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>
62 #include <asm/pgalloc.h>
63 #include <asm/uaccess.h>
65 #include <asm/tlbflush.h>
66 #include <asm/pgtable.h>
70 #ifndef CONFIG_NEED_MULTIPLE_NODES
71 /* use the per-pgdat data instead for discontigmem - mbligh */
72 unsigned long max_mapnr
;
75 EXPORT_SYMBOL(max_mapnr
);
76 EXPORT_SYMBOL(mem_map
);
79 unsigned long num_physpages
;
81 * A number of key systems in x86 including ioremap() rely on the assumption
82 * that high_memory defines the upper bound on direct map memory, then end
83 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
84 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
89 EXPORT_SYMBOL(num_physpages
);
90 EXPORT_SYMBOL(high_memory
);
93 * Randomize the address space (stacks, mmaps, brk, etc.).
95 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96 * as ancient (libc5 based) binaries can segfault. )
98 int randomize_va_space __read_mostly
=
99 #ifdef CONFIG_COMPAT_BRK
105 static int __init
disable_randmaps(char *s
)
107 randomize_va_space
= 0;
110 __setup("norandmaps", disable_randmaps
);
112 unsigned long zero_pfn __read_mostly
;
113 unsigned long highest_memmap_pfn __read_mostly
;
116 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
118 static int __init
init_zero_pfn(void)
120 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
123 core_initcall(init_zero_pfn
);
126 #if defined(SPLIT_RSS_COUNTING)
128 void sync_mm_rss(struct mm_struct
*mm
)
132 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
133 if (current
->rss_stat
.count
[i
]) {
134 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
135 current
->rss_stat
.count
[i
] = 0;
138 current
->rss_stat
.events
= 0;
141 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
143 struct task_struct
*task
= current
;
145 if (likely(task
->mm
== mm
))
146 task
->rss_stat
.count
[member
] += val
;
148 add_mm_counter(mm
, member
, val
);
150 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
153 /* sync counter once per 64 page faults */
154 #define TASK_RSS_EVENTS_THRESH (64)
155 static void check_sync_rss_stat(struct task_struct
*task
)
157 if (unlikely(task
!= current
))
159 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
160 sync_mm_rss(task
->mm
);
162 #else /* SPLIT_RSS_COUNTING */
164 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
165 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
167 static void check_sync_rss_stat(struct task_struct
*task
)
171 #endif /* SPLIT_RSS_COUNTING */
173 #ifdef HAVE_GENERIC_MMU_GATHER
175 static int tlb_next_batch(struct mmu_gather
*tlb
)
177 struct mmu_gather_batch
*batch
;
181 tlb
->active
= batch
->next
;
185 batch
= (void *)__get_free_pages(GFP_NOWAIT
| __GFP_NOWARN
, 0);
191 batch
->max
= MAX_GATHER_BATCH
;
193 tlb
->active
->next
= batch
;
200 * Called to initialize an (on-stack) mmu_gather structure for page-table
201 * tear-down from @mm. The @fullmm argument is used when @mm is without
202 * users and we're going to destroy the full address space (exit/execve).
204 void tlb_gather_mmu(struct mmu_gather
*tlb
, struct mm_struct
*mm
, bool fullmm
)
208 tlb
->fullmm
= fullmm
;
210 tlb
->fast_mode
= (num_possible_cpus() == 1);
211 tlb
->local
.next
= NULL
;
213 tlb
->local
.max
= ARRAY_SIZE(tlb
->__pages
);
214 tlb
->active
= &tlb
->local
;
216 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
221 void tlb_flush_mmu(struct mmu_gather
*tlb
)
223 struct mmu_gather_batch
*batch
;
225 if (!tlb
->need_flush
)
229 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
230 tlb_table_flush(tlb
);
233 if (tlb_fast_mode(tlb
))
236 for (batch
= &tlb
->local
; batch
; batch
= batch
->next
) {
237 free_pages_and_swap_cache(batch
->pages
, batch
->nr
);
240 tlb
->active
= &tlb
->local
;
244 * Called at the end of the shootdown operation to free up any resources
245 * that were required.
247 void tlb_finish_mmu(struct mmu_gather
*tlb
, unsigned long start
, unsigned long end
)
249 struct mmu_gather_batch
*batch
, *next
;
253 /* keep the page table cache within bounds */
256 for (batch
= tlb
->local
.next
; batch
; batch
= next
) {
258 free_pages((unsigned long)batch
, 0);
260 tlb
->local
.next
= NULL
;
264 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
265 * handling the additional races in SMP caused by other CPUs caching valid
266 * mappings in their TLBs. Returns the number of free page slots left.
267 * When out of page slots we must call tlb_flush_mmu().
269 int __tlb_remove_page(struct mmu_gather
*tlb
, struct page
*page
)
271 struct mmu_gather_batch
*batch
;
273 VM_BUG_ON(!tlb
->need_flush
);
275 if (tlb_fast_mode(tlb
)) {
276 free_page_and_swap_cache(page
);
277 return 1; /* avoid calling tlb_flush_mmu() */
281 batch
->pages
[batch
->nr
++] = page
;
282 if (batch
->nr
== batch
->max
) {
283 if (!tlb_next_batch(tlb
))
287 VM_BUG_ON(batch
->nr
> batch
->max
);
289 return batch
->max
- batch
->nr
;
292 #endif /* HAVE_GENERIC_MMU_GATHER */
294 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
297 * See the comment near struct mmu_table_batch.
300 static void tlb_remove_table_smp_sync(void *arg
)
302 /* Simply deliver the interrupt */
305 static void tlb_remove_table_one(void *table
)
308 * This isn't an RCU grace period and hence the page-tables cannot be
309 * assumed to be actually RCU-freed.
311 * It is however sufficient for software page-table walkers that rely on
312 * IRQ disabling. See the comment near struct mmu_table_batch.
314 smp_call_function(tlb_remove_table_smp_sync
, NULL
, 1);
315 __tlb_remove_table(table
);
318 static void tlb_remove_table_rcu(struct rcu_head
*head
)
320 struct mmu_table_batch
*batch
;
323 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
325 for (i
= 0; i
< batch
->nr
; i
++)
326 __tlb_remove_table(batch
->tables
[i
]);
328 free_page((unsigned long)batch
);
331 void tlb_table_flush(struct mmu_gather
*tlb
)
333 struct mmu_table_batch
**batch
= &tlb
->batch
;
336 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
341 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
343 struct mmu_table_batch
**batch
= &tlb
->batch
;
348 * When there's less then two users of this mm there cannot be a
349 * concurrent page-table walk.
351 if (atomic_read(&tlb
->mm
->mm_users
) < 2) {
352 __tlb_remove_table(table
);
356 if (*batch
== NULL
) {
357 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
358 if (*batch
== NULL
) {
359 tlb_remove_table_one(table
);
364 (*batch
)->tables
[(*batch
)->nr
++] = table
;
365 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
366 tlb_table_flush(tlb
);
369 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
372 * If a p?d_bad entry is found while walking page tables, report
373 * the error, before resetting entry to p?d_none. Usually (but
374 * very seldom) called out from the p?d_none_or_clear_bad macros.
377 void pgd_clear_bad(pgd_t
*pgd
)
383 void pud_clear_bad(pud_t
*pud
)
389 void pmd_clear_bad(pmd_t
*pmd
)
396 * Note: this doesn't free the actual pages themselves. That
397 * has been handled earlier when unmapping all the memory regions.
399 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
402 pgtable_t token
= pmd_pgtable(*pmd
);
404 pte_free_tlb(tlb
, token
, addr
);
408 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
409 unsigned long addr
, unsigned long end
,
410 unsigned long floor
, unsigned long ceiling
)
417 pmd
= pmd_offset(pud
, addr
);
419 next
= pmd_addr_end(addr
, end
);
420 if (pmd_none_or_clear_bad(pmd
))
422 free_pte_range(tlb
, pmd
, addr
);
423 } while (pmd
++, addr
= next
, addr
!= end
);
433 if (end
- 1 > ceiling
- 1)
436 pmd
= pmd_offset(pud
, start
);
438 pmd_free_tlb(tlb
, pmd
, start
);
441 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
442 unsigned long addr
, unsigned long end
,
443 unsigned long floor
, unsigned long ceiling
)
450 pud
= pud_offset(pgd
, addr
);
452 next
= pud_addr_end(addr
, end
);
453 if (pud_none_or_clear_bad(pud
))
455 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
456 } while (pud
++, addr
= next
, addr
!= end
);
462 ceiling
&= PGDIR_MASK
;
466 if (end
- 1 > ceiling
- 1)
469 pud
= pud_offset(pgd
, start
);
471 pud_free_tlb(tlb
, pud
, start
);
475 * This function frees user-level page tables of a process.
477 * Must be called with pagetable lock held.
479 void free_pgd_range(struct mmu_gather
*tlb
,
480 unsigned long addr
, unsigned long end
,
481 unsigned long floor
, unsigned long ceiling
)
487 * The next few lines have given us lots of grief...
489 * Why are we testing PMD* at this top level? Because often
490 * there will be no work to do at all, and we'd prefer not to
491 * go all the way down to the bottom just to discover that.
493 * Why all these "- 1"s? Because 0 represents both the bottom
494 * of the address space and the top of it (using -1 for the
495 * top wouldn't help much: the masks would do the wrong thing).
496 * The rule is that addr 0 and floor 0 refer to the bottom of
497 * the address space, but end 0 and ceiling 0 refer to the top
498 * Comparisons need to use "end - 1" and "ceiling - 1" (though
499 * that end 0 case should be mythical).
501 * Wherever addr is brought up or ceiling brought down, we must
502 * be careful to reject "the opposite 0" before it confuses the
503 * subsequent tests. But what about where end is brought down
504 * by PMD_SIZE below? no, end can't go down to 0 there.
506 * Whereas we round start (addr) and ceiling down, by different
507 * masks at different levels, in order to test whether a table
508 * now has no other vmas using it, so can be freed, we don't
509 * bother to round floor or end up - the tests don't need that.
523 if (end
- 1 > ceiling
- 1)
528 pgd
= pgd_offset(tlb
->mm
, addr
);
530 next
= pgd_addr_end(addr
, end
);
531 if (pgd_none_or_clear_bad(pgd
))
533 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
534 } while (pgd
++, addr
= next
, addr
!= end
);
537 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
538 unsigned long floor
, unsigned long ceiling
)
541 struct vm_area_struct
*next
= vma
->vm_next
;
542 unsigned long addr
= vma
->vm_start
;
545 * Hide vma from rmap and truncate_pagecache before freeing
548 unlink_anon_vmas(vma
);
549 unlink_file_vma(vma
);
551 if (is_vm_hugetlb_page(vma
)) {
552 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
553 floor
, next
? next
->vm_start
: ceiling
);
556 * Optimization: gather nearby vmas into one call down
558 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
559 && !is_vm_hugetlb_page(next
)) {
562 unlink_anon_vmas(vma
);
563 unlink_file_vma(vma
);
565 free_pgd_range(tlb
, addr
, vma
->vm_end
,
566 floor
, next
? next
->vm_start
: ceiling
);
572 int __pte_alloc(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
573 pmd_t
*pmd
, unsigned long address
)
575 pgtable_t
new = pte_alloc_one(mm
, address
);
576 int wait_split_huge_page
;
581 * Ensure all pte setup (eg. pte page lock and page clearing) are
582 * visible before the pte is made visible to other CPUs by being
583 * put into page tables.
585 * The other side of the story is the pointer chasing in the page
586 * table walking code (when walking the page table without locking;
587 * ie. most of the time). Fortunately, these data accesses consist
588 * of a chain of data-dependent loads, meaning most CPUs (alpha
589 * being the notable exception) will already guarantee loads are
590 * seen in-order. See the alpha page table accessors for the
591 * smp_read_barrier_depends() barriers in page table walking code.
593 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
595 spin_lock(&mm
->page_table_lock
);
596 wait_split_huge_page
= 0;
597 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
599 pmd_populate(mm
, pmd
, new);
601 } else if (unlikely(pmd_trans_splitting(*pmd
)))
602 wait_split_huge_page
= 1;
603 spin_unlock(&mm
->page_table_lock
);
606 if (wait_split_huge_page
)
607 wait_split_huge_page(vma
->anon_vma
, pmd
);
611 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
613 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
617 smp_wmb(); /* See comment in __pte_alloc */
619 spin_lock(&init_mm
.page_table_lock
);
620 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
621 pmd_populate_kernel(&init_mm
, pmd
, new);
624 VM_BUG_ON(pmd_trans_splitting(*pmd
));
625 spin_unlock(&init_mm
.page_table_lock
);
627 pte_free_kernel(&init_mm
, new);
631 static inline void init_rss_vec(int *rss
)
633 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
636 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
640 if (current
->mm
== mm
)
642 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
644 add_mm_counter(mm
, i
, rss
[i
]);
648 * This function is called to print an error when a bad pte
649 * is found. For example, we might have a PFN-mapped pte in
650 * a region that doesn't allow it.
652 * The calling function must still handle the error.
654 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
655 pte_t pte
, struct page
*page
)
657 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
658 pud_t
*pud
= pud_offset(pgd
, addr
);
659 pmd_t
*pmd
= pmd_offset(pud
, addr
);
660 struct address_space
*mapping
;
662 static unsigned long resume
;
663 static unsigned long nr_shown
;
664 static unsigned long nr_unshown
;
667 * Allow a burst of 60 reports, then keep quiet for that minute;
668 * or allow a steady drip of one report per second.
670 if (nr_shown
== 60) {
671 if (time_before(jiffies
, resume
)) {
677 "BUG: Bad page map: %lu messages suppressed\n",
684 resume
= jiffies
+ 60 * HZ
;
686 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
687 index
= linear_page_index(vma
, addr
);
690 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
692 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
696 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
697 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
699 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
702 print_symbol(KERN_ALERT
"vma->vm_ops->fault: %s\n",
703 (unsigned long)vma
->vm_ops
->fault
);
704 if (vma
->vm_file
&& vma
->vm_file
->f_op
)
705 print_symbol(KERN_ALERT
"vma->vm_file->f_op->mmap: %s\n",
706 (unsigned long)vma
->vm_file
->f_op
->mmap
);
708 add_taint(TAINT_BAD_PAGE
);
711 static inline int is_cow_mapping(vm_flags_t flags
)
713 return (flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
717 static inline int is_zero_pfn(unsigned long pfn
)
719 return pfn
== zero_pfn
;
724 static inline unsigned long my_zero_pfn(unsigned long addr
)
731 * vm_normal_page -- This function gets the "struct page" associated with a pte.
733 * "Special" mappings do not wish to be associated with a "struct page" (either
734 * it doesn't exist, or it exists but they don't want to touch it). In this
735 * case, NULL is returned here. "Normal" mappings do have a struct page.
737 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
738 * pte bit, in which case this function is trivial. Secondly, an architecture
739 * may not have a spare pte bit, which requires a more complicated scheme,
742 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
743 * special mapping (even if there are underlying and valid "struct pages").
744 * COWed pages of a VM_PFNMAP are always normal.
746 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
747 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
748 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
749 * mapping will always honor the rule
751 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
753 * And for normal mappings this is false.
755 * This restricts such mappings to be a linear translation from virtual address
756 * to pfn. To get around this restriction, we allow arbitrary mappings so long
757 * as the vma is not a COW mapping; in that case, we know that all ptes are
758 * special (because none can have been COWed).
761 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
763 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
764 * page" backing, however the difference is that _all_ pages with a struct
765 * page (that is, those where pfn_valid is true) are refcounted and considered
766 * normal pages by the VM. The disadvantage is that pages are refcounted
767 * (which can be slower and simply not an option for some PFNMAP users). The
768 * advantage is that we don't have to follow the strict linearity rule of
769 * PFNMAP mappings in order to support COWable mappings.
772 #ifdef __HAVE_ARCH_PTE_SPECIAL
773 # define HAVE_PTE_SPECIAL 1
775 # define HAVE_PTE_SPECIAL 0
777 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
780 unsigned long pfn
= pte_pfn(pte
);
782 if (HAVE_PTE_SPECIAL
) {
783 if (likely(!pte_special(pte
)))
785 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
787 if (!is_zero_pfn(pfn
))
788 print_bad_pte(vma
, addr
, pte
, NULL
);
792 /* !HAVE_PTE_SPECIAL case follows: */
794 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
795 if (vma
->vm_flags
& VM_MIXEDMAP
) {
801 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
802 if (pfn
== vma
->vm_pgoff
+ off
)
804 if (!is_cow_mapping(vma
->vm_flags
))
809 if (is_zero_pfn(pfn
))
812 if (unlikely(pfn
> highest_memmap_pfn
)) {
813 print_bad_pte(vma
, addr
, pte
, NULL
);
818 * NOTE! We still have PageReserved() pages in the page tables.
819 * eg. VDSO mappings can cause them to exist.
822 return pfn_to_page(pfn
);
826 * copy one vm_area from one task to the other. Assumes the page tables
827 * already present in the new task to be cleared in the whole range
828 * covered by this vma.
831 static inline unsigned long
832 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
833 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
834 unsigned long addr
, int *rss
)
836 unsigned long vm_flags
= vma
->vm_flags
;
837 pte_t pte
= *src_pte
;
840 /* pte contains position in swap or file, so copy. */
841 if (unlikely(!pte_present(pte
))) {
842 if (!pte_file(pte
)) {
843 swp_entry_t entry
= pte_to_swp_entry(pte
);
845 if (swap_duplicate(entry
) < 0)
848 /* make sure dst_mm is on swapoff's mmlist. */
849 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
850 spin_lock(&mmlist_lock
);
851 if (list_empty(&dst_mm
->mmlist
))
852 list_add(&dst_mm
->mmlist
,
854 spin_unlock(&mmlist_lock
);
856 if (likely(!non_swap_entry(entry
)))
858 else if (is_migration_entry(entry
)) {
859 page
= migration_entry_to_page(entry
);
866 if (is_write_migration_entry(entry
) &&
867 is_cow_mapping(vm_flags
)) {
869 * COW mappings require pages in both
870 * parent and child to be set to read.
872 make_migration_entry_read(&entry
);
873 pte
= swp_entry_to_pte(entry
);
874 set_pte_at(src_mm
, addr
, src_pte
, pte
);
882 * If it's a COW mapping, write protect it both
883 * in the parent and the child
885 if (is_cow_mapping(vm_flags
)) {
886 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
887 pte
= pte_wrprotect(pte
);
891 * If it's a shared mapping, mark it clean in
894 if (vm_flags
& VM_SHARED
)
895 pte
= pte_mkclean(pte
);
896 pte
= pte_mkold(pte
);
898 page
= vm_normal_page(vma
, addr
, pte
);
909 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
913 int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
914 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
915 unsigned long addr
, unsigned long end
)
917 pte_t
*orig_src_pte
, *orig_dst_pte
;
918 pte_t
*src_pte
, *dst_pte
;
919 spinlock_t
*src_ptl
, *dst_ptl
;
921 int rss
[NR_MM_COUNTERS
];
922 swp_entry_t entry
= (swp_entry_t
){0};
927 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
930 src_pte
= pte_offset_map(src_pmd
, addr
);
931 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
932 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
933 orig_src_pte
= src_pte
;
934 orig_dst_pte
= dst_pte
;
935 arch_enter_lazy_mmu_mode();
939 * We are holding two locks at this point - either of them
940 * could generate latencies in another task on another CPU.
942 if (progress
>= 32) {
944 if (need_resched() ||
945 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
948 if (pte_none(*src_pte
)) {
952 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
957 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
959 arch_leave_lazy_mmu_mode();
960 spin_unlock(src_ptl
);
961 pte_unmap(orig_src_pte
);
962 add_mm_rss_vec(dst_mm
, rss
);
963 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
967 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
976 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
977 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
978 unsigned long addr
, unsigned long end
)
980 pmd_t
*src_pmd
, *dst_pmd
;
983 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
986 src_pmd
= pmd_offset(src_pud
, addr
);
988 next
= pmd_addr_end(addr
, end
);
989 if (pmd_trans_huge(*src_pmd
)) {
991 VM_BUG_ON(next
-addr
!= HPAGE_PMD_SIZE
);
992 err
= copy_huge_pmd(dst_mm
, src_mm
,
993 dst_pmd
, src_pmd
, addr
, vma
);
1000 if (pmd_none_or_clear_bad(src_pmd
))
1002 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
1005 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1009 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1010 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
1011 unsigned long addr
, unsigned long end
)
1013 pud_t
*src_pud
, *dst_pud
;
1016 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
1019 src_pud
= pud_offset(src_pgd
, addr
);
1021 next
= pud_addr_end(addr
, end
);
1022 if (pud_none_or_clear_bad(src_pud
))
1024 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
1027 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1031 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1032 struct vm_area_struct
*vma
)
1034 pgd_t
*src_pgd
, *dst_pgd
;
1036 unsigned long addr
= vma
->vm_start
;
1037 unsigned long end
= vma
->vm_end
;
1041 * Don't copy ptes where a page fault will fill them correctly.
1042 * Fork becomes much lighter when there are big shared or private
1043 * readonly mappings. The tradeoff is that copy_page_range is more
1044 * efficient than faulting.
1046 if (!(vma
->vm_flags
& (VM_HUGETLB
|VM_NONLINEAR
|VM_PFNMAP
|VM_INSERTPAGE
))) {
1051 if (is_vm_hugetlb_page(vma
))
1052 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1054 if (unlikely(is_pfn_mapping(vma
))) {
1056 * We do not free on error cases below as remove_vma
1057 * gets called on error from higher level routine
1059 ret
= track_pfn_vma_copy(vma
);
1065 * We need to invalidate the secondary MMU mappings only when
1066 * there could be a permission downgrade on the ptes of the
1067 * parent mm. And a permission downgrade will only happen if
1068 * is_cow_mapping() returns true.
1070 if (is_cow_mapping(vma
->vm_flags
))
1071 mmu_notifier_invalidate_range_start(src_mm
, addr
, end
);
1074 dst_pgd
= pgd_offset(dst_mm
, addr
);
1075 src_pgd
= pgd_offset(src_mm
, addr
);
1077 next
= pgd_addr_end(addr
, end
);
1078 if (pgd_none_or_clear_bad(src_pgd
))
1080 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1081 vma
, addr
, next
))) {
1085 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1087 if (is_cow_mapping(vma
->vm_flags
))
1088 mmu_notifier_invalidate_range_end(src_mm
,
1089 vma
->vm_start
, end
);
1093 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1094 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1095 unsigned long addr
, unsigned long end
,
1096 struct zap_details
*details
)
1098 struct mm_struct
*mm
= tlb
->mm
;
1099 int force_flush
= 0;
1100 int rss
[NR_MM_COUNTERS
];
1107 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1109 arch_enter_lazy_mmu_mode();
1112 if (pte_none(ptent
)) {
1116 if (pte_present(ptent
)) {
1119 page
= vm_normal_page(vma
, addr
, ptent
);
1120 if (unlikely(details
) && page
) {
1122 * unmap_shared_mapping_pages() wants to
1123 * invalidate cache without truncating:
1124 * unmap shared but keep private pages.
1126 if (details
->check_mapping
&&
1127 details
->check_mapping
!= page
->mapping
)
1130 * Each page->index must be checked when
1131 * invalidating or truncating nonlinear.
1133 if (details
->nonlinear_vma
&&
1134 (page
->index
< details
->first_index
||
1135 page
->index
> details
->last_index
))
1138 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1140 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1141 if (unlikely(!page
))
1143 if (unlikely(details
) && details
->nonlinear_vma
1144 && linear_page_index(details
->nonlinear_vma
,
1145 addr
) != page
->index
)
1146 set_pte_at(mm
, addr
, pte
,
1147 pgoff_to_pte(page
->index
));
1149 rss
[MM_ANONPAGES
]--;
1151 if (pte_dirty(ptent
))
1152 set_page_dirty(page
);
1153 if (pte_young(ptent
) &&
1154 likely(!VM_SequentialReadHint(vma
)))
1155 mark_page_accessed(page
);
1156 rss
[MM_FILEPAGES
]--;
1158 page_remove_rmap(page
);
1159 if (unlikely(page_mapcount(page
) < 0))
1160 print_bad_pte(vma
, addr
, ptent
, page
);
1161 force_flush
= !__tlb_remove_page(tlb
, page
);
1167 * If details->check_mapping, we leave swap entries;
1168 * if details->nonlinear_vma, we leave file entries.
1170 if (unlikely(details
))
1172 if (pte_file(ptent
)) {
1173 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
)))
1174 print_bad_pte(vma
, addr
, ptent
, NULL
);
1176 swp_entry_t entry
= pte_to_swp_entry(ptent
);
1178 if (!non_swap_entry(entry
))
1180 else if (is_migration_entry(entry
)) {
1183 page
= migration_entry_to_page(entry
);
1186 rss
[MM_ANONPAGES
]--;
1188 rss
[MM_FILEPAGES
]--;
1190 if (unlikely(!free_swap_and_cache(entry
)))
1191 print_bad_pte(vma
, addr
, ptent
, NULL
);
1193 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1194 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1196 add_mm_rss_vec(mm
, rss
);
1197 arch_leave_lazy_mmu_mode();
1198 pte_unmap_unlock(start_pte
, ptl
);
1201 * mmu_gather ran out of room to batch pages, we break out of
1202 * the PTE lock to avoid doing the potential expensive TLB invalidate
1203 * and page-free while holding it.
1215 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1216 struct vm_area_struct
*vma
, pud_t
*pud
,
1217 unsigned long addr
, unsigned long end
,
1218 struct zap_details
*details
)
1223 pmd
= pmd_offset(pud
, addr
);
1225 next
= pmd_addr_end(addr
, end
);
1226 if (pmd_trans_huge(*pmd
)) {
1227 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1228 VM_BUG_ON(!rwsem_is_locked(&tlb
->mm
->mmap_sem
));
1229 split_huge_page_pmd(vma
->vm_mm
, pmd
);
1230 } else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1235 * Here there can be other concurrent MADV_DONTNEED or
1236 * trans huge page faults running, and if the pmd is
1237 * none or trans huge it can change under us. This is
1238 * because MADV_DONTNEED holds the mmap_sem in read
1241 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1243 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1246 } while (pmd
++, addr
= next
, addr
!= end
);
1251 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1252 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1253 unsigned long addr
, unsigned long end
,
1254 struct zap_details
*details
)
1259 pud
= pud_offset(pgd
, addr
);
1261 next
= pud_addr_end(addr
, end
);
1262 if (pud_none_or_clear_bad(pud
))
1264 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1265 } while (pud
++, addr
= next
, addr
!= end
);
1270 static void unmap_page_range(struct mmu_gather
*tlb
,
1271 struct vm_area_struct
*vma
,
1272 unsigned long addr
, unsigned long end
,
1273 struct zap_details
*details
)
1278 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
1281 BUG_ON(addr
>= end
);
1282 mem_cgroup_uncharge_start();
1283 tlb_start_vma(tlb
, vma
);
1284 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1286 next
= pgd_addr_end(addr
, end
);
1287 if (pgd_none_or_clear_bad(pgd
))
1289 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
1290 } while (pgd
++, addr
= next
, addr
!= end
);
1291 tlb_end_vma(tlb
, vma
);
1292 mem_cgroup_uncharge_end();
1296 static void unmap_single_vma(struct mmu_gather
*tlb
,
1297 struct vm_area_struct
*vma
, unsigned long start_addr
,
1298 unsigned long end_addr
,
1299 struct zap_details
*details
)
1301 unsigned long start
= max(vma
->vm_start
, start_addr
);
1304 if (start
>= vma
->vm_end
)
1306 end
= min(vma
->vm_end
, end_addr
);
1307 if (end
<= vma
->vm_start
)
1310 if (unlikely(is_pfn_mapping(vma
)))
1311 untrack_pfn_vma(vma
, 0, 0);
1314 if (unlikely(is_vm_hugetlb_page(vma
))) {
1316 * It is undesirable to test vma->vm_file as it
1317 * should be non-null for valid hugetlb area.
1318 * However, vm_file will be NULL in the error
1319 * cleanup path of do_mmap_pgoff. When
1320 * hugetlbfs ->mmap method fails,
1321 * do_mmap_pgoff() nullifies vma->vm_file
1322 * before calling this function to clean up.
1323 * Since no pte has actually been setup, it is
1324 * safe to do nothing in this case.
1327 unmap_hugepage_range(vma
, start
, end
, NULL
);
1329 unmap_page_range(tlb
, vma
, start
, end
, details
);
1334 * unmap_vmas - unmap a range of memory covered by a list of vma's
1335 * @tlb: address of the caller's struct mmu_gather
1336 * @vma: the starting vma
1337 * @start_addr: virtual address at which to start unmapping
1338 * @end_addr: virtual address at which to end unmapping
1340 * Unmap all pages in the vma list.
1342 * Only addresses between `start' and `end' will be unmapped.
1344 * The VMA list must be sorted in ascending virtual address order.
1346 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1347 * range after unmap_vmas() returns. So the only responsibility here is to
1348 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1349 * drops the lock and schedules.
1351 void unmap_vmas(struct mmu_gather
*tlb
,
1352 struct vm_area_struct
*vma
, unsigned long start_addr
,
1353 unsigned long end_addr
)
1355 struct mm_struct
*mm
= vma
->vm_mm
;
1357 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1358 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1359 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1360 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1364 * zap_page_range - remove user pages in a given range
1365 * @vma: vm_area_struct holding the applicable pages
1366 * @address: starting address of pages to zap
1367 * @size: number of bytes to zap
1368 * @details: details of nonlinear truncation or shared cache invalidation
1370 * Caller must protect the VMA list
1372 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1373 unsigned long size
, struct zap_details
*details
)
1375 struct mm_struct
*mm
= vma
->vm_mm
;
1376 struct mmu_gather tlb
;
1377 unsigned long end
= start
+ size
;
1380 tlb_gather_mmu(&tlb
, mm
, 0);
1381 update_hiwater_rss(mm
);
1382 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1383 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
)
1384 unmap_single_vma(&tlb
, vma
, start
, end
, details
);
1385 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1386 tlb_finish_mmu(&tlb
, start
, end
);
1390 * zap_page_range_single - remove user pages in a given range
1391 * @vma: vm_area_struct holding the applicable pages
1392 * @address: starting address of pages to zap
1393 * @size: number of bytes to zap
1394 * @details: details of nonlinear truncation or shared cache invalidation
1396 * The range must fit into one VMA.
1398 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1399 unsigned long size
, struct zap_details
*details
)
1401 struct mm_struct
*mm
= vma
->vm_mm
;
1402 struct mmu_gather tlb
;
1403 unsigned long end
= address
+ size
;
1406 tlb_gather_mmu(&tlb
, mm
, 0);
1407 update_hiwater_rss(mm
);
1408 mmu_notifier_invalidate_range_start(mm
, address
, end
);
1409 unmap_single_vma(&tlb
, vma
, address
, end
, details
);
1410 mmu_notifier_invalidate_range_end(mm
, address
, end
);
1411 tlb_finish_mmu(&tlb
, address
, end
);
1415 * zap_vma_ptes - remove ptes mapping the vma
1416 * @vma: vm_area_struct holding ptes to be zapped
1417 * @address: starting address of pages to zap
1418 * @size: number of bytes to zap
1420 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1422 * The entire address range must be fully contained within the vma.
1424 * Returns 0 if successful.
1426 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1429 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1430 !(vma
->vm_flags
& VM_PFNMAP
))
1432 zap_page_range_single(vma
, address
, size
, NULL
);
1435 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1438 * follow_page - look up a page descriptor from a user-virtual address
1439 * @vma: vm_area_struct mapping @address
1440 * @address: virtual address to look up
1441 * @flags: flags modifying lookup behaviour
1443 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1445 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1446 * an error pointer if there is a mapping to something not represented
1447 * by a page descriptor (see also vm_normal_page()).
1449 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
1458 struct mm_struct
*mm
= vma
->vm_mm
;
1460 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
1461 if (!IS_ERR(page
)) {
1462 BUG_ON(flags
& FOLL_GET
);
1467 pgd
= pgd_offset(mm
, address
);
1468 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
1471 pud
= pud_offset(pgd
, address
);
1474 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
1475 BUG_ON(flags
& FOLL_GET
);
1476 page
= follow_huge_pud(mm
, address
, pud
, flags
& FOLL_WRITE
);
1479 if (unlikely(pud_bad(*pud
)))
1482 pmd
= pmd_offset(pud
, address
);
1485 if (pmd_huge(*pmd
) && vma
->vm_flags
& VM_HUGETLB
) {
1486 BUG_ON(flags
& FOLL_GET
);
1487 page
= follow_huge_pmd(mm
, address
, pmd
, flags
& FOLL_WRITE
);
1490 if (pmd_trans_huge(*pmd
)) {
1491 if (flags
& FOLL_SPLIT
) {
1492 split_huge_page_pmd(mm
, pmd
);
1493 goto split_fallthrough
;
1495 spin_lock(&mm
->page_table_lock
);
1496 if (likely(pmd_trans_huge(*pmd
))) {
1497 if (unlikely(pmd_trans_splitting(*pmd
))) {
1498 spin_unlock(&mm
->page_table_lock
);
1499 wait_split_huge_page(vma
->anon_vma
, pmd
);
1501 page
= follow_trans_huge_pmd(mm
, address
,
1503 spin_unlock(&mm
->page_table_lock
);
1507 spin_unlock(&mm
->page_table_lock
);
1511 if (unlikely(pmd_bad(*pmd
)))
1514 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1517 if (!pte_present(pte
))
1519 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
1522 page
= vm_normal_page(vma
, address
, pte
);
1523 if (unlikely(!page
)) {
1524 if ((flags
& FOLL_DUMP
) ||
1525 !is_zero_pfn(pte_pfn(pte
)))
1527 page
= pte_page(pte
);
1530 if (flags
& FOLL_GET
)
1531 get_page_foll(page
);
1532 if (flags
& FOLL_TOUCH
) {
1533 if ((flags
& FOLL_WRITE
) &&
1534 !pte_dirty(pte
) && !PageDirty(page
))
1535 set_page_dirty(page
);
1537 * pte_mkyoung() would be more correct here, but atomic care
1538 * is needed to avoid losing the dirty bit: it is easier to use
1539 * mark_page_accessed().
1541 mark_page_accessed(page
);
1543 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1545 * The preliminary mapping check is mainly to avoid the
1546 * pointless overhead of lock_page on the ZERO_PAGE
1547 * which might bounce very badly if there is contention.
1549 * If the page is already locked, we don't need to
1550 * handle it now - vmscan will handle it later if and
1551 * when it attempts to reclaim the page.
1553 if (page
->mapping
&& trylock_page(page
)) {
1554 lru_add_drain(); /* push cached pages to LRU */
1556 * Because we lock page here and migration is
1557 * blocked by the pte's page reference, we need
1558 * only check for file-cache page truncation.
1561 mlock_vma_page(page
);
1566 pte_unmap_unlock(ptep
, ptl
);
1571 pte_unmap_unlock(ptep
, ptl
);
1572 return ERR_PTR(-EFAULT
);
1575 pte_unmap_unlock(ptep
, ptl
);
1581 * When core dumping an enormous anonymous area that nobody
1582 * has touched so far, we don't want to allocate unnecessary pages or
1583 * page tables. Return error instead of NULL to skip handle_mm_fault,
1584 * then get_dump_page() will return NULL to leave a hole in the dump.
1585 * But we can only make this optimization where a hole would surely
1586 * be zero-filled if handle_mm_fault() actually did handle it.
1588 if ((flags
& FOLL_DUMP
) &&
1589 (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
1590 return ERR_PTR(-EFAULT
);
1594 static inline int stack_guard_page(struct vm_area_struct
*vma
, unsigned long addr
)
1596 return stack_guard_page_start(vma
, addr
) ||
1597 stack_guard_page_end(vma
, addr
+PAGE_SIZE
);
1601 * __get_user_pages() - pin user pages in memory
1602 * @tsk: task_struct of target task
1603 * @mm: mm_struct of target mm
1604 * @start: starting user address
1605 * @nr_pages: number of pages from start to pin
1606 * @gup_flags: flags modifying pin behaviour
1607 * @pages: array that receives pointers to the pages pinned.
1608 * Should be at least nr_pages long. Or NULL, if caller
1609 * only intends to ensure the pages are faulted in.
1610 * @vmas: array of pointers to vmas corresponding to each page.
1611 * Or NULL if the caller does not require them.
1612 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1614 * Returns number of pages pinned. This may be fewer than the number
1615 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1616 * were pinned, returns -errno. Each page returned must be released
1617 * with a put_page() call when it is finished with. vmas will only
1618 * remain valid while mmap_sem is held.
1620 * Must be called with mmap_sem held for read or write.
1622 * __get_user_pages walks a process's page tables and takes a reference to
1623 * each struct page that each user address corresponds to at a given
1624 * instant. That is, it takes the page that would be accessed if a user
1625 * thread accesses the given user virtual address at that instant.
1627 * This does not guarantee that the page exists in the user mappings when
1628 * __get_user_pages returns, and there may even be a completely different
1629 * page there in some cases (eg. if mmapped pagecache has been invalidated
1630 * and subsequently re faulted). However it does guarantee that the page
1631 * won't be freed completely. And mostly callers simply care that the page
1632 * contains data that was valid *at some point in time*. Typically, an IO
1633 * or similar operation cannot guarantee anything stronger anyway because
1634 * locks can't be held over the syscall boundary.
1636 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1637 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1638 * appropriate) must be called after the page is finished with, and
1639 * before put_page is called.
1641 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1642 * or mmap_sem contention, and if waiting is needed to pin all pages,
1643 * *@nonblocking will be set to 0.
1645 * In most cases, get_user_pages or get_user_pages_fast should be used
1646 * instead of __get_user_pages. __get_user_pages should be used only if
1647 * you need some special @gup_flags.
1649 int __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1650 unsigned long start
, int nr_pages
, unsigned int gup_flags
,
1651 struct page
**pages
, struct vm_area_struct
**vmas
,
1655 unsigned long vm_flags
;
1660 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
1663 * Require read or write permissions.
1664 * If FOLL_FORCE is set, we only require the "MAY" flags.
1666 vm_flags
= (gup_flags
& FOLL_WRITE
) ?
1667 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1668 vm_flags
&= (gup_flags
& FOLL_FORCE
) ?
1669 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1673 struct vm_area_struct
*vma
;
1675 vma
= find_extend_vma(mm
, start
);
1676 if (!vma
&& in_gate_area(mm
, start
)) {
1677 unsigned long pg
= start
& PAGE_MASK
;
1683 /* user gate pages are read-only */
1684 if (gup_flags
& FOLL_WRITE
)
1685 return i
? : -EFAULT
;
1687 pgd
= pgd_offset_k(pg
);
1689 pgd
= pgd_offset_gate(mm
, pg
);
1690 BUG_ON(pgd_none(*pgd
));
1691 pud
= pud_offset(pgd
, pg
);
1692 BUG_ON(pud_none(*pud
));
1693 pmd
= pmd_offset(pud
, pg
);
1695 return i
? : -EFAULT
;
1696 VM_BUG_ON(pmd_trans_huge(*pmd
));
1697 pte
= pte_offset_map(pmd
, pg
);
1698 if (pte_none(*pte
)) {
1700 return i
? : -EFAULT
;
1702 vma
= get_gate_vma(mm
);
1706 page
= vm_normal_page(vma
, start
, *pte
);
1708 if (!(gup_flags
& FOLL_DUMP
) &&
1709 is_zero_pfn(pte_pfn(*pte
)))
1710 page
= pte_page(*pte
);
1713 return i
? : -EFAULT
;
1724 (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1725 !(vm_flags
& vma
->vm_flags
))
1726 return i
? : -EFAULT
;
1728 if (is_vm_hugetlb_page(vma
)) {
1729 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1730 &start
, &nr_pages
, i
, gup_flags
);
1736 unsigned int foll_flags
= gup_flags
;
1739 * If we have a pending SIGKILL, don't keep faulting
1740 * pages and potentially allocating memory.
1742 if (unlikely(fatal_signal_pending(current
)))
1743 return i
? i
: -ERESTARTSYS
;
1746 while (!(page
= follow_page(vma
, start
, foll_flags
))) {
1748 unsigned int fault_flags
= 0;
1750 /* For mlock, just skip the stack guard page. */
1751 if (foll_flags
& FOLL_MLOCK
) {
1752 if (stack_guard_page(vma
, start
))
1755 if (foll_flags
& FOLL_WRITE
)
1756 fault_flags
|= FAULT_FLAG_WRITE
;
1758 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
1759 if (foll_flags
& FOLL_NOWAIT
)
1760 fault_flags
|= (FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
);
1762 ret
= handle_mm_fault(mm
, vma
, start
,
1765 if (ret
& VM_FAULT_ERROR
) {
1766 if (ret
& VM_FAULT_OOM
)
1767 return i
? i
: -ENOMEM
;
1768 if (ret
& (VM_FAULT_HWPOISON
|
1769 VM_FAULT_HWPOISON_LARGE
)) {
1772 else if (gup_flags
& FOLL_HWPOISON
)
1777 if (ret
& VM_FAULT_SIGBUS
)
1778 return i
? i
: -EFAULT
;
1783 if (ret
& VM_FAULT_MAJOR
)
1789 if (ret
& VM_FAULT_RETRY
) {
1796 * The VM_FAULT_WRITE bit tells us that
1797 * do_wp_page has broken COW when necessary,
1798 * even if maybe_mkwrite decided not to set
1799 * pte_write. We can thus safely do subsequent
1800 * page lookups as if they were reads. But only
1801 * do so when looping for pte_write is futile:
1802 * in some cases userspace may also be wanting
1803 * to write to the gotten user page, which a
1804 * read fault here might prevent (a readonly
1805 * page might get reCOWed by userspace write).
1807 if ((ret
& VM_FAULT_WRITE
) &&
1808 !(vma
->vm_flags
& VM_WRITE
))
1809 foll_flags
&= ~FOLL_WRITE
;
1814 return i
? i
: PTR_ERR(page
);
1818 flush_anon_page(vma
, page
, start
);
1819 flush_dcache_page(page
);
1827 } while (nr_pages
&& start
< vma
->vm_end
);
1831 EXPORT_SYMBOL(__get_user_pages
);
1834 * fixup_user_fault() - manually resolve a user page fault
1835 * @tsk: the task_struct to use for page fault accounting, or
1836 * NULL if faults are not to be recorded.
1837 * @mm: mm_struct of target mm
1838 * @address: user address
1839 * @fault_flags:flags to pass down to handle_mm_fault()
1841 * This is meant to be called in the specific scenario where for locking reasons
1842 * we try to access user memory in atomic context (within a pagefault_disable()
1843 * section), this returns -EFAULT, and we want to resolve the user fault before
1846 * Typically this is meant to be used by the futex code.
1848 * The main difference with get_user_pages() is that this function will
1849 * unconditionally call handle_mm_fault() which will in turn perform all the
1850 * necessary SW fixup of the dirty and young bits in the PTE, while
1851 * handle_mm_fault() only guarantees to update these in the struct page.
1853 * This is important for some architectures where those bits also gate the
1854 * access permission to the page because they are maintained in software. On
1855 * such architectures, gup() will not be enough to make a subsequent access
1858 * This should be called with the mm_sem held for read.
1860 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
1861 unsigned long address
, unsigned int fault_flags
)
1863 struct vm_area_struct
*vma
;
1866 vma
= find_extend_vma(mm
, address
);
1867 if (!vma
|| address
< vma
->vm_start
)
1870 ret
= handle_mm_fault(mm
, vma
, address
, fault_flags
);
1871 if (ret
& VM_FAULT_ERROR
) {
1872 if (ret
& VM_FAULT_OOM
)
1874 if (ret
& (VM_FAULT_HWPOISON
| VM_FAULT_HWPOISON_LARGE
))
1876 if (ret
& VM_FAULT_SIGBUS
)
1881 if (ret
& VM_FAULT_MAJOR
)
1890 * get_user_pages() - pin user pages in memory
1891 * @tsk: the task_struct to use for page fault accounting, or
1892 * NULL if faults are not to be recorded.
1893 * @mm: mm_struct of target mm
1894 * @start: starting user address
1895 * @nr_pages: number of pages from start to pin
1896 * @write: whether pages will be written to by the caller
1897 * @force: whether to force write access even if user mapping is
1898 * readonly. This will result in the page being COWed even
1899 * in MAP_SHARED mappings. You do not want this.
1900 * @pages: array that receives pointers to the pages pinned.
1901 * Should be at least nr_pages long. Or NULL, if caller
1902 * only intends to ensure the pages are faulted in.
1903 * @vmas: array of pointers to vmas corresponding to each page.
1904 * Or NULL if the caller does not require them.
1906 * Returns number of pages pinned. This may be fewer than the number
1907 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1908 * were pinned, returns -errno. Each page returned must be released
1909 * with a put_page() call when it is finished with. vmas will only
1910 * remain valid while mmap_sem is held.
1912 * Must be called with mmap_sem held for read or write.
1914 * get_user_pages walks a process's page tables and takes a reference to
1915 * each struct page that each user address corresponds to at a given
1916 * instant. That is, it takes the page that would be accessed if a user
1917 * thread accesses the given user virtual address at that instant.
1919 * This does not guarantee that the page exists in the user mappings when
1920 * get_user_pages returns, and there may even be a completely different
1921 * page there in some cases (eg. if mmapped pagecache has been invalidated
1922 * and subsequently re faulted). However it does guarantee that the page
1923 * won't be freed completely. And mostly callers simply care that the page
1924 * contains data that was valid *at some point in time*. Typically, an IO
1925 * or similar operation cannot guarantee anything stronger anyway because
1926 * locks can't be held over the syscall boundary.
1928 * If write=0, the page must not be written to. If the page is written to,
1929 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1930 * after the page is finished with, and before put_page is called.
1932 * get_user_pages is typically used for fewer-copy IO operations, to get a
1933 * handle on the memory by some means other than accesses via the user virtual
1934 * addresses. The pages may be submitted for DMA to devices or accessed via
1935 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1936 * use the correct cache flushing APIs.
1938 * See also get_user_pages_fast, for performance critical applications.
1940 int get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1941 unsigned long start
, int nr_pages
, int write
, int force
,
1942 struct page
**pages
, struct vm_area_struct
**vmas
)
1944 int flags
= FOLL_TOUCH
;
1949 flags
|= FOLL_WRITE
;
1951 flags
|= FOLL_FORCE
;
1953 return __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
, vmas
,
1956 EXPORT_SYMBOL(get_user_pages
);
1959 * get_dump_page() - pin user page in memory while writing it to core dump
1960 * @addr: user address
1962 * Returns struct page pointer of user page pinned for dump,
1963 * to be freed afterwards by page_cache_release() or put_page().
1965 * Returns NULL on any kind of failure - a hole must then be inserted into
1966 * the corefile, to preserve alignment with its headers; and also returns
1967 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1968 * allowing a hole to be left in the corefile to save diskspace.
1970 * Called without mmap_sem, but after all other threads have been killed.
1972 #ifdef CONFIG_ELF_CORE
1973 struct page
*get_dump_page(unsigned long addr
)
1975 struct vm_area_struct
*vma
;
1978 if (__get_user_pages(current
, current
->mm
, addr
, 1,
1979 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
1982 flush_cache_page(vma
, addr
, page_to_pfn(page
));
1985 #endif /* CONFIG_ELF_CORE */
1987 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1990 pgd_t
* pgd
= pgd_offset(mm
, addr
);
1991 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
1993 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
1995 VM_BUG_ON(pmd_trans_huge(*pmd
));
1996 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
2003 * This is the old fallback for page remapping.
2005 * For historical reasons, it only allows reserved pages. Only
2006 * old drivers should use this, and they needed to mark their
2007 * pages reserved for the old functions anyway.
2009 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2010 struct page
*page
, pgprot_t prot
)
2012 struct mm_struct
*mm
= vma
->vm_mm
;
2021 flush_dcache_page(page
);
2022 pte
= get_locked_pte(mm
, addr
, &ptl
);
2026 if (!pte_none(*pte
))
2029 /* Ok, finally just insert the thing.. */
2031 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
2032 page_add_file_rmap(page
);
2033 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
2036 pte_unmap_unlock(pte
, ptl
);
2039 pte_unmap_unlock(pte
, ptl
);
2045 * vm_insert_page - insert single page into user vma
2046 * @vma: user vma to map to
2047 * @addr: target user address of this page
2048 * @page: source kernel page
2050 * This allows drivers to insert individual pages they've allocated
2053 * The page has to be a nice clean _individual_ kernel allocation.
2054 * If you allocate a compound page, you need to have marked it as
2055 * such (__GFP_COMP), or manually just split the page up yourself
2056 * (see split_page()).
2058 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2059 * took an arbitrary page protection parameter. This doesn't allow
2060 * that. Your vma protection will have to be set up correctly, which
2061 * means that if you want a shared writable mapping, you'd better
2062 * ask for a shared writable mapping!
2064 * The page does not need to be reserved.
2066 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2069 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2071 if (!page_count(page
))
2073 vma
->vm_flags
|= VM_INSERTPAGE
;
2074 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2076 EXPORT_SYMBOL(vm_insert_page
);
2078 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2079 unsigned long pfn
, pgprot_t prot
)
2081 struct mm_struct
*mm
= vma
->vm_mm
;
2087 pte
= get_locked_pte(mm
, addr
, &ptl
);
2091 if (!pte_none(*pte
))
2094 /* Ok, finally just insert the thing.. */
2095 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
2096 set_pte_at(mm
, addr
, pte
, entry
);
2097 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
2101 pte_unmap_unlock(pte
, ptl
);
2107 * vm_insert_pfn - insert single pfn into user vma
2108 * @vma: user vma to map to
2109 * @addr: target user address of this page
2110 * @pfn: source kernel pfn
2112 * Similar to vm_inert_page, this allows drivers to insert individual pages
2113 * they've allocated into a user vma. Same comments apply.
2115 * This function should only be called from a vm_ops->fault handler, and
2116 * in that case the handler should return NULL.
2118 * vma cannot be a COW mapping.
2120 * As this is called only for pages that do not currently exist, we
2121 * do not need to flush old virtual caches or the TLB.
2123 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2127 pgprot_t pgprot
= vma
->vm_page_prot
;
2129 * Technically, architectures with pte_special can avoid all these
2130 * restrictions (same for remap_pfn_range). However we would like
2131 * consistency in testing and feature parity among all, so we should
2132 * try to keep these invariants in place for everybody.
2134 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
2135 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
2136 (VM_PFNMAP
|VM_MIXEDMAP
));
2137 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
2138 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
2140 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2142 if (track_pfn_vma_new(vma
, &pgprot
, pfn
, PAGE_SIZE
))
2145 ret
= insert_pfn(vma
, addr
, pfn
, pgprot
);
2148 untrack_pfn_vma(vma
, pfn
, PAGE_SIZE
);
2152 EXPORT_SYMBOL(vm_insert_pfn
);
2154 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2157 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
2159 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2163 * If we don't have pte special, then we have to use the pfn_valid()
2164 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2165 * refcount the page if pfn_valid is true (hence insert_page rather
2166 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2167 * without pte special, it would there be refcounted as a normal page.
2169 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
2172 page
= pfn_to_page(pfn
);
2173 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2175 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
2177 EXPORT_SYMBOL(vm_insert_mixed
);
2180 * maps a range of physical memory into the requested pages. the old
2181 * mappings are removed. any references to nonexistent pages results
2182 * in null mappings (currently treated as "copy-on-access")
2184 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2185 unsigned long addr
, unsigned long end
,
2186 unsigned long pfn
, pgprot_t prot
)
2191 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2194 arch_enter_lazy_mmu_mode();
2196 BUG_ON(!pte_none(*pte
));
2197 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2199 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2200 arch_leave_lazy_mmu_mode();
2201 pte_unmap_unlock(pte
- 1, ptl
);
2205 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2206 unsigned long addr
, unsigned long end
,
2207 unsigned long pfn
, pgprot_t prot
)
2212 pfn
-= addr
>> PAGE_SHIFT
;
2213 pmd
= pmd_alloc(mm
, pud
, addr
);
2216 VM_BUG_ON(pmd_trans_huge(*pmd
));
2218 next
= pmd_addr_end(addr
, end
);
2219 if (remap_pte_range(mm
, pmd
, addr
, next
,
2220 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
2222 } while (pmd
++, addr
= next
, addr
!= end
);
2226 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2227 unsigned long addr
, unsigned long end
,
2228 unsigned long pfn
, pgprot_t prot
)
2233 pfn
-= addr
>> PAGE_SHIFT
;
2234 pud
= pud_alloc(mm
, pgd
, addr
);
2238 next
= pud_addr_end(addr
, end
);
2239 if (remap_pmd_range(mm
, pud
, addr
, next
,
2240 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
2242 } while (pud
++, addr
= next
, addr
!= end
);
2247 * remap_pfn_range - remap kernel memory to userspace
2248 * @vma: user vma to map to
2249 * @addr: target user address to start at
2250 * @pfn: physical address of kernel memory
2251 * @size: size of map area
2252 * @prot: page protection flags for this mapping
2254 * Note: this is only safe if the mm semaphore is held when called.
2256 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2257 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2261 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2262 struct mm_struct
*mm
= vma
->vm_mm
;
2266 * Physically remapped pages are special. Tell the
2267 * rest of the world about it:
2268 * VM_IO tells people not to look at these pages
2269 * (accesses can have side effects).
2270 * VM_RESERVED is specified all over the place, because
2271 * in 2.4 it kept swapout's vma scan off this vma; but
2272 * in 2.6 the LRU scan won't even find its pages, so this
2273 * flag means no more than count its pages in reserved_vm,
2274 * and omit it from core dump, even when VM_IO turned off.
2275 * VM_PFNMAP tells the core MM that the base pages are just
2276 * raw PFN mappings, and do not have a "struct page" associated
2279 * There's a horrible special case to handle copy-on-write
2280 * behaviour that some programs depend on. We mark the "original"
2281 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2283 if (addr
== vma
->vm_start
&& end
== vma
->vm_end
) {
2284 vma
->vm_pgoff
= pfn
;
2285 vma
->vm_flags
|= VM_PFN_AT_MMAP
;
2286 } else if (is_cow_mapping(vma
->vm_flags
))
2289 vma
->vm_flags
|= VM_IO
| VM_RESERVED
| VM_PFNMAP
;
2291 err
= track_pfn_vma_new(vma
, &prot
, pfn
, PAGE_ALIGN(size
));
2294 * To indicate that track_pfn related cleanup is not
2295 * needed from higher level routine calling unmap_vmas
2297 vma
->vm_flags
&= ~(VM_IO
| VM_RESERVED
| VM_PFNMAP
);
2298 vma
->vm_flags
&= ~VM_PFN_AT_MMAP
;
2302 BUG_ON(addr
>= end
);
2303 pfn
-= addr
>> PAGE_SHIFT
;
2304 pgd
= pgd_offset(mm
, addr
);
2305 flush_cache_range(vma
, addr
, end
);
2307 next
= pgd_addr_end(addr
, end
);
2308 err
= remap_pud_range(mm
, pgd
, addr
, next
,
2309 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2312 } while (pgd
++, addr
= next
, addr
!= end
);
2315 untrack_pfn_vma(vma
, pfn
, PAGE_ALIGN(size
));
2319 EXPORT_SYMBOL(remap_pfn_range
);
2321 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2322 unsigned long addr
, unsigned long end
,
2323 pte_fn_t fn
, void *data
)
2328 spinlock_t
*uninitialized_var(ptl
);
2330 pte
= (mm
== &init_mm
) ?
2331 pte_alloc_kernel(pmd
, addr
) :
2332 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2336 BUG_ON(pmd_huge(*pmd
));
2338 arch_enter_lazy_mmu_mode();
2340 token
= pmd_pgtable(*pmd
);
2343 err
= fn(pte
++, token
, addr
, data
);
2346 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2348 arch_leave_lazy_mmu_mode();
2351 pte_unmap_unlock(pte
-1, ptl
);
2355 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2356 unsigned long addr
, unsigned long end
,
2357 pte_fn_t fn
, void *data
)
2363 BUG_ON(pud_huge(*pud
));
2365 pmd
= pmd_alloc(mm
, pud
, addr
);
2369 next
= pmd_addr_end(addr
, end
);
2370 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
2373 } while (pmd
++, addr
= next
, addr
!= end
);
2377 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2378 unsigned long addr
, unsigned long end
,
2379 pte_fn_t fn
, void *data
)
2385 pud
= pud_alloc(mm
, pgd
, addr
);
2389 next
= pud_addr_end(addr
, end
);
2390 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
2393 } while (pud
++, addr
= next
, addr
!= end
);
2398 * Scan a region of virtual memory, filling in page tables as necessary
2399 * and calling a provided function on each leaf page table.
2401 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2402 unsigned long size
, pte_fn_t fn
, void *data
)
2406 unsigned long end
= addr
+ size
;
2409 BUG_ON(addr
>= end
);
2410 pgd
= pgd_offset(mm
, addr
);
2412 next
= pgd_addr_end(addr
, end
);
2413 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
2416 } while (pgd
++, addr
= next
, addr
!= end
);
2420 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2423 * handle_pte_fault chooses page fault handler according to an entry
2424 * which was read non-atomically. Before making any commitment, on
2425 * those architectures or configurations (e.g. i386 with PAE) which
2426 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2427 * must check under lock before unmapping the pte and proceeding
2428 * (but do_wp_page is only called after already making such a check;
2429 * and do_anonymous_page can safely check later on).
2431 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2432 pte_t
*page_table
, pte_t orig_pte
)
2435 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2436 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2437 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2439 same
= pte_same(*page_table
, orig_pte
);
2443 pte_unmap(page_table
);
2447 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
2450 * If the source page was a PFN mapping, we don't have
2451 * a "struct page" for it. We do a best-effort copy by
2452 * just copying from the original user address. If that
2453 * fails, we just zero-fill it. Live with it.
2455 if (unlikely(!src
)) {
2456 void *kaddr
= kmap_atomic(dst
);
2457 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
2460 * This really shouldn't fail, because the page is there
2461 * in the page tables. But it might just be unreadable,
2462 * in which case we just give up and fill the result with
2465 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
2467 kunmap_atomic(kaddr
);
2468 flush_dcache_page(dst
);
2470 copy_user_highpage(dst
, src
, va
, vma
);
2474 * This routine handles present pages, when users try to write
2475 * to a shared page. It is done by copying the page to a new address
2476 * and decrementing the shared-page counter for the old page.
2478 * Note that this routine assumes that the protection checks have been
2479 * done by the caller (the low-level page fault routine in most cases).
2480 * Thus we can safely just mark it writable once we've done any necessary
2483 * We also mark the page dirty at this point even though the page will
2484 * change only once the write actually happens. This avoids a few races,
2485 * and potentially makes it more efficient.
2487 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2488 * but allow concurrent faults), with pte both mapped and locked.
2489 * We return with mmap_sem still held, but pte unmapped and unlocked.
2491 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2492 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2493 spinlock_t
*ptl
, pte_t orig_pte
)
2496 struct page
*old_page
, *new_page
;
2499 int page_mkwrite
= 0;
2500 struct page
*dirty_page
= NULL
;
2502 old_page
= vm_normal_page(vma
, address
, orig_pte
);
2505 * VM_MIXEDMAP !pfn_valid() case
2507 * We should not cow pages in a shared writeable mapping.
2508 * Just mark the pages writable as we can't do any dirty
2509 * accounting on raw pfn maps.
2511 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2512 (VM_WRITE
|VM_SHARED
))
2518 * Take out anonymous pages first, anonymous shared vmas are
2519 * not dirty accountable.
2521 if (PageAnon(old_page
) && !PageKsm(old_page
)) {
2522 if (!trylock_page(old_page
)) {
2523 page_cache_get(old_page
);
2524 pte_unmap_unlock(page_table
, ptl
);
2525 lock_page(old_page
);
2526 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2528 if (!pte_same(*page_table
, orig_pte
)) {
2529 unlock_page(old_page
);
2532 page_cache_release(old_page
);
2534 if (reuse_swap_page(old_page
)) {
2536 * The page is all ours. Move it to our anon_vma so
2537 * the rmap code will not search our parent or siblings.
2538 * Protected against the rmap code by the page lock.
2540 page_move_anon_rmap(old_page
, vma
, address
);
2541 unlock_page(old_page
);
2544 unlock_page(old_page
);
2545 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2546 (VM_WRITE
|VM_SHARED
))) {
2548 * Only catch write-faults on shared writable pages,
2549 * read-only shared pages can get COWed by
2550 * get_user_pages(.write=1, .force=1).
2552 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2553 struct vm_fault vmf
;
2556 vmf
.virtual_address
= (void __user
*)(address
&
2558 vmf
.pgoff
= old_page
->index
;
2559 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2560 vmf
.page
= old_page
;
2563 * Notify the address space that the page is about to
2564 * become writable so that it can prohibit this or wait
2565 * for the page to get into an appropriate state.
2567 * We do this without the lock held, so that it can
2568 * sleep if it needs to.
2570 page_cache_get(old_page
);
2571 pte_unmap_unlock(page_table
, ptl
);
2573 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
2575 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
2577 goto unwritable_page
;
2579 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
2580 lock_page(old_page
);
2581 if (!old_page
->mapping
) {
2582 ret
= 0; /* retry the fault */
2583 unlock_page(old_page
);
2584 goto unwritable_page
;
2587 VM_BUG_ON(!PageLocked(old_page
));
2590 * Since we dropped the lock we need to revalidate
2591 * the PTE as someone else may have changed it. If
2592 * they did, we just return, as we can count on the
2593 * MMU to tell us if they didn't also make it writable.
2595 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2597 if (!pte_same(*page_table
, orig_pte
)) {
2598 unlock_page(old_page
);
2604 dirty_page
= old_page
;
2605 get_page(dirty_page
);
2608 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2609 entry
= pte_mkyoung(orig_pte
);
2610 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2611 if (ptep_set_access_flags(vma
, address
, page_table
, entry
,1))
2612 update_mmu_cache(vma
, address
, page_table
);
2613 pte_unmap_unlock(page_table
, ptl
);
2614 ret
|= VM_FAULT_WRITE
;
2620 * Yes, Virginia, this is actually required to prevent a race
2621 * with clear_page_dirty_for_io() from clearing the page dirty
2622 * bit after it clear all dirty ptes, but before a racing
2623 * do_wp_page installs a dirty pte.
2625 * __do_fault is protected similarly.
2627 if (!page_mkwrite
) {
2628 wait_on_page_locked(dirty_page
);
2629 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2631 put_page(dirty_page
);
2633 struct address_space
*mapping
= dirty_page
->mapping
;
2635 set_page_dirty(dirty_page
);
2636 unlock_page(dirty_page
);
2637 page_cache_release(dirty_page
);
2640 * Some device drivers do not set page.mapping
2641 * but still dirty their pages
2643 balance_dirty_pages_ratelimited(mapping
);
2647 /* file_update_time outside page_lock */
2649 file_update_time(vma
->vm_file
);
2655 * Ok, we need to copy. Oh, well..
2657 page_cache_get(old_page
);
2659 pte_unmap_unlock(page_table
, ptl
);
2661 if (unlikely(anon_vma_prepare(vma
)))
2664 if (is_zero_pfn(pte_pfn(orig_pte
))) {
2665 new_page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2669 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2672 cow_user_page(new_page
, old_page
, address
, vma
);
2674 __SetPageUptodate(new_page
);
2676 if (mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
))
2680 * Re-check the pte - we dropped the lock
2682 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2683 if (likely(pte_same(*page_table
, orig_pte
))) {
2685 if (!PageAnon(old_page
)) {
2686 dec_mm_counter_fast(mm
, MM_FILEPAGES
);
2687 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2690 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2691 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2692 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2693 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2695 * Clear the pte entry and flush it first, before updating the
2696 * pte with the new entry. This will avoid a race condition
2697 * seen in the presence of one thread doing SMC and another
2700 ptep_clear_flush(vma
, address
, page_table
);
2701 page_add_new_anon_rmap(new_page
, vma
, address
);
2703 * We call the notify macro here because, when using secondary
2704 * mmu page tables (such as kvm shadow page tables), we want the
2705 * new page to be mapped directly into the secondary page table.
2707 set_pte_at_notify(mm
, address
, page_table
, entry
);
2708 update_mmu_cache(vma
, address
, page_table
);
2711 * Only after switching the pte to the new page may
2712 * we remove the mapcount here. Otherwise another
2713 * process may come and find the rmap count decremented
2714 * before the pte is switched to the new page, and
2715 * "reuse" the old page writing into it while our pte
2716 * here still points into it and can be read by other
2719 * The critical issue is to order this
2720 * page_remove_rmap with the ptp_clear_flush above.
2721 * Those stores are ordered by (if nothing else,)
2722 * the barrier present in the atomic_add_negative
2723 * in page_remove_rmap.
2725 * Then the TLB flush in ptep_clear_flush ensures that
2726 * no process can access the old page before the
2727 * decremented mapcount is visible. And the old page
2728 * cannot be reused until after the decremented
2729 * mapcount is visible. So transitively, TLBs to
2730 * old page will be flushed before it can be reused.
2732 page_remove_rmap(old_page
);
2735 /* Free the old page.. */
2736 new_page
= old_page
;
2737 ret
|= VM_FAULT_WRITE
;
2739 mem_cgroup_uncharge_page(new_page
);
2742 page_cache_release(new_page
);
2744 pte_unmap_unlock(page_table
, ptl
);
2747 * Don't let another task, with possibly unlocked vma,
2748 * keep the mlocked page.
2750 if ((ret
& VM_FAULT_WRITE
) && (vma
->vm_flags
& VM_LOCKED
)) {
2751 lock_page(old_page
); /* LRU manipulation */
2752 munlock_vma_page(old_page
);
2753 unlock_page(old_page
);
2755 page_cache_release(old_page
);
2759 page_cache_release(new_page
);
2763 unlock_page(old_page
);
2764 page_cache_release(old_page
);
2766 page_cache_release(old_page
);
2768 return VM_FAULT_OOM
;
2771 page_cache_release(old_page
);
2775 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2776 unsigned long start_addr
, unsigned long end_addr
,
2777 struct zap_details
*details
)
2779 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2782 static inline void unmap_mapping_range_tree(struct prio_tree_root
*root
,
2783 struct zap_details
*details
)
2785 struct vm_area_struct
*vma
;
2786 struct prio_tree_iter iter
;
2787 pgoff_t vba
, vea
, zba
, zea
;
2789 vma_prio_tree_foreach(vma
, &iter
, root
,
2790 details
->first_index
, details
->last_index
) {
2792 vba
= vma
->vm_pgoff
;
2793 vea
= vba
+ ((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) - 1;
2794 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2795 zba
= details
->first_index
;
2798 zea
= details
->last_index
;
2802 unmap_mapping_range_vma(vma
,
2803 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2804 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2809 static inline void unmap_mapping_range_list(struct list_head
*head
,
2810 struct zap_details
*details
)
2812 struct vm_area_struct
*vma
;
2815 * In nonlinear VMAs there is no correspondence between virtual address
2816 * offset and file offset. So we must perform an exhaustive search
2817 * across *all* the pages in each nonlinear VMA, not just the pages
2818 * whose virtual address lies outside the file truncation point.
2820 list_for_each_entry(vma
, head
, shared
.vm_set
.list
) {
2821 details
->nonlinear_vma
= vma
;
2822 unmap_mapping_range_vma(vma
, vma
->vm_start
, vma
->vm_end
, details
);
2827 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2828 * @mapping: the address space containing mmaps to be unmapped.
2829 * @holebegin: byte in first page to unmap, relative to the start of
2830 * the underlying file. This will be rounded down to a PAGE_SIZE
2831 * boundary. Note that this is different from truncate_pagecache(), which
2832 * must keep the partial page. In contrast, we must get rid of
2834 * @holelen: size of prospective hole in bytes. This will be rounded
2835 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2837 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2838 * but 0 when invalidating pagecache, don't throw away private data.
2840 void unmap_mapping_range(struct address_space
*mapping
,
2841 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2843 struct zap_details details
;
2844 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2845 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2847 /* Check for overflow. */
2848 if (sizeof(holelen
) > sizeof(hlen
)) {
2850 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2851 if (holeend
& ~(long long)ULONG_MAX
)
2852 hlen
= ULONG_MAX
- hba
+ 1;
2855 details
.check_mapping
= even_cows
? NULL
: mapping
;
2856 details
.nonlinear_vma
= NULL
;
2857 details
.first_index
= hba
;
2858 details
.last_index
= hba
+ hlen
- 1;
2859 if (details
.last_index
< details
.first_index
)
2860 details
.last_index
= ULONG_MAX
;
2863 mutex_lock(&mapping
->i_mmap_mutex
);
2864 if (unlikely(!prio_tree_empty(&mapping
->i_mmap
)))
2865 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2866 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
2867 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
2868 mutex_unlock(&mapping
->i_mmap_mutex
);
2870 EXPORT_SYMBOL(unmap_mapping_range
);
2873 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2874 * but allow concurrent faults), and pte mapped but not yet locked.
2875 * We return with mmap_sem still held, but pte unmapped and unlocked.
2877 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2878 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2879 unsigned int flags
, pte_t orig_pte
)
2882 struct page
*page
, *swapcache
= NULL
;
2886 struct mem_cgroup
*ptr
;
2890 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2893 entry
= pte_to_swp_entry(orig_pte
);
2894 if (unlikely(non_swap_entry(entry
))) {
2895 if (is_migration_entry(entry
)) {
2896 migration_entry_wait(mm
, pmd
, address
);
2897 } else if (is_hwpoison_entry(entry
)) {
2898 ret
= VM_FAULT_HWPOISON
;
2900 print_bad_pte(vma
, address
, orig_pte
, NULL
);
2901 ret
= VM_FAULT_SIGBUS
;
2905 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2906 page
= lookup_swap_cache(entry
);
2908 grab_swap_token(mm
); /* Contend for token _before_ read-in */
2909 page
= swapin_readahead(entry
,
2910 GFP_HIGHUSER_MOVABLE
, vma
, address
);
2913 * Back out if somebody else faulted in this pte
2914 * while we released the pte lock.
2916 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2917 if (likely(pte_same(*page_table
, orig_pte
)))
2919 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2923 /* Had to read the page from swap area: Major fault */
2924 ret
= VM_FAULT_MAJOR
;
2925 count_vm_event(PGMAJFAULT
);
2926 mem_cgroup_count_vm_event(mm
, PGMAJFAULT
);
2927 } else if (PageHWPoison(page
)) {
2929 * hwpoisoned dirty swapcache pages are kept for killing
2930 * owner processes (which may be unknown at hwpoison time)
2932 ret
= VM_FAULT_HWPOISON
;
2933 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2937 locked
= lock_page_or_retry(page
, mm
, flags
);
2938 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2940 ret
|= VM_FAULT_RETRY
;
2945 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2946 * release the swapcache from under us. The page pin, and pte_same
2947 * test below, are not enough to exclude that. Even if it is still
2948 * swapcache, we need to check that the page's swap has not changed.
2950 if (unlikely(!PageSwapCache(page
) || page_private(page
) != entry
.val
))
2953 if (ksm_might_need_to_copy(page
, vma
, address
)) {
2955 page
= ksm_does_need_to_copy(page
, vma
, address
);
2957 if (unlikely(!page
)) {
2965 if (mem_cgroup_try_charge_swapin(mm
, page
, GFP_KERNEL
, &ptr
)) {
2971 * Back out if somebody else already faulted in this pte.
2973 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2974 if (unlikely(!pte_same(*page_table
, orig_pte
)))
2977 if (unlikely(!PageUptodate(page
))) {
2978 ret
= VM_FAULT_SIGBUS
;
2983 * The page isn't present yet, go ahead with the fault.
2985 * Be careful about the sequence of operations here.
2986 * To get its accounting right, reuse_swap_page() must be called
2987 * while the page is counted on swap but not yet in mapcount i.e.
2988 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2989 * must be called after the swap_free(), or it will never succeed.
2990 * Because delete_from_swap_page() may be called by reuse_swap_page(),
2991 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2992 * in page->private. In this case, a record in swap_cgroup is silently
2993 * discarded at swap_free().
2996 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2997 dec_mm_counter_fast(mm
, MM_SWAPENTS
);
2998 pte
= mk_pte(page
, vma
->vm_page_prot
);
2999 if ((flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
)) {
3000 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3001 flags
&= ~FAULT_FLAG_WRITE
;
3002 ret
|= VM_FAULT_WRITE
;
3005 flush_icache_page(vma
, page
);
3006 set_pte_at(mm
, address
, page_table
, pte
);
3007 do_page_add_anon_rmap(page
, vma
, address
, exclusive
);
3008 /* It's better to call commit-charge after rmap is established */
3009 mem_cgroup_commit_charge_swapin(page
, ptr
);
3012 if (vm_swap_full() || (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
3013 try_to_free_swap(page
);
3017 * Hold the lock to avoid the swap entry to be reused
3018 * until we take the PT lock for the pte_same() check
3019 * (to avoid false positives from pte_same). For
3020 * further safety release the lock after the swap_free
3021 * so that the swap count won't change under a
3022 * parallel locked swapcache.
3024 unlock_page(swapcache
);
3025 page_cache_release(swapcache
);
3028 if (flags
& FAULT_FLAG_WRITE
) {
3029 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
3030 if (ret
& VM_FAULT_ERROR
)
3031 ret
&= VM_FAULT_ERROR
;
3035 /* No need to invalidate - it was non-present before */
3036 update_mmu_cache(vma
, address
, page_table
);
3038 pte_unmap_unlock(page_table
, ptl
);
3042 mem_cgroup_cancel_charge_swapin(ptr
);
3043 pte_unmap_unlock(page_table
, ptl
);
3047 page_cache_release(page
);
3049 unlock_page(swapcache
);
3050 page_cache_release(swapcache
);
3056 * This is like a special single-page "expand_{down|up}wards()",
3057 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3058 * doesn't hit another vma.
3060 static inline int check_stack_guard_page(struct vm_area_struct
*vma
, unsigned long address
)
3062 address
&= PAGE_MASK
;
3063 if ((vma
->vm_flags
& VM_GROWSDOWN
) && address
== vma
->vm_start
) {
3064 struct vm_area_struct
*prev
= vma
->vm_prev
;
3067 * Is there a mapping abutting this one below?
3069 * That's only ok if it's the same stack mapping
3070 * that has gotten split..
3072 if (prev
&& prev
->vm_end
== address
)
3073 return prev
->vm_flags
& VM_GROWSDOWN
? 0 : -ENOMEM
;
3075 expand_downwards(vma
, address
- PAGE_SIZE
);
3077 if ((vma
->vm_flags
& VM_GROWSUP
) && address
+ PAGE_SIZE
== vma
->vm_end
) {
3078 struct vm_area_struct
*next
= vma
->vm_next
;
3080 /* As VM_GROWSDOWN but s/below/above/ */
3081 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
)
3082 return next
->vm_flags
& VM_GROWSUP
? 0 : -ENOMEM
;
3084 expand_upwards(vma
, address
+ PAGE_SIZE
);
3090 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3091 * but allow concurrent faults), and pte mapped but not yet locked.
3092 * We return with mmap_sem still held, but pte unmapped and unlocked.
3094 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3095 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3102 pte_unmap(page_table
);
3104 /* Check if we need to add a guard page to the stack */
3105 if (check_stack_guard_page(vma
, address
) < 0)
3106 return VM_FAULT_SIGBUS
;
3108 /* Use the zero-page for reads */
3109 if (!(flags
& FAULT_FLAG_WRITE
)) {
3110 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(address
),
3111 vma
->vm_page_prot
));
3112 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3113 if (!pte_none(*page_table
))
3118 /* Allocate our own private page. */
3119 if (unlikely(anon_vma_prepare(vma
)))
3121 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
3124 __SetPageUptodate(page
);
3126 if (mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
))
3129 entry
= mk_pte(page
, vma
->vm_page_prot
);
3130 if (vma
->vm_flags
& VM_WRITE
)
3131 entry
= pte_mkwrite(pte_mkdirty(entry
));
3133 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3134 if (!pte_none(*page_table
))
3137 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3138 page_add_new_anon_rmap(page
, vma
, address
);
3140 set_pte_at(mm
, address
, page_table
, entry
);
3142 /* No need to invalidate - it was non-present before */
3143 update_mmu_cache(vma
, address
, page_table
);
3145 pte_unmap_unlock(page_table
, ptl
);
3148 mem_cgroup_uncharge_page(page
);
3149 page_cache_release(page
);
3152 page_cache_release(page
);
3154 return VM_FAULT_OOM
;
3158 * __do_fault() tries to create a new page mapping. It aggressively
3159 * tries to share with existing pages, but makes a separate copy if
3160 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3161 * the next page fault.
3163 * As this is called only for pages that do not currently exist, we
3164 * do not need to flush old virtual caches or the TLB.
3166 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3167 * but allow concurrent faults), and pte neither mapped nor locked.
3168 * We return with mmap_sem still held, but pte unmapped and unlocked.
3170 static int __do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3171 unsigned long address
, pmd_t
*pmd
,
3172 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3177 struct page
*cow_page
;
3180 struct page
*dirty_page
= NULL
;
3181 struct vm_fault vmf
;
3183 int page_mkwrite
= 0;
3186 * If we do COW later, allocate page befor taking lock_page()
3187 * on the file cache page. This will reduce lock holding time.
3189 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3191 if (unlikely(anon_vma_prepare(vma
)))
3192 return VM_FAULT_OOM
;
3194 cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
3196 return VM_FAULT_OOM
;
3198 if (mem_cgroup_newpage_charge(cow_page
, mm
, GFP_KERNEL
)) {
3199 page_cache_release(cow_page
);
3200 return VM_FAULT_OOM
;
3205 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
3210 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
3211 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
3215 if (unlikely(PageHWPoison(vmf
.page
))) {
3216 if (ret
& VM_FAULT_LOCKED
)
3217 unlock_page(vmf
.page
);
3218 ret
= VM_FAULT_HWPOISON
;
3223 * For consistency in subsequent calls, make the faulted page always
3226 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3227 lock_page(vmf
.page
);
3229 VM_BUG_ON(!PageLocked(vmf
.page
));
3232 * Should we do an early C-O-W break?
3235 if (flags
& FAULT_FLAG_WRITE
) {
3236 if (!(vma
->vm_flags
& VM_SHARED
)) {
3239 copy_user_highpage(page
, vmf
.page
, address
, vma
);
3240 __SetPageUptodate(page
);
3243 * If the page will be shareable, see if the backing
3244 * address space wants to know that the page is about
3245 * to become writable
3247 if (vma
->vm_ops
->page_mkwrite
) {
3251 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
3252 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
3254 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3256 goto unwritable_page
;
3258 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
3260 if (!page
->mapping
) {
3261 ret
= 0; /* retry the fault */
3263 goto unwritable_page
;
3266 VM_BUG_ON(!PageLocked(page
));
3273 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3276 * This silly early PAGE_DIRTY setting removes a race
3277 * due to the bad i386 page protection. But it's valid
3278 * for other architectures too.
3280 * Note that if FAULT_FLAG_WRITE is set, we either now have
3281 * an exclusive copy of the page, or this is a shared mapping,
3282 * so we can make it writable and dirty to avoid having to
3283 * handle that later.
3285 /* Only go through if we didn't race with anybody else... */
3286 if (likely(pte_same(*page_table
, orig_pte
))) {
3287 flush_icache_page(vma
, page
);
3288 entry
= mk_pte(page
, vma
->vm_page_prot
);
3289 if (flags
& FAULT_FLAG_WRITE
)
3290 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3292 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3293 page_add_new_anon_rmap(page
, vma
, address
);
3295 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
3296 page_add_file_rmap(page
);
3297 if (flags
& FAULT_FLAG_WRITE
) {
3299 get_page(dirty_page
);
3302 set_pte_at(mm
, address
, page_table
, entry
);
3304 /* no need to invalidate: a not-present page won't be cached */
3305 update_mmu_cache(vma
, address
, page_table
);
3308 mem_cgroup_uncharge_page(cow_page
);
3310 page_cache_release(page
);
3312 anon
= 1; /* no anon but release faulted_page */
3315 pte_unmap_unlock(page_table
, ptl
);
3318 struct address_space
*mapping
= page
->mapping
;
3320 if (set_page_dirty(dirty_page
))
3322 unlock_page(dirty_page
);
3323 put_page(dirty_page
);
3324 if (page_mkwrite
&& mapping
) {
3326 * Some device drivers do not set page.mapping but still
3329 balance_dirty_pages_ratelimited(mapping
);
3332 /* file_update_time outside page_lock */
3334 file_update_time(vma
->vm_file
);
3336 unlock_page(vmf
.page
);
3338 page_cache_release(vmf
.page
);
3344 page_cache_release(page
);
3347 /* fs's fault handler get error */
3349 mem_cgroup_uncharge_page(cow_page
);
3350 page_cache_release(cow_page
);
3355 static int do_linear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3356 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3357 unsigned int flags
, pte_t orig_pte
)
3359 pgoff_t pgoff
= (((address
& PAGE_MASK
)
3360 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
3362 pte_unmap(page_table
);
3363 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3367 * Fault of a previously existing named mapping. Repopulate the pte
3368 * from the encoded file_pte if possible. This enables swappable
3371 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3372 * but allow concurrent faults), and pte mapped but not yet locked.
3373 * We return with mmap_sem still held, but pte unmapped and unlocked.
3375 static int do_nonlinear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3376 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3377 unsigned int flags
, pte_t orig_pte
)
3381 flags
|= FAULT_FLAG_NONLINEAR
;
3383 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
3386 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
))) {
3388 * Page table corrupted: show pte and kill process.
3390 print_bad_pte(vma
, address
, orig_pte
, NULL
);
3391 return VM_FAULT_SIGBUS
;
3394 pgoff
= pte_to_pgoff(orig_pte
);
3395 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3399 * These routines also need to handle stuff like marking pages dirty
3400 * and/or accessed for architectures that don't do it in hardware (most
3401 * RISC architectures). The early dirtying is also good on the i386.
3403 * There is also a hook called "update_mmu_cache()" that architectures
3404 * with external mmu caches can use to update those (ie the Sparc or
3405 * PowerPC hashed page tables that act as extended TLBs).
3407 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3408 * but allow concurrent faults), and pte mapped but not yet locked.
3409 * We return with mmap_sem still held, but pte unmapped and unlocked.
3411 int handle_pte_fault(struct mm_struct
*mm
,
3412 struct vm_area_struct
*vma
, unsigned long address
,
3413 pte_t
*pte
, pmd_t
*pmd
, unsigned int flags
)
3419 if (!pte_present(entry
)) {
3420 if (pte_none(entry
)) {
3422 if (likely(vma
->vm_ops
->fault
))
3423 return do_linear_fault(mm
, vma
, address
,
3424 pte
, pmd
, flags
, entry
);
3426 return do_anonymous_page(mm
, vma
, address
,
3429 if (pte_file(entry
))
3430 return do_nonlinear_fault(mm
, vma
, address
,
3431 pte
, pmd
, flags
, entry
);
3432 return do_swap_page(mm
, vma
, address
,
3433 pte
, pmd
, flags
, entry
);
3436 ptl
= pte_lockptr(mm
, pmd
);
3438 if (unlikely(!pte_same(*pte
, entry
)))
3440 if (flags
& FAULT_FLAG_WRITE
) {
3441 if (!pte_write(entry
))
3442 return do_wp_page(mm
, vma
, address
,
3443 pte
, pmd
, ptl
, entry
);
3444 entry
= pte_mkdirty(entry
);
3446 entry
= pte_mkyoung(entry
);
3447 if (ptep_set_access_flags(vma
, address
, pte
, entry
, flags
& FAULT_FLAG_WRITE
)) {
3448 update_mmu_cache(vma
, address
, pte
);
3451 * This is needed only for protection faults but the arch code
3452 * is not yet telling us if this is a protection fault or not.
3453 * This still avoids useless tlb flushes for .text page faults
3456 if (flags
& FAULT_FLAG_WRITE
)
3457 flush_tlb_fix_spurious_fault(vma
, address
);
3460 pte_unmap_unlock(pte
, ptl
);
3465 * By the time we get here, we already hold the mm semaphore
3467 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3468 unsigned long address
, unsigned int flags
)
3475 __set_current_state(TASK_RUNNING
);
3477 count_vm_event(PGFAULT
);
3478 mem_cgroup_count_vm_event(mm
, PGFAULT
);
3480 /* do counter updates before entering really critical section. */
3481 check_sync_rss_stat(current
);
3483 if (unlikely(is_vm_hugetlb_page(vma
)))
3484 return hugetlb_fault(mm
, vma
, address
, flags
);
3486 pgd
= pgd_offset(mm
, address
);
3487 pud
= pud_alloc(mm
, pgd
, address
);
3489 return VM_FAULT_OOM
;
3490 pmd
= pmd_alloc(mm
, pud
, address
);
3492 return VM_FAULT_OOM
;
3493 if (pmd_none(*pmd
) && transparent_hugepage_enabled(vma
)) {
3495 return do_huge_pmd_anonymous_page(mm
, vma
, address
,
3498 pmd_t orig_pmd
= *pmd
;
3500 if (pmd_trans_huge(orig_pmd
)) {
3501 if (flags
& FAULT_FLAG_WRITE
&&
3502 !pmd_write(orig_pmd
) &&
3503 !pmd_trans_splitting(orig_pmd
))
3504 return do_huge_pmd_wp_page(mm
, vma
, address
,
3511 * Use __pte_alloc instead of pte_alloc_map, because we can't
3512 * run pte_offset_map on the pmd, if an huge pmd could
3513 * materialize from under us from a different thread.
3515 if (unlikely(pmd_none(*pmd
)) && __pte_alloc(mm
, vma
, pmd
, address
))
3516 return VM_FAULT_OOM
;
3517 /* if an huge pmd materialized from under us just retry later */
3518 if (unlikely(pmd_trans_huge(*pmd
)))
3521 * A regular pmd is established and it can't morph into a huge pmd
3522 * from under us anymore at this point because we hold the mmap_sem
3523 * read mode and khugepaged takes it in write mode. So now it's
3524 * safe to run pte_offset_map().
3526 pte
= pte_offset_map(pmd
, address
);
3528 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
3531 #ifndef __PAGETABLE_PUD_FOLDED
3533 * Allocate page upper directory.
3534 * We've already handled the fast-path in-line.
3536 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
3538 pud_t
*new = pud_alloc_one(mm
, address
);
3542 smp_wmb(); /* See comment in __pte_alloc */
3544 spin_lock(&mm
->page_table_lock
);
3545 if (pgd_present(*pgd
)) /* Another has populated it */
3548 pgd_populate(mm
, pgd
, new);
3549 spin_unlock(&mm
->page_table_lock
);
3552 #endif /* __PAGETABLE_PUD_FOLDED */
3554 #ifndef __PAGETABLE_PMD_FOLDED
3556 * Allocate page middle directory.
3557 * We've already handled the fast-path in-line.
3559 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
3561 pmd_t
*new = pmd_alloc_one(mm
, address
);
3565 smp_wmb(); /* See comment in __pte_alloc */
3567 spin_lock(&mm
->page_table_lock
);
3568 #ifndef __ARCH_HAS_4LEVEL_HACK
3569 if (pud_present(*pud
)) /* Another has populated it */
3572 pud_populate(mm
, pud
, new);
3574 if (pgd_present(*pud
)) /* Another has populated it */
3577 pgd_populate(mm
, pud
, new);
3578 #endif /* __ARCH_HAS_4LEVEL_HACK */
3579 spin_unlock(&mm
->page_table_lock
);
3582 #endif /* __PAGETABLE_PMD_FOLDED */
3584 int make_pages_present(unsigned long addr
, unsigned long end
)
3586 int ret
, len
, write
;
3587 struct vm_area_struct
* vma
;
3589 vma
= find_vma(current
->mm
, addr
);
3593 * We want to touch writable mappings with a write fault in order
3594 * to break COW, except for shared mappings because these don't COW
3595 * and we would not want to dirty them for nothing.
3597 write
= (vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
;
3598 BUG_ON(addr
>= end
);
3599 BUG_ON(end
> vma
->vm_end
);
3600 len
= DIV_ROUND_UP(end
, PAGE_SIZE
) - addr
/PAGE_SIZE
;
3601 ret
= get_user_pages(current
, current
->mm
, addr
,
3602 len
, write
, 0, NULL
, NULL
);
3605 return ret
== len
? 0 : -EFAULT
;
3608 #if !defined(__HAVE_ARCH_GATE_AREA)
3610 #if defined(AT_SYSINFO_EHDR)
3611 static struct vm_area_struct gate_vma
;
3613 static int __init
gate_vma_init(void)
3615 gate_vma
.vm_mm
= NULL
;
3616 gate_vma
.vm_start
= FIXADDR_USER_START
;
3617 gate_vma
.vm_end
= FIXADDR_USER_END
;
3618 gate_vma
.vm_flags
= VM_READ
| VM_MAYREAD
| VM_EXEC
| VM_MAYEXEC
;
3619 gate_vma
.vm_page_prot
= __P101
;
3623 __initcall(gate_vma_init
);
3626 struct vm_area_struct
*get_gate_vma(struct mm_struct
*mm
)
3628 #ifdef AT_SYSINFO_EHDR
3635 int in_gate_area_no_mm(unsigned long addr
)
3637 #ifdef AT_SYSINFO_EHDR
3638 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
3644 #endif /* __HAVE_ARCH_GATE_AREA */
3646 static int __follow_pte(struct mm_struct
*mm
, unsigned long address
,
3647 pte_t
**ptepp
, spinlock_t
**ptlp
)
3654 pgd
= pgd_offset(mm
, address
);
3655 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
3658 pud
= pud_offset(pgd
, address
);
3659 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
3662 pmd
= pmd_offset(pud
, address
);
3663 VM_BUG_ON(pmd_trans_huge(*pmd
));
3664 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
3667 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3671 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
3674 if (!pte_present(*ptep
))
3679 pte_unmap_unlock(ptep
, *ptlp
);
3684 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
3685 pte_t
**ptepp
, spinlock_t
**ptlp
)
3689 /* (void) is needed to make gcc happy */
3690 (void) __cond_lock(*ptlp
,
3691 !(res
= __follow_pte(mm
, address
, ptepp
, ptlp
)));
3696 * follow_pfn - look up PFN at a user virtual address
3697 * @vma: memory mapping
3698 * @address: user virtual address
3699 * @pfn: location to store found PFN
3701 * Only IO mappings and raw PFN mappings are allowed.
3703 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3705 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
3712 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3715 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
3718 *pfn
= pte_pfn(*ptep
);
3719 pte_unmap_unlock(ptep
, ptl
);
3722 EXPORT_SYMBOL(follow_pfn
);
3724 #ifdef CONFIG_HAVE_IOREMAP_PROT
3725 int follow_phys(struct vm_area_struct
*vma
,
3726 unsigned long address
, unsigned int flags
,
3727 unsigned long *prot
, resource_size_t
*phys
)
3733 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3736 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
3740 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
3743 *prot
= pgprot_val(pte_pgprot(pte
));
3744 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
3748 pte_unmap_unlock(ptep
, ptl
);
3753 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
3754 void *buf
, int len
, int write
)
3756 resource_size_t phys_addr
;
3757 unsigned long prot
= 0;
3758 void __iomem
*maddr
;
3759 int offset
= addr
& (PAGE_SIZE
-1);
3761 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
3764 maddr
= ioremap_prot(phys_addr
, PAGE_SIZE
, prot
);
3766 memcpy_toio(maddr
+ offset
, buf
, len
);
3768 memcpy_fromio(buf
, maddr
+ offset
, len
);
3776 * Access another process' address space as given in mm. If non-NULL, use the
3777 * given task for page fault accounting.
3779 static int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
3780 unsigned long addr
, void *buf
, int len
, int write
)
3782 struct vm_area_struct
*vma
;
3783 void *old_buf
= buf
;
3785 down_read(&mm
->mmap_sem
);
3786 /* ignore errors, just check how much was successfully transferred */
3788 int bytes
, ret
, offset
;
3790 struct page
*page
= NULL
;
3792 ret
= get_user_pages(tsk
, mm
, addr
, 1,
3793 write
, 1, &page
, &vma
);
3796 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3797 * we can access using slightly different code.
3799 #ifdef CONFIG_HAVE_IOREMAP_PROT
3800 vma
= find_vma(mm
, addr
);
3801 if (!vma
|| vma
->vm_start
> addr
)
3803 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
3804 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
3812 offset
= addr
& (PAGE_SIZE
-1);
3813 if (bytes
> PAGE_SIZE
-offset
)
3814 bytes
= PAGE_SIZE
-offset
;
3818 copy_to_user_page(vma
, page
, addr
,
3819 maddr
+ offset
, buf
, bytes
);
3820 set_page_dirty_lock(page
);
3822 copy_from_user_page(vma
, page
, addr
,
3823 buf
, maddr
+ offset
, bytes
);
3826 page_cache_release(page
);
3832 up_read(&mm
->mmap_sem
);
3834 return buf
- old_buf
;
3838 * access_remote_vm - access another process' address space
3839 * @mm: the mm_struct of the target address space
3840 * @addr: start address to access
3841 * @buf: source or destination buffer
3842 * @len: number of bytes to transfer
3843 * @write: whether the access is a write
3845 * The caller must hold a reference on @mm.
3847 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
3848 void *buf
, int len
, int write
)
3850 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, write
);
3854 * Access another process' address space.
3855 * Source/target buffer must be kernel space,
3856 * Do not walk the page table directly, use get_user_pages
3858 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
3859 void *buf
, int len
, int write
)
3861 struct mm_struct
*mm
;
3864 mm
= get_task_mm(tsk
);
3868 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, write
);
3875 * Print the name of a VMA.
3877 void print_vma_addr(char *prefix
, unsigned long ip
)
3879 struct mm_struct
*mm
= current
->mm
;
3880 struct vm_area_struct
*vma
;
3883 * Do not print if we are in atomic
3884 * contexts (in exception stacks, etc.):
3886 if (preempt_count())
3889 down_read(&mm
->mmap_sem
);
3890 vma
= find_vma(mm
, ip
);
3891 if (vma
&& vma
->vm_file
) {
3892 struct file
*f
= vma
->vm_file
;
3893 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
3897 p
= d_path(&f
->f_path
, buf
, PAGE_SIZE
);
3900 s
= strrchr(p
, '/');
3903 printk("%s%s[%lx+%lx]", prefix
, p
,
3905 vma
->vm_end
- vma
->vm_start
);
3906 free_page((unsigned long)buf
);
3909 up_read(¤t
->mm
->mmap_sem
);
3912 #ifdef CONFIG_PROVE_LOCKING
3913 void might_fault(void)
3916 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3917 * holding the mmap_sem, this is safe because kernel memory doesn't
3918 * get paged out, therefore we'll never actually fault, and the
3919 * below annotations will generate false positives.
3921 if (segment_eq(get_fs(), KERNEL_DS
))
3926 * it would be nicer only to annotate paths which are not under
3927 * pagefault_disable, however that requires a larger audit and
3928 * providing helpers like get_user_atomic.
3930 if (!in_atomic() && current
->mm
)
3931 might_lock_read(¤t
->mm
->mmap_sem
);
3933 EXPORT_SYMBOL(might_fault
);
3936 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3937 static void clear_gigantic_page(struct page
*page
,
3939 unsigned int pages_per_huge_page
)
3942 struct page
*p
= page
;
3945 for (i
= 0; i
< pages_per_huge_page
;
3946 i
++, p
= mem_map_next(p
, page
, i
)) {
3948 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
3951 void clear_huge_page(struct page
*page
,
3952 unsigned long addr
, unsigned int pages_per_huge_page
)
3956 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3957 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
3962 for (i
= 0; i
< pages_per_huge_page
; i
++) {
3964 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
3968 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
3970 struct vm_area_struct
*vma
,
3971 unsigned int pages_per_huge_page
)
3974 struct page
*dst_base
= dst
;
3975 struct page
*src_base
= src
;
3977 for (i
= 0; i
< pages_per_huge_page
; ) {
3979 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
3982 dst
= mem_map_next(dst
, dst_base
, i
);
3983 src
= mem_map_next(src
, src_base
, i
);
3987 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
3988 unsigned long addr
, struct vm_area_struct
*vma
,
3989 unsigned int pages_per_huge_page
)
3993 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3994 copy_user_gigantic_page(dst
, src
, addr
, vma
,
3995 pages_per_huge_page
);
4000 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4002 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
4005 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */