1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * demand-loading started 01.12.91 - seems it is high on the list of
10 * things wanted, and it should be easy to implement. - Linus
14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
15 * pages started 02.12.91, seems to work. - Linus.
17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
18 * would have taken more than the 6M I have free, but it worked well as
21 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
25 * Real VM (paging to/from disk) started 18.12.91. Much more work and
26 * thought has to go into this. Oh, well..
27 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
28 * Found it. Everything seems to work now.
29 * 20.12.91 - Ok, making the swap-device changeable like the root.
33 * 05.04.94 - Multi-page memory management added for v1.1.
34 * Idea by Alex Bligh (alex@cconcepts.co.uk)
36 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
37 * (Gerhard.Wichert@pdb.siemens.de)
39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
42 #include <linux/kernel_stat.h>
44 #include <linux/mm_inline.h>
45 #include <linux/sched/mm.h>
46 #include <linux/sched/coredump.h>
47 #include <linux/sched/numa_balancing.h>
48 #include <linux/sched/task.h>
49 #include <linux/hugetlb.h>
50 #include <linux/mman.h>
51 #include <linux/swap.h>
52 #include <linux/highmem.h>
53 #include <linux/pagemap.h>
54 #include <linux/memremap.h>
55 #include <linux/ksm.h>
56 #include <linux/rmap.h>
57 #include <linux/export.h>
58 #include <linux/delayacct.h>
59 #include <linux/init.h>
60 #include <linux/pfn_t.h>
61 #include <linux/writeback.h>
62 #include <linux/memcontrol.h>
63 #include <linux/mmu_notifier.h>
64 #include <linux/swapops.h>
65 #include <linux/elf.h>
66 #include <linux/gfp.h>
67 #include <linux/migrate.h>
68 #include <linux/string.h>
69 #include <linux/debugfs.h>
70 #include <linux/userfaultfd_k.h>
71 #include <linux/dax.h>
72 #include <linux/oom.h>
73 #include <linux/numa.h>
74 #include <linux/perf_event.h>
75 #include <linux/ptrace.h>
76 #include <linux/vmalloc.h>
78 #include <trace/events/kmem.h>
81 #include <asm/mmu_context.h>
82 #include <asm/pgalloc.h>
83 #include <linux/uaccess.h>
85 #include <asm/tlbflush.h>
87 #include "pgalloc-track.h"
91 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
92 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
96 unsigned long max_mapnr
;
97 EXPORT_SYMBOL(max_mapnr
);
100 EXPORT_SYMBOL(mem_map
);
103 static vm_fault_t
do_fault(struct vm_fault
*vmf
);
106 * A number of key systems in x86 including ioremap() rely on the assumption
107 * that high_memory defines the upper bound on direct map memory, then end
108 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
109 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
113 EXPORT_SYMBOL(high_memory
);
116 * Randomize the address space (stacks, mmaps, brk, etc.).
118 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
119 * as ancient (libc5 based) binaries can segfault. )
121 int randomize_va_space __read_mostly
=
122 #ifdef CONFIG_COMPAT_BRK
128 #ifndef arch_faults_on_old_pte
129 static inline bool arch_faults_on_old_pte(void)
132 * Those arches which don't have hw access flag feature need to
133 * implement their own helper. By default, "true" means pagefault
134 * will be hit on old pte.
140 #ifndef arch_wants_old_prefaulted_pte
141 static inline bool arch_wants_old_prefaulted_pte(void)
144 * Transitioning a PTE from 'old' to 'young' can be expensive on
145 * some architectures, even if it's performed in hardware. By
146 * default, "false" means prefaulted entries will be 'young'.
152 static int __init
disable_randmaps(char *s
)
154 randomize_va_space
= 0;
157 __setup("norandmaps", disable_randmaps
);
159 unsigned long zero_pfn __read_mostly
;
160 EXPORT_SYMBOL(zero_pfn
);
162 unsigned long highest_memmap_pfn __read_mostly
;
165 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
167 static int __init
init_zero_pfn(void)
169 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
172 early_initcall(init_zero_pfn
);
174 void mm_trace_rss_stat(struct mm_struct
*mm
, int member
, long count
)
176 trace_rss_stat(mm
, member
, count
);
179 #if defined(SPLIT_RSS_COUNTING)
181 void sync_mm_rss(struct mm_struct
*mm
)
185 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
186 if (current
->rss_stat
.count
[i
]) {
187 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
188 current
->rss_stat
.count
[i
] = 0;
191 current
->rss_stat
.events
= 0;
194 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
196 struct task_struct
*task
= current
;
198 if (likely(task
->mm
== mm
))
199 task
->rss_stat
.count
[member
] += val
;
201 add_mm_counter(mm
, member
, val
);
203 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
204 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
206 /* sync counter once per 64 page faults */
207 #define TASK_RSS_EVENTS_THRESH (64)
208 static void check_sync_rss_stat(struct task_struct
*task
)
210 if (unlikely(task
!= current
))
212 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
213 sync_mm_rss(task
->mm
);
215 #else /* SPLIT_RSS_COUNTING */
217 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
218 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
220 static void check_sync_rss_stat(struct task_struct
*task
)
224 #endif /* SPLIT_RSS_COUNTING */
227 * Note: this doesn't free the actual pages themselves. That
228 * has been handled earlier when unmapping all the memory regions.
230 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
233 pgtable_t token
= pmd_pgtable(*pmd
);
235 pte_free_tlb(tlb
, token
, addr
);
236 mm_dec_nr_ptes(tlb
->mm
);
239 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
240 unsigned long addr
, unsigned long end
,
241 unsigned long floor
, unsigned long ceiling
)
248 pmd
= pmd_offset(pud
, addr
);
250 next
= pmd_addr_end(addr
, end
);
251 if (pmd_none_or_clear_bad(pmd
))
253 free_pte_range(tlb
, pmd
, addr
);
254 } while (pmd
++, addr
= next
, addr
!= end
);
264 if (end
- 1 > ceiling
- 1)
267 pmd
= pmd_offset(pud
, start
);
269 pmd_free_tlb(tlb
, pmd
, start
);
270 mm_dec_nr_pmds(tlb
->mm
);
273 static inline void free_pud_range(struct mmu_gather
*tlb
, p4d_t
*p4d
,
274 unsigned long addr
, unsigned long end
,
275 unsigned long floor
, unsigned long ceiling
)
282 pud
= pud_offset(p4d
, addr
);
284 next
= pud_addr_end(addr
, end
);
285 if (pud_none_or_clear_bad(pud
))
287 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
288 } while (pud
++, addr
= next
, addr
!= end
);
298 if (end
- 1 > ceiling
- 1)
301 pud
= pud_offset(p4d
, start
);
303 pud_free_tlb(tlb
, pud
, start
);
304 mm_dec_nr_puds(tlb
->mm
);
307 static inline void free_p4d_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
308 unsigned long addr
, unsigned long end
,
309 unsigned long floor
, unsigned long ceiling
)
316 p4d
= p4d_offset(pgd
, addr
);
318 next
= p4d_addr_end(addr
, end
);
319 if (p4d_none_or_clear_bad(p4d
))
321 free_pud_range(tlb
, p4d
, addr
, next
, floor
, ceiling
);
322 } while (p4d
++, addr
= next
, addr
!= end
);
328 ceiling
&= PGDIR_MASK
;
332 if (end
- 1 > ceiling
- 1)
335 p4d
= p4d_offset(pgd
, start
);
337 p4d_free_tlb(tlb
, p4d
, start
);
341 * This function frees user-level page tables of a process.
343 void free_pgd_range(struct mmu_gather
*tlb
,
344 unsigned long addr
, unsigned long end
,
345 unsigned long floor
, unsigned long ceiling
)
351 * The next few lines have given us lots of grief...
353 * Why are we testing PMD* at this top level? Because often
354 * there will be no work to do at all, and we'd prefer not to
355 * go all the way down to the bottom just to discover that.
357 * Why all these "- 1"s? Because 0 represents both the bottom
358 * of the address space and the top of it (using -1 for the
359 * top wouldn't help much: the masks would do the wrong thing).
360 * The rule is that addr 0 and floor 0 refer to the bottom of
361 * the address space, but end 0 and ceiling 0 refer to the top
362 * Comparisons need to use "end - 1" and "ceiling - 1" (though
363 * that end 0 case should be mythical).
365 * Wherever addr is brought up or ceiling brought down, we must
366 * be careful to reject "the opposite 0" before it confuses the
367 * subsequent tests. But what about where end is brought down
368 * by PMD_SIZE below? no, end can't go down to 0 there.
370 * Whereas we round start (addr) and ceiling down, by different
371 * masks at different levels, in order to test whether a table
372 * now has no other vmas using it, so can be freed, we don't
373 * bother to round floor or end up - the tests don't need that.
387 if (end
- 1 > ceiling
- 1)
392 * We add page table cache pages with PAGE_SIZE,
393 * (see pte_free_tlb()), flush the tlb if we need
395 tlb_change_page_size(tlb
, PAGE_SIZE
);
396 pgd
= pgd_offset(tlb
->mm
, addr
);
398 next
= pgd_addr_end(addr
, end
);
399 if (pgd_none_or_clear_bad(pgd
))
401 free_p4d_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
402 } while (pgd
++, addr
= next
, addr
!= end
);
405 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
406 unsigned long floor
, unsigned long ceiling
)
409 struct vm_area_struct
*next
= vma
->vm_next
;
410 unsigned long addr
= vma
->vm_start
;
413 * Hide vma from rmap and truncate_pagecache before freeing
416 unlink_anon_vmas(vma
);
417 unlink_file_vma(vma
);
419 if (is_vm_hugetlb_page(vma
)) {
420 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
421 floor
, next
? next
->vm_start
: ceiling
);
424 * Optimization: gather nearby vmas into one call down
426 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
427 && !is_vm_hugetlb_page(next
)) {
430 unlink_anon_vmas(vma
);
431 unlink_file_vma(vma
);
433 free_pgd_range(tlb
, addr
, vma
->vm_end
,
434 floor
, next
? next
->vm_start
: ceiling
);
440 void pmd_install(struct mm_struct
*mm
, pmd_t
*pmd
, pgtable_t
*pte
)
442 spinlock_t
*ptl
= pmd_lock(mm
, pmd
);
444 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
447 * Ensure all pte setup (eg. pte page lock and page clearing) are
448 * visible before the pte is made visible to other CPUs by being
449 * put into page tables.
451 * The other side of the story is the pointer chasing in the page
452 * table walking code (when walking the page table without locking;
453 * ie. most of the time). Fortunately, these data accesses consist
454 * of a chain of data-dependent loads, meaning most CPUs (alpha
455 * being the notable exception) will already guarantee loads are
456 * seen in-order. See the alpha page table accessors for the
457 * smp_rmb() barriers in page table walking code.
459 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
460 pmd_populate(mm
, pmd
, *pte
);
466 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
)
468 pgtable_t
new = pte_alloc_one(mm
);
472 pmd_install(mm
, pmd
, &new);
478 int __pte_alloc_kernel(pmd_t
*pmd
)
480 pte_t
*new = pte_alloc_one_kernel(&init_mm
);
484 spin_lock(&init_mm
.page_table_lock
);
485 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
486 smp_wmb(); /* See comment in pmd_install() */
487 pmd_populate_kernel(&init_mm
, pmd
, new);
490 spin_unlock(&init_mm
.page_table_lock
);
492 pte_free_kernel(&init_mm
, new);
496 static inline void init_rss_vec(int *rss
)
498 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
501 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
505 if (current
->mm
== mm
)
507 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
509 add_mm_counter(mm
, i
, rss
[i
]);
513 * This function is called to print an error when a bad pte
514 * is found. For example, we might have a PFN-mapped pte in
515 * a region that doesn't allow it.
517 * The calling function must still handle the error.
519 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
520 pte_t pte
, struct page
*page
)
522 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
523 p4d_t
*p4d
= p4d_offset(pgd
, addr
);
524 pud_t
*pud
= pud_offset(p4d
, addr
);
525 pmd_t
*pmd
= pmd_offset(pud
, addr
);
526 struct address_space
*mapping
;
528 static unsigned long resume
;
529 static unsigned long nr_shown
;
530 static unsigned long nr_unshown
;
533 * Allow a burst of 60 reports, then keep quiet for that minute;
534 * or allow a steady drip of one report per second.
536 if (nr_shown
== 60) {
537 if (time_before(jiffies
, resume
)) {
542 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
549 resume
= jiffies
+ 60 * HZ
;
551 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
552 index
= linear_page_index(vma
, addr
);
554 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
556 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
558 dump_page(page
, "bad pte");
559 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
560 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
561 pr_alert("file:%pD fault:%ps mmap:%ps read_folio:%ps\n",
563 vma
->vm_ops
? vma
->vm_ops
->fault
: NULL
,
564 vma
->vm_file
? vma
->vm_file
->f_op
->mmap
: NULL
,
565 mapping
? mapping
->a_ops
->read_folio
: NULL
);
567 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
571 * vm_normal_page -- This function gets the "struct page" associated with a pte.
573 * "Special" mappings do not wish to be associated with a "struct page" (either
574 * it doesn't exist, or it exists but they don't want to touch it). In this
575 * case, NULL is returned here. "Normal" mappings do have a struct page.
577 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
578 * pte bit, in which case this function is trivial. Secondly, an architecture
579 * may not have a spare pte bit, which requires a more complicated scheme,
582 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
583 * special mapping (even if there are underlying and valid "struct pages").
584 * COWed pages of a VM_PFNMAP are always normal.
586 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
587 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
588 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
589 * mapping will always honor the rule
591 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
593 * And for normal mappings this is false.
595 * This restricts such mappings to be a linear translation from virtual address
596 * to pfn. To get around this restriction, we allow arbitrary mappings so long
597 * as the vma is not a COW mapping; in that case, we know that all ptes are
598 * special (because none can have been COWed).
601 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
603 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
604 * page" backing, however the difference is that _all_ pages with a struct
605 * page (that is, those where pfn_valid is true) are refcounted and considered
606 * normal pages by the VM. The disadvantage is that pages are refcounted
607 * (which can be slower and simply not an option for some PFNMAP users). The
608 * advantage is that we don't have to follow the strict linearity rule of
609 * PFNMAP mappings in order to support COWable mappings.
612 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
615 unsigned long pfn
= pte_pfn(pte
);
617 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
)) {
618 if (likely(!pte_special(pte
)))
620 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
621 return vma
->vm_ops
->find_special_page(vma
, addr
);
622 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
624 if (is_zero_pfn(pfn
))
628 * NOTE: New users of ZONE_DEVICE will not set pte_devmap()
629 * and will have refcounts incremented on their struct pages
630 * when they are inserted into PTEs, thus they are safe to
631 * return here. Legacy ZONE_DEVICE pages that set pte_devmap()
632 * do not have refcounts. Example of legacy ZONE_DEVICE is
633 * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers.
637 print_bad_pte(vma
, addr
, pte
, NULL
);
641 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
643 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
644 if (vma
->vm_flags
& VM_MIXEDMAP
) {
650 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
651 if (pfn
== vma
->vm_pgoff
+ off
)
653 if (!is_cow_mapping(vma
->vm_flags
))
658 if (is_zero_pfn(pfn
))
662 if (unlikely(pfn
> highest_memmap_pfn
)) {
663 print_bad_pte(vma
, addr
, pte
, NULL
);
668 * NOTE! We still have PageReserved() pages in the page tables.
669 * eg. VDSO mappings can cause them to exist.
672 return pfn_to_page(pfn
);
675 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
676 struct page
*vm_normal_page_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
679 unsigned long pfn
= pmd_pfn(pmd
);
682 * There is no pmd_special() but there may be special pmds, e.g.
683 * in a direct-access (dax) mapping, so let's just replicate the
684 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
686 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
687 if (vma
->vm_flags
& VM_MIXEDMAP
) {
693 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
694 if (pfn
== vma
->vm_pgoff
+ off
)
696 if (!is_cow_mapping(vma
->vm_flags
))
703 if (is_huge_zero_pmd(pmd
))
705 if (unlikely(pfn
> highest_memmap_pfn
))
709 * NOTE! We still have PageReserved() pages in the page tables.
710 * eg. VDSO mappings can cause them to exist.
713 return pfn_to_page(pfn
);
717 static void restore_exclusive_pte(struct vm_area_struct
*vma
,
718 struct page
*page
, unsigned long address
,
724 pte
= pte_mkold(mk_pte(page
, READ_ONCE(vma
->vm_page_prot
)));
725 if (pte_swp_soft_dirty(*ptep
))
726 pte
= pte_mksoft_dirty(pte
);
728 entry
= pte_to_swp_entry(*ptep
);
729 if (pte_swp_uffd_wp(*ptep
))
730 pte
= pte_mkuffd_wp(pte
);
731 else if (is_writable_device_exclusive_entry(entry
))
732 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
734 VM_BUG_ON(pte_write(pte
) && !(PageAnon(page
) && PageAnonExclusive(page
)));
737 * No need to take a page reference as one was already
738 * created when the swap entry was made.
741 page_add_anon_rmap(page
, vma
, address
, RMAP_NONE
);
744 * Currently device exclusive access only supports anonymous
745 * memory so the entry shouldn't point to a filebacked page.
749 set_pte_at(vma
->vm_mm
, address
, ptep
, pte
);
752 * No need to invalidate - it was non-present before. However
753 * secondary CPUs may have mappings that need invalidating.
755 update_mmu_cache(vma
, address
, ptep
);
759 * Tries to restore an exclusive pte if the page lock can be acquired without
763 try_restore_exclusive_pte(pte_t
*src_pte
, struct vm_area_struct
*vma
,
766 swp_entry_t entry
= pte_to_swp_entry(*src_pte
);
767 struct page
*page
= pfn_swap_entry_to_page(entry
);
769 if (trylock_page(page
)) {
770 restore_exclusive_pte(vma
, page
, addr
, src_pte
);
779 * copy one vm_area from one task to the other. Assumes the page tables
780 * already present in the new task to be cleared in the whole range
781 * covered by this vma.
785 copy_nonpresent_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
786 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*dst_vma
,
787 struct vm_area_struct
*src_vma
, unsigned long addr
, int *rss
)
789 unsigned long vm_flags
= dst_vma
->vm_flags
;
790 pte_t pte
= *src_pte
;
792 swp_entry_t entry
= pte_to_swp_entry(pte
);
794 if (likely(!non_swap_entry(entry
))) {
795 if (swap_duplicate(entry
) < 0)
798 /* make sure dst_mm is on swapoff's mmlist. */
799 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
800 spin_lock(&mmlist_lock
);
801 if (list_empty(&dst_mm
->mmlist
))
802 list_add(&dst_mm
->mmlist
,
804 spin_unlock(&mmlist_lock
);
806 /* Mark the swap entry as shared. */
807 if (pte_swp_exclusive(*src_pte
)) {
808 pte
= pte_swp_clear_exclusive(*src_pte
);
809 set_pte_at(src_mm
, addr
, src_pte
, pte
);
812 } else if (is_migration_entry(entry
)) {
813 page
= pfn_swap_entry_to_page(entry
);
815 rss
[mm_counter(page
)]++;
817 if (!is_readable_migration_entry(entry
) &&
818 is_cow_mapping(vm_flags
)) {
820 * COW mappings require pages in both parent and child
821 * to be set to read. A previously exclusive entry is
824 entry
= make_readable_migration_entry(
826 pte
= swp_entry_to_pte(entry
);
827 if (pte_swp_soft_dirty(*src_pte
))
828 pte
= pte_swp_mksoft_dirty(pte
);
829 if (pte_swp_uffd_wp(*src_pte
))
830 pte
= pte_swp_mkuffd_wp(pte
);
831 set_pte_at(src_mm
, addr
, src_pte
, pte
);
833 } else if (is_device_private_entry(entry
)) {
834 page
= pfn_swap_entry_to_page(entry
);
837 * Update rss count even for unaddressable pages, as
838 * they should treated just like normal pages in this
841 * We will likely want to have some new rss counters
842 * for unaddressable pages, at some point. But for now
843 * keep things as they are.
846 rss
[mm_counter(page
)]++;
847 /* Cannot fail as these pages cannot get pinned. */
848 BUG_ON(page_try_dup_anon_rmap(page
, false, src_vma
));
851 * We do not preserve soft-dirty information, because so
852 * far, checkpoint/restore is the only feature that
853 * requires that. And checkpoint/restore does not work
854 * when a device driver is involved (you cannot easily
855 * save and restore device driver state).
857 if (is_writable_device_private_entry(entry
) &&
858 is_cow_mapping(vm_flags
)) {
859 entry
= make_readable_device_private_entry(
861 pte
= swp_entry_to_pte(entry
);
862 if (pte_swp_uffd_wp(*src_pte
))
863 pte
= pte_swp_mkuffd_wp(pte
);
864 set_pte_at(src_mm
, addr
, src_pte
, pte
);
866 } else if (is_device_exclusive_entry(entry
)) {
868 * Make device exclusive entries present by restoring the
869 * original entry then copying as for a present pte. Device
870 * exclusive entries currently only support private writable
871 * (ie. COW) mappings.
873 VM_BUG_ON(!is_cow_mapping(src_vma
->vm_flags
));
874 if (try_restore_exclusive_pte(src_pte
, src_vma
, addr
))
877 } else if (is_pte_marker_entry(entry
)) {
879 * We're copying the pgtable should only because dst_vma has
880 * uffd-wp enabled, do sanity check.
882 WARN_ON_ONCE(!userfaultfd_wp(dst_vma
));
883 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
886 if (!userfaultfd_wp(dst_vma
))
887 pte
= pte_swp_clear_uffd_wp(pte
);
888 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
893 * Copy a present and normal page.
895 * NOTE! The usual case is that this isn't required;
896 * instead, the caller can just increase the page refcount
897 * and re-use the pte the traditional way.
899 * And if we need a pre-allocated page but don't yet have
900 * one, return a negative error to let the preallocation
901 * code know so that it can do so outside the page table
905 copy_present_page(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
906 pte_t
*dst_pte
, pte_t
*src_pte
, unsigned long addr
, int *rss
,
907 struct page
**prealloc
, struct page
*page
)
909 struct page
*new_page
;
912 new_page
= *prealloc
;
917 * We have a prealloc page, all good! Take it
918 * over and copy the page & arm it.
921 copy_user_highpage(new_page
, page
, addr
, src_vma
);
922 __SetPageUptodate(new_page
);
923 page_add_new_anon_rmap(new_page
, dst_vma
, addr
);
924 lru_cache_add_inactive_or_unevictable(new_page
, dst_vma
);
925 rss
[mm_counter(new_page
)]++;
927 /* All done, just insert the new page copy in the child */
928 pte
= mk_pte(new_page
, dst_vma
->vm_page_prot
);
929 pte
= maybe_mkwrite(pte_mkdirty(pte
), dst_vma
);
930 if (userfaultfd_pte_wp(dst_vma
, *src_pte
))
931 /* Uffd-wp needs to be delivered to dest pte as well */
932 pte
= pte_wrprotect(pte_mkuffd_wp(pte
));
933 set_pte_at(dst_vma
->vm_mm
, addr
, dst_pte
, pte
);
938 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
939 * is required to copy this pte.
942 copy_present_pte(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
943 pte_t
*dst_pte
, pte_t
*src_pte
, unsigned long addr
, int *rss
,
944 struct page
**prealloc
)
946 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
947 unsigned long vm_flags
= src_vma
->vm_flags
;
948 pte_t pte
= *src_pte
;
951 page
= vm_normal_page(src_vma
, addr
, pte
);
952 if (page
&& PageAnon(page
)) {
954 * If this page may have been pinned by the parent process,
955 * copy the page immediately for the child so that we'll always
956 * guarantee the pinned page won't be randomly replaced in the
960 if (unlikely(page_try_dup_anon_rmap(page
, false, src_vma
))) {
961 /* Page maybe pinned, we have to copy. */
963 return copy_present_page(dst_vma
, src_vma
, dst_pte
, src_pte
,
964 addr
, rss
, prealloc
, page
);
966 rss
[mm_counter(page
)]++;
969 page_dup_file_rmap(page
, false);
970 rss
[mm_counter(page
)]++;
974 * If it's a COW mapping, write protect it both
975 * in the parent and the child
977 if (is_cow_mapping(vm_flags
) && pte_write(pte
)) {
978 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
979 pte
= pte_wrprotect(pte
);
981 VM_BUG_ON(page
&& PageAnon(page
) && PageAnonExclusive(page
));
984 * If it's a shared mapping, mark it clean in
987 if (vm_flags
& VM_SHARED
)
988 pte
= pte_mkclean(pte
);
989 pte
= pte_mkold(pte
);
991 if (!userfaultfd_wp(dst_vma
))
992 pte
= pte_clear_uffd_wp(pte
);
994 set_pte_at(dst_vma
->vm_mm
, addr
, dst_pte
, pte
);
998 static inline struct page
*
999 page_copy_prealloc(struct mm_struct
*src_mm
, struct vm_area_struct
*vma
,
1002 struct page
*new_page
;
1004 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, addr
);
1008 if (mem_cgroup_charge(page_folio(new_page
), src_mm
, GFP_KERNEL
)) {
1012 cgroup_throttle_swaprate(new_page
, GFP_KERNEL
);
1018 copy_pte_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1019 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
1022 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1023 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1024 pte_t
*orig_src_pte
, *orig_dst_pte
;
1025 pte_t
*src_pte
, *dst_pte
;
1026 spinlock_t
*src_ptl
, *dst_ptl
;
1027 int progress
, ret
= 0;
1028 int rss
[NR_MM_COUNTERS
];
1029 swp_entry_t entry
= (swp_entry_t
){0};
1030 struct page
*prealloc
= NULL
;
1036 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
1041 src_pte
= pte_offset_map(src_pmd
, addr
);
1042 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
1043 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1044 orig_src_pte
= src_pte
;
1045 orig_dst_pte
= dst_pte
;
1046 arch_enter_lazy_mmu_mode();
1050 * We are holding two locks at this point - either of them
1051 * could generate latencies in another task on another CPU.
1053 if (progress
>= 32) {
1055 if (need_resched() ||
1056 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
1059 if (pte_none(*src_pte
)) {
1063 if (unlikely(!pte_present(*src_pte
))) {
1064 ret
= copy_nonpresent_pte(dst_mm
, src_mm
,
1069 entry
= pte_to_swp_entry(*src_pte
);
1071 } else if (ret
== -EBUSY
) {
1079 * Device exclusive entry restored, continue by copying
1080 * the now present pte.
1082 WARN_ON_ONCE(ret
!= -ENOENT
);
1084 /* copy_present_pte() will clear `*prealloc' if consumed */
1085 ret
= copy_present_pte(dst_vma
, src_vma
, dst_pte
, src_pte
,
1086 addr
, rss
, &prealloc
);
1088 * If we need a pre-allocated page for this pte, drop the
1089 * locks, allocate, and try again.
1091 if (unlikely(ret
== -EAGAIN
))
1093 if (unlikely(prealloc
)) {
1095 * pre-alloc page cannot be reused by next time so as
1096 * to strictly follow mempolicy (e.g., alloc_page_vma()
1097 * will allocate page according to address). This
1098 * could only happen if one pinned pte changed.
1104 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1106 arch_leave_lazy_mmu_mode();
1107 spin_unlock(src_ptl
);
1108 pte_unmap(orig_src_pte
);
1109 add_mm_rss_vec(dst_mm
, rss
);
1110 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
1114 VM_WARN_ON_ONCE(!entry
.val
);
1115 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0) {
1120 } else if (ret
== -EBUSY
) {
1122 } else if (ret
== -EAGAIN
) {
1123 prealloc
= page_copy_prealloc(src_mm
, src_vma
, addr
);
1130 /* We've captured and resolved the error. Reset, try again. */
1136 if (unlikely(prealloc
))
1142 copy_pmd_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1143 pud_t
*dst_pud
, pud_t
*src_pud
, unsigned long addr
,
1146 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1147 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1148 pmd_t
*src_pmd
, *dst_pmd
;
1151 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
1154 src_pmd
= pmd_offset(src_pud
, addr
);
1156 next
= pmd_addr_end(addr
, end
);
1157 if (is_swap_pmd(*src_pmd
) || pmd_trans_huge(*src_pmd
)
1158 || pmd_devmap(*src_pmd
)) {
1160 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PMD_SIZE
, src_vma
);
1161 err
= copy_huge_pmd(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
1162 addr
, dst_vma
, src_vma
);
1169 if (pmd_none_or_clear_bad(src_pmd
))
1171 if (copy_pte_range(dst_vma
, src_vma
, dst_pmd
, src_pmd
,
1174 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1179 copy_pud_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1180 p4d_t
*dst_p4d
, p4d_t
*src_p4d
, unsigned long addr
,
1183 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1184 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1185 pud_t
*src_pud
, *dst_pud
;
1188 dst_pud
= pud_alloc(dst_mm
, dst_p4d
, addr
);
1191 src_pud
= pud_offset(src_p4d
, addr
);
1193 next
= pud_addr_end(addr
, end
);
1194 if (pud_trans_huge(*src_pud
) || pud_devmap(*src_pud
)) {
1197 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PUD_SIZE
, src_vma
);
1198 err
= copy_huge_pud(dst_mm
, src_mm
,
1199 dst_pud
, src_pud
, addr
, src_vma
);
1206 if (pud_none_or_clear_bad(src_pud
))
1208 if (copy_pmd_range(dst_vma
, src_vma
, dst_pud
, src_pud
,
1211 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1216 copy_p4d_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1217 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, unsigned long addr
,
1220 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1221 p4d_t
*src_p4d
, *dst_p4d
;
1224 dst_p4d
= p4d_alloc(dst_mm
, dst_pgd
, addr
);
1227 src_p4d
= p4d_offset(src_pgd
, addr
);
1229 next
= p4d_addr_end(addr
, end
);
1230 if (p4d_none_or_clear_bad(src_p4d
))
1232 if (copy_pud_range(dst_vma
, src_vma
, dst_p4d
, src_p4d
,
1235 } while (dst_p4d
++, src_p4d
++, addr
= next
, addr
!= end
);
1240 * Return true if the vma needs to copy the pgtable during this fork(). Return
1241 * false when we can speed up fork() by allowing lazy page faults later until
1242 * when the child accesses the memory range.
1245 vma_needs_copy(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
)
1248 * Always copy pgtables when dst_vma has uffd-wp enabled even if it's
1249 * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable
1250 * contains uffd-wp protection information, that's something we can't
1251 * retrieve from page cache, and skip copying will lose those info.
1253 if (userfaultfd_wp(dst_vma
))
1256 if (src_vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
1259 if (src_vma
->anon_vma
)
1263 * Don't copy ptes where a page fault will fill them correctly. Fork
1264 * becomes much lighter when there are big shared or private readonly
1265 * mappings. The tradeoff is that copy_page_range is more efficient
1272 copy_page_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
)
1274 pgd_t
*src_pgd
, *dst_pgd
;
1276 unsigned long addr
= src_vma
->vm_start
;
1277 unsigned long end
= src_vma
->vm_end
;
1278 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1279 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1280 struct mmu_notifier_range range
;
1284 if (!vma_needs_copy(dst_vma
, src_vma
))
1287 if (is_vm_hugetlb_page(src_vma
))
1288 return copy_hugetlb_page_range(dst_mm
, src_mm
, dst_vma
, src_vma
);
1290 if (unlikely(src_vma
->vm_flags
& VM_PFNMAP
)) {
1292 * We do not free on error cases below as remove_vma
1293 * gets called on error from higher level routine
1295 ret
= track_pfn_copy(src_vma
);
1301 * We need to invalidate the secondary MMU mappings only when
1302 * there could be a permission downgrade on the ptes of the
1303 * parent mm. And a permission downgrade will only happen if
1304 * is_cow_mapping() returns true.
1306 is_cow
= is_cow_mapping(src_vma
->vm_flags
);
1309 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_PAGE
,
1310 0, src_vma
, src_mm
, addr
, end
);
1311 mmu_notifier_invalidate_range_start(&range
);
1313 * Disabling preemption is not needed for the write side, as
1314 * the read side doesn't spin, but goes to the mmap_lock.
1316 * Use the raw variant of the seqcount_t write API to avoid
1317 * lockdep complaining about preemptibility.
1319 mmap_assert_write_locked(src_mm
);
1320 raw_write_seqcount_begin(&src_mm
->write_protect_seq
);
1324 dst_pgd
= pgd_offset(dst_mm
, addr
);
1325 src_pgd
= pgd_offset(src_mm
, addr
);
1327 next
= pgd_addr_end(addr
, end
);
1328 if (pgd_none_or_clear_bad(src_pgd
))
1330 if (unlikely(copy_p4d_range(dst_vma
, src_vma
, dst_pgd
, src_pgd
,
1335 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1338 raw_write_seqcount_end(&src_mm
->write_protect_seq
);
1339 mmu_notifier_invalidate_range_end(&range
);
1345 * Parameter block passed down to zap_pte_range in exceptional cases.
1347 struct zap_details
{
1348 struct folio
*single_folio
; /* Locked folio to be unmapped */
1349 bool even_cows
; /* Zap COWed private pages too? */
1350 zap_flags_t zap_flags
; /* Extra flags for zapping */
1353 /* Whether we should zap all COWed (private) pages too */
1354 static inline bool should_zap_cows(struct zap_details
*details
)
1356 /* By default, zap all pages */
1360 /* Or, we zap COWed pages only if the caller wants to */
1361 return details
->even_cows
;
1364 /* Decides whether we should zap this page with the page pointer specified */
1365 static inline bool should_zap_page(struct zap_details
*details
, struct page
*page
)
1367 /* If we can make a decision without *page.. */
1368 if (should_zap_cows(details
))
1371 /* E.g. the caller passes NULL for the case of a zero page */
1375 /* Otherwise we should only zap non-anon pages */
1376 return !PageAnon(page
);
1379 static inline bool zap_drop_file_uffd_wp(struct zap_details
*details
)
1384 return details
->zap_flags
& ZAP_FLAG_DROP_MARKER
;
1388 * This function makes sure that we'll replace the none pte with an uffd-wp
1389 * swap special pte marker when necessary. Must be with the pgtable lock held.
1392 zap_install_uffd_wp_if_needed(struct vm_area_struct
*vma
,
1393 unsigned long addr
, pte_t
*pte
,
1394 struct zap_details
*details
, pte_t pteval
)
1396 if (zap_drop_file_uffd_wp(details
))
1399 pte_install_uffd_wp_if_needed(vma
, addr
, pte
, pteval
);
1402 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1403 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1404 unsigned long addr
, unsigned long end
,
1405 struct zap_details
*details
)
1407 struct mm_struct
*mm
= tlb
->mm
;
1408 int force_flush
= 0;
1409 int rss
[NR_MM_COUNTERS
];
1415 tlb_change_page_size(tlb
, PAGE_SIZE
);
1418 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1420 flush_tlb_batched_pending(mm
);
1421 arch_enter_lazy_mmu_mode();
1426 if (pte_none(ptent
))
1432 if (pte_present(ptent
)) {
1433 page
= vm_normal_page(vma
, addr
, ptent
);
1434 if (unlikely(!should_zap_page(details
, page
)))
1436 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1438 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1439 zap_install_uffd_wp_if_needed(vma
, addr
, pte
, details
,
1441 if (unlikely(!page
))
1444 if (!PageAnon(page
)) {
1445 if (pte_dirty(ptent
)) {
1447 set_page_dirty(page
);
1449 if (pte_young(ptent
) &&
1450 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1451 mark_page_accessed(page
);
1453 rss
[mm_counter(page
)]--;
1454 page_remove_rmap(page
, vma
, false);
1455 if (unlikely(page_mapcount(page
) < 0))
1456 print_bad_pte(vma
, addr
, ptent
, page
);
1457 if (unlikely(__tlb_remove_page(tlb
, page
))) {
1465 entry
= pte_to_swp_entry(ptent
);
1466 if (is_device_private_entry(entry
) ||
1467 is_device_exclusive_entry(entry
)) {
1468 page
= pfn_swap_entry_to_page(entry
);
1469 if (unlikely(!should_zap_page(details
, page
)))
1472 * Both device private/exclusive mappings should only
1473 * work with anonymous page so far, so we don't need to
1474 * consider uffd-wp bit when zap. For more information,
1475 * see zap_install_uffd_wp_if_needed().
1477 WARN_ON_ONCE(!vma_is_anonymous(vma
));
1478 rss
[mm_counter(page
)]--;
1479 if (is_device_private_entry(entry
))
1480 page_remove_rmap(page
, vma
, false);
1482 } else if (!non_swap_entry(entry
)) {
1483 /* Genuine swap entry, hence a private anon page */
1484 if (!should_zap_cows(details
))
1487 if (unlikely(!free_swap_and_cache(entry
)))
1488 print_bad_pte(vma
, addr
, ptent
, NULL
);
1489 } else if (is_migration_entry(entry
)) {
1490 page
= pfn_swap_entry_to_page(entry
);
1491 if (!should_zap_page(details
, page
))
1493 rss
[mm_counter(page
)]--;
1494 } else if (pte_marker_entry_uffd_wp(entry
)) {
1495 /* Only drop the uffd-wp marker if explicitly requested */
1496 if (!zap_drop_file_uffd_wp(details
))
1498 } else if (is_hwpoison_entry(entry
) ||
1499 is_swapin_error_entry(entry
)) {
1500 if (!should_zap_cows(details
))
1503 /* We should have covered all the swap entry types */
1506 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1507 zap_install_uffd_wp_if_needed(vma
, addr
, pte
, details
, ptent
);
1508 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1510 add_mm_rss_vec(mm
, rss
);
1511 arch_leave_lazy_mmu_mode();
1513 /* Do the actual TLB flush before dropping ptl */
1515 tlb_flush_mmu_tlbonly(tlb
);
1516 pte_unmap_unlock(start_pte
, ptl
);
1519 * If we forced a TLB flush (either due to running out of
1520 * batch buffers or because we needed to flush dirty TLB
1521 * entries before releasing the ptl), free the batched
1522 * memory too. Restart if we didn't do everything.
1537 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1538 struct vm_area_struct
*vma
, pud_t
*pud
,
1539 unsigned long addr
, unsigned long end
,
1540 struct zap_details
*details
)
1545 pmd
= pmd_offset(pud
, addr
);
1547 next
= pmd_addr_end(addr
, end
);
1548 if (is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1549 if (next
- addr
!= HPAGE_PMD_SIZE
)
1550 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
1551 else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1554 } else if (details
&& details
->single_folio
&&
1555 folio_test_pmd_mappable(details
->single_folio
) &&
1556 next
- addr
== HPAGE_PMD_SIZE
&& pmd_none(*pmd
)) {
1557 spinlock_t
*ptl
= pmd_lock(tlb
->mm
, pmd
);
1559 * Take and drop THP pmd lock so that we cannot return
1560 * prematurely, while zap_huge_pmd() has cleared *pmd,
1561 * but not yet decremented compound_mapcount().
1567 * Here there can be other concurrent MADV_DONTNEED or
1568 * trans huge page faults running, and if the pmd is
1569 * none or trans huge it can change under us. This is
1570 * because MADV_DONTNEED holds the mmap_lock in read
1573 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1575 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1578 } while (pmd
++, addr
= next
, addr
!= end
);
1583 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1584 struct vm_area_struct
*vma
, p4d_t
*p4d
,
1585 unsigned long addr
, unsigned long end
,
1586 struct zap_details
*details
)
1591 pud
= pud_offset(p4d
, addr
);
1593 next
= pud_addr_end(addr
, end
);
1594 if (pud_trans_huge(*pud
) || pud_devmap(*pud
)) {
1595 if (next
- addr
!= HPAGE_PUD_SIZE
) {
1596 mmap_assert_locked(tlb
->mm
);
1597 split_huge_pud(vma
, pud
, addr
);
1598 } else if (zap_huge_pud(tlb
, vma
, pud
, addr
))
1602 if (pud_none_or_clear_bad(pud
))
1604 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1607 } while (pud
++, addr
= next
, addr
!= end
);
1612 static inline unsigned long zap_p4d_range(struct mmu_gather
*tlb
,
1613 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1614 unsigned long addr
, unsigned long end
,
1615 struct zap_details
*details
)
1620 p4d
= p4d_offset(pgd
, addr
);
1622 next
= p4d_addr_end(addr
, end
);
1623 if (p4d_none_or_clear_bad(p4d
))
1625 next
= zap_pud_range(tlb
, vma
, p4d
, addr
, next
, details
);
1626 } while (p4d
++, addr
= next
, addr
!= end
);
1631 void unmap_page_range(struct mmu_gather
*tlb
,
1632 struct vm_area_struct
*vma
,
1633 unsigned long addr
, unsigned long end
,
1634 struct zap_details
*details
)
1639 BUG_ON(addr
>= end
);
1640 tlb_start_vma(tlb
, vma
);
1641 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1643 next
= pgd_addr_end(addr
, end
);
1644 if (pgd_none_or_clear_bad(pgd
))
1646 next
= zap_p4d_range(tlb
, vma
, pgd
, addr
, next
, details
);
1647 } while (pgd
++, addr
= next
, addr
!= end
);
1648 tlb_end_vma(tlb
, vma
);
1652 static void unmap_single_vma(struct mmu_gather
*tlb
,
1653 struct vm_area_struct
*vma
, unsigned long start_addr
,
1654 unsigned long end_addr
,
1655 struct zap_details
*details
)
1657 unsigned long start
= max(vma
->vm_start
, start_addr
);
1660 if (start
>= vma
->vm_end
)
1662 end
= min(vma
->vm_end
, end_addr
);
1663 if (end
<= vma
->vm_start
)
1667 uprobe_munmap(vma
, start
, end
);
1669 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1670 untrack_pfn(vma
, 0, 0);
1673 if (unlikely(is_vm_hugetlb_page(vma
))) {
1675 * It is undesirable to test vma->vm_file as it
1676 * should be non-null for valid hugetlb area.
1677 * However, vm_file will be NULL in the error
1678 * cleanup path of mmap_region. When
1679 * hugetlbfs ->mmap method fails,
1680 * mmap_region() nullifies vma->vm_file
1681 * before calling this function to clean up.
1682 * Since no pte has actually been setup, it is
1683 * safe to do nothing in this case.
1686 zap_flags_t zap_flags
= details
?
1687 details
->zap_flags
: 0;
1688 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
1689 __unmap_hugepage_range_final(tlb
, vma
, start
, end
,
1691 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
1694 unmap_page_range(tlb
, vma
, start
, end
, details
);
1699 * unmap_vmas - unmap a range of memory covered by a list of vma's
1700 * @tlb: address of the caller's struct mmu_gather
1701 * @vma: the starting vma
1702 * @start_addr: virtual address at which to start unmapping
1703 * @end_addr: virtual address at which to end unmapping
1705 * Unmap all pages in the vma list.
1707 * Only addresses between `start' and `end' will be unmapped.
1709 * The VMA list must be sorted in ascending virtual address order.
1711 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1712 * range after unmap_vmas() returns. So the only responsibility here is to
1713 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1714 * drops the lock and schedules.
1716 void unmap_vmas(struct mmu_gather
*tlb
,
1717 struct vm_area_struct
*vma
, unsigned long start_addr
,
1718 unsigned long end_addr
)
1720 struct mmu_notifier_range range
;
1721 struct zap_details details
= {
1722 .zap_flags
= ZAP_FLAG_DROP_MARKER
,
1723 /* Careful - we need to zap private pages too! */
1727 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
, vma
->vm_mm
,
1728 start_addr
, end_addr
);
1729 mmu_notifier_invalidate_range_start(&range
);
1730 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1731 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, &details
);
1732 mmu_notifier_invalidate_range_end(&range
);
1736 * zap_page_range - remove user pages in a given range
1737 * @vma: vm_area_struct holding the applicable pages
1738 * @start: starting address of pages to zap
1739 * @size: number of bytes to zap
1741 * Caller must protect the VMA list
1743 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1746 struct mmu_notifier_range range
;
1747 struct mmu_gather tlb
;
1750 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1751 start
, start
+ size
);
1752 tlb_gather_mmu(&tlb
, vma
->vm_mm
);
1753 update_hiwater_rss(vma
->vm_mm
);
1754 mmu_notifier_invalidate_range_start(&range
);
1755 for ( ; vma
&& vma
->vm_start
< range
.end
; vma
= vma
->vm_next
)
1756 unmap_single_vma(&tlb
, vma
, start
, range
.end
, NULL
);
1757 mmu_notifier_invalidate_range_end(&range
);
1758 tlb_finish_mmu(&tlb
);
1762 * zap_page_range_single - remove user pages in a given range
1763 * @vma: vm_area_struct holding the applicable pages
1764 * @address: starting address of pages to zap
1765 * @size: number of bytes to zap
1766 * @details: details of shared cache invalidation
1768 * The range must fit into one VMA.
1770 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1771 unsigned long size
, struct zap_details
*details
)
1773 struct mmu_notifier_range range
;
1774 struct mmu_gather tlb
;
1777 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1778 address
, address
+ size
);
1779 tlb_gather_mmu(&tlb
, vma
->vm_mm
);
1780 update_hiwater_rss(vma
->vm_mm
);
1781 mmu_notifier_invalidate_range_start(&range
);
1782 unmap_single_vma(&tlb
, vma
, address
, range
.end
, details
);
1783 mmu_notifier_invalidate_range_end(&range
);
1784 tlb_finish_mmu(&tlb
);
1788 * zap_vma_ptes - remove ptes mapping the vma
1789 * @vma: vm_area_struct holding ptes to be zapped
1790 * @address: starting address of pages to zap
1791 * @size: number of bytes to zap
1793 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1795 * The entire address range must be fully contained within the vma.
1798 void zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1801 if (!range_in_vma(vma
, address
, address
+ size
) ||
1802 !(vma
->vm_flags
& VM_PFNMAP
))
1805 zap_page_range_single(vma
, address
, size
, NULL
);
1807 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1809 static pmd_t
*walk_to_pmd(struct mm_struct
*mm
, unsigned long addr
)
1816 pgd
= pgd_offset(mm
, addr
);
1817 p4d
= p4d_alloc(mm
, pgd
, addr
);
1820 pud
= pud_alloc(mm
, p4d
, addr
);
1823 pmd
= pmd_alloc(mm
, pud
, addr
);
1827 VM_BUG_ON(pmd_trans_huge(*pmd
));
1831 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1834 pmd_t
*pmd
= walk_to_pmd(mm
, addr
);
1838 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1841 static int validate_page_before_insert(struct page
*page
)
1843 if (PageAnon(page
) || PageSlab(page
) || page_has_type(page
))
1845 flush_dcache_page(page
);
1849 static int insert_page_into_pte_locked(struct vm_area_struct
*vma
, pte_t
*pte
,
1850 unsigned long addr
, struct page
*page
, pgprot_t prot
)
1852 if (!pte_none(*pte
))
1854 /* Ok, finally just insert the thing.. */
1856 inc_mm_counter_fast(vma
->vm_mm
, mm_counter_file(page
));
1857 page_add_file_rmap(page
, vma
, false);
1858 set_pte_at(vma
->vm_mm
, addr
, pte
, mk_pte(page
, prot
));
1863 * This is the old fallback for page remapping.
1865 * For historical reasons, it only allows reserved pages. Only
1866 * old drivers should use this, and they needed to mark their
1867 * pages reserved for the old functions anyway.
1869 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1870 struct page
*page
, pgprot_t prot
)
1876 retval
= validate_page_before_insert(page
);
1880 pte
= get_locked_pte(vma
->vm_mm
, addr
, &ptl
);
1883 retval
= insert_page_into_pte_locked(vma
, pte
, addr
, page
, prot
);
1884 pte_unmap_unlock(pte
, ptl
);
1890 static int insert_page_in_batch_locked(struct vm_area_struct
*vma
, pte_t
*pte
,
1891 unsigned long addr
, struct page
*page
, pgprot_t prot
)
1895 if (!page_count(page
))
1897 err
= validate_page_before_insert(page
);
1900 return insert_page_into_pte_locked(vma
, pte
, addr
, page
, prot
);
1903 /* insert_pages() amortizes the cost of spinlock operations
1904 * when inserting pages in a loop. Arch *must* define pte_index.
1906 static int insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
1907 struct page
**pages
, unsigned long *num
, pgprot_t prot
)
1910 pte_t
*start_pte
, *pte
;
1911 spinlock_t
*pte_lock
;
1912 struct mm_struct
*const mm
= vma
->vm_mm
;
1913 unsigned long curr_page_idx
= 0;
1914 unsigned long remaining_pages_total
= *num
;
1915 unsigned long pages_to_write_in_pmd
;
1919 pmd
= walk_to_pmd(mm
, addr
);
1923 pages_to_write_in_pmd
= min_t(unsigned long,
1924 remaining_pages_total
, PTRS_PER_PTE
- pte_index(addr
));
1926 /* Allocate the PTE if necessary; takes PMD lock once only. */
1928 if (pte_alloc(mm
, pmd
))
1931 while (pages_to_write_in_pmd
) {
1933 const int batch_size
= min_t(int, pages_to_write_in_pmd
, 8);
1935 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &pte_lock
);
1936 for (pte
= start_pte
; pte_idx
< batch_size
; ++pte
, ++pte_idx
) {
1937 int err
= insert_page_in_batch_locked(vma
, pte
,
1938 addr
, pages
[curr_page_idx
], prot
);
1939 if (unlikely(err
)) {
1940 pte_unmap_unlock(start_pte
, pte_lock
);
1942 remaining_pages_total
-= pte_idx
;
1948 pte_unmap_unlock(start_pte
, pte_lock
);
1949 pages_to_write_in_pmd
-= batch_size
;
1950 remaining_pages_total
-= batch_size
;
1952 if (remaining_pages_total
)
1956 *num
= remaining_pages_total
;
1959 #endif /* ifdef pte_index */
1962 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1963 * @vma: user vma to map to
1964 * @addr: target start user address of these pages
1965 * @pages: source kernel pages
1966 * @num: in: number of pages to map. out: number of pages that were *not*
1967 * mapped. (0 means all pages were successfully mapped).
1969 * Preferred over vm_insert_page() when inserting multiple pages.
1971 * In case of error, we may have mapped a subset of the provided
1972 * pages. It is the caller's responsibility to account for this case.
1974 * The same restrictions apply as in vm_insert_page().
1976 int vm_insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
1977 struct page
**pages
, unsigned long *num
)
1980 const unsigned long end_addr
= addr
+ (*num
* PAGE_SIZE
) - 1;
1982 if (addr
< vma
->vm_start
|| end_addr
>= vma
->vm_end
)
1984 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1985 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
1986 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1987 vma
->vm_flags
|= VM_MIXEDMAP
;
1989 /* Defer page refcount checking till we're about to map that page. */
1990 return insert_pages(vma
, addr
, pages
, num
, vma
->vm_page_prot
);
1992 unsigned long idx
= 0, pgcount
= *num
;
1995 for (; idx
< pgcount
; ++idx
) {
1996 err
= vm_insert_page(vma
, addr
+ (PAGE_SIZE
* idx
), pages
[idx
]);
2000 *num
= pgcount
- idx
;
2002 #endif /* ifdef pte_index */
2004 EXPORT_SYMBOL(vm_insert_pages
);
2007 * vm_insert_page - insert single page into user vma
2008 * @vma: user vma to map to
2009 * @addr: target user address of this page
2010 * @page: source kernel page
2012 * This allows drivers to insert individual pages they've allocated
2015 * The page has to be a nice clean _individual_ kernel allocation.
2016 * If you allocate a compound page, you need to have marked it as
2017 * such (__GFP_COMP), or manually just split the page up yourself
2018 * (see split_page()).
2020 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2021 * took an arbitrary page protection parameter. This doesn't allow
2022 * that. Your vma protection will have to be set up correctly, which
2023 * means that if you want a shared writable mapping, you'd better
2024 * ask for a shared writable mapping!
2026 * The page does not need to be reserved.
2028 * Usually this function is called from f_op->mmap() handler
2029 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
2030 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2031 * function from other places, for example from page-fault handler.
2033 * Return: %0 on success, negative error code otherwise.
2035 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2038 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2040 if (!page_count(page
))
2042 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
2043 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
2044 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
2045 vma
->vm_flags
|= VM_MIXEDMAP
;
2047 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2049 EXPORT_SYMBOL(vm_insert_page
);
2052 * __vm_map_pages - maps range of kernel pages into user vma
2053 * @vma: user vma to map to
2054 * @pages: pointer to array of source kernel pages
2055 * @num: number of pages in page array
2056 * @offset: user's requested vm_pgoff
2058 * This allows drivers to map range of kernel pages into a user vma.
2060 * Return: 0 on success and error code otherwise.
2062 static int __vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
2063 unsigned long num
, unsigned long offset
)
2065 unsigned long count
= vma_pages(vma
);
2066 unsigned long uaddr
= vma
->vm_start
;
2069 /* Fail if the user requested offset is beyond the end of the object */
2073 /* Fail if the user requested size exceeds available object size */
2074 if (count
> num
- offset
)
2077 for (i
= 0; i
< count
; i
++) {
2078 ret
= vm_insert_page(vma
, uaddr
, pages
[offset
+ i
]);
2088 * vm_map_pages - maps range of kernel pages starts with non zero offset
2089 * @vma: user vma to map to
2090 * @pages: pointer to array of source kernel pages
2091 * @num: number of pages in page array
2093 * Maps an object consisting of @num pages, catering for the user's
2094 * requested vm_pgoff
2096 * If we fail to insert any page into the vma, the function will return
2097 * immediately leaving any previously inserted pages present. Callers
2098 * from the mmap handler may immediately return the error as their caller
2099 * will destroy the vma, removing any successfully inserted pages. Other
2100 * callers should make their own arrangements for calling unmap_region().
2102 * Context: Process context. Called by mmap handlers.
2103 * Return: 0 on success and error code otherwise.
2105 int vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
2108 return __vm_map_pages(vma
, pages
, num
, vma
->vm_pgoff
);
2110 EXPORT_SYMBOL(vm_map_pages
);
2113 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2114 * @vma: user vma to map to
2115 * @pages: pointer to array of source kernel pages
2116 * @num: number of pages in page array
2118 * Similar to vm_map_pages(), except that it explicitly sets the offset
2119 * to 0. This function is intended for the drivers that did not consider
2122 * Context: Process context. Called by mmap handlers.
2123 * Return: 0 on success and error code otherwise.
2125 int vm_map_pages_zero(struct vm_area_struct
*vma
, struct page
**pages
,
2128 return __vm_map_pages(vma
, pages
, num
, 0);
2130 EXPORT_SYMBOL(vm_map_pages_zero
);
2132 static vm_fault_t
insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2133 pfn_t pfn
, pgprot_t prot
, bool mkwrite
)
2135 struct mm_struct
*mm
= vma
->vm_mm
;
2139 pte
= get_locked_pte(mm
, addr
, &ptl
);
2141 return VM_FAULT_OOM
;
2142 if (!pte_none(*pte
)) {
2145 * For read faults on private mappings the PFN passed
2146 * in may not match the PFN we have mapped if the
2147 * mapped PFN is a writeable COW page. In the mkwrite
2148 * case we are creating a writable PTE for a shared
2149 * mapping and we expect the PFNs to match. If they
2150 * don't match, we are likely racing with block
2151 * allocation and mapping invalidation so just skip the
2154 if (pte_pfn(*pte
) != pfn_t_to_pfn(pfn
)) {
2155 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte
)));
2158 entry
= pte_mkyoung(*pte
);
2159 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2160 if (ptep_set_access_flags(vma
, addr
, pte
, entry
, 1))
2161 update_mmu_cache(vma
, addr
, pte
);
2166 /* Ok, finally just insert the thing.. */
2167 if (pfn_t_devmap(pfn
))
2168 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
2170 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
2173 entry
= pte_mkyoung(entry
);
2174 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2177 set_pte_at(mm
, addr
, pte
, entry
);
2178 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
2181 pte_unmap_unlock(pte
, ptl
);
2182 return VM_FAULT_NOPAGE
;
2186 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2187 * @vma: user vma to map to
2188 * @addr: target user address of this page
2189 * @pfn: source kernel pfn
2190 * @pgprot: pgprot flags for the inserted page
2192 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2193 * to override pgprot on a per-page basis.
2195 * This only makes sense for IO mappings, and it makes no sense for
2196 * COW mappings. In general, using multiple vmas is preferable;
2197 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2200 * See vmf_insert_mixed_prot() for a discussion of the implication of using
2201 * a value of @pgprot different from that of @vma->vm_page_prot.
2203 * Context: Process context. May allocate using %GFP_KERNEL.
2204 * Return: vm_fault_t value.
2206 vm_fault_t
vmf_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
2207 unsigned long pfn
, pgprot_t pgprot
)
2210 * Technically, architectures with pte_special can avoid all these
2211 * restrictions (same for remap_pfn_range). However we would like
2212 * consistency in testing and feature parity among all, so we should
2213 * try to keep these invariants in place for everybody.
2215 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
2216 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
2217 (VM_PFNMAP
|VM_MIXEDMAP
));
2218 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
2219 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
2221 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2222 return VM_FAULT_SIGBUS
;
2224 if (!pfn_modify_allowed(pfn
, pgprot
))
2225 return VM_FAULT_SIGBUS
;
2227 track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
));
2229 return insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
,
2232 EXPORT_SYMBOL(vmf_insert_pfn_prot
);
2235 * vmf_insert_pfn - insert single pfn into user vma
2236 * @vma: user vma to map to
2237 * @addr: target user address of this page
2238 * @pfn: source kernel pfn
2240 * Similar to vm_insert_page, this allows drivers to insert individual pages
2241 * they've allocated into a user vma. Same comments apply.
2243 * This function should only be called from a vm_ops->fault handler, and
2244 * in that case the handler should return the result of this function.
2246 * vma cannot be a COW mapping.
2248 * As this is called only for pages that do not currently exist, we
2249 * do not need to flush old virtual caches or the TLB.
2251 * Context: Process context. May allocate using %GFP_KERNEL.
2252 * Return: vm_fault_t value.
2254 vm_fault_t
vmf_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2257 return vmf_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
2259 EXPORT_SYMBOL(vmf_insert_pfn
);
2261 static bool vm_mixed_ok(struct vm_area_struct
*vma
, pfn_t pfn
)
2263 /* these checks mirror the abort conditions in vm_normal_page */
2264 if (vma
->vm_flags
& VM_MIXEDMAP
)
2266 if (pfn_t_devmap(pfn
))
2268 if (pfn_t_special(pfn
))
2270 if (is_zero_pfn(pfn_t_to_pfn(pfn
)))
2275 static vm_fault_t
__vm_insert_mixed(struct vm_area_struct
*vma
,
2276 unsigned long addr
, pfn_t pfn
, pgprot_t pgprot
,
2281 BUG_ON(!vm_mixed_ok(vma
, pfn
));
2283 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2284 return VM_FAULT_SIGBUS
;
2286 track_pfn_insert(vma
, &pgprot
, pfn
);
2288 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn
), pgprot
))
2289 return VM_FAULT_SIGBUS
;
2292 * If we don't have pte special, then we have to use the pfn_valid()
2293 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2294 * refcount the page if pfn_valid is true (hence insert_page rather
2295 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2296 * without pte special, it would there be refcounted as a normal page.
2298 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
) &&
2299 !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
2303 * At this point we are committed to insert_page()
2304 * regardless of whether the caller specified flags that
2305 * result in pfn_t_has_page() == false.
2307 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
2308 err
= insert_page(vma
, addr
, page
, pgprot
);
2310 return insert_pfn(vma
, addr
, pfn
, pgprot
, mkwrite
);
2314 return VM_FAULT_OOM
;
2315 if (err
< 0 && err
!= -EBUSY
)
2316 return VM_FAULT_SIGBUS
;
2318 return VM_FAULT_NOPAGE
;
2322 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2323 * @vma: user vma to map to
2324 * @addr: target user address of this page
2325 * @pfn: source kernel pfn
2326 * @pgprot: pgprot flags for the inserted page
2328 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2329 * to override pgprot on a per-page basis.
2331 * Typically this function should be used by drivers to set caching- and
2332 * encryption bits different than those of @vma->vm_page_prot, because
2333 * 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()).
2339 * Also when new page-table entries are created, this is only done using the
2340 * fault() callback, and never using the value of vma->vm_page_prot,
2341 * except for page-table entries that point to anonymous pages as the result
2344 * Context: Process context. May allocate using %GFP_KERNEL.
2345 * Return: vm_fault_t value.
2347 vm_fault_t
vmf_insert_mixed_prot(struct vm_area_struct
*vma
, unsigned long addr
,
2348 pfn_t pfn
, pgprot_t pgprot
)
2350 return __vm_insert_mixed(vma
, addr
, pfn
, pgprot
, false);
2352 EXPORT_SYMBOL(vmf_insert_mixed_prot
);
2354 vm_fault_t
vmf_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2357 return __vm_insert_mixed(vma
, addr
, pfn
, vma
->vm_page_prot
, false);
2359 EXPORT_SYMBOL(vmf_insert_mixed
);
2362 * If the insertion of PTE failed because someone else already added a
2363 * different entry in the mean time, we treat that as success as we assume
2364 * the same entry was actually inserted.
2366 vm_fault_t
vmf_insert_mixed_mkwrite(struct vm_area_struct
*vma
,
2367 unsigned long addr
, pfn_t pfn
)
2369 return __vm_insert_mixed(vma
, addr
, pfn
, vma
->vm_page_prot
, true);
2371 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite
);
2374 * maps a range of physical memory into the requested pages. the old
2375 * mappings are removed. any references to nonexistent pages results
2376 * in null mappings (currently treated as "copy-on-access")
2378 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2379 unsigned long addr
, unsigned long end
,
2380 unsigned long pfn
, pgprot_t prot
)
2382 pte_t
*pte
, *mapped_pte
;
2386 mapped_pte
= pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2389 arch_enter_lazy_mmu_mode();
2391 BUG_ON(!pte_none(*pte
));
2392 if (!pfn_modify_allowed(pfn
, prot
)) {
2396 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2398 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2399 arch_leave_lazy_mmu_mode();
2400 pte_unmap_unlock(mapped_pte
, ptl
);
2404 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2405 unsigned long addr
, unsigned long end
,
2406 unsigned long pfn
, pgprot_t prot
)
2412 pfn
-= addr
>> PAGE_SHIFT
;
2413 pmd
= pmd_alloc(mm
, pud
, addr
);
2416 VM_BUG_ON(pmd_trans_huge(*pmd
));
2418 next
= pmd_addr_end(addr
, end
);
2419 err
= remap_pte_range(mm
, pmd
, addr
, next
,
2420 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2423 } while (pmd
++, addr
= next
, addr
!= end
);
2427 static inline int remap_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2428 unsigned long addr
, unsigned long end
,
2429 unsigned long pfn
, pgprot_t prot
)
2435 pfn
-= addr
>> PAGE_SHIFT
;
2436 pud
= pud_alloc(mm
, p4d
, addr
);
2440 next
= pud_addr_end(addr
, end
);
2441 err
= remap_pmd_range(mm
, pud
, addr
, next
,
2442 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2445 } while (pud
++, addr
= next
, addr
!= end
);
2449 static inline int remap_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2450 unsigned long addr
, unsigned long end
,
2451 unsigned long pfn
, pgprot_t prot
)
2457 pfn
-= addr
>> PAGE_SHIFT
;
2458 p4d
= p4d_alloc(mm
, pgd
, addr
);
2462 next
= p4d_addr_end(addr
, end
);
2463 err
= remap_pud_range(mm
, p4d
, addr
, next
,
2464 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2467 } while (p4d
++, addr
= next
, addr
!= end
);
2472 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2473 * must have pre-validated the caching bits of the pgprot_t.
2475 int remap_pfn_range_notrack(struct vm_area_struct
*vma
, unsigned long addr
,
2476 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2480 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2481 struct mm_struct
*mm
= vma
->vm_mm
;
2484 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr
)))
2488 * Physically remapped pages are special. Tell the
2489 * rest of the world about it:
2490 * VM_IO tells people not to look at these pages
2491 * (accesses can have side effects).
2492 * VM_PFNMAP tells the core MM that the base pages are just
2493 * raw PFN mappings, and do not have a "struct page" associated
2496 * Disable vma merging and expanding with mremap().
2498 * Omit vma from core dump, even when VM_IO turned off.
2500 * There's a horrible special case to handle copy-on-write
2501 * behaviour that some programs depend on. We mark the "original"
2502 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2503 * See vm_normal_page() for details.
2505 if (is_cow_mapping(vma
->vm_flags
)) {
2506 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2508 vma
->vm_pgoff
= pfn
;
2511 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
2513 BUG_ON(addr
>= end
);
2514 pfn
-= addr
>> PAGE_SHIFT
;
2515 pgd
= pgd_offset(mm
, addr
);
2516 flush_cache_range(vma
, addr
, end
);
2518 next
= pgd_addr_end(addr
, end
);
2519 err
= remap_p4d_range(mm
, pgd
, addr
, next
,
2520 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2523 } while (pgd
++, addr
= next
, addr
!= end
);
2529 * remap_pfn_range - remap kernel memory to userspace
2530 * @vma: user vma to map to
2531 * @addr: target page aligned user address to start at
2532 * @pfn: page frame number of kernel physical memory address
2533 * @size: size of mapping area
2534 * @prot: page protection flags for this mapping
2536 * Note: this is only safe if the mm semaphore is held when called.
2538 * Return: %0 on success, negative error code otherwise.
2540 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2541 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2545 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
2549 err
= remap_pfn_range_notrack(vma
, addr
, pfn
, size
, prot
);
2551 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
));
2554 EXPORT_SYMBOL(remap_pfn_range
);
2557 * vm_iomap_memory - remap memory to userspace
2558 * @vma: user vma to map to
2559 * @start: start of the physical memory to be mapped
2560 * @len: size of area
2562 * This is a simplified io_remap_pfn_range() for common driver use. The
2563 * driver just needs to give us the physical memory range to be mapped,
2564 * we'll figure out the rest from the vma information.
2566 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2567 * whatever write-combining details or similar.
2569 * Return: %0 on success, negative error code otherwise.
2571 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2573 unsigned long vm_len
, pfn
, pages
;
2575 /* Check that the physical memory area passed in looks valid */
2576 if (start
+ len
< start
)
2579 * You *really* shouldn't map things that aren't page-aligned,
2580 * but we've historically allowed it because IO memory might
2581 * just have smaller alignment.
2583 len
+= start
& ~PAGE_MASK
;
2584 pfn
= start
>> PAGE_SHIFT
;
2585 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2586 if (pfn
+ pages
< pfn
)
2589 /* We start the mapping 'vm_pgoff' pages into the area */
2590 if (vma
->vm_pgoff
> pages
)
2592 pfn
+= vma
->vm_pgoff
;
2593 pages
-= vma
->vm_pgoff
;
2595 /* Can we fit all of the mapping? */
2596 vm_len
= vma
->vm_end
- vma
->vm_start
;
2597 if (vm_len
>> PAGE_SHIFT
> pages
)
2600 /* Ok, let it rip */
2601 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2603 EXPORT_SYMBOL(vm_iomap_memory
);
2605 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2606 unsigned long addr
, unsigned long end
,
2607 pte_fn_t fn
, void *data
, bool create
,
2608 pgtbl_mod_mask
*mask
)
2610 pte_t
*pte
, *mapped_pte
;
2615 mapped_pte
= pte
= (mm
== &init_mm
) ?
2616 pte_alloc_kernel_track(pmd
, addr
, mask
) :
2617 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2621 mapped_pte
= pte
= (mm
== &init_mm
) ?
2622 pte_offset_kernel(pmd
, addr
) :
2623 pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
2626 BUG_ON(pmd_huge(*pmd
));
2628 arch_enter_lazy_mmu_mode();
2632 if (create
|| !pte_none(*pte
)) {
2633 err
= fn(pte
++, addr
, data
);
2637 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2639 *mask
|= PGTBL_PTE_MODIFIED
;
2641 arch_leave_lazy_mmu_mode();
2644 pte_unmap_unlock(mapped_pte
, ptl
);
2648 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2649 unsigned long addr
, unsigned long end
,
2650 pte_fn_t fn
, void *data
, bool create
,
2651 pgtbl_mod_mask
*mask
)
2657 BUG_ON(pud_huge(*pud
));
2660 pmd
= pmd_alloc_track(mm
, pud
, addr
, mask
);
2664 pmd
= pmd_offset(pud
, addr
);
2667 next
= pmd_addr_end(addr
, end
);
2668 if (pmd_none(*pmd
) && !create
)
2670 if (WARN_ON_ONCE(pmd_leaf(*pmd
)))
2672 if (!pmd_none(*pmd
) && WARN_ON_ONCE(pmd_bad(*pmd
))) {
2677 err
= apply_to_pte_range(mm
, pmd
, addr
, next
,
2678 fn
, data
, create
, mask
);
2681 } while (pmd
++, addr
= next
, addr
!= end
);
2686 static int apply_to_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2687 unsigned long addr
, unsigned long end
,
2688 pte_fn_t fn
, void *data
, bool create
,
2689 pgtbl_mod_mask
*mask
)
2696 pud
= pud_alloc_track(mm
, p4d
, addr
, mask
);
2700 pud
= pud_offset(p4d
, addr
);
2703 next
= pud_addr_end(addr
, end
);
2704 if (pud_none(*pud
) && !create
)
2706 if (WARN_ON_ONCE(pud_leaf(*pud
)))
2708 if (!pud_none(*pud
) && WARN_ON_ONCE(pud_bad(*pud
))) {
2713 err
= apply_to_pmd_range(mm
, pud
, addr
, next
,
2714 fn
, data
, create
, mask
);
2717 } while (pud
++, addr
= next
, addr
!= end
);
2722 static int apply_to_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2723 unsigned long addr
, unsigned long end
,
2724 pte_fn_t fn
, void *data
, bool create
,
2725 pgtbl_mod_mask
*mask
)
2732 p4d
= p4d_alloc_track(mm
, pgd
, addr
, mask
);
2736 p4d
= p4d_offset(pgd
, addr
);
2739 next
= p4d_addr_end(addr
, end
);
2740 if (p4d_none(*p4d
) && !create
)
2742 if (WARN_ON_ONCE(p4d_leaf(*p4d
)))
2744 if (!p4d_none(*p4d
) && WARN_ON_ONCE(p4d_bad(*p4d
))) {
2749 err
= apply_to_pud_range(mm
, p4d
, addr
, next
,
2750 fn
, data
, create
, mask
);
2753 } while (p4d
++, addr
= next
, addr
!= end
);
2758 static int __apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2759 unsigned long size
, pte_fn_t fn
,
2760 void *data
, bool create
)
2763 unsigned long start
= addr
, next
;
2764 unsigned long end
= addr
+ size
;
2765 pgtbl_mod_mask mask
= 0;
2768 if (WARN_ON(addr
>= end
))
2771 pgd
= pgd_offset(mm
, addr
);
2773 next
= pgd_addr_end(addr
, end
);
2774 if (pgd_none(*pgd
) && !create
)
2776 if (WARN_ON_ONCE(pgd_leaf(*pgd
)))
2778 if (!pgd_none(*pgd
) && WARN_ON_ONCE(pgd_bad(*pgd
))) {
2783 err
= apply_to_p4d_range(mm
, pgd
, addr
, next
,
2784 fn
, data
, create
, &mask
);
2787 } while (pgd
++, addr
= next
, addr
!= end
);
2789 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
2790 arch_sync_kernel_mappings(start
, start
+ size
);
2796 * Scan a region of virtual memory, filling in page tables as necessary
2797 * and calling a provided function on each leaf page table.
2799 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2800 unsigned long size
, pte_fn_t fn
, void *data
)
2802 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, true);
2804 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2807 * Scan a region of virtual memory, calling a provided function on
2808 * each leaf page table where it exists.
2810 * Unlike apply_to_page_range, this does _not_ fill in page tables
2811 * where they are absent.
2813 int apply_to_existing_page_range(struct mm_struct
*mm
, unsigned long addr
,
2814 unsigned long size
, pte_fn_t fn
, void *data
)
2816 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, false);
2818 EXPORT_SYMBOL_GPL(apply_to_existing_page_range
);
2821 * handle_pte_fault chooses page fault handler according to an entry which was
2822 * read non-atomically. Before making any commitment, on those architectures
2823 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2824 * parts, do_swap_page must check under lock before unmapping the pte and
2825 * proceeding (but do_wp_page is only called after already making such a check;
2826 * and do_anonymous_page can safely check later on).
2828 static inline int pte_unmap_same(struct vm_fault
*vmf
)
2831 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2832 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2833 spinlock_t
*ptl
= pte_lockptr(vmf
->vma
->vm_mm
, vmf
->pmd
);
2835 same
= pte_same(*vmf
->pte
, vmf
->orig_pte
);
2839 pte_unmap(vmf
->pte
);
2844 static inline bool __wp_page_copy_user(struct page
*dst
, struct page
*src
,
2845 struct vm_fault
*vmf
)
2850 bool locked
= false;
2851 struct vm_area_struct
*vma
= vmf
->vma
;
2852 struct mm_struct
*mm
= vma
->vm_mm
;
2853 unsigned long addr
= vmf
->address
;
2856 copy_user_highpage(dst
, src
, addr
, vma
);
2861 * If the source page was a PFN mapping, we don't have
2862 * a "struct page" for it. We do a best-effort copy by
2863 * just copying from the original user address. If that
2864 * fails, we just zero-fill it. Live with it.
2866 kaddr
= kmap_atomic(dst
);
2867 uaddr
= (void __user
*)(addr
& PAGE_MASK
);
2870 * On architectures with software "accessed" bits, we would
2871 * take a double page fault, so mark it accessed here.
2873 if (arch_faults_on_old_pte() && !pte_young(vmf
->orig_pte
)) {
2876 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2878 if (!likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2880 * Other thread has already handled the fault
2881 * and update local tlb only
2883 update_mmu_tlb(vma
, addr
, vmf
->pte
);
2888 entry
= pte_mkyoung(vmf
->orig_pte
);
2889 if (ptep_set_access_flags(vma
, addr
, vmf
->pte
, entry
, 0))
2890 update_mmu_cache(vma
, addr
, vmf
->pte
);
2894 * This really shouldn't fail, because the page is there
2895 * in the page tables. But it might just be unreadable,
2896 * in which case we just give up and fill the result with
2899 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
2903 /* Re-validate under PTL if the page is still mapped */
2904 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2906 if (!likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2907 /* The PTE changed under us, update local tlb */
2908 update_mmu_tlb(vma
, addr
, vmf
->pte
);
2914 * The same page can be mapped back since last copy attempt.
2915 * Try to copy again under PTL.
2917 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
2919 * Give a warn in case there can be some obscure
2932 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2933 kunmap_atomic(kaddr
);
2934 flush_dcache_page(dst
);
2939 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
2941 struct file
*vm_file
= vma
->vm_file
;
2944 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
2947 * Special mappings (e.g. VDSO) do not have any file so fake
2948 * a default GFP_KERNEL for them.
2954 * Notify the address space that the page is about to become writable so that
2955 * it can prohibit this or wait for the page to get into an appropriate state.
2957 * We do this without the lock held, so that it can sleep if it needs to.
2959 static vm_fault_t
do_page_mkwrite(struct vm_fault
*vmf
)
2962 struct page
*page
= vmf
->page
;
2963 unsigned int old_flags
= vmf
->flags
;
2965 vmf
->flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2967 if (vmf
->vma
->vm_file
&&
2968 IS_SWAPFILE(vmf
->vma
->vm_file
->f_mapping
->host
))
2969 return VM_FAULT_SIGBUS
;
2971 ret
= vmf
->vma
->vm_ops
->page_mkwrite(vmf
);
2972 /* Restore original flags so that caller is not surprised */
2973 vmf
->flags
= old_flags
;
2974 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2976 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
2978 if (!page
->mapping
) {
2980 return 0; /* retry */
2982 ret
|= VM_FAULT_LOCKED
;
2984 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2989 * Handle dirtying of a page in shared file mapping on a write fault.
2991 * The function expects the page to be locked and unlocks it.
2993 static vm_fault_t
fault_dirty_shared_page(struct vm_fault
*vmf
)
2995 struct vm_area_struct
*vma
= vmf
->vma
;
2996 struct address_space
*mapping
;
2997 struct page
*page
= vmf
->page
;
2999 bool page_mkwrite
= vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
;
3001 dirtied
= set_page_dirty(page
);
3002 VM_BUG_ON_PAGE(PageAnon(page
), page
);
3004 * Take a local copy of the address_space - page.mapping may be zeroed
3005 * by truncate after unlock_page(). The address_space itself remains
3006 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3007 * release semantics to prevent the compiler from undoing this copying.
3009 mapping
= page_rmapping(page
);
3013 file_update_time(vma
->vm_file
);
3016 * Throttle page dirtying rate down to writeback speed.
3018 * mapping may be NULL here because some device drivers do not
3019 * set page.mapping but still dirty their pages
3021 * Drop the mmap_lock before waiting on IO, if we can. The file
3022 * is pinning the mapping, as per above.
3024 if ((dirtied
|| page_mkwrite
) && mapping
) {
3027 fpin
= maybe_unlock_mmap_for_io(vmf
, NULL
);
3028 balance_dirty_pages_ratelimited(mapping
);
3031 return VM_FAULT_COMPLETED
;
3039 * Handle write page faults for pages that can be reused in the current vma
3041 * This can happen either due to the mapping being with the VM_SHARED flag,
3042 * or due to us being the last reference standing to the page. In either
3043 * case, all we need to do here is to mark the page as writable and update
3044 * any related book-keeping.
3046 static inline void wp_page_reuse(struct vm_fault
*vmf
)
3047 __releases(vmf
->ptl
)
3049 struct vm_area_struct
*vma
= vmf
->vma
;
3050 struct page
*page
= vmf
->page
;
3053 VM_BUG_ON(!(vmf
->flags
& FAULT_FLAG_WRITE
));
3054 VM_BUG_ON(page
&& PageAnon(page
) && !PageAnonExclusive(page
));
3057 * Clear the pages cpupid information as the existing
3058 * information potentially belongs to a now completely
3059 * unrelated process.
3062 page_cpupid_xchg_last(page
, (1 << LAST_CPUPID_SHIFT
) - 1);
3064 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
3065 entry
= pte_mkyoung(vmf
->orig_pte
);
3066 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3067 if (ptep_set_access_flags(vma
, vmf
->address
, vmf
->pte
, entry
, 1))
3068 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3069 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3070 count_vm_event(PGREUSE
);
3074 * Handle the case of a page which we actually need to copy to a new page,
3075 * either due to COW or unsharing.
3077 * Called with mmap_lock locked and the old page referenced, but
3078 * without the ptl held.
3080 * High level logic flow:
3082 * - Allocate a page, copy the content of the old page to the new one.
3083 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
3084 * - Take the PTL. If the pte changed, bail out and release the allocated page
3085 * - If the pte is still the way we remember it, update the page table and all
3086 * relevant references. This includes dropping the reference the page-table
3087 * held to the old page, as well as updating the rmap.
3088 * - In any case, unlock the PTL and drop the reference we took to the old page.
3090 static vm_fault_t
wp_page_copy(struct vm_fault
*vmf
)
3092 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
3093 struct vm_area_struct
*vma
= vmf
->vma
;
3094 struct mm_struct
*mm
= vma
->vm_mm
;
3095 struct page
*old_page
= vmf
->page
;
3096 struct page
*new_page
= NULL
;
3098 int page_copied
= 0;
3099 struct mmu_notifier_range range
;
3101 delayacct_wpcopy_start();
3103 if (unlikely(anon_vma_prepare(vma
)))
3106 if (is_zero_pfn(pte_pfn(vmf
->orig_pte
))) {
3107 new_page
= alloc_zeroed_user_highpage_movable(vma
,
3112 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
3117 if (!__wp_page_copy_user(new_page
, old_page
, vmf
)) {
3119 * COW failed, if the fault was solved by other,
3120 * it's fine. If not, userspace would re-fault on
3121 * the same address and we will handle the fault
3122 * from the second attempt.
3128 delayacct_wpcopy_end();
3133 if (mem_cgroup_charge(page_folio(new_page
), mm
, GFP_KERNEL
))
3135 cgroup_throttle_swaprate(new_page
, GFP_KERNEL
);
3137 __SetPageUptodate(new_page
);
3139 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
,
3140 vmf
->address
& PAGE_MASK
,
3141 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
);
3142 mmu_notifier_invalidate_range_start(&range
);
3145 * Re-check the pte - we dropped the lock
3147 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, vmf
->address
, &vmf
->ptl
);
3148 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
3150 if (!PageAnon(old_page
)) {
3151 dec_mm_counter_fast(mm
,
3152 mm_counter_file(old_page
));
3153 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3156 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3158 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
3159 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
3160 entry
= pte_sw_mkyoung(entry
);
3161 if (unlikely(unshare
)) {
3162 if (pte_soft_dirty(vmf
->orig_pte
))
3163 entry
= pte_mksoft_dirty(entry
);
3164 if (pte_uffd_wp(vmf
->orig_pte
))
3165 entry
= pte_mkuffd_wp(entry
);
3167 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3171 * Clear the pte entry and flush it first, before updating the
3172 * pte with the new entry, to keep TLBs on different CPUs in
3173 * sync. This code used to set the new PTE then flush TLBs, but
3174 * that left a window where the new PTE could be loaded into
3175 * some TLBs while the old PTE remains in others.
3177 ptep_clear_flush_notify(vma
, vmf
->address
, vmf
->pte
);
3178 page_add_new_anon_rmap(new_page
, vma
, vmf
->address
);
3179 lru_cache_add_inactive_or_unevictable(new_page
, vma
);
3181 * We call the notify macro here because, when using secondary
3182 * mmu page tables (such as kvm shadow page tables), we want the
3183 * new page to be mapped directly into the secondary page table.
3185 BUG_ON(unshare
&& pte_write(entry
));
3186 set_pte_at_notify(mm
, vmf
->address
, vmf
->pte
, entry
);
3187 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3190 * Only after switching the pte to the new page may
3191 * we remove the mapcount here. Otherwise another
3192 * process may come and find the rmap count decremented
3193 * before the pte is switched to the new page, and
3194 * "reuse" the old page writing into it while our pte
3195 * here still points into it and can be read by other
3198 * The critical issue is to order this
3199 * page_remove_rmap with the ptp_clear_flush above.
3200 * Those stores are ordered by (if nothing else,)
3201 * the barrier present in the atomic_add_negative
3202 * in page_remove_rmap.
3204 * Then the TLB flush in ptep_clear_flush ensures that
3205 * no process can access the old page before the
3206 * decremented mapcount is visible. And the old page
3207 * cannot be reused until after the decremented
3208 * mapcount is visible. So transitively, TLBs to
3209 * old page will be flushed before it can be reused.
3211 page_remove_rmap(old_page
, vma
, false);
3214 /* Free the old page.. */
3215 new_page
= old_page
;
3218 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
3224 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3226 * No need to double call mmu_notifier->invalidate_range() callback as
3227 * the above ptep_clear_flush_notify() did already call it.
3229 mmu_notifier_invalidate_range_only_end(&range
);
3232 free_swap_cache(old_page
);
3236 delayacct_wpcopy_end();
3237 return (page_copied
&& !unshare
) ? VM_FAULT_WRITE
: 0;
3244 delayacct_wpcopy_end();
3245 return VM_FAULT_OOM
;
3249 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3250 * writeable once the page is prepared
3252 * @vmf: structure describing the fault
3254 * This function handles all that is needed to finish a write page fault in a
3255 * shared mapping due to PTE being read-only once the mapped page is prepared.
3256 * It handles locking of PTE and modifying it.
3258 * The function expects the page to be locked or other protection against
3259 * concurrent faults / writeback (such as DAX radix tree locks).
3261 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3262 * we acquired PTE lock.
3264 vm_fault_t
finish_mkwrite_fault(struct vm_fault
*vmf
)
3266 WARN_ON_ONCE(!(vmf
->vma
->vm_flags
& VM_SHARED
));
3267 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3270 * We might have raced with another page fault while we released the
3271 * pte_offset_map_lock.
3273 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
3274 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
3275 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3276 return VM_FAULT_NOPAGE
;
3283 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3286 static vm_fault_t
wp_pfn_shared(struct vm_fault
*vmf
)
3288 struct vm_area_struct
*vma
= vmf
->vma
;
3290 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
3293 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3294 vmf
->flags
|= FAULT_FLAG_MKWRITE
;
3295 ret
= vma
->vm_ops
->pfn_mkwrite(vmf
);
3296 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))
3298 return finish_mkwrite_fault(vmf
);
3301 return VM_FAULT_WRITE
;
3304 static vm_fault_t
wp_page_shared(struct vm_fault
*vmf
)
3305 __releases(vmf
->ptl
)
3307 struct vm_area_struct
*vma
= vmf
->vma
;
3308 vm_fault_t ret
= VM_FAULT_WRITE
;
3310 get_page(vmf
->page
);
3312 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
3315 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3316 tmp
= do_page_mkwrite(vmf
);
3317 if (unlikely(!tmp
|| (tmp
&
3318 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3319 put_page(vmf
->page
);
3322 tmp
= finish_mkwrite_fault(vmf
);
3323 if (unlikely(tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3324 unlock_page(vmf
->page
);
3325 put_page(vmf
->page
);
3330 lock_page(vmf
->page
);
3332 ret
|= fault_dirty_shared_page(vmf
);
3333 put_page(vmf
->page
);
3339 * This routine handles present pages, when
3340 * * users try to write to a shared page (FAULT_FLAG_WRITE)
3341 * * GUP wants to take a R/O pin on a possibly shared anonymous page
3342 * (FAULT_FLAG_UNSHARE)
3344 * It is done by copying the page to a new address and decrementing the
3345 * shared-page counter for the old page.
3347 * Note that this routine assumes that the protection checks have been
3348 * done by the caller (the low-level page fault routine in most cases).
3349 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've
3350 * done any necessary COW.
3352 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even
3353 * though the page will change only once the write actually happens. This
3354 * avoids a few races, and potentially makes it more efficient.
3356 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3357 * but allow concurrent faults), with pte both mapped and locked.
3358 * We return with mmap_lock still held, but pte unmapped and unlocked.
3360 static vm_fault_t
do_wp_page(struct vm_fault
*vmf
)
3361 __releases(vmf
->ptl
)
3363 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
3364 struct vm_area_struct
*vma
= vmf
->vma
;
3366 VM_BUG_ON(unshare
&& (vmf
->flags
& FAULT_FLAG_WRITE
));
3367 VM_BUG_ON(!unshare
&& !(vmf
->flags
& FAULT_FLAG_WRITE
));
3369 if (likely(!unshare
)) {
3370 if (userfaultfd_pte_wp(vma
, *vmf
->pte
)) {
3371 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3372 return handle_userfault(vmf
, VM_UFFD_WP
);
3376 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3377 * is flushed in this case before copying.
3379 if (unlikely(userfaultfd_wp(vmf
->vma
) &&
3380 mm_tlb_flush_pending(vmf
->vma
->vm_mm
)))
3381 flush_tlb_page(vmf
->vma
, vmf
->address
);
3384 vmf
->page
= vm_normal_page(vma
, vmf
->address
, vmf
->orig_pte
);
3386 if (unlikely(unshare
)) {
3387 /* No anonymous page -> nothing to do. */
3388 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3393 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3396 * We should not cow pages in a shared writeable mapping.
3397 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3399 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
3400 (VM_WRITE
|VM_SHARED
))
3401 return wp_pfn_shared(vmf
);
3403 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3404 return wp_page_copy(vmf
);
3408 * Take out anonymous pages first, anonymous shared vmas are
3409 * not dirty accountable.
3411 if (PageAnon(vmf
->page
)) {
3412 struct page
*page
= vmf
->page
;
3415 * If the page is exclusive to this process we must reuse the
3416 * page without further checks.
3418 if (PageAnonExclusive(page
))
3422 * We have to verify under page lock: these early checks are
3423 * just an optimization to avoid locking the page and freeing
3424 * the swapcache if there is little hope that we can reuse.
3426 * PageKsm() doesn't necessarily raise the page refcount.
3428 if (PageKsm(page
) || page_count(page
) > 3)
3432 * Note: We cannot easily detect+handle references from
3433 * remote LRU pagevecs or references to PageLRU() pages.
3436 if (page_count(page
) > 1 + PageSwapCache(page
))
3438 if (!trylock_page(page
))
3440 if (PageSwapCache(page
))
3441 try_to_free_swap(page
);
3442 if (PageKsm(page
) || page_count(page
) != 1) {
3447 * Ok, we've got the only page reference from our mapping
3448 * and the page is locked, it's dark out, and we're wearing
3449 * sunglasses. Hit it.
3451 page_move_anon_rmap(page
, vma
);
3454 if (unlikely(unshare
)) {
3455 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3459 return VM_FAULT_WRITE
;
3460 } else if (unshare
) {
3461 /* No anonymous page -> nothing to do. */
3462 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3464 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
3465 (VM_WRITE
|VM_SHARED
))) {
3466 return wp_page_shared(vmf
);
3470 * Ok, we need to copy. Oh, well..
3472 get_page(vmf
->page
);
3474 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3476 if (PageKsm(vmf
->page
))
3477 count_vm_event(COW_KSM
);
3479 return wp_page_copy(vmf
);
3482 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
3483 unsigned long start_addr
, unsigned long end_addr
,
3484 struct zap_details
*details
)
3486 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
3489 static inline void unmap_mapping_range_tree(struct rb_root_cached
*root
,
3490 pgoff_t first_index
,
3492 struct zap_details
*details
)
3494 struct vm_area_struct
*vma
;
3495 pgoff_t vba
, vea
, zba
, zea
;
3497 vma_interval_tree_foreach(vma
, root
, first_index
, last_index
) {
3498 vba
= vma
->vm_pgoff
;
3499 vea
= vba
+ vma_pages(vma
) - 1;
3500 zba
= max(first_index
, vba
);
3501 zea
= min(last_index
, vea
);
3503 unmap_mapping_range_vma(vma
,
3504 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
3505 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
3511 * unmap_mapping_folio() - Unmap single folio from processes.
3512 * @folio: The locked folio to be unmapped.
3514 * Unmap this folio from any userspace process which still has it mmaped.
3515 * Typically, for efficiency, the range of nearby pages has already been
3516 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3517 * truncation or invalidation holds the lock on a folio, it may find that
3518 * the page has been remapped again: and then uses unmap_mapping_folio()
3519 * to unmap it finally.
3521 void unmap_mapping_folio(struct folio
*folio
)
3523 struct address_space
*mapping
= folio
->mapping
;
3524 struct zap_details details
= { };
3525 pgoff_t first_index
;
3528 VM_BUG_ON(!folio_test_locked(folio
));
3530 first_index
= folio
->index
;
3531 last_index
= folio
->index
+ folio_nr_pages(folio
) - 1;
3533 details
.even_cows
= false;
3534 details
.single_folio
= folio
;
3535 details
.zap_flags
= ZAP_FLAG_DROP_MARKER
;
3537 i_mmap_lock_read(mapping
);
3538 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3539 unmap_mapping_range_tree(&mapping
->i_mmap
, first_index
,
3540 last_index
, &details
);
3541 i_mmap_unlock_read(mapping
);
3545 * unmap_mapping_pages() - Unmap pages from processes.
3546 * @mapping: The address space containing pages to be unmapped.
3547 * @start: Index of first page to be unmapped.
3548 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3549 * @even_cows: Whether to unmap even private COWed pages.
3551 * Unmap the pages in this address space from any userspace process which
3552 * has them mmaped. Generally, you want to remove COWed pages as well when
3553 * a file is being truncated, but not when invalidating pages from the page
3556 void unmap_mapping_pages(struct address_space
*mapping
, pgoff_t start
,
3557 pgoff_t nr
, bool even_cows
)
3559 struct zap_details details
= { };
3560 pgoff_t first_index
= start
;
3561 pgoff_t last_index
= start
+ nr
- 1;
3563 details
.even_cows
= even_cows
;
3564 if (last_index
< first_index
)
3565 last_index
= ULONG_MAX
;
3567 i_mmap_lock_read(mapping
);
3568 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3569 unmap_mapping_range_tree(&mapping
->i_mmap
, first_index
,
3570 last_index
, &details
);
3571 i_mmap_unlock_read(mapping
);
3573 EXPORT_SYMBOL_GPL(unmap_mapping_pages
);
3576 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3577 * address_space corresponding to the specified byte range in the underlying
3580 * @mapping: the address space containing mmaps to be unmapped.
3581 * @holebegin: byte in first page to unmap, relative to the start of
3582 * the underlying file. This will be rounded down to a PAGE_SIZE
3583 * boundary. Note that this is different from truncate_pagecache(), which
3584 * must keep the partial page. In contrast, we must get rid of
3586 * @holelen: size of prospective hole in bytes. This will be rounded
3587 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3589 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3590 * but 0 when invalidating pagecache, don't throw away private data.
3592 void unmap_mapping_range(struct address_space
*mapping
,
3593 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
3595 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
3596 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3598 /* Check for overflow. */
3599 if (sizeof(holelen
) > sizeof(hlen
)) {
3601 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3602 if (holeend
& ~(long long)ULONG_MAX
)
3603 hlen
= ULONG_MAX
- hba
+ 1;
3606 unmap_mapping_pages(mapping
, hba
, hlen
, even_cows
);
3608 EXPORT_SYMBOL(unmap_mapping_range
);
3611 * Restore a potential device exclusive pte to a working pte entry
3613 static vm_fault_t
remove_device_exclusive_entry(struct vm_fault
*vmf
)
3615 struct page
*page
= vmf
->page
;
3616 struct vm_area_struct
*vma
= vmf
->vma
;
3617 struct mmu_notifier_range range
;
3619 if (!lock_page_or_retry(page
, vma
->vm_mm
, vmf
->flags
))
3620 return VM_FAULT_RETRY
;
3621 mmu_notifier_range_init_owner(&range
, MMU_NOTIFY_EXCLUSIVE
, 0, vma
,
3622 vma
->vm_mm
, vmf
->address
& PAGE_MASK
,
3623 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
, NULL
);
3624 mmu_notifier_invalidate_range_start(&range
);
3626 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3628 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3629 restore_exclusive_pte(vma
, page
, vmf
->address
, vmf
->pte
);
3631 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3634 mmu_notifier_invalidate_range_end(&range
);
3638 static inline bool should_try_to_free_swap(struct page
*page
,
3639 struct vm_area_struct
*vma
,
3640 unsigned int fault_flags
)
3642 if (!PageSwapCache(page
))
3644 if (mem_cgroup_swap_full(page
) || (vma
->vm_flags
& VM_LOCKED
) ||
3648 * If we want to map a page that's in the swapcache writable, we
3649 * have to detect via the refcount if we're really the exclusive
3650 * user. Try freeing the swapcache to get rid of the swapcache
3651 * reference only in case it's likely that we'll be the exlusive user.
3653 return (fault_flags
& FAULT_FLAG_WRITE
) && !PageKsm(page
) &&
3654 page_count(page
) == 2;
3657 static vm_fault_t
pte_marker_clear(struct vm_fault
*vmf
)
3659 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
,
3660 vmf
->address
, &vmf
->ptl
);
3662 * Be careful so that we will only recover a special uffd-wp pte into a
3663 * none pte. Otherwise it means the pte could have changed, so retry.
3665 if (is_pte_marker(*vmf
->pte
))
3666 pte_clear(vmf
->vma
->vm_mm
, vmf
->address
, vmf
->pte
);
3667 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3672 * This is actually a page-missing access, but with uffd-wp special pte
3673 * installed. It means this pte was wr-protected before being unmapped.
3675 static vm_fault_t
pte_marker_handle_uffd_wp(struct vm_fault
*vmf
)
3678 * Just in case there're leftover special ptes even after the region
3679 * got unregistered - we can simply clear them. We can also do that
3680 * proactively when e.g. when we do UFFDIO_UNREGISTER upon some uffd-wp
3681 * ranges, but it should be more efficient to be done lazily here.
3683 if (unlikely(!userfaultfd_wp(vmf
->vma
) || vma_is_anonymous(vmf
->vma
)))
3684 return pte_marker_clear(vmf
);
3686 /* do_fault() can handle pte markers too like none pte */
3687 return do_fault(vmf
);
3690 static vm_fault_t
handle_pte_marker(struct vm_fault
*vmf
)
3692 swp_entry_t entry
= pte_to_swp_entry(vmf
->orig_pte
);
3693 unsigned long marker
= pte_marker_get(entry
);
3696 * PTE markers should always be with file-backed memories, and the
3697 * marker should never be empty. If anything weird happened, the best
3698 * thing to do is to kill the process along with its mm.
3700 if (WARN_ON_ONCE(vma_is_anonymous(vmf
->vma
) || !marker
))
3701 return VM_FAULT_SIGBUS
;
3703 if (pte_marker_entry_uffd_wp(entry
))
3704 return pte_marker_handle_uffd_wp(vmf
);
3706 /* This is an unknown pte marker */
3707 return VM_FAULT_SIGBUS
;
3711 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3712 * but allow concurrent faults), and pte mapped but not yet locked.
3713 * We return with pte unmapped and unlocked.
3715 * We return with the mmap_lock locked or unlocked in the same cases
3716 * as does filemap_fault().
3718 vm_fault_t
do_swap_page(struct vm_fault
*vmf
)
3720 struct vm_area_struct
*vma
= vmf
->vma
;
3721 struct page
*page
= NULL
, *swapcache
;
3722 struct swap_info_struct
*si
= NULL
;
3723 rmap_t rmap_flags
= RMAP_NONE
;
3724 bool exclusive
= false;
3729 void *shadow
= NULL
;
3731 if (!pte_unmap_same(vmf
))
3734 entry
= pte_to_swp_entry(vmf
->orig_pte
);
3735 if (unlikely(non_swap_entry(entry
))) {
3736 if (is_migration_entry(entry
)) {
3737 migration_entry_wait(vma
->vm_mm
, vmf
->pmd
,
3739 } else if (is_device_exclusive_entry(entry
)) {
3740 vmf
->page
= pfn_swap_entry_to_page(entry
);
3741 ret
= remove_device_exclusive_entry(vmf
);
3742 } else if (is_device_private_entry(entry
)) {
3743 vmf
->page
= pfn_swap_entry_to_page(entry
);
3744 ret
= vmf
->page
->pgmap
->ops
->migrate_to_ram(vmf
);
3745 } else if (is_hwpoison_entry(entry
)) {
3746 ret
= VM_FAULT_HWPOISON
;
3747 } else if (is_swapin_error_entry(entry
)) {
3748 ret
= VM_FAULT_SIGBUS
;
3749 } else if (is_pte_marker_entry(entry
)) {
3750 ret
= handle_pte_marker(vmf
);
3752 print_bad_pte(vma
, vmf
->address
, vmf
->orig_pte
, NULL
);
3753 ret
= VM_FAULT_SIGBUS
;
3758 /* Prevent swapoff from happening to us. */
3759 si
= get_swap_device(entry
);
3763 page
= lookup_swap_cache(entry
, vma
, vmf
->address
);
3767 if (data_race(si
->flags
& SWP_SYNCHRONOUS_IO
) &&
3768 __swap_count(entry
) == 1) {
3769 /* skip swapcache */
3770 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
3773 __SetPageLocked(page
);
3774 __SetPageSwapBacked(page
);
3776 if (mem_cgroup_swapin_charge_page(page
,
3777 vma
->vm_mm
, GFP_KERNEL
, entry
)) {
3781 mem_cgroup_swapin_uncharge_swap(entry
);
3783 shadow
= get_shadow_from_swap_cache(entry
);
3785 workingset_refault(page_folio(page
),
3788 lru_cache_add(page
);
3790 /* To provide entry to swap_readpage() */
3791 set_page_private(page
, entry
.val
);
3792 swap_readpage(page
, true, NULL
);
3793 set_page_private(page
, 0);
3796 page
= swapin_readahead(entry
, GFP_HIGHUSER_MOVABLE
,
3803 * Back out if somebody else faulted in this pte
3804 * while we released the pte lock.
3806 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3807 vmf
->address
, &vmf
->ptl
);
3808 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3813 /* Had to read the page from swap area: Major fault */
3814 ret
= VM_FAULT_MAJOR
;
3815 count_vm_event(PGMAJFAULT
);
3816 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
3817 } else if (PageHWPoison(page
)) {
3819 * hwpoisoned dirty swapcache pages are kept for killing
3820 * owner processes (which may be unknown at hwpoison time)
3822 ret
= VM_FAULT_HWPOISON
;
3826 locked
= lock_page_or_retry(page
, vma
->vm_mm
, vmf
->flags
);
3829 ret
|= VM_FAULT_RETRY
;
3835 * Make sure try_to_free_swap or swapoff did not release the
3836 * swapcache from under us. The page pin, and pte_same test
3837 * below, are not enough to exclude that. Even if it is still
3838 * swapcache, we need to check that the page's swap has not
3841 if (unlikely(!PageSwapCache(page
) ||
3842 page_private(page
) != entry
.val
))
3846 * KSM sometimes has to copy on read faults, for example, if
3847 * page->index of !PageKSM() pages would be nonlinear inside the
3848 * anon VMA -- PageKSM() is lost on actual swapout.
3850 page
= ksm_might_need_to_copy(page
, vma
, vmf
->address
);
3851 if (unlikely(!page
)) {
3858 * If we want to map a page that's in the swapcache writable, we
3859 * have to detect via the refcount if we're really the exclusive
3860 * owner. Try removing the extra reference from the local LRU
3861 * pagevecs if required.
3863 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && page
== swapcache
&&
3864 !PageKsm(page
) && !PageLRU(page
))
3868 cgroup_throttle_swaprate(page
, GFP_KERNEL
);
3871 * Back out if somebody else already faulted in this pte.
3873 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3875 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3878 if (unlikely(!PageUptodate(page
))) {
3879 ret
= VM_FAULT_SIGBUS
;
3884 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte
3885 * must never point at an anonymous page in the swapcache that is
3886 * PG_anon_exclusive. Sanity check that this holds and especially, that
3887 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity
3888 * check after taking the PT lock and making sure that nobody
3889 * concurrently faulted in this page and set PG_anon_exclusive.
3891 BUG_ON(!PageAnon(page
) && PageMappedToDisk(page
));
3892 BUG_ON(PageAnon(page
) && PageAnonExclusive(page
));
3895 * Check under PT lock (to protect against concurrent fork() sharing
3896 * the swap entry concurrently) for certainly exclusive pages.
3898 if (!PageKsm(page
)) {
3900 * Note that pte_swp_exclusive() == false for architectures
3901 * without __HAVE_ARCH_PTE_SWP_EXCLUSIVE.
3903 exclusive
= pte_swp_exclusive(vmf
->orig_pte
);
3904 if (page
!= swapcache
) {
3906 * We have a fresh page that is not exposed to the
3907 * swapcache -> certainly exclusive.
3910 } else if (exclusive
&& PageWriteback(page
) &&
3911 data_race(si
->flags
& SWP_STABLE_WRITES
)) {
3913 * This is tricky: not all swap backends support
3914 * concurrent page modifications while under writeback.
3916 * So if we stumble over such a page in the swapcache
3917 * we must not set the page exclusive, otherwise we can
3918 * map it writable without further checks and modify it
3919 * while still under writeback.
3921 * For these problematic swap backends, simply drop the
3922 * exclusive marker: this is perfectly fine as we start
3923 * writeback only if we fully unmapped the page and
3924 * there are no unexpected references on the page after
3925 * unmapping succeeded. After fully unmapped, no
3926 * further GUP references (FOLL_GET and FOLL_PIN) can
3927 * appear, so dropping the exclusive marker and mapping
3928 * it only R/O is fine.
3935 * Remove the swap entry and conditionally try to free up the swapcache.
3936 * We're already holding a reference on the page but haven't mapped it
3940 if (should_try_to_free_swap(page
, vma
, vmf
->flags
))
3941 try_to_free_swap(page
);
3943 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3944 dec_mm_counter_fast(vma
->vm_mm
, MM_SWAPENTS
);
3945 pte
= mk_pte(page
, vma
->vm_page_prot
);
3948 * Same logic as in do_wp_page(); however, optimize for pages that are
3949 * certainly not shared either because we just allocated them without
3950 * exposing them to the swapcache or because the swap entry indicates
3953 if (!PageKsm(page
) && (exclusive
|| page_count(page
) == 1)) {
3954 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
3955 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3956 vmf
->flags
&= ~FAULT_FLAG_WRITE
;
3957 ret
|= VM_FAULT_WRITE
;
3959 rmap_flags
|= RMAP_EXCLUSIVE
;
3961 flush_icache_page(vma
, page
);
3962 if (pte_swp_soft_dirty(vmf
->orig_pte
))
3963 pte
= pte_mksoft_dirty(pte
);
3964 if (pte_swp_uffd_wp(vmf
->orig_pte
)) {
3965 pte
= pte_mkuffd_wp(pte
);
3966 pte
= pte_wrprotect(pte
);
3968 vmf
->orig_pte
= pte
;
3970 /* ksm created a completely new copy */
3971 if (unlikely(page
!= swapcache
&& swapcache
)) {
3972 page_add_new_anon_rmap(page
, vma
, vmf
->address
);
3973 lru_cache_add_inactive_or_unevictable(page
, vma
);
3975 page_add_anon_rmap(page
, vma
, vmf
->address
, rmap_flags
);
3978 VM_BUG_ON(!PageAnon(page
) || (pte_write(pte
) && !PageAnonExclusive(page
)));
3979 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3980 arch_do_swap_page(vma
->vm_mm
, vma
, vmf
->address
, pte
, vmf
->orig_pte
);
3983 if (page
!= swapcache
&& swapcache
) {
3985 * Hold the lock to avoid the swap entry to be reused
3986 * until we take the PT lock for the pte_same() check
3987 * (to avoid false positives from pte_same). For
3988 * further safety release the lock after the swap_free
3989 * so that the swap count won't change under a
3990 * parallel locked swapcache.
3992 unlock_page(swapcache
);
3993 put_page(swapcache
);
3996 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
3997 ret
|= do_wp_page(vmf
);
3998 if (ret
& VM_FAULT_ERROR
)
3999 ret
&= VM_FAULT_ERROR
;
4003 /* No need to invalidate - it was non-present before */
4004 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
4006 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4009 put_swap_device(si
);
4012 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4017 if (page
!= swapcache
&& swapcache
) {
4018 unlock_page(swapcache
);
4019 put_page(swapcache
);
4022 put_swap_device(si
);
4027 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4028 * but allow concurrent faults), and pte mapped but not yet locked.
4029 * We return with mmap_lock still held, but pte unmapped and unlocked.
4031 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
)
4033 struct vm_area_struct
*vma
= vmf
->vma
;
4038 /* File mapping without ->vm_ops ? */
4039 if (vma
->vm_flags
& VM_SHARED
)
4040 return VM_FAULT_SIGBUS
;
4043 * Use pte_alloc() instead of pte_alloc_map(). We can't run
4044 * pte_offset_map() on pmds where a huge pmd might be created
4045 * from a different thread.
4047 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
4048 * parallel threads are excluded by other means.
4050 * Here we only have mmap_read_lock(mm).
4052 if (pte_alloc(vma
->vm_mm
, vmf
->pmd
))
4053 return VM_FAULT_OOM
;
4055 /* See comment in handle_pte_fault() */
4056 if (unlikely(pmd_trans_unstable(vmf
->pmd
)))
4059 /* Use the zero-page for reads */
4060 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
4061 !mm_forbids_zeropage(vma
->vm_mm
)) {
4062 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(vmf
->address
),
4063 vma
->vm_page_prot
));
4064 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4065 vmf
->address
, &vmf
->ptl
);
4066 if (!pte_none(*vmf
->pte
)) {
4067 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
4070 ret
= check_stable_address_space(vma
->vm_mm
);
4073 /* Deliver the page fault to userland, check inside PT lock */
4074 if (userfaultfd_missing(vma
)) {
4075 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4076 return handle_userfault(vmf
, VM_UFFD_MISSING
);
4081 /* Allocate our own private page. */
4082 if (unlikely(anon_vma_prepare(vma
)))
4084 page
= alloc_zeroed_user_highpage_movable(vma
, vmf
->address
);
4088 if (mem_cgroup_charge(page_folio(page
), vma
->vm_mm
, GFP_KERNEL
))
4090 cgroup_throttle_swaprate(page
, GFP_KERNEL
);
4093 * The memory barrier inside __SetPageUptodate makes sure that
4094 * preceding stores to the page contents become visible before
4095 * the set_pte_at() write.
4097 __SetPageUptodate(page
);
4099 entry
= mk_pte(page
, vma
->vm_page_prot
);
4100 entry
= pte_sw_mkyoung(entry
);
4101 if (vma
->vm_flags
& VM_WRITE
)
4102 entry
= pte_mkwrite(pte_mkdirty(entry
));
4104 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
4106 if (!pte_none(*vmf
->pte
)) {
4107 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
4111 ret
= check_stable_address_space(vma
->vm_mm
);
4115 /* Deliver the page fault to userland, check inside PT lock */
4116 if (userfaultfd_missing(vma
)) {
4117 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4119 return handle_userfault(vmf
, VM_UFFD_MISSING
);
4122 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
4123 page_add_new_anon_rmap(page
, vma
, vmf
->address
);
4124 lru_cache_add_inactive_or_unevictable(page
, vma
);
4126 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
4128 /* No need to invalidate - it was non-present before */
4129 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
4131 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4139 return VM_FAULT_OOM
;
4143 * The mmap_lock must have been held on entry, and may have been
4144 * released depending on flags and vma->vm_ops->fault() return value.
4145 * See filemap_fault() and __lock_page_retry().
4147 static vm_fault_t
__do_fault(struct vm_fault
*vmf
)
4149 struct vm_area_struct
*vma
= vmf
->vma
;
4153 * Preallocate pte before we take page_lock because this might lead to
4154 * deadlocks for memcg reclaim which waits for pages under writeback:
4156 * SetPageWriteback(A)
4162 * wait_on_page_writeback(A)
4163 * SetPageWriteback(B)
4165 * # flush A, B to clear the writeback
4167 if (pmd_none(*vmf
->pmd
) && !vmf
->prealloc_pte
) {
4168 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
4169 if (!vmf
->prealloc_pte
)
4170 return VM_FAULT_OOM
;
4173 ret
= vma
->vm_ops
->fault(vmf
);
4174 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
|
4175 VM_FAULT_DONE_COW
)))
4178 if (unlikely(PageHWPoison(vmf
->page
))) {
4179 struct page
*page
= vmf
->page
;
4180 vm_fault_t poisonret
= VM_FAULT_HWPOISON
;
4181 if (ret
& VM_FAULT_LOCKED
) {
4182 if (page_mapped(page
))
4183 unmap_mapping_pages(page_mapping(page
),
4184 page
->index
, 1, false);
4185 /* Retry if a clean page was removed from the cache. */
4186 if (invalidate_inode_page(page
))
4187 poisonret
= VM_FAULT_NOPAGE
;
4195 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
4196 lock_page(vmf
->page
);
4198 VM_BUG_ON_PAGE(!PageLocked(vmf
->page
), vmf
->page
);
4203 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4204 static void deposit_prealloc_pte(struct vm_fault
*vmf
)
4206 struct vm_area_struct
*vma
= vmf
->vma
;
4208 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
4210 * We are going to consume the prealloc table,
4211 * count that as nr_ptes.
4213 mm_inc_nr_ptes(vma
->vm_mm
);
4214 vmf
->prealloc_pte
= NULL
;
4217 vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
4219 struct vm_area_struct
*vma
= vmf
->vma
;
4220 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
4221 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
4224 vm_fault_t ret
= VM_FAULT_FALLBACK
;
4226 if (!transhuge_vma_suitable(vma
, haddr
))
4229 page
= compound_head(page
);
4230 if (compound_order(page
) != HPAGE_PMD_ORDER
)
4234 * Just backoff if any subpage of a THP is corrupted otherwise
4235 * the corrupted page may mapped by PMD silently to escape the
4236 * check. This kind of THP just can be PTE mapped. Access to
4237 * the corrupted subpage should trigger SIGBUS as expected.
4239 if (unlikely(PageHasHWPoisoned(page
)))
4243 * Archs like ppc64 need additional space to store information
4244 * related to pte entry. Use the preallocated table for that.
4246 if (arch_needs_pgtable_deposit() && !vmf
->prealloc_pte
) {
4247 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
4248 if (!vmf
->prealloc_pte
)
4249 return VM_FAULT_OOM
;
4252 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
4253 if (unlikely(!pmd_none(*vmf
->pmd
)))
4256 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
4257 flush_icache_page(vma
, page
+ i
);
4259 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
4261 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
4263 add_mm_counter(vma
->vm_mm
, mm_counter_file(page
), HPAGE_PMD_NR
);
4264 page_add_file_rmap(page
, vma
, true);
4267 * deposit and withdraw with pmd lock held
4269 if (arch_needs_pgtable_deposit())
4270 deposit_prealloc_pte(vmf
);
4272 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
4274 update_mmu_cache_pmd(vma
, haddr
, vmf
->pmd
);
4276 /* fault is handled */
4278 count_vm_event(THP_FILE_MAPPED
);
4280 spin_unlock(vmf
->ptl
);
4284 vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
4286 return VM_FAULT_FALLBACK
;
4290 void do_set_pte(struct vm_fault
*vmf
, struct page
*page
, unsigned long addr
)
4292 struct vm_area_struct
*vma
= vmf
->vma
;
4293 bool uffd_wp
= pte_marker_uffd_wp(vmf
->orig_pte
);
4294 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
4295 bool prefault
= vmf
->address
!= addr
;
4298 flush_icache_page(vma
, page
);
4299 entry
= mk_pte(page
, vma
->vm_page_prot
);
4301 if (prefault
&& arch_wants_old_prefaulted_pte())
4302 entry
= pte_mkold(entry
);
4304 entry
= pte_sw_mkyoung(entry
);
4307 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
4308 if (unlikely(uffd_wp
))
4309 entry
= pte_mkuffd_wp(pte_wrprotect(entry
));
4310 /* copy-on-write page */
4311 if (write
&& !(vma
->vm_flags
& VM_SHARED
)) {
4312 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
4313 page_add_new_anon_rmap(page
, vma
, addr
);
4314 lru_cache_add_inactive_or_unevictable(page
, vma
);
4316 inc_mm_counter_fast(vma
->vm_mm
, mm_counter_file(page
));
4317 page_add_file_rmap(page
, vma
, false);
4319 set_pte_at(vma
->vm_mm
, addr
, vmf
->pte
, entry
);
4322 static bool vmf_pte_changed(struct vm_fault
*vmf
)
4324 if (vmf
->flags
& FAULT_FLAG_ORIG_PTE_VALID
)
4325 return !pte_same(*vmf
->pte
, vmf
->orig_pte
);
4327 return !pte_none(*vmf
->pte
);
4331 * finish_fault - finish page fault once we have prepared the page to fault
4333 * @vmf: structure describing the fault
4335 * This function handles all that is needed to finish a page fault once the
4336 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4337 * given page, adds reverse page mapping, handles memcg charges and LRU
4340 * The function expects the page to be locked and on success it consumes a
4341 * reference of a page being mapped (for the PTE which maps it).
4343 * Return: %0 on success, %VM_FAULT_ code in case of error.
4345 vm_fault_t
finish_fault(struct vm_fault
*vmf
)
4347 struct vm_area_struct
*vma
= vmf
->vma
;
4351 /* Did we COW the page? */
4352 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
4353 page
= vmf
->cow_page
;
4358 * check even for read faults because we might have lost our CoWed
4361 if (!(vma
->vm_flags
& VM_SHARED
)) {
4362 ret
= check_stable_address_space(vma
->vm_mm
);
4367 if (pmd_none(*vmf
->pmd
)) {
4368 if (PageTransCompound(page
)) {
4369 ret
= do_set_pmd(vmf
, page
);
4370 if (ret
!= VM_FAULT_FALLBACK
)
4374 if (vmf
->prealloc_pte
)
4375 pmd_install(vma
->vm_mm
, vmf
->pmd
, &vmf
->prealloc_pte
);
4376 else if (unlikely(pte_alloc(vma
->vm_mm
, vmf
->pmd
)))
4377 return VM_FAULT_OOM
;
4381 * See comment in handle_pte_fault() for how this scenario happens, we
4382 * need to return NOPAGE so that we drop this page.
4384 if (pmd_devmap_trans_unstable(vmf
->pmd
))
4385 return VM_FAULT_NOPAGE
;
4387 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4388 vmf
->address
, &vmf
->ptl
);
4390 /* Re-check under ptl */
4391 if (likely(!vmf_pte_changed(vmf
)))
4392 do_set_pte(vmf
, page
, vmf
->address
);
4394 ret
= VM_FAULT_NOPAGE
;
4396 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
4397 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4401 static unsigned long fault_around_bytes __read_mostly
=
4402 rounddown_pow_of_two(65536);
4404 #ifdef CONFIG_DEBUG_FS
4405 static int fault_around_bytes_get(void *data
, u64
*val
)
4407 *val
= fault_around_bytes
;
4412 * fault_around_bytes must be rounded down to the nearest page order as it's
4413 * what do_fault_around() expects to see.
4415 static int fault_around_bytes_set(void *data
, u64 val
)
4417 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
4419 if (val
> PAGE_SIZE
)
4420 fault_around_bytes
= rounddown_pow_of_two(val
);
4422 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
4425 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops
,
4426 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
4428 static int __init
fault_around_debugfs(void)
4430 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL
, NULL
,
4431 &fault_around_bytes_fops
);
4434 late_initcall(fault_around_debugfs
);
4438 * do_fault_around() tries to map few pages around the fault address. The hope
4439 * is that the pages will be needed soon and this will lower the number of
4442 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4443 * not ready to be mapped: not up-to-date, locked, etc.
4445 * This function doesn't cross the VMA boundaries, in order to call map_pages()
4448 * fault_around_bytes defines how many bytes we'll try to map.
4449 * do_fault_around() expects it to be set to a power of two less than or equal
4452 * The virtual address of the area that we map is naturally aligned to
4453 * fault_around_bytes rounded down to the machine page size
4454 * (and therefore to page order). This way it's easier to guarantee
4455 * that we don't cross page table boundaries.
4457 static vm_fault_t
do_fault_around(struct vm_fault
*vmf
)
4459 unsigned long address
= vmf
->address
, nr_pages
, mask
;
4460 pgoff_t start_pgoff
= vmf
->pgoff
;
4464 nr_pages
= READ_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
4465 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
4467 address
= max(address
& mask
, vmf
->vma
->vm_start
);
4468 off
= ((vmf
->address
- address
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
4472 * end_pgoff is either the end of the page table, the end of
4473 * the vma or nr_pages from start_pgoff, depending what is nearest.
4475 end_pgoff
= start_pgoff
-
4476 ((address
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
4478 end_pgoff
= min3(end_pgoff
, vma_pages(vmf
->vma
) + vmf
->vma
->vm_pgoff
- 1,
4479 start_pgoff
+ nr_pages
- 1);
4481 if (pmd_none(*vmf
->pmd
)) {
4482 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
);
4483 if (!vmf
->prealloc_pte
)
4484 return VM_FAULT_OOM
;
4487 return vmf
->vma
->vm_ops
->map_pages(vmf
, start_pgoff
, end_pgoff
);
4490 /* Return true if we should do read fault-around, false otherwise */
4491 static inline bool should_fault_around(struct vm_fault
*vmf
)
4493 /* No ->map_pages? No way to fault around... */
4494 if (!vmf
->vma
->vm_ops
->map_pages
)
4497 if (uffd_disable_fault_around(vmf
->vma
))
4500 return fault_around_bytes
>> PAGE_SHIFT
> 1;
4503 static vm_fault_t
do_read_fault(struct vm_fault
*vmf
)
4508 * Let's call ->map_pages() first and use ->fault() as fallback
4509 * if page by the offset is not ready to be mapped (cold cache or
4512 if (should_fault_around(vmf
)) {
4513 ret
= do_fault_around(vmf
);
4518 ret
= __do_fault(vmf
);
4519 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4522 ret
|= finish_fault(vmf
);
4523 unlock_page(vmf
->page
);
4524 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4525 put_page(vmf
->page
);
4529 static vm_fault_t
do_cow_fault(struct vm_fault
*vmf
)
4531 struct vm_area_struct
*vma
= vmf
->vma
;
4534 if (unlikely(anon_vma_prepare(vma
)))
4535 return VM_FAULT_OOM
;
4537 vmf
->cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, vmf
->address
);
4539 return VM_FAULT_OOM
;
4541 if (mem_cgroup_charge(page_folio(vmf
->cow_page
), vma
->vm_mm
,
4543 put_page(vmf
->cow_page
);
4544 return VM_FAULT_OOM
;
4546 cgroup_throttle_swaprate(vmf
->cow_page
, GFP_KERNEL
);
4548 ret
= __do_fault(vmf
);
4549 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4551 if (ret
& VM_FAULT_DONE_COW
)
4554 copy_user_highpage(vmf
->cow_page
, vmf
->page
, vmf
->address
, vma
);
4555 __SetPageUptodate(vmf
->cow_page
);
4557 ret
|= finish_fault(vmf
);
4558 unlock_page(vmf
->page
);
4559 put_page(vmf
->page
);
4560 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4564 put_page(vmf
->cow_page
);
4568 static vm_fault_t
do_shared_fault(struct vm_fault
*vmf
)
4570 struct vm_area_struct
*vma
= vmf
->vma
;
4571 vm_fault_t ret
, tmp
;
4573 ret
= __do_fault(vmf
);
4574 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4578 * Check if the backing address space wants to know that the page is
4579 * about to become writable
4581 if (vma
->vm_ops
->page_mkwrite
) {
4582 unlock_page(vmf
->page
);
4583 tmp
= do_page_mkwrite(vmf
);
4584 if (unlikely(!tmp
||
4585 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
4586 put_page(vmf
->page
);
4591 ret
|= finish_fault(vmf
);
4592 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
4594 unlock_page(vmf
->page
);
4595 put_page(vmf
->page
);
4599 ret
|= fault_dirty_shared_page(vmf
);
4604 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4605 * but allow concurrent faults).
4606 * The mmap_lock may have been released depending on flags and our
4607 * return value. See filemap_fault() and __folio_lock_or_retry().
4608 * If mmap_lock is released, vma may become invalid (for example
4609 * by other thread calling munmap()).
4611 static vm_fault_t
do_fault(struct vm_fault
*vmf
)
4613 struct vm_area_struct
*vma
= vmf
->vma
;
4614 struct mm_struct
*vm_mm
= vma
->vm_mm
;
4618 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4620 if (!vma
->vm_ops
->fault
) {
4622 * If we find a migration pmd entry or a none pmd entry, which
4623 * should never happen, return SIGBUS
4625 if (unlikely(!pmd_present(*vmf
->pmd
)))
4626 ret
= VM_FAULT_SIGBUS
;
4628 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
,
4633 * Make sure this is not a temporary clearing of pte
4634 * by holding ptl and checking again. A R/M/W update
4635 * of pte involves: take ptl, clearing the pte so that
4636 * we don't have concurrent modification by hardware
4637 * followed by an update.
4639 if (unlikely(pte_none(*vmf
->pte
)))
4640 ret
= VM_FAULT_SIGBUS
;
4642 ret
= VM_FAULT_NOPAGE
;
4644 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4646 } else if (!(vmf
->flags
& FAULT_FLAG_WRITE
))
4647 ret
= do_read_fault(vmf
);
4648 else if (!(vma
->vm_flags
& VM_SHARED
))
4649 ret
= do_cow_fault(vmf
);
4651 ret
= do_shared_fault(vmf
);
4653 /* preallocated pagetable is unused: free it */
4654 if (vmf
->prealloc_pte
) {
4655 pte_free(vm_mm
, vmf
->prealloc_pte
);
4656 vmf
->prealloc_pte
= NULL
;
4661 int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
4662 unsigned long addr
, int page_nid
, int *flags
)
4666 count_vm_numa_event(NUMA_HINT_FAULTS
);
4667 if (page_nid
== numa_node_id()) {
4668 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
4669 *flags
|= TNF_FAULT_LOCAL
;
4672 return mpol_misplaced(page
, vma
, addr
);
4675 static vm_fault_t
do_numa_page(struct vm_fault
*vmf
)
4677 struct vm_area_struct
*vma
= vmf
->vma
;
4678 struct page
*page
= NULL
;
4679 int page_nid
= NUMA_NO_NODE
;
4683 bool was_writable
= pte_savedwrite(vmf
->orig_pte
);
4687 * The "pte" at this point cannot be used safely without
4688 * validation through pte_unmap_same(). It's of NUMA type but
4689 * the pfn may be screwed if the read is non atomic.
4691 vmf
->ptl
= pte_lockptr(vma
->vm_mm
, vmf
->pmd
);
4692 spin_lock(vmf
->ptl
);
4693 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
4694 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4698 /* Get the normal PTE */
4699 old_pte
= ptep_get(vmf
->pte
);
4700 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
4702 page
= vm_normal_page(vma
, vmf
->address
, pte
);
4703 if (!page
|| is_zone_device_page(page
))
4706 /* TODO: handle PTE-mapped THP */
4707 if (PageCompound(page
))
4711 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4712 * much anyway since they can be in shared cache state. This misses
4713 * the case where a mapping is writable but the process never writes
4714 * to it but pte_write gets cleared during protection updates and
4715 * pte_dirty has unpredictable behaviour between PTE scan updates,
4716 * background writeback, dirty balancing and application behaviour.
4719 flags
|= TNF_NO_GROUP
;
4722 * Flag if the page is shared between multiple address spaces. This
4723 * is later used when determining whether to group tasks together
4725 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
4726 flags
|= TNF_SHARED
;
4728 last_cpupid
= page_cpupid_last(page
);
4729 page_nid
= page_to_nid(page
);
4730 target_nid
= numa_migrate_prep(page
, vma
, vmf
->address
, page_nid
,
4732 if (target_nid
== NUMA_NO_NODE
) {
4736 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4738 /* Migrate to the requested node */
4739 if (migrate_misplaced_page(page
, vma
, target_nid
)) {
4740 page_nid
= target_nid
;
4741 flags
|= TNF_MIGRATED
;
4743 flags
|= TNF_MIGRATE_FAIL
;
4744 vmf
->pte
= pte_offset_map(vmf
->pmd
, vmf
->address
);
4745 spin_lock(vmf
->ptl
);
4746 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
4747 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4754 if (page_nid
!= NUMA_NO_NODE
)
4755 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
4759 * Make it present again, depending on how arch implements
4760 * non-accessible ptes, some can allow access by kernel mode.
4762 old_pte
= ptep_modify_prot_start(vma
, vmf
->address
, vmf
->pte
);
4763 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
4764 pte
= pte_mkyoung(pte
);
4766 pte
= pte_mkwrite(pte
);
4767 ptep_modify_prot_commit(vma
, vmf
->address
, vmf
->pte
, old_pte
, pte
);
4768 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
4769 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4773 static inline vm_fault_t
create_huge_pmd(struct vm_fault
*vmf
)
4775 if (vma_is_anonymous(vmf
->vma
))
4776 return do_huge_pmd_anonymous_page(vmf
);
4777 if (vmf
->vma
->vm_ops
->huge_fault
)
4778 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
4779 return VM_FAULT_FALLBACK
;
4782 /* `inline' is required to avoid gcc 4.1.2 build error */
4783 static inline vm_fault_t
wp_huge_pmd(struct vm_fault
*vmf
)
4785 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
4787 if (vma_is_anonymous(vmf
->vma
)) {
4788 if (likely(!unshare
) &&
4789 userfaultfd_huge_pmd_wp(vmf
->vma
, vmf
->orig_pmd
))
4790 return handle_userfault(vmf
, VM_UFFD_WP
);
4791 return do_huge_pmd_wp_page(vmf
);
4793 if (vmf
->vma
->vm_ops
->huge_fault
) {
4794 vm_fault_t ret
= vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
4796 if (!(ret
& VM_FAULT_FALLBACK
))
4800 /* COW or write-notify handled on pte level: split pmd. */
4801 __split_huge_pmd(vmf
->vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
4803 return VM_FAULT_FALLBACK
;
4806 static vm_fault_t
create_huge_pud(struct vm_fault
*vmf
)
4808 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4809 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4810 /* No support for anonymous transparent PUD pages yet */
4811 if (vma_is_anonymous(vmf
->vma
))
4812 return VM_FAULT_FALLBACK
;
4813 if (vmf
->vma
->vm_ops
->huge_fault
)
4814 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
4815 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4816 return VM_FAULT_FALLBACK
;
4819 static vm_fault_t
wp_huge_pud(struct vm_fault
*vmf
, pud_t orig_pud
)
4821 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4822 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4823 /* No support for anonymous transparent PUD pages yet */
4824 if (vma_is_anonymous(vmf
->vma
))
4826 if (vmf
->vma
->vm_ops
->huge_fault
) {
4827 vm_fault_t ret
= vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
4829 if (!(ret
& VM_FAULT_FALLBACK
))
4833 /* COW or write-notify not handled on PUD level: split pud.*/
4834 __split_huge_pud(vmf
->vma
, vmf
->pud
, vmf
->address
);
4835 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
4836 return VM_FAULT_FALLBACK
;
4840 * These routines also need to handle stuff like marking pages dirty
4841 * and/or accessed for architectures that don't do it in hardware (most
4842 * RISC architectures). The early dirtying is also good on the i386.
4844 * There is also a hook called "update_mmu_cache()" that architectures
4845 * with external mmu caches can use to update those (ie the Sparc or
4846 * PowerPC hashed page tables that act as extended TLBs).
4848 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4849 * concurrent faults).
4851 * The mmap_lock may have been released depending on flags and our return value.
4852 * See filemap_fault() and __folio_lock_or_retry().
4854 static vm_fault_t
handle_pte_fault(struct vm_fault
*vmf
)
4858 if (unlikely(pmd_none(*vmf
->pmd
))) {
4860 * Leave __pte_alloc() until later: because vm_ops->fault may
4861 * want to allocate huge page, and if we expose page table
4862 * for an instant, it will be difficult to retract from
4863 * concurrent faults and from rmap lookups.
4866 vmf
->flags
&= ~FAULT_FLAG_ORIG_PTE_VALID
;
4869 * If a huge pmd materialized under us just retry later. Use
4870 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead
4871 * of pmd_trans_huge() to ensure the pmd didn't become
4872 * pmd_trans_huge under us and then back to pmd_none, as a
4873 * result of MADV_DONTNEED running immediately after a huge pmd
4874 * fault in a different thread of this mm, in turn leading to a
4875 * misleading pmd_trans_huge() retval. All we have to ensure is
4876 * that it is a regular pmd that we can walk with
4877 * pte_offset_map() and we can do that through an atomic read
4878 * in C, which is what pmd_trans_unstable() provides.
4880 if (pmd_devmap_trans_unstable(vmf
->pmd
))
4883 * A regular pmd is established and it can't morph into a huge
4884 * pmd from under us anymore at this point because we hold the
4885 * mmap_lock read mode and khugepaged takes it in write mode.
4886 * So now it's safe to run pte_offset_map().
4888 vmf
->pte
= pte_offset_map(vmf
->pmd
, vmf
->address
);
4889 vmf
->orig_pte
= *vmf
->pte
;
4890 vmf
->flags
|= FAULT_FLAG_ORIG_PTE_VALID
;
4893 * some architectures can have larger ptes than wordsize,
4894 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4895 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4896 * accesses. The code below just needs a consistent view
4897 * for the ifs and we later double check anyway with the
4898 * ptl lock held. So here a barrier will do.
4901 if (pte_none(vmf
->orig_pte
)) {
4902 pte_unmap(vmf
->pte
);
4908 if (vma_is_anonymous(vmf
->vma
))
4909 return do_anonymous_page(vmf
);
4911 return do_fault(vmf
);
4914 if (!pte_present(vmf
->orig_pte
))
4915 return do_swap_page(vmf
);
4917 if (pte_protnone(vmf
->orig_pte
) && vma_is_accessible(vmf
->vma
))
4918 return do_numa_page(vmf
);
4920 vmf
->ptl
= pte_lockptr(vmf
->vma
->vm_mm
, vmf
->pmd
);
4921 spin_lock(vmf
->ptl
);
4922 entry
= vmf
->orig_pte
;
4923 if (unlikely(!pte_same(*vmf
->pte
, entry
))) {
4924 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
4927 if (vmf
->flags
& (FAULT_FLAG_WRITE
|FAULT_FLAG_UNSHARE
)) {
4928 if (!pte_write(entry
))
4929 return do_wp_page(vmf
);
4930 else if (likely(vmf
->flags
& FAULT_FLAG_WRITE
))
4931 entry
= pte_mkdirty(entry
);
4933 entry
= pte_mkyoung(entry
);
4934 if (ptep_set_access_flags(vmf
->vma
, vmf
->address
, vmf
->pte
, entry
,
4935 vmf
->flags
& FAULT_FLAG_WRITE
)) {
4936 update_mmu_cache(vmf
->vma
, vmf
->address
, vmf
->pte
);
4938 /* Skip spurious TLB flush for retried page fault */
4939 if (vmf
->flags
& FAULT_FLAG_TRIED
)
4942 * This is needed only for protection faults but the arch code
4943 * is not yet telling us if this is a protection fault or not.
4944 * This still avoids useless tlb flushes for .text page faults
4947 if (vmf
->flags
& FAULT_FLAG_WRITE
)
4948 flush_tlb_fix_spurious_fault(vmf
->vma
, vmf
->address
);
4951 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4956 * By the time we get here, we already hold the mm semaphore
4958 * The mmap_lock may have been released depending on flags and our
4959 * return value. See filemap_fault() and __folio_lock_or_retry().
4961 static vm_fault_t
__handle_mm_fault(struct vm_area_struct
*vma
,
4962 unsigned long address
, unsigned int flags
)
4964 struct vm_fault vmf
= {
4966 .address
= address
& PAGE_MASK
,
4967 .real_address
= address
,
4969 .pgoff
= linear_page_index(vma
, address
),
4970 .gfp_mask
= __get_fault_gfp_mask(vma
),
4972 struct mm_struct
*mm
= vma
->vm_mm
;
4973 unsigned long vm_flags
= vma
->vm_flags
;
4978 pgd
= pgd_offset(mm
, address
);
4979 p4d
= p4d_alloc(mm
, pgd
, address
);
4981 return VM_FAULT_OOM
;
4983 vmf
.pud
= pud_alloc(mm
, p4d
, address
);
4985 return VM_FAULT_OOM
;
4987 if (pud_none(*vmf
.pud
) &&
4988 hugepage_vma_check(vma
, vm_flags
, false, true)) {
4989 ret
= create_huge_pud(&vmf
);
4990 if (!(ret
& VM_FAULT_FALLBACK
))
4993 pud_t orig_pud
= *vmf
.pud
;
4996 if (pud_trans_huge(orig_pud
) || pud_devmap(orig_pud
)) {
4999 * TODO once we support anonymous PUDs: NUMA case and
5000 * FAULT_FLAG_UNSHARE handling.
5002 if ((flags
& FAULT_FLAG_WRITE
) && !pud_write(orig_pud
)) {
5003 ret
= wp_huge_pud(&vmf
, orig_pud
);
5004 if (!(ret
& VM_FAULT_FALLBACK
))
5007 huge_pud_set_accessed(&vmf
, orig_pud
);
5013 vmf
.pmd
= pmd_alloc(mm
, vmf
.pud
, address
);
5015 return VM_FAULT_OOM
;
5017 /* Huge pud page fault raced with pmd_alloc? */
5018 if (pud_trans_unstable(vmf
.pud
))
5021 if (pmd_none(*vmf
.pmd
) &&
5022 hugepage_vma_check(vma
, vm_flags
, false, true)) {
5023 ret
= create_huge_pmd(&vmf
);
5024 if (!(ret
& VM_FAULT_FALLBACK
))
5027 vmf
.orig_pmd
= *vmf
.pmd
;
5030 if (unlikely(is_swap_pmd(vmf
.orig_pmd
))) {
5031 VM_BUG_ON(thp_migration_supported() &&
5032 !is_pmd_migration_entry(vmf
.orig_pmd
));
5033 if (is_pmd_migration_entry(vmf
.orig_pmd
))
5034 pmd_migration_entry_wait(mm
, vmf
.pmd
);
5037 if (pmd_trans_huge(vmf
.orig_pmd
) || pmd_devmap(vmf
.orig_pmd
)) {
5038 if (pmd_protnone(vmf
.orig_pmd
) && vma_is_accessible(vma
))
5039 return do_huge_pmd_numa_page(&vmf
);
5041 if ((flags
& (FAULT_FLAG_WRITE
|FAULT_FLAG_UNSHARE
)) &&
5042 !pmd_write(vmf
.orig_pmd
)) {
5043 ret
= wp_huge_pmd(&vmf
);
5044 if (!(ret
& VM_FAULT_FALLBACK
))
5047 huge_pmd_set_accessed(&vmf
);
5053 return handle_pte_fault(&vmf
);
5057 * mm_account_fault - Do page fault accounting
5059 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
5060 * of perf event counters, but we'll still do the per-task accounting to
5061 * the task who triggered this page fault.
5062 * @address: the faulted address.
5063 * @flags: the fault flags.
5064 * @ret: the fault retcode.
5066 * This will take care of most of the page fault accounting. Meanwhile, it
5067 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
5068 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
5069 * still be in per-arch page fault handlers at the entry of page fault.
5071 static inline void mm_account_fault(struct pt_regs
*regs
,
5072 unsigned long address
, unsigned int flags
,
5078 * We don't do accounting for some specific faults:
5080 * - Unsuccessful faults (e.g. when the address wasn't valid). That
5081 * includes arch_vma_access_permitted() failing before reaching here.
5082 * So this is not a "this many hardware page faults" counter. We
5083 * should use the hw profiling for that.
5085 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
5086 * once they're completed.
5088 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_RETRY
))
5092 * We define the fault as a major fault when the final successful fault
5093 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
5094 * handle it immediately previously).
5096 major
= (ret
& VM_FAULT_MAJOR
) || (flags
& FAULT_FLAG_TRIED
);
5104 * If the fault is done for GUP, regs will be NULL. We only do the
5105 * accounting for the per thread fault counters who triggered the
5106 * fault, and we skip the perf event updates.
5112 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ
, 1, regs
, address
);
5114 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN
, 1, regs
, address
);
5118 * By the time we get here, we already hold the mm semaphore
5120 * The mmap_lock may have been released depending on flags and our
5121 * return value. See filemap_fault() and __folio_lock_or_retry().
5123 vm_fault_t
handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
5124 unsigned int flags
, struct pt_regs
*regs
)
5128 __set_current_state(TASK_RUNNING
);
5130 count_vm_event(PGFAULT
);
5131 count_memcg_event_mm(vma
->vm_mm
, PGFAULT
);
5133 /* do counter updates before entering really critical section. */
5134 check_sync_rss_stat(current
);
5136 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
5137 flags
& FAULT_FLAG_INSTRUCTION
,
5138 flags
& FAULT_FLAG_REMOTE
))
5139 return VM_FAULT_SIGSEGV
;
5142 * Enable the memcg OOM handling for faults triggered in user
5143 * space. Kernel faults are handled more gracefully.
5145 if (flags
& FAULT_FLAG_USER
)
5146 mem_cgroup_enter_user_fault();
5148 if (unlikely(is_vm_hugetlb_page(vma
)))
5149 ret
= hugetlb_fault(vma
->vm_mm
, vma
, address
, flags
);
5151 ret
= __handle_mm_fault(vma
, address
, flags
);
5153 if (flags
& FAULT_FLAG_USER
) {
5154 mem_cgroup_exit_user_fault();
5156 * The task may have entered a memcg OOM situation but
5157 * if the allocation error was handled gracefully (no
5158 * VM_FAULT_OOM), there is no need to kill anything.
5159 * Just clean up the OOM state peacefully.
5161 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
5162 mem_cgroup_oom_synchronize(false);
5165 mm_account_fault(regs
, address
, flags
, ret
);
5169 EXPORT_SYMBOL_GPL(handle_mm_fault
);
5171 #ifndef __PAGETABLE_P4D_FOLDED
5173 * Allocate p4d page table.
5174 * We've already handled the fast-path in-line.
5176 int __p4d_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
5178 p4d_t
*new = p4d_alloc_one(mm
, address
);
5182 spin_lock(&mm
->page_table_lock
);
5183 if (pgd_present(*pgd
)) { /* Another has populated it */
5186 smp_wmb(); /* See comment in pmd_install() */
5187 pgd_populate(mm
, pgd
, new);
5189 spin_unlock(&mm
->page_table_lock
);
5192 #endif /* __PAGETABLE_P4D_FOLDED */
5194 #ifndef __PAGETABLE_PUD_FOLDED
5196 * Allocate page upper directory.
5197 * We've already handled the fast-path in-line.
5199 int __pud_alloc(struct mm_struct
*mm
, p4d_t
*p4d
, unsigned long address
)
5201 pud_t
*new = pud_alloc_one(mm
, address
);
5205 spin_lock(&mm
->page_table_lock
);
5206 if (!p4d_present(*p4d
)) {
5208 smp_wmb(); /* See comment in pmd_install() */
5209 p4d_populate(mm
, p4d
, new);
5210 } else /* Another has populated it */
5212 spin_unlock(&mm
->page_table_lock
);
5215 #endif /* __PAGETABLE_PUD_FOLDED */
5217 #ifndef __PAGETABLE_PMD_FOLDED
5219 * Allocate page middle directory.
5220 * We've already handled the fast-path in-line.
5222 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
5225 pmd_t
*new = pmd_alloc_one(mm
, address
);
5229 ptl
= pud_lock(mm
, pud
);
5230 if (!pud_present(*pud
)) {
5232 smp_wmb(); /* See comment in pmd_install() */
5233 pud_populate(mm
, pud
, new);
5234 } else { /* Another has populated it */
5240 #endif /* __PAGETABLE_PMD_FOLDED */
5243 * follow_pte - look up PTE at a user virtual address
5244 * @mm: the mm_struct of the target address space
5245 * @address: user virtual address
5246 * @ptepp: location to store found PTE
5247 * @ptlp: location to store the lock for the PTE
5249 * On a successful return, the pointer to the PTE is stored in @ptepp;
5250 * the corresponding lock is taken and its location is stored in @ptlp.
5251 * The contents of the PTE are only stable until @ptlp is released;
5252 * any further use, if any, must be protected against invalidation
5253 * with MMU notifiers.
5255 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
5256 * should be taken for read.
5258 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
5259 * it is not a good general-purpose API.
5261 * Return: zero on success, -ve otherwise.
5263 int follow_pte(struct mm_struct
*mm
, unsigned long address
,
5264 pte_t
**ptepp
, spinlock_t
**ptlp
)
5272 pgd
= pgd_offset(mm
, address
);
5273 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
5276 p4d
= p4d_offset(pgd
, address
);
5277 if (p4d_none(*p4d
) || unlikely(p4d_bad(*p4d
)))
5280 pud
= pud_offset(p4d
, address
);
5281 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
5284 pmd
= pmd_offset(pud
, address
);
5285 VM_BUG_ON(pmd_trans_huge(*pmd
));
5287 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
5290 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
5291 if (!pte_present(*ptep
))
5296 pte_unmap_unlock(ptep
, *ptlp
);
5300 EXPORT_SYMBOL_GPL(follow_pte
);
5303 * follow_pfn - look up PFN at a user virtual address
5304 * @vma: memory mapping
5305 * @address: user virtual address
5306 * @pfn: location to store found PFN
5308 * Only IO mappings and raw PFN mappings are allowed.
5310 * This function does not allow the caller to read the permissions
5311 * of the PTE. Do not use it.
5313 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5315 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
5322 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
5325 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
5328 *pfn
= pte_pfn(*ptep
);
5329 pte_unmap_unlock(ptep
, ptl
);
5332 EXPORT_SYMBOL(follow_pfn
);
5334 #ifdef CONFIG_HAVE_IOREMAP_PROT
5335 int follow_phys(struct vm_area_struct
*vma
,
5336 unsigned long address
, unsigned int flags
,
5337 unsigned long *prot
, resource_size_t
*phys
)
5343 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
5346 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
5350 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
5353 *prot
= pgprot_val(pte_pgprot(pte
));
5354 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
5358 pte_unmap_unlock(ptep
, ptl
);
5364 * generic_access_phys - generic implementation for iomem mmap access
5365 * @vma: the vma to access
5366 * @addr: userspace address, not relative offset within @vma
5367 * @buf: buffer to read/write
5368 * @len: length of transfer
5369 * @write: set to FOLL_WRITE when writing, otherwise reading
5371 * This is a generic implementation for &vm_operations_struct.access for an
5372 * iomem mapping. This callback is used by access_process_vm() when the @vma is
5375 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
5376 void *buf
, int len
, int write
)
5378 resource_size_t phys_addr
;
5379 unsigned long prot
= 0;
5380 void __iomem
*maddr
;
5383 int offset
= offset_in_page(addr
);
5386 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
5390 if (follow_pte(vma
->vm_mm
, addr
, &ptep
, &ptl
))
5393 pte_unmap_unlock(ptep
, ptl
);
5395 prot
= pgprot_val(pte_pgprot(pte
));
5396 phys_addr
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
5398 if ((write
& FOLL_WRITE
) && !pte_write(pte
))
5401 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
5405 if (follow_pte(vma
->vm_mm
, addr
, &ptep
, &ptl
))
5408 if (!pte_same(pte
, *ptep
)) {
5409 pte_unmap_unlock(ptep
, ptl
);
5416 memcpy_toio(maddr
+ offset
, buf
, len
);
5418 memcpy_fromio(buf
, maddr
+ offset
, len
);
5420 pte_unmap_unlock(ptep
, ptl
);
5426 EXPORT_SYMBOL_GPL(generic_access_phys
);
5430 * Access another process' address space as given in mm.
5432 int __access_remote_vm(struct mm_struct
*mm
, unsigned long addr
, void *buf
,
5433 int len
, unsigned int gup_flags
)
5435 struct vm_area_struct
*vma
;
5436 void *old_buf
= buf
;
5437 int write
= gup_flags
& FOLL_WRITE
;
5439 if (mmap_read_lock_killable(mm
))
5442 /* ignore errors, just check how much was successfully transferred */
5444 int bytes
, ret
, offset
;
5446 struct page
*page
= NULL
;
5448 ret
= get_user_pages_remote(mm
, addr
, 1,
5449 gup_flags
, &page
, &vma
, NULL
);
5451 #ifndef CONFIG_HAVE_IOREMAP_PROT
5455 * Check if this is a VM_IO | VM_PFNMAP VMA, which
5456 * we can access using slightly different code.
5458 vma
= vma_lookup(mm
, addr
);
5461 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
5462 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
5470 offset
= addr
& (PAGE_SIZE
-1);
5471 if (bytes
> PAGE_SIZE
-offset
)
5472 bytes
= PAGE_SIZE
-offset
;
5476 copy_to_user_page(vma
, page
, addr
,
5477 maddr
+ offset
, buf
, bytes
);
5478 set_page_dirty_lock(page
);
5480 copy_from_user_page(vma
, page
, addr
,
5481 buf
, maddr
+ offset
, bytes
);
5490 mmap_read_unlock(mm
);
5492 return buf
- old_buf
;
5496 * access_remote_vm - access another process' address space
5497 * @mm: the mm_struct of the target address space
5498 * @addr: start address to access
5499 * @buf: source or destination buffer
5500 * @len: number of bytes to transfer
5501 * @gup_flags: flags modifying lookup behaviour
5503 * The caller must hold a reference on @mm.
5505 * Return: number of bytes copied from source to destination.
5507 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
5508 void *buf
, int len
, unsigned int gup_flags
)
5510 return __access_remote_vm(mm
, addr
, buf
, len
, gup_flags
);
5514 * Access another process' address space.
5515 * Source/target buffer must be kernel space,
5516 * Do not walk the page table directly, use get_user_pages
5518 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
5519 void *buf
, int len
, unsigned int gup_flags
)
5521 struct mm_struct
*mm
;
5524 mm
= get_task_mm(tsk
);
5528 ret
= __access_remote_vm(mm
, addr
, buf
, len
, gup_flags
);
5534 EXPORT_SYMBOL_GPL(access_process_vm
);
5537 * Print the name of a VMA.
5539 void print_vma_addr(char *prefix
, unsigned long ip
)
5541 struct mm_struct
*mm
= current
->mm
;
5542 struct vm_area_struct
*vma
;
5545 * we might be running from an atomic context so we cannot sleep
5547 if (!mmap_read_trylock(mm
))
5550 vma
= find_vma(mm
, ip
);
5551 if (vma
&& vma
->vm_file
) {
5552 struct file
*f
= vma
->vm_file
;
5553 char *buf
= (char *)__get_free_page(GFP_NOWAIT
);
5557 p
= file_path(f
, buf
, PAGE_SIZE
);
5560 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
5562 vma
->vm_end
- vma
->vm_start
);
5563 free_page((unsigned long)buf
);
5566 mmap_read_unlock(mm
);
5569 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5570 void __might_fault(const char *file
, int line
)
5572 if (pagefault_disabled())
5574 __might_sleep(file
, line
);
5575 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5577 might_lock_read(¤t
->mm
->mmap_lock
);
5580 EXPORT_SYMBOL(__might_fault
);
5583 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5585 * Process all subpages of the specified huge page with the specified
5586 * operation. The target subpage will be processed last to keep its
5589 static inline void process_huge_page(
5590 unsigned long addr_hint
, unsigned int pages_per_huge_page
,
5591 void (*process_subpage
)(unsigned long addr
, int idx
, void *arg
),
5595 unsigned long addr
= addr_hint
&
5596 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5598 /* Process target subpage last to keep its cache lines hot */
5600 n
= (addr_hint
- addr
) / PAGE_SIZE
;
5601 if (2 * n
<= pages_per_huge_page
) {
5602 /* If target subpage in first half of huge page */
5605 /* Process subpages at the end of huge page */
5606 for (i
= pages_per_huge_page
- 1; i
>= 2 * n
; i
--) {
5608 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
5611 /* If target subpage in second half of huge page */
5612 base
= pages_per_huge_page
- 2 * (pages_per_huge_page
- n
);
5613 l
= pages_per_huge_page
- n
;
5614 /* Process subpages at the begin of huge page */
5615 for (i
= 0; i
< base
; i
++) {
5617 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
5621 * Process remaining subpages in left-right-left-right pattern
5622 * towards the target subpage
5624 for (i
= 0; i
< l
; i
++) {
5625 int left_idx
= base
+ i
;
5626 int right_idx
= base
+ 2 * l
- 1 - i
;
5629 process_subpage(addr
+ left_idx
* PAGE_SIZE
, left_idx
, arg
);
5631 process_subpage(addr
+ right_idx
* PAGE_SIZE
, right_idx
, arg
);
5635 static void clear_gigantic_page(struct page
*page
,
5637 unsigned int pages_per_huge_page
)
5640 struct page
*p
= page
;
5643 for (i
= 0; i
< pages_per_huge_page
;
5644 i
++, p
= mem_map_next(p
, page
, i
)) {
5646 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
5650 static void clear_subpage(unsigned long addr
, int idx
, void *arg
)
5652 struct page
*page
= arg
;
5654 clear_user_highpage(page
+ idx
, addr
);
5657 void clear_huge_page(struct page
*page
,
5658 unsigned long addr_hint
, unsigned int pages_per_huge_page
)
5660 unsigned long addr
= addr_hint
&
5661 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5663 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
5664 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
5668 process_huge_page(addr_hint
, pages_per_huge_page
, clear_subpage
, page
);
5671 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
5673 struct vm_area_struct
*vma
,
5674 unsigned int pages_per_huge_page
)
5677 struct page
*dst_base
= dst
;
5678 struct page
*src_base
= src
;
5680 for (i
= 0; i
< pages_per_huge_page
; ) {
5682 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
5685 dst
= mem_map_next(dst
, dst_base
, i
);
5686 src
= mem_map_next(src
, src_base
, i
);
5690 struct copy_subpage_arg
{
5693 struct vm_area_struct
*vma
;
5696 static void copy_subpage(unsigned long addr
, int idx
, void *arg
)
5698 struct copy_subpage_arg
*copy_arg
= arg
;
5700 copy_user_highpage(copy_arg
->dst
+ idx
, copy_arg
->src
+ idx
,
5701 addr
, copy_arg
->vma
);
5704 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
5705 unsigned long addr_hint
, struct vm_area_struct
*vma
,
5706 unsigned int pages_per_huge_page
)
5708 unsigned long addr
= addr_hint
&
5709 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5710 struct copy_subpage_arg arg
= {
5716 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
5717 copy_user_gigantic_page(dst
, src
, addr
, vma
,
5718 pages_per_huge_page
);
5722 process_huge_page(addr_hint
, pages_per_huge_page
, copy_subpage
, &arg
);
5725 long copy_huge_page_from_user(struct page
*dst_page
,
5726 const void __user
*usr_src
,
5727 unsigned int pages_per_huge_page
,
5728 bool allow_pagefault
)
5731 unsigned long i
, rc
= 0;
5732 unsigned long ret_val
= pages_per_huge_page
* PAGE_SIZE
;
5733 struct page
*subpage
= dst_page
;
5735 for (i
= 0; i
< pages_per_huge_page
;
5736 i
++, subpage
= mem_map_next(subpage
, dst_page
, i
)) {
5737 if (allow_pagefault
)
5738 page_kaddr
= kmap(subpage
);
5740 page_kaddr
= kmap_atomic(subpage
);
5741 rc
= copy_from_user(page_kaddr
,
5742 usr_src
+ i
* PAGE_SIZE
, PAGE_SIZE
);
5743 if (allow_pagefault
)
5746 kunmap_atomic(page_kaddr
);
5748 ret_val
-= (PAGE_SIZE
- rc
);
5752 flush_dcache_page(subpage
);
5758 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5760 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5762 static struct kmem_cache
*page_ptl_cachep
;
5764 void __init
ptlock_cache_init(void)
5766 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
5770 bool ptlock_alloc(struct page
*page
)
5774 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
5781 void ptlock_free(struct page
*page
)
5783 kmem_cache_free(page_ptl_cachep
, page
->ptl
);