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 mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
798 pte_t
*dst_pte
, pte_t
*src_pte
,
799 struct vm_area_struct
*vma
, struct vm_area_struct
*new,
800 unsigned long addr
, int *rss
, struct page
**prealloc
,
801 pte_t pte
, struct page
*page
)
803 struct page
*new_page
;
805 if (!is_cow_mapping(vma
->vm_flags
))
809 * What we want to do is to check whether this page may
810 * have been pinned by the parent process. If so,
811 * instead of wrprotect the pte on both sides, we copy
812 * the page immediately so that we'll always guarantee
813 * the pinned page won't be randomly replaced in the
816 * The page pinning checks are just "has this mm ever
817 * seen pinning", along with the (inexact) check of
818 * the page count. That might give false positives for
819 * for pinning, but it will work correctly.
821 if (likely(!atomic_read(&src_mm
->has_pinned
)))
823 if (likely(!page_maybe_dma_pinned(page
)))
826 new_page
= *prealloc
;
831 * We have a prealloc page, all good! Take it
832 * over and copy the page & arm it.
835 copy_user_highpage(new_page
, page
, addr
, vma
);
836 __SetPageUptodate(new_page
);
837 page_add_new_anon_rmap(new_page
, new, addr
, false);
838 lru_cache_add_inactive_or_unevictable(new_page
, new);
839 rss
[mm_counter(new_page
)]++;
841 /* All done, just insert the new page copy in the child */
842 pte
= mk_pte(new_page
, new->vm_page_prot
);
843 pte
= maybe_mkwrite(pte_mkdirty(pte
), new);
844 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
849 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
850 * is required to copy this pte.
853 copy_present_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
854 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
855 struct vm_area_struct
*new,
856 unsigned long addr
, int *rss
, struct page
**prealloc
)
858 unsigned long vm_flags
= vma
->vm_flags
;
859 pte_t pte
= *src_pte
;
862 page
= vm_normal_page(vma
, addr
, pte
);
866 retval
= copy_present_page(dst_mm
, src_mm
,
875 page_dup_rmap(page
, false);
876 rss
[mm_counter(page
)]++;
880 * If it's a COW mapping, write protect it both
881 * in the parent and the child
883 if (is_cow_mapping(vm_flags
) && pte_write(pte
)) {
884 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
885 pte
= pte_wrprotect(pte
);
889 * If it's a shared mapping, mark it clean in
892 if (vm_flags
& VM_SHARED
)
893 pte
= pte_mkclean(pte
);
894 pte
= pte_mkold(pte
);
897 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
898 * does not have the VM_UFFD_WP, which means that the uffd
899 * fork event is not enabled.
901 if (!(vm_flags
& VM_UFFD_WP
))
902 pte
= pte_clear_uffd_wp(pte
);
904 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
908 static inline struct page
*
909 page_copy_prealloc(struct mm_struct
*src_mm
, struct vm_area_struct
*vma
,
912 struct page
*new_page
;
914 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, addr
);
918 if (mem_cgroup_charge(new_page
, src_mm
, GFP_KERNEL
)) {
922 cgroup_throttle_swaprate(new_page
, GFP_KERNEL
);
927 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
928 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
929 struct vm_area_struct
*new,
930 unsigned long addr
, unsigned long end
)
932 pte_t
*orig_src_pte
, *orig_dst_pte
;
933 pte_t
*src_pte
, *dst_pte
;
934 spinlock_t
*src_ptl
, *dst_ptl
;
935 int progress
, ret
= 0;
936 int rss
[NR_MM_COUNTERS
];
937 swp_entry_t entry
= (swp_entry_t
){0};
938 struct page
*prealloc
= NULL
;
944 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
949 src_pte
= pte_offset_map(src_pmd
, addr
);
950 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
951 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
952 orig_src_pte
= src_pte
;
953 orig_dst_pte
= dst_pte
;
954 arch_enter_lazy_mmu_mode();
958 * We are holding two locks at this point - either of them
959 * could generate latencies in another task on another CPU.
961 if (progress
>= 32) {
963 if (need_resched() ||
964 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
967 if (pte_none(*src_pte
)) {
971 if (unlikely(!pte_present(*src_pte
))) {
972 entry
.val
= copy_nonpresent_pte(dst_mm
, src_mm
,
980 /* copy_present_pte() will clear `*prealloc' if consumed */
981 ret
= copy_present_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
982 vma
, new, addr
, rss
, &prealloc
);
984 * If we need a pre-allocated page for this pte, drop the
985 * locks, allocate, and try again.
987 if (unlikely(ret
== -EAGAIN
))
989 if (unlikely(prealloc
)) {
991 * pre-alloc page cannot be reused by next time so as
992 * to strictly follow mempolicy (e.g., alloc_page_vma()
993 * will allocate page according to address). This
994 * could only happen if one pinned pte changed.
1000 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1002 arch_leave_lazy_mmu_mode();
1003 spin_unlock(src_ptl
);
1004 pte_unmap(orig_src_pte
);
1005 add_mm_rss_vec(dst_mm
, rss
);
1006 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
1010 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0) {
1016 WARN_ON_ONCE(ret
!= -EAGAIN
);
1017 prealloc
= page_copy_prealloc(src_mm
, vma
, addr
);
1020 /* We've captured and resolved the error. Reset, try again. */
1026 if (unlikely(prealloc
))
1031 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1032 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
1033 struct vm_area_struct
*new,
1034 unsigned long addr
, unsigned long end
)
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
, vma
);
1049 err
= copy_huge_pmd(dst_mm
, src_mm
,
1050 dst_pmd
, src_pmd
, addr
, vma
);
1057 if (pmd_none_or_clear_bad(src_pmd
))
1059 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
1060 vma
, new, addr
, next
))
1062 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1066 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1067 p4d_t
*dst_p4d
, p4d_t
*src_p4d
, struct vm_area_struct
*vma
,
1068 struct vm_area_struct
*new,
1069 unsigned long addr
, unsigned long end
)
1071 pud_t
*src_pud
, *dst_pud
;
1074 dst_pud
= pud_alloc(dst_mm
, dst_p4d
, addr
);
1077 src_pud
= pud_offset(src_p4d
, addr
);
1079 next
= pud_addr_end(addr
, end
);
1080 if (pud_trans_huge(*src_pud
) || pud_devmap(*src_pud
)) {
1083 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PUD_SIZE
, vma
);
1084 err
= copy_huge_pud(dst_mm
, src_mm
,
1085 dst_pud
, src_pud
, addr
, vma
);
1092 if (pud_none_or_clear_bad(src_pud
))
1094 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
1095 vma
, new, addr
, next
))
1097 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1101 static inline int copy_p4d_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1102 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
1103 struct vm_area_struct
*new,
1104 unsigned long addr
, unsigned long end
)
1106 p4d_t
*src_p4d
, *dst_p4d
;
1109 dst_p4d
= p4d_alloc(dst_mm
, dst_pgd
, addr
);
1112 src_p4d
= p4d_offset(src_pgd
, addr
);
1114 next
= p4d_addr_end(addr
, end
);
1115 if (p4d_none_or_clear_bad(src_p4d
))
1117 if (copy_pud_range(dst_mm
, src_mm
, dst_p4d
, src_p4d
,
1118 vma
, new, addr
, next
))
1120 } while (dst_p4d
++, src_p4d
++, addr
= next
, addr
!= end
);
1124 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1125 struct vm_area_struct
*vma
, struct vm_area_struct
*new)
1127 pgd_t
*src_pgd
, *dst_pgd
;
1129 unsigned long addr
= vma
->vm_start
;
1130 unsigned long end
= vma
->vm_end
;
1131 struct mmu_notifier_range range
;
1136 * Don't copy ptes where a page fault will fill them correctly.
1137 * Fork becomes much lighter when there are big shared or private
1138 * readonly mappings. The tradeoff is that copy_page_range is more
1139 * efficient than faulting.
1141 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_PFNMAP
| VM_MIXEDMAP
)) &&
1145 if (is_vm_hugetlb_page(vma
))
1146 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1148 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
1150 * We do not free on error cases below as remove_vma
1151 * gets called on error from higher level routine
1153 ret
= track_pfn_copy(vma
);
1159 * We need to invalidate the secondary MMU mappings only when
1160 * there could be a permission downgrade on the ptes of the
1161 * parent mm. And a permission downgrade will only happen if
1162 * is_cow_mapping() returns true.
1164 is_cow
= is_cow_mapping(vma
->vm_flags
);
1167 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_PAGE
,
1168 0, vma
, src_mm
, addr
, end
);
1169 mmu_notifier_invalidate_range_start(&range
);
1173 dst_pgd
= pgd_offset(dst_mm
, addr
);
1174 src_pgd
= pgd_offset(src_mm
, addr
);
1176 next
= pgd_addr_end(addr
, end
);
1177 if (pgd_none_or_clear_bad(src_pgd
))
1179 if (unlikely(copy_p4d_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1180 vma
, new, addr
, next
))) {
1184 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1187 mmu_notifier_invalidate_range_end(&range
);
1191 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1192 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1193 unsigned long addr
, unsigned long end
,
1194 struct zap_details
*details
)
1196 struct mm_struct
*mm
= tlb
->mm
;
1197 int force_flush
= 0;
1198 int rss
[NR_MM_COUNTERS
];
1204 tlb_change_page_size(tlb
, PAGE_SIZE
);
1207 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1209 flush_tlb_batched_pending(mm
);
1210 arch_enter_lazy_mmu_mode();
1213 if (pte_none(ptent
))
1219 if (pte_present(ptent
)) {
1222 page
= vm_normal_page(vma
, addr
, ptent
);
1223 if (unlikely(details
) && page
) {
1225 * unmap_shared_mapping_pages() wants to
1226 * invalidate cache without truncating:
1227 * unmap shared but keep private pages.
1229 if (details
->check_mapping
&&
1230 details
->check_mapping
!= page_rmapping(page
))
1233 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1235 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1236 if (unlikely(!page
))
1239 if (!PageAnon(page
)) {
1240 if (pte_dirty(ptent
)) {
1242 set_page_dirty(page
);
1244 if (pte_young(ptent
) &&
1245 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1246 mark_page_accessed(page
);
1248 rss
[mm_counter(page
)]--;
1249 page_remove_rmap(page
, false);
1250 if (unlikely(page_mapcount(page
) < 0))
1251 print_bad_pte(vma
, addr
, ptent
, page
);
1252 if (unlikely(__tlb_remove_page(tlb
, page
))) {
1260 entry
= pte_to_swp_entry(ptent
);
1261 if (is_device_private_entry(entry
)) {
1262 struct page
*page
= device_private_entry_to_page(entry
);
1264 if (unlikely(details
&& details
->check_mapping
)) {
1266 * unmap_shared_mapping_pages() wants to
1267 * invalidate cache without truncating:
1268 * unmap shared but keep private pages.
1270 if (details
->check_mapping
!=
1271 page_rmapping(page
))
1275 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1276 rss
[mm_counter(page
)]--;
1277 page_remove_rmap(page
, false);
1282 /* If details->check_mapping, we leave swap entries. */
1283 if (unlikely(details
))
1286 if (!non_swap_entry(entry
))
1288 else if (is_migration_entry(entry
)) {
1291 page
= migration_entry_to_page(entry
);
1292 rss
[mm_counter(page
)]--;
1294 if (unlikely(!free_swap_and_cache(entry
)))
1295 print_bad_pte(vma
, addr
, ptent
, NULL
);
1296 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1297 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1299 add_mm_rss_vec(mm
, rss
);
1300 arch_leave_lazy_mmu_mode();
1302 /* Do the actual TLB flush before dropping ptl */
1304 tlb_flush_mmu_tlbonly(tlb
);
1305 pte_unmap_unlock(start_pte
, ptl
);
1308 * If we forced a TLB flush (either due to running out of
1309 * batch buffers or because we needed to flush dirty TLB
1310 * entries before releasing the ptl), free the batched
1311 * memory too. Restart if we didn't do everything.
1326 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1327 struct vm_area_struct
*vma
, pud_t
*pud
,
1328 unsigned long addr
, unsigned long end
,
1329 struct zap_details
*details
)
1334 pmd
= pmd_offset(pud
, addr
);
1336 next
= pmd_addr_end(addr
, end
);
1337 if (is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1338 if (next
- addr
!= HPAGE_PMD_SIZE
)
1339 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
1340 else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1345 * Here there can be other concurrent MADV_DONTNEED or
1346 * trans huge page faults running, and if the pmd is
1347 * none or trans huge it can change under us. This is
1348 * because MADV_DONTNEED holds the mmap_lock in read
1351 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1353 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1356 } while (pmd
++, addr
= next
, addr
!= end
);
1361 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1362 struct vm_area_struct
*vma
, p4d_t
*p4d
,
1363 unsigned long addr
, unsigned long end
,
1364 struct zap_details
*details
)
1369 pud
= pud_offset(p4d
, addr
);
1371 next
= pud_addr_end(addr
, end
);
1372 if (pud_trans_huge(*pud
) || pud_devmap(*pud
)) {
1373 if (next
- addr
!= HPAGE_PUD_SIZE
) {
1374 mmap_assert_locked(tlb
->mm
);
1375 split_huge_pud(vma
, pud
, addr
);
1376 } else if (zap_huge_pud(tlb
, vma
, pud
, addr
))
1380 if (pud_none_or_clear_bad(pud
))
1382 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1385 } while (pud
++, addr
= next
, addr
!= end
);
1390 static inline unsigned long zap_p4d_range(struct mmu_gather
*tlb
,
1391 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1392 unsigned long addr
, unsigned long end
,
1393 struct zap_details
*details
)
1398 p4d
= p4d_offset(pgd
, addr
);
1400 next
= p4d_addr_end(addr
, end
);
1401 if (p4d_none_or_clear_bad(p4d
))
1403 next
= zap_pud_range(tlb
, vma
, p4d
, addr
, next
, details
);
1404 } while (p4d
++, addr
= next
, addr
!= end
);
1409 void unmap_page_range(struct mmu_gather
*tlb
,
1410 struct vm_area_struct
*vma
,
1411 unsigned long addr
, unsigned long end
,
1412 struct zap_details
*details
)
1417 BUG_ON(addr
>= end
);
1418 tlb_start_vma(tlb
, vma
);
1419 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1421 next
= pgd_addr_end(addr
, end
);
1422 if (pgd_none_or_clear_bad(pgd
))
1424 next
= zap_p4d_range(tlb
, vma
, pgd
, addr
, next
, details
);
1425 } while (pgd
++, addr
= next
, addr
!= end
);
1426 tlb_end_vma(tlb
, vma
);
1430 static void unmap_single_vma(struct mmu_gather
*tlb
,
1431 struct vm_area_struct
*vma
, unsigned long start_addr
,
1432 unsigned long end_addr
,
1433 struct zap_details
*details
)
1435 unsigned long start
= max(vma
->vm_start
, start_addr
);
1438 if (start
>= vma
->vm_end
)
1440 end
= min(vma
->vm_end
, end_addr
);
1441 if (end
<= vma
->vm_start
)
1445 uprobe_munmap(vma
, start
, end
);
1447 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1448 untrack_pfn(vma
, 0, 0);
1451 if (unlikely(is_vm_hugetlb_page(vma
))) {
1453 * It is undesirable to test vma->vm_file as it
1454 * should be non-null for valid hugetlb area.
1455 * However, vm_file will be NULL in the error
1456 * cleanup path of mmap_region. When
1457 * hugetlbfs ->mmap method fails,
1458 * mmap_region() nullifies vma->vm_file
1459 * before calling this function to clean up.
1460 * Since no pte has actually been setup, it is
1461 * safe to do nothing in this case.
1464 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
1465 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1466 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
1469 unmap_page_range(tlb
, vma
, start
, end
, details
);
1474 * unmap_vmas - unmap a range of memory covered by a list of vma's
1475 * @tlb: address of the caller's struct mmu_gather
1476 * @vma: the starting vma
1477 * @start_addr: virtual address at which to start unmapping
1478 * @end_addr: virtual address at which to end unmapping
1480 * Unmap all pages in the vma list.
1482 * Only addresses between `start' and `end' will be unmapped.
1484 * The VMA list must be sorted in ascending virtual address order.
1486 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1487 * range after unmap_vmas() returns. So the only responsibility here is to
1488 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1489 * drops the lock and schedules.
1491 void unmap_vmas(struct mmu_gather
*tlb
,
1492 struct vm_area_struct
*vma
, unsigned long start_addr
,
1493 unsigned long end_addr
)
1495 struct mmu_notifier_range range
;
1497 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
, vma
->vm_mm
,
1498 start_addr
, end_addr
);
1499 mmu_notifier_invalidate_range_start(&range
);
1500 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1501 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1502 mmu_notifier_invalidate_range_end(&range
);
1506 * zap_page_range - remove user pages in a given range
1507 * @vma: vm_area_struct holding the applicable pages
1508 * @start: starting address of pages to zap
1509 * @size: number of bytes to zap
1511 * Caller must protect the VMA list
1513 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1516 struct mmu_notifier_range range
;
1517 struct mmu_gather tlb
;
1520 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1521 start
, start
+ size
);
1522 tlb_gather_mmu(&tlb
, vma
->vm_mm
, start
, range
.end
);
1523 update_hiwater_rss(vma
->vm_mm
);
1524 mmu_notifier_invalidate_range_start(&range
);
1525 for ( ; vma
&& vma
->vm_start
< range
.end
; vma
= vma
->vm_next
)
1526 unmap_single_vma(&tlb
, vma
, start
, range
.end
, NULL
);
1527 mmu_notifier_invalidate_range_end(&range
);
1528 tlb_finish_mmu(&tlb
, start
, range
.end
);
1532 * zap_page_range_single - remove user pages in a given range
1533 * @vma: vm_area_struct holding the applicable pages
1534 * @address: starting address of pages to zap
1535 * @size: number of bytes to zap
1536 * @details: details of shared cache invalidation
1538 * The range must fit into one VMA.
1540 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1541 unsigned long size
, struct zap_details
*details
)
1543 struct mmu_notifier_range range
;
1544 struct mmu_gather tlb
;
1547 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1548 address
, address
+ size
);
1549 tlb_gather_mmu(&tlb
, vma
->vm_mm
, address
, range
.end
);
1550 update_hiwater_rss(vma
->vm_mm
);
1551 mmu_notifier_invalidate_range_start(&range
);
1552 unmap_single_vma(&tlb
, vma
, address
, range
.end
, details
);
1553 mmu_notifier_invalidate_range_end(&range
);
1554 tlb_finish_mmu(&tlb
, address
, range
.end
);
1558 * zap_vma_ptes - remove ptes mapping the vma
1559 * @vma: vm_area_struct holding ptes to be zapped
1560 * @address: starting address of pages to zap
1561 * @size: number of bytes to zap
1563 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1565 * The entire address range must be fully contained within the vma.
1568 void zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1571 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1572 !(vma
->vm_flags
& VM_PFNMAP
))
1575 zap_page_range_single(vma
, address
, size
, NULL
);
1577 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1579 static pmd_t
*walk_to_pmd(struct mm_struct
*mm
, unsigned long addr
)
1586 pgd
= pgd_offset(mm
, addr
);
1587 p4d
= p4d_alloc(mm
, pgd
, addr
);
1590 pud
= pud_alloc(mm
, p4d
, addr
);
1593 pmd
= pmd_alloc(mm
, pud
, addr
);
1597 VM_BUG_ON(pmd_trans_huge(*pmd
));
1601 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1604 pmd_t
*pmd
= walk_to_pmd(mm
, addr
);
1608 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1611 static int validate_page_before_insert(struct page
*page
)
1613 if (PageAnon(page
) || PageSlab(page
) || page_has_type(page
))
1615 flush_dcache_page(page
);
1619 static int insert_page_into_pte_locked(struct mm_struct
*mm
, pte_t
*pte
,
1620 unsigned long addr
, struct page
*page
, pgprot_t prot
)
1622 if (!pte_none(*pte
))
1624 /* Ok, finally just insert the thing.. */
1626 inc_mm_counter_fast(mm
, mm_counter_file(page
));
1627 page_add_file_rmap(page
, false);
1628 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1633 * This is the old fallback for page remapping.
1635 * For historical reasons, it only allows reserved pages. Only
1636 * old drivers should use this, and they needed to mark their
1637 * pages reserved for the old functions anyway.
1639 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1640 struct page
*page
, pgprot_t prot
)
1642 struct mm_struct
*mm
= vma
->vm_mm
;
1647 retval
= validate_page_before_insert(page
);
1651 pte
= get_locked_pte(mm
, addr
, &ptl
);
1654 retval
= insert_page_into_pte_locked(mm
, pte
, addr
, page
, prot
);
1655 pte_unmap_unlock(pte
, ptl
);
1661 static int insert_page_in_batch_locked(struct mm_struct
*mm
, pte_t
*pte
,
1662 unsigned long addr
, struct page
*page
, pgprot_t prot
)
1666 if (!page_count(page
))
1668 err
= validate_page_before_insert(page
);
1671 return insert_page_into_pte_locked(mm
, pte
, addr
, page
, prot
);
1674 /* insert_pages() amortizes the cost of spinlock operations
1675 * when inserting pages in a loop. Arch *must* define pte_index.
1677 static int insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
1678 struct page
**pages
, unsigned long *num
, pgprot_t prot
)
1681 pte_t
*start_pte
, *pte
;
1682 spinlock_t
*pte_lock
;
1683 struct mm_struct
*const mm
= vma
->vm_mm
;
1684 unsigned long curr_page_idx
= 0;
1685 unsigned long remaining_pages_total
= *num
;
1686 unsigned long pages_to_write_in_pmd
;
1690 pmd
= walk_to_pmd(mm
, addr
);
1694 pages_to_write_in_pmd
= min_t(unsigned long,
1695 remaining_pages_total
, PTRS_PER_PTE
- pte_index(addr
));
1697 /* Allocate the PTE if necessary; takes PMD lock once only. */
1699 if (pte_alloc(mm
, pmd
))
1702 while (pages_to_write_in_pmd
) {
1704 const int batch_size
= min_t(int, pages_to_write_in_pmd
, 8);
1706 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &pte_lock
);
1707 for (pte
= start_pte
; pte_idx
< batch_size
; ++pte
, ++pte_idx
) {
1708 int err
= insert_page_in_batch_locked(mm
, pte
,
1709 addr
, pages
[curr_page_idx
], prot
);
1710 if (unlikely(err
)) {
1711 pte_unmap_unlock(start_pte
, pte_lock
);
1713 remaining_pages_total
-= pte_idx
;
1719 pte_unmap_unlock(start_pte
, pte_lock
);
1720 pages_to_write_in_pmd
-= batch_size
;
1721 remaining_pages_total
-= batch_size
;
1723 if (remaining_pages_total
)
1727 *num
= remaining_pages_total
;
1730 #endif /* ifdef pte_index */
1733 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1734 * @vma: user vma to map to
1735 * @addr: target start user address of these pages
1736 * @pages: source kernel pages
1737 * @num: in: number of pages to map. out: number of pages that were *not*
1738 * mapped. (0 means all pages were successfully mapped).
1740 * Preferred over vm_insert_page() when inserting multiple pages.
1742 * In case of error, we may have mapped a subset of the provided
1743 * pages. It is the caller's responsibility to account for this case.
1745 * The same restrictions apply as in vm_insert_page().
1747 int vm_insert_pages(struct vm_area_struct
*vma
, unsigned long addr
,
1748 struct page
**pages
, unsigned long *num
)
1751 const unsigned long end_addr
= addr
+ (*num
* PAGE_SIZE
) - 1;
1753 if (addr
< vma
->vm_start
|| end_addr
>= vma
->vm_end
)
1755 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1756 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
1757 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1758 vma
->vm_flags
|= VM_MIXEDMAP
;
1760 /* Defer page refcount checking till we're about to map that page. */
1761 return insert_pages(vma
, addr
, pages
, num
, vma
->vm_page_prot
);
1763 unsigned long idx
= 0, pgcount
= *num
;
1766 for (; idx
< pgcount
; ++idx
) {
1767 err
= vm_insert_page(vma
, addr
+ (PAGE_SIZE
* idx
), pages
[idx
]);
1771 *num
= pgcount
- idx
;
1773 #endif /* ifdef pte_index */
1775 EXPORT_SYMBOL(vm_insert_pages
);
1778 * vm_insert_page - insert single page into user vma
1779 * @vma: user vma to map to
1780 * @addr: target user address of this page
1781 * @page: source kernel page
1783 * This allows drivers to insert individual pages they've allocated
1786 * The page has to be a nice clean _individual_ kernel allocation.
1787 * If you allocate a compound page, you need to have marked it as
1788 * such (__GFP_COMP), or manually just split the page up yourself
1789 * (see split_page()).
1791 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1792 * took an arbitrary page protection parameter. This doesn't allow
1793 * that. Your vma protection will have to be set up correctly, which
1794 * means that if you want a shared writable mapping, you'd better
1795 * ask for a shared writable mapping!
1797 * The page does not need to be reserved.
1799 * Usually this function is called from f_op->mmap() handler
1800 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1801 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1802 * function from other places, for example from page-fault handler.
1804 * Return: %0 on success, negative error code otherwise.
1806 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1809 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1811 if (!page_count(page
))
1813 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1814 BUG_ON(mmap_read_trylock(vma
->vm_mm
));
1815 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1816 vma
->vm_flags
|= VM_MIXEDMAP
;
1818 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1820 EXPORT_SYMBOL(vm_insert_page
);
1823 * __vm_map_pages - maps range of kernel pages into user vma
1824 * @vma: user vma to map to
1825 * @pages: pointer to array of source kernel pages
1826 * @num: number of pages in page array
1827 * @offset: user's requested vm_pgoff
1829 * This allows drivers to map range of kernel pages into a user vma.
1831 * Return: 0 on success and error code otherwise.
1833 static int __vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
1834 unsigned long num
, unsigned long offset
)
1836 unsigned long count
= vma_pages(vma
);
1837 unsigned long uaddr
= vma
->vm_start
;
1840 /* Fail if the user requested offset is beyond the end of the object */
1844 /* Fail if the user requested size exceeds available object size */
1845 if (count
> num
- offset
)
1848 for (i
= 0; i
< count
; i
++) {
1849 ret
= vm_insert_page(vma
, uaddr
, pages
[offset
+ i
]);
1859 * vm_map_pages - maps range of kernel pages starts with non zero offset
1860 * @vma: user vma to map to
1861 * @pages: pointer to array of source kernel pages
1862 * @num: number of pages in page array
1864 * Maps an object consisting of @num pages, catering for the user's
1865 * requested vm_pgoff
1867 * If we fail to insert any page into the vma, the function will return
1868 * immediately leaving any previously inserted pages present. Callers
1869 * from the mmap handler may immediately return the error as their caller
1870 * will destroy the vma, removing any successfully inserted pages. Other
1871 * callers should make their own arrangements for calling unmap_region().
1873 * Context: Process context. Called by mmap handlers.
1874 * Return: 0 on success and error code otherwise.
1876 int vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
1879 return __vm_map_pages(vma
, pages
, num
, vma
->vm_pgoff
);
1881 EXPORT_SYMBOL(vm_map_pages
);
1884 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1885 * @vma: user vma to map to
1886 * @pages: pointer to array of source kernel pages
1887 * @num: number of pages in page array
1889 * Similar to vm_map_pages(), except that it explicitly sets the offset
1890 * to 0. This function is intended for the drivers that did not consider
1893 * Context: Process context. Called by mmap handlers.
1894 * Return: 0 on success and error code otherwise.
1896 int vm_map_pages_zero(struct vm_area_struct
*vma
, struct page
**pages
,
1899 return __vm_map_pages(vma
, pages
, num
, 0);
1901 EXPORT_SYMBOL(vm_map_pages_zero
);
1903 static vm_fault_t
insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1904 pfn_t pfn
, pgprot_t prot
, bool mkwrite
)
1906 struct mm_struct
*mm
= vma
->vm_mm
;
1910 pte
= get_locked_pte(mm
, addr
, &ptl
);
1912 return VM_FAULT_OOM
;
1913 if (!pte_none(*pte
)) {
1916 * For read faults on private mappings the PFN passed
1917 * in may not match the PFN we have mapped if the
1918 * mapped PFN is a writeable COW page. In the mkwrite
1919 * case we are creating a writable PTE for a shared
1920 * mapping and we expect the PFNs to match. If they
1921 * don't match, we are likely racing with block
1922 * allocation and mapping invalidation so just skip the
1925 if (pte_pfn(*pte
) != pfn_t_to_pfn(pfn
)) {
1926 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte
)));
1929 entry
= pte_mkyoung(*pte
);
1930 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1931 if (ptep_set_access_flags(vma
, addr
, pte
, entry
, 1))
1932 update_mmu_cache(vma
, addr
, pte
);
1937 /* Ok, finally just insert the thing.. */
1938 if (pfn_t_devmap(pfn
))
1939 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
1941 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
1944 entry
= pte_mkyoung(entry
);
1945 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1948 set_pte_at(mm
, addr
, pte
, entry
);
1949 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1952 pte_unmap_unlock(pte
, ptl
);
1953 return VM_FAULT_NOPAGE
;
1957 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1958 * @vma: user vma to map to
1959 * @addr: target user address of this page
1960 * @pfn: source kernel pfn
1961 * @pgprot: pgprot flags for the inserted page
1963 * This is exactly like vmf_insert_pfn(), except that it allows drivers
1964 * to override pgprot on a per-page basis.
1966 * This only makes sense for IO mappings, and it makes no sense for
1967 * COW mappings. In general, using multiple vmas is preferable;
1968 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1971 * See vmf_insert_mixed_prot() for a discussion of the implication of using
1972 * a value of @pgprot different from that of @vma->vm_page_prot.
1974 * Context: Process context. May allocate using %GFP_KERNEL.
1975 * Return: vm_fault_t value.
1977 vm_fault_t
vmf_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
1978 unsigned long pfn
, pgprot_t pgprot
)
1981 * Technically, architectures with pte_special can avoid all these
1982 * restrictions (same for remap_pfn_range). However we would like
1983 * consistency in testing and feature parity among all, so we should
1984 * try to keep these invariants in place for everybody.
1986 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1987 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1988 (VM_PFNMAP
|VM_MIXEDMAP
));
1989 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1990 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1992 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1993 return VM_FAULT_SIGBUS
;
1995 if (!pfn_modify_allowed(pfn
, pgprot
))
1996 return VM_FAULT_SIGBUS
;
1998 track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
));
2000 return insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
,
2003 EXPORT_SYMBOL(vmf_insert_pfn_prot
);
2006 * vmf_insert_pfn - insert single pfn into user vma
2007 * @vma: user vma to map to
2008 * @addr: target user address of this page
2009 * @pfn: source kernel pfn
2011 * Similar to vm_insert_page, this allows drivers to insert individual pages
2012 * they've allocated into a user vma. Same comments apply.
2014 * This function should only be called from a vm_ops->fault handler, and
2015 * in that case the handler should return the result of this function.
2017 * vma cannot be a COW mapping.
2019 * As this is called only for pages that do not currently exist, we
2020 * do not need to flush old virtual caches or the TLB.
2022 * Context: Process context. May allocate using %GFP_KERNEL.
2023 * Return: vm_fault_t value.
2025 vm_fault_t
vmf_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2028 return vmf_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
2030 EXPORT_SYMBOL(vmf_insert_pfn
);
2032 static bool vm_mixed_ok(struct vm_area_struct
*vma
, pfn_t pfn
)
2034 /* these checks mirror the abort conditions in vm_normal_page */
2035 if (vma
->vm_flags
& VM_MIXEDMAP
)
2037 if (pfn_t_devmap(pfn
))
2039 if (pfn_t_special(pfn
))
2041 if (is_zero_pfn(pfn_t_to_pfn(pfn
)))
2046 static vm_fault_t
__vm_insert_mixed(struct vm_area_struct
*vma
,
2047 unsigned long addr
, pfn_t pfn
, pgprot_t pgprot
,
2052 BUG_ON(!vm_mixed_ok(vma
, pfn
));
2054 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2055 return VM_FAULT_SIGBUS
;
2057 track_pfn_insert(vma
, &pgprot
, pfn
);
2059 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn
), pgprot
))
2060 return VM_FAULT_SIGBUS
;
2063 * If we don't have pte special, then we have to use the pfn_valid()
2064 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2065 * refcount the page if pfn_valid is true (hence insert_page rather
2066 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2067 * without pte special, it would there be refcounted as a normal page.
2069 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
) &&
2070 !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
2074 * At this point we are committed to insert_page()
2075 * regardless of whether the caller specified flags that
2076 * result in pfn_t_has_page() == false.
2078 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
2079 err
= insert_page(vma
, addr
, page
, pgprot
);
2081 return insert_pfn(vma
, addr
, pfn
, pgprot
, mkwrite
);
2085 return VM_FAULT_OOM
;
2086 if (err
< 0 && err
!= -EBUSY
)
2087 return VM_FAULT_SIGBUS
;
2089 return VM_FAULT_NOPAGE
;
2093 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2094 * @vma: user vma to map to
2095 * @addr: target user address of this page
2096 * @pfn: source kernel pfn
2097 * @pgprot: pgprot flags for the inserted page
2099 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2100 * to override pgprot on a per-page basis.
2102 * Typically this function should be used by drivers to set caching- and
2103 * encryption bits different than those of @vma->vm_page_prot, because
2104 * the caching- or encryption mode may not be known at mmap() time.
2105 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2106 * to set caching and encryption bits for those vmas (except for COW pages).
2107 * This is ensured by core vm only modifying these page table entries using
2108 * functions that don't touch caching- or encryption bits, using pte_modify()
2109 * if needed. (See for example mprotect()).
2110 * Also when new page-table entries are created, this is only done using the
2111 * fault() callback, and never using the value of vma->vm_page_prot,
2112 * except for page-table entries that point to anonymous pages as the result
2115 * Context: Process context. May allocate using %GFP_KERNEL.
2116 * Return: vm_fault_t value.
2118 vm_fault_t
vmf_insert_mixed_prot(struct vm_area_struct
*vma
, unsigned long addr
,
2119 pfn_t pfn
, pgprot_t pgprot
)
2121 return __vm_insert_mixed(vma
, addr
, pfn
, pgprot
, false);
2123 EXPORT_SYMBOL(vmf_insert_mixed_prot
);
2125 vm_fault_t
vmf_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2128 return __vm_insert_mixed(vma
, addr
, pfn
, vma
->vm_page_prot
, false);
2130 EXPORT_SYMBOL(vmf_insert_mixed
);
2133 * If the insertion of PTE failed because someone else already added a
2134 * different entry in the mean time, we treat that as success as we assume
2135 * the same entry was actually inserted.
2137 vm_fault_t
vmf_insert_mixed_mkwrite(struct vm_area_struct
*vma
,
2138 unsigned long addr
, pfn_t pfn
)
2140 return __vm_insert_mixed(vma
, addr
, pfn
, vma
->vm_page_prot
, true);
2142 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite
);
2145 * maps a range of physical memory into the requested pages. the old
2146 * mappings are removed. any references to nonexistent pages results
2147 * in null mappings (currently treated as "copy-on-access")
2149 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2150 unsigned long addr
, unsigned long end
,
2151 unsigned long pfn
, pgprot_t prot
)
2157 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2160 arch_enter_lazy_mmu_mode();
2162 BUG_ON(!pte_none(*pte
));
2163 if (!pfn_modify_allowed(pfn
, prot
)) {
2167 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2169 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2170 arch_leave_lazy_mmu_mode();
2171 pte_unmap_unlock(pte
- 1, ptl
);
2175 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2176 unsigned long addr
, unsigned long end
,
2177 unsigned long pfn
, pgprot_t prot
)
2183 pfn
-= addr
>> PAGE_SHIFT
;
2184 pmd
= pmd_alloc(mm
, pud
, addr
);
2187 VM_BUG_ON(pmd_trans_huge(*pmd
));
2189 next
= pmd_addr_end(addr
, end
);
2190 err
= remap_pte_range(mm
, pmd
, addr
, next
,
2191 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2194 } while (pmd
++, addr
= next
, addr
!= end
);
2198 static inline int remap_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2199 unsigned long addr
, unsigned long end
,
2200 unsigned long pfn
, pgprot_t prot
)
2206 pfn
-= addr
>> PAGE_SHIFT
;
2207 pud
= pud_alloc(mm
, p4d
, addr
);
2211 next
= pud_addr_end(addr
, end
);
2212 err
= remap_pmd_range(mm
, pud
, addr
, next
,
2213 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2216 } while (pud
++, addr
= next
, addr
!= end
);
2220 static inline int remap_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2221 unsigned long addr
, unsigned long end
,
2222 unsigned long pfn
, pgprot_t prot
)
2228 pfn
-= addr
>> PAGE_SHIFT
;
2229 p4d
= p4d_alloc(mm
, pgd
, addr
);
2233 next
= p4d_addr_end(addr
, end
);
2234 err
= remap_pud_range(mm
, p4d
, addr
, next
,
2235 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2238 } while (p4d
++, addr
= next
, addr
!= end
);
2243 * remap_pfn_range - remap kernel memory to userspace
2244 * @vma: user vma to map to
2245 * @addr: target page aligned user address to start at
2246 * @pfn: page frame number of kernel physical memory address
2247 * @size: size of mapping area
2248 * @prot: page protection flags for this mapping
2250 * Note: this is only safe if the mm semaphore is held when called.
2252 * Return: %0 on success, negative error code otherwise.
2254 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2255 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2259 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2260 struct mm_struct
*mm
= vma
->vm_mm
;
2261 unsigned long remap_pfn
= pfn
;
2264 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr
)))
2268 * Physically remapped pages are special. Tell the
2269 * rest of the world about it:
2270 * VM_IO tells people not to look at these pages
2271 * (accesses can have side effects).
2272 * VM_PFNMAP tells the core MM that the base pages are just
2273 * raw PFN mappings, and do not have a "struct page" associated
2276 * Disable vma merging and expanding with mremap().
2278 * Omit vma from core dump, even when VM_IO turned off.
2280 * There's a horrible special case to handle copy-on-write
2281 * behaviour that some programs depend on. We mark the "original"
2282 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2283 * See vm_normal_page() for details.
2285 if (is_cow_mapping(vma
->vm_flags
)) {
2286 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2288 vma
->vm_pgoff
= pfn
;
2291 err
= track_pfn_remap(vma
, &prot
, remap_pfn
, addr
, PAGE_ALIGN(size
));
2295 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
2297 BUG_ON(addr
>= end
);
2298 pfn
-= addr
>> PAGE_SHIFT
;
2299 pgd
= pgd_offset(mm
, addr
);
2300 flush_cache_range(vma
, addr
, end
);
2302 next
= pgd_addr_end(addr
, end
);
2303 err
= remap_p4d_range(mm
, pgd
, addr
, next
,
2304 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2307 } while (pgd
++, addr
= next
, addr
!= end
);
2310 untrack_pfn(vma
, remap_pfn
, PAGE_ALIGN(size
));
2314 EXPORT_SYMBOL(remap_pfn_range
);
2317 * vm_iomap_memory - remap memory to userspace
2318 * @vma: user vma to map to
2319 * @start: start of the physical memory to be mapped
2320 * @len: size of area
2322 * This is a simplified io_remap_pfn_range() for common driver use. The
2323 * driver just needs to give us the physical memory range to be mapped,
2324 * we'll figure out the rest from the vma information.
2326 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2327 * whatever write-combining details or similar.
2329 * Return: %0 on success, negative error code otherwise.
2331 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2333 unsigned long vm_len
, pfn
, pages
;
2335 /* Check that the physical memory area passed in looks valid */
2336 if (start
+ len
< start
)
2339 * You *really* shouldn't map things that aren't page-aligned,
2340 * but we've historically allowed it because IO memory might
2341 * just have smaller alignment.
2343 len
+= start
& ~PAGE_MASK
;
2344 pfn
= start
>> PAGE_SHIFT
;
2345 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2346 if (pfn
+ pages
< pfn
)
2349 /* We start the mapping 'vm_pgoff' pages into the area */
2350 if (vma
->vm_pgoff
> pages
)
2352 pfn
+= vma
->vm_pgoff
;
2353 pages
-= vma
->vm_pgoff
;
2355 /* Can we fit all of the mapping? */
2356 vm_len
= vma
->vm_end
- vma
->vm_start
;
2357 if (vm_len
>> PAGE_SHIFT
> pages
)
2360 /* Ok, let it rip */
2361 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2363 EXPORT_SYMBOL(vm_iomap_memory
);
2365 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2366 unsigned long addr
, unsigned long end
,
2367 pte_fn_t fn
, void *data
, bool create
,
2368 pgtbl_mod_mask
*mask
)
2375 pte
= (mm
== &init_mm
) ?
2376 pte_alloc_kernel_track(pmd
, addr
, mask
) :
2377 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2381 pte
= (mm
== &init_mm
) ?
2382 pte_offset_kernel(pmd
, addr
) :
2383 pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
2386 BUG_ON(pmd_huge(*pmd
));
2388 arch_enter_lazy_mmu_mode();
2391 if (create
|| !pte_none(*pte
)) {
2392 err
= fn(pte
++, addr
, data
);
2396 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2397 *mask
|= PGTBL_PTE_MODIFIED
;
2399 arch_leave_lazy_mmu_mode();
2402 pte_unmap_unlock(pte
-1, ptl
);
2406 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2407 unsigned long addr
, unsigned long end
,
2408 pte_fn_t fn
, void *data
, bool create
,
2409 pgtbl_mod_mask
*mask
)
2415 BUG_ON(pud_huge(*pud
));
2418 pmd
= pmd_alloc_track(mm
, pud
, addr
, mask
);
2422 pmd
= pmd_offset(pud
, addr
);
2425 next
= pmd_addr_end(addr
, end
);
2426 if (create
|| !pmd_none_or_clear_bad(pmd
)) {
2427 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
,
2432 } while (pmd
++, addr
= next
, addr
!= end
);
2436 static int apply_to_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2437 unsigned long addr
, unsigned long end
,
2438 pte_fn_t fn
, void *data
, bool create
,
2439 pgtbl_mod_mask
*mask
)
2446 pud
= pud_alloc_track(mm
, p4d
, addr
, mask
);
2450 pud
= pud_offset(p4d
, addr
);
2453 next
= pud_addr_end(addr
, end
);
2454 if (create
|| !pud_none_or_clear_bad(pud
)) {
2455 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
,
2460 } while (pud
++, addr
= next
, addr
!= end
);
2464 static int apply_to_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2465 unsigned long addr
, unsigned long end
,
2466 pte_fn_t fn
, void *data
, bool create
,
2467 pgtbl_mod_mask
*mask
)
2474 p4d
= p4d_alloc_track(mm
, pgd
, addr
, mask
);
2478 p4d
= p4d_offset(pgd
, addr
);
2481 next
= p4d_addr_end(addr
, end
);
2482 if (create
|| !p4d_none_or_clear_bad(p4d
)) {
2483 err
= apply_to_pud_range(mm
, p4d
, addr
, next
, fn
, data
,
2488 } while (p4d
++, addr
= next
, addr
!= end
);
2492 static int __apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2493 unsigned long size
, pte_fn_t fn
,
2494 void *data
, bool create
)
2497 unsigned long start
= addr
, next
;
2498 unsigned long end
= addr
+ size
;
2499 pgtbl_mod_mask mask
= 0;
2502 if (WARN_ON(addr
>= end
))
2505 pgd
= pgd_offset(mm
, addr
);
2507 next
= pgd_addr_end(addr
, end
);
2508 if (!create
&& pgd_none_or_clear_bad(pgd
))
2510 err
= apply_to_p4d_range(mm
, pgd
, addr
, next
, fn
, data
, create
, &mask
);
2513 } while (pgd
++, addr
= next
, addr
!= end
);
2515 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
2516 arch_sync_kernel_mappings(start
, start
+ size
);
2522 * Scan a region of virtual memory, filling in page tables as necessary
2523 * and calling a provided function on each leaf page table.
2525 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2526 unsigned long size
, pte_fn_t fn
, void *data
)
2528 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, true);
2530 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2533 * Scan a region of virtual memory, calling a provided function on
2534 * each leaf page table where it exists.
2536 * Unlike apply_to_page_range, this does _not_ fill in page tables
2537 * where they are absent.
2539 int apply_to_existing_page_range(struct mm_struct
*mm
, unsigned long addr
,
2540 unsigned long size
, pte_fn_t fn
, void *data
)
2542 return __apply_to_page_range(mm
, addr
, size
, fn
, data
, false);
2544 EXPORT_SYMBOL_GPL(apply_to_existing_page_range
);
2547 * handle_pte_fault chooses page fault handler according to an entry which was
2548 * read non-atomically. Before making any commitment, on those architectures
2549 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2550 * parts, do_swap_page must check under lock before unmapping the pte and
2551 * proceeding (but do_wp_page is only called after already making such a check;
2552 * and do_anonymous_page can safely check later on).
2554 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2555 pte_t
*page_table
, pte_t orig_pte
)
2558 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2559 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2560 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2562 same
= pte_same(*page_table
, orig_pte
);
2566 pte_unmap(page_table
);
2570 static inline bool cow_user_page(struct page
*dst
, struct page
*src
,
2571 struct vm_fault
*vmf
)
2576 bool locked
= false;
2577 struct vm_area_struct
*vma
= vmf
->vma
;
2578 struct mm_struct
*mm
= vma
->vm_mm
;
2579 unsigned long addr
= vmf
->address
;
2582 copy_user_highpage(dst
, src
, addr
, vma
);
2587 * If the source page was a PFN mapping, we don't have
2588 * a "struct page" for it. We do a best-effort copy by
2589 * just copying from the original user address. If that
2590 * fails, we just zero-fill it. Live with it.
2592 kaddr
= kmap_atomic(dst
);
2593 uaddr
= (void __user
*)(addr
& PAGE_MASK
);
2596 * On architectures with software "accessed" bits, we would
2597 * take a double page fault, so mark it accessed here.
2599 if (arch_faults_on_old_pte() && !pte_young(vmf
->orig_pte
)) {
2602 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2604 if (!likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2606 * Other thread has already handled the fault
2607 * and update local tlb only
2609 update_mmu_tlb(vma
, addr
, vmf
->pte
);
2614 entry
= pte_mkyoung(vmf
->orig_pte
);
2615 if (ptep_set_access_flags(vma
, addr
, vmf
->pte
, entry
, 0))
2616 update_mmu_cache(vma
, addr
, vmf
->pte
);
2620 * This really shouldn't fail, because the page is there
2621 * in the page tables. But it might just be unreadable,
2622 * in which case we just give up and fill the result with
2625 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
2629 /* Re-validate under PTL if the page is still mapped */
2630 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, addr
, &vmf
->ptl
);
2632 if (!likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2633 /* The PTE changed under us, update local tlb */
2634 update_mmu_tlb(vma
, addr
, vmf
->pte
);
2640 * The same page can be mapped back since last copy attempt.
2641 * Try to copy again under PTL.
2643 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
)) {
2645 * Give a warn in case there can be some obscure
2658 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2659 kunmap_atomic(kaddr
);
2660 flush_dcache_page(dst
);
2665 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
2667 struct file
*vm_file
= vma
->vm_file
;
2670 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
2673 * Special mappings (e.g. VDSO) do not have any file so fake
2674 * a default GFP_KERNEL for them.
2680 * Notify the address space that the page is about to become writable so that
2681 * it can prohibit this or wait for the page to get into an appropriate state.
2683 * We do this without the lock held, so that it can sleep if it needs to.
2685 static vm_fault_t
do_page_mkwrite(struct vm_fault
*vmf
)
2688 struct page
*page
= vmf
->page
;
2689 unsigned int old_flags
= vmf
->flags
;
2691 vmf
->flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2693 if (vmf
->vma
->vm_file
&&
2694 IS_SWAPFILE(vmf
->vma
->vm_file
->f_mapping
->host
))
2695 return VM_FAULT_SIGBUS
;
2697 ret
= vmf
->vma
->vm_ops
->page_mkwrite(vmf
);
2698 /* Restore original flags so that caller is not surprised */
2699 vmf
->flags
= old_flags
;
2700 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2702 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
2704 if (!page
->mapping
) {
2706 return 0; /* retry */
2708 ret
|= VM_FAULT_LOCKED
;
2710 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2715 * Handle dirtying of a page in shared file mapping on a write fault.
2717 * The function expects the page to be locked and unlocks it.
2719 static vm_fault_t
fault_dirty_shared_page(struct vm_fault
*vmf
)
2721 struct vm_area_struct
*vma
= vmf
->vma
;
2722 struct address_space
*mapping
;
2723 struct page
*page
= vmf
->page
;
2725 bool page_mkwrite
= vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
;
2727 dirtied
= set_page_dirty(page
);
2728 VM_BUG_ON_PAGE(PageAnon(page
), page
);
2730 * Take a local copy of the address_space - page.mapping may be zeroed
2731 * by truncate after unlock_page(). The address_space itself remains
2732 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2733 * release semantics to prevent the compiler from undoing this copying.
2735 mapping
= page_rmapping(page
);
2739 file_update_time(vma
->vm_file
);
2742 * Throttle page dirtying rate down to writeback speed.
2744 * mapping may be NULL here because some device drivers do not
2745 * set page.mapping but still dirty their pages
2747 * Drop the mmap_lock before waiting on IO, if we can. The file
2748 * is pinning the mapping, as per above.
2750 if ((dirtied
|| page_mkwrite
) && mapping
) {
2753 fpin
= maybe_unlock_mmap_for_io(vmf
, NULL
);
2754 balance_dirty_pages_ratelimited(mapping
);
2757 return VM_FAULT_RETRY
;
2765 * Handle write page faults for pages that can be reused in the current vma
2767 * This can happen either due to the mapping being with the VM_SHARED flag,
2768 * or due to us being the last reference standing to the page. In either
2769 * case, all we need to do here is to mark the page as writable and update
2770 * any related book-keeping.
2772 static inline void wp_page_reuse(struct vm_fault
*vmf
)
2773 __releases(vmf
->ptl
)
2775 struct vm_area_struct
*vma
= vmf
->vma
;
2776 struct page
*page
= vmf
->page
;
2779 * Clear the pages cpupid information as the existing
2780 * information potentially belongs to a now completely
2781 * unrelated process.
2784 page_cpupid_xchg_last(page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2786 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2787 entry
= pte_mkyoung(vmf
->orig_pte
);
2788 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2789 if (ptep_set_access_flags(vma
, vmf
->address
, vmf
->pte
, entry
, 1))
2790 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2791 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2792 count_vm_event(PGREUSE
);
2796 * Handle the case of a page which we actually need to copy to a new page.
2798 * Called with mmap_lock locked and the old page referenced, but
2799 * without the ptl held.
2801 * High level logic flow:
2803 * - Allocate a page, copy the content of the old page to the new one.
2804 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2805 * - Take the PTL. If the pte changed, bail out and release the allocated page
2806 * - If the pte is still the way we remember it, update the page table and all
2807 * relevant references. This includes dropping the reference the page-table
2808 * held to the old page, as well as updating the rmap.
2809 * - In any case, unlock the PTL and drop the reference we took to the old page.
2811 static vm_fault_t
wp_page_copy(struct vm_fault
*vmf
)
2813 struct vm_area_struct
*vma
= vmf
->vma
;
2814 struct mm_struct
*mm
= vma
->vm_mm
;
2815 struct page
*old_page
= vmf
->page
;
2816 struct page
*new_page
= NULL
;
2818 int page_copied
= 0;
2819 struct mmu_notifier_range range
;
2821 if (unlikely(anon_vma_prepare(vma
)))
2824 if (is_zero_pfn(pte_pfn(vmf
->orig_pte
))) {
2825 new_page
= alloc_zeroed_user_highpage_movable(vma
,
2830 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
2835 if (!cow_user_page(new_page
, old_page
, vmf
)) {
2837 * COW failed, if the fault was solved by other,
2838 * it's fine. If not, userspace would re-fault on
2839 * the same address and we will handle the fault
2840 * from the second attempt.
2849 if (mem_cgroup_charge(new_page
, mm
, GFP_KERNEL
))
2851 cgroup_throttle_swaprate(new_page
, GFP_KERNEL
);
2853 __SetPageUptodate(new_page
);
2855 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
,
2856 vmf
->address
& PAGE_MASK
,
2857 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
);
2858 mmu_notifier_invalidate_range_start(&range
);
2861 * Re-check the pte - we dropped the lock
2863 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, vmf
->address
, &vmf
->ptl
);
2864 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2866 if (!PageAnon(old_page
)) {
2867 dec_mm_counter_fast(mm
,
2868 mm_counter_file(old_page
));
2869 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2872 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2874 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2875 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2876 entry
= pte_sw_mkyoung(entry
);
2877 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2879 * Clear the pte entry and flush it first, before updating the
2880 * pte with the new entry. This will avoid a race condition
2881 * seen in the presence of one thread doing SMC and another
2884 ptep_clear_flush_notify(vma
, vmf
->address
, vmf
->pte
);
2885 page_add_new_anon_rmap(new_page
, vma
, vmf
->address
, false);
2886 lru_cache_add_inactive_or_unevictable(new_page
, vma
);
2888 * We call the notify macro here because, when using secondary
2889 * mmu page tables (such as kvm shadow page tables), we want the
2890 * new page to be mapped directly into the secondary page table.
2892 set_pte_at_notify(mm
, vmf
->address
, vmf
->pte
, entry
);
2893 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2896 * Only after switching the pte to the new page may
2897 * we remove the mapcount here. Otherwise another
2898 * process may come and find the rmap count decremented
2899 * before the pte is switched to the new page, and
2900 * "reuse" the old page writing into it while our pte
2901 * here still points into it and can be read by other
2904 * The critical issue is to order this
2905 * page_remove_rmap with the ptp_clear_flush above.
2906 * Those stores are ordered by (if nothing else,)
2907 * the barrier present in the atomic_add_negative
2908 * in page_remove_rmap.
2910 * Then the TLB flush in ptep_clear_flush ensures that
2911 * no process can access the old page before the
2912 * decremented mapcount is visible. And the old page
2913 * cannot be reused until after the decremented
2914 * mapcount is visible. So transitively, TLBs to
2915 * old page will be flushed before it can be reused.
2917 page_remove_rmap(old_page
, false);
2920 /* Free the old page.. */
2921 new_page
= old_page
;
2924 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
2930 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2932 * No need to double call mmu_notifier->invalidate_range() callback as
2933 * the above ptep_clear_flush_notify() did already call it.
2935 mmu_notifier_invalidate_range_only_end(&range
);
2938 * Don't let another task, with possibly unlocked vma,
2939 * keep the mlocked page.
2941 if (page_copied
&& (vma
->vm_flags
& VM_LOCKED
)) {
2942 lock_page(old_page
); /* LRU manipulation */
2943 if (PageMlocked(old_page
))
2944 munlock_vma_page(old_page
);
2945 unlock_page(old_page
);
2949 return page_copied
? VM_FAULT_WRITE
: 0;
2955 return VM_FAULT_OOM
;
2959 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2960 * writeable once the page is prepared
2962 * @vmf: structure describing the fault
2964 * This function handles all that is needed to finish a write page fault in a
2965 * shared mapping due to PTE being read-only once the mapped page is prepared.
2966 * It handles locking of PTE and modifying it.
2968 * The function expects the page to be locked or other protection against
2969 * concurrent faults / writeback (such as DAX radix tree locks).
2971 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2972 * we acquired PTE lock.
2974 vm_fault_t
finish_mkwrite_fault(struct vm_fault
*vmf
)
2976 WARN_ON_ONCE(!(vmf
->vma
->vm_flags
& VM_SHARED
));
2977 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
2980 * We might have raced with another page fault while we released the
2981 * pte_offset_map_lock.
2983 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
2984 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
2985 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2986 return VM_FAULT_NOPAGE
;
2993 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2996 static vm_fault_t
wp_pfn_shared(struct vm_fault
*vmf
)
2998 struct vm_area_struct
*vma
= vmf
->vma
;
3000 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
3003 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3004 vmf
->flags
|= FAULT_FLAG_MKWRITE
;
3005 ret
= vma
->vm_ops
->pfn_mkwrite(vmf
);
3006 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))
3008 return finish_mkwrite_fault(vmf
);
3011 return VM_FAULT_WRITE
;
3014 static vm_fault_t
wp_page_shared(struct vm_fault
*vmf
)
3015 __releases(vmf
->ptl
)
3017 struct vm_area_struct
*vma
= vmf
->vma
;
3018 vm_fault_t ret
= VM_FAULT_WRITE
;
3020 get_page(vmf
->page
);
3022 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
3025 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3026 tmp
= do_page_mkwrite(vmf
);
3027 if (unlikely(!tmp
|| (tmp
&
3028 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3029 put_page(vmf
->page
);
3032 tmp
= finish_mkwrite_fault(vmf
);
3033 if (unlikely(tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3034 unlock_page(vmf
->page
);
3035 put_page(vmf
->page
);
3040 lock_page(vmf
->page
);
3042 ret
|= fault_dirty_shared_page(vmf
);
3043 put_page(vmf
->page
);
3049 * This routine handles present pages, when users try to write
3050 * to a shared page. It is done by copying the page to a new address
3051 * and decrementing the shared-page counter for the old page.
3053 * Note that this routine assumes that the protection checks have been
3054 * done by the caller (the low-level page fault routine in most cases).
3055 * Thus we can safely just mark it writable once we've done any necessary
3058 * We also mark the page dirty at this point even though the page will
3059 * change only once the write actually happens. This avoids a few races,
3060 * and potentially makes it more efficient.
3062 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3063 * but allow concurrent faults), with pte both mapped and locked.
3064 * We return with mmap_lock still held, but pte unmapped and unlocked.
3066 static vm_fault_t
do_wp_page(struct vm_fault
*vmf
)
3067 __releases(vmf
->ptl
)
3069 struct vm_area_struct
*vma
= vmf
->vma
;
3071 if (userfaultfd_pte_wp(vma
, *vmf
->pte
)) {
3072 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3073 return handle_userfault(vmf
, VM_UFFD_WP
);
3076 vmf
->page
= vm_normal_page(vma
, vmf
->address
, vmf
->orig_pte
);
3079 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3082 * We should not cow pages in a shared writeable mapping.
3083 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3085 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
3086 (VM_WRITE
|VM_SHARED
))
3087 return wp_pfn_shared(vmf
);
3089 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3090 return wp_page_copy(vmf
);
3094 * Take out anonymous pages first, anonymous shared vmas are
3095 * not dirty accountable.
3097 if (PageAnon(vmf
->page
)) {
3098 struct page
*page
= vmf
->page
;
3100 /* PageKsm() doesn't necessarily raise the page refcount */
3101 if (PageKsm(page
) || page_count(page
) != 1)
3103 if (!trylock_page(page
))
3105 if (PageKsm(page
) || page_mapcount(page
) != 1 || page_count(page
) != 1) {
3110 * Ok, we've got the only map reference, and the only
3111 * page count reference, and the page is locked,
3112 * it's dark out, and we're wearing sunglasses. Hit it.
3116 return VM_FAULT_WRITE
;
3117 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
3118 (VM_WRITE
|VM_SHARED
))) {
3119 return wp_page_shared(vmf
);
3123 * Ok, we need to copy. Oh, well..
3125 get_page(vmf
->page
);
3127 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3128 return wp_page_copy(vmf
);
3131 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
3132 unsigned long start_addr
, unsigned long end_addr
,
3133 struct zap_details
*details
)
3135 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
3138 static inline void unmap_mapping_range_tree(struct rb_root_cached
*root
,
3139 struct zap_details
*details
)
3141 struct vm_area_struct
*vma
;
3142 pgoff_t vba
, vea
, zba
, zea
;
3144 vma_interval_tree_foreach(vma
, root
,
3145 details
->first_index
, details
->last_index
) {
3147 vba
= vma
->vm_pgoff
;
3148 vea
= vba
+ vma_pages(vma
) - 1;
3149 zba
= details
->first_index
;
3152 zea
= details
->last_index
;
3156 unmap_mapping_range_vma(vma
,
3157 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
3158 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
3164 * unmap_mapping_pages() - Unmap pages from processes.
3165 * @mapping: The address space containing pages to be unmapped.
3166 * @start: Index of first page to be unmapped.
3167 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3168 * @even_cows: Whether to unmap even private COWed pages.
3170 * Unmap the pages in this address space from any userspace process which
3171 * has them mmaped. Generally, you want to remove COWed pages as well when
3172 * a file is being truncated, but not when invalidating pages from the page
3175 void unmap_mapping_pages(struct address_space
*mapping
, pgoff_t start
,
3176 pgoff_t nr
, bool even_cows
)
3178 struct zap_details details
= { };
3180 details
.check_mapping
= even_cows
? NULL
: mapping
;
3181 details
.first_index
= start
;
3182 details
.last_index
= start
+ nr
- 1;
3183 if (details
.last_index
< details
.first_index
)
3184 details
.last_index
= ULONG_MAX
;
3186 i_mmap_lock_write(mapping
);
3187 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
3188 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
3189 i_mmap_unlock_write(mapping
);
3193 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3194 * address_space corresponding to the specified byte range in the underlying
3197 * @mapping: the address space containing mmaps to be unmapped.
3198 * @holebegin: byte in first page to unmap, relative to the start of
3199 * the underlying file. This will be rounded down to a PAGE_SIZE
3200 * boundary. Note that this is different from truncate_pagecache(), which
3201 * must keep the partial page. In contrast, we must get rid of
3203 * @holelen: size of prospective hole in bytes. This will be rounded
3204 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3206 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3207 * but 0 when invalidating pagecache, don't throw away private data.
3209 void unmap_mapping_range(struct address_space
*mapping
,
3210 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
3212 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
3213 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3215 /* Check for overflow. */
3216 if (sizeof(holelen
) > sizeof(hlen
)) {
3218 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3219 if (holeend
& ~(long long)ULONG_MAX
)
3220 hlen
= ULONG_MAX
- hba
+ 1;
3223 unmap_mapping_pages(mapping
, hba
, hlen
, even_cows
);
3225 EXPORT_SYMBOL(unmap_mapping_range
);
3228 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3229 * but allow concurrent faults), and pte mapped but not yet locked.
3230 * We return with pte unmapped and unlocked.
3232 * We return with the mmap_lock locked or unlocked in the same cases
3233 * as does filemap_fault().
3235 vm_fault_t
do_swap_page(struct vm_fault
*vmf
)
3237 struct vm_area_struct
*vma
= vmf
->vma
;
3238 struct page
*page
= NULL
, *swapcache
;
3244 void *shadow
= NULL
;
3246 if (!pte_unmap_same(vma
->vm_mm
, vmf
->pmd
, vmf
->pte
, vmf
->orig_pte
))
3249 entry
= pte_to_swp_entry(vmf
->orig_pte
);
3250 if (unlikely(non_swap_entry(entry
))) {
3251 if (is_migration_entry(entry
)) {
3252 migration_entry_wait(vma
->vm_mm
, vmf
->pmd
,
3254 } else if (is_device_private_entry(entry
)) {
3255 vmf
->page
= device_private_entry_to_page(entry
);
3256 ret
= vmf
->page
->pgmap
->ops
->migrate_to_ram(vmf
);
3257 } else if (is_hwpoison_entry(entry
)) {
3258 ret
= VM_FAULT_HWPOISON
;
3260 print_bad_pte(vma
, vmf
->address
, vmf
->orig_pte
, NULL
);
3261 ret
= VM_FAULT_SIGBUS
;
3267 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
3268 page
= lookup_swap_cache(entry
, vma
, vmf
->address
);
3272 struct swap_info_struct
*si
= swp_swap_info(entry
);
3274 if (data_race(si
->flags
& SWP_SYNCHRONOUS_IO
) &&
3275 __swap_count(entry
) == 1) {
3276 /* skip swapcache */
3277 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
3282 __SetPageLocked(page
);
3283 __SetPageSwapBacked(page
);
3284 set_page_private(page
, entry
.val
);
3286 /* Tell memcg to use swap ownership records */
3287 SetPageSwapCache(page
);
3288 err
= mem_cgroup_charge(page
, vma
->vm_mm
,
3290 ClearPageSwapCache(page
);
3296 shadow
= get_shadow_from_swap_cache(entry
);
3298 workingset_refault(page
, shadow
);
3300 lru_cache_add(page
);
3301 swap_readpage(page
, true);
3304 page
= swapin_readahead(entry
, GFP_HIGHUSER_MOVABLE
,
3311 * Back out if somebody else faulted in this pte
3312 * while we released the pte lock.
3314 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3315 vmf
->address
, &vmf
->ptl
);
3316 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3318 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3322 /* Had to read the page from swap area: Major fault */
3323 ret
= VM_FAULT_MAJOR
;
3324 count_vm_event(PGMAJFAULT
);
3325 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
3326 } else if (PageHWPoison(page
)) {
3328 * hwpoisoned dirty swapcache pages are kept for killing
3329 * owner processes (which may be unknown at hwpoison time)
3331 ret
= VM_FAULT_HWPOISON
;
3332 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3336 locked
= lock_page_or_retry(page
, vma
->vm_mm
, vmf
->flags
);
3338 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3340 ret
|= VM_FAULT_RETRY
;
3345 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3346 * release the swapcache from under us. The page pin, and pte_same
3347 * test below, are not enough to exclude that. Even if it is still
3348 * swapcache, we need to check that the page's swap has not changed.
3350 if (unlikely((!PageSwapCache(page
) ||
3351 page_private(page
) != entry
.val
)) && swapcache
)
3354 page
= ksm_might_need_to_copy(page
, vma
, vmf
->address
);
3355 if (unlikely(!page
)) {
3361 cgroup_throttle_swaprate(page
, GFP_KERNEL
);
3364 * Back out if somebody else already faulted in this pte.
3366 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3368 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
)))
3371 if (unlikely(!PageUptodate(page
))) {
3372 ret
= VM_FAULT_SIGBUS
;
3377 * The page isn't present yet, go ahead with the fault.
3379 * Be careful about the sequence of operations here.
3380 * To get its accounting right, reuse_swap_page() must be called
3381 * while the page is counted on swap but not yet in mapcount i.e.
3382 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3383 * must be called after the swap_free(), or it will never succeed.
3386 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3387 dec_mm_counter_fast(vma
->vm_mm
, MM_SWAPENTS
);
3388 pte
= mk_pte(page
, vma
->vm_page_prot
);
3389 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
, NULL
)) {
3390 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3391 vmf
->flags
&= ~FAULT_FLAG_WRITE
;
3392 ret
|= VM_FAULT_WRITE
;
3393 exclusive
= RMAP_EXCLUSIVE
;
3395 flush_icache_page(vma
, page
);
3396 if (pte_swp_soft_dirty(vmf
->orig_pte
))
3397 pte
= pte_mksoft_dirty(pte
);
3398 if (pte_swp_uffd_wp(vmf
->orig_pte
)) {
3399 pte
= pte_mkuffd_wp(pte
);
3400 pte
= pte_wrprotect(pte
);
3402 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
3403 arch_do_swap_page(vma
->vm_mm
, vma
, vmf
->address
, pte
, vmf
->orig_pte
);
3404 vmf
->orig_pte
= pte
;
3406 /* ksm created a completely new copy */
3407 if (unlikely(page
!= swapcache
&& swapcache
)) {
3408 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3409 lru_cache_add_inactive_or_unevictable(page
, vma
);
3411 do_page_add_anon_rmap(page
, vma
, vmf
->address
, exclusive
);
3415 if (mem_cgroup_swap_full(page
) ||
3416 (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
3417 try_to_free_swap(page
);
3419 if (page
!= swapcache
&& swapcache
) {
3421 * Hold the lock to avoid the swap entry to be reused
3422 * until we take the PT lock for the pte_same() check
3423 * (to avoid false positives from pte_same). For
3424 * further safety release the lock after the swap_free
3425 * so that the swap count won't change under a
3426 * parallel locked swapcache.
3428 unlock_page(swapcache
);
3429 put_page(swapcache
);
3432 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
3433 ret
|= do_wp_page(vmf
);
3434 if (ret
& VM_FAULT_ERROR
)
3435 ret
&= VM_FAULT_ERROR
;
3439 /* No need to invalidate - it was non-present before */
3440 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3442 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3446 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3451 if (page
!= swapcache
&& swapcache
) {
3452 unlock_page(swapcache
);
3453 put_page(swapcache
);
3459 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3460 * but allow concurrent faults), and pte mapped but not yet locked.
3461 * We return with mmap_lock still held, but pte unmapped and unlocked.
3463 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
)
3465 struct vm_area_struct
*vma
= vmf
->vma
;
3470 /* File mapping without ->vm_ops ? */
3471 if (vma
->vm_flags
& VM_SHARED
)
3472 return VM_FAULT_SIGBUS
;
3475 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3476 * pte_offset_map() on pmds where a huge pmd might be created
3477 * from a different thread.
3479 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3480 * parallel threads are excluded by other means.
3482 * Here we only have mmap_read_lock(mm).
3484 if (pte_alloc(vma
->vm_mm
, vmf
->pmd
))
3485 return VM_FAULT_OOM
;
3487 /* See the comment in pte_alloc_one_map() */
3488 if (unlikely(pmd_trans_unstable(vmf
->pmd
)))
3491 /* Use the zero-page for reads */
3492 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
3493 !mm_forbids_zeropage(vma
->vm_mm
)) {
3494 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(vmf
->address
),
3495 vma
->vm_page_prot
));
3496 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3497 vmf
->address
, &vmf
->ptl
);
3498 if (!pte_none(*vmf
->pte
)) {
3499 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
3502 ret
= check_stable_address_space(vma
->vm_mm
);
3505 /* Deliver the page fault to userland, check inside PT lock */
3506 if (userfaultfd_missing(vma
)) {
3507 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3508 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3513 /* Allocate our own private page. */
3514 if (unlikely(anon_vma_prepare(vma
)))
3516 page
= alloc_zeroed_user_highpage_movable(vma
, vmf
->address
);
3520 if (mem_cgroup_charge(page
, vma
->vm_mm
, GFP_KERNEL
))
3522 cgroup_throttle_swaprate(page
, GFP_KERNEL
);
3525 * The memory barrier inside __SetPageUptodate makes sure that
3526 * preceding stores to the page contents become visible before
3527 * the set_pte_at() write.
3529 __SetPageUptodate(page
);
3531 entry
= mk_pte(page
, vma
->vm_page_prot
);
3532 entry
= pte_sw_mkyoung(entry
);
3533 if (vma
->vm_flags
& VM_WRITE
)
3534 entry
= pte_mkwrite(pte_mkdirty(entry
));
3536 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3538 if (!pte_none(*vmf
->pte
)) {
3539 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3543 ret
= check_stable_address_space(vma
->vm_mm
);
3547 /* Deliver the page fault to userland, check inside PT lock */
3548 if (userfaultfd_missing(vma
)) {
3549 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3551 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3554 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3555 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3556 lru_cache_add_inactive_or_unevictable(page
, vma
);
3558 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3560 /* No need to invalidate - it was non-present before */
3561 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3563 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3571 return VM_FAULT_OOM
;
3575 * The mmap_lock must have been held on entry, and may have been
3576 * released depending on flags and vma->vm_ops->fault() return value.
3577 * See filemap_fault() and __lock_page_retry().
3579 static vm_fault_t
__do_fault(struct vm_fault
*vmf
)
3581 struct vm_area_struct
*vma
= vmf
->vma
;
3585 * Preallocate pte before we take page_lock because this might lead to
3586 * deadlocks for memcg reclaim which waits for pages under writeback:
3588 * SetPageWriteback(A)
3594 * wait_on_page_writeback(A)
3595 * SetPageWriteback(B)
3597 * # flush A, B to clear the writeback
3599 if (pmd_none(*vmf
->pmd
) && !vmf
->prealloc_pte
) {
3600 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
3601 if (!vmf
->prealloc_pte
)
3602 return VM_FAULT_OOM
;
3603 smp_wmb(); /* See comment in __pte_alloc() */
3606 ret
= vma
->vm_ops
->fault(vmf
);
3607 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
|
3608 VM_FAULT_DONE_COW
)))
3611 if (unlikely(PageHWPoison(vmf
->page
))) {
3612 if (ret
& VM_FAULT_LOCKED
)
3613 unlock_page(vmf
->page
);
3614 put_page(vmf
->page
);
3616 return VM_FAULT_HWPOISON
;
3619 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3620 lock_page(vmf
->page
);
3622 VM_BUG_ON_PAGE(!PageLocked(vmf
->page
), vmf
->page
);
3628 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3629 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3630 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3631 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3633 static int pmd_devmap_trans_unstable(pmd_t
*pmd
)
3635 return pmd_devmap(*pmd
) || pmd_trans_unstable(pmd
);
3638 static vm_fault_t
pte_alloc_one_map(struct vm_fault
*vmf
)
3640 struct vm_area_struct
*vma
= vmf
->vma
;
3642 if (!pmd_none(*vmf
->pmd
))
3644 if (vmf
->prealloc_pte
) {
3645 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3646 if (unlikely(!pmd_none(*vmf
->pmd
))) {
3647 spin_unlock(vmf
->ptl
);
3651 mm_inc_nr_ptes(vma
->vm_mm
);
3652 pmd_populate(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3653 spin_unlock(vmf
->ptl
);
3654 vmf
->prealloc_pte
= NULL
;
3655 } else if (unlikely(pte_alloc(vma
->vm_mm
, vmf
->pmd
))) {
3656 return VM_FAULT_OOM
;
3660 * If a huge pmd materialized under us just retry later. Use
3661 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3662 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3663 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3664 * running immediately after a huge pmd fault in a different thread of
3665 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3666 * All we have to ensure is that it is a regular pmd that we can walk
3667 * with pte_offset_map() and we can do that through an atomic read in
3668 * C, which is what pmd_trans_unstable() provides.
3670 if (pmd_devmap_trans_unstable(vmf
->pmd
))
3671 return VM_FAULT_NOPAGE
;
3674 * At this point we know that our vmf->pmd points to a page of ptes
3675 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3676 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3677 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3678 * be valid and we will re-check to make sure the vmf->pte isn't
3679 * pte_none() under vmf->ptl protection when we return to
3682 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3687 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3688 static void deposit_prealloc_pte(struct vm_fault
*vmf
)
3690 struct vm_area_struct
*vma
= vmf
->vma
;
3692 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3694 * We are going to consume the prealloc table,
3695 * count that as nr_ptes.
3697 mm_inc_nr_ptes(vma
->vm_mm
);
3698 vmf
->prealloc_pte
= NULL
;
3701 static vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3703 struct vm_area_struct
*vma
= vmf
->vma
;
3704 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3705 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
3710 if (!transhuge_vma_suitable(vma
, haddr
))
3711 return VM_FAULT_FALLBACK
;
3713 ret
= VM_FAULT_FALLBACK
;
3714 page
= compound_head(page
);
3717 * Archs like ppc64 need additonal space to store information
3718 * related to pte entry. Use the preallocated table for that.
3720 if (arch_needs_pgtable_deposit() && !vmf
->prealloc_pte
) {
3721 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
3722 if (!vmf
->prealloc_pte
)
3723 return VM_FAULT_OOM
;
3724 smp_wmb(); /* See comment in __pte_alloc() */
3727 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3728 if (unlikely(!pmd_none(*vmf
->pmd
)))
3731 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
3732 flush_icache_page(vma
, page
+ i
);
3734 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
3736 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
3738 add_mm_counter(vma
->vm_mm
, mm_counter_file(page
), HPAGE_PMD_NR
);
3739 page_add_file_rmap(page
, true);
3741 * deposit and withdraw with pmd lock held
3743 if (arch_needs_pgtable_deposit())
3744 deposit_prealloc_pte(vmf
);
3746 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
3748 update_mmu_cache_pmd(vma
, haddr
, vmf
->pmd
);
3750 /* fault is handled */
3752 count_vm_event(THP_FILE_MAPPED
);
3754 spin_unlock(vmf
->ptl
);
3758 static vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3766 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3767 * mapping. If needed, the function allocates page table or use pre-allocated.
3769 * @vmf: fault environment
3770 * @page: page to map
3772 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3775 * Target users are page handler itself and implementations of
3776 * vm_ops->map_pages.
3778 * Return: %0 on success, %VM_FAULT_ code in case of error.
3780 vm_fault_t
alloc_set_pte(struct vm_fault
*vmf
, struct page
*page
)
3782 struct vm_area_struct
*vma
= vmf
->vma
;
3783 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3787 if (pmd_none(*vmf
->pmd
) && PageTransCompound(page
)) {
3788 ret
= do_set_pmd(vmf
, page
);
3789 if (ret
!= VM_FAULT_FALLBACK
)
3794 ret
= pte_alloc_one_map(vmf
);
3799 /* Re-check under ptl */
3800 if (unlikely(!pte_none(*vmf
->pte
))) {
3801 update_mmu_tlb(vma
, vmf
->address
, vmf
->pte
);
3802 return VM_FAULT_NOPAGE
;
3805 flush_icache_page(vma
, page
);
3806 entry
= mk_pte(page
, vma
->vm_page_prot
);
3807 entry
= pte_sw_mkyoung(entry
);
3809 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3810 /* copy-on-write page */
3811 if (write
&& !(vma
->vm_flags
& VM_SHARED
)) {
3812 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3813 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3814 lru_cache_add_inactive_or_unevictable(page
, vma
);
3816 inc_mm_counter_fast(vma
->vm_mm
, mm_counter_file(page
));
3817 page_add_file_rmap(page
, false);
3819 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3821 /* no need to invalidate: a not-present page won't be cached */
3822 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3829 * finish_fault - finish page fault once we have prepared the page to fault
3831 * @vmf: structure describing the fault
3833 * This function handles all that is needed to finish a page fault once the
3834 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3835 * given page, adds reverse page mapping, handles memcg charges and LRU
3838 * The function expects the page to be locked and on success it consumes a
3839 * reference of a page being mapped (for the PTE which maps it).
3841 * Return: %0 on success, %VM_FAULT_ code in case of error.
3843 vm_fault_t
finish_fault(struct vm_fault
*vmf
)
3848 /* Did we COW the page? */
3849 if ((vmf
->flags
& FAULT_FLAG_WRITE
) &&
3850 !(vmf
->vma
->vm_flags
& VM_SHARED
))
3851 page
= vmf
->cow_page
;
3856 * check even for read faults because we might have lost our CoWed
3859 if (!(vmf
->vma
->vm_flags
& VM_SHARED
))
3860 ret
= check_stable_address_space(vmf
->vma
->vm_mm
);
3862 ret
= alloc_set_pte(vmf
, page
);
3864 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3868 static unsigned long fault_around_bytes __read_mostly
=
3869 rounddown_pow_of_two(65536);
3871 #ifdef CONFIG_DEBUG_FS
3872 static int fault_around_bytes_get(void *data
, u64
*val
)
3874 *val
= fault_around_bytes
;
3879 * fault_around_bytes must be rounded down to the nearest page order as it's
3880 * what do_fault_around() expects to see.
3882 static int fault_around_bytes_set(void *data
, u64 val
)
3884 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
3886 if (val
> PAGE_SIZE
)
3887 fault_around_bytes
= rounddown_pow_of_two(val
);
3889 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
3892 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops
,
3893 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
3895 static int __init
fault_around_debugfs(void)
3897 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL
, NULL
,
3898 &fault_around_bytes_fops
);
3901 late_initcall(fault_around_debugfs
);
3905 * do_fault_around() tries to map few pages around the fault address. The hope
3906 * is that the pages will be needed soon and this will lower the number of
3909 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3910 * not ready to be mapped: not up-to-date, locked, etc.
3912 * This function is called with the page table lock taken. In the split ptlock
3913 * case the page table lock only protects only those entries which belong to
3914 * the page table corresponding to the fault address.
3916 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3919 * fault_around_bytes defines how many bytes we'll try to map.
3920 * do_fault_around() expects it to be set to a power of two less than or equal
3923 * The virtual address of the area that we map is naturally aligned to
3924 * fault_around_bytes rounded down to the machine page size
3925 * (and therefore to page order). This way it's easier to guarantee
3926 * that we don't cross page table boundaries.
3928 static vm_fault_t
do_fault_around(struct vm_fault
*vmf
)
3930 unsigned long address
= vmf
->address
, nr_pages
, mask
;
3931 pgoff_t start_pgoff
= vmf
->pgoff
;
3936 nr_pages
= READ_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
3937 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
3939 vmf
->address
= max(address
& mask
, vmf
->vma
->vm_start
);
3940 off
= ((address
- vmf
->address
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
3944 * end_pgoff is either the end of the page table, the end of
3945 * the vma or nr_pages from start_pgoff, depending what is nearest.
3947 end_pgoff
= start_pgoff
-
3948 ((vmf
->address
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
3950 end_pgoff
= min3(end_pgoff
, vma_pages(vmf
->vma
) + vmf
->vma
->vm_pgoff
- 1,
3951 start_pgoff
+ nr_pages
- 1);
3953 if (pmd_none(*vmf
->pmd
)) {
3954 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
);
3955 if (!vmf
->prealloc_pte
)
3957 smp_wmb(); /* See comment in __pte_alloc() */
3960 vmf
->vma
->vm_ops
->map_pages(vmf
, start_pgoff
, end_pgoff
);
3962 /* Huge page is mapped? Page fault is solved */
3963 if (pmd_trans_huge(*vmf
->pmd
)) {
3964 ret
= VM_FAULT_NOPAGE
;
3968 /* ->map_pages() haven't done anything useful. Cold page cache? */
3972 /* check if the page fault is solved */
3973 vmf
->pte
-= (vmf
->address
>> PAGE_SHIFT
) - (address
>> PAGE_SHIFT
);
3974 if (!pte_none(*vmf
->pte
))
3975 ret
= VM_FAULT_NOPAGE
;
3976 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3978 vmf
->address
= address
;
3983 static vm_fault_t
do_read_fault(struct vm_fault
*vmf
)
3985 struct vm_area_struct
*vma
= vmf
->vma
;
3989 * Let's call ->map_pages() first and use ->fault() as fallback
3990 * if page by the offset is not ready to be mapped (cold cache or
3993 if (vma
->vm_ops
->map_pages
&& fault_around_bytes
>> PAGE_SHIFT
> 1) {
3994 ret
= do_fault_around(vmf
);
3999 ret
= __do_fault(vmf
);
4000 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4003 ret
|= finish_fault(vmf
);
4004 unlock_page(vmf
->page
);
4005 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4006 put_page(vmf
->page
);
4010 static vm_fault_t
do_cow_fault(struct vm_fault
*vmf
)
4012 struct vm_area_struct
*vma
= vmf
->vma
;
4015 if (unlikely(anon_vma_prepare(vma
)))
4016 return VM_FAULT_OOM
;
4018 vmf
->cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, vmf
->address
);
4020 return VM_FAULT_OOM
;
4022 if (mem_cgroup_charge(vmf
->cow_page
, vma
->vm_mm
, GFP_KERNEL
)) {
4023 put_page(vmf
->cow_page
);
4024 return VM_FAULT_OOM
;
4026 cgroup_throttle_swaprate(vmf
->cow_page
, GFP_KERNEL
);
4028 ret
= __do_fault(vmf
);
4029 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4031 if (ret
& VM_FAULT_DONE_COW
)
4034 copy_user_highpage(vmf
->cow_page
, vmf
->page
, vmf
->address
, vma
);
4035 __SetPageUptodate(vmf
->cow_page
);
4037 ret
|= finish_fault(vmf
);
4038 unlock_page(vmf
->page
);
4039 put_page(vmf
->page
);
4040 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4044 put_page(vmf
->cow_page
);
4048 static vm_fault_t
do_shared_fault(struct vm_fault
*vmf
)
4050 struct vm_area_struct
*vma
= vmf
->vma
;
4051 vm_fault_t ret
, tmp
;
4053 ret
= __do_fault(vmf
);
4054 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
4058 * Check if the backing address space wants to know that the page is
4059 * about to become writable
4061 if (vma
->vm_ops
->page_mkwrite
) {
4062 unlock_page(vmf
->page
);
4063 tmp
= do_page_mkwrite(vmf
);
4064 if (unlikely(!tmp
||
4065 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
4066 put_page(vmf
->page
);
4071 ret
|= finish_fault(vmf
);
4072 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
4074 unlock_page(vmf
->page
);
4075 put_page(vmf
->page
);
4079 ret
|= fault_dirty_shared_page(vmf
);
4084 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4085 * but allow concurrent faults).
4086 * The mmap_lock may have been released depending on flags and our
4087 * return value. See filemap_fault() and __lock_page_or_retry().
4088 * If mmap_lock is released, vma may become invalid (for example
4089 * by other thread calling munmap()).
4091 static vm_fault_t
do_fault(struct vm_fault
*vmf
)
4093 struct vm_area_struct
*vma
= vmf
->vma
;
4094 struct mm_struct
*vm_mm
= vma
->vm_mm
;
4098 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4100 if (!vma
->vm_ops
->fault
) {
4102 * If we find a migration pmd entry or a none pmd entry, which
4103 * should never happen, return SIGBUS
4105 if (unlikely(!pmd_present(*vmf
->pmd
)))
4106 ret
= VM_FAULT_SIGBUS
;
4108 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
,
4113 * Make sure this is not a temporary clearing of pte
4114 * by holding ptl and checking again. A R/M/W update
4115 * of pte involves: take ptl, clearing the pte so that
4116 * we don't have concurrent modification by hardware
4117 * followed by an update.
4119 if (unlikely(pte_none(*vmf
->pte
)))
4120 ret
= VM_FAULT_SIGBUS
;
4122 ret
= VM_FAULT_NOPAGE
;
4124 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4126 } else if (!(vmf
->flags
& FAULT_FLAG_WRITE
))
4127 ret
= do_read_fault(vmf
);
4128 else if (!(vma
->vm_flags
& VM_SHARED
))
4129 ret
= do_cow_fault(vmf
);
4131 ret
= do_shared_fault(vmf
);
4133 /* preallocated pagetable is unused: free it */
4134 if (vmf
->prealloc_pte
) {
4135 pte_free(vm_mm
, vmf
->prealloc_pte
);
4136 vmf
->prealloc_pte
= NULL
;
4141 static int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
4142 unsigned long addr
, int page_nid
,
4147 count_vm_numa_event(NUMA_HINT_FAULTS
);
4148 if (page_nid
== numa_node_id()) {
4149 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
4150 *flags
|= TNF_FAULT_LOCAL
;
4153 return mpol_misplaced(page
, vma
, addr
);
4156 static vm_fault_t
do_numa_page(struct vm_fault
*vmf
)
4158 struct vm_area_struct
*vma
= vmf
->vma
;
4159 struct page
*page
= NULL
;
4160 int page_nid
= NUMA_NO_NODE
;
4163 bool migrated
= false;
4165 bool was_writable
= pte_savedwrite(vmf
->orig_pte
);
4169 * The "pte" at this point cannot be used safely without
4170 * validation through pte_unmap_same(). It's of NUMA type but
4171 * the pfn may be screwed if the read is non atomic.
4173 vmf
->ptl
= pte_lockptr(vma
->vm_mm
, vmf
->pmd
);
4174 spin_lock(vmf
->ptl
);
4175 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
4176 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4181 * Make it present again, Depending on how arch implementes non
4182 * accessible ptes, some can allow access by kernel mode.
4184 old_pte
= ptep_modify_prot_start(vma
, vmf
->address
, vmf
->pte
);
4185 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
4186 pte
= pte_mkyoung(pte
);
4188 pte
= pte_mkwrite(pte
);
4189 ptep_modify_prot_commit(vma
, vmf
->address
, vmf
->pte
, old_pte
, pte
);
4190 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
4192 page
= vm_normal_page(vma
, vmf
->address
, pte
);
4194 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4198 /* TODO: handle PTE-mapped THP */
4199 if (PageCompound(page
)) {
4200 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4205 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4206 * much anyway since they can be in shared cache state. This misses
4207 * the case where a mapping is writable but the process never writes
4208 * to it but pte_write gets cleared during protection updates and
4209 * pte_dirty has unpredictable behaviour between PTE scan updates,
4210 * background writeback, dirty balancing and application behaviour.
4212 if (!pte_write(pte
))
4213 flags
|= TNF_NO_GROUP
;
4216 * Flag if the page is shared between multiple address spaces. This
4217 * is later used when determining whether to group tasks together
4219 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
4220 flags
|= TNF_SHARED
;
4222 last_cpupid
= page_cpupid_last(page
);
4223 page_nid
= page_to_nid(page
);
4224 target_nid
= numa_migrate_prep(page
, vma
, vmf
->address
, page_nid
,
4226 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4227 if (target_nid
== NUMA_NO_NODE
) {
4232 /* Migrate to the requested node */
4233 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
4235 page_nid
= target_nid
;
4236 flags
|= TNF_MIGRATED
;
4238 flags
|= TNF_MIGRATE_FAIL
;
4241 if (page_nid
!= NUMA_NO_NODE
)
4242 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
4246 static inline vm_fault_t
create_huge_pmd(struct vm_fault
*vmf
)
4248 if (vma_is_anonymous(vmf
->vma
))
4249 return do_huge_pmd_anonymous_page(vmf
);
4250 if (vmf
->vma
->vm_ops
->huge_fault
)
4251 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
4252 return VM_FAULT_FALLBACK
;
4255 /* `inline' is required to avoid gcc 4.1.2 build error */
4256 static inline vm_fault_t
wp_huge_pmd(struct vm_fault
*vmf
, pmd_t orig_pmd
)
4258 if (vma_is_anonymous(vmf
->vma
)) {
4259 if (userfaultfd_huge_pmd_wp(vmf
->vma
, orig_pmd
))
4260 return handle_userfault(vmf
, VM_UFFD_WP
);
4261 return do_huge_pmd_wp_page(vmf
, orig_pmd
);
4263 if (vmf
->vma
->vm_ops
->huge_fault
) {
4264 vm_fault_t ret
= vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
4266 if (!(ret
& VM_FAULT_FALLBACK
))
4270 /* COW or write-notify handled on pte level: split pmd. */
4271 __split_huge_pmd(vmf
->vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
4273 return VM_FAULT_FALLBACK
;
4276 static vm_fault_t
create_huge_pud(struct vm_fault
*vmf
)
4278 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4279 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4280 /* No support for anonymous transparent PUD pages yet */
4281 if (vma_is_anonymous(vmf
->vma
))
4283 if (vmf
->vma
->vm_ops
->huge_fault
) {
4284 vm_fault_t ret
= vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
4286 if (!(ret
& VM_FAULT_FALLBACK
))
4290 /* COW or write-notify not handled on PUD level: split pud.*/
4291 __split_huge_pud(vmf
->vma
, vmf
->pud
, vmf
->address
);
4292 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4293 return VM_FAULT_FALLBACK
;
4296 static vm_fault_t
wp_huge_pud(struct vm_fault
*vmf
, pud_t orig_pud
)
4298 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4299 /* No support for anonymous transparent PUD pages yet */
4300 if (vma_is_anonymous(vmf
->vma
))
4301 return VM_FAULT_FALLBACK
;
4302 if (vmf
->vma
->vm_ops
->huge_fault
)
4303 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
4304 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4305 return VM_FAULT_FALLBACK
;
4309 * These routines also need to handle stuff like marking pages dirty
4310 * and/or accessed for architectures that don't do it in hardware (most
4311 * RISC architectures). The early dirtying is also good on the i386.
4313 * There is also a hook called "update_mmu_cache()" that architectures
4314 * with external mmu caches can use to update those (ie the Sparc or
4315 * PowerPC hashed page tables that act as extended TLBs).
4317 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4318 * concurrent faults).
4320 * The mmap_lock may have been released depending on flags and our return value.
4321 * See filemap_fault() and __lock_page_or_retry().
4323 static vm_fault_t
handle_pte_fault(struct vm_fault
*vmf
)
4327 if (unlikely(pmd_none(*vmf
->pmd
))) {
4329 * Leave __pte_alloc() until later: because vm_ops->fault may
4330 * want to allocate huge page, and if we expose page table
4331 * for an instant, it will be difficult to retract from
4332 * concurrent faults and from rmap lookups.
4336 /* See comment in pte_alloc_one_map() */
4337 if (pmd_devmap_trans_unstable(vmf
->pmd
))
4340 * A regular pmd is established and it can't morph into a huge
4341 * pmd from under us anymore at this point because we hold the
4342 * mmap_lock read mode and khugepaged takes it in write mode.
4343 * So now it's safe to run pte_offset_map().
4345 vmf
->pte
= pte_offset_map(vmf
->pmd
, vmf
->address
);
4346 vmf
->orig_pte
= *vmf
->pte
;
4349 * some architectures can have larger ptes than wordsize,
4350 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4351 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4352 * accesses. The code below just needs a consistent view
4353 * for the ifs and we later double check anyway with the
4354 * ptl lock held. So here a barrier will do.
4357 if (pte_none(vmf
->orig_pte
)) {
4358 pte_unmap(vmf
->pte
);
4364 if (vma_is_anonymous(vmf
->vma
))
4365 return do_anonymous_page(vmf
);
4367 return do_fault(vmf
);
4370 if (!pte_present(vmf
->orig_pte
))
4371 return do_swap_page(vmf
);
4373 if (pte_protnone(vmf
->orig_pte
) && vma_is_accessible(vmf
->vma
))
4374 return do_numa_page(vmf
);
4376 vmf
->ptl
= pte_lockptr(vmf
->vma
->vm_mm
, vmf
->pmd
);
4377 spin_lock(vmf
->ptl
);
4378 entry
= vmf
->orig_pte
;
4379 if (unlikely(!pte_same(*vmf
->pte
, entry
))) {
4380 update_mmu_tlb(vmf
->vma
, vmf
->address
, vmf
->pte
);
4383 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
4384 if (!pte_write(entry
))
4385 return do_wp_page(vmf
);
4386 entry
= pte_mkdirty(entry
);
4388 entry
= pte_mkyoung(entry
);
4389 if (ptep_set_access_flags(vmf
->vma
, vmf
->address
, vmf
->pte
, entry
,
4390 vmf
->flags
& FAULT_FLAG_WRITE
)) {
4391 update_mmu_cache(vmf
->vma
, vmf
->address
, vmf
->pte
);
4393 /* Skip spurious TLB flush for retried page fault */
4394 if (vmf
->flags
& FAULT_FLAG_TRIED
)
4397 * This is needed only for protection faults but the arch code
4398 * is not yet telling us if this is a protection fault or not.
4399 * This still avoids useless tlb flushes for .text page faults
4402 if (vmf
->flags
& FAULT_FLAG_WRITE
)
4403 flush_tlb_fix_spurious_fault(vmf
->vma
, vmf
->address
);
4406 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
4411 * By the time we get here, we already hold the mm semaphore
4413 * The mmap_lock may have been released depending on flags and our
4414 * return value. See filemap_fault() and __lock_page_or_retry().
4416 static vm_fault_t
__handle_mm_fault(struct vm_area_struct
*vma
,
4417 unsigned long address
, unsigned int flags
)
4419 struct vm_fault vmf
= {
4421 .address
= address
& PAGE_MASK
,
4423 .pgoff
= linear_page_index(vma
, address
),
4424 .gfp_mask
= __get_fault_gfp_mask(vma
),
4426 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
4427 struct mm_struct
*mm
= vma
->vm_mm
;
4432 pgd
= pgd_offset(mm
, address
);
4433 p4d
= p4d_alloc(mm
, pgd
, address
);
4435 return VM_FAULT_OOM
;
4437 vmf
.pud
= pud_alloc(mm
, p4d
, address
);
4439 return VM_FAULT_OOM
;
4441 if (pud_none(*vmf
.pud
) && __transparent_hugepage_enabled(vma
)) {
4442 ret
= create_huge_pud(&vmf
);
4443 if (!(ret
& VM_FAULT_FALLBACK
))
4446 pud_t orig_pud
= *vmf
.pud
;
4449 if (pud_trans_huge(orig_pud
) || pud_devmap(orig_pud
)) {
4451 /* NUMA case for anonymous PUDs would go here */
4453 if (dirty
&& !pud_write(orig_pud
)) {
4454 ret
= wp_huge_pud(&vmf
, orig_pud
);
4455 if (!(ret
& VM_FAULT_FALLBACK
))
4458 huge_pud_set_accessed(&vmf
, orig_pud
);
4464 vmf
.pmd
= pmd_alloc(mm
, vmf
.pud
, address
);
4466 return VM_FAULT_OOM
;
4468 /* Huge pud page fault raced with pmd_alloc? */
4469 if (pud_trans_unstable(vmf
.pud
))
4472 if (pmd_none(*vmf
.pmd
) && __transparent_hugepage_enabled(vma
)) {
4473 ret
= create_huge_pmd(&vmf
);
4474 if (!(ret
& VM_FAULT_FALLBACK
))
4477 pmd_t orig_pmd
= *vmf
.pmd
;
4480 if (unlikely(is_swap_pmd(orig_pmd
))) {
4481 VM_BUG_ON(thp_migration_supported() &&
4482 !is_pmd_migration_entry(orig_pmd
));
4483 if (is_pmd_migration_entry(orig_pmd
))
4484 pmd_migration_entry_wait(mm
, vmf
.pmd
);
4487 if (pmd_trans_huge(orig_pmd
) || pmd_devmap(orig_pmd
)) {
4488 if (pmd_protnone(orig_pmd
) && vma_is_accessible(vma
))
4489 return do_huge_pmd_numa_page(&vmf
, orig_pmd
);
4491 if (dirty
&& !pmd_write(orig_pmd
)) {
4492 ret
= wp_huge_pmd(&vmf
, orig_pmd
);
4493 if (!(ret
& VM_FAULT_FALLBACK
))
4496 huge_pmd_set_accessed(&vmf
, orig_pmd
);
4502 return handle_pte_fault(&vmf
);
4506 * mm_account_fault - Do page fault accountings
4508 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4509 * of perf event counters, but we'll still do the per-task accounting to
4510 * the task who triggered this page fault.
4511 * @address: the faulted address.
4512 * @flags: the fault flags.
4513 * @ret: the fault retcode.
4515 * This will take care of most of the page fault accountings. Meanwhile, it
4516 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4517 * updates. However note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4518 * still be in per-arch page fault handlers at the entry of page fault.
4520 static inline void mm_account_fault(struct pt_regs
*regs
,
4521 unsigned long address
, unsigned int flags
,
4527 * We don't do accounting for some specific faults:
4529 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4530 * includes arch_vma_access_permitted() failing before reaching here.
4531 * So this is not a "this many hardware page faults" counter. We
4532 * should use the hw profiling for that.
4534 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4535 * once they're completed.
4537 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_RETRY
))
4541 * We define the fault as a major fault when the final successful fault
4542 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4543 * handle it immediately previously).
4545 major
= (ret
& VM_FAULT_MAJOR
) || (flags
& FAULT_FLAG_TRIED
);
4553 * If the fault is done for GUP, regs will be NULL. We only do the
4554 * accounting for the per thread fault counters who triggered the
4555 * fault, and we skip the perf event updates.
4561 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ
, 1, regs
, address
);
4563 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN
, 1, regs
, address
);
4567 * By the time we get here, we already hold the mm semaphore
4569 * The mmap_lock may have been released depending on flags and our
4570 * return value. See filemap_fault() and __lock_page_or_retry().
4572 vm_fault_t
handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
4573 unsigned int flags
, struct pt_regs
*regs
)
4577 __set_current_state(TASK_RUNNING
);
4579 count_vm_event(PGFAULT
);
4580 count_memcg_event_mm(vma
->vm_mm
, PGFAULT
);
4582 /* do counter updates before entering really critical section. */
4583 check_sync_rss_stat(current
);
4585 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
4586 flags
& FAULT_FLAG_INSTRUCTION
,
4587 flags
& FAULT_FLAG_REMOTE
))
4588 return VM_FAULT_SIGSEGV
;
4591 * Enable the memcg OOM handling for faults triggered in user
4592 * space. Kernel faults are handled more gracefully.
4594 if (flags
& FAULT_FLAG_USER
)
4595 mem_cgroup_enter_user_fault();
4597 if (unlikely(is_vm_hugetlb_page(vma
)))
4598 ret
= hugetlb_fault(vma
->vm_mm
, vma
, address
, flags
);
4600 ret
= __handle_mm_fault(vma
, address
, flags
);
4602 if (flags
& FAULT_FLAG_USER
) {
4603 mem_cgroup_exit_user_fault();
4605 * The task may have entered a memcg OOM situation but
4606 * if the allocation error was handled gracefully (no
4607 * VM_FAULT_OOM), there is no need to kill anything.
4608 * Just clean up the OOM state peacefully.
4610 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
4611 mem_cgroup_oom_synchronize(false);
4614 mm_account_fault(regs
, address
, flags
, ret
);
4618 EXPORT_SYMBOL_GPL(handle_mm_fault
);
4620 #ifndef __PAGETABLE_P4D_FOLDED
4622 * Allocate p4d page table.
4623 * We've already handled the fast-path in-line.
4625 int __p4d_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
4627 p4d_t
*new = p4d_alloc_one(mm
, address
);
4631 smp_wmb(); /* See comment in __pte_alloc */
4633 spin_lock(&mm
->page_table_lock
);
4634 if (pgd_present(*pgd
)) /* Another has populated it */
4637 pgd_populate(mm
, pgd
, new);
4638 spin_unlock(&mm
->page_table_lock
);
4641 #endif /* __PAGETABLE_P4D_FOLDED */
4643 #ifndef __PAGETABLE_PUD_FOLDED
4645 * Allocate page upper directory.
4646 * We've already handled the fast-path in-line.
4648 int __pud_alloc(struct mm_struct
*mm
, p4d_t
*p4d
, unsigned long address
)
4650 pud_t
*new = pud_alloc_one(mm
, address
);
4654 smp_wmb(); /* See comment in __pte_alloc */
4656 spin_lock(&mm
->page_table_lock
);
4657 if (!p4d_present(*p4d
)) {
4659 p4d_populate(mm
, p4d
, new);
4660 } else /* Another has populated it */
4662 spin_unlock(&mm
->page_table_lock
);
4665 #endif /* __PAGETABLE_PUD_FOLDED */
4667 #ifndef __PAGETABLE_PMD_FOLDED
4669 * Allocate page middle directory.
4670 * We've already handled the fast-path in-line.
4672 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
4675 pmd_t
*new = pmd_alloc_one(mm
, address
);
4679 smp_wmb(); /* See comment in __pte_alloc */
4681 ptl
= pud_lock(mm
, pud
);
4682 if (!pud_present(*pud
)) {
4684 pud_populate(mm
, pud
, new);
4685 } else /* Another has populated it */
4690 #endif /* __PAGETABLE_PMD_FOLDED */
4692 static int __follow_pte_pmd(struct mm_struct
*mm
, unsigned long address
,
4693 struct mmu_notifier_range
*range
,
4694 pte_t
**ptepp
, pmd_t
**pmdpp
, spinlock_t
**ptlp
)
4702 pgd
= pgd_offset(mm
, address
);
4703 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
4706 p4d
= p4d_offset(pgd
, address
);
4707 if (p4d_none(*p4d
) || unlikely(p4d_bad(*p4d
)))
4710 pud
= pud_offset(p4d
, address
);
4711 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
4714 pmd
= pmd_offset(pud
, address
);
4715 VM_BUG_ON(pmd_trans_huge(*pmd
));
4717 if (pmd_huge(*pmd
)) {
4722 mmu_notifier_range_init(range
, MMU_NOTIFY_CLEAR
, 0,
4723 NULL
, mm
, address
& PMD_MASK
,
4724 (address
& PMD_MASK
) + PMD_SIZE
);
4725 mmu_notifier_invalidate_range_start(range
);
4727 *ptlp
= pmd_lock(mm
, pmd
);
4728 if (pmd_huge(*pmd
)) {
4734 mmu_notifier_invalidate_range_end(range
);
4737 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
4741 mmu_notifier_range_init(range
, MMU_NOTIFY_CLEAR
, 0, NULL
, mm
,
4742 address
& PAGE_MASK
,
4743 (address
& PAGE_MASK
) + PAGE_SIZE
);
4744 mmu_notifier_invalidate_range_start(range
);
4746 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
4747 if (!pte_present(*ptep
))
4752 pte_unmap_unlock(ptep
, *ptlp
);
4754 mmu_notifier_invalidate_range_end(range
);
4759 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
4760 pte_t
**ptepp
, spinlock_t
**ptlp
)
4764 /* (void) is needed to make gcc happy */
4765 (void) __cond_lock(*ptlp
,
4766 !(res
= __follow_pte_pmd(mm
, address
, NULL
,
4767 ptepp
, NULL
, ptlp
)));
4771 int follow_pte_pmd(struct mm_struct
*mm
, unsigned long address
,
4772 struct mmu_notifier_range
*range
,
4773 pte_t
**ptepp
, pmd_t
**pmdpp
, spinlock_t
**ptlp
)
4777 /* (void) is needed to make gcc happy */
4778 (void) __cond_lock(*ptlp
,
4779 !(res
= __follow_pte_pmd(mm
, address
, range
,
4780 ptepp
, pmdpp
, ptlp
)));
4783 EXPORT_SYMBOL(follow_pte_pmd
);
4786 * follow_pfn - look up PFN at a user virtual address
4787 * @vma: memory mapping
4788 * @address: user virtual address
4789 * @pfn: location to store found PFN
4791 * Only IO mappings and raw PFN mappings are allowed.
4793 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4795 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
4802 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4805 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
4808 *pfn
= pte_pfn(*ptep
);
4809 pte_unmap_unlock(ptep
, ptl
);
4812 EXPORT_SYMBOL(follow_pfn
);
4814 #ifdef CONFIG_HAVE_IOREMAP_PROT
4815 int follow_phys(struct vm_area_struct
*vma
,
4816 unsigned long address
, unsigned int flags
,
4817 unsigned long *prot
, resource_size_t
*phys
)
4823 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4826 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
4830 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
4833 *prot
= pgprot_val(pte_pgprot(pte
));
4834 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
4838 pte_unmap_unlock(ptep
, ptl
);
4843 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
4844 void *buf
, int len
, int write
)
4846 resource_size_t phys_addr
;
4847 unsigned long prot
= 0;
4848 void __iomem
*maddr
;
4849 int offset
= addr
& (PAGE_SIZE
-1);
4851 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
4854 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
4859 memcpy_toio(maddr
+ offset
, buf
, len
);
4861 memcpy_fromio(buf
, maddr
+ offset
, len
);
4866 EXPORT_SYMBOL_GPL(generic_access_phys
);
4870 * Access another process' address space as given in mm. If non-NULL, use the
4871 * given task for page fault accounting.
4873 int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
4874 unsigned long addr
, void *buf
, int len
, unsigned int gup_flags
)
4876 struct vm_area_struct
*vma
;
4877 void *old_buf
= buf
;
4878 int write
= gup_flags
& FOLL_WRITE
;
4880 if (mmap_read_lock_killable(mm
))
4883 /* ignore errors, just check how much was successfully transferred */
4885 int bytes
, ret
, offset
;
4887 struct page
*page
= NULL
;
4889 ret
= get_user_pages_remote(mm
, addr
, 1,
4890 gup_flags
, &page
, &vma
, NULL
);
4892 #ifndef CONFIG_HAVE_IOREMAP_PROT
4896 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4897 * we can access using slightly different code.
4899 vma
= find_vma(mm
, addr
);
4900 if (!vma
|| vma
->vm_start
> addr
)
4902 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
4903 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
4911 offset
= addr
& (PAGE_SIZE
-1);
4912 if (bytes
> PAGE_SIZE
-offset
)
4913 bytes
= PAGE_SIZE
-offset
;
4917 copy_to_user_page(vma
, page
, addr
,
4918 maddr
+ offset
, buf
, bytes
);
4919 set_page_dirty_lock(page
);
4921 copy_from_user_page(vma
, page
, addr
,
4922 buf
, maddr
+ offset
, bytes
);
4931 mmap_read_unlock(mm
);
4933 return buf
- old_buf
;
4937 * access_remote_vm - access another process' address space
4938 * @mm: the mm_struct of the target address space
4939 * @addr: start address to access
4940 * @buf: source or destination buffer
4941 * @len: number of bytes to transfer
4942 * @gup_flags: flags modifying lookup behaviour
4944 * The caller must hold a reference on @mm.
4946 * Return: number of bytes copied from source to destination.
4948 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
4949 void *buf
, int len
, unsigned int gup_flags
)
4951 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, gup_flags
);
4955 * Access another process' address space.
4956 * Source/target buffer must be kernel space,
4957 * Do not walk the page table directly, use get_user_pages
4959 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
4960 void *buf
, int len
, unsigned int gup_flags
)
4962 struct mm_struct
*mm
;
4965 mm
= get_task_mm(tsk
);
4969 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, gup_flags
);
4975 EXPORT_SYMBOL_GPL(access_process_vm
);
4978 * Print the name of a VMA.
4980 void print_vma_addr(char *prefix
, unsigned long ip
)
4982 struct mm_struct
*mm
= current
->mm
;
4983 struct vm_area_struct
*vma
;
4986 * we might be running from an atomic context so we cannot sleep
4988 if (!mmap_read_trylock(mm
))
4991 vma
= find_vma(mm
, ip
);
4992 if (vma
&& vma
->vm_file
) {
4993 struct file
*f
= vma
->vm_file
;
4994 char *buf
= (char *)__get_free_page(GFP_NOWAIT
);
4998 p
= file_path(f
, buf
, PAGE_SIZE
);
5001 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
5003 vma
->vm_end
- vma
->vm_start
);
5004 free_page((unsigned long)buf
);
5007 mmap_read_unlock(mm
);
5010 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5011 void __might_fault(const char *file
, int line
)
5014 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5015 * holding the mmap_lock, this is safe because kernel memory doesn't
5016 * get paged out, therefore we'll never actually fault, and the
5017 * below annotations will generate false positives.
5019 if (uaccess_kernel())
5021 if (pagefault_disabled())
5023 __might_sleep(file
, line
, 0);
5024 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5026 might_lock_read(¤t
->mm
->mmap_lock
);
5029 EXPORT_SYMBOL(__might_fault
);
5032 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5034 * Process all subpages of the specified huge page with the specified
5035 * operation. The target subpage will be processed last to keep its
5038 static inline void process_huge_page(
5039 unsigned long addr_hint
, unsigned int pages_per_huge_page
,
5040 void (*process_subpage
)(unsigned long addr
, int idx
, void *arg
),
5044 unsigned long addr
= addr_hint
&
5045 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5047 /* Process target subpage last to keep its cache lines hot */
5049 n
= (addr_hint
- addr
) / PAGE_SIZE
;
5050 if (2 * n
<= pages_per_huge_page
) {
5051 /* If target subpage in first half of huge page */
5054 /* Process subpages at the end of huge page */
5055 for (i
= pages_per_huge_page
- 1; i
>= 2 * n
; i
--) {
5057 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
5060 /* If target subpage in second half of huge page */
5061 base
= pages_per_huge_page
- 2 * (pages_per_huge_page
- n
);
5062 l
= pages_per_huge_page
- n
;
5063 /* Process subpages at the begin of huge page */
5064 for (i
= 0; i
< base
; i
++) {
5066 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
5070 * Process remaining subpages in left-right-left-right pattern
5071 * towards the target subpage
5073 for (i
= 0; i
< l
; i
++) {
5074 int left_idx
= base
+ i
;
5075 int right_idx
= base
+ 2 * l
- 1 - i
;
5078 process_subpage(addr
+ left_idx
* PAGE_SIZE
, left_idx
, arg
);
5080 process_subpage(addr
+ right_idx
* PAGE_SIZE
, right_idx
, arg
);
5084 static void clear_gigantic_page(struct page
*page
,
5086 unsigned int pages_per_huge_page
)
5089 struct page
*p
= page
;
5092 for (i
= 0; i
< pages_per_huge_page
;
5093 i
++, p
= mem_map_next(p
, page
, i
)) {
5095 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
5099 static void clear_subpage(unsigned long addr
, int idx
, void *arg
)
5101 struct page
*page
= arg
;
5103 clear_user_highpage(page
+ idx
, addr
);
5106 void clear_huge_page(struct page
*page
,
5107 unsigned long addr_hint
, unsigned int pages_per_huge_page
)
5109 unsigned long addr
= addr_hint
&
5110 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5112 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
5113 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
5117 process_huge_page(addr_hint
, pages_per_huge_page
, clear_subpage
, page
);
5120 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
5122 struct vm_area_struct
*vma
,
5123 unsigned int pages_per_huge_page
)
5126 struct page
*dst_base
= dst
;
5127 struct page
*src_base
= src
;
5129 for (i
= 0; i
< pages_per_huge_page
; ) {
5131 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
5134 dst
= mem_map_next(dst
, dst_base
, i
);
5135 src
= mem_map_next(src
, src_base
, i
);
5139 struct copy_subpage_arg
{
5142 struct vm_area_struct
*vma
;
5145 static void copy_subpage(unsigned long addr
, int idx
, void *arg
)
5147 struct copy_subpage_arg
*copy_arg
= arg
;
5149 copy_user_highpage(copy_arg
->dst
+ idx
, copy_arg
->src
+ idx
,
5150 addr
, copy_arg
->vma
);
5153 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
5154 unsigned long addr_hint
, struct vm_area_struct
*vma
,
5155 unsigned int pages_per_huge_page
)
5157 unsigned long addr
= addr_hint
&
5158 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
5159 struct copy_subpage_arg arg
= {
5165 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
5166 copy_user_gigantic_page(dst
, src
, addr
, vma
,
5167 pages_per_huge_page
);
5171 process_huge_page(addr_hint
, pages_per_huge_page
, copy_subpage
, &arg
);
5174 long copy_huge_page_from_user(struct page
*dst_page
,
5175 const void __user
*usr_src
,
5176 unsigned int pages_per_huge_page
,
5177 bool allow_pagefault
)
5179 void *src
= (void *)usr_src
;
5181 unsigned long i
, rc
= 0;
5182 unsigned long ret_val
= pages_per_huge_page
* PAGE_SIZE
;
5184 for (i
= 0; i
< pages_per_huge_page
; i
++) {
5185 if (allow_pagefault
)
5186 page_kaddr
= kmap(dst_page
+ i
);
5188 page_kaddr
= kmap_atomic(dst_page
+ i
);
5189 rc
= copy_from_user(page_kaddr
,
5190 (const void __user
*)(src
+ i
* PAGE_SIZE
),
5192 if (allow_pagefault
)
5193 kunmap(dst_page
+ i
);
5195 kunmap_atomic(page_kaddr
);
5197 ret_val
-= (PAGE_SIZE
- rc
);
5205 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5207 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5209 static struct kmem_cache
*page_ptl_cachep
;
5211 void __init
ptlock_cache_init(void)
5213 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
5217 bool ptlock_alloc(struct page
*page
)
5221 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
5228 void ptlock_free(struct page
*page
)
5230 kmem_cache_free(page_ptl_cachep
, page
->ptl
);