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.
93 #ifndef CONFIG_NEED_MULTIPLE_NODES
94 /* use the per-pgdat data instead for discontigmem - mbligh */
95 unsigned long max_mapnr
;
96 EXPORT_SYMBOL(max_mapnr
);
99 EXPORT_SYMBOL(mem_map
);
103 * A number of key systems in x86 including ioremap() rely on the assumption
104 * that high_memory defines the upper bound on direct map memory, then end
105 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
106 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
110 EXPORT_SYMBOL(high_memory
);
113 * Randomize the address space (stacks, mmaps, brk, etc.).
115 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
116 * as ancient (libc5 based) binaries can segfault. )
118 int randomize_va_space __read_mostly
=
119 #ifdef CONFIG_COMPAT_BRK
125 #ifndef arch_faults_on_old_pte
126 static inline bool arch_faults_on_old_pte(void)
129 * Those arches which don't have hw access flag feature need to
130 * implement their own helper. By default, "true" means pagefault
131 * will be hit on old pte.
137 static int __init
disable_randmaps(char *s
)
139 randomize_va_space
= 0;
142 __setup("norandmaps", disable_randmaps
);
144 unsigned long zero_pfn __read_mostly
;
145 EXPORT_SYMBOL(zero_pfn
);
147 unsigned long highest_memmap_pfn __read_mostly
;
150 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
152 static int __init
init_zero_pfn(void)
154 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
157 early_initcall(init_zero_pfn
);
159 void mm_trace_rss_stat(struct mm_struct
*mm
, int member
, long count
)
161 trace_rss_stat(mm
, member
, count
);
164 #if defined(SPLIT_RSS_COUNTING)
166 void sync_mm_rss(struct mm_struct
*mm
)
170 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
171 if (current
->rss_stat
.count
[i
]) {
172 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
173 current
->rss_stat
.count
[i
] = 0;
176 current
->rss_stat
.events
= 0;
179 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
181 struct task_struct
*task
= current
;
183 if (likely(task
->mm
== mm
))
184 task
->rss_stat
.count
[member
] += val
;
186 add_mm_counter(mm
, member
, val
);
188 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
189 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
191 /* sync counter once per 64 page faults */
192 #define TASK_RSS_EVENTS_THRESH (64)
193 static void check_sync_rss_stat(struct task_struct
*task
)
195 if (unlikely(task
!= current
))
197 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
198 sync_mm_rss(task
->mm
);
200 #else /* SPLIT_RSS_COUNTING */
202 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
203 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
205 static void check_sync_rss_stat(struct task_struct
*task
)
209 #endif /* SPLIT_RSS_COUNTING */
212 * Note: this doesn't free the actual pages themselves. That
213 * has been handled earlier when unmapping all the memory regions.
215 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
218 pgtable_t token
= pmd_pgtable(*pmd
);
220 pte_free_tlb(tlb
, token
, addr
);
221 mm_dec_nr_ptes(tlb
->mm
);
224 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
225 unsigned long addr
, unsigned long end
,
226 unsigned long floor
, unsigned long ceiling
)
233 pmd
= pmd_offset(pud
, addr
);
235 next
= pmd_addr_end(addr
, end
);
236 if (pmd_none_or_clear_bad(pmd
))
238 free_pte_range(tlb
, pmd
, addr
);
239 } while (pmd
++, addr
= next
, addr
!= end
);
249 if (end
- 1 > ceiling
- 1)
252 pmd
= pmd_offset(pud
, start
);
254 pmd_free_tlb(tlb
, pmd
, start
);
255 mm_dec_nr_pmds(tlb
->mm
);
258 static inline void free_pud_range(struct mmu_gather
*tlb
, p4d_t
*p4d
,
259 unsigned long addr
, unsigned long end
,
260 unsigned long floor
, unsigned long ceiling
)
267 pud
= pud_offset(p4d
, addr
);
269 next
= pud_addr_end(addr
, end
);
270 if (pud_none_or_clear_bad(pud
))
272 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
273 } while (pud
++, addr
= next
, addr
!= end
);
283 if (end
- 1 > ceiling
- 1)
286 pud
= pud_offset(p4d
, start
);
288 pud_free_tlb(tlb
, pud
, start
);
289 mm_dec_nr_puds(tlb
->mm
);
292 static inline void free_p4d_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
293 unsigned long addr
, unsigned long end
,
294 unsigned long floor
, unsigned long ceiling
)
301 p4d
= p4d_offset(pgd
, addr
);
303 next
= p4d_addr_end(addr
, end
);
304 if (p4d_none_or_clear_bad(p4d
))
306 free_pud_range(tlb
, p4d
, addr
, next
, floor
, ceiling
);
307 } while (p4d
++, addr
= next
, addr
!= end
);
313 ceiling
&= PGDIR_MASK
;
317 if (end
- 1 > ceiling
- 1)
320 p4d
= p4d_offset(pgd
, start
);
322 p4d_free_tlb(tlb
, p4d
, start
);
326 * This function frees user-level page tables of a process.
328 void free_pgd_range(struct mmu_gather
*tlb
,
329 unsigned long addr
, unsigned long end
,
330 unsigned long floor
, unsigned long ceiling
)
336 * The next few lines have given us lots of grief...
338 * Why are we testing PMD* at this top level? Because often
339 * there will be no work to do at all, and we'd prefer not to
340 * go all the way down to the bottom just to discover that.
342 * Why all these "- 1"s? Because 0 represents both the bottom
343 * of the address space and the top of it (using -1 for the
344 * top wouldn't help much: the masks would do the wrong thing).
345 * The rule is that addr 0 and floor 0 refer to the bottom of
346 * the address space, but end 0 and ceiling 0 refer to the top
347 * Comparisons need to use "end - 1" and "ceiling - 1" (though
348 * that end 0 case should be mythical).
350 * Wherever addr is brought up or ceiling brought down, we must
351 * be careful to reject "the opposite 0" before it confuses the
352 * subsequent tests. But what about where end is brought down
353 * by PMD_SIZE below? no, end can't go down to 0 there.
355 * Whereas we round start (addr) and ceiling down, by different
356 * masks at different levels, in order to test whether a table
357 * now has no other vmas using it, so can be freed, we don't
358 * bother to round floor or end up - the tests don't need that.
372 if (end
- 1 > ceiling
- 1)
377 * We add page table cache pages with PAGE_SIZE,
378 * (see pte_free_tlb()), flush the tlb if we need
380 tlb_change_page_size(tlb
, PAGE_SIZE
);
381 pgd
= pgd_offset(tlb
->mm
, addr
);
383 next
= pgd_addr_end(addr
, end
);
384 if (pgd_none_or_clear_bad(pgd
))
386 free_p4d_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
387 } while (pgd
++, addr
= next
, addr
!= end
);
390 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
391 unsigned long floor
, unsigned long ceiling
)
394 struct vm_area_struct
*next
= vma
->vm_next
;
395 unsigned long addr
= vma
->vm_start
;
398 * Hide vma from rmap and truncate_pagecache before freeing
401 unlink_anon_vmas(vma
);
402 unlink_file_vma(vma
);
404 if (is_vm_hugetlb_page(vma
)) {
405 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
406 floor
, next
? next
->vm_start
: ceiling
);
409 * Optimization: gather nearby vmas into one call down
411 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
412 && !is_vm_hugetlb_page(next
)) {
415 unlink_anon_vmas(vma
);
416 unlink_file_vma(vma
);
418 free_pgd_range(tlb
, addr
, vma
->vm_end
,
419 floor
, next
? next
->vm_start
: ceiling
);
425 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
)
428 pgtable_t
new = pte_alloc_one(mm
);
433 * Ensure all pte setup (eg. pte page lock and page clearing) are
434 * visible before the pte is made visible to other CPUs by being
435 * put into page tables.
437 * The other side of the story is the pointer chasing in the page
438 * table walking code (when walking the page table without locking;
439 * ie. most of the time). Fortunately, these data accesses consist
440 * of a chain of data-dependent loads, meaning most CPUs (alpha
441 * being the notable exception) will already guarantee loads are
442 * seen in-order. See the alpha page table accessors for the
443 * smp_rmb() barriers in page table walking code.
445 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
447 ptl
= pmd_lock(mm
, pmd
);
448 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
450 pmd_populate(mm
, pmd
, new);
459 int __pte_alloc_kernel(pmd_t
*pmd
)
461 pte_t
*new = pte_alloc_one_kernel(&init_mm
);
465 smp_wmb(); /* See comment in __pte_alloc */
467 spin_lock(&init_mm
.page_table_lock
);
468 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
469 pmd_populate_kernel(&init_mm
, pmd
, new);
472 spin_unlock(&init_mm
.page_table_lock
);
474 pte_free_kernel(&init_mm
, new);
478 static inline void init_rss_vec(int *rss
)
480 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
483 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
487 if (current
->mm
== mm
)
489 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
491 add_mm_counter(mm
, i
, rss
[i
]);
495 * This function is called to print an error when a bad pte
496 * is found. For example, we might have a PFN-mapped pte in
497 * a region that doesn't allow it.
499 * The calling function must still handle the error.
501 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
502 pte_t pte
, struct page
*page
)
504 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
505 p4d_t
*p4d
= p4d_offset(pgd
, addr
);
506 pud_t
*pud
= pud_offset(p4d
, addr
);
507 pmd_t
*pmd
= pmd_offset(pud
, addr
);
508 struct address_space
*mapping
;
510 static unsigned long resume
;
511 static unsigned long nr_shown
;
512 static unsigned long nr_unshown
;
515 * Allow a burst of 60 reports, then keep quiet for that minute;
516 * or allow a steady drip of one report per second.
518 if (nr_shown
== 60) {
519 if (time_before(jiffies
, resume
)) {
524 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
531 resume
= jiffies
+ 60 * HZ
;
533 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
534 index
= linear_page_index(vma
, addr
);
536 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
538 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
540 dump_page(page
, "bad pte");
541 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
542 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
543 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
545 vma
->vm_ops
? vma
->vm_ops
->fault
: NULL
,
546 vma
->vm_file
? vma
->vm_file
->f_op
->mmap
: NULL
,
547 mapping
? mapping
->a_ops
->readpage
: NULL
);
549 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
553 * vm_normal_page -- This function gets the "struct page" associated with a pte.
555 * "Special" mappings do not wish to be associated with a "struct page" (either
556 * it doesn't exist, or it exists but they don't want to touch it). In this
557 * case, NULL is returned here. "Normal" mappings do have a struct page.
559 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
560 * pte bit, in which case this function is trivial. Secondly, an architecture
561 * may not have a spare pte bit, which requires a more complicated scheme,
564 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
565 * special mapping (even if there are underlying and valid "struct pages").
566 * COWed pages of a VM_PFNMAP are always normal.
568 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
569 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
570 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
571 * mapping will always honor the rule
573 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
575 * And for normal mappings this is false.
577 * This restricts such mappings to be a linear translation from virtual address
578 * to pfn. To get around this restriction, we allow arbitrary mappings so long
579 * as the vma is not a COW mapping; in that case, we know that all ptes are
580 * special (because none can have been COWed).
583 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
585 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
586 * page" backing, however the difference is that _all_ pages with a struct
587 * page (that is, those where pfn_valid is true) are refcounted and considered
588 * normal pages by the VM. The disadvantage is that pages are refcounted
589 * (which can be slower and simply not an option for some PFNMAP users). The
590 * advantage is that we don't have to follow the strict linearity rule of
591 * PFNMAP mappings in order to support COWable mappings.
594 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
597 unsigned long pfn
= pte_pfn(pte
);
599 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
)) {
600 if (likely(!pte_special(pte
)))
602 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
603 return vma
->vm_ops
->find_special_page(vma
, addr
);
604 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
606 if (is_zero_pfn(pfn
))
611 print_bad_pte(vma
, addr
, pte
, NULL
);
615 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
617 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
618 if (vma
->vm_flags
& VM_MIXEDMAP
) {
624 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
625 if (pfn
== vma
->vm_pgoff
+ off
)
627 if (!is_cow_mapping(vma
->vm_flags
))
632 if (is_zero_pfn(pfn
))
636 if (unlikely(pfn
> highest_memmap_pfn
)) {
637 print_bad_pte(vma
, addr
, pte
, NULL
);
642 * NOTE! We still have PageReserved() pages in the page tables.
643 * eg. VDSO mappings can cause them to exist.
646 return pfn_to_page(pfn
);
649 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
650 struct page
*vm_normal_page_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
653 unsigned long pfn
= pmd_pfn(pmd
);
656 * There is no pmd_special() but there may be special pmds, e.g.
657 * in a direct-access (dax) mapping, so let's just replicate the
658 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
660 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
661 if (vma
->vm_flags
& VM_MIXEDMAP
) {
667 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
668 if (pfn
== vma
->vm_pgoff
+ off
)
670 if (!is_cow_mapping(vma
->vm_flags
))
677 if (is_huge_zero_pmd(pmd
))
679 if (unlikely(pfn
> highest_memmap_pfn
))
683 * NOTE! We still have PageReserved() pages in the page tables.
684 * eg. VDSO mappings can cause them to exist.
687 return pfn_to_page(pfn
);
692 * copy one vm_area from one task to the other. Assumes the page tables
693 * already present in the new task to be cleared in the whole range
694 * covered by this vma.
698 copy_nonpresent_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
699 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
700 unsigned long addr
, int *rss
)
702 unsigned long vm_flags
= vma
->vm_flags
;
703 pte_t pte
= *src_pte
;
705 swp_entry_t entry
= pte_to_swp_entry(pte
);
707 if (likely(!non_swap_entry(entry
))) {
708 if (swap_duplicate(entry
) < 0)
711 /* make sure dst_mm is on swapoff's mmlist. */
712 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
713 spin_lock(&mmlist_lock
);
714 if (list_empty(&dst_mm
->mmlist
))
715 list_add(&dst_mm
->mmlist
,
717 spin_unlock(&mmlist_lock
);
720 } else if (is_migration_entry(entry
)) {
721 page
= migration_entry_to_page(entry
);
723 rss
[mm_counter(page
)]++;
725 if (is_write_migration_entry(entry
) &&
726 is_cow_mapping(vm_flags
)) {
728 * COW mappings require pages in both
729 * parent and child to be set to read.
731 make_migration_entry_read(&entry
);
732 pte
= swp_entry_to_pte(entry
);
733 if (pte_swp_soft_dirty(*src_pte
))
734 pte
= pte_swp_mksoft_dirty(pte
);
735 if (pte_swp_uffd_wp(*src_pte
))
736 pte
= pte_swp_mkuffd_wp(pte
);
737 set_pte_at(src_mm
, addr
, src_pte
, pte
);
739 } else if (is_device_private_entry(entry
)) {
740 page
= device_private_entry_to_page(entry
);
743 * Update rss count even for unaddressable pages, as
744 * they should treated just like normal pages in this
747 * We will likely want to have some new rss counters
748 * for unaddressable pages, at some point. But for now
749 * keep things as they are.
752 rss
[mm_counter(page
)]++;
753 page_dup_rmap(page
, false);
756 * We do not preserve soft-dirty information, because so
757 * far, checkpoint/restore is the only feature that
758 * requires that. And checkpoint/restore does not work
759 * when a device driver is involved (you cannot easily
760 * save and restore device driver state).
762 if (is_write_device_private_entry(entry
) &&
763 is_cow_mapping(vm_flags
)) {
764 make_device_private_entry_read(&entry
);
765 pte
= swp_entry_to_pte(entry
);
766 if (pte_swp_uffd_wp(*src_pte
))
767 pte
= pte_swp_mkuffd_wp(pte
);
768 set_pte_at(src_mm
, addr
, src_pte
, pte
);
771 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
776 * Copy a present and normal page if necessary.
778 * NOTE! The usual case is that this doesn't need to do
779 * anything, and can just return a positive value. That
780 * will let the caller know that it can just increase
781 * the page refcount and re-use the pte the traditional
784 * But _if_ we need to copy it because it needs to be
785 * pinned in the parent (and the child should get its own
786 * copy rather than just a reference to the same page),
787 * we'll do that here and return zero to let the caller
790 * And if we need a pre-allocated page but don't yet have
791 * one, return a negative error to let the preallocation
792 * code know so that it can do so outside the page table
796 copy_present_page(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
797 pte_t
*dst_pte
, pte_t
*src_pte
, unsigned long addr
, int *rss
,
798 struct page
**prealloc
, pte_t pte
, struct page
*page
)
800 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
801 struct page
*new_page
;
803 if (!is_cow_mapping(src_vma
->vm_flags
))
807 * What we want to do is to check whether this page may
808 * have been pinned by the parent process. If so,
809 * instead of wrprotect the pte on both sides, we copy
810 * the page immediately so that we'll always guarantee
811 * the pinned page won't be randomly replaced in the
814 * The page pinning checks are just "has this mm ever
815 * seen pinning", along with the (inexact) check of
816 * the page count. That might give false positives for
817 * for pinning, but it will work correctly.
819 if (likely(!atomic_read(&src_mm
->has_pinned
)))
821 if (likely(!page_maybe_dma_pinned(page
)))
824 new_page
= *prealloc
;
829 * We have a prealloc page, all good! Take it
830 * over and copy the page & arm it.
833 copy_user_highpage(new_page
, page
, addr
, src_vma
);
834 __SetPageUptodate(new_page
);
835 page_add_new_anon_rmap(new_page
, dst_vma
, addr
, false);
836 lru_cache_add_inactive_or_unevictable(new_page
, dst_vma
);
837 rss
[mm_counter(new_page
)]++;
839 /* All done, just insert the new page copy in the child */
840 pte
= mk_pte(new_page
, dst_vma
->vm_page_prot
);
841 pte
= maybe_mkwrite(pte_mkdirty(pte
), dst_vma
);
842 set_pte_at(dst_vma
->vm_mm
, addr
, dst_pte
, pte
);
847 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
848 * is required to copy this pte.
851 copy_present_pte(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
852 pte_t
*dst_pte
, pte_t
*src_pte
, unsigned long addr
, int *rss
,
853 struct page
**prealloc
)
855 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
856 unsigned long vm_flags
= src_vma
->vm_flags
;
857 pte_t pte
= *src_pte
;
860 page
= vm_normal_page(src_vma
, addr
, pte
);
864 retval
= copy_present_page(dst_vma
, src_vma
, dst_pte
, src_pte
,
865 addr
, rss
, prealloc
, pte
, page
);
870 page_dup_rmap(page
, false);
871 rss
[mm_counter(page
)]++;
875 * If it's a COW mapping, write protect it both
876 * in the parent and the child
878 if (is_cow_mapping(vm_flags
) && pte_write(pte
)) {
879 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
880 pte
= pte_wrprotect(pte
);
884 * If it's a shared mapping, mark it clean in
887 if (vm_flags
& VM_SHARED
)
888 pte
= pte_mkclean(pte
);
889 pte
= pte_mkold(pte
);
892 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
893 * does not have the VM_UFFD_WP, which means that the uffd
894 * fork event is not enabled.
896 if (!(vm_flags
& VM_UFFD_WP
))
897 pte
= pte_clear_uffd_wp(pte
);
899 set_pte_at(dst_vma
->vm_mm
, addr
, dst_pte
, pte
);
903 static inline struct page
*
904 page_copy_prealloc(struct mm_struct
*src_mm
, struct vm_area_struct
*vma
,
907 struct page
*new_page
;
909 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, addr
);
913 if (mem_cgroup_charge(new_page
, src_mm
, GFP_KERNEL
)) {
917 cgroup_throttle_swaprate(new_page
, GFP_KERNEL
);
923 copy_pte_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
924 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
927 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
928 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
929 pte_t
*orig_src_pte
, *orig_dst_pte
;
930 pte_t
*src_pte
, *dst_pte
;
931 spinlock_t
*src_ptl
, *dst_ptl
;
932 int progress
, ret
= 0;
933 int rss
[NR_MM_COUNTERS
];
934 swp_entry_t entry
= (swp_entry_t
){0};
935 struct page
*prealloc
= NULL
;
941 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
946 src_pte
= pte_offset_map(src_pmd
, addr
);
947 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
948 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
949 orig_src_pte
= src_pte
;
950 orig_dst_pte
= dst_pte
;
951 arch_enter_lazy_mmu_mode();
955 * We are holding two locks at this point - either of them
956 * could generate latencies in another task on another CPU.
958 if (progress
>= 32) {
960 if (need_resched() ||
961 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
964 if (pte_none(*src_pte
)) {
968 if (unlikely(!pte_present(*src_pte
))) {
969 entry
.val
= copy_nonpresent_pte(dst_mm
, src_mm
,
977 /* copy_present_pte() will clear `*prealloc' if consumed */
978 ret
= copy_present_pte(dst_vma
, src_vma
, dst_pte
, src_pte
,
979 addr
, rss
, &prealloc
);
981 * If we need a pre-allocated page for this pte, drop the
982 * locks, allocate, and try again.
984 if (unlikely(ret
== -EAGAIN
))
986 if (unlikely(prealloc
)) {
988 * pre-alloc page cannot be reused by next time so as
989 * to strictly follow mempolicy (e.g., alloc_page_vma()
990 * will allocate page according to address). This
991 * could only happen if one pinned pte changed.
997 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
999 arch_leave_lazy_mmu_mode();
1000 spin_unlock(src_ptl
);
1001 pte_unmap(orig_src_pte
);
1002 add_mm_rss_vec(dst_mm
, rss
);
1003 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
1007 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0) {
1013 WARN_ON_ONCE(ret
!= -EAGAIN
);
1014 prealloc
= page_copy_prealloc(src_mm
, src_vma
, addr
);
1017 /* We've captured and resolved the error. Reset, try again. */
1023 if (unlikely(prealloc
))
1029 copy_pmd_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1030 pud_t
*dst_pud
, pud_t
*src_pud
, unsigned long addr
,
1033 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1034 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1035 pmd_t
*src_pmd
, *dst_pmd
;
1038 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
1041 src_pmd
= pmd_offset(src_pud
, addr
);
1043 next
= pmd_addr_end(addr
, end
);
1044 if (is_swap_pmd(*src_pmd
) || pmd_trans_huge(*src_pmd
)
1045 || pmd_devmap(*src_pmd
)) {
1047 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PMD_SIZE
, src_vma
);
1048 err
= copy_huge_pmd(dst_mm
, src_mm
,
1049 dst_pmd
, src_pmd
, addr
, src_vma
);
1056 if (pmd_none_or_clear_bad(src_pmd
))
1058 if (copy_pte_range(dst_vma
, src_vma
, dst_pmd
, src_pmd
,
1061 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1066 copy_pud_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1067 p4d_t
*dst_p4d
, p4d_t
*src_p4d
, unsigned long addr
,
1070 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1071 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1072 pud_t
*src_pud
, *dst_pud
;
1075 dst_pud
= pud_alloc(dst_mm
, dst_p4d
, addr
);
1078 src_pud
= pud_offset(src_p4d
, addr
);
1080 next
= pud_addr_end(addr
, end
);
1081 if (pud_trans_huge(*src_pud
) || pud_devmap(*src_pud
)) {
1084 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PUD_SIZE
, src_vma
);
1085 err
= copy_huge_pud(dst_mm
, src_mm
,
1086 dst_pud
, src_pud
, addr
, src_vma
);
1093 if (pud_none_or_clear_bad(src_pud
))
1095 if (copy_pmd_range(dst_vma
, src_vma
, dst_pud
, src_pud
,
1098 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1103 copy_p4d_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1104 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, unsigned long addr
,
1107 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1108 p4d_t
*src_p4d
, *dst_p4d
;
1111 dst_p4d
= p4d_alloc(dst_mm
, dst_pgd
, addr
);
1114 src_p4d
= p4d_offset(src_pgd
, addr
);
1116 next
= p4d_addr_end(addr
, end
);
1117 if (p4d_none_or_clear_bad(src_p4d
))
1119 if (copy_pud_range(dst_vma
, src_vma
, dst_p4d
, src_p4d
,
1122 } while (dst_p4d
++, src_p4d
++, addr
= next
, addr
!= end
);
1127 copy_page_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
)
1129 pgd_t
*src_pgd
, *dst_pgd
;
1131 unsigned long addr
= src_vma
->vm_start
;
1132 unsigned long end
= src_vma
->vm_end
;
1133 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1134 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1135 struct mmu_notifier_range range
;
1140 * Don't copy ptes where a page fault will fill them correctly.
1141 * Fork becomes much lighter when there are big shared or private
1142 * readonly mappings. The tradeoff is that copy_page_range is more
1143 * efficient than faulting.
1145 if (!(src_vma
->vm_flags
& (VM_HUGETLB
| VM_PFNMAP
| VM_MIXEDMAP
)) &&
1149 if (is_vm_hugetlb_page(src_vma
))
1150 return copy_hugetlb_page_range(dst_mm
, src_mm
, src_vma
);
1152 if (unlikely(src_vma
->vm_flags
& VM_PFNMAP
)) {
1154 * We do not free on error cases below as remove_vma
1155 * gets called on error from higher level routine
1157 ret
= track_pfn_copy(src_vma
);
1163 * We need to invalidate the secondary MMU mappings only when
1164 * there could be a permission downgrade on the ptes of the
1165 * parent mm. And a permission downgrade will only happen if
1166 * is_cow_mapping() returns true.
1168 is_cow
= is_cow_mapping(src_vma
->vm_flags
);
1171 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_PAGE
,
1172 0, src_vma
, src_mm
, addr
, end
);
1173 mmu_notifier_invalidate_range_start(&range
);
1175 * Disabling preemption is not needed for the write side, as
1176 * the read side doesn't spin, but goes to the mmap_lock.
1178 * Use the raw variant of the seqcount_t write API to avoid
1179 * lockdep complaining about preemptibility.
1181 mmap_assert_write_locked(src_mm
);
1182 raw_write_seqcount_begin(&src_mm
->write_protect_seq
);
1186 dst_pgd
= pgd_offset(dst_mm
, addr
);
1187 src_pgd
= pgd_offset(src_mm
, addr
);
1189 next
= pgd_addr_end(addr
, end
);
1190 if (pgd_none_or_clear_bad(src_pgd
))
1192 if (unlikely(copy_p4d_range(dst_vma
, src_vma
, dst_pgd
, src_pgd
,
1197 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1200 raw_write_seqcount_end(&src_mm
->write_protect_seq
);
1201 mmu_notifier_invalidate_range_end(&range
);
1206 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1207 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1208 unsigned long addr
, unsigned long end
,
1209 struct zap_details
*details
)
1211 struct mm_struct
*mm
= tlb
->mm
;
1212 int force_flush
= 0;
1213 int rss
[NR_MM_COUNTERS
];
1219 tlb_change_page_size(tlb
, PAGE_SIZE
);
1222 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1224 flush_tlb_batched_pending(mm
);
1225 arch_enter_lazy_mmu_mode();
1228 if (pte_none(ptent
))
1234 if (pte_present(ptent
)) {
1237 page
= vm_normal_page(vma
, addr
, ptent
);
1238 if (unlikely(details
) && page
) {
1240 * unmap_shared_mapping_pages() wants to
1241 * invalidate cache without truncating:
1242 * unmap shared but keep private pages.
1244 if (details
->check_mapping
&&
1245 details
->check_mapping
!= page_rmapping(page
))
1248 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1250 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1251 if (unlikely(!page
))
1254 if (!PageAnon(page
)) {
1255 if (pte_dirty(ptent
)) {
1257 set_page_dirty(page
);
1259 if (pte_young(ptent
) &&
1260 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1261 mark_page_accessed(page
);
1263 rss
[mm_counter(page
)]--;
1264 page_remove_rmap(page
, false);
1265 if (unlikely(page_mapcount(page
) < 0))
1266 print_bad_pte(vma
, addr
, ptent
, page
);
1267 if (unlikely(__tlb_remove_page(tlb
, page
))) {
1275 entry
= pte_to_swp_entry(ptent
);
1276 if (is_device_private_entry(entry
)) {
1277 struct page
*page
= device_private_entry_to_page(entry
);
1279 if (unlikely(details
&& details
->check_mapping
)) {
1281 * unmap_shared_mapping_pages() wants to
1282 * invalidate cache without truncating:
1283 * unmap shared but keep private pages.
1285 if (details
->check_mapping
!=
1286 page_rmapping(page
))
1290 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1291 rss
[mm_counter(page
)]--;
1292 page_remove_rmap(page
, false);
1297 /* If details->check_mapping, we leave swap entries. */
1298 if (unlikely(details
))
1301 if (!non_swap_entry(entry
))
1303 else if (is_migration_entry(entry
)) {
1306 page
= migration_entry_to_page(entry
);
1307 rss
[mm_counter(page
)]--;
1309 if (unlikely(!free_swap_and_cache(entry
)))
1310 print_bad_pte(vma
, addr
, ptent
, NULL
);
1311 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1312 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1314 add_mm_rss_vec(mm
, rss
);
1315 arch_leave_lazy_mmu_mode();
1317 /* Do the actual TLB flush before dropping ptl */
1319 tlb_flush_mmu_tlbonly(tlb
);
1320 pte_unmap_unlock(start_pte
, ptl
);
1323 * If we forced a TLB flush (either due to running out of
1324 * batch buffers or because we needed to flush dirty TLB
1325 * entries before releasing the ptl), free the batched
1326 * memory too. Restart if we didn't do everything.
1341 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1342 struct vm_area_struct
*vma
, pud_t
*pud
,
1343 unsigned long addr
, unsigned long end
,
1344 struct zap_details
*details
)
1349 pmd
= pmd_offset(pud
, addr
);
1351 next
= pmd_addr_end(addr
, end
);
1352 if (is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1353 if (next
- addr
!= HPAGE_PMD_SIZE
)
1354 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
1355 else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1360 * Here there can be other concurrent MADV_DONTNEED or
1361 * trans huge page faults running, and if the pmd is
1362 * none or trans huge it can change under us. This is
1363 * because MADV_DONTNEED holds the mmap_lock in read
1366 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1368 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1371 } while (pmd
++, addr
= next
, addr
!= end
);
1376 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1377 struct vm_area_struct
*vma
, p4d_t
*p4d
,
1378 unsigned long addr
, unsigned long end
,
1379 struct zap_details
*details
)
1384 pud
= pud_offset(p4d
, addr
);
1386 next
= pud_addr_end(addr
, end
);
1387 if (pud_trans_huge(*pud
) || pud_devmap(*pud
)) {
1388 if (next
- addr
!= HPAGE_PUD_SIZE
) {
1389 mmap_assert_locked(tlb
->mm
);
1390 split_huge_pud(vma
, pud
, addr
);
1391 } else if (zap_huge_pud(tlb
, vma
, pud
, addr
))
1395 if (pud_none_or_clear_bad(pud
))
1397 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1400 } while (pud
++, addr
= next
, addr
!= end
);
1405 static inline unsigned long zap_p4d_range(struct mmu_gather
*tlb
,
1406 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1407 unsigned long addr
, unsigned long end
,
1408 struct zap_details
*details
)
1413 p4d
= p4d_offset(pgd
, addr
);
1415 next
= p4d_addr_end(addr
, end
);
1416 if (p4d_none_or_clear_bad(p4d
))
1418 next
= zap_pud_range(tlb
, vma
, p4d
, addr
, next
, details
);
1419 } while (p4d
++, addr
= next
, addr
!= end
);
1424 void unmap_page_range(struct mmu_gather
*tlb
,
1425 struct vm_area_struct
*vma
,
1426 unsigned long addr
, unsigned long end
,
1427 struct zap_details
*details
)
1432 BUG_ON(addr
>= end
);
1433 tlb_start_vma(tlb
, vma
);
1434 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1436 next
= pgd_addr_end(addr
, end
);
1437 if (pgd_none_or_clear_bad(pgd
))
1439 next
= zap_p4d_range(tlb
, vma
, pgd
, addr
, next
, details
);
1440 } while (pgd
++, addr
= next
, addr
!= end
);
1441 tlb_end_vma(tlb
, vma
);
1445 static void unmap_single_vma(struct mmu_gather
*tlb
,
1446 struct vm_area_struct
*vma
, unsigned long start_addr
,
1447 unsigned long end_addr
,
1448 struct zap_details
*details
)
1450 unsigned long start
= max(vma
->vm_start
, start_addr
);
1453 if (start
>= vma
->vm_end
)
1455 end
= min(vma
->vm_end
, end_addr
);
1456 if (end
<= vma
->vm_start
)
1460 uprobe_munmap(vma
, start
, end
);
1462 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1463 untrack_pfn(vma
, 0, 0);
1466 if (unlikely(is_vm_hugetlb_page(vma
))) {
1468 * It is undesirable to test vma->vm_file as it
1469 * should be non-null for valid hugetlb area.
1470 * However, vm_file will be NULL in the error
1471 * cleanup path of mmap_region. When
1472 * hugetlbfs ->mmap method fails,
1473 * mmap_region() nullifies vma->vm_file
1474 * before calling this function to clean up.
1475 * Since no pte has actually been setup, it is
1476 * safe to do nothing in this case.
1479 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
1480 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1481 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
1484 unmap_page_range(tlb
, vma
, start
, end
, details
);
1489 * unmap_vmas - unmap a range of memory covered by a list of vma's
1490 * @tlb: address of the caller's struct mmu_gather
1491 * @vma: the starting vma
1492 * @start_addr: virtual address at which to start unmapping
1493 * @end_addr: virtual address at which to end unmapping
1495 * Unmap all pages in the vma list.
1497 * Only addresses between `start' and `end' will be unmapped.
1499 * The VMA list must be sorted in ascending virtual address order.
1501 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1502 * range after unmap_vmas() returns. So the only responsibility here is to
1503 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1504 * drops the lock and schedules.
1506 void unmap_vmas(struct mmu_gather
*tlb
,
1507 struct vm_area_struct
*vma
, unsigned long start_addr
,
1508 unsigned long end_addr
)
1510 struct mmu_notifier_range range
;
1512 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
, vma
->vm_mm
,
1513 start_addr
, end_addr
);
1514 mmu_notifier_invalidate_range_start(&range
);
1515 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1516 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1517 mmu_notifier_invalidate_range_end(&range
);
1521 * zap_page_range - remove user pages in a given range
1522 * @vma: vm_area_struct holding the applicable pages
1523 * @start: starting address of pages to zap
1524 * @size: number of bytes to zap
1526 * Caller must protect the VMA list
1528 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1531 struct mmu_notifier_range range
;
1532 struct mmu_gather tlb
;
1535 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1536 start
, start
+ size
);
1537 tlb_gather_mmu(&tlb
, vma
->vm_mm
, start
, range
.end
);
1538 update_hiwater_rss(vma
->vm_mm
);
1539 mmu_notifier_invalidate_range_start(&range
);
1540 for ( ; vma
&& vma
->vm_start
< range
.end
; vma
= vma
->vm_next
)
1541 unmap_single_vma(&tlb
, vma
, start
, range
.end
, NULL
);
1542 mmu_notifier_invalidate_range_end(&range
);
1543 tlb_finish_mmu(&tlb
, start
, range
.end
);
1545 EXPORT_SYMBOL(zap_page_range
);
1548 * zap_page_range_single - remove user pages in a given range
1549 * @vma: vm_area_struct holding the applicable pages
1550 * @address: starting address of pages to zap
1551 * @size: number of bytes to zap
1552 * @details: details of shared cache invalidation
1554 * The range must fit into one VMA.
1556 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1557 unsigned long size
, struct zap_details
*details
)
1559 struct mmu_notifier_range range
;
1560 struct mmu_gather tlb
;
1563 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1564 address
, address
+ size
);
1565 tlb_gather_mmu(&tlb
, vma
->vm_mm
, address
, range
.end
);
1566 update_hiwater_rss(vma
->vm_mm
);
1567 mmu_notifier_invalidate_range_start(&range
);
1568 unmap_single_vma(&tlb
, vma
, address
, range
.end
, details
);
1569 mmu_notifier_invalidate_range_end(&range
);
1570 tlb_finish_mmu(&tlb
, address
, range
.end
);
1574 * zap_vma_ptes - remove ptes mapping the vma
1575 * @vma: vm_area_struct holding ptes to be zapped
1576 * @address: starting address of pages to zap
1577 * @size: number of bytes to zap
1579 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1581 * The entire address range must be fully contained within the vma.
1584 void zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1587 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1588 !(vma
->vm_flags
& VM_PFNMAP
))
1591 zap_page_range_single(vma
, address
, size
, NULL
);
1593 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1595 static pmd_t
*walk_to_pmd(struct mm_struct
*mm
, unsigned long addr
)
1602 pgd
= pgd_offset(mm
, addr
);
1603 p4d
= p4d_alloc(mm
, pgd
, addr
);
1606 pud
= pud_alloc(mm
, p4d
, addr
);
1609 pmd
= pmd_alloc(mm
, pud
, addr
);
1613 VM_BUG_ON(pmd_trans_huge(*pmd
));
1617 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1620 pmd_t
*pmd
= walk_to_pmd(mm
, addr
);
1624 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1627 static int validate_page_before_insert(struct page
*page
)
1629 if (PageAnon(page
) || PageSlab(page
) || page_has_type(page
))
1631 flush_dcache_page(page
);
1635 static int insert_page_into_pte_locked(struct mm_struct
*mm
, pte_t
*pte
,
1636 unsigned long addr
, struct page
*page
, pgprot_t prot
)
1638 if (!pte_none(*pte
))
1640 /* Ok, finally just insert the thing.. */
1642 inc_mm_counter_fast(mm
, mm_counter_file(page
));
1643 page_add_file_rmap(page
, false);
1644 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1649 * This is the old fallback for page remapping.
1651 * For historical reasons, it only allows reserved pages. Only
1652 * old drivers should use this, and they needed to mark their
1653 * pages reserved for the old functions anyway.
1655 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1656 struct page
*page
, pgprot_t prot
)
1658 struct mm_struct
*mm
= vma
->vm_mm
;
1663 retval
= validate_page_before_insert(page
);
1667 pte
= get_locked_pte(mm
, addr
, &ptl
);
1670 retval
= insert_page_into_pte_locked(mm
, pte
, addr
, page
, prot
);
1671 pte_unmap_unlock(pte
, ptl
);
1677 static int insert_page_in_batch_locked(struct mm_struct
*mm
, pte_t
*pte
,
1678 unsigned long addr
, struct page
*page
, pgprot_t prot
)
1682 if (!page_count(page
))
1684 err
= validate_page_before_insert(page
);
1687 return insert_page_into_pte_locked(mm
, pte
, addr
, page
, prot
);
1690 /* insert_pages() amortizes the cost of spinlock operations
1691 * when inserting pages in a loop. Arch *must* define pte_index.
1693 static int insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
1694 struct page
**pages
, unsigned long *num
, pgprot_t prot
)
1697 pte_t
*start_pte
, *pte
;
1698 spinlock_t
*pte_lock
;
1699 struct mm_struct
*const mm
= vma
->vm_mm
;
1700 unsigned long curr_page_idx
= 0;
1701 unsigned long remaining_pages_total
= *num
;
1702 unsigned long pages_to_write_in_pmd
;
1706 pmd
= walk_to_pmd(mm
, addr
);
1710 pages_to_write_in_pmd
= min_t(unsigned long,
1711 remaining_pages_total
, PTRS_PER_PTE
- pte_index(addr
));
1713 /* Allocate the PTE if necessary; takes PMD lock once only. */
1715 if (pte_alloc(mm
, pmd
))
1718 while (pages_to_write_in_pmd
) {
1720 const int batch_size
= min_t(int, pages_to_write_in_pmd
, 8);
1722 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &pte_lock
);
1723 for (pte
= start_pte
; pte_idx
< batch_size
; ++pte
, ++pte_idx
) {
1724 int err
= insert_page_in_batch_locked(mm
, pte
,
1725 addr
, pages
[curr_page_idx
], prot
);
1726 if (unlikely(err
)) {
1727 pte_unmap_unlock(start_pte
, pte_lock
);
1729 remaining_pages_total
-= pte_idx
;
1735 pte_unmap_unlock(start_pte
, pte_lock
);
1736 pages_to_write_in_pmd
-= batch_size
;
1737 remaining_pages_total
-= batch_size
;
1739 if (remaining_pages_total
)
1743 *num
= remaining_pages_total
;
1746 #endif /* ifdef pte_index */
1749 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1750 * @vma: user vma to map to
1751 * @addr: target start user address of these pages
1752 * @pages: source kernel pages
1753 * @num: in: number of pages to map. out: number of pages that were *not*
1754 * mapped. (0 means all pages were successfully mapped).
1756 * Preferred over vm_insert_page() when inserting multiple pages.
1758 * In case of error, we may have mapped a subset of the provided
1759 * pages. It is the caller's responsibility to account for this case.
1761 * The same restrictions apply as in vm_insert_page().
1763 int vm_insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
1764 struct page
**pages
, unsigned long *num
)
1767 const unsigned long end_addr
= addr
+ (*num
* PAGE_SIZE
) - 1;
1769 if (addr
< vma
->vm_start
|| end_addr
>= vma
->vm_end
)
1771 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1772 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
1773 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1774 vma
->vm_flags
|= VM_MIXEDMAP
;
1776 /* Defer page refcount checking till we're about to map that page. */
1777 return insert_pages(vma
, addr
, pages
, num
, vma
->vm_page_prot
);
1779 unsigned long idx
= 0, pgcount
= *num
;
1782 for (; idx
< pgcount
; ++idx
) {
1783 err
= vm_insert_page(vma
, addr
+ (PAGE_SIZE
* idx
), pages
[idx
]);
1787 *num
= pgcount
- idx
;
1789 #endif /* ifdef pte_index */
1791 EXPORT_SYMBOL(vm_insert_pages
);
1794 * vm_insert_page - insert single page into user vma
1795 * @vma: user vma to map to
1796 * @addr: target user address of this page
1797 * @page: source kernel page
1799 * This allows drivers to insert individual pages they've allocated
1802 * The page has to be a nice clean _individual_ kernel allocation.
1803 * If you allocate a compound page, you need to have marked it as
1804 * such (__GFP_COMP), or manually just split the page up yourself
1805 * (see split_page()).
1807 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1808 * took an arbitrary page protection parameter. This doesn't allow
1809 * that. Your vma protection will have to be set up correctly, which
1810 * means that if you want a shared writable mapping, you'd better
1811 * ask for a shared writable mapping!
1813 * The page does not need to be reserved.
1815 * Usually this function is called from f_op->mmap() handler
1816 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1817 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1818 * function from other places, for example from page-fault handler.
1820 * Return: %0 on success, negative error code otherwise.
1822 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1825 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1827 if (!page_count(page
))
1829 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1830 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
1831 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1832 vma
->vm_flags
|= VM_MIXEDMAP
;
1834 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1836 EXPORT_SYMBOL(vm_insert_page
);
1839 * __vm_map_pages - maps range of kernel pages into user vma
1840 * @vma: user vma to map to
1841 * @pages: pointer to array of source kernel pages
1842 * @num: number of pages in page array
1843 * @offset: user's requested vm_pgoff
1845 * This allows drivers to map range of kernel pages into a user vma.
1847 * Return: 0 on success and error code otherwise.
1849 static int __vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
1850 unsigned long num
, unsigned long offset
)
1852 unsigned long count
= vma_pages(vma
);
1853 unsigned long uaddr
= vma
->vm_start
;
1856 /* Fail if the user requested offset is beyond the end of the object */
1860 /* Fail if the user requested size exceeds available object size */
1861 if (count
> num
- offset
)
1864 for (i
= 0; i
< count
; i
++) {
1865 ret
= vm_insert_page(vma
, uaddr
, pages
[offset
+ i
]);
1875 * vm_map_pages - maps range of kernel pages starts with non zero offset
1876 * @vma: user vma to map to
1877 * @pages: pointer to array of source kernel pages
1878 * @num: number of pages in page array
1880 * Maps an object consisting of @num pages, catering for the user's
1881 * requested vm_pgoff
1883 * If we fail to insert any page into the vma, the function will return
1884 * immediately leaving any previously inserted pages present. Callers
1885 * from the mmap handler may immediately return the error as their caller
1886 * will destroy the vma, removing any successfully inserted pages. Other
1887 * callers should make their own arrangements for calling unmap_region().
1889 * Context: Process context. Called by mmap handlers.
1890 * Return: 0 on success and error code otherwise.
1892 int vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
1895 return __vm_map_pages(vma
, pages
, num
, vma
->vm_pgoff
);
1897 EXPORT_SYMBOL(vm_map_pages
);
1900 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1901 * @vma: user vma to map to
1902 * @pages: pointer to array of source kernel pages
1903 * @num: number of pages in page array
1905 * Similar to vm_map_pages(), except that it explicitly sets the offset
1906 * to 0. This function is intended for the drivers that did not consider
1909 * Context: Process context. Called by mmap handlers.
1910 * Return: 0 on success and error code otherwise.
1912 int vm_map_pages_zero(struct vm_area_struct
*vma
, struct page
**pages
,
1915 return __vm_map_pages(vma
, pages
, num
, 0);
1917 EXPORT_SYMBOL(vm_map_pages_zero
);
1919 static vm_fault_t
insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1920 pfn_t pfn
, pgprot_t prot
, bool mkwrite
)
1922 struct mm_struct
*mm
= vma
->vm_mm
;
1926 pte
= get_locked_pte(mm
, addr
, &ptl
);
1928 return VM_FAULT_OOM
;
1929 if (!pte_none(*pte
)) {
1932 * For read faults on private mappings the PFN passed
1933 * in may not match the PFN we have mapped if the
1934 * mapped PFN is a writeable COW page. In the mkwrite
1935 * case we are creating a writable PTE for a shared
1936 * mapping and we expect the PFNs to match. If they
1937 * don't match, we are likely racing with block
1938 * allocation and mapping invalidation so just skip the
1941 if (pte_pfn(*pte
) != pfn_t_to_pfn(pfn
)) {
1942 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte
)));
1945 entry
= pte_mkyoung(*pte
);
1946 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1947 if (ptep_set_access_flags(vma
, addr
, pte
, entry
, 1))
1948 update_mmu_cache(vma
, addr
, pte
);
1953 /* Ok, finally just insert the thing.. */
1954 if (pfn_t_devmap(pfn
))
1955 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
1957 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
1960 entry
= pte_mkyoung(entry
);
1961 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1964 set_pte_at(mm
, addr
, pte
, entry
);
1965 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1968 pte_unmap_unlock(pte
, ptl
);
1969 return VM_FAULT_NOPAGE
;
1973 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1974 * @vma: user vma to map to
1975 * @addr: target user address of this page
1976 * @pfn: source kernel pfn
1977 * @pgprot: pgprot flags for the inserted page
1979 * This is exactly like vmf_insert_pfn(), except that it allows drivers
1980 * to override pgprot on a per-page basis.
1982 * This only makes sense for IO mappings, and it makes no sense for
1983 * COW mappings. In general, using multiple vmas is preferable;
1984 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1987 * See vmf_insert_mixed_prot() for a discussion of the implication of using
1988 * a value of @pgprot different from that of @vma->vm_page_prot.
1990 * Context: Process context. May allocate using %GFP_KERNEL.
1991 * Return: vm_fault_t value.
1993 vm_fault_t
vmf_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
1994 unsigned long pfn
, pgprot_t pgprot
)
1997 * Technically, architectures with pte_special can avoid all these
1998 * restrictions (same for remap_pfn_range). However we would like
1999 * consistency in testing and feature parity among all, so we should
2000 * try to keep these invariants in place for everybody.
2002 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
2003 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
2004 (VM_PFNMAP
|VM_MIXEDMAP
));
2005 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
2006 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
2008 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2009 return VM_FAULT_SIGBUS
;
2011 if (!pfn_modify_allowed(pfn
, pgprot
))
2012 return VM_FAULT_SIGBUS
;
2014 track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
));
2016 return insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
,
2019 EXPORT_SYMBOL(vmf_insert_pfn_prot
);
2022 * vmf_insert_pfn - insert single pfn into user vma
2023 * @vma: user vma to map to
2024 * @addr: target user address of this page
2025 * @pfn: source kernel pfn
2027 * Similar to vm_insert_page, this allows drivers to insert individual pages
2028 * they've allocated into a user vma. Same comments apply.
2030 * This function should only be called from a vm_ops->fault handler, and
2031 * in that case the handler should return the result of this function.
2033 * vma cannot be a COW mapping.
2035 * As this is called only for pages that do not currently exist, we
2036 * do not need to flush old virtual caches or the TLB.
2038 * Context: Process context. May allocate using %GFP_KERNEL.
2039 * Return: vm_fault_t value.
2041 vm_fault_t
vmf_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2044 return vmf_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
2046 EXPORT_SYMBOL(vmf_insert_pfn
);
2048 static bool vm_mixed_ok(struct vm_area_struct
*vma
, pfn_t pfn
)
2050 /* these checks mirror the abort conditions in vm_normal_page */
2051 if (vma
->vm_flags
& VM_MIXEDMAP
)
2053 if (pfn_t_devmap(pfn
))
2055 if (pfn_t_special(pfn
))
2057 if (is_zero_pfn(pfn_t_to_pfn(pfn
)))
2062 static vm_fault_t
__vm_insert_mixed(struct vm_area_struct
*vma
,
2063 unsigned long addr
, pfn_t pfn
, pgprot_t pgprot
,
2068 BUG_ON(!vm_mixed_ok(vma
, pfn
));
2070 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2071 return VM_FAULT_SIGBUS
;
2073 track_pfn_insert(vma
, &pgprot
, pfn
);
2075 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn
), pgprot
))
2076 return VM_FAULT_SIGBUS
;
2079 * If we don't have pte special, then we have to use the pfn_valid()
2080 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2081 * refcount the page if pfn_valid is true (hence insert_page rather
2082 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2083 * without pte special, it would there be refcounted as a normal page.
2085 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
) &&
2086 !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
2090 * At this point we are committed to insert_page()
2091 * regardless of whether the caller specified flags that
2092 * result in pfn_t_has_page() == false.
2094 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
2095 err
= insert_page(vma
, addr
, page
, pgprot
);
2097 return insert_pfn(vma
, addr
, pfn
, pgprot
, mkwrite
);
2101 return VM_FAULT_OOM
;
2102 if (err
< 0 && err
!= -EBUSY
)
2103 return VM_FAULT_SIGBUS
;
2105 return VM_FAULT_NOPAGE
;
2109 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2110 * @vma: user vma to map to
2111 * @addr: target user address of this page
2112 * @pfn: source kernel pfn
2113 * @pgprot: pgprot flags for the inserted page
2115 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2116 * to override pgprot on a per-page basis.
2118 * Typically this function should be used by drivers to set caching- and
2119 * encryption bits different than those of @vma->vm_page_prot, because
2120 * the caching- or encryption mode may not be known at mmap() time.
2121 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2122 * to set caching and encryption bits for those vmas (except for COW pages).
2123 * This is ensured by core vm only modifying these page table entries using
2124 * functions that don't touch caching- or encryption bits, using pte_modify()
2125 * if needed. (See for example mprotect()).
2126 * Also when new page-table entries are created, this is only done using the
2127 * fault() callback, and never using the value of vma->vm_page_prot,
2128 * except for page-table entries that point to anonymous pages as the result
2131 * Context: Process context. May allocate using %GFP_KERNEL.
2132 * Return: vm_fault_t value.
2134 vm_fault_t
vmf_insert_mixed_prot(struct vm_area_struct
*vma
, unsigned long addr
,
2135 pfn_t pfn
, pgprot_t pgprot
)
2137 return __vm_insert_mixed(vma
, addr
, pfn
, pgprot
, false);
2139 EXPORT_SYMBOL(vmf_insert_mixed_prot
);
2141 vm_fault_t
vmf_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2144 return __vm_insert_mixed(vma
, addr
, pfn
, vma
->vm_page_prot
, false);
2146 EXPORT_SYMBOL(vmf_insert_mixed
);
2149 * If the insertion of PTE failed because someone else already added a
2150 * different entry in the mean time, we treat that as success as we assume
2151 * the same entry was actually inserted.
2153 vm_fault_t
vmf_insert_mixed_mkwrite(struct vm_area_struct
*vma
,
2154 unsigned long addr
, pfn_t pfn
)
2156 return __vm_insert_mixed(vma
, addr
, pfn
, vma
->vm_page_prot
, true);
2158 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite
);
2161 * maps a range of physical memory into the requested pages. the old
2162 * mappings are removed. any references to nonexistent pages results
2163 * in null mappings (currently treated as "copy-on-access")
2165 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2166 unsigned long addr
, unsigned long end
,
2167 unsigned long pfn
, pgprot_t prot
)
2169 pte_t
*pte
, *mapped_pte
;
2173 mapped_pte
= pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2176 arch_enter_lazy_mmu_mode();
2178 BUG_ON(!pte_none(*pte
));
2179 if (!pfn_modify_allowed(pfn
, prot
)) {
2183 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2185 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2186 arch_leave_lazy_mmu_mode();
2187 pte_unmap_unlock(mapped_pte
, ptl
);
2191 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2192 unsigned long addr
, unsigned long end
,
2193 unsigned long pfn
, pgprot_t prot
)
2199 pfn
-= addr
>> PAGE_SHIFT
;
2200 pmd
= pmd_alloc(mm
, pud
, addr
);
2203 VM_BUG_ON(pmd_trans_huge(*pmd
));
2205 next
= pmd_addr_end(addr
, end
);
2206 err
= remap_pte_range(mm
, pmd
, addr
, next
,
2207 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2210 } while (pmd
++, addr
= next
, addr
!= end
);
2214 static inline int remap_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2215 unsigned long addr
, unsigned long end
,
2216 unsigned long pfn
, pgprot_t prot
)
2222 pfn
-= addr
>> PAGE_SHIFT
;
2223 pud
= pud_alloc(mm
, p4d
, addr
);
2227 next
= pud_addr_end(addr
, end
);
2228 err
= remap_pmd_range(mm
, pud
, addr
, next
,
2229 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2232 } while (pud
++, addr
= next
, addr
!= end
);
2236 static inline int remap_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2237 unsigned long addr
, unsigned long end
,
2238 unsigned long pfn
, pgprot_t prot
)
2244 pfn
-= addr
>> PAGE_SHIFT
;
2245 p4d
= p4d_alloc(mm
, pgd
, addr
);
2249 next
= p4d_addr_end(addr
, end
);
2250 err
= remap_pud_range(mm
, p4d
, addr
, next
,
2251 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2254 } while (p4d
++, addr
= next
, addr
!= end
);
2259 * remap_pfn_range - remap kernel memory to userspace
2260 * @vma: user vma to map to
2261 * @addr: target page aligned user address to start at
2262 * @pfn: page frame number of kernel physical memory address
2263 * @size: size of mapping area
2264 * @prot: page protection flags for this mapping
2266 * Note: this is only safe if the mm semaphore is held when called.
2268 * Return: %0 on success, negative error code otherwise.
2270 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2271 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2275 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2276 struct mm_struct
*mm
= vma
->vm_mm
;
2277 unsigned long remap_pfn
= pfn
;
2280 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr
)))
2284 * Physically remapped pages are special. Tell the
2285 * rest of the world about it:
2286 * VM_IO tells people not to look at these pages
2287 * (accesses can have side effects).
2288 * VM_PFNMAP tells the core MM that the base pages are just
2289 * raw PFN mappings, and do not have a "struct page" associated
2292 * Disable vma merging and expanding with mremap().
2294 * Omit vma from core dump, even when VM_IO turned off.
2296 * There's a horrible special case to handle copy-on-write
2297 * behaviour that some programs depend on. We mark the "original"
2298 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2299 * See vm_normal_page() for details.
2301 if (is_cow_mapping(vma
->vm_flags
)) {
2302 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2304 vma
->vm_pgoff
= pfn
;
2307 err
= track_pfn_remap(vma
, &prot
, remap_pfn
, addr
, PAGE_ALIGN(size
));
2311 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
2313 BUG_ON(addr
>= end
);
2314 pfn
-= addr
>> PAGE_SHIFT
;
2315 pgd
= pgd_offset(mm
, addr
);
2316 flush_cache_range(vma
, addr
, end
);
2318 next
= pgd_addr_end(addr
, end
);
2319 err
= remap_p4d_range(mm
, pgd
, addr
, next
,
2320 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2323 } while (pgd
++, addr
= next
, addr
!= end
);
2326 untrack_pfn(vma
, remap_pfn
, PAGE_ALIGN(size
));
2330 EXPORT_SYMBOL(remap_pfn_range
);
2333 * vm_iomap_memory - remap memory to userspace
2334 * @vma: user vma to map to
2335 * @start: start of the physical memory to be mapped
2336 * @len: size of area
2338 * This is a simplified io_remap_pfn_range() for common driver use. The
2339 * driver just needs to give us the physical memory range to be mapped,
2340 * we'll figure out the rest from the vma information.
2342 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2343 * whatever write-combining details or similar.
2345 * Return: %0 on success, negative error code otherwise.
2347 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2349 unsigned long vm_len
, pfn
, pages
;
2351 /* Check that the physical memory area passed in looks valid */
2352 if (start
+ len
< start
)
2355 * You *really* shouldn't map things that aren't page-aligned,
2356 * but we've historically allowed it because IO memory might
2357 * just have smaller alignment.
2359 len
+= start
& ~PAGE_MASK
;
2360 pfn
= start
>> PAGE_SHIFT
;
2361 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2362 if (pfn
+ pages
< pfn
)
2365 /* We start the mapping 'vm_pgoff' pages into the area */
2366 if (vma
->vm_pgoff
> pages
)
2368 pfn
+= vma
->vm_pgoff
;
2369 pages
-= vma
->vm_pgoff
;
2371 /* Can we fit all of the mapping? */
2372 vm_len
= vma
->vm_end
- vma
->vm_start
;
2373 if (vm_len
>> PAGE_SHIFT
> pages
)
2376 /* Ok, let it rip */
2377 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2379 EXPORT_SYMBOL(vm_iomap_memory
);
2381 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2382 unsigned long addr
, unsigned long end
,
2383 pte_fn_t fn
, void *data
, bool create
,
2384 pgtbl_mod_mask
*mask
)
2391 pte
= (mm
== &init_mm
) ?
2392 pte_alloc_kernel_track(pmd
, addr
, mask
) :
2393 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2397 pte
= (mm
== &init_mm
) ?
2398 pte_offset_kernel(pmd
, addr
) :
2399 pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
2402 BUG_ON(pmd_huge(*pmd
));
2404 arch_enter_lazy_mmu_mode();
2408 if (create
|| !pte_none(*pte
)) {
2409 err
= fn(pte
++, addr
, data
);
2413 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2415 *mask
|= PGTBL_PTE_MODIFIED
;
2417 arch_leave_lazy_mmu_mode();
2420 pte_unmap_unlock(pte
-1, ptl
);
2424 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2425 unsigned long addr
, unsigned long end
,
2426 pte_fn_t fn
, void *data
, bool create
,
2427 pgtbl_mod_mask
*mask
)
2433 BUG_ON(pud_huge(*pud
));
2436 pmd
= pmd_alloc_track(mm
, pud
, addr
, mask
);
2440 pmd
= pmd_offset(pud
, addr
);
2443 next
= pmd_addr_end(addr
, end
);
2444 if (create
|| !pmd_none_or_clear_bad(pmd
)) {
2445 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
,
2450 } while (pmd
++, addr
= next
, addr
!= end
);
2454 static int apply_to_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2455 unsigned long addr
, unsigned long end
,
2456 pte_fn_t fn
, void *data
, bool create
,
2457 pgtbl_mod_mask
*mask
)
2464 pud
= pud_alloc_track(mm
, p4d
, addr
, mask
);
2468 pud
= pud_offset(p4d
, addr
);
2471 next
= pud_addr_end(addr
, end
);
2472 if (create
|| !pud_none_or_clear_bad(pud
)) {
2473 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
,
2478 } while (pud
++, addr
= next
, addr
!= end
);
2482 static int apply_to_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2483 unsigned long addr
, unsigned long end
,
2484 pte_fn_t fn
, void *data
, bool create
,
2485 pgtbl_mod_mask
*mask
)
2492 p4d
= p4d_alloc_track(mm
, pgd
, addr
, mask
);
2496 p4d
= p4d_offset(pgd
, addr
);
2499 next
= p4d_addr_end(addr
, end
);
2500 if (create
|| !p4d_none_or_clear_bad(p4d
)) {
2501 err
= apply_to_pud_range(mm
, p4d
, addr
, next
, fn
, data
,
2506 } while (p4d
++, addr
= next
, addr
!= end
);
2510 static int __apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2511 unsigned long size
, pte_fn_t fn
,
2512 void *data
, bool create
)
2515 unsigned long start
= addr
, next
;
2516 unsigned long end
= addr
+ size
;
2517 pgtbl_mod_mask mask
= 0;
2520 if (WARN_ON(addr
>= end
))
2523 pgd
= pgd_offset(mm
, addr
);
2525 next
= pgd_addr_end(addr
, end
);
2526 if (!create
&& pgd_none_or_clear_bad(pgd
))
2528 err
= apply_to_p4d_range(mm
, pgd
, addr
, next
, fn
, data
, create
, &mask
);
2531 } while (pgd
++, addr
= next
, addr
!= end
);
2533 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
2534 arch_sync_kernel_mappings(start
, start
+ size
);
2540 * Scan a region of virtual memory, filling in page tables as necessary
2541 * and calling a provided function on each leaf page table.
2543 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2544 unsigned long size
, pte_fn_t fn
, void *data
)
2546 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, true);
2548 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2551 * Scan a region of virtual memory, calling a provided function on
2552 * each leaf page table where it exists.
2554 * Unlike apply_to_page_range, this does _not_ fill in page tables
2555 * where they are absent.
2557 int apply_to_existing_page_range(struct mm_struct
*mm
, unsigned long addr
,
2558 unsigned long size
, pte_fn_t fn
, void *data
)
2560 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, false);
2562 EXPORT_SYMBOL_GPL(apply_to_existing_page_range
);
2565 * handle_pte_fault chooses page fault handler according to an entry which was
2566 * read non-atomically. Before making any commitment, on those architectures
2567 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2568 * parts, do_swap_page must check under lock before unmapping the pte and
2569 * proceeding (but do_wp_page is only called after already making such a check;
2570 * and do_anonymous_page can safely check later on).
2572 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2573 pte_t
*page_table
, pte_t orig_pte
)
2576 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2577 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2578 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2580 same
= pte_same(*page_table
, orig_pte
);
2584 pte_unmap(page_table
);
2588 static inline bool cow_user_page(struct page
*dst
, struct page
*src
,
2589 struct vm_fault
*vmf
)
2594 bool locked
= false;
2595 struct vm_area_struct
*vma
= vmf
->vma
;
2596 struct mm_struct
*mm
= vma
->vm_mm
;
2597 unsigned long addr
= vmf
->address
;
2600 copy_user_highpage(dst
, src
, addr
, vma
);
2605 * If the source page was a PFN mapping, we don't have
2606 * a "struct page" for it. We do a best-effort copy by
2607 * just copying from the original user address. If that
2608 * fails, we just zero-fill it. Live with it.
2610 kaddr
= kmap_atomic(dst
);
2611 uaddr
= (void __user
*)(addr
& PAGE_MASK
);
2614 * On architectures with software "accessed" bits, we would
2615 * take a double page fault, so mark it accessed here.
2617 if (arch_faults_on_old_pte() && !pte_young(vmf
->orig_pte
)) {
2620 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2622 if (!likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2624 * Other thread has already handled the fault
2625 * and update local tlb only
2627 update_mmu_tlb(vma
, addr
, vmf
->pte
);
2632 entry
= pte_mkyoung(vmf
->orig_pte
);
2633 if (ptep_set_access_flags(vma
, addr
, vmf
->pte
, entry
, 0))
2634 update_mmu_cache(vma
, addr
, vmf
->pte
);
2638 * This really shouldn't fail, because the page is there
2639 * in the page tables. But it might just be unreadable,
2640 * in which case we just give up and fill the result with
2643 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
2647 /* Re-validate under PTL if the page is still mapped */
2648 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2650 if (!likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2651 /* The PTE changed under us, update local tlb */
2652 update_mmu_tlb(vma
, addr
, vmf
->pte
);
2658 * The same page can be mapped back since last copy attempt.
2659 * Try to copy again under PTL.
2661 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
2663 * Give a warn in case there can be some obscure
2676 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2677 kunmap_atomic(kaddr
);
2678 flush_dcache_page(dst
);
2683 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
2685 struct file
*vm_file
= vma
->vm_file
;
2688 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
2691 * Special mappings (e.g. VDSO) do not have any file so fake
2692 * a default GFP_KERNEL for them.
2698 * Notify the address space that the page is about to become writable so that
2699 * it can prohibit this or wait for the page to get into an appropriate state.
2701 * We do this without the lock held, so that it can sleep if it needs to.
2703 static vm_fault_t
do_page_mkwrite(struct vm_fault
*vmf
)
2706 struct page
*page
= vmf
->page
;
2707 unsigned int old_flags
= vmf
->flags
;
2709 vmf
->flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2711 if (vmf
->vma
->vm_file
&&
2712 IS_SWAPFILE(vmf
->vma
->vm_file
->f_mapping
->host
))
2713 return VM_FAULT_SIGBUS
;
2715 ret
= vmf
->vma
->vm_ops
->page_mkwrite(vmf
);
2716 /* Restore original flags so that caller is not surprised */
2717 vmf
->flags
= old_flags
;
2718 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2720 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
2722 if (!page
->mapping
) {
2724 return 0; /* retry */
2726 ret
|= VM_FAULT_LOCKED
;
2728 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2733 * Handle dirtying of a page in shared file mapping on a write fault.
2735 * The function expects the page to be locked and unlocks it.
2737 static vm_fault_t
fault_dirty_shared_page(struct vm_fault
*vmf
)
2739 struct vm_area_struct
*vma
= vmf
->vma
;
2740 struct address_space
*mapping
;
2741 struct page
*page
= vmf
->page
;
2743 bool page_mkwrite
= vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
;
2745 dirtied
= set_page_dirty(page
);
2746 VM_BUG_ON_PAGE(PageAnon(page
), page
);
2748 * Take a local copy of the address_space - page.mapping may be zeroed
2749 * by truncate after unlock_page(). The address_space itself remains
2750 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2751 * release semantics to prevent the compiler from undoing this copying.
2753 mapping
= page_rmapping(page
);
2757 file_update_time(vma
->vm_file
);
2760 * Throttle page dirtying rate down to writeback speed.
2762 * mapping may be NULL here because some device drivers do not
2763 * set page.mapping but still dirty their pages
2765 * Drop the mmap_lock before waiting on IO, if we can. The file
2766 * is pinning the mapping, as per above.
2768 if ((dirtied
|| page_mkwrite
) && mapping
) {
2771 fpin
= maybe_unlock_mmap_for_io(vmf
, NULL
);
2772 balance_dirty_pages_ratelimited(mapping
);
2775 return VM_FAULT_RETRY
;
2783 * Handle write page faults for pages that can be reused in the current vma
2785 * This can happen either due to the mapping being with the VM_SHARED flag,
2786 * or due to us being the last reference standing to the page. In either
2787 * case, all we need to do here is to mark the page as writable and update
2788 * any related book-keeping.
2790 static inline void wp_page_reuse(struct vm_fault
*vmf
)
2791 __releases(vmf
->ptl
)
2793 struct vm_area_struct
*vma
= vmf
->vma
;
2794 struct page
*page
= vmf
->page
;
2797 * Clear the pages cpupid information as the existing
2798 * information potentially belongs to a now completely
2799 * unrelated process.
2802 page_cpupid_xchg_last(page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2804 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2805 entry
= pte_mkyoung(vmf
->orig_pte
);
2806 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2807 if (ptep_set_access_flags(vma
, vmf
->address
, vmf
->pte
, entry
, 1))
2808 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2809 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2810 count_vm_event(PGREUSE
);
2814 * Handle the case of a page which we actually need to copy to a new page.
2816 * Called with mmap_lock locked and the old page referenced, but
2817 * without the ptl held.
2819 * High level logic flow:
2821 * - Allocate a page, copy the content of the old page to the new one.
2822 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2823 * - Take the PTL. If the pte changed, bail out and release the allocated page
2824 * - If the pte is still the way we remember it, update the page table and all
2825 * relevant references. This includes dropping the reference the page-table
2826 * held to the old page, as well as updating the rmap.
2827 * - In any case, unlock the PTL and drop the reference we took to the old page.
2829 static vm_fault_t
wp_page_copy(struct vm_fault
*vmf
)
2831 struct vm_area_struct
*vma
= vmf
->vma
;
2832 struct mm_struct
*mm
= vma
->vm_mm
;
2833 struct page
*old_page
= vmf
->page
;
2834 struct page
*new_page
= NULL
;
2836 int page_copied
= 0;
2837 struct mmu_notifier_range range
;
2839 if (unlikely(anon_vma_prepare(vma
)))
2842 if (is_zero_pfn(pte_pfn(vmf
->orig_pte
))) {
2843 new_page
= alloc_zeroed_user_highpage_movable(vma
,
2848 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
2853 if (!cow_user_page(new_page
, old_page
, vmf
)) {
2855 * COW failed, if the fault was solved by other,
2856 * it's fine. If not, userspace would re-fault on
2857 * the same address and we will handle the fault
2858 * from the second attempt.
2867 if (mem_cgroup_charge(new_page
, mm
, GFP_KERNEL
))
2869 cgroup_throttle_swaprate(new_page
, GFP_KERNEL
);
2871 __SetPageUptodate(new_page
);
2873 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
,
2874 vmf
->address
& PAGE_MASK
,
2875 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
);
2876 mmu_notifier_invalidate_range_start(&range
);
2879 * Re-check the pte - we dropped the lock
2881 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, vmf
->address
, &vmf
->ptl
);
2882 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2884 if (!PageAnon(old_page
)) {
2885 dec_mm_counter_fast(mm
,
2886 mm_counter_file(old_page
));
2887 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2890 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2892 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2893 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2894 entry
= pte_sw_mkyoung(entry
);
2895 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2898 * Clear the pte entry and flush it first, before updating the
2899 * pte with the new entry, to keep TLBs on different CPUs in
2900 * sync. This code used to set the new PTE then flush TLBs, but
2901 * that left a window where the new PTE could be loaded into
2902 * some TLBs while the old PTE remains in others.
2904 ptep_clear_flush_notify(vma
, vmf
->address
, vmf
->pte
);
2905 page_add_new_anon_rmap(new_page
, vma
, vmf
->address
, false);
2906 lru_cache_add_inactive_or_unevictable(new_page
, vma
);
2908 * We call the notify macro here because, when using secondary
2909 * mmu page tables (such as kvm shadow page tables), we want the
2910 * new page to be mapped directly into the secondary page table.
2912 set_pte_at_notify(mm
, vmf
->address
, vmf
->pte
, entry
);
2913 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2916 * Only after switching the pte to the new page may
2917 * we remove the mapcount here. Otherwise another
2918 * process may come and find the rmap count decremented
2919 * before the pte is switched to the new page, and
2920 * "reuse" the old page writing into it while our pte
2921 * here still points into it and can be read by other
2924 * The critical issue is to order this
2925 * page_remove_rmap with the ptp_clear_flush above.
2926 * Those stores are ordered by (if nothing else,)
2927 * the barrier present in the atomic_add_negative
2928 * in page_remove_rmap.
2930 * Then the TLB flush in ptep_clear_flush ensures that
2931 * no process can access the old page before the
2932 * decremented mapcount is visible. And the old page
2933 * cannot be reused until after the decremented
2934 * mapcount is visible. So transitively, TLBs to
2935 * old page will be flushed before it can be reused.
2937 page_remove_rmap(old_page
, false);
2940 /* Free the old page.. */
2941 new_page
= old_page
;
2944 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
2950 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2952 * No need to double call mmu_notifier->invalidate_range() callback as
2953 * the above ptep_clear_flush_notify() did already call it.
2955 mmu_notifier_invalidate_range_only_end(&range
);
2958 * Don't let another task, with possibly unlocked vma,
2959 * keep the mlocked page.
2961 if (page_copied
&& (vma
->vm_flags
& VM_LOCKED
)) {
2962 lock_page(old_page
); /* LRU manipulation */
2963 if (PageMlocked(old_page
))
2964 munlock_vma_page(old_page
);
2965 unlock_page(old_page
);
2969 return page_copied
? VM_FAULT_WRITE
: 0;
2975 return VM_FAULT_OOM
;
2979 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2980 * writeable once the page is prepared
2982 * @vmf: structure describing the fault
2984 * This function handles all that is needed to finish a write page fault in a
2985 * shared mapping due to PTE being read-only once the mapped page is prepared.
2986 * It handles locking of PTE and modifying it.
2988 * The function expects the page to be locked or other protection against
2989 * concurrent faults / writeback (such as DAX radix tree locks).
2991 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2992 * we acquired PTE lock.
2994 vm_fault_t
finish_mkwrite_fault(struct vm_fault
*vmf
)
2996 WARN_ON_ONCE(!(vmf
->vma
->vm_flags
& VM_SHARED
));
2997 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3000 * We might have raced with another page fault while we released the
3001 * pte_offset_map_lock.
3003 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
3004 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
3005 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3006 return VM_FAULT_NOPAGE
;
3013 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3016 static vm_fault_t
wp_pfn_shared(struct vm_fault
*vmf
)
3018 struct vm_area_struct
*vma
= vmf
->vma
;
3020 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
3023 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3024 vmf
->flags
|= FAULT_FLAG_MKWRITE
;
3025 ret
= vma
->vm_ops
->pfn_mkwrite(vmf
);
3026 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))
3028 return finish_mkwrite_fault(vmf
);
3031 return VM_FAULT_WRITE
;
3034 static vm_fault_t
wp_page_shared(struct vm_fault
*vmf
)
3035 __releases(vmf
->ptl
)
3037 struct vm_area_struct
*vma
= vmf
->vma
;
3038 vm_fault_t ret
= VM_FAULT_WRITE
;
3040 get_page(vmf
->page
);
3042 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
3045 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3046 tmp
= do_page_mkwrite(vmf
);
3047 if (unlikely(!tmp
|| (tmp
&
3048 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3049 put_page(vmf
->page
);
3052 tmp
= finish_mkwrite_fault(vmf
);
3053 if (unlikely(tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3054 unlock_page(vmf
->page
);
3055 put_page(vmf
->page
);
3060 lock_page(vmf
->page
);
3062 ret
|= fault_dirty_shared_page(vmf
);
3063 put_page(vmf
->page
);
3069 * This routine handles present pages, when users try to write
3070 * to a shared page. It is done by copying the page to a new address
3071 * and decrementing the shared-page counter for the old page.
3073 * Note that this routine assumes that the protection checks have been
3074 * done by the caller (the low-level page fault routine in most cases).
3075 * Thus we can safely just mark it writable once we've done any necessary
3078 * We also mark the page dirty at this point even though the page will
3079 * change only once the write actually happens. This avoids a few races,
3080 * and potentially makes it more efficient.
3082 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3083 * but allow concurrent faults), with pte both mapped and locked.
3084 * We return with mmap_lock still held, but pte unmapped and unlocked.
3086 static vm_fault_t
do_wp_page(struct vm_fault
*vmf
)
3087 __releases(vmf
->ptl
)
3089 struct vm_area_struct
*vma
= vmf
->vma
;
3091 if (userfaultfd_pte_wp(vma
, *vmf
->pte
)) {
3092 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3093 return handle_userfault(vmf
, VM_UFFD_WP
);
3097 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3098 * is flushed in this case before copying.
3100 if (unlikely(userfaultfd_wp(vmf
->vma
) &&
3101 mm_tlb_flush_pending(vmf
->vma
->vm_mm
)))
3102 flush_tlb_page(vmf
->vma
, vmf
->address
);
3104 vmf
->page
= vm_normal_page(vma
, vmf
->address
, vmf
->orig_pte
);
3107 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3110 * We should not cow pages in a shared writeable mapping.
3111 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3113 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
3114 (VM_WRITE
|VM_SHARED
))
3115 return wp_pfn_shared(vmf
);
3117 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3118 return wp_page_copy(vmf
);
3122 * Take out anonymous pages first, anonymous shared vmas are
3123 * not dirty accountable.
3125 if (PageAnon(vmf
->page
)) {
3126 struct page
*page
= vmf
->page
;
3128 /* PageKsm() doesn't necessarily raise the page refcount */
3129 if (PageKsm(page
) || page_count(page
) != 1)
3131 if (!trylock_page(page
))
3133 if (PageKsm(page
) || page_mapcount(page
) != 1 || page_count(page
) != 1) {
3138 * Ok, we've got the only map reference, and the only
3139 * page count reference, and the page is locked,
3140 * it's dark out, and we're wearing sunglasses. Hit it.
3144 return VM_FAULT_WRITE
;
3145 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
3146 (VM_WRITE
|VM_SHARED
))) {
3147 return wp_page_shared(vmf
);
3151 * Ok, we need to copy. Oh, well..
3153 get_page(vmf
->page
);
3155 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3156 return wp_page_copy(vmf
);
3159 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
3160 unsigned long start_addr
, unsigned long end_addr
,
3161 struct zap_details
*details
)
3163 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
3166 static inline void unmap_mapping_range_tree(struct rb_root_cached
*root
,
3167 struct zap_details
*details
)
3169 struct vm_area_struct
*vma
;
3170 pgoff_t vba
, vea
, zba
, zea
;
3172 vma_interval_tree_foreach(vma
, root
,
3173 details
->first_index
, details
->last_index
) {
3175 vba
= vma
->vm_pgoff
;
3176 vea
= vba
+ vma_pages(vma
) - 1;
3177 zba
= details
->first_index
;
3180 zea
= details
->last_index
;
3184 unmap_mapping_range_vma(vma
,
3185 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
3186 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
3192 * unmap_mapping_pages() - Unmap pages from processes.
3193 * @mapping: The address space containing pages to be unmapped.
3194 * @start: Index of first page to be unmapped.
3195 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3196 * @even_cows: Whether to unmap even private COWed pages.
3198 * Unmap the pages in this address space from any userspace process which
3199 * has them mmaped. Generally, you want to remove COWed pages as well when
3200 * a file is being truncated, but not when invalidating pages from the page
3203 void unmap_mapping_pages(struct address_space
*mapping
, pgoff_t start
,
3204 pgoff_t nr
, bool even_cows
)
3206 struct zap_details details
= { };
3208 details
.check_mapping
= even_cows
? NULL
: mapping
;
3209 details
.first_index
= start
;
3210 details
.last_index
= start
+ nr
- 1;
3211 if (details
.last_index
< details
.first_index
)
3212 details
.last_index
= ULONG_MAX
;
3214 i_mmap_lock_write(mapping
);
3215 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3216 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
3217 i_mmap_unlock_write(mapping
);
3221 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3222 * address_space corresponding to the specified byte range in the underlying
3225 * @mapping: the address space containing mmaps to be unmapped.
3226 * @holebegin: byte in first page to unmap, relative to the start of
3227 * the underlying file. This will be rounded down to a PAGE_SIZE
3228 * boundary. Note that this is different from truncate_pagecache(), which
3229 * must keep the partial page. In contrast, we must get rid of
3231 * @holelen: size of prospective hole in bytes. This will be rounded
3232 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3234 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3235 * but 0 when invalidating pagecache, don't throw away private data.
3237 void unmap_mapping_range(struct address_space
*mapping
,
3238 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
3240 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
3241 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3243 /* Check for overflow. */
3244 if (sizeof(holelen
) > sizeof(hlen
)) {
3246 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3247 if (holeend
& ~(long long)ULONG_MAX
)
3248 hlen
= ULONG_MAX
- hba
+ 1;
3251 unmap_mapping_pages(mapping
, hba
, hlen
, even_cows
);
3253 EXPORT_SYMBOL(unmap_mapping_range
);
3256 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3257 * but allow concurrent faults), and pte mapped but not yet locked.
3258 * We return with pte unmapped and unlocked.
3260 * We return with the mmap_lock locked or unlocked in the same cases
3261 * as does filemap_fault().
3263 vm_fault_t
do_swap_page(struct vm_fault
*vmf
)
3265 struct vm_area_struct
*vma
= vmf
->vma
;
3266 struct page
*page
= NULL
, *swapcache
;
3272 void *shadow
= NULL
;
3274 if (!pte_unmap_same(vma
->vm_mm
, vmf
->pmd
, vmf
->pte
, vmf
->orig_pte
))
3277 entry
= pte_to_swp_entry(vmf
->orig_pte
);
3278 if (unlikely(non_swap_entry(entry
))) {
3279 if (is_migration_entry(entry
)) {
3280 migration_entry_wait(vma
->vm_mm
, vmf
->pmd
,
3282 } else if (is_device_private_entry(entry
)) {
3283 vmf
->page
= device_private_entry_to_page(entry
);
3284 ret
= vmf
->page
->pgmap
->ops
->migrate_to_ram(vmf
);
3285 } else if (is_hwpoison_entry(entry
)) {
3286 ret
= VM_FAULT_HWPOISON
;
3288 print_bad_pte(vma
, vmf
->address
, vmf
->orig_pte
, NULL
);
3289 ret
= VM_FAULT_SIGBUS
;
3295 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
3296 page
= lookup_swap_cache(entry
, vma
, vmf
->address
);
3300 struct swap_info_struct
*si
= swp_swap_info(entry
);
3302 if (data_race(si
->flags
& SWP_SYNCHRONOUS_IO
) &&
3303 __swap_count(entry
) == 1) {
3304 /* skip swapcache */
3305 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
3310 __SetPageLocked(page
);
3311 __SetPageSwapBacked(page
);
3312 set_page_private(page
, entry
.val
);
3314 /* Tell memcg to use swap ownership records */
3315 SetPageSwapCache(page
);
3316 err
= mem_cgroup_charge(page
, vma
->vm_mm
,
3318 ClearPageSwapCache(page
);
3324 shadow
= get_shadow_from_swap_cache(entry
);
3326 workingset_refault(page
, shadow
);
3328 lru_cache_add(page
);
3329 swap_readpage(page
, true);
3332 page
= swapin_readahead(entry
, GFP_HIGHUSER_MOVABLE
,
3339 * Back out if somebody else faulted in this pte
3340 * while we released the pte lock.
3342 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3343 vmf
->address
, &vmf
->ptl
);
3344 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3346 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3350 /* Had to read the page from swap area: Major fault */
3351 ret
= VM_FAULT_MAJOR
;
3352 count_vm_event(PGMAJFAULT
);
3353 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
3354 } else if (PageHWPoison(page
)) {
3356 * hwpoisoned dirty swapcache pages are kept for killing
3357 * owner processes (which may be unknown at hwpoison time)
3359 ret
= VM_FAULT_HWPOISON
;
3360 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3364 locked
= lock_page_or_retry(page
, vma
->vm_mm
, vmf
->flags
);
3366 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3368 ret
|= VM_FAULT_RETRY
;
3373 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3374 * release the swapcache from under us. The page pin, and pte_same
3375 * test below, are not enough to exclude that. Even if it is still
3376 * swapcache, we need to check that the page's swap has not changed.
3378 if (unlikely((!PageSwapCache(page
) ||
3379 page_private(page
) != entry
.val
)) && swapcache
)
3382 page
= ksm_might_need_to_copy(page
, vma
, vmf
->address
);
3383 if (unlikely(!page
)) {
3389 cgroup_throttle_swaprate(page
, GFP_KERNEL
);
3392 * Back out if somebody else already faulted in this pte.
3394 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3396 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3399 if (unlikely(!PageUptodate(page
))) {
3400 ret
= VM_FAULT_SIGBUS
;
3405 * The page isn't present yet, go ahead with the fault.
3407 * Be careful about the sequence of operations here.
3408 * To get its accounting right, reuse_swap_page() must be called
3409 * while the page is counted on swap but not yet in mapcount i.e.
3410 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3411 * must be called after the swap_free(), or it will never succeed.
3414 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3415 dec_mm_counter_fast(vma
->vm_mm
, MM_SWAPENTS
);
3416 pte
= mk_pte(page
, vma
->vm_page_prot
);
3417 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
, NULL
)) {
3418 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3419 vmf
->flags
&= ~FAULT_FLAG_WRITE
;
3420 ret
|= VM_FAULT_WRITE
;
3421 exclusive
= RMAP_EXCLUSIVE
;
3423 flush_icache_page(vma
, page
);
3424 if (pte_swp_soft_dirty(vmf
->orig_pte
))
3425 pte
= pte_mksoft_dirty(pte
);
3426 if (pte_swp_uffd_wp(vmf
->orig_pte
)) {
3427 pte
= pte_mkuffd_wp(pte
);
3428 pte
= pte_wrprotect(pte
);
3430 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3431 arch_do_swap_page(vma
->vm_mm
, vma
, vmf
->address
, pte
, vmf
->orig_pte
);
3432 vmf
->orig_pte
= pte
;
3434 /* ksm created a completely new copy */
3435 if (unlikely(page
!= swapcache
&& swapcache
)) {
3436 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3437 lru_cache_add_inactive_or_unevictable(page
, vma
);
3439 do_page_add_anon_rmap(page
, vma
, vmf
->address
, exclusive
);
3443 if (mem_cgroup_swap_full(page
) ||
3444 (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
3445 try_to_free_swap(page
);
3447 if (page
!= swapcache
&& swapcache
) {
3449 * Hold the lock to avoid the swap entry to be reused
3450 * until we take the PT lock for the pte_same() check
3451 * (to avoid false positives from pte_same). For
3452 * further safety release the lock after the swap_free
3453 * so that the swap count won't change under a
3454 * parallel locked swapcache.
3456 unlock_page(swapcache
);
3457 put_page(swapcache
);
3460 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
3461 ret
|= do_wp_page(vmf
);
3462 if (ret
& VM_FAULT_ERROR
)
3463 ret
&= VM_FAULT_ERROR
;
3467 /* No need to invalidate - it was non-present before */
3468 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3470 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3474 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3479 if (page
!= swapcache
&& swapcache
) {
3480 unlock_page(swapcache
);
3481 put_page(swapcache
);
3487 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3488 * but allow concurrent faults), and pte mapped but not yet locked.
3489 * We return with mmap_lock still held, but pte unmapped and unlocked.
3491 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
)
3493 struct vm_area_struct
*vma
= vmf
->vma
;
3498 /* File mapping without ->vm_ops ? */
3499 if (vma
->vm_flags
& VM_SHARED
)
3500 return VM_FAULT_SIGBUS
;
3503 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3504 * pte_offset_map() on pmds where a huge pmd might be created
3505 * from a different thread.
3507 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3508 * parallel threads are excluded by other means.
3510 * Here we only have mmap_read_lock(mm).
3512 if (pte_alloc(vma
->vm_mm
, vmf
->pmd
))
3513 return VM_FAULT_OOM
;
3515 /* See the comment in pte_alloc_one_map() */
3516 if (unlikely(pmd_trans_unstable(vmf
->pmd
)))
3519 /* Use the zero-page for reads */
3520 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
3521 !mm_forbids_zeropage(vma
->vm_mm
)) {
3522 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(vmf
->address
),
3523 vma
->vm_page_prot
));
3524 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3525 vmf
->address
, &vmf
->ptl
);
3526 if (!pte_none(*vmf
->pte
)) {
3527 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
3530 ret
= check_stable_address_space(vma
->vm_mm
);
3533 /* Deliver the page fault to userland, check inside PT lock */
3534 if (userfaultfd_missing(vma
)) {
3535 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3536 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3541 /* Allocate our own private page. */
3542 if (unlikely(anon_vma_prepare(vma
)))
3544 page
= alloc_zeroed_user_highpage_movable(vma
, vmf
->address
);
3548 if (mem_cgroup_charge(page
, vma
->vm_mm
, GFP_KERNEL
))
3550 cgroup_throttle_swaprate(page
, GFP_KERNEL
);
3553 * The memory barrier inside __SetPageUptodate makes sure that
3554 * preceding stores to the page contents become visible before
3555 * the set_pte_at() write.
3557 __SetPageUptodate(page
);
3559 entry
= mk_pte(page
, vma
->vm_page_prot
);
3560 entry
= pte_sw_mkyoung(entry
);
3561 if (vma
->vm_flags
& VM_WRITE
)
3562 entry
= pte_mkwrite(pte_mkdirty(entry
));
3564 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3566 if (!pte_none(*vmf
->pte
)) {
3567 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3571 ret
= check_stable_address_space(vma
->vm_mm
);
3575 /* Deliver the page fault to userland, check inside PT lock */
3576 if (userfaultfd_missing(vma
)) {
3577 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3579 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3582 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3583 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3584 lru_cache_add_inactive_or_unevictable(page
, vma
);
3586 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3588 /* No need to invalidate - it was non-present before */
3589 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3591 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3599 return VM_FAULT_OOM
;
3603 * The mmap_lock must have been held on entry, and may have been
3604 * released depending on flags and vma->vm_ops->fault() return value.
3605 * See filemap_fault() and __lock_page_retry().
3607 static vm_fault_t
__do_fault(struct vm_fault
*vmf
)
3609 struct vm_area_struct
*vma
= vmf
->vma
;
3613 * Preallocate pte before we take page_lock because this might lead to
3614 * deadlocks for memcg reclaim which waits for pages under writeback:
3616 * SetPageWriteback(A)
3622 * wait_on_page_writeback(A)
3623 * SetPageWriteback(B)
3625 * # flush A, B to clear the writeback
3627 if (pmd_none(*vmf
->pmd
) && !vmf
->prealloc_pte
) {
3628 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
3629 if (!vmf
->prealloc_pte
)
3630 return VM_FAULT_OOM
;
3631 smp_wmb(); /* See comment in __pte_alloc() */
3634 ret
= vma
->vm_ops
->fault(vmf
);
3635 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
|
3636 VM_FAULT_DONE_COW
)))
3639 if (unlikely(PageHWPoison(vmf
->page
))) {
3640 if (ret
& VM_FAULT_LOCKED
)
3641 unlock_page(vmf
->page
);
3642 put_page(vmf
->page
);
3644 return VM_FAULT_HWPOISON
;
3647 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3648 lock_page(vmf
->page
);
3650 VM_BUG_ON_PAGE(!PageLocked(vmf
->page
), vmf
->page
);
3656 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3657 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3658 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3659 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3661 static int pmd_devmap_trans_unstable(pmd_t
*pmd
)
3663 return pmd_devmap(*pmd
) || pmd_trans_unstable(pmd
);
3666 static vm_fault_t
pte_alloc_one_map(struct vm_fault
*vmf
)
3668 struct vm_area_struct
*vma
= vmf
->vma
;
3670 if (!pmd_none(*vmf
->pmd
))
3672 if (vmf
->prealloc_pte
) {
3673 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3674 if (unlikely(!pmd_none(*vmf
->pmd
))) {
3675 spin_unlock(vmf
->ptl
);
3679 mm_inc_nr_ptes(vma
->vm_mm
);
3680 pmd_populate(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3681 spin_unlock(vmf
->ptl
);
3682 vmf
->prealloc_pte
= NULL
;
3683 } else if (unlikely(pte_alloc(vma
->vm_mm
, vmf
->pmd
))) {
3684 return VM_FAULT_OOM
;
3688 * If a huge pmd materialized under us just retry later. Use
3689 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3690 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3691 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3692 * running immediately after a huge pmd fault in a different thread of
3693 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3694 * All we have to ensure is that it is a regular pmd that we can walk
3695 * with pte_offset_map() and we can do that through an atomic read in
3696 * C, which is what pmd_trans_unstable() provides.
3698 if (pmd_devmap_trans_unstable(vmf
->pmd
))
3699 return VM_FAULT_NOPAGE
;
3702 * At this point we know that our vmf->pmd points to a page of ptes
3703 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3704 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3705 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3706 * be valid and we will re-check to make sure the vmf->pte isn't
3707 * pte_none() under vmf->ptl protection when we return to
3710 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3715 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3716 static void deposit_prealloc_pte(struct vm_fault
*vmf
)
3718 struct vm_area_struct
*vma
= vmf
->vma
;
3720 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3722 * We are going to consume the prealloc table,
3723 * count that as nr_ptes.
3725 mm_inc_nr_ptes(vma
->vm_mm
);
3726 vmf
->prealloc_pte
= NULL
;
3729 static vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3731 struct vm_area_struct
*vma
= vmf
->vma
;
3732 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3733 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
3736 vm_fault_t ret
= VM_FAULT_FALLBACK
;
3738 if (!transhuge_vma_suitable(vma
, haddr
))
3741 page
= compound_head(page
);
3742 if (compound_order(page
) != HPAGE_PMD_ORDER
)
3746 * Archs like ppc64 need additonal space to store information
3747 * related to pte entry. Use the preallocated table for that.
3749 if (arch_needs_pgtable_deposit() && !vmf
->prealloc_pte
) {
3750 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
3751 if (!vmf
->prealloc_pte
)
3752 return VM_FAULT_OOM
;
3753 smp_wmb(); /* See comment in __pte_alloc() */
3756 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3757 if (unlikely(!pmd_none(*vmf
->pmd
)))
3760 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
3761 flush_icache_page(vma
, page
+ i
);
3763 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
3765 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
3767 add_mm_counter(vma
->vm_mm
, mm_counter_file(page
), HPAGE_PMD_NR
);
3768 page_add_file_rmap(page
, true);
3770 * deposit and withdraw with pmd lock held
3772 if (arch_needs_pgtable_deposit())
3773 deposit_prealloc_pte(vmf
);
3775 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
3777 update_mmu_cache_pmd(vma
, haddr
, vmf
->pmd
);
3779 /* fault is handled */
3781 count_vm_event(THP_FILE_MAPPED
);
3783 spin_unlock(vmf
->ptl
);
3787 static vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3795 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3796 * mapping. If needed, the function allocates page table or use pre-allocated.
3798 * @vmf: fault environment
3799 * @page: page to map
3801 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3804 * Target users are page handler itself and implementations of
3805 * vm_ops->map_pages.
3807 * Return: %0 on success, %VM_FAULT_ code in case of error.
3809 vm_fault_t
alloc_set_pte(struct vm_fault
*vmf
, struct page
*page
)
3811 struct vm_area_struct
*vma
= vmf
->vma
;
3812 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3816 if (pmd_none(*vmf
->pmd
) && PageTransCompound(page
)) {
3817 ret
= do_set_pmd(vmf
, page
);
3818 if (ret
!= VM_FAULT_FALLBACK
)
3823 ret
= pte_alloc_one_map(vmf
);
3828 /* Re-check under ptl */
3829 if (unlikely(!pte_none(*vmf
->pte
))) {
3830 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
3831 return VM_FAULT_NOPAGE
;
3834 flush_icache_page(vma
, page
);
3835 entry
= mk_pte(page
, vma
->vm_page_prot
);
3836 entry
= pte_sw_mkyoung(entry
);
3838 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3839 /* copy-on-write page */
3840 if (write
&& !(vma
->vm_flags
& VM_SHARED
)) {
3841 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3842 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3843 lru_cache_add_inactive_or_unevictable(page
, vma
);
3845 inc_mm_counter_fast(vma
->vm_mm
, mm_counter_file(page
));
3846 page_add_file_rmap(page
, false);
3848 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3850 /* no need to invalidate: a not-present page won't be cached */
3851 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3858 * finish_fault - finish page fault once we have prepared the page to fault
3860 * @vmf: structure describing the fault
3862 * This function handles all that is needed to finish a page fault once the
3863 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3864 * given page, adds reverse page mapping, handles memcg charges and LRU
3867 * The function expects the page to be locked and on success it consumes a
3868 * reference of a page being mapped (for the PTE which maps it).
3870 * Return: %0 on success, %VM_FAULT_ code in case of error.
3872 vm_fault_t
finish_fault(struct vm_fault
*vmf
)
3877 /* Did we COW the page? */
3878 if ((vmf
->flags
& FAULT_FLAG_WRITE
) &&
3879 !(vmf
->vma
->vm_flags
& VM_SHARED
))
3880 page
= vmf
->cow_page
;
3885 * check even for read faults because we might have lost our CoWed
3888 if (!(vmf
->vma
->vm_flags
& VM_SHARED
))
3889 ret
= check_stable_address_space(vmf
->vma
->vm_mm
);
3891 ret
= alloc_set_pte(vmf
, page
);
3893 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3897 static unsigned long fault_around_bytes __read_mostly
=
3898 rounddown_pow_of_two(65536);
3900 #ifdef CONFIG_DEBUG_FS
3901 static int fault_around_bytes_get(void *data
, u64
*val
)
3903 *val
= fault_around_bytes
;
3908 * fault_around_bytes must be rounded down to the nearest page order as it's
3909 * what do_fault_around() expects to see.
3911 static int fault_around_bytes_set(void *data
, u64 val
)
3913 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
3915 if (val
> PAGE_SIZE
)
3916 fault_around_bytes
= rounddown_pow_of_two(val
);
3918 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
3921 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops
,
3922 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
3924 static int __init
fault_around_debugfs(void)
3926 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL
, NULL
,
3927 &fault_around_bytes_fops
);
3930 late_initcall(fault_around_debugfs
);
3934 * do_fault_around() tries to map few pages around the fault address. The hope
3935 * is that the pages will be needed soon and this will lower the number of
3938 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3939 * not ready to be mapped: not up-to-date, locked, etc.
3941 * This function is called with the page table lock taken. In the split ptlock
3942 * case the page table lock only protects only those entries which belong to
3943 * the page table corresponding to the fault address.
3945 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3948 * fault_around_bytes defines how many bytes we'll try to map.
3949 * do_fault_around() expects it to be set to a power of two less than or equal
3952 * The virtual address of the area that we map is naturally aligned to
3953 * fault_around_bytes rounded down to the machine page size
3954 * (and therefore to page order). This way it's easier to guarantee
3955 * that we don't cross page table boundaries.
3957 static vm_fault_t
do_fault_around(struct vm_fault
*vmf
)
3959 unsigned long address
= vmf
->address
, nr_pages
, mask
;
3960 pgoff_t start_pgoff
= vmf
->pgoff
;
3965 nr_pages
= READ_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
3966 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
3968 vmf
->address
= max(address
& mask
, vmf
->vma
->vm_start
);
3969 off
= ((address
- vmf
->address
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
3973 * end_pgoff is either the end of the page table, the end of
3974 * the vma or nr_pages from start_pgoff, depending what is nearest.
3976 end_pgoff
= start_pgoff
-
3977 ((vmf
->address
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
3979 end_pgoff
= min3(end_pgoff
, vma_pages(vmf
->vma
) + vmf
->vma
->vm_pgoff
- 1,
3980 start_pgoff
+ nr_pages
- 1);
3982 if (pmd_none(*vmf
->pmd
)) {
3983 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
);
3984 if (!vmf
->prealloc_pte
)
3986 smp_wmb(); /* See comment in __pte_alloc() */
3989 vmf
->vma
->vm_ops
->map_pages(vmf
, start_pgoff
, end_pgoff
);
3991 /* Huge page is mapped? Page fault is solved */
3992 if (pmd_trans_huge(*vmf
->pmd
)) {
3993 ret
= VM_FAULT_NOPAGE
;
3997 /* ->map_pages() haven't done anything useful. Cold page cache? */
4001 /* check if the page fault is solved */
4002 vmf
->pte
-= (vmf
->address
>> PAGE_SHIFT
) - (address
>> PAGE_SHIFT
);
4003 if (!pte_none(*vmf
->pte
))
4004 ret
= VM_FAULT_NOPAGE
;
4005 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4007 vmf
->address
= address
;
4012 static vm_fault_t
do_read_fault(struct vm_fault
*vmf
)
4014 struct vm_area_struct
*vma
= vmf
->vma
;
4018 * Let's call ->map_pages() first and use ->fault() as fallback
4019 * if page by the offset is not ready to be mapped (cold cache or
4022 if (vma
->vm_ops
->map_pages
&& fault_around_bytes
>> PAGE_SHIFT
> 1) {
4023 ret
= do_fault_around(vmf
);
4028 ret
= __do_fault(vmf
);
4029 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4032 ret
|= finish_fault(vmf
);
4033 unlock_page(vmf
->page
);
4034 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4035 put_page(vmf
->page
);
4039 static vm_fault_t
do_cow_fault(struct vm_fault
*vmf
)
4041 struct vm_area_struct
*vma
= vmf
->vma
;
4044 if (unlikely(anon_vma_prepare(vma
)))
4045 return VM_FAULT_OOM
;
4047 vmf
->cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, vmf
->address
);
4049 return VM_FAULT_OOM
;
4051 if (mem_cgroup_charge(vmf
->cow_page
, vma
->vm_mm
, GFP_KERNEL
)) {
4052 put_page(vmf
->cow_page
);
4053 return VM_FAULT_OOM
;
4055 cgroup_throttle_swaprate(vmf
->cow_page
, GFP_KERNEL
);
4057 ret
= __do_fault(vmf
);
4058 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4060 if (ret
& VM_FAULT_DONE_COW
)
4063 copy_user_highpage(vmf
->cow_page
, vmf
->page
, vmf
->address
, vma
);
4064 __SetPageUptodate(vmf
->cow_page
);
4066 ret
|= finish_fault(vmf
);
4067 unlock_page(vmf
->page
);
4068 put_page(vmf
->page
);
4069 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4073 put_page(vmf
->cow_page
);
4077 static vm_fault_t
do_shared_fault(struct vm_fault
*vmf
)
4079 struct vm_area_struct
*vma
= vmf
->vma
;
4080 vm_fault_t ret
, tmp
;
4082 ret
= __do_fault(vmf
);
4083 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4087 * Check if the backing address space wants to know that the page is
4088 * about to become writable
4090 if (vma
->vm_ops
->page_mkwrite
) {
4091 unlock_page(vmf
->page
);
4092 tmp
= do_page_mkwrite(vmf
);
4093 if (unlikely(!tmp
||
4094 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
4095 put_page(vmf
->page
);
4100 ret
|= finish_fault(vmf
);
4101 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
4103 unlock_page(vmf
->page
);
4104 put_page(vmf
->page
);
4108 ret
|= fault_dirty_shared_page(vmf
);
4113 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4114 * but allow concurrent faults).
4115 * The mmap_lock may have been released depending on flags and our
4116 * return value. See filemap_fault() and __lock_page_or_retry().
4117 * If mmap_lock is released, vma may become invalid (for example
4118 * by other thread calling munmap()).
4120 static vm_fault_t
do_fault(struct vm_fault
*vmf
)
4122 struct vm_area_struct
*vma
= vmf
->vma
;
4123 struct mm_struct
*vm_mm
= vma
->vm_mm
;
4127 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4129 if (!vma
->vm_ops
->fault
) {
4131 * If we find a migration pmd entry or a none pmd entry, which
4132 * should never happen, return SIGBUS
4134 if (unlikely(!pmd_present(*vmf
->pmd
)))
4135 ret
= VM_FAULT_SIGBUS
;
4137 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
,
4142 * Make sure this is not a temporary clearing of pte
4143 * by holding ptl and checking again. A R/M/W update
4144 * of pte involves: take ptl, clearing the pte so that
4145 * we don't have concurrent modification by hardware
4146 * followed by an update.
4148 if (unlikely(pte_none(*vmf
->pte
)))
4149 ret
= VM_FAULT_SIGBUS
;
4151 ret
= VM_FAULT_NOPAGE
;
4153 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4155 } else if (!(vmf
->flags
& FAULT_FLAG_WRITE
))
4156 ret
= do_read_fault(vmf
);
4157 else if (!(vma
->vm_flags
& VM_SHARED
))
4158 ret
= do_cow_fault(vmf
);
4160 ret
= do_shared_fault(vmf
);
4162 /* preallocated pagetable is unused: free it */
4163 if (vmf
->prealloc_pte
) {
4164 pte_free(vm_mm
, vmf
->prealloc_pte
);
4165 vmf
->prealloc_pte
= NULL
;
4170 static int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
4171 unsigned long addr
, int page_nid
,
4176 count_vm_numa_event(NUMA_HINT_FAULTS
);
4177 if (page_nid
== numa_node_id()) {
4178 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
4179 *flags
|= TNF_FAULT_LOCAL
;
4182 return mpol_misplaced(page
, vma
, addr
);
4185 static vm_fault_t
do_numa_page(struct vm_fault
*vmf
)
4187 struct vm_area_struct
*vma
= vmf
->vma
;
4188 struct page
*page
= NULL
;
4189 int page_nid
= NUMA_NO_NODE
;
4192 bool migrated
= false;
4194 bool was_writable
= pte_savedwrite(vmf
->orig_pte
);
4198 * The "pte" at this point cannot be used safely without
4199 * validation through pte_unmap_same(). It's of NUMA type but
4200 * the pfn may be screwed if the read is non atomic.
4202 vmf
->ptl
= pte_lockptr(vma
->vm_mm
, vmf
->pmd
);
4203 spin_lock(vmf
->ptl
);
4204 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
4205 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4210 * Make it present again, Depending on how arch implementes non
4211 * accessible ptes, some can allow access by kernel mode.
4213 old_pte
= ptep_modify_prot_start(vma
, vmf
->address
, vmf
->pte
);
4214 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
4215 pte
= pte_mkyoung(pte
);
4217 pte
= pte_mkwrite(pte
);
4218 ptep_modify_prot_commit(vma
, vmf
->address
, vmf
->pte
, old_pte
, pte
);
4219 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
4221 page
= vm_normal_page(vma
, vmf
->address
, pte
);
4223 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4227 /* TODO: handle PTE-mapped THP */
4228 if (PageCompound(page
)) {
4229 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4234 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4235 * much anyway since they can be in shared cache state. This misses
4236 * the case where a mapping is writable but the process never writes
4237 * to it but pte_write gets cleared during protection updates and
4238 * pte_dirty has unpredictable behaviour between PTE scan updates,
4239 * background writeback, dirty balancing and application behaviour.
4241 if (!pte_write(pte
))
4242 flags
|= TNF_NO_GROUP
;
4245 * Flag if the page is shared between multiple address spaces. This
4246 * is later used when determining whether to group tasks together
4248 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
4249 flags
|= TNF_SHARED
;
4251 last_cpupid
= page_cpupid_last(page
);
4252 page_nid
= page_to_nid(page
);
4253 target_nid
= numa_migrate_prep(page
, vma
, vmf
->address
, page_nid
,
4255 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4256 if (target_nid
== NUMA_NO_NODE
) {
4261 /* Migrate to the requested node */
4262 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
4264 page_nid
= target_nid
;
4265 flags
|= TNF_MIGRATED
;
4267 flags
|= TNF_MIGRATE_FAIL
;
4270 if (page_nid
!= NUMA_NO_NODE
)
4271 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
4275 static inline vm_fault_t
create_huge_pmd(struct vm_fault
*vmf
)
4277 if (vma_is_anonymous(vmf
->vma
))
4278 return do_huge_pmd_anonymous_page(vmf
);
4279 if (vmf
->vma
->vm_ops
->huge_fault
)
4280 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
4281 return VM_FAULT_FALLBACK
;
4284 /* `inline' is required to avoid gcc 4.1.2 build error */
4285 static inline vm_fault_t
wp_huge_pmd(struct vm_fault
*vmf
, pmd_t orig_pmd
)
4287 if (vma_is_anonymous(vmf
->vma
)) {
4288 if (userfaultfd_huge_pmd_wp(vmf
->vma
, orig_pmd
))
4289 return handle_userfault(vmf
, VM_UFFD_WP
);
4290 return do_huge_pmd_wp_page(vmf
, orig_pmd
);
4292 if (vmf
->vma
->vm_ops
->huge_fault
) {
4293 vm_fault_t ret
= vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
4295 if (!(ret
& VM_FAULT_FALLBACK
))
4299 /* COW or write-notify handled on pte level: split pmd. */
4300 __split_huge_pmd(vmf
->vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
4302 return VM_FAULT_FALLBACK
;
4305 static vm_fault_t
create_huge_pud(struct vm_fault
*vmf
)
4307 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4308 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4309 /* No support for anonymous transparent PUD pages yet */
4310 if (vma_is_anonymous(vmf
->vma
))
4312 if (vmf
->vma
->vm_ops
->huge_fault
) {
4313 vm_fault_t ret
= vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
4315 if (!(ret
& VM_FAULT_FALLBACK
))
4319 /* COW or write-notify not handled on PUD level: split pud.*/
4320 __split_huge_pud(vmf
->vma
, vmf
->pud
, vmf
->address
);
4321 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4322 return VM_FAULT_FALLBACK
;
4325 static vm_fault_t
wp_huge_pud(struct vm_fault
*vmf
, pud_t orig_pud
)
4327 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4328 /* No support for anonymous transparent PUD pages yet */
4329 if (vma_is_anonymous(vmf
->vma
))
4330 return VM_FAULT_FALLBACK
;
4331 if (vmf
->vma
->vm_ops
->huge_fault
)
4332 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
4333 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4334 return VM_FAULT_FALLBACK
;
4338 * These routines also need to handle stuff like marking pages dirty
4339 * and/or accessed for architectures that don't do it in hardware (most
4340 * RISC architectures). The early dirtying is also good on the i386.
4342 * There is also a hook called "update_mmu_cache()" that architectures
4343 * with external mmu caches can use to update those (ie the Sparc or
4344 * PowerPC hashed page tables that act as extended TLBs).
4346 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4347 * concurrent faults).
4349 * The mmap_lock may have been released depending on flags and our return value.
4350 * See filemap_fault() and __lock_page_or_retry().
4352 static vm_fault_t
handle_pte_fault(struct vm_fault
*vmf
)
4356 if (unlikely(pmd_none(*vmf
->pmd
))) {
4358 * Leave __pte_alloc() until later: because vm_ops->fault may
4359 * want to allocate huge page, and if we expose page table
4360 * for an instant, it will be difficult to retract from
4361 * concurrent faults and from rmap lookups.
4365 /* See comment in pte_alloc_one_map() */
4366 if (pmd_devmap_trans_unstable(vmf
->pmd
))
4369 * A regular pmd is established and it can't morph into a huge
4370 * pmd from under us anymore at this point because we hold the
4371 * mmap_lock read mode and khugepaged takes it in write mode.
4372 * So now it's safe to run pte_offset_map().
4374 vmf
->pte
= pte_offset_map(vmf
->pmd
, vmf
->address
);
4375 vmf
->orig_pte
= *vmf
->pte
;
4378 * some architectures can have larger ptes than wordsize,
4379 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4380 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4381 * accesses. The code below just needs a consistent view
4382 * for the ifs and we later double check anyway with the
4383 * ptl lock held. So here a barrier will do.
4386 if (pte_none(vmf
->orig_pte
)) {
4387 pte_unmap(vmf
->pte
);
4393 if (vma_is_anonymous(vmf
->vma
))
4394 return do_anonymous_page(vmf
);
4396 return do_fault(vmf
);
4399 if (!pte_present(vmf
->orig_pte
))
4400 return do_swap_page(vmf
);
4402 if (pte_protnone(vmf
->orig_pte
) && vma_is_accessible(vmf
->vma
))
4403 return do_numa_page(vmf
);
4405 vmf
->ptl
= pte_lockptr(vmf
->vma
->vm_mm
, vmf
->pmd
);
4406 spin_lock(vmf
->ptl
);
4407 entry
= vmf
->orig_pte
;
4408 if (unlikely(!pte_same(*vmf
->pte
, entry
))) {
4409 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
4412 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4413 if (!pte_write(entry
))
4414 return do_wp_page(vmf
);
4415 entry
= pte_mkdirty(entry
);
4417 entry
= pte_mkyoung(entry
);
4418 if (ptep_set_access_flags(vmf
->vma
, vmf
->address
, vmf
->pte
, entry
,
4419 vmf
->flags
& FAULT_FLAG_WRITE
)) {
4420 update_mmu_cache(vmf
->vma
, vmf
->address
, vmf
->pte
);
4422 /* Skip spurious TLB flush for retried page fault */
4423 if (vmf
->flags
& FAULT_FLAG_TRIED
)
4426 * This is needed only for protection faults but the arch code
4427 * is not yet telling us if this is a protection fault or not.
4428 * This still avoids useless tlb flushes for .text page faults
4431 if (vmf
->flags
& FAULT_FLAG_WRITE
)
4432 flush_tlb_fix_spurious_fault(vmf
->vma
, vmf
->address
);
4435 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4440 * By the time we get here, we already hold the mm semaphore
4442 * The mmap_lock may have been released depending on flags and our
4443 * return value. See filemap_fault() and __lock_page_or_retry().
4445 static vm_fault_t
__handle_mm_fault(struct vm_area_struct
*vma
,
4446 unsigned long address
, unsigned int flags
)
4448 struct vm_fault vmf
= {
4450 .address
= address
& PAGE_MASK
,
4452 .pgoff
= linear_page_index(vma
, address
),
4453 .gfp_mask
= __get_fault_gfp_mask(vma
),
4455 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
4456 struct mm_struct
*mm
= vma
->vm_mm
;
4461 pgd
= pgd_offset(mm
, address
);
4462 p4d
= p4d_alloc(mm
, pgd
, address
);
4464 return VM_FAULT_OOM
;
4466 vmf
.pud
= pud_alloc(mm
, p4d
, address
);
4468 return VM_FAULT_OOM
;
4470 if (pud_none(*vmf
.pud
) && __transparent_hugepage_enabled(vma
)) {
4471 ret
= create_huge_pud(&vmf
);
4472 if (!(ret
& VM_FAULT_FALLBACK
))
4475 pud_t orig_pud
= *vmf
.pud
;
4478 if (pud_trans_huge(orig_pud
) || pud_devmap(orig_pud
)) {
4480 /* NUMA case for anonymous PUDs would go here */
4482 if (dirty
&& !pud_write(orig_pud
)) {
4483 ret
= wp_huge_pud(&vmf
, orig_pud
);
4484 if (!(ret
& VM_FAULT_FALLBACK
))
4487 huge_pud_set_accessed(&vmf
, orig_pud
);
4493 vmf
.pmd
= pmd_alloc(mm
, vmf
.pud
, address
);
4495 return VM_FAULT_OOM
;
4497 /* Huge pud page fault raced with pmd_alloc? */
4498 if (pud_trans_unstable(vmf
.pud
))
4501 if (pmd_none(*vmf
.pmd
) && __transparent_hugepage_enabled(vma
)) {
4502 ret
= create_huge_pmd(&vmf
);
4503 if (!(ret
& VM_FAULT_FALLBACK
))
4506 pmd_t orig_pmd
= *vmf
.pmd
;
4509 if (unlikely(is_swap_pmd(orig_pmd
))) {
4510 VM_BUG_ON(thp_migration_supported() &&
4511 !is_pmd_migration_entry(orig_pmd
));
4512 if (is_pmd_migration_entry(orig_pmd
))
4513 pmd_migration_entry_wait(mm
, vmf
.pmd
);
4516 if (pmd_trans_huge(orig_pmd
) || pmd_devmap(orig_pmd
)) {
4517 if (pmd_protnone(orig_pmd
) && vma_is_accessible(vma
))
4518 return do_huge_pmd_numa_page(&vmf
, orig_pmd
);
4520 if (dirty
&& !pmd_write(orig_pmd
)) {
4521 ret
= wp_huge_pmd(&vmf
, orig_pmd
);
4522 if (!(ret
& VM_FAULT_FALLBACK
))
4525 huge_pmd_set_accessed(&vmf
, orig_pmd
);
4531 return handle_pte_fault(&vmf
);
4535 * mm_account_fault - Do page fault accountings
4537 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4538 * of perf event counters, but we'll still do the per-task accounting to
4539 * the task who triggered this page fault.
4540 * @address: the faulted address.
4541 * @flags: the fault flags.
4542 * @ret: the fault retcode.
4544 * This will take care of most of the page fault accountings. Meanwhile, it
4545 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4546 * updates. However note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4547 * still be in per-arch page fault handlers at the entry of page fault.
4549 static inline void mm_account_fault(struct pt_regs
*regs
,
4550 unsigned long address
, unsigned int flags
,
4556 * We don't do accounting for some specific faults:
4558 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4559 * includes arch_vma_access_permitted() failing before reaching here.
4560 * So this is not a "this many hardware page faults" counter. We
4561 * should use the hw profiling for that.
4563 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4564 * once they're completed.
4566 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_RETRY
))
4570 * We define the fault as a major fault when the final successful fault
4571 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4572 * handle it immediately previously).
4574 major
= (ret
& VM_FAULT_MAJOR
) || (flags
& FAULT_FLAG_TRIED
);
4582 * If the fault is done for GUP, regs will be NULL. We only do the
4583 * accounting for the per thread fault counters who triggered the
4584 * fault, and we skip the perf event updates.
4590 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ
, 1, regs
, address
);
4592 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN
, 1, regs
, address
);
4596 * By the time we get here, we already hold the mm semaphore
4598 * The mmap_lock may have been released depending on flags and our
4599 * return value. See filemap_fault() and __lock_page_or_retry().
4601 vm_fault_t
handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
4602 unsigned int flags
, struct pt_regs
*regs
)
4606 __set_current_state(TASK_RUNNING
);
4608 count_vm_event(PGFAULT
);
4609 count_memcg_event_mm(vma
->vm_mm
, PGFAULT
);
4611 /* do counter updates before entering really critical section. */
4612 check_sync_rss_stat(current
);
4614 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
4615 flags
& FAULT_FLAG_INSTRUCTION
,
4616 flags
& FAULT_FLAG_REMOTE
))
4617 return VM_FAULT_SIGSEGV
;
4620 * Enable the memcg OOM handling for faults triggered in user
4621 * space. Kernel faults are handled more gracefully.
4623 if (flags
& FAULT_FLAG_USER
)
4624 mem_cgroup_enter_user_fault();
4626 if (unlikely(is_vm_hugetlb_page(vma
)))
4627 ret
= hugetlb_fault(vma
->vm_mm
, vma
, address
, flags
);
4629 ret
= __handle_mm_fault(vma
, address
, flags
);
4631 if (flags
& FAULT_FLAG_USER
) {
4632 mem_cgroup_exit_user_fault();
4634 * The task may have entered a memcg OOM situation but
4635 * if the allocation error was handled gracefully (no
4636 * VM_FAULT_OOM), there is no need to kill anything.
4637 * Just clean up the OOM state peacefully.
4639 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
4640 mem_cgroup_oom_synchronize(false);
4643 mm_account_fault(regs
, address
, flags
, ret
);
4647 EXPORT_SYMBOL_GPL(handle_mm_fault
);
4649 #ifndef __PAGETABLE_P4D_FOLDED
4651 * Allocate p4d page table.
4652 * We've already handled the fast-path in-line.
4654 int __p4d_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
4656 p4d_t
*new = p4d_alloc_one(mm
, address
);
4660 smp_wmb(); /* See comment in __pte_alloc */
4662 spin_lock(&mm
->page_table_lock
);
4663 if (pgd_present(*pgd
)) /* Another has populated it */
4666 pgd_populate(mm
, pgd
, new);
4667 spin_unlock(&mm
->page_table_lock
);
4670 #endif /* __PAGETABLE_P4D_FOLDED */
4672 #ifndef __PAGETABLE_PUD_FOLDED
4674 * Allocate page upper directory.
4675 * We've already handled the fast-path in-line.
4677 int __pud_alloc(struct mm_struct
*mm
, p4d_t
*p4d
, unsigned long address
)
4679 pud_t
*new = pud_alloc_one(mm
, address
);
4683 smp_wmb(); /* See comment in __pte_alloc */
4685 spin_lock(&mm
->page_table_lock
);
4686 if (!p4d_present(*p4d
)) {
4688 p4d_populate(mm
, p4d
, new);
4689 } else /* Another has populated it */
4691 spin_unlock(&mm
->page_table_lock
);
4694 #endif /* __PAGETABLE_PUD_FOLDED */
4696 #ifndef __PAGETABLE_PMD_FOLDED
4698 * Allocate page middle directory.
4699 * We've already handled the fast-path in-line.
4701 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
4704 pmd_t
*new = pmd_alloc_one(mm
, address
);
4708 smp_wmb(); /* See comment in __pte_alloc */
4710 ptl
= pud_lock(mm
, pud
);
4711 if (!pud_present(*pud
)) {
4713 pud_populate(mm
, pud
, new);
4714 } else /* Another has populated it */
4719 #endif /* __PAGETABLE_PMD_FOLDED */
4721 int follow_invalidate_pte(struct mm_struct
*mm
, unsigned long address
,
4722 struct mmu_notifier_range
*range
, pte_t
**ptepp
,
4723 pmd_t
**pmdpp
, spinlock_t
**ptlp
)
4731 pgd
= pgd_offset(mm
, address
);
4732 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
4735 p4d
= p4d_offset(pgd
, address
);
4736 if (p4d_none(*p4d
) || unlikely(p4d_bad(*p4d
)))
4739 pud
= pud_offset(p4d
, address
);
4740 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
4743 pmd
= pmd_offset(pud
, address
);
4744 VM_BUG_ON(pmd_trans_huge(*pmd
));
4746 if (pmd_huge(*pmd
)) {
4751 mmu_notifier_range_init(range
, MMU_NOTIFY_CLEAR
, 0,
4752 NULL
, mm
, address
& PMD_MASK
,
4753 (address
& PMD_MASK
) + PMD_SIZE
);
4754 mmu_notifier_invalidate_range_start(range
);
4756 *ptlp
= pmd_lock(mm
, pmd
);
4757 if (pmd_huge(*pmd
)) {
4763 mmu_notifier_invalidate_range_end(range
);
4766 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
4770 mmu_notifier_range_init(range
, MMU_NOTIFY_CLEAR
, 0, NULL
, mm
,
4771 address
& PAGE_MASK
,
4772 (address
& PAGE_MASK
) + PAGE_SIZE
);
4773 mmu_notifier_invalidate_range_start(range
);
4775 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
4776 if (!pte_present(*ptep
))
4781 pte_unmap_unlock(ptep
, *ptlp
);
4783 mmu_notifier_invalidate_range_end(range
);
4789 * follow_pte - look up PTE at a user virtual address
4790 * @mm: the mm_struct of the target address space
4791 * @address: user virtual address
4792 * @ptepp: location to store found PTE
4793 * @ptlp: location to store the lock for the PTE
4795 * On a successful return, the pointer to the PTE is stored in @ptepp;
4796 * the corresponding lock is taken and its location is stored in @ptlp.
4797 * The contents of the PTE are only stable until @ptlp is released;
4798 * any further use, if any, must be protected against invalidation
4799 * with MMU notifiers.
4801 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
4802 * should be taken for read.
4804 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
4805 * it is not a good general-purpose API.
4807 * Return: zero on success, -ve otherwise.
4809 int follow_pte(struct mm_struct
*mm
, unsigned long address
,
4810 pte_t
**ptepp
, spinlock_t
**ptlp
)
4812 return follow_invalidate_pte(mm
, address
, NULL
, ptepp
, NULL
, ptlp
);
4814 EXPORT_SYMBOL_GPL(follow_pte
);
4817 * follow_pfn - look up PFN at a user virtual address
4818 * @vma: memory mapping
4819 * @address: user virtual address
4820 * @pfn: location to store found PFN
4822 * Only IO mappings and raw PFN mappings are allowed.
4824 * This function does not allow the caller to read the permissions
4825 * of the PTE. Do not use it.
4827 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4829 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
4836 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4839 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
4842 *pfn
= pte_pfn(*ptep
);
4843 pte_unmap_unlock(ptep
, ptl
);
4846 EXPORT_SYMBOL(follow_pfn
);
4848 #ifdef CONFIG_HAVE_IOREMAP_PROT
4849 int follow_phys(struct vm_area_struct
*vma
,
4850 unsigned long address
, unsigned int flags
,
4851 unsigned long *prot
, resource_size_t
*phys
)
4857 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4860 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
4864 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
4867 *prot
= pgprot_val(pte_pgprot(pte
));
4868 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
4872 pte_unmap_unlock(ptep
, ptl
);
4877 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
4878 void *buf
, int len
, int write
)
4880 resource_size_t phys_addr
;
4881 unsigned long prot
= 0;
4882 void __iomem
*maddr
;
4883 int offset
= addr
& (PAGE_SIZE
-1);
4885 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
4888 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
4893 memcpy_toio(maddr
+ offset
, buf
, len
);
4895 memcpy_fromio(buf
, maddr
+ offset
, len
);
4900 EXPORT_SYMBOL_GPL(generic_access_phys
);
4904 * Access another process' address space as given in mm.
4906 int __access_remote_vm(struct mm_struct
*mm
, unsigned long addr
, void *buf
,
4907 int len
, unsigned int gup_flags
)
4909 struct vm_area_struct
*vma
;
4910 void *old_buf
= buf
;
4911 int write
= gup_flags
& FOLL_WRITE
;
4913 if (mmap_read_lock_killable(mm
))
4916 /* ignore errors, just check how much was successfully transferred */
4918 int bytes
, ret
, offset
;
4920 struct page
*page
= NULL
;
4922 ret
= get_user_pages_remote(mm
, addr
, 1,
4923 gup_flags
, &page
, &vma
, NULL
);
4925 #ifndef CONFIG_HAVE_IOREMAP_PROT
4929 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4930 * we can access using slightly different code.
4932 vma
= find_vma(mm
, addr
);
4933 if (!vma
|| vma
->vm_start
> addr
)
4935 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
4936 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
4944 offset
= addr
& (PAGE_SIZE
-1);
4945 if (bytes
> PAGE_SIZE
-offset
)
4946 bytes
= PAGE_SIZE
-offset
;
4950 copy_to_user_page(vma
, page
, addr
,
4951 maddr
+ offset
, buf
, bytes
);
4952 set_page_dirty_lock(page
);
4954 copy_from_user_page(vma
, page
, addr
,
4955 buf
, maddr
+ offset
, bytes
);
4964 mmap_read_unlock(mm
);
4966 return buf
- old_buf
;
4970 * access_remote_vm - access another process' address space
4971 * @mm: the mm_struct of the target address space
4972 * @addr: start address to access
4973 * @buf: source or destination buffer
4974 * @len: number of bytes to transfer
4975 * @gup_flags: flags modifying lookup behaviour
4977 * The caller must hold a reference on @mm.
4979 * Return: number of bytes copied from source to destination.
4981 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
4982 void *buf
, int len
, unsigned int gup_flags
)
4984 return __access_remote_vm(mm
, addr
, buf
, len
, gup_flags
);
4988 * Access another process' address space.
4989 * Source/target buffer must be kernel space,
4990 * Do not walk the page table directly, use get_user_pages
4992 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
4993 void *buf
, int len
, unsigned int gup_flags
)
4995 struct mm_struct
*mm
;
4998 mm
= get_task_mm(tsk
);
5002 ret
= __access_remote_vm(mm
, addr
, buf
, len
, gup_flags
);
5008 EXPORT_SYMBOL_GPL(access_process_vm
);
5011 * Print the name of a VMA.
5013 void print_vma_addr(char *prefix
, unsigned long ip
)
5015 struct mm_struct
*mm
= current
->mm
;
5016 struct vm_area_struct
*vma
;
5019 * we might be running from an atomic context so we cannot sleep
5021 if (!mmap_read_trylock(mm
))
5024 vma
= find_vma(mm
, ip
);
5025 if (vma
&& vma
->vm_file
) {
5026 struct file
*f
= vma
->vm_file
;
5027 char *buf
= (char *)__get_free_page(GFP_NOWAIT
);
5031 p
= file_path(f
, buf
, PAGE_SIZE
);
5034 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
5036 vma
->vm_end
- vma
->vm_start
);
5037 free_page((unsigned long)buf
);
5040 mmap_read_unlock(mm
);
5043 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5044 void __might_fault(const char *file
, int line
)
5047 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5048 * holding the mmap_lock, this is safe because kernel memory doesn't
5049 * get paged out, therefore we'll never actually fault, and the
5050 * below annotations will generate false positives.
5052 if (uaccess_kernel())
5054 if (pagefault_disabled())
5056 __might_sleep(file
, line
, 0);
5057 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5059 might_lock_read(¤t
->mm
->mmap_lock
);
5062 EXPORT_SYMBOL(__might_fault
);
5065 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5067 * Process all subpages of the specified huge page with the specified
5068 * operation. The target subpage will be processed last to keep its
5071 static inline void process_huge_page(
5072 unsigned long addr_hint
, unsigned int pages_per_huge_page
,
5073 void (*process_subpage
)(unsigned long addr
, int idx
, void *arg
),
5077 unsigned long addr
= addr_hint
&
5078 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5080 /* Process target subpage last to keep its cache lines hot */
5082 n
= (addr_hint
- addr
) / PAGE_SIZE
;
5083 if (2 * n
<= pages_per_huge_page
) {
5084 /* If target subpage in first half of huge page */
5087 /* Process subpages at the end of huge page */
5088 for (i
= pages_per_huge_page
- 1; i
>= 2 * n
; i
--) {
5090 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
5093 /* If target subpage in second half of huge page */
5094 base
= pages_per_huge_page
- 2 * (pages_per_huge_page
- n
);
5095 l
= pages_per_huge_page
- n
;
5096 /* Process subpages at the begin of huge page */
5097 for (i
= 0; i
< base
; i
++) {
5099 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
5103 * Process remaining subpages in left-right-left-right pattern
5104 * towards the target subpage
5106 for (i
= 0; i
< l
; i
++) {
5107 int left_idx
= base
+ i
;
5108 int right_idx
= base
+ 2 * l
- 1 - i
;
5111 process_subpage(addr
+ left_idx
* PAGE_SIZE
, left_idx
, arg
);
5113 process_subpage(addr
+ right_idx
* PAGE_SIZE
, right_idx
, arg
);
5117 static void clear_gigantic_page(struct page
*page
,
5119 unsigned int pages_per_huge_page
)
5122 struct page
*p
= page
;
5125 for (i
= 0; i
< pages_per_huge_page
;
5126 i
++, p
= mem_map_next(p
, page
, i
)) {
5128 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
5132 static void clear_subpage(unsigned long addr
, int idx
, void *arg
)
5134 struct page
*page
= arg
;
5136 clear_user_highpage(page
+ idx
, addr
);
5139 void clear_huge_page(struct page
*page
,
5140 unsigned long addr_hint
, unsigned int pages_per_huge_page
)
5142 unsigned long addr
= addr_hint
&
5143 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5145 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
5146 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
5150 process_huge_page(addr_hint
, pages_per_huge_page
, clear_subpage
, page
);
5153 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
5155 struct vm_area_struct
*vma
,
5156 unsigned int pages_per_huge_page
)
5159 struct page
*dst_base
= dst
;
5160 struct page
*src_base
= src
;
5162 for (i
= 0; i
< pages_per_huge_page
; ) {
5164 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
5167 dst
= mem_map_next(dst
, dst_base
, i
);
5168 src
= mem_map_next(src
, src_base
, i
);
5172 struct copy_subpage_arg
{
5175 struct vm_area_struct
*vma
;
5178 static void copy_subpage(unsigned long addr
, int idx
, void *arg
)
5180 struct copy_subpage_arg
*copy_arg
= arg
;
5182 copy_user_highpage(copy_arg
->dst
+ idx
, copy_arg
->src
+ idx
,
5183 addr
, copy_arg
->vma
);
5186 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
5187 unsigned long addr_hint
, struct vm_area_struct
*vma
,
5188 unsigned int pages_per_huge_page
)
5190 unsigned long addr
= addr_hint
&
5191 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5192 struct copy_subpage_arg arg
= {
5198 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
5199 copy_user_gigantic_page(dst
, src
, addr
, vma
,
5200 pages_per_huge_page
);
5204 process_huge_page(addr_hint
, pages_per_huge_page
, copy_subpage
, &arg
);
5207 long copy_huge_page_from_user(struct page
*dst_page
,
5208 const void __user
*usr_src
,
5209 unsigned int pages_per_huge_page
,
5210 bool allow_pagefault
)
5212 void *src
= (void *)usr_src
;
5214 unsigned long i
, rc
= 0;
5215 unsigned long ret_val
= pages_per_huge_page
* PAGE_SIZE
;
5216 struct page
*subpage
= dst_page
;
5218 for (i
= 0; i
< pages_per_huge_page
;
5219 i
++, subpage
= mem_map_next(subpage
, dst_page
, i
)) {
5220 if (allow_pagefault
)
5221 page_kaddr
= kmap(subpage
);
5223 page_kaddr
= kmap_atomic(subpage
);
5224 rc
= copy_from_user(page_kaddr
,
5225 (const void __user
*)(src
+ i
* PAGE_SIZE
),
5227 if (allow_pagefault
)
5230 kunmap_atomic(page_kaddr
);
5232 ret_val
-= (PAGE_SIZE
- rc
);
5240 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5242 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5244 static struct kmem_cache
*page_ptl_cachep
;
5246 void __init
ptlock_cache_init(void)
5248 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
5252 bool ptlock_alloc(struct page
*page
)
5256 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
5263 void ptlock_free(struct page
*page
)
5265 kmem_cache_free(page_ptl_cachep
, page
->ptl
);