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 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1017 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1018 unsigned long addr
, unsigned long end
,
1019 struct zap_details
*details
)
1021 struct mm_struct
*mm
= tlb
->mm
;
1022 int force_flush
= 0;
1023 int rss
[NR_MM_COUNTERS
];
1029 tlb_change_page_size(tlb
, PAGE_SIZE
);
1032 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1034 flush_tlb_batched_pending(mm
);
1035 arch_enter_lazy_mmu_mode();
1038 if (pte_none(ptent
))
1044 if (pte_present(ptent
)) {
1047 page
= vm_normal_page(vma
, addr
, ptent
);
1048 if (unlikely(details
) && page
) {
1050 * unmap_shared_mapping_pages() wants to
1051 * invalidate cache without truncating:
1052 * unmap shared but keep private pages.
1054 if (details
->check_mapping
&&
1055 details
->check_mapping
!= page_rmapping(page
))
1058 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1060 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1061 if (unlikely(!page
))
1064 if (!PageAnon(page
)) {
1065 if (pte_dirty(ptent
)) {
1067 set_page_dirty(page
);
1069 if (pte_young(ptent
) &&
1070 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1071 mark_page_accessed(page
);
1073 rss
[mm_counter(page
)]--;
1074 page_remove_rmap(page
, false);
1075 if (unlikely(page_mapcount(page
) < 0))
1076 print_bad_pte(vma
, addr
, ptent
, page
);
1077 if (unlikely(__tlb_remove_page(tlb
, page
))) {
1085 entry
= pte_to_swp_entry(ptent
);
1086 if (non_swap_entry(entry
) && is_device_private_entry(entry
)) {
1087 struct page
*page
= device_private_entry_to_page(entry
);
1089 if (unlikely(details
&& details
->check_mapping
)) {
1091 * unmap_shared_mapping_pages() wants to
1092 * invalidate cache without truncating:
1093 * unmap shared but keep private pages.
1095 if (details
->check_mapping
!=
1096 page_rmapping(page
))
1100 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1101 rss
[mm_counter(page
)]--;
1102 page_remove_rmap(page
, false);
1107 /* If details->check_mapping, we leave swap entries. */
1108 if (unlikely(details
))
1111 if (!non_swap_entry(entry
))
1113 else if (is_migration_entry(entry
)) {
1116 page
= migration_entry_to_page(entry
);
1117 rss
[mm_counter(page
)]--;
1119 if (unlikely(!free_swap_and_cache(entry
)))
1120 print_bad_pte(vma
, addr
, ptent
, NULL
);
1121 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1122 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1124 add_mm_rss_vec(mm
, rss
);
1125 arch_leave_lazy_mmu_mode();
1127 /* Do the actual TLB flush before dropping ptl */
1129 tlb_flush_mmu_tlbonly(tlb
);
1130 pte_unmap_unlock(start_pte
, ptl
);
1133 * If we forced a TLB flush (either due to running out of
1134 * batch buffers or because we needed to flush dirty TLB
1135 * entries before releasing the ptl), free the batched
1136 * memory too. Restart if we didn't do everything.
1151 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1152 struct vm_area_struct
*vma
, pud_t
*pud
,
1153 unsigned long addr
, unsigned long end
,
1154 struct zap_details
*details
)
1159 pmd
= pmd_offset(pud
, addr
);
1161 next
= pmd_addr_end(addr
, end
);
1162 if (is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1163 if (next
- addr
!= HPAGE_PMD_SIZE
)
1164 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
1165 else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1168 } else if (details
&& details
->single_page
&&
1169 PageTransCompound(details
->single_page
) &&
1170 next
- addr
== HPAGE_PMD_SIZE
&& pmd_none(*pmd
)) {
1171 spinlock_t
*ptl
= pmd_lock(tlb
->mm
, pmd
);
1173 * Take and drop THP pmd lock so that we cannot return
1174 * prematurely, while zap_huge_pmd() has cleared *pmd,
1175 * but not yet decremented compound_mapcount().
1181 * Here there can be other concurrent MADV_DONTNEED or
1182 * trans huge page faults running, and if the pmd is
1183 * none or trans huge it can change under us. This is
1184 * because MADV_DONTNEED holds the mmap_sem in read
1187 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1189 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1192 } while (pmd
++, addr
= next
, addr
!= end
);
1197 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1198 struct vm_area_struct
*vma
, p4d_t
*p4d
,
1199 unsigned long addr
, unsigned long end
,
1200 struct zap_details
*details
)
1205 pud
= pud_offset(p4d
, addr
);
1207 next
= pud_addr_end(addr
, end
);
1208 if (pud_trans_huge(*pud
) || pud_devmap(*pud
)) {
1209 if (next
- addr
!= HPAGE_PUD_SIZE
) {
1210 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb
->mm
->mmap_sem
), vma
);
1211 split_huge_pud(vma
, pud
, addr
);
1212 } else if (zap_huge_pud(tlb
, vma
, pud
, addr
))
1216 if (pud_none_or_clear_bad(pud
))
1218 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1221 } while (pud
++, addr
= next
, addr
!= end
);
1226 static inline unsigned long zap_p4d_range(struct mmu_gather
*tlb
,
1227 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1228 unsigned long addr
, unsigned long end
,
1229 struct zap_details
*details
)
1234 p4d
= p4d_offset(pgd
, addr
);
1236 next
= p4d_addr_end(addr
, end
);
1237 if (p4d_none_or_clear_bad(p4d
))
1239 next
= zap_pud_range(tlb
, vma
, p4d
, addr
, next
, details
);
1240 } while (p4d
++, addr
= next
, addr
!= end
);
1245 void unmap_page_range(struct mmu_gather
*tlb
,
1246 struct vm_area_struct
*vma
,
1247 unsigned long addr
, unsigned long end
,
1248 struct zap_details
*details
)
1253 BUG_ON(addr
>= end
);
1254 tlb_start_vma(tlb
, vma
);
1255 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1257 next
= pgd_addr_end(addr
, end
);
1258 if (pgd_none_or_clear_bad(pgd
))
1260 next
= zap_p4d_range(tlb
, vma
, pgd
, addr
, next
, details
);
1261 } while (pgd
++, addr
= next
, addr
!= end
);
1262 tlb_end_vma(tlb
, vma
);
1266 static void unmap_single_vma(struct mmu_gather
*tlb
,
1267 struct vm_area_struct
*vma
, unsigned long start_addr
,
1268 unsigned long end_addr
,
1269 struct zap_details
*details
)
1271 unsigned long start
= max(vma
->vm_start
, start_addr
);
1274 if (start
>= vma
->vm_end
)
1276 end
= min(vma
->vm_end
, end_addr
);
1277 if (end
<= vma
->vm_start
)
1281 uprobe_munmap(vma
, start
, end
);
1283 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1284 untrack_pfn(vma
, 0, 0);
1287 if (unlikely(is_vm_hugetlb_page(vma
))) {
1289 * It is undesirable to test vma->vm_file as it
1290 * should be non-null for valid hugetlb area.
1291 * However, vm_file will be NULL in the error
1292 * cleanup path of mmap_region. When
1293 * hugetlbfs ->mmap method fails,
1294 * mmap_region() nullifies vma->vm_file
1295 * before calling this function to clean up.
1296 * Since no pte has actually been setup, it is
1297 * safe to do nothing in this case.
1300 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
1301 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1302 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
1305 unmap_page_range(tlb
, vma
, start
, end
, details
);
1310 * unmap_vmas - unmap a range of memory covered by a list of vma's
1311 * @tlb: address of the caller's struct mmu_gather
1312 * @vma: the starting vma
1313 * @start_addr: virtual address at which to start unmapping
1314 * @end_addr: virtual address at which to end unmapping
1316 * Unmap all pages in the vma list.
1318 * Only addresses between `start' and `end' will be unmapped.
1320 * The VMA list must be sorted in ascending virtual address order.
1322 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1323 * range after unmap_vmas() returns. So the only responsibility here is to
1324 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1325 * drops the lock and schedules.
1327 void unmap_vmas(struct mmu_gather
*tlb
,
1328 struct vm_area_struct
*vma
, unsigned long start_addr
,
1329 unsigned long end_addr
)
1331 struct mmu_notifier_range range
;
1333 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
, vma
->vm_mm
,
1334 start_addr
, end_addr
);
1335 mmu_notifier_invalidate_range_start(&range
);
1336 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1337 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1338 mmu_notifier_invalidate_range_end(&range
);
1342 * zap_page_range - remove user pages in a given range
1343 * @vma: vm_area_struct holding the applicable pages
1344 * @start: starting address of pages to zap
1345 * @size: number of bytes to zap
1347 * Caller must protect the VMA list
1349 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1352 struct mmu_notifier_range range
;
1353 struct mmu_gather tlb
;
1356 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1357 start
, start
+ size
);
1358 tlb_gather_mmu(&tlb
, vma
->vm_mm
, start
, range
.end
);
1359 update_hiwater_rss(vma
->vm_mm
);
1360 mmu_notifier_invalidate_range_start(&range
);
1361 for ( ; vma
&& vma
->vm_start
< range
.end
; vma
= vma
->vm_next
)
1362 unmap_single_vma(&tlb
, vma
, start
, range
.end
, NULL
);
1363 mmu_notifier_invalidate_range_end(&range
);
1364 tlb_finish_mmu(&tlb
, start
, range
.end
);
1366 EXPORT_SYMBOL(zap_page_range
);
1369 * zap_page_range_single - remove user pages in a given range
1370 * @vma: vm_area_struct holding the applicable pages
1371 * @address: starting address of pages to zap
1372 * @size: number of bytes to zap
1373 * @details: details of shared cache invalidation
1375 * The range must fit into one VMA.
1377 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1378 unsigned long size
, struct zap_details
*details
)
1380 struct mmu_notifier_range range
;
1381 struct mmu_gather tlb
;
1384 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1385 address
, address
+ size
);
1386 tlb_gather_mmu(&tlb
, vma
->vm_mm
, address
, range
.end
);
1387 update_hiwater_rss(vma
->vm_mm
);
1388 mmu_notifier_invalidate_range_start(&range
);
1389 unmap_single_vma(&tlb
, vma
, address
, range
.end
, details
);
1390 mmu_notifier_invalidate_range_end(&range
);
1391 tlb_finish_mmu(&tlb
, address
, range
.end
);
1395 * zap_vma_ptes - remove ptes mapping the vma
1396 * @vma: vm_area_struct holding ptes to be zapped
1397 * @address: starting address of pages to zap
1398 * @size: number of bytes to zap
1400 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1402 * The entire address range must be fully contained within the vma.
1405 void zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1408 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1409 !(vma
->vm_flags
& VM_PFNMAP
))
1412 zap_page_range_single(vma
, address
, size
, NULL
);
1414 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1416 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1424 pgd
= pgd_offset(mm
, addr
);
1425 p4d
= p4d_alloc(mm
, pgd
, addr
);
1428 pud
= pud_alloc(mm
, p4d
, addr
);
1431 pmd
= pmd_alloc(mm
, pud
, addr
);
1435 VM_BUG_ON(pmd_trans_huge(*pmd
));
1436 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1440 * This is the old fallback for page remapping.
1442 * For historical reasons, it only allows reserved pages. Only
1443 * old drivers should use this, and they needed to mark their
1444 * pages reserved for the old functions anyway.
1446 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1447 struct page
*page
, pgprot_t prot
)
1449 struct mm_struct
*mm
= vma
->vm_mm
;
1455 if (PageAnon(page
) || PageSlab(page
) || page_has_type(page
))
1458 flush_dcache_page(page
);
1459 pte
= get_locked_pte(mm
, addr
, &ptl
);
1463 if (!pte_none(*pte
))
1466 /* Ok, finally just insert the thing.. */
1468 inc_mm_counter_fast(mm
, mm_counter_file(page
));
1469 page_add_file_rmap(page
, false);
1470 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1474 pte_unmap_unlock(pte
, ptl
);
1480 * vm_insert_page - insert single page into user vma
1481 * @vma: user vma to map to
1482 * @addr: target user address of this page
1483 * @page: source kernel page
1485 * This allows drivers to insert individual pages they've allocated
1488 * The page has to be a nice clean _individual_ kernel allocation.
1489 * If you allocate a compound page, you need to have marked it as
1490 * such (__GFP_COMP), or manually just split the page up yourself
1491 * (see split_page()).
1493 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1494 * took an arbitrary page protection parameter. This doesn't allow
1495 * that. Your vma protection will have to be set up correctly, which
1496 * means that if you want a shared writable mapping, you'd better
1497 * ask for a shared writable mapping!
1499 * The page does not need to be reserved.
1501 * Usually this function is called from f_op->mmap() handler
1502 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1503 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1504 * function from other places, for example from page-fault handler.
1506 * Return: %0 on success, negative error code otherwise.
1508 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1511 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1513 if (!page_count(page
))
1515 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1516 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
1517 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1518 vma
->vm_flags
|= VM_MIXEDMAP
;
1520 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1522 EXPORT_SYMBOL(vm_insert_page
);
1525 * __vm_map_pages - maps range of kernel pages into user vma
1526 * @vma: user vma to map to
1527 * @pages: pointer to array of source kernel pages
1528 * @num: number of pages in page array
1529 * @offset: user's requested vm_pgoff
1531 * This allows drivers to map range of kernel pages into a user vma.
1533 * Return: 0 on success and error code otherwise.
1535 static int __vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
1536 unsigned long num
, unsigned long offset
)
1538 unsigned long count
= vma_pages(vma
);
1539 unsigned long uaddr
= vma
->vm_start
;
1542 /* Fail if the user requested offset is beyond the end of the object */
1546 /* Fail if the user requested size exceeds available object size */
1547 if (count
> num
- offset
)
1550 for (i
= 0; i
< count
; i
++) {
1551 ret
= vm_insert_page(vma
, uaddr
, pages
[offset
+ i
]);
1561 * vm_map_pages - maps range of kernel pages starts with non zero offset
1562 * @vma: user vma to map to
1563 * @pages: pointer to array of source kernel pages
1564 * @num: number of pages in page array
1566 * Maps an object consisting of @num pages, catering for the user's
1567 * requested vm_pgoff
1569 * If we fail to insert any page into the vma, the function will return
1570 * immediately leaving any previously inserted pages present. Callers
1571 * from the mmap handler may immediately return the error as their caller
1572 * will destroy the vma, removing any successfully inserted pages. Other
1573 * callers should make their own arrangements for calling unmap_region().
1575 * Context: Process context. Called by mmap handlers.
1576 * Return: 0 on success and error code otherwise.
1578 int vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
1581 return __vm_map_pages(vma
, pages
, num
, vma
->vm_pgoff
);
1583 EXPORT_SYMBOL(vm_map_pages
);
1586 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1587 * @vma: user vma to map to
1588 * @pages: pointer to array of source kernel pages
1589 * @num: number of pages in page array
1591 * Similar to vm_map_pages(), except that it explicitly sets the offset
1592 * to 0. This function is intended for the drivers that did not consider
1595 * Context: Process context. Called by mmap handlers.
1596 * Return: 0 on success and error code otherwise.
1598 int vm_map_pages_zero(struct vm_area_struct
*vma
, struct page
**pages
,
1601 return __vm_map_pages(vma
, pages
, num
, 0);
1603 EXPORT_SYMBOL(vm_map_pages_zero
);
1605 static vm_fault_t
insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1606 pfn_t pfn
, pgprot_t prot
, bool mkwrite
)
1608 struct mm_struct
*mm
= vma
->vm_mm
;
1612 pte
= get_locked_pte(mm
, addr
, &ptl
);
1614 return VM_FAULT_OOM
;
1615 if (!pte_none(*pte
)) {
1618 * For read faults on private mappings the PFN passed
1619 * in may not match the PFN we have mapped if the
1620 * mapped PFN is a writeable COW page. In the mkwrite
1621 * case we are creating a writable PTE for a shared
1622 * mapping and we expect the PFNs to match. If they
1623 * don't match, we are likely racing with block
1624 * allocation and mapping invalidation so just skip the
1627 if (pte_pfn(*pte
) != pfn_t_to_pfn(pfn
)) {
1628 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte
)));
1631 entry
= pte_mkyoung(*pte
);
1632 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1633 if (ptep_set_access_flags(vma
, addr
, pte
, entry
, 1))
1634 update_mmu_cache(vma
, addr
, pte
);
1639 /* Ok, finally just insert the thing.. */
1640 if (pfn_t_devmap(pfn
))
1641 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
1643 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
1646 entry
= pte_mkyoung(entry
);
1647 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1650 set_pte_at(mm
, addr
, pte
, entry
);
1651 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1654 pte_unmap_unlock(pte
, ptl
);
1655 return VM_FAULT_NOPAGE
;
1659 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1660 * @vma: user vma to map to
1661 * @addr: target user address of this page
1662 * @pfn: source kernel pfn
1663 * @pgprot: pgprot flags for the inserted page
1665 * This is exactly like vmf_insert_pfn(), except that it allows drivers to
1666 * to override pgprot on a per-page basis.
1668 * This only makes sense for IO mappings, and it makes no sense for
1669 * COW mappings. In general, using multiple vmas is preferable;
1670 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1673 * Context: Process context. May allocate using %GFP_KERNEL.
1674 * Return: vm_fault_t value.
1676 vm_fault_t
vmf_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
1677 unsigned long pfn
, pgprot_t pgprot
)
1680 * Technically, architectures with pte_special can avoid all these
1681 * restrictions (same for remap_pfn_range). However we would like
1682 * consistency in testing and feature parity among all, so we should
1683 * try to keep these invariants in place for everybody.
1685 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1686 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1687 (VM_PFNMAP
|VM_MIXEDMAP
));
1688 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1689 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1691 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1692 return VM_FAULT_SIGBUS
;
1694 if (!pfn_modify_allowed(pfn
, pgprot
))
1695 return VM_FAULT_SIGBUS
;
1697 track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
));
1699 return insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
,
1702 EXPORT_SYMBOL(vmf_insert_pfn_prot
);
1705 * vmf_insert_pfn - insert single pfn into user vma
1706 * @vma: user vma to map to
1707 * @addr: target user address of this page
1708 * @pfn: source kernel pfn
1710 * Similar to vm_insert_page, this allows drivers to insert individual pages
1711 * they've allocated into a user vma. Same comments apply.
1713 * This function should only be called from a vm_ops->fault handler, and
1714 * in that case the handler should return the result of this function.
1716 * vma cannot be a COW mapping.
1718 * As this is called only for pages that do not currently exist, we
1719 * do not need to flush old virtual caches or the TLB.
1721 * Context: Process context. May allocate using %GFP_KERNEL.
1722 * Return: vm_fault_t value.
1724 vm_fault_t
vmf_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1727 return vmf_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
1729 EXPORT_SYMBOL(vmf_insert_pfn
);
1731 static bool vm_mixed_ok(struct vm_area_struct
*vma
, pfn_t pfn
)
1733 /* these checks mirror the abort conditions in vm_normal_page */
1734 if (vma
->vm_flags
& VM_MIXEDMAP
)
1736 if (pfn_t_devmap(pfn
))
1738 if (pfn_t_special(pfn
))
1740 if (is_zero_pfn(pfn_t_to_pfn(pfn
)))
1745 static vm_fault_t
__vm_insert_mixed(struct vm_area_struct
*vma
,
1746 unsigned long addr
, pfn_t pfn
, bool mkwrite
)
1748 pgprot_t pgprot
= vma
->vm_page_prot
;
1751 BUG_ON(!vm_mixed_ok(vma
, pfn
));
1753 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1754 return VM_FAULT_SIGBUS
;
1756 track_pfn_insert(vma
, &pgprot
, pfn
);
1758 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn
), pgprot
))
1759 return VM_FAULT_SIGBUS
;
1762 * If we don't have pte special, then we have to use the pfn_valid()
1763 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1764 * refcount the page if pfn_valid is true (hence insert_page rather
1765 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1766 * without pte special, it would there be refcounted as a normal page.
1768 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
) &&
1769 !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
1773 * At this point we are committed to insert_page()
1774 * regardless of whether the caller specified flags that
1775 * result in pfn_t_has_page() == false.
1777 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
1778 err
= insert_page(vma
, addr
, page
, pgprot
);
1780 return insert_pfn(vma
, addr
, pfn
, pgprot
, mkwrite
);
1784 return VM_FAULT_OOM
;
1785 if (err
< 0 && err
!= -EBUSY
)
1786 return VM_FAULT_SIGBUS
;
1788 return VM_FAULT_NOPAGE
;
1791 vm_fault_t
vmf_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1794 return __vm_insert_mixed(vma
, addr
, pfn
, false);
1796 EXPORT_SYMBOL(vmf_insert_mixed
);
1799 * If the insertion of PTE failed because someone else already added a
1800 * different entry in the mean time, we treat that as success as we assume
1801 * the same entry was actually inserted.
1803 vm_fault_t
vmf_insert_mixed_mkwrite(struct vm_area_struct
*vma
,
1804 unsigned long addr
, pfn_t pfn
)
1806 return __vm_insert_mixed(vma
, addr
, pfn
, true);
1808 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite
);
1811 * maps a range of physical memory into the requested pages. the old
1812 * mappings are removed. any references to nonexistent pages results
1813 * in null mappings (currently treated as "copy-on-access")
1815 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1816 unsigned long addr
, unsigned long end
,
1817 unsigned long pfn
, pgprot_t prot
)
1819 pte_t
*pte
, *mapped_pte
;
1823 mapped_pte
= pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1826 arch_enter_lazy_mmu_mode();
1828 BUG_ON(!pte_none(*pte
));
1829 if (!pfn_modify_allowed(pfn
, prot
)) {
1833 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
1835 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1836 arch_leave_lazy_mmu_mode();
1837 pte_unmap_unlock(mapped_pte
, ptl
);
1841 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1842 unsigned long addr
, unsigned long end
,
1843 unsigned long pfn
, pgprot_t prot
)
1849 pfn
-= addr
>> PAGE_SHIFT
;
1850 pmd
= pmd_alloc(mm
, pud
, addr
);
1853 VM_BUG_ON(pmd_trans_huge(*pmd
));
1855 next
= pmd_addr_end(addr
, end
);
1856 err
= remap_pte_range(mm
, pmd
, addr
, next
,
1857 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1860 } while (pmd
++, addr
= next
, addr
!= end
);
1864 static inline int remap_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
1865 unsigned long addr
, unsigned long end
,
1866 unsigned long pfn
, pgprot_t prot
)
1872 pfn
-= addr
>> PAGE_SHIFT
;
1873 pud
= pud_alloc(mm
, p4d
, addr
);
1877 next
= pud_addr_end(addr
, end
);
1878 err
= remap_pmd_range(mm
, pud
, addr
, next
,
1879 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1882 } while (pud
++, addr
= next
, addr
!= end
);
1886 static inline int remap_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1887 unsigned long addr
, unsigned long end
,
1888 unsigned long pfn
, pgprot_t prot
)
1894 pfn
-= addr
>> PAGE_SHIFT
;
1895 p4d
= p4d_alloc(mm
, pgd
, addr
);
1899 next
= p4d_addr_end(addr
, end
);
1900 err
= remap_pud_range(mm
, p4d
, addr
, next
,
1901 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1904 } while (p4d
++, addr
= next
, addr
!= end
);
1909 * remap_pfn_range - remap kernel memory to userspace
1910 * @vma: user vma to map to
1911 * @addr: target user address to start at
1912 * @pfn: physical address of kernel memory
1913 * @size: size of map area
1914 * @prot: page protection flags for this mapping
1916 * Note: this is only safe if the mm semaphore is held when called.
1918 * Return: %0 on success, negative error code otherwise.
1920 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1921 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1925 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1926 struct mm_struct
*mm
= vma
->vm_mm
;
1927 unsigned long remap_pfn
= pfn
;
1931 * Physically remapped pages are special. Tell the
1932 * rest of the world about it:
1933 * VM_IO tells people not to look at these pages
1934 * (accesses can have side effects).
1935 * VM_PFNMAP tells the core MM that the base pages are just
1936 * raw PFN mappings, and do not have a "struct page" associated
1939 * Disable vma merging and expanding with mremap().
1941 * Omit vma from core dump, even when VM_IO turned off.
1943 * There's a horrible special case to handle copy-on-write
1944 * behaviour that some programs depend on. We mark the "original"
1945 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1946 * See vm_normal_page() for details.
1948 if (is_cow_mapping(vma
->vm_flags
)) {
1949 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
1951 vma
->vm_pgoff
= pfn
;
1954 err
= track_pfn_remap(vma
, &prot
, remap_pfn
, addr
, PAGE_ALIGN(size
));
1958 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
1960 BUG_ON(addr
>= end
);
1961 pfn
-= addr
>> PAGE_SHIFT
;
1962 pgd
= pgd_offset(mm
, addr
);
1963 flush_cache_range(vma
, addr
, end
);
1965 next
= pgd_addr_end(addr
, end
);
1966 err
= remap_p4d_range(mm
, pgd
, addr
, next
,
1967 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1970 } while (pgd
++, addr
= next
, addr
!= end
);
1973 untrack_pfn(vma
, remap_pfn
, PAGE_ALIGN(size
));
1977 EXPORT_SYMBOL(remap_pfn_range
);
1980 * vm_iomap_memory - remap memory to userspace
1981 * @vma: user vma to map to
1982 * @start: start of area
1983 * @len: size of area
1985 * This is a simplified io_remap_pfn_range() for common driver use. The
1986 * driver just needs to give us the physical memory range to be mapped,
1987 * we'll figure out the rest from the vma information.
1989 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1990 * whatever write-combining details or similar.
1992 * Return: %0 on success, negative error code otherwise.
1994 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
1996 unsigned long vm_len
, pfn
, pages
;
1998 /* Check that the physical memory area passed in looks valid */
1999 if (start
+ len
< start
)
2002 * You *really* shouldn't map things that aren't page-aligned,
2003 * but we've historically allowed it because IO memory might
2004 * just have smaller alignment.
2006 len
+= start
& ~PAGE_MASK
;
2007 pfn
= start
>> PAGE_SHIFT
;
2008 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2009 if (pfn
+ pages
< pfn
)
2012 /* We start the mapping 'vm_pgoff' pages into the area */
2013 if (vma
->vm_pgoff
> pages
)
2015 pfn
+= vma
->vm_pgoff
;
2016 pages
-= vma
->vm_pgoff
;
2018 /* Can we fit all of the mapping? */
2019 vm_len
= vma
->vm_end
- vma
->vm_start
;
2020 if (vm_len
>> PAGE_SHIFT
> pages
)
2023 /* Ok, let it rip */
2024 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2026 EXPORT_SYMBOL(vm_iomap_memory
);
2028 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2029 unsigned long addr
, unsigned long end
,
2030 pte_fn_t fn
, void *data
)
2034 spinlock_t
*uninitialized_var(ptl
);
2036 pte
= (mm
== &init_mm
) ?
2037 pte_alloc_kernel(pmd
, addr
) :
2038 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2042 BUG_ON(pmd_huge(*pmd
));
2044 arch_enter_lazy_mmu_mode();
2047 err
= fn(pte
++, addr
, data
);
2050 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2052 arch_leave_lazy_mmu_mode();
2055 pte_unmap_unlock(pte
-1, ptl
);
2059 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2060 unsigned long addr
, unsigned long end
,
2061 pte_fn_t fn
, void *data
)
2067 BUG_ON(pud_huge(*pud
));
2069 pmd
= pmd_alloc(mm
, pud
, addr
);
2073 next
= pmd_addr_end(addr
, end
);
2074 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
2077 } while (pmd
++, addr
= next
, addr
!= end
);
2081 static int apply_to_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2082 unsigned long addr
, unsigned long end
,
2083 pte_fn_t fn
, void *data
)
2089 pud
= pud_alloc(mm
, p4d
, addr
);
2093 next
= pud_addr_end(addr
, end
);
2094 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
2097 } while (pud
++, addr
= next
, addr
!= end
);
2101 static int apply_to_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2102 unsigned long addr
, unsigned long end
,
2103 pte_fn_t fn
, void *data
)
2109 p4d
= p4d_alloc(mm
, pgd
, addr
);
2113 next
= p4d_addr_end(addr
, end
);
2114 err
= apply_to_pud_range(mm
, p4d
, addr
, next
, fn
, data
);
2117 } while (p4d
++, addr
= next
, addr
!= end
);
2122 * Scan a region of virtual memory, filling in page tables as necessary
2123 * and calling a provided function on each leaf page table.
2125 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2126 unsigned long size
, pte_fn_t fn
, void *data
)
2130 unsigned long end
= addr
+ size
;
2133 if (WARN_ON(addr
>= end
))
2136 pgd
= pgd_offset(mm
, addr
);
2138 next
= pgd_addr_end(addr
, end
);
2139 err
= apply_to_p4d_range(mm
, pgd
, addr
, next
, fn
, data
);
2142 } while (pgd
++, addr
= next
, addr
!= end
);
2146 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2149 * handle_pte_fault chooses page fault handler according to an entry which was
2150 * read non-atomically. Before making any commitment, on those architectures
2151 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2152 * parts, do_swap_page must check under lock before unmapping the pte and
2153 * proceeding (but do_wp_page is only called after already making such a check;
2154 * and do_anonymous_page can safely check later on).
2156 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2157 pte_t
*page_table
, pte_t orig_pte
)
2160 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2161 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2162 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2164 same
= pte_same(*page_table
, orig_pte
);
2168 pte_unmap(page_table
);
2172 static inline bool cow_user_page(struct page
*dst
, struct page
*src
,
2173 struct vm_fault
*vmf
)
2178 bool locked
= false;
2179 struct vm_area_struct
*vma
= vmf
->vma
;
2180 struct mm_struct
*mm
= vma
->vm_mm
;
2181 unsigned long addr
= vmf
->address
;
2183 debug_dma_assert_idle(src
);
2186 copy_user_highpage(dst
, src
, addr
, vma
);
2191 * If the source page was a PFN mapping, we don't have
2192 * a "struct page" for it. We do a best-effort copy by
2193 * just copying from the original user address. If that
2194 * fails, we just zero-fill it. Live with it.
2196 kaddr
= kmap_atomic(dst
);
2197 uaddr
= (void __user
*)(addr
& PAGE_MASK
);
2200 * On architectures with software "accessed" bits, we would
2201 * take a double page fault, so mark it accessed here.
2203 if (arch_faults_on_old_pte() && !pte_young(vmf
->orig_pte
)) {
2206 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2208 if (!likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2210 * Other thread has already handled the fault
2211 * and we don't need to do anything. If it's
2212 * not the case, the fault will be triggered
2213 * again on the same address.
2219 entry
= pte_mkyoung(vmf
->orig_pte
);
2220 if (ptep_set_access_flags(vma
, addr
, vmf
->pte
, entry
, 0))
2221 update_mmu_cache(vma
, addr
, vmf
->pte
);
2225 * This really shouldn't fail, because the page is there
2226 * in the page tables. But it might just be unreadable,
2227 * in which case we just give up and fill the result with
2230 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
2234 /* Re-validate under PTL if the page is still mapped */
2235 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2237 if (!likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2238 /* The PTE changed under us. Retry page fault. */
2244 * The same page can be mapped back since last copy attampt.
2245 * Try to copy again under PTL.
2247 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
2249 * Give a warn in case there can be some obscure
2262 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2263 kunmap_atomic(kaddr
);
2264 flush_dcache_page(dst
);
2269 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
2271 struct file
*vm_file
= vma
->vm_file
;
2274 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
2277 * Special mappings (e.g. VDSO) do not have any file so fake
2278 * a default GFP_KERNEL for them.
2284 * Notify the address space that the page is about to become writable so that
2285 * it can prohibit this or wait for the page to get into an appropriate state.
2287 * We do this without the lock held, so that it can sleep if it needs to.
2289 static vm_fault_t
do_page_mkwrite(struct vm_fault
*vmf
)
2292 struct page
*page
= vmf
->page
;
2293 unsigned int old_flags
= vmf
->flags
;
2295 vmf
->flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2297 if (vmf
->vma
->vm_file
&&
2298 IS_SWAPFILE(vmf
->vma
->vm_file
->f_mapping
->host
))
2299 return VM_FAULT_SIGBUS
;
2301 ret
= vmf
->vma
->vm_ops
->page_mkwrite(vmf
);
2302 /* Restore original flags so that caller is not surprised */
2303 vmf
->flags
= old_flags
;
2304 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2306 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
2308 if (!page
->mapping
) {
2310 return 0; /* retry */
2312 ret
|= VM_FAULT_LOCKED
;
2314 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2319 * Handle dirtying of a page in shared file mapping on a write fault.
2321 * The function expects the page to be locked and unlocks it.
2323 static vm_fault_t
fault_dirty_shared_page(struct vm_fault
*vmf
)
2325 struct vm_area_struct
*vma
= vmf
->vma
;
2326 struct address_space
*mapping
;
2327 struct page
*page
= vmf
->page
;
2329 bool page_mkwrite
= vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
;
2331 dirtied
= set_page_dirty(page
);
2332 VM_BUG_ON_PAGE(PageAnon(page
), page
);
2334 * Take a local copy of the address_space - page.mapping may be zeroed
2335 * by truncate after unlock_page(). The address_space itself remains
2336 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2337 * release semantics to prevent the compiler from undoing this copying.
2339 mapping
= page_rmapping(page
);
2343 file_update_time(vma
->vm_file
);
2346 * Throttle page dirtying rate down to writeback speed.
2348 * mapping may be NULL here because some device drivers do not
2349 * set page.mapping but still dirty their pages
2351 * Drop the mmap_sem before waiting on IO, if we can. The file
2352 * is pinning the mapping, as per above.
2354 if ((dirtied
|| page_mkwrite
) && mapping
) {
2357 fpin
= maybe_unlock_mmap_for_io(vmf
, NULL
);
2358 balance_dirty_pages_ratelimited(mapping
);
2361 return VM_FAULT_RETRY
;
2369 * Handle write page faults for pages that can be reused in the current vma
2371 * This can happen either due to the mapping being with the VM_SHARED flag,
2372 * or due to us being the last reference standing to the page. In either
2373 * case, all we need to do here is to mark the page as writable and update
2374 * any related book-keeping.
2376 static inline void wp_page_reuse(struct vm_fault
*vmf
)
2377 __releases(vmf
->ptl
)
2379 struct vm_area_struct
*vma
= vmf
->vma
;
2380 struct page
*page
= vmf
->page
;
2383 * Clear the pages cpupid information as the existing
2384 * information potentially belongs to a now completely
2385 * unrelated process.
2388 page_cpupid_xchg_last(page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2390 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2391 entry
= pte_mkyoung(vmf
->orig_pte
);
2392 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2393 if (ptep_set_access_flags(vma
, vmf
->address
, vmf
->pte
, entry
, 1))
2394 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2395 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2399 * Handle the case of a page which we actually need to copy to a new page.
2401 * Called with mmap_sem locked and the old page referenced, but
2402 * without the ptl held.
2404 * High level logic flow:
2406 * - Allocate a page, copy the content of the old page to the new one.
2407 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2408 * - Take the PTL. If the pte changed, bail out and release the allocated page
2409 * - If the pte is still the way we remember it, update the page table and all
2410 * relevant references. This includes dropping the reference the page-table
2411 * held to the old page, as well as updating the rmap.
2412 * - In any case, unlock the PTL and drop the reference we took to the old page.
2414 static vm_fault_t
wp_page_copy(struct vm_fault
*vmf
)
2416 struct vm_area_struct
*vma
= vmf
->vma
;
2417 struct mm_struct
*mm
= vma
->vm_mm
;
2418 struct page
*old_page
= vmf
->page
;
2419 struct page
*new_page
= NULL
;
2421 int page_copied
= 0;
2422 struct mem_cgroup
*memcg
;
2423 struct mmu_notifier_range range
;
2425 if (unlikely(anon_vma_prepare(vma
)))
2428 if (is_zero_pfn(pte_pfn(vmf
->orig_pte
))) {
2429 new_page
= alloc_zeroed_user_highpage_movable(vma
,
2434 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
2439 if (!cow_user_page(new_page
, old_page
, vmf
)) {
2441 * COW failed, if the fault was solved by other,
2442 * it's fine. If not, userspace would re-fault on
2443 * the same address and we will handle the fault
2444 * from the second attempt.
2453 if (mem_cgroup_try_charge_delay(new_page
, mm
, GFP_KERNEL
, &memcg
, false))
2456 __SetPageUptodate(new_page
);
2458 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
,
2459 vmf
->address
& PAGE_MASK
,
2460 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
);
2461 mmu_notifier_invalidate_range_start(&range
);
2464 * Re-check the pte - we dropped the lock
2466 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, vmf
->address
, &vmf
->ptl
);
2467 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2469 if (!PageAnon(old_page
)) {
2470 dec_mm_counter_fast(mm
,
2471 mm_counter_file(old_page
));
2472 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2475 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2477 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2478 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2479 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2481 * Clear the pte entry and flush it first, before updating the
2482 * pte with the new entry. This will avoid a race condition
2483 * seen in the presence of one thread doing SMC and another
2486 ptep_clear_flush_notify(vma
, vmf
->address
, vmf
->pte
);
2487 page_add_new_anon_rmap(new_page
, vma
, vmf
->address
, false);
2488 mem_cgroup_commit_charge(new_page
, memcg
, false, false);
2489 lru_cache_add_active_or_unevictable(new_page
, vma
);
2491 * We call the notify macro here because, when using secondary
2492 * mmu page tables (such as kvm shadow page tables), we want the
2493 * new page to be mapped directly into the secondary page table.
2495 set_pte_at_notify(mm
, vmf
->address
, vmf
->pte
, entry
);
2496 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2499 * Only after switching the pte to the new page may
2500 * we remove the mapcount here. Otherwise another
2501 * process may come and find the rmap count decremented
2502 * before the pte is switched to the new page, and
2503 * "reuse" the old page writing into it while our pte
2504 * here still points into it and can be read by other
2507 * The critical issue is to order this
2508 * page_remove_rmap with the ptp_clear_flush above.
2509 * Those stores are ordered by (if nothing else,)
2510 * the barrier present in the atomic_add_negative
2511 * in page_remove_rmap.
2513 * Then the TLB flush in ptep_clear_flush ensures that
2514 * no process can access the old page before the
2515 * decremented mapcount is visible. And the old page
2516 * cannot be reused until after the decremented
2517 * mapcount is visible. So transitively, TLBs to
2518 * old page will be flushed before it can be reused.
2520 page_remove_rmap(old_page
, false);
2523 /* Free the old page.. */
2524 new_page
= old_page
;
2527 mem_cgroup_cancel_charge(new_page
, memcg
, false);
2533 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2535 * No need to double call mmu_notifier->invalidate_range() callback as
2536 * the above ptep_clear_flush_notify() did already call it.
2538 mmu_notifier_invalidate_range_only_end(&range
);
2541 * Don't let another task, with possibly unlocked vma,
2542 * keep the mlocked page.
2544 if (page_copied
&& (vma
->vm_flags
& VM_LOCKED
)) {
2545 lock_page(old_page
); /* LRU manipulation */
2546 if (PageMlocked(old_page
))
2547 munlock_vma_page(old_page
);
2548 unlock_page(old_page
);
2552 return page_copied
? VM_FAULT_WRITE
: 0;
2558 return VM_FAULT_OOM
;
2562 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2563 * writeable once the page is prepared
2565 * @vmf: structure describing the fault
2567 * This function handles all that is needed to finish a write page fault in a
2568 * shared mapping due to PTE being read-only once the mapped page is prepared.
2569 * It handles locking of PTE and modifying it.
2571 * The function expects the page to be locked or other protection against
2572 * concurrent faults / writeback (such as DAX radix tree locks).
2574 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2575 * we acquired PTE lock.
2577 vm_fault_t
finish_mkwrite_fault(struct vm_fault
*vmf
)
2579 WARN_ON_ONCE(!(vmf
->vma
->vm_flags
& VM_SHARED
));
2580 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
2583 * We might have raced with another page fault while we released the
2584 * pte_offset_map_lock.
2586 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
2587 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2588 return VM_FAULT_NOPAGE
;
2595 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2598 static vm_fault_t
wp_pfn_shared(struct vm_fault
*vmf
)
2600 struct vm_area_struct
*vma
= vmf
->vma
;
2602 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
2605 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2606 vmf
->flags
|= FAULT_FLAG_MKWRITE
;
2607 ret
= vma
->vm_ops
->pfn_mkwrite(vmf
);
2608 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))
2610 return finish_mkwrite_fault(vmf
);
2613 return VM_FAULT_WRITE
;
2616 static vm_fault_t
wp_page_shared(struct vm_fault
*vmf
)
2617 __releases(vmf
->ptl
)
2619 struct vm_area_struct
*vma
= vmf
->vma
;
2620 vm_fault_t ret
= VM_FAULT_WRITE
;
2622 get_page(vmf
->page
);
2624 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2627 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2628 tmp
= do_page_mkwrite(vmf
);
2629 if (unlikely(!tmp
|| (tmp
&
2630 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
2631 put_page(vmf
->page
);
2634 tmp
= finish_mkwrite_fault(vmf
);
2635 if (unlikely(tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
2636 unlock_page(vmf
->page
);
2637 put_page(vmf
->page
);
2642 lock_page(vmf
->page
);
2644 ret
|= fault_dirty_shared_page(vmf
);
2645 put_page(vmf
->page
);
2651 * This routine handles present pages, when users try to write
2652 * to a shared page. It is done by copying the page to a new address
2653 * and decrementing the shared-page counter for the old page.
2655 * Note that this routine assumes that the protection checks have been
2656 * done by the caller (the low-level page fault routine in most cases).
2657 * Thus we can safely just mark it writable once we've done any necessary
2660 * We also mark the page dirty at this point even though the page will
2661 * change only once the write actually happens. This avoids a few races,
2662 * and potentially makes it more efficient.
2664 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2665 * but allow concurrent faults), with pte both mapped and locked.
2666 * We return with mmap_sem still held, but pte unmapped and unlocked.
2668 static vm_fault_t
do_wp_page(struct vm_fault
*vmf
)
2669 __releases(vmf
->ptl
)
2671 struct vm_area_struct
*vma
= vmf
->vma
;
2673 vmf
->page
= vm_normal_page(vma
, vmf
->address
, vmf
->orig_pte
);
2676 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2679 * We should not cow pages in a shared writeable mapping.
2680 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2682 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2683 (VM_WRITE
|VM_SHARED
))
2684 return wp_pfn_shared(vmf
);
2686 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2687 return wp_page_copy(vmf
);
2691 * Take out anonymous pages first, anonymous shared vmas are
2692 * not dirty accountable.
2694 if (PageAnon(vmf
->page
)) {
2695 int total_map_swapcount
;
2696 if (PageKsm(vmf
->page
) && (PageSwapCache(vmf
->page
) ||
2697 page_count(vmf
->page
) != 1))
2699 if (!trylock_page(vmf
->page
)) {
2700 get_page(vmf
->page
);
2701 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2702 lock_page(vmf
->page
);
2703 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
2704 vmf
->address
, &vmf
->ptl
);
2705 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
2706 unlock_page(vmf
->page
);
2707 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2708 put_page(vmf
->page
);
2711 put_page(vmf
->page
);
2713 if (PageKsm(vmf
->page
)) {
2714 bool reused
= reuse_ksm_page(vmf
->page
, vmf
->vma
,
2716 unlock_page(vmf
->page
);
2720 return VM_FAULT_WRITE
;
2722 if (reuse_swap_page(vmf
->page
, &total_map_swapcount
)) {
2723 if (total_map_swapcount
== 1) {
2725 * The page is all ours. Move it to
2726 * our anon_vma so the rmap code will
2727 * not search our parent or siblings.
2728 * Protected against the rmap code by
2731 page_move_anon_rmap(vmf
->page
, vma
);
2733 unlock_page(vmf
->page
);
2735 return VM_FAULT_WRITE
;
2737 unlock_page(vmf
->page
);
2738 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2739 (VM_WRITE
|VM_SHARED
))) {
2740 return wp_page_shared(vmf
);
2744 * Ok, we need to copy. Oh, well..
2746 get_page(vmf
->page
);
2748 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2749 return wp_page_copy(vmf
);
2752 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2753 unsigned long start_addr
, unsigned long end_addr
,
2754 struct zap_details
*details
)
2756 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2759 static inline void unmap_mapping_range_tree(struct rb_root_cached
*root
,
2760 struct zap_details
*details
)
2762 struct vm_area_struct
*vma
;
2763 pgoff_t vba
, vea
, zba
, zea
;
2765 vma_interval_tree_foreach(vma
, root
,
2766 details
->first_index
, details
->last_index
) {
2768 vba
= vma
->vm_pgoff
;
2769 vea
= vba
+ vma_pages(vma
) - 1;
2770 zba
= details
->first_index
;
2773 zea
= details
->last_index
;
2777 unmap_mapping_range_vma(vma
,
2778 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2779 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2785 * unmap_mapping_page() - Unmap single page from processes.
2786 * @page: The locked page to be unmapped.
2788 * Unmap this page from any userspace process which still has it mmaped.
2789 * Typically, for efficiency, the range of nearby pages has already been
2790 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
2791 * truncation or invalidation holds the lock on a page, it may find that
2792 * the page has been remapped again: and then uses unmap_mapping_page()
2793 * to unmap it finally.
2795 void unmap_mapping_page(struct page
*page
)
2797 struct address_space
*mapping
= page
->mapping
;
2798 struct zap_details details
= { };
2800 VM_BUG_ON(!PageLocked(page
));
2801 VM_BUG_ON(PageTail(page
));
2803 details
.check_mapping
= mapping
;
2804 details
.first_index
= page
->index
;
2805 details
.last_index
= page
->index
+ hpage_nr_pages(page
) - 1;
2806 details
.single_page
= page
;
2808 i_mmap_lock_write(mapping
);
2809 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
2810 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2811 i_mmap_unlock_write(mapping
);
2815 * unmap_mapping_pages() - Unmap pages from processes.
2816 * @mapping: The address space containing pages to be unmapped.
2817 * @start: Index of first page to be unmapped.
2818 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2819 * @even_cows: Whether to unmap even private COWed pages.
2821 * Unmap the pages in this address space from any userspace process which
2822 * has them mmaped. Generally, you want to remove COWed pages as well when
2823 * a file is being truncated, but not when invalidating pages from the page
2826 void unmap_mapping_pages(struct address_space
*mapping
, pgoff_t start
,
2827 pgoff_t nr
, bool even_cows
)
2829 struct zap_details details
= { };
2831 details
.check_mapping
= even_cows
? NULL
: mapping
;
2832 details
.first_index
= start
;
2833 details
.last_index
= start
+ nr
- 1;
2834 if (details
.last_index
< details
.first_index
)
2835 details
.last_index
= ULONG_MAX
;
2837 i_mmap_lock_write(mapping
);
2838 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
2839 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2840 i_mmap_unlock_write(mapping
);
2844 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2845 * address_space corresponding to the specified byte range in the underlying
2848 * @mapping: the address space containing mmaps to be unmapped.
2849 * @holebegin: byte in first page to unmap, relative to the start of
2850 * the underlying file. This will be rounded down to a PAGE_SIZE
2851 * boundary. Note that this is different from truncate_pagecache(), which
2852 * must keep the partial page. In contrast, we must get rid of
2854 * @holelen: size of prospective hole in bytes. This will be rounded
2855 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2857 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2858 * but 0 when invalidating pagecache, don't throw away private data.
2860 void unmap_mapping_range(struct address_space
*mapping
,
2861 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2863 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2864 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2866 /* Check for overflow. */
2867 if (sizeof(holelen
) > sizeof(hlen
)) {
2869 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2870 if (holeend
& ~(long long)ULONG_MAX
)
2871 hlen
= ULONG_MAX
- hba
+ 1;
2874 unmap_mapping_pages(mapping
, hba
, hlen
, even_cows
);
2876 EXPORT_SYMBOL(unmap_mapping_range
);
2879 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2880 * but allow concurrent faults), and pte mapped but not yet locked.
2881 * We return with pte unmapped and unlocked.
2883 * We return with the mmap_sem locked or unlocked in the same cases
2884 * as does filemap_fault().
2886 vm_fault_t
do_swap_page(struct vm_fault
*vmf
)
2888 struct vm_area_struct
*vma
= vmf
->vma
;
2889 struct page
*page
= NULL
, *swapcache
;
2890 struct mem_cgroup
*memcg
;
2897 if (!pte_unmap_same(vma
->vm_mm
, vmf
->pmd
, vmf
->pte
, vmf
->orig_pte
))
2900 entry
= pte_to_swp_entry(vmf
->orig_pte
);
2901 if (unlikely(non_swap_entry(entry
))) {
2902 if (is_migration_entry(entry
)) {
2903 migration_entry_wait(vma
->vm_mm
, vmf
->pmd
,
2905 } else if (is_device_private_entry(entry
)) {
2906 vmf
->page
= device_private_entry_to_page(entry
);
2907 ret
= vmf
->page
->pgmap
->ops
->migrate_to_ram(vmf
);
2908 } else if (is_hwpoison_entry(entry
)) {
2909 ret
= VM_FAULT_HWPOISON
;
2911 print_bad_pte(vma
, vmf
->address
, vmf
->orig_pte
, NULL
);
2912 ret
= VM_FAULT_SIGBUS
;
2918 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2919 page
= lookup_swap_cache(entry
, vma
, vmf
->address
);
2923 struct swap_info_struct
*si
= swp_swap_info(entry
);
2925 if (si
->flags
& SWP_SYNCHRONOUS_IO
&&
2926 __swap_count(entry
) == 1) {
2927 /* skip swapcache */
2928 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
2931 __SetPageLocked(page
);
2932 __SetPageSwapBacked(page
);
2933 set_page_private(page
, entry
.val
);
2934 lru_cache_add_anon(page
);
2935 swap_readpage(page
, true);
2938 page
= swapin_readahead(entry
, GFP_HIGHUSER_MOVABLE
,
2945 * Back out if somebody else faulted in this pte
2946 * while we released the pte lock.
2948 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
2949 vmf
->address
, &vmf
->ptl
);
2950 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
)))
2952 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2956 /* Had to read the page from swap area: Major fault */
2957 ret
= VM_FAULT_MAJOR
;
2958 count_vm_event(PGMAJFAULT
);
2959 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
2960 } else if (PageHWPoison(page
)) {
2962 * hwpoisoned dirty swapcache pages are kept for killing
2963 * owner processes (which may be unknown at hwpoison time)
2965 ret
= VM_FAULT_HWPOISON
;
2966 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2970 locked
= lock_page_or_retry(page
, vma
->vm_mm
, vmf
->flags
);
2972 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2974 ret
|= VM_FAULT_RETRY
;
2979 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2980 * release the swapcache from under us. The page pin, and pte_same
2981 * test below, are not enough to exclude that. Even if it is still
2982 * swapcache, we need to check that the page's swap has not changed.
2984 if (unlikely((!PageSwapCache(page
) ||
2985 page_private(page
) != entry
.val
)) && swapcache
)
2988 page
= ksm_might_need_to_copy(page
, vma
, vmf
->address
);
2989 if (unlikely(!page
)) {
2995 if (mem_cgroup_try_charge_delay(page
, vma
->vm_mm
, GFP_KERNEL
,
3002 * Back out if somebody else already faulted in this pte.
3004 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3006 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3009 if (unlikely(!PageUptodate(page
))) {
3010 ret
= VM_FAULT_SIGBUS
;
3015 * The page isn't present yet, go ahead with the fault.
3017 * Be careful about the sequence of operations here.
3018 * To get its accounting right, reuse_swap_page() must be called
3019 * while the page is counted on swap but not yet in mapcount i.e.
3020 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3021 * must be called after the swap_free(), or it will never succeed.
3024 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3025 dec_mm_counter_fast(vma
->vm_mm
, MM_SWAPENTS
);
3026 pte
= mk_pte(page
, vma
->vm_page_prot
);
3027 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
, NULL
)) {
3028 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3029 vmf
->flags
&= ~FAULT_FLAG_WRITE
;
3030 ret
|= VM_FAULT_WRITE
;
3031 exclusive
= RMAP_EXCLUSIVE
;
3033 flush_icache_page(vma
, page
);
3034 if (pte_swp_soft_dirty(vmf
->orig_pte
))
3035 pte
= pte_mksoft_dirty(pte
);
3036 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3037 arch_do_swap_page(vma
->vm_mm
, vma
, vmf
->address
, pte
, vmf
->orig_pte
);
3038 vmf
->orig_pte
= pte
;
3040 /* ksm created a completely new copy */
3041 if (unlikely(page
!= swapcache
&& swapcache
)) {
3042 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3043 mem_cgroup_commit_charge(page
, memcg
, false, false);
3044 lru_cache_add_active_or_unevictable(page
, vma
);
3046 do_page_add_anon_rmap(page
, vma
, vmf
->address
, exclusive
);
3047 mem_cgroup_commit_charge(page
, memcg
, true, false);
3048 activate_page(page
);
3052 if (mem_cgroup_swap_full(page
) ||
3053 (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
3054 try_to_free_swap(page
);
3056 if (page
!= swapcache
&& swapcache
) {
3058 * Hold the lock to avoid the swap entry to be reused
3059 * until we take the PT lock for the pte_same() check
3060 * (to avoid false positives from pte_same). For
3061 * further safety release the lock after the swap_free
3062 * so that the swap count won't change under a
3063 * parallel locked swapcache.
3065 unlock_page(swapcache
);
3066 put_page(swapcache
);
3069 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
3070 ret
|= do_wp_page(vmf
);
3071 if (ret
& VM_FAULT_ERROR
)
3072 ret
&= VM_FAULT_ERROR
;
3076 /* No need to invalidate - it was non-present before */
3077 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3079 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3083 mem_cgroup_cancel_charge(page
, memcg
, false);
3084 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3089 if (page
!= swapcache
&& swapcache
) {
3090 unlock_page(swapcache
);
3091 put_page(swapcache
);
3097 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3098 * but allow concurrent faults), and pte mapped but not yet locked.
3099 * We return with mmap_sem still held, but pte unmapped and unlocked.
3101 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
)
3103 struct vm_area_struct
*vma
= vmf
->vma
;
3104 struct mem_cgroup
*memcg
;
3109 /* File mapping without ->vm_ops ? */
3110 if (vma
->vm_flags
& VM_SHARED
)
3111 return VM_FAULT_SIGBUS
;
3114 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3115 * pte_offset_map() on pmds where a huge pmd might be created
3116 * from a different thread.
3118 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3119 * parallel threads are excluded by other means.
3121 * Here we only have down_read(mmap_sem).
3123 if (pte_alloc(vma
->vm_mm
, vmf
->pmd
))
3124 return VM_FAULT_OOM
;
3126 /* See the comment in pte_alloc_one_map() */
3127 if (unlikely(pmd_trans_unstable(vmf
->pmd
)))
3130 /* Use the zero-page for reads */
3131 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
3132 !mm_forbids_zeropage(vma
->vm_mm
)) {
3133 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(vmf
->address
),
3134 vma
->vm_page_prot
));
3135 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3136 vmf
->address
, &vmf
->ptl
);
3137 if (!pte_none(*vmf
->pte
))
3139 ret
= check_stable_address_space(vma
->vm_mm
);
3142 /* Deliver the page fault to userland, check inside PT lock */
3143 if (userfaultfd_missing(vma
)) {
3144 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3145 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3150 /* Allocate our own private page. */
3151 if (unlikely(anon_vma_prepare(vma
)))
3153 page
= alloc_zeroed_user_highpage_movable(vma
, vmf
->address
);
3157 if (mem_cgroup_try_charge_delay(page
, vma
->vm_mm
, GFP_KERNEL
, &memcg
,
3162 * The memory barrier inside __SetPageUptodate makes sure that
3163 * preceeding stores to the page contents become visible before
3164 * the set_pte_at() write.
3166 __SetPageUptodate(page
);
3168 entry
= mk_pte(page
, vma
->vm_page_prot
);
3169 if (vma
->vm_flags
& VM_WRITE
)
3170 entry
= pte_mkwrite(pte_mkdirty(entry
));
3172 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3174 if (!pte_none(*vmf
->pte
))
3177 ret
= check_stable_address_space(vma
->vm_mm
);
3181 /* Deliver the page fault to userland, check inside PT lock */
3182 if (userfaultfd_missing(vma
)) {
3183 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3184 mem_cgroup_cancel_charge(page
, memcg
, false);
3186 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3189 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3190 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3191 mem_cgroup_commit_charge(page
, memcg
, false, false);
3192 lru_cache_add_active_or_unevictable(page
, vma
);
3194 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3196 /* No need to invalidate - it was non-present before */
3197 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3199 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3202 mem_cgroup_cancel_charge(page
, memcg
, false);
3208 return VM_FAULT_OOM
;
3212 * The mmap_sem must have been held on entry, and may have been
3213 * released depending on flags and vma->vm_ops->fault() return value.
3214 * See filemap_fault() and __lock_page_retry().
3216 static vm_fault_t
__do_fault(struct vm_fault
*vmf
)
3218 struct vm_area_struct
*vma
= vmf
->vma
;
3222 * Preallocate pte before we take page_lock because this might lead to
3223 * deadlocks for memcg reclaim which waits for pages under writeback:
3225 * SetPageWriteback(A)
3231 * wait_on_page_writeback(A)
3232 * SetPageWriteback(B)
3234 * # flush A, B to clear the writeback
3236 if (pmd_none(*vmf
->pmd
) && !vmf
->prealloc_pte
) {
3237 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
);
3238 if (!vmf
->prealloc_pte
)
3239 return VM_FAULT_OOM
;
3240 smp_wmb(); /* See comment in __pte_alloc() */
3243 ret
= vma
->vm_ops
->fault(vmf
);
3244 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
|
3245 VM_FAULT_DONE_COW
)))
3248 if (unlikely(PageHWPoison(vmf
->page
))) {
3249 if (ret
& VM_FAULT_LOCKED
)
3250 unlock_page(vmf
->page
);
3251 put_page(vmf
->page
);
3253 return VM_FAULT_HWPOISON
;
3256 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3257 lock_page(vmf
->page
);
3259 VM_BUG_ON_PAGE(!PageLocked(vmf
->page
), vmf
->page
);
3265 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3266 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3267 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3268 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3270 static int pmd_devmap_trans_unstable(pmd_t
*pmd
)
3272 return pmd_devmap(*pmd
) || pmd_trans_unstable(pmd
);
3275 static vm_fault_t
pte_alloc_one_map(struct vm_fault
*vmf
)
3277 struct vm_area_struct
*vma
= vmf
->vma
;
3279 if (!pmd_none(*vmf
->pmd
))
3281 if (vmf
->prealloc_pte
) {
3282 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3283 if (unlikely(!pmd_none(*vmf
->pmd
))) {
3284 spin_unlock(vmf
->ptl
);
3288 mm_inc_nr_ptes(vma
->vm_mm
);
3289 pmd_populate(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3290 spin_unlock(vmf
->ptl
);
3291 vmf
->prealloc_pte
= NULL
;
3292 } else if (unlikely(pte_alloc(vma
->vm_mm
, vmf
->pmd
))) {
3293 return VM_FAULT_OOM
;
3297 * If a huge pmd materialized under us just retry later. Use
3298 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3299 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3300 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3301 * running immediately after a huge pmd fault in a different thread of
3302 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3303 * All we have to ensure is that it is a regular pmd that we can walk
3304 * with pte_offset_map() and we can do that through an atomic read in
3305 * C, which is what pmd_trans_unstable() provides.
3307 if (pmd_devmap_trans_unstable(vmf
->pmd
))
3308 return VM_FAULT_NOPAGE
;
3311 * At this point we know that our vmf->pmd points to a page of ptes
3312 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3313 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3314 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3315 * be valid and we will re-check to make sure the vmf->pte isn't
3316 * pte_none() under vmf->ptl protection when we return to
3319 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3324 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3325 static void deposit_prealloc_pte(struct vm_fault
*vmf
)
3327 struct vm_area_struct
*vma
= vmf
->vma
;
3329 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3331 * We are going to consume the prealloc table,
3332 * count that as nr_ptes.
3334 mm_inc_nr_ptes(vma
->vm_mm
);
3335 vmf
->prealloc_pte
= NULL
;
3338 static vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3340 struct vm_area_struct
*vma
= vmf
->vma
;
3341 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3342 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
3347 if (!transhuge_vma_suitable(vma
, haddr
))
3348 return VM_FAULT_FALLBACK
;
3350 ret
= VM_FAULT_FALLBACK
;
3351 page
= compound_head(page
);
3354 * Archs like ppc64 need additonal space to store information
3355 * related to pte entry. Use the preallocated table for that.
3357 if (arch_needs_pgtable_deposit() && !vmf
->prealloc_pte
) {
3358 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
3359 if (!vmf
->prealloc_pte
)
3360 return VM_FAULT_OOM
;
3361 smp_wmb(); /* See comment in __pte_alloc() */
3364 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3365 if (unlikely(!pmd_none(*vmf
->pmd
)))
3368 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
3369 flush_icache_page(vma
, page
+ i
);
3371 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
3373 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
3375 add_mm_counter(vma
->vm_mm
, mm_counter_file(page
), HPAGE_PMD_NR
);
3376 page_add_file_rmap(page
, true);
3378 * deposit and withdraw with pmd lock held
3380 if (arch_needs_pgtable_deposit())
3381 deposit_prealloc_pte(vmf
);
3383 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
3385 update_mmu_cache_pmd(vma
, haddr
, vmf
->pmd
);
3387 /* fault is handled */
3389 count_vm_event(THP_FILE_MAPPED
);
3391 spin_unlock(vmf
->ptl
);
3395 static vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3403 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3404 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3406 * @vmf: fault environment
3407 * @memcg: memcg to charge page (only for private mappings)
3408 * @page: page to map
3410 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3413 * Target users are page handler itself and implementations of
3414 * vm_ops->map_pages.
3416 * Return: %0 on success, %VM_FAULT_ code in case of error.
3418 vm_fault_t
alloc_set_pte(struct vm_fault
*vmf
, struct mem_cgroup
*memcg
,
3421 struct vm_area_struct
*vma
= vmf
->vma
;
3422 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3426 if (pmd_none(*vmf
->pmd
) && PageTransCompound(page
) &&
3427 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE
)) {
3429 VM_BUG_ON_PAGE(memcg
, page
);
3431 ret
= do_set_pmd(vmf
, page
);
3432 if (ret
!= VM_FAULT_FALLBACK
)
3437 ret
= pte_alloc_one_map(vmf
);
3442 /* Re-check under ptl */
3443 if (unlikely(!pte_none(*vmf
->pte
)))
3444 return VM_FAULT_NOPAGE
;
3446 flush_icache_page(vma
, page
);
3447 entry
= mk_pte(page
, vma
->vm_page_prot
);
3449 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3450 /* copy-on-write page */
3451 if (write
&& !(vma
->vm_flags
& VM_SHARED
)) {
3452 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3453 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3454 mem_cgroup_commit_charge(page
, memcg
, false, false);
3455 lru_cache_add_active_or_unevictable(page
, vma
);
3457 inc_mm_counter_fast(vma
->vm_mm
, mm_counter_file(page
));
3458 page_add_file_rmap(page
, false);
3460 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3462 /* no need to invalidate: a not-present page won't be cached */
3463 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3470 * finish_fault - finish page fault once we have prepared the page to fault
3472 * @vmf: structure describing the fault
3474 * This function handles all that is needed to finish a page fault once the
3475 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3476 * given page, adds reverse page mapping, handles memcg charges and LRU
3479 * The function expects the page to be locked and on success it consumes a
3480 * reference of a page being mapped (for the PTE which maps it).
3482 * Return: %0 on success, %VM_FAULT_ code in case of error.
3484 vm_fault_t
finish_fault(struct vm_fault
*vmf
)
3489 /* Did we COW the page? */
3490 if ((vmf
->flags
& FAULT_FLAG_WRITE
) &&
3491 !(vmf
->vma
->vm_flags
& VM_SHARED
))
3492 page
= vmf
->cow_page
;
3497 * check even for read faults because we might have lost our CoWed
3500 if (!(vmf
->vma
->vm_flags
& VM_SHARED
))
3501 ret
= check_stable_address_space(vmf
->vma
->vm_mm
);
3503 ret
= alloc_set_pte(vmf
, vmf
->memcg
, page
);
3505 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3509 static unsigned long fault_around_bytes __read_mostly
=
3510 rounddown_pow_of_two(65536);
3512 #ifdef CONFIG_DEBUG_FS
3513 static int fault_around_bytes_get(void *data
, u64
*val
)
3515 *val
= fault_around_bytes
;
3520 * fault_around_bytes must be rounded down to the nearest page order as it's
3521 * what do_fault_around() expects to see.
3523 static int fault_around_bytes_set(void *data
, u64 val
)
3525 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
3527 if (val
> PAGE_SIZE
)
3528 fault_around_bytes
= rounddown_pow_of_two(val
);
3530 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
3533 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops
,
3534 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
3536 static int __init
fault_around_debugfs(void)
3538 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL
, NULL
,
3539 &fault_around_bytes_fops
);
3542 late_initcall(fault_around_debugfs
);
3546 * do_fault_around() tries to map few pages around the fault address. The hope
3547 * is that the pages will be needed soon and this will lower the number of
3550 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3551 * not ready to be mapped: not up-to-date, locked, etc.
3553 * This function is called with the page table lock taken. In the split ptlock
3554 * case the page table lock only protects only those entries which belong to
3555 * the page table corresponding to the fault address.
3557 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3560 * fault_around_bytes defines how many bytes we'll try to map.
3561 * do_fault_around() expects it to be set to a power of two less than or equal
3564 * The virtual address of the area that we map is naturally aligned to
3565 * fault_around_bytes rounded down to the machine page size
3566 * (and therefore to page order). This way it's easier to guarantee
3567 * that we don't cross page table boundaries.
3569 static vm_fault_t
do_fault_around(struct vm_fault
*vmf
)
3571 unsigned long address
= vmf
->address
, nr_pages
, mask
;
3572 pgoff_t start_pgoff
= vmf
->pgoff
;
3577 nr_pages
= READ_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
3578 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
3580 vmf
->address
= max(address
& mask
, vmf
->vma
->vm_start
);
3581 off
= ((address
- vmf
->address
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
3585 * end_pgoff is either the end of the page table, the end of
3586 * the vma or nr_pages from start_pgoff, depending what is nearest.
3588 end_pgoff
= start_pgoff
-
3589 ((vmf
->address
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
3591 end_pgoff
= min3(end_pgoff
, vma_pages(vmf
->vma
) + vmf
->vma
->vm_pgoff
- 1,
3592 start_pgoff
+ nr_pages
- 1);
3594 if (pmd_none(*vmf
->pmd
)) {
3595 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
);
3596 if (!vmf
->prealloc_pte
)
3598 smp_wmb(); /* See comment in __pte_alloc() */
3601 vmf
->vma
->vm_ops
->map_pages(vmf
, start_pgoff
, end_pgoff
);
3603 /* Huge page is mapped? Page fault is solved */
3604 if (pmd_trans_huge(*vmf
->pmd
)) {
3605 ret
= VM_FAULT_NOPAGE
;
3609 /* ->map_pages() haven't done anything useful. Cold page cache? */
3613 /* check if the page fault is solved */
3614 vmf
->pte
-= (vmf
->address
>> PAGE_SHIFT
) - (address
>> PAGE_SHIFT
);
3615 if (!pte_none(*vmf
->pte
))
3616 ret
= VM_FAULT_NOPAGE
;
3617 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3619 vmf
->address
= address
;
3624 static vm_fault_t
do_read_fault(struct vm_fault
*vmf
)
3626 struct vm_area_struct
*vma
= vmf
->vma
;
3630 * Let's call ->map_pages() first and use ->fault() as fallback
3631 * if page by the offset is not ready to be mapped (cold cache or
3634 if (vma
->vm_ops
->map_pages
&& fault_around_bytes
>> PAGE_SHIFT
> 1) {
3635 ret
= do_fault_around(vmf
);
3640 ret
= __do_fault(vmf
);
3641 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3644 ret
|= finish_fault(vmf
);
3645 unlock_page(vmf
->page
);
3646 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3647 put_page(vmf
->page
);
3651 static vm_fault_t
do_cow_fault(struct vm_fault
*vmf
)
3653 struct vm_area_struct
*vma
= vmf
->vma
;
3656 if (unlikely(anon_vma_prepare(vma
)))
3657 return VM_FAULT_OOM
;
3659 vmf
->cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, vmf
->address
);
3661 return VM_FAULT_OOM
;
3663 if (mem_cgroup_try_charge_delay(vmf
->cow_page
, vma
->vm_mm
, GFP_KERNEL
,
3664 &vmf
->memcg
, false)) {
3665 put_page(vmf
->cow_page
);
3666 return VM_FAULT_OOM
;
3669 ret
= __do_fault(vmf
);
3670 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3672 if (ret
& VM_FAULT_DONE_COW
)
3675 copy_user_highpage(vmf
->cow_page
, vmf
->page
, vmf
->address
, vma
);
3676 __SetPageUptodate(vmf
->cow_page
);
3678 ret
|= finish_fault(vmf
);
3679 unlock_page(vmf
->page
);
3680 put_page(vmf
->page
);
3681 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3685 mem_cgroup_cancel_charge(vmf
->cow_page
, vmf
->memcg
, false);
3686 put_page(vmf
->cow_page
);
3690 static vm_fault_t
do_shared_fault(struct vm_fault
*vmf
)
3692 struct vm_area_struct
*vma
= vmf
->vma
;
3693 vm_fault_t ret
, tmp
;
3695 ret
= __do_fault(vmf
);
3696 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3700 * Check if the backing address space wants to know that the page is
3701 * about to become writable
3703 if (vma
->vm_ops
->page_mkwrite
) {
3704 unlock_page(vmf
->page
);
3705 tmp
= do_page_mkwrite(vmf
);
3706 if (unlikely(!tmp
||
3707 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3708 put_page(vmf
->page
);
3713 ret
|= finish_fault(vmf
);
3714 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
3716 unlock_page(vmf
->page
);
3717 put_page(vmf
->page
);
3721 ret
|= fault_dirty_shared_page(vmf
);
3726 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3727 * but allow concurrent faults).
3728 * The mmap_sem may have been released depending on flags and our
3729 * return value. See filemap_fault() and __lock_page_or_retry().
3730 * If mmap_sem is released, vma may become invalid (for example
3731 * by other thread calling munmap()).
3733 static vm_fault_t
do_fault(struct vm_fault
*vmf
)
3735 struct vm_area_struct
*vma
= vmf
->vma
;
3736 struct mm_struct
*vm_mm
= vma
->vm_mm
;
3740 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3742 if (!vma
->vm_ops
->fault
) {
3744 * If we find a migration pmd entry or a none pmd entry, which
3745 * should never happen, return SIGBUS
3747 if (unlikely(!pmd_present(*vmf
->pmd
)))
3748 ret
= VM_FAULT_SIGBUS
;
3750 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
,
3755 * Make sure this is not a temporary clearing of pte
3756 * by holding ptl and checking again. A R/M/W update
3757 * of pte involves: take ptl, clearing the pte so that
3758 * we don't have concurrent modification by hardware
3759 * followed by an update.
3761 if (unlikely(pte_none(*vmf
->pte
)))
3762 ret
= VM_FAULT_SIGBUS
;
3764 ret
= VM_FAULT_NOPAGE
;
3766 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3768 } else if (!(vmf
->flags
& FAULT_FLAG_WRITE
))
3769 ret
= do_read_fault(vmf
);
3770 else if (!(vma
->vm_flags
& VM_SHARED
))
3771 ret
= do_cow_fault(vmf
);
3773 ret
= do_shared_fault(vmf
);
3775 /* preallocated pagetable is unused: free it */
3776 if (vmf
->prealloc_pte
) {
3777 pte_free(vm_mm
, vmf
->prealloc_pte
);
3778 vmf
->prealloc_pte
= NULL
;
3783 static int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3784 unsigned long addr
, int page_nid
,
3789 count_vm_numa_event(NUMA_HINT_FAULTS
);
3790 if (page_nid
== numa_node_id()) {
3791 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3792 *flags
|= TNF_FAULT_LOCAL
;
3795 return mpol_misplaced(page
, vma
, addr
);
3798 static vm_fault_t
do_numa_page(struct vm_fault
*vmf
)
3800 struct vm_area_struct
*vma
= vmf
->vma
;
3801 struct page
*page
= NULL
;
3802 int page_nid
= NUMA_NO_NODE
;
3805 bool migrated
= false;
3807 bool was_writable
= pte_savedwrite(vmf
->orig_pte
);
3811 * The "pte" at this point cannot be used safely without
3812 * validation through pte_unmap_same(). It's of NUMA type but
3813 * the pfn may be screwed if the read is non atomic.
3815 vmf
->ptl
= pte_lockptr(vma
->vm_mm
, vmf
->pmd
);
3816 spin_lock(vmf
->ptl
);
3817 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
3818 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3823 * Make it present again, Depending on how arch implementes non
3824 * accessible ptes, some can allow access by kernel mode.
3826 old_pte
= ptep_modify_prot_start(vma
, vmf
->address
, vmf
->pte
);
3827 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
3828 pte
= pte_mkyoung(pte
);
3830 pte
= pte_mkwrite(pte
);
3831 ptep_modify_prot_commit(vma
, vmf
->address
, vmf
->pte
, old_pte
, pte
);
3832 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3834 page
= vm_normal_page(vma
, vmf
->address
, pte
);
3836 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3840 /* TODO: handle PTE-mapped THP */
3841 if (PageCompound(page
)) {
3842 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3847 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3848 * much anyway since they can be in shared cache state. This misses
3849 * the case where a mapping is writable but the process never writes
3850 * to it but pte_write gets cleared during protection updates and
3851 * pte_dirty has unpredictable behaviour between PTE scan updates,
3852 * background writeback, dirty balancing and application behaviour.
3854 if (!pte_write(pte
))
3855 flags
|= TNF_NO_GROUP
;
3858 * Flag if the page is shared between multiple address spaces. This
3859 * is later used when determining whether to group tasks together
3861 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
3862 flags
|= TNF_SHARED
;
3864 last_cpupid
= page_cpupid_last(page
);
3865 page_nid
= page_to_nid(page
);
3866 target_nid
= numa_migrate_prep(page
, vma
, vmf
->address
, page_nid
,
3868 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3869 if (target_nid
== NUMA_NO_NODE
) {
3874 /* Migrate to the requested node */
3875 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
3877 page_nid
= target_nid
;
3878 flags
|= TNF_MIGRATED
;
3880 flags
|= TNF_MIGRATE_FAIL
;
3883 if (page_nid
!= NUMA_NO_NODE
)
3884 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
3888 static inline vm_fault_t
create_huge_pmd(struct vm_fault
*vmf
)
3890 if (vma_is_anonymous(vmf
->vma
))
3891 return do_huge_pmd_anonymous_page(vmf
);
3892 if (vmf
->vma
->vm_ops
->huge_fault
)
3893 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
3894 return VM_FAULT_FALLBACK
;
3897 /* `inline' is required to avoid gcc 4.1.2 build error */
3898 static inline vm_fault_t
wp_huge_pmd(struct vm_fault
*vmf
, pmd_t orig_pmd
)
3900 if (vma_is_anonymous(vmf
->vma
))
3901 return do_huge_pmd_wp_page(vmf
, orig_pmd
);
3902 if (vmf
->vma
->vm_ops
->huge_fault
)
3903 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
3905 /* COW handled on pte level: split pmd */
3906 VM_BUG_ON_VMA(vmf
->vma
->vm_flags
& VM_SHARED
, vmf
->vma
);
3907 __split_huge_pmd(vmf
->vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
3909 return VM_FAULT_FALLBACK
;
3912 static inline bool vma_is_accessible(struct vm_area_struct
*vma
)
3914 return vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
);
3917 static vm_fault_t
create_huge_pud(struct vm_fault
*vmf
)
3919 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3920 /* No support for anonymous transparent PUD pages yet */
3921 if (vma_is_anonymous(vmf
->vma
))
3922 return VM_FAULT_FALLBACK
;
3923 if (vmf
->vma
->vm_ops
->huge_fault
)
3924 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
3925 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3926 return VM_FAULT_FALLBACK
;
3929 static vm_fault_t
wp_huge_pud(struct vm_fault
*vmf
, pud_t orig_pud
)
3931 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3932 /* No support for anonymous transparent PUD pages yet */
3933 if (vma_is_anonymous(vmf
->vma
))
3934 return VM_FAULT_FALLBACK
;
3935 if (vmf
->vma
->vm_ops
->huge_fault
)
3936 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
3937 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3938 return VM_FAULT_FALLBACK
;
3942 * These routines also need to handle stuff like marking pages dirty
3943 * and/or accessed for architectures that don't do it in hardware (most
3944 * RISC architectures). The early dirtying is also good on the i386.
3946 * There is also a hook called "update_mmu_cache()" that architectures
3947 * with external mmu caches can use to update those (ie the Sparc or
3948 * PowerPC hashed page tables that act as extended TLBs).
3950 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3951 * concurrent faults).
3953 * The mmap_sem may have been released depending on flags and our return value.
3954 * See filemap_fault() and __lock_page_or_retry().
3956 static vm_fault_t
handle_pte_fault(struct vm_fault
*vmf
)
3960 if (unlikely(pmd_none(*vmf
->pmd
))) {
3962 * Leave __pte_alloc() until later: because vm_ops->fault may
3963 * want to allocate huge page, and if we expose page table
3964 * for an instant, it will be difficult to retract from
3965 * concurrent faults and from rmap lookups.
3969 /* See comment in pte_alloc_one_map() */
3970 if (pmd_devmap_trans_unstable(vmf
->pmd
))
3973 * A regular pmd is established and it can't morph into a huge
3974 * pmd from under us anymore at this point because we hold the
3975 * mmap_sem read mode and khugepaged takes it in write mode.
3976 * So now it's safe to run pte_offset_map().
3978 vmf
->pte
= pte_offset_map(vmf
->pmd
, vmf
->address
);
3979 vmf
->orig_pte
= *vmf
->pte
;
3982 * some architectures can have larger ptes than wordsize,
3983 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3984 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3985 * accesses. The code below just needs a consistent view
3986 * for the ifs and we later double check anyway with the
3987 * ptl lock held. So here a barrier will do.
3990 if (pte_none(vmf
->orig_pte
)) {
3991 pte_unmap(vmf
->pte
);
3997 if (vma_is_anonymous(vmf
->vma
))
3998 return do_anonymous_page(vmf
);
4000 return do_fault(vmf
);
4003 if (!pte_present(vmf
->orig_pte
))
4004 return do_swap_page(vmf
);
4006 if (pte_protnone(vmf
->orig_pte
) && vma_is_accessible(vmf
->vma
))
4007 return do_numa_page(vmf
);
4009 vmf
->ptl
= pte_lockptr(vmf
->vma
->vm_mm
, vmf
->pmd
);
4010 spin_lock(vmf
->ptl
);
4011 entry
= vmf
->orig_pte
;
4012 if (unlikely(!pte_same(*vmf
->pte
, entry
)))
4014 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4015 if (!pte_write(entry
))
4016 return do_wp_page(vmf
);
4017 entry
= pte_mkdirty(entry
);
4019 entry
= pte_mkyoung(entry
);
4020 if (ptep_set_access_flags(vmf
->vma
, vmf
->address
, vmf
->pte
, entry
,
4021 vmf
->flags
& FAULT_FLAG_WRITE
)) {
4022 update_mmu_cache(vmf
->vma
, vmf
->address
, vmf
->pte
);
4025 * This is needed only for protection faults but the arch code
4026 * is not yet telling us if this is a protection fault or not.
4027 * This still avoids useless tlb flushes for .text page faults
4030 if (vmf
->flags
& FAULT_FLAG_WRITE
)
4031 flush_tlb_fix_spurious_fault(vmf
->vma
, vmf
->address
);
4034 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4039 * By the time we get here, we already hold the mm semaphore
4041 * The mmap_sem may have been released depending on flags and our
4042 * return value. See filemap_fault() and __lock_page_or_retry().
4044 static vm_fault_t
__handle_mm_fault(struct vm_area_struct
*vma
,
4045 unsigned long address
, unsigned int flags
)
4047 struct vm_fault vmf
= {
4049 .address
= address
& PAGE_MASK
,
4051 .pgoff
= linear_page_index(vma
, address
),
4052 .gfp_mask
= __get_fault_gfp_mask(vma
),
4054 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
4055 struct mm_struct
*mm
= vma
->vm_mm
;
4060 pgd
= pgd_offset(mm
, address
);
4061 p4d
= p4d_alloc(mm
, pgd
, address
);
4063 return VM_FAULT_OOM
;
4065 vmf
.pud
= pud_alloc(mm
, p4d
, address
);
4067 return VM_FAULT_OOM
;
4068 if (pud_none(*vmf
.pud
) && __transparent_hugepage_enabled(vma
)) {
4069 ret
= create_huge_pud(&vmf
);
4070 if (!(ret
& VM_FAULT_FALLBACK
))
4073 pud_t orig_pud
= *vmf
.pud
;
4076 if (pud_trans_huge(orig_pud
) || pud_devmap(orig_pud
)) {
4078 /* NUMA case for anonymous PUDs would go here */
4080 if (dirty
&& !pud_write(orig_pud
)) {
4081 ret
= wp_huge_pud(&vmf
, orig_pud
);
4082 if (!(ret
& VM_FAULT_FALLBACK
))
4085 huge_pud_set_accessed(&vmf
, orig_pud
);
4091 vmf
.pmd
= pmd_alloc(mm
, vmf
.pud
, address
);
4093 return VM_FAULT_OOM
;
4094 if (pmd_none(*vmf
.pmd
) && __transparent_hugepage_enabled(vma
)) {
4095 ret
= create_huge_pmd(&vmf
);
4096 if (!(ret
& VM_FAULT_FALLBACK
))
4099 pmd_t orig_pmd
= *vmf
.pmd
;
4102 if (unlikely(is_swap_pmd(orig_pmd
))) {
4103 VM_BUG_ON(thp_migration_supported() &&
4104 !is_pmd_migration_entry(orig_pmd
));
4105 if (is_pmd_migration_entry(orig_pmd
))
4106 pmd_migration_entry_wait(mm
, vmf
.pmd
);
4109 if (pmd_trans_huge(orig_pmd
) || pmd_devmap(orig_pmd
)) {
4110 if (pmd_protnone(orig_pmd
) && vma_is_accessible(vma
))
4111 return do_huge_pmd_numa_page(&vmf
, orig_pmd
);
4113 if (dirty
&& !pmd_write(orig_pmd
)) {
4114 ret
= wp_huge_pmd(&vmf
, orig_pmd
);
4115 if (!(ret
& VM_FAULT_FALLBACK
))
4118 huge_pmd_set_accessed(&vmf
, orig_pmd
);
4124 return handle_pte_fault(&vmf
);
4128 * By the time we get here, we already hold the mm semaphore
4130 * The mmap_sem may have been released depending on flags and our
4131 * return value. See filemap_fault() and __lock_page_or_retry().
4133 vm_fault_t
handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
4138 __set_current_state(TASK_RUNNING
);
4140 count_vm_event(PGFAULT
);
4141 count_memcg_event_mm(vma
->vm_mm
, PGFAULT
);
4143 /* do counter updates before entering really critical section. */
4144 check_sync_rss_stat(current
);
4146 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
4147 flags
& FAULT_FLAG_INSTRUCTION
,
4148 flags
& FAULT_FLAG_REMOTE
))
4149 return VM_FAULT_SIGSEGV
;
4152 * Enable the memcg OOM handling for faults triggered in user
4153 * space. Kernel faults are handled more gracefully.
4155 if (flags
& FAULT_FLAG_USER
)
4156 mem_cgroup_enter_user_fault();
4158 if (unlikely(is_vm_hugetlb_page(vma
)))
4159 ret
= hugetlb_fault(vma
->vm_mm
, vma
, address
, flags
);
4161 ret
= __handle_mm_fault(vma
, address
, flags
);
4163 if (flags
& FAULT_FLAG_USER
) {
4164 mem_cgroup_exit_user_fault();
4166 * The task may have entered a memcg OOM situation but
4167 * if the allocation error was handled gracefully (no
4168 * VM_FAULT_OOM), there is no need to kill anything.
4169 * Just clean up the OOM state peacefully.
4171 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
4172 mem_cgroup_oom_synchronize(false);
4177 EXPORT_SYMBOL_GPL(handle_mm_fault
);
4179 #ifndef __PAGETABLE_P4D_FOLDED
4181 * Allocate p4d page table.
4182 * We've already handled the fast-path in-line.
4184 int __p4d_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
4186 p4d_t
*new = p4d_alloc_one(mm
, address
);
4190 smp_wmb(); /* See comment in __pte_alloc */
4192 spin_lock(&mm
->page_table_lock
);
4193 if (pgd_present(*pgd
)) /* Another has populated it */
4196 pgd_populate(mm
, pgd
, new);
4197 spin_unlock(&mm
->page_table_lock
);
4200 #endif /* __PAGETABLE_P4D_FOLDED */
4202 #ifndef __PAGETABLE_PUD_FOLDED
4204 * Allocate page upper directory.
4205 * We've already handled the fast-path in-line.
4207 int __pud_alloc(struct mm_struct
*mm
, p4d_t
*p4d
, unsigned long address
)
4209 pud_t
*new = pud_alloc_one(mm
, address
);
4213 smp_wmb(); /* See comment in __pte_alloc */
4215 spin_lock(&mm
->page_table_lock
);
4216 #ifndef __ARCH_HAS_5LEVEL_HACK
4217 if (!p4d_present(*p4d
)) {
4219 p4d_populate(mm
, p4d
, new);
4220 } else /* Another has populated it */
4223 if (!pgd_present(*p4d
)) {
4225 pgd_populate(mm
, p4d
, new);
4226 } else /* Another has populated it */
4228 #endif /* __ARCH_HAS_5LEVEL_HACK */
4229 spin_unlock(&mm
->page_table_lock
);
4232 #endif /* __PAGETABLE_PUD_FOLDED */
4234 #ifndef __PAGETABLE_PMD_FOLDED
4236 * Allocate page middle directory.
4237 * We've already handled the fast-path in-line.
4239 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
4242 pmd_t
*new = pmd_alloc_one(mm
, address
);
4246 smp_wmb(); /* See comment in __pte_alloc */
4248 ptl
= pud_lock(mm
, pud
);
4249 #ifndef __ARCH_HAS_4LEVEL_HACK
4250 if (!pud_present(*pud
)) {
4252 pud_populate(mm
, pud
, new);
4253 } else /* Another has populated it */
4256 if (!pgd_present(*pud
)) {
4258 pgd_populate(mm
, pud
, new);
4259 } else /* Another has populated it */
4261 #endif /* __ARCH_HAS_4LEVEL_HACK */
4265 #endif /* __PAGETABLE_PMD_FOLDED */
4267 int follow_invalidate_pte(struct mm_struct
*mm
, unsigned long address
,
4268 struct mmu_notifier_range
*range
, pte_t
**ptepp
,
4269 pmd_t
**pmdpp
, spinlock_t
**ptlp
)
4277 pgd
= pgd_offset(mm
, address
);
4278 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
4281 p4d
= p4d_offset(pgd
, address
);
4282 if (p4d_none(*p4d
) || unlikely(p4d_bad(*p4d
)))
4285 pud
= pud_offset(p4d
, address
);
4286 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
4289 pmd
= pmd_offset(pud
, address
);
4290 VM_BUG_ON(pmd_trans_huge(*pmd
));
4292 if (pmd_huge(*pmd
)) {
4297 mmu_notifier_range_init(range
, MMU_NOTIFY_CLEAR
, 0,
4298 NULL
, mm
, address
& PMD_MASK
,
4299 (address
& PMD_MASK
) + PMD_SIZE
);
4300 mmu_notifier_invalidate_range_start(range
);
4302 *ptlp
= pmd_lock(mm
, pmd
);
4303 if (pmd_huge(*pmd
)) {
4309 mmu_notifier_invalidate_range_end(range
);
4312 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
4316 mmu_notifier_range_init(range
, MMU_NOTIFY_CLEAR
, 0, NULL
, mm
,
4317 address
& PAGE_MASK
,
4318 (address
& PAGE_MASK
) + PAGE_SIZE
);
4319 mmu_notifier_invalidate_range_start(range
);
4321 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
4322 if (!pte_present(*ptep
))
4327 pte_unmap_unlock(ptep
, *ptlp
);
4329 mmu_notifier_invalidate_range_end(range
);
4335 * follow_pte - look up PTE at a user virtual address
4336 * @mm: the mm_struct of the target address space
4337 * @address: user virtual address
4338 * @ptepp: location to store found PTE
4339 * @ptlp: location to store the lock for the PTE
4341 * On a successful return, the pointer to the PTE is stored in @ptepp;
4342 * the corresponding lock is taken and its location is stored in @ptlp.
4343 * The contents of the PTE are only stable until @ptlp is released;
4344 * any further use, if any, must be protected against invalidation
4345 * with MMU notifiers.
4347 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
4348 * should be taken for read.
4350 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
4351 * it is not a good general-purpose API.
4353 * Return: zero on success, -ve otherwise.
4355 int follow_pte(struct mm_struct
*mm
, unsigned long address
,
4356 pte_t
**ptepp
, spinlock_t
**ptlp
)
4358 return follow_invalidate_pte(mm
, address
, NULL
, ptepp
, NULL
, ptlp
);
4360 EXPORT_SYMBOL_GPL(follow_pte
);
4363 * follow_pfn - look up PFN at a user virtual address
4364 * @vma: memory mapping
4365 * @address: user virtual address
4366 * @pfn: location to store found PFN
4368 * Only IO mappings and raw PFN mappings are allowed.
4370 * This function does not allow the caller to read the permissions
4371 * of the PTE. Do not use it.
4373 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4375 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
4382 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4385 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
4388 *pfn
= pte_pfn(*ptep
);
4389 pte_unmap_unlock(ptep
, ptl
);
4392 EXPORT_SYMBOL(follow_pfn
);
4394 #ifdef CONFIG_HAVE_IOREMAP_PROT
4395 int follow_phys(struct vm_area_struct
*vma
,
4396 unsigned long address
, unsigned int flags
,
4397 unsigned long *prot
, resource_size_t
*phys
)
4403 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4406 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
4410 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
4413 *prot
= pgprot_val(pte_pgprot(pte
));
4414 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
4418 pte_unmap_unlock(ptep
, ptl
);
4423 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
4424 void *buf
, int len
, int write
)
4426 resource_size_t phys_addr
;
4427 unsigned long prot
= 0;
4428 void __iomem
*maddr
;
4429 int offset
= addr
& (PAGE_SIZE
-1);
4431 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
4434 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
4439 memcpy_toio(maddr
+ offset
, buf
, len
);
4441 memcpy_fromio(buf
, maddr
+ offset
, len
);
4446 EXPORT_SYMBOL_GPL(generic_access_phys
);
4450 * Access another process' address space as given in mm. If non-NULL, use the
4451 * given task for page fault accounting.
4453 int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
4454 unsigned long addr
, void *buf
, int len
, unsigned int gup_flags
)
4456 struct vm_area_struct
*vma
;
4457 void *old_buf
= buf
;
4458 int write
= gup_flags
& FOLL_WRITE
;
4460 if (down_read_killable(&mm
->mmap_sem
))
4463 /* ignore errors, just check how much was successfully transferred */
4465 int bytes
, ret
, offset
;
4467 struct page
*page
= NULL
;
4469 ret
= get_user_pages_remote(tsk
, mm
, addr
, 1,
4470 gup_flags
, &page
, &vma
, NULL
);
4472 #ifndef CONFIG_HAVE_IOREMAP_PROT
4476 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4477 * we can access using slightly different code.
4479 vma
= find_vma(mm
, addr
);
4480 if (!vma
|| vma
->vm_start
> addr
)
4482 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
4483 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
4491 offset
= addr
& (PAGE_SIZE
-1);
4492 if (bytes
> PAGE_SIZE
-offset
)
4493 bytes
= PAGE_SIZE
-offset
;
4497 copy_to_user_page(vma
, page
, addr
,
4498 maddr
+ offset
, buf
, bytes
);
4499 set_page_dirty_lock(page
);
4501 copy_from_user_page(vma
, page
, addr
,
4502 buf
, maddr
+ offset
, bytes
);
4511 up_read(&mm
->mmap_sem
);
4513 return buf
- old_buf
;
4517 * access_remote_vm - access another process' address space
4518 * @mm: the mm_struct of the target address space
4519 * @addr: start address to access
4520 * @buf: source or destination buffer
4521 * @len: number of bytes to transfer
4522 * @gup_flags: flags modifying lookup behaviour
4524 * The caller must hold a reference on @mm.
4526 * Return: number of bytes copied from source to destination.
4528 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
4529 void *buf
, int len
, unsigned int gup_flags
)
4531 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, gup_flags
);
4535 * Access another process' address space.
4536 * Source/target buffer must be kernel space,
4537 * Do not walk the page table directly, use get_user_pages
4539 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
4540 void *buf
, int len
, unsigned int gup_flags
)
4542 struct mm_struct
*mm
;
4545 mm
= get_task_mm(tsk
);
4549 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, gup_flags
);
4555 EXPORT_SYMBOL_GPL(access_process_vm
);
4558 * Print the name of a VMA.
4560 void print_vma_addr(char *prefix
, unsigned long ip
)
4562 struct mm_struct
*mm
= current
->mm
;
4563 struct vm_area_struct
*vma
;
4566 * we might be running from an atomic context so we cannot sleep
4568 if (!down_read_trylock(&mm
->mmap_sem
))
4571 vma
= find_vma(mm
, ip
);
4572 if (vma
&& vma
->vm_file
) {
4573 struct file
*f
= vma
->vm_file
;
4574 char *buf
= (char *)__get_free_page(GFP_NOWAIT
);
4578 p
= file_path(f
, buf
, PAGE_SIZE
);
4581 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
4583 vma
->vm_end
- vma
->vm_start
);
4584 free_page((unsigned long)buf
);
4587 up_read(&mm
->mmap_sem
);
4590 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4591 void __might_fault(const char *file
, int line
)
4594 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4595 * holding the mmap_sem, this is safe because kernel memory doesn't
4596 * get paged out, therefore we'll never actually fault, and the
4597 * below annotations will generate false positives.
4599 if (uaccess_kernel())
4601 if (pagefault_disabled())
4603 __might_sleep(file
, line
, 0);
4604 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4606 might_lock_read(¤t
->mm
->mmap_sem
);
4609 EXPORT_SYMBOL(__might_fault
);
4612 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4614 * Process all subpages of the specified huge page with the specified
4615 * operation. The target subpage will be processed last to keep its
4618 static inline void process_huge_page(
4619 unsigned long addr_hint
, unsigned int pages_per_huge_page
,
4620 void (*process_subpage
)(unsigned long addr
, int idx
, void *arg
),
4624 unsigned long addr
= addr_hint
&
4625 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
4627 /* Process target subpage last to keep its cache lines hot */
4629 n
= (addr_hint
- addr
) / PAGE_SIZE
;
4630 if (2 * n
<= pages_per_huge_page
) {
4631 /* If target subpage in first half of huge page */
4634 /* Process subpages at the end of huge page */
4635 for (i
= pages_per_huge_page
- 1; i
>= 2 * n
; i
--) {
4637 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
4640 /* If target subpage in second half of huge page */
4641 base
= pages_per_huge_page
- 2 * (pages_per_huge_page
- n
);
4642 l
= pages_per_huge_page
- n
;
4643 /* Process subpages at the begin of huge page */
4644 for (i
= 0; i
< base
; i
++) {
4646 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
4650 * Process remaining subpages in left-right-left-right pattern
4651 * towards the target subpage
4653 for (i
= 0; i
< l
; i
++) {
4654 int left_idx
= base
+ i
;
4655 int right_idx
= base
+ 2 * l
- 1 - i
;
4658 process_subpage(addr
+ left_idx
* PAGE_SIZE
, left_idx
, arg
);
4660 process_subpage(addr
+ right_idx
* PAGE_SIZE
, right_idx
, arg
);
4664 static void clear_gigantic_page(struct page
*page
,
4666 unsigned int pages_per_huge_page
)
4669 struct page
*p
= page
;
4672 for (i
= 0; i
< pages_per_huge_page
;
4673 i
++, p
= mem_map_next(p
, page
, i
)) {
4675 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
4679 static void clear_subpage(unsigned long addr
, int idx
, void *arg
)
4681 struct page
*page
= arg
;
4683 clear_user_highpage(page
+ idx
, addr
);
4686 void clear_huge_page(struct page
*page
,
4687 unsigned long addr_hint
, unsigned int pages_per_huge_page
)
4689 unsigned long addr
= addr_hint
&
4690 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
4692 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4693 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
4697 process_huge_page(addr_hint
, pages_per_huge_page
, clear_subpage
, page
);
4700 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
4702 struct vm_area_struct
*vma
,
4703 unsigned int pages_per_huge_page
)
4706 struct page
*dst_base
= dst
;
4707 struct page
*src_base
= src
;
4709 for (i
= 0; i
< pages_per_huge_page
; ) {
4711 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
4714 dst
= mem_map_next(dst
, dst_base
, i
);
4715 src
= mem_map_next(src
, src_base
, i
);
4719 struct copy_subpage_arg
{
4722 struct vm_area_struct
*vma
;
4725 static void copy_subpage(unsigned long addr
, int idx
, void *arg
)
4727 struct copy_subpage_arg
*copy_arg
= arg
;
4729 copy_user_highpage(copy_arg
->dst
+ idx
, copy_arg
->src
+ idx
,
4730 addr
, copy_arg
->vma
);
4733 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
4734 unsigned long addr_hint
, struct vm_area_struct
*vma
,
4735 unsigned int pages_per_huge_page
)
4737 unsigned long addr
= addr_hint
&
4738 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
4739 struct copy_subpage_arg arg
= {
4745 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4746 copy_user_gigantic_page(dst
, src
, addr
, vma
,
4747 pages_per_huge_page
);
4751 process_huge_page(addr_hint
, pages_per_huge_page
, copy_subpage
, &arg
);
4754 long copy_huge_page_from_user(struct page
*dst_page
,
4755 const void __user
*usr_src
,
4756 unsigned int pages_per_huge_page
,
4757 bool allow_pagefault
)
4759 void *src
= (void *)usr_src
;
4761 unsigned long i
, rc
= 0;
4762 unsigned long ret_val
= pages_per_huge_page
* PAGE_SIZE
;
4763 struct page
*subpage
= dst_page
;
4765 for (i
= 0; i
< pages_per_huge_page
;
4766 i
++, subpage
= mem_map_next(subpage
, dst_page
, i
)) {
4767 if (allow_pagefault
)
4768 page_kaddr
= kmap(subpage
);
4770 page_kaddr
= kmap_atomic(subpage
);
4771 rc
= copy_from_user(page_kaddr
,
4772 (const void __user
*)(src
+ i
* PAGE_SIZE
),
4774 if (allow_pagefault
)
4777 kunmap_atomic(page_kaddr
);
4779 ret_val
-= (PAGE_SIZE
- rc
);
4787 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4789 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4791 static struct kmem_cache
*page_ptl_cachep
;
4793 void __init
ptlock_cache_init(void)
4795 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
4799 bool ptlock_alloc(struct page
*page
)
4803 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
4810 void ptlock_free(struct page
*page
)
4812 kmem_cache_free(page_ptl_cachep
, page
->ptl
);