1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * demand-loading started 01.12.91 - seems it is high on the list of
10 * things wanted, and it should be easy to implement. - Linus
14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
15 * pages started 02.12.91, seems to work. - Linus.
17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
18 * would have taken more than the 6M I have free, but it worked well as
21 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
25 * Real VM (paging to/from disk) started 18.12.91. Much more work and
26 * thought has to go into this. Oh, well..
27 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
28 * Found it. Everything seems to work now.
29 * 20.12.91 - Ok, making the swap-device changeable like the root.
33 * 05.04.94 - Multi-page memory management added for v1.1.
34 * Idea by Alex Bligh (alex@cconcepts.co.uk)
36 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
37 * (Gerhard.Wichert@pdb.siemens.de)
39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
42 #include <linux/kernel_stat.h>
44 #include <linux/sched/mm.h>
45 #include <linux/sched/coredump.h>
46 #include <linux/sched/numa_balancing.h>
47 #include <linux/sched/task.h>
48 #include <linux/hugetlb.h>
49 #include <linux/mman.h>
50 #include <linux/swap.h>
51 #include <linux/highmem.h>
52 #include <linux/pagemap.h>
53 #include <linux/memremap.h>
54 #include <linux/ksm.h>
55 #include <linux/rmap.h>
56 #include <linux/export.h>
57 #include <linux/delayacct.h>
58 #include <linux/init.h>
59 #include <linux/pfn_t.h>
60 #include <linux/writeback.h>
61 #include <linux/memcontrol.h>
62 #include <linux/mmu_notifier.h>
63 #include <linux/swapops.h>
64 #include <linux/elf.h>
65 #include <linux/gfp.h>
66 #include <linux/migrate.h>
67 #include <linux/string.h>
68 #include <linux/dma-debug.h>
69 #include <linux/debugfs.h>
70 #include <linux/userfaultfd_k.h>
71 #include <linux/dax.h>
72 #include <linux/oom.h>
73 #include <linux/numa.h>
74 #include <linux/perf_event.h>
75 #include <linux/ptrace.h>
76 #include <linux/vmalloc.h>
78 #include <trace/events/kmem.h>
81 #include <asm/mmu_context.h>
82 #include <asm/pgalloc.h>
83 #include <linux/uaccess.h>
85 #include <asm/tlbflush.h>
87 #include "pgalloc-track.h"
90 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
91 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
94 #ifndef CONFIG_NEED_MULTIPLE_NODES
95 /* use the per-pgdat data instead for discontigmem - mbligh */
96 unsigned long max_mapnr
;
97 EXPORT_SYMBOL(max_mapnr
);
100 EXPORT_SYMBOL(mem_map
);
104 * A number of key systems in x86 including ioremap() rely on the assumption
105 * that high_memory defines the upper bound on direct map memory, then end
106 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
107 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
111 EXPORT_SYMBOL(high_memory
);
114 * Randomize the address space (stacks, mmaps, brk, etc.).
116 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
117 * as ancient (libc5 based) binaries can segfault. )
119 int randomize_va_space __read_mostly
=
120 #ifdef CONFIG_COMPAT_BRK
126 #ifndef arch_faults_on_old_pte
127 static inline bool arch_faults_on_old_pte(void)
130 * Those arches which don't have hw access flag feature need to
131 * implement their own helper. By default, "true" means pagefault
132 * will be hit on old pte.
138 static int __init
disable_randmaps(char *s
)
140 randomize_va_space
= 0;
143 __setup("norandmaps", disable_randmaps
);
145 unsigned long zero_pfn __read_mostly
;
146 EXPORT_SYMBOL(zero_pfn
);
148 unsigned long highest_memmap_pfn __read_mostly
;
151 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
153 static int __init
init_zero_pfn(void)
155 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
158 core_initcall(init_zero_pfn
);
160 void mm_trace_rss_stat(struct mm_struct
*mm
, int member
, long count
)
162 trace_rss_stat(mm
, member
, count
);
165 #if defined(SPLIT_RSS_COUNTING)
167 void sync_mm_rss(struct mm_struct
*mm
)
171 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
172 if (current
->rss_stat
.count
[i
]) {
173 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
174 current
->rss_stat
.count
[i
] = 0;
177 current
->rss_stat
.events
= 0;
180 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
182 struct task_struct
*task
= current
;
184 if (likely(task
->mm
== mm
))
185 task
->rss_stat
.count
[member
] += val
;
187 add_mm_counter(mm
, member
, val
);
189 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
190 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
192 /* sync counter once per 64 page faults */
193 #define TASK_RSS_EVENTS_THRESH (64)
194 static void check_sync_rss_stat(struct task_struct
*task
)
196 if (unlikely(task
!= current
))
198 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
199 sync_mm_rss(task
->mm
);
201 #else /* SPLIT_RSS_COUNTING */
203 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
204 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
206 static void check_sync_rss_stat(struct task_struct
*task
)
210 #endif /* SPLIT_RSS_COUNTING */
213 * Note: this doesn't free the actual pages themselves. That
214 * has been handled earlier when unmapping all the memory regions.
216 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
219 pgtable_t token
= pmd_pgtable(*pmd
);
221 pte_free_tlb(tlb
, token
, addr
);
222 mm_dec_nr_ptes(tlb
->mm
);
225 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
226 unsigned long addr
, unsigned long end
,
227 unsigned long floor
, unsigned long ceiling
)
234 pmd
= pmd_offset(pud
, addr
);
236 next
= pmd_addr_end(addr
, end
);
237 if (pmd_none_or_clear_bad(pmd
))
239 free_pte_range(tlb
, pmd
, addr
);
240 } while (pmd
++, addr
= next
, addr
!= end
);
250 if (end
- 1 > ceiling
- 1)
253 pmd
= pmd_offset(pud
, start
);
255 pmd_free_tlb(tlb
, pmd
, start
);
256 mm_dec_nr_pmds(tlb
->mm
);
259 static inline void free_pud_range(struct mmu_gather
*tlb
, p4d_t
*p4d
,
260 unsigned long addr
, unsigned long end
,
261 unsigned long floor
, unsigned long ceiling
)
268 pud
= pud_offset(p4d
, addr
);
270 next
= pud_addr_end(addr
, end
);
271 if (pud_none_or_clear_bad(pud
))
273 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
274 } while (pud
++, addr
= next
, addr
!= end
);
284 if (end
- 1 > ceiling
- 1)
287 pud
= pud_offset(p4d
, start
);
289 pud_free_tlb(tlb
, pud
, start
);
290 mm_dec_nr_puds(tlb
->mm
);
293 static inline void free_p4d_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
294 unsigned long addr
, unsigned long end
,
295 unsigned long floor
, unsigned long ceiling
)
302 p4d
= p4d_offset(pgd
, addr
);
304 next
= p4d_addr_end(addr
, end
);
305 if (p4d_none_or_clear_bad(p4d
))
307 free_pud_range(tlb
, p4d
, addr
, next
, floor
, ceiling
);
308 } while (p4d
++, addr
= next
, addr
!= end
);
314 ceiling
&= PGDIR_MASK
;
318 if (end
- 1 > ceiling
- 1)
321 p4d
= p4d_offset(pgd
, start
);
323 p4d_free_tlb(tlb
, p4d
, start
);
327 * This function frees user-level page tables of a process.
329 void free_pgd_range(struct mmu_gather
*tlb
,
330 unsigned long addr
, unsigned long end
,
331 unsigned long floor
, unsigned long ceiling
)
337 * The next few lines have given us lots of grief...
339 * Why are we testing PMD* at this top level? Because often
340 * there will be no work to do at all, and we'd prefer not to
341 * go all the way down to the bottom just to discover that.
343 * Why all these "- 1"s? Because 0 represents both the bottom
344 * of the address space and the top of it (using -1 for the
345 * top wouldn't help much: the masks would do the wrong thing).
346 * The rule is that addr 0 and floor 0 refer to the bottom of
347 * the address space, but end 0 and ceiling 0 refer to the top
348 * Comparisons need to use "end - 1" and "ceiling - 1" (though
349 * that end 0 case should be mythical).
351 * Wherever addr is brought up or ceiling brought down, we must
352 * be careful to reject "the opposite 0" before it confuses the
353 * subsequent tests. But what about where end is brought down
354 * by PMD_SIZE below? no, end can't go down to 0 there.
356 * Whereas we round start (addr) and ceiling down, by different
357 * masks at different levels, in order to test whether a table
358 * now has no other vmas using it, so can be freed, we don't
359 * bother to round floor or end up - the tests don't need that.
373 if (end
- 1 > ceiling
- 1)
378 * We add page table cache pages with PAGE_SIZE,
379 * (see pte_free_tlb()), flush the tlb if we need
381 tlb_change_page_size(tlb
, PAGE_SIZE
);
382 pgd
= pgd_offset(tlb
->mm
, addr
);
384 next
= pgd_addr_end(addr
, end
);
385 if (pgd_none_or_clear_bad(pgd
))
387 free_p4d_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
388 } while (pgd
++, addr
= next
, addr
!= end
);
391 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
392 unsigned long floor
, unsigned long ceiling
)
395 struct vm_area_struct
*next
= vma
->vm_next
;
396 unsigned long addr
= vma
->vm_start
;
399 * Hide vma from rmap and truncate_pagecache before freeing
402 unlink_anon_vmas(vma
);
403 unlink_file_vma(vma
);
405 if (is_vm_hugetlb_page(vma
)) {
406 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
407 floor
, next
? next
->vm_start
: ceiling
);
410 * Optimization: gather nearby vmas into one call down
412 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
413 && !is_vm_hugetlb_page(next
)) {
416 unlink_anon_vmas(vma
);
417 unlink_file_vma(vma
);
419 free_pgd_range(tlb
, addr
, vma
->vm_end
,
420 floor
, next
? next
->vm_start
: ceiling
);
426 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
)
429 pgtable_t
new = pte_alloc_one(mm
);
434 * Ensure all pte setup (eg. pte page lock and page clearing) are
435 * visible before the pte is made visible to other CPUs by being
436 * put into page tables.
438 * The other side of the story is the pointer chasing in the page
439 * table walking code (when walking the page table without locking;
440 * ie. most of the time). Fortunately, these data accesses consist
441 * of a chain of data-dependent loads, meaning most CPUs (alpha
442 * being the notable exception) will already guarantee loads are
443 * seen in-order. See the alpha page table accessors for the
444 * smp_rmb() barriers in page table walking code.
446 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
448 ptl
= pmd_lock(mm
, pmd
);
449 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
451 pmd_populate(mm
, pmd
, new);
460 int __pte_alloc_kernel(pmd_t
*pmd
)
462 pte_t
*new = pte_alloc_one_kernel(&init_mm
);
466 smp_wmb(); /* See comment in __pte_alloc */
468 spin_lock(&init_mm
.page_table_lock
);
469 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
470 pmd_populate_kernel(&init_mm
, pmd
, new);
473 spin_unlock(&init_mm
.page_table_lock
);
475 pte_free_kernel(&init_mm
, new);
479 static inline void init_rss_vec(int *rss
)
481 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
484 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
488 if (current
->mm
== mm
)
490 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
492 add_mm_counter(mm
, i
, rss
[i
]);
496 * This function is called to print an error when a bad pte
497 * is found. For example, we might have a PFN-mapped pte in
498 * a region that doesn't allow it.
500 * The calling function must still handle the error.
502 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
503 pte_t pte
, struct page
*page
)
505 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
506 p4d_t
*p4d
= p4d_offset(pgd
, addr
);
507 pud_t
*pud
= pud_offset(p4d
, addr
);
508 pmd_t
*pmd
= pmd_offset(pud
, addr
);
509 struct address_space
*mapping
;
511 static unsigned long resume
;
512 static unsigned long nr_shown
;
513 static unsigned long nr_unshown
;
516 * Allow a burst of 60 reports, then keep quiet for that minute;
517 * or allow a steady drip of one report per second.
519 if (nr_shown
== 60) {
520 if (time_before(jiffies
, resume
)) {
525 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
532 resume
= jiffies
+ 60 * HZ
;
534 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
535 index
= linear_page_index(vma
, addr
);
537 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
539 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
541 dump_page(page
, "bad pte");
542 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
543 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
544 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
546 vma
->vm_ops
? vma
->vm_ops
->fault
: NULL
,
547 vma
->vm_file
? vma
->vm_file
->f_op
->mmap
: NULL
,
548 mapping
? mapping
->a_ops
->readpage
: NULL
);
550 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
554 * vm_normal_page -- This function gets the "struct page" associated with a pte.
556 * "Special" mappings do not wish to be associated with a "struct page" (either
557 * it doesn't exist, or it exists but they don't want to touch it). In this
558 * case, NULL is returned here. "Normal" mappings do have a struct page.
560 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
561 * pte bit, in which case this function is trivial. Secondly, an architecture
562 * may not have a spare pte bit, which requires a more complicated scheme,
565 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
566 * special mapping (even if there are underlying and valid "struct pages").
567 * COWed pages of a VM_PFNMAP are always normal.
569 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
570 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
571 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
572 * mapping will always honor the rule
574 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
576 * And for normal mappings this is false.
578 * This restricts such mappings to be a linear translation from virtual address
579 * to pfn. To get around this restriction, we allow arbitrary mappings so long
580 * as the vma is not a COW mapping; in that case, we know that all ptes are
581 * special (because none can have been COWed).
584 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
586 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
587 * page" backing, however the difference is that _all_ pages with a struct
588 * page (that is, those where pfn_valid is true) are refcounted and considered
589 * normal pages by the VM. The disadvantage is that pages are refcounted
590 * (which can be slower and simply not an option for some PFNMAP users). The
591 * advantage is that we don't have to follow the strict linearity rule of
592 * PFNMAP mappings in order to support COWable mappings.
595 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
598 unsigned long pfn
= pte_pfn(pte
);
600 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
)) {
601 if (likely(!pte_special(pte
)))
603 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
604 return vma
->vm_ops
->find_special_page(vma
, addr
);
605 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
607 if (is_zero_pfn(pfn
))
612 print_bad_pte(vma
, addr
, pte
, NULL
);
616 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
618 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
619 if (vma
->vm_flags
& VM_MIXEDMAP
) {
625 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
626 if (pfn
== vma
->vm_pgoff
+ off
)
628 if (!is_cow_mapping(vma
->vm_flags
))
633 if (is_zero_pfn(pfn
))
637 if (unlikely(pfn
> highest_memmap_pfn
)) {
638 print_bad_pte(vma
, addr
, pte
, NULL
);
643 * NOTE! We still have PageReserved() pages in the page tables.
644 * eg. VDSO mappings can cause them to exist.
647 return pfn_to_page(pfn
);
650 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
651 struct page
*vm_normal_page_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
654 unsigned long pfn
= pmd_pfn(pmd
);
657 * There is no pmd_special() but there may be special pmds, e.g.
658 * in a direct-access (dax) mapping, so let's just replicate the
659 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
661 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
662 if (vma
->vm_flags
& VM_MIXEDMAP
) {
668 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
669 if (pfn
== vma
->vm_pgoff
+ off
)
671 if (!is_cow_mapping(vma
->vm_flags
))
678 if (is_huge_zero_pmd(pmd
))
680 if (unlikely(pfn
> highest_memmap_pfn
))
684 * NOTE! We still have PageReserved() pages in the page tables.
685 * eg. VDSO mappings can cause them to exist.
688 return pfn_to_page(pfn
);
693 * copy one vm_area from one task to the other. Assumes the page tables
694 * already present in the new task to be cleared in the whole range
695 * covered by this vma.
699 copy_nonpresent_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
700 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
701 unsigned long addr
, int *rss
)
703 unsigned long vm_flags
= vma
->vm_flags
;
704 pte_t pte
= *src_pte
;
706 swp_entry_t entry
= pte_to_swp_entry(pte
);
708 if (likely(!non_swap_entry(entry
))) {
709 if (swap_duplicate(entry
) < 0)
712 /* make sure dst_mm is on swapoff's mmlist. */
713 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
714 spin_lock(&mmlist_lock
);
715 if (list_empty(&dst_mm
->mmlist
))
716 list_add(&dst_mm
->mmlist
,
718 spin_unlock(&mmlist_lock
);
721 } else if (is_migration_entry(entry
)) {
722 page
= migration_entry_to_page(entry
);
724 rss
[mm_counter(page
)]++;
726 if (is_write_migration_entry(entry
) &&
727 is_cow_mapping(vm_flags
)) {
729 * COW mappings require pages in both
730 * parent and child to be set to read.
732 make_migration_entry_read(&entry
);
733 pte
= swp_entry_to_pte(entry
);
734 if (pte_swp_soft_dirty(*src_pte
))
735 pte
= pte_swp_mksoft_dirty(pte
);
736 if (pte_swp_uffd_wp(*src_pte
))
737 pte
= pte_swp_mkuffd_wp(pte
);
738 set_pte_at(src_mm
, addr
, src_pte
, pte
);
740 } else if (is_device_private_entry(entry
)) {
741 page
= device_private_entry_to_page(entry
);
744 * Update rss count even for unaddressable pages, as
745 * they should treated just like normal pages in this
748 * We will likely want to have some new rss counters
749 * for unaddressable pages, at some point. But for now
750 * keep things as they are.
753 rss
[mm_counter(page
)]++;
754 page_dup_rmap(page
, false);
757 * We do not preserve soft-dirty information, because so
758 * far, checkpoint/restore is the only feature that
759 * requires that. And checkpoint/restore does not work
760 * when a device driver is involved (you cannot easily
761 * save and restore device driver state).
763 if (is_write_device_private_entry(entry
) &&
764 is_cow_mapping(vm_flags
)) {
765 make_device_private_entry_read(&entry
);
766 pte
= swp_entry_to_pte(entry
);
767 if (pte_swp_uffd_wp(*src_pte
))
768 pte
= pte_swp_mkuffd_wp(pte
);
769 set_pte_at(src_mm
, addr
, src_pte
, pte
);
772 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
777 * Copy a present and normal page if necessary.
779 * NOTE! The usual case is that this doesn't need to do
780 * anything, and can just return a positive value. That
781 * will let the caller know that it can just increase
782 * the page refcount and re-use the pte the traditional
785 * But _if_ we need to copy it because it needs to be
786 * pinned in the parent (and the child should get its own
787 * copy rather than just a reference to the same page),
788 * we'll do that here and return zero to let the caller
791 * And if we need a pre-allocated page but don't yet have
792 * one, return a negative error to let the preallocation
793 * code know so that it can do so outside the page table
797 copy_present_page(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
798 pte_t
*dst_pte
, pte_t
*src_pte
, unsigned long addr
, int *rss
,
799 struct page
**prealloc
, pte_t pte
, struct page
*page
)
801 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
802 struct page
*new_page
;
804 if (!is_cow_mapping(src_vma
->vm_flags
))
808 * What we want to do is to check whether this page may
809 * have been pinned by the parent process. If so,
810 * instead of wrprotect the pte on both sides, we copy
811 * the page immediately so that we'll always guarantee
812 * the pinned page won't be randomly replaced in the
815 * The page pinning checks are just "has this mm ever
816 * seen pinning", along with the (inexact) check of
817 * the page count. That might give false positives for
818 * for pinning, but it will work correctly.
820 if (likely(!atomic_read(&src_mm
->has_pinned
)))
822 if (likely(!page_maybe_dma_pinned(page
)))
825 new_page
= *prealloc
;
830 * We have a prealloc page, all good! Take it
831 * over and copy the page & arm it.
834 copy_user_highpage(new_page
, page
, addr
, src_vma
);
835 __SetPageUptodate(new_page
);
836 page_add_new_anon_rmap(new_page
, dst_vma
, addr
, false);
837 lru_cache_add_inactive_or_unevictable(new_page
, dst_vma
);
838 rss
[mm_counter(new_page
)]++;
840 /* All done, just insert the new page copy in the child */
841 pte
= mk_pte(new_page
, dst_vma
->vm_page_prot
);
842 pte
= maybe_mkwrite(pte_mkdirty(pte
), dst_vma
);
843 set_pte_at(dst_vma
->vm_mm
, addr
, dst_pte
, pte
);
848 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
849 * is required to copy this pte.
852 copy_present_pte(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
853 pte_t
*dst_pte
, pte_t
*src_pte
, unsigned long addr
, int *rss
,
854 struct page
**prealloc
)
856 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
857 unsigned long vm_flags
= src_vma
->vm_flags
;
858 pte_t pte
= *src_pte
;
861 page
= vm_normal_page(src_vma
, addr
, pte
);
865 retval
= copy_present_page(dst_vma
, src_vma
, dst_pte
, src_pte
,
866 addr
, rss
, prealloc
, pte
, page
);
871 page_dup_rmap(page
, false);
872 rss
[mm_counter(page
)]++;
876 * If it's a COW mapping, write protect it both
877 * in the parent and the child
879 if (is_cow_mapping(vm_flags
) && pte_write(pte
)) {
880 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
881 pte
= pte_wrprotect(pte
);
885 * If it's a shared mapping, mark it clean in
888 if (vm_flags
& VM_SHARED
)
889 pte
= pte_mkclean(pte
);
890 pte
= pte_mkold(pte
);
893 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
894 * does not have the VM_UFFD_WP, which means that the uffd
895 * fork event is not enabled.
897 if (!(vm_flags
& VM_UFFD_WP
))
898 pte
= pte_clear_uffd_wp(pte
);
900 set_pte_at(dst_vma
->vm_mm
, addr
, dst_pte
, pte
);
904 static inline struct page
*
905 page_copy_prealloc(struct mm_struct
*src_mm
, struct vm_area_struct
*vma
,
908 struct page
*new_page
;
910 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, addr
);
914 if (mem_cgroup_charge(new_page
, src_mm
, GFP_KERNEL
)) {
918 cgroup_throttle_swaprate(new_page
, GFP_KERNEL
);
924 copy_pte_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
925 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
928 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
929 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
930 pte_t
*orig_src_pte
, *orig_dst_pte
;
931 pte_t
*src_pte
, *dst_pte
;
932 spinlock_t
*src_ptl
, *dst_ptl
;
933 int progress
, ret
= 0;
934 int rss
[NR_MM_COUNTERS
];
935 swp_entry_t entry
= (swp_entry_t
){0};
936 struct page
*prealloc
= NULL
;
942 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
947 src_pte
= pte_offset_map(src_pmd
, addr
);
948 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
949 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
950 orig_src_pte
= src_pte
;
951 orig_dst_pte
= dst_pte
;
952 arch_enter_lazy_mmu_mode();
956 * We are holding two locks at this point - either of them
957 * could generate latencies in another task on another CPU.
959 if (progress
>= 32) {
961 if (need_resched() ||
962 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
965 if (pte_none(*src_pte
)) {
969 if (unlikely(!pte_present(*src_pte
))) {
970 entry
.val
= copy_nonpresent_pte(dst_mm
, src_mm
,
978 /* copy_present_pte() will clear `*prealloc' if consumed */
979 ret
= copy_present_pte(dst_vma
, src_vma
, dst_pte
, src_pte
,
980 addr
, rss
, &prealloc
);
982 * If we need a pre-allocated page for this pte, drop the
983 * locks, allocate, and try again.
985 if (unlikely(ret
== -EAGAIN
))
987 if (unlikely(prealloc
)) {
989 * pre-alloc page cannot be reused by next time so as
990 * to strictly follow mempolicy (e.g., alloc_page_vma()
991 * will allocate page according to address). This
992 * could only happen if one pinned pte changed.
998 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1000 arch_leave_lazy_mmu_mode();
1001 spin_unlock(src_ptl
);
1002 pte_unmap(orig_src_pte
);
1003 add_mm_rss_vec(dst_mm
, rss
);
1004 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
1008 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0) {
1014 WARN_ON_ONCE(ret
!= -EAGAIN
);
1015 prealloc
= page_copy_prealloc(src_mm
, src_vma
, addr
);
1018 /* We've captured and resolved the error. Reset, try again. */
1024 if (unlikely(prealloc
))
1030 copy_pmd_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1031 pud_t
*dst_pud
, pud_t
*src_pud
, unsigned long addr
,
1034 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1035 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1036 pmd_t
*src_pmd
, *dst_pmd
;
1039 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
1042 src_pmd
= pmd_offset(src_pud
, addr
);
1044 next
= pmd_addr_end(addr
, end
);
1045 if (is_swap_pmd(*src_pmd
) || pmd_trans_huge(*src_pmd
)
1046 || pmd_devmap(*src_pmd
)) {
1048 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PMD_SIZE
, src_vma
);
1049 err
= copy_huge_pmd(dst_mm
, src_mm
,
1050 dst_pmd
, src_pmd
, addr
, src_vma
);
1057 if (pmd_none_or_clear_bad(src_pmd
))
1059 if (copy_pte_range(dst_vma
, src_vma
, dst_pmd
, src_pmd
,
1062 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1067 copy_pud_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1068 p4d_t
*dst_p4d
, p4d_t
*src_p4d
, unsigned long addr
,
1071 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1072 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1073 pud_t
*src_pud
, *dst_pud
;
1076 dst_pud
= pud_alloc(dst_mm
, dst_p4d
, addr
);
1079 src_pud
= pud_offset(src_p4d
, addr
);
1081 next
= pud_addr_end(addr
, end
);
1082 if (pud_trans_huge(*src_pud
) || pud_devmap(*src_pud
)) {
1085 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PUD_SIZE
, src_vma
);
1086 err
= copy_huge_pud(dst_mm
, src_mm
,
1087 dst_pud
, src_pud
, addr
, src_vma
);
1094 if (pud_none_or_clear_bad(src_pud
))
1096 if (copy_pmd_range(dst_vma
, src_vma
, dst_pud
, src_pud
,
1099 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1104 copy_p4d_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
,
1105 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, unsigned long addr
,
1108 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1109 p4d_t
*src_p4d
, *dst_p4d
;
1112 dst_p4d
= p4d_alloc(dst_mm
, dst_pgd
, addr
);
1115 src_p4d
= p4d_offset(src_pgd
, addr
);
1117 next
= p4d_addr_end(addr
, end
);
1118 if (p4d_none_or_clear_bad(src_p4d
))
1120 if (copy_pud_range(dst_vma
, src_vma
, dst_p4d
, src_p4d
,
1123 } while (dst_p4d
++, src_p4d
++, addr
= next
, addr
!= end
);
1128 copy_page_range(struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
)
1130 pgd_t
*src_pgd
, *dst_pgd
;
1132 unsigned long addr
= src_vma
->vm_start
;
1133 unsigned long end
= src_vma
->vm_end
;
1134 struct mm_struct
*dst_mm
= dst_vma
->vm_mm
;
1135 struct mm_struct
*src_mm
= src_vma
->vm_mm
;
1136 struct mmu_notifier_range range
;
1141 * Don't copy ptes where a page fault will fill them correctly.
1142 * Fork becomes much lighter when there are big shared or private
1143 * readonly mappings. The tradeoff is that copy_page_range is more
1144 * efficient than faulting.
1146 if (!(src_vma
->vm_flags
& (VM_HUGETLB
| VM_PFNMAP
| VM_MIXEDMAP
)) &&
1150 if (is_vm_hugetlb_page(src_vma
))
1151 return copy_hugetlb_page_range(dst_mm
, src_mm
, src_vma
);
1153 if (unlikely(src_vma
->vm_flags
& VM_PFNMAP
)) {
1155 * We do not free on error cases below as remove_vma
1156 * gets called on error from higher level routine
1158 ret
= track_pfn_copy(src_vma
);
1164 * We need to invalidate the secondary MMU mappings only when
1165 * there could be a permission downgrade on the ptes of the
1166 * parent mm. And a permission downgrade will only happen if
1167 * is_cow_mapping() returns true.
1169 is_cow
= is_cow_mapping(src_vma
->vm_flags
);
1172 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_PAGE
,
1173 0, src_vma
, src_mm
, addr
, end
);
1174 mmu_notifier_invalidate_range_start(&range
);
1178 dst_pgd
= pgd_offset(dst_mm
, addr
);
1179 src_pgd
= pgd_offset(src_mm
, addr
);
1181 next
= pgd_addr_end(addr
, end
);
1182 if (pgd_none_or_clear_bad(src_pgd
))
1184 if (unlikely(copy_p4d_range(dst_vma
, src_vma
, dst_pgd
, src_pgd
,
1189 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1192 mmu_notifier_invalidate_range_end(&range
);
1196 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1197 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1198 unsigned long addr
, unsigned long end
,
1199 struct zap_details
*details
)
1201 struct mm_struct
*mm
= tlb
->mm
;
1202 int force_flush
= 0;
1203 int rss
[NR_MM_COUNTERS
];
1209 tlb_change_page_size(tlb
, PAGE_SIZE
);
1212 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1214 flush_tlb_batched_pending(mm
);
1215 arch_enter_lazy_mmu_mode();
1218 if (pte_none(ptent
))
1224 if (pte_present(ptent
)) {
1227 page
= vm_normal_page(vma
, addr
, ptent
);
1228 if (unlikely(details
) && page
) {
1230 * unmap_shared_mapping_pages() wants to
1231 * invalidate cache without truncating:
1232 * unmap shared but keep private pages.
1234 if (details
->check_mapping
&&
1235 details
->check_mapping
!= page_rmapping(page
))
1238 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1240 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1241 if (unlikely(!page
))
1244 if (!PageAnon(page
)) {
1245 if (pte_dirty(ptent
)) {
1247 set_page_dirty(page
);
1249 if (pte_young(ptent
) &&
1250 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1251 mark_page_accessed(page
);
1253 rss
[mm_counter(page
)]--;
1254 page_remove_rmap(page
, false);
1255 if (unlikely(page_mapcount(page
) < 0))
1256 print_bad_pte(vma
, addr
, ptent
, page
);
1257 if (unlikely(__tlb_remove_page(tlb
, page
))) {
1265 entry
= pte_to_swp_entry(ptent
);
1266 if (is_device_private_entry(entry
)) {
1267 struct page
*page
= device_private_entry_to_page(entry
);
1269 if (unlikely(details
&& details
->check_mapping
)) {
1271 * unmap_shared_mapping_pages() wants to
1272 * invalidate cache without truncating:
1273 * unmap shared but keep private pages.
1275 if (details
->check_mapping
!=
1276 page_rmapping(page
))
1280 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1281 rss
[mm_counter(page
)]--;
1282 page_remove_rmap(page
, false);
1287 /* If details->check_mapping, we leave swap entries. */
1288 if (unlikely(details
))
1291 if (!non_swap_entry(entry
))
1293 else if (is_migration_entry(entry
)) {
1296 page
= migration_entry_to_page(entry
);
1297 rss
[mm_counter(page
)]--;
1299 if (unlikely(!free_swap_and_cache(entry
)))
1300 print_bad_pte(vma
, addr
, ptent
, NULL
);
1301 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1302 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1304 add_mm_rss_vec(mm
, rss
);
1305 arch_leave_lazy_mmu_mode();
1307 /* Do the actual TLB flush before dropping ptl */
1309 tlb_flush_mmu_tlbonly(tlb
);
1310 pte_unmap_unlock(start_pte
, ptl
);
1313 * If we forced a TLB flush (either due to running out of
1314 * batch buffers or because we needed to flush dirty TLB
1315 * entries before releasing the ptl), free the batched
1316 * memory too. Restart if we didn't do everything.
1331 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1332 struct vm_area_struct
*vma
, pud_t
*pud
,
1333 unsigned long addr
, unsigned long end
,
1334 struct zap_details
*details
)
1339 pmd
= pmd_offset(pud
, addr
);
1341 next
= pmd_addr_end(addr
, end
);
1342 if (is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1343 if (next
- addr
!= HPAGE_PMD_SIZE
)
1344 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
1345 else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1350 * Here there can be other concurrent MADV_DONTNEED or
1351 * trans huge page faults running, and if the pmd is
1352 * none or trans huge it can change under us. This is
1353 * because MADV_DONTNEED holds the mmap_lock in read
1356 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1358 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1361 } while (pmd
++, addr
= next
, addr
!= end
);
1366 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1367 struct vm_area_struct
*vma
, p4d_t
*p4d
,
1368 unsigned long addr
, unsigned long end
,
1369 struct zap_details
*details
)
1374 pud
= pud_offset(p4d
, addr
);
1376 next
= pud_addr_end(addr
, end
);
1377 if (pud_trans_huge(*pud
) || pud_devmap(*pud
)) {
1378 if (next
- addr
!= HPAGE_PUD_SIZE
) {
1379 mmap_assert_locked(tlb
->mm
);
1380 split_huge_pud(vma
, pud
, addr
);
1381 } else if (zap_huge_pud(tlb
, vma
, pud
, addr
))
1385 if (pud_none_or_clear_bad(pud
))
1387 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1390 } while (pud
++, addr
= next
, addr
!= end
);
1395 static inline unsigned long zap_p4d_range(struct mmu_gather
*tlb
,
1396 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1397 unsigned long addr
, unsigned long end
,
1398 struct zap_details
*details
)
1403 p4d
= p4d_offset(pgd
, addr
);
1405 next
= p4d_addr_end(addr
, end
);
1406 if (p4d_none_or_clear_bad(p4d
))
1408 next
= zap_pud_range(tlb
, vma
, p4d
, addr
, next
, details
);
1409 } while (p4d
++, addr
= next
, addr
!= end
);
1414 void unmap_page_range(struct mmu_gather
*tlb
,
1415 struct vm_area_struct
*vma
,
1416 unsigned long addr
, unsigned long end
,
1417 struct zap_details
*details
)
1422 BUG_ON(addr
>= end
);
1423 tlb_start_vma(tlb
, vma
);
1424 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1426 next
= pgd_addr_end(addr
, end
);
1427 if (pgd_none_or_clear_bad(pgd
))
1429 next
= zap_p4d_range(tlb
, vma
, pgd
, addr
, next
, details
);
1430 } while (pgd
++, addr
= next
, addr
!= end
);
1431 tlb_end_vma(tlb
, vma
);
1435 static void unmap_single_vma(struct mmu_gather
*tlb
,
1436 struct vm_area_struct
*vma
, unsigned long start_addr
,
1437 unsigned long end_addr
,
1438 struct zap_details
*details
)
1440 unsigned long start
= max(vma
->vm_start
, start_addr
);
1443 if (start
>= vma
->vm_end
)
1445 end
= min(vma
->vm_end
, end_addr
);
1446 if (end
<= vma
->vm_start
)
1450 uprobe_munmap(vma
, start
, end
);
1452 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1453 untrack_pfn(vma
, 0, 0);
1456 if (unlikely(is_vm_hugetlb_page(vma
))) {
1458 * It is undesirable to test vma->vm_file as it
1459 * should be non-null for valid hugetlb area.
1460 * However, vm_file will be NULL in the error
1461 * cleanup path of mmap_region. When
1462 * hugetlbfs ->mmap method fails,
1463 * mmap_region() nullifies vma->vm_file
1464 * before calling this function to clean up.
1465 * Since no pte has actually been setup, it is
1466 * safe to do nothing in this case.
1469 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
1470 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1471 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
1474 unmap_page_range(tlb
, vma
, start
, end
, details
);
1479 * unmap_vmas - unmap a range of memory covered by a list of vma's
1480 * @tlb: address of the caller's struct mmu_gather
1481 * @vma: the starting vma
1482 * @start_addr: virtual address at which to start unmapping
1483 * @end_addr: virtual address at which to end unmapping
1485 * Unmap all pages in the vma list.
1487 * Only addresses between `start' and `end' will be unmapped.
1489 * The VMA list must be sorted in ascending virtual address order.
1491 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1492 * range after unmap_vmas() returns. So the only responsibility here is to
1493 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1494 * drops the lock and schedules.
1496 void unmap_vmas(struct mmu_gather
*tlb
,
1497 struct vm_area_struct
*vma
, unsigned long start_addr
,
1498 unsigned long end_addr
)
1500 struct mmu_notifier_range range
;
1502 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
, vma
->vm_mm
,
1503 start_addr
, end_addr
);
1504 mmu_notifier_invalidate_range_start(&range
);
1505 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1506 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1507 mmu_notifier_invalidate_range_end(&range
);
1511 * zap_page_range - remove user pages in a given range
1512 * @vma: vm_area_struct holding the applicable pages
1513 * @start: starting address of pages to zap
1514 * @size: number of bytes to zap
1516 * Caller must protect the VMA list
1518 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1521 struct mmu_notifier_range range
;
1522 struct mmu_gather tlb
;
1525 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1526 start
, start
+ size
);
1527 tlb_gather_mmu(&tlb
, vma
->vm_mm
, start
, range
.end
);
1528 update_hiwater_rss(vma
->vm_mm
);
1529 mmu_notifier_invalidate_range_start(&range
);
1530 for ( ; vma
&& vma
->vm_start
< range
.end
; vma
= vma
->vm_next
)
1531 unmap_single_vma(&tlb
, vma
, start
, range
.end
, NULL
);
1532 mmu_notifier_invalidate_range_end(&range
);
1533 tlb_finish_mmu(&tlb
, start
, range
.end
);
1537 * zap_page_range_single - remove user pages in a given range
1538 * @vma: vm_area_struct holding the applicable pages
1539 * @address: starting address of pages to zap
1540 * @size: number of bytes to zap
1541 * @details: details of shared cache invalidation
1543 * The range must fit into one VMA.
1545 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1546 unsigned long size
, struct zap_details
*details
)
1548 struct mmu_notifier_range range
;
1549 struct mmu_gather tlb
;
1552 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1553 address
, address
+ size
);
1554 tlb_gather_mmu(&tlb
, vma
->vm_mm
, address
, range
.end
);
1555 update_hiwater_rss(vma
->vm_mm
);
1556 mmu_notifier_invalidate_range_start(&range
);
1557 unmap_single_vma(&tlb
, vma
, address
, range
.end
, details
);
1558 mmu_notifier_invalidate_range_end(&range
);
1559 tlb_finish_mmu(&tlb
, address
, range
.end
);
1563 * zap_vma_ptes - remove ptes mapping the vma
1564 * @vma: vm_area_struct holding ptes to be zapped
1565 * @address: starting address of pages to zap
1566 * @size: number of bytes to zap
1568 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1570 * The entire address range must be fully contained within the vma.
1573 void zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1576 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1577 !(vma
->vm_flags
& VM_PFNMAP
))
1580 zap_page_range_single(vma
, address
, size
, NULL
);
1582 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1584 static pmd_t
*walk_to_pmd(struct mm_struct
*mm
, unsigned long addr
)
1591 pgd
= pgd_offset(mm
, addr
);
1592 p4d
= p4d_alloc(mm
, pgd
, addr
);
1595 pud
= pud_alloc(mm
, p4d
, addr
);
1598 pmd
= pmd_alloc(mm
, pud
, addr
);
1602 VM_BUG_ON(pmd_trans_huge(*pmd
));
1606 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1609 pmd_t
*pmd
= walk_to_pmd(mm
, addr
);
1613 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1616 static int validate_page_before_insert(struct page
*page
)
1618 if (PageAnon(page
) || PageSlab(page
) || page_has_type(page
))
1620 flush_dcache_page(page
);
1624 static int insert_page_into_pte_locked(struct mm_struct
*mm
, pte_t
*pte
,
1625 unsigned long addr
, struct page
*page
, pgprot_t prot
)
1627 if (!pte_none(*pte
))
1629 /* Ok, finally just insert the thing.. */
1631 inc_mm_counter_fast(mm
, mm_counter_file(page
));
1632 page_add_file_rmap(page
, false);
1633 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1638 * This is the old fallback for page remapping.
1640 * For historical reasons, it only allows reserved pages. Only
1641 * old drivers should use this, and they needed to mark their
1642 * pages reserved for the old functions anyway.
1644 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1645 struct page
*page
, pgprot_t prot
)
1647 struct mm_struct
*mm
= vma
->vm_mm
;
1652 retval
= validate_page_before_insert(page
);
1656 pte
= get_locked_pte(mm
, addr
, &ptl
);
1659 retval
= insert_page_into_pte_locked(mm
, pte
, addr
, page
, prot
);
1660 pte_unmap_unlock(pte
, ptl
);
1666 static int insert_page_in_batch_locked(struct mm_struct
*mm
, pte_t
*pte
,
1667 unsigned long addr
, struct page
*page
, pgprot_t prot
)
1671 if (!page_count(page
))
1673 err
= validate_page_before_insert(page
);
1676 return insert_page_into_pte_locked(mm
, pte
, addr
, page
, prot
);
1679 /* insert_pages() amortizes the cost of spinlock operations
1680 * when inserting pages in a loop. Arch *must* define pte_index.
1682 static int insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
1683 struct page
**pages
, unsigned long *num
, pgprot_t prot
)
1686 pte_t
*start_pte
, *pte
;
1687 spinlock_t
*pte_lock
;
1688 struct mm_struct
*const mm
= vma
->vm_mm
;
1689 unsigned long curr_page_idx
= 0;
1690 unsigned long remaining_pages_total
= *num
;
1691 unsigned long pages_to_write_in_pmd
;
1695 pmd
= walk_to_pmd(mm
, addr
);
1699 pages_to_write_in_pmd
= min_t(unsigned long,
1700 remaining_pages_total
, PTRS_PER_PTE
- pte_index(addr
));
1702 /* Allocate the PTE if necessary; takes PMD lock once only. */
1704 if (pte_alloc(mm
, pmd
))
1707 while (pages_to_write_in_pmd
) {
1709 const int batch_size
= min_t(int, pages_to_write_in_pmd
, 8);
1711 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &pte_lock
);
1712 for (pte
= start_pte
; pte_idx
< batch_size
; ++pte
, ++pte_idx
) {
1713 int err
= insert_page_in_batch_locked(mm
, pte
,
1714 addr
, pages
[curr_page_idx
], prot
);
1715 if (unlikely(err
)) {
1716 pte_unmap_unlock(start_pte
, pte_lock
);
1718 remaining_pages_total
-= pte_idx
;
1724 pte_unmap_unlock(start_pte
, pte_lock
);
1725 pages_to_write_in_pmd
-= batch_size
;
1726 remaining_pages_total
-= batch_size
;
1728 if (remaining_pages_total
)
1732 *num
= remaining_pages_total
;
1735 #endif /* ifdef pte_index */
1738 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1739 * @vma: user vma to map to
1740 * @addr: target start user address of these pages
1741 * @pages: source kernel pages
1742 * @num: in: number of pages to map. out: number of pages that were *not*
1743 * mapped. (0 means all pages were successfully mapped).
1745 * Preferred over vm_insert_page() when inserting multiple pages.
1747 * In case of error, we may have mapped a subset of the provided
1748 * pages. It is the caller's responsibility to account for this case.
1750 * The same restrictions apply as in vm_insert_page().
1752 int vm_insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
1753 struct page
**pages
, unsigned long *num
)
1756 const unsigned long end_addr
= addr
+ (*num
* PAGE_SIZE
) - 1;
1758 if (addr
< vma
->vm_start
|| end_addr
>= vma
->vm_end
)
1760 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1761 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
1762 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1763 vma
->vm_flags
|= VM_MIXEDMAP
;
1765 /* Defer page refcount checking till we're about to map that page. */
1766 return insert_pages(vma
, addr
, pages
, num
, vma
->vm_page_prot
);
1768 unsigned long idx
= 0, pgcount
= *num
;
1771 for (; idx
< pgcount
; ++idx
) {
1772 err
= vm_insert_page(vma
, addr
+ (PAGE_SIZE
* idx
), pages
[idx
]);
1776 *num
= pgcount
- idx
;
1778 #endif /* ifdef pte_index */
1780 EXPORT_SYMBOL(vm_insert_pages
);
1783 * vm_insert_page - insert single page into user vma
1784 * @vma: user vma to map to
1785 * @addr: target user address of this page
1786 * @page: source kernel page
1788 * This allows drivers to insert individual pages they've allocated
1791 * The page has to be a nice clean _individual_ kernel allocation.
1792 * If you allocate a compound page, you need to have marked it as
1793 * such (__GFP_COMP), or manually just split the page up yourself
1794 * (see split_page()).
1796 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1797 * took an arbitrary page protection parameter. This doesn't allow
1798 * that. Your vma protection will have to be set up correctly, which
1799 * means that if you want a shared writable mapping, you'd better
1800 * ask for a shared writable mapping!
1802 * The page does not need to be reserved.
1804 * Usually this function is called from f_op->mmap() handler
1805 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1806 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1807 * function from other places, for example from page-fault handler.
1809 * Return: %0 on success, negative error code otherwise.
1811 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1814 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1816 if (!page_count(page
))
1818 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1819 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
1820 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1821 vma
->vm_flags
|= VM_MIXEDMAP
;
1823 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1825 EXPORT_SYMBOL(vm_insert_page
);
1828 * __vm_map_pages - maps range of kernel pages into user vma
1829 * @vma: user vma to map to
1830 * @pages: pointer to array of source kernel pages
1831 * @num: number of pages in page array
1832 * @offset: user's requested vm_pgoff
1834 * This allows drivers to map range of kernel pages into a user vma.
1836 * Return: 0 on success and error code otherwise.
1838 static int __vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
1839 unsigned long num
, unsigned long offset
)
1841 unsigned long count
= vma_pages(vma
);
1842 unsigned long uaddr
= vma
->vm_start
;
1845 /* Fail if the user requested offset is beyond the end of the object */
1849 /* Fail if the user requested size exceeds available object size */
1850 if (count
> num
- offset
)
1853 for (i
= 0; i
< count
; i
++) {
1854 ret
= vm_insert_page(vma
, uaddr
, pages
[offset
+ i
]);
1864 * vm_map_pages - maps range of kernel pages starts with non zero offset
1865 * @vma: user vma to map to
1866 * @pages: pointer to array of source kernel pages
1867 * @num: number of pages in page array
1869 * Maps an object consisting of @num pages, catering for the user's
1870 * requested vm_pgoff
1872 * If we fail to insert any page into the vma, the function will return
1873 * immediately leaving any previously inserted pages present. Callers
1874 * from the mmap handler may immediately return the error as their caller
1875 * will destroy the vma, removing any successfully inserted pages. Other
1876 * callers should make their own arrangements for calling unmap_region().
1878 * Context: Process context. Called by mmap handlers.
1879 * Return: 0 on success and error code otherwise.
1881 int vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
1884 return __vm_map_pages(vma
, pages
, num
, vma
->vm_pgoff
);
1886 EXPORT_SYMBOL(vm_map_pages
);
1889 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1890 * @vma: user vma to map to
1891 * @pages: pointer to array of source kernel pages
1892 * @num: number of pages in page array
1894 * Similar to vm_map_pages(), except that it explicitly sets the offset
1895 * to 0. This function is intended for the drivers that did not consider
1898 * Context: Process context. Called by mmap handlers.
1899 * Return: 0 on success and error code otherwise.
1901 int vm_map_pages_zero(struct vm_area_struct
*vma
, struct page
**pages
,
1904 return __vm_map_pages(vma
, pages
, num
, 0);
1906 EXPORT_SYMBOL(vm_map_pages_zero
);
1908 static vm_fault_t
insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1909 pfn_t pfn
, pgprot_t prot
, bool mkwrite
)
1911 struct mm_struct
*mm
= vma
->vm_mm
;
1915 pte
= get_locked_pte(mm
, addr
, &ptl
);
1917 return VM_FAULT_OOM
;
1918 if (!pte_none(*pte
)) {
1921 * For read faults on private mappings the PFN passed
1922 * in may not match the PFN we have mapped if the
1923 * mapped PFN is a writeable COW page. In the mkwrite
1924 * case we are creating a writable PTE for a shared
1925 * mapping and we expect the PFNs to match. If they
1926 * don't match, we are likely racing with block
1927 * allocation and mapping invalidation so just skip the
1930 if (pte_pfn(*pte
) != pfn_t_to_pfn(pfn
)) {
1931 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte
)));
1934 entry
= pte_mkyoung(*pte
);
1935 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1936 if (ptep_set_access_flags(vma
, addr
, pte
, entry
, 1))
1937 update_mmu_cache(vma
, addr
, pte
);
1942 /* Ok, finally just insert the thing.. */
1943 if (pfn_t_devmap(pfn
))
1944 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
1946 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
1949 entry
= pte_mkyoung(entry
);
1950 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1953 set_pte_at(mm
, addr
, pte
, entry
);
1954 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1957 pte_unmap_unlock(pte
, ptl
);
1958 return VM_FAULT_NOPAGE
;
1962 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1963 * @vma: user vma to map to
1964 * @addr: target user address of this page
1965 * @pfn: source kernel pfn
1966 * @pgprot: pgprot flags for the inserted page
1968 * This is exactly like vmf_insert_pfn(), except that it allows drivers
1969 * to override pgprot on a per-page basis.
1971 * This only makes sense for IO mappings, and it makes no sense for
1972 * COW mappings. In general, using multiple vmas is preferable;
1973 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1976 * See vmf_insert_mixed_prot() for a discussion of the implication of using
1977 * a value of @pgprot different from that of @vma->vm_page_prot.
1979 * Context: Process context. May allocate using %GFP_KERNEL.
1980 * Return: vm_fault_t value.
1982 vm_fault_t
vmf_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
1983 unsigned long pfn
, pgprot_t pgprot
)
1986 * Technically, architectures with pte_special can avoid all these
1987 * restrictions (same for remap_pfn_range). However we would like
1988 * consistency in testing and feature parity among all, so we should
1989 * try to keep these invariants in place for everybody.
1991 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1992 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1993 (VM_PFNMAP
|VM_MIXEDMAP
));
1994 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1995 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1997 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1998 return VM_FAULT_SIGBUS
;
2000 if (!pfn_modify_allowed(pfn
, pgprot
))
2001 return VM_FAULT_SIGBUS
;
2003 track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
));
2005 return insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
,
2008 EXPORT_SYMBOL(vmf_insert_pfn_prot
);
2011 * vmf_insert_pfn - insert single pfn into user vma
2012 * @vma: user vma to map to
2013 * @addr: target user address of this page
2014 * @pfn: source kernel pfn
2016 * Similar to vm_insert_page, this allows drivers to insert individual pages
2017 * they've allocated into a user vma. Same comments apply.
2019 * This function should only be called from a vm_ops->fault handler, and
2020 * in that case the handler should return the result of this function.
2022 * vma cannot be a COW mapping.
2024 * As this is called only for pages that do not currently exist, we
2025 * do not need to flush old virtual caches or the TLB.
2027 * Context: Process context. May allocate using %GFP_KERNEL.
2028 * Return: vm_fault_t value.
2030 vm_fault_t
vmf_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2033 return vmf_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
2035 EXPORT_SYMBOL(vmf_insert_pfn
);
2037 static bool vm_mixed_ok(struct vm_area_struct
*vma
, pfn_t pfn
)
2039 /* these checks mirror the abort conditions in vm_normal_page */
2040 if (vma
->vm_flags
& VM_MIXEDMAP
)
2042 if (pfn_t_devmap(pfn
))
2044 if (pfn_t_special(pfn
))
2046 if (is_zero_pfn(pfn_t_to_pfn(pfn
)))
2051 static vm_fault_t
__vm_insert_mixed(struct vm_area_struct
*vma
,
2052 unsigned long addr
, pfn_t pfn
, pgprot_t pgprot
,
2057 BUG_ON(!vm_mixed_ok(vma
, pfn
));
2059 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2060 return VM_FAULT_SIGBUS
;
2062 track_pfn_insert(vma
, &pgprot
, pfn
);
2064 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn
), pgprot
))
2065 return VM_FAULT_SIGBUS
;
2068 * If we don't have pte special, then we have to use the pfn_valid()
2069 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2070 * refcount the page if pfn_valid is true (hence insert_page rather
2071 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2072 * without pte special, it would there be refcounted as a normal page.
2074 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
) &&
2075 !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
2079 * At this point we are committed to insert_page()
2080 * regardless of whether the caller specified flags that
2081 * result in pfn_t_has_page() == false.
2083 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
2084 err
= insert_page(vma
, addr
, page
, pgprot
);
2086 return insert_pfn(vma
, addr
, pfn
, pgprot
, mkwrite
);
2090 return VM_FAULT_OOM
;
2091 if (err
< 0 && err
!= -EBUSY
)
2092 return VM_FAULT_SIGBUS
;
2094 return VM_FAULT_NOPAGE
;
2098 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2099 * @vma: user vma to map to
2100 * @addr: target user address of this page
2101 * @pfn: source kernel pfn
2102 * @pgprot: pgprot flags for the inserted page
2104 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2105 * to override pgprot on a per-page basis.
2107 * Typically this function should be used by drivers to set caching- and
2108 * encryption bits different than those of @vma->vm_page_prot, because
2109 * the caching- or encryption mode may not be known at mmap() time.
2110 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2111 * to set caching and encryption bits for those vmas (except for COW pages).
2112 * This is ensured by core vm only modifying these page table entries using
2113 * functions that don't touch caching- or encryption bits, using pte_modify()
2114 * if needed. (See for example mprotect()).
2115 * Also when new page-table entries are created, this is only done using the
2116 * fault() callback, and never using the value of vma->vm_page_prot,
2117 * except for page-table entries that point to anonymous pages as the result
2120 * Context: Process context. May allocate using %GFP_KERNEL.
2121 * Return: vm_fault_t value.
2123 vm_fault_t
vmf_insert_mixed_prot(struct vm_area_struct
*vma
, unsigned long addr
,
2124 pfn_t pfn
, pgprot_t pgprot
)
2126 return __vm_insert_mixed(vma
, addr
, pfn
, pgprot
, false);
2128 EXPORT_SYMBOL(vmf_insert_mixed_prot
);
2130 vm_fault_t
vmf_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2133 return __vm_insert_mixed(vma
, addr
, pfn
, vma
->vm_page_prot
, false);
2135 EXPORT_SYMBOL(vmf_insert_mixed
);
2138 * If the insertion of PTE failed because someone else already added a
2139 * different entry in the mean time, we treat that as success as we assume
2140 * the same entry was actually inserted.
2142 vm_fault_t
vmf_insert_mixed_mkwrite(struct vm_area_struct
*vma
,
2143 unsigned long addr
, pfn_t pfn
)
2145 return __vm_insert_mixed(vma
, addr
, pfn
, vma
->vm_page_prot
, true);
2147 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite
);
2150 * maps a range of physical memory into the requested pages. the old
2151 * mappings are removed. any references to nonexistent pages results
2152 * in null mappings (currently treated as "copy-on-access")
2154 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2155 unsigned long addr
, unsigned long end
,
2156 unsigned long pfn
, pgprot_t prot
)
2162 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2165 arch_enter_lazy_mmu_mode();
2167 BUG_ON(!pte_none(*pte
));
2168 if (!pfn_modify_allowed(pfn
, prot
)) {
2172 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2174 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2175 arch_leave_lazy_mmu_mode();
2176 pte_unmap_unlock(pte
- 1, ptl
);
2180 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2181 unsigned long addr
, unsigned long end
,
2182 unsigned long pfn
, pgprot_t prot
)
2188 pfn
-= addr
>> PAGE_SHIFT
;
2189 pmd
= pmd_alloc(mm
, pud
, addr
);
2192 VM_BUG_ON(pmd_trans_huge(*pmd
));
2194 next
= pmd_addr_end(addr
, end
);
2195 err
= remap_pte_range(mm
, pmd
, addr
, next
,
2196 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2199 } while (pmd
++, addr
= next
, addr
!= end
);
2203 static inline int remap_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2204 unsigned long addr
, unsigned long end
,
2205 unsigned long pfn
, pgprot_t prot
)
2211 pfn
-= addr
>> PAGE_SHIFT
;
2212 pud
= pud_alloc(mm
, p4d
, addr
);
2216 next
= pud_addr_end(addr
, end
);
2217 err
= remap_pmd_range(mm
, pud
, addr
, next
,
2218 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2221 } while (pud
++, addr
= next
, addr
!= end
);
2225 static inline int remap_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2226 unsigned long addr
, unsigned long end
,
2227 unsigned long pfn
, pgprot_t prot
)
2233 pfn
-= addr
>> PAGE_SHIFT
;
2234 p4d
= p4d_alloc(mm
, pgd
, addr
);
2238 next
= p4d_addr_end(addr
, end
);
2239 err
= remap_pud_range(mm
, p4d
, addr
, next
,
2240 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2243 } while (p4d
++, addr
= next
, addr
!= end
);
2248 * remap_pfn_range - remap kernel memory to userspace
2249 * @vma: user vma to map to
2250 * @addr: target page aligned user address to start at
2251 * @pfn: page frame number of kernel physical memory address
2252 * @size: size of mapping area
2253 * @prot: page protection flags for this mapping
2255 * Note: this is only safe if the mm semaphore is held when called.
2257 * Return: %0 on success, negative error code otherwise.
2259 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2260 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2264 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2265 struct mm_struct
*mm
= vma
->vm_mm
;
2266 unsigned long remap_pfn
= pfn
;
2269 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr
)))
2273 * Physically remapped pages are special. Tell the
2274 * rest of the world about it:
2275 * VM_IO tells people not to look at these pages
2276 * (accesses can have side effects).
2277 * VM_PFNMAP tells the core MM that the base pages are just
2278 * raw PFN mappings, and do not have a "struct page" associated
2281 * Disable vma merging and expanding with mremap().
2283 * Omit vma from core dump, even when VM_IO turned off.
2285 * There's a horrible special case to handle copy-on-write
2286 * behaviour that some programs depend on. We mark the "original"
2287 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2288 * See vm_normal_page() for details.
2290 if (is_cow_mapping(vma
->vm_flags
)) {
2291 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2293 vma
->vm_pgoff
= pfn
;
2296 err
= track_pfn_remap(vma
, &prot
, remap_pfn
, addr
, PAGE_ALIGN(size
));
2300 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
2302 BUG_ON(addr
>= end
);
2303 pfn
-= addr
>> PAGE_SHIFT
;
2304 pgd
= pgd_offset(mm
, addr
);
2305 flush_cache_range(vma
, addr
, end
);
2307 next
= pgd_addr_end(addr
, end
);
2308 err
= remap_p4d_range(mm
, pgd
, addr
, next
,
2309 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2312 } while (pgd
++, addr
= next
, addr
!= end
);
2315 untrack_pfn(vma
, remap_pfn
, PAGE_ALIGN(size
));
2319 EXPORT_SYMBOL(remap_pfn_range
);
2322 * vm_iomap_memory - remap memory to userspace
2323 * @vma: user vma to map to
2324 * @start: start of the physical memory to be mapped
2325 * @len: size of area
2327 * This is a simplified io_remap_pfn_range() for common driver use. The
2328 * driver just needs to give us the physical memory range to be mapped,
2329 * we'll figure out the rest from the vma information.
2331 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2332 * whatever write-combining details or similar.
2334 * Return: %0 on success, negative error code otherwise.
2336 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2338 unsigned long vm_len
, pfn
, pages
;
2340 /* Check that the physical memory area passed in looks valid */
2341 if (start
+ len
< start
)
2344 * You *really* shouldn't map things that aren't page-aligned,
2345 * but we've historically allowed it because IO memory might
2346 * just have smaller alignment.
2348 len
+= start
& ~PAGE_MASK
;
2349 pfn
= start
>> PAGE_SHIFT
;
2350 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2351 if (pfn
+ pages
< pfn
)
2354 /* We start the mapping 'vm_pgoff' pages into the area */
2355 if (vma
->vm_pgoff
> pages
)
2357 pfn
+= vma
->vm_pgoff
;
2358 pages
-= vma
->vm_pgoff
;
2360 /* Can we fit all of the mapping? */
2361 vm_len
= vma
->vm_end
- vma
->vm_start
;
2362 if (vm_len
>> PAGE_SHIFT
> pages
)
2365 /* Ok, let it rip */
2366 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2368 EXPORT_SYMBOL(vm_iomap_memory
);
2370 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2371 unsigned long addr
, unsigned long end
,
2372 pte_fn_t fn
, void *data
, bool create
,
2373 pgtbl_mod_mask
*mask
)
2380 pte
= (mm
== &init_mm
) ?
2381 pte_alloc_kernel_track(pmd
, addr
, mask
) :
2382 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2386 pte
= (mm
== &init_mm
) ?
2387 pte_offset_kernel(pmd
, addr
) :
2388 pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
2391 BUG_ON(pmd_huge(*pmd
));
2393 arch_enter_lazy_mmu_mode();
2396 if (create
|| !pte_none(*pte
)) {
2397 err
= fn(pte
++, addr
, data
);
2401 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2402 *mask
|= PGTBL_PTE_MODIFIED
;
2404 arch_leave_lazy_mmu_mode();
2407 pte_unmap_unlock(pte
-1, ptl
);
2411 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2412 unsigned long addr
, unsigned long end
,
2413 pte_fn_t fn
, void *data
, bool create
,
2414 pgtbl_mod_mask
*mask
)
2420 BUG_ON(pud_huge(*pud
));
2423 pmd
= pmd_alloc_track(mm
, pud
, addr
, mask
);
2427 pmd
= pmd_offset(pud
, addr
);
2430 next
= pmd_addr_end(addr
, end
);
2431 if (create
|| !pmd_none_or_clear_bad(pmd
)) {
2432 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
,
2437 } while (pmd
++, addr
= next
, addr
!= end
);
2441 static int apply_to_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2442 unsigned long addr
, unsigned long end
,
2443 pte_fn_t fn
, void *data
, bool create
,
2444 pgtbl_mod_mask
*mask
)
2451 pud
= pud_alloc_track(mm
, p4d
, addr
, mask
);
2455 pud
= pud_offset(p4d
, addr
);
2458 next
= pud_addr_end(addr
, end
);
2459 if (create
|| !pud_none_or_clear_bad(pud
)) {
2460 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
,
2465 } while (pud
++, addr
= next
, addr
!= end
);
2469 static int apply_to_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2470 unsigned long addr
, unsigned long end
,
2471 pte_fn_t fn
, void *data
, bool create
,
2472 pgtbl_mod_mask
*mask
)
2479 p4d
= p4d_alloc_track(mm
, pgd
, addr
, mask
);
2483 p4d
= p4d_offset(pgd
, addr
);
2486 next
= p4d_addr_end(addr
, end
);
2487 if (create
|| !p4d_none_or_clear_bad(p4d
)) {
2488 err
= apply_to_pud_range(mm
, p4d
, addr
, next
, fn
, data
,
2493 } while (p4d
++, addr
= next
, addr
!= end
);
2497 static int __apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2498 unsigned long size
, pte_fn_t fn
,
2499 void *data
, bool create
)
2502 unsigned long start
= addr
, next
;
2503 unsigned long end
= addr
+ size
;
2504 pgtbl_mod_mask mask
= 0;
2507 if (WARN_ON(addr
>= end
))
2510 pgd
= pgd_offset(mm
, addr
);
2512 next
= pgd_addr_end(addr
, end
);
2513 if (!create
&& pgd_none_or_clear_bad(pgd
))
2515 err
= apply_to_p4d_range(mm
, pgd
, addr
, next
, fn
, data
, create
, &mask
);
2518 } while (pgd
++, addr
= next
, addr
!= end
);
2520 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
2521 arch_sync_kernel_mappings(start
, start
+ size
);
2527 * Scan a region of virtual memory, filling in page tables as necessary
2528 * and calling a provided function on each leaf page table.
2530 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2531 unsigned long size
, pte_fn_t fn
, void *data
)
2533 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, true);
2535 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2538 * Scan a region of virtual memory, calling a provided function on
2539 * each leaf page table where it exists.
2541 * Unlike apply_to_page_range, this does _not_ fill in page tables
2542 * where they are absent.
2544 int apply_to_existing_page_range(struct mm_struct
*mm
, unsigned long addr
,
2545 unsigned long size
, pte_fn_t fn
, void *data
)
2547 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, false);
2549 EXPORT_SYMBOL_GPL(apply_to_existing_page_range
);
2552 * handle_pte_fault chooses page fault handler according to an entry which was
2553 * read non-atomically. Before making any commitment, on those architectures
2554 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2555 * parts, do_swap_page must check under lock before unmapping the pte and
2556 * proceeding (but do_wp_page is only called after already making such a check;
2557 * and do_anonymous_page can safely check later on).
2559 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2560 pte_t
*page_table
, pte_t orig_pte
)
2563 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2564 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2565 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2567 same
= pte_same(*page_table
, orig_pte
);
2571 pte_unmap(page_table
);
2575 static inline bool cow_user_page(struct page
*dst
, struct page
*src
,
2576 struct vm_fault
*vmf
)
2581 bool locked
= false;
2582 struct vm_area_struct
*vma
= vmf
->vma
;
2583 struct mm_struct
*mm
= vma
->vm_mm
;
2584 unsigned long addr
= vmf
->address
;
2587 copy_user_highpage(dst
, src
, addr
, vma
);
2592 * If the source page was a PFN mapping, we don't have
2593 * a "struct page" for it. We do a best-effort copy by
2594 * just copying from the original user address. If that
2595 * fails, we just zero-fill it. Live with it.
2597 kaddr
= kmap_atomic(dst
);
2598 uaddr
= (void __user
*)(addr
& PAGE_MASK
);
2601 * On architectures with software "accessed" bits, we would
2602 * take a double page fault, so mark it accessed here.
2604 if (arch_faults_on_old_pte() && !pte_young(vmf
->orig_pte
)) {
2607 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2609 if (!likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2611 * Other thread has already handled the fault
2612 * and update local tlb only
2614 update_mmu_tlb(vma
, addr
, vmf
->pte
);
2619 entry
= pte_mkyoung(vmf
->orig_pte
);
2620 if (ptep_set_access_flags(vma
, addr
, vmf
->pte
, entry
, 0))
2621 update_mmu_cache(vma
, addr
, vmf
->pte
);
2625 * This really shouldn't fail, because the page is there
2626 * in the page tables. But it might just be unreadable,
2627 * in which case we just give up and fill the result with
2630 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
2634 /* Re-validate under PTL if the page is still mapped */
2635 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2637 if (!likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2638 /* The PTE changed under us, update local tlb */
2639 update_mmu_tlb(vma
, addr
, vmf
->pte
);
2645 * The same page can be mapped back since last copy attempt.
2646 * Try to copy again under PTL.
2648 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
2650 * Give a warn in case there can be some obscure
2663 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2664 kunmap_atomic(kaddr
);
2665 flush_dcache_page(dst
);
2670 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
2672 struct file
*vm_file
= vma
->vm_file
;
2675 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
2678 * Special mappings (e.g. VDSO) do not have any file so fake
2679 * a default GFP_KERNEL for them.
2685 * Notify the address space that the page is about to become writable so that
2686 * it can prohibit this or wait for the page to get into an appropriate state.
2688 * We do this without the lock held, so that it can sleep if it needs to.
2690 static vm_fault_t
do_page_mkwrite(struct vm_fault
*vmf
)
2693 struct page
*page
= vmf
->page
;
2694 unsigned int old_flags
= vmf
->flags
;
2696 vmf
->flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2698 if (vmf
->vma
->vm_file
&&
2699 IS_SWAPFILE(vmf
->vma
->vm_file
->f_mapping
->host
))
2700 return VM_FAULT_SIGBUS
;
2702 ret
= vmf
->vma
->vm_ops
->page_mkwrite(vmf
);
2703 /* Restore original flags so that caller is not surprised */
2704 vmf
->flags
= old_flags
;
2705 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2707 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
2709 if (!page
->mapping
) {
2711 return 0; /* retry */
2713 ret
|= VM_FAULT_LOCKED
;
2715 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2720 * Handle dirtying of a page in shared file mapping on a write fault.
2722 * The function expects the page to be locked and unlocks it.
2724 static vm_fault_t
fault_dirty_shared_page(struct vm_fault
*vmf
)
2726 struct vm_area_struct
*vma
= vmf
->vma
;
2727 struct address_space
*mapping
;
2728 struct page
*page
= vmf
->page
;
2730 bool page_mkwrite
= vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
;
2732 dirtied
= set_page_dirty(page
);
2733 VM_BUG_ON_PAGE(PageAnon(page
), page
);
2735 * Take a local copy of the address_space - page.mapping may be zeroed
2736 * by truncate after unlock_page(). The address_space itself remains
2737 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2738 * release semantics to prevent the compiler from undoing this copying.
2740 mapping
= page_rmapping(page
);
2744 file_update_time(vma
->vm_file
);
2747 * Throttle page dirtying rate down to writeback speed.
2749 * mapping may be NULL here because some device drivers do not
2750 * set page.mapping but still dirty their pages
2752 * Drop the mmap_lock before waiting on IO, if we can. The file
2753 * is pinning the mapping, as per above.
2755 if ((dirtied
|| page_mkwrite
) && mapping
) {
2758 fpin
= maybe_unlock_mmap_for_io(vmf
, NULL
);
2759 balance_dirty_pages_ratelimited(mapping
);
2762 return VM_FAULT_RETRY
;
2770 * Handle write page faults for pages that can be reused in the current vma
2772 * This can happen either due to the mapping being with the VM_SHARED flag,
2773 * or due to us being the last reference standing to the page. In either
2774 * case, all we need to do here is to mark the page as writable and update
2775 * any related book-keeping.
2777 static inline void wp_page_reuse(struct vm_fault
*vmf
)
2778 __releases(vmf
->ptl
)
2780 struct vm_area_struct
*vma
= vmf
->vma
;
2781 struct page
*page
= vmf
->page
;
2784 * Clear the pages cpupid information as the existing
2785 * information potentially belongs to a now completely
2786 * unrelated process.
2789 page_cpupid_xchg_last(page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2791 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2792 entry
= pte_mkyoung(vmf
->orig_pte
);
2793 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2794 if (ptep_set_access_flags(vma
, vmf
->address
, vmf
->pte
, entry
, 1))
2795 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2796 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2797 count_vm_event(PGREUSE
);
2801 * Handle the case of a page which we actually need to copy to a new page.
2803 * Called with mmap_lock locked and the old page referenced, but
2804 * without the ptl held.
2806 * High level logic flow:
2808 * - Allocate a page, copy the content of the old page to the new one.
2809 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2810 * - Take the PTL. If the pte changed, bail out and release the allocated page
2811 * - If the pte is still the way we remember it, update the page table and all
2812 * relevant references. This includes dropping the reference the page-table
2813 * held to the old page, as well as updating the rmap.
2814 * - In any case, unlock the PTL and drop the reference we took to the old page.
2816 static vm_fault_t
wp_page_copy(struct vm_fault
*vmf
)
2818 struct vm_area_struct
*vma
= vmf
->vma
;
2819 struct mm_struct
*mm
= vma
->vm_mm
;
2820 struct page
*old_page
= vmf
->page
;
2821 struct page
*new_page
= NULL
;
2823 int page_copied
= 0;
2824 struct mmu_notifier_range range
;
2826 if (unlikely(anon_vma_prepare(vma
)))
2829 if (is_zero_pfn(pte_pfn(vmf
->orig_pte
))) {
2830 new_page
= alloc_zeroed_user_highpage_movable(vma
,
2835 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
2840 if (!cow_user_page(new_page
, old_page
, vmf
)) {
2842 * COW failed, if the fault was solved by other,
2843 * it's fine. If not, userspace would re-fault on
2844 * the same address and we will handle the fault
2845 * from the second attempt.
2854 if (mem_cgroup_charge(new_page
, mm
, GFP_KERNEL
))
2856 cgroup_throttle_swaprate(new_page
, GFP_KERNEL
);
2858 __SetPageUptodate(new_page
);
2860 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
,
2861 vmf
->address
& PAGE_MASK
,
2862 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
);
2863 mmu_notifier_invalidate_range_start(&range
);
2866 * Re-check the pte - we dropped the lock
2868 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, vmf
->address
, &vmf
->ptl
);
2869 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2871 if (!PageAnon(old_page
)) {
2872 dec_mm_counter_fast(mm
,
2873 mm_counter_file(old_page
));
2874 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2877 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2879 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2880 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2881 entry
= pte_sw_mkyoung(entry
);
2882 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2884 * Clear the pte entry and flush it first, before updating the
2885 * pte with the new entry. This will avoid a race condition
2886 * seen in the presence of one thread doing SMC and another
2889 ptep_clear_flush_notify(vma
, vmf
->address
, vmf
->pte
);
2890 page_add_new_anon_rmap(new_page
, vma
, vmf
->address
, false);
2891 lru_cache_add_inactive_or_unevictable(new_page
, vma
);
2893 * We call the notify macro here because, when using secondary
2894 * mmu page tables (such as kvm shadow page tables), we want the
2895 * new page to be mapped directly into the secondary page table.
2897 set_pte_at_notify(mm
, vmf
->address
, vmf
->pte
, entry
);
2898 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2901 * Only after switching the pte to the new page may
2902 * we remove the mapcount here. Otherwise another
2903 * process may come and find the rmap count decremented
2904 * before the pte is switched to the new page, and
2905 * "reuse" the old page writing into it while our pte
2906 * here still points into it and can be read by other
2909 * The critical issue is to order this
2910 * page_remove_rmap with the ptp_clear_flush above.
2911 * Those stores are ordered by (if nothing else,)
2912 * the barrier present in the atomic_add_negative
2913 * in page_remove_rmap.
2915 * Then the TLB flush in ptep_clear_flush ensures that
2916 * no process can access the old page before the
2917 * decremented mapcount is visible. And the old page
2918 * cannot be reused until after the decremented
2919 * mapcount is visible. So transitively, TLBs to
2920 * old page will be flushed before it can be reused.
2922 page_remove_rmap(old_page
, false);
2925 /* Free the old page.. */
2926 new_page
= old_page
;
2929 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
2935 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2937 * No need to double call mmu_notifier->invalidate_range() callback as
2938 * the above ptep_clear_flush_notify() did already call it.
2940 mmu_notifier_invalidate_range_only_end(&range
);
2943 * Don't let another task, with possibly unlocked vma,
2944 * keep the mlocked page.
2946 if (page_copied
&& (vma
->vm_flags
& VM_LOCKED
)) {
2947 lock_page(old_page
); /* LRU manipulation */
2948 if (PageMlocked(old_page
))
2949 munlock_vma_page(old_page
);
2950 unlock_page(old_page
);
2954 return page_copied
? VM_FAULT_WRITE
: 0;
2960 return VM_FAULT_OOM
;
2964 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2965 * writeable once the page is prepared
2967 * @vmf: structure describing the fault
2969 * This function handles all that is needed to finish a write page fault in a
2970 * shared mapping due to PTE being read-only once the mapped page is prepared.
2971 * It handles locking of PTE and modifying it.
2973 * The function expects the page to be locked or other protection against
2974 * concurrent faults / writeback (such as DAX radix tree locks).
2976 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2977 * we acquired PTE lock.
2979 vm_fault_t
finish_mkwrite_fault(struct vm_fault
*vmf
)
2981 WARN_ON_ONCE(!(vmf
->vma
->vm_flags
& VM_SHARED
));
2982 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
2985 * We might have raced with another page fault while we released the
2986 * pte_offset_map_lock.
2988 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
2989 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
2990 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2991 return VM_FAULT_NOPAGE
;
2998 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3001 static vm_fault_t
wp_pfn_shared(struct vm_fault
*vmf
)
3003 struct vm_area_struct
*vma
= vmf
->vma
;
3005 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
3008 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3009 vmf
->flags
|= FAULT_FLAG_MKWRITE
;
3010 ret
= vma
->vm_ops
->pfn_mkwrite(vmf
);
3011 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))
3013 return finish_mkwrite_fault(vmf
);
3016 return VM_FAULT_WRITE
;
3019 static vm_fault_t
wp_page_shared(struct vm_fault
*vmf
)
3020 __releases(vmf
->ptl
)
3022 struct vm_area_struct
*vma
= vmf
->vma
;
3023 vm_fault_t ret
= VM_FAULT_WRITE
;
3025 get_page(vmf
->page
);
3027 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
3030 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3031 tmp
= do_page_mkwrite(vmf
);
3032 if (unlikely(!tmp
|| (tmp
&
3033 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3034 put_page(vmf
->page
);
3037 tmp
= finish_mkwrite_fault(vmf
);
3038 if (unlikely(tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3039 unlock_page(vmf
->page
);
3040 put_page(vmf
->page
);
3045 lock_page(vmf
->page
);
3047 ret
|= fault_dirty_shared_page(vmf
);
3048 put_page(vmf
->page
);
3054 * This routine handles present pages, when users try to write
3055 * to a shared page. It is done by copying the page to a new address
3056 * and decrementing the shared-page counter for the old page.
3058 * Note that this routine assumes that the protection checks have been
3059 * done by the caller (the low-level page fault routine in most cases).
3060 * Thus we can safely just mark it writable once we've done any necessary
3063 * We also mark the page dirty at this point even though the page will
3064 * change only once the write actually happens. This avoids a few races,
3065 * and potentially makes it more efficient.
3067 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3068 * but allow concurrent faults), with pte both mapped and locked.
3069 * We return with mmap_lock still held, but pte unmapped and unlocked.
3071 static vm_fault_t
do_wp_page(struct vm_fault
*vmf
)
3072 __releases(vmf
->ptl
)
3074 struct vm_area_struct
*vma
= vmf
->vma
;
3076 if (userfaultfd_pte_wp(vma
, *vmf
->pte
)) {
3077 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3078 return handle_userfault(vmf
, VM_UFFD_WP
);
3081 vmf
->page
= vm_normal_page(vma
, vmf
->address
, vmf
->orig_pte
);
3084 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3087 * We should not cow pages in a shared writeable mapping.
3088 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3090 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
3091 (VM_WRITE
|VM_SHARED
))
3092 return wp_pfn_shared(vmf
);
3094 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3095 return wp_page_copy(vmf
);
3099 * Take out anonymous pages first, anonymous shared vmas are
3100 * not dirty accountable.
3102 if (PageAnon(vmf
->page
)) {
3103 struct page
*page
= vmf
->page
;
3105 /* PageKsm() doesn't necessarily raise the page refcount */
3106 if (PageKsm(page
) || page_count(page
) != 1)
3108 if (!trylock_page(page
))
3110 if (PageKsm(page
) || page_mapcount(page
) != 1 || page_count(page
) != 1) {
3115 * Ok, we've got the only map reference, and the only
3116 * page count reference, and the page is locked,
3117 * it's dark out, and we're wearing sunglasses. Hit it.
3121 return VM_FAULT_WRITE
;
3122 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
3123 (VM_WRITE
|VM_SHARED
))) {
3124 return wp_page_shared(vmf
);
3128 * Ok, we need to copy. Oh, well..
3130 get_page(vmf
->page
);
3132 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3133 return wp_page_copy(vmf
);
3136 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
3137 unsigned long start_addr
, unsigned long end_addr
,
3138 struct zap_details
*details
)
3140 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
3143 static inline void unmap_mapping_range_tree(struct rb_root_cached
*root
,
3144 struct zap_details
*details
)
3146 struct vm_area_struct
*vma
;
3147 pgoff_t vba
, vea
, zba
, zea
;
3149 vma_interval_tree_foreach(vma
, root
,
3150 details
->first_index
, details
->last_index
) {
3152 vba
= vma
->vm_pgoff
;
3153 vea
= vba
+ vma_pages(vma
) - 1;
3154 zba
= details
->first_index
;
3157 zea
= details
->last_index
;
3161 unmap_mapping_range_vma(vma
,
3162 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
3163 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
3169 * unmap_mapping_pages() - Unmap pages from processes.
3170 * @mapping: The address space containing pages to be unmapped.
3171 * @start: Index of first page to be unmapped.
3172 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3173 * @even_cows: Whether to unmap even private COWed pages.
3175 * Unmap the pages in this address space from any userspace process which
3176 * has them mmaped. Generally, you want to remove COWed pages as well when
3177 * a file is being truncated, but not when invalidating pages from the page
3180 void unmap_mapping_pages(struct address_space
*mapping
, pgoff_t start
,
3181 pgoff_t nr
, bool even_cows
)
3183 struct zap_details details
= { };
3185 details
.check_mapping
= even_cows
? NULL
: mapping
;
3186 details
.first_index
= start
;
3187 details
.last_index
= start
+ nr
- 1;
3188 if (details
.last_index
< details
.first_index
)
3189 details
.last_index
= ULONG_MAX
;
3191 i_mmap_lock_write(mapping
);
3192 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3193 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
3194 i_mmap_unlock_write(mapping
);
3198 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3199 * address_space corresponding to the specified byte range in the underlying
3202 * @mapping: the address space containing mmaps to be unmapped.
3203 * @holebegin: byte in first page to unmap, relative to the start of
3204 * the underlying file. This will be rounded down to a PAGE_SIZE
3205 * boundary. Note that this is different from truncate_pagecache(), which
3206 * must keep the partial page. In contrast, we must get rid of
3208 * @holelen: size of prospective hole in bytes. This will be rounded
3209 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3211 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3212 * but 0 when invalidating pagecache, don't throw away private data.
3214 void unmap_mapping_range(struct address_space
*mapping
,
3215 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
3217 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
3218 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3220 /* Check for overflow. */
3221 if (sizeof(holelen
) > sizeof(hlen
)) {
3223 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3224 if (holeend
& ~(long long)ULONG_MAX
)
3225 hlen
= ULONG_MAX
- hba
+ 1;
3228 unmap_mapping_pages(mapping
, hba
, hlen
, even_cows
);
3230 EXPORT_SYMBOL(unmap_mapping_range
);
3233 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3234 * but allow concurrent faults), and pte mapped but not yet locked.
3235 * We return with pte unmapped and unlocked.
3237 * We return with the mmap_lock locked or unlocked in the same cases
3238 * as does filemap_fault().
3240 vm_fault_t
do_swap_page(struct vm_fault
*vmf
)
3242 struct vm_area_struct
*vma
= vmf
->vma
;
3243 struct page
*page
= NULL
, *swapcache
;
3249 void *shadow
= NULL
;
3251 if (!pte_unmap_same(vma
->vm_mm
, vmf
->pmd
, vmf
->pte
, vmf
->orig_pte
))
3254 entry
= pte_to_swp_entry(vmf
->orig_pte
);
3255 if (unlikely(non_swap_entry(entry
))) {
3256 if (is_migration_entry(entry
)) {
3257 migration_entry_wait(vma
->vm_mm
, vmf
->pmd
,
3259 } else if (is_device_private_entry(entry
)) {
3260 vmf
->page
= device_private_entry_to_page(entry
);
3261 ret
= vmf
->page
->pgmap
->ops
->migrate_to_ram(vmf
);
3262 } else if (is_hwpoison_entry(entry
)) {
3263 ret
= VM_FAULT_HWPOISON
;
3265 print_bad_pte(vma
, vmf
->address
, vmf
->orig_pte
, NULL
);
3266 ret
= VM_FAULT_SIGBUS
;
3272 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
3273 page
= lookup_swap_cache(entry
, vma
, vmf
->address
);
3277 struct swap_info_struct
*si
= swp_swap_info(entry
);
3279 if (data_race(si
->flags
& SWP_SYNCHRONOUS_IO
) &&
3280 __swap_count(entry
) == 1) {
3281 /* skip swapcache */
3282 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
3287 __SetPageLocked(page
);
3288 __SetPageSwapBacked(page
);
3289 set_page_private(page
, entry
.val
);
3291 /* Tell memcg to use swap ownership records */
3292 SetPageSwapCache(page
);
3293 err
= mem_cgroup_charge(page
, vma
->vm_mm
,
3295 ClearPageSwapCache(page
);
3301 shadow
= get_shadow_from_swap_cache(entry
);
3303 workingset_refault(page
, shadow
);
3305 lru_cache_add(page
);
3306 swap_readpage(page
, true);
3309 page
= swapin_readahead(entry
, GFP_HIGHUSER_MOVABLE
,
3316 * Back out if somebody else faulted in this pte
3317 * while we released the pte lock.
3319 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3320 vmf
->address
, &vmf
->ptl
);
3321 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3323 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3327 /* Had to read the page from swap area: Major fault */
3328 ret
= VM_FAULT_MAJOR
;
3329 count_vm_event(PGMAJFAULT
);
3330 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
3331 } else if (PageHWPoison(page
)) {
3333 * hwpoisoned dirty swapcache pages are kept for killing
3334 * owner processes (which may be unknown at hwpoison time)
3336 ret
= VM_FAULT_HWPOISON
;
3337 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3341 locked
= lock_page_or_retry(page
, vma
->vm_mm
, vmf
->flags
);
3343 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3345 ret
|= VM_FAULT_RETRY
;
3350 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3351 * release the swapcache from under us. The page pin, and pte_same
3352 * test below, are not enough to exclude that. Even if it is still
3353 * swapcache, we need to check that the page's swap has not changed.
3355 if (unlikely((!PageSwapCache(page
) ||
3356 page_private(page
) != entry
.val
)) && swapcache
)
3359 page
= ksm_might_need_to_copy(page
, vma
, vmf
->address
);
3360 if (unlikely(!page
)) {
3366 cgroup_throttle_swaprate(page
, GFP_KERNEL
);
3369 * Back out if somebody else already faulted in this pte.
3371 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3373 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3376 if (unlikely(!PageUptodate(page
))) {
3377 ret
= VM_FAULT_SIGBUS
;
3382 * The page isn't present yet, go ahead with the fault.
3384 * Be careful about the sequence of operations here.
3385 * To get its accounting right, reuse_swap_page() must be called
3386 * while the page is counted on swap but not yet in mapcount i.e.
3387 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3388 * must be called after the swap_free(), or it will never succeed.
3391 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3392 dec_mm_counter_fast(vma
->vm_mm
, MM_SWAPENTS
);
3393 pte
= mk_pte(page
, vma
->vm_page_prot
);
3394 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
, NULL
)) {
3395 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3396 vmf
->flags
&= ~FAULT_FLAG_WRITE
;
3397 ret
|= VM_FAULT_WRITE
;
3398 exclusive
= RMAP_EXCLUSIVE
;
3400 flush_icache_page(vma
, page
);
3401 if (pte_swp_soft_dirty(vmf
->orig_pte
))
3402 pte
= pte_mksoft_dirty(pte
);
3403 if (pte_swp_uffd_wp(vmf
->orig_pte
)) {
3404 pte
= pte_mkuffd_wp(pte
);
3405 pte
= pte_wrprotect(pte
);
3407 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3408 arch_do_swap_page(vma
->vm_mm
, vma
, vmf
->address
, pte
, vmf
->orig_pte
);
3409 vmf
->orig_pte
= pte
;
3411 /* ksm created a completely new copy */
3412 if (unlikely(page
!= swapcache
&& swapcache
)) {
3413 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3414 lru_cache_add_inactive_or_unevictable(page
, vma
);
3416 do_page_add_anon_rmap(page
, vma
, vmf
->address
, exclusive
);
3420 if (mem_cgroup_swap_full(page
) ||
3421 (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
3422 try_to_free_swap(page
);
3424 if (page
!= swapcache
&& swapcache
) {
3426 * Hold the lock to avoid the swap entry to be reused
3427 * until we take the PT lock for the pte_same() check
3428 * (to avoid false positives from pte_same). For
3429 * further safety release the lock after the swap_free
3430 * so that the swap count won't change under a
3431 * parallel locked swapcache.
3433 unlock_page(swapcache
);
3434 put_page(swapcache
);
3437 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
3438 ret
|= do_wp_page(vmf
);
3439 if (ret
& VM_FAULT_ERROR
)
3440 ret
&= VM_FAULT_ERROR
;
3444 /* No need to invalidate - it was non-present before */
3445 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3447 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3451 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3456 if (page
!= swapcache
&& swapcache
) {
3457 unlock_page(swapcache
);
3458 put_page(swapcache
);
3464 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3465 * but allow concurrent faults), and pte mapped but not yet locked.
3466 * We return with mmap_lock still held, but pte unmapped and unlocked.
3468 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
)
3470 struct vm_area_struct
*vma
= vmf
->vma
;
3475 /* File mapping without ->vm_ops ? */
3476 if (vma
->vm_flags
& VM_SHARED
)
3477 return VM_FAULT_SIGBUS
;
3480 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3481 * pte_offset_map() on pmds where a huge pmd might be created
3482 * from a different thread.
3484 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3485 * parallel threads are excluded by other means.
3487 * Here we only have mmap_read_lock(mm).
3489 if (pte_alloc(vma
->vm_mm
, vmf
->pmd
))
3490 return VM_FAULT_OOM
;
3492 /* See the comment in pte_alloc_one_map() */
3493 if (unlikely(pmd_trans_unstable(vmf
->pmd
)))
3496 /* Use the zero-page for reads */
3497 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
3498 !mm_forbids_zeropage(vma
->vm_mm
)) {
3499 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(vmf
->address
),
3500 vma
->vm_page_prot
));
3501 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3502 vmf
->address
, &vmf
->ptl
);
3503 if (!pte_none(*vmf
->pte
)) {
3504 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
3507 ret
= check_stable_address_space(vma
->vm_mm
);
3510 /* Deliver the page fault to userland, check inside PT lock */
3511 if (userfaultfd_missing(vma
)) {
3512 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3513 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3518 /* Allocate our own private page. */
3519 if (unlikely(anon_vma_prepare(vma
)))
3521 page
= alloc_zeroed_user_highpage_movable(vma
, vmf
->address
);
3525 if (mem_cgroup_charge(page
, vma
->vm_mm
, GFP_KERNEL
))
3527 cgroup_throttle_swaprate(page
, GFP_KERNEL
);
3530 * The memory barrier inside __SetPageUptodate makes sure that
3531 * preceding stores to the page contents become visible before
3532 * the set_pte_at() write.
3534 __SetPageUptodate(page
);
3536 entry
= mk_pte(page
, vma
->vm_page_prot
);
3537 entry
= pte_sw_mkyoung(entry
);
3538 if (vma
->vm_flags
& VM_WRITE
)
3539 entry
= pte_mkwrite(pte_mkdirty(entry
));
3541 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3543 if (!pte_none(*vmf
->pte
)) {
3544 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3548 ret
= check_stable_address_space(vma
->vm_mm
);
3552 /* Deliver the page fault to userland, check inside PT lock */
3553 if (userfaultfd_missing(vma
)) {
3554 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3556 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3559 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3560 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3561 lru_cache_add_inactive_or_unevictable(page
, vma
);
3563 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3565 /* No need to invalidate - it was non-present before */
3566 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3568 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3576 return VM_FAULT_OOM
;
3580 * The mmap_lock must have been held on entry, and may have been
3581 * released depending on flags and vma->vm_ops->fault() return value.
3582 * See filemap_fault() and __lock_page_retry().
3584 static vm_fault_t
__do_fault(struct vm_fault
*vmf
)
3586 struct vm_area_struct
*vma
= vmf
->vma
;
3590 * Preallocate pte before we take page_lock because this might lead to
3591 * deadlocks for memcg reclaim which waits for pages under writeback:
3593 * SetPageWriteback(A)
3599 * wait_on_page_writeback(A)
3600 * SetPageWriteback(B)
3602 * # flush A, B to clear the writeback
3604 if (pmd_none(*vmf
->pmd
) && !vmf
->prealloc_pte
) {
3605 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
3606 if (!vmf
->prealloc_pte
)
3607 return VM_FAULT_OOM
;
3608 smp_wmb(); /* See comment in __pte_alloc() */
3611 ret
= vma
->vm_ops
->fault(vmf
);
3612 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
|
3613 VM_FAULT_DONE_COW
)))
3616 if (unlikely(PageHWPoison(vmf
->page
))) {
3617 if (ret
& VM_FAULT_LOCKED
)
3618 unlock_page(vmf
->page
);
3619 put_page(vmf
->page
);
3621 return VM_FAULT_HWPOISON
;
3624 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3625 lock_page(vmf
->page
);
3627 VM_BUG_ON_PAGE(!PageLocked(vmf
->page
), vmf
->page
);
3633 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3634 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3635 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3636 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3638 static int pmd_devmap_trans_unstable(pmd_t
*pmd
)
3640 return pmd_devmap(*pmd
) || pmd_trans_unstable(pmd
);
3643 static vm_fault_t
pte_alloc_one_map(struct vm_fault
*vmf
)
3645 struct vm_area_struct
*vma
= vmf
->vma
;
3647 if (!pmd_none(*vmf
->pmd
))
3649 if (vmf
->prealloc_pte
) {
3650 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3651 if (unlikely(!pmd_none(*vmf
->pmd
))) {
3652 spin_unlock(vmf
->ptl
);
3656 mm_inc_nr_ptes(vma
->vm_mm
);
3657 pmd_populate(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3658 spin_unlock(vmf
->ptl
);
3659 vmf
->prealloc_pte
= NULL
;
3660 } else if (unlikely(pte_alloc(vma
->vm_mm
, vmf
->pmd
))) {
3661 return VM_FAULT_OOM
;
3665 * If a huge pmd materialized under us just retry later. Use
3666 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3667 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3668 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3669 * running immediately after a huge pmd fault in a different thread of
3670 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3671 * All we have to ensure is that it is a regular pmd that we can walk
3672 * with pte_offset_map() and we can do that through an atomic read in
3673 * C, which is what pmd_trans_unstable() provides.
3675 if (pmd_devmap_trans_unstable(vmf
->pmd
))
3676 return VM_FAULT_NOPAGE
;
3679 * At this point we know that our vmf->pmd points to a page of ptes
3680 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3681 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3682 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3683 * be valid and we will re-check to make sure the vmf->pte isn't
3684 * pte_none() under vmf->ptl protection when we return to
3687 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3692 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3693 static void deposit_prealloc_pte(struct vm_fault
*vmf
)
3695 struct vm_area_struct
*vma
= vmf
->vma
;
3697 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3699 * We are going to consume the prealloc table,
3700 * count that as nr_ptes.
3702 mm_inc_nr_ptes(vma
->vm_mm
);
3703 vmf
->prealloc_pte
= NULL
;
3706 static vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3708 struct vm_area_struct
*vma
= vmf
->vma
;
3709 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3710 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
3715 if (!transhuge_vma_suitable(vma
, haddr
))
3716 return VM_FAULT_FALLBACK
;
3718 ret
= VM_FAULT_FALLBACK
;
3719 page
= compound_head(page
);
3722 * Archs like ppc64 need additonal space to store information
3723 * related to pte entry. Use the preallocated table for that.
3725 if (arch_needs_pgtable_deposit() && !vmf
->prealloc_pte
) {
3726 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
3727 if (!vmf
->prealloc_pte
)
3728 return VM_FAULT_OOM
;
3729 smp_wmb(); /* See comment in __pte_alloc() */
3732 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3733 if (unlikely(!pmd_none(*vmf
->pmd
)))
3736 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
3737 flush_icache_page(vma
, page
+ i
);
3739 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
3741 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
3743 add_mm_counter(vma
->vm_mm
, mm_counter_file(page
), HPAGE_PMD_NR
);
3744 page_add_file_rmap(page
, true);
3746 * deposit and withdraw with pmd lock held
3748 if (arch_needs_pgtable_deposit())
3749 deposit_prealloc_pte(vmf
);
3751 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
3753 update_mmu_cache_pmd(vma
, haddr
, vmf
->pmd
);
3755 /* fault is handled */
3757 count_vm_event(THP_FILE_MAPPED
);
3759 spin_unlock(vmf
->ptl
);
3763 static vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3771 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3772 * mapping. If needed, the function allocates page table or use pre-allocated.
3774 * @vmf: fault environment
3775 * @page: page to map
3777 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3780 * Target users are page handler itself and implementations of
3781 * vm_ops->map_pages.
3783 * Return: %0 on success, %VM_FAULT_ code in case of error.
3785 vm_fault_t
alloc_set_pte(struct vm_fault
*vmf
, struct page
*page
)
3787 struct vm_area_struct
*vma
= vmf
->vma
;
3788 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3792 if (pmd_none(*vmf
->pmd
) && PageTransCompound(page
)) {
3793 ret
= do_set_pmd(vmf
, page
);
3794 if (ret
!= VM_FAULT_FALLBACK
)
3799 ret
= pte_alloc_one_map(vmf
);
3804 /* Re-check under ptl */
3805 if (unlikely(!pte_none(*vmf
->pte
))) {
3806 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
3807 return VM_FAULT_NOPAGE
;
3810 flush_icache_page(vma
, page
);
3811 entry
= mk_pte(page
, vma
->vm_page_prot
);
3812 entry
= pte_sw_mkyoung(entry
);
3814 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3815 /* copy-on-write page */
3816 if (write
&& !(vma
->vm_flags
& VM_SHARED
)) {
3817 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3818 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3819 lru_cache_add_inactive_or_unevictable(page
, vma
);
3821 inc_mm_counter_fast(vma
->vm_mm
, mm_counter_file(page
));
3822 page_add_file_rmap(page
, false);
3824 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3826 /* no need to invalidate: a not-present page won't be cached */
3827 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3834 * finish_fault - finish page fault once we have prepared the page to fault
3836 * @vmf: structure describing the fault
3838 * This function handles all that is needed to finish a page fault once the
3839 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3840 * given page, adds reverse page mapping, handles memcg charges and LRU
3843 * The function expects the page to be locked and on success it consumes a
3844 * reference of a page being mapped (for the PTE which maps it).
3846 * Return: %0 on success, %VM_FAULT_ code in case of error.
3848 vm_fault_t
finish_fault(struct vm_fault
*vmf
)
3853 /* Did we COW the page? */
3854 if ((vmf
->flags
& FAULT_FLAG_WRITE
) &&
3855 !(vmf
->vma
->vm_flags
& VM_SHARED
))
3856 page
= vmf
->cow_page
;
3861 * check even for read faults because we might have lost our CoWed
3864 if (!(vmf
->vma
->vm_flags
& VM_SHARED
))
3865 ret
= check_stable_address_space(vmf
->vma
->vm_mm
);
3867 ret
= alloc_set_pte(vmf
, page
);
3869 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3873 static unsigned long fault_around_bytes __read_mostly
=
3874 rounddown_pow_of_two(65536);
3876 #ifdef CONFIG_DEBUG_FS
3877 static int fault_around_bytes_get(void *data
, u64
*val
)
3879 *val
= fault_around_bytes
;
3884 * fault_around_bytes must be rounded down to the nearest page order as it's
3885 * what do_fault_around() expects to see.
3887 static int fault_around_bytes_set(void *data
, u64 val
)
3889 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
3891 if (val
> PAGE_SIZE
)
3892 fault_around_bytes
= rounddown_pow_of_two(val
);
3894 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
3897 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops
,
3898 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
3900 static int __init
fault_around_debugfs(void)
3902 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL
, NULL
,
3903 &fault_around_bytes_fops
);
3906 late_initcall(fault_around_debugfs
);
3910 * do_fault_around() tries to map few pages around the fault address. The hope
3911 * is that the pages will be needed soon and this will lower the number of
3914 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3915 * not ready to be mapped: not up-to-date, locked, etc.
3917 * This function is called with the page table lock taken. In the split ptlock
3918 * case the page table lock only protects only those entries which belong to
3919 * the page table corresponding to the fault address.
3921 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3924 * fault_around_bytes defines how many bytes we'll try to map.
3925 * do_fault_around() expects it to be set to a power of two less than or equal
3928 * The virtual address of the area that we map is naturally aligned to
3929 * fault_around_bytes rounded down to the machine page size
3930 * (and therefore to page order). This way it's easier to guarantee
3931 * that we don't cross page table boundaries.
3933 static vm_fault_t
do_fault_around(struct vm_fault
*vmf
)
3935 unsigned long address
= vmf
->address
, nr_pages
, mask
;
3936 pgoff_t start_pgoff
= vmf
->pgoff
;
3941 nr_pages
= READ_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
3942 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
3944 vmf
->address
= max(address
& mask
, vmf
->vma
->vm_start
);
3945 off
= ((address
- vmf
->address
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
3949 * end_pgoff is either the end of the page table, the end of
3950 * the vma or nr_pages from start_pgoff, depending what is nearest.
3952 end_pgoff
= start_pgoff
-
3953 ((vmf
->address
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
3955 end_pgoff
= min3(end_pgoff
, vma_pages(vmf
->vma
) + vmf
->vma
->vm_pgoff
- 1,
3956 start_pgoff
+ nr_pages
- 1);
3958 if (pmd_none(*vmf
->pmd
)) {
3959 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
);
3960 if (!vmf
->prealloc_pte
)
3962 smp_wmb(); /* See comment in __pte_alloc() */
3965 vmf
->vma
->vm_ops
->map_pages(vmf
, start_pgoff
, end_pgoff
);
3967 /* Huge page is mapped? Page fault is solved */
3968 if (pmd_trans_huge(*vmf
->pmd
)) {
3969 ret
= VM_FAULT_NOPAGE
;
3973 /* ->map_pages() haven't done anything useful. Cold page cache? */
3977 /* check if the page fault is solved */
3978 vmf
->pte
-= (vmf
->address
>> PAGE_SHIFT
) - (address
>> PAGE_SHIFT
);
3979 if (!pte_none(*vmf
->pte
))
3980 ret
= VM_FAULT_NOPAGE
;
3981 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3983 vmf
->address
= address
;
3988 static vm_fault_t
do_read_fault(struct vm_fault
*vmf
)
3990 struct vm_area_struct
*vma
= vmf
->vma
;
3994 * Let's call ->map_pages() first and use ->fault() as fallback
3995 * if page by the offset is not ready to be mapped (cold cache or
3998 if (vma
->vm_ops
->map_pages
&& fault_around_bytes
>> PAGE_SHIFT
> 1) {
3999 ret
= do_fault_around(vmf
);
4004 ret
= __do_fault(vmf
);
4005 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4008 ret
|= finish_fault(vmf
);
4009 unlock_page(vmf
->page
);
4010 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4011 put_page(vmf
->page
);
4015 static vm_fault_t
do_cow_fault(struct vm_fault
*vmf
)
4017 struct vm_area_struct
*vma
= vmf
->vma
;
4020 if (unlikely(anon_vma_prepare(vma
)))
4021 return VM_FAULT_OOM
;
4023 vmf
->cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, vmf
->address
);
4025 return VM_FAULT_OOM
;
4027 if (mem_cgroup_charge(vmf
->cow_page
, vma
->vm_mm
, GFP_KERNEL
)) {
4028 put_page(vmf
->cow_page
);
4029 return VM_FAULT_OOM
;
4031 cgroup_throttle_swaprate(vmf
->cow_page
, GFP_KERNEL
);
4033 ret
= __do_fault(vmf
);
4034 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4036 if (ret
& VM_FAULT_DONE_COW
)
4039 copy_user_highpage(vmf
->cow_page
, vmf
->page
, vmf
->address
, vma
);
4040 __SetPageUptodate(vmf
->cow_page
);
4042 ret
|= finish_fault(vmf
);
4043 unlock_page(vmf
->page
);
4044 put_page(vmf
->page
);
4045 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4049 put_page(vmf
->cow_page
);
4053 static vm_fault_t
do_shared_fault(struct vm_fault
*vmf
)
4055 struct vm_area_struct
*vma
= vmf
->vma
;
4056 vm_fault_t ret
, tmp
;
4058 ret
= __do_fault(vmf
);
4059 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4063 * Check if the backing address space wants to know that the page is
4064 * about to become writable
4066 if (vma
->vm_ops
->page_mkwrite
) {
4067 unlock_page(vmf
->page
);
4068 tmp
= do_page_mkwrite(vmf
);
4069 if (unlikely(!tmp
||
4070 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
4071 put_page(vmf
->page
);
4076 ret
|= finish_fault(vmf
);
4077 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
4079 unlock_page(vmf
->page
);
4080 put_page(vmf
->page
);
4084 ret
|= fault_dirty_shared_page(vmf
);
4089 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4090 * but allow concurrent faults).
4091 * The mmap_lock may have been released depending on flags and our
4092 * return value. See filemap_fault() and __lock_page_or_retry().
4093 * If mmap_lock is released, vma may become invalid (for example
4094 * by other thread calling munmap()).
4096 static vm_fault_t
do_fault(struct vm_fault
*vmf
)
4098 struct vm_area_struct
*vma
= vmf
->vma
;
4099 struct mm_struct
*vm_mm
= vma
->vm_mm
;
4103 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4105 if (!vma
->vm_ops
->fault
) {
4107 * If we find a migration pmd entry or a none pmd entry, which
4108 * should never happen, return SIGBUS
4110 if (unlikely(!pmd_present(*vmf
->pmd
)))
4111 ret
= VM_FAULT_SIGBUS
;
4113 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
,
4118 * Make sure this is not a temporary clearing of pte
4119 * by holding ptl and checking again. A R/M/W update
4120 * of pte involves: take ptl, clearing the pte so that
4121 * we don't have concurrent modification by hardware
4122 * followed by an update.
4124 if (unlikely(pte_none(*vmf
->pte
)))
4125 ret
= VM_FAULT_SIGBUS
;
4127 ret
= VM_FAULT_NOPAGE
;
4129 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4131 } else if (!(vmf
->flags
& FAULT_FLAG_WRITE
))
4132 ret
= do_read_fault(vmf
);
4133 else if (!(vma
->vm_flags
& VM_SHARED
))
4134 ret
= do_cow_fault(vmf
);
4136 ret
= do_shared_fault(vmf
);
4138 /* preallocated pagetable is unused: free it */
4139 if (vmf
->prealloc_pte
) {
4140 pte_free(vm_mm
, vmf
->prealloc_pte
);
4141 vmf
->prealloc_pte
= NULL
;
4146 static int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
4147 unsigned long addr
, int page_nid
,
4152 count_vm_numa_event(NUMA_HINT_FAULTS
);
4153 if (page_nid
== numa_node_id()) {
4154 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
4155 *flags
|= TNF_FAULT_LOCAL
;
4158 return mpol_misplaced(page
, vma
, addr
);
4161 static vm_fault_t
do_numa_page(struct vm_fault
*vmf
)
4163 struct vm_area_struct
*vma
= vmf
->vma
;
4164 struct page
*page
= NULL
;
4165 int page_nid
= NUMA_NO_NODE
;
4168 bool migrated
= false;
4170 bool was_writable
= pte_savedwrite(vmf
->orig_pte
);
4174 * The "pte" at this point cannot be used safely without
4175 * validation through pte_unmap_same(). It's of NUMA type but
4176 * the pfn may be screwed if the read is non atomic.
4178 vmf
->ptl
= pte_lockptr(vma
->vm_mm
, vmf
->pmd
);
4179 spin_lock(vmf
->ptl
);
4180 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
4181 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4186 * Make it present again, Depending on how arch implementes non
4187 * accessible ptes, some can allow access by kernel mode.
4189 old_pte
= ptep_modify_prot_start(vma
, vmf
->address
, vmf
->pte
);
4190 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
4191 pte
= pte_mkyoung(pte
);
4193 pte
= pte_mkwrite(pte
);
4194 ptep_modify_prot_commit(vma
, vmf
->address
, vmf
->pte
, old_pte
, pte
);
4195 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
4197 page
= vm_normal_page(vma
, vmf
->address
, pte
);
4199 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4203 /* TODO: handle PTE-mapped THP */
4204 if (PageCompound(page
)) {
4205 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4210 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4211 * much anyway since they can be in shared cache state. This misses
4212 * the case where a mapping is writable but the process never writes
4213 * to it but pte_write gets cleared during protection updates and
4214 * pte_dirty has unpredictable behaviour between PTE scan updates,
4215 * background writeback, dirty balancing and application behaviour.
4217 if (!pte_write(pte
))
4218 flags
|= TNF_NO_GROUP
;
4221 * Flag if the page is shared between multiple address spaces. This
4222 * is later used when determining whether to group tasks together
4224 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
4225 flags
|= TNF_SHARED
;
4227 last_cpupid
= page_cpupid_last(page
);
4228 page_nid
= page_to_nid(page
);
4229 target_nid
= numa_migrate_prep(page
, vma
, vmf
->address
, page_nid
,
4231 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4232 if (target_nid
== NUMA_NO_NODE
) {
4237 /* Migrate to the requested node */
4238 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
4240 page_nid
= target_nid
;
4241 flags
|= TNF_MIGRATED
;
4243 flags
|= TNF_MIGRATE_FAIL
;
4246 if (page_nid
!= NUMA_NO_NODE
)
4247 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
4251 static inline vm_fault_t
create_huge_pmd(struct vm_fault
*vmf
)
4253 if (vma_is_anonymous(vmf
->vma
))
4254 return do_huge_pmd_anonymous_page(vmf
);
4255 if (vmf
->vma
->vm_ops
->huge_fault
)
4256 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
4257 return VM_FAULT_FALLBACK
;
4260 /* `inline' is required to avoid gcc 4.1.2 build error */
4261 static inline vm_fault_t
wp_huge_pmd(struct vm_fault
*vmf
, pmd_t orig_pmd
)
4263 if (vma_is_anonymous(vmf
->vma
)) {
4264 if (userfaultfd_huge_pmd_wp(vmf
->vma
, orig_pmd
))
4265 return handle_userfault(vmf
, VM_UFFD_WP
);
4266 return do_huge_pmd_wp_page(vmf
, orig_pmd
);
4268 if (vmf
->vma
->vm_ops
->huge_fault
) {
4269 vm_fault_t ret
= vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
4271 if (!(ret
& VM_FAULT_FALLBACK
))
4275 /* COW or write-notify handled on pte level: split pmd. */
4276 __split_huge_pmd(vmf
->vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
4278 return VM_FAULT_FALLBACK
;
4281 static vm_fault_t
create_huge_pud(struct vm_fault
*vmf
)
4283 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4284 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4285 /* No support for anonymous transparent PUD pages yet */
4286 if (vma_is_anonymous(vmf
->vma
))
4288 if (vmf
->vma
->vm_ops
->huge_fault
) {
4289 vm_fault_t ret
= vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
4291 if (!(ret
& VM_FAULT_FALLBACK
))
4295 /* COW or write-notify not handled on PUD level: split pud.*/
4296 __split_huge_pud(vmf
->vma
, vmf
->pud
, vmf
->address
);
4297 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4298 return VM_FAULT_FALLBACK
;
4301 static vm_fault_t
wp_huge_pud(struct vm_fault
*vmf
, pud_t orig_pud
)
4303 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4304 /* No support for anonymous transparent PUD pages yet */
4305 if (vma_is_anonymous(vmf
->vma
))
4306 return VM_FAULT_FALLBACK
;
4307 if (vmf
->vma
->vm_ops
->huge_fault
)
4308 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
4309 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4310 return VM_FAULT_FALLBACK
;
4314 * These routines also need to handle stuff like marking pages dirty
4315 * and/or accessed for architectures that don't do it in hardware (most
4316 * RISC architectures). The early dirtying is also good on the i386.
4318 * There is also a hook called "update_mmu_cache()" that architectures
4319 * with external mmu caches can use to update those (ie the Sparc or
4320 * PowerPC hashed page tables that act as extended TLBs).
4322 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4323 * concurrent faults).
4325 * The mmap_lock may have been released depending on flags and our return value.
4326 * See filemap_fault() and __lock_page_or_retry().
4328 static vm_fault_t
handle_pte_fault(struct vm_fault
*vmf
)
4332 if (unlikely(pmd_none(*vmf
->pmd
))) {
4334 * Leave __pte_alloc() until later: because vm_ops->fault may
4335 * want to allocate huge page, and if we expose page table
4336 * for an instant, it will be difficult to retract from
4337 * concurrent faults and from rmap lookups.
4341 /* See comment in pte_alloc_one_map() */
4342 if (pmd_devmap_trans_unstable(vmf
->pmd
))
4345 * A regular pmd is established and it can't morph into a huge
4346 * pmd from under us anymore at this point because we hold the
4347 * mmap_lock read mode and khugepaged takes it in write mode.
4348 * So now it's safe to run pte_offset_map().
4350 vmf
->pte
= pte_offset_map(vmf
->pmd
, vmf
->address
);
4351 vmf
->orig_pte
= *vmf
->pte
;
4354 * some architectures can have larger ptes than wordsize,
4355 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4356 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4357 * accesses. The code below just needs a consistent view
4358 * for the ifs and we later double check anyway with the
4359 * ptl lock held. So here a barrier will do.
4362 if (pte_none(vmf
->orig_pte
)) {
4363 pte_unmap(vmf
->pte
);
4369 if (vma_is_anonymous(vmf
->vma
))
4370 return do_anonymous_page(vmf
);
4372 return do_fault(vmf
);
4375 if (!pte_present(vmf
->orig_pte
))
4376 return do_swap_page(vmf
);
4378 if (pte_protnone(vmf
->orig_pte
) && vma_is_accessible(vmf
->vma
))
4379 return do_numa_page(vmf
);
4381 vmf
->ptl
= pte_lockptr(vmf
->vma
->vm_mm
, vmf
->pmd
);
4382 spin_lock(vmf
->ptl
);
4383 entry
= vmf
->orig_pte
;
4384 if (unlikely(!pte_same(*vmf
->pte
, entry
))) {
4385 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
4388 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4389 if (!pte_write(entry
))
4390 return do_wp_page(vmf
);
4391 entry
= pte_mkdirty(entry
);
4393 entry
= pte_mkyoung(entry
);
4394 if (ptep_set_access_flags(vmf
->vma
, vmf
->address
, vmf
->pte
, entry
,
4395 vmf
->flags
& FAULT_FLAG_WRITE
)) {
4396 update_mmu_cache(vmf
->vma
, vmf
->address
, vmf
->pte
);
4398 /* Skip spurious TLB flush for retried page fault */
4399 if (vmf
->flags
& FAULT_FLAG_TRIED
)
4402 * This is needed only for protection faults but the arch code
4403 * is not yet telling us if this is a protection fault or not.
4404 * This still avoids useless tlb flushes for .text page faults
4407 if (vmf
->flags
& FAULT_FLAG_WRITE
)
4408 flush_tlb_fix_spurious_fault(vmf
->vma
, vmf
->address
);
4411 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4416 * By the time we get here, we already hold the mm semaphore
4418 * The mmap_lock may have been released depending on flags and our
4419 * return value. See filemap_fault() and __lock_page_or_retry().
4421 static vm_fault_t
__handle_mm_fault(struct vm_area_struct
*vma
,
4422 unsigned long address
, unsigned int flags
)
4424 struct vm_fault vmf
= {
4426 .address
= address
& PAGE_MASK
,
4428 .pgoff
= linear_page_index(vma
, address
),
4429 .gfp_mask
= __get_fault_gfp_mask(vma
),
4431 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
4432 struct mm_struct
*mm
= vma
->vm_mm
;
4437 pgd
= pgd_offset(mm
, address
);
4438 p4d
= p4d_alloc(mm
, pgd
, address
);
4440 return VM_FAULT_OOM
;
4442 vmf
.pud
= pud_alloc(mm
, p4d
, address
);
4444 return VM_FAULT_OOM
;
4446 if (pud_none(*vmf
.pud
) && __transparent_hugepage_enabled(vma
)) {
4447 ret
= create_huge_pud(&vmf
);
4448 if (!(ret
& VM_FAULT_FALLBACK
))
4451 pud_t orig_pud
= *vmf
.pud
;
4454 if (pud_trans_huge(orig_pud
) || pud_devmap(orig_pud
)) {
4456 /* NUMA case for anonymous PUDs would go here */
4458 if (dirty
&& !pud_write(orig_pud
)) {
4459 ret
= wp_huge_pud(&vmf
, orig_pud
);
4460 if (!(ret
& VM_FAULT_FALLBACK
))
4463 huge_pud_set_accessed(&vmf
, orig_pud
);
4469 vmf
.pmd
= pmd_alloc(mm
, vmf
.pud
, address
);
4471 return VM_FAULT_OOM
;
4473 /* Huge pud page fault raced with pmd_alloc? */
4474 if (pud_trans_unstable(vmf
.pud
))
4477 if (pmd_none(*vmf
.pmd
) && __transparent_hugepage_enabled(vma
)) {
4478 ret
= create_huge_pmd(&vmf
);
4479 if (!(ret
& VM_FAULT_FALLBACK
))
4482 pmd_t orig_pmd
= *vmf
.pmd
;
4485 if (unlikely(is_swap_pmd(orig_pmd
))) {
4486 VM_BUG_ON(thp_migration_supported() &&
4487 !is_pmd_migration_entry(orig_pmd
));
4488 if (is_pmd_migration_entry(orig_pmd
))
4489 pmd_migration_entry_wait(mm
, vmf
.pmd
);
4492 if (pmd_trans_huge(orig_pmd
) || pmd_devmap(orig_pmd
)) {
4493 if (pmd_protnone(orig_pmd
) && vma_is_accessible(vma
))
4494 return do_huge_pmd_numa_page(&vmf
, orig_pmd
);
4496 if (dirty
&& !pmd_write(orig_pmd
)) {
4497 ret
= wp_huge_pmd(&vmf
, orig_pmd
);
4498 if (!(ret
& VM_FAULT_FALLBACK
))
4501 huge_pmd_set_accessed(&vmf
, orig_pmd
);
4507 return handle_pte_fault(&vmf
);
4511 * mm_account_fault - Do page fault accountings
4513 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4514 * of perf event counters, but we'll still do the per-task accounting to
4515 * the task who triggered this page fault.
4516 * @address: the faulted address.
4517 * @flags: the fault flags.
4518 * @ret: the fault retcode.
4520 * This will take care of most of the page fault accountings. Meanwhile, it
4521 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4522 * updates. However note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4523 * still be in per-arch page fault handlers at the entry of page fault.
4525 static inline void mm_account_fault(struct pt_regs
*regs
,
4526 unsigned long address
, unsigned int flags
,
4532 * We don't do accounting for some specific faults:
4534 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4535 * includes arch_vma_access_permitted() failing before reaching here.
4536 * So this is not a "this many hardware page faults" counter. We
4537 * should use the hw profiling for that.
4539 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4540 * once they're completed.
4542 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_RETRY
))
4546 * We define the fault as a major fault when the final successful fault
4547 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4548 * handle it immediately previously).
4550 major
= (ret
& VM_FAULT_MAJOR
) || (flags
& FAULT_FLAG_TRIED
);
4558 * If the fault is done for GUP, regs will be NULL. We only do the
4559 * accounting for the per thread fault counters who triggered the
4560 * fault, and we skip the perf event updates.
4566 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ
, 1, regs
, address
);
4568 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN
, 1, regs
, address
);
4572 * By the time we get here, we already hold the mm semaphore
4574 * The mmap_lock may have been released depending on flags and our
4575 * return value. See filemap_fault() and __lock_page_or_retry().
4577 vm_fault_t
handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
4578 unsigned int flags
, struct pt_regs
*regs
)
4582 __set_current_state(TASK_RUNNING
);
4584 count_vm_event(PGFAULT
);
4585 count_memcg_event_mm(vma
->vm_mm
, PGFAULT
);
4587 /* do counter updates before entering really critical section. */
4588 check_sync_rss_stat(current
);
4590 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
4591 flags
& FAULT_FLAG_INSTRUCTION
,
4592 flags
& FAULT_FLAG_REMOTE
))
4593 return VM_FAULT_SIGSEGV
;
4596 * Enable the memcg OOM handling for faults triggered in user
4597 * space. Kernel faults are handled more gracefully.
4599 if (flags
& FAULT_FLAG_USER
)
4600 mem_cgroup_enter_user_fault();
4602 if (unlikely(is_vm_hugetlb_page(vma
)))
4603 ret
= hugetlb_fault(vma
->vm_mm
, vma
, address
, flags
);
4605 ret
= __handle_mm_fault(vma
, address
, flags
);
4607 if (flags
& FAULT_FLAG_USER
) {
4608 mem_cgroup_exit_user_fault();
4610 * The task may have entered a memcg OOM situation but
4611 * if the allocation error was handled gracefully (no
4612 * VM_FAULT_OOM), there is no need to kill anything.
4613 * Just clean up the OOM state peacefully.
4615 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
4616 mem_cgroup_oom_synchronize(false);
4619 mm_account_fault(regs
, address
, flags
, ret
);
4623 EXPORT_SYMBOL_GPL(handle_mm_fault
);
4625 #ifndef __PAGETABLE_P4D_FOLDED
4627 * Allocate p4d page table.
4628 * We've already handled the fast-path in-line.
4630 int __p4d_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
4632 p4d_t
*new = p4d_alloc_one(mm
, address
);
4636 smp_wmb(); /* See comment in __pte_alloc */
4638 spin_lock(&mm
->page_table_lock
);
4639 if (pgd_present(*pgd
)) /* Another has populated it */
4642 pgd_populate(mm
, pgd
, new);
4643 spin_unlock(&mm
->page_table_lock
);
4646 #endif /* __PAGETABLE_P4D_FOLDED */
4648 #ifndef __PAGETABLE_PUD_FOLDED
4650 * Allocate page upper directory.
4651 * We've already handled the fast-path in-line.
4653 int __pud_alloc(struct mm_struct
*mm
, p4d_t
*p4d
, unsigned long address
)
4655 pud_t
*new = pud_alloc_one(mm
, address
);
4659 smp_wmb(); /* See comment in __pte_alloc */
4661 spin_lock(&mm
->page_table_lock
);
4662 if (!p4d_present(*p4d
)) {
4664 p4d_populate(mm
, p4d
, new);
4665 } else /* Another has populated it */
4667 spin_unlock(&mm
->page_table_lock
);
4670 #endif /* __PAGETABLE_PUD_FOLDED */
4672 #ifndef __PAGETABLE_PMD_FOLDED
4674 * Allocate page middle directory.
4675 * We've already handled the fast-path in-line.
4677 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
4680 pmd_t
*new = pmd_alloc_one(mm
, address
);
4684 smp_wmb(); /* See comment in __pte_alloc */
4686 ptl
= pud_lock(mm
, pud
);
4687 if (!pud_present(*pud
)) {
4689 pud_populate(mm
, pud
, new);
4690 } else /* Another has populated it */
4695 #endif /* __PAGETABLE_PMD_FOLDED */
4697 static int __follow_pte_pmd(struct mm_struct
*mm
, unsigned long address
,
4698 struct mmu_notifier_range
*range
,
4699 pte_t
**ptepp
, pmd_t
**pmdpp
, spinlock_t
**ptlp
)
4707 pgd
= pgd_offset(mm
, address
);
4708 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
4711 p4d
= p4d_offset(pgd
, address
);
4712 if (p4d_none(*p4d
) || unlikely(p4d_bad(*p4d
)))
4715 pud
= pud_offset(p4d
, address
);
4716 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
4719 pmd
= pmd_offset(pud
, address
);
4720 VM_BUG_ON(pmd_trans_huge(*pmd
));
4722 if (pmd_huge(*pmd
)) {
4727 mmu_notifier_range_init(range
, MMU_NOTIFY_CLEAR
, 0,
4728 NULL
, mm
, address
& PMD_MASK
,
4729 (address
& PMD_MASK
) + PMD_SIZE
);
4730 mmu_notifier_invalidate_range_start(range
);
4732 *ptlp
= pmd_lock(mm
, pmd
);
4733 if (pmd_huge(*pmd
)) {
4739 mmu_notifier_invalidate_range_end(range
);
4742 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
4746 mmu_notifier_range_init(range
, MMU_NOTIFY_CLEAR
, 0, NULL
, mm
,
4747 address
& PAGE_MASK
,
4748 (address
& PAGE_MASK
) + PAGE_SIZE
);
4749 mmu_notifier_invalidate_range_start(range
);
4751 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
4752 if (!pte_present(*ptep
))
4757 pte_unmap_unlock(ptep
, *ptlp
);
4759 mmu_notifier_invalidate_range_end(range
);
4764 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
4765 pte_t
**ptepp
, spinlock_t
**ptlp
)
4769 /* (void) is needed to make gcc happy */
4770 (void) __cond_lock(*ptlp
,
4771 !(res
= __follow_pte_pmd(mm
, address
, NULL
,
4772 ptepp
, NULL
, ptlp
)));
4776 int follow_pte_pmd(struct mm_struct
*mm
, unsigned long address
,
4777 struct mmu_notifier_range
*range
,
4778 pte_t
**ptepp
, pmd_t
**pmdpp
, spinlock_t
**ptlp
)
4782 /* (void) is needed to make gcc happy */
4783 (void) __cond_lock(*ptlp
,
4784 !(res
= __follow_pte_pmd(mm
, address
, range
,
4785 ptepp
, pmdpp
, ptlp
)));
4788 EXPORT_SYMBOL(follow_pte_pmd
);
4791 * follow_pfn - look up PFN at a user virtual address
4792 * @vma: memory mapping
4793 * @address: user virtual address
4794 * @pfn: location to store found PFN
4796 * Only IO mappings and raw PFN mappings are allowed.
4798 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4800 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
4807 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4810 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
4813 *pfn
= pte_pfn(*ptep
);
4814 pte_unmap_unlock(ptep
, ptl
);
4817 EXPORT_SYMBOL(follow_pfn
);
4819 #ifdef CONFIG_HAVE_IOREMAP_PROT
4820 int follow_phys(struct vm_area_struct
*vma
,
4821 unsigned long address
, unsigned int flags
,
4822 unsigned long *prot
, resource_size_t
*phys
)
4828 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4831 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
4835 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
4838 *prot
= pgprot_val(pte_pgprot(pte
));
4839 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
4843 pte_unmap_unlock(ptep
, ptl
);
4848 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
4849 void *buf
, int len
, int write
)
4851 resource_size_t phys_addr
;
4852 unsigned long prot
= 0;
4853 void __iomem
*maddr
;
4854 int offset
= addr
& (PAGE_SIZE
-1);
4856 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
4859 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
4864 memcpy_toio(maddr
+ offset
, buf
, len
);
4866 memcpy_fromio(buf
, maddr
+ offset
, len
);
4871 EXPORT_SYMBOL_GPL(generic_access_phys
);
4875 * Access another process' address space as given in mm. If non-NULL, use the
4876 * given task for page fault accounting.
4878 int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
4879 unsigned long addr
, void *buf
, int len
, unsigned int gup_flags
)
4881 struct vm_area_struct
*vma
;
4882 void *old_buf
= buf
;
4883 int write
= gup_flags
& FOLL_WRITE
;
4885 if (mmap_read_lock_killable(mm
))
4888 /* ignore errors, just check how much was successfully transferred */
4890 int bytes
, ret
, offset
;
4892 struct page
*page
= NULL
;
4894 ret
= get_user_pages_remote(mm
, addr
, 1,
4895 gup_flags
, &page
, &vma
, NULL
);
4897 #ifndef CONFIG_HAVE_IOREMAP_PROT
4901 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4902 * we can access using slightly different code.
4904 vma
= find_vma(mm
, addr
);
4905 if (!vma
|| vma
->vm_start
> addr
)
4907 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
4908 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
4916 offset
= addr
& (PAGE_SIZE
-1);
4917 if (bytes
> PAGE_SIZE
-offset
)
4918 bytes
= PAGE_SIZE
-offset
;
4922 copy_to_user_page(vma
, page
, addr
,
4923 maddr
+ offset
, buf
, bytes
);
4924 set_page_dirty_lock(page
);
4926 copy_from_user_page(vma
, page
, addr
,
4927 buf
, maddr
+ offset
, bytes
);
4936 mmap_read_unlock(mm
);
4938 return buf
- old_buf
;
4942 * access_remote_vm - access another process' address space
4943 * @mm: the mm_struct of the target address space
4944 * @addr: start address to access
4945 * @buf: source or destination buffer
4946 * @len: number of bytes to transfer
4947 * @gup_flags: flags modifying lookup behaviour
4949 * The caller must hold a reference on @mm.
4951 * Return: number of bytes copied from source to destination.
4953 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
4954 void *buf
, int len
, unsigned int gup_flags
)
4956 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, gup_flags
);
4960 * Access another process' address space.
4961 * Source/target buffer must be kernel space,
4962 * Do not walk the page table directly, use get_user_pages
4964 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
4965 void *buf
, int len
, unsigned int gup_flags
)
4967 struct mm_struct
*mm
;
4970 mm
= get_task_mm(tsk
);
4974 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, gup_flags
);
4980 EXPORT_SYMBOL_GPL(access_process_vm
);
4983 * Print the name of a VMA.
4985 void print_vma_addr(char *prefix
, unsigned long ip
)
4987 struct mm_struct
*mm
= current
->mm
;
4988 struct vm_area_struct
*vma
;
4991 * we might be running from an atomic context so we cannot sleep
4993 if (!mmap_read_trylock(mm
))
4996 vma
= find_vma(mm
, ip
);
4997 if (vma
&& vma
->vm_file
) {
4998 struct file
*f
= vma
->vm_file
;
4999 char *buf
= (char *)__get_free_page(GFP_NOWAIT
);
5003 p
= file_path(f
, buf
, PAGE_SIZE
);
5006 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
5008 vma
->vm_end
- vma
->vm_start
);
5009 free_page((unsigned long)buf
);
5012 mmap_read_unlock(mm
);
5015 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5016 void __might_fault(const char *file
, int line
)
5019 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5020 * holding the mmap_lock, this is safe because kernel memory doesn't
5021 * get paged out, therefore we'll never actually fault, and the
5022 * below annotations will generate false positives.
5024 if (uaccess_kernel())
5026 if (pagefault_disabled())
5028 __might_sleep(file
, line
, 0);
5029 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5031 might_lock_read(¤t
->mm
->mmap_lock
);
5034 EXPORT_SYMBOL(__might_fault
);
5037 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5039 * Process all subpages of the specified huge page with the specified
5040 * operation. The target subpage will be processed last to keep its
5043 static inline void process_huge_page(
5044 unsigned long addr_hint
, unsigned int pages_per_huge_page
,
5045 void (*process_subpage
)(unsigned long addr
, int idx
, void *arg
),
5049 unsigned long addr
= addr_hint
&
5050 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5052 /* Process target subpage last to keep its cache lines hot */
5054 n
= (addr_hint
- addr
) / PAGE_SIZE
;
5055 if (2 * n
<= pages_per_huge_page
) {
5056 /* If target subpage in first half of huge page */
5059 /* Process subpages at the end of huge page */
5060 for (i
= pages_per_huge_page
- 1; i
>= 2 * n
; i
--) {
5062 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
5065 /* If target subpage in second half of huge page */
5066 base
= pages_per_huge_page
- 2 * (pages_per_huge_page
- n
);
5067 l
= pages_per_huge_page
- n
;
5068 /* Process subpages at the begin of huge page */
5069 for (i
= 0; i
< base
; i
++) {
5071 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
5075 * Process remaining subpages in left-right-left-right pattern
5076 * towards the target subpage
5078 for (i
= 0; i
< l
; i
++) {
5079 int left_idx
= base
+ i
;
5080 int right_idx
= base
+ 2 * l
- 1 - i
;
5083 process_subpage(addr
+ left_idx
* PAGE_SIZE
, left_idx
, arg
);
5085 process_subpage(addr
+ right_idx
* PAGE_SIZE
, right_idx
, arg
);
5089 static void clear_gigantic_page(struct page
*page
,
5091 unsigned int pages_per_huge_page
)
5094 struct page
*p
= page
;
5097 for (i
= 0; i
< pages_per_huge_page
;
5098 i
++, p
= mem_map_next(p
, page
, i
)) {
5100 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
5104 static void clear_subpage(unsigned long addr
, int idx
, void *arg
)
5106 struct page
*page
= arg
;
5108 clear_user_highpage(page
+ idx
, addr
);
5111 void clear_huge_page(struct page
*page
,
5112 unsigned long addr_hint
, unsigned int pages_per_huge_page
)
5114 unsigned long addr
= addr_hint
&
5115 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5117 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
5118 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
5122 process_huge_page(addr_hint
, pages_per_huge_page
, clear_subpage
, page
);
5125 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
5127 struct vm_area_struct
*vma
,
5128 unsigned int pages_per_huge_page
)
5131 struct page
*dst_base
= dst
;
5132 struct page
*src_base
= src
;
5134 for (i
= 0; i
< pages_per_huge_page
; ) {
5136 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
5139 dst
= mem_map_next(dst
, dst_base
, i
);
5140 src
= mem_map_next(src
, src_base
, i
);
5144 struct copy_subpage_arg
{
5147 struct vm_area_struct
*vma
;
5150 static void copy_subpage(unsigned long addr
, int idx
, void *arg
)
5152 struct copy_subpage_arg
*copy_arg
= arg
;
5154 copy_user_highpage(copy_arg
->dst
+ idx
, copy_arg
->src
+ idx
,
5155 addr
, copy_arg
->vma
);
5158 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
5159 unsigned long addr_hint
, struct vm_area_struct
*vma
,
5160 unsigned int pages_per_huge_page
)
5162 unsigned long addr
= addr_hint
&
5163 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5164 struct copy_subpage_arg arg
= {
5170 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
5171 copy_user_gigantic_page(dst
, src
, addr
, vma
,
5172 pages_per_huge_page
);
5176 process_huge_page(addr_hint
, pages_per_huge_page
, copy_subpage
, &arg
);
5179 long copy_huge_page_from_user(struct page
*dst_page
,
5180 const void __user
*usr_src
,
5181 unsigned int pages_per_huge_page
,
5182 bool allow_pagefault
)
5184 void *src
= (void *)usr_src
;
5186 unsigned long i
, rc
= 0;
5187 unsigned long ret_val
= pages_per_huge_page
* PAGE_SIZE
;
5189 for (i
= 0; i
< pages_per_huge_page
; i
++) {
5190 if (allow_pagefault
)
5191 page_kaddr
= kmap(dst_page
+ i
);
5193 page_kaddr
= kmap_atomic(dst_page
+ i
);
5194 rc
= copy_from_user(page_kaddr
,
5195 (const void __user
*)(src
+ i
* PAGE_SIZE
),
5197 if (allow_pagefault
)
5198 kunmap(dst_page
+ i
);
5200 kunmap_atomic(page_kaddr
);
5202 ret_val
-= (PAGE_SIZE
- rc
);
5210 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5212 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5214 static struct kmem_cache
*page_ptl_cachep
;
5216 void __init
ptlock_cache_init(void)
5218 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
5222 bool ptlock_alloc(struct page
*page
)
5226 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
5233 void ptlock_free(struct page
*page
)
5235 kmem_cache_free(page_ptl_cachep
, page
->ptl
);