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/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
72 #include <linux/numa.h>
73 #include <linux/perf_event.h>
74 #include <linux/ptrace.h>
75 #include <linux/vmalloc.h>
77 #include <trace/events/kmem.h>
80 #include <asm/mmu_context.h>
81 #include <asm/pgalloc.h>
82 #include <linux/uaccess.h>
84 #include <asm/tlbflush.h>
86 #include "pgalloc-track.h"
89 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
90 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
94 unsigned long max_mapnr
;
95 EXPORT_SYMBOL(max_mapnr
);
98 EXPORT_SYMBOL(mem_map
);
102 * A number of key systems in x86 including ioremap() rely on the assumption
103 * that high_memory defines the upper bound on direct map memory, then end
104 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
105 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
109 EXPORT_SYMBOL(high_memory
);
112 * Randomize the address space (stacks, mmaps, brk, etc.).
114 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
115 * as ancient (libc5 based) binaries can segfault. )
117 int randomize_va_space __read_mostly
=
118 #ifdef CONFIG_COMPAT_BRK
124 #ifndef arch_faults_on_old_pte
125 static inline bool arch_faults_on_old_pte(void)
128 * Those arches which don't have hw access flag feature need to
129 * implement their own helper. By default, "true" means pagefault
130 * will be hit on old pte.
136 #ifndef arch_wants_old_prefaulted_pte
137 static inline bool arch_wants_old_prefaulted_pte(void)
140 * Transitioning a PTE from 'old' to 'young' can be expensive on
141 * some architectures, even if it's performed in hardware. By
142 * default, "false" means prefaulted entries will be 'young'.
148 static int __init
disable_randmaps(char *s
)
150 randomize_va_space
= 0;
153 __setup("norandmaps", disable_randmaps
);
155 unsigned long zero_pfn __read_mostly
;
156 EXPORT_SYMBOL(zero_pfn
);
158 unsigned long highest_memmap_pfn __read_mostly
;
161 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
163 static int __init
init_zero_pfn(void)
165 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
168 early_initcall(init_zero_pfn
);
170 void mm_trace_rss_stat(struct mm_struct
*mm
, int member
, long count
)
172 trace_rss_stat(mm
, member
, count
);
175 #if defined(SPLIT_RSS_COUNTING)
177 void sync_mm_rss(struct mm_struct
*mm
)
181 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
182 if (current
->rss_stat
.count
[i
]) {
183 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
184 current
->rss_stat
.count
[i
] = 0;
187 current
->rss_stat
.events
= 0;
190 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
192 struct task_struct
*task
= current
;
194 if (likely(task
->mm
== mm
))
195 task
->rss_stat
.count
[member
] += val
;
197 add_mm_counter(mm
, member
, val
);
199 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
200 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
202 /* sync counter once per 64 page faults */
203 #define TASK_RSS_EVENTS_THRESH (64)
204 static void check_sync_rss_stat(struct task_struct
*task
)
206 if (unlikely(task
!= current
))
208 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
209 sync_mm_rss(task
->mm
);
211 #else /* SPLIT_RSS_COUNTING */
213 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
214 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
216 static void check_sync_rss_stat(struct task_struct
*task
)
220 #endif /* SPLIT_RSS_COUNTING */
223 * Note: this doesn't free the actual pages themselves. That
224 * has been handled earlier when unmapping all the memory regions.
226 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
229 pgtable_t token
= pmd_pgtable(*pmd
);
231 pte_free_tlb(tlb
, token
, addr
);
232 mm_dec_nr_ptes(tlb
->mm
);
235 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
236 unsigned long addr
, unsigned long end
,
237 unsigned long floor
, unsigned long ceiling
)
244 pmd
= pmd_offset(pud
, addr
);
246 next
= pmd_addr_end(addr
, end
);
247 if (pmd_none_or_clear_bad(pmd
))
249 free_pte_range(tlb
, pmd
, addr
);
250 } while (pmd
++, addr
= next
, addr
!= end
);
260 if (end
- 1 > ceiling
- 1)
263 pmd
= pmd_offset(pud
, start
);
265 pmd_free_tlb(tlb
, pmd
, start
);
266 mm_dec_nr_pmds(tlb
->mm
);
269 static inline void free_pud_range(struct mmu_gather
*tlb
, p4d_t
*p4d
,
270 unsigned long addr
, unsigned long end
,
271 unsigned long floor
, unsigned long ceiling
)
278 pud
= pud_offset(p4d
, addr
);
280 next
= pud_addr_end(addr
, end
);
281 if (pud_none_or_clear_bad(pud
))
283 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
284 } while (pud
++, addr
= next
, addr
!= end
);
294 if (end
- 1 > ceiling
- 1)
297 pud
= pud_offset(p4d
, start
);
299 pud_free_tlb(tlb
, pud
, start
);
300 mm_dec_nr_puds(tlb
->mm
);
303 static inline void free_p4d_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
304 unsigned long addr
, unsigned long end
,
305 unsigned long floor
, unsigned long ceiling
)
312 p4d
= p4d_offset(pgd
, addr
);
314 next
= p4d_addr_end(addr
, end
);
315 if (p4d_none_or_clear_bad(p4d
))
317 free_pud_range(tlb
, p4d
, addr
, next
, floor
, ceiling
);
318 } while (p4d
++, addr
= next
, addr
!= end
);
324 ceiling
&= PGDIR_MASK
;
328 if (end
- 1 > ceiling
- 1)
331 p4d
= p4d_offset(pgd
, start
);
333 p4d_free_tlb(tlb
, p4d
, start
);
337 * This function frees user-level page tables of a process.
339 void free_pgd_range(struct mmu_gather
*tlb
,
340 unsigned long addr
, unsigned long end
,
341 unsigned long floor
, unsigned long ceiling
)
347 * The next few lines have given us lots of grief...
349 * Why are we testing PMD* at this top level? Because often
350 * there will be no work to do at all, and we'd prefer not to
351 * go all the way down to the bottom just to discover that.
353 * Why all these "- 1"s? Because 0 represents both the bottom
354 * of the address space and the top of it (using -1 for the
355 * top wouldn't help much: the masks would do the wrong thing).
356 * The rule is that addr 0 and floor 0 refer to the bottom of
357 * the address space, but end 0 and ceiling 0 refer to the top
358 * Comparisons need to use "end - 1" and "ceiling - 1" (though
359 * that end 0 case should be mythical).
361 * Wherever addr is brought up or ceiling brought down, we must
362 * be careful to reject "the opposite 0" before it confuses the
363 * subsequent tests. But what about where end is brought down
364 * by PMD_SIZE below? no, end can't go down to 0 there.
366 * Whereas we round start (addr) and ceiling down, by different
367 * masks at different levels, in order to test whether a table
368 * now has no other vmas using it, so can be freed, we don't
369 * bother to round floor or end up - the tests don't need that.
383 if (end
- 1 > ceiling
- 1)
388 * We add page table cache pages with PAGE_SIZE,
389 * (see pte_free_tlb()), flush the tlb if we need
391 tlb_change_page_size(tlb
, PAGE_SIZE
);
392 pgd
= pgd_offset(tlb
->mm
, addr
);
394 next
= pgd_addr_end(addr
, end
);
395 if (pgd_none_or_clear_bad(pgd
))
397 free_p4d_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
398 } while (pgd
++, addr
= next
, addr
!= end
);
401 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
402 unsigned long floor
, unsigned long ceiling
)
405 struct vm_area_struct
*next
= vma
->vm_next
;
406 unsigned long addr
= vma
->vm_start
;
409 * Hide vma from rmap and truncate_pagecache before freeing
412 unlink_anon_vmas(vma
);
413 unlink_file_vma(vma
);
415 if (is_vm_hugetlb_page(vma
)) {
416 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
417 floor
, next
? next
->vm_start
: ceiling
);
420 * Optimization: gather nearby vmas into one call down
422 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
423 && !is_vm_hugetlb_page(next
)) {
426 unlink_anon_vmas(vma
);
427 unlink_file_vma(vma
);
429 free_pgd_range(tlb
, addr
, vma
->vm_end
,
430 floor
, next
? next
->vm_start
: ceiling
);
436 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
)
439 pgtable_t
new = pte_alloc_one(mm
);
444 * Ensure all pte setup (eg. pte page lock and page clearing) are
445 * visible before the pte is made visible to other CPUs by being
446 * put into page tables.
448 * The other side of the story is the pointer chasing in the page
449 * table walking code (when walking the page table without locking;
450 * ie. most of the time). Fortunately, these data accesses consist
451 * of a chain of data-dependent loads, meaning most CPUs (alpha
452 * being the notable exception) will already guarantee loads are
453 * seen in-order. See the alpha page table accessors for the
454 * smp_rmb() barriers in page table walking code.
456 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
458 ptl
= pmd_lock(mm
, pmd
);
459 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
461 pmd_populate(mm
, pmd
, new);
470 int __pte_alloc_kernel(pmd_t
*pmd
)
472 pte_t
*new = pte_alloc_one_kernel(&init_mm
);
476 smp_wmb(); /* See comment in __pte_alloc */
478 spin_lock(&init_mm
.page_table_lock
);
479 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
480 pmd_populate_kernel(&init_mm
, pmd
, new);
483 spin_unlock(&init_mm
.page_table_lock
);
485 pte_free_kernel(&init_mm
, new);
489 static inline void init_rss_vec(int *rss
)
491 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
494 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
498 if (current
->mm
== mm
)
500 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
502 add_mm_counter(mm
, i
, rss
[i
]);
506 * This function is called to print an error when a bad pte
507 * is found. For example, we might have a PFN-mapped pte in
508 * a region that doesn't allow it.
510 * The calling function must still handle the error.
512 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
513 pte_t pte
, struct page
*page
)
515 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
516 p4d_t
*p4d
= p4d_offset(pgd
, addr
);
517 pud_t
*pud
= pud_offset(p4d
, addr
);
518 pmd_t
*pmd
= pmd_offset(pud
, addr
);
519 struct address_space
*mapping
;
521 static unsigned long resume
;
522 static unsigned long nr_shown
;
523 static unsigned long nr_unshown
;
526 * Allow a burst of 60 reports, then keep quiet for that minute;
527 * or allow a steady drip of one report per second.
529 if (nr_shown
== 60) {
530 if (time_before(jiffies
, resume
)) {
535 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
542 resume
= jiffies
+ 60 * HZ
;
544 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
545 index
= linear_page_index(vma
, addr
);
547 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
549 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
551 dump_page(page
, "bad pte");
552 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
553 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
554 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
556 vma
->vm_ops
? vma
->vm_ops
->fault
: NULL
,
557 vma
->vm_file
? vma
->vm_file
->f_op
->mmap
: NULL
,
558 mapping
? mapping
->a_ops
->readpage
: NULL
);
560 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
564 * vm_normal_page -- This function gets the "struct page" associated with a pte.
566 * "Special" mappings do not wish to be associated with a "struct page" (either
567 * it doesn't exist, or it exists but they don't want to touch it). In this
568 * case, NULL is returned here. "Normal" mappings do have a struct page.
570 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
571 * pte bit, in which case this function is trivial. Secondly, an architecture
572 * may not have a spare pte bit, which requires a more complicated scheme,
575 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
576 * special mapping (even if there are underlying and valid "struct pages").
577 * COWed pages of a VM_PFNMAP are always normal.
579 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
580 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
581 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
582 * mapping will always honor the rule
584 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
586 * And for normal mappings this is false.
588 * This restricts such mappings to be a linear translation from virtual address
589 * to pfn. To get around this restriction, we allow arbitrary mappings so long
590 * as the vma is not a COW mapping; in that case, we know that all ptes are
591 * special (because none can have been COWed).
594 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
596 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
597 * page" backing, however the difference is that _all_ pages with a struct
598 * page (that is, those where pfn_valid is true) are refcounted and considered
599 * normal pages by the VM. The disadvantage is that pages are refcounted
600 * (which can be slower and simply not an option for some PFNMAP users). The
601 * advantage is that we don't have to follow the strict linearity rule of
602 * PFNMAP mappings in order to support COWable mappings.
605 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
608 unsigned long pfn
= pte_pfn(pte
);
610 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
)) {
611 if (likely(!pte_special(pte
)))
613 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
614 return vma
->vm_ops
->find_special_page(vma
, addr
);
615 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
617 if (is_zero_pfn(pfn
))
622 print_bad_pte(vma
, addr
, pte
, NULL
);
626 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
628 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
629 if (vma
->vm_flags
& VM_MIXEDMAP
) {
635 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
636 if (pfn
== vma
->vm_pgoff
+ off
)
638 if (!is_cow_mapping(vma
->vm_flags
))
643 if (is_zero_pfn(pfn
))
647 if (unlikely(pfn
> highest_memmap_pfn
)) {
648 print_bad_pte(vma
, addr
, pte
, NULL
);
653 * NOTE! We still have PageReserved() pages in the page tables.
654 * eg. VDSO mappings can cause them to exist.
657 return pfn_to_page(pfn
);
660 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
661 struct page
*vm_normal_page_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
664 unsigned long pfn
= pmd_pfn(pmd
);
667 * There is no pmd_special() but there may be special pmds, e.g.
668 * in a direct-access (dax) mapping, so let's just replicate the
669 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
671 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
672 if (vma
->vm_flags
& VM_MIXEDMAP
) {
678 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
679 if (pfn
== vma
->vm_pgoff
+ off
)
681 if (!is_cow_mapping(vma
->vm_flags
))
688 if (is_huge_zero_pmd(pmd
))
690 if (unlikely(pfn
> highest_memmap_pfn
))
694 * NOTE! We still have PageReserved() pages in the page tables.
695 * eg. VDSO mappings can cause them to exist.
698 return pfn_to_page(pfn
);
702 static void restore_exclusive_pte(struct vm_area_struct
*vma
,
703 struct page
*page
, unsigned long address
,
709 pte
= pte_mkold(mk_pte(page
, READ_ONCE(vma
->vm_page_prot
)));
710 if (pte_swp_soft_dirty(*ptep
))
711 pte
= pte_mksoft_dirty(pte
);
713 entry
= pte_to_swp_entry(*ptep
);
714 if (pte_swp_uffd_wp(*ptep
))
715 pte
= pte_mkuffd_wp(pte
);
716 else if (is_writable_device_exclusive_entry(entry
))
717 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
719 set_pte_at(vma
->vm_mm
, address
, ptep
, pte
);
722 * No need to take a page reference as one was already
723 * created when the swap entry was made.
726 page_add_anon_rmap(page
, vma
, address
, false);
729 * Currently device exclusive access only supports anonymous
730 * memory so the entry shouldn't point to a filebacked page.
732 WARN_ON_ONCE(!PageAnon(page
));
734 if (vma
->vm_flags
& VM_LOCKED
)
735 mlock_vma_page(page
);
738 * No need to invalidate - it was non-present before. However
739 * secondary CPUs may have mappings that need invalidating.
741 update_mmu_cache(vma
, address
, ptep
);
745 * Tries to restore an exclusive pte if the page lock can be acquired without
749 try_restore_exclusive_pte(pte_t
*src_pte
, struct vm_area_struct
*vma
,
752 swp_entry_t entry
= pte_to_swp_entry(*src_pte
);
753 struct page
*page
= pfn_swap_entry_to_page(entry
);
755 if (trylock_page(page
)) {
756 restore_exclusive_pte(vma
, page
, addr
, src_pte
);
765 * copy one vm_area from one task to the other. Assumes the page tables
766 * already present in the new task to be cleared in the whole range
767 * covered by this vma.
771 copy_nonpresent_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
772 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*dst_vma
,
773 struct vm_area_struct
*src_vma
, unsigned long addr
, int *rss
)
775 unsigned long vm_flags
= dst_vma
->vm_flags
;
776 pte_t pte
= *src_pte
;
778 swp_entry_t entry
= pte_to_swp_entry(pte
);
780 if (likely(!non_swap_entry(entry
))) {
781 if (swap_duplicate(entry
) < 0)
784 /* make sure dst_mm is on swapoff's mmlist. */
785 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
786 spin_lock(&mmlist_lock
);
787 if (list_empty(&dst_mm
->mmlist
))
788 list_add(&dst_mm
->mmlist
,
790 spin_unlock(&mmlist_lock
);
793 } else if (is_migration_entry(entry
)) {
794 page
= pfn_swap_entry_to_page(entry
);
796 rss
[mm_counter(page
)]++;
798 if (is_writable_migration_entry(entry
) &&
799 is_cow_mapping(vm_flags
)) {
801 * COW mappings require pages in both
802 * parent and child to be set to read.
804 entry
= make_readable_migration_entry(
806 pte
= swp_entry_to_pte(entry
);
807 if (pte_swp_soft_dirty(*src_pte
))
808 pte
= pte_swp_mksoft_dirty(pte
);
809 if (pte_swp_uffd_wp(*src_pte
))
810 pte
= pte_swp_mkuffd_wp(pte
);
811 set_pte_at(src_mm
, addr
, src_pte
, pte
);
813 } else if (is_device_private_entry(entry
)) {
814 page
= pfn_swap_entry_to_page(entry
);
817 * Update rss count even for unaddressable pages, as
818 * they should treated just like normal pages in this
821 * We will likely want to have some new rss counters
822 * for unaddressable pages, at some point. But for now
823 * keep things as they are.
826 rss
[mm_counter(page
)]++;
827 page_dup_rmap(page
, false);
830 * We do not preserve soft-dirty information, because so
831 * far, checkpoint/restore is the only feature that
832 * requires that. And checkpoint/restore does not work
833 * when a device driver is involved (you cannot easily
834 * save and restore device driver state).
836 if (is_writable_device_private_entry(entry
) &&
837 is_cow_mapping(vm_flags
)) {
838 entry
= make_readable_device_private_entry(
840 pte
= swp_entry_to_pte(entry
);
841 if (pte_swp_uffd_wp(*src_pte
))
842 pte
= pte_swp_mkuffd_wp(pte
);
843 set_pte_at(src_mm
, addr
, src_pte
, pte
);
845 } else if (is_device_exclusive_entry(entry
)) {
847 * Make device exclusive entries present by restoring the
848 * original entry then copying as for a present pte. Device
849 * exclusive entries currently only support private writable
850 * (ie. COW) mappings.
852 VM_BUG_ON(!is_cow_mapping(src_vma
->vm_flags
));
853 if (try_restore_exclusive_pte(src_pte
, src_vma
, addr
))
857 if (!userfaultfd_wp(dst_vma
))
858 pte
= pte_swp_clear_uffd_wp(pte
);
859 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
864 * Copy a present and normal page if necessary.
866 * NOTE! The usual case is that this doesn't need to do
867 * anything, and can just return a positive value. That
868 * will let the caller know that it can just increase
869 * the page refcount and re-use the pte the traditional
872 * But _if_ we need to copy it because it needs to be
873 * pinned in the parent (and the child should get its own
874 * copy rather than just a reference to the same page),
875 * we'll do that here and return zero to let the caller
878 * And if we need a pre-allocated page but don't yet have
879 * one, return a negative error to let the preallocation
880 * code know so that it can do so outside the page table
884 copy_present_page(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
885 pte_t
*dst_pte
, pte_t
*src_pte
, unsigned long addr
, int *rss
,
886 struct page
**prealloc
, pte_t pte
, struct page
*page
)
888 struct page
*new_page
;
891 * What we want to do is to check whether this page may
892 * have been pinned by the parent process. If so,
893 * instead of wrprotect the pte on both sides, we copy
894 * the page immediately so that we'll always guarantee
895 * the pinned page won't be randomly replaced in the
898 * The page pinning checks are just "has this mm ever
899 * seen pinning", along with the (inexact) check of
900 * the page count. That might give false positives for
901 * for pinning, but it will work correctly.
903 if (likely(!page_needs_cow_for_dma(src_vma
, page
)))
906 new_page
= *prealloc
;
911 * We have a prealloc page, all good! Take it
912 * over and copy the page & arm it.
915 copy_user_highpage(new_page
, page
, addr
, src_vma
);
916 __SetPageUptodate(new_page
);
917 page_add_new_anon_rmap(new_page
, dst_vma
, addr
, false);
918 lru_cache_add_inactive_or_unevictable(new_page
, dst_vma
);
919 rss
[mm_counter(new_page
)]++;
921 /* All done, just insert the new page copy in the child */
922 pte
= mk_pte(new_page
, dst_vma
->vm_page_prot
);
923 pte
= maybe_mkwrite(pte_mkdirty(pte
), dst_vma
);
924 if (userfaultfd_pte_wp(dst_vma
, *src_pte
))
925 /* Uffd-wp needs to be delivered to dest pte as well */
926 pte
= pte_wrprotect(pte_mkuffd_wp(pte
));
927 set_pte_at(dst_vma
->vm_mm
, addr
, dst_pte
, pte
);
932 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
933 * is required to copy this pte.
936 copy_present_pte(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
937 pte_t
*dst_pte
, pte_t
*src_pte
, unsigned long addr
, int *rss
,
938 struct page
**prealloc
)
940 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
941 unsigned long vm_flags
= src_vma
->vm_flags
;
942 pte_t pte
= *src_pte
;
945 page
= vm_normal_page(src_vma
, addr
, pte
);
949 retval
= copy_present_page(dst_vma
, src_vma
, dst_pte
, src_pte
,
950 addr
, rss
, prealloc
, pte
, page
);
955 page_dup_rmap(page
, false);
956 rss
[mm_counter(page
)]++;
960 * If it's a COW mapping, write protect it both
961 * in the parent and the child
963 if (is_cow_mapping(vm_flags
) && pte_write(pte
)) {
964 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
965 pte
= pte_wrprotect(pte
);
969 * If it's a shared mapping, mark it clean in
972 if (vm_flags
& VM_SHARED
)
973 pte
= pte_mkclean(pte
);
974 pte
= pte_mkold(pte
);
976 if (!userfaultfd_wp(dst_vma
))
977 pte
= pte_clear_uffd_wp(pte
);
979 set_pte_at(dst_vma
->vm_mm
, addr
, dst_pte
, pte
);
983 static inline struct page
*
984 page_copy_prealloc(struct mm_struct
*src_mm
, struct vm_area_struct
*vma
,
987 struct page
*new_page
;
989 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, addr
);
993 if (mem_cgroup_charge(new_page
, src_mm
, GFP_KERNEL
)) {
997 cgroup_throttle_swaprate(new_page
, GFP_KERNEL
);
1003 copy_pte_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1004 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
1007 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1008 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1009 pte_t
*orig_src_pte
, *orig_dst_pte
;
1010 pte_t
*src_pte
, *dst_pte
;
1011 spinlock_t
*src_ptl
, *dst_ptl
;
1012 int progress
, ret
= 0;
1013 int rss
[NR_MM_COUNTERS
];
1014 swp_entry_t entry
= (swp_entry_t
){0};
1015 struct page
*prealloc
= NULL
;
1021 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
1026 src_pte
= pte_offset_map(src_pmd
, addr
);
1027 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
1028 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1029 orig_src_pte
= src_pte
;
1030 orig_dst_pte
= dst_pte
;
1031 arch_enter_lazy_mmu_mode();
1035 * We are holding two locks at this point - either of them
1036 * could generate latencies in another task on another CPU.
1038 if (progress
>= 32) {
1040 if (need_resched() ||
1041 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
1044 if (pte_none(*src_pte
)) {
1048 if (unlikely(!pte_present(*src_pte
))) {
1049 ret
= copy_nonpresent_pte(dst_mm
, src_mm
,
1054 entry
= pte_to_swp_entry(*src_pte
);
1056 } else if (ret
== -EBUSY
) {
1064 * Device exclusive entry restored, continue by copying
1065 * the now present pte.
1067 WARN_ON_ONCE(ret
!= -ENOENT
);
1069 /* copy_present_pte() will clear `*prealloc' if consumed */
1070 ret
= copy_present_pte(dst_vma
, src_vma
, dst_pte
, src_pte
,
1071 addr
, rss
, &prealloc
);
1073 * If we need a pre-allocated page for this pte, drop the
1074 * locks, allocate, and try again.
1076 if (unlikely(ret
== -EAGAIN
))
1078 if (unlikely(prealloc
)) {
1080 * pre-alloc page cannot be reused by next time so as
1081 * to strictly follow mempolicy (e.g., alloc_page_vma()
1082 * will allocate page according to address). This
1083 * could only happen if one pinned pte changed.
1089 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1091 arch_leave_lazy_mmu_mode();
1092 spin_unlock(src_ptl
);
1093 pte_unmap(orig_src_pte
);
1094 add_mm_rss_vec(dst_mm
, rss
);
1095 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
1099 VM_WARN_ON_ONCE(!entry
.val
);
1100 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0) {
1105 } else if (ret
== -EBUSY
) {
1107 } else if (ret
== -EAGAIN
) {
1108 prealloc
= page_copy_prealloc(src_mm
, src_vma
, addr
);
1115 /* We've captured and resolved the error. Reset, try again. */
1121 if (unlikely(prealloc
))
1127 copy_pmd_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1128 pud_t
*dst_pud
, pud_t
*src_pud
, unsigned long addr
,
1131 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1132 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1133 pmd_t
*src_pmd
, *dst_pmd
;
1136 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
1139 src_pmd
= pmd_offset(src_pud
, addr
);
1141 next
= pmd_addr_end(addr
, end
);
1142 if (is_swap_pmd(*src_pmd
) || pmd_trans_huge(*src_pmd
)
1143 || pmd_devmap(*src_pmd
)) {
1145 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PMD_SIZE
, src_vma
);
1146 err
= copy_huge_pmd(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
1147 addr
, dst_vma
, src_vma
);
1154 if (pmd_none_or_clear_bad(src_pmd
))
1156 if (copy_pte_range(dst_vma
, src_vma
, dst_pmd
, src_pmd
,
1159 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1164 copy_pud_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1165 p4d_t
*dst_p4d
, p4d_t
*src_p4d
, unsigned long addr
,
1168 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1169 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1170 pud_t
*src_pud
, *dst_pud
;
1173 dst_pud
= pud_alloc(dst_mm
, dst_p4d
, addr
);
1176 src_pud
= pud_offset(src_p4d
, addr
);
1178 next
= pud_addr_end(addr
, end
);
1179 if (pud_trans_huge(*src_pud
) || pud_devmap(*src_pud
)) {
1182 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PUD_SIZE
, src_vma
);
1183 err
= copy_huge_pud(dst_mm
, src_mm
,
1184 dst_pud
, src_pud
, addr
, src_vma
);
1191 if (pud_none_or_clear_bad(src_pud
))
1193 if (copy_pmd_range(dst_vma
, src_vma
, dst_pud
, src_pud
,
1196 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1201 copy_p4d_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1202 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, unsigned long addr
,
1205 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1206 p4d_t
*src_p4d
, *dst_p4d
;
1209 dst_p4d
= p4d_alloc(dst_mm
, dst_pgd
, addr
);
1212 src_p4d
= p4d_offset(src_pgd
, addr
);
1214 next
= p4d_addr_end(addr
, end
);
1215 if (p4d_none_or_clear_bad(src_p4d
))
1217 if (copy_pud_range(dst_vma
, src_vma
, dst_p4d
, src_p4d
,
1220 } while (dst_p4d
++, src_p4d
++, addr
= next
, addr
!= end
);
1225 copy_page_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
)
1227 pgd_t
*src_pgd
, *dst_pgd
;
1229 unsigned long addr
= src_vma
->vm_start
;
1230 unsigned long end
= src_vma
->vm_end
;
1231 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1232 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1233 struct mmu_notifier_range range
;
1238 * Don't copy ptes where a page fault will fill them correctly.
1239 * Fork becomes much lighter when there are big shared or private
1240 * readonly mappings. The tradeoff is that copy_page_range is more
1241 * efficient than faulting.
1243 if (!(src_vma
->vm_flags
& (VM_HUGETLB
| VM_PFNMAP
| VM_MIXEDMAP
)) &&
1247 if (is_vm_hugetlb_page(src_vma
))
1248 return copy_hugetlb_page_range(dst_mm
, src_mm
, src_vma
);
1250 if (unlikely(src_vma
->vm_flags
& VM_PFNMAP
)) {
1252 * We do not free on error cases below as remove_vma
1253 * gets called on error from higher level routine
1255 ret
= track_pfn_copy(src_vma
);
1261 * We need to invalidate the secondary MMU mappings only when
1262 * there could be a permission downgrade on the ptes of the
1263 * parent mm. And a permission downgrade will only happen if
1264 * is_cow_mapping() returns true.
1266 is_cow
= is_cow_mapping(src_vma
->vm_flags
);
1269 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_PAGE
,
1270 0, src_vma
, src_mm
, addr
, end
);
1271 mmu_notifier_invalidate_range_start(&range
);
1273 * Disabling preemption is not needed for the write side, as
1274 * the read side doesn't spin, but goes to the mmap_lock.
1276 * Use the raw variant of the seqcount_t write API to avoid
1277 * lockdep complaining about preemptibility.
1279 mmap_assert_write_locked(src_mm
);
1280 raw_write_seqcount_begin(&src_mm
->write_protect_seq
);
1284 dst_pgd
= pgd_offset(dst_mm
, addr
);
1285 src_pgd
= pgd_offset(src_mm
, addr
);
1287 next
= pgd_addr_end(addr
, end
);
1288 if (pgd_none_or_clear_bad(src_pgd
))
1290 if (unlikely(copy_p4d_range(dst_vma
, src_vma
, dst_pgd
, src_pgd
,
1295 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1298 raw_write_seqcount_end(&src_mm
->write_protect_seq
);
1299 mmu_notifier_invalidate_range_end(&range
);
1304 /* Whether we should zap all COWed (private) pages too */
1305 static inline bool should_zap_cows(struct zap_details
*details
)
1307 /* By default, zap all pages */
1311 /* Or, we zap COWed pages only if the caller wants to */
1312 return !details
->check_mapping
;
1315 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1316 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1317 unsigned long addr
, unsigned long end
,
1318 struct zap_details
*details
)
1320 struct mm_struct
*mm
= tlb
->mm
;
1321 int force_flush
= 0;
1322 int rss
[NR_MM_COUNTERS
];
1328 tlb_change_page_size(tlb
, PAGE_SIZE
);
1331 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1333 flush_tlb_batched_pending(mm
);
1334 arch_enter_lazy_mmu_mode();
1337 if (pte_none(ptent
))
1343 if (pte_present(ptent
)) {
1346 page
= vm_normal_page(vma
, addr
, ptent
);
1347 if (unlikely(details
) && page
) {
1349 * unmap_shared_mapping_pages() wants to
1350 * invalidate cache without truncating:
1351 * unmap shared but keep private pages.
1353 if (details
->check_mapping
&&
1354 details
->check_mapping
!= page_rmapping(page
))
1357 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1359 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1360 if (unlikely(!page
))
1363 if (!PageAnon(page
)) {
1364 if (pte_dirty(ptent
)) {
1366 set_page_dirty(page
);
1368 if (pte_young(ptent
) &&
1369 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1370 mark_page_accessed(page
);
1372 rss
[mm_counter(page
)]--;
1373 page_remove_rmap(page
, false);
1374 if (unlikely(page_mapcount(page
) < 0))
1375 print_bad_pte(vma
, addr
, ptent
, page
);
1376 if (unlikely(__tlb_remove_page(tlb
, page
))) {
1384 entry
= pte_to_swp_entry(ptent
);
1385 if (is_device_private_entry(entry
) ||
1386 is_device_exclusive_entry(entry
)) {
1387 struct page
*page
= pfn_swap_entry_to_page(entry
);
1389 if (unlikely(details
&& details
->check_mapping
)) {
1391 * unmap_shared_mapping_pages() wants to
1392 * invalidate cache without truncating:
1393 * unmap shared but keep private pages.
1395 if (details
->check_mapping
!=
1396 page_rmapping(page
))
1400 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1401 rss
[mm_counter(page
)]--;
1403 if (is_device_private_entry(entry
))
1404 page_remove_rmap(page
, false);
1410 if (!non_swap_entry(entry
)) {
1411 /* Genuine swap entry, hence a private anon page */
1412 if (!should_zap_cows(details
))
1415 } else if (is_migration_entry(entry
)) {
1418 page
= pfn_swap_entry_to_page(entry
);
1419 if (details
&& details
->check_mapping
&&
1420 details
->check_mapping
!= page_rmapping(page
))
1422 rss
[mm_counter(page
)]--;
1424 if (unlikely(!free_swap_and_cache(entry
)))
1425 print_bad_pte(vma
, addr
, ptent
, NULL
);
1426 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1427 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1429 add_mm_rss_vec(mm
, rss
);
1430 arch_leave_lazy_mmu_mode();
1432 /* Do the actual TLB flush before dropping ptl */
1434 tlb_flush_mmu_tlbonly(tlb
);
1435 pte_unmap_unlock(start_pte
, ptl
);
1438 * If we forced a TLB flush (either due to running out of
1439 * batch buffers or because we needed to flush dirty TLB
1440 * entries before releasing the ptl), free the batched
1441 * memory too. Restart if we didn't do everything.
1456 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1457 struct vm_area_struct
*vma
, pud_t
*pud
,
1458 unsigned long addr
, unsigned long end
,
1459 struct zap_details
*details
)
1464 pmd
= pmd_offset(pud
, addr
);
1466 next
= pmd_addr_end(addr
, end
);
1467 if (is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1468 if (next
- addr
!= HPAGE_PMD_SIZE
)
1469 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
1470 else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1473 } else if (details
&& details
->single_page
&&
1474 PageTransCompound(details
->single_page
) &&
1475 next
- addr
== HPAGE_PMD_SIZE
&& pmd_none(*pmd
)) {
1476 spinlock_t
*ptl
= pmd_lock(tlb
->mm
, pmd
);
1478 * Take and drop THP pmd lock so that we cannot return
1479 * prematurely, while zap_huge_pmd() has cleared *pmd,
1480 * but not yet decremented compound_mapcount().
1486 * Here there can be other concurrent MADV_DONTNEED or
1487 * trans huge page faults running, and if the pmd is
1488 * none or trans huge it can change under us. This is
1489 * because MADV_DONTNEED holds the mmap_lock in read
1492 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1494 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1497 } while (pmd
++, addr
= next
, addr
!= end
);
1502 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1503 struct vm_area_struct
*vma
, p4d_t
*p4d
,
1504 unsigned long addr
, unsigned long end
,
1505 struct zap_details
*details
)
1510 pud
= pud_offset(p4d
, addr
);
1512 next
= pud_addr_end(addr
, end
);
1513 if (pud_trans_huge(*pud
) || pud_devmap(*pud
)) {
1514 if (next
- addr
!= HPAGE_PUD_SIZE
) {
1515 mmap_assert_locked(tlb
->mm
);
1516 split_huge_pud(vma
, pud
, addr
);
1517 } else if (zap_huge_pud(tlb
, vma
, pud
, addr
))
1521 if (pud_none_or_clear_bad(pud
))
1523 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1526 } while (pud
++, addr
= next
, addr
!= end
);
1531 static inline unsigned long zap_p4d_range(struct mmu_gather
*tlb
,
1532 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1533 unsigned long addr
, unsigned long end
,
1534 struct zap_details
*details
)
1539 p4d
= p4d_offset(pgd
, addr
);
1541 next
= p4d_addr_end(addr
, end
);
1542 if (p4d_none_or_clear_bad(p4d
))
1544 next
= zap_pud_range(tlb
, vma
, p4d
, addr
, next
, details
);
1545 } while (p4d
++, addr
= next
, addr
!= end
);
1550 void unmap_page_range(struct mmu_gather
*tlb
,
1551 struct vm_area_struct
*vma
,
1552 unsigned long addr
, unsigned long end
,
1553 struct zap_details
*details
)
1558 BUG_ON(addr
>= end
);
1559 tlb_start_vma(tlb
, vma
);
1560 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1562 next
= pgd_addr_end(addr
, end
);
1563 if (pgd_none_or_clear_bad(pgd
))
1565 next
= zap_p4d_range(tlb
, vma
, pgd
, addr
, next
, details
);
1566 } while (pgd
++, addr
= next
, addr
!= end
);
1567 tlb_end_vma(tlb
, vma
);
1571 static void unmap_single_vma(struct mmu_gather
*tlb
,
1572 struct vm_area_struct
*vma
, unsigned long start_addr
,
1573 unsigned long end_addr
,
1574 struct zap_details
*details
)
1576 unsigned long start
= max(vma
->vm_start
, start_addr
);
1579 if (start
>= vma
->vm_end
)
1581 end
= min(vma
->vm_end
, end_addr
);
1582 if (end
<= vma
->vm_start
)
1586 uprobe_munmap(vma
, start
, end
);
1588 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1589 untrack_pfn(vma
, 0, 0);
1592 if (unlikely(is_vm_hugetlb_page(vma
))) {
1594 * It is undesirable to test vma->vm_file as it
1595 * should be non-null for valid hugetlb area.
1596 * However, vm_file will be NULL in the error
1597 * cleanup path of mmap_region. When
1598 * hugetlbfs ->mmap method fails,
1599 * mmap_region() nullifies vma->vm_file
1600 * before calling this function to clean up.
1601 * Since no pte has actually been setup, it is
1602 * safe to do nothing in this case.
1605 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
1606 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1607 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
1610 unmap_page_range(tlb
, vma
, start
, end
, details
);
1615 * unmap_vmas - unmap a range of memory covered by a list of vma's
1616 * @tlb: address of the caller's struct mmu_gather
1617 * @vma: the starting vma
1618 * @start_addr: virtual address at which to start unmapping
1619 * @end_addr: virtual address at which to end unmapping
1621 * Unmap all pages in the vma list.
1623 * Only addresses between `start' and `end' will be unmapped.
1625 * The VMA list must be sorted in ascending virtual address order.
1627 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1628 * range after unmap_vmas() returns. So the only responsibility here is to
1629 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1630 * drops the lock and schedules.
1632 void unmap_vmas(struct mmu_gather
*tlb
,
1633 struct vm_area_struct
*vma
, unsigned long start_addr
,
1634 unsigned long end_addr
)
1636 struct mmu_notifier_range range
;
1638 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
, vma
->vm_mm
,
1639 start_addr
, end_addr
);
1640 mmu_notifier_invalidate_range_start(&range
);
1641 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1642 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1643 mmu_notifier_invalidate_range_end(&range
);
1647 * zap_page_range - remove user pages in a given range
1648 * @vma: vm_area_struct holding the applicable pages
1649 * @start: starting address of pages to zap
1650 * @size: number of bytes to zap
1652 * Caller must protect the VMA list
1654 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1657 struct mmu_notifier_range range
;
1658 struct mmu_gather tlb
;
1661 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1662 start
, start
+ size
);
1663 tlb_gather_mmu(&tlb
, vma
->vm_mm
);
1664 update_hiwater_rss(vma
->vm_mm
);
1665 mmu_notifier_invalidate_range_start(&range
);
1666 for ( ; vma
&& vma
->vm_start
< range
.end
; vma
= vma
->vm_next
)
1667 unmap_single_vma(&tlb
, vma
, start
, range
.end
, NULL
);
1668 mmu_notifier_invalidate_range_end(&range
);
1669 tlb_finish_mmu(&tlb
);
1671 EXPORT_SYMBOL(zap_page_range
);
1674 * zap_page_range_single - remove user pages in a given range
1675 * @vma: vm_area_struct holding the applicable pages
1676 * @address: starting address of pages to zap
1677 * @size: number of bytes to zap
1678 * @details: details of shared cache invalidation
1680 * The range must fit into one VMA.
1682 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1683 unsigned long size
, struct zap_details
*details
)
1685 struct mmu_notifier_range range
;
1686 struct mmu_gather tlb
;
1689 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1690 address
, address
+ size
);
1691 tlb_gather_mmu(&tlb
, vma
->vm_mm
);
1692 update_hiwater_rss(vma
->vm_mm
);
1693 mmu_notifier_invalidate_range_start(&range
);
1694 unmap_single_vma(&tlb
, vma
, address
, range
.end
, details
);
1695 mmu_notifier_invalidate_range_end(&range
);
1696 tlb_finish_mmu(&tlb
);
1700 * zap_vma_ptes - remove ptes mapping the vma
1701 * @vma: vm_area_struct holding ptes to be zapped
1702 * @address: starting address of pages to zap
1703 * @size: number of bytes to zap
1705 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1707 * The entire address range must be fully contained within the vma.
1710 void zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1713 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1714 !(vma
->vm_flags
& VM_PFNMAP
))
1717 zap_page_range_single(vma
, address
, size
, NULL
);
1719 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1721 static pmd_t
*walk_to_pmd(struct mm_struct
*mm
, unsigned long addr
)
1728 pgd
= pgd_offset(mm
, addr
);
1729 p4d
= p4d_alloc(mm
, pgd
, addr
);
1732 pud
= pud_alloc(mm
, p4d
, addr
);
1735 pmd
= pmd_alloc(mm
, pud
, addr
);
1739 VM_BUG_ON(pmd_trans_huge(*pmd
));
1743 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1746 pmd_t
*pmd
= walk_to_pmd(mm
, addr
);
1750 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1753 static int validate_page_before_insert(struct page
*page
)
1755 if (PageAnon(page
) || PageSlab(page
) || page_has_type(page
))
1757 flush_dcache_page(page
);
1761 static int insert_page_into_pte_locked(struct mm_struct
*mm
, pte_t
*pte
,
1762 unsigned long addr
, struct page
*page
, pgprot_t prot
)
1764 if (!pte_none(*pte
))
1766 /* Ok, finally just insert the thing.. */
1768 inc_mm_counter_fast(mm
, mm_counter_file(page
));
1769 page_add_file_rmap(page
, false);
1770 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1775 * This is the old fallback for page remapping.
1777 * For historical reasons, it only allows reserved pages. Only
1778 * old drivers should use this, and they needed to mark their
1779 * pages reserved for the old functions anyway.
1781 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1782 struct page
*page
, pgprot_t prot
)
1784 struct mm_struct
*mm
= vma
->vm_mm
;
1789 retval
= validate_page_before_insert(page
);
1793 pte
= get_locked_pte(mm
, addr
, &ptl
);
1796 retval
= insert_page_into_pte_locked(mm
, pte
, addr
, page
, prot
);
1797 pte_unmap_unlock(pte
, ptl
);
1803 static int insert_page_in_batch_locked(struct mm_struct
*mm
, pte_t
*pte
,
1804 unsigned long addr
, struct page
*page
, pgprot_t prot
)
1808 if (!page_count(page
))
1810 err
= validate_page_before_insert(page
);
1813 return insert_page_into_pte_locked(mm
, pte
, addr
, page
, prot
);
1816 /* insert_pages() amortizes the cost of spinlock operations
1817 * when inserting pages in a loop. Arch *must* define pte_index.
1819 static int insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
1820 struct page
**pages
, unsigned long *num
, pgprot_t prot
)
1823 pte_t
*start_pte
, *pte
;
1824 spinlock_t
*pte_lock
;
1825 struct mm_struct
*const mm
= vma
->vm_mm
;
1826 unsigned long curr_page_idx
= 0;
1827 unsigned long remaining_pages_total
= *num
;
1828 unsigned long pages_to_write_in_pmd
;
1832 pmd
= walk_to_pmd(mm
, addr
);
1836 pages_to_write_in_pmd
= min_t(unsigned long,
1837 remaining_pages_total
, PTRS_PER_PTE
- pte_index(addr
));
1839 /* Allocate the PTE if necessary; takes PMD lock once only. */
1841 if (pte_alloc(mm
, pmd
))
1844 while (pages_to_write_in_pmd
) {
1846 const int batch_size
= min_t(int, pages_to_write_in_pmd
, 8);
1848 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &pte_lock
);
1849 for (pte
= start_pte
; pte_idx
< batch_size
; ++pte
, ++pte_idx
) {
1850 int err
= insert_page_in_batch_locked(mm
, pte
,
1851 addr
, pages
[curr_page_idx
], prot
);
1852 if (unlikely(err
)) {
1853 pte_unmap_unlock(start_pte
, pte_lock
);
1855 remaining_pages_total
-= pte_idx
;
1861 pte_unmap_unlock(start_pte
, pte_lock
);
1862 pages_to_write_in_pmd
-= batch_size
;
1863 remaining_pages_total
-= batch_size
;
1865 if (remaining_pages_total
)
1869 *num
= remaining_pages_total
;
1872 #endif /* ifdef pte_index */
1875 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1876 * @vma: user vma to map to
1877 * @addr: target start user address of these pages
1878 * @pages: source kernel pages
1879 * @num: in: number of pages to map. out: number of pages that were *not*
1880 * mapped. (0 means all pages were successfully mapped).
1882 * Preferred over vm_insert_page() when inserting multiple pages.
1884 * In case of error, we may have mapped a subset of the provided
1885 * pages. It is the caller's responsibility to account for this case.
1887 * The same restrictions apply as in vm_insert_page().
1889 int vm_insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
1890 struct page
**pages
, unsigned long *num
)
1893 const unsigned long end_addr
= addr
+ (*num
* PAGE_SIZE
) - 1;
1895 if (addr
< vma
->vm_start
|| end_addr
>= vma
->vm_end
)
1897 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1898 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
1899 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1900 vma
->vm_flags
|= VM_MIXEDMAP
;
1902 /* Defer page refcount checking till we're about to map that page. */
1903 return insert_pages(vma
, addr
, pages
, num
, vma
->vm_page_prot
);
1905 unsigned long idx
= 0, pgcount
= *num
;
1908 for (; idx
< pgcount
; ++idx
) {
1909 err
= vm_insert_page(vma
, addr
+ (PAGE_SIZE
* idx
), pages
[idx
]);
1913 *num
= pgcount
- idx
;
1915 #endif /* ifdef pte_index */
1917 EXPORT_SYMBOL(vm_insert_pages
);
1920 * vm_insert_page - insert single page into user vma
1921 * @vma: user vma to map to
1922 * @addr: target user address of this page
1923 * @page: source kernel page
1925 * This allows drivers to insert individual pages they've allocated
1928 * The page has to be a nice clean _individual_ kernel allocation.
1929 * If you allocate a compound page, you need to have marked it as
1930 * such (__GFP_COMP), or manually just split the page up yourself
1931 * (see split_page()).
1933 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1934 * took an arbitrary page protection parameter. This doesn't allow
1935 * that. Your vma protection will have to be set up correctly, which
1936 * means that if you want a shared writable mapping, you'd better
1937 * ask for a shared writable mapping!
1939 * The page does not need to be reserved.
1941 * Usually this function is called from f_op->mmap() handler
1942 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1943 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1944 * function from other places, for example from page-fault handler.
1946 * Return: %0 on success, negative error code otherwise.
1948 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1951 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1953 if (!page_count(page
))
1955 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1956 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
1957 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1958 vma
->vm_flags
|= VM_MIXEDMAP
;
1960 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1962 EXPORT_SYMBOL(vm_insert_page
);
1965 * __vm_map_pages - maps range of kernel pages into user vma
1966 * @vma: user vma to map to
1967 * @pages: pointer to array of source kernel pages
1968 * @num: number of pages in page array
1969 * @offset: user's requested vm_pgoff
1971 * This allows drivers to map range of kernel pages into a user vma.
1973 * Return: 0 on success and error code otherwise.
1975 static int __vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
1976 unsigned long num
, unsigned long offset
)
1978 unsigned long count
= vma_pages(vma
);
1979 unsigned long uaddr
= vma
->vm_start
;
1982 /* Fail if the user requested offset is beyond the end of the object */
1986 /* Fail if the user requested size exceeds available object size */
1987 if (count
> num
- offset
)
1990 for (i
= 0; i
< count
; i
++) {
1991 ret
= vm_insert_page(vma
, uaddr
, pages
[offset
+ i
]);
2001 * vm_map_pages - maps range of kernel pages starts with non zero offset
2002 * @vma: user vma to map to
2003 * @pages: pointer to array of source kernel pages
2004 * @num: number of pages in page array
2006 * Maps an object consisting of @num pages, catering for the user's
2007 * requested vm_pgoff
2009 * If we fail to insert any page into the vma, the function will return
2010 * immediately leaving any previously inserted pages present. Callers
2011 * from the mmap handler may immediately return the error as their caller
2012 * will destroy the vma, removing any successfully inserted pages. Other
2013 * callers should make their own arrangements for calling unmap_region().
2015 * Context: Process context. Called by mmap handlers.
2016 * Return: 0 on success and error code otherwise.
2018 int vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
2021 return __vm_map_pages(vma
, pages
, num
, vma
->vm_pgoff
);
2023 EXPORT_SYMBOL(vm_map_pages
);
2026 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2027 * @vma: user vma to map to
2028 * @pages: pointer to array of source kernel pages
2029 * @num: number of pages in page array
2031 * Similar to vm_map_pages(), except that it explicitly sets the offset
2032 * to 0. This function is intended for the drivers that did not consider
2035 * Context: Process context. Called by mmap handlers.
2036 * Return: 0 on success and error code otherwise.
2038 int vm_map_pages_zero(struct vm_area_struct
*vma
, struct page
**pages
,
2041 return __vm_map_pages(vma
, pages
, num
, 0);
2043 EXPORT_SYMBOL(vm_map_pages_zero
);
2045 static vm_fault_t
insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2046 pfn_t pfn
, pgprot_t prot
, bool mkwrite
)
2048 struct mm_struct
*mm
= vma
->vm_mm
;
2052 pte
= get_locked_pte(mm
, addr
, &ptl
);
2054 return VM_FAULT_OOM
;
2055 if (!pte_none(*pte
)) {
2058 * For read faults on private mappings the PFN passed
2059 * in may not match the PFN we have mapped if the
2060 * mapped PFN is a writeable COW page. In the mkwrite
2061 * case we are creating a writable PTE for a shared
2062 * mapping and we expect the PFNs to match. If they
2063 * don't match, we are likely racing with block
2064 * allocation and mapping invalidation so just skip the
2067 if (pte_pfn(*pte
) != pfn_t_to_pfn(pfn
)) {
2068 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte
)));
2071 entry
= pte_mkyoung(*pte
);
2072 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2073 if (ptep_set_access_flags(vma
, addr
, pte
, entry
, 1))
2074 update_mmu_cache(vma
, addr
, pte
);
2079 /* Ok, finally just insert the thing.. */
2080 if (pfn_t_devmap(pfn
))
2081 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
2083 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
2086 entry
= pte_mkyoung(entry
);
2087 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2090 set_pte_at(mm
, addr
, pte
, entry
);
2091 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
2094 pte_unmap_unlock(pte
, ptl
);
2095 return VM_FAULT_NOPAGE
;
2099 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2100 * @vma: user vma to map to
2101 * @addr: target user address of this page
2102 * @pfn: source kernel pfn
2103 * @pgprot: pgprot flags for the inserted page
2105 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2106 * to override pgprot on a per-page basis.
2108 * This only makes sense for IO mappings, and it makes no sense for
2109 * COW mappings. In general, using multiple vmas is preferable;
2110 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2113 * See vmf_insert_mixed_prot() for a discussion of the implication of using
2114 * a value of @pgprot different from that of @vma->vm_page_prot.
2116 * Context: Process context. May allocate using %GFP_KERNEL.
2117 * Return: vm_fault_t value.
2119 vm_fault_t
vmf_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
2120 unsigned long pfn
, pgprot_t pgprot
)
2123 * Technically, architectures with pte_special can avoid all these
2124 * restrictions (same for remap_pfn_range). However we would like
2125 * consistency in testing and feature parity among all, so we should
2126 * try to keep these invariants in place for everybody.
2128 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
2129 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
2130 (VM_PFNMAP
|VM_MIXEDMAP
));
2131 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
2132 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
2134 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2135 return VM_FAULT_SIGBUS
;
2137 if (!pfn_modify_allowed(pfn
, pgprot
))
2138 return VM_FAULT_SIGBUS
;
2140 track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
));
2142 return insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
,
2145 EXPORT_SYMBOL(vmf_insert_pfn_prot
);
2148 * vmf_insert_pfn - insert single pfn into user vma
2149 * @vma: user vma to map to
2150 * @addr: target user address of this page
2151 * @pfn: source kernel pfn
2153 * Similar to vm_insert_page, this allows drivers to insert individual pages
2154 * they've allocated into a user vma. Same comments apply.
2156 * This function should only be called from a vm_ops->fault handler, and
2157 * in that case the handler should return the result of this function.
2159 * vma cannot be a COW mapping.
2161 * As this is called only for pages that do not currently exist, we
2162 * do not need to flush old virtual caches or the TLB.
2164 * Context: Process context. May allocate using %GFP_KERNEL.
2165 * Return: vm_fault_t value.
2167 vm_fault_t
vmf_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2170 return vmf_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
2172 EXPORT_SYMBOL(vmf_insert_pfn
);
2174 static bool vm_mixed_ok(struct vm_area_struct
*vma
, pfn_t pfn
)
2176 /* these checks mirror the abort conditions in vm_normal_page */
2177 if (vma
->vm_flags
& VM_MIXEDMAP
)
2179 if (pfn_t_devmap(pfn
))
2181 if (pfn_t_special(pfn
))
2183 if (is_zero_pfn(pfn_t_to_pfn(pfn
)))
2188 static vm_fault_t
__vm_insert_mixed(struct vm_area_struct
*vma
,
2189 unsigned long addr
, pfn_t pfn
, pgprot_t pgprot
,
2194 BUG_ON(!vm_mixed_ok(vma
, pfn
));
2196 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2197 return VM_FAULT_SIGBUS
;
2199 track_pfn_insert(vma
, &pgprot
, pfn
);
2201 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn
), pgprot
))
2202 return VM_FAULT_SIGBUS
;
2205 * If we don't have pte special, then we have to use the pfn_valid()
2206 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2207 * refcount the page if pfn_valid is true (hence insert_page rather
2208 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2209 * without pte special, it would there be refcounted as a normal page.
2211 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
) &&
2212 !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
2216 * At this point we are committed to insert_page()
2217 * regardless of whether the caller specified flags that
2218 * result in pfn_t_has_page() == false.
2220 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
2221 err
= insert_page(vma
, addr
, page
, pgprot
);
2223 return insert_pfn(vma
, addr
, pfn
, pgprot
, mkwrite
);
2227 return VM_FAULT_OOM
;
2228 if (err
< 0 && err
!= -EBUSY
)
2229 return VM_FAULT_SIGBUS
;
2231 return VM_FAULT_NOPAGE
;
2235 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2236 * @vma: user vma to map to
2237 * @addr: target user address of this page
2238 * @pfn: source kernel pfn
2239 * @pgprot: pgprot flags for the inserted page
2241 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2242 * to override pgprot on a per-page basis.
2244 * Typically this function should be used by drivers to set caching- and
2245 * encryption bits different than those of @vma->vm_page_prot, because
2246 * the caching- or encryption mode may not be known at mmap() time.
2247 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2248 * to set caching and encryption bits for those vmas (except for COW pages).
2249 * This is ensured by core vm only modifying these page table entries using
2250 * functions that don't touch caching- or encryption bits, using pte_modify()
2251 * if needed. (See for example mprotect()).
2252 * Also when new page-table entries are created, this is only done using the
2253 * fault() callback, and never using the value of vma->vm_page_prot,
2254 * except for page-table entries that point to anonymous pages as the result
2257 * Context: Process context. May allocate using %GFP_KERNEL.
2258 * Return: vm_fault_t value.
2260 vm_fault_t
vmf_insert_mixed_prot(struct vm_area_struct
*vma
, unsigned long addr
,
2261 pfn_t pfn
, pgprot_t pgprot
)
2263 return __vm_insert_mixed(vma
, addr
, pfn
, pgprot
, false);
2265 EXPORT_SYMBOL(vmf_insert_mixed_prot
);
2267 vm_fault_t
vmf_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2270 return __vm_insert_mixed(vma
, addr
, pfn
, vma
->vm_page_prot
, false);
2272 EXPORT_SYMBOL(vmf_insert_mixed
);
2275 * If the insertion of PTE failed because someone else already added a
2276 * different entry in the mean time, we treat that as success as we assume
2277 * the same entry was actually inserted.
2279 vm_fault_t
vmf_insert_mixed_mkwrite(struct vm_area_struct
*vma
,
2280 unsigned long addr
, pfn_t pfn
)
2282 return __vm_insert_mixed(vma
, addr
, pfn
, vma
->vm_page_prot
, true);
2284 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite
);
2287 * maps a range of physical memory into the requested pages. the old
2288 * mappings are removed. any references to nonexistent pages results
2289 * in null mappings (currently treated as "copy-on-access")
2291 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2292 unsigned long addr
, unsigned long end
,
2293 unsigned long pfn
, pgprot_t prot
)
2295 pte_t
*pte
, *mapped_pte
;
2299 mapped_pte
= pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2302 arch_enter_lazy_mmu_mode();
2304 BUG_ON(!pte_none(*pte
));
2305 if (!pfn_modify_allowed(pfn
, prot
)) {
2309 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2311 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2312 arch_leave_lazy_mmu_mode();
2313 pte_unmap_unlock(mapped_pte
, ptl
);
2317 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2318 unsigned long addr
, unsigned long end
,
2319 unsigned long pfn
, pgprot_t prot
)
2325 pfn
-= addr
>> PAGE_SHIFT
;
2326 pmd
= pmd_alloc(mm
, pud
, addr
);
2329 VM_BUG_ON(pmd_trans_huge(*pmd
));
2331 next
= pmd_addr_end(addr
, end
);
2332 err
= remap_pte_range(mm
, pmd
, addr
, next
,
2333 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2336 } while (pmd
++, addr
= next
, addr
!= end
);
2340 static inline int remap_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2341 unsigned long addr
, unsigned long end
,
2342 unsigned long pfn
, pgprot_t prot
)
2348 pfn
-= addr
>> PAGE_SHIFT
;
2349 pud
= pud_alloc(mm
, p4d
, addr
);
2353 next
= pud_addr_end(addr
, end
);
2354 err
= remap_pmd_range(mm
, pud
, addr
, next
,
2355 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2358 } while (pud
++, addr
= next
, addr
!= end
);
2362 static inline int remap_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2363 unsigned long addr
, unsigned long end
,
2364 unsigned long pfn
, pgprot_t prot
)
2370 pfn
-= addr
>> PAGE_SHIFT
;
2371 p4d
= p4d_alloc(mm
, pgd
, addr
);
2375 next
= p4d_addr_end(addr
, end
);
2376 err
= remap_pud_range(mm
, p4d
, addr
, next
,
2377 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2380 } while (p4d
++, addr
= next
, addr
!= end
);
2385 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2386 * must have pre-validated the caching bits of the pgprot_t.
2388 int remap_pfn_range_notrack(struct vm_area_struct
*vma
, unsigned long addr
,
2389 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2393 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2394 struct mm_struct
*mm
= vma
->vm_mm
;
2397 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr
)))
2401 * Physically remapped pages are special. Tell the
2402 * rest of the world about it:
2403 * VM_IO tells people not to look at these pages
2404 * (accesses can have side effects).
2405 * VM_PFNMAP tells the core MM that the base pages are just
2406 * raw PFN mappings, and do not have a "struct page" associated
2409 * Disable vma merging and expanding with mremap().
2411 * Omit vma from core dump, even when VM_IO turned off.
2413 * There's a horrible special case to handle copy-on-write
2414 * behaviour that some programs depend on. We mark the "original"
2415 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2416 * See vm_normal_page() for details.
2418 if (is_cow_mapping(vma
->vm_flags
)) {
2419 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2421 vma
->vm_pgoff
= pfn
;
2424 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
2426 BUG_ON(addr
>= end
);
2427 pfn
-= addr
>> PAGE_SHIFT
;
2428 pgd
= pgd_offset(mm
, addr
);
2429 flush_cache_range(vma
, addr
, end
);
2431 next
= pgd_addr_end(addr
, end
);
2432 err
= remap_p4d_range(mm
, pgd
, addr
, next
,
2433 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2436 } while (pgd
++, addr
= next
, addr
!= end
);
2442 * remap_pfn_range - remap kernel memory to userspace
2443 * @vma: user vma to map to
2444 * @addr: target page aligned user address to start at
2445 * @pfn: page frame number of kernel physical memory address
2446 * @size: size of mapping area
2447 * @prot: page protection flags for this mapping
2449 * Note: this is only safe if the mm semaphore is held when called.
2451 * Return: %0 on success, negative error code otherwise.
2453 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2454 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2458 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
2462 err
= remap_pfn_range_notrack(vma
, addr
, pfn
, size
, prot
);
2464 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
));
2467 EXPORT_SYMBOL(remap_pfn_range
);
2470 * vm_iomap_memory - remap memory to userspace
2471 * @vma: user vma to map to
2472 * @start: start of the physical memory to be mapped
2473 * @len: size of area
2475 * This is a simplified io_remap_pfn_range() for common driver use. The
2476 * driver just needs to give us the physical memory range to be mapped,
2477 * we'll figure out the rest from the vma information.
2479 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2480 * whatever write-combining details or similar.
2482 * Return: %0 on success, negative error code otherwise.
2484 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2486 unsigned long vm_len
, pfn
, pages
;
2488 /* Check that the physical memory area passed in looks valid */
2489 if (start
+ len
< start
)
2492 * You *really* shouldn't map things that aren't page-aligned,
2493 * but we've historically allowed it because IO memory might
2494 * just have smaller alignment.
2496 len
+= start
& ~PAGE_MASK
;
2497 pfn
= start
>> PAGE_SHIFT
;
2498 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2499 if (pfn
+ pages
< pfn
)
2502 /* We start the mapping 'vm_pgoff' pages into the area */
2503 if (vma
->vm_pgoff
> pages
)
2505 pfn
+= vma
->vm_pgoff
;
2506 pages
-= vma
->vm_pgoff
;
2508 /* Can we fit all of the mapping? */
2509 vm_len
= vma
->vm_end
- vma
->vm_start
;
2510 if (vm_len
>> PAGE_SHIFT
> pages
)
2513 /* Ok, let it rip */
2514 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2516 EXPORT_SYMBOL(vm_iomap_memory
);
2518 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2519 unsigned long addr
, unsigned long end
,
2520 pte_fn_t fn
, void *data
, bool create
,
2521 pgtbl_mod_mask
*mask
)
2523 pte_t
*pte
, *mapped_pte
;
2528 mapped_pte
= pte
= (mm
== &init_mm
) ?
2529 pte_alloc_kernel_track(pmd
, addr
, mask
) :
2530 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2534 mapped_pte
= pte
= (mm
== &init_mm
) ?
2535 pte_offset_kernel(pmd
, addr
) :
2536 pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
2539 BUG_ON(pmd_huge(*pmd
));
2541 arch_enter_lazy_mmu_mode();
2545 if (create
|| !pte_none(*pte
)) {
2546 err
= fn(pte
++, addr
, data
);
2550 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2552 *mask
|= PGTBL_PTE_MODIFIED
;
2554 arch_leave_lazy_mmu_mode();
2557 pte_unmap_unlock(mapped_pte
, ptl
);
2561 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2562 unsigned long addr
, unsigned long end
,
2563 pte_fn_t fn
, void *data
, bool create
,
2564 pgtbl_mod_mask
*mask
)
2570 BUG_ON(pud_huge(*pud
));
2573 pmd
= pmd_alloc_track(mm
, pud
, addr
, mask
);
2577 pmd
= pmd_offset(pud
, addr
);
2580 next
= pmd_addr_end(addr
, end
);
2581 if (pmd_none(*pmd
) && !create
)
2583 if (WARN_ON_ONCE(pmd_leaf(*pmd
)))
2585 if (!pmd_none(*pmd
) && WARN_ON_ONCE(pmd_bad(*pmd
))) {
2590 err
= apply_to_pte_range(mm
, pmd
, addr
, next
,
2591 fn
, data
, create
, mask
);
2594 } while (pmd
++, addr
= next
, addr
!= end
);
2599 static int apply_to_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2600 unsigned long addr
, unsigned long end
,
2601 pte_fn_t fn
, void *data
, bool create
,
2602 pgtbl_mod_mask
*mask
)
2609 pud
= pud_alloc_track(mm
, p4d
, addr
, mask
);
2613 pud
= pud_offset(p4d
, addr
);
2616 next
= pud_addr_end(addr
, end
);
2617 if (pud_none(*pud
) && !create
)
2619 if (WARN_ON_ONCE(pud_leaf(*pud
)))
2621 if (!pud_none(*pud
) && WARN_ON_ONCE(pud_bad(*pud
))) {
2626 err
= apply_to_pmd_range(mm
, pud
, addr
, next
,
2627 fn
, data
, create
, mask
);
2630 } while (pud
++, addr
= next
, addr
!= end
);
2635 static int apply_to_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2636 unsigned long addr
, unsigned long end
,
2637 pte_fn_t fn
, void *data
, bool create
,
2638 pgtbl_mod_mask
*mask
)
2645 p4d
= p4d_alloc_track(mm
, pgd
, addr
, mask
);
2649 p4d
= p4d_offset(pgd
, addr
);
2652 next
= p4d_addr_end(addr
, end
);
2653 if (p4d_none(*p4d
) && !create
)
2655 if (WARN_ON_ONCE(p4d_leaf(*p4d
)))
2657 if (!p4d_none(*p4d
) && WARN_ON_ONCE(p4d_bad(*p4d
))) {
2662 err
= apply_to_pud_range(mm
, p4d
, addr
, next
,
2663 fn
, data
, create
, mask
);
2666 } while (p4d
++, addr
= next
, addr
!= end
);
2671 static int __apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2672 unsigned long size
, pte_fn_t fn
,
2673 void *data
, bool create
)
2676 unsigned long start
= addr
, next
;
2677 unsigned long end
= addr
+ size
;
2678 pgtbl_mod_mask mask
= 0;
2681 if (WARN_ON(addr
>= end
))
2684 pgd
= pgd_offset(mm
, addr
);
2686 next
= pgd_addr_end(addr
, end
);
2687 if (pgd_none(*pgd
) && !create
)
2689 if (WARN_ON_ONCE(pgd_leaf(*pgd
)))
2691 if (!pgd_none(*pgd
) && WARN_ON_ONCE(pgd_bad(*pgd
))) {
2696 err
= apply_to_p4d_range(mm
, pgd
, addr
, next
,
2697 fn
, data
, create
, &mask
);
2700 } while (pgd
++, addr
= next
, addr
!= end
);
2702 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
2703 arch_sync_kernel_mappings(start
, start
+ size
);
2709 * Scan a region of virtual memory, filling in page tables as necessary
2710 * and calling a provided function on each leaf page table.
2712 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2713 unsigned long size
, pte_fn_t fn
, void *data
)
2715 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, true);
2717 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2720 * Scan a region of virtual memory, calling a provided function on
2721 * each leaf page table where it exists.
2723 * Unlike apply_to_page_range, this does _not_ fill in page tables
2724 * where they are absent.
2726 int apply_to_existing_page_range(struct mm_struct
*mm
, unsigned long addr
,
2727 unsigned long size
, pte_fn_t fn
, void *data
)
2729 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, false);
2731 EXPORT_SYMBOL_GPL(apply_to_existing_page_range
);
2734 * handle_pte_fault chooses page fault handler according to an entry which was
2735 * read non-atomically. Before making any commitment, on those architectures
2736 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2737 * parts, do_swap_page must check under lock before unmapping the pte and
2738 * proceeding (but do_wp_page is only called after already making such a check;
2739 * and do_anonymous_page can safely check later on).
2741 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2742 pte_t
*page_table
, pte_t orig_pte
)
2745 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2746 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2747 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2749 same
= pte_same(*page_table
, orig_pte
);
2753 pte_unmap(page_table
);
2757 static inline bool cow_user_page(struct page
*dst
, struct page
*src
,
2758 struct vm_fault
*vmf
)
2763 bool locked
= false;
2764 struct vm_area_struct
*vma
= vmf
->vma
;
2765 struct mm_struct
*mm
= vma
->vm_mm
;
2766 unsigned long addr
= vmf
->address
;
2769 copy_user_highpage(dst
, src
, addr
, vma
);
2774 * If the source page was a PFN mapping, we don't have
2775 * a "struct page" for it. We do a best-effort copy by
2776 * just copying from the original user address. If that
2777 * fails, we just zero-fill it. Live with it.
2779 kaddr
= kmap_atomic(dst
);
2780 uaddr
= (void __user
*)(addr
& PAGE_MASK
);
2783 * On architectures with software "accessed" bits, we would
2784 * take a double page fault, so mark it accessed here.
2786 if (arch_faults_on_old_pte() && !pte_young(vmf
->orig_pte
)) {
2789 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2791 if (!likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2793 * Other thread has already handled the fault
2794 * and update local tlb only
2796 update_mmu_tlb(vma
, addr
, vmf
->pte
);
2801 entry
= pte_mkyoung(vmf
->orig_pte
);
2802 if (ptep_set_access_flags(vma
, addr
, vmf
->pte
, entry
, 0))
2803 update_mmu_cache(vma
, addr
, vmf
->pte
);
2807 * This really shouldn't fail, because the page is there
2808 * in the page tables. But it might just be unreadable,
2809 * in which case we just give up and fill the result with
2812 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
2816 /* Re-validate under PTL if the page is still mapped */
2817 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2819 if (!likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2820 /* The PTE changed under us, update local tlb */
2821 update_mmu_tlb(vma
, addr
, vmf
->pte
);
2827 * The same page can be mapped back since last copy attempt.
2828 * Try to copy again under PTL.
2830 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
2832 * Give a warn in case there can be some obscure
2845 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2846 kunmap_atomic(kaddr
);
2847 flush_dcache_page(dst
);
2852 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
2854 struct file
*vm_file
= vma
->vm_file
;
2857 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
2860 * Special mappings (e.g. VDSO) do not have any file so fake
2861 * a default GFP_KERNEL for them.
2867 * Notify the address space that the page is about to become writable so that
2868 * it can prohibit this or wait for the page to get into an appropriate state.
2870 * We do this without the lock held, so that it can sleep if it needs to.
2872 static vm_fault_t
do_page_mkwrite(struct vm_fault
*vmf
)
2875 struct page
*page
= vmf
->page
;
2876 unsigned int old_flags
= vmf
->flags
;
2878 vmf
->flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2880 if (vmf
->vma
->vm_file
&&
2881 IS_SWAPFILE(vmf
->vma
->vm_file
->f_mapping
->host
))
2882 return VM_FAULT_SIGBUS
;
2884 ret
= vmf
->vma
->vm_ops
->page_mkwrite(vmf
);
2885 /* Restore original flags so that caller is not surprised */
2886 vmf
->flags
= old_flags
;
2887 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2889 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
2891 if (!page
->mapping
) {
2893 return 0; /* retry */
2895 ret
|= VM_FAULT_LOCKED
;
2897 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2902 * Handle dirtying of a page in shared file mapping on a write fault.
2904 * The function expects the page to be locked and unlocks it.
2906 static vm_fault_t
fault_dirty_shared_page(struct vm_fault
*vmf
)
2908 struct vm_area_struct
*vma
= vmf
->vma
;
2909 struct address_space
*mapping
;
2910 struct page
*page
= vmf
->page
;
2912 bool page_mkwrite
= vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
;
2914 dirtied
= set_page_dirty(page
);
2915 VM_BUG_ON_PAGE(PageAnon(page
), page
);
2917 * Take a local copy of the address_space - page.mapping may be zeroed
2918 * by truncate after unlock_page(). The address_space itself remains
2919 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2920 * release semantics to prevent the compiler from undoing this copying.
2922 mapping
= page_rmapping(page
);
2926 file_update_time(vma
->vm_file
);
2929 * Throttle page dirtying rate down to writeback speed.
2931 * mapping may be NULL here because some device drivers do not
2932 * set page.mapping but still dirty their pages
2934 * Drop the mmap_lock before waiting on IO, if we can. The file
2935 * is pinning the mapping, as per above.
2937 if ((dirtied
|| page_mkwrite
) && mapping
) {
2940 fpin
= maybe_unlock_mmap_for_io(vmf
, NULL
);
2941 balance_dirty_pages_ratelimited(mapping
);
2944 return VM_FAULT_RETRY
;
2952 * Handle write page faults for pages that can be reused in the current vma
2954 * This can happen either due to the mapping being with the VM_SHARED flag,
2955 * or due to us being the last reference standing to the page. In either
2956 * case, all we need to do here is to mark the page as writable and update
2957 * any related book-keeping.
2959 static inline void wp_page_reuse(struct vm_fault
*vmf
)
2960 __releases(vmf
->ptl
)
2962 struct vm_area_struct
*vma
= vmf
->vma
;
2963 struct page
*page
= vmf
->page
;
2966 * Clear the pages cpupid information as the existing
2967 * information potentially belongs to a now completely
2968 * unrelated process.
2971 page_cpupid_xchg_last(page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2973 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2974 entry
= pte_mkyoung(vmf
->orig_pte
);
2975 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2976 if (ptep_set_access_flags(vma
, vmf
->address
, vmf
->pte
, entry
, 1))
2977 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2978 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2979 count_vm_event(PGREUSE
);
2983 * Handle the case of a page which we actually need to copy to a new page.
2985 * Called with mmap_lock locked and the old page referenced, but
2986 * without the ptl held.
2988 * High level logic flow:
2990 * - Allocate a page, copy the content of the old page to the new one.
2991 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2992 * - Take the PTL. If the pte changed, bail out and release the allocated page
2993 * - If the pte is still the way we remember it, update the page table and all
2994 * relevant references. This includes dropping the reference the page-table
2995 * held to the old page, as well as updating the rmap.
2996 * - In any case, unlock the PTL and drop the reference we took to the old page.
2998 static vm_fault_t
wp_page_copy(struct vm_fault
*vmf
)
3000 struct vm_area_struct
*vma
= vmf
->vma
;
3001 struct mm_struct
*mm
= vma
->vm_mm
;
3002 struct page
*old_page
= vmf
->page
;
3003 struct page
*new_page
= NULL
;
3005 int page_copied
= 0;
3006 struct mmu_notifier_range range
;
3008 if (unlikely(anon_vma_prepare(vma
)))
3011 if (is_zero_pfn(pte_pfn(vmf
->orig_pte
))) {
3012 new_page
= alloc_zeroed_user_highpage_movable(vma
,
3017 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
3022 if (!cow_user_page(new_page
, old_page
, vmf
)) {
3024 * COW failed, if the fault was solved by other,
3025 * it's fine. If not, userspace would re-fault on
3026 * the same address and we will handle the fault
3027 * from the second attempt.
3036 if (mem_cgroup_charge(new_page
, mm
, GFP_KERNEL
))
3038 cgroup_throttle_swaprate(new_page
, GFP_KERNEL
);
3040 __SetPageUptodate(new_page
);
3042 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
,
3043 vmf
->address
& PAGE_MASK
,
3044 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
);
3045 mmu_notifier_invalidate_range_start(&range
);
3048 * Re-check the pte - we dropped the lock
3050 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, vmf
->address
, &vmf
->ptl
);
3051 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
3053 if (!PageAnon(old_page
)) {
3054 dec_mm_counter_fast(mm
,
3055 mm_counter_file(old_page
));
3056 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3059 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3061 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
3062 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
3063 entry
= pte_sw_mkyoung(entry
);
3064 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3067 * Clear the pte entry and flush it first, before updating the
3068 * pte with the new entry, to keep TLBs on different CPUs in
3069 * sync. This code used to set the new PTE then flush TLBs, but
3070 * that left a window where the new PTE could be loaded into
3071 * some TLBs while the old PTE remains in others.
3073 ptep_clear_flush_notify(vma
, vmf
->address
, vmf
->pte
);
3074 page_add_new_anon_rmap(new_page
, vma
, vmf
->address
, false);
3075 lru_cache_add_inactive_or_unevictable(new_page
, vma
);
3077 * We call the notify macro here because, when using secondary
3078 * mmu page tables (such as kvm shadow page tables), we want the
3079 * new page to be mapped directly into the secondary page table.
3081 set_pte_at_notify(mm
, vmf
->address
, vmf
->pte
, entry
);
3082 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3085 * Only after switching the pte to the new page may
3086 * we remove the mapcount here. Otherwise another
3087 * process may come and find the rmap count decremented
3088 * before the pte is switched to the new page, and
3089 * "reuse" the old page writing into it while our pte
3090 * here still points into it and can be read by other
3093 * The critical issue is to order this
3094 * page_remove_rmap with the ptp_clear_flush above.
3095 * Those stores are ordered by (if nothing else,)
3096 * the barrier present in the atomic_add_negative
3097 * in page_remove_rmap.
3099 * Then the TLB flush in ptep_clear_flush ensures that
3100 * no process can access the old page before the
3101 * decremented mapcount is visible. And the old page
3102 * cannot be reused until after the decremented
3103 * mapcount is visible. So transitively, TLBs to
3104 * old page will be flushed before it can be reused.
3106 page_remove_rmap(old_page
, false);
3109 /* Free the old page.. */
3110 new_page
= old_page
;
3113 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
3119 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3121 * No need to double call mmu_notifier->invalidate_range() callback as
3122 * the above ptep_clear_flush_notify() did already call it.
3124 mmu_notifier_invalidate_range_only_end(&range
);
3127 * Don't let another task, with possibly unlocked vma,
3128 * keep the mlocked page.
3130 if (page_copied
&& (vma
->vm_flags
& VM_LOCKED
)) {
3131 lock_page(old_page
); /* LRU manipulation */
3132 if (PageMlocked(old_page
))
3133 munlock_vma_page(old_page
);
3134 unlock_page(old_page
);
3137 free_swap_cache(old_page
);
3140 return page_copied
? VM_FAULT_WRITE
: 0;
3146 return VM_FAULT_OOM
;
3150 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3151 * writeable once the page is prepared
3153 * @vmf: structure describing the fault
3155 * This function handles all that is needed to finish a write page fault in a
3156 * shared mapping due to PTE being read-only once the mapped page is prepared.
3157 * It handles locking of PTE and modifying it.
3159 * The function expects the page to be locked or other protection against
3160 * concurrent faults / writeback (such as DAX radix tree locks).
3162 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3163 * we acquired PTE lock.
3165 vm_fault_t
finish_mkwrite_fault(struct vm_fault
*vmf
)
3167 WARN_ON_ONCE(!(vmf
->vma
->vm_flags
& VM_SHARED
));
3168 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3171 * We might have raced with another page fault while we released the
3172 * pte_offset_map_lock.
3174 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
3175 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
3176 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3177 return VM_FAULT_NOPAGE
;
3184 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3187 static vm_fault_t
wp_pfn_shared(struct vm_fault
*vmf
)
3189 struct vm_area_struct
*vma
= vmf
->vma
;
3191 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
3194 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3195 vmf
->flags
|= FAULT_FLAG_MKWRITE
;
3196 ret
= vma
->vm_ops
->pfn_mkwrite(vmf
);
3197 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))
3199 return finish_mkwrite_fault(vmf
);
3202 return VM_FAULT_WRITE
;
3205 static vm_fault_t
wp_page_shared(struct vm_fault
*vmf
)
3206 __releases(vmf
->ptl
)
3208 struct vm_area_struct
*vma
= vmf
->vma
;
3209 vm_fault_t ret
= VM_FAULT_WRITE
;
3211 get_page(vmf
->page
);
3213 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
3216 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3217 tmp
= do_page_mkwrite(vmf
);
3218 if (unlikely(!tmp
|| (tmp
&
3219 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3220 put_page(vmf
->page
);
3223 tmp
= finish_mkwrite_fault(vmf
);
3224 if (unlikely(tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3225 unlock_page(vmf
->page
);
3226 put_page(vmf
->page
);
3231 lock_page(vmf
->page
);
3233 ret
|= fault_dirty_shared_page(vmf
);
3234 put_page(vmf
->page
);
3240 * This routine handles present pages, when users try to write
3241 * to a shared page. It is done by copying the page to a new address
3242 * and decrementing the shared-page counter for the old page.
3244 * Note that this routine assumes that the protection checks have been
3245 * done by the caller (the low-level page fault routine in most cases).
3246 * Thus we can safely just mark it writable once we've done any necessary
3249 * We also mark the page dirty at this point even though the page will
3250 * change only once the write actually happens. This avoids a few races,
3251 * and potentially makes it more efficient.
3253 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3254 * but allow concurrent faults), with pte both mapped and locked.
3255 * We return with mmap_lock still held, but pte unmapped and unlocked.
3257 static vm_fault_t
do_wp_page(struct vm_fault
*vmf
)
3258 __releases(vmf
->ptl
)
3260 struct vm_area_struct
*vma
= vmf
->vma
;
3262 if (userfaultfd_pte_wp(vma
, *vmf
->pte
)) {
3263 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3264 return handle_userfault(vmf
, VM_UFFD_WP
);
3268 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3269 * is flushed in this case before copying.
3271 if (unlikely(userfaultfd_wp(vmf
->vma
) &&
3272 mm_tlb_flush_pending(vmf
->vma
->vm_mm
)))
3273 flush_tlb_page(vmf
->vma
, vmf
->address
);
3275 vmf
->page
= vm_normal_page(vma
, vmf
->address
, vmf
->orig_pte
);
3278 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3281 * We should not cow pages in a shared writeable mapping.
3282 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3284 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
3285 (VM_WRITE
|VM_SHARED
))
3286 return wp_pfn_shared(vmf
);
3288 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3289 return wp_page_copy(vmf
);
3293 * Take out anonymous pages first, anonymous shared vmas are
3294 * not dirty accountable.
3296 if (PageAnon(vmf
->page
)) {
3297 struct page
*page
= vmf
->page
;
3299 /* PageKsm() doesn't necessarily raise the page refcount */
3300 if (PageKsm(page
) || page_count(page
) != 1)
3302 if (!trylock_page(page
))
3304 if (PageKsm(page
) || page_mapcount(page
) != 1 || page_count(page
) != 1) {
3309 * Ok, we've got the only map reference, and the only
3310 * page count reference, and the page is locked,
3311 * it's dark out, and we're wearing sunglasses. Hit it.
3315 return VM_FAULT_WRITE
;
3316 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
3317 (VM_WRITE
|VM_SHARED
))) {
3318 return wp_page_shared(vmf
);
3322 * Ok, we need to copy. Oh, well..
3324 get_page(vmf
->page
);
3326 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3327 return wp_page_copy(vmf
);
3330 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
3331 unsigned long start_addr
, unsigned long end_addr
,
3332 struct zap_details
*details
)
3334 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
3337 static inline void unmap_mapping_range_tree(struct rb_root_cached
*root
,
3338 struct zap_details
*details
)
3340 struct vm_area_struct
*vma
;
3341 pgoff_t vba
, vea
, zba
, zea
;
3343 vma_interval_tree_foreach(vma
, root
,
3344 details
->first_index
, details
->last_index
) {
3346 vba
= vma
->vm_pgoff
;
3347 vea
= vba
+ vma_pages(vma
) - 1;
3348 zba
= details
->first_index
;
3351 zea
= details
->last_index
;
3355 unmap_mapping_range_vma(vma
,
3356 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
3357 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
3363 * unmap_mapping_page() - Unmap single page from processes.
3364 * @page: The locked page to be unmapped.
3366 * Unmap this page from any userspace process which still has it mmaped.
3367 * Typically, for efficiency, the range of nearby pages has already been
3368 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3369 * truncation or invalidation holds the lock on a page, it may find that
3370 * the page has been remapped again: and then uses unmap_mapping_page()
3371 * to unmap it finally.
3373 void unmap_mapping_page(struct page
*page
)
3375 struct address_space
*mapping
= page
->mapping
;
3376 struct zap_details details
= { };
3378 VM_BUG_ON(!PageLocked(page
));
3379 VM_BUG_ON(PageTail(page
));
3381 details
.check_mapping
= mapping
;
3382 details
.first_index
= page
->index
;
3383 details
.last_index
= page
->index
+ thp_nr_pages(page
) - 1;
3384 details
.single_page
= page
;
3386 i_mmap_lock_write(mapping
);
3387 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3388 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
3389 i_mmap_unlock_write(mapping
);
3393 * unmap_mapping_pages() - Unmap pages from processes.
3394 * @mapping: The address space containing pages to be unmapped.
3395 * @start: Index of first page to be unmapped.
3396 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3397 * @even_cows: Whether to unmap even private COWed pages.
3399 * Unmap the pages in this address space from any userspace process which
3400 * has them mmaped. Generally, you want to remove COWed pages as well when
3401 * a file is being truncated, but not when invalidating pages from the page
3404 void unmap_mapping_pages(struct address_space
*mapping
, pgoff_t start
,
3405 pgoff_t nr
, bool even_cows
)
3407 struct zap_details details
= { };
3409 details
.check_mapping
= even_cows
? NULL
: mapping
;
3410 details
.first_index
= start
;
3411 details
.last_index
= start
+ nr
- 1;
3412 if (details
.last_index
< details
.first_index
)
3413 details
.last_index
= ULONG_MAX
;
3415 i_mmap_lock_write(mapping
);
3416 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3417 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
3418 i_mmap_unlock_write(mapping
);
3420 EXPORT_SYMBOL_GPL(unmap_mapping_pages
);
3423 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3424 * address_space corresponding to the specified byte range in the underlying
3427 * @mapping: the address space containing mmaps to be unmapped.
3428 * @holebegin: byte in first page to unmap, relative to the start of
3429 * the underlying file. This will be rounded down to a PAGE_SIZE
3430 * boundary. Note that this is different from truncate_pagecache(), which
3431 * must keep the partial page. In contrast, we must get rid of
3433 * @holelen: size of prospective hole in bytes. This will be rounded
3434 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3436 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3437 * but 0 when invalidating pagecache, don't throw away private data.
3439 void unmap_mapping_range(struct address_space
*mapping
,
3440 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
3442 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
3443 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3445 /* Check for overflow. */
3446 if (sizeof(holelen
) > sizeof(hlen
)) {
3448 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3449 if (holeend
& ~(long long)ULONG_MAX
)
3450 hlen
= ULONG_MAX
- hba
+ 1;
3453 unmap_mapping_pages(mapping
, hba
, hlen
, even_cows
);
3455 EXPORT_SYMBOL(unmap_mapping_range
);
3458 * Restore a potential device exclusive pte to a working pte entry
3460 static vm_fault_t
remove_device_exclusive_entry(struct vm_fault
*vmf
)
3462 struct page
*page
= vmf
->page
;
3463 struct vm_area_struct
*vma
= vmf
->vma
;
3464 struct mmu_notifier_range range
;
3466 if (!lock_page_or_retry(page
, vma
->vm_mm
, vmf
->flags
))
3467 return VM_FAULT_RETRY
;
3468 mmu_notifier_range_init_owner(&range
, MMU_NOTIFY_EXCLUSIVE
, 0, vma
,
3469 vma
->vm_mm
, vmf
->address
& PAGE_MASK
,
3470 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
, NULL
);
3471 mmu_notifier_invalidate_range_start(&range
);
3473 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3475 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3476 restore_exclusive_pte(vma
, page
, vmf
->address
, vmf
->pte
);
3478 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3481 mmu_notifier_invalidate_range_end(&range
);
3486 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3487 * but allow concurrent faults), and pte mapped but not yet locked.
3488 * We return with pte unmapped and unlocked.
3490 * We return with the mmap_lock locked or unlocked in the same cases
3491 * as does filemap_fault().
3493 vm_fault_t
do_swap_page(struct vm_fault
*vmf
)
3495 struct vm_area_struct
*vma
= vmf
->vma
;
3496 struct page
*page
= NULL
, *swapcache
;
3497 struct swap_info_struct
*si
= NULL
;
3503 void *shadow
= NULL
;
3505 if (!pte_unmap_same(vma
->vm_mm
, vmf
->pmd
, vmf
->pte
, vmf
->orig_pte
))
3508 entry
= pte_to_swp_entry(vmf
->orig_pte
);
3509 if (unlikely(non_swap_entry(entry
))) {
3510 if (is_migration_entry(entry
)) {
3511 migration_entry_wait(vma
->vm_mm
, vmf
->pmd
,
3513 } else if (is_device_exclusive_entry(entry
)) {
3514 vmf
->page
= pfn_swap_entry_to_page(entry
);
3515 ret
= remove_device_exclusive_entry(vmf
);
3516 } else if (is_device_private_entry(entry
)) {
3517 vmf
->page
= pfn_swap_entry_to_page(entry
);
3518 ret
= vmf
->page
->pgmap
->ops
->migrate_to_ram(vmf
);
3519 } else if (is_hwpoison_entry(entry
)) {
3520 ret
= VM_FAULT_HWPOISON
;
3522 print_bad_pte(vma
, vmf
->address
, vmf
->orig_pte
, NULL
);
3523 ret
= VM_FAULT_SIGBUS
;
3528 /* Prevent swapoff from happening to us. */
3529 si
= get_swap_device(entry
);
3533 delayacct_set_flag(current
, DELAYACCT_PF_SWAPIN
);
3534 page
= lookup_swap_cache(entry
, vma
, vmf
->address
);
3538 if (data_race(si
->flags
& SWP_SYNCHRONOUS_IO
) &&
3539 __swap_count(entry
) == 1) {
3540 /* skip swapcache */
3541 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
3544 __SetPageLocked(page
);
3545 __SetPageSwapBacked(page
);
3547 if (mem_cgroup_swapin_charge_page(page
,
3548 vma
->vm_mm
, GFP_KERNEL
, entry
)) {
3552 mem_cgroup_swapin_uncharge_swap(entry
);
3554 shadow
= get_shadow_from_swap_cache(entry
);
3556 workingset_refault(page
, shadow
);
3558 lru_cache_add(page
);
3560 /* To provide entry to swap_readpage() */
3561 set_page_private(page
, entry
.val
);
3562 swap_readpage(page
, true);
3563 set_page_private(page
, 0);
3566 page
= swapin_readahead(entry
, GFP_HIGHUSER_MOVABLE
,
3573 * Back out if somebody else faulted in this pte
3574 * while we released the pte lock.
3576 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3577 vmf
->address
, &vmf
->ptl
);
3578 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3580 delayacct_clear_flag(current
, DELAYACCT_PF_SWAPIN
);
3584 /* Had to read the page from swap area: Major fault */
3585 ret
= VM_FAULT_MAJOR
;
3586 count_vm_event(PGMAJFAULT
);
3587 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
3588 } else if (PageHWPoison(page
)) {
3590 * hwpoisoned dirty swapcache pages are kept for killing
3591 * owner processes (which may be unknown at hwpoison time)
3593 ret
= VM_FAULT_HWPOISON
;
3594 delayacct_clear_flag(current
, DELAYACCT_PF_SWAPIN
);
3598 locked
= lock_page_or_retry(page
, vma
->vm_mm
, vmf
->flags
);
3600 delayacct_clear_flag(current
, DELAYACCT_PF_SWAPIN
);
3602 ret
|= VM_FAULT_RETRY
;
3607 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3608 * release the swapcache from under us. The page pin, and pte_same
3609 * test below, are not enough to exclude that. Even if it is still
3610 * swapcache, we need to check that the page's swap has not changed.
3612 if (unlikely((!PageSwapCache(page
) ||
3613 page_private(page
) != entry
.val
)) && swapcache
)
3616 page
= ksm_might_need_to_copy(page
, vma
, vmf
->address
);
3617 if (unlikely(!page
)) {
3623 cgroup_throttle_swaprate(page
, GFP_KERNEL
);
3626 * Back out if somebody else already faulted in this pte.
3628 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3630 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3633 if (unlikely(!PageUptodate(page
))) {
3634 ret
= VM_FAULT_SIGBUS
;
3639 * The page isn't present yet, go ahead with the fault.
3641 * Be careful about the sequence of operations here.
3642 * To get its accounting right, reuse_swap_page() must be called
3643 * while the page is counted on swap but not yet in mapcount i.e.
3644 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3645 * must be called after the swap_free(), or it will never succeed.
3648 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3649 dec_mm_counter_fast(vma
->vm_mm
, MM_SWAPENTS
);
3650 pte
= mk_pte(page
, vma
->vm_page_prot
);
3651 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
, NULL
)) {
3652 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3653 vmf
->flags
&= ~FAULT_FLAG_WRITE
;
3654 ret
|= VM_FAULT_WRITE
;
3655 exclusive
= RMAP_EXCLUSIVE
;
3657 flush_icache_page(vma
, page
);
3658 if (pte_swp_soft_dirty(vmf
->orig_pte
))
3659 pte
= pte_mksoft_dirty(pte
);
3660 if (pte_swp_uffd_wp(vmf
->orig_pte
)) {
3661 pte
= pte_mkuffd_wp(pte
);
3662 pte
= pte_wrprotect(pte
);
3664 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3665 arch_do_swap_page(vma
->vm_mm
, vma
, vmf
->address
, pte
, vmf
->orig_pte
);
3666 vmf
->orig_pte
= pte
;
3668 /* ksm created a completely new copy */
3669 if (unlikely(page
!= swapcache
&& swapcache
)) {
3670 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3671 lru_cache_add_inactive_or_unevictable(page
, vma
);
3673 do_page_add_anon_rmap(page
, vma
, vmf
->address
, exclusive
);
3677 if (mem_cgroup_swap_full(page
) ||
3678 (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
3679 try_to_free_swap(page
);
3681 if (page
!= swapcache
&& swapcache
) {
3683 * Hold the lock to avoid the swap entry to be reused
3684 * until we take the PT lock for the pte_same() check
3685 * (to avoid false positives from pte_same). For
3686 * further safety release the lock after the swap_free
3687 * so that the swap count won't change under a
3688 * parallel locked swapcache.
3690 unlock_page(swapcache
);
3691 put_page(swapcache
);
3694 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
3695 ret
|= do_wp_page(vmf
);
3696 if (ret
& VM_FAULT_ERROR
)
3697 ret
&= VM_FAULT_ERROR
;
3701 /* No need to invalidate - it was non-present before */
3702 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3704 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3707 put_swap_device(si
);
3710 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3715 if (page
!= swapcache
&& swapcache
) {
3716 unlock_page(swapcache
);
3717 put_page(swapcache
);
3720 put_swap_device(si
);
3725 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3726 * but allow concurrent faults), and pte mapped but not yet locked.
3727 * We return with mmap_lock still held, but pte unmapped and unlocked.
3729 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
)
3731 struct vm_area_struct
*vma
= vmf
->vma
;
3736 /* File mapping without ->vm_ops ? */
3737 if (vma
->vm_flags
& VM_SHARED
)
3738 return VM_FAULT_SIGBUS
;
3741 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3742 * pte_offset_map() on pmds where a huge pmd might be created
3743 * from a different thread.
3745 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3746 * parallel threads are excluded by other means.
3748 * Here we only have mmap_read_lock(mm).
3750 if (pte_alloc(vma
->vm_mm
, vmf
->pmd
))
3751 return VM_FAULT_OOM
;
3753 /* See comment in handle_pte_fault() */
3754 if (unlikely(pmd_trans_unstable(vmf
->pmd
)))
3757 /* Use the zero-page for reads */
3758 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
3759 !mm_forbids_zeropage(vma
->vm_mm
)) {
3760 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(vmf
->address
),
3761 vma
->vm_page_prot
));
3762 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3763 vmf
->address
, &vmf
->ptl
);
3764 if (!pte_none(*vmf
->pte
)) {
3765 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
3768 ret
= check_stable_address_space(vma
->vm_mm
);
3771 /* Deliver the page fault to userland, check inside PT lock */
3772 if (userfaultfd_missing(vma
)) {
3773 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3774 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3779 /* Allocate our own private page. */
3780 if (unlikely(anon_vma_prepare(vma
)))
3782 page
= alloc_zeroed_user_highpage_movable(vma
, vmf
->address
);
3786 if (mem_cgroup_charge(page
, vma
->vm_mm
, GFP_KERNEL
))
3788 cgroup_throttle_swaprate(page
, GFP_KERNEL
);
3791 * The memory barrier inside __SetPageUptodate makes sure that
3792 * preceding stores to the page contents become visible before
3793 * the set_pte_at() write.
3795 __SetPageUptodate(page
);
3797 entry
= mk_pte(page
, vma
->vm_page_prot
);
3798 entry
= pte_sw_mkyoung(entry
);
3799 if (vma
->vm_flags
& VM_WRITE
)
3800 entry
= pte_mkwrite(pte_mkdirty(entry
));
3802 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3804 if (!pte_none(*vmf
->pte
)) {
3805 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3809 ret
= check_stable_address_space(vma
->vm_mm
);
3813 /* Deliver the page fault to userland, check inside PT lock */
3814 if (userfaultfd_missing(vma
)) {
3815 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3817 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3820 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3821 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3822 lru_cache_add_inactive_or_unevictable(page
, vma
);
3824 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3826 /* No need to invalidate - it was non-present before */
3827 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3829 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3837 return VM_FAULT_OOM
;
3841 * The mmap_lock must have been held on entry, and may have been
3842 * released depending on flags and vma->vm_ops->fault() return value.
3843 * See filemap_fault() and __lock_page_retry().
3845 static vm_fault_t
__do_fault(struct vm_fault
*vmf
)
3847 struct vm_area_struct
*vma
= vmf
->vma
;
3851 * Preallocate pte before we take page_lock because this might lead to
3852 * deadlocks for memcg reclaim which waits for pages under writeback:
3854 * SetPageWriteback(A)
3860 * wait_on_page_writeback(A)
3861 * SetPageWriteback(B)
3863 * # flush A, B to clear the writeback
3865 if (pmd_none(*vmf
->pmd
) && !vmf
->prealloc_pte
) {
3866 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
3867 if (!vmf
->prealloc_pte
)
3868 return VM_FAULT_OOM
;
3869 smp_wmb(); /* See comment in __pte_alloc() */
3872 ret
= vma
->vm_ops
->fault(vmf
);
3873 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
|
3874 VM_FAULT_DONE_COW
)))
3877 if (unlikely(PageHWPoison(vmf
->page
))) {
3878 struct page
*page
= vmf
->page
;
3879 vm_fault_t poisonret
= VM_FAULT_HWPOISON
;
3880 if (ret
& VM_FAULT_LOCKED
) {
3881 if (page_mapped(page
))
3882 unmap_mapping_pages(page_mapping(page
),
3883 page
->index
, 1, false);
3884 /* Retry if a clean page was removed from the cache. */
3885 if (invalidate_inode_page(page
))
3886 poisonret
= VM_FAULT_NOPAGE
;
3894 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3895 lock_page(vmf
->page
);
3897 VM_BUG_ON_PAGE(!PageLocked(vmf
->page
), vmf
->page
);
3902 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3903 static void deposit_prealloc_pte(struct vm_fault
*vmf
)
3905 struct vm_area_struct
*vma
= vmf
->vma
;
3907 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3909 * We are going to consume the prealloc table,
3910 * count that as nr_ptes.
3912 mm_inc_nr_ptes(vma
->vm_mm
);
3913 vmf
->prealloc_pte
= NULL
;
3916 vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3918 struct vm_area_struct
*vma
= vmf
->vma
;
3919 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3920 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
3923 vm_fault_t ret
= VM_FAULT_FALLBACK
;
3925 if (!transhuge_vma_suitable(vma
, haddr
))
3928 page
= compound_head(page
);
3929 if (compound_order(page
) != HPAGE_PMD_ORDER
)
3933 * Just backoff if any subpage of a THP is corrupted otherwise
3934 * the corrupted page may mapped by PMD silently to escape the
3935 * check. This kind of THP just can be PTE mapped. Access to
3936 * the corrupted subpage should trigger SIGBUS as expected.
3938 if (unlikely(PageHasHWPoisoned(page
)))
3942 * Archs like ppc64 need additional space to store information
3943 * related to pte entry. Use the preallocated table for that.
3945 if (arch_needs_pgtable_deposit() && !vmf
->prealloc_pte
) {
3946 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
3947 if (!vmf
->prealloc_pte
)
3948 return VM_FAULT_OOM
;
3949 smp_wmb(); /* See comment in __pte_alloc() */
3952 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3953 if (unlikely(!pmd_none(*vmf
->pmd
)))
3956 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
3957 flush_icache_page(vma
, page
+ i
);
3959 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
3961 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
3963 add_mm_counter(vma
->vm_mm
, mm_counter_file(page
), HPAGE_PMD_NR
);
3964 page_add_file_rmap(page
, true);
3966 * deposit and withdraw with pmd lock held
3968 if (arch_needs_pgtable_deposit())
3969 deposit_prealloc_pte(vmf
);
3971 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
3973 update_mmu_cache_pmd(vma
, haddr
, vmf
->pmd
);
3975 /* fault is handled */
3977 count_vm_event(THP_FILE_MAPPED
);
3979 spin_unlock(vmf
->ptl
);
3983 vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3985 return VM_FAULT_FALLBACK
;
3989 void do_set_pte(struct vm_fault
*vmf
, struct page
*page
, unsigned long addr
)
3991 struct vm_area_struct
*vma
= vmf
->vma
;
3992 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3993 bool prefault
= vmf
->address
!= addr
;
3996 flush_icache_page(vma
, page
);
3997 entry
= mk_pte(page
, vma
->vm_page_prot
);
3999 if (prefault
&& arch_wants_old_prefaulted_pte())
4000 entry
= pte_mkold(entry
);
4002 entry
= pte_sw_mkyoung(entry
);
4005 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
4006 /* copy-on-write page */
4007 if (write
&& !(vma
->vm_flags
& VM_SHARED
)) {
4008 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
4009 page_add_new_anon_rmap(page
, vma
, addr
, false);
4010 lru_cache_add_inactive_or_unevictable(page
, vma
);
4012 inc_mm_counter_fast(vma
->vm_mm
, mm_counter_file(page
));
4013 page_add_file_rmap(page
, false);
4015 set_pte_at(vma
->vm_mm
, addr
, vmf
->pte
, entry
);
4019 * finish_fault - finish page fault once we have prepared the page to fault
4021 * @vmf: structure describing the fault
4023 * This function handles all that is needed to finish a page fault once the
4024 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4025 * given page, adds reverse page mapping, handles memcg charges and LRU
4028 * The function expects the page to be locked and on success it consumes a
4029 * reference of a page being mapped (for the PTE which maps it).
4031 * Return: %0 on success, %VM_FAULT_ code in case of error.
4033 vm_fault_t
finish_fault(struct vm_fault
*vmf
)
4035 struct vm_area_struct
*vma
= vmf
->vma
;
4039 /* Did we COW the page? */
4040 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
4041 page
= vmf
->cow_page
;
4046 * check even for read faults because we might have lost our CoWed
4049 if (!(vma
->vm_flags
& VM_SHARED
)) {
4050 ret
= check_stable_address_space(vma
->vm_mm
);
4055 if (pmd_none(*vmf
->pmd
)) {
4056 if (PageTransCompound(page
)) {
4057 ret
= do_set_pmd(vmf
, page
);
4058 if (ret
!= VM_FAULT_FALLBACK
)
4062 if (vmf
->prealloc_pte
) {
4063 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
4064 if (likely(pmd_none(*vmf
->pmd
))) {
4065 mm_inc_nr_ptes(vma
->vm_mm
);
4066 pmd_populate(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
4067 vmf
->prealloc_pte
= NULL
;
4069 spin_unlock(vmf
->ptl
);
4070 } else if (unlikely(pte_alloc(vma
->vm_mm
, vmf
->pmd
))) {
4071 return VM_FAULT_OOM
;
4075 /* See comment in handle_pte_fault() */
4076 if (pmd_devmap_trans_unstable(vmf
->pmd
))
4079 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
4080 vmf
->address
, &vmf
->ptl
);
4082 /* Re-check under ptl */
4083 if (likely(pte_none(*vmf
->pte
)))
4084 do_set_pte(vmf
, page
, vmf
->address
);
4086 ret
= VM_FAULT_NOPAGE
;
4088 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
4089 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4093 static unsigned long fault_around_bytes __read_mostly
=
4094 rounddown_pow_of_two(65536);
4096 #ifdef CONFIG_DEBUG_FS
4097 static int fault_around_bytes_get(void *data
, u64
*val
)
4099 *val
= fault_around_bytes
;
4104 * fault_around_bytes must be rounded down to the nearest page order as it's
4105 * what do_fault_around() expects to see.
4107 static int fault_around_bytes_set(void *data
, u64 val
)
4109 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
4111 if (val
> PAGE_SIZE
)
4112 fault_around_bytes
= rounddown_pow_of_two(val
);
4114 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
4117 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops
,
4118 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
4120 static int __init
fault_around_debugfs(void)
4122 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL
, NULL
,
4123 &fault_around_bytes_fops
);
4126 late_initcall(fault_around_debugfs
);
4130 * do_fault_around() tries to map few pages around the fault address. The hope
4131 * is that the pages will be needed soon and this will lower the number of
4134 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4135 * not ready to be mapped: not up-to-date, locked, etc.
4137 * This function is called with the page table lock taken. In the split ptlock
4138 * case the page table lock only protects only those entries which belong to
4139 * the page table corresponding to the fault address.
4141 * This function doesn't cross the VMA boundaries, in order to call map_pages()
4144 * fault_around_bytes defines how many bytes we'll try to map.
4145 * do_fault_around() expects it to be set to a power of two less than or equal
4148 * The virtual address of the area that we map is naturally aligned to
4149 * fault_around_bytes rounded down to the machine page size
4150 * (and therefore to page order). This way it's easier to guarantee
4151 * that we don't cross page table boundaries.
4153 static vm_fault_t
do_fault_around(struct vm_fault
*vmf
)
4155 unsigned long address
= vmf
->address
, nr_pages
, mask
;
4156 pgoff_t start_pgoff
= vmf
->pgoff
;
4160 nr_pages
= READ_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
4161 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
4163 address
= max(address
& mask
, vmf
->vma
->vm_start
);
4164 off
= ((vmf
->address
- address
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
4168 * end_pgoff is either the end of the page table, the end of
4169 * the vma or nr_pages from start_pgoff, depending what is nearest.
4171 end_pgoff
= start_pgoff
-
4172 ((address
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
4174 end_pgoff
= min3(end_pgoff
, vma_pages(vmf
->vma
) + vmf
->vma
->vm_pgoff
- 1,
4175 start_pgoff
+ nr_pages
- 1);
4177 if (pmd_none(*vmf
->pmd
)) {
4178 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
);
4179 if (!vmf
->prealloc_pte
)
4180 return VM_FAULT_OOM
;
4181 smp_wmb(); /* See comment in __pte_alloc() */
4184 return vmf
->vma
->vm_ops
->map_pages(vmf
, start_pgoff
, end_pgoff
);
4187 static vm_fault_t
do_read_fault(struct vm_fault
*vmf
)
4189 struct vm_area_struct
*vma
= vmf
->vma
;
4193 * Let's call ->map_pages() first and use ->fault() as fallback
4194 * if page by the offset is not ready to be mapped (cold cache or
4197 if (vma
->vm_ops
->map_pages
&& fault_around_bytes
>> PAGE_SHIFT
> 1) {
4198 if (likely(!userfaultfd_minor(vmf
->vma
))) {
4199 ret
= do_fault_around(vmf
);
4205 ret
= __do_fault(vmf
);
4206 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4209 ret
|= finish_fault(vmf
);
4210 unlock_page(vmf
->page
);
4211 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4212 put_page(vmf
->page
);
4216 static vm_fault_t
do_cow_fault(struct vm_fault
*vmf
)
4218 struct vm_area_struct
*vma
= vmf
->vma
;
4221 if (unlikely(anon_vma_prepare(vma
)))
4222 return VM_FAULT_OOM
;
4224 vmf
->cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, vmf
->address
);
4226 return VM_FAULT_OOM
;
4228 if (mem_cgroup_charge(vmf
->cow_page
, vma
->vm_mm
, GFP_KERNEL
)) {
4229 put_page(vmf
->cow_page
);
4230 return VM_FAULT_OOM
;
4232 cgroup_throttle_swaprate(vmf
->cow_page
, GFP_KERNEL
);
4234 ret
= __do_fault(vmf
);
4235 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4237 if (ret
& VM_FAULT_DONE_COW
)
4240 copy_user_highpage(vmf
->cow_page
, vmf
->page
, vmf
->address
, vma
);
4241 __SetPageUptodate(vmf
->cow_page
);
4243 ret
|= finish_fault(vmf
);
4244 unlock_page(vmf
->page
);
4245 put_page(vmf
->page
);
4246 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4250 put_page(vmf
->cow_page
);
4254 static vm_fault_t
do_shared_fault(struct vm_fault
*vmf
)
4256 struct vm_area_struct
*vma
= vmf
->vma
;
4257 vm_fault_t ret
, tmp
;
4259 ret
= __do_fault(vmf
);
4260 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4264 * Check if the backing address space wants to know that the page is
4265 * about to become writable
4267 if (vma
->vm_ops
->page_mkwrite
) {
4268 unlock_page(vmf
->page
);
4269 tmp
= do_page_mkwrite(vmf
);
4270 if (unlikely(!tmp
||
4271 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
4272 put_page(vmf
->page
);
4277 ret
|= finish_fault(vmf
);
4278 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
4280 unlock_page(vmf
->page
);
4281 put_page(vmf
->page
);
4285 ret
|= fault_dirty_shared_page(vmf
);
4290 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4291 * but allow concurrent faults).
4292 * The mmap_lock may have been released depending on flags and our
4293 * return value. See filemap_fault() and __lock_page_or_retry().
4294 * If mmap_lock is released, vma may become invalid (for example
4295 * by other thread calling munmap()).
4297 static vm_fault_t
do_fault(struct vm_fault
*vmf
)
4299 struct vm_area_struct
*vma
= vmf
->vma
;
4300 struct mm_struct
*vm_mm
= vma
->vm_mm
;
4304 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4306 if (!vma
->vm_ops
->fault
) {
4308 * If we find a migration pmd entry or a none pmd entry, which
4309 * should never happen, return SIGBUS
4311 if (unlikely(!pmd_present(*vmf
->pmd
)))
4312 ret
= VM_FAULT_SIGBUS
;
4314 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
,
4319 * Make sure this is not a temporary clearing of pte
4320 * by holding ptl and checking again. A R/M/W update
4321 * of pte involves: take ptl, clearing the pte so that
4322 * we don't have concurrent modification by hardware
4323 * followed by an update.
4325 if (unlikely(pte_none(*vmf
->pte
)))
4326 ret
= VM_FAULT_SIGBUS
;
4328 ret
= VM_FAULT_NOPAGE
;
4330 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4332 } else if (!(vmf
->flags
& FAULT_FLAG_WRITE
))
4333 ret
= do_read_fault(vmf
);
4334 else if (!(vma
->vm_flags
& VM_SHARED
))
4335 ret
= do_cow_fault(vmf
);
4337 ret
= do_shared_fault(vmf
);
4339 /* preallocated pagetable is unused: free it */
4340 if (vmf
->prealloc_pte
) {
4341 pte_free(vm_mm
, vmf
->prealloc_pte
);
4342 vmf
->prealloc_pte
= NULL
;
4347 int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
4348 unsigned long addr
, int page_nid
, int *flags
)
4352 count_vm_numa_event(NUMA_HINT_FAULTS
);
4353 if (page_nid
== numa_node_id()) {
4354 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
4355 *flags
|= TNF_FAULT_LOCAL
;
4358 return mpol_misplaced(page
, vma
, addr
);
4361 static vm_fault_t
do_numa_page(struct vm_fault
*vmf
)
4363 struct vm_area_struct
*vma
= vmf
->vma
;
4364 struct page
*page
= NULL
;
4365 int page_nid
= NUMA_NO_NODE
;
4369 bool was_writable
= pte_savedwrite(vmf
->orig_pte
);
4373 * The "pte" at this point cannot be used safely without
4374 * validation through pte_unmap_same(). It's of NUMA type but
4375 * the pfn may be screwed if the read is non atomic.
4377 vmf
->ptl
= pte_lockptr(vma
->vm_mm
, vmf
->pmd
);
4378 spin_lock(vmf
->ptl
);
4379 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
4380 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4384 /* Get the normal PTE */
4385 old_pte
= ptep_get(vmf
->pte
);
4386 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
4388 page
= vm_normal_page(vma
, vmf
->address
, pte
);
4392 /* TODO: handle PTE-mapped THP */
4393 if (PageCompound(page
))
4397 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4398 * much anyway since they can be in shared cache state. This misses
4399 * the case where a mapping is writable but the process never writes
4400 * to it but pte_write gets cleared during protection updates and
4401 * pte_dirty has unpredictable behaviour between PTE scan updates,
4402 * background writeback, dirty balancing and application behaviour.
4405 flags
|= TNF_NO_GROUP
;
4408 * Flag if the page is shared between multiple address spaces. This
4409 * is later used when determining whether to group tasks together
4411 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
4412 flags
|= TNF_SHARED
;
4414 last_cpupid
= page_cpupid_last(page
);
4415 page_nid
= page_to_nid(page
);
4416 target_nid
= numa_migrate_prep(page
, vma
, vmf
->address
, page_nid
,
4418 if (target_nid
== NUMA_NO_NODE
) {
4422 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4424 /* Migrate to the requested node */
4425 if (migrate_misplaced_page(page
, vma
, target_nid
)) {
4426 page_nid
= target_nid
;
4427 flags
|= TNF_MIGRATED
;
4429 flags
|= TNF_MIGRATE_FAIL
;
4430 vmf
->pte
= pte_offset_map(vmf
->pmd
, vmf
->address
);
4431 spin_lock(vmf
->ptl
);
4432 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
4433 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4440 if (page_nid
!= NUMA_NO_NODE
)
4441 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
4445 * Make it present again, depending on how arch implements
4446 * non-accessible ptes, some can allow access by kernel mode.
4448 old_pte
= ptep_modify_prot_start(vma
, vmf
->address
, vmf
->pte
);
4449 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
4450 pte
= pte_mkyoung(pte
);
4452 pte
= pte_mkwrite(pte
);
4453 ptep_modify_prot_commit(vma
, vmf
->address
, vmf
->pte
, old_pte
, pte
);
4454 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
4455 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4459 static inline vm_fault_t
create_huge_pmd(struct vm_fault
*vmf
)
4461 if (vma_is_anonymous(vmf
->vma
))
4462 return do_huge_pmd_anonymous_page(vmf
);
4463 if (vmf
->vma
->vm_ops
->huge_fault
)
4464 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
4465 return VM_FAULT_FALLBACK
;
4468 /* `inline' is required to avoid gcc 4.1.2 build error */
4469 static inline vm_fault_t
wp_huge_pmd(struct vm_fault
*vmf
)
4471 if (vma_is_anonymous(vmf
->vma
)) {
4472 if (userfaultfd_huge_pmd_wp(vmf
->vma
, vmf
->orig_pmd
))
4473 return handle_userfault(vmf
, VM_UFFD_WP
);
4474 return do_huge_pmd_wp_page(vmf
);
4476 if (vmf
->vma
->vm_ops
->huge_fault
) {
4477 vm_fault_t ret
= vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
4479 if (!(ret
& VM_FAULT_FALLBACK
))
4483 /* COW or write-notify handled on pte level: split pmd. */
4484 __split_huge_pmd(vmf
->vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
4486 return VM_FAULT_FALLBACK
;
4489 static vm_fault_t
create_huge_pud(struct vm_fault
*vmf
)
4491 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4492 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4493 /* No support for anonymous transparent PUD pages yet */
4494 if (vma_is_anonymous(vmf
->vma
))
4496 if (vmf
->vma
->vm_ops
->huge_fault
) {
4497 vm_fault_t ret
= vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
4499 if (!(ret
& VM_FAULT_FALLBACK
))
4503 /* COW or write-notify not handled on PUD level: split pud.*/
4504 __split_huge_pud(vmf
->vma
, vmf
->pud
, vmf
->address
);
4505 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4506 return VM_FAULT_FALLBACK
;
4509 static vm_fault_t
wp_huge_pud(struct vm_fault
*vmf
, pud_t orig_pud
)
4511 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4512 /* No support for anonymous transparent PUD pages yet */
4513 if (vma_is_anonymous(vmf
->vma
))
4514 return VM_FAULT_FALLBACK
;
4515 if (vmf
->vma
->vm_ops
->huge_fault
)
4516 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
4517 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4518 return VM_FAULT_FALLBACK
;
4522 * These routines also need to handle stuff like marking pages dirty
4523 * and/or accessed for architectures that don't do it in hardware (most
4524 * RISC architectures). The early dirtying is also good on the i386.
4526 * There is also a hook called "update_mmu_cache()" that architectures
4527 * with external mmu caches can use to update those (ie the Sparc or
4528 * PowerPC hashed page tables that act as extended TLBs).
4530 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4531 * concurrent faults).
4533 * The mmap_lock may have been released depending on flags and our return value.
4534 * See filemap_fault() and __lock_page_or_retry().
4536 static vm_fault_t
handle_pte_fault(struct vm_fault
*vmf
)
4540 if (unlikely(pmd_none(*vmf
->pmd
))) {
4542 * Leave __pte_alloc() until later: because vm_ops->fault may
4543 * want to allocate huge page, and if we expose page table
4544 * for an instant, it will be difficult to retract from
4545 * concurrent faults and from rmap lookups.
4550 * If a huge pmd materialized under us just retry later. Use
4551 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead
4552 * of pmd_trans_huge() to ensure the pmd didn't become
4553 * pmd_trans_huge under us and then back to pmd_none, as a
4554 * result of MADV_DONTNEED running immediately after a huge pmd
4555 * fault in a different thread of this mm, in turn leading to a
4556 * misleading pmd_trans_huge() retval. All we have to ensure is
4557 * that it is a regular pmd that we can walk with
4558 * pte_offset_map() and we can do that through an atomic read
4559 * in C, which is what pmd_trans_unstable() provides.
4561 if (pmd_devmap_trans_unstable(vmf
->pmd
))
4564 * A regular pmd is established and it can't morph into a huge
4565 * pmd from under us anymore at this point because we hold the
4566 * mmap_lock read mode and khugepaged takes it in write mode.
4567 * So now it's safe to run pte_offset_map().
4569 vmf
->pte
= pte_offset_map(vmf
->pmd
, vmf
->address
);
4570 vmf
->orig_pte
= *vmf
->pte
;
4573 * some architectures can have larger ptes than wordsize,
4574 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4575 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4576 * accesses. The code below just needs a consistent view
4577 * for the ifs and we later double check anyway with the
4578 * ptl lock held. So here a barrier will do.
4581 if (pte_none(vmf
->orig_pte
)) {
4582 pte_unmap(vmf
->pte
);
4588 if (vma_is_anonymous(vmf
->vma
))
4589 return do_anonymous_page(vmf
);
4591 return do_fault(vmf
);
4594 if (!pte_present(vmf
->orig_pte
))
4595 return do_swap_page(vmf
);
4597 if (pte_protnone(vmf
->orig_pte
) && vma_is_accessible(vmf
->vma
))
4598 return do_numa_page(vmf
);
4600 vmf
->ptl
= pte_lockptr(vmf
->vma
->vm_mm
, vmf
->pmd
);
4601 spin_lock(vmf
->ptl
);
4602 entry
= vmf
->orig_pte
;
4603 if (unlikely(!pte_same(*vmf
->pte
, entry
))) {
4604 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
4607 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4608 if (!pte_write(entry
))
4609 return do_wp_page(vmf
);
4610 entry
= pte_mkdirty(entry
);
4612 entry
= pte_mkyoung(entry
);
4613 if (ptep_set_access_flags(vmf
->vma
, vmf
->address
, vmf
->pte
, entry
,
4614 vmf
->flags
& FAULT_FLAG_WRITE
)) {
4615 update_mmu_cache(vmf
->vma
, vmf
->address
, vmf
->pte
);
4617 /* Skip spurious TLB flush for retried page fault */
4618 if (vmf
->flags
& FAULT_FLAG_TRIED
)
4621 * This is needed only for protection faults but the arch code
4622 * is not yet telling us if this is a protection fault or not.
4623 * This still avoids useless tlb flushes for .text page faults
4626 if (vmf
->flags
& FAULT_FLAG_WRITE
)
4627 flush_tlb_fix_spurious_fault(vmf
->vma
, vmf
->address
);
4630 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4635 * By the time we get here, we already hold the mm semaphore
4637 * The mmap_lock may have been released depending on flags and our
4638 * return value. See filemap_fault() and __lock_page_or_retry().
4640 static vm_fault_t
__handle_mm_fault(struct vm_area_struct
*vma
,
4641 unsigned long address
, unsigned int flags
)
4643 struct vm_fault vmf
= {
4645 .address
= address
& PAGE_MASK
,
4647 .pgoff
= linear_page_index(vma
, address
),
4648 .gfp_mask
= __get_fault_gfp_mask(vma
),
4650 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
4651 struct mm_struct
*mm
= vma
->vm_mm
;
4656 pgd
= pgd_offset(mm
, address
);
4657 p4d
= p4d_alloc(mm
, pgd
, address
);
4659 return VM_FAULT_OOM
;
4661 vmf
.pud
= pud_alloc(mm
, p4d
, address
);
4663 return VM_FAULT_OOM
;
4665 if (pud_none(*vmf
.pud
) && __transparent_hugepage_enabled(vma
)) {
4666 ret
= create_huge_pud(&vmf
);
4667 if (!(ret
& VM_FAULT_FALLBACK
))
4670 pud_t orig_pud
= *vmf
.pud
;
4673 if (pud_trans_huge(orig_pud
) || pud_devmap(orig_pud
)) {
4675 /* NUMA case for anonymous PUDs would go here */
4677 if (dirty
&& !pud_write(orig_pud
)) {
4678 ret
= wp_huge_pud(&vmf
, orig_pud
);
4679 if (!(ret
& VM_FAULT_FALLBACK
))
4682 huge_pud_set_accessed(&vmf
, orig_pud
);
4688 vmf
.pmd
= pmd_alloc(mm
, vmf
.pud
, address
);
4690 return VM_FAULT_OOM
;
4692 /* Huge pud page fault raced with pmd_alloc? */
4693 if (pud_trans_unstable(vmf
.pud
))
4696 if (pmd_none(*vmf
.pmd
) && __transparent_hugepage_enabled(vma
)) {
4697 ret
= create_huge_pmd(&vmf
);
4698 if (!(ret
& VM_FAULT_FALLBACK
))
4701 vmf
.orig_pmd
= *vmf
.pmd
;
4704 if (unlikely(is_swap_pmd(vmf
.orig_pmd
))) {
4705 VM_BUG_ON(thp_migration_supported() &&
4706 !is_pmd_migration_entry(vmf
.orig_pmd
));
4707 if (is_pmd_migration_entry(vmf
.orig_pmd
))
4708 pmd_migration_entry_wait(mm
, vmf
.pmd
);
4711 if (pmd_trans_huge(vmf
.orig_pmd
) || pmd_devmap(vmf
.orig_pmd
)) {
4712 if (pmd_protnone(vmf
.orig_pmd
) && vma_is_accessible(vma
))
4713 return do_huge_pmd_numa_page(&vmf
);
4715 if (dirty
&& !pmd_write(vmf
.orig_pmd
)) {
4716 ret
= wp_huge_pmd(&vmf
);
4717 if (!(ret
& VM_FAULT_FALLBACK
))
4720 huge_pmd_set_accessed(&vmf
);
4726 return handle_pte_fault(&vmf
);
4730 * mm_account_fault - Do page fault accounting
4732 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4733 * of perf event counters, but we'll still do the per-task accounting to
4734 * the task who triggered this page fault.
4735 * @address: the faulted address.
4736 * @flags: the fault flags.
4737 * @ret: the fault retcode.
4739 * This will take care of most of the page fault accounting. Meanwhile, it
4740 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4741 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4742 * still be in per-arch page fault handlers at the entry of page fault.
4744 static inline void mm_account_fault(struct pt_regs
*regs
,
4745 unsigned long address
, unsigned int flags
,
4751 * We don't do accounting for some specific faults:
4753 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4754 * includes arch_vma_access_permitted() failing before reaching here.
4755 * So this is not a "this many hardware page faults" counter. We
4756 * should use the hw profiling for that.
4758 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4759 * once they're completed.
4761 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_RETRY
))
4765 * We define the fault as a major fault when the final successful fault
4766 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4767 * handle it immediately previously).
4769 major
= (ret
& VM_FAULT_MAJOR
) || (flags
& FAULT_FLAG_TRIED
);
4777 * If the fault is done for GUP, regs will be NULL. We only do the
4778 * accounting for the per thread fault counters who triggered the
4779 * fault, and we skip the perf event updates.
4785 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ
, 1, regs
, address
);
4787 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN
, 1, regs
, address
);
4791 * By the time we get here, we already hold the mm semaphore
4793 * The mmap_lock may have been released depending on flags and our
4794 * return value. See filemap_fault() and __lock_page_or_retry().
4796 vm_fault_t
handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
4797 unsigned int flags
, struct pt_regs
*regs
)
4801 __set_current_state(TASK_RUNNING
);
4803 count_vm_event(PGFAULT
);
4804 count_memcg_event_mm(vma
->vm_mm
, PGFAULT
);
4806 /* do counter updates before entering really critical section. */
4807 check_sync_rss_stat(current
);
4809 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
4810 flags
& FAULT_FLAG_INSTRUCTION
,
4811 flags
& FAULT_FLAG_REMOTE
))
4812 return VM_FAULT_SIGSEGV
;
4815 * Enable the memcg OOM handling for faults triggered in user
4816 * space. Kernel faults are handled more gracefully.
4818 if (flags
& FAULT_FLAG_USER
)
4819 mem_cgroup_enter_user_fault();
4821 if (unlikely(is_vm_hugetlb_page(vma
)))
4822 ret
= hugetlb_fault(vma
->vm_mm
, vma
, address
, flags
);
4824 ret
= __handle_mm_fault(vma
, address
, flags
);
4826 if (flags
& FAULT_FLAG_USER
) {
4827 mem_cgroup_exit_user_fault();
4829 * The task may have entered a memcg OOM situation but
4830 * if the allocation error was handled gracefully (no
4831 * VM_FAULT_OOM), there is no need to kill anything.
4832 * Just clean up the OOM state peacefully.
4834 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
4835 mem_cgroup_oom_synchronize(false);
4838 mm_account_fault(regs
, address
, flags
, ret
);
4842 EXPORT_SYMBOL_GPL(handle_mm_fault
);
4844 #ifndef __PAGETABLE_P4D_FOLDED
4846 * Allocate p4d page table.
4847 * We've already handled the fast-path in-line.
4849 int __p4d_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
4851 p4d_t
*new = p4d_alloc_one(mm
, address
);
4855 smp_wmb(); /* See comment in __pte_alloc */
4857 spin_lock(&mm
->page_table_lock
);
4858 if (pgd_present(*pgd
)) /* Another has populated it */
4861 pgd_populate(mm
, pgd
, new);
4862 spin_unlock(&mm
->page_table_lock
);
4865 #endif /* __PAGETABLE_P4D_FOLDED */
4867 #ifndef __PAGETABLE_PUD_FOLDED
4869 * Allocate page upper directory.
4870 * We've already handled the fast-path in-line.
4872 int __pud_alloc(struct mm_struct
*mm
, p4d_t
*p4d
, unsigned long address
)
4874 pud_t
*new = pud_alloc_one(mm
, address
);
4878 smp_wmb(); /* See comment in __pte_alloc */
4880 spin_lock(&mm
->page_table_lock
);
4881 if (!p4d_present(*p4d
)) {
4883 p4d_populate(mm
, p4d
, new);
4884 } else /* Another has populated it */
4886 spin_unlock(&mm
->page_table_lock
);
4889 #endif /* __PAGETABLE_PUD_FOLDED */
4891 #ifndef __PAGETABLE_PMD_FOLDED
4893 * Allocate page middle directory.
4894 * We've already handled the fast-path in-line.
4896 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
4899 pmd_t
*new = pmd_alloc_one(mm
, address
);
4903 smp_wmb(); /* See comment in __pte_alloc */
4905 ptl
= pud_lock(mm
, pud
);
4906 if (!pud_present(*pud
)) {
4908 pud_populate(mm
, pud
, new);
4909 } else /* Another has populated it */
4914 #endif /* __PAGETABLE_PMD_FOLDED */
4916 int follow_invalidate_pte(struct mm_struct
*mm
, unsigned long address
,
4917 struct mmu_notifier_range
*range
, pte_t
**ptepp
,
4918 pmd_t
**pmdpp
, spinlock_t
**ptlp
)
4926 pgd
= pgd_offset(mm
, address
);
4927 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
4930 p4d
= p4d_offset(pgd
, address
);
4931 if (p4d_none(*p4d
) || unlikely(p4d_bad(*p4d
)))
4934 pud
= pud_offset(p4d
, address
);
4935 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
4938 pmd
= pmd_offset(pud
, address
);
4939 VM_BUG_ON(pmd_trans_huge(*pmd
));
4941 if (pmd_huge(*pmd
)) {
4946 mmu_notifier_range_init(range
, MMU_NOTIFY_CLEAR
, 0,
4947 NULL
, mm
, address
& PMD_MASK
,
4948 (address
& PMD_MASK
) + PMD_SIZE
);
4949 mmu_notifier_invalidate_range_start(range
);
4951 *ptlp
= pmd_lock(mm
, pmd
);
4952 if (pmd_huge(*pmd
)) {
4958 mmu_notifier_invalidate_range_end(range
);
4961 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
4965 mmu_notifier_range_init(range
, MMU_NOTIFY_CLEAR
, 0, NULL
, mm
,
4966 address
& PAGE_MASK
,
4967 (address
& PAGE_MASK
) + PAGE_SIZE
);
4968 mmu_notifier_invalidate_range_start(range
);
4970 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
4971 if (!pte_present(*ptep
))
4976 pte_unmap_unlock(ptep
, *ptlp
);
4978 mmu_notifier_invalidate_range_end(range
);
4984 * follow_pte - look up PTE at a user virtual address
4985 * @mm: the mm_struct of the target address space
4986 * @address: user virtual address
4987 * @ptepp: location to store found PTE
4988 * @ptlp: location to store the lock for the PTE
4990 * On a successful return, the pointer to the PTE is stored in @ptepp;
4991 * the corresponding lock is taken and its location is stored in @ptlp.
4992 * The contents of the PTE are only stable until @ptlp is released;
4993 * any further use, if any, must be protected against invalidation
4994 * with MMU notifiers.
4996 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
4997 * should be taken for read.
4999 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
5000 * it is not a good general-purpose API.
5002 * Return: zero on success, -ve otherwise.
5004 int follow_pte(struct mm_struct
*mm
, unsigned long address
,
5005 pte_t
**ptepp
, spinlock_t
**ptlp
)
5007 return follow_invalidate_pte(mm
, address
, NULL
, ptepp
, NULL
, ptlp
);
5009 EXPORT_SYMBOL_GPL(follow_pte
);
5012 * follow_pfn - look up PFN at a user virtual address
5013 * @vma: memory mapping
5014 * @address: user virtual address
5015 * @pfn: location to store found PFN
5017 * Only IO mappings and raw PFN mappings are allowed.
5019 * This function does not allow the caller to read the permissions
5020 * of the PTE. Do not use it.
5022 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5024 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
5031 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
5034 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
5037 *pfn
= pte_pfn(*ptep
);
5038 pte_unmap_unlock(ptep
, ptl
);
5041 EXPORT_SYMBOL(follow_pfn
);
5043 #ifdef CONFIG_HAVE_IOREMAP_PROT
5044 int follow_phys(struct vm_area_struct
*vma
,
5045 unsigned long address
, unsigned int flags
,
5046 unsigned long *prot
, resource_size_t
*phys
)
5052 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
5055 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
5059 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
5062 *prot
= pgprot_val(pte_pgprot(pte
));
5063 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
5067 pte_unmap_unlock(ptep
, ptl
);
5073 * generic_access_phys - generic implementation for iomem mmap access
5074 * @vma: the vma to access
5075 * @addr: userspace address, not relative offset within @vma
5076 * @buf: buffer to read/write
5077 * @len: length of transfer
5078 * @write: set to FOLL_WRITE when writing, otherwise reading
5080 * This is a generic implementation for &vm_operations_struct.access for an
5081 * iomem mapping. This callback is used by access_process_vm() when the @vma is
5084 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
5085 void *buf
, int len
, int write
)
5087 resource_size_t phys_addr
;
5088 unsigned long prot
= 0;
5089 void __iomem
*maddr
;
5092 int offset
= offset_in_page(addr
);
5095 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
5099 if (follow_pte(vma
->vm_mm
, addr
, &ptep
, &ptl
))
5102 pte_unmap_unlock(ptep
, ptl
);
5104 prot
= pgprot_val(pte_pgprot(pte
));
5105 phys_addr
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
5107 if ((write
& FOLL_WRITE
) && !pte_write(pte
))
5110 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
5114 if (follow_pte(vma
->vm_mm
, addr
, &ptep
, &ptl
))
5117 if (!pte_same(pte
, *ptep
)) {
5118 pte_unmap_unlock(ptep
, ptl
);
5125 memcpy_toio(maddr
+ offset
, buf
, len
);
5127 memcpy_fromio(buf
, maddr
+ offset
, len
);
5129 pte_unmap_unlock(ptep
, ptl
);
5135 EXPORT_SYMBOL_GPL(generic_access_phys
);
5139 * Access another process' address space as given in mm.
5141 int __access_remote_vm(struct mm_struct
*mm
, unsigned long addr
, void *buf
,
5142 int len
, unsigned int gup_flags
)
5144 struct vm_area_struct
*vma
;
5145 void *old_buf
= buf
;
5146 int write
= gup_flags
& FOLL_WRITE
;
5148 if (mmap_read_lock_killable(mm
))
5151 /* ignore errors, just check how much was successfully transferred */
5153 int bytes
, ret
, offset
;
5155 struct page
*page
= NULL
;
5157 ret
= get_user_pages_remote(mm
, addr
, 1,
5158 gup_flags
, &page
, &vma
, NULL
);
5160 #ifndef CONFIG_HAVE_IOREMAP_PROT
5164 * Check if this is a VM_IO | VM_PFNMAP VMA, which
5165 * we can access using slightly different code.
5167 vma
= vma_lookup(mm
, addr
);
5170 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
5171 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
5179 offset
= addr
& (PAGE_SIZE
-1);
5180 if (bytes
> PAGE_SIZE
-offset
)
5181 bytes
= PAGE_SIZE
-offset
;
5185 copy_to_user_page(vma
, page
, addr
,
5186 maddr
+ offset
, buf
, bytes
);
5187 set_page_dirty_lock(page
);
5189 copy_from_user_page(vma
, page
, addr
,
5190 buf
, maddr
+ offset
, bytes
);
5199 mmap_read_unlock(mm
);
5201 return buf
- old_buf
;
5205 * access_remote_vm - access another process' address space
5206 * @mm: the mm_struct of the target address space
5207 * @addr: start address to access
5208 * @buf: source or destination buffer
5209 * @len: number of bytes to transfer
5210 * @gup_flags: flags modifying lookup behaviour
5212 * The caller must hold a reference on @mm.
5214 * Return: number of bytes copied from source to destination.
5216 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
5217 void *buf
, int len
, unsigned int gup_flags
)
5219 return __access_remote_vm(mm
, addr
, buf
, len
, gup_flags
);
5223 * Access another process' address space.
5224 * Source/target buffer must be kernel space,
5225 * Do not walk the page table directly, use get_user_pages
5227 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
5228 void *buf
, int len
, unsigned int gup_flags
)
5230 struct mm_struct
*mm
;
5233 mm
= get_task_mm(tsk
);
5237 ret
= __access_remote_vm(mm
, addr
, buf
, len
, gup_flags
);
5243 EXPORT_SYMBOL_GPL(access_process_vm
);
5246 * Print the name of a VMA.
5248 void print_vma_addr(char *prefix
, unsigned long ip
)
5250 struct mm_struct
*mm
= current
->mm
;
5251 struct vm_area_struct
*vma
;
5254 * we might be running from an atomic context so we cannot sleep
5256 if (!mmap_read_trylock(mm
))
5259 vma
= find_vma(mm
, ip
);
5260 if (vma
&& vma
->vm_file
) {
5261 struct file
*f
= vma
->vm_file
;
5262 char *buf
= (char *)__get_free_page(GFP_NOWAIT
);
5266 p
= file_path(f
, buf
, PAGE_SIZE
);
5269 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
5271 vma
->vm_end
- vma
->vm_start
);
5272 free_page((unsigned long)buf
);
5275 mmap_read_unlock(mm
);
5278 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5279 void __might_fault(const char *file
, int line
)
5282 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5283 * holding the mmap_lock, this is safe because kernel memory doesn't
5284 * get paged out, therefore we'll never actually fault, and the
5285 * below annotations will generate false positives.
5287 if (uaccess_kernel())
5289 if (pagefault_disabled())
5291 __might_sleep(file
, line
, 0);
5292 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5294 might_lock_read(¤t
->mm
->mmap_lock
);
5297 EXPORT_SYMBOL(__might_fault
);
5300 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5302 * Process all subpages of the specified huge page with the specified
5303 * operation. The target subpage will be processed last to keep its
5306 static inline void process_huge_page(
5307 unsigned long addr_hint
, unsigned int pages_per_huge_page
,
5308 void (*process_subpage
)(unsigned long addr
, int idx
, void *arg
),
5312 unsigned long addr
= addr_hint
&
5313 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5315 /* Process target subpage last to keep its cache lines hot */
5317 n
= (addr_hint
- addr
) / PAGE_SIZE
;
5318 if (2 * n
<= pages_per_huge_page
) {
5319 /* If target subpage in first half of huge page */
5322 /* Process subpages at the end of huge page */
5323 for (i
= pages_per_huge_page
- 1; i
>= 2 * n
; i
--) {
5325 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
5328 /* If target subpage in second half of huge page */
5329 base
= pages_per_huge_page
- 2 * (pages_per_huge_page
- n
);
5330 l
= pages_per_huge_page
- n
;
5331 /* Process subpages at the begin of huge page */
5332 for (i
= 0; i
< base
; i
++) {
5334 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
5338 * Process remaining subpages in left-right-left-right pattern
5339 * towards the target subpage
5341 for (i
= 0; i
< l
; i
++) {
5342 int left_idx
= base
+ i
;
5343 int right_idx
= base
+ 2 * l
- 1 - i
;
5346 process_subpage(addr
+ left_idx
* PAGE_SIZE
, left_idx
, arg
);
5348 process_subpage(addr
+ right_idx
* PAGE_SIZE
, right_idx
, arg
);
5352 static void clear_gigantic_page(struct page
*page
,
5354 unsigned int pages_per_huge_page
)
5357 struct page
*p
= page
;
5360 for (i
= 0; i
< pages_per_huge_page
;
5361 i
++, p
= mem_map_next(p
, page
, i
)) {
5363 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
5367 static void clear_subpage(unsigned long addr
, int idx
, void *arg
)
5369 struct page
*page
= arg
;
5371 clear_user_highpage(page
+ idx
, addr
);
5374 void clear_huge_page(struct page
*page
,
5375 unsigned long addr_hint
, unsigned int pages_per_huge_page
)
5377 unsigned long addr
= addr_hint
&
5378 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5380 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
5381 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
5385 process_huge_page(addr_hint
, pages_per_huge_page
, clear_subpage
, page
);
5388 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
5390 struct vm_area_struct
*vma
,
5391 unsigned int pages_per_huge_page
)
5394 struct page
*dst_base
= dst
;
5395 struct page
*src_base
= src
;
5397 for (i
= 0; i
< pages_per_huge_page
; ) {
5399 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
5402 dst
= mem_map_next(dst
, dst_base
, i
);
5403 src
= mem_map_next(src
, src_base
, i
);
5407 struct copy_subpage_arg
{
5410 struct vm_area_struct
*vma
;
5413 static void copy_subpage(unsigned long addr
, int idx
, void *arg
)
5415 struct copy_subpage_arg
*copy_arg
= arg
;
5417 copy_user_highpage(copy_arg
->dst
+ idx
, copy_arg
->src
+ idx
,
5418 addr
, copy_arg
->vma
);
5421 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
5422 unsigned long addr_hint
, struct vm_area_struct
*vma
,
5423 unsigned int pages_per_huge_page
)
5425 unsigned long addr
= addr_hint
&
5426 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5427 struct copy_subpage_arg arg
= {
5433 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
5434 copy_user_gigantic_page(dst
, src
, addr
, vma
,
5435 pages_per_huge_page
);
5439 process_huge_page(addr_hint
, pages_per_huge_page
, copy_subpage
, &arg
);
5442 long copy_huge_page_from_user(struct page
*dst_page
,
5443 const void __user
*usr_src
,
5444 unsigned int pages_per_huge_page
,
5445 bool allow_pagefault
)
5447 void *src
= (void *)usr_src
;
5449 unsigned long i
, rc
= 0;
5450 unsigned long ret_val
= pages_per_huge_page
* PAGE_SIZE
;
5451 struct page
*subpage
= dst_page
;
5453 for (i
= 0; i
< pages_per_huge_page
;
5454 i
++, subpage
= mem_map_next(subpage
, dst_page
, i
)) {
5455 if (allow_pagefault
)
5456 page_kaddr
= kmap(subpage
);
5458 page_kaddr
= kmap_atomic(subpage
);
5459 rc
= copy_from_user(page_kaddr
,
5460 (const void __user
*)(src
+ i
* PAGE_SIZE
),
5462 if (allow_pagefault
)
5465 kunmap_atomic(page_kaddr
);
5467 ret_val
-= (PAGE_SIZE
- rc
);
5475 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5477 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5479 static struct kmem_cache
*page_ptl_cachep
;
5481 void __init
ptlock_cache_init(void)
5483 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
5487 bool ptlock_alloc(struct page
*page
)
5491 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
5498 void ptlock_free(struct page
*page
)
5500 kmem_cache_free(page_ptl_cachep
, page
->ptl
);