma
;
4907 struct folio
*folio
;
4910 ret
= vmf_can_call_fault(vmf
);
4912 ret
= vmf_anon_prepare(vmf
);
4916 folio
= folio_prealloc(vma
->vm_mm
, vma
, vmf
->address
, false);
4918 return VM_FAULT_OOM
;
4920 vmf
->cow_page
= &folio
->page
;
4922 ret
= __do_fault(vmf
);
4923 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4925 if (ret
& VM_FAULT_DONE_COW
)
4928 copy_user_highpage(vmf
->cow_page
, vmf
->page
, vmf
->address
, vma
);
4929 __folio_mark_uptodate(folio
);
4931 ret
|= finish_fault(vmf
);
4932 unlock_page(vmf
->page
);
4933 put_page(vmf
->page
);
4934 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4942 static vm_fault_t
do_shared_fault(struct vm_fault
*vmf
)
4944 struct vm_area_struct
*vma
= vmf
->vma
;
4945 vm_fault_t ret
, tmp
;
4946 struct folio
*folio
;
4948 ret
= vmf_can_call_fault(vmf
);
4952 ret
= __do_fault(vmf
);
4953 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4956 folio
= page_folio(vmf
->page
);
4959 * Check if the backing address space wants to know that the page is
4960 * about to become writable
4962 if (vma
->vm_ops
->page_mkwrite
) {
4963 folio_unlock(folio
);
4964 tmp
= do_page_mkwrite(vmf
, folio
);
4965 if (unlikely(!tmp
||
4966 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
4972 ret
|= finish_fault(vmf
);
4973 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
4975 folio_unlock(folio
);
4980 ret
|= fault_dirty_shared_page(vmf
);
4985 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4986 * but allow concurrent faults).
4987 * The mmap_lock may have been released depending on flags and our
4988 * return value. See filemap_fault() and __folio_lock_or_retry().
4989 * If mmap_lock is released, vma may become invalid (for example
4990 * by other thread calling munmap()).
4992 static vm_fault_t
do_fault(struct vm_fault
*vmf
)
4994 struct vm_area_struct
*vma
= vmf
->vma
;
4995 struct mm_struct
*vm_mm
= vma
->vm_mm
;
4999 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
5001 if (!vma
->vm_ops
->fault
) {
5002 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
,
5003 vmf
->address
, &vmf
->ptl
);
5004 if (unlikely(!vmf
->pte
))
5005 ret
= VM_FAULT_SIGBUS
;
5008 * Make sure this is not a temporary clearing of pte
5009 * by holding ptl and checking again. A R/M/W update
5010 * of pte involves: take ptl, clearing the pte so that
5011 * we don't have concurrent modification by hardware
5012 * followed by an update.
5014 if (unlikely(pte_none(ptep_get(vmf
->pte
))))
5015 ret
= VM_FAULT_SIGBUS
;
5017 ret
= VM_FAULT_NOPAGE
;
5019 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5021 } else if (!(vmf
->flags
& FAULT_FLAG_WRITE
))
5022 ret
= do_read_fault(vmf
);
5023 else if (!(vma
->vm_flags
& VM_SHARED
))
5024 ret
= do_cow_fault(vmf
);
5026 ret
= do_shared_fault(vmf
);
5028 /* preallocated pagetable is unused: free it */
5029 if (vmf
->prealloc_pte
) {
5030 pte_free(vm_mm
, vmf
->prealloc_pte
);
5031 vmf
->prealloc_pte
= NULL
;
5036 int numa_migrate_prep(struct folio
*folio
, struct vm_area_struct
*vma
,
5037 unsigned long addr
, int page_nid
, int *flags
)
5041 /* Record the current PID acceesing VMA */
5042 vma_set_access_pid_bit(vma
);
5044 count_vm_numa_event(NUMA_HINT_FAULTS
);
5045 if (page_nid
== numa_node_id()) {
5046 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
5047 *flags
|= TNF_FAULT_LOCAL
;
5050 return mpol_misplaced(folio
, vma
, addr
);
5053 static vm_fault_t
do_numa_page(struct vm_fault
*vmf
)
5055 struct vm_area_struct
*vma
= vmf
->vma
;
5056 struct folio
*folio
= NULL
;
5057 int nid
= NUMA_NO_NODE
;
5058 bool writable
= false;
5065 * The pte cannot be used safely until we verify, while holding the page
5066 * table lock, that its contents have not changed during fault handling.
5068 spin_lock(vmf
->ptl
);
5069 /* Read the live PTE from the page tables: */
5070 old_pte
= ptep_get(vmf
->pte
);
5072 if (unlikely(!pte_same(old_pte
, vmf
->orig_pte
))) {
5073 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5077 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
5080 * Detect now whether the PTE could be writable; this information
5081 * is only valid while holding the PT lock.
5083 writable
= pte_write(pte
);
5084 if (!writable
&& vma_wants_manual_pte_write_upgrade(vma
) &&
5085 can_change_pte_writable(vma
, vmf
->address
, pte
))
5088 folio
= vm_normal_folio(vma
, vmf
->address
, pte
);
5089 if (!folio
|| folio_is_zone_device(folio
))
5092 /* TODO: handle PTE-mapped THP */
5093 if (folio_test_large(folio
))
5097 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
5098 * much anyway since they can be in shared cache state. This misses
5099 * the case where a mapping is writable but the process never writes
5100 * to it but pte_write gets cleared during protection updates and
5101 * pte_dirty has unpredictable behaviour between PTE scan updates,
5102 * background writeback, dirty balancing and application behaviour.
5105 flags
|= TNF_NO_GROUP
;
5108 * Flag if the folio is shared between multiple address spaces. This
5109 * is later used when determining whether to group tasks together
5111 if (folio_estimated_sharers(folio
) > 1 && (vma
->vm_flags
& VM_SHARED
))
5112 flags
|= TNF_SHARED
;
5114 nid
= folio_nid(folio
);
5116 * For memory tiering mode, cpupid of slow memory page is used
5117 * to record page access time. So use default value.
5119 if ((sysctl_numa_balancing_mode
& NUMA_BALANCING_MEMORY_TIERING
) &&
5120 !node_is_toptier(nid
))
5121 last_cpupid
= (-1 & LAST_CPUPID_MASK
);
5123 last_cpupid
= folio_last_cpupid(folio
);
5124 target_nid
= numa_migrate_prep(folio
, vma
, vmf
->address
, nid
, &flags
);
5125 if (target_nid
== NUMA_NO_NODE
) {
5129 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5132 /* Migrate to the requested node */
5133 if (migrate_misplaced_folio(folio
, vma
, target_nid
)) {
5135 flags
|= TNF_MIGRATED
;
5137 flags
|= TNF_MIGRATE_FAIL
;
5138 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
5139 vmf
->address
, &vmf
->ptl
);
5140 if (unlikely(!vmf
->pte
))
5142 if (unlikely(!pte_same(ptep_get(vmf
->pte
), vmf
->orig_pte
))) {
5143 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5150 if (nid
!= NUMA_NO_NODE
)
5151 task_numa_fault(last_cpupid
, nid
, 1, flags
);
5155 * Make it present again, depending on how arch implements
5156 * non-accessible ptes, some can allow access by kernel mode.
5158 old_pte
= ptep_modify_prot_start(vma
, vmf
->address
, vmf
->pte
);
5159 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
5160 pte
= pte_mkyoung(pte
);
5162 pte
= pte_mkwrite(pte
, vma
);
5163 ptep_modify_prot_commit(vma
, vmf
->address
, vmf
->pte
, old_pte
, pte
);
5164 update_mmu_cache_range(vmf
, vma
, vmf
->address
, vmf
->pte
, 1);
5165 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5169 static inline vm_fault_t
create_huge_pmd(struct vm_fault
*vmf
)
5171 struct vm_area_struct
*vma
= vmf
->vma
;
5172 if (vma_is_anonymous(vma
))
5173 return do_huge_pmd_anonymous_page(vmf
);
5174 if (vma
->vm_ops
->huge_fault
)
5175 return vma
->vm_ops
->huge_fault(vmf
, PMD_ORDER
);
5176 return VM_FAULT_FALLBACK
;
5179 /* `inline' is required to avoid gcc 4.1.2 build error */
5180 static inline vm_fault_t
wp_huge_pmd(struct vm_fault
*vmf
)
5182 struct vm_area_struct
*vma
= vmf
->vma
;
5183 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
5186 if (vma_is_anonymous(vma
)) {
5187 if (likely(!unshare
) &&
5188 userfaultfd_huge_pmd_wp(vma
, vmf
->orig_pmd
)) {
5189 if (userfaultfd_wp_async(vmf
->vma
))
5191 return handle_userfault(vmf
, VM_UFFD_WP
);
5193 return do_huge_pmd_wp_page(vmf
);
5196 if (vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
)) {
5197 if (vma
->vm_ops
->huge_fault
) {
5198 ret
= vma
->vm_ops
->huge_fault(vmf
, PMD_ORDER
);
5199 if (!(ret
& VM_FAULT_FALLBACK
))
5205 /* COW or write-notify handled on pte level: split pmd. */
5206 __split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
5208 return VM_FAULT_FALLBACK
;
5211 static vm_fault_t
create_huge_pud(struct vm_fault
*vmf
)
5213 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
5214 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5215 struct vm_area_struct
*vma
= vmf
->vma
;
5216 /* No support for anonymous transparent PUD pages yet */
5217 if (vma_is_anonymous(vma
))
5218 return VM_FAULT_FALLBACK
;
5219 if (vma
->vm_ops
->huge_fault
)
5220 return vma
->vm_ops
->huge_fault(vmf
, PUD_ORDER
);
5221 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
5222 return VM_FAULT_FALLBACK
;
5225 static vm_fault_t
wp_huge_pud(struct vm_fault
*vmf
, pud_t orig_pud
)
5227 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
5228 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5229 struct vm_area_struct
*vma
= vmf
->vma
;
5232 /* No support for anonymous transparent PUD pages yet */
5233 if (vma_is_anonymous(vma
))
5235 if (vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
)) {
5236 if (vma
->vm_ops
->huge_fault
) {
5237 ret
= vma
->vm_ops
->huge_fault(vmf
, PUD_ORDER
);
5238 if (!(ret
& VM_FAULT_FALLBACK
))
5243 /* COW or write-notify not handled on PUD level: split pud.*/
5244 __split_huge_pud(vma
, vmf
->pud
, vmf
->address
);
5245 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
5246 return VM_FAULT_FALLBACK
;
5250 * These routines also need to handle stuff like marking pages dirty
5251 * and/or accessed for architectures that don't do it in hardware (most
5252 * RISC architectures). The early dirtying is also good on the i386.
5254 * There is also a hook called "update_mmu_cache()" that architectures
5255 * with external mmu caches can use to update those (ie the Sparc or
5256 * PowerPC hashed page tables that act as extended TLBs).
5258 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
5259 * concurrent faults).
5261 * The mmap_lock may have been released depending on flags and our return value.
5262 * See filemap_fault() and __folio_lock_or_retry().
5264 static vm_fault_t
handle_pte_fault(struct vm_fault
*vmf
)
5268 if (unlikely(pmd_none(*vmf
->pmd
))) {
5270 * Leave __pte_alloc() until later: because vm_ops->fault may
5271 * want to allocate huge page, and if we expose page table
5272 * for an instant, it will be difficult to retract from
5273 * concurrent faults and from rmap lookups.
5276 vmf
->flags
&= ~FAULT_FLAG_ORIG_PTE_VALID
;
5279 * A regular pmd is established and it can't morph into a huge
5280 * pmd by anon khugepaged, since that takes mmap_lock in write
5281 * mode; but shmem or file collapse to THP could still morph
5282 * it into a huge pmd: just retry later if so.
5284 vmf
->pte
= pte_offset_map_nolock(vmf
->vma
->vm_mm
, vmf
->pmd
,
5285 vmf
->address
, &vmf
->ptl
);
5286 if (unlikely(!vmf
->pte
))
5288 vmf
->orig_pte
= ptep_get_lockless(vmf
->pte
);
5289 vmf
->flags
|= FAULT_FLAG_ORIG_PTE_VALID
;
5291 if (pte_none(vmf
->orig_pte
)) {
5292 pte_unmap(vmf
->pte
);
5298 return do_pte_missing(vmf
);
5300 if (!pte_present(vmf
->orig_pte
))
5301 return do_swap_page(vmf
);
5303 if (pte_protnone(vmf
->orig_pte
) && vma_is_accessible(vmf
->vma
))
5304 return do_numa_page(vmf
);
5306 spin_lock(vmf
->ptl
);
5307 entry
= vmf
->orig_pte
;
5308 if (unlikely(!pte_same(ptep_get(vmf
->pte
), entry
))) {
5309 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
5312 if (vmf
->flags
& (FAULT_FLAG_WRITE
|FAULT_FLAG_UNSHARE
)) {
5313 if (!pte_write(entry
))
5314 return do_wp_page(vmf
);
5315 else if (likely(vmf
->flags
& FAULT_FLAG_WRITE
))
5316 entry
= pte_mkdirty(entry
);
5318 entry
= pte_mkyoung(entry
);
5319 if (ptep_set_access_flags(vmf
->vma
, vmf
->address
, vmf
->pte
, entry
,
5320 vmf
->flags
& FAULT_FLAG_WRITE
)) {
5321 update_mmu_cache_range(vmf
, vmf
->vma
, vmf
->address
,
5324 /* Skip spurious TLB flush for retried page fault */
5325 if (vmf
->flags
& FAULT_FLAG_TRIED
)
5328 * This is needed only for protection faults but the arch code
5329 * is not yet telling us if this is a protection fault or not.
5330 * This still avoids useless tlb flushes for .text page faults
5333 if (vmf
->flags
& FAULT_FLAG_WRITE
)
5334 flush_tlb_fix_spurious_fault(vmf
->vma
, vmf
->address
,
5338 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
5343 * On entry, we hold either the VMA lock or the mmap_lock
5344 * (FAULT_FLAG_VMA_LOCK tells you which). If VM_FAULT_RETRY is set in
5345 * the result, the mmap_lock is not held on exit. See filemap_fault()
5346 * and __folio_lock_or_retry().
5348 static vm_fault_t
__handle_mm_fault(struct vm_area_struct
*vma
,
5349 unsigned long address
, unsigned int flags
)
5351 struct vm_fault vmf
= {
5353 .address
= address
& PAGE_MASK
,
5354 .real_address
= address
,
5356 .pgoff
= linear_page_index(vma
, address
),
5357 .gfp_mask
= __get_fault_gfp_mask(vma
),
5359 struct mm_struct
*mm
= vma
->vm_mm
;
5360 unsigned long vm_flags
= vma
->vm_flags
;
5365 pgd
= pgd_offset(mm
, address
);
5366 p4d
= p4d_alloc(mm
, pgd
, address
);
5368 return VM_FAULT_OOM
;
5370 vmf
.pud
= pud_alloc(mm
, p4d
, address
);
5372 return VM_FAULT_OOM
;
5374 if (pud_none(*vmf
.pud
) &&
5375 thp_vma_allowable_order(vma
, vm_flags
, false, true, true, PUD_ORDER
)) {
5376 ret
= create_huge_pud(&vmf
);
5377 if (!(ret
& VM_FAULT_FALLBACK
))
5380 pud_t orig_pud
= *vmf
.pud
;
5383 if (pud_trans_huge(orig_pud
) || pud_devmap(orig_pud
)) {
5386 * TODO once we support anonymous PUDs: NUMA case and
5387 * FAULT_FLAG_UNSHARE handling.
5389 if ((flags
& FAULT_FLAG_WRITE
) && !pud_write(orig_pud
)) {
5390 ret
= wp_huge_pud(&vmf
, orig_pud
);
5391 if (!(ret
& VM_FAULT_FALLBACK
))
5394 huge_pud_set_accessed(&vmf
, orig_pud
);
5400 vmf
.pmd
= pmd_alloc(mm
, vmf
.pud
, address
);
5402 return VM_FAULT_OOM
;
5404 /* Huge pud page fault raced with pmd_alloc? */
5405 if (pud_trans_unstable(vmf
.pud
))
5408 if (pmd_none(*vmf
.pmd
) &&
5409 thp_vma_allowable_order(vma
, vm_flags
, false, true, true, PMD_ORDER
)) {
5410 ret
= create_huge_pmd(&vmf
);
5411 if (!(ret
& VM_FAULT_FALLBACK
))
5414 vmf
.orig_pmd
= pmdp_get_lockless(vmf
.pmd
);
5416 if (unlikely(is_swap_pmd(vmf
.orig_pmd
))) {
5417 VM_BUG_ON(thp_migration_supported() &&
5418 !is_pmd_migration_entry(vmf
.orig_pmd
));
5419 if (is_pmd_migration_entry(vmf
.orig_pmd
))
5420 pmd_migration_entry_wait(mm
, vmf
.pmd
);
5423 if (pmd_trans_huge(vmf
.orig_pmd
) || pmd_devmap(vmf
.orig_pmd
)) {
5424 if (pmd_protnone(vmf
.orig_pmd
) && vma_is_accessible(vma
))
5425 return do_huge_pmd_numa_page(&vmf
);
5427 if ((flags
& (FAULT_FLAG_WRITE
|FAULT_FLAG_UNSHARE
)) &&
5428 !pmd_write(vmf
.orig_pmd
)) {
5429 ret
= wp_huge_pmd(&vmf
);
5430 if (!(ret
& VM_FAULT_FALLBACK
))
5433 huge_pmd_set_accessed(&vmf
);
5439 return handle_pte_fault(&vmf
);
5443 * mm_account_fault - Do page fault accounting
5444 * @mm: mm from which memcg should be extracted. It can be NULL.
5445 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
5446 * of perf event counters, but we'll still do the per-task accounting to
5447 * the task who triggered this page fault.
5448 * @address: the faulted address.
5449 * @flags: the fault flags.
5450 * @ret: the fault retcode.
5452 * This will take care of most of the page fault accounting. Meanwhile, it
5453 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
5454 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
5455 * still be in per-arch page fault handlers at the entry of page fault.
5457 static inline void mm_account_fault(struct mm_struct
*mm
, struct pt_regs
*regs
,
5458 unsigned long address
, unsigned int flags
,
5463 /* Incomplete faults will be accounted upon completion. */
5464 if (ret
& VM_FAULT_RETRY
)
5468 * To preserve the behavior of older kernels, PGFAULT counters record
5469 * both successful and failed faults, as opposed to perf counters,
5470 * which ignore failed cases.
5472 count_vm_event(PGFAULT
);
5473 count_memcg_event_mm(mm
, PGFAULT
);
5476 * Do not account for unsuccessful faults (e.g. when the address wasn't
5477 * valid). That includes arch_vma_access_permitted() failing before
5478 * reaching here. So this is not a "this many hardware page faults"
5479 * counter. We should use the hw profiling for that.
5481 if (ret
& VM_FAULT_ERROR
)
5485 * We define the fault as a major fault when the final successful fault
5486 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
5487 * handle it immediately previously).
5489 major
= (ret
& VM_FAULT_MAJOR
) || (flags
& FAULT_FLAG_TRIED
);
5497 * If the fault is done for GUP, regs will be NULL. We only do the
5498 * accounting for the per thread fault counters who triggered the
5499 * fault, and we skip the perf event updates.
5505 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ
, 1, regs
, address
);
5507 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN
, 1, regs
, address
);
5510 #ifdef CONFIG_LRU_GEN
5511 static void lru_gen_enter_fault(struct vm_area_struct
*vma
)
5513 /* the LRU algorithm only applies to accesses with recency */
5514 current
->in_lru_fault
= vma_has_recency(vma
);
5517 static void lru_gen_exit_fault(void)
5519 current
->in_lru_fault
= false;
5522 static void lru_gen_enter_fault(struct vm_area_struct
*vma
)
5526 static void lru_gen_exit_fault(void)
5529 #endif /* CONFIG_LRU_GEN */
5531 static vm_fault_t
sanitize_fault_flags(struct vm_area_struct
*vma
,
5532 unsigned int *flags
)
5534 if (unlikely(*flags
& FAULT_FLAG_UNSHARE
)) {
5535 if (WARN_ON_ONCE(*flags
& FAULT_FLAG_WRITE
))
5536 return VM_FAULT_SIGSEGV
;
5538 * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's
5539 * just treat it like an ordinary read-fault otherwise.
5541 if (!is_cow_mapping(vma
->vm_flags
))
5542 *flags
&= ~FAULT_FLAG_UNSHARE
;
5543 } else if (*flags
& FAULT_FLAG_WRITE
) {
5544 /* Write faults on read-only mappings are impossible ... */
5545 if (WARN_ON_ONCE(!(vma
->vm_flags
& VM_MAYWRITE
)))
5546 return VM_FAULT_SIGSEGV
;
5547 /* ... and FOLL_FORCE only applies to COW mappings. */
5548 if (WARN_ON_ONCE(!(vma
->vm_flags
& VM_WRITE
) &&
5549 !is_cow_mapping(vma
->vm_flags
)))
5550 return VM_FAULT_SIGSEGV
;
5552 #ifdef CONFIG_PER_VMA_LOCK
5554 * Per-VMA locks can't be used with FAULT_FLAG_RETRY_NOWAIT because of
5555 * the assumption that lock is dropped on VM_FAULT_RETRY.
5557 if (WARN_ON_ONCE((*flags
&
5558 (FAULT_FLAG_VMA_LOCK
| FAULT_FLAG_RETRY_NOWAIT
)) ==
5559 (FAULT_FLAG_VMA_LOCK
| FAULT_FLAG_RETRY_NOWAIT
)))
5560 return VM_FAULT_SIGSEGV
;
5567 * By the time we get here, we already hold the mm semaphore
5569 * The mmap_lock may have been released depending on flags and our
5570 * return value. See filemap_fault() and __folio_lock_or_retry().
5572 vm_fault_t
handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
5573 unsigned int flags
, struct pt_regs
*regs
)
5575 /* If the fault handler drops the mmap_lock, vma may be freed */
5576 struct mm_struct
*mm
= vma
->vm_mm
;
5579 __set_current_state(TASK_RUNNING
);
5581 ret
= sanitize_fault_flags(vma
, &flags
);
5585 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
5586 flags
& FAULT_FLAG_INSTRUCTION
,
5587 flags
& FAULT_FLAG_REMOTE
)) {
5588 ret
= VM_FAULT_SIGSEGV
;
5593 * Enable the memcg OOM handling for faults triggered in user
5594 * space. Kernel faults are handled more gracefully.
5596 if (flags
& FAULT_FLAG_USER
)
5597 mem_cgroup_enter_user_fault();
5599 lru_gen_enter_fault(vma
);
5601 if (unlikely(is_vm_hugetlb_page(vma
)))
5602 ret
= hugetlb_fault(vma
->vm_mm
, vma
, address
, flags
);
5604 ret
= __handle_mm_fault(vma
, address
, flags
);
5606 lru_gen_exit_fault();
5608 if (flags
& FAULT_FLAG_USER
) {
5609 mem_cgroup_exit_user_fault();
5611 * The task may have entered a memcg OOM situation but
5612 * if the allocation error was handled gracefully (no
5613 * VM_FAULT_OOM), there is no need to kill anything.
5614 * Just clean up the OOM state peacefully.
5616 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
5617 mem_cgroup_oom_synchronize(false);
5620 mm_account_fault(mm
, regs
, address
, flags
, ret
);
5624 EXPORT_SYMBOL_GPL(handle_mm_fault
);
5626 #ifdef CONFIG_LOCK_MM_AND_FIND_VMA
5627 #include <linux/extable.h>
5629 static inline bool get_mmap_lock_carefully(struct mm_struct
*mm
, struct pt_regs
*regs
)
5631 if (likely(mmap_read_trylock(mm
)))
5634 if (regs
&& !user_mode(regs
)) {
5635 unsigned long ip
= exception_ip(regs
);
5636 if (!search_exception_tables(ip
))
5640 return !mmap_read_lock_killable(mm
);
5643 static inline bool mmap_upgrade_trylock(struct mm_struct
*mm
)
5646 * We don't have this operation yet.
5648 * It should be easy enough to do: it's basically a
5649 * atomic_long_try_cmpxchg_acquire()
5650 * from RWSEM_READER_BIAS -> RWSEM_WRITER_LOCKED, but
5651 * it also needs the proper lockdep magic etc.
5656 static inline bool upgrade_mmap_lock_carefully(struct mm_struct
*mm
, struct pt_regs
*regs
)
5658 mmap_read_unlock(mm
);
5659 if (regs
&& !user_mode(regs
)) {
5660 unsigned long ip
= exception_ip(regs
);
5661 if (!search_exception_tables(ip
))
5664 return !mmap_write_lock_killable(mm
);
5668 * Helper for page fault handling.
5670 * This is kind of equivalend to "mmap_read_lock()" followed
5671 * by "find_extend_vma()", except it's a lot more careful about
5672 * the locking (and will drop the lock on failure).
5674 * For example, if we have a kernel bug that causes a page
5675 * fault, we don't want to just use mmap_read_lock() to get
5676 * the mm lock, because that would deadlock if the bug were
5677 * to happen while we're holding the mm lock for writing.
5679 * So this checks the exception tables on kernel faults in
5680 * order to only do this all for instructions that are actually
5681 * expected to fault.
5683 * We can also actually take the mm lock for writing if we
5684 * need to extend the vma, which helps the VM layer a lot.
5686 struct vm_area_struct
*lock_mm_and_find_vma(struct mm_struct
*mm
,
5687 unsigned long addr
, struct pt_regs
*regs
)
5689 struct vm_area_struct
*vma
;
5691 if (!get_mmap_lock_carefully(mm
, regs
))
5694 vma
= find_vma(mm
, addr
);
5695 if (likely(vma
&& (vma
->vm_start
<= addr
)))
5699 * Well, dang. We might still be successful, but only
5700 * if we can extend a vma to do so.
5702 if (!vma
|| !(vma
->vm_flags
& VM_GROWSDOWN
)) {
5703 mmap_read_unlock(mm
);
5708 * We can try to upgrade the mmap lock atomically,
5709 * in which case we can continue to use the vma
5710 * we already looked up.
5712 * Otherwise we'll have to drop the mmap lock and
5713 * re-take it, and also look up the vma again,
5716 if (!mmap_upgrade_trylock(mm
)) {
5717 if (!upgrade_mmap_lock_carefully(mm
, regs
))
5720 vma
= find_vma(mm
, addr
);
5723 if (vma
->vm_start
<= addr
)
5725 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
5729 if (expand_stack_locked(vma
, addr
))
5733 mmap_write_downgrade(mm
);
5737 mmap_write_unlock(mm
);
5742 #ifdef CONFIG_PER_VMA_LOCK
5744 * Lookup and lock a VMA under RCU protection. Returned VMA is guaranteed to be
5745 * stable and not isolated. If the VMA is not found or is being modified the
5746 * function returns NULL.
5748 struct vm_area_struct
*lock_vma_under_rcu(struct mm_struct
*mm
,
5749 unsigned long address
)
5751 MA_STATE(mas
, &mm
->mm_mt
, address
, address
);
5752 struct vm_area_struct
*vma
;
5756 vma
= mas_walk(&mas
);
5760 if (!vma_start_read(vma
))
5764 * find_mergeable_anon_vma uses adjacent vmas which are not locked.
5765 * This check must happen after vma_start_read(); otherwise, a
5766 * concurrent mremap() with MREMAP_DONTUNMAP could dissociate the VMA
5767 * from its anon_vma.
5769 if (unlikely(vma_is_anonymous(vma
) && !vma
->anon_vma
))
5770 goto inval_end_read
;
5772 /* Check since vm_start/vm_end might change before we lock the VMA */
5773 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
))
5774 goto inval_end_read
;
5776 /* Check if the VMA got isolated after we found it */
5777 if (vma
->detached
) {
5779 count_vm_vma_lock_event(VMA_LOCK_MISS
);
5780 /* The area was replaced with another one */
5791 count_vm_vma_lock_event(VMA_LOCK_ABORT
);
5794 #endif /* CONFIG_PER_VMA_LOCK */
5796 #ifndef __PAGETABLE_P4D_FOLDED
5798 * Allocate p4d page table.
5799 * We've already handled the fast-path in-line.
5801 int __p4d_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
5803 p4d_t
*new = p4d_alloc_one(mm
, address
);
5807 spin_lock(&mm
->page_table_lock
);
5808 if (pgd_present(*pgd
)) { /* Another has populated it */
5811 smp_wmb(); /* See comment in pmd_install() */
5812 pgd_populate(mm
, pgd
, new);
5814 spin_unlock(&mm
->page_table_lock
);
5817 #endif /* __PAGETABLE_P4D_FOLDED */
5819 #ifndef __PAGETABLE_PUD_FOLDED
5821 * Allocate page upper directory.
5822 * We've already handled the fast-path in-line.
5824 int __pud_alloc(struct mm_struct
*mm
, p4d_t
*p4d
, unsigned long address
)
5826 pud_t
*new = pud_alloc_one(mm
, address
);
5830 spin_lock(&mm
->page_table_lock
);
5831 if (!p4d_present(*p4d
)) {
5833 smp_wmb(); /* See comment in pmd_install() */
5834 p4d_populate(mm
, p4d
, new);
5835 } else /* Another has populated it */
5837 spin_unlock(&mm
->page_table_lock
);
5840 #endif /* __PAGETABLE_PUD_FOLDED */
5842 #ifndef __PAGETABLE_PMD_FOLDED
5844 * Allocate page middle directory.
5845 * We've already handled the fast-path in-line.
5847 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
5850 pmd_t
*new = pmd_alloc_one(mm
, address
);
5854 ptl
= pud_lock(mm
, pud
);
5855 if (!pud_present(*pud
)) {
5857 smp_wmb(); /* See comment in pmd_install() */
5858 pud_populate(mm
, pud
, new);
5859 } else { /* Another has populated it */
5865 #endif /* __PAGETABLE_PMD_FOLDED */
5868 * follow_pte - look up PTE at a user virtual address
5869 * @mm: the mm_struct of the target address space
5870 * @address: user virtual address
5871 * @ptepp: location to store found PTE
5872 * @ptlp: location to store the lock for the PTE
5874 * On a successful return, the pointer to the PTE is stored in @ptepp;
5875 * the corresponding lock is taken and its location is stored in @ptlp.
5876 * The contents of the PTE are only stable until @ptlp is released;
5877 * any further use, if any, must be protected against invalidation
5878 * with MMU notifiers.
5880 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
5881 * should be taken for read.
5883 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
5884 * it is not a good general-purpose API.
5886 * Return: zero on success, -ve otherwise.
5888 int follow_pte(struct mm_struct
*mm
, unsigned long address
,
5889 pte_t
**ptepp
, spinlock_t
**ptlp
)
5897 pgd
= pgd_offset(mm
, address
);
5898 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
5901 p4d
= p4d_offset(pgd
, address
);
5902 if (p4d_none(*p4d
) || unlikely(p4d_bad(*p4d
)))
5905 pud
= pud_offset(p4d
, address
);
5906 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
5909 pmd
= pmd_offset(pud
, address
);
5910 VM_BUG_ON(pmd_trans_huge(*pmd
));
5912 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
5915 if (!pte_present(ptep_get(ptep
)))
5920 pte_unmap_unlock(ptep
, *ptlp
);
5924 EXPORT_SYMBOL_GPL(follow_pte
);
5927 * follow_pfn - look up PFN at a user virtual address
5928 * @vma: memory mapping
5929 * @address: user virtual address
5930 * @pfn: location to store found PFN
5932 * Only IO mappings and raw PFN mappings are allowed.
5934 * This function does not allow the caller to read the permissions
5935 * of the PTE. Do not use it.
5937 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5939 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
5946 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
5949 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
5952 *pfn
= pte_pfn(ptep_get(ptep
));
5953 pte_unmap_unlock(ptep
, ptl
);
5956 EXPORT_SYMBOL(follow_pfn
);
5958 #ifdef CONFIG_HAVE_IOREMAP_PROT
5959 int follow_phys(struct vm_area_struct
*vma
,
5960 unsigned long address
, unsigned int flags
,
5961 unsigned long *prot
, resource_size_t
*phys
)
5967 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
5970 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
5972 pte
= ptep_get(ptep
);
5974 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
5977 *prot
= pgprot_val(pte_pgprot(pte
));
5978 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
5982 pte_unmap_unlock(ptep
, ptl
);
5988 * generic_access_phys - generic implementation for iomem mmap access
5989 * @vma: the vma to access
5990 * @addr: userspace address, not relative offset within @vma
5991 * @buf: buffer to read/write
5992 * @len: length of transfer
5993 * @write: set to FOLL_WRITE when writing, otherwise reading
5995 * This is a generic implementation for &vm_operations_struct.access for an
5996 * iomem mapping. This callback is used by access_process_vm() when the @vma is
5999 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
6000 void *buf
, int len
, int write
)
6002 resource_size_t phys_addr
;
6003 unsigned long prot
= 0;
6004 void __iomem
*maddr
;
6007 int offset
= offset_in_page(addr
);
6010 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
6014 if (follow_pte(vma
->vm_mm
, addr
, &ptep
, &ptl
))
6016 pte
= ptep_get(ptep
);
6017 pte_unmap_unlock(ptep
, ptl
);
6019 prot
= pgprot_val(pte_pgprot(pte
));
6020 phys_addr
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
6022 if ((write
& FOLL_WRITE
) && !pte_write(pte
))
6025 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
6029 if (follow_pte(vma
->vm_mm
, addr
, &ptep
, &ptl
))
6032 if (!pte_same(pte
, ptep_get(ptep
))) {
6033 pte_unmap_unlock(ptep
, ptl
);
6040 memcpy_toio(maddr
+ offset
, buf
, len
);
6042 memcpy_fromio(buf
, maddr
+ offset
, len
);
6044 pte_unmap_unlock(ptep
, ptl
);
6050 EXPORT_SYMBOL_GPL(generic_access_phys
);
6054 * Access another process' address space as given in mm.
6056 static int __access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
6057 void *buf
, int len
, unsigned int gup_flags
)
6059 void *old_buf
= buf
;
6060 int write
= gup_flags
& FOLL_WRITE
;
6062 if (mmap_read_lock_killable(mm
))
6065 /* Untag the address before looking up the VMA */
6066 addr
= untagged_addr_remote(mm
, addr
);
6068 /* Avoid triggering the temporary warning in __get_user_pages */
6069 if (!vma_lookup(mm
, addr
) && !expand_stack(mm
, addr
))
6072 /* ignore errors, just check how much was successfully transferred */
6076 struct vm_area_struct
*vma
= NULL
;
6077 struct page
*page
= get_user_page_vma_remote(mm
, addr
,
6081 /* We might need to expand the stack to access it */
6082 vma
= vma_lookup(mm
, addr
);
6084 vma
= expand_stack(mm
, addr
);
6086 /* mmap_lock was dropped on failure */
6088 return buf
- old_buf
;
6090 /* Try again if stack expansion worked */
6095 * Check if this is a VM_IO | VM_PFNMAP VMA, which
6096 * we can access using slightly different code.
6099 #ifdef CONFIG_HAVE_IOREMAP_PROT
6100 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
6101 bytes
= vma
->vm_ops
->access(vma
, addr
, buf
,
6108 offset
= addr
& (PAGE_SIZE
-1);
6109 if (bytes
> PAGE_SIZE
-offset
)
6110 bytes
= PAGE_SIZE
-offset
;
6112 maddr
= kmap_local_page(page
);
6114 copy_to_user_page(vma
, page
, addr
,
6115 maddr
+ offset
, buf
, bytes
);
6116 set_page_dirty_lock(page
);
6118 copy_from_user_page(vma
, page
, addr
,
6119 buf
, maddr
+ offset
, bytes
);
6121 unmap_and_put_page(page
, maddr
);
6127 mmap_read_unlock(mm
);
6129 return buf
- old_buf
;
6133 * access_remote_vm - access another process' address space
6134 * @mm: the mm_struct of the target address space
6135 * @addr: start address to access
6136 * @buf: source or destination buffer
6137 * @len: number of bytes to transfer
6138 * @gup_flags: flags modifying lookup behaviour
6140 * The caller must hold a reference on @mm.
6142 * Return: number of bytes copied from source to destination.
6144 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
6145 void *buf
, int len
, unsigned int gup_flags
)
6147 return __access_remote_vm(mm
, addr
, buf
, len
, gup_flags
);
6151 * Access another process' address space.
6152 * Source/target buffer must be kernel space,
6153 * Do not walk the page table directly, use get_user_pages
6155 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
6156 void *buf
, int len
, unsigned int gup_flags
)
6158 struct mm_struct
*mm
;
6161 mm
= get_task_mm(tsk
);
6165 ret
= __access_remote_vm(mm
, addr
, buf
, len
, gup_flags
);
6171 EXPORT_SYMBOL_GPL(access_process_vm
);
6174 * Print the name of a VMA.
6176 void print_vma_addr(char *prefix
, unsigned long ip
)
6178 struct mm_struct
*mm
= current
->mm
;
6179 struct vm_area_struct
*vma
;
6182 * we might be running from an atomic context so we cannot sleep
6184 if (!mmap_read_trylock(mm
))
6187 vma
= find_vma(mm
, ip
);
6188 if (vma
&& vma
->vm_file
) {
6189 struct file
*f
= vma
->vm_file
;
6190 char *buf
= (char *)__get_free_page(GFP_NOWAIT
);
6194 p
= file_path(f
, buf
, PAGE_SIZE
);
6197 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
6199 vma
->vm_end
- vma
->vm_start
);
6200 free_page((unsigned long)buf
);
6203 mmap_read_unlock(mm
);
6206 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6207 void __might_fault(const char *file
, int line
)
6209 if (pagefault_disabled())
6211 __might_sleep(file
, line
);
6212 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6214 might_lock_read(¤t
->mm
->mmap_lock
);
6217 EXPORT_SYMBOL(__might_fault
);
6220 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
6222 * Process all subpages of the specified huge page with the specified
6223 * operation. The target subpage will be processed last to keep its
6226 static inline int process_huge_page(
6227 unsigned long addr_hint
, unsigned int pages_per_huge_page
,
6228 int (*process_subpage
)(unsigned long addr
, int idx
, void *arg
),
6231 int i
, n
, base
, l
, ret
;
6232 unsigned long addr
= addr_hint
&
6233 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
6235 /* Process target subpage last to keep its cache lines hot */
6237 n
= (addr_hint
- addr
) / PAGE_SIZE
;
6238 if (2 * n
<= pages_per_huge_page
) {
6239 /* If target subpage in first half of huge page */
6242 /* Process subpages at the end of huge page */
6243 for (i
= pages_per_huge_page
- 1; i
>= 2 * n
; i
--) {
6245 ret
= process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
6250 /* If target subpage in second half of huge page */
6251 base
= pages_per_huge_page
- 2 * (pages_per_huge_page
- n
);
6252 l
= pages_per_huge_page
- n
;
6253 /* Process subpages at the begin of huge page */
6254 for (i
= 0; i
< base
; i
++) {
6256 ret
= process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
6262 * Process remaining subpages in left-right-left-right pattern
6263 * towards the target subpage
6265 for (i
= 0; i
< l
; i
++) {
6266 int left_idx
= base
+ i
;
6267 int right_idx
= base
+ 2 * l
- 1 - i
;
6270 ret
= process_subpage(addr
+ left_idx
* PAGE_SIZE
, left_idx
, arg
);
6274 ret
= process_subpage(addr
+ right_idx
* PAGE_SIZE
, right_idx
, arg
);
6281 static void clear_gigantic_page(struct page
*page
,
6283 unsigned int pages_per_huge_page
)
6289 for (i
= 0; i
< pages_per_huge_page
; i
++) {
6290 p
= nth_page(page
, i
);
6292 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
6296 static int clear_subpage(unsigned long addr
, int idx
, void *arg
)
6298 struct page
*page
= arg
;
6300 clear_user_highpage(nth_page(page
, idx
), addr
);
6304 void clear_huge_page(struct page
*page
,
6305 unsigned long addr_hint
, unsigned int pages_per_huge_page
)
6307 unsigned long addr
= addr_hint
&
6308 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
6310 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
6311 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
6315 process_huge_page(addr_hint
, pages_per_huge_page
, clear_subpage
, page
);
6318 static int copy_user_gigantic_page(struct folio
*dst
, struct folio
*src
,
6320 struct vm_area_struct
*vma
,
6321 unsigned int pages_per_huge_page
)
6324 struct page
*dst_page
;
6325 struct page
*src_page
;
6327 for (i
= 0; i
< pages_per_huge_page
; i
++) {
6328 dst_page
= folio_page(dst
, i
);
6329 src_page
= folio_page(src
, i
);
6332 if (copy_mc_user_highpage(dst_page
, src_page
,
6333 addr
+ i
*PAGE_SIZE
, vma
)) {
6334 memory_failure_queue(page_to_pfn(src_page
), 0);
6341 struct copy_subpage_arg
{
6344 struct vm_area_struct
*vma
;
6347 static int copy_subpage(unsigned long addr
, int idx
, void *arg
)
6349 struct copy_subpage_arg
*copy_arg
= arg
;
6350 struct page
*dst
= nth_page(copy_arg
->dst
, idx
);
6351 struct page
*src
= nth_page(copy_arg
->src
, idx
);
6353 if (copy_mc_user_highpage(dst
, src
, addr
, copy_arg
->vma
)) {
6354 memory_failure_queue(page_to_pfn(src
), 0);
6360 int copy_user_large_folio(struct folio
*dst
, struct folio
*src
,
6361 unsigned long addr_hint
, struct vm_area_struct
*vma
)
6363 unsigned int pages_per_huge_page
= folio_nr_pages(dst
);
6364 unsigned long addr
= addr_hint
&
6365 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
6366 struct copy_subpage_arg arg
= {
6372 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
))
6373 return copy_user_gigantic_page(dst
, src
, addr
, vma
,
6374 pages_per_huge_page
);
6376 return process_huge_page(addr_hint
, pages_per_huge_page
, copy_subpage
, &arg
);
6379 long copy_folio_from_user(struct folio
*dst_folio
,
6380 const void __user
*usr_src
,
6381 bool allow_pagefault
)
6384 unsigned long i
, rc
= 0;
6385 unsigned int nr_pages
= folio_nr_pages(dst_folio
);
6386 unsigned long ret_val
= nr_pages
* PAGE_SIZE
;
6387 struct page
*subpage
;
6389 for (i
= 0; i
< nr_pages
; i
++) {
6390 subpage
= folio_page(dst_folio
, i
);
6391 kaddr
= kmap_local_page(subpage
);
6392 if (!allow_pagefault
)
6393 pagefault_disable();
6394 rc
= copy_from_user(kaddr
, usr_src
+ i
* PAGE_SIZE
, PAGE_SIZE
);
6395 if (!allow_pagefault
)
6397 kunmap_local(kaddr
);
6399 ret_val
-= (PAGE_SIZE
- rc
);
6403 flush_dcache_page(subpage
);
6409 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
6411 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
6413 static struct kmem_cache
*page_ptl_cachep
;
6415 void __init
ptlock_cache_init(void)
6417 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
6421 bool ptlock_alloc(struct ptdesc
*ptdesc
)
6425 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
6432 void ptlock_free(struct ptdesc
*ptdesc
)
6434 kmem_cache_free(page_ptl_cachep
, ptdesc
->ptl
);