4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 #include <linux/kernel_stat.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/export.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
60 #include <linux/migrate.h>
61 #include <linux/string.h>
62 #include <linux/dma-debug.h>
65 #include <asm/pgalloc.h>
66 #include <asm/uaccess.h>
68 #include <asm/tlbflush.h>
69 #include <asm/pgtable.h>
73 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
74 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
77 #ifndef CONFIG_NEED_MULTIPLE_NODES
78 /* use the per-pgdat data instead for discontigmem - mbligh */
79 unsigned long max_mapnr
;
82 EXPORT_SYMBOL(max_mapnr
);
83 EXPORT_SYMBOL(mem_map
);
87 * A number of key systems in x86 including ioremap() rely on the assumption
88 * that high_memory defines the upper bound on direct map memory, then end
89 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
90 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
95 EXPORT_SYMBOL(high_memory
);
98 * Randomize the address space (stacks, mmaps, brk, etc.).
100 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
101 * as ancient (libc5 based) binaries can segfault. )
103 int randomize_va_space __read_mostly
=
104 #ifdef CONFIG_COMPAT_BRK
110 static int __init
disable_randmaps(char *s
)
112 randomize_va_space
= 0;
115 __setup("norandmaps", disable_randmaps
);
117 unsigned long zero_pfn __read_mostly
;
118 unsigned long highest_memmap_pfn __read_mostly
;
121 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
123 static int __init
init_zero_pfn(void)
125 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
128 core_initcall(init_zero_pfn
);
131 #if defined(SPLIT_RSS_COUNTING)
133 void sync_mm_rss(struct mm_struct
*mm
)
137 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
138 if (current
->rss_stat
.count
[i
]) {
139 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
140 current
->rss_stat
.count
[i
] = 0;
143 current
->rss_stat
.events
= 0;
146 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
148 struct task_struct
*task
= current
;
150 if (likely(task
->mm
== mm
))
151 task
->rss_stat
.count
[member
] += val
;
153 add_mm_counter(mm
, member
, val
);
155 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
156 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
158 /* sync counter once per 64 page faults */
159 #define TASK_RSS_EVENTS_THRESH (64)
160 static void check_sync_rss_stat(struct task_struct
*task
)
162 if (unlikely(task
!= current
))
164 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
165 sync_mm_rss(task
->mm
);
167 #else /* SPLIT_RSS_COUNTING */
169 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
170 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
172 static void check_sync_rss_stat(struct task_struct
*task
)
176 #endif /* SPLIT_RSS_COUNTING */
178 #ifdef HAVE_GENERIC_MMU_GATHER
180 static int tlb_next_batch(struct mmu_gather
*tlb
)
182 struct mmu_gather_batch
*batch
;
186 tlb
->active
= batch
->next
;
190 if (tlb
->batch_count
== MAX_GATHER_BATCH_COUNT
)
193 batch
= (void *)__get_free_pages(GFP_NOWAIT
| __GFP_NOWARN
, 0);
200 batch
->max
= MAX_GATHER_BATCH
;
202 tlb
->active
->next
= batch
;
209 * Called to initialize an (on-stack) mmu_gather structure for page-table
210 * tear-down from @mm. The @fullmm argument is used when @mm is without
211 * users and we're going to destroy the full address space (exit/execve).
213 void tlb_gather_mmu(struct mmu_gather
*tlb
, struct mm_struct
*mm
, unsigned long start
, unsigned long end
)
217 /* Is it from 0 to ~0? */
218 tlb
->fullmm
= !(start
| (end
+1));
219 tlb
->need_flush_all
= 0;
223 tlb
->local
.next
= NULL
;
225 tlb
->local
.max
= ARRAY_SIZE(tlb
->__pages
);
226 tlb
->active
= &tlb
->local
;
227 tlb
->batch_count
= 0;
229 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
234 void tlb_flush_mmu(struct mmu_gather
*tlb
)
236 struct mmu_gather_batch
*batch
;
238 if (!tlb
->need_flush
)
242 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
243 tlb_table_flush(tlb
);
246 for (batch
= &tlb
->local
; batch
; batch
= batch
->next
) {
247 free_pages_and_swap_cache(batch
->pages
, batch
->nr
);
250 tlb
->active
= &tlb
->local
;
254 * Called at the end of the shootdown operation to free up any resources
255 * that were required.
257 void tlb_finish_mmu(struct mmu_gather
*tlb
, unsigned long start
, unsigned long end
)
259 struct mmu_gather_batch
*batch
, *next
;
263 /* keep the page table cache within bounds */
266 for (batch
= tlb
->local
.next
; batch
; batch
= next
) {
268 free_pages((unsigned long)batch
, 0);
270 tlb
->local
.next
= NULL
;
274 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
275 * handling the additional races in SMP caused by other CPUs caching valid
276 * mappings in their TLBs. Returns the number of free page slots left.
277 * When out of page slots we must call tlb_flush_mmu().
279 int __tlb_remove_page(struct mmu_gather
*tlb
, struct page
*page
)
281 struct mmu_gather_batch
*batch
;
283 VM_BUG_ON(!tlb
->need_flush
);
286 batch
->pages
[batch
->nr
++] = page
;
287 if (batch
->nr
== batch
->max
) {
288 if (!tlb_next_batch(tlb
))
292 VM_BUG_ON_PAGE(batch
->nr
> batch
->max
, page
);
294 return batch
->max
- batch
->nr
;
297 #endif /* HAVE_GENERIC_MMU_GATHER */
299 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
302 * See the comment near struct mmu_table_batch.
305 static void tlb_remove_table_smp_sync(void *arg
)
307 /* Simply deliver the interrupt */
310 static void tlb_remove_table_one(void *table
)
313 * This isn't an RCU grace period and hence the page-tables cannot be
314 * assumed to be actually RCU-freed.
316 * It is however sufficient for software page-table walkers that rely on
317 * IRQ disabling. See the comment near struct mmu_table_batch.
319 smp_call_function(tlb_remove_table_smp_sync
, NULL
, 1);
320 __tlb_remove_table(table
);
323 static void tlb_remove_table_rcu(struct rcu_head
*head
)
325 struct mmu_table_batch
*batch
;
328 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
330 for (i
= 0; i
< batch
->nr
; i
++)
331 __tlb_remove_table(batch
->tables
[i
]);
333 free_page((unsigned long)batch
);
336 void tlb_table_flush(struct mmu_gather
*tlb
)
338 struct mmu_table_batch
**batch
= &tlb
->batch
;
341 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
346 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
348 struct mmu_table_batch
**batch
= &tlb
->batch
;
353 * When there's less then two users of this mm there cannot be a
354 * concurrent page-table walk.
356 if (atomic_read(&tlb
->mm
->mm_users
) < 2) {
357 __tlb_remove_table(table
);
361 if (*batch
== NULL
) {
362 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
363 if (*batch
== NULL
) {
364 tlb_remove_table_one(table
);
369 (*batch
)->tables
[(*batch
)->nr
++] = table
;
370 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
371 tlb_table_flush(tlb
);
374 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
377 * Note: this doesn't free the actual pages themselves. That
378 * has been handled earlier when unmapping all the memory regions.
380 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
383 pgtable_t token
= pmd_pgtable(*pmd
);
385 pte_free_tlb(tlb
, token
, addr
);
386 atomic_long_dec(&tlb
->mm
->nr_ptes
);
389 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
390 unsigned long addr
, unsigned long end
,
391 unsigned long floor
, unsigned long ceiling
)
398 pmd
= pmd_offset(pud
, addr
);
400 next
= pmd_addr_end(addr
, end
);
401 if (pmd_none_or_clear_bad(pmd
))
403 free_pte_range(tlb
, pmd
, addr
);
404 } while (pmd
++, addr
= next
, addr
!= end
);
414 if (end
- 1 > ceiling
- 1)
417 pmd
= pmd_offset(pud
, start
);
419 pmd_free_tlb(tlb
, pmd
, start
);
422 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
423 unsigned long addr
, unsigned long end
,
424 unsigned long floor
, unsigned long ceiling
)
431 pud
= pud_offset(pgd
, addr
);
433 next
= pud_addr_end(addr
, end
);
434 if (pud_none_or_clear_bad(pud
))
436 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
437 } while (pud
++, addr
= next
, addr
!= end
);
443 ceiling
&= PGDIR_MASK
;
447 if (end
- 1 > ceiling
- 1)
450 pud
= pud_offset(pgd
, start
);
452 pud_free_tlb(tlb
, pud
, start
);
456 * This function frees user-level page tables of a process.
458 void free_pgd_range(struct mmu_gather
*tlb
,
459 unsigned long addr
, unsigned long end
,
460 unsigned long floor
, unsigned long ceiling
)
466 * The next few lines have given us lots of grief...
468 * Why are we testing PMD* at this top level? Because often
469 * there will be no work to do at all, and we'd prefer not to
470 * go all the way down to the bottom just to discover that.
472 * Why all these "- 1"s? Because 0 represents both the bottom
473 * of the address space and the top of it (using -1 for the
474 * top wouldn't help much: the masks would do the wrong thing).
475 * The rule is that addr 0 and floor 0 refer to the bottom of
476 * the address space, but end 0 and ceiling 0 refer to the top
477 * Comparisons need to use "end - 1" and "ceiling - 1" (though
478 * that end 0 case should be mythical).
480 * Wherever addr is brought up or ceiling brought down, we must
481 * be careful to reject "the opposite 0" before it confuses the
482 * subsequent tests. But what about where end is brought down
483 * by PMD_SIZE below? no, end can't go down to 0 there.
485 * Whereas we round start (addr) and ceiling down, by different
486 * masks at different levels, in order to test whether a table
487 * now has no other vmas using it, so can be freed, we don't
488 * bother to round floor or end up - the tests don't need that.
502 if (end
- 1 > ceiling
- 1)
507 pgd
= pgd_offset(tlb
->mm
, addr
);
509 next
= pgd_addr_end(addr
, end
);
510 if (pgd_none_or_clear_bad(pgd
))
512 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
513 } while (pgd
++, addr
= next
, addr
!= end
);
516 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
517 unsigned long floor
, unsigned long ceiling
)
520 struct vm_area_struct
*next
= vma
->vm_next
;
521 unsigned long addr
= vma
->vm_start
;
524 * Hide vma from rmap and truncate_pagecache before freeing
527 unlink_anon_vmas(vma
);
528 unlink_file_vma(vma
);
530 if (is_vm_hugetlb_page(vma
)) {
531 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
532 floor
, next
? next
->vm_start
: ceiling
);
535 * Optimization: gather nearby vmas into one call down
537 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
538 && !is_vm_hugetlb_page(next
)) {
541 unlink_anon_vmas(vma
);
542 unlink_file_vma(vma
);
544 free_pgd_range(tlb
, addr
, vma
->vm_end
,
545 floor
, next
? next
->vm_start
: ceiling
);
551 int __pte_alloc(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
552 pmd_t
*pmd
, unsigned long address
)
555 pgtable_t
new = pte_alloc_one(mm
, address
);
556 int wait_split_huge_page
;
561 * Ensure all pte setup (eg. pte page lock and page clearing) are
562 * visible before the pte is made visible to other CPUs by being
563 * put into page tables.
565 * The other side of the story is the pointer chasing in the page
566 * table walking code (when walking the page table without locking;
567 * ie. most of the time). Fortunately, these data accesses consist
568 * of a chain of data-dependent loads, meaning most CPUs (alpha
569 * being the notable exception) will already guarantee loads are
570 * seen in-order. See the alpha page table accessors for the
571 * smp_read_barrier_depends() barriers in page table walking code.
573 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
575 ptl
= pmd_lock(mm
, pmd
);
576 wait_split_huge_page
= 0;
577 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
578 atomic_long_inc(&mm
->nr_ptes
);
579 pmd_populate(mm
, pmd
, new);
581 } else if (unlikely(pmd_trans_splitting(*pmd
)))
582 wait_split_huge_page
= 1;
586 if (wait_split_huge_page
)
587 wait_split_huge_page(vma
->anon_vma
, pmd
);
591 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
593 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
597 smp_wmb(); /* See comment in __pte_alloc */
599 spin_lock(&init_mm
.page_table_lock
);
600 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
601 pmd_populate_kernel(&init_mm
, pmd
, new);
604 VM_BUG_ON(pmd_trans_splitting(*pmd
));
605 spin_unlock(&init_mm
.page_table_lock
);
607 pte_free_kernel(&init_mm
, new);
611 static inline void init_rss_vec(int *rss
)
613 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
616 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
620 if (current
->mm
== mm
)
622 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
624 add_mm_counter(mm
, i
, rss
[i
]);
628 * This function is called to print an error when a bad pte
629 * is found. For example, we might have a PFN-mapped pte in
630 * a region that doesn't allow it.
632 * The calling function must still handle the error.
634 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
635 pte_t pte
, struct page
*page
)
637 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
638 pud_t
*pud
= pud_offset(pgd
, addr
);
639 pmd_t
*pmd
= pmd_offset(pud
, addr
);
640 struct address_space
*mapping
;
642 static unsigned long resume
;
643 static unsigned long nr_shown
;
644 static unsigned long nr_unshown
;
647 * Allow a burst of 60 reports, then keep quiet for that minute;
648 * or allow a steady drip of one report per second.
650 if (nr_shown
== 60) {
651 if (time_before(jiffies
, resume
)) {
657 "BUG: Bad page map: %lu messages suppressed\n",
664 resume
= jiffies
+ 60 * HZ
;
666 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
667 index
= linear_page_index(vma
, addr
);
670 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
672 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
674 dump_page(page
, "bad pte");
676 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
677 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
679 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
682 printk(KERN_ALERT
"vma->vm_ops->fault: %pSR\n",
685 printk(KERN_ALERT
"vma->vm_file->f_op->mmap: %pSR\n",
686 vma
->vm_file
->f_op
->mmap
);
688 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
691 static inline bool is_cow_mapping(vm_flags_t flags
)
693 return (flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
697 * vm_normal_page -- This function gets the "struct page" associated with a pte.
699 * "Special" mappings do not wish to be associated with a "struct page" (either
700 * it doesn't exist, or it exists but they don't want to touch it). In this
701 * case, NULL is returned here. "Normal" mappings do have a struct page.
703 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
704 * pte bit, in which case this function is trivial. Secondly, an architecture
705 * may not have a spare pte bit, which requires a more complicated scheme,
708 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
709 * special mapping (even if there are underlying and valid "struct pages").
710 * COWed pages of a VM_PFNMAP are always normal.
712 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
713 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
714 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
715 * mapping will always honor the rule
717 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
719 * And for normal mappings this is false.
721 * This restricts such mappings to be a linear translation from virtual address
722 * to pfn. To get around this restriction, we allow arbitrary mappings so long
723 * as the vma is not a COW mapping; in that case, we know that all ptes are
724 * special (because none can have been COWed).
727 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
729 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
730 * page" backing, however the difference is that _all_ pages with a struct
731 * page (that is, those where pfn_valid is true) are refcounted and considered
732 * normal pages by the VM. The disadvantage is that pages are refcounted
733 * (which can be slower and simply not an option for some PFNMAP users). The
734 * advantage is that we don't have to follow the strict linearity rule of
735 * PFNMAP mappings in order to support COWable mappings.
738 #ifdef __HAVE_ARCH_PTE_SPECIAL
739 # define HAVE_PTE_SPECIAL 1
741 # define HAVE_PTE_SPECIAL 0
743 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
746 unsigned long pfn
= pte_pfn(pte
);
748 if (HAVE_PTE_SPECIAL
) {
749 if (likely(!pte_special(pte
)))
751 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
753 if (!is_zero_pfn(pfn
))
754 print_bad_pte(vma
, addr
, pte
, NULL
);
758 /* !HAVE_PTE_SPECIAL case follows: */
760 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
761 if (vma
->vm_flags
& VM_MIXEDMAP
) {
767 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
768 if (pfn
== vma
->vm_pgoff
+ off
)
770 if (!is_cow_mapping(vma
->vm_flags
))
775 if (is_zero_pfn(pfn
))
778 if (unlikely(pfn
> highest_memmap_pfn
)) {
779 print_bad_pte(vma
, addr
, pte
, NULL
);
784 * NOTE! We still have PageReserved() pages in the page tables.
785 * eg. VDSO mappings can cause them to exist.
788 return pfn_to_page(pfn
);
792 * copy one vm_area from one task to the other. Assumes the page tables
793 * already present in the new task to be cleared in the whole range
794 * covered by this vma.
797 static inline unsigned long
798 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
799 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
800 unsigned long addr
, int *rss
)
802 unsigned long vm_flags
= vma
->vm_flags
;
803 pte_t pte
= *src_pte
;
806 /* pte contains position in swap or file, so copy. */
807 if (unlikely(!pte_present(pte
))) {
808 if (!pte_file(pte
)) {
809 swp_entry_t entry
= pte_to_swp_entry(pte
);
811 if (swap_duplicate(entry
) < 0)
814 /* make sure dst_mm is on swapoff's mmlist. */
815 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
816 spin_lock(&mmlist_lock
);
817 if (list_empty(&dst_mm
->mmlist
))
818 list_add(&dst_mm
->mmlist
,
820 spin_unlock(&mmlist_lock
);
822 if (likely(!non_swap_entry(entry
)))
824 else if (is_migration_entry(entry
)) {
825 page
= migration_entry_to_page(entry
);
832 if (is_write_migration_entry(entry
) &&
833 is_cow_mapping(vm_flags
)) {
835 * COW mappings require pages in both
836 * parent and child to be set to read.
838 make_migration_entry_read(&entry
);
839 pte
= swp_entry_to_pte(entry
);
840 if (pte_swp_soft_dirty(*src_pte
))
841 pte
= pte_swp_mksoft_dirty(pte
);
842 set_pte_at(src_mm
, addr
, src_pte
, pte
);
850 * If it's a COW mapping, write protect it both
851 * in the parent and the child
853 if (is_cow_mapping(vm_flags
)) {
854 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
855 pte
= pte_wrprotect(pte
);
859 * If it's a shared mapping, mark it clean in
862 if (vm_flags
& VM_SHARED
)
863 pte
= pte_mkclean(pte
);
864 pte
= pte_mkold(pte
);
866 page
= vm_normal_page(vma
, addr
, pte
);
877 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
881 int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
882 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
883 unsigned long addr
, unsigned long end
)
885 pte_t
*orig_src_pte
, *orig_dst_pte
;
886 pte_t
*src_pte
, *dst_pte
;
887 spinlock_t
*src_ptl
, *dst_ptl
;
889 int rss
[NR_MM_COUNTERS
];
890 swp_entry_t entry
= (swp_entry_t
){0};
895 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
898 src_pte
= pte_offset_map(src_pmd
, addr
);
899 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
900 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
901 orig_src_pte
= src_pte
;
902 orig_dst_pte
= dst_pte
;
903 arch_enter_lazy_mmu_mode();
907 * We are holding two locks at this point - either of them
908 * could generate latencies in another task on another CPU.
910 if (progress
>= 32) {
912 if (need_resched() ||
913 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
916 if (pte_none(*src_pte
)) {
920 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
925 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
927 arch_leave_lazy_mmu_mode();
928 spin_unlock(src_ptl
);
929 pte_unmap(orig_src_pte
);
930 add_mm_rss_vec(dst_mm
, rss
);
931 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
935 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
944 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
945 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
946 unsigned long addr
, unsigned long end
)
948 pmd_t
*src_pmd
, *dst_pmd
;
951 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
954 src_pmd
= pmd_offset(src_pud
, addr
);
956 next
= pmd_addr_end(addr
, end
);
957 if (pmd_trans_huge(*src_pmd
)) {
959 VM_BUG_ON(next
-addr
!= HPAGE_PMD_SIZE
);
960 err
= copy_huge_pmd(dst_mm
, src_mm
,
961 dst_pmd
, src_pmd
, addr
, vma
);
968 if (pmd_none_or_clear_bad(src_pmd
))
970 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
973 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
977 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
978 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
979 unsigned long addr
, unsigned long end
)
981 pud_t
*src_pud
, *dst_pud
;
984 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
987 src_pud
= pud_offset(src_pgd
, addr
);
989 next
= pud_addr_end(addr
, end
);
990 if (pud_none_or_clear_bad(src_pud
))
992 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
995 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
999 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1000 struct vm_area_struct
*vma
)
1002 pgd_t
*src_pgd
, *dst_pgd
;
1004 unsigned long addr
= vma
->vm_start
;
1005 unsigned long end
= vma
->vm_end
;
1006 unsigned long mmun_start
; /* For mmu_notifiers */
1007 unsigned long mmun_end
; /* For mmu_notifiers */
1012 * Don't copy ptes where a page fault will fill them correctly.
1013 * Fork becomes much lighter when there are big shared or private
1014 * readonly mappings. The tradeoff is that copy_page_range is more
1015 * efficient than faulting.
1017 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_NONLINEAR
|
1018 VM_PFNMAP
| VM_MIXEDMAP
))) {
1023 if (is_vm_hugetlb_page(vma
))
1024 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1026 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
1028 * We do not free on error cases below as remove_vma
1029 * gets called on error from higher level routine
1031 ret
= track_pfn_copy(vma
);
1037 * We need to invalidate the secondary MMU mappings only when
1038 * there could be a permission downgrade on the ptes of the
1039 * parent mm. And a permission downgrade will only happen if
1040 * is_cow_mapping() returns true.
1042 is_cow
= is_cow_mapping(vma
->vm_flags
);
1046 mmu_notifier_invalidate_range_start(src_mm
, mmun_start
,
1050 dst_pgd
= pgd_offset(dst_mm
, addr
);
1051 src_pgd
= pgd_offset(src_mm
, addr
);
1053 next
= pgd_addr_end(addr
, end
);
1054 if (pgd_none_or_clear_bad(src_pgd
))
1056 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1057 vma
, addr
, next
))) {
1061 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1064 mmu_notifier_invalidate_range_end(src_mm
, mmun_start
, mmun_end
);
1068 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1069 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1070 unsigned long addr
, unsigned long end
,
1071 struct zap_details
*details
)
1073 struct mm_struct
*mm
= tlb
->mm
;
1074 int force_flush
= 0;
1075 int rss
[NR_MM_COUNTERS
];
1082 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1084 arch_enter_lazy_mmu_mode();
1087 if (pte_none(ptent
)) {
1091 if (pte_present(ptent
)) {
1094 page
= vm_normal_page(vma
, addr
, ptent
);
1095 if (unlikely(details
) && page
) {
1097 * unmap_shared_mapping_pages() wants to
1098 * invalidate cache without truncating:
1099 * unmap shared but keep private pages.
1101 if (details
->check_mapping
&&
1102 details
->check_mapping
!= page
->mapping
)
1105 * Each page->index must be checked when
1106 * invalidating or truncating nonlinear.
1108 if (details
->nonlinear_vma
&&
1109 (page
->index
< details
->first_index
||
1110 page
->index
> details
->last_index
))
1113 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1115 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1116 if (unlikely(!page
))
1118 if (unlikely(details
) && details
->nonlinear_vma
1119 && linear_page_index(details
->nonlinear_vma
,
1120 addr
) != page
->index
) {
1121 pte_t ptfile
= pgoff_to_pte(page
->index
);
1122 if (pte_soft_dirty(ptent
))
1123 pte_file_mksoft_dirty(ptfile
);
1124 set_pte_at(mm
, addr
, pte
, ptfile
);
1127 rss
[MM_ANONPAGES
]--;
1129 if (pte_dirty(ptent
))
1130 set_page_dirty(page
);
1131 if (pte_young(ptent
) &&
1132 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1133 mark_page_accessed(page
);
1134 rss
[MM_FILEPAGES
]--;
1136 page_remove_rmap(page
);
1137 if (unlikely(page_mapcount(page
) < 0))
1138 print_bad_pte(vma
, addr
, ptent
, page
);
1139 force_flush
= !__tlb_remove_page(tlb
, page
);
1145 * If details->check_mapping, we leave swap entries;
1146 * if details->nonlinear_vma, we leave file entries.
1148 if (unlikely(details
))
1150 if (pte_file(ptent
)) {
1151 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
)))
1152 print_bad_pte(vma
, addr
, ptent
, NULL
);
1154 swp_entry_t entry
= pte_to_swp_entry(ptent
);
1156 if (!non_swap_entry(entry
))
1158 else if (is_migration_entry(entry
)) {
1161 page
= migration_entry_to_page(entry
);
1164 rss
[MM_ANONPAGES
]--;
1166 rss
[MM_FILEPAGES
]--;
1168 if (unlikely(!free_swap_and_cache(entry
)))
1169 print_bad_pte(vma
, addr
, ptent
, NULL
);
1171 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1172 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1174 add_mm_rss_vec(mm
, rss
);
1175 arch_leave_lazy_mmu_mode();
1176 pte_unmap_unlock(start_pte
, ptl
);
1179 * mmu_gather ran out of room to batch pages, we break out of
1180 * the PTE lock to avoid doing the potential expensive TLB invalidate
1181 * and page-free while holding it.
1184 unsigned long old_end
;
1189 * Flush the TLB just for the previous segment,
1190 * then update the range to be the remaining
1208 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1209 struct vm_area_struct
*vma
, pud_t
*pud
,
1210 unsigned long addr
, unsigned long end
,
1211 struct zap_details
*details
)
1216 pmd
= pmd_offset(pud
, addr
);
1218 next
= pmd_addr_end(addr
, end
);
1219 if (pmd_trans_huge(*pmd
)) {
1220 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1221 #ifdef CONFIG_DEBUG_VM
1222 if (!rwsem_is_locked(&tlb
->mm
->mmap_sem
)) {
1223 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1224 __func__
, addr
, end
,
1230 split_huge_page_pmd(vma
, addr
, pmd
);
1231 } else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1236 * Here there can be other concurrent MADV_DONTNEED or
1237 * trans huge page faults running, and if the pmd is
1238 * none or trans huge it can change under us. This is
1239 * because MADV_DONTNEED holds the mmap_sem in read
1242 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1244 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1247 } while (pmd
++, addr
= next
, addr
!= end
);
1252 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1253 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1254 unsigned long addr
, unsigned long end
,
1255 struct zap_details
*details
)
1260 pud
= pud_offset(pgd
, addr
);
1262 next
= pud_addr_end(addr
, end
);
1263 if (pud_none_or_clear_bad(pud
))
1265 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1266 } while (pud
++, addr
= next
, addr
!= end
);
1271 static void unmap_page_range(struct mmu_gather
*tlb
,
1272 struct vm_area_struct
*vma
,
1273 unsigned long addr
, unsigned long end
,
1274 struct zap_details
*details
)
1279 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
1282 BUG_ON(addr
>= end
);
1283 mem_cgroup_uncharge_start();
1284 tlb_start_vma(tlb
, vma
);
1285 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1287 next
= pgd_addr_end(addr
, end
);
1288 if (pgd_none_or_clear_bad(pgd
))
1290 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
1291 } while (pgd
++, addr
= next
, addr
!= end
);
1292 tlb_end_vma(tlb
, vma
);
1293 mem_cgroup_uncharge_end();
1297 static void unmap_single_vma(struct mmu_gather
*tlb
,
1298 struct vm_area_struct
*vma
, unsigned long start_addr
,
1299 unsigned long end_addr
,
1300 struct zap_details
*details
)
1302 unsigned long start
= max(vma
->vm_start
, start_addr
);
1305 if (start
>= vma
->vm_end
)
1307 end
= min(vma
->vm_end
, end_addr
);
1308 if (end
<= vma
->vm_start
)
1312 uprobe_munmap(vma
, start
, end
);
1314 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1315 untrack_pfn(vma
, 0, 0);
1318 if (unlikely(is_vm_hugetlb_page(vma
))) {
1320 * It is undesirable to test vma->vm_file as it
1321 * should be non-null for valid hugetlb area.
1322 * However, vm_file will be NULL in the error
1323 * cleanup path of do_mmap_pgoff. When
1324 * hugetlbfs ->mmap method fails,
1325 * do_mmap_pgoff() nullifies vma->vm_file
1326 * before calling this function to clean up.
1327 * Since no pte has actually been setup, it is
1328 * safe to do nothing in this case.
1331 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
1332 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1333 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
1336 unmap_page_range(tlb
, vma
, start
, end
, details
);
1341 * unmap_vmas - unmap a range of memory covered by a list of vma's
1342 * @tlb: address of the caller's struct mmu_gather
1343 * @vma: the starting vma
1344 * @start_addr: virtual address at which to start unmapping
1345 * @end_addr: virtual address at which to end unmapping
1347 * Unmap all pages in the vma list.
1349 * Only addresses between `start' and `end' will be unmapped.
1351 * The VMA list must be sorted in ascending virtual address order.
1353 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1354 * range after unmap_vmas() returns. So the only responsibility here is to
1355 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1356 * drops the lock and schedules.
1358 void unmap_vmas(struct mmu_gather
*tlb
,
1359 struct vm_area_struct
*vma
, unsigned long start_addr
,
1360 unsigned long end_addr
)
1362 struct mm_struct
*mm
= vma
->vm_mm
;
1364 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1365 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1366 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1367 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1371 * zap_page_range - remove user pages in a given range
1372 * @vma: vm_area_struct holding the applicable pages
1373 * @start: starting address of pages to zap
1374 * @size: number of bytes to zap
1375 * @details: details of nonlinear truncation or shared cache invalidation
1377 * Caller must protect the VMA list
1379 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1380 unsigned long size
, struct zap_details
*details
)
1382 struct mm_struct
*mm
= vma
->vm_mm
;
1383 struct mmu_gather tlb
;
1384 unsigned long end
= start
+ size
;
1387 tlb_gather_mmu(&tlb
, mm
, start
, end
);
1388 update_hiwater_rss(mm
);
1389 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1390 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
)
1391 unmap_single_vma(&tlb
, vma
, start
, end
, details
);
1392 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1393 tlb_finish_mmu(&tlb
, start
, end
);
1397 * zap_page_range_single - remove user pages in a given range
1398 * @vma: vm_area_struct holding the applicable pages
1399 * @address: starting address of pages to zap
1400 * @size: number of bytes to zap
1401 * @details: details of nonlinear truncation or shared cache invalidation
1403 * The range must fit into one VMA.
1405 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1406 unsigned long size
, struct zap_details
*details
)
1408 struct mm_struct
*mm
= vma
->vm_mm
;
1409 struct mmu_gather tlb
;
1410 unsigned long end
= address
+ size
;
1413 tlb_gather_mmu(&tlb
, mm
, address
, end
);
1414 update_hiwater_rss(mm
);
1415 mmu_notifier_invalidate_range_start(mm
, address
, end
);
1416 unmap_single_vma(&tlb
, vma
, address
, end
, details
);
1417 mmu_notifier_invalidate_range_end(mm
, address
, end
);
1418 tlb_finish_mmu(&tlb
, address
, end
);
1422 * zap_vma_ptes - remove ptes mapping the vma
1423 * @vma: vm_area_struct holding ptes to be zapped
1424 * @address: starting address of pages to zap
1425 * @size: number of bytes to zap
1427 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1429 * The entire address range must be fully contained within the vma.
1431 * Returns 0 if successful.
1433 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1436 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1437 !(vma
->vm_flags
& VM_PFNMAP
))
1439 zap_page_range_single(vma
, address
, size
, NULL
);
1442 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1445 * follow_page_mask - look up a page descriptor from a user-virtual address
1446 * @vma: vm_area_struct mapping @address
1447 * @address: virtual address to look up
1448 * @flags: flags modifying lookup behaviour
1449 * @page_mask: on output, *page_mask is set according to the size of the page
1451 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1453 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1454 * an error pointer if there is a mapping to something not represented
1455 * by a page descriptor (see also vm_normal_page()).
1457 struct page
*follow_page_mask(struct vm_area_struct
*vma
,
1458 unsigned long address
, unsigned int flags
,
1459 unsigned int *page_mask
)
1467 struct mm_struct
*mm
= vma
->vm_mm
;
1471 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
1472 if (!IS_ERR(page
)) {
1473 BUG_ON(flags
& FOLL_GET
);
1478 pgd
= pgd_offset(mm
, address
);
1479 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
1482 pud
= pud_offset(pgd
, address
);
1485 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
1486 if (flags
& FOLL_GET
)
1488 page
= follow_huge_pud(mm
, address
, pud
, flags
& FOLL_WRITE
);
1491 if (unlikely(pud_bad(*pud
)))
1494 pmd
= pmd_offset(pud
, address
);
1497 if (pmd_huge(*pmd
) && vma
->vm_flags
& VM_HUGETLB
) {
1498 page
= follow_huge_pmd(mm
, address
, pmd
, flags
& FOLL_WRITE
);
1499 if (flags
& FOLL_GET
) {
1501 * Refcount on tail pages are not well-defined and
1502 * shouldn't be taken. The caller should handle a NULL
1503 * return when trying to follow tail pages.
1514 if ((flags
& FOLL_NUMA
) && pmd_numa(*pmd
))
1516 if (pmd_trans_huge(*pmd
)) {
1517 if (flags
& FOLL_SPLIT
) {
1518 split_huge_page_pmd(vma
, address
, pmd
);
1519 goto split_fallthrough
;
1521 ptl
= pmd_lock(mm
, pmd
);
1522 if (likely(pmd_trans_huge(*pmd
))) {
1523 if (unlikely(pmd_trans_splitting(*pmd
))) {
1525 wait_split_huge_page(vma
->anon_vma
, pmd
);
1527 page
= follow_trans_huge_pmd(vma
, address
,
1530 *page_mask
= HPAGE_PMD_NR
- 1;
1538 if (unlikely(pmd_bad(*pmd
)))
1541 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1544 if (!pte_present(pte
)) {
1547 * KSM's break_ksm() relies upon recognizing a ksm page
1548 * even while it is being migrated, so for that case we
1549 * need migration_entry_wait().
1551 if (likely(!(flags
& FOLL_MIGRATION
)))
1553 if (pte_none(pte
) || pte_file(pte
))
1555 entry
= pte_to_swp_entry(pte
);
1556 if (!is_migration_entry(entry
))
1558 pte_unmap_unlock(ptep
, ptl
);
1559 migration_entry_wait(mm
, pmd
, address
);
1560 goto split_fallthrough
;
1562 if ((flags
& FOLL_NUMA
) && pte_numa(pte
))
1564 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
1567 page
= vm_normal_page(vma
, address
, pte
);
1568 if (unlikely(!page
)) {
1569 if ((flags
& FOLL_DUMP
) ||
1570 !is_zero_pfn(pte_pfn(pte
)))
1572 page
= pte_page(pte
);
1575 if (flags
& FOLL_GET
)
1576 get_page_foll(page
);
1577 if (flags
& FOLL_TOUCH
) {
1578 if ((flags
& FOLL_WRITE
) &&
1579 !pte_dirty(pte
) && !PageDirty(page
))
1580 set_page_dirty(page
);
1582 * pte_mkyoung() would be more correct here, but atomic care
1583 * is needed to avoid losing the dirty bit: it is easier to use
1584 * mark_page_accessed().
1586 mark_page_accessed(page
);
1588 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1590 * The preliminary mapping check is mainly to avoid the
1591 * pointless overhead of lock_page on the ZERO_PAGE
1592 * which might bounce very badly if there is contention.
1594 * If the page is already locked, we don't need to
1595 * handle it now - vmscan will handle it later if and
1596 * when it attempts to reclaim the page.
1598 if (page
->mapping
&& trylock_page(page
)) {
1599 lru_add_drain(); /* push cached pages to LRU */
1601 * Because we lock page here, and migration is
1602 * blocked by the pte's page reference, and we
1603 * know the page is still mapped, we don't even
1604 * need to check for file-cache page truncation.
1606 mlock_vma_page(page
);
1611 pte_unmap_unlock(ptep
, ptl
);
1616 pte_unmap_unlock(ptep
, ptl
);
1617 return ERR_PTR(-EFAULT
);
1620 pte_unmap_unlock(ptep
, ptl
);
1626 * When core dumping an enormous anonymous area that nobody
1627 * has touched so far, we don't want to allocate unnecessary pages or
1628 * page tables. Return error instead of NULL to skip handle_mm_fault,
1629 * then get_dump_page() will return NULL to leave a hole in the dump.
1630 * But we can only make this optimization where a hole would surely
1631 * be zero-filled if handle_mm_fault() actually did handle it.
1633 if ((flags
& FOLL_DUMP
) &&
1634 (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
1635 return ERR_PTR(-EFAULT
);
1639 static inline int stack_guard_page(struct vm_area_struct
*vma
, unsigned long addr
)
1641 return stack_guard_page_start(vma
, addr
) ||
1642 stack_guard_page_end(vma
, addr
+PAGE_SIZE
);
1646 * __get_user_pages() - pin user pages in memory
1647 * @tsk: task_struct of target task
1648 * @mm: mm_struct of target mm
1649 * @start: starting user address
1650 * @nr_pages: number of pages from start to pin
1651 * @gup_flags: flags modifying pin behaviour
1652 * @pages: array that receives pointers to the pages pinned.
1653 * Should be at least nr_pages long. Or NULL, if caller
1654 * only intends to ensure the pages are faulted in.
1655 * @vmas: array of pointers to vmas corresponding to each page.
1656 * Or NULL if the caller does not require them.
1657 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1659 * Returns number of pages pinned. This may be fewer than the number
1660 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1661 * were pinned, returns -errno. Each page returned must be released
1662 * with a put_page() call when it is finished with. vmas will only
1663 * remain valid while mmap_sem is held.
1665 * Must be called with mmap_sem held for read or write.
1667 * __get_user_pages walks a process's page tables and takes a reference to
1668 * each struct page that each user address corresponds to at a given
1669 * instant. That is, it takes the page that would be accessed if a user
1670 * thread accesses the given user virtual address at that instant.
1672 * This does not guarantee that the page exists in the user mappings when
1673 * __get_user_pages returns, and there may even be a completely different
1674 * page there in some cases (eg. if mmapped pagecache has been invalidated
1675 * and subsequently re faulted). However it does guarantee that the page
1676 * won't be freed completely. And mostly callers simply care that the page
1677 * contains data that was valid *at some point in time*. Typically, an IO
1678 * or similar operation cannot guarantee anything stronger anyway because
1679 * locks can't be held over the syscall boundary.
1681 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1682 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1683 * appropriate) must be called after the page is finished with, and
1684 * before put_page is called.
1686 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1687 * or mmap_sem contention, and if waiting is needed to pin all pages,
1688 * *@nonblocking will be set to 0.
1690 * In most cases, get_user_pages or get_user_pages_fast should be used
1691 * instead of __get_user_pages. __get_user_pages should be used only if
1692 * you need some special @gup_flags.
1694 long __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1695 unsigned long start
, unsigned long nr_pages
,
1696 unsigned int gup_flags
, struct page
**pages
,
1697 struct vm_area_struct
**vmas
, int *nonblocking
)
1700 unsigned long vm_flags
;
1701 unsigned int page_mask
;
1706 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
1709 * Require read or write permissions.
1710 * If FOLL_FORCE is set, we only require the "MAY" flags.
1712 vm_flags
= (gup_flags
& FOLL_WRITE
) ?
1713 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1714 vm_flags
&= (gup_flags
& FOLL_FORCE
) ?
1715 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1718 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1719 * would be called on PROT_NONE ranges. We must never invoke
1720 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1721 * page faults would unprotect the PROT_NONE ranges if
1722 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1723 * bitflag. So to avoid that, don't set FOLL_NUMA if
1724 * FOLL_FORCE is set.
1726 if (!(gup_flags
& FOLL_FORCE
))
1727 gup_flags
|= FOLL_NUMA
;
1732 struct vm_area_struct
*vma
;
1734 vma
= find_extend_vma(mm
, start
);
1735 if (!vma
&& in_gate_area(mm
, start
)) {
1736 unsigned long pg
= start
& PAGE_MASK
;
1742 /* user gate pages are read-only */
1743 if (gup_flags
& FOLL_WRITE
)
1744 return i
? : -EFAULT
;
1746 pgd
= pgd_offset_k(pg
);
1748 pgd
= pgd_offset_gate(mm
, pg
);
1749 BUG_ON(pgd_none(*pgd
));
1750 pud
= pud_offset(pgd
, pg
);
1751 BUG_ON(pud_none(*pud
));
1752 pmd
= pmd_offset(pud
, pg
);
1754 return i
? : -EFAULT
;
1755 VM_BUG_ON(pmd_trans_huge(*pmd
));
1756 pte
= pte_offset_map(pmd
, pg
);
1757 if (pte_none(*pte
)) {
1759 return i
? : -EFAULT
;
1761 vma
= get_gate_vma(mm
);
1765 page
= vm_normal_page(vma
, start
, *pte
);
1767 if (!(gup_flags
& FOLL_DUMP
) &&
1768 is_zero_pfn(pte_pfn(*pte
)))
1769 page
= pte_page(*pte
);
1772 return i
? : -EFAULT
;
1784 (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1785 !(vm_flags
& vma
->vm_flags
))
1786 return i
? : -EFAULT
;
1788 if (is_vm_hugetlb_page(vma
)) {
1789 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1790 &start
, &nr_pages
, i
, gup_flags
);
1796 unsigned int foll_flags
= gup_flags
;
1797 unsigned int page_increm
;
1800 * If we have a pending SIGKILL, don't keep faulting
1801 * pages and potentially allocating memory.
1803 if (unlikely(fatal_signal_pending(current
)))
1804 return i
? i
: -ERESTARTSYS
;
1807 while (!(page
= follow_page_mask(vma
, start
,
1808 foll_flags
, &page_mask
))) {
1810 unsigned int fault_flags
= 0;
1812 /* For mlock, just skip the stack guard page. */
1813 if (foll_flags
& FOLL_MLOCK
) {
1814 if (stack_guard_page(vma
, start
))
1817 if (foll_flags
& FOLL_WRITE
)
1818 fault_flags
|= FAULT_FLAG_WRITE
;
1820 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
1821 if (foll_flags
& FOLL_NOWAIT
)
1822 fault_flags
|= (FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
);
1824 ret
= handle_mm_fault(mm
, vma
, start
,
1827 if (ret
& VM_FAULT_ERROR
) {
1828 if (ret
& VM_FAULT_OOM
)
1829 return i
? i
: -ENOMEM
;
1830 if (ret
& (VM_FAULT_HWPOISON
|
1831 VM_FAULT_HWPOISON_LARGE
)) {
1834 else if (gup_flags
& FOLL_HWPOISON
)
1839 if (ret
& VM_FAULT_SIGBUS
)
1840 return i
? i
: -EFAULT
;
1845 if (ret
& VM_FAULT_MAJOR
)
1851 if (ret
& VM_FAULT_RETRY
) {
1858 * The VM_FAULT_WRITE bit tells us that
1859 * do_wp_page has broken COW when necessary,
1860 * even if maybe_mkwrite decided not to set
1861 * pte_write. We can thus safely do subsequent
1862 * page lookups as if they were reads. But only
1863 * do so when looping for pte_write is futile:
1864 * in some cases userspace may also be wanting
1865 * to write to the gotten user page, which a
1866 * read fault here might prevent (a readonly
1867 * page might get reCOWed by userspace write).
1869 if ((ret
& VM_FAULT_WRITE
) &&
1870 !(vma
->vm_flags
& VM_WRITE
))
1871 foll_flags
&= ~FOLL_WRITE
;
1876 return i
? i
: PTR_ERR(page
);
1880 flush_anon_page(vma
, page
, start
);
1881 flush_dcache_page(page
);
1889 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & page_mask
);
1890 if (page_increm
> nr_pages
)
1891 page_increm
= nr_pages
;
1893 start
+= page_increm
* PAGE_SIZE
;
1894 nr_pages
-= page_increm
;
1895 } while (nr_pages
&& start
< vma
->vm_end
);
1899 EXPORT_SYMBOL(__get_user_pages
);
1902 * fixup_user_fault() - manually resolve a user page fault
1903 * @tsk: the task_struct to use for page fault accounting, or
1904 * NULL if faults are not to be recorded.
1905 * @mm: mm_struct of target mm
1906 * @address: user address
1907 * @fault_flags:flags to pass down to handle_mm_fault()
1909 * This is meant to be called in the specific scenario where for locking reasons
1910 * we try to access user memory in atomic context (within a pagefault_disable()
1911 * section), this returns -EFAULT, and we want to resolve the user fault before
1914 * Typically this is meant to be used by the futex code.
1916 * The main difference with get_user_pages() is that this function will
1917 * unconditionally call handle_mm_fault() which will in turn perform all the
1918 * necessary SW fixup of the dirty and young bits in the PTE, while
1919 * handle_mm_fault() only guarantees to update these in the struct page.
1921 * This is important for some architectures where those bits also gate the
1922 * access permission to the page because they are maintained in software. On
1923 * such architectures, gup() will not be enough to make a subsequent access
1926 * This should be called with the mm_sem held for read.
1928 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
1929 unsigned long address
, unsigned int fault_flags
)
1931 struct vm_area_struct
*vma
;
1934 vma
= find_extend_vma(mm
, address
);
1935 if (!vma
|| address
< vma
->vm_start
)
1938 ret
= handle_mm_fault(mm
, vma
, address
, fault_flags
);
1939 if (ret
& VM_FAULT_ERROR
) {
1940 if (ret
& VM_FAULT_OOM
)
1942 if (ret
& (VM_FAULT_HWPOISON
| VM_FAULT_HWPOISON_LARGE
))
1944 if (ret
& VM_FAULT_SIGBUS
)
1949 if (ret
& VM_FAULT_MAJOR
)
1958 * get_user_pages() - pin user pages in memory
1959 * @tsk: the task_struct to use for page fault accounting, or
1960 * NULL if faults are not to be recorded.
1961 * @mm: mm_struct of target mm
1962 * @start: starting user address
1963 * @nr_pages: number of pages from start to pin
1964 * @write: whether pages will be written to by the caller
1965 * @force: whether to force write access even if user mapping is
1966 * readonly. This will result in the page being COWed even
1967 * in MAP_SHARED mappings. You do not want this.
1968 * @pages: array that receives pointers to the pages pinned.
1969 * Should be at least nr_pages long. Or NULL, if caller
1970 * only intends to ensure the pages are faulted in.
1971 * @vmas: array of pointers to vmas corresponding to each page.
1972 * Or NULL if the caller does not require them.
1974 * Returns number of pages pinned. This may be fewer than the number
1975 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1976 * were pinned, returns -errno. Each page returned must be released
1977 * with a put_page() call when it is finished with. vmas will only
1978 * remain valid while mmap_sem is held.
1980 * Must be called with mmap_sem held for read or write.
1982 * get_user_pages walks a process's page tables and takes a reference to
1983 * each struct page that each user address corresponds to at a given
1984 * instant. That is, it takes the page that would be accessed if a user
1985 * thread accesses the given user virtual address at that instant.
1987 * This does not guarantee that the page exists in the user mappings when
1988 * get_user_pages returns, and there may even be a completely different
1989 * page there in some cases (eg. if mmapped pagecache has been invalidated
1990 * and subsequently re faulted). However it does guarantee that the page
1991 * won't be freed completely. And mostly callers simply care that the page
1992 * contains data that was valid *at some point in time*. Typically, an IO
1993 * or similar operation cannot guarantee anything stronger anyway because
1994 * locks can't be held over the syscall boundary.
1996 * If write=0, the page must not be written to. If the page is written to,
1997 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1998 * after the page is finished with, and before put_page is called.
2000 * get_user_pages is typically used for fewer-copy IO operations, to get a
2001 * handle on the memory by some means other than accesses via the user virtual
2002 * addresses. The pages may be submitted for DMA to devices or accessed via
2003 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2004 * use the correct cache flushing APIs.
2006 * See also get_user_pages_fast, for performance critical applications.
2008 long get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
2009 unsigned long start
, unsigned long nr_pages
, int write
,
2010 int force
, struct page
**pages
, struct vm_area_struct
**vmas
)
2012 int flags
= FOLL_TOUCH
;
2017 flags
|= FOLL_WRITE
;
2019 flags
|= FOLL_FORCE
;
2021 return __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
, vmas
,
2024 EXPORT_SYMBOL(get_user_pages
);
2027 * get_dump_page() - pin user page in memory while writing it to core dump
2028 * @addr: user address
2030 * Returns struct page pointer of user page pinned for dump,
2031 * to be freed afterwards by page_cache_release() or put_page().
2033 * Returns NULL on any kind of failure - a hole must then be inserted into
2034 * the corefile, to preserve alignment with its headers; and also returns
2035 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2036 * allowing a hole to be left in the corefile to save diskspace.
2038 * Called without mmap_sem, but after all other threads have been killed.
2040 #ifdef CONFIG_ELF_CORE
2041 struct page
*get_dump_page(unsigned long addr
)
2043 struct vm_area_struct
*vma
;
2046 if (__get_user_pages(current
, current
->mm
, addr
, 1,
2047 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
2050 flush_cache_page(vma
, addr
, page_to_pfn(page
));
2053 #endif /* CONFIG_ELF_CORE */
2055 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
2058 pgd_t
* pgd
= pgd_offset(mm
, addr
);
2059 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
2061 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
2063 VM_BUG_ON(pmd_trans_huge(*pmd
));
2064 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
2071 * This is the old fallback for page remapping.
2073 * For historical reasons, it only allows reserved pages. Only
2074 * old drivers should use this, and they needed to mark their
2075 * pages reserved for the old functions anyway.
2077 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2078 struct page
*page
, pgprot_t prot
)
2080 struct mm_struct
*mm
= vma
->vm_mm
;
2089 flush_dcache_page(page
);
2090 pte
= get_locked_pte(mm
, addr
, &ptl
);
2094 if (!pte_none(*pte
))
2097 /* Ok, finally just insert the thing.. */
2099 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
2100 page_add_file_rmap(page
);
2101 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
2104 pte_unmap_unlock(pte
, ptl
);
2107 pte_unmap_unlock(pte
, ptl
);
2113 * vm_insert_page - insert single page into user vma
2114 * @vma: user vma to map to
2115 * @addr: target user address of this page
2116 * @page: source kernel page
2118 * This allows drivers to insert individual pages they've allocated
2121 * The page has to be a nice clean _individual_ kernel allocation.
2122 * If you allocate a compound page, you need to have marked it as
2123 * such (__GFP_COMP), or manually just split the page up yourself
2124 * (see split_page()).
2126 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2127 * took an arbitrary page protection parameter. This doesn't allow
2128 * that. Your vma protection will have to be set up correctly, which
2129 * means that if you want a shared writable mapping, you'd better
2130 * ask for a shared writable mapping!
2132 * The page does not need to be reserved.
2134 * Usually this function is called from f_op->mmap() handler
2135 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2136 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2137 * function from other places, for example from page-fault handler.
2139 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2142 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2144 if (!page_count(page
))
2146 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
2147 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
2148 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
2149 vma
->vm_flags
|= VM_MIXEDMAP
;
2151 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2153 EXPORT_SYMBOL(vm_insert_page
);
2155 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2156 unsigned long pfn
, pgprot_t prot
)
2158 struct mm_struct
*mm
= vma
->vm_mm
;
2164 pte
= get_locked_pte(mm
, addr
, &ptl
);
2168 if (!pte_none(*pte
))
2171 /* Ok, finally just insert the thing.. */
2172 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
2173 set_pte_at(mm
, addr
, pte
, entry
);
2174 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
2178 pte_unmap_unlock(pte
, ptl
);
2184 * vm_insert_pfn - insert single pfn into user vma
2185 * @vma: user vma to map to
2186 * @addr: target user address of this page
2187 * @pfn: source kernel pfn
2189 * Similar to vm_insert_page, this allows drivers to insert individual pages
2190 * they've allocated into a user vma. Same comments apply.
2192 * This function should only be called from a vm_ops->fault handler, and
2193 * in that case the handler should return NULL.
2195 * vma cannot be a COW mapping.
2197 * As this is called only for pages that do not currently exist, we
2198 * do not need to flush old virtual caches or the TLB.
2200 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2204 pgprot_t pgprot
= vma
->vm_page_prot
;
2206 * Technically, architectures with pte_special can avoid all these
2207 * restrictions (same for remap_pfn_range). However we would like
2208 * consistency in testing and feature parity among all, so we should
2209 * try to keep these invariants in place for everybody.
2211 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
2212 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
2213 (VM_PFNMAP
|VM_MIXEDMAP
));
2214 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
2215 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
2217 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2219 if (track_pfn_insert(vma
, &pgprot
, pfn
))
2222 ret
= insert_pfn(vma
, addr
, pfn
, pgprot
);
2226 EXPORT_SYMBOL(vm_insert_pfn
);
2228 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2231 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
2233 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2237 * If we don't have pte special, then we have to use the pfn_valid()
2238 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2239 * refcount the page if pfn_valid is true (hence insert_page rather
2240 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2241 * without pte special, it would there be refcounted as a normal page.
2243 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
2246 page
= pfn_to_page(pfn
);
2247 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2249 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
2251 EXPORT_SYMBOL(vm_insert_mixed
);
2254 * maps a range of physical memory into the requested pages. the old
2255 * mappings are removed. any references to nonexistent pages results
2256 * in null mappings (currently treated as "copy-on-access")
2258 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2259 unsigned long addr
, unsigned long end
,
2260 unsigned long pfn
, pgprot_t prot
)
2265 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2268 arch_enter_lazy_mmu_mode();
2270 BUG_ON(!pte_none(*pte
));
2271 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2273 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2274 arch_leave_lazy_mmu_mode();
2275 pte_unmap_unlock(pte
- 1, ptl
);
2279 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2280 unsigned long addr
, unsigned long end
,
2281 unsigned long pfn
, pgprot_t prot
)
2286 pfn
-= addr
>> PAGE_SHIFT
;
2287 pmd
= pmd_alloc(mm
, pud
, addr
);
2290 VM_BUG_ON(pmd_trans_huge(*pmd
));
2292 next
= pmd_addr_end(addr
, end
);
2293 if (remap_pte_range(mm
, pmd
, addr
, next
,
2294 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
2296 } while (pmd
++, addr
= next
, addr
!= end
);
2300 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2301 unsigned long addr
, unsigned long end
,
2302 unsigned long pfn
, pgprot_t prot
)
2307 pfn
-= addr
>> PAGE_SHIFT
;
2308 pud
= pud_alloc(mm
, pgd
, addr
);
2312 next
= pud_addr_end(addr
, end
);
2313 if (remap_pmd_range(mm
, pud
, addr
, next
,
2314 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
2316 } while (pud
++, addr
= next
, addr
!= end
);
2321 * remap_pfn_range - remap kernel memory to userspace
2322 * @vma: user vma to map to
2323 * @addr: target user address to start at
2324 * @pfn: physical address of kernel memory
2325 * @size: size of map area
2326 * @prot: page protection flags for this mapping
2328 * Note: this is only safe if the mm semaphore is held when called.
2330 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2331 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2335 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2336 struct mm_struct
*mm
= vma
->vm_mm
;
2340 * Physically remapped pages are special. Tell the
2341 * rest of the world about it:
2342 * VM_IO tells people not to look at these pages
2343 * (accesses can have side effects).
2344 * VM_PFNMAP tells the core MM that the base pages are just
2345 * raw PFN mappings, and do not have a "struct page" associated
2348 * Disable vma merging and expanding with mremap().
2350 * Omit vma from core dump, even when VM_IO turned off.
2352 * There's a horrible special case to handle copy-on-write
2353 * behaviour that some programs depend on. We mark the "original"
2354 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2355 * See vm_normal_page() for details.
2357 if (is_cow_mapping(vma
->vm_flags
)) {
2358 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2360 vma
->vm_pgoff
= pfn
;
2363 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
2367 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
2369 BUG_ON(addr
>= end
);
2370 pfn
-= addr
>> PAGE_SHIFT
;
2371 pgd
= pgd_offset(mm
, addr
);
2372 flush_cache_range(vma
, addr
, end
);
2374 next
= pgd_addr_end(addr
, end
);
2375 err
= remap_pud_range(mm
, pgd
, addr
, next
,
2376 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2379 } while (pgd
++, addr
= next
, addr
!= end
);
2382 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
));
2386 EXPORT_SYMBOL(remap_pfn_range
);
2389 * vm_iomap_memory - remap memory to userspace
2390 * @vma: user vma to map to
2391 * @start: start of area
2392 * @len: size of area
2394 * This is a simplified io_remap_pfn_range() for common driver use. The
2395 * driver just needs to give us the physical memory range to be mapped,
2396 * we'll figure out the rest from the vma information.
2398 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2399 * whatever write-combining details or similar.
2401 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2403 unsigned long vm_len
, pfn
, pages
;
2405 /* Check that the physical memory area passed in looks valid */
2406 if (start
+ len
< start
)
2409 * You *really* shouldn't map things that aren't page-aligned,
2410 * but we've historically allowed it because IO memory might
2411 * just have smaller alignment.
2413 len
+= start
& ~PAGE_MASK
;
2414 pfn
= start
>> PAGE_SHIFT
;
2415 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2416 if (pfn
+ pages
< pfn
)
2419 /* We start the mapping 'vm_pgoff' pages into the area */
2420 if (vma
->vm_pgoff
> pages
)
2422 pfn
+= vma
->vm_pgoff
;
2423 pages
-= vma
->vm_pgoff
;
2425 /* Can we fit all of the mapping? */
2426 vm_len
= vma
->vm_end
- vma
->vm_start
;
2427 if (vm_len
>> PAGE_SHIFT
> pages
)
2430 /* Ok, let it rip */
2431 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2433 EXPORT_SYMBOL(vm_iomap_memory
);
2435 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2436 unsigned long addr
, unsigned long end
,
2437 pte_fn_t fn
, void *data
)
2442 spinlock_t
*uninitialized_var(ptl
);
2444 pte
= (mm
== &init_mm
) ?
2445 pte_alloc_kernel(pmd
, addr
) :
2446 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2450 BUG_ON(pmd_huge(*pmd
));
2452 arch_enter_lazy_mmu_mode();
2454 token
= pmd_pgtable(*pmd
);
2457 err
= fn(pte
++, token
, addr
, data
);
2460 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2462 arch_leave_lazy_mmu_mode();
2465 pte_unmap_unlock(pte
-1, ptl
);
2469 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2470 unsigned long addr
, unsigned long end
,
2471 pte_fn_t fn
, void *data
)
2477 BUG_ON(pud_huge(*pud
));
2479 pmd
= pmd_alloc(mm
, pud
, addr
);
2483 next
= pmd_addr_end(addr
, end
);
2484 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
2487 } while (pmd
++, addr
= next
, addr
!= end
);
2491 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2492 unsigned long addr
, unsigned long end
,
2493 pte_fn_t fn
, void *data
)
2499 pud
= pud_alloc(mm
, pgd
, addr
);
2503 next
= pud_addr_end(addr
, end
);
2504 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
2507 } while (pud
++, addr
= next
, addr
!= end
);
2512 * Scan a region of virtual memory, filling in page tables as necessary
2513 * and calling a provided function on each leaf page table.
2515 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2516 unsigned long size
, pte_fn_t fn
, void *data
)
2520 unsigned long end
= addr
+ size
;
2523 BUG_ON(addr
>= end
);
2524 pgd
= pgd_offset(mm
, addr
);
2526 next
= pgd_addr_end(addr
, end
);
2527 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
2530 } while (pgd
++, addr
= next
, addr
!= end
);
2534 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2537 * handle_pte_fault chooses page fault handler according to an entry
2538 * which was read non-atomically. Before making any commitment, on
2539 * those architectures or configurations (e.g. i386 with PAE) which
2540 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2541 * must check under lock before unmapping the pte and proceeding
2542 * (but do_wp_page is only called after already making such a check;
2543 * and do_anonymous_page can safely check later on).
2545 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2546 pte_t
*page_table
, pte_t orig_pte
)
2549 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2550 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2551 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2553 same
= pte_same(*page_table
, orig_pte
);
2557 pte_unmap(page_table
);
2561 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
2563 debug_dma_assert_idle(src
);
2566 * If the source page was a PFN mapping, we don't have
2567 * a "struct page" for it. We do a best-effort copy by
2568 * just copying from the original user address. If that
2569 * fails, we just zero-fill it. Live with it.
2571 if (unlikely(!src
)) {
2572 void *kaddr
= kmap_atomic(dst
);
2573 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
2576 * This really shouldn't fail, because the page is there
2577 * in the page tables. But it might just be unreadable,
2578 * in which case we just give up and fill the result with
2581 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
2583 kunmap_atomic(kaddr
);
2584 flush_dcache_page(dst
);
2586 copy_user_highpage(dst
, src
, va
, vma
);
2590 * This routine handles present pages, when users try to write
2591 * to a shared page. It is done by copying the page to a new address
2592 * and decrementing the shared-page counter for the old page.
2594 * Note that this routine assumes that the protection checks have been
2595 * done by the caller (the low-level page fault routine in most cases).
2596 * Thus we can safely just mark it writable once we've done any necessary
2599 * We also mark the page dirty at this point even though the page will
2600 * change only once the write actually happens. This avoids a few races,
2601 * and potentially makes it more efficient.
2603 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2604 * but allow concurrent faults), with pte both mapped and locked.
2605 * We return with mmap_sem still held, but pte unmapped and unlocked.
2607 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2608 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2609 spinlock_t
*ptl
, pte_t orig_pte
)
2612 struct page
*old_page
, *new_page
= NULL
;
2615 int page_mkwrite
= 0;
2616 struct page
*dirty_page
= NULL
;
2617 unsigned long mmun_start
= 0; /* For mmu_notifiers */
2618 unsigned long mmun_end
= 0; /* For mmu_notifiers */
2620 old_page
= vm_normal_page(vma
, address
, orig_pte
);
2623 * VM_MIXEDMAP !pfn_valid() case
2625 * We should not cow pages in a shared writeable mapping.
2626 * Just mark the pages writable as we can't do any dirty
2627 * accounting on raw pfn maps.
2629 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2630 (VM_WRITE
|VM_SHARED
))
2636 * Take out anonymous pages first, anonymous shared vmas are
2637 * not dirty accountable.
2639 if (PageAnon(old_page
) && !PageKsm(old_page
)) {
2640 if (!trylock_page(old_page
)) {
2641 page_cache_get(old_page
);
2642 pte_unmap_unlock(page_table
, ptl
);
2643 lock_page(old_page
);
2644 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2646 if (!pte_same(*page_table
, orig_pte
)) {
2647 unlock_page(old_page
);
2650 page_cache_release(old_page
);
2652 if (reuse_swap_page(old_page
)) {
2654 * The page is all ours. Move it to our anon_vma so
2655 * the rmap code will not search our parent or siblings.
2656 * Protected against the rmap code by the page lock.
2658 page_move_anon_rmap(old_page
, vma
, address
);
2659 unlock_page(old_page
);
2662 unlock_page(old_page
);
2663 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2664 (VM_WRITE
|VM_SHARED
))) {
2666 * Only catch write-faults on shared writable pages,
2667 * read-only shared pages can get COWed by
2668 * get_user_pages(.write=1, .force=1).
2670 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2671 struct vm_fault vmf
;
2674 vmf
.virtual_address
= (void __user
*)(address
&
2676 vmf
.pgoff
= old_page
->index
;
2677 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2678 vmf
.page
= old_page
;
2681 * Notify the address space that the page is about to
2682 * become writable so that it can prohibit this or wait
2683 * for the page to get into an appropriate state.
2685 * We do this without the lock held, so that it can
2686 * sleep if it needs to.
2688 page_cache_get(old_page
);
2689 pte_unmap_unlock(page_table
, ptl
);
2691 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
2693 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
2695 goto unwritable_page
;
2697 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
2698 lock_page(old_page
);
2699 if (!old_page
->mapping
) {
2700 ret
= 0; /* retry the fault */
2701 unlock_page(old_page
);
2702 goto unwritable_page
;
2705 VM_BUG_ON_PAGE(!PageLocked(old_page
), old_page
);
2708 * Since we dropped the lock we need to revalidate
2709 * the PTE as someone else may have changed it. If
2710 * they did, we just return, as we can count on the
2711 * MMU to tell us if they didn't also make it writable.
2713 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2715 if (!pte_same(*page_table
, orig_pte
)) {
2716 unlock_page(old_page
);
2722 dirty_page
= old_page
;
2723 get_page(dirty_page
);
2727 * Clear the pages cpupid information as the existing
2728 * information potentially belongs to a now completely
2729 * unrelated process.
2732 page_cpupid_xchg_last(old_page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2734 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2735 entry
= pte_mkyoung(orig_pte
);
2736 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2737 if (ptep_set_access_flags(vma
, address
, page_table
, entry
,1))
2738 update_mmu_cache(vma
, address
, page_table
);
2739 pte_unmap_unlock(page_table
, ptl
);
2740 ret
|= VM_FAULT_WRITE
;
2746 * Yes, Virginia, this is actually required to prevent a race
2747 * with clear_page_dirty_for_io() from clearing the page dirty
2748 * bit after it clear all dirty ptes, but before a racing
2749 * do_wp_page installs a dirty pte.
2751 * __do_fault is protected similarly.
2753 if (!page_mkwrite
) {
2754 wait_on_page_locked(dirty_page
);
2755 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2756 /* file_update_time outside page_lock */
2758 file_update_time(vma
->vm_file
);
2760 put_page(dirty_page
);
2762 struct address_space
*mapping
= dirty_page
->mapping
;
2764 set_page_dirty(dirty_page
);
2765 unlock_page(dirty_page
);
2766 page_cache_release(dirty_page
);
2769 * Some device drivers do not set page.mapping
2770 * but still dirty their pages
2772 balance_dirty_pages_ratelimited(mapping
);
2780 * Ok, we need to copy. Oh, well..
2782 page_cache_get(old_page
);
2784 pte_unmap_unlock(page_table
, ptl
);
2786 if (unlikely(anon_vma_prepare(vma
)))
2789 if (is_zero_pfn(pte_pfn(orig_pte
))) {
2790 new_page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2794 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2797 cow_user_page(new_page
, old_page
, address
, vma
);
2799 __SetPageUptodate(new_page
);
2801 if (mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
))
2804 mmun_start
= address
& PAGE_MASK
;
2805 mmun_end
= mmun_start
+ PAGE_SIZE
;
2806 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2809 * Re-check the pte - we dropped the lock
2811 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2812 if (likely(pte_same(*page_table
, orig_pte
))) {
2814 if (!PageAnon(old_page
)) {
2815 dec_mm_counter_fast(mm
, MM_FILEPAGES
);
2816 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2819 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2820 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2821 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2822 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2824 * Clear the pte entry and flush it first, before updating the
2825 * pte with the new entry. This will avoid a race condition
2826 * seen in the presence of one thread doing SMC and another
2829 ptep_clear_flush(vma
, address
, page_table
);
2830 page_add_new_anon_rmap(new_page
, vma
, address
);
2832 * We call the notify macro here because, when using secondary
2833 * mmu page tables (such as kvm shadow page tables), we want the
2834 * new page to be mapped directly into the secondary page table.
2836 set_pte_at_notify(mm
, address
, page_table
, entry
);
2837 update_mmu_cache(vma
, address
, page_table
);
2840 * Only after switching the pte to the new page may
2841 * we remove the mapcount here. Otherwise another
2842 * process may come and find the rmap count decremented
2843 * before the pte is switched to the new page, and
2844 * "reuse" the old page writing into it while our pte
2845 * here still points into it and can be read by other
2848 * The critical issue is to order this
2849 * page_remove_rmap with the ptp_clear_flush above.
2850 * Those stores are ordered by (if nothing else,)
2851 * the barrier present in the atomic_add_negative
2852 * in page_remove_rmap.
2854 * Then the TLB flush in ptep_clear_flush ensures that
2855 * no process can access the old page before the
2856 * decremented mapcount is visible. And the old page
2857 * cannot be reused until after the decremented
2858 * mapcount is visible. So transitively, TLBs to
2859 * old page will be flushed before it can be reused.
2861 page_remove_rmap(old_page
);
2864 /* Free the old page.. */
2865 new_page
= old_page
;
2866 ret
|= VM_FAULT_WRITE
;
2868 mem_cgroup_uncharge_page(new_page
);
2871 page_cache_release(new_page
);
2873 pte_unmap_unlock(page_table
, ptl
);
2874 if (mmun_end
> mmun_start
)
2875 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2878 * Don't let another task, with possibly unlocked vma,
2879 * keep the mlocked page.
2881 if ((ret
& VM_FAULT_WRITE
) && (vma
->vm_flags
& VM_LOCKED
)) {
2882 lock_page(old_page
); /* LRU manipulation */
2883 munlock_vma_page(old_page
);
2884 unlock_page(old_page
);
2886 page_cache_release(old_page
);
2890 page_cache_release(new_page
);
2893 page_cache_release(old_page
);
2894 return VM_FAULT_OOM
;
2897 page_cache_release(old_page
);
2901 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2902 unsigned long start_addr
, unsigned long end_addr
,
2903 struct zap_details
*details
)
2905 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2908 static inline void unmap_mapping_range_tree(struct rb_root
*root
,
2909 struct zap_details
*details
)
2911 struct vm_area_struct
*vma
;
2912 pgoff_t vba
, vea
, zba
, zea
;
2914 vma_interval_tree_foreach(vma
, root
,
2915 details
->first_index
, details
->last_index
) {
2917 vba
= vma
->vm_pgoff
;
2918 vea
= vba
+ vma_pages(vma
) - 1;
2919 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2920 zba
= details
->first_index
;
2923 zea
= details
->last_index
;
2927 unmap_mapping_range_vma(vma
,
2928 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2929 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2934 static inline void unmap_mapping_range_list(struct list_head
*head
,
2935 struct zap_details
*details
)
2937 struct vm_area_struct
*vma
;
2940 * In nonlinear VMAs there is no correspondence between virtual address
2941 * offset and file offset. So we must perform an exhaustive search
2942 * across *all* the pages in each nonlinear VMA, not just the pages
2943 * whose virtual address lies outside the file truncation point.
2945 list_for_each_entry(vma
, head
, shared
.nonlinear
) {
2946 details
->nonlinear_vma
= vma
;
2947 unmap_mapping_range_vma(vma
, vma
->vm_start
, vma
->vm_end
, details
);
2952 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2953 * @mapping: the address space containing mmaps to be unmapped.
2954 * @holebegin: byte in first page to unmap, relative to the start of
2955 * the underlying file. This will be rounded down to a PAGE_SIZE
2956 * boundary. Note that this is different from truncate_pagecache(), which
2957 * must keep the partial page. In contrast, we must get rid of
2959 * @holelen: size of prospective hole in bytes. This will be rounded
2960 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2962 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2963 * but 0 when invalidating pagecache, don't throw away private data.
2965 void unmap_mapping_range(struct address_space
*mapping
,
2966 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2968 struct zap_details details
;
2969 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2970 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2972 /* Check for overflow. */
2973 if (sizeof(holelen
) > sizeof(hlen
)) {
2975 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2976 if (holeend
& ~(long long)ULONG_MAX
)
2977 hlen
= ULONG_MAX
- hba
+ 1;
2980 details
.check_mapping
= even_cows
? NULL
: mapping
;
2981 details
.nonlinear_vma
= NULL
;
2982 details
.first_index
= hba
;
2983 details
.last_index
= hba
+ hlen
- 1;
2984 if (details
.last_index
< details
.first_index
)
2985 details
.last_index
= ULONG_MAX
;
2988 mutex_lock(&mapping
->i_mmap_mutex
);
2989 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
)))
2990 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2991 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
2992 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
2993 mutex_unlock(&mapping
->i_mmap_mutex
);
2995 EXPORT_SYMBOL(unmap_mapping_range
);
2998 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2999 * but allow concurrent faults), and pte mapped but not yet locked.
3000 * We return with mmap_sem still held, but pte unmapped and unlocked.
3002 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3003 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3004 unsigned int flags
, pte_t orig_pte
)
3007 struct page
*page
, *swapcache
;
3011 struct mem_cgroup
*ptr
;
3015 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
3018 entry
= pte_to_swp_entry(orig_pte
);
3019 if (unlikely(non_swap_entry(entry
))) {
3020 if (is_migration_entry(entry
)) {
3021 migration_entry_wait(mm
, pmd
, address
);
3022 } else if (is_hwpoison_entry(entry
)) {
3023 ret
= VM_FAULT_HWPOISON
;
3025 print_bad_pte(vma
, address
, orig_pte
, NULL
);
3026 ret
= VM_FAULT_SIGBUS
;
3030 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
3031 page
= lookup_swap_cache(entry
);
3033 page
= swapin_readahead(entry
,
3034 GFP_HIGHUSER_MOVABLE
, vma
, address
);
3037 * Back out if somebody else faulted in this pte
3038 * while we released the pte lock.
3040 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3041 if (likely(pte_same(*page_table
, orig_pte
)))
3043 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3047 /* Had to read the page from swap area: Major fault */
3048 ret
= VM_FAULT_MAJOR
;
3049 count_vm_event(PGMAJFAULT
);
3050 mem_cgroup_count_vm_event(mm
, PGMAJFAULT
);
3051 } else if (PageHWPoison(page
)) {
3053 * hwpoisoned dirty swapcache pages are kept for killing
3054 * owner processes (which may be unknown at hwpoison time)
3056 ret
= VM_FAULT_HWPOISON
;
3057 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3063 locked
= lock_page_or_retry(page
, mm
, flags
);
3065 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3067 ret
|= VM_FAULT_RETRY
;
3072 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3073 * release the swapcache from under us. The page pin, and pte_same
3074 * test below, are not enough to exclude that. Even if it is still
3075 * swapcache, we need to check that the page's swap has not changed.
3077 if (unlikely(!PageSwapCache(page
) || page_private(page
) != entry
.val
))
3080 page
= ksm_might_need_to_copy(page
, vma
, address
);
3081 if (unlikely(!page
)) {
3087 if (mem_cgroup_try_charge_swapin(mm
, page
, GFP_KERNEL
, &ptr
)) {
3093 * Back out if somebody else already faulted in this pte.
3095 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3096 if (unlikely(!pte_same(*page_table
, orig_pte
)))
3099 if (unlikely(!PageUptodate(page
))) {
3100 ret
= VM_FAULT_SIGBUS
;
3105 * The page isn't present yet, go ahead with the fault.
3107 * Be careful about the sequence of operations here.
3108 * To get its accounting right, reuse_swap_page() must be called
3109 * while the page is counted on swap but not yet in mapcount i.e.
3110 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3111 * must be called after the swap_free(), or it will never succeed.
3112 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3113 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3114 * in page->private. In this case, a record in swap_cgroup is silently
3115 * discarded at swap_free().
3118 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3119 dec_mm_counter_fast(mm
, MM_SWAPENTS
);
3120 pte
= mk_pte(page
, vma
->vm_page_prot
);
3121 if ((flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
)) {
3122 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3123 flags
&= ~FAULT_FLAG_WRITE
;
3124 ret
|= VM_FAULT_WRITE
;
3127 flush_icache_page(vma
, page
);
3128 if (pte_swp_soft_dirty(orig_pte
))
3129 pte
= pte_mksoft_dirty(pte
);
3130 set_pte_at(mm
, address
, page_table
, pte
);
3131 if (page
== swapcache
)
3132 do_page_add_anon_rmap(page
, vma
, address
, exclusive
);
3133 else /* ksm created a completely new copy */
3134 page_add_new_anon_rmap(page
, vma
, address
);
3135 /* It's better to call commit-charge after rmap is established */
3136 mem_cgroup_commit_charge_swapin(page
, ptr
);
3139 if (vm_swap_full() || (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
3140 try_to_free_swap(page
);
3142 if (page
!= swapcache
) {
3144 * Hold the lock to avoid the swap entry to be reused
3145 * until we take the PT lock for the pte_same() check
3146 * (to avoid false positives from pte_same). For
3147 * further safety release the lock after the swap_free
3148 * so that the swap count won't change under a
3149 * parallel locked swapcache.
3151 unlock_page(swapcache
);
3152 page_cache_release(swapcache
);
3155 if (flags
& FAULT_FLAG_WRITE
) {
3156 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
3157 if (ret
& VM_FAULT_ERROR
)
3158 ret
&= VM_FAULT_ERROR
;
3162 /* No need to invalidate - it was non-present before */
3163 update_mmu_cache(vma
, address
, page_table
);
3165 pte_unmap_unlock(page_table
, ptl
);
3169 mem_cgroup_cancel_charge_swapin(ptr
);
3170 pte_unmap_unlock(page_table
, ptl
);
3174 page_cache_release(page
);
3175 if (page
!= swapcache
) {
3176 unlock_page(swapcache
);
3177 page_cache_release(swapcache
);
3183 * This is like a special single-page "expand_{down|up}wards()",
3184 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3185 * doesn't hit another vma.
3187 static inline int check_stack_guard_page(struct vm_area_struct
*vma
, unsigned long address
)
3189 address
&= PAGE_MASK
;
3190 if ((vma
->vm_flags
& VM_GROWSDOWN
) && address
== vma
->vm_start
) {
3191 struct vm_area_struct
*prev
= vma
->vm_prev
;
3194 * Is there a mapping abutting this one below?
3196 * That's only ok if it's the same stack mapping
3197 * that has gotten split..
3199 if (prev
&& prev
->vm_end
== address
)
3200 return prev
->vm_flags
& VM_GROWSDOWN
? 0 : -ENOMEM
;
3202 expand_downwards(vma
, address
- PAGE_SIZE
);
3204 if ((vma
->vm_flags
& VM_GROWSUP
) && address
+ PAGE_SIZE
== vma
->vm_end
) {
3205 struct vm_area_struct
*next
= vma
->vm_next
;
3207 /* As VM_GROWSDOWN but s/below/above/ */
3208 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
)
3209 return next
->vm_flags
& VM_GROWSUP
? 0 : -ENOMEM
;
3211 expand_upwards(vma
, address
+ PAGE_SIZE
);
3217 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3218 * but allow concurrent faults), and pte mapped but not yet locked.
3219 * We return with mmap_sem still held, but pte unmapped and unlocked.
3221 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3222 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3229 pte_unmap(page_table
);
3231 /* Check if we need to add a guard page to the stack */
3232 if (check_stack_guard_page(vma
, address
) < 0)
3233 return VM_FAULT_SIGBUS
;
3235 /* Use the zero-page for reads */
3236 if (!(flags
& FAULT_FLAG_WRITE
)) {
3237 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(address
),
3238 vma
->vm_page_prot
));
3239 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3240 if (!pte_none(*page_table
))
3245 /* Allocate our own private page. */
3246 if (unlikely(anon_vma_prepare(vma
)))
3248 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
3252 * The memory barrier inside __SetPageUptodate makes sure that
3253 * preceeding stores to the page contents become visible before
3254 * the set_pte_at() write.
3256 __SetPageUptodate(page
);
3258 if (mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
))
3261 entry
= mk_pte(page
, vma
->vm_page_prot
);
3262 if (vma
->vm_flags
& VM_WRITE
)
3263 entry
= pte_mkwrite(pte_mkdirty(entry
));
3265 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3266 if (!pte_none(*page_table
))
3269 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3270 page_add_new_anon_rmap(page
, vma
, address
);
3272 set_pte_at(mm
, address
, page_table
, entry
);
3274 /* No need to invalidate - it was non-present before */
3275 update_mmu_cache(vma
, address
, page_table
);
3277 pte_unmap_unlock(page_table
, ptl
);
3280 mem_cgroup_uncharge_page(page
);
3281 page_cache_release(page
);
3284 page_cache_release(page
);
3286 return VM_FAULT_OOM
;
3290 * __do_fault() tries to create a new page mapping. It aggressively
3291 * tries to share with existing pages, but makes a separate copy if
3292 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3293 * the next page fault.
3295 * As this is called only for pages that do not currently exist, we
3296 * do not need to flush old virtual caches or the TLB.
3298 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3299 * but allow concurrent faults), and pte neither mapped nor locked.
3300 * We return with mmap_sem still held, but pte unmapped and unlocked.
3302 static int __do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3303 unsigned long address
, pmd_t
*pmd
,
3304 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3309 struct page
*cow_page
;
3312 struct page
*dirty_page
= NULL
;
3313 struct vm_fault vmf
;
3315 int page_mkwrite
= 0;
3318 * If we do COW later, allocate page befor taking lock_page()
3319 * on the file cache page. This will reduce lock holding time.
3321 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3323 if (unlikely(anon_vma_prepare(vma
)))
3324 return VM_FAULT_OOM
;
3326 cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
3328 return VM_FAULT_OOM
;
3330 if (mem_cgroup_newpage_charge(cow_page
, mm
, GFP_KERNEL
)) {
3331 page_cache_release(cow_page
);
3332 return VM_FAULT_OOM
;
3337 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
3342 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
3343 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
3347 if (unlikely(PageHWPoison(vmf
.page
))) {
3348 if (ret
& VM_FAULT_LOCKED
)
3349 unlock_page(vmf
.page
);
3350 ret
= VM_FAULT_HWPOISON
;
3355 * For consistency in subsequent calls, make the faulted page always
3358 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3359 lock_page(vmf
.page
);
3361 VM_BUG_ON_PAGE(!PageLocked(vmf
.page
), vmf
.page
);
3364 * Should we do an early C-O-W break?
3367 if (flags
& FAULT_FLAG_WRITE
) {
3368 if (!(vma
->vm_flags
& VM_SHARED
)) {
3371 copy_user_highpage(page
, vmf
.page
, address
, vma
);
3372 __SetPageUptodate(page
);
3375 * If the page will be shareable, see if the backing
3376 * address space wants to know that the page is about
3377 * to become writable
3379 if (vma
->vm_ops
->page_mkwrite
) {
3383 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
3384 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
3386 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3388 goto unwritable_page
;
3390 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
3392 if (!page
->mapping
) {
3393 ret
= 0; /* retry the fault */
3395 goto unwritable_page
;
3398 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
3405 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3408 * This silly early PAGE_DIRTY setting removes a race
3409 * due to the bad i386 page protection. But it's valid
3410 * for other architectures too.
3412 * Note that if FAULT_FLAG_WRITE is set, we either now have
3413 * an exclusive copy of the page, or this is a shared mapping,
3414 * so we can make it writable and dirty to avoid having to
3415 * handle that later.
3417 /* Only go through if we didn't race with anybody else... */
3418 if (likely(pte_same(*page_table
, orig_pte
))) {
3419 flush_icache_page(vma
, page
);
3420 entry
= mk_pte(page
, vma
->vm_page_prot
);
3421 if (flags
& FAULT_FLAG_WRITE
)
3422 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3423 else if (pte_file(orig_pte
) && pte_file_soft_dirty(orig_pte
))
3424 pte_mksoft_dirty(entry
);
3426 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3427 page_add_new_anon_rmap(page
, vma
, address
);
3429 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
3430 page_add_file_rmap(page
);
3431 if (flags
& FAULT_FLAG_WRITE
) {
3433 get_page(dirty_page
);
3436 set_pte_at(mm
, address
, page_table
, entry
);
3438 /* no need to invalidate: a not-present page won't be cached */
3439 update_mmu_cache(vma
, address
, page_table
);
3442 mem_cgroup_uncharge_page(cow_page
);
3444 page_cache_release(page
);
3446 anon
= 1; /* no anon but release faulted_page */
3449 pte_unmap_unlock(page_table
, ptl
);
3452 struct address_space
*mapping
= page
->mapping
;
3455 if (set_page_dirty(dirty_page
))
3457 unlock_page(dirty_page
);
3458 put_page(dirty_page
);
3459 if ((dirtied
|| page_mkwrite
) && mapping
) {
3461 * Some device drivers do not set page.mapping but still
3464 balance_dirty_pages_ratelimited(mapping
);
3467 /* file_update_time outside page_lock */
3468 if (vma
->vm_file
&& !page_mkwrite
)
3469 file_update_time(vma
->vm_file
);
3471 unlock_page(vmf
.page
);
3473 page_cache_release(vmf
.page
);
3479 page_cache_release(page
);
3482 /* fs's fault handler get error */
3484 mem_cgroup_uncharge_page(cow_page
);
3485 page_cache_release(cow_page
);
3490 static int do_linear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3491 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3492 unsigned int flags
, pte_t orig_pte
)
3494 pgoff_t pgoff
= (((address
& PAGE_MASK
)
3495 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
3497 pte_unmap(page_table
);
3498 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3502 * Fault of a previously existing named mapping. Repopulate the pte
3503 * from the encoded file_pte if possible. This enables swappable
3506 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3507 * but allow concurrent faults), and pte mapped but not yet locked.
3508 * We return with mmap_sem still held, but pte unmapped and unlocked.
3510 static int do_nonlinear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3511 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3512 unsigned int flags
, pte_t orig_pte
)
3516 flags
|= FAULT_FLAG_NONLINEAR
;
3518 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
3521 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
))) {
3523 * Page table corrupted: show pte and kill process.
3525 print_bad_pte(vma
, address
, orig_pte
, NULL
);
3526 return VM_FAULT_SIGBUS
;
3529 pgoff
= pte_to_pgoff(orig_pte
);
3530 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3533 int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3534 unsigned long addr
, int page_nid
,
3539 count_vm_numa_event(NUMA_HINT_FAULTS
);
3540 if (page_nid
== numa_node_id()) {
3541 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3542 *flags
|= TNF_FAULT_LOCAL
;
3545 return mpol_misplaced(page
, vma
, addr
);
3548 int do_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3549 unsigned long addr
, pte_t pte
, pte_t
*ptep
, pmd_t
*pmd
)
3551 struct page
*page
= NULL
;
3556 bool migrated
= false;
3560 * The "pte" at this point cannot be used safely without
3561 * validation through pte_unmap_same(). It's of NUMA type but
3562 * the pfn may be screwed if the read is non atomic.
3564 * ptep_modify_prot_start is not called as this is clearing
3565 * the _PAGE_NUMA bit and it is not really expected that there
3566 * would be concurrent hardware modifications to the PTE.
3568 ptl
= pte_lockptr(mm
, pmd
);
3570 if (unlikely(!pte_same(*ptep
, pte
))) {
3571 pte_unmap_unlock(ptep
, ptl
);
3575 pte
= pte_mknonnuma(pte
);
3576 set_pte_at(mm
, addr
, ptep
, pte
);
3577 update_mmu_cache(vma
, addr
, ptep
);
3579 page
= vm_normal_page(vma
, addr
, pte
);
3581 pte_unmap_unlock(ptep
, ptl
);
3584 BUG_ON(is_zero_pfn(page_to_pfn(page
)));
3587 * Avoid grouping on DSO/COW pages in specific and RO pages
3588 * in general, RO pages shouldn't hurt as much anyway since
3589 * they can be in shared cache state.
3591 if (!pte_write(pte
))
3592 flags
|= TNF_NO_GROUP
;
3595 * Flag if the page is shared between multiple address spaces. This
3596 * is later used when determining whether to group tasks together
3598 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
3599 flags
|= TNF_SHARED
;
3601 last_cpupid
= page_cpupid_last(page
);
3602 page_nid
= page_to_nid(page
);
3603 target_nid
= numa_migrate_prep(page
, vma
, addr
, page_nid
, &flags
);
3604 pte_unmap_unlock(ptep
, ptl
);
3605 if (target_nid
== -1) {
3610 /* Migrate to the requested node */
3611 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
3613 page_nid
= target_nid
;
3614 flags
|= TNF_MIGRATED
;
3619 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
3624 * These routines also need to handle stuff like marking pages dirty
3625 * and/or accessed for architectures that don't do it in hardware (most
3626 * RISC architectures). The early dirtying is also good on the i386.
3628 * There is also a hook called "update_mmu_cache()" that architectures
3629 * with external mmu caches can use to update those (ie the Sparc or
3630 * PowerPC hashed page tables that act as extended TLBs).
3632 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3633 * but allow concurrent faults), and pte mapped but not yet locked.
3634 * We return with mmap_sem still held, but pte unmapped and unlocked.
3636 static int handle_pte_fault(struct mm_struct
*mm
,
3637 struct vm_area_struct
*vma
, unsigned long address
,
3638 pte_t
*pte
, pmd_t
*pmd
, unsigned int flags
)
3644 if (!pte_present(entry
)) {
3645 if (pte_none(entry
)) {
3647 if (likely(vma
->vm_ops
->fault
))
3648 return do_linear_fault(mm
, vma
, address
,
3649 pte
, pmd
, flags
, entry
);
3651 return do_anonymous_page(mm
, vma
, address
,
3654 if (pte_file(entry
))
3655 return do_nonlinear_fault(mm
, vma
, address
,
3656 pte
, pmd
, flags
, entry
);
3657 return do_swap_page(mm
, vma
, address
,
3658 pte
, pmd
, flags
, entry
);
3661 if (pte_numa(entry
))
3662 return do_numa_page(mm
, vma
, address
, entry
, pte
, pmd
);
3664 ptl
= pte_lockptr(mm
, pmd
);
3666 if (unlikely(!pte_same(*pte
, entry
)))
3668 if (flags
& FAULT_FLAG_WRITE
) {
3669 if (!pte_write(entry
))
3670 return do_wp_page(mm
, vma
, address
,
3671 pte
, pmd
, ptl
, entry
);
3672 entry
= pte_mkdirty(entry
);
3674 entry
= pte_mkyoung(entry
);
3675 if (ptep_set_access_flags(vma
, address
, pte
, entry
, flags
& FAULT_FLAG_WRITE
)) {
3676 update_mmu_cache(vma
, address
, pte
);
3679 * This is needed only for protection faults but the arch code
3680 * is not yet telling us if this is a protection fault or not.
3681 * This still avoids useless tlb flushes for .text page faults
3684 if (flags
& FAULT_FLAG_WRITE
)
3685 flush_tlb_fix_spurious_fault(vma
, address
);
3688 pte_unmap_unlock(pte
, ptl
);
3693 * By the time we get here, we already hold the mm semaphore
3695 static int __handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3696 unsigned long address
, unsigned int flags
)
3703 if (unlikely(is_vm_hugetlb_page(vma
)))
3704 return hugetlb_fault(mm
, vma
, address
, flags
);
3707 pgd
= pgd_offset(mm
, address
);
3708 pud
= pud_alloc(mm
, pgd
, address
);
3710 return VM_FAULT_OOM
;
3711 pmd
= pmd_alloc(mm
, pud
, address
);
3713 return VM_FAULT_OOM
;
3714 if (pmd_none(*pmd
) && transparent_hugepage_enabled(vma
)) {
3715 int ret
= VM_FAULT_FALLBACK
;
3717 ret
= do_huge_pmd_anonymous_page(mm
, vma
, address
,
3719 if (!(ret
& VM_FAULT_FALLBACK
))
3722 pmd_t orig_pmd
= *pmd
;
3726 if (pmd_trans_huge(orig_pmd
)) {
3727 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
3730 * If the pmd is splitting, return and retry the
3731 * the fault. Alternative: wait until the split
3732 * is done, and goto retry.
3734 if (pmd_trans_splitting(orig_pmd
))
3737 if (pmd_numa(orig_pmd
))
3738 return do_huge_pmd_numa_page(mm
, vma
, address
,
3741 if (dirty
&& !pmd_write(orig_pmd
)) {
3742 ret
= do_huge_pmd_wp_page(mm
, vma
, address
, pmd
,
3745 * If COW results in an oom, the huge pmd will
3746 * have been split, so retry the fault on the
3747 * pte for a smaller charge.
3749 if (unlikely(ret
& VM_FAULT_OOM
))
3753 huge_pmd_set_accessed(mm
, vma
, address
, pmd
,
3761 /* THP should already have been handled */
3762 BUG_ON(pmd_numa(*pmd
));
3765 * Use __pte_alloc instead of pte_alloc_map, because we can't
3766 * run pte_offset_map on the pmd, if an huge pmd could
3767 * materialize from under us from a different thread.
3769 if (unlikely(pmd_none(*pmd
)) &&
3770 unlikely(__pte_alloc(mm
, vma
, pmd
, address
)))
3771 return VM_FAULT_OOM
;
3772 /* if an huge pmd materialized from under us just retry later */
3773 if (unlikely(pmd_trans_huge(*pmd
)))
3776 * A regular pmd is established and it can't morph into a huge pmd
3777 * from under us anymore at this point because we hold the mmap_sem
3778 * read mode and khugepaged takes it in write mode. So now it's
3779 * safe to run pte_offset_map().
3781 pte
= pte_offset_map(pmd
, address
);
3783 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
3786 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3787 unsigned long address
, unsigned int flags
)
3791 __set_current_state(TASK_RUNNING
);
3793 count_vm_event(PGFAULT
);
3794 mem_cgroup_count_vm_event(mm
, PGFAULT
);
3796 /* do counter updates before entering really critical section. */
3797 check_sync_rss_stat(current
);
3800 * Enable the memcg OOM handling for faults triggered in user
3801 * space. Kernel faults are handled more gracefully.
3803 if (flags
& FAULT_FLAG_USER
)
3804 mem_cgroup_oom_enable();
3806 ret
= __handle_mm_fault(mm
, vma
, address
, flags
);
3808 if (flags
& FAULT_FLAG_USER
) {
3809 mem_cgroup_oom_disable();
3811 * The task may have entered a memcg OOM situation but
3812 * if the allocation error was handled gracefully (no
3813 * VM_FAULT_OOM), there is no need to kill anything.
3814 * Just clean up the OOM state peacefully.
3816 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
3817 mem_cgroup_oom_synchronize(false);
3823 #ifndef __PAGETABLE_PUD_FOLDED
3825 * Allocate page upper directory.
3826 * We've already handled the fast-path in-line.
3828 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
3830 pud_t
*new = pud_alloc_one(mm
, address
);
3834 smp_wmb(); /* See comment in __pte_alloc */
3836 spin_lock(&mm
->page_table_lock
);
3837 if (pgd_present(*pgd
)) /* Another has populated it */
3840 pgd_populate(mm
, pgd
, new);
3841 spin_unlock(&mm
->page_table_lock
);
3844 #endif /* __PAGETABLE_PUD_FOLDED */
3846 #ifndef __PAGETABLE_PMD_FOLDED
3848 * Allocate page middle directory.
3849 * We've already handled the fast-path in-line.
3851 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
3853 pmd_t
*new = pmd_alloc_one(mm
, address
);
3857 smp_wmb(); /* See comment in __pte_alloc */
3859 spin_lock(&mm
->page_table_lock
);
3860 #ifndef __ARCH_HAS_4LEVEL_HACK
3861 if (pud_present(*pud
)) /* Another has populated it */
3864 pud_populate(mm
, pud
, new);
3866 if (pgd_present(*pud
)) /* Another has populated it */
3869 pgd_populate(mm
, pud
, new);
3870 #endif /* __ARCH_HAS_4LEVEL_HACK */
3871 spin_unlock(&mm
->page_table_lock
);
3874 #endif /* __PAGETABLE_PMD_FOLDED */
3876 #if !defined(__HAVE_ARCH_GATE_AREA)
3878 #if defined(AT_SYSINFO_EHDR)
3879 static struct vm_area_struct gate_vma
;
3881 static int __init
gate_vma_init(void)
3883 gate_vma
.vm_mm
= NULL
;
3884 gate_vma
.vm_start
= FIXADDR_USER_START
;
3885 gate_vma
.vm_end
= FIXADDR_USER_END
;
3886 gate_vma
.vm_flags
= VM_READ
| VM_MAYREAD
| VM_EXEC
| VM_MAYEXEC
;
3887 gate_vma
.vm_page_prot
= __P101
;
3891 __initcall(gate_vma_init
);
3894 struct vm_area_struct
*get_gate_vma(struct mm_struct
*mm
)
3896 #ifdef AT_SYSINFO_EHDR
3903 int in_gate_area_no_mm(unsigned long addr
)
3905 #ifdef AT_SYSINFO_EHDR
3906 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
3912 #endif /* __HAVE_ARCH_GATE_AREA */
3914 static int __follow_pte(struct mm_struct
*mm
, unsigned long address
,
3915 pte_t
**ptepp
, spinlock_t
**ptlp
)
3922 pgd
= pgd_offset(mm
, address
);
3923 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
3926 pud
= pud_offset(pgd
, address
);
3927 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
3930 pmd
= pmd_offset(pud
, address
);
3931 VM_BUG_ON(pmd_trans_huge(*pmd
));
3932 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
3935 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3939 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
3942 if (!pte_present(*ptep
))
3947 pte_unmap_unlock(ptep
, *ptlp
);
3952 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
3953 pte_t
**ptepp
, spinlock_t
**ptlp
)
3957 /* (void) is needed to make gcc happy */
3958 (void) __cond_lock(*ptlp
,
3959 !(res
= __follow_pte(mm
, address
, ptepp
, ptlp
)));
3964 * follow_pfn - look up PFN at a user virtual address
3965 * @vma: memory mapping
3966 * @address: user virtual address
3967 * @pfn: location to store found PFN
3969 * Only IO mappings and raw PFN mappings are allowed.
3971 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3973 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
3980 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3983 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
3986 *pfn
= pte_pfn(*ptep
);
3987 pte_unmap_unlock(ptep
, ptl
);
3990 EXPORT_SYMBOL(follow_pfn
);
3992 #ifdef CONFIG_HAVE_IOREMAP_PROT
3993 int follow_phys(struct vm_area_struct
*vma
,
3994 unsigned long address
, unsigned int flags
,
3995 unsigned long *prot
, resource_size_t
*phys
)
4001 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4004 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
4008 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
4011 *prot
= pgprot_val(pte_pgprot(pte
));
4012 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
4016 pte_unmap_unlock(ptep
, ptl
);
4021 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
4022 void *buf
, int len
, int write
)
4024 resource_size_t phys_addr
;
4025 unsigned long prot
= 0;
4026 void __iomem
*maddr
;
4027 int offset
= addr
& (PAGE_SIZE
-1);
4029 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
4032 maddr
= ioremap_prot(phys_addr
, PAGE_SIZE
, prot
);
4034 memcpy_toio(maddr
+ offset
, buf
, len
);
4036 memcpy_fromio(buf
, maddr
+ offset
, len
);
4041 EXPORT_SYMBOL_GPL(generic_access_phys
);
4045 * Access another process' address space as given in mm. If non-NULL, use the
4046 * given task for page fault accounting.
4048 static int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
4049 unsigned long addr
, void *buf
, int len
, int write
)
4051 struct vm_area_struct
*vma
;
4052 void *old_buf
= buf
;
4054 down_read(&mm
->mmap_sem
);
4055 /* ignore errors, just check how much was successfully transferred */
4057 int bytes
, ret
, offset
;
4059 struct page
*page
= NULL
;
4061 ret
= get_user_pages(tsk
, mm
, addr
, 1,
4062 write
, 1, &page
, &vma
);
4065 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4066 * we can access using slightly different code.
4068 #ifdef CONFIG_HAVE_IOREMAP_PROT
4069 vma
= find_vma(mm
, addr
);
4070 if (!vma
|| vma
->vm_start
> addr
)
4072 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
4073 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
4081 offset
= addr
& (PAGE_SIZE
-1);
4082 if (bytes
> PAGE_SIZE
-offset
)
4083 bytes
= PAGE_SIZE
-offset
;
4087 copy_to_user_page(vma
, page
, addr
,
4088 maddr
+ offset
, buf
, bytes
);
4089 set_page_dirty_lock(page
);
4091 copy_from_user_page(vma
, page
, addr
,
4092 buf
, maddr
+ offset
, bytes
);
4095 page_cache_release(page
);
4101 up_read(&mm
->mmap_sem
);
4103 return buf
- old_buf
;
4107 * access_remote_vm - access another process' address space
4108 * @mm: the mm_struct of the target address space
4109 * @addr: start address to access
4110 * @buf: source or destination buffer
4111 * @len: number of bytes to transfer
4112 * @write: whether the access is a write
4114 * The caller must hold a reference on @mm.
4116 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
4117 void *buf
, int len
, int write
)
4119 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, write
);
4123 * Access another process' address space.
4124 * Source/target buffer must be kernel space,
4125 * Do not walk the page table directly, use get_user_pages
4127 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
4128 void *buf
, int len
, int write
)
4130 struct mm_struct
*mm
;
4133 mm
= get_task_mm(tsk
);
4137 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, write
);
4144 * Print the name of a VMA.
4146 void print_vma_addr(char *prefix
, unsigned long ip
)
4148 struct mm_struct
*mm
= current
->mm
;
4149 struct vm_area_struct
*vma
;
4152 * Do not print if we are in atomic
4153 * contexts (in exception stacks, etc.):
4155 if (preempt_count())
4158 down_read(&mm
->mmap_sem
);
4159 vma
= find_vma(mm
, ip
);
4160 if (vma
&& vma
->vm_file
) {
4161 struct file
*f
= vma
->vm_file
;
4162 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
4166 p
= d_path(&f
->f_path
, buf
, PAGE_SIZE
);
4169 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
4171 vma
->vm_end
- vma
->vm_start
);
4172 free_page((unsigned long)buf
);
4175 up_read(&mm
->mmap_sem
);
4178 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4179 void might_fault(void)
4182 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4183 * holding the mmap_sem, this is safe because kernel memory doesn't
4184 * get paged out, therefore we'll never actually fault, and the
4185 * below annotations will generate false positives.
4187 if (segment_eq(get_fs(), KERNEL_DS
))
4191 * it would be nicer only to annotate paths which are not under
4192 * pagefault_disable, however that requires a larger audit and
4193 * providing helpers like get_user_atomic.
4198 __might_sleep(__FILE__
, __LINE__
, 0);
4201 might_lock_read(¤t
->mm
->mmap_sem
);
4203 EXPORT_SYMBOL(might_fault
);
4206 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4207 static void clear_gigantic_page(struct page
*page
,
4209 unsigned int pages_per_huge_page
)
4212 struct page
*p
= page
;
4215 for (i
= 0; i
< pages_per_huge_page
;
4216 i
++, p
= mem_map_next(p
, page
, i
)) {
4218 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
4221 void clear_huge_page(struct page
*page
,
4222 unsigned long addr
, unsigned int pages_per_huge_page
)
4226 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4227 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
4232 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4234 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
4238 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
4240 struct vm_area_struct
*vma
,
4241 unsigned int pages_per_huge_page
)
4244 struct page
*dst_base
= dst
;
4245 struct page
*src_base
= src
;
4247 for (i
= 0; i
< pages_per_huge_page
; ) {
4249 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
4252 dst
= mem_map_next(dst
, dst_base
, i
);
4253 src
= mem_map_next(src
, src_base
, i
);
4257 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
4258 unsigned long addr
, struct vm_area_struct
*vma
,
4259 unsigned int pages_per_huge_page
)
4263 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4264 copy_user_gigantic_page(dst
, src
, addr
, vma
,
4265 pages_per_huge_page
);
4270 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4272 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
4275 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4277 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4279 static struct kmem_cache
*page_ptl_cachep
;
4281 void __init
ptlock_cache_init(void)
4283 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
4287 bool ptlock_alloc(struct page
*page
)
4291 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
4298 void ptlock_free(struct page
*page
)
4300 kmem_cache_free(page_ptl_cachep
, page
->ptl
);