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/sched/mm.h>
45 #include <linux/sched/coredump.h>
46 #include <linux/sched/numa_balancing.h>
47 #include <linux/sched/task.h>
48 #include <linux/hugetlb.h>
49 #include <linux/mman.h>
50 #include <linux/swap.h>
51 #include <linux/highmem.h>
52 #include <linux/pagemap.h>
53 #include <linux/memremap.h>
54 #include <linux/ksm.h>
55 #include <linux/rmap.h>
56 #include <linux/export.h>
57 #include <linux/delayacct.h>
58 #include <linux/init.h>
59 #include <linux/pfn_t.h>
60 #include <linux/writeback.h>
61 #include <linux/memcontrol.h>
62 #include <linux/mmu_notifier.h>
63 #include <linux/swapops.h>
64 #include <linux/elf.h>
65 #include <linux/gfp.h>
66 #include <linux/migrate.h>
67 #include <linux/string.h>
68 #include <linux/dma-debug.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>
76 #include <asm/mmu_context.h>
77 #include <asm/pgalloc.h>
78 #include <linux/uaccess.h>
80 #include <asm/tlbflush.h>
81 #include <asm/pgtable.h>
85 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
86 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
89 #ifndef CONFIG_NEED_MULTIPLE_NODES
90 /* use the per-pgdat data instead for discontigmem - mbligh */
91 unsigned long max_mapnr
;
92 EXPORT_SYMBOL(max_mapnr
);
95 EXPORT_SYMBOL(mem_map
);
99 * A number of key systems in x86 including ioremap() rely on the assumption
100 * that high_memory defines the upper bound on direct map memory, then end
101 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
102 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
106 EXPORT_SYMBOL(high_memory
);
109 * Randomize the address space (stacks, mmaps, brk, etc.).
111 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
112 * as ancient (libc5 based) binaries can segfault. )
114 int randomize_va_space __read_mostly
=
115 #ifdef CONFIG_COMPAT_BRK
121 #ifndef arch_faults_on_old_pte
122 static inline bool arch_faults_on_old_pte(void)
125 * Those arches which don't have hw access flag feature need to
126 * implement their own helper. By default, "true" means pagefault
127 * will be hit on old pte.
133 static int __init
disable_randmaps(char *s
)
135 randomize_va_space
= 0;
138 __setup("norandmaps", disable_randmaps
);
140 unsigned long zero_pfn __read_mostly
;
141 EXPORT_SYMBOL(zero_pfn
);
143 unsigned long highest_memmap_pfn __read_mostly
;
146 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
148 static int __init
init_zero_pfn(void)
150 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
153 early_initcall(init_zero_pfn
);
156 #if defined(SPLIT_RSS_COUNTING)
158 void sync_mm_rss(struct mm_struct
*mm
)
162 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
163 if (current
->rss_stat
.count
[i
]) {
164 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
165 current
->rss_stat
.count
[i
] = 0;
168 current
->rss_stat
.events
= 0;
171 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
173 struct task_struct
*task
= current
;
175 if (likely(task
->mm
== mm
))
176 task
->rss_stat
.count
[member
] += val
;
178 add_mm_counter(mm
, member
, val
);
180 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
181 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
183 /* sync counter once per 64 page faults */
184 #define TASK_RSS_EVENTS_THRESH (64)
185 static void check_sync_rss_stat(struct task_struct
*task
)
187 if (unlikely(task
!= current
))
189 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
190 sync_mm_rss(task
->mm
);
192 #else /* SPLIT_RSS_COUNTING */
194 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
195 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
197 static void check_sync_rss_stat(struct task_struct
*task
)
201 #endif /* SPLIT_RSS_COUNTING */
204 * Note: this doesn't free the actual pages themselves. That
205 * has been handled earlier when unmapping all the memory regions.
207 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
210 pgtable_t token
= pmd_pgtable(*pmd
);
212 pte_free_tlb(tlb
, token
, addr
);
213 mm_dec_nr_ptes(tlb
->mm
);
216 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
217 unsigned long addr
, unsigned long end
,
218 unsigned long floor
, unsigned long ceiling
)
225 pmd
= pmd_offset(pud
, addr
);
227 next
= pmd_addr_end(addr
, end
);
228 if (pmd_none_or_clear_bad(pmd
))
230 free_pte_range(tlb
, pmd
, addr
);
231 } while (pmd
++, addr
= next
, addr
!= end
);
241 if (end
- 1 > ceiling
- 1)
244 pmd
= pmd_offset(pud
, start
);
246 pmd_free_tlb(tlb
, pmd
, start
);
247 mm_dec_nr_pmds(tlb
->mm
);
250 static inline void free_pud_range(struct mmu_gather
*tlb
, p4d_t
*p4d
,
251 unsigned long addr
, unsigned long end
,
252 unsigned long floor
, unsigned long ceiling
)
259 pud
= pud_offset(p4d
, addr
);
261 next
= pud_addr_end(addr
, end
);
262 if (pud_none_or_clear_bad(pud
))
264 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
265 } while (pud
++, addr
= next
, addr
!= end
);
275 if (end
- 1 > ceiling
- 1)
278 pud
= pud_offset(p4d
, start
);
280 pud_free_tlb(tlb
, pud
, start
);
281 mm_dec_nr_puds(tlb
->mm
);
284 static inline void free_p4d_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
285 unsigned long addr
, unsigned long end
,
286 unsigned long floor
, unsigned long ceiling
)
293 p4d
= p4d_offset(pgd
, addr
);
295 next
= p4d_addr_end(addr
, end
);
296 if (p4d_none_or_clear_bad(p4d
))
298 free_pud_range(tlb
, p4d
, addr
, next
, floor
, ceiling
);
299 } while (p4d
++, addr
= next
, addr
!= end
);
305 ceiling
&= PGDIR_MASK
;
309 if (end
- 1 > ceiling
- 1)
312 p4d
= p4d_offset(pgd
, start
);
314 p4d_free_tlb(tlb
, p4d
, start
);
318 * This function frees user-level page tables of a process.
320 void free_pgd_range(struct mmu_gather
*tlb
,
321 unsigned long addr
, unsigned long end
,
322 unsigned long floor
, unsigned long ceiling
)
328 * The next few lines have given us lots of grief...
330 * Why are we testing PMD* at this top level? Because often
331 * there will be no work to do at all, and we'd prefer not to
332 * go all the way down to the bottom just to discover that.
334 * Why all these "- 1"s? Because 0 represents both the bottom
335 * of the address space and the top of it (using -1 for the
336 * top wouldn't help much: the masks would do the wrong thing).
337 * The rule is that addr 0 and floor 0 refer to the bottom of
338 * the address space, but end 0 and ceiling 0 refer to the top
339 * Comparisons need to use "end - 1" and "ceiling - 1" (though
340 * that end 0 case should be mythical).
342 * Wherever addr is brought up or ceiling brought down, we must
343 * be careful to reject "the opposite 0" before it confuses the
344 * subsequent tests. But what about where end is brought down
345 * by PMD_SIZE below? no, end can't go down to 0 there.
347 * Whereas we round start (addr) and ceiling down, by different
348 * masks at different levels, in order to test whether a table
349 * now has no other vmas using it, so can be freed, we don't
350 * bother to round floor or end up - the tests don't need that.
364 if (end
- 1 > ceiling
- 1)
369 * We add page table cache pages with PAGE_SIZE,
370 * (see pte_free_tlb()), flush the tlb if we need
372 tlb_change_page_size(tlb
, PAGE_SIZE
);
373 pgd
= pgd_offset(tlb
->mm
, addr
);
375 next
= pgd_addr_end(addr
, end
);
376 if (pgd_none_or_clear_bad(pgd
))
378 free_p4d_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
379 } while (pgd
++, addr
= next
, addr
!= end
);
382 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
383 unsigned long floor
, unsigned long ceiling
)
386 struct vm_area_struct
*next
= vma
->vm_next
;
387 unsigned long addr
= vma
->vm_start
;
390 * Hide vma from rmap and truncate_pagecache before freeing
393 unlink_anon_vmas(vma
);
394 unlink_file_vma(vma
);
396 if (is_vm_hugetlb_page(vma
)) {
397 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
398 floor
, next
? next
->vm_start
: ceiling
);
401 * Optimization: gather nearby vmas into one call down
403 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
404 && !is_vm_hugetlb_page(next
)) {
407 unlink_anon_vmas(vma
);
408 unlink_file_vma(vma
);
410 free_pgd_range(tlb
, addr
, vma
->vm_end
,
411 floor
, next
? next
->vm_start
: ceiling
);
417 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
)
420 pgtable_t
new = pte_alloc_one(mm
);
425 * Ensure all pte setup (eg. pte page lock and page clearing) are
426 * visible before the pte is made visible to other CPUs by being
427 * put into page tables.
429 * The other side of the story is the pointer chasing in the page
430 * table walking code (when walking the page table without locking;
431 * ie. most of the time). Fortunately, these data accesses consist
432 * of a chain of data-dependent loads, meaning most CPUs (alpha
433 * being the notable exception) will already guarantee loads are
434 * seen in-order. See the alpha page table accessors for the
435 * smp_read_barrier_depends() barriers in page table walking code.
437 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
439 ptl
= pmd_lock(mm
, pmd
);
440 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
442 pmd_populate(mm
, pmd
, new);
451 int __pte_alloc_kernel(pmd_t
*pmd
)
453 pte_t
*new = pte_alloc_one_kernel(&init_mm
);
457 smp_wmb(); /* See comment in __pte_alloc */
459 spin_lock(&init_mm
.page_table_lock
);
460 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
461 pmd_populate_kernel(&init_mm
, pmd
, new);
464 spin_unlock(&init_mm
.page_table_lock
);
466 pte_free_kernel(&init_mm
, new);
470 static inline void init_rss_vec(int *rss
)
472 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
475 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
479 if (current
->mm
== mm
)
481 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
483 add_mm_counter(mm
, i
, rss
[i
]);
487 * This function is called to print an error when a bad pte
488 * is found. For example, we might have a PFN-mapped pte in
489 * a region that doesn't allow it.
491 * The calling function must still handle the error.
493 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
494 pte_t pte
, struct page
*page
)
496 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
497 p4d_t
*p4d
= p4d_offset(pgd
, addr
);
498 pud_t
*pud
= pud_offset(p4d
, addr
);
499 pmd_t
*pmd
= pmd_offset(pud
, addr
);
500 struct address_space
*mapping
;
502 static unsigned long resume
;
503 static unsigned long nr_shown
;
504 static unsigned long nr_unshown
;
507 * Allow a burst of 60 reports, then keep quiet for that minute;
508 * or allow a steady drip of one report per second.
510 if (nr_shown
== 60) {
511 if (time_before(jiffies
, resume
)) {
516 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
523 resume
= jiffies
+ 60 * HZ
;
525 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
526 index
= linear_page_index(vma
, addr
);
528 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
530 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
532 dump_page(page
, "bad pte");
533 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
534 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
535 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
537 vma
->vm_ops
? vma
->vm_ops
->fault
: NULL
,
538 vma
->vm_file
? vma
->vm_file
->f_op
->mmap
: NULL
,
539 mapping
? mapping
->a_ops
->readpage
: NULL
);
541 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
545 * vm_normal_page -- This function gets the "struct page" associated with a pte.
547 * "Special" mappings do not wish to be associated with a "struct page" (either
548 * it doesn't exist, or it exists but they don't want to touch it). In this
549 * case, NULL is returned here. "Normal" mappings do have a struct page.
551 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
552 * pte bit, in which case this function is trivial. Secondly, an architecture
553 * may not have a spare pte bit, which requires a more complicated scheme,
556 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
557 * special mapping (even if there are underlying and valid "struct pages").
558 * COWed pages of a VM_PFNMAP are always normal.
560 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
561 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
562 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
563 * mapping will always honor the rule
565 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
567 * And for normal mappings this is false.
569 * This restricts such mappings to be a linear translation from virtual address
570 * to pfn. To get around this restriction, we allow arbitrary mappings so long
571 * as the vma is not a COW mapping; in that case, we know that all ptes are
572 * special (because none can have been COWed).
575 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
577 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
578 * page" backing, however the difference is that _all_ pages with a struct
579 * page (that is, those where pfn_valid is true) are refcounted and considered
580 * normal pages by the VM. The disadvantage is that pages are refcounted
581 * (which can be slower and simply not an option for some PFNMAP users). The
582 * advantage is that we don't have to follow the strict linearity rule of
583 * PFNMAP mappings in order to support COWable mappings.
586 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
589 unsigned long pfn
= pte_pfn(pte
);
591 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
)) {
592 if (likely(!pte_special(pte
)))
594 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
595 return vma
->vm_ops
->find_special_page(vma
, addr
);
596 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
598 if (is_zero_pfn(pfn
))
603 print_bad_pte(vma
, addr
, pte
, NULL
);
607 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
609 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
610 if (vma
->vm_flags
& VM_MIXEDMAP
) {
616 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
617 if (pfn
== vma
->vm_pgoff
+ off
)
619 if (!is_cow_mapping(vma
->vm_flags
))
624 if (is_zero_pfn(pfn
))
628 if (unlikely(pfn
> highest_memmap_pfn
)) {
629 print_bad_pte(vma
, addr
, pte
, NULL
);
634 * NOTE! We still have PageReserved() pages in the page tables.
635 * eg. VDSO mappings can cause them to exist.
638 return pfn_to_page(pfn
);
641 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
642 struct page
*vm_normal_page_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
645 unsigned long pfn
= pmd_pfn(pmd
);
648 * There is no pmd_special() but there may be special pmds, e.g.
649 * in a direct-access (dax) mapping, so let's just replicate the
650 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
652 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
653 if (vma
->vm_flags
& VM_MIXEDMAP
) {
659 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
660 if (pfn
== vma
->vm_pgoff
+ off
)
662 if (!is_cow_mapping(vma
->vm_flags
))
669 if (is_zero_pfn(pfn
))
671 if (unlikely(pfn
> highest_memmap_pfn
))
675 * NOTE! We still have PageReserved() pages in the page tables.
676 * eg. VDSO mappings can cause them to exist.
679 return pfn_to_page(pfn
);
684 * copy one vm_area from one task to the other. Assumes the page tables
685 * already present in the new task to be cleared in the whole range
686 * covered by this vma.
689 static inline unsigned long
690 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
691 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
692 unsigned long addr
, int *rss
)
694 unsigned long vm_flags
= vma
->vm_flags
;
695 pte_t pte
= *src_pte
;
698 /* pte contains position in swap or file, so copy. */
699 if (unlikely(!pte_present(pte
))) {
700 swp_entry_t entry
= pte_to_swp_entry(pte
);
702 if (likely(!non_swap_entry(entry
))) {
703 if (swap_duplicate(entry
) < 0)
706 /* make sure dst_mm is on swapoff's mmlist. */
707 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
708 spin_lock(&mmlist_lock
);
709 if (list_empty(&dst_mm
->mmlist
))
710 list_add(&dst_mm
->mmlist
,
712 spin_unlock(&mmlist_lock
);
715 } else if (is_migration_entry(entry
)) {
716 page
= migration_entry_to_page(entry
);
718 rss
[mm_counter(page
)]++;
720 if (is_write_migration_entry(entry
) &&
721 is_cow_mapping(vm_flags
)) {
723 * COW mappings require pages in both
724 * parent and child to be set to read.
726 make_migration_entry_read(&entry
);
727 pte
= swp_entry_to_pte(entry
);
728 if (pte_swp_soft_dirty(*src_pte
))
729 pte
= pte_swp_mksoft_dirty(pte
);
730 set_pte_at(src_mm
, addr
, src_pte
, pte
);
732 } else if (is_device_private_entry(entry
)) {
733 page
= device_private_entry_to_page(entry
);
736 * Update rss count even for unaddressable pages, as
737 * they should treated just like normal pages in this
740 * We will likely want to have some new rss counters
741 * for unaddressable pages, at some point. But for now
742 * keep things as they are.
745 rss
[mm_counter(page
)]++;
746 page_dup_rmap(page
, false);
749 * We do not preserve soft-dirty information, because so
750 * far, checkpoint/restore is the only feature that
751 * requires that. And checkpoint/restore does not work
752 * when a device driver is involved (you cannot easily
753 * save and restore device driver state).
755 if (is_write_device_private_entry(entry
) &&
756 is_cow_mapping(vm_flags
)) {
757 make_device_private_entry_read(&entry
);
758 pte
= swp_entry_to_pte(entry
);
759 set_pte_at(src_mm
, addr
, src_pte
, pte
);
766 * If it's a COW mapping, write protect it both
767 * in the parent and the child
769 if (is_cow_mapping(vm_flags
) && pte_write(pte
)) {
770 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
771 pte
= pte_wrprotect(pte
);
775 * If it's a shared mapping, mark it clean in
778 if (vm_flags
& VM_SHARED
)
779 pte
= pte_mkclean(pte
);
780 pte
= pte_mkold(pte
);
782 page
= vm_normal_page(vma
, addr
, pte
);
785 page_dup_rmap(page
, false);
786 rss
[mm_counter(page
)]++;
787 } else if (pte_devmap(pte
)) {
788 page
= pte_page(pte
);
792 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
796 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
797 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
798 unsigned long addr
, unsigned long end
)
800 pte_t
*orig_src_pte
, *orig_dst_pte
;
801 pte_t
*src_pte
, *dst_pte
;
802 spinlock_t
*src_ptl
, *dst_ptl
;
804 int rss
[NR_MM_COUNTERS
];
805 swp_entry_t entry
= (swp_entry_t
){0};
810 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
813 src_pte
= pte_offset_map(src_pmd
, addr
);
814 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
815 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
816 orig_src_pte
= src_pte
;
817 orig_dst_pte
= dst_pte
;
818 arch_enter_lazy_mmu_mode();
822 * We are holding two locks at this point - either of them
823 * could generate latencies in another task on another CPU.
825 if (progress
>= 32) {
827 if (need_resched() ||
828 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
831 if (pte_none(*src_pte
)) {
835 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
840 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
842 arch_leave_lazy_mmu_mode();
843 spin_unlock(src_ptl
);
844 pte_unmap(orig_src_pte
);
845 add_mm_rss_vec(dst_mm
, rss
);
846 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
850 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
859 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
860 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
861 unsigned long addr
, unsigned long end
)
863 pmd_t
*src_pmd
, *dst_pmd
;
866 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
869 src_pmd
= pmd_offset(src_pud
, addr
);
871 next
= pmd_addr_end(addr
, end
);
872 if (is_swap_pmd(*src_pmd
) || pmd_trans_huge(*src_pmd
)
873 || pmd_devmap(*src_pmd
)) {
875 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PMD_SIZE
, vma
);
876 err
= copy_huge_pmd(dst_mm
, src_mm
,
877 dst_pmd
, src_pmd
, addr
, vma
);
884 if (pmd_none_or_clear_bad(src_pmd
))
886 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
889 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
893 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
894 p4d_t
*dst_p4d
, p4d_t
*src_p4d
, struct vm_area_struct
*vma
,
895 unsigned long addr
, unsigned long end
)
897 pud_t
*src_pud
, *dst_pud
;
900 dst_pud
= pud_alloc(dst_mm
, dst_p4d
, addr
);
903 src_pud
= pud_offset(src_p4d
, addr
);
905 next
= pud_addr_end(addr
, end
);
906 if (pud_trans_huge(*src_pud
) || pud_devmap(*src_pud
)) {
909 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PUD_SIZE
, vma
);
910 err
= copy_huge_pud(dst_mm
, src_mm
,
911 dst_pud
, src_pud
, addr
, vma
);
918 if (pud_none_or_clear_bad(src_pud
))
920 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
923 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
927 static inline int copy_p4d_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
928 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
929 unsigned long addr
, unsigned long end
)
931 p4d_t
*src_p4d
, *dst_p4d
;
934 dst_p4d
= p4d_alloc(dst_mm
, dst_pgd
, addr
);
937 src_p4d
= p4d_offset(src_pgd
, addr
);
939 next
= p4d_addr_end(addr
, end
);
940 if (p4d_none_or_clear_bad(src_p4d
))
942 if (copy_pud_range(dst_mm
, src_mm
, dst_p4d
, src_p4d
,
945 } while (dst_p4d
++, src_p4d
++, addr
= next
, addr
!= end
);
949 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
950 struct vm_area_struct
*vma
)
952 pgd_t
*src_pgd
, *dst_pgd
;
954 unsigned long addr
= vma
->vm_start
;
955 unsigned long end
= vma
->vm_end
;
956 struct mmu_notifier_range range
;
961 * Don't copy ptes where a page fault will fill them correctly.
962 * Fork becomes much lighter when there are big shared or private
963 * readonly mappings. The tradeoff is that copy_page_range is more
964 * efficient than faulting.
966 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_PFNMAP
| VM_MIXEDMAP
)) &&
970 if (is_vm_hugetlb_page(vma
))
971 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
973 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
975 * We do not free on error cases below as remove_vma
976 * gets called on error from higher level routine
978 ret
= track_pfn_copy(vma
);
984 * We need to invalidate the secondary MMU mappings only when
985 * there could be a permission downgrade on the ptes of the
986 * parent mm. And a permission downgrade will only happen if
987 * is_cow_mapping() returns true.
989 is_cow
= is_cow_mapping(vma
->vm_flags
);
992 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_PAGE
,
993 0, vma
, src_mm
, addr
, end
);
994 mmu_notifier_invalidate_range_start(&range
);
998 dst_pgd
= pgd_offset(dst_mm
, addr
);
999 src_pgd
= pgd_offset(src_mm
, addr
);
1001 next
= pgd_addr_end(addr
, end
);
1002 if (pgd_none_or_clear_bad(src_pgd
))
1004 if (unlikely(copy_p4d_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1005 vma
, addr
, next
))) {
1009 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1012 mmu_notifier_invalidate_range_end(&range
);
1016 /* Whether we should zap all COWed (private) pages too */
1017 static inline bool should_zap_cows(struct zap_details
*details
)
1019 /* By default, zap all pages */
1023 /* Or, we zap COWed pages only if the caller wants to */
1024 return !details
->check_mapping
;
1027 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1028 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1029 unsigned long addr
, unsigned long end
,
1030 struct zap_details
*details
)
1032 struct mm_struct
*mm
= tlb
->mm
;
1033 int force_flush
= 0;
1034 int rss
[NR_MM_COUNTERS
];
1040 tlb_change_page_size(tlb
, PAGE_SIZE
);
1043 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1045 flush_tlb_batched_pending(mm
);
1046 arch_enter_lazy_mmu_mode();
1049 if (pte_none(ptent
))
1055 if (pte_present(ptent
)) {
1058 page
= vm_normal_page(vma
, addr
, ptent
);
1059 if (unlikely(details
) && page
) {
1061 * unmap_shared_mapping_pages() wants to
1062 * invalidate cache without truncating:
1063 * unmap shared but keep private pages.
1065 if (details
->check_mapping
&&
1066 details
->check_mapping
!= page_rmapping(page
))
1069 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1071 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1072 if (unlikely(!page
))
1075 if (!PageAnon(page
)) {
1076 if (pte_dirty(ptent
)) {
1078 set_page_dirty(page
);
1080 if (pte_young(ptent
) &&
1081 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1082 mark_page_accessed(page
);
1084 rss
[mm_counter(page
)]--;
1085 page_remove_rmap(page
, false);
1086 if (unlikely(page_mapcount(page
) < 0))
1087 print_bad_pte(vma
, addr
, ptent
, page
);
1088 if (unlikely(__tlb_remove_page(tlb
, page
))) {
1096 entry
= pte_to_swp_entry(ptent
);
1097 if (non_swap_entry(entry
) && is_device_private_entry(entry
)) {
1098 struct page
*page
= device_private_entry_to_page(entry
);
1100 if (unlikely(details
&& details
->check_mapping
)) {
1102 * unmap_shared_mapping_pages() wants to
1103 * invalidate cache without truncating:
1104 * unmap shared but keep private pages.
1106 if (details
->check_mapping
!=
1107 page_rmapping(page
))
1111 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1112 rss
[mm_counter(page
)]--;
1113 page_remove_rmap(page
, false);
1118 if (!non_swap_entry(entry
)) {
1119 /* Genuine swap entry, hence a private anon page */
1120 if (!should_zap_cows(details
))
1123 } else if (is_migration_entry(entry
)) {
1126 page
= migration_entry_to_page(entry
);
1127 if (details
&& details
->check_mapping
&&
1128 details
->check_mapping
!= page_rmapping(page
))
1130 rss
[mm_counter(page
)]--;
1132 if (unlikely(!free_swap_and_cache(entry
)))
1133 print_bad_pte(vma
, addr
, ptent
, NULL
);
1134 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1135 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1137 add_mm_rss_vec(mm
, rss
);
1138 arch_leave_lazy_mmu_mode();
1140 /* Do the actual TLB flush before dropping ptl */
1142 tlb_flush_mmu_tlbonly(tlb
);
1143 pte_unmap_unlock(start_pte
, ptl
);
1146 * If we forced a TLB flush (either due to running out of
1147 * batch buffers or because we needed to flush dirty TLB
1148 * entries before releasing the ptl), free the batched
1149 * memory too. Restart if we didn't do everything.
1164 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1165 struct vm_area_struct
*vma
, pud_t
*pud
,
1166 unsigned long addr
, unsigned long end
,
1167 struct zap_details
*details
)
1172 pmd
= pmd_offset(pud
, addr
);
1174 next
= pmd_addr_end(addr
, end
);
1175 if (is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1176 if (next
- addr
!= HPAGE_PMD_SIZE
)
1177 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
1178 else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1181 } else if (details
&& details
->single_page
&&
1182 PageTransCompound(details
->single_page
) &&
1183 next
- addr
== HPAGE_PMD_SIZE
&& pmd_none(*pmd
)) {
1184 spinlock_t
*ptl
= pmd_lock(tlb
->mm
, pmd
);
1186 * Take and drop THP pmd lock so that we cannot return
1187 * prematurely, while zap_huge_pmd() has cleared *pmd,
1188 * but not yet decremented compound_mapcount().
1194 * Here there can be other concurrent MADV_DONTNEED or
1195 * trans huge page faults running, and if the pmd is
1196 * none or trans huge it can change under us. This is
1197 * because MADV_DONTNEED holds the mmap_sem in read
1200 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1202 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1205 } while (pmd
++, addr
= next
, addr
!= end
);
1210 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1211 struct vm_area_struct
*vma
, p4d_t
*p4d
,
1212 unsigned long addr
, unsigned long end
,
1213 struct zap_details
*details
)
1218 pud
= pud_offset(p4d
, addr
);
1220 next
= pud_addr_end(addr
, end
);
1221 if (pud_trans_huge(*pud
) || pud_devmap(*pud
)) {
1222 if (next
- addr
!= HPAGE_PUD_SIZE
) {
1223 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb
->mm
->mmap_sem
), vma
);
1224 split_huge_pud(vma
, pud
, addr
);
1225 } else if (zap_huge_pud(tlb
, vma
, pud
, addr
))
1229 if (pud_none_or_clear_bad(pud
))
1231 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1234 } while (pud
++, addr
= next
, addr
!= end
);
1239 static inline unsigned long zap_p4d_range(struct mmu_gather
*tlb
,
1240 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1241 unsigned long addr
, unsigned long end
,
1242 struct zap_details
*details
)
1247 p4d
= p4d_offset(pgd
, addr
);
1249 next
= p4d_addr_end(addr
, end
);
1250 if (p4d_none_or_clear_bad(p4d
))
1252 next
= zap_pud_range(tlb
, vma
, p4d
, addr
, next
, details
);
1253 } while (p4d
++, addr
= next
, addr
!= end
);
1258 void unmap_page_range(struct mmu_gather
*tlb
,
1259 struct vm_area_struct
*vma
,
1260 unsigned long addr
, unsigned long end
,
1261 struct zap_details
*details
)
1266 BUG_ON(addr
>= end
);
1267 tlb_start_vma(tlb
, vma
);
1268 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1270 next
= pgd_addr_end(addr
, end
);
1271 if (pgd_none_or_clear_bad(pgd
))
1273 next
= zap_p4d_range(tlb
, vma
, pgd
, addr
, next
, details
);
1274 } while (pgd
++, addr
= next
, addr
!= end
);
1275 tlb_end_vma(tlb
, vma
);
1279 static void unmap_single_vma(struct mmu_gather
*tlb
,
1280 struct vm_area_struct
*vma
, unsigned long start_addr
,
1281 unsigned long end_addr
,
1282 struct zap_details
*details
)
1284 unsigned long start
= max(vma
->vm_start
, start_addr
);
1287 if (start
>= vma
->vm_end
)
1289 end
= min(vma
->vm_end
, end_addr
);
1290 if (end
<= vma
->vm_start
)
1294 uprobe_munmap(vma
, start
, end
);
1296 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1297 untrack_pfn(vma
, 0, 0);
1300 if (unlikely(is_vm_hugetlb_page(vma
))) {
1302 * It is undesirable to test vma->vm_file as it
1303 * should be non-null for valid hugetlb area.
1304 * However, vm_file will be NULL in the error
1305 * cleanup path of mmap_region. When
1306 * hugetlbfs ->mmap method fails,
1307 * mmap_region() nullifies vma->vm_file
1308 * before calling this function to clean up.
1309 * Since no pte has actually been setup, it is
1310 * safe to do nothing in this case.
1313 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
1314 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1315 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
1318 unmap_page_range(tlb
, vma
, start
, end
, details
);
1323 * unmap_vmas - unmap a range of memory covered by a list of vma's
1324 * @tlb: address of the caller's struct mmu_gather
1325 * @vma: the starting vma
1326 * @start_addr: virtual address at which to start unmapping
1327 * @end_addr: virtual address at which to end unmapping
1329 * Unmap all pages in the vma list.
1331 * Only addresses between `start' and `end' will be unmapped.
1333 * The VMA list must be sorted in ascending virtual address order.
1335 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1336 * range after unmap_vmas() returns. So the only responsibility here is to
1337 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1338 * drops the lock and schedules.
1340 void unmap_vmas(struct mmu_gather
*tlb
,
1341 struct vm_area_struct
*vma
, unsigned long start_addr
,
1342 unsigned long end_addr
)
1344 struct mmu_notifier_range range
;
1346 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
, vma
->vm_mm
,
1347 start_addr
, end_addr
);
1348 mmu_notifier_invalidate_range_start(&range
);
1349 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1350 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1351 mmu_notifier_invalidate_range_end(&range
);
1355 * zap_page_range - remove user pages in a given range
1356 * @vma: vm_area_struct holding the applicable pages
1357 * @start: starting address of pages to zap
1358 * @size: number of bytes to zap
1360 * Caller must protect the VMA list
1362 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1365 struct mmu_notifier_range range
;
1366 struct mmu_gather tlb
;
1369 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1370 start
, start
+ size
);
1371 tlb_gather_mmu(&tlb
, vma
->vm_mm
, start
, range
.end
);
1372 update_hiwater_rss(vma
->vm_mm
);
1373 mmu_notifier_invalidate_range_start(&range
);
1374 for ( ; vma
&& vma
->vm_start
< range
.end
; vma
= vma
->vm_next
)
1375 unmap_single_vma(&tlb
, vma
, start
, range
.end
, NULL
);
1376 mmu_notifier_invalidate_range_end(&range
);
1377 tlb_finish_mmu(&tlb
, start
, range
.end
);
1379 EXPORT_SYMBOL(zap_page_range
);
1382 * zap_page_range_single - remove user pages in a given range
1383 * @vma: vm_area_struct holding the applicable pages
1384 * @address: starting address of pages to zap
1385 * @size: number of bytes to zap
1386 * @details: details of shared cache invalidation
1388 * The range must fit into one VMA.
1390 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1391 unsigned long size
, struct zap_details
*details
)
1393 struct mmu_notifier_range range
;
1394 struct mmu_gather tlb
;
1397 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1398 address
, address
+ size
);
1399 tlb_gather_mmu(&tlb
, vma
->vm_mm
, address
, range
.end
);
1400 update_hiwater_rss(vma
->vm_mm
);
1401 mmu_notifier_invalidate_range_start(&range
);
1402 unmap_single_vma(&tlb
, vma
, address
, range
.end
, details
);
1403 mmu_notifier_invalidate_range_end(&range
);
1404 tlb_finish_mmu(&tlb
, address
, range
.end
);
1408 * zap_vma_ptes - remove ptes mapping the vma
1409 * @vma: vm_area_struct holding ptes to be zapped
1410 * @address: starting address of pages to zap
1411 * @size: number of bytes to zap
1413 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1415 * The entire address range must be fully contained within the vma.
1418 void zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1421 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1422 !(vma
->vm_flags
& VM_PFNMAP
))
1425 zap_page_range_single(vma
, address
, size
, NULL
);
1427 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1429 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1437 pgd
= pgd_offset(mm
, addr
);
1438 p4d
= p4d_alloc(mm
, pgd
, addr
);
1441 pud
= pud_alloc(mm
, p4d
, addr
);
1444 pmd
= pmd_alloc(mm
, pud
, addr
);
1448 VM_BUG_ON(pmd_trans_huge(*pmd
));
1449 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1453 * This is the old fallback for page remapping.
1455 * For historical reasons, it only allows reserved pages. Only
1456 * old drivers should use this, and they needed to mark their
1457 * pages reserved for the old functions anyway.
1459 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1460 struct page
*page
, pgprot_t prot
)
1462 struct mm_struct
*mm
= vma
->vm_mm
;
1468 if (PageAnon(page
) || PageSlab(page
) || page_has_type(page
))
1471 flush_dcache_page(page
);
1472 pte
= get_locked_pte(mm
, addr
, &ptl
);
1476 if (!pte_none(*pte
))
1479 /* Ok, finally just insert the thing.. */
1481 inc_mm_counter_fast(mm
, mm_counter_file(page
));
1482 page_add_file_rmap(page
, false);
1483 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1487 pte_unmap_unlock(pte
, ptl
);
1493 * vm_insert_page - insert single page into user vma
1494 * @vma: user vma to map to
1495 * @addr: target user address of this page
1496 * @page: source kernel page
1498 * This allows drivers to insert individual pages they've allocated
1501 * The page has to be a nice clean _individual_ kernel allocation.
1502 * If you allocate a compound page, you need to have marked it as
1503 * such (__GFP_COMP), or manually just split the page up yourself
1504 * (see split_page()).
1506 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1507 * took an arbitrary page protection parameter. This doesn't allow
1508 * that. Your vma protection will have to be set up correctly, which
1509 * means that if you want a shared writable mapping, you'd better
1510 * ask for a shared writable mapping!
1512 * The page does not need to be reserved.
1514 * Usually this function is called from f_op->mmap() handler
1515 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1516 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1517 * function from other places, for example from page-fault handler.
1519 * Return: %0 on success, negative error code otherwise.
1521 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1524 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1526 if (!page_count(page
))
1528 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1529 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
1530 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1531 vma
->vm_flags
|= VM_MIXEDMAP
;
1533 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1535 EXPORT_SYMBOL(vm_insert_page
);
1538 * __vm_map_pages - maps range of kernel pages into user vma
1539 * @vma: user vma to map to
1540 * @pages: pointer to array of source kernel pages
1541 * @num: number of pages in page array
1542 * @offset: user's requested vm_pgoff
1544 * This allows drivers to map range of kernel pages into a user vma.
1546 * Return: 0 on success and error code otherwise.
1548 static int __vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
1549 unsigned long num
, unsigned long offset
)
1551 unsigned long count
= vma_pages(vma
);
1552 unsigned long uaddr
= vma
->vm_start
;
1555 /* Fail if the user requested offset is beyond the end of the object */
1559 /* Fail if the user requested size exceeds available object size */
1560 if (count
> num
- offset
)
1563 for (i
= 0; i
< count
; i
++) {
1564 ret
= vm_insert_page(vma
, uaddr
, pages
[offset
+ i
]);
1574 * vm_map_pages - maps range of kernel pages starts with non zero offset
1575 * @vma: user vma to map to
1576 * @pages: pointer to array of source kernel pages
1577 * @num: number of pages in page array
1579 * Maps an object consisting of @num pages, catering for the user's
1580 * requested vm_pgoff
1582 * If we fail to insert any page into the vma, the function will return
1583 * immediately leaving any previously inserted pages present. Callers
1584 * from the mmap handler may immediately return the error as their caller
1585 * will destroy the vma, removing any successfully inserted pages. Other
1586 * callers should make their own arrangements for calling unmap_region().
1588 * Context: Process context. Called by mmap handlers.
1589 * Return: 0 on success and error code otherwise.
1591 int vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
1594 return __vm_map_pages(vma
, pages
, num
, vma
->vm_pgoff
);
1596 EXPORT_SYMBOL(vm_map_pages
);
1599 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1600 * @vma: user vma to map to
1601 * @pages: pointer to array of source kernel pages
1602 * @num: number of pages in page array
1604 * Similar to vm_map_pages(), except that it explicitly sets the offset
1605 * to 0. This function is intended for the drivers that did not consider
1608 * Context: Process context. Called by mmap handlers.
1609 * Return: 0 on success and error code otherwise.
1611 int vm_map_pages_zero(struct vm_area_struct
*vma
, struct page
**pages
,
1614 return __vm_map_pages(vma
, pages
, num
, 0);
1616 EXPORT_SYMBOL(vm_map_pages_zero
);
1618 static vm_fault_t
insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1619 pfn_t pfn
, pgprot_t prot
, bool mkwrite
)
1621 struct mm_struct
*mm
= vma
->vm_mm
;
1625 pte
= get_locked_pte(mm
, addr
, &ptl
);
1627 return VM_FAULT_OOM
;
1628 if (!pte_none(*pte
)) {
1631 * For read faults on private mappings the PFN passed
1632 * in may not match the PFN we have mapped if the
1633 * mapped PFN is a writeable COW page. In the mkwrite
1634 * case we are creating a writable PTE for a shared
1635 * mapping and we expect the PFNs to match. If they
1636 * don't match, we are likely racing with block
1637 * allocation and mapping invalidation so just skip the
1640 if (pte_pfn(*pte
) != pfn_t_to_pfn(pfn
)) {
1641 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte
)));
1644 entry
= pte_mkyoung(*pte
);
1645 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1646 if (ptep_set_access_flags(vma
, addr
, pte
, entry
, 1))
1647 update_mmu_cache(vma
, addr
, pte
);
1652 /* Ok, finally just insert the thing.. */
1653 if (pfn_t_devmap(pfn
))
1654 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
1656 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
1659 entry
= pte_mkyoung(entry
);
1660 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1663 set_pte_at(mm
, addr
, pte
, entry
);
1664 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1667 pte_unmap_unlock(pte
, ptl
);
1668 return VM_FAULT_NOPAGE
;
1672 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1673 * @vma: user vma to map to
1674 * @addr: target user address of this page
1675 * @pfn: source kernel pfn
1676 * @pgprot: pgprot flags for the inserted page
1678 * This is exactly like vmf_insert_pfn(), except that it allows drivers to
1679 * to override pgprot on a per-page basis.
1681 * This only makes sense for IO mappings, and it makes no sense for
1682 * COW mappings. In general, using multiple vmas is preferable;
1683 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1686 * Context: Process context. May allocate using %GFP_KERNEL.
1687 * Return: vm_fault_t value.
1689 vm_fault_t
vmf_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
1690 unsigned long pfn
, pgprot_t pgprot
)
1693 * Technically, architectures with pte_special can avoid all these
1694 * restrictions (same for remap_pfn_range). However we would like
1695 * consistency in testing and feature parity among all, so we should
1696 * try to keep these invariants in place for everybody.
1698 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1699 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1700 (VM_PFNMAP
|VM_MIXEDMAP
));
1701 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1702 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1704 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1705 return VM_FAULT_SIGBUS
;
1707 if (!pfn_modify_allowed(pfn
, pgprot
))
1708 return VM_FAULT_SIGBUS
;
1710 track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
));
1712 return insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
,
1715 EXPORT_SYMBOL(vmf_insert_pfn_prot
);
1718 * vmf_insert_pfn - insert single pfn into user vma
1719 * @vma: user vma to map to
1720 * @addr: target user address of this page
1721 * @pfn: source kernel pfn
1723 * Similar to vm_insert_page, this allows drivers to insert individual pages
1724 * they've allocated into a user vma. Same comments apply.
1726 * This function should only be called from a vm_ops->fault handler, and
1727 * in that case the handler should return the result of this function.
1729 * vma cannot be a COW mapping.
1731 * As this is called only for pages that do not currently exist, we
1732 * do not need to flush old virtual caches or the TLB.
1734 * Context: Process context. May allocate using %GFP_KERNEL.
1735 * Return: vm_fault_t value.
1737 vm_fault_t
vmf_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1740 return vmf_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
1742 EXPORT_SYMBOL(vmf_insert_pfn
);
1744 static bool vm_mixed_ok(struct vm_area_struct
*vma
, pfn_t pfn
)
1746 /* these checks mirror the abort conditions in vm_normal_page */
1747 if (vma
->vm_flags
& VM_MIXEDMAP
)
1749 if (pfn_t_devmap(pfn
))
1751 if (pfn_t_special(pfn
))
1753 if (is_zero_pfn(pfn_t_to_pfn(pfn
)))
1758 static vm_fault_t
__vm_insert_mixed(struct vm_area_struct
*vma
,
1759 unsigned long addr
, pfn_t pfn
, bool mkwrite
)
1761 pgprot_t pgprot
= vma
->vm_page_prot
;
1764 BUG_ON(!vm_mixed_ok(vma
, pfn
));
1766 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1767 return VM_FAULT_SIGBUS
;
1769 track_pfn_insert(vma
, &pgprot
, pfn
);
1771 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn
), pgprot
))
1772 return VM_FAULT_SIGBUS
;
1775 * If we don't have pte special, then we have to use the pfn_valid()
1776 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1777 * refcount the page if pfn_valid is true (hence insert_page rather
1778 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1779 * without pte special, it would there be refcounted as a normal page.
1781 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
) &&
1782 !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
1786 * At this point we are committed to insert_page()
1787 * regardless of whether the caller specified flags that
1788 * result in pfn_t_has_page() == false.
1790 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
1791 err
= insert_page(vma
, addr
, page
, pgprot
);
1793 return insert_pfn(vma
, addr
, pfn
, pgprot
, mkwrite
);
1797 return VM_FAULT_OOM
;
1798 if (err
< 0 && err
!= -EBUSY
)
1799 return VM_FAULT_SIGBUS
;
1801 return VM_FAULT_NOPAGE
;
1804 vm_fault_t
vmf_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1807 return __vm_insert_mixed(vma
, addr
, pfn
, false);
1809 EXPORT_SYMBOL(vmf_insert_mixed
);
1812 * If the insertion of PTE failed because someone else already added a
1813 * different entry in the mean time, we treat that as success as we assume
1814 * the same entry was actually inserted.
1816 vm_fault_t
vmf_insert_mixed_mkwrite(struct vm_area_struct
*vma
,
1817 unsigned long addr
, pfn_t pfn
)
1819 return __vm_insert_mixed(vma
, addr
, pfn
, true);
1821 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite
);
1824 * maps a range of physical memory into the requested pages. the old
1825 * mappings are removed. any references to nonexistent pages results
1826 * in null mappings (currently treated as "copy-on-access")
1828 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1829 unsigned long addr
, unsigned long end
,
1830 unsigned long pfn
, pgprot_t prot
)
1832 pte_t
*pte
, *mapped_pte
;
1836 mapped_pte
= pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1839 arch_enter_lazy_mmu_mode();
1841 BUG_ON(!pte_none(*pte
));
1842 if (!pfn_modify_allowed(pfn
, prot
)) {
1846 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
1848 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1849 arch_leave_lazy_mmu_mode();
1850 pte_unmap_unlock(mapped_pte
, ptl
);
1854 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1855 unsigned long addr
, unsigned long end
,
1856 unsigned long pfn
, pgprot_t prot
)
1862 pfn
-= addr
>> PAGE_SHIFT
;
1863 pmd
= pmd_alloc(mm
, pud
, addr
);
1866 VM_BUG_ON(pmd_trans_huge(*pmd
));
1868 next
= pmd_addr_end(addr
, end
);
1869 err
= remap_pte_range(mm
, pmd
, addr
, next
,
1870 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1873 } while (pmd
++, addr
= next
, addr
!= end
);
1877 static inline int remap_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
1878 unsigned long addr
, unsigned long end
,
1879 unsigned long pfn
, pgprot_t prot
)
1885 pfn
-= addr
>> PAGE_SHIFT
;
1886 pud
= pud_alloc(mm
, p4d
, addr
);
1890 next
= pud_addr_end(addr
, end
);
1891 err
= remap_pmd_range(mm
, pud
, addr
, next
,
1892 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1895 } while (pud
++, addr
= next
, addr
!= end
);
1899 static inline int remap_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1900 unsigned long addr
, unsigned long end
,
1901 unsigned long pfn
, pgprot_t prot
)
1907 pfn
-= addr
>> PAGE_SHIFT
;
1908 p4d
= p4d_alloc(mm
, pgd
, addr
);
1912 next
= p4d_addr_end(addr
, end
);
1913 err
= remap_pud_range(mm
, p4d
, addr
, next
,
1914 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1917 } while (p4d
++, addr
= next
, addr
!= end
);
1922 * remap_pfn_range - remap kernel memory to userspace
1923 * @vma: user vma to map to
1924 * @addr: target user address to start at
1925 * @pfn: physical address of kernel memory
1926 * @size: size of map area
1927 * @prot: page protection flags for this mapping
1929 * Note: this is only safe if the mm semaphore is held when called.
1931 * Return: %0 on success, negative error code otherwise.
1933 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1934 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1938 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1939 struct mm_struct
*mm
= vma
->vm_mm
;
1940 unsigned long remap_pfn
= pfn
;
1944 * Physically remapped pages are special. Tell the
1945 * rest of the world about it:
1946 * VM_IO tells people not to look at these pages
1947 * (accesses can have side effects).
1948 * VM_PFNMAP tells the core MM that the base pages are just
1949 * raw PFN mappings, and do not have a "struct page" associated
1952 * Disable vma merging and expanding with mremap().
1954 * Omit vma from core dump, even when VM_IO turned off.
1956 * There's a horrible special case to handle copy-on-write
1957 * behaviour that some programs depend on. We mark the "original"
1958 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1959 * See vm_normal_page() for details.
1961 if (is_cow_mapping(vma
->vm_flags
)) {
1962 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
1964 vma
->vm_pgoff
= pfn
;
1967 err
= track_pfn_remap(vma
, &prot
, remap_pfn
, addr
, PAGE_ALIGN(size
));
1971 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
1973 BUG_ON(addr
>= end
);
1974 pfn
-= addr
>> PAGE_SHIFT
;
1975 pgd
= pgd_offset(mm
, addr
);
1976 flush_cache_range(vma
, addr
, end
);
1978 next
= pgd_addr_end(addr
, end
);
1979 err
= remap_p4d_range(mm
, pgd
, addr
, next
,
1980 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1983 } while (pgd
++, addr
= next
, addr
!= end
);
1986 untrack_pfn(vma
, remap_pfn
, PAGE_ALIGN(size
));
1990 EXPORT_SYMBOL(remap_pfn_range
);
1993 * vm_iomap_memory - remap memory to userspace
1994 * @vma: user vma to map to
1995 * @start: start of area
1996 * @len: size of area
1998 * This is a simplified io_remap_pfn_range() for common driver use. The
1999 * driver just needs to give us the physical memory range to be mapped,
2000 * we'll figure out the rest from the vma information.
2002 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2003 * whatever write-combining details or similar.
2005 * Return: %0 on success, negative error code otherwise.
2007 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2009 unsigned long vm_len
, pfn
, pages
;
2011 /* Check that the physical memory area passed in looks valid */
2012 if (start
+ len
< start
)
2015 * You *really* shouldn't map things that aren't page-aligned,
2016 * but we've historically allowed it because IO memory might
2017 * just have smaller alignment.
2019 len
+= start
& ~PAGE_MASK
;
2020 pfn
= start
>> PAGE_SHIFT
;
2021 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2022 if (pfn
+ pages
< pfn
)
2025 /* We start the mapping 'vm_pgoff' pages into the area */
2026 if (vma
->vm_pgoff
> pages
)
2028 pfn
+= vma
->vm_pgoff
;
2029 pages
-= vma
->vm_pgoff
;
2031 /* Can we fit all of the mapping? */
2032 vm_len
= vma
->vm_end
- vma
->vm_start
;
2033 if (vm_len
>> PAGE_SHIFT
> pages
)
2036 /* Ok, let it rip */
2037 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2039 EXPORT_SYMBOL(vm_iomap_memory
);
2041 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2042 unsigned long addr
, unsigned long end
,
2043 pte_fn_t fn
, void *data
)
2047 spinlock_t
*uninitialized_var(ptl
);
2049 pte
= (mm
== &init_mm
) ?
2050 pte_alloc_kernel(pmd
, addr
) :
2051 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2055 BUG_ON(pmd_huge(*pmd
));
2057 arch_enter_lazy_mmu_mode();
2060 err
= fn(pte
++, addr
, data
);
2063 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2065 arch_leave_lazy_mmu_mode();
2068 pte_unmap_unlock(pte
-1, ptl
);
2072 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2073 unsigned long addr
, unsigned long end
,
2074 pte_fn_t fn
, void *data
)
2080 BUG_ON(pud_huge(*pud
));
2082 pmd
= pmd_alloc(mm
, pud
, addr
);
2086 next
= pmd_addr_end(addr
, end
);
2087 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
2090 } while (pmd
++, addr
= next
, addr
!= end
);
2094 static int apply_to_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2095 unsigned long addr
, unsigned long end
,
2096 pte_fn_t fn
, void *data
)
2102 pud
= pud_alloc(mm
, p4d
, addr
);
2106 next
= pud_addr_end(addr
, end
);
2107 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
2110 } while (pud
++, addr
= next
, addr
!= end
);
2114 static int apply_to_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2115 unsigned long addr
, unsigned long end
,
2116 pte_fn_t fn
, void *data
)
2122 p4d
= p4d_alloc(mm
, pgd
, addr
);
2126 next
= p4d_addr_end(addr
, end
);
2127 err
= apply_to_pud_range(mm
, p4d
, addr
, next
, fn
, data
);
2130 } while (p4d
++, addr
= next
, addr
!= end
);
2135 * Scan a region of virtual memory, filling in page tables as necessary
2136 * and calling a provided function on each leaf page table.
2138 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2139 unsigned long size
, pte_fn_t fn
, void *data
)
2143 unsigned long end
= addr
+ size
;
2146 if (WARN_ON(addr
>= end
))
2149 pgd
= pgd_offset(mm
, addr
);
2151 next
= pgd_addr_end(addr
, end
);
2152 err
= apply_to_p4d_range(mm
, pgd
, addr
, next
, fn
, data
);
2155 } while (pgd
++, addr
= next
, addr
!= end
);
2159 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2162 * handle_pte_fault chooses page fault handler according to an entry which was
2163 * read non-atomically. Before making any commitment, on those architectures
2164 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2165 * parts, do_swap_page must check under lock before unmapping the pte and
2166 * proceeding (but do_wp_page is only called after already making such a check;
2167 * and do_anonymous_page can safely check later on).
2169 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2170 pte_t
*page_table
, pte_t orig_pte
)
2173 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2174 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2175 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2177 same
= pte_same(*page_table
, orig_pte
);
2181 pte_unmap(page_table
);
2185 static inline bool cow_user_page(struct page
*dst
, struct page
*src
,
2186 struct vm_fault
*vmf
)
2191 bool locked
= false;
2192 struct vm_area_struct
*vma
= vmf
->vma
;
2193 struct mm_struct
*mm
= vma
->vm_mm
;
2194 unsigned long addr
= vmf
->address
;
2196 debug_dma_assert_idle(src
);
2199 copy_user_highpage(dst
, src
, addr
, vma
);
2204 * If the source page was a PFN mapping, we don't have
2205 * a "struct page" for it. We do a best-effort copy by
2206 * just copying from the original user address. If that
2207 * fails, we just zero-fill it. Live with it.
2209 kaddr
= kmap_atomic(dst
);
2210 uaddr
= (void __user
*)(addr
& PAGE_MASK
);
2213 * On architectures with software "accessed" bits, we would
2214 * take a double page fault, so mark it accessed here.
2216 if (arch_faults_on_old_pte() && !pte_young(vmf
->orig_pte
)) {
2219 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2221 if (!likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2223 * Other thread has already handled the fault
2224 * and we don't need to do anything. If it's
2225 * not the case, the fault will be triggered
2226 * again on the same address.
2232 entry
= pte_mkyoung(vmf
->orig_pte
);
2233 if (ptep_set_access_flags(vma
, addr
, vmf
->pte
, entry
, 0))
2234 update_mmu_cache(vma
, addr
, vmf
->pte
);
2238 * This really shouldn't fail, because the page is there
2239 * in the page tables. But it might just be unreadable,
2240 * in which case we just give up and fill the result with
2243 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
2247 /* Re-validate under PTL if the page is still mapped */
2248 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2250 if (!likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2251 /* The PTE changed under us. Retry page fault. */
2257 * The same page can be mapped back since last copy attampt.
2258 * Try to copy again under PTL.
2260 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
2262 * Give a warn in case there can be some obscure
2275 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2276 kunmap_atomic(kaddr
);
2277 flush_dcache_page(dst
);
2282 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
2284 struct file
*vm_file
= vma
->vm_file
;
2287 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
2290 * Special mappings (e.g. VDSO) do not have any file so fake
2291 * a default GFP_KERNEL for them.
2297 * Notify the address space that the page is about to become writable so that
2298 * it can prohibit this or wait for the page to get into an appropriate state.
2300 * We do this without the lock held, so that it can sleep if it needs to.
2302 static vm_fault_t
do_page_mkwrite(struct vm_fault
*vmf
)
2305 struct page
*page
= vmf
->page
;
2306 unsigned int old_flags
= vmf
->flags
;
2308 vmf
->flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2310 if (vmf
->vma
->vm_file
&&
2311 IS_SWAPFILE(vmf
->vma
->vm_file
->f_mapping
->host
))
2312 return VM_FAULT_SIGBUS
;
2314 ret
= vmf
->vma
->vm_ops
->page_mkwrite(vmf
);
2315 /* Restore original flags so that caller is not surprised */
2316 vmf
->flags
= old_flags
;
2317 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2319 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
2321 if (!page
->mapping
) {
2323 return 0; /* retry */
2325 ret
|= VM_FAULT_LOCKED
;
2327 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2332 * Handle dirtying of a page in shared file mapping on a write fault.
2334 * The function expects the page to be locked and unlocks it.
2336 static vm_fault_t
fault_dirty_shared_page(struct vm_fault
*vmf
)
2338 struct vm_area_struct
*vma
= vmf
->vma
;
2339 struct address_space
*mapping
;
2340 struct page
*page
= vmf
->page
;
2342 bool page_mkwrite
= vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
;
2344 dirtied
= set_page_dirty(page
);
2345 VM_BUG_ON_PAGE(PageAnon(page
), page
);
2347 * Take a local copy of the address_space - page.mapping may be zeroed
2348 * by truncate after unlock_page(). The address_space itself remains
2349 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2350 * release semantics to prevent the compiler from undoing this copying.
2352 mapping
= page_rmapping(page
);
2356 file_update_time(vma
->vm_file
);
2359 * Throttle page dirtying rate down to writeback speed.
2361 * mapping may be NULL here because some device drivers do not
2362 * set page.mapping but still dirty their pages
2364 * Drop the mmap_sem before waiting on IO, if we can. The file
2365 * is pinning the mapping, as per above.
2367 if ((dirtied
|| page_mkwrite
) && mapping
) {
2370 fpin
= maybe_unlock_mmap_for_io(vmf
, NULL
);
2371 balance_dirty_pages_ratelimited(mapping
);
2374 return VM_FAULT_RETRY
;
2382 * Handle write page faults for pages that can be reused in the current vma
2384 * This can happen either due to the mapping being with the VM_SHARED flag,
2385 * or due to us being the last reference standing to the page. In either
2386 * case, all we need to do here is to mark the page as writable and update
2387 * any related book-keeping.
2389 static inline void wp_page_reuse(struct vm_fault
*vmf
)
2390 __releases(vmf
->ptl
)
2392 struct vm_area_struct
*vma
= vmf
->vma
;
2393 struct page
*page
= vmf
->page
;
2396 * Clear the pages cpupid information as the existing
2397 * information potentially belongs to a now completely
2398 * unrelated process.
2401 page_cpupid_xchg_last(page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2403 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2404 entry
= pte_mkyoung(vmf
->orig_pte
);
2405 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2406 if (ptep_set_access_flags(vma
, vmf
->address
, vmf
->pte
, entry
, 1))
2407 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2408 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2412 * Handle the case of a page which we actually need to copy to a new page.
2414 * Called with mmap_sem locked and the old page referenced, but
2415 * without the ptl held.
2417 * High level logic flow:
2419 * - Allocate a page, copy the content of the old page to the new one.
2420 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2421 * - Take the PTL. If the pte changed, bail out and release the allocated page
2422 * - If the pte is still the way we remember it, update the page table and all
2423 * relevant references. This includes dropping the reference the page-table
2424 * held to the old page, as well as updating the rmap.
2425 * - In any case, unlock the PTL and drop the reference we took to the old page.
2427 static vm_fault_t
wp_page_copy(struct vm_fault
*vmf
)
2429 struct vm_area_struct
*vma
= vmf
->vma
;
2430 struct mm_struct
*mm
= vma
->vm_mm
;
2431 struct page
*old_page
= vmf
->page
;
2432 struct page
*new_page
= NULL
;
2434 int page_copied
= 0;
2435 struct mem_cgroup
*memcg
;
2436 struct mmu_notifier_range range
;
2438 if (unlikely(anon_vma_prepare(vma
)))
2441 if (is_zero_pfn(pte_pfn(vmf
->orig_pte
))) {
2442 new_page
= alloc_zeroed_user_highpage_movable(vma
,
2447 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
2452 if (!cow_user_page(new_page
, old_page
, vmf
)) {
2454 * COW failed, if the fault was solved by other,
2455 * it's fine. If not, userspace would re-fault on
2456 * the same address and we will handle the fault
2457 * from the second attempt.
2466 if (mem_cgroup_try_charge_delay(new_page
, mm
, GFP_KERNEL
, &memcg
, false))
2469 __SetPageUptodate(new_page
);
2471 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
,
2472 vmf
->address
& PAGE_MASK
,
2473 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
);
2474 mmu_notifier_invalidate_range_start(&range
);
2477 * Re-check the pte - we dropped the lock
2479 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, vmf
->address
, &vmf
->ptl
);
2480 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2482 if (!PageAnon(old_page
)) {
2483 dec_mm_counter_fast(mm
,
2484 mm_counter_file(old_page
));
2485 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2488 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2490 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2491 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2492 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2494 * Clear the pte entry and flush it first, before updating the
2495 * pte with the new entry. This will avoid a race condition
2496 * seen in the presence of one thread doing SMC and another
2499 ptep_clear_flush_notify(vma
, vmf
->address
, vmf
->pte
);
2500 page_add_new_anon_rmap(new_page
, vma
, vmf
->address
, false);
2501 mem_cgroup_commit_charge(new_page
, memcg
, false, false);
2502 lru_cache_add_active_or_unevictable(new_page
, vma
);
2504 * We call the notify macro here because, when using secondary
2505 * mmu page tables (such as kvm shadow page tables), we want the
2506 * new page to be mapped directly into the secondary page table.
2508 set_pte_at_notify(mm
, vmf
->address
, vmf
->pte
, entry
);
2509 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2512 * Only after switching the pte to the new page may
2513 * we remove the mapcount here. Otherwise another
2514 * process may come and find the rmap count decremented
2515 * before the pte is switched to the new page, and
2516 * "reuse" the old page writing into it while our pte
2517 * here still points into it and can be read by other
2520 * The critical issue is to order this
2521 * page_remove_rmap with the ptp_clear_flush above.
2522 * Those stores are ordered by (if nothing else,)
2523 * the barrier present in the atomic_add_negative
2524 * in page_remove_rmap.
2526 * Then the TLB flush in ptep_clear_flush ensures that
2527 * no process can access the old page before the
2528 * decremented mapcount is visible. And the old page
2529 * cannot be reused until after the decremented
2530 * mapcount is visible. So transitively, TLBs to
2531 * old page will be flushed before it can be reused.
2533 page_remove_rmap(old_page
, false);
2536 /* Free the old page.. */
2537 new_page
= old_page
;
2540 mem_cgroup_cancel_charge(new_page
, memcg
, false);
2546 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2548 * No need to double call mmu_notifier->invalidate_range() callback as
2549 * the above ptep_clear_flush_notify() did already call it.
2551 mmu_notifier_invalidate_range_only_end(&range
);
2554 * Don't let another task, with possibly unlocked vma,
2555 * keep the mlocked page.
2557 if (page_copied
&& (vma
->vm_flags
& VM_LOCKED
)) {
2558 lock_page(old_page
); /* LRU manipulation */
2559 if (PageMlocked(old_page
))
2560 munlock_vma_page(old_page
);
2561 unlock_page(old_page
);
2565 return page_copied
? VM_FAULT_WRITE
: 0;
2571 return VM_FAULT_OOM
;
2575 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2576 * writeable once the page is prepared
2578 * @vmf: structure describing the fault
2580 * This function handles all that is needed to finish a write page fault in a
2581 * shared mapping due to PTE being read-only once the mapped page is prepared.
2582 * It handles locking of PTE and modifying it.
2584 * The function expects the page to be locked or other protection against
2585 * concurrent faults / writeback (such as DAX radix tree locks).
2587 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2588 * we acquired PTE lock.
2590 vm_fault_t
finish_mkwrite_fault(struct vm_fault
*vmf
)
2592 WARN_ON_ONCE(!(vmf
->vma
->vm_flags
& VM_SHARED
));
2593 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
2596 * We might have raced with another page fault while we released the
2597 * pte_offset_map_lock.
2599 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
2600 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2601 return VM_FAULT_NOPAGE
;
2608 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2611 static vm_fault_t
wp_pfn_shared(struct vm_fault
*vmf
)
2613 struct vm_area_struct
*vma
= vmf
->vma
;
2615 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
2618 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2619 vmf
->flags
|= FAULT_FLAG_MKWRITE
;
2620 ret
= vma
->vm_ops
->pfn_mkwrite(vmf
);
2621 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))
2623 return finish_mkwrite_fault(vmf
);
2626 return VM_FAULT_WRITE
;
2629 static vm_fault_t
wp_page_shared(struct vm_fault
*vmf
)
2630 __releases(vmf
->ptl
)
2632 struct vm_area_struct
*vma
= vmf
->vma
;
2633 vm_fault_t ret
= VM_FAULT_WRITE
;
2635 get_page(vmf
->page
);
2637 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2640 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2641 tmp
= do_page_mkwrite(vmf
);
2642 if (unlikely(!tmp
|| (tmp
&
2643 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
2644 put_page(vmf
->page
);
2647 tmp
= finish_mkwrite_fault(vmf
);
2648 if (unlikely(tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
2649 unlock_page(vmf
->page
);
2650 put_page(vmf
->page
);
2655 lock_page(vmf
->page
);
2657 ret
|= fault_dirty_shared_page(vmf
);
2658 put_page(vmf
->page
);
2664 * This routine handles present pages, when users try to write
2665 * to a shared page. It is done by copying the page to a new address
2666 * and decrementing the shared-page counter for the old page.
2668 * Note that this routine assumes that the protection checks have been
2669 * done by the caller (the low-level page fault routine in most cases).
2670 * Thus we can safely just mark it writable once we've done any necessary
2673 * We also mark the page dirty at this point even though the page will
2674 * change only once the write actually happens. This avoids a few races,
2675 * and potentially makes it more efficient.
2677 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2678 * but allow concurrent faults), with pte both mapped and locked.
2679 * We return with mmap_sem still held, but pte unmapped and unlocked.
2681 static vm_fault_t
do_wp_page(struct vm_fault
*vmf
)
2682 __releases(vmf
->ptl
)
2684 struct vm_area_struct
*vma
= vmf
->vma
;
2686 vmf
->page
= vm_normal_page(vma
, vmf
->address
, vmf
->orig_pte
);
2689 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2692 * We should not cow pages in a shared writeable mapping.
2693 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2695 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2696 (VM_WRITE
|VM_SHARED
))
2697 return wp_pfn_shared(vmf
);
2699 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2700 return wp_page_copy(vmf
);
2704 * Take out anonymous pages first, anonymous shared vmas are
2705 * not dirty accountable.
2707 if (PageAnon(vmf
->page
)) {
2708 int total_map_swapcount
;
2709 if (PageKsm(vmf
->page
) && (PageSwapCache(vmf
->page
) ||
2710 page_count(vmf
->page
) != 1))
2712 if (!trylock_page(vmf
->page
)) {
2713 get_page(vmf
->page
);
2714 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2715 lock_page(vmf
->page
);
2716 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
2717 vmf
->address
, &vmf
->ptl
);
2718 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
2719 unlock_page(vmf
->page
);
2720 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2721 put_page(vmf
->page
);
2724 put_page(vmf
->page
);
2726 if (PageKsm(vmf
->page
)) {
2727 bool reused
= reuse_ksm_page(vmf
->page
, vmf
->vma
,
2729 unlock_page(vmf
->page
);
2733 return VM_FAULT_WRITE
;
2735 if (reuse_swap_page(vmf
->page
, &total_map_swapcount
)) {
2736 if (total_map_swapcount
== 1) {
2738 * The page is all ours. Move it to
2739 * our anon_vma so the rmap code will
2740 * not search our parent or siblings.
2741 * Protected against the rmap code by
2744 page_move_anon_rmap(vmf
->page
, vma
);
2746 unlock_page(vmf
->page
);
2748 return VM_FAULT_WRITE
;
2750 unlock_page(vmf
->page
);
2751 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2752 (VM_WRITE
|VM_SHARED
))) {
2753 return wp_page_shared(vmf
);
2757 * Ok, we need to copy. Oh, well..
2759 get_page(vmf
->page
);
2761 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2762 return wp_page_copy(vmf
);
2765 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2766 unsigned long start_addr
, unsigned long end_addr
,
2767 struct zap_details
*details
)
2769 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2772 static inline void unmap_mapping_range_tree(struct rb_root_cached
*root
,
2773 struct zap_details
*details
)
2775 struct vm_area_struct
*vma
;
2776 pgoff_t vba
, vea
, zba
, zea
;
2778 vma_interval_tree_foreach(vma
, root
,
2779 details
->first_index
, details
->last_index
) {
2781 vba
= vma
->vm_pgoff
;
2782 vea
= vba
+ vma_pages(vma
) - 1;
2783 zba
= details
->first_index
;
2786 zea
= details
->last_index
;
2790 unmap_mapping_range_vma(vma
,
2791 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2792 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2798 * unmap_mapping_page() - Unmap single page from processes.
2799 * @page: The locked page to be unmapped.
2801 * Unmap this page from any userspace process which still has it mmaped.
2802 * Typically, for efficiency, the range of nearby pages has already been
2803 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
2804 * truncation or invalidation holds the lock on a page, it may find that
2805 * the page has been remapped again: and then uses unmap_mapping_page()
2806 * to unmap it finally.
2808 void unmap_mapping_page(struct page
*page
)
2810 struct address_space
*mapping
= page
->mapping
;
2811 struct zap_details details
= { };
2813 VM_BUG_ON(!PageLocked(page
));
2814 VM_BUG_ON(PageTail(page
));
2816 details
.check_mapping
= mapping
;
2817 details
.first_index
= page
->index
;
2818 details
.last_index
= page
->index
+ hpage_nr_pages(page
) - 1;
2819 details
.single_page
= page
;
2821 i_mmap_lock_write(mapping
);
2822 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
2823 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2824 i_mmap_unlock_write(mapping
);
2828 * unmap_mapping_pages() - Unmap pages from processes.
2829 * @mapping: The address space containing pages to be unmapped.
2830 * @start: Index of first page to be unmapped.
2831 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2832 * @even_cows: Whether to unmap even private COWed pages.
2834 * Unmap the pages in this address space from any userspace process which
2835 * has them mmaped. Generally, you want to remove COWed pages as well when
2836 * a file is being truncated, but not when invalidating pages from the page
2839 void unmap_mapping_pages(struct address_space
*mapping
, pgoff_t start
,
2840 pgoff_t nr
, bool even_cows
)
2842 struct zap_details details
= { };
2844 details
.check_mapping
= even_cows
? NULL
: mapping
;
2845 details
.first_index
= start
;
2846 details
.last_index
= start
+ nr
- 1;
2847 if (details
.last_index
< details
.first_index
)
2848 details
.last_index
= ULONG_MAX
;
2850 i_mmap_lock_write(mapping
);
2851 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
2852 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2853 i_mmap_unlock_write(mapping
);
2857 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2858 * address_space corresponding to the specified byte range in the underlying
2861 * @mapping: the address space containing mmaps to be unmapped.
2862 * @holebegin: byte in first page to unmap, relative to the start of
2863 * the underlying file. This will be rounded down to a PAGE_SIZE
2864 * boundary. Note that this is different from truncate_pagecache(), which
2865 * must keep the partial page. In contrast, we must get rid of
2867 * @holelen: size of prospective hole in bytes. This will be rounded
2868 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2870 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2871 * but 0 when invalidating pagecache, don't throw away private data.
2873 void unmap_mapping_range(struct address_space
*mapping
,
2874 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2876 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2877 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2879 /* Check for overflow. */
2880 if (sizeof(holelen
) > sizeof(hlen
)) {
2882 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2883 if (holeend
& ~(long long)ULONG_MAX
)
2884 hlen
= ULONG_MAX
- hba
+ 1;
2887 unmap_mapping_pages(mapping
, hba
, hlen
, even_cows
);
2889 EXPORT_SYMBOL(unmap_mapping_range
);
2892 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2893 * but allow concurrent faults), and pte mapped but not yet locked.
2894 * We return with pte unmapped and unlocked.
2896 * We return with the mmap_sem locked or unlocked in the same cases
2897 * as does filemap_fault().
2899 vm_fault_t
do_swap_page(struct vm_fault
*vmf
)
2901 struct vm_area_struct
*vma
= vmf
->vma
;
2902 struct page
*page
= NULL
, *swapcache
;
2903 struct mem_cgroup
*memcg
;
2910 if (!pte_unmap_same(vma
->vm_mm
, vmf
->pmd
, vmf
->pte
, vmf
->orig_pte
))
2913 entry
= pte_to_swp_entry(vmf
->orig_pte
);
2914 if (unlikely(non_swap_entry(entry
))) {
2915 if (is_migration_entry(entry
)) {
2916 migration_entry_wait(vma
->vm_mm
, vmf
->pmd
,
2918 } else if (is_device_private_entry(entry
)) {
2919 vmf
->page
= device_private_entry_to_page(entry
);
2920 ret
= vmf
->page
->pgmap
->ops
->migrate_to_ram(vmf
);
2921 } else if (is_hwpoison_entry(entry
)) {
2922 ret
= VM_FAULT_HWPOISON
;
2924 print_bad_pte(vma
, vmf
->address
, vmf
->orig_pte
, NULL
);
2925 ret
= VM_FAULT_SIGBUS
;
2931 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2932 page
= lookup_swap_cache(entry
, vma
, vmf
->address
);
2936 struct swap_info_struct
*si
= swp_swap_info(entry
);
2938 if (si
->flags
& SWP_SYNCHRONOUS_IO
&&
2939 __swap_count(entry
) == 1) {
2940 /* skip swapcache */
2941 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
2944 __SetPageLocked(page
);
2945 __SetPageSwapBacked(page
);
2946 set_page_private(page
, entry
.val
);
2947 lru_cache_add_anon(page
);
2948 swap_readpage(page
, true);
2951 page
= swapin_readahead(entry
, GFP_HIGHUSER_MOVABLE
,
2958 * Back out if somebody else faulted in this pte
2959 * while we released the pte lock.
2961 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
2962 vmf
->address
, &vmf
->ptl
);
2963 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
)))
2965 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2969 /* Had to read the page from swap area: Major fault */
2970 ret
= VM_FAULT_MAJOR
;
2971 count_vm_event(PGMAJFAULT
);
2972 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
2973 } else if (PageHWPoison(page
)) {
2975 * hwpoisoned dirty swapcache pages are kept for killing
2976 * owner processes (which may be unknown at hwpoison time)
2978 ret
= VM_FAULT_HWPOISON
;
2979 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2983 locked
= lock_page_or_retry(page
, vma
->vm_mm
, vmf
->flags
);
2985 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2987 ret
|= VM_FAULT_RETRY
;
2992 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2993 * release the swapcache from under us. The page pin, and pte_same
2994 * test below, are not enough to exclude that. Even if it is still
2995 * swapcache, we need to check that the page's swap has not changed.
2997 if (unlikely((!PageSwapCache(page
) ||
2998 page_private(page
) != entry
.val
)) && swapcache
)
3001 page
= ksm_might_need_to_copy(page
, vma
, vmf
->address
);
3002 if (unlikely(!page
)) {
3008 if (mem_cgroup_try_charge_delay(page
, vma
->vm_mm
, GFP_KERNEL
,
3015 * Back out if somebody else already faulted in this pte.
3017 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3019 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3022 if (unlikely(!PageUptodate(page
))) {
3023 ret
= VM_FAULT_SIGBUS
;
3028 * The page isn't present yet, go ahead with the fault.
3030 * Be careful about the sequence of operations here.
3031 * To get its accounting right, reuse_swap_page() must be called
3032 * while the page is counted on swap but not yet in mapcount i.e.
3033 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3034 * must be called after the swap_free(), or it will never succeed.
3037 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3038 dec_mm_counter_fast(vma
->vm_mm
, MM_SWAPENTS
);
3039 pte
= mk_pte(page
, vma
->vm_page_prot
);
3040 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
, NULL
)) {
3041 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3042 vmf
->flags
&= ~FAULT_FLAG_WRITE
;
3043 ret
|= VM_FAULT_WRITE
;
3044 exclusive
= RMAP_EXCLUSIVE
;
3046 flush_icache_page(vma
, page
);
3047 if (pte_swp_soft_dirty(vmf
->orig_pte
))
3048 pte
= pte_mksoft_dirty(pte
);
3049 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3050 arch_do_swap_page(vma
->vm_mm
, vma
, vmf
->address
, pte
, vmf
->orig_pte
);
3051 vmf
->orig_pte
= pte
;
3053 /* ksm created a completely new copy */
3054 if (unlikely(page
!= swapcache
&& swapcache
)) {
3055 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3056 mem_cgroup_commit_charge(page
, memcg
, false, false);
3057 lru_cache_add_active_or_unevictable(page
, vma
);
3059 do_page_add_anon_rmap(page
, vma
, vmf
->address
, exclusive
);
3060 mem_cgroup_commit_charge(page
, memcg
, true, false);
3061 activate_page(page
);
3065 if (mem_cgroup_swap_full(page
) ||
3066 (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
3067 try_to_free_swap(page
);
3069 if (page
!= swapcache
&& swapcache
) {
3071 * Hold the lock to avoid the swap entry to be reused
3072 * until we take the PT lock for the pte_same() check
3073 * (to avoid false positives from pte_same). For
3074 * further safety release the lock after the swap_free
3075 * so that the swap count won't change under a
3076 * parallel locked swapcache.
3078 unlock_page(swapcache
);
3079 put_page(swapcache
);
3082 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
3083 ret
|= do_wp_page(vmf
);
3084 if (ret
& VM_FAULT_ERROR
)
3085 ret
&= VM_FAULT_ERROR
;
3089 /* No need to invalidate - it was non-present before */
3090 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3092 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3096 mem_cgroup_cancel_charge(page
, memcg
, false);
3097 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3102 if (page
!= swapcache
&& swapcache
) {
3103 unlock_page(swapcache
);
3104 put_page(swapcache
);
3110 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3111 * but allow concurrent faults), and pte mapped but not yet locked.
3112 * We return with mmap_sem still held, but pte unmapped and unlocked.
3114 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
)
3116 struct vm_area_struct
*vma
= vmf
->vma
;
3117 struct mem_cgroup
*memcg
;
3122 /* File mapping without ->vm_ops ? */
3123 if (vma
->vm_flags
& VM_SHARED
)
3124 return VM_FAULT_SIGBUS
;
3127 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3128 * pte_offset_map() on pmds where a huge pmd might be created
3129 * from a different thread.
3131 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3132 * parallel threads are excluded by other means.
3134 * Here we only have down_read(mmap_sem).
3136 if (pte_alloc(vma
->vm_mm
, vmf
->pmd
))
3137 return VM_FAULT_OOM
;
3139 /* See the comment in pte_alloc_one_map() */
3140 if (unlikely(pmd_trans_unstable(vmf
->pmd
)))
3143 /* Use the zero-page for reads */
3144 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
3145 !mm_forbids_zeropage(vma
->vm_mm
)) {
3146 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(vmf
->address
),
3147 vma
->vm_page_prot
));
3148 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3149 vmf
->address
, &vmf
->ptl
);
3150 if (!pte_none(*vmf
->pte
))
3152 ret
= check_stable_address_space(vma
->vm_mm
);
3155 /* Deliver the page fault to userland, check inside PT lock */
3156 if (userfaultfd_missing(vma
)) {
3157 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3158 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3163 /* Allocate our own private page. */
3164 if (unlikely(anon_vma_prepare(vma
)))
3166 page
= alloc_zeroed_user_highpage_movable(vma
, vmf
->address
);
3170 if (mem_cgroup_try_charge_delay(page
, vma
->vm_mm
, GFP_KERNEL
, &memcg
,
3175 * The memory barrier inside __SetPageUptodate makes sure that
3176 * preceeding stores to the page contents become visible before
3177 * the set_pte_at() write.
3179 __SetPageUptodate(page
);
3181 entry
= mk_pte(page
, vma
->vm_page_prot
);
3182 if (vma
->vm_flags
& VM_WRITE
)
3183 entry
= pte_mkwrite(pte_mkdirty(entry
));
3185 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3187 if (!pte_none(*vmf
->pte
))
3190 ret
= check_stable_address_space(vma
->vm_mm
);
3194 /* Deliver the page fault to userland, check inside PT lock */
3195 if (userfaultfd_missing(vma
)) {
3196 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3197 mem_cgroup_cancel_charge(page
, memcg
, false);
3199 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3202 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3203 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3204 mem_cgroup_commit_charge(page
, memcg
, false, false);
3205 lru_cache_add_active_or_unevictable(page
, vma
);
3207 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3209 /* No need to invalidate - it was non-present before */
3210 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3212 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3215 mem_cgroup_cancel_charge(page
, memcg
, false);
3221 return VM_FAULT_OOM
;
3225 * The mmap_sem must have been held on entry, and may have been
3226 * released depending on flags and vma->vm_ops->fault() return value.
3227 * See filemap_fault() and __lock_page_retry().
3229 static vm_fault_t
__do_fault(struct vm_fault
*vmf
)
3231 struct vm_area_struct
*vma
= vmf
->vma
;
3235 * Preallocate pte before we take page_lock because this might lead to
3236 * deadlocks for memcg reclaim which waits for pages under writeback:
3238 * SetPageWriteback(A)
3244 * wait_on_page_writeback(A)
3245 * SetPageWriteback(B)
3247 * # flush A, B to clear the writeback
3249 if (pmd_none(*vmf
->pmd
) && !vmf
->prealloc_pte
) {
3250 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
);
3251 if (!vmf
->prealloc_pte
)
3252 return VM_FAULT_OOM
;
3253 smp_wmb(); /* See comment in __pte_alloc() */
3256 ret
= vma
->vm_ops
->fault(vmf
);
3257 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
|
3258 VM_FAULT_DONE_COW
)))
3261 if (unlikely(PageHWPoison(vmf
->page
))) {
3262 struct page
*page
= vmf
->page
;
3263 vm_fault_t poisonret
= VM_FAULT_HWPOISON
;
3264 if (ret
& VM_FAULT_LOCKED
) {
3265 if (page_mapped(page
))
3266 unmap_mapping_pages(page_mapping(page
),
3267 page
->index
, 1, false);
3268 /* Retry if a clean page was removed from the cache. */
3269 if (invalidate_inode_page(page
))
3270 poisonret
= VM_FAULT_NOPAGE
;
3278 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3279 lock_page(vmf
->page
);
3281 VM_BUG_ON_PAGE(!PageLocked(vmf
->page
), vmf
->page
);
3287 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3288 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3289 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3290 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3292 static int pmd_devmap_trans_unstable(pmd_t
*pmd
)
3294 return pmd_devmap(*pmd
) || pmd_trans_unstable(pmd
);
3297 static vm_fault_t
pte_alloc_one_map(struct vm_fault
*vmf
)
3299 struct vm_area_struct
*vma
= vmf
->vma
;
3301 if (!pmd_none(*vmf
->pmd
))
3303 if (vmf
->prealloc_pte
) {
3304 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3305 if (unlikely(!pmd_none(*vmf
->pmd
))) {
3306 spin_unlock(vmf
->ptl
);
3310 mm_inc_nr_ptes(vma
->vm_mm
);
3311 pmd_populate(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3312 spin_unlock(vmf
->ptl
);
3313 vmf
->prealloc_pte
= NULL
;
3314 } else if (unlikely(pte_alloc(vma
->vm_mm
, vmf
->pmd
))) {
3315 return VM_FAULT_OOM
;
3319 * If a huge pmd materialized under us just retry later. Use
3320 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3321 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3322 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3323 * running immediately after a huge pmd fault in a different thread of
3324 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3325 * All we have to ensure is that it is a regular pmd that we can walk
3326 * with pte_offset_map() and we can do that through an atomic read in
3327 * C, which is what pmd_trans_unstable() provides.
3329 if (pmd_devmap_trans_unstable(vmf
->pmd
))
3330 return VM_FAULT_NOPAGE
;
3333 * At this point we know that our vmf->pmd points to a page of ptes
3334 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3335 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3336 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3337 * be valid and we will re-check to make sure the vmf->pte isn't
3338 * pte_none() under vmf->ptl protection when we return to
3341 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3346 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3347 static void deposit_prealloc_pte(struct vm_fault
*vmf
)
3349 struct vm_area_struct
*vma
= vmf
->vma
;
3351 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3353 * We are going to consume the prealloc table,
3354 * count that as nr_ptes.
3356 mm_inc_nr_ptes(vma
->vm_mm
);
3357 vmf
->prealloc_pte
= NULL
;
3360 static vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3362 struct vm_area_struct
*vma
= vmf
->vma
;
3363 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3364 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
3369 if (!transhuge_vma_suitable(vma
, haddr
))
3370 return VM_FAULT_FALLBACK
;
3372 ret
= VM_FAULT_FALLBACK
;
3373 page
= compound_head(page
);
3376 * Archs like ppc64 need additonal space to store information
3377 * related to pte entry. Use the preallocated table for that.
3379 if (arch_needs_pgtable_deposit() && !vmf
->prealloc_pte
) {
3380 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
3381 if (!vmf
->prealloc_pte
)
3382 return VM_FAULT_OOM
;
3383 smp_wmb(); /* See comment in __pte_alloc() */
3386 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3387 if (unlikely(!pmd_none(*vmf
->pmd
)))
3390 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
3391 flush_icache_page(vma
, page
+ i
);
3393 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
3395 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
3397 add_mm_counter(vma
->vm_mm
, mm_counter_file(page
), HPAGE_PMD_NR
);
3398 page_add_file_rmap(page
, true);
3400 * deposit and withdraw with pmd lock held
3402 if (arch_needs_pgtable_deposit())
3403 deposit_prealloc_pte(vmf
);
3405 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
3407 update_mmu_cache_pmd(vma
, haddr
, vmf
->pmd
);
3409 /* fault is handled */
3411 count_vm_event(THP_FILE_MAPPED
);
3413 spin_unlock(vmf
->ptl
);
3417 static vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3425 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3426 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3428 * @vmf: fault environment
3429 * @memcg: memcg to charge page (only for private mappings)
3430 * @page: page to map
3432 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3435 * Target users are page handler itself and implementations of
3436 * vm_ops->map_pages.
3438 * Return: %0 on success, %VM_FAULT_ code in case of error.
3440 vm_fault_t
alloc_set_pte(struct vm_fault
*vmf
, struct mem_cgroup
*memcg
,
3443 struct vm_area_struct
*vma
= vmf
->vma
;
3444 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3448 if (pmd_none(*vmf
->pmd
) && PageTransCompound(page
) &&
3449 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE
)) {
3451 VM_BUG_ON_PAGE(memcg
, page
);
3453 ret
= do_set_pmd(vmf
, page
);
3454 if (ret
!= VM_FAULT_FALLBACK
)
3459 ret
= pte_alloc_one_map(vmf
);
3464 /* Re-check under ptl */
3465 if (unlikely(!pte_none(*vmf
->pte
)))
3466 return VM_FAULT_NOPAGE
;
3468 flush_icache_page(vma
, page
);
3469 entry
= mk_pte(page
, vma
->vm_page_prot
);
3471 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3472 /* copy-on-write page */
3473 if (write
&& !(vma
->vm_flags
& VM_SHARED
)) {
3474 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3475 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3476 mem_cgroup_commit_charge(page
, memcg
, false, false);
3477 lru_cache_add_active_or_unevictable(page
, vma
);
3479 inc_mm_counter_fast(vma
->vm_mm
, mm_counter_file(page
));
3480 page_add_file_rmap(page
, false);
3482 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3484 /* no need to invalidate: a not-present page won't be cached */
3485 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3492 * finish_fault - finish page fault once we have prepared the page to fault
3494 * @vmf: structure describing the fault
3496 * This function handles all that is needed to finish a page fault once the
3497 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3498 * given page, adds reverse page mapping, handles memcg charges and LRU
3501 * The function expects the page to be locked and on success it consumes a
3502 * reference of a page being mapped (for the PTE which maps it).
3504 * Return: %0 on success, %VM_FAULT_ code in case of error.
3506 vm_fault_t
finish_fault(struct vm_fault
*vmf
)
3511 /* Did we COW the page? */
3512 if ((vmf
->flags
& FAULT_FLAG_WRITE
) &&
3513 !(vmf
->vma
->vm_flags
& VM_SHARED
))
3514 page
= vmf
->cow_page
;
3519 * check even for read faults because we might have lost our CoWed
3522 if (!(vmf
->vma
->vm_flags
& VM_SHARED
))
3523 ret
= check_stable_address_space(vmf
->vma
->vm_mm
);
3525 ret
= alloc_set_pte(vmf
, vmf
->memcg
, page
);
3527 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3531 static unsigned long fault_around_bytes __read_mostly
=
3532 rounddown_pow_of_two(65536);
3534 #ifdef CONFIG_DEBUG_FS
3535 static int fault_around_bytes_get(void *data
, u64
*val
)
3537 *val
= fault_around_bytes
;
3542 * fault_around_bytes must be rounded down to the nearest page order as it's
3543 * what do_fault_around() expects to see.
3545 static int fault_around_bytes_set(void *data
, u64 val
)
3547 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
3549 if (val
> PAGE_SIZE
)
3550 fault_around_bytes
= rounddown_pow_of_two(val
);
3552 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
3555 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops
,
3556 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
3558 static int __init
fault_around_debugfs(void)
3560 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL
, NULL
,
3561 &fault_around_bytes_fops
);
3564 late_initcall(fault_around_debugfs
);
3568 * do_fault_around() tries to map few pages around the fault address. The hope
3569 * is that the pages will be needed soon and this will lower the number of
3572 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3573 * not ready to be mapped: not up-to-date, locked, etc.
3575 * This function is called with the page table lock taken. In the split ptlock
3576 * case the page table lock only protects only those entries which belong to
3577 * the page table corresponding to the fault address.
3579 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3582 * fault_around_bytes defines how many bytes we'll try to map.
3583 * do_fault_around() expects it to be set to a power of two less than or equal
3586 * The virtual address of the area that we map is naturally aligned to
3587 * fault_around_bytes rounded down to the machine page size
3588 * (and therefore to page order). This way it's easier to guarantee
3589 * that we don't cross page table boundaries.
3591 static vm_fault_t
do_fault_around(struct vm_fault
*vmf
)
3593 unsigned long address
= vmf
->address
, nr_pages
, mask
;
3594 pgoff_t start_pgoff
= vmf
->pgoff
;
3599 nr_pages
= READ_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
3600 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
3602 vmf
->address
= max(address
& mask
, vmf
->vma
->vm_start
);
3603 off
= ((address
- vmf
->address
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
3607 * end_pgoff is either the end of the page table, the end of
3608 * the vma or nr_pages from start_pgoff, depending what is nearest.
3610 end_pgoff
= start_pgoff
-
3611 ((vmf
->address
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
3613 end_pgoff
= min3(end_pgoff
, vma_pages(vmf
->vma
) + vmf
->vma
->vm_pgoff
- 1,
3614 start_pgoff
+ nr_pages
- 1);
3616 if (pmd_none(*vmf
->pmd
)) {
3617 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
);
3618 if (!vmf
->prealloc_pte
)
3620 smp_wmb(); /* See comment in __pte_alloc() */
3623 vmf
->vma
->vm_ops
->map_pages(vmf
, start_pgoff
, end_pgoff
);
3625 /* Huge page is mapped? Page fault is solved */
3626 if (pmd_trans_huge(*vmf
->pmd
)) {
3627 ret
= VM_FAULT_NOPAGE
;
3631 /* ->map_pages() haven't done anything useful. Cold page cache? */
3635 /* check if the page fault is solved */
3636 vmf
->pte
-= (vmf
->address
>> PAGE_SHIFT
) - (address
>> PAGE_SHIFT
);
3637 if (!pte_none(*vmf
->pte
))
3638 ret
= VM_FAULT_NOPAGE
;
3639 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3641 vmf
->address
= address
;
3646 static vm_fault_t
do_read_fault(struct vm_fault
*vmf
)
3648 struct vm_area_struct
*vma
= vmf
->vma
;
3652 * Let's call ->map_pages() first and use ->fault() as fallback
3653 * if page by the offset is not ready to be mapped (cold cache or
3656 if (vma
->vm_ops
->map_pages
&& fault_around_bytes
>> PAGE_SHIFT
> 1) {
3657 ret
= do_fault_around(vmf
);
3662 ret
= __do_fault(vmf
);
3663 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3666 ret
|= finish_fault(vmf
);
3667 unlock_page(vmf
->page
);
3668 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3669 put_page(vmf
->page
);
3673 static vm_fault_t
do_cow_fault(struct vm_fault
*vmf
)
3675 struct vm_area_struct
*vma
= vmf
->vma
;
3678 if (unlikely(anon_vma_prepare(vma
)))
3679 return VM_FAULT_OOM
;
3681 vmf
->cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, vmf
->address
);
3683 return VM_FAULT_OOM
;
3685 if (mem_cgroup_try_charge_delay(vmf
->cow_page
, vma
->vm_mm
, GFP_KERNEL
,
3686 &vmf
->memcg
, false)) {
3687 put_page(vmf
->cow_page
);
3688 return VM_FAULT_OOM
;
3691 ret
= __do_fault(vmf
);
3692 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3694 if (ret
& VM_FAULT_DONE_COW
)
3697 copy_user_highpage(vmf
->cow_page
, vmf
->page
, vmf
->address
, vma
);
3698 __SetPageUptodate(vmf
->cow_page
);
3700 ret
|= finish_fault(vmf
);
3701 unlock_page(vmf
->page
);
3702 put_page(vmf
->page
);
3703 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3707 mem_cgroup_cancel_charge(vmf
->cow_page
, vmf
->memcg
, false);
3708 put_page(vmf
->cow_page
);
3712 static vm_fault_t
do_shared_fault(struct vm_fault
*vmf
)
3714 struct vm_area_struct
*vma
= vmf
->vma
;
3715 vm_fault_t ret
, tmp
;
3717 ret
= __do_fault(vmf
);
3718 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3722 * Check if the backing address space wants to know that the page is
3723 * about to become writable
3725 if (vma
->vm_ops
->page_mkwrite
) {
3726 unlock_page(vmf
->page
);
3727 tmp
= do_page_mkwrite(vmf
);
3728 if (unlikely(!tmp
||
3729 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3730 put_page(vmf
->page
);
3735 ret
|= finish_fault(vmf
);
3736 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
3738 unlock_page(vmf
->page
);
3739 put_page(vmf
->page
);
3743 ret
|= fault_dirty_shared_page(vmf
);
3748 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3749 * but allow concurrent faults).
3750 * The mmap_sem may have been released depending on flags and our
3751 * return value. See filemap_fault() and __lock_page_or_retry().
3752 * If mmap_sem is released, vma may become invalid (for example
3753 * by other thread calling munmap()).
3755 static vm_fault_t
do_fault(struct vm_fault
*vmf
)
3757 struct vm_area_struct
*vma
= vmf
->vma
;
3758 struct mm_struct
*vm_mm
= vma
->vm_mm
;
3762 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3764 if (!vma
->vm_ops
->fault
) {
3766 * If we find a migration pmd entry or a none pmd entry, which
3767 * should never happen, return SIGBUS
3769 if (unlikely(!pmd_present(*vmf
->pmd
)))
3770 ret
= VM_FAULT_SIGBUS
;
3772 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
,
3777 * Make sure this is not a temporary clearing of pte
3778 * by holding ptl and checking again. A R/M/W update
3779 * of pte involves: take ptl, clearing the pte so that
3780 * we don't have concurrent modification by hardware
3781 * followed by an update.
3783 if (unlikely(pte_none(*vmf
->pte
)))
3784 ret
= VM_FAULT_SIGBUS
;
3786 ret
= VM_FAULT_NOPAGE
;
3788 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3790 } else if (!(vmf
->flags
& FAULT_FLAG_WRITE
))
3791 ret
= do_read_fault(vmf
);
3792 else if (!(vma
->vm_flags
& VM_SHARED
))
3793 ret
= do_cow_fault(vmf
);
3795 ret
= do_shared_fault(vmf
);
3797 /* preallocated pagetable is unused: free it */
3798 if (vmf
->prealloc_pte
) {
3799 pte_free(vm_mm
, vmf
->prealloc_pte
);
3800 vmf
->prealloc_pte
= NULL
;
3805 static int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3806 unsigned long addr
, int page_nid
,
3811 count_vm_numa_event(NUMA_HINT_FAULTS
);
3812 if (page_nid
== numa_node_id()) {
3813 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3814 *flags
|= TNF_FAULT_LOCAL
;
3817 return mpol_misplaced(page
, vma
, addr
);
3820 static vm_fault_t
do_numa_page(struct vm_fault
*vmf
)
3822 struct vm_area_struct
*vma
= vmf
->vma
;
3823 struct page
*page
= NULL
;
3824 int page_nid
= NUMA_NO_NODE
;
3827 bool migrated
= false;
3829 bool was_writable
= pte_savedwrite(vmf
->orig_pte
);
3833 * The "pte" at this point cannot be used safely without
3834 * validation through pte_unmap_same(). It's of NUMA type but
3835 * the pfn may be screwed if the read is non atomic.
3837 vmf
->ptl
= pte_lockptr(vma
->vm_mm
, vmf
->pmd
);
3838 spin_lock(vmf
->ptl
);
3839 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
3840 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3845 * Make it present again, Depending on how arch implementes non
3846 * accessible ptes, some can allow access by kernel mode.
3848 old_pte
= ptep_modify_prot_start(vma
, vmf
->address
, vmf
->pte
);
3849 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
3850 pte
= pte_mkyoung(pte
);
3852 pte
= pte_mkwrite(pte
);
3853 ptep_modify_prot_commit(vma
, vmf
->address
, vmf
->pte
, old_pte
, pte
);
3854 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3856 page
= vm_normal_page(vma
, vmf
->address
, pte
);
3858 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3862 /* TODO: handle PTE-mapped THP */
3863 if (PageCompound(page
)) {
3864 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3869 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3870 * much anyway since they can be in shared cache state. This misses
3871 * the case where a mapping is writable but the process never writes
3872 * to it but pte_write gets cleared during protection updates and
3873 * pte_dirty has unpredictable behaviour between PTE scan updates,
3874 * background writeback, dirty balancing and application behaviour.
3876 if (!pte_write(pte
))
3877 flags
|= TNF_NO_GROUP
;
3880 * Flag if the page is shared between multiple address spaces. This
3881 * is later used when determining whether to group tasks together
3883 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
3884 flags
|= TNF_SHARED
;
3886 last_cpupid
= page_cpupid_last(page
);
3887 page_nid
= page_to_nid(page
);
3888 target_nid
= numa_migrate_prep(page
, vma
, vmf
->address
, page_nid
,
3890 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3891 if (target_nid
== NUMA_NO_NODE
) {
3896 /* Migrate to the requested node */
3897 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
3899 page_nid
= target_nid
;
3900 flags
|= TNF_MIGRATED
;
3902 flags
|= TNF_MIGRATE_FAIL
;
3905 if (page_nid
!= NUMA_NO_NODE
)
3906 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
3910 static inline vm_fault_t
create_huge_pmd(struct vm_fault
*vmf
)
3912 if (vma_is_anonymous(vmf
->vma
))
3913 return do_huge_pmd_anonymous_page(vmf
);
3914 if (vmf
->vma
->vm_ops
->huge_fault
)
3915 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
3916 return VM_FAULT_FALLBACK
;
3919 /* `inline' is required to avoid gcc 4.1.2 build error */
3920 static inline vm_fault_t
wp_huge_pmd(struct vm_fault
*vmf
, pmd_t orig_pmd
)
3922 if (vma_is_anonymous(vmf
->vma
))
3923 return do_huge_pmd_wp_page(vmf
, orig_pmd
);
3924 if (vmf
->vma
->vm_ops
->huge_fault
)
3925 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
3927 /* COW handled on pte level: split pmd */
3928 VM_BUG_ON_VMA(vmf
->vma
->vm_flags
& VM_SHARED
, vmf
->vma
);
3929 __split_huge_pmd(vmf
->vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
3931 return VM_FAULT_FALLBACK
;
3934 static inline bool vma_is_accessible(struct vm_area_struct
*vma
)
3936 return vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
);
3939 static vm_fault_t
create_huge_pud(struct vm_fault
*vmf
)
3941 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3942 /* No support for anonymous transparent PUD pages yet */
3943 if (vma_is_anonymous(vmf
->vma
))
3944 return VM_FAULT_FALLBACK
;
3945 if (vmf
->vma
->vm_ops
->huge_fault
)
3946 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
3947 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3948 return VM_FAULT_FALLBACK
;
3951 static vm_fault_t
wp_huge_pud(struct vm_fault
*vmf
, pud_t orig_pud
)
3953 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3954 /* No support for anonymous transparent PUD pages yet */
3955 if (vma_is_anonymous(vmf
->vma
))
3956 return VM_FAULT_FALLBACK
;
3957 if (vmf
->vma
->vm_ops
->huge_fault
)
3958 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
3959 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3960 return VM_FAULT_FALLBACK
;
3964 * These routines also need to handle stuff like marking pages dirty
3965 * and/or accessed for architectures that don't do it in hardware (most
3966 * RISC architectures). The early dirtying is also good on the i386.
3968 * There is also a hook called "update_mmu_cache()" that architectures
3969 * with external mmu caches can use to update those (ie the Sparc or
3970 * PowerPC hashed page tables that act as extended TLBs).
3972 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3973 * concurrent faults).
3975 * The mmap_sem may have been released depending on flags and our return value.
3976 * See filemap_fault() and __lock_page_or_retry().
3978 static vm_fault_t
handle_pte_fault(struct vm_fault
*vmf
)
3982 if (unlikely(pmd_none(*vmf
->pmd
))) {
3984 * Leave __pte_alloc() until later: because vm_ops->fault may
3985 * want to allocate huge page, and if we expose page table
3986 * for an instant, it will be difficult to retract from
3987 * concurrent faults and from rmap lookups.
3991 /* See comment in pte_alloc_one_map() */
3992 if (pmd_devmap_trans_unstable(vmf
->pmd
))
3995 * A regular pmd is established and it can't morph into a huge
3996 * pmd from under us anymore at this point because we hold the
3997 * mmap_sem read mode and khugepaged takes it in write mode.
3998 * So now it's safe to run pte_offset_map().
4000 vmf
->pte
= pte_offset_map(vmf
->pmd
, vmf
->address
);
4001 vmf
->orig_pte
= *vmf
->pte
;
4004 * some architectures can have larger ptes than wordsize,
4005 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4006 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4007 * accesses. The code below just needs a consistent view
4008 * for the ifs and we later double check anyway with the
4009 * ptl lock held. So here a barrier will do.
4012 if (pte_none(vmf
->orig_pte
)) {
4013 pte_unmap(vmf
->pte
);
4019 if (vma_is_anonymous(vmf
->vma
))
4020 return do_anonymous_page(vmf
);
4022 return do_fault(vmf
);
4025 if (!pte_present(vmf
->orig_pte
))
4026 return do_swap_page(vmf
);
4028 if (pte_protnone(vmf
->orig_pte
) && vma_is_accessible(vmf
->vma
))
4029 return do_numa_page(vmf
);
4031 vmf
->ptl
= pte_lockptr(vmf
->vma
->vm_mm
, vmf
->pmd
);
4032 spin_lock(vmf
->ptl
);
4033 entry
= vmf
->orig_pte
;
4034 if (unlikely(!pte_same(*vmf
->pte
, entry
)))
4036 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4037 if (!pte_write(entry
))
4038 return do_wp_page(vmf
);
4039 entry
= pte_mkdirty(entry
);
4041 entry
= pte_mkyoung(entry
);
4042 if (ptep_set_access_flags(vmf
->vma
, vmf
->address
, vmf
->pte
, entry
,
4043 vmf
->flags
& FAULT_FLAG_WRITE
)) {
4044 update_mmu_cache(vmf
->vma
, vmf
->address
, vmf
->pte
);
4047 * This is needed only for protection faults but the arch code
4048 * is not yet telling us if this is a protection fault or not.
4049 * This still avoids useless tlb flushes for .text page faults
4052 if (vmf
->flags
& FAULT_FLAG_WRITE
)
4053 flush_tlb_fix_spurious_fault(vmf
->vma
, vmf
->address
);
4056 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4061 * By the time we get here, we already hold the mm semaphore
4063 * The mmap_sem may have been released depending on flags and our
4064 * return value. See filemap_fault() and __lock_page_or_retry().
4066 static vm_fault_t
__handle_mm_fault(struct vm_area_struct
*vma
,
4067 unsigned long address
, unsigned int flags
)
4069 struct vm_fault vmf
= {
4071 .address
= address
& PAGE_MASK
,
4073 .pgoff
= linear_page_index(vma
, address
),
4074 .gfp_mask
= __get_fault_gfp_mask(vma
),
4076 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
4077 struct mm_struct
*mm
= vma
->vm_mm
;
4082 pgd
= pgd_offset(mm
, address
);
4083 p4d
= p4d_alloc(mm
, pgd
, address
);
4085 return VM_FAULT_OOM
;
4087 vmf
.pud
= pud_alloc(mm
, p4d
, address
);
4089 return VM_FAULT_OOM
;
4090 if (pud_none(*vmf
.pud
) && __transparent_hugepage_enabled(vma
)) {
4091 ret
= create_huge_pud(&vmf
);
4092 if (!(ret
& VM_FAULT_FALLBACK
))
4095 pud_t orig_pud
= *vmf
.pud
;
4098 if (pud_trans_huge(orig_pud
) || pud_devmap(orig_pud
)) {
4100 /* NUMA case for anonymous PUDs would go here */
4102 if (dirty
&& !pud_write(orig_pud
)) {
4103 ret
= wp_huge_pud(&vmf
, orig_pud
);
4104 if (!(ret
& VM_FAULT_FALLBACK
))
4107 huge_pud_set_accessed(&vmf
, orig_pud
);
4113 vmf
.pmd
= pmd_alloc(mm
, vmf
.pud
, address
);
4115 return VM_FAULT_OOM
;
4116 if (pmd_none(*vmf
.pmd
) && __transparent_hugepage_enabled(vma
)) {
4117 ret
= create_huge_pmd(&vmf
);
4118 if (!(ret
& VM_FAULT_FALLBACK
))
4121 pmd_t orig_pmd
= *vmf
.pmd
;
4124 if (unlikely(is_swap_pmd(orig_pmd
))) {
4125 VM_BUG_ON(thp_migration_supported() &&
4126 !is_pmd_migration_entry(orig_pmd
));
4127 if (is_pmd_migration_entry(orig_pmd
))
4128 pmd_migration_entry_wait(mm
, vmf
.pmd
);
4131 if (pmd_trans_huge(orig_pmd
) || pmd_devmap(orig_pmd
)) {
4132 if (pmd_protnone(orig_pmd
) && vma_is_accessible(vma
))
4133 return do_huge_pmd_numa_page(&vmf
, orig_pmd
);
4135 if (dirty
&& !pmd_write(orig_pmd
)) {
4136 ret
= wp_huge_pmd(&vmf
, orig_pmd
);
4137 if (!(ret
& VM_FAULT_FALLBACK
))
4140 huge_pmd_set_accessed(&vmf
, orig_pmd
);
4146 return handle_pte_fault(&vmf
);
4150 * By the time we get here, we already hold the mm semaphore
4152 * The mmap_sem may have been released depending on flags and our
4153 * return value. See filemap_fault() and __lock_page_or_retry().
4155 vm_fault_t
handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
4160 __set_current_state(TASK_RUNNING
);
4162 count_vm_event(PGFAULT
);
4163 count_memcg_event_mm(vma
->vm_mm
, PGFAULT
);
4165 /* do counter updates before entering really critical section. */
4166 check_sync_rss_stat(current
);
4168 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
4169 flags
& FAULT_FLAG_INSTRUCTION
,
4170 flags
& FAULT_FLAG_REMOTE
))
4171 return VM_FAULT_SIGSEGV
;
4174 * Enable the memcg OOM handling for faults triggered in user
4175 * space. Kernel faults are handled more gracefully.
4177 if (flags
& FAULT_FLAG_USER
)
4178 mem_cgroup_enter_user_fault();
4180 if (unlikely(is_vm_hugetlb_page(vma
)))
4181 ret
= hugetlb_fault(vma
->vm_mm
, vma
, address
, flags
);
4183 ret
= __handle_mm_fault(vma
, address
, flags
);
4185 if (flags
& FAULT_FLAG_USER
) {
4186 mem_cgroup_exit_user_fault();
4188 * The task may have entered a memcg OOM situation but
4189 * if the allocation error was handled gracefully (no
4190 * VM_FAULT_OOM), there is no need to kill anything.
4191 * Just clean up the OOM state peacefully.
4193 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
4194 mem_cgroup_oom_synchronize(false);
4199 EXPORT_SYMBOL_GPL(handle_mm_fault
);
4201 #ifndef __PAGETABLE_P4D_FOLDED
4203 * Allocate p4d page table.
4204 * We've already handled the fast-path in-line.
4206 int __p4d_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
4208 p4d_t
*new = p4d_alloc_one(mm
, address
);
4212 smp_wmb(); /* See comment in __pte_alloc */
4214 spin_lock(&mm
->page_table_lock
);
4215 if (pgd_present(*pgd
)) /* Another has populated it */
4218 pgd_populate(mm
, pgd
, new);
4219 spin_unlock(&mm
->page_table_lock
);
4222 #endif /* __PAGETABLE_P4D_FOLDED */
4224 #ifndef __PAGETABLE_PUD_FOLDED
4226 * Allocate page upper directory.
4227 * We've already handled the fast-path in-line.
4229 int __pud_alloc(struct mm_struct
*mm
, p4d_t
*p4d
, unsigned long address
)
4231 pud_t
*new = pud_alloc_one(mm
, address
);
4235 smp_wmb(); /* See comment in __pte_alloc */
4237 spin_lock(&mm
->page_table_lock
);
4238 #ifndef __ARCH_HAS_5LEVEL_HACK
4239 if (!p4d_present(*p4d
)) {
4241 p4d_populate(mm
, p4d
, new);
4242 } else /* Another has populated it */
4245 if (!pgd_present(*p4d
)) {
4247 pgd_populate(mm
, p4d
, new);
4248 } else /* Another has populated it */
4250 #endif /* __ARCH_HAS_5LEVEL_HACK */
4251 spin_unlock(&mm
->page_table_lock
);
4254 #endif /* __PAGETABLE_PUD_FOLDED */
4256 #ifndef __PAGETABLE_PMD_FOLDED
4258 * Allocate page middle directory.
4259 * We've already handled the fast-path in-line.
4261 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
4264 pmd_t
*new = pmd_alloc_one(mm
, address
);
4268 smp_wmb(); /* See comment in __pte_alloc */
4270 ptl
= pud_lock(mm
, pud
);
4271 #ifndef __ARCH_HAS_4LEVEL_HACK
4272 if (!pud_present(*pud
)) {
4274 pud_populate(mm
, pud
, new);
4275 } else /* Another has populated it */
4278 if (!pgd_present(*pud
)) {
4280 pgd_populate(mm
, pud
, new);
4281 } else /* Another has populated it */
4283 #endif /* __ARCH_HAS_4LEVEL_HACK */
4287 #endif /* __PAGETABLE_PMD_FOLDED */
4289 int follow_invalidate_pte(struct mm_struct
*mm
, unsigned long address
,
4290 struct mmu_notifier_range
*range
, pte_t
**ptepp
,
4291 pmd_t
**pmdpp
, spinlock_t
**ptlp
)
4299 pgd
= pgd_offset(mm
, address
);
4300 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
4303 p4d
= p4d_offset(pgd
, address
);
4304 if (p4d_none(*p4d
) || unlikely(p4d_bad(*p4d
)))
4307 pud
= pud_offset(p4d
, address
);
4308 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
4311 pmd
= pmd_offset(pud
, address
);
4312 VM_BUG_ON(pmd_trans_huge(*pmd
));
4314 if (pmd_huge(*pmd
)) {
4319 mmu_notifier_range_init(range
, MMU_NOTIFY_CLEAR
, 0,
4320 NULL
, mm
, address
& PMD_MASK
,
4321 (address
& PMD_MASK
) + PMD_SIZE
);
4322 mmu_notifier_invalidate_range_start(range
);
4324 *ptlp
= pmd_lock(mm
, pmd
);
4325 if (pmd_huge(*pmd
)) {
4331 mmu_notifier_invalidate_range_end(range
);
4334 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
4338 mmu_notifier_range_init(range
, MMU_NOTIFY_CLEAR
, 0, NULL
, mm
,
4339 address
& PAGE_MASK
,
4340 (address
& PAGE_MASK
) + PAGE_SIZE
);
4341 mmu_notifier_invalidate_range_start(range
);
4343 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
4344 if (!pte_present(*ptep
))
4349 pte_unmap_unlock(ptep
, *ptlp
);
4351 mmu_notifier_invalidate_range_end(range
);
4357 * follow_pte - look up PTE at a user virtual address
4358 * @mm: the mm_struct of the target address space
4359 * @address: user virtual address
4360 * @ptepp: location to store found PTE
4361 * @ptlp: location to store the lock for the PTE
4363 * On a successful return, the pointer to the PTE is stored in @ptepp;
4364 * the corresponding lock is taken and its location is stored in @ptlp.
4365 * The contents of the PTE are only stable until @ptlp is released;
4366 * any further use, if any, must be protected against invalidation
4367 * with MMU notifiers.
4369 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
4370 * should be taken for read.
4372 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
4373 * it is not a good general-purpose API.
4375 * Return: zero on success, -ve otherwise.
4377 int follow_pte(struct mm_struct
*mm
, unsigned long address
,
4378 pte_t
**ptepp
, spinlock_t
**ptlp
)
4380 return follow_invalidate_pte(mm
, address
, NULL
, ptepp
, NULL
, ptlp
);
4382 EXPORT_SYMBOL_GPL(follow_pte
);
4385 * follow_pfn - look up PFN at a user virtual address
4386 * @vma: memory mapping
4387 * @address: user virtual address
4388 * @pfn: location to store found PFN
4390 * Only IO mappings and raw PFN mappings are allowed.
4392 * This function does not allow the caller to read the permissions
4393 * of the PTE. Do not use it.
4395 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4397 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
4404 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4407 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
4410 *pfn
= pte_pfn(*ptep
);
4411 pte_unmap_unlock(ptep
, ptl
);
4414 EXPORT_SYMBOL(follow_pfn
);
4416 #ifdef CONFIG_HAVE_IOREMAP_PROT
4417 int follow_phys(struct vm_area_struct
*vma
,
4418 unsigned long address
, unsigned int flags
,
4419 unsigned long *prot
, resource_size_t
*phys
)
4425 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4428 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
4432 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
4435 *prot
= pgprot_val(pte_pgprot(pte
));
4436 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
4440 pte_unmap_unlock(ptep
, ptl
);
4445 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
4446 void *buf
, int len
, int write
)
4448 resource_size_t phys_addr
;
4449 unsigned long prot
= 0;
4450 void __iomem
*maddr
;
4451 int offset
= addr
& (PAGE_SIZE
-1);
4453 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
4456 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
4461 memcpy_toio(maddr
+ offset
, buf
, len
);
4463 memcpy_fromio(buf
, maddr
+ offset
, len
);
4468 EXPORT_SYMBOL_GPL(generic_access_phys
);
4472 * Access another process' address space as given in mm. If non-NULL, use the
4473 * given task for page fault accounting.
4475 int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
4476 unsigned long addr
, void *buf
, int len
, unsigned int gup_flags
)
4478 struct vm_area_struct
*vma
;
4479 void *old_buf
= buf
;
4480 int write
= gup_flags
& FOLL_WRITE
;
4482 if (down_read_killable(&mm
->mmap_sem
))
4485 /* ignore errors, just check how much was successfully transferred */
4487 int bytes
, ret
, offset
;
4489 struct page
*page
= NULL
;
4491 ret
= get_user_pages_remote(tsk
, mm
, addr
, 1,
4492 gup_flags
, &page
, &vma
, NULL
);
4494 #ifndef CONFIG_HAVE_IOREMAP_PROT
4498 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4499 * we can access using slightly different code.
4501 vma
= find_vma(mm
, addr
);
4502 if (!vma
|| vma
->vm_start
> addr
)
4504 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
4505 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
4513 offset
= addr
& (PAGE_SIZE
-1);
4514 if (bytes
> PAGE_SIZE
-offset
)
4515 bytes
= PAGE_SIZE
-offset
;
4519 copy_to_user_page(vma
, page
, addr
,
4520 maddr
+ offset
, buf
, bytes
);
4521 set_page_dirty_lock(page
);
4523 copy_from_user_page(vma
, page
, addr
,
4524 buf
, maddr
+ offset
, bytes
);
4533 up_read(&mm
->mmap_sem
);
4535 return buf
- old_buf
;
4539 * access_remote_vm - access another process' address space
4540 * @mm: the mm_struct of the target address space
4541 * @addr: start address to access
4542 * @buf: source or destination buffer
4543 * @len: number of bytes to transfer
4544 * @gup_flags: flags modifying lookup behaviour
4546 * The caller must hold a reference on @mm.
4548 * Return: number of bytes copied from source to destination.
4550 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
4551 void *buf
, int len
, unsigned int gup_flags
)
4553 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, gup_flags
);
4557 * Access another process' address space.
4558 * Source/target buffer must be kernel space,
4559 * Do not walk the page table directly, use get_user_pages
4561 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
4562 void *buf
, int len
, unsigned int gup_flags
)
4564 struct mm_struct
*mm
;
4567 mm
= get_task_mm(tsk
);
4571 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, gup_flags
);
4577 EXPORT_SYMBOL_GPL(access_process_vm
);
4580 * Print the name of a VMA.
4582 void print_vma_addr(char *prefix
, unsigned long ip
)
4584 struct mm_struct
*mm
= current
->mm
;
4585 struct vm_area_struct
*vma
;
4588 * we might be running from an atomic context so we cannot sleep
4590 if (!down_read_trylock(&mm
->mmap_sem
))
4593 vma
= find_vma(mm
, ip
);
4594 if (vma
&& vma
->vm_file
) {
4595 struct file
*f
= vma
->vm_file
;
4596 char *buf
= (char *)__get_free_page(GFP_NOWAIT
);
4600 p
= file_path(f
, buf
, PAGE_SIZE
);
4603 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
4605 vma
->vm_end
- vma
->vm_start
);
4606 free_page((unsigned long)buf
);
4609 up_read(&mm
->mmap_sem
);
4612 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4613 void __might_fault(const char *file
, int line
)
4616 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4617 * holding the mmap_sem, this is safe because kernel memory doesn't
4618 * get paged out, therefore we'll never actually fault, and the
4619 * below annotations will generate false positives.
4621 if (uaccess_kernel())
4623 if (pagefault_disabled())
4625 __might_sleep(file
, line
, 0);
4626 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4628 might_lock_read(¤t
->mm
->mmap_sem
);
4631 EXPORT_SYMBOL(__might_fault
);
4634 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4636 * Process all subpages of the specified huge page with the specified
4637 * operation. The target subpage will be processed last to keep its
4640 static inline void process_huge_page(
4641 unsigned long addr_hint
, unsigned int pages_per_huge_page
,
4642 void (*process_subpage
)(unsigned long addr
, int idx
, void *arg
),
4646 unsigned long addr
= addr_hint
&
4647 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
4649 /* Process target subpage last to keep its cache lines hot */
4651 n
= (addr_hint
- addr
) / PAGE_SIZE
;
4652 if (2 * n
<= pages_per_huge_page
) {
4653 /* If target subpage in first half of huge page */
4656 /* Process subpages at the end of huge page */
4657 for (i
= pages_per_huge_page
- 1; i
>= 2 * n
; i
--) {
4659 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
4662 /* If target subpage in second half of huge page */
4663 base
= pages_per_huge_page
- 2 * (pages_per_huge_page
- n
);
4664 l
= pages_per_huge_page
- n
;
4665 /* Process subpages at the begin of huge page */
4666 for (i
= 0; i
< base
; i
++) {
4668 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
4672 * Process remaining subpages in left-right-left-right pattern
4673 * towards the target subpage
4675 for (i
= 0; i
< l
; i
++) {
4676 int left_idx
= base
+ i
;
4677 int right_idx
= base
+ 2 * l
- 1 - i
;
4680 process_subpage(addr
+ left_idx
* PAGE_SIZE
, left_idx
, arg
);
4682 process_subpage(addr
+ right_idx
* PAGE_SIZE
, right_idx
, arg
);
4686 static void clear_gigantic_page(struct page
*page
,
4688 unsigned int pages_per_huge_page
)
4691 struct page
*p
= page
;
4694 for (i
= 0; i
< pages_per_huge_page
;
4695 i
++, p
= mem_map_next(p
, page
, i
)) {
4697 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
4701 static void clear_subpage(unsigned long addr
, int idx
, void *arg
)
4703 struct page
*page
= arg
;
4705 clear_user_highpage(page
+ idx
, addr
);
4708 void clear_huge_page(struct page
*page
,
4709 unsigned long addr_hint
, unsigned int pages_per_huge_page
)
4711 unsigned long addr
= addr_hint
&
4712 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
4714 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4715 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
4719 process_huge_page(addr_hint
, pages_per_huge_page
, clear_subpage
, page
);
4722 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
4724 struct vm_area_struct
*vma
,
4725 unsigned int pages_per_huge_page
)
4728 struct page
*dst_base
= dst
;
4729 struct page
*src_base
= src
;
4731 for (i
= 0; i
< pages_per_huge_page
; ) {
4733 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
4736 dst
= mem_map_next(dst
, dst_base
, i
);
4737 src
= mem_map_next(src
, src_base
, i
);
4741 struct copy_subpage_arg
{
4744 struct vm_area_struct
*vma
;
4747 static void copy_subpage(unsigned long addr
, int idx
, void *arg
)
4749 struct copy_subpage_arg
*copy_arg
= arg
;
4751 copy_user_highpage(copy_arg
->dst
+ idx
, copy_arg
->src
+ idx
,
4752 addr
, copy_arg
->vma
);
4755 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
4756 unsigned long addr_hint
, struct vm_area_struct
*vma
,
4757 unsigned int pages_per_huge_page
)
4759 unsigned long addr
= addr_hint
&
4760 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
4761 struct copy_subpage_arg arg
= {
4767 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4768 copy_user_gigantic_page(dst
, src
, addr
, vma
,
4769 pages_per_huge_page
);
4773 process_huge_page(addr_hint
, pages_per_huge_page
, copy_subpage
, &arg
);
4776 long copy_huge_page_from_user(struct page
*dst_page
,
4777 const void __user
*usr_src
,
4778 unsigned int pages_per_huge_page
,
4779 bool allow_pagefault
)
4781 void *src
= (void *)usr_src
;
4783 unsigned long i
, rc
= 0;
4784 unsigned long ret_val
= pages_per_huge_page
* PAGE_SIZE
;
4785 struct page
*subpage
= dst_page
;
4787 for (i
= 0; i
< pages_per_huge_page
;
4788 i
++, subpage
= mem_map_next(subpage
, dst_page
, i
)) {
4789 if (allow_pagefault
)
4790 page_kaddr
= kmap(subpage
);
4792 page_kaddr
= kmap_atomic(subpage
);
4793 rc
= copy_from_user(page_kaddr
,
4794 (const void __user
*)(src
+ i
* PAGE_SIZE
),
4796 if (allow_pagefault
)
4799 kunmap_atomic(page_kaddr
);
4801 ret_val
-= (PAGE_SIZE
- rc
);
4809 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4811 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4813 static struct kmem_cache
*page_ptl_cachep
;
4815 void __init
ptlock_cache_init(void)
4817 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
4821 bool ptlock_alloc(struct page
*page
)
4825 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
4832 void ptlock_free(struct page
*page
)
4834 kmem_cache_free(page_ptl_cachep
, page
->ptl
);