1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * demand-loading started 01.12.91 - seems it is high on the list of
10 * things wanted, and it should be easy to implement. - Linus
14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
15 * pages started 02.12.91, seems to work. - Linus.
17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
18 * would have taken more than the 6M I have free, but it worked well as
21 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
25 * Real VM (paging to/from disk) started 18.12.91. Much more work and
26 * thought has to go into this. Oh, well..
27 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
28 * Found it. Everything seems to work now.
29 * 20.12.91 - Ok, making the swap-device changeable like the root.
33 * 05.04.94 - Multi-page memory management added for v1.1.
34 * Idea by Alex Bligh (alex@cconcepts.co.uk)
36 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
37 * (Gerhard.Wichert@pdb.siemens.de)
39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
42 #include <linux/kernel_stat.h>
44 #include <linux/sched/mm.h>
45 #include <linux/sched/coredump.h>
46 #include <linux/sched/numa_balancing.h>
47 #include <linux/sched/task.h>
48 #include <linux/hugetlb.h>
49 #include <linux/mman.h>
50 #include <linux/swap.h>
51 #include <linux/highmem.h>
52 #include <linux/pagemap.h>
53 #include <linux/memremap.h>
54 #include <linux/ksm.h>
55 #include <linux/rmap.h>
56 #include <linux/export.h>
57 #include <linux/delayacct.h>
58 #include <linux/init.h>
59 #include <linux/pfn_t.h>
60 #include <linux/writeback.h>
61 #include <linux/memcontrol.h>
62 #include <linux/mmu_notifier.h>
63 #include <linux/swapops.h>
64 #include <linux/elf.h>
65 #include <linux/gfp.h>
66 #include <linux/migrate.h>
67 #include <linux/string.h>
68 #include <linux/dma-debug.h>
69 #include <linux/debugfs.h>
70 #include <linux/userfaultfd_k.h>
71 #include <linux/dax.h>
72 #include <linux/oom.h>
73 #include <linux/numa.h>
76 #include <asm/mmu_context.h>
77 #include <asm/pgalloc.h>
78 #include <linux/uaccess.h>
80 #include <asm/tlbflush.h>
81 #include <asm/pgtable.h>
85 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
86 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
89 #ifndef CONFIG_NEED_MULTIPLE_NODES
90 /* use the per-pgdat data instead for discontigmem - mbligh */
91 unsigned long max_mapnr
;
92 EXPORT_SYMBOL(max_mapnr
);
95 EXPORT_SYMBOL(mem_map
);
99 * A number of key systems in x86 including ioremap() rely on the assumption
100 * that high_memory defines the upper bound on direct map memory, then end
101 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
102 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
106 EXPORT_SYMBOL(high_memory
);
109 * Randomize the address space (stacks, mmaps, brk, etc.).
111 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
112 * as ancient (libc5 based) binaries can segfault. )
114 int randomize_va_space __read_mostly
=
115 #ifdef CONFIG_COMPAT_BRK
121 static int __init
disable_randmaps(char *s
)
123 randomize_va_space
= 0;
126 __setup("norandmaps", disable_randmaps
);
128 unsigned long zero_pfn __read_mostly
;
129 EXPORT_SYMBOL(zero_pfn
);
131 unsigned long highest_memmap_pfn __read_mostly
;
134 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
136 static int __init
init_zero_pfn(void)
138 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
141 core_initcall(init_zero_pfn
);
144 #if defined(SPLIT_RSS_COUNTING)
146 void sync_mm_rss(struct mm_struct
*mm
)
150 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
151 if (current
->rss_stat
.count
[i
]) {
152 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
153 current
->rss_stat
.count
[i
] = 0;
156 current
->rss_stat
.events
= 0;
159 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
161 struct task_struct
*task
= current
;
163 if (likely(task
->mm
== mm
))
164 task
->rss_stat
.count
[member
] += val
;
166 add_mm_counter(mm
, member
, val
);
168 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
169 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
171 /* sync counter once per 64 page faults */
172 #define TASK_RSS_EVENTS_THRESH (64)
173 static void check_sync_rss_stat(struct task_struct
*task
)
175 if (unlikely(task
!= current
))
177 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
178 sync_mm_rss(task
->mm
);
180 #else /* SPLIT_RSS_COUNTING */
182 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
183 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
185 static void check_sync_rss_stat(struct task_struct
*task
)
189 #endif /* SPLIT_RSS_COUNTING */
192 * Note: this doesn't free the actual pages themselves. That
193 * has been handled earlier when unmapping all the memory regions.
195 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
198 pgtable_t token
= pmd_pgtable(*pmd
);
200 pte_free_tlb(tlb
, token
, addr
);
201 mm_dec_nr_ptes(tlb
->mm
);
204 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
205 unsigned long addr
, unsigned long end
,
206 unsigned long floor
, unsigned long ceiling
)
213 pmd
= pmd_offset(pud
, addr
);
215 next
= pmd_addr_end(addr
, end
);
216 if (pmd_none_or_clear_bad(pmd
))
218 free_pte_range(tlb
, pmd
, addr
);
219 } while (pmd
++, addr
= next
, addr
!= end
);
229 if (end
- 1 > ceiling
- 1)
232 pmd
= pmd_offset(pud
, start
);
234 pmd_free_tlb(tlb
, pmd
, start
);
235 mm_dec_nr_pmds(tlb
->mm
);
238 static inline void free_pud_range(struct mmu_gather
*tlb
, p4d_t
*p4d
,
239 unsigned long addr
, unsigned long end
,
240 unsigned long floor
, unsigned long ceiling
)
247 pud
= pud_offset(p4d
, addr
);
249 next
= pud_addr_end(addr
, end
);
250 if (pud_none_or_clear_bad(pud
))
252 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
253 } while (pud
++, addr
= next
, addr
!= end
);
263 if (end
- 1 > ceiling
- 1)
266 pud
= pud_offset(p4d
, start
);
268 pud_free_tlb(tlb
, pud
, start
);
269 mm_dec_nr_puds(tlb
->mm
);
272 static inline void free_p4d_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
273 unsigned long addr
, unsigned long end
,
274 unsigned long floor
, unsigned long ceiling
)
281 p4d
= p4d_offset(pgd
, addr
);
283 next
= p4d_addr_end(addr
, end
);
284 if (p4d_none_or_clear_bad(p4d
))
286 free_pud_range(tlb
, p4d
, addr
, next
, floor
, ceiling
);
287 } while (p4d
++, addr
= next
, addr
!= end
);
293 ceiling
&= PGDIR_MASK
;
297 if (end
- 1 > ceiling
- 1)
300 p4d
= p4d_offset(pgd
, start
);
302 p4d_free_tlb(tlb
, p4d
, start
);
306 * This function frees user-level page tables of a process.
308 void free_pgd_range(struct mmu_gather
*tlb
,
309 unsigned long addr
, unsigned long end
,
310 unsigned long floor
, unsigned long ceiling
)
316 * The next few lines have given us lots of grief...
318 * Why are we testing PMD* at this top level? Because often
319 * there will be no work to do at all, and we'd prefer not to
320 * go all the way down to the bottom just to discover that.
322 * Why all these "- 1"s? Because 0 represents both the bottom
323 * of the address space and the top of it (using -1 for the
324 * top wouldn't help much: the masks would do the wrong thing).
325 * The rule is that addr 0 and floor 0 refer to the bottom of
326 * the address space, but end 0 and ceiling 0 refer to the top
327 * Comparisons need to use "end - 1" and "ceiling - 1" (though
328 * that end 0 case should be mythical).
330 * Wherever addr is brought up or ceiling brought down, we must
331 * be careful to reject "the opposite 0" before it confuses the
332 * subsequent tests. But what about where end is brought down
333 * by PMD_SIZE below? no, end can't go down to 0 there.
335 * Whereas we round start (addr) and ceiling down, by different
336 * masks at different levels, in order to test whether a table
337 * now has no other vmas using it, so can be freed, we don't
338 * bother to round floor or end up - the tests don't need that.
352 if (end
- 1 > ceiling
- 1)
357 * We add page table cache pages with PAGE_SIZE,
358 * (see pte_free_tlb()), flush the tlb if we need
360 tlb_change_page_size(tlb
, PAGE_SIZE
);
361 pgd
= pgd_offset(tlb
->mm
, addr
);
363 next
= pgd_addr_end(addr
, end
);
364 if (pgd_none_or_clear_bad(pgd
))
366 free_p4d_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
367 } while (pgd
++, addr
= next
, addr
!= end
);
370 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
371 unsigned long floor
, unsigned long ceiling
)
374 struct vm_area_struct
*next
= vma
->vm_next
;
375 unsigned long addr
= vma
->vm_start
;
378 * Hide vma from rmap and truncate_pagecache before freeing
381 unlink_anon_vmas(vma
);
382 unlink_file_vma(vma
);
384 if (is_vm_hugetlb_page(vma
)) {
385 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
386 floor
, next
? next
->vm_start
: ceiling
);
389 * Optimization: gather nearby vmas into one call down
391 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
392 && !is_vm_hugetlb_page(next
)) {
395 unlink_anon_vmas(vma
);
396 unlink_file_vma(vma
);
398 free_pgd_range(tlb
, addr
, vma
->vm_end
,
399 floor
, next
? next
->vm_start
: ceiling
);
405 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
)
408 pgtable_t
new = pte_alloc_one(mm
);
413 * Ensure all pte setup (eg. pte page lock and page clearing) are
414 * visible before the pte is made visible to other CPUs by being
415 * put into page tables.
417 * The other side of the story is the pointer chasing in the page
418 * table walking code (when walking the page table without locking;
419 * ie. most of the time). Fortunately, these data accesses consist
420 * of a chain of data-dependent loads, meaning most CPUs (alpha
421 * being the notable exception) will already guarantee loads are
422 * seen in-order. See the alpha page table accessors for the
423 * smp_read_barrier_depends() barriers in page table walking code.
425 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
427 ptl
= pmd_lock(mm
, pmd
);
428 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
430 pmd_populate(mm
, pmd
, new);
439 int __pte_alloc_kernel(pmd_t
*pmd
)
441 pte_t
*new = pte_alloc_one_kernel(&init_mm
);
445 smp_wmb(); /* See comment in __pte_alloc */
447 spin_lock(&init_mm
.page_table_lock
);
448 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
449 pmd_populate_kernel(&init_mm
, pmd
, new);
452 spin_unlock(&init_mm
.page_table_lock
);
454 pte_free_kernel(&init_mm
, new);
458 static inline void init_rss_vec(int *rss
)
460 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
463 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
467 if (current
->mm
== mm
)
469 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
471 add_mm_counter(mm
, i
, rss
[i
]);
475 * This function is called to print an error when a bad pte
476 * is found. For example, we might have a PFN-mapped pte in
477 * a region that doesn't allow it.
479 * The calling function must still handle the error.
481 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
482 pte_t pte
, struct page
*page
)
484 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
485 p4d_t
*p4d
= p4d_offset(pgd
, addr
);
486 pud_t
*pud
= pud_offset(p4d
, addr
);
487 pmd_t
*pmd
= pmd_offset(pud
, addr
);
488 struct address_space
*mapping
;
490 static unsigned long resume
;
491 static unsigned long nr_shown
;
492 static unsigned long nr_unshown
;
495 * Allow a burst of 60 reports, then keep quiet for that minute;
496 * or allow a steady drip of one report per second.
498 if (nr_shown
== 60) {
499 if (time_before(jiffies
, resume
)) {
504 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
511 resume
= jiffies
+ 60 * HZ
;
513 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
514 index
= linear_page_index(vma
, addr
);
516 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
518 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
520 dump_page(page
, "bad pte");
521 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
522 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
523 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
525 vma
->vm_ops
? vma
->vm_ops
->fault
: NULL
,
526 vma
->vm_file
? vma
->vm_file
->f_op
->mmap
: NULL
,
527 mapping
? mapping
->a_ops
->readpage
: NULL
);
529 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
533 * vm_normal_page -- This function gets the "struct page" associated with a pte.
535 * "Special" mappings do not wish to be associated with a "struct page" (either
536 * it doesn't exist, or it exists but they don't want to touch it). In this
537 * case, NULL is returned here. "Normal" mappings do have a struct page.
539 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
540 * pte bit, in which case this function is trivial. Secondly, an architecture
541 * may not have a spare pte bit, which requires a more complicated scheme,
544 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
545 * special mapping (even if there are underlying and valid "struct pages").
546 * COWed pages of a VM_PFNMAP are always normal.
548 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
549 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
550 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
551 * mapping will always honor the rule
553 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
555 * And for normal mappings this is false.
557 * This restricts such mappings to be a linear translation from virtual address
558 * to pfn. To get around this restriction, we allow arbitrary mappings so long
559 * as the vma is not a COW mapping; in that case, we know that all ptes are
560 * special (because none can have been COWed).
563 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
565 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
566 * page" backing, however the difference is that _all_ pages with a struct
567 * page (that is, those where pfn_valid is true) are refcounted and considered
568 * normal pages by the VM. The disadvantage is that pages are refcounted
569 * (which can be slower and simply not an option for some PFNMAP users). The
570 * advantage is that we don't have to follow the strict linearity rule of
571 * PFNMAP mappings in order to support COWable mappings.
574 struct page
*_vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
575 pte_t pte
, bool with_public_device
)
577 unsigned long pfn
= pte_pfn(pte
);
579 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
)) {
580 if (likely(!pte_special(pte
)))
582 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
583 return vma
->vm_ops
->find_special_page(vma
, addr
);
584 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
586 if (is_zero_pfn(pfn
))
590 * Device public pages are special pages (they are ZONE_DEVICE
591 * pages but different from persistent memory). They behave
592 * allmost like normal pages. The difference is that they are
593 * not on the lru and thus should never be involve with any-
594 * thing that involve lru manipulation (mlock, numa balancing,
597 * This is why we still want to return NULL for such page from
598 * vm_normal_page() so that we do not have to special case all
599 * call site of vm_normal_page().
601 if (likely(pfn
<= highest_memmap_pfn
)) {
602 struct page
*page
= pfn_to_page(pfn
);
604 if (is_device_public_page(page
)) {
605 if (with_public_device
)
614 print_bad_pte(vma
, addr
, pte
, NULL
);
618 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
620 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
621 if (vma
->vm_flags
& VM_MIXEDMAP
) {
627 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
628 if (pfn
== vma
->vm_pgoff
+ off
)
630 if (!is_cow_mapping(vma
->vm_flags
))
635 if (is_zero_pfn(pfn
))
639 if (unlikely(pfn
> highest_memmap_pfn
)) {
640 print_bad_pte(vma
, addr
, pte
, NULL
);
645 * NOTE! We still have PageReserved() pages in the page tables.
646 * eg. VDSO mappings can cause them to exist.
649 return pfn_to_page(pfn
);
652 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
653 struct page
*vm_normal_page_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
656 unsigned long pfn
= pmd_pfn(pmd
);
659 * There is no pmd_special() but there may be special pmds, e.g.
660 * in a direct-access (dax) mapping, so let's just replicate the
661 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
663 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
664 if (vma
->vm_flags
& VM_MIXEDMAP
) {
670 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
671 if (pfn
== vma
->vm_pgoff
+ off
)
673 if (!is_cow_mapping(vma
->vm_flags
))
680 if (is_zero_pfn(pfn
))
682 if (unlikely(pfn
> highest_memmap_pfn
))
686 * NOTE! We still have PageReserved() pages in the page tables.
687 * eg. VDSO mappings can cause them to exist.
690 return pfn_to_page(pfn
);
695 * copy one vm_area from one task to the other. Assumes the page tables
696 * already present in the new task to be cleared in the whole range
697 * covered by this vma.
700 static inline unsigned long
701 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
702 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
703 unsigned long addr
, int *rss
)
705 unsigned long vm_flags
= vma
->vm_flags
;
706 pte_t pte
= *src_pte
;
709 /* pte contains position in swap or file, so copy. */
710 if (unlikely(!pte_present(pte
))) {
711 swp_entry_t entry
= pte_to_swp_entry(pte
);
713 if (likely(!non_swap_entry(entry
))) {
714 if (swap_duplicate(entry
) < 0)
717 /* make sure dst_mm is on swapoff's mmlist. */
718 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
719 spin_lock(&mmlist_lock
);
720 if (list_empty(&dst_mm
->mmlist
))
721 list_add(&dst_mm
->mmlist
,
723 spin_unlock(&mmlist_lock
);
726 } else if (is_migration_entry(entry
)) {
727 page
= migration_entry_to_page(entry
);
729 rss
[mm_counter(page
)]++;
731 if (is_write_migration_entry(entry
) &&
732 is_cow_mapping(vm_flags
)) {
734 * COW mappings require pages in both
735 * parent and child to be set to read.
737 make_migration_entry_read(&entry
);
738 pte
= swp_entry_to_pte(entry
);
739 if (pte_swp_soft_dirty(*src_pte
))
740 pte
= pte_swp_mksoft_dirty(pte
);
741 set_pte_at(src_mm
, addr
, src_pte
, pte
);
743 } else if (is_device_private_entry(entry
)) {
744 page
= device_private_entry_to_page(entry
);
747 * Update rss count even for unaddressable pages, as
748 * they should treated just like normal pages in this
751 * We will likely want to have some new rss counters
752 * for unaddressable pages, at some point. But for now
753 * keep things as they are.
756 rss
[mm_counter(page
)]++;
757 page_dup_rmap(page
, false);
760 * We do not preserve soft-dirty information, because so
761 * far, checkpoint/restore is the only feature that
762 * requires that. And checkpoint/restore does not work
763 * when a device driver is involved (you cannot easily
764 * save and restore device driver state).
766 if (is_write_device_private_entry(entry
) &&
767 is_cow_mapping(vm_flags
)) {
768 make_device_private_entry_read(&entry
);
769 pte
= swp_entry_to_pte(entry
);
770 set_pte_at(src_mm
, addr
, src_pte
, pte
);
777 * If it's a COW mapping, write protect it both
778 * in the parent and the child
780 if (is_cow_mapping(vm_flags
) && pte_write(pte
)) {
781 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
782 pte
= pte_wrprotect(pte
);
786 * If it's a shared mapping, mark it clean in
789 if (vm_flags
& VM_SHARED
)
790 pte
= pte_mkclean(pte
);
791 pte
= pte_mkold(pte
);
793 page
= vm_normal_page(vma
, addr
, pte
);
796 page_dup_rmap(page
, false);
797 rss
[mm_counter(page
)]++;
798 } else if (pte_devmap(pte
)) {
799 page
= pte_page(pte
);
802 * Cache coherent device memory behave like regular page and
803 * not like persistent memory page. For more informations see
804 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
806 if (is_device_public_page(page
)) {
808 page_dup_rmap(page
, false);
809 rss
[mm_counter(page
)]++;
814 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
818 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
819 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
820 unsigned long addr
, unsigned long end
)
822 pte_t
*orig_src_pte
, *orig_dst_pte
;
823 pte_t
*src_pte
, *dst_pte
;
824 spinlock_t
*src_ptl
, *dst_ptl
;
826 int rss
[NR_MM_COUNTERS
];
827 swp_entry_t entry
= (swp_entry_t
){0};
832 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
835 src_pte
= pte_offset_map(src_pmd
, addr
);
836 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
837 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
838 orig_src_pte
= src_pte
;
839 orig_dst_pte
= dst_pte
;
840 arch_enter_lazy_mmu_mode();
844 * We are holding two locks at this point - either of them
845 * could generate latencies in another task on another CPU.
847 if (progress
>= 32) {
849 if (need_resched() ||
850 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
853 if (pte_none(*src_pte
)) {
857 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
862 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
864 arch_leave_lazy_mmu_mode();
865 spin_unlock(src_ptl
);
866 pte_unmap(orig_src_pte
);
867 add_mm_rss_vec(dst_mm
, rss
);
868 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
872 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
881 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
882 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
883 unsigned long addr
, unsigned long end
)
885 pmd_t
*src_pmd
, *dst_pmd
;
888 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
891 src_pmd
= pmd_offset(src_pud
, addr
);
893 next
= pmd_addr_end(addr
, end
);
894 if (is_swap_pmd(*src_pmd
) || pmd_trans_huge(*src_pmd
)
895 || pmd_devmap(*src_pmd
)) {
897 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PMD_SIZE
, vma
);
898 err
= copy_huge_pmd(dst_mm
, src_mm
,
899 dst_pmd
, src_pmd
, addr
, vma
);
906 if (pmd_none_or_clear_bad(src_pmd
))
908 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
911 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
915 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
916 p4d_t
*dst_p4d
, p4d_t
*src_p4d
, struct vm_area_struct
*vma
,
917 unsigned long addr
, unsigned long end
)
919 pud_t
*src_pud
, *dst_pud
;
922 dst_pud
= pud_alloc(dst_mm
, dst_p4d
, addr
);
925 src_pud
= pud_offset(src_p4d
, addr
);
927 next
= pud_addr_end(addr
, end
);
928 if (pud_trans_huge(*src_pud
) || pud_devmap(*src_pud
)) {
931 VM_BUG_ON_VMA(next
-addr
!= HPAGE_PUD_SIZE
, vma
);
932 err
= copy_huge_pud(dst_mm
, src_mm
,
933 dst_pud
, src_pud
, addr
, vma
);
940 if (pud_none_or_clear_bad(src_pud
))
942 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
945 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
949 static inline int copy_p4d_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
950 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
951 unsigned long addr
, unsigned long end
)
953 p4d_t
*src_p4d
, *dst_p4d
;
956 dst_p4d
= p4d_alloc(dst_mm
, dst_pgd
, addr
);
959 src_p4d
= p4d_offset(src_pgd
, addr
);
961 next
= p4d_addr_end(addr
, end
);
962 if (p4d_none_or_clear_bad(src_p4d
))
964 if (copy_pud_range(dst_mm
, src_mm
, dst_p4d
, src_p4d
,
967 } while (dst_p4d
++, src_p4d
++, addr
= next
, addr
!= end
);
971 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
972 struct vm_area_struct
*vma
)
974 pgd_t
*src_pgd
, *dst_pgd
;
976 unsigned long addr
= vma
->vm_start
;
977 unsigned long end
= vma
->vm_end
;
978 struct mmu_notifier_range range
;
983 * Don't copy ptes where a page fault will fill them correctly.
984 * Fork becomes much lighter when there are big shared or private
985 * readonly mappings. The tradeoff is that copy_page_range is more
986 * efficient than faulting.
988 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_PFNMAP
| VM_MIXEDMAP
)) &&
992 if (is_vm_hugetlb_page(vma
))
993 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
995 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
997 * We do not free on error cases below as remove_vma
998 * gets called on error from higher level routine
1000 ret
= track_pfn_copy(vma
);
1006 * We need to invalidate the secondary MMU mappings only when
1007 * there could be a permission downgrade on the ptes of the
1008 * parent mm. And a permission downgrade will only happen if
1009 * is_cow_mapping() returns true.
1011 is_cow
= is_cow_mapping(vma
->vm_flags
);
1014 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_PAGE
,
1015 0, vma
, src_mm
, addr
, end
);
1016 mmu_notifier_invalidate_range_start(&range
);
1020 dst_pgd
= pgd_offset(dst_mm
, addr
);
1021 src_pgd
= pgd_offset(src_mm
, addr
);
1023 next
= pgd_addr_end(addr
, end
);
1024 if (pgd_none_or_clear_bad(src_pgd
))
1026 if (unlikely(copy_p4d_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1027 vma
, addr
, next
))) {
1031 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1034 mmu_notifier_invalidate_range_end(&range
);
1038 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1039 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1040 unsigned long addr
, unsigned long end
,
1041 struct zap_details
*details
)
1043 struct mm_struct
*mm
= tlb
->mm
;
1044 int force_flush
= 0;
1045 int rss
[NR_MM_COUNTERS
];
1051 tlb_change_page_size(tlb
, PAGE_SIZE
);
1054 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1056 flush_tlb_batched_pending(mm
);
1057 arch_enter_lazy_mmu_mode();
1060 if (pte_none(ptent
))
1063 if (pte_present(ptent
)) {
1066 page
= _vm_normal_page(vma
, addr
, ptent
, true);
1067 if (unlikely(details
) && page
) {
1069 * unmap_shared_mapping_pages() wants to
1070 * invalidate cache without truncating:
1071 * unmap shared but keep private pages.
1073 if (details
->check_mapping
&&
1074 details
->check_mapping
!= page_rmapping(page
))
1077 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1079 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1080 if (unlikely(!page
))
1083 if (!PageAnon(page
)) {
1084 if (pte_dirty(ptent
)) {
1086 set_page_dirty(page
);
1088 if (pte_young(ptent
) &&
1089 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1090 mark_page_accessed(page
);
1092 rss
[mm_counter(page
)]--;
1093 page_remove_rmap(page
, false);
1094 if (unlikely(page_mapcount(page
) < 0))
1095 print_bad_pte(vma
, addr
, ptent
, page
);
1096 if (unlikely(__tlb_remove_page(tlb
, page
))) {
1104 entry
= pte_to_swp_entry(ptent
);
1105 if (non_swap_entry(entry
) && is_device_private_entry(entry
)) {
1106 struct page
*page
= device_private_entry_to_page(entry
);
1108 if (unlikely(details
&& details
->check_mapping
)) {
1110 * unmap_shared_mapping_pages() wants to
1111 * invalidate cache without truncating:
1112 * unmap shared but keep private pages.
1114 if (details
->check_mapping
!=
1115 page_rmapping(page
))
1119 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1120 rss
[mm_counter(page
)]--;
1121 page_remove_rmap(page
, false);
1126 /* If details->check_mapping, we leave swap entries. */
1127 if (unlikely(details
))
1130 entry
= pte_to_swp_entry(ptent
);
1131 if (!non_swap_entry(entry
))
1133 else if (is_migration_entry(entry
)) {
1136 page
= migration_entry_to_page(entry
);
1137 rss
[mm_counter(page
)]--;
1139 if (unlikely(!free_swap_and_cache(entry
)))
1140 print_bad_pte(vma
, addr
, ptent
, NULL
);
1141 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1142 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1144 add_mm_rss_vec(mm
, rss
);
1145 arch_leave_lazy_mmu_mode();
1147 /* Do the actual TLB flush before dropping ptl */
1149 tlb_flush_mmu_tlbonly(tlb
);
1150 pte_unmap_unlock(start_pte
, ptl
);
1153 * If we forced a TLB flush (either due to running out of
1154 * batch buffers or because we needed to flush dirty TLB
1155 * entries before releasing the ptl), free the batched
1156 * memory too. Restart if we didn't do everything.
1168 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1169 struct vm_area_struct
*vma
, pud_t
*pud
,
1170 unsigned long addr
, unsigned long end
,
1171 struct zap_details
*details
)
1176 pmd
= pmd_offset(pud
, addr
);
1178 next
= pmd_addr_end(addr
, end
);
1179 if (is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1180 if (next
- addr
!= HPAGE_PMD_SIZE
)
1181 __split_huge_pmd(vma
, pmd
, addr
, false, NULL
);
1182 else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1187 * Here there can be other concurrent MADV_DONTNEED or
1188 * trans huge page faults running, and if the pmd is
1189 * none or trans huge it can change under us. This is
1190 * because MADV_DONTNEED holds the mmap_sem in read
1193 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1195 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1198 } while (pmd
++, addr
= next
, addr
!= end
);
1203 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1204 struct vm_area_struct
*vma
, p4d_t
*p4d
,
1205 unsigned long addr
, unsigned long end
,
1206 struct zap_details
*details
)
1211 pud
= pud_offset(p4d
, addr
);
1213 next
= pud_addr_end(addr
, end
);
1214 if (pud_trans_huge(*pud
) || pud_devmap(*pud
)) {
1215 if (next
- addr
!= HPAGE_PUD_SIZE
) {
1216 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb
->mm
->mmap_sem
), vma
);
1217 split_huge_pud(vma
, pud
, addr
);
1218 } else if (zap_huge_pud(tlb
, vma
, pud
, addr
))
1222 if (pud_none_or_clear_bad(pud
))
1224 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1227 } while (pud
++, addr
= next
, addr
!= end
);
1232 static inline unsigned long zap_p4d_range(struct mmu_gather
*tlb
,
1233 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1234 unsigned long addr
, unsigned long end
,
1235 struct zap_details
*details
)
1240 p4d
= p4d_offset(pgd
, addr
);
1242 next
= p4d_addr_end(addr
, end
);
1243 if (p4d_none_or_clear_bad(p4d
))
1245 next
= zap_pud_range(tlb
, vma
, p4d
, addr
, next
, details
);
1246 } while (p4d
++, addr
= next
, addr
!= end
);
1251 void unmap_page_range(struct mmu_gather
*tlb
,
1252 struct vm_area_struct
*vma
,
1253 unsigned long addr
, unsigned long end
,
1254 struct zap_details
*details
)
1259 BUG_ON(addr
>= end
);
1260 tlb_start_vma(tlb
, vma
);
1261 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1263 next
= pgd_addr_end(addr
, end
);
1264 if (pgd_none_or_clear_bad(pgd
))
1266 next
= zap_p4d_range(tlb
, vma
, pgd
, addr
, next
, details
);
1267 } while (pgd
++, addr
= next
, addr
!= end
);
1268 tlb_end_vma(tlb
, vma
);
1272 static void unmap_single_vma(struct mmu_gather
*tlb
,
1273 struct vm_area_struct
*vma
, unsigned long start_addr
,
1274 unsigned long end_addr
,
1275 struct zap_details
*details
)
1277 unsigned long start
= max(vma
->vm_start
, start_addr
);
1280 if (start
>= vma
->vm_end
)
1282 end
= min(vma
->vm_end
, end_addr
);
1283 if (end
<= vma
->vm_start
)
1287 uprobe_munmap(vma
, start
, end
);
1289 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1290 untrack_pfn(vma
, 0, 0);
1293 if (unlikely(is_vm_hugetlb_page(vma
))) {
1295 * It is undesirable to test vma->vm_file as it
1296 * should be non-null for valid hugetlb area.
1297 * However, vm_file will be NULL in the error
1298 * cleanup path of mmap_region. When
1299 * hugetlbfs ->mmap method fails,
1300 * mmap_region() nullifies vma->vm_file
1301 * before calling this function to clean up.
1302 * Since no pte has actually been setup, it is
1303 * safe to do nothing in this case.
1306 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
1307 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1308 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
1311 unmap_page_range(tlb
, vma
, start
, end
, details
);
1316 * unmap_vmas - unmap a range of memory covered by a list of vma's
1317 * @tlb: address of the caller's struct mmu_gather
1318 * @vma: the starting vma
1319 * @start_addr: virtual address at which to start unmapping
1320 * @end_addr: virtual address at which to end unmapping
1322 * Unmap all pages in the vma list.
1324 * Only addresses between `start' and `end' will be unmapped.
1326 * The VMA list must be sorted in ascending virtual address order.
1328 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1329 * range after unmap_vmas() returns. So the only responsibility here is to
1330 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1331 * drops the lock and schedules.
1333 void unmap_vmas(struct mmu_gather
*tlb
,
1334 struct vm_area_struct
*vma
, unsigned long start_addr
,
1335 unsigned long end_addr
)
1337 struct mmu_notifier_range range
;
1339 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
, vma
->vm_mm
,
1340 start_addr
, end_addr
);
1341 mmu_notifier_invalidate_range_start(&range
);
1342 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1343 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1344 mmu_notifier_invalidate_range_end(&range
);
1348 * zap_page_range - remove user pages in a given range
1349 * @vma: vm_area_struct holding the applicable pages
1350 * @start: starting address of pages to zap
1351 * @size: number of bytes to zap
1353 * Caller must protect the VMA list
1355 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1358 struct mmu_notifier_range range
;
1359 struct mmu_gather tlb
;
1362 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1363 start
, start
+ size
);
1364 tlb_gather_mmu(&tlb
, vma
->vm_mm
, start
, range
.end
);
1365 update_hiwater_rss(vma
->vm_mm
);
1366 mmu_notifier_invalidate_range_start(&range
);
1367 for ( ; vma
&& vma
->vm_start
< range
.end
; vma
= vma
->vm_next
)
1368 unmap_single_vma(&tlb
, vma
, start
, range
.end
, NULL
);
1369 mmu_notifier_invalidate_range_end(&range
);
1370 tlb_finish_mmu(&tlb
, start
, range
.end
);
1374 * zap_page_range_single - remove user pages in a given range
1375 * @vma: vm_area_struct holding the applicable pages
1376 * @address: starting address of pages to zap
1377 * @size: number of bytes to zap
1378 * @details: details of shared cache invalidation
1380 * The range must fit into one VMA.
1382 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1383 unsigned long size
, struct zap_details
*details
)
1385 struct mmu_notifier_range range
;
1386 struct mmu_gather tlb
;
1389 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1390 address
, address
+ size
);
1391 tlb_gather_mmu(&tlb
, vma
->vm_mm
, address
, range
.end
);
1392 update_hiwater_rss(vma
->vm_mm
);
1393 mmu_notifier_invalidate_range_start(&range
);
1394 unmap_single_vma(&tlb
, vma
, address
, range
.end
, details
);
1395 mmu_notifier_invalidate_range_end(&range
);
1396 tlb_finish_mmu(&tlb
, address
, range
.end
);
1400 * zap_vma_ptes - remove ptes mapping the vma
1401 * @vma: vm_area_struct holding ptes to be zapped
1402 * @address: starting address of pages to zap
1403 * @size: number of bytes to zap
1405 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1407 * The entire address range must be fully contained within the vma.
1410 void zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1413 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1414 !(vma
->vm_flags
& VM_PFNMAP
))
1417 zap_page_range_single(vma
, address
, size
, NULL
);
1419 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1421 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1429 pgd
= pgd_offset(mm
, addr
);
1430 p4d
= p4d_alloc(mm
, pgd
, addr
);
1433 pud
= pud_alloc(mm
, p4d
, addr
);
1436 pmd
= pmd_alloc(mm
, pud
, addr
);
1440 VM_BUG_ON(pmd_trans_huge(*pmd
));
1441 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1445 * This is the old fallback for page remapping.
1447 * For historical reasons, it only allows reserved pages. Only
1448 * old drivers should use this, and they needed to mark their
1449 * pages reserved for the old functions anyway.
1451 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1452 struct page
*page
, pgprot_t prot
)
1454 struct mm_struct
*mm
= vma
->vm_mm
;
1460 if (PageAnon(page
) || PageSlab(page
) || page_has_type(page
))
1463 flush_dcache_page(page
);
1464 pte
= get_locked_pte(mm
, addr
, &ptl
);
1468 if (!pte_none(*pte
))
1471 /* Ok, finally just insert the thing.. */
1473 inc_mm_counter_fast(mm
, mm_counter_file(page
));
1474 page_add_file_rmap(page
, false);
1475 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1479 pte_unmap_unlock(pte
, ptl
);
1485 * vm_insert_page - insert single page into user vma
1486 * @vma: user vma to map to
1487 * @addr: target user address of this page
1488 * @page: source kernel page
1490 * This allows drivers to insert individual pages they've allocated
1493 * The page has to be a nice clean _individual_ kernel allocation.
1494 * If you allocate a compound page, you need to have marked it as
1495 * such (__GFP_COMP), or manually just split the page up yourself
1496 * (see split_page()).
1498 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1499 * took an arbitrary page protection parameter. This doesn't allow
1500 * that. Your vma protection will have to be set up correctly, which
1501 * means that if you want a shared writable mapping, you'd better
1502 * ask for a shared writable mapping!
1504 * The page does not need to be reserved.
1506 * Usually this function is called from f_op->mmap() handler
1507 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1508 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1509 * function from other places, for example from page-fault handler.
1511 * Return: %0 on success, negative error code otherwise.
1513 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1516 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1518 if (!page_count(page
))
1520 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1521 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
1522 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1523 vma
->vm_flags
|= VM_MIXEDMAP
;
1525 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1527 EXPORT_SYMBOL(vm_insert_page
);
1530 * __vm_map_pages - maps range of kernel pages into user vma
1531 * @vma: user vma to map to
1532 * @pages: pointer to array of source kernel pages
1533 * @num: number of pages in page array
1534 * @offset: user's requested vm_pgoff
1536 * This allows drivers to map range of kernel pages into a user vma.
1538 * Return: 0 on success and error code otherwise.
1540 static int __vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
1541 unsigned long num
, unsigned long offset
)
1543 unsigned long count
= vma_pages(vma
);
1544 unsigned long uaddr
= vma
->vm_start
;
1547 /* Fail if the user requested offset is beyond the end of the object */
1551 /* Fail if the user requested size exceeds available object size */
1552 if (count
> num
- offset
)
1555 for (i
= 0; i
< count
; i
++) {
1556 ret
= vm_insert_page(vma
, uaddr
, pages
[offset
+ i
]);
1566 * vm_map_pages - maps range of kernel pages starts with non zero offset
1567 * @vma: user vma to map to
1568 * @pages: pointer to array of source kernel pages
1569 * @num: number of pages in page array
1571 * Maps an object consisting of @num pages, catering for the user's
1572 * requested vm_pgoff
1574 * If we fail to insert any page into the vma, the function will return
1575 * immediately leaving any previously inserted pages present. Callers
1576 * from the mmap handler may immediately return the error as their caller
1577 * will destroy the vma, removing any successfully inserted pages. Other
1578 * callers should make their own arrangements for calling unmap_region().
1580 * Context: Process context. Called by mmap handlers.
1581 * Return: 0 on success and error code otherwise.
1583 int vm_map_pages(struct vm_area_struct
*vma
, struct page
**pages
,
1586 return __vm_map_pages(vma
, pages
, num
, vma
->vm_pgoff
);
1588 EXPORT_SYMBOL(vm_map_pages
);
1591 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1592 * @vma: user vma to map to
1593 * @pages: pointer to array of source kernel pages
1594 * @num: number of pages in page array
1596 * Similar to vm_map_pages(), except that it explicitly sets the offset
1597 * to 0. This function is intended for the drivers that did not consider
1600 * Context: Process context. Called by mmap handlers.
1601 * Return: 0 on success and error code otherwise.
1603 int vm_map_pages_zero(struct vm_area_struct
*vma
, struct page
**pages
,
1606 return __vm_map_pages(vma
, pages
, num
, 0);
1608 EXPORT_SYMBOL(vm_map_pages_zero
);
1610 static vm_fault_t
insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1611 pfn_t pfn
, pgprot_t prot
, bool mkwrite
)
1613 struct mm_struct
*mm
= vma
->vm_mm
;
1617 pte
= get_locked_pte(mm
, addr
, &ptl
);
1619 return VM_FAULT_OOM
;
1620 if (!pte_none(*pte
)) {
1623 * For read faults on private mappings the PFN passed
1624 * in may not match the PFN we have mapped if the
1625 * mapped PFN is a writeable COW page. In the mkwrite
1626 * case we are creating a writable PTE for a shared
1627 * mapping and we expect the PFNs to match. If they
1628 * don't match, we are likely racing with block
1629 * allocation and mapping invalidation so just skip the
1632 if (pte_pfn(*pte
) != pfn_t_to_pfn(pfn
)) {
1633 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte
)));
1636 entry
= pte_mkyoung(*pte
);
1637 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1638 if (ptep_set_access_flags(vma
, addr
, pte
, entry
, 1))
1639 update_mmu_cache(vma
, addr
, pte
);
1644 /* Ok, finally just insert the thing.. */
1645 if (pfn_t_devmap(pfn
))
1646 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
1648 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
1651 entry
= pte_mkyoung(entry
);
1652 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1655 set_pte_at(mm
, addr
, pte
, entry
);
1656 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1659 pte_unmap_unlock(pte
, ptl
);
1660 return VM_FAULT_NOPAGE
;
1664 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1665 * @vma: user vma to map to
1666 * @addr: target user address of this page
1667 * @pfn: source kernel pfn
1668 * @pgprot: pgprot flags for the inserted page
1670 * This is exactly like vmf_insert_pfn(), except that it allows drivers to
1671 * to override pgprot on a per-page basis.
1673 * This only makes sense for IO mappings, and it makes no sense for
1674 * COW mappings. In general, using multiple vmas is preferable;
1675 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1678 * Context: Process context. May allocate using %GFP_KERNEL.
1679 * Return: vm_fault_t value.
1681 vm_fault_t
vmf_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
1682 unsigned long pfn
, pgprot_t pgprot
)
1685 * Technically, architectures with pte_special can avoid all these
1686 * restrictions (same for remap_pfn_range). However we would like
1687 * consistency in testing and feature parity among all, so we should
1688 * try to keep these invariants in place for everybody.
1690 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1691 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1692 (VM_PFNMAP
|VM_MIXEDMAP
));
1693 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1694 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1696 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1697 return VM_FAULT_SIGBUS
;
1699 if (!pfn_modify_allowed(pfn
, pgprot
))
1700 return VM_FAULT_SIGBUS
;
1702 track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
));
1704 return insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
,
1707 EXPORT_SYMBOL(vmf_insert_pfn_prot
);
1710 * vmf_insert_pfn - insert single pfn into user vma
1711 * @vma: user vma to map to
1712 * @addr: target user address of this page
1713 * @pfn: source kernel pfn
1715 * Similar to vm_insert_page, this allows drivers to insert individual pages
1716 * they've allocated into a user vma. Same comments apply.
1718 * This function should only be called from a vm_ops->fault handler, and
1719 * in that case the handler should return the result of this function.
1721 * vma cannot be a COW mapping.
1723 * As this is called only for pages that do not currently exist, we
1724 * do not need to flush old virtual caches or the TLB.
1726 * Context: Process context. May allocate using %GFP_KERNEL.
1727 * Return: vm_fault_t value.
1729 vm_fault_t
vmf_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1732 return vmf_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
1734 EXPORT_SYMBOL(vmf_insert_pfn
);
1736 static bool vm_mixed_ok(struct vm_area_struct
*vma
, pfn_t pfn
)
1738 /* these checks mirror the abort conditions in vm_normal_page */
1739 if (vma
->vm_flags
& VM_MIXEDMAP
)
1741 if (pfn_t_devmap(pfn
))
1743 if (pfn_t_special(pfn
))
1745 if (is_zero_pfn(pfn_t_to_pfn(pfn
)))
1750 static vm_fault_t
__vm_insert_mixed(struct vm_area_struct
*vma
,
1751 unsigned long addr
, pfn_t pfn
, bool mkwrite
)
1753 pgprot_t pgprot
= vma
->vm_page_prot
;
1756 BUG_ON(!vm_mixed_ok(vma
, pfn
));
1758 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1759 return VM_FAULT_SIGBUS
;
1761 track_pfn_insert(vma
, &pgprot
, pfn
);
1763 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn
), pgprot
))
1764 return VM_FAULT_SIGBUS
;
1767 * If we don't have pte special, then we have to use the pfn_valid()
1768 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1769 * refcount the page if pfn_valid is true (hence insert_page rather
1770 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1771 * without pte special, it would there be refcounted as a normal page.
1773 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL
) &&
1774 !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
1778 * At this point we are committed to insert_page()
1779 * regardless of whether the caller specified flags that
1780 * result in pfn_t_has_page() == false.
1782 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
1783 err
= insert_page(vma
, addr
, page
, pgprot
);
1785 return insert_pfn(vma
, addr
, pfn
, pgprot
, mkwrite
);
1789 return VM_FAULT_OOM
;
1790 if (err
< 0 && err
!= -EBUSY
)
1791 return VM_FAULT_SIGBUS
;
1793 return VM_FAULT_NOPAGE
;
1796 vm_fault_t
vmf_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1799 return __vm_insert_mixed(vma
, addr
, pfn
, false);
1801 EXPORT_SYMBOL(vmf_insert_mixed
);
1804 * If the insertion of PTE failed because someone else already added a
1805 * different entry in the mean time, we treat that as success as we assume
1806 * the same entry was actually inserted.
1808 vm_fault_t
vmf_insert_mixed_mkwrite(struct vm_area_struct
*vma
,
1809 unsigned long addr
, pfn_t pfn
)
1811 return __vm_insert_mixed(vma
, addr
, pfn
, true);
1813 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite
);
1816 * maps a range of physical memory into the requested pages. the old
1817 * mappings are removed. any references to nonexistent pages results
1818 * in null mappings (currently treated as "copy-on-access")
1820 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1821 unsigned long addr
, unsigned long end
,
1822 unsigned long pfn
, pgprot_t prot
)
1828 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1831 arch_enter_lazy_mmu_mode();
1833 BUG_ON(!pte_none(*pte
));
1834 if (!pfn_modify_allowed(pfn
, prot
)) {
1838 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
1840 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1841 arch_leave_lazy_mmu_mode();
1842 pte_unmap_unlock(pte
- 1, ptl
);
1846 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1847 unsigned long addr
, unsigned long end
,
1848 unsigned long pfn
, pgprot_t prot
)
1854 pfn
-= addr
>> PAGE_SHIFT
;
1855 pmd
= pmd_alloc(mm
, pud
, addr
);
1858 VM_BUG_ON(pmd_trans_huge(*pmd
));
1860 next
= pmd_addr_end(addr
, end
);
1861 err
= remap_pte_range(mm
, pmd
, addr
, next
,
1862 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1865 } while (pmd
++, addr
= next
, addr
!= end
);
1869 static inline int remap_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
1870 unsigned long addr
, unsigned long end
,
1871 unsigned long pfn
, pgprot_t prot
)
1877 pfn
-= addr
>> PAGE_SHIFT
;
1878 pud
= pud_alloc(mm
, p4d
, addr
);
1882 next
= pud_addr_end(addr
, end
);
1883 err
= remap_pmd_range(mm
, pud
, addr
, next
,
1884 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1887 } while (pud
++, addr
= next
, addr
!= end
);
1891 static inline int remap_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1892 unsigned long addr
, unsigned long end
,
1893 unsigned long pfn
, pgprot_t prot
)
1899 pfn
-= addr
>> PAGE_SHIFT
;
1900 p4d
= p4d_alloc(mm
, pgd
, addr
);
1904 next
= p4d_addr_end(addr
, end
);
1905 err
= remap_pud_range(mm
, p4d
, addr
, next
,
1906 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1909 } while (p4d
++, addr
= next
, addr
!= end
);
1914 * remap_pfn_range - remap kernel memory to userspace
1915 * @vma: user vma to map to
1916 * @addr: target user address to start at
1917 * @pfn: physical address of kernel memory
1918 * @size: size of map area
1919 * @prot: page protection flags for this mapping
1921 * Note: this is only safe if the mm semaphore is held when called.
1923 * Return: %0 on success, negative error code otherwise.
1925 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1926 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1930 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1931 struct mm_struct
*mm
= vma
->vm_mm
;
1932 unsigned long remap_pfn
= pfn
;
1936 * Physically remapped pages are special. Tell the
1937 * rest of the world about it:
1938 * VM_IO tells people not to look at these pages
1939 * (accesses can have side effects).
1940 * VM_PFNMAP tells the core MM that the base pages are just
1941 * raw PFN mappings, and do not have a "struct page" associated
1944 * Disable vma merging and expanding with mremap().
1946 * Omit vma from core dump, even when VM_IO turned off.
1948 * There's a horrible special case to handle copy-on-write
1949 * behaviour that some programs depend on. We mark the "original"
1950 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1951 * See vm_normal_page() for details.
1953 if (is_cow_mapping(vma
->vm_flags
)) {
1954 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
1956 vma
->vm_pgoff
= pfn
;
1959 err
= track_pfn_remap(vma
, &prot
, remap_pfn
, addr
, PAGE_ALIGN(size
));
1963 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
1965 BUG_ON(addr
>= end
);
1966 pfn
-= addr
>> PAGE_SHIFT
;
1967 pgd
= pgd_offset(mm
, addr
);
1968 flush_cache_range(vma
, addr
, end
);
1970 next
= pgd_addr_end(addr
, end
);
1971 err
= remap_p4d_range(mm
, pgd
, addr
, next
,
1972 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1975 } while (pgd
++, addr
= next
, addr
!= end
);
1978 untrack_pfn(vma
, remap_pfn
, PAGE_ALIGN(size
));
1982 EXPORT_SYMBOL(remap_pfn_range
);
1985 * vm_iomap_memory - remap memory to userspace
1986 * @vma: user vma to map to
1987 * @start: start of area
1988 * @len: size of area
1990 * This is a simplified io_remap_pfn_range() for common driver use. The
1991 * driver just needs to give us the physical memory range to be mapped,
1992 * we'll figure out the rest from the vma information.
1994 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1995 * whatever write-combining details or similar.
1997 * Return: %0 on success, negative error code otherwise.
1999 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2001 unsigned long vm_len
, pfn
, pages
;
2003 /* Check that the physical memory area passed in looks valid */
2004 if (start
+ len
< start
)
2007 * You *really* shouldn't map things that aren't page-aligned,
2008 * but we've historically allowed it because IO memory might
2009 * just have smaller alignment.
2011 len
+= start
& ~PAGE_MASK
;
2012 pfn
= start
>> PAGE_SHIFT
;
2013 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2014 if (pfn
+ pages
< pfn
)
2017 /* We start the mapping 'vm_pgoff' pages into the area */
2018 if (vma
->vm_pgoff
> pages
)
2020 pfn
+= vma
->vm_pgoff
;
2021 pages
-= vma
->vm_pgoff
;
2023 /* Can we fit all of the mapping? */
2024 vm_len
= vma
->vm_end
- vma
->vm_start
;
2025 if (vm_len
>> PAGE_SHIFT
> pages
)
2028 /* Ok, let it rip */
2029 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2031 EXPORT_SYMBOL(vm_iomap_memory
);
2033 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2034 unsigned long addr
, unsigned long end
,
2035 pte_fn_t fn
, void *data
)
2039 spinlock_t
*uninitialized_var(ptl
);
2041 pte
= (mm
== &init_mm
) ?
2042 pte_alloc_kernel(pmd
, addr
) :
2043 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2047 BUG_ON(pmd_huge(*pmd
));
2049 arch_enter_lazy_mmu_mode();
2052 err
= fn(pte
++, addr
, data
);
2055 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2057 arch_leave_lazy_mmu_mode();
2060 pte_unmap_unlock(pte
-1, ptl
);
2064 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2065 unsigned long addr
, unsigned long end
,
2066 pte_fn_t fn
, void *data
)
2072 BUG_ON(pud_huge(*pud
));
2074 pmd
= pmd_alloc(mm
, pud
, addr
);
2078 next
= pmd_addr_end(addr
, end
);
2079 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
2082 } while (pmd
++, addr
= next
, addr
!= end
);
2086 static int apply_to_pud_range(struct mm_struct
*mm
, p4d_t
*p4d
,
2087 unsigned long addr
, unsigned long end
,
2088 pte_fn_t fn
, void *data
)
2094 pud
= pud_alloc(mm
, p4d
, addr
);
2098 next
= pud_addr_end(addr
, end
);
2099 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
2102 } while (pud
++, addr
= next
, addr
!= end
);
2106 static int apply_to_p4d_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2107 unsigned long addr
, unsigned long end
,
2108 pte_fn_t fn
, void *data
)
2114 p4d
= p4d_alloc(mm
, pgd
, addr
);
2118 next
= p4d_addr_end(addr
, end
);
2119 err
= apply_to_pud_range(mm
, p4d
, addr
, next
, fn
, data
);
2122 } while (p4d
++, addr
= next
, addr
!= end
);
2127 * Scan a region of virtual memory, filling in page tables as necessary
2128 * and calling a provided function on each leaf page table.
2130 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2131 unsigned long size
, pte_fn_t fn
, void *data
)
2135 unsigned long end
= addr
+ size
;
2138 if (WARN_ON(addr
>= end
))
2141 pgd
= pgd_offset(mm
, addr
);
2143 next
= pgd_addr_end(addr
, end
);
2144 err
= apply_to_p4d_range(mm
, pgd
, addr
, next
, fn
, data
);
2147 } while (pgd
++, addr
= next
, addr
!= end
);
2151 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2154 * handle_pte_fault chooses page fault handler according to an entry which was
2155 * read non-atomically. Before making any commitment, on those architectures
2156 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2157 * parts, do_swap_page must check under lock before unmapping the pte and
2158 * proceeding (but do_wp_page is only called after already making such a check;
2159 * and do_anonymous_page can safely check later on).
2161 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2162 pte_t
*page_table
, pte_t orig_pte
)
2165 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2166 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2167 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2169 same
= pte_same(*page_table
, orig_pte
);
2173 pte_unmap(page_table
);
2177 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
2179 debug_dma_assert_idle(src
);
2182 * If the source page was a PFN mapping, we don't have
2183 * a "struct page" for it. We do a best-effort copy by
2184 * just copying from the original user address. If that
2185 * fails, we just zero-fill it. Live with it.
2187 if (unlikely(!src
)) {
2188 void *kaddr
= kmap_atomic(dst
);
2189 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
2192 * This really shouldn't fail, because the page is there
2193 * in the page tables. But it might just be unreadable,
2194 * in which case we just give up and fill the result with
2197 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
2199 kunmap_atomic(kaddr
);
2200 flush_dcache_page(dst
);
2202 copy_user_highpage(dst
, src
, va
, vma
);
2205 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
2207 struct file
*vm_file
= vma
->vm_file
;
2210 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
2213 * Special mappings (e.g. VDSO) do not have any file so fake
2214 * a default GFP_KERNEL for them.
2220 * Notify the address space that the page is about to become writable so that
2221 * it can prohibit this or wait for the page to get into an appropriate state.
2223 * We do this without the lock held, so that it can sleep if it needs to.
2225 static vm_fault_t
do_page_mkwrite(struct vm_fault
*vmf
)
2228 struct page
*page
= vmf
->page
;
2229 unsigned int old_flags
= vmf
->flags
;
2231 vmf
->flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2233 ret
= vmf
->vma
->vm_ops
->page_mkwrite(vmf
);
2234 /* Restore original flags so that caller is not surprised */
2235 vmf
->flags
= old_flags
;
2236 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2238 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
2240 if (!page
->mapping
) {
2242 return 0; /* retry */
2244 ret
|= VM_FAULT_LOCKED
;
2246 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2251 * Handle dirtying of a page in shared file mapping on a write fault.
2253 * The function expects the page to be locked and unlocks it.
2255 static void fault_dirty_shared_page(struct vm_area_struct
*vma
,
2258 struct address_space
*mapping
;
2260 bool page_mkwrite
= vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
;
2262 dirtied
= set_page_dirty(page
);
2263 VM_BUG_ON_PAGE(PageAnon(page
), page
);
2265 * Take a local copy of the address_space - page.mapping may be zeroed
2266 * by truncate after unlock_page(). The address_space itself remains
2267 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2268 * release semantics to prevent the compiler from undoing this copying.
2270 mapping
= page_rmapping(page
);
2273 if ((dirtied
|| page_mkwrite
) && mapping
) {
2275 * Some device drivers do not set page.mapping
2276 * but still dirty their pages
2278 balance_dirty_pages_ratelimited(mapping
);
2282 file_update_time(vma
->vm_file
);
2286 * Handle write page faults for pages that can be reused in the current vma
2288 * This can happen either due to the mapping being with the VM_SHARED flag,
2289 * or due to us being the last reference standing to the page. In either
2290 * case, all we need to do here is to mark the page as writable and update
2291 * any related book-keeping.
2293 static inline void wp_page_reuse(struct vm_fault
*vmf
)
2294 __releases(vmf
->ptl
)
2296 struct vm_area_struct
*vma
= vmf
->vma
;
2297 struct page
*page
= vmf
->page
;
2300 * Clear the pages cpupid information as the existing
2301 * information potentially belongs to a now completely
2302 * unrelated process.
2305 page_cpupid_xchg_last(page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2307 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2308 entry
= pte_mkyoung(vmf
->orig_pte
);
2309 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2310 if (ptep_set_access_flags(vma
, vmf
->address
, vmf
->pte
, entry
, 1))
2311 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2312 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2316 * Handle the case of a page which we actually need to copy to a new page.
2318 * Called with mmap_sem locked and the old page referenced, but
2319 * without the ptl held.
2321 * High level logic flow:
2323 * - Allocate a page, copy the content of the old page to the new one.
2324 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2325 * - Take the PTL. If the pte changed, bail out and release the allocated page
2326 * - If the pte is still the way we remember it, update the page table and all
2327 * relevant references. This includes dropping the reference the page-table
2328 * held to the old page, as well as updating the rmap.
2329 * - In any case, unlock the PTL and drop the reference we took to the old page.
2331 static vm_fault_t
wp_page_copy(struct vm_fault
*vmf
)
2333 struct vm_area_struct
*vma
= vmf
->vma
;
2334 struct mm_struct
*mm
= vma
->vm_mm
;
2335 struct page
*old_page
= vmf
->page
;
2336 struct page
*new_page
= NULL
;
2338 int page_copied
= 0;
2339 struct mem_cgroup
*memcg
;
2340 struct mmu_notifier_range range
;
2342 if (unlikely(anon_vma_prepare(vma
)))
2345 if (is_zero_pfn(pte_pfn(vmf
->orig_pte
))) {
2346 new_page
= alloc_zeroed_user_highpage_movable(vma
,
2351 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
2355 cow_user_page(new_page
, old_page
, vmf
->address
, vma
);
2358 if (mem_cgroup_try_charge_delay(new_page
, mm
, GFP_KERNEL
, &memcg
, false))
2361 __SetPageUptodate(new_page
);
2363 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
,
2364 vmf
->address
& PAGE_MASK
,
2365 (vmf
->address
& PAGE_MASK
) + PAGE_SIZE
);
2366 mmu_notifier_invalidate_range_start(&range
);
2369 * Re-check the pte - we dropped the lock
2371 vmf
->pte
= pte_offset_map_lock(mm
, vmf
->pmd
, vmf
->address
, &vmf
->ptl
);
2372 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
2374 if (!PageAnon(old_page
)) {
2375 dec_mm_counter_fast(mm
,
2376 mm_counter_file(old_page
));
2377 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2380 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2382 flush_cache_page(vma
, vmf
->address
, pte_pfn(vmf
->orig_pte
));
2383 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2384 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2386 * Clear the pte entry and flush it first, before updating the
2387 * pte with the new entry. This will avoid a race condition
2388 * seen in the presence of one thread doing SMC and another
2391 ptep_clear_flush_notify(vma
, vmf
->address
, vmf
->pte
);
2392 page_add_new_anon_rmap(new_page
, vma
, vmf
->address
, false);
2393 mem_cgroup_commit_charge(new_page
, memcg
, false, false);
2394 lru_cache_add_active_or_unevictable(new_page
, vma
);
2396 * We call the notify macro here because, when using secondary
2397 * mmu page tables (such as kvm shadow page tables), we want the
2398 * new page to be mapped directly into the secondary page table.
2400 set_pte_at_notify(mm
, vmf
->address
, vmf
->pte
, entry
);
2401 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2404 * Only after switching the pte to the new page may
2405 * we remove the mapcount here. Otherwise another
2406 * process may come and find the rmap count decremented
2407 * before the pte is switched to the new page, and
2408 * "reuse" the old page writing into it while our pte
2409 * here still points into it and can be read by other
2412 * The critical issue is to order this
2413 * page_remove_rmap with the ptp_clear_flush above.
2414 * Those stores are ordered by (if nothing else,)
2415 * the barrier present in the atomic_add_negative
2416 * in page_remove_rmap.
2418 * Then the TLB flush in ptep_clear_flush ensures that
2419 * no process can access the old page before the
2420 * decremented mapcount is visible. And the old page
2421 * cannot be reused until after the decremented
2422 * mapcount is visible. So transitively, TLBs to
2423 * old page will be flushed before it can be reused.
2425 page_remove_rmap(old_page
, false);
2428 /* Free the old page.. */
2429 new_page
= old_page
;
2432 mem_cgroup_cancel_charge(new_page
, memcg
, false);
2438 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2440 * No need to double call mmu_notifier->invalidate_range() callback as
2441 * the above ptep_clear_flush_notify() did already call it.
2443 mmu_notifier_invalidate_range_only_end(&range
);
2446 * Don't let another task, with possibly unlocked vma,
2447 * keep the mlocked page.
2449 if (page_copied
&& (vma
->vm_flags
& VM_LOCKED
)) {
2450 lock_page(old_page
); /* LRU manipulation */
2451 if (PageMlocked(old_page
))
2452 munlock_vma_page(old_page
);
2453 unlock_page(old_page
);
2457 return page_copied
? VM_FAULT_WRITE
: 0;
2463 return VM_FAULT_OOM
;
2467 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2468 * writeable once the page is prepared
2470 * @vmf: structure describing the fault
2472 * This function handles all that is needed to finish a write page fault in a
2473 * shared mapping due to PTE being read-only once the mapped page is prepared.
2474 * It handles locking of PTE and modifying it.
2476 * The function expects the page to be locked or other protection against
2477 * concurrent faults / writeback (such as DAX radix tree locks).
2479 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2480 * we acquired PTE lock.
2482 vm_fault_t
finish_mkwrite_fault(struct vm_fault
*vmf
)
2484 WARN_ON_ONCE(!(vmf
->vma
->vm_flags
& VM_SHARED
));
2485 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
2488 * We might have raced with another page fault while we released the
2489 * pte_offset_map_lock.
2491 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
2492 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2493 return VM_FAULT_NOPAGE
;
2500 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2503 static vm_fault_t
wp_pfn_shared(struct vm_fault
*vmf
)
2505 struct vm_area_struct
*vma
= vmf
->vma
;
2507 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
2510 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2511 vmf
->flags
|= FAULT_FLAG_MKWRITE
;
2512 ret
= vma
->vm_ops
->pfn_mkwrite(vmf
);
2513 if (ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))
2515 return finish_mkwrite_fault(vmf
);
2518 return VM_FAULT_WRITE
;
2521 static vm_fault_t
wp_page_shared(struct vm_fault
*vmf
)
2522 __releases(vmf
->ptl
)
2524 struct vm_area_struct
*vma
= vmf
->vma
;
2526 get_page(vmf
->page
);
2528 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2531 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2532 tmp
= do_page_mkwrite(vmf
);
2533 if (unlikely(!tmp
|| (tmp
&
2534 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
2535 put_page(vmf
->page
);
2538 tmp
= finish_mkwrite_fault(vmf
);
2539 if (unlikely(tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
2540 unlock_page(vmf
->page
);
2541 put_page(vmf
->page
);
2546 lock_page(vmf
->page
);
2548 fault_dirty_shared_page(vma
, vmf
->page
);
2549 put_page(vmf
->page
);
2551 return VM_FAULT_WRITE
;
2555 * This routine handles present pages, when users try to write
2556 * to a shared page. It is done by copying the page to a new address
2557 * and decrementing the shared-page counter for the old page.
2559 * Note that this routine assumes that the protection checks have been
2560 * done by the caller (the low-level page fault routine in most cases).
2561 * Thus we can safely just mark it writable once we've done any necessary
2564 * We also mark the page dirty at this point even though the page will
2565 * change only once the write actually happens. This avoids a few races,
2566 * and potentially makes it more efficient.
2568 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2569 * but allow concurrent faults), with pte both mapped and locked.
2570 * We return with mmap_sem still held, but pte unmapped and unlocked.
2572 static vm_fault_t
do_wp_page(struct vm_fault
*vmf
)
2573 __releases(vmf
->ptl
)
2575 struct vm_area_struct
*vma
= vmf
->vma
;
2577 vmf
->page
= vm_normal_page(vma
, vmf
->address
, vmf
->orig_pte
);
2580 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2583 * We should not cow pages in a shared writeable mapping.
2584 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2586 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2587 (VM_WRITE
|VM_SHARED
))
2588 return wp_pfn_shared(vmf
);
2590 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2591 return wp_page_copy(vmf
);
2595 * Take out anonymous pages first, anonymous shared vmas are
2596 * not dirty accountable.
2598 if (PageAnon(vmf
->page
)) {
2599 int total_map_swapcount
;
2600 if (PageKsm(vmf
->page
) && (PageSwapCache(vmf
->page
) ||
2601 page_count(vmf
->page
) != 1))
2603 if (!trylock_page(vmf
->page
)) {
2604 get_page(vmf
->page
);
2605 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2606 lock_page(vmf
->page
);
2607 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
2608 vmf
->address
, &vmf
->ptl
);
2609 if (!pte_same(*vmf
->pte
, vmf
->orig_pte
)) {
2610 unlock_page(vmf
->page
);
2611 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2612 put_page(vmf
->page
);
2615 put_page(vmf
->page
);
2617 if (PageKsm(vmf
->page
)) {
2618 bool reused
= reuse_ksm_page(vmf
->page
, vmf
->vma
,
2620 unlock_page(vmf
->page
);
2624 return VM_FAULT_WRITE
;
2626 if (reuse_swap_page(vmf
->page
, &total_map_swapcount
)) {
2627 if (total_map_swapcount
== 1) {
2629 * The page is all ours. Move it to
2630 * our anon_vma so the rmap code will
2631 * not search our parent or siblings.
2632 * Protected against the rmap code by
2635 page_move_anon_rmap(vmf
->page
, vma
);
2637 unlock_page(vmf
->page
);
2639 return VM_FAULT_WRITE
;
2641 unlock_page(vmf
->page
);
2642 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2643 (VM_WRITE
|VM_SHARED
))) {
2644 return wp_page_shared(vmf
);
2648 * Ok, we need to copy. Oh, well..
2650 get_page(vmf
->page
);
2652 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2653 return wp_page_copy(vmf
);
2656 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2657 unsigned long start_addr
, unsigned long end_addr
,
2658 struct zap_details
*details
)
2660 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2663 static inline void unmap_mapping_range_tree(struct rb_root_cached
*root
,
2664 struct zap_details
*details
)
2666 struct vm_area_struct
*vma
;
2667 pgoff_t vba
, vea
, zba
, zea
;
2669 vma_interval_tree_foreach(vma
, root
,
2670 details
->first_index
, details
->last_index
) {
2672 vba
= vma
->vm_pgoff
;
2673 vea
= vba
+ vma_pages(vma
) - 1;
2674 zba
= details
->first_index
;
2677 zea
= details
->last_index
;
2681 unmap_mapping_range_vma(vma
,
2682 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2683 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2689 * unmap_mapping_pages() - Unmap pages from processes.
2690 * @mapping: The address space containing pages to be unmapped.
2691 * @start: Index of first page to be unmapped.
2692 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2693 * @even_cows: Whether to unmap even private COWed pages.
2695 * Unmap the pages in this address space from any userspace process which
2696 * has them mmaped. Generally, you want to remove COWed pages as well when
2697 * a file is being truncated, but not when invalidating pages from the page
2700 void unmap_mapping_pages(struct address_space
*mapping
, pgoff_t start
,
2701 pgoff_t nr
, bool even_cows
)
2703 struct zap_details details
= { };
2705 details
.check_mapping
= even_cows
? NULL
: mapping
;
2706 details
.first_index
= start
;
2707 details
.last_index
= start
+ nr
- 1;
2708 if (details
.last_index
< details
.first_index
)
2709 details
.last_index
= ULONG_MAX
;
2711 i_mmap_lock_write(mapping
);
2712 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
.rb_root
)))
2713 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2714 i_mmap_unlock_write(mapping
);
2718 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2719 * address_space corresponding to the specified byte range in the underlying
2722 * @mapping: the address space containing mmaps to be unmapped.
2723 * @holebegin: byte in first page to unmap, relative to the start of
2724 * the underlying file. This will be rounded down to a PAGE_SIZE
2725 * boundary. Note that this is different from truncate_pagecache(), which
2726 * must keep the partial page. In contrast, we must get rid of
2728 * @holelen: size of prospective hole in bytes. This will be rounded
2729 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2731 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2732 * but 0 when invalidating pagecache, don't throw away private data.
2734 void unmap_mapping_range(struct address_space
*mapping
,
2735 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2737 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2738 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2740 /* Check for overflow. */
2741 if (sizeof(holelen
) > sizeof(hlen
)) {
2743 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2744 if (holeend
& ~(long long)ULONG_MAX
)
2745 hlen
= ULONG_MAX
- hba
+ 1;
2748 unmap_mapping_pages(mapping
, hba
, hlen
, even_cows
);
2750 EXPORT_SYMBOL(unmap_mapping_range
);
2753 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2754 * but allow concurrent faults), and pte mapped but not yet locked.
2755 * We return with pte unmapped and unlocked.
2757 * We return with the mmap_sem locked or unlocked in the same cases
2758 * as does filemap_fault().
2760 vm_fault_t
do_swap_page(struct vm_fault
*vmf
)
2762 struct vm_area_struct
*vma
= vmf
->vma
;
2763 struct page
*page
= NULL
, *swapcache
;
2764 struct mem_cgroup
*memcg
;
2771 if (!pte_unmap_same(vma
->vm_mm
, vmf
->pmd
, vmf
->pte
, vmf
->orig_pte
))
2774 entry
= pte_to_swp_entry(vmf
->orig_pte
);
2775 if (unlikely(non_swap_entry(entry
))) {
2776 if (is_migration_entry(entry
)) {
2777 migration_entry_wait(vma
->vm_mm
, vmf
->pmd
,
2779 } else if (is_device_private_entry(entry
)) {
2781 * For un-addressable device memory we call the pgmap
2782 * fault handler callback. The callback must migrate
2783 * the page back to some CPU accessible page.
2785 ret
= device_private_entry_fault(vma
, vmf
->address
, entry
,
2786 vmf
->flags
, vmf
->pmd
);
2787 } else if (is_hwpoison_entry(entry
)) {
2788 ret
= VM_FAULT_HWPOISON
;
2790 print_bad_pte(vma
, vmf
->address
, vmf
->orig_pte
, NULL
);
2791 ret
= VM_FAULT_SIGBUS
;
2797 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2798 page
= lookup_swap_cache(entry
, vma
, vmf
->address
);
2802 struct swap_info_struct
*si
= swp_swap_info(entry
);
2804 if (si
->flags
& SWP_SYNCHRONOUS_IO
&&
2805 __swap_count(entry
) == 1) {
2806 /* skip swapcache */
2807 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
,
2810 __SetPageLocked(page
);
2811 __SetPageSwapBacked(page
);
2812 set_page_private(page
, entry
.val
);
2813 lru_cache_add_anon(page
);
2814 swap_readpage(page
, true);
2817 page
= swapin_readahead(entry
, GFP_HIGHUSER_MOVABLE
,
2824 * Back out if somebody else faulted in this pte
2825 * while we released the pte lock.
2827 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
2828 vmf
->address
, &vmf
->ptl
);
2829 if (likely(pte_same(*vmf
->pte
, vmf
->orig_pte
)))
2831 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2835 /* Had to read the page from swap area: Major fault */
2836 ret
= VM_FAULT_MAJOR
;
2837 count_vm_event(PGMAJFAULT
);
2838 count_memcg_event_mm(vma
->vm_mm
, PGMAJFAULT
);
2839 } else if (PageHWPoison(page
)) {
2841 * hwpoisoned dirty swapcache pages are kept for killing
2842 * owner processes (which may be unknown at hwpoison time)
2844 ret
= VM_FAULT_HWPOISON
;
2845 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2849 locked
= lock_page_or_retry(page
, vma
->vm_mm
, vmf
->flags
);
2851 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2853 ret
|= VM_FAULT_RETRY
;
2858 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2859 * release the swapcache from under us. The page pin, and pte_same
2860 * test below, are not enough to exclude that. Even if it is still
2861 * swapcache, we need to check that the page's swap has not changed.
2863 if (unlikely((!PageSwapCache(page
) ||
2864 page_private(page
) != entry
.val
)) && swapcache
)
2867 page
= ksm_might_need_to_copy(page
, vma
, vmf
->address
);
2868 if (unlikely(!page
)) {
2874 if (mem_cgroup_try_charge_delay(page
, vma
->vm_mm
, GFP_KERNEL
,
2881 * Back out if somebody else already faulted in this pte.
2883 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
2885 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
)))
2888 if (unlikely(!PageUptodate(page
))) {
2889 ret
= VM_FAULT_SIGBUS
;
2894 * The page isn't present yet, go ahead with the fault.
2896 * Be careful about the sequence of operations here.
2897 * To get its accounting right, reuse_swap_page() must be called
2898 * while the page is counted on swap but not yet in mapcount i.e.
2899 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2900 * must be called after the swap_free(), or it will never succeed.
2903 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
2904 dec_mm_counter_fast(vma
->vm_mm
, MM_SWAPENTS
);
2905 pte
= mk_pte(page
, vma
->vm_page_prot
);
2906 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
, NULL
)) {
2907 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
2908 vmf
->flags
&= ~FAULT_FLAG_WRITE
;
2909 ret
|= VM_FAULT_WRITE
;
2910 exclusive
= RMAP_EXCLUSIVE
;
2912 flush_icache_page(vma
, page
);
2913 if (pte_swp_soft_dirty(vmf
->orig_pte
))
2914 pte
= pte_mksoft_dirty(pte
);
2915 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, pte
);
2916 arch_do_swap_page(vma
->vm_mm
, vma
, vmf
->address
, pte
, vmf
->orig_pte
);
2917 vmf
->orig_pte
= pte
;
2919 /* ksm created a completely new copy */
2920 if (unlikely(page
!= swapcache
&& swapcache
)) {
2921 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
2922 mem_cgroup_commit_charge(page
, memcg
, false, false);
2923 lru_cache_add_active_or_unevictable(page
, vma
);
2925 do_page_add_anon_rmap(page
, vma
, vmf
->address
, exclusive
);
2926 mem_cgroup_commit_charge(page
, memcg
, true, false);
2927 activate_page(page
);
2931 if (mem_cgroup_swap_full(page
) ||
2932 (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
2933 try_to_free_swap(page
);
2935 if (page
!= swapcache
&& swapcache
) {
2937 * Hold the lock to avoid the swap entry to be reused
2938 * until we take the PT lock for the pte_same() check
2939 * (to avoid false positives from pte_same). For
2940 * further safety release the lock after the swap_free
2941 * so that the swap count won't change under a
2942 * parallel locked swapcache.
2944 unlock_page(swapcache
);
2945 put_page(swapcache
);
2948 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
2949 ret
|= do_wp_page(vmf
);
2950 if (ret
& VM_FAULT_ERROR
)
2951 ret
&= VM_FAULT_ERROR
;
2955 /* No need to invalidate - it was non-present before */
2956 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
2958 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2962 mem_cgroup_cancel_charge(page
, memcg
, false);
2963 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
2968 if (page
!= swapcache
&& swapcache
) {
2969 unlock_page(swapcache
);
2970 put_page(swapcache
);
2976 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2977 * but allow concurrent faults), and pte mapped but not yet locked.
2978 * We return with mmap_sem still held, but pte unmapped and unlocked.
2980 static vm_fault_t
do_anonymous_page(struct vm_fault
*vmf
)
2982 struct vm_area_struct
*vma
= vmf
->vma
;
2983 struct mem_cgroup
*memcg
;
2988 /* File mapping without ->vm_ops ? */
2989 if (vma
->vm_flags
& VM_SHARED
)
2990 return VM_FAULT_SIGBUS
;
2993 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2994 * pte_offset_map() on pmds where a huge pmd might be created
2995 * from a different thread.
2997 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2998 * parallel threads are excluded by other means.
3000 * Here we only have down_read(mmap_sem).
3002 if (pte_alloc(vma
->vm_mm
, vmf
->pmd
))
3003 return VM_FAULT_OOM
;
3005 /* See the comment in pte_alloc_one_map() */
3006 if (unlikely(pmd_trans_unstable(vmf
->pmd
)))
3009 /* Use the zero-page for reads */
3010 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
3011 !mm_forbids_zeropage(vma
->vm_mm
)) {
3012 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(vmf
->address
),
3013 vma
->vm_page_prot
));
3014 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
,
3015 vmf
->address
, &vmf
->ptl
);
3016 if (!pte_none(*vmf
->pte
))
3018 ret
= check_stable_address_space(vma
->vm_mm
);
3021 /* Deliver the page fault to userland, check inside PT lock */
3022 if (userfaultfd_missing(vma
)) {
3023 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3024 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3029 /* Allocate our own private page. */
3030 if (unlikely(anon_vma_prepare(vma
)))
3032 page
= alloc_zeroed_user_highpage_movable(vma
, vmf
->address
);
3036 if (mem_cgroup_try_charge_delay(page
, vma
->vm_mm
, GFP_KERNEL
, &memcg
,
3041 * The memory barrier inside __SetPageUptodate makes sure that
3042 * preceeding stores to the page contents become visible before
3043 * the set_pte_at() write.
3045 __SetPageUptodate(page
);
3047 entry
= mk_pte(page
, vma
->vm_page_prot
);
3048 if (vma
->vm_flags
& VM_WRITE
)
3049 entry
= pte_mkwrite(pte_mkdirty(entry
));
3051 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3053 if (!pte_none(*vmf
->pte
))
3056 ret
= check_stable_address_space(vma
->vm_mm
);
3060 /* Deliver the page fault to userland, check inside PT lock */
3061 if (userfaultfd_missing(vma
)) {
3062 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3063 mem_cgroup_cancel_charge(page
, memcg
, false);
3065 return handle_userfault(vmf
, VM_UFFD_MISSING
);
3068 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3069 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3070 mem_cgroup_commit_charge(page
, memcg
, false, false);
3071 lru_cache_add_active_or_unevictable(page
, vma
);
3073 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3075 /* No need to invalidate - it was non-present before */
3076 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3078 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3081 mem_cgroup_cancel_charge(page
, memcg
, false);
3087 return VM_FAULT_OOM
;
3091 * The mmap_sem must have been held on entry, and may have been
3092 * released depending on flags and vma->vm_ops->fault() return value.
3093 * See filemap_fault() and __lock_page_retry().
3095 static vm_fault_t
__do_fault(struct vm_fault
*vmf
)
3097 struct vm_area_struct
*vma
= vmf
->vma
;
3101 * Preallocate pte before we take page_lock because this might lead to
3102 * deadlocks for memcg reclaim which waits for pages under writeback:
3104 * SetPageWriteback(A)
3110 * wait_on_page_writeback(A)
3111 * SetPageWriteback(B)
3113 * # flush A, B to clear the writeback
3115 if (pmd_none(*vmf
->pmd
) && !vmf
->prealloc_pte
) {
3116 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
);
3117 if (!vmf
->prealloc_pte
)
3118 return VM_FAULT_OOM
;
3119 smp_wmb(); /* See comment in __pte_alloc() */
3122 ret
= vma
->vm_ops
->fault(vmf
);
3123 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
|
3124 VM_FAULT_DONE_COW
)))
3127 if (unlikely(PageHWPoison(vmf
->page
))) {
3128 if (ret
& VM_FAULT_LOCKED
)
3129 unlock_page(vmf
->page
);
3130 put_page(vmf
->page
);
3132 return VM_FAULT_HWPOISON
;
3135 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3136 lock_page(vmf
->page
);
3138 VM_BUG_ON_PAGE(!PageLocked(vmf
->page
), vmf
->page
);
3144 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3145 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3146 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3147 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3149 static int pmd_devmap_trans_unstable(pmd_t
*pmd
)
3151 return pmd_devmap(*pmd
) || pmd_trans_unstable(pmd
);
3154 static vm_fault_t
pte_alloc_one_map(struct vm_fault
*vmf
)
3156 struct vm_area_struct
*vma
= vmf
->vma
;
3158 if (!pmd_none(*vmf
->pmd
))
3160 if (vmf
->prealloc_pte
) {
3161 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3162 if (unlikely(!pmd_none(*vmf
->pmd
))) {
3163 spin_unlock(vmf
->ptl
);
3167 mm_inc_nr_ptes(vma
->vm_mm
);
3168 pmd_populate(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3169 spin_unlock(vmf
->ptl
);
3170 vmf
->prealloc_pte
= NULL
;
3171 } else if (unlikely(pte_alloc(vma
->vm_mm
, vmf
->pmd
))) {
3172 return VM_FAULT_OOM
;
3176 * If a huge pmd materialized under us just retry later. Use
3177 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3178 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3179 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3180 * running immediately after a huge pmd fault in a different thread of
3181 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3182 * All we have to ensure is that it is a regular pmd that we can walk
3183 * with pte_offset_map() and we can do that through an atomic read in
3184 * C, which is what pmd_trans_unstable() provides.
3186 if (pmd_devmap_trans_unstable(vmf
->pmd
))
3187 return VM_FAULT_NOPAGE
;
3190 * At this point we know that our vmf->pmd points to a page of ptes
3191 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3192 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3193 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3194 * be valid and we will re-check to make sure the vmf->pte isn't
3195 * pte_none() under vmf->ptl protection when we return to
3198 vmf
->pte
= pte_offset_map_lock(vma
->vm_mm
, vmf
->pmd
, vmf
->address
,
3203 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3205 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3206 static inline bool transhuge_vma_suitable(struct vm_area_struct
*vma
,
3207 unsigned long haddr
)
3209 if (((vma
->vm_start
>> PAGE_SHIFT
) & HPAGE_CACHE_INDEX_MASK
) !=
3210 (vma
->vm_pgoff
& HPAGE_CACHE_INDEX_MASK
))
3212 if (haddr
< vma
->vm_start
|| haddr
+ HPAGE_PMD_SIZE
> vma
->vm_end
)
3217 static void deposit_prealloc_pte(struct vm_fault
*vmf
)
3219 struct vm_area_struct
*vma
= vmf
->vma
;
3221 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, vmf
->prealloc_pte
);
3223 * We are going to consume the prealloc table,
3224 * count that as nr_ptes.
3226 mm_inc_nr_ptes(vma
->vm_mm
);
3227 vmf
->prealloc_pte
= NULL
;
3230 static vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3232 struct vm_area_struct
*vma
= vmf
->vma
;
3233 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3234 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
3239 if (!transhuge_vma_suitable(vma
, haddr
))
3240 return VM_FAULT_FALLBACK
;
3242 ret
= VM_FAULT_FALLBACK
;
3243 page
= compound_head(page
);
3246 * Archs like ppc64 need additonal space to store information
3247 * related to pte entry. Use the preallocated table for that.
3249 if (arch_needs_pgtable_deposit() && !vmf
->prealloc_pte
) {
3250 vmf
->prealloc_pte
= pte_alloc_one(vma
->vm_mm
);
3251 if (!vmf
->prealloc_pte
)
3252 return VM_FAULT_OOM
;
3253 smp_wmb(); /* See comment in __pte_alloc() */
3256 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
3257 if (unlikely(!pmd_none(*vmf
->pmd
)))
3260 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
3261 flush_icache_page(vma
, page
+ i
);
3263 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
3265 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
3267 add_mm_counter(vma
->vm_mm
, mm_counter_file(page
), HPAGE_PMD_NR
);
3268 page_add_file_rmap(page
, true);
3270 * deposit and withdraw with pmd lock held
3272 if (arch_needs_pgtable_deposit())
3273 deposit_prealloc_pte(vmf
);
3275 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
3277 update_mmu_cache_pmd(vma
, haddr
, vmf
->pmd
);
3279 /* fault is handled */
3281 count_vm_event(THP_FILE_MAPPED
);
3283 spin_unlock(vmf
->ptl
);
3287 static vm_fault_t
do_set_pmd(struct vm_fault
*vmf
, struct page
*page
)
3295 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3296 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3298 * @vmf: fault environment
3299 * @memcg: memcg to charge page (only for private mappings)
3300 * @page: page to map
3302 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3305 * Target users are page handler itself and implementations of
3306 * vm_ops->map_pages.
3308 * Return: %0 on success, %VM_FAULT_ code in case of error.
3310 vm_fault_t
alloc_set_pte(struct vm_fault
*vmf
, struct mem_cgroup
*memcg
,
3313 struct vm_area_struct
*vma
= vmf
->vma
;
3314 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
3318 if (pmd_none(*vmf
->pmd
) && PageTransCompound(page
) &&
3319 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE
)) {
3321 VM_BUG_ON_PAGE(memcg
, page
);
3323 ret
= do_set_pmd(vmf
, page
);
3324 if (ret
!= VM_FAULT_FALLBACK
)
3329 ret
= pte_alloc_one_map(vmf
);
3334 /* Re-check under ptl */
3335 if (unlikely(!pte_none(*vmf
->pte
)))
3336 return VM_FAULT_NOPAGE
;
3338 flush_icache_page(vma
, page
);
3339 entry
= mk_pte(page
, vma
->vm_page_prot
);
3341 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3342 /* copy-on-write page */
3343 if (write
&& !(vma
->vm_flags
& VM_SHARED
)) {
3344 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
3345 page_add_new_anon_rmap(page
, vma
, vmf
->address
, false);
3346 mem_cgroup_commit_charge(page
, memcg
, false, false);
3347 lru_cache_add_active_or_unevictable(page
, vma
);
3349 inc_mm_counter_fast(vma
->vm_mm
, mm_counter_file(page
));
3350 page_add_file_rmap(page
, false);
3352 set_pte_at(vma
->vm_mm
, vmf
->address
, vmf
->pte
, entry
);
3354 /* no need to invalidate: a not-present page won't be cached */
3355 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3362 * finish_fault - finish page fault once we have prepared the page to fault
3364 * @vmf: structure describing the fault
3366 * This function handles all that is needed to finish a page fault once the
3367 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3368 * given page, adds reverse page mapping, handles memcg charges and LRU
3371 * The function expects the page to be locked and on success it consumes a
3372 * reference of a page being mapped (for the PTE which maps it).
3374 * Return: %0 on success, %VM_FAULT_ code in case of error.
3376 vm_fault_t
finish_fault(struct vm_fault
*vmf
)
3381 /* Did we COW the page? */
3382 if ((vmf
->flags
& FAULT_FLAG_WRITE
) &&
3383 !(vmf
->vma
->vm_flags
& VM_SHARED
))
3384 page
= vmf
->cow_page
;
3389 * check even for read faults because we might have lost our CoWed
3392 if (!(vmf
->vma
->vm_flags
& VM_SHARED
))
3393 ret
= check_stable_address_space(vmf
->vma
->vm_mm
);
3395 ret
= alloc_set_pte(vmf
, vmf
->memcg
, page
);
3397 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3401 static unsigned long fault_around_bytes __read_mostly
=
3402 rounddown_pow_of_two(65536);
3404 #ifdef CONFIG_DEBUG_FS
3405 static int fault_around_bytes_get(void *data
, u64
*val
)
3407 *val
= fault_around_bytes
;
3412 * fault_around_bytes must be rounded down to the nearest page order as it's
3413 * what do_fault_around() expects to see.
3415 static int fault_around_bytes_set(void *data
, u64 val
)
3417 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
3419 if (val
> PAGE_SIZE
)
3420 fault_around_bytes
= rounddown_pow_of_two(val
);
3422 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
3425 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops
,
3426 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
3428 static int __init
fault_around_debugfs(void)
3430 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL
, NULL
,
3431 &fault_around_bytes_fops
);
3434 late_initcall(fault_around_debugfs
);
3438 * do_fault_around() tries to map few pages around the fault address. The hope
3439 * is that the pages will be needed soon and this will lower the number of
3442 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3443 * not ready to be mapped: not up-to-date, locked, etc.
3445 * This function is called with the page table lock taken. In the split ptlock
3446 * case the page table lock only protects only those entries which belong to
3447 * the page table corresponding to the fault address.
3449 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3452 * fault_around_bytes defines how many bytes we'll try to map.
3453 * do_fault_around() expects it to be set to a power of two less than or equal
3456 * The virtual address of the area that we map is naturally aligned to
3457 * fault_around_bytes rounded down to the machine page size
3458 * (and therefore to page order). This way it's easier to guarantee
3459 * that we don't cross page table boundaries.
3461 static vm_fault_t
do_fault_around(struct vm_fault
*vmf
)
3463 unsigned long address
= vmf
->address
, nr_pages
, mask
;
3464 pgoff_t start_pgoff
= vmf
->pgoff
;
3469 nr_pages
= READ_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
3470 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
3472 vmf
->address
= max(address
& mask
, vmf
->vma
->vm_start
);
3473 off
= ((address
- vmf
->address
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
3477 * end_pgoff is either the end of the page table, the end of
3478 * the vma or nr_pages from start_pgoff, depending what is nearest.
3480 end_pgoff
= start_pgoff
-
3481 ((vmf
->address
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
3483 end_pgoff
= min3(end_pgoff
, vma_pages(vmf
->vma
) + vmf
->vma
->vm_pgoff
- 1,
3484 start_pgoff
+ nr_pages
- 1);
3486 if (pmd_none(*vmf
->pmd
)) {
3487 vmf
->prealloc_pte
= pte_alloc_one(vmf
->vma
->vm_mm
);
3488 if (!vmf
->prealloc_pte
)
3490 smp_wmb(); /* See comment in __pte_alloc() */
3493 vmf
->vma
->vm_ops
->map_pages(vmf
, start_pgoff
, end_pgoff
);
3495 /* Huge page is mapped? Page fault is solved */
3496 if (pmd_trans_huge(*vmf
->pmd
)) {
3497 ret
= VM_FAULT_NOPAGE
;
3501 /* ->map_pages() haven't done anything useful. Cold page cache? */
3505 /* check if the page fault is solved */
3506 vmf
->pte
-= (vmf
->address
>> PAGE_SHIFT
) - (address
>> PAGE_SHIFT
);
3507 if (!pte_none(*vmf
->pte
))
3508 ret
= VM_FAULT_NOPAGE
;
3509 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3511 vmf
->address
= address
;
3516 static vm_fault_t
do_read_fault(struct vm_fault
*vmf
)
3518 struct vm_area_struct
*vma
= vmf
->vma
;
3522 * Let's call ->map_pages() first and use ->fault() as fallback
3523 * if page by the offset is not ready to be mapped (cold cache or
3526 if (vma
->vm_ops
->map_pages
&& fault_around_bytes
>> PAGE_SHIFT
> 1) {
3527 ret
= do_fault_around(vmf
);
3532 ret
= __do_fault(vmf
);
3533 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3536 ret
|= finish_fault(vmf
);
3537 unlock_page(vmf
->page
);
3538 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3539 put_page(vmf
->page
);
3543 static vm_fault_t
do_cow_fault(struct vm_fault
*vmf
)
3545 struct vm_area_struct
*vma
= vmf
->vma
;
3548 if (unlikely(anon_vma_prepare(vma
)))
3549 return VM_FAULT_OOM
;
3551 vmf
->cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, vmf
->address
);
3553 return VM_FAULT_OOM
;
3555 if (mem_cgroup_try_charge_delay(vmf
->cow_page
, vma
->vm_mm
, GFP_KERNEL
,
3556 &vmf
->memcg
, false)) {
3557 put_page(vmf
->cow_page
);
3558 return VM_FAULT_OOM
;
3561 ret
= __do_fault(vmf
);
3562 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3564 if (ret
& VM_FAULT_DONE_COW
)
3567 copy_user_highpage(vmf
->cow_page
, vmf
->page
, vmf
->address
, vma
);
3568 __SetPageUptodate(vmf
->cow_page
);
3570 ret
|= finish_fault(vmf
);
3571 unlock_page(vmf
->page
);
3572 put_page(vmf
->page
);
3573 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3577 mem_cgroup_cancel_charge(vmf
->cow_page
, vmf
->memcg
, false);
3578 put_page(vmf
->cow_page
);
3582 static vm_fault_t
do_shared_fault(struct vm_fault
*vmf
)
3584 struct vm_area_struct
*vma
= vmf
->vma
;
3585 vm_fault_t ret
, tmp
;
3587 ret
= __do_fault(vmf
);
3588 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3592 * Check if the backing address space wants to know that the page is
3593 * about to become writable
3595 if (vma
->vm_ops
->page_mkwrite
) {
3596 unlock_page(vmf
->page
);
3597 tmp
= do_page_mkwrite(vmf
);
3598 if (unlikely(!tmp
||
3599 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3600 put_page(vmf
->page
);
3605 ret
|= finish_fault(vmf
);
3606 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
3608 unlock_page(vmf
->page
);
3609 put_page(vmf
->page
);
3613 fault_dirty_shared_page(vma
, vmf
->page
);
3618 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3619 * but allow concurrent faults).
3620 * The mmap_sem may have been released depending on flags and our
3621 * return value. See filemap_fault() and __lock_page_or_retry().
3622 * If mmap_sem is released, vma may become invalid (for example
3623 * by other thread calling munmap()).
3625 static vm_fault_t
do_fault(struct vm_fault
*vmf
)
3627 struct vm_area_struct
*vma
= vmf
->vma
;
3628 struct mm_struct
*vm_mm
= vma
->vm_mm
;
3632 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3634 if (!vma
->vm_ops
->fault
) {
3636 * If we find a migration pmd entry or a none pmd entry, which
3637 * should never happen, return SIGBUS
3639 if (unlikely(!pmd_present(*vmf
->pmd
)))
3640 ret
= VM_FAULT_SIGBUS
;
3642 vmf
->pte
= pte_offset_map_lock(vmf
->vma
->vm_mm
,
3647 * Make sure this is not a temporary clearing of pte
3648 * by holding ptl and checking again. A R/M/W update
3649 * of pte involves: take ptl, clearing the pte so that
3650 * we don't have concurrent modification by hardware
3651 * followed by an update.
3653 if (unlikely(pte_none(*vmf
->pte
)))
3654 ret
= VM_FAULT_SIGBUS
;
3656 ret
= VM_FAULT_NOPAGE
;
3658 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3660 } else if (!(vmf
->flags
& FAULT_FLAG_WRITE
))
3661 ret
= do_read_fault(vmf
);
3662 else if (!(vma
->vm_flags
& VM_SHARED
))
3663 ret
= do_cow_fault(vmf
);
3665 ret
= do_shared_fault(vmf
);
3667 /* preallocated pagetable is unused: free it */
3668 if (vmf
->prealloc_pte
) {
3669 pte_free(vm_mm
, vmf
->prealloc_pte
);
3670 vmf
->prealloc_pte
= NULL
;
3675 static int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3676 unsigned long addr
, int page_nid
,
3681 count_vm_numa_event(NUMA_HINT_FAULTS
);
3682 if (page_nid
== numa_node_id()) {
3683 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3684 *flags
|= TNF_FAULT_LOCAL
;
3687 return mpol_misplaced(page
, vma
, addr
);
3690 static vm_fault_t
do_numa_page(struct vm_fault
*vmf
)
3692 struct vm_area_struct
*vma
= vmf
->vma
;
3693 struct page
*page
= NULL
;
3694 int page_nid
= NUMA_NO_NODE
;
3697 bool migrated
= false;
3699 bool was_writable
= pte_savedwrite(vmf
->orig_pte
);
3703 * The "pte" at this point cannot be used safely without
3704 * validation through pte_unmap_same(). It's of NUMA type but
3705 * the pfn may be screwed if the read is non atomic.
3707 vmf
->ptl
= pte_lockptr(vma
->vm_mm
, vmf
->pmd
);
3708 spin_lock(vmf
->ptl
);
3709 if (unlikely(!pte_same(*vmf
->pte
, vmf
->orig_pte
))) {
3710 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3715 * Make it present again, Depending on how arch implementes non
3716 * accessible ptes, some can allow access by kernel mode.
3718 old_pte
= ptep_modify_prot_start(vma
, vmf
->address
, vmf
->pte
);
3719 pte
= pte_modify(old_pte
, vma
->vm_page_prot
);
3720 pte
= pte_mkyoung(pte
);
3722 pte
= pte_mkwrite(pte
);
3723 ptep_modify_prot_commit(vma
, vmf
->address
, vmf
->pte
, old_pte
, pte
);
3724 update_mmu_cache(vma
, vmf
->address
, vmf
->pte
);
3726 page
= vm_normal_page(vma
, vmf
->address
, pte
);
3728 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3732 /* TODO: handle PTE-mapped THP */
3733 if (PageCompound(page
)) {
3734 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3739 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3740 * much anyway since they can be in shared cache state. This misses
3741 * the case where a mapping is writable but the process never writes
3742 * to it but pte_write gets cleared during protection updates and
3743 * pte_dirty has unpredictable behaviour between PTE scan updates,
3744 * background writeback, dirty balancing and application behaviour.
3746 if (!pte_write(pte
))
3747 flags
|= TNF_NO_GROUP
;
3750 * Flag if the page is shared between multiple address spaces. This
3751 * is later used when determining whether to group tasks together
3753 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
3754 flags
|= TNF_SHARED
;
3756 last_cpupid
= page_cpupid_last(page
);
3757 page_nid
= page_to_nid(page
);
3758 target_nid
= numa_migrate_prep(page
, vma
, vmf
->address
, page_nid
,
3760 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3761 if (target_nid
== NUMA_NO_NODE
) {
3766 /* Migrate to the requested node */
3767 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
3769 page_nid
= target_nid
;
3770 flags
|= TNF_MIGRATED
;
3772 flags
|= TNF_MIGRATE_FAIL
;
3775 if (page_nid
!= NUMA_NO_NODE
)
3776 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
3780 static inline vm_fault_t
create_huge_pmd(struct vm_fault
*vmf
)
3782 if (vma_is_anonymous(vmf
->vma
))
3783 return do_huge_pmd_anonymous_page(vmf
);
3784 if (vmf
->vma
->vm_ops
->huge_fault
)
3785 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
3786 return VM_FAULT_FALLBACK
;
3789 /* `inline' is required to avoid gcc 4.1.2 build error */
3790 static inline vm_fault_t
wp_huge_pmd(struct vm_fault
*vmf
, pmd_t orig_pmd
)
3792 if (vma_is_anonymous(vmf
->vma
))
3793 return do_huge_pmd_wp_page(vmf
, orig_pmd
);
3794 if (vmf
->vma
->vm_ops
->huge_fault
)
3795 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PMD
);
3797 /* COW handled on pte level: split pmd */
3798 VM_BUG_ON_VMA(vmf
->vma
->vm_flags
& VM_SHARED
, vmf
->vma
);
3799 __split_huge_pmd(vmf
->vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
3801 return VM_FAULT_FALLBACK
;
3804 static inline bool vma_is_accessible(struct vm_area_struct
*vma
)
3806 return vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
);
3809 static vm_fault_t
create_huge_pud(struct vm_fault
*vmf
)
3811 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3812 /* No support for anonymous transparent PUD pages yet */
3813 if (vma_is_anonymous(vmf
->vma
))
3814 return VM_FAULT_FALLBACK
;
3815 if (vmf
->vma
->vm_ops
->huge_fault
)
3816 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
3817 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3818 return VM_FAULT_FALLBACK
;
3821 static vm_fault_t
wp_huge_pud(struct vm_fault
*vmf
, pud_t orig_pud
)
3823 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3824 /* No support for anonymous transparent PUD pages yet */
3825 if (vma_is_anonymous(vmf
->vma
))
3826 return VM_FAULT_FALLBACK
;
3827 if (vmf
->vma
->vm_ops
->huge_fault
)
3828 return vmf
->vma
->vm_ops
->huge_fault(vmf
, PE_SIZE_PUD
);
3829 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3830 return VM_FAULT_FALLBACK
;
3834 * These routines also need to handle stuff like marking pages dirty
3835 * and/or accessed for architectures that don't do it in hardware (most
3836 * RISC architectures). The early dirtying is also good on the i386.
3838 * There is also a hook called "update_mmu_cache()" that architectures
3839 * with external mmu caches can use to update those (ie the Sparc or
3840 * PowerPC hashed page tables that act as extended TLBs).
3842 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3843 * concurrent faults).
3845 * The mmap_sem may have been released depending on flags and our return value.
3846 * See filemap_fault() and __lock_page_or_retry().
3848 static vm_fault_t
handle_pte_fault(struct vm_fault
*vmf
)
3852 if (unlikely(pmd_none(*vmf
->pmd
))) {
3854 * Leave __pte_alloc() until later: because vm_ops->fault may
3855 * want to allocate huge page, and if we expose page table
3856 * for an instant, it will be difficult to retract from
3857 * concurrent faults and from rmap lookups.
3861 /* See comment in pte_alloc_one_map() */
3862 if (pmd_devmap_trans_unstable(vmf
->pmd
))
3865 * A regular pmd is established and it can't morph into a huge
3866 * pmd from under us anymore at this point because we hold the
3867 * mmap_sem read mode and khugepaged takes it in write mode.
3868 * So now it's safe to run pte_offset_map().
3870 vmf
->pte
= pte_offset_map(vmf
->pmd
, vmf
->address
);
3871 vmf
->orig_pte
= *vmf
->pte
;
3874 * some architectures can have larger ptes than wordsize,
3875 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3876 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3877 * accesses. The code below just needs a consistent view
3878 * for the ifs and we later double check anyway with the
3879 * ptl lock held. So here a barrier will do.
3882 if (pte_none(vmf
->orig_pte
)) {
3883 pte_unmap(vmf
->pte
);
3889 if (vma_is_anonymous(vmf
->vma
))
3890 return do_anonymous_page(vmf
);
3892 return do_fault(vmf
);
3895 if (!pte_present(vmf
->orig_pte
))
3896 return do_swap_page(vmf
);
3898 if (pte_protnone(vmf
->orig_pte
) && vma_is_accessible(vmf
->vma
))
3899 return do_numa_page(vmf
);
3901 vmf
->ptl
= pte_lockptr(vmf
->vma
->vm_mm
, vmf
->pmd
);
3902 spin_lock(vmf
->ptl
);
3903 entry
= vmf
->orig_pte
;
3904 if (unlikely(!pte_same(*vmf
->pte
, entry
)))
3906 if (vmf
->flags
& FAULT_FLAG_WRITE
) {
3907 if (!pte_write(entry
))
3908 return do_wp_page(vmf
);
3909 entry
= pte_mkdirty(entry
);
3911 entry
= pte_mkyoung(entry
);
3912 if (ptep_set_access_flags(vmf
->vma
, vmf
->address
, vmf
->pte
, entry
,
3913 vmf
->flags
& FAULT_FLAG_WRITE
)) {
3914 update_mmu_cache(vmf
->vma
, vmf
->address
, vmf
->pte
);
3917 * This is needed only for protection faults but the arch code
3918 * is not yet telling us if this is a protection fault or not.
3919 * This still avoids useless tlb flushes for .text page faults
3922 if (vmf
->flags
& FAULT_FLAG_WRITE
)
3923 flush_tlb_fix_spurious_fault(vmf
->vma
, vmf
->address
);
3926 pte_unmap_unlock(vmf
->pte
, vmf
->ptl
);
3931 * By the time we get here, we already hold the mm semaphore
3933 * The mmap_sem may have been released depending on flags and our
3934 * return value. See filemap_fault() and __lock_page_or_retry().
3936 static vm_fault_t
__handle_mm_fault(struct vm_area_struct
*vma
,
3937 unsigned long address
, unsigned int flags
)
3939 struct vm_fault vmf
= {
3941 .address
= address
& PAGE_MASK
,
3943 .pgoff
= linear_page_index(vma
, address
),
3944 .gfp_mask
= __get_fault_gfp_mask(vma
),
3946 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
3947 struct mm_struct
*mm
= vma
->vm_mm
;
3952 pgd
= pgd_offset(mm
, address
);
3953 p4d
= p4d_alloc(mm
, pgd
, address
);
3955 return VM_FAULT_OOM
;
3957 vmf
.pud
= pud_alloc(mm
, p4d
, address
);
3959 return VM_FAULT_OOM
;
3960 if (pud_none(*vmf
.pud
) && __transparent_hugepage_enabled(vma
)) {
3961 ret
= create_huge_pud(&vmf
);
3962 if (!(ret
& VM_FAULT_FALLBACK
))
3965 pud_t orig_pud
= *vmf
.pud
;
3968 if (pud_trans_huge(orig_pud
) || pud_devmap(orig_pud
)) {
3970 /* NUMA case for anonymous PUDs would go here */
3972 if (dirty
&& !pud_write(orig_pud
)) {
3973 ret
= wp_huge_pud(&vmf
, orig_pud
);
3974 if (!(ret
& VM_FAULT_FALLBACK
))
3977 huge_pud_set_accessed(&vmf
, orig_pud
);
3983 vmf
.pmd
= pmd_alloc(mm
, vmf
.pud
, address
);
3985 return VM_FAULT_OOM
;
3986 if (pmd_none(*vmf
.pmd
) && __transparent_hugepage_enabled(vma
)) {
3987 ret
= create_huge_pmd(&vmf
);
3988 if (!(ret
& VM_FAULT_FALLBACK
))
3991 pmd_t orig_pmd
= *vmf
.pmd
;
3994 if (unlikely(is_swap_pmd(orig_pmd
))) {
3995 VM_BUG_ON(thp_migration_supported() &&
3996 !is_pmd_migration_entry(orig_pmd
));
3997 if (is_pmd_migration_entry(orig_pmd
))
3998 pmd_migration_entry_wait(mm
, vmf
.pmd
);
4001 if (pmd_trans_huge(orig_pmd
) || pmd_devmap(orig_pmd
)) {
4002 if (pmd_protnone(orig_pmd
) && vma_is_accessible(vma
))
4003 return do_huge_pmd_numa_page(&vmf
, orig_pmd
);
4005 if (dirty
&& !pmd_write(orig_pmd
)) {
4006 ret
= wp_huge_pmd(&vmf
, orig_pmd
);
4007 if (!(ret
& VM_FAULT_FALLBACK
))
4010 huge_pmd_set_accessed(&vmf
, orig_pmd
);
4016 return handle_pte_fault(&vmf
);
4020 * By the time we get here, we already hold the mm semaphore
4022 * The mmap_sem may have been released depending on flags and our
4023 * return value. See filemap_fault() and __lock_page_or_retry().
4025 vm_fault_t
handle_mm_fault(struct vm_area_struct
*vma
, unsigned long address
,
4030 __set_current_state(TASK_RUNNING
);
4032 count_vm_event(PGFAULT
);
4033 count_memcg_event_mm(vma
->vm_mm
, PGFAULT
);
4035 /* do counter updates before entering really critical section. */
4036 check_sync_rss_stat(current
);
4038 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
4039 flags
& FAULT_FLAG_INSTRUCTION
,
4040 flags
& FAULT_FLAG_REMOTE
))
4041 return VM_FAULT_SIGSEGV
;
4044 * Enable the memcg OOM handling for faults triggered in user
4045 * space. Kernel faults are handled more gracefully.
4047 if (flags
& FAULT_FLAG_USER
)
4048 mem_cgroup_enter_user_fault();
4050 if (unlikely(is_vm_hugetlb_page(vma
)))
4051 ret
= hugetlb_fault(vma
->vm_mm
, vma
, address
, flags
);
4053 ret
= __handle_mm_fault(vma
, address
, flags
);
4055 if (flags
& FAULT_FLAG_USER
) {
4056 mem_cgroup_exit_user_fault();
4058 * The task may have entered a memcg OOM situation but
4059 * if the allocation error was handled gracefully (no
4060 * VM_FAULT_OOM), there is no need to kill anything.
4061 * Just clean up the OOM state peacefully.
4063 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
4064 mem_cgroup_oom_synchronize(false);
4069 EXPORT_SYMBOL_GPL(handle_mm_fault
);
4071 #ifndef __PAGETABLE_P4D_FOLDED
4073 * Allocate p4d page table.
4074 * We've already handled the fast-path in-line.
4076 int __p4d_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
4078 p4d_t
*new = p4d_alloc_one(mm
, address
);
4082 smp_wmb(); /* See comment in __pte_alloc */
4084 spin_lock(&mm
->page_table_lock
);
4085 if (pgd_present(*pgd
)) /* Another has populated it */
4088 pgd_populate(mm
, pgd
, new);
4089 spin_unlock(&mm
->page_table_lock
);
4092 #endif /* __PAGETABLE_P4D_FOLDED */
4094 #ifndef __PAGETABLE_PUD_FOLDED
4096 * Allocate page upper directory.
4097 * We've already handled the fast-path in-line.
4099 int __pud_alloc(struct mm_struct
*mm
, p4d_t
*p4d
, unsigned long address
)
4101 pud_t
*new = pud_alloc_one(mm
, address
);
4105 smp_wmb(); /* See comment in __pte_alloc */
4107 spin_lock(&mm
->page_table_lock
);
4108 #ifndef __ARCH_HAS_5LEVEL_HACK
4109 if (!p4d_present(*p4d
)) {
4111 p4d_populate(mm
, p4d
, new);
4112 } else /* Another has populated it */
4115 if (!pgd_present(*p4d
)) {
4117 pgd_populate(mm
, p4d
, new);
4118 } else /* Another has populated it */
4120 #endif /* __ARCH_HAS_5LEVEL_HACK */
4121 spin_unlock(&mm
->page_table_lock
);
4124 #endif /* __PAGETABLE_PUD_FOLDED */
4126 #ifndef __PAGETABLE_PMD_FOLDED
4128 * Allocate page middle directory.
4129 * We've already handled the fast-path in-line.
4131 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
4134 pmd_t
*new = pmd_alloc_one(mm
, address
);
4138 smp_wmb(); /* See comment in __pte_alloc */
4140 ptl
= pud_lock(mm
, pud
);
4141 #ifndef __ARCH_HAS_4LEVEL_HACK
4142 if (!pud_present(*pud
)) {
4144 pud_populate(mm
, pud
, new);
4145 } else /* Another has populated it */
4148 if (!pgd_present(*pud
)) {
4150 pgd_populate(mm
, pud
, new);
4151 } else /* Another has populated it */
4153 #endif /* __ARCH_HAS_4LEVEL_HACK */
4157 #endif /* __PAGETABLE_PMD_FOLDED */
4159 static int __follow_pte_pmd(struct mm_struct
*mm
, unsigned long address
,
4160 struct mmu_notifier_range
*range
,
4161 pte_t
**ptepp
, pmd_t
**pmdpp
, spinlock_t
**ptlp
)
4169 pgd
= pgd_offset(mm
, address
);
4170 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
4173 p4d
= p4d_offset(pgd
, address
);
4174 if (p4d_none(*p4d
) || unlikely(p4d_bad(*p4d
)))
4177 pud
= pud_offset(p4d
, address
);
4178 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
4181 pmd
= pmd_offset(pud
, address
);
4182 VM_BUG_ON(pmd_trans_huge(*pmd
));
4184 if (pmd_huge(*pmd
)) {
4189 mmu_notifier_range_init(range
, MMU_NOTIFY_CLEAR
, 0,
4190 NULL
, mm
, address
& PMD_MASK
,
4191 (address
& PMD_MASK
) + PMD_SIZE
);
4192 mmu_notifier_invalidate_range_start(range
);
4194 *ptlp
= pmd_lock(mm
, pmd
);
4195 if (pmd_huge(*pmd
)) {
4201 mmu_notifier_invalidate_range_end(range
);
4204 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
4208 mmu_notifier_range_init(range
, MMU_NOTIFY_CLEAR
, 0, NULL
, mm
,
4209 address
& PAGE_MASK
,
4210 (address
& PAGE_MASK
) + PAGE_SIZE
);
4211 mmu_notifier_invalidate_range_start(range
);
4213 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
4214 if (!pte_present(*ptep
))
4219 pte_unmap_unlock(ptep
, *ptlp
);
4221 mmu_notifier_invalidate_range_end(range
);
4226 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
4227 pte_t
**ptepp
, spinlock_t
**ptlp
)
4231 /* (void) is needed to make gcc happy */
4232 (void) __cond_lock(*ptlp
,
4233 !(res
= __follow_pte_pmd(mm
, address
, NULL
,
4234 ptepp
, NULL
, ptlp
)));
4238 int follow_pte_pmd(struct mm_struct
*mm
, unsigned long address
,
4239 struct mmu_notifier_range
*range
,
4240 pte_t
**ptepp
, pmd_t
**pmdpp
, spinlock_t
**ptlp
)
4244 /* (void) is needed to make gcc happy */
4245 (void) __cond_lock(*ptlp
,
4246 !(res
= __follow_pte_pmd(mm
, address
, range
,
4247 ptepp
, pmdpp
, ptlp
)));
4250 EXPORT_SYMBOL(follow_pte_pmd
);
4253 * follow_pfn - look up PFN at a user virtual address
4254 * @vma: memory mapping
4255 * @address: user virtual address
4256 * @pfn: location to store found PFN
4258 * Only IO mappings and raw PFN mappings are allowed.
4260 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4262 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
4269 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4272 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
4275 *pfn
= pte_pfn(*ptep
);
4276 pte_unmap_unlock(ptep
, ptl
);
4279 EXPORT_SYMBOL(follow_pfn
);
4281 #ifdef CONFIG_HAVE_IOREMAP_PROT
4282 int follow_phys(struct vm_area_struct
*vma
,
4283 unsigned long address
, unsigned int flags
,
4284 unsigned long *prot
, resource_size_t
*phys
)
4290 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4293 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
4297 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
4300 *prot
= pgprot_val(pte_pgprot(pte
));
4301 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
4305 pte_unmap_unlock(ptep
, ptl
);
4310 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
4311 void *buf
, int len
, int write
)
4313 resource_size_t phys_addr
;
4314 unsigned long prot
= 0;
4315 void __iomem
*maddr
;
4316 int offset
= addr
& (PAGE_SIZE
-1);
4318 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
4321 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
4326 memcpy_toio(maddr
+ offset
, buf
, len
);
4328 memcpy_fromio(buf
, maddr
+ offset
, len
);
4333 EXPORT_SYMBOL_GPL(generic_access_phys
);
4337 * Access another process' address space as given in mm. If non-NULL, use the
4338 * given task for page fault accounting.
4340 int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
4341 unsigned long addr
, void *buf
, int len
, unsigned int gup_flags
)
4343 struct vm_area_struct
*vma
;
4344 void *old_buf
= buf
;
4345 int write
= gup_flags
& FOLL_WRITE
;
4347 down_read(&mm
->mmap_sem
);
4348 /* ignore errors, just check how much was successfully transferred */
4350 int bytes
, ret
, offset
;
4352 struct page
*page
= NULL
;
4354 ret
= get_user_pages_remote(tsk
, mm
, addr
, 1,
4355 gup_flags
, &page
, &vma
, NULL
);
4357 #ifndef CONFIG_HAVE_IOREMAP_PROT
4361 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4362 * we can access using slightly different code.
4364 vma
= find_vma(mm
, addr
);
4365 if (!vma
|| vma
->vm_start
> addr
)
4367 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
4368 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
4376 offset
= addr
& (PAGE_SIZE
-1);
4377 if (bytes
> PAGE_SIZE
-offset
)
4378 bytes
= PAGE_SIZE
-offset
;
4382 copy_to_user_page(vma
, page
, addr
,
4383 maddr
+ offset
, buf
, bytes
);
4384 set_page_dirty_lock(page
);
4386 copy_from_user_page(vma
, page
, addr
,
4387 buf
, maddr
+ offset
, bytes
);
4396 up_read(&mm
->mmap_sem
);
4398 return buf
- old_buf
;
4402 * access_remote_vm - access another process' address space
4403 * @mm: the mm_struct of the target address space
4404 * @addr: start address to access
4405 * @buf: source or destination buffer
4406 * @len: number of bytes to transfer
4407 * @gup_flags: flags modifying lookup behaviour
4409 * The caller must hold a reference on @mm.
4411 * Return: number of bytes copied from source to destination.
4413 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
4414 void *buf
, int len
, unsigned int gup_flags
)
4416 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, gup_flags
);
4420 * Access another process' address space.
4421 * Source/target buffer must be kernel space,
4422 * Do not walk the page table directly, use get_user_pages
4424 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
4425 void *buf
, int len
, unsigned int gup_flags
)
4427 struct mm_struct
*mm
;
4430 mm
= get_task_mm(tsk
);
4434 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, gup_flags
);
4440 EXPORT_SYMBOL_GPL(access_process_vm
);
4443 * Print the name of a VMA.
4445 void print_vma_addr(char *prefix
, unsigned long ip
)
4447 struct mm_struct
*mm
= current
->mm
;
4448 struct vm_area_struct
*vma
;
4451 * we might be running from an atomic context so we cannot sleep
4453 if (!down_read_trylock(&mm
->mmap_sem
))
4456 vma
= find_vma(mm
, ip
);
4457 if (vma
&& vma
->vm_file
) {
4458 struct file
*f
= vma
->vm_file
;
4459 char *buf
= (char *)__get_free_page(GFP_NOWAIT
);
4463 p
= file_path(f
, buf
, PAGE_SIZE
);
4466 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
4468 vma
->vm_end
- vma
->vm_start
);
4469 free_page((unsigned long)buf
);
4472 up_read(&mm
->mmap_sem
);
4475 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4476 void __might_fault(const char *file
, int line
)
4479 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4480 * holding the mmap_sem, this is safe because kernel memory doesn't
4481 * get paged out, therefore we'll never actually fault, and the
4482 * below annotations will generate false positives.
4484 if (uaccess_kernel())
4486 if (pagefault_disabled())
4488 __might_sleep(file
, line
, 0);
4489 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4491 might_lock_read(¤t
->mm
->mmap_sem
);
4494 EXPORT_SYMBOL(__might_fault
);
4497 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4499 * Process all subpages of the specified huge page with the specified
4500 * operation. The target subpage will be processed last to keep its
4503 static inline void process_huge_page(
4504 unsigned long addr_hint
, unsigned int pages_per_huge_page
,
4505 void (*process_subpage
)(unsigned long addr
, int idx
, void *arg
),
4509 unsigned long addr
= addr_hint
&
4510 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
4512 /* Process target subpage last to keep its cache lines hot */
4514 n
= (addr_hint
- addr
) / PAGE_SIZE
;
4515 if (2 * n
<= pages_per_huge_page
) {
4516 /* If target subpage in first half of huge page */
4519 /* Process subpages at the end of huge page */
4520 for (i
= pages_per_huge_page
- 1; i
>= 2 * n
; i
--) {
4522 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
4525 /* If target subpage in second half of huge page */
4526 base
= pages_per_huge_page
- 2 * (pages_per_huge_page
- n
);
4527 l
= pages_per_huge_page
- n
;
4528 /* Process subpages at the begin of huge page */
4529 for (i
= 0; i
< base
; i
++) {
4531 process_subpage(addr
+ i
* PAGE_SIZE
, i
, arg
);
4535 * Process remaining subpages in left-right-left-right pattern
4536 * towards the target subpage
4538 for (i
= 0; i
< l
; i
++) {
4539 int left_idx
= base
+ i
;
4540 int right_idx
= base
+ 2 * l
- 1 - i
;
4543 process_subpage(addr
+ left_idx
* PAGE_SIZE
, left_idx
, arg
);
4545 process_subpage(addr
+ right_idx
* PAGE_SIZE
, right_idx
, arg
);
4549 static void clear_gigantic_page(struct page
*page
,
4551 unsigned int pages_per_huge_page
)
4554 struct page
*p
= page
;
4557 for (i
= 0; i
< pages_per_huge_page
;
4558 i
++, p
= mem_map_next(p
, page
, i
)) {
4560 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
4564 static void clear_subpage(unsigned long addr
, int idx
, void *arg
)
4566 struct page
*page
= arg
;
4568 clear_user_highpage(page
+ idx
, addr
);
4571 void clear_huge_page(struct page
*page
,
4572 unsigned long addr_hint
, unsigned int pages_per_huge_page
)
4574 unsigned long addr
= addr_hint
&
4575 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
4577 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4578 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
4582 process_huge_page(addr_hint
, pages_per_huge_page
, clear_subpage
, page
);
4585 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
4587 struct vm_area_struct
*vma
,
4588 unsigned int pages_per_huge_page
)
4591 struct page
*dst_base
= dst
;
4592 struct page
*src_base
= src
;
4594 for (i
= 0; i
< pages_per_huge_page
; ) {
4596 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
4599 dst
= mem_map_next(dst
, dst_base
, i
);
4600 src
= mem_map_next(src
, src_base
, i
);
4604 struct copy_subpage_arg
{
4607 struct vm_area_struct
*vma
;
4610 static void copy_subpage(unsigned long addr
, int idx
, void *arg
)
4612 struct copy_subpage_arg
*copy_arg
= arg
;
4614 copy_user_highpage(copy_arg
->dst
+ idx
, copy_arg
->src
+ idx
,
4615 addr
, copy_arg
->vma
);
4618 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
4619 unsigned long addr_hint
, struct vm_area_struct
*vma
,
4620 unsigned int pages_per_huge_page
)
4622 unsigned long addr
= addr_hint
&
4623 ~(((unsigned long)pages_per_huge_page
<< PAGE_SHIFT
) - 1);
4624 struct copy_subpage_arg arg
= {
4630 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4631 copy_user_gigantic_page(dst
, src
, addr
, vma
,
4632 pages_per_huge_page
);
4636 process_huge_page(addr_hint
, pages_per_huge_page
, copy_subpage
, &arg
);
4639 long copy_huge_page_from_user(struct page
*dst_page
,
4640 const void __user
*usr_src
,
4641 unsigned int pages_per_huge_page
,
4642 bool allow_pagefault
)
4644 void *src
= (void *)usr_src
;
4646 unsigned long i
, rc
= 0;
4647 unsigned long ret_val
= pages_per_huge_page
* PAGE_SIZE
;
4649 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4650 if (allow_pagefault
)
4651 page_kaddr
= kmap(dst_page
+ i
);
4653 page_kaddr
= kmap_atomic(dst_page
+ i
);
4654 rc
= copy_from_user(page_kaddr
,
4655 (const void __user
*)(src
+ i
* PAGE_SIZE
),
4657 if (allow_pagefault
)
4658 kunmap(dst_page
+ i
);
4660 kunmap_atomic(page_kaddr
);
4662 ret_val
-= (PAGE_SIZE
- rc
);
4670 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4672 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4674 static struct kmem_cache
*page_ptl_cachep
;
4676 void __init
ptlock_cache_init(void)
4678 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
4682 bool ptlock_alloc(struct page
*page
)
4686 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
4693 void ptlock_free(struct page
*page
)
4695 kmem_cache_free(page_ptl_cachep
, page
->ptl
);