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
;
212 tlb
->fast_mode
= (num_possible_cpus() == 1);
213 tlb
->local
.next
= NULL
;
215 tlb
->local
.max
= ARRAY_SIZE(tlb
->__pages
);
216 tlb
->active
= &tlb
->local
;
218 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
223 void tlb_flush_mmu(struct mmu_gather
*tlb
)
225 struct mmu_gather_batch
*batch
;
227 if (!tlb
->need_flush
)
231 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
232 tlb_table_flush(tlb
);
235 if (tlb_fast_mode(tlb
))
238 for (batch
= &tlb
->local
; batch
; batch
= batch
->next
) {
239 free_pages_and_swap_cache(batch
->pages
, batch
->nr
);
242 tlb
->active
= &tlb
->local
;
246 * Called at the end of the shootdown operation to free up any resources
247 * that were required.
249 void tlb_finish_mmu(struct mmu_gather
*tlb
, unsigned long start
, unsigned long end
)
251 struct mmu_gather_batch
*batch
, *next
;
257 /* keep the page table cache within bounds */
260 for (batch
= tlb
->local
.next
; batch
; batch
= next
) {
262 free_pages((unsigned long)batch
, 0);
264 tlb
->local
.next
= NULL
;
268 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
269 * handling the additional races in SMP caused by other CPUs caching valid
270 * mappings in their TLBs. Returns the number of free page slots left.
271 * When out of page slots we must call tlb_flush_mmu().
273 int __tlb_remove_page(struct mmu_gather
*tlb
, struct page
*page
)
275 struct mmu_gather_batch
*batch
;
277 VM_BUG_ON(!tlb
->need_flush
);
279 if (tlb_fast_mode(tlb
)) {
280 free_page_and_swap_cache(page
);
281 return 1; /* avoid calling tlb_flush_mmu() */
285 batch
->pages
[batch
->nr
++] = page
;
286 if (batch
->nr
== batch
->max
) {
287 if (!tlb_next_batch(tlb
))
291 VM_BUG_ON(batch
->nr
> batch
->max
);
293 return batch
->max
- batch
->nr
;
296 #endif /* HAVE_GENERIC_MMU_GATHER */
298 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
301 * See the comment near struct mmu_table_batch.
304 static void tlb_remove_table_smp_sync(void *arg
)
306 /* Simply deliver the interrupt */
309 static void tlb_remove_table_one(void *table
)
312 * This isn't an RCU grace period and hence the page-tables cannot be
313 * assumed to be actually RCU-freed.
315 * It is however sufficient for software page-table walkers that rely on
316 * IRQ disabling. See the comment near struct mmu_table_batch.
318 smp_call_function(tlb_remove_table_smp_sync
, NULL
, 1);
319 __tlb_remove_table(table
);
322 static void tlb_remove_table_rcu(struct rcu_head
*head
)
324 struct mmu_table_batch
*batch
;
327 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
329 for (i
= 0; i
< batch
->nr
; i
++)
330 __tlb_remove_table(batch
->tables
[i
]);
332 free_page((unsigned long)batch
);
335 void tlb_table_flush(struct mmu_gather
*tlb
)
337 struct mmu_table_batch
**batch
= &tlb
->batch
;
340 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
345 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
347 struct mmu_table_batch
**batch
= &tlb
->batch
;
352 * When there's less then two users of this mm there cannot be a
353 * concurrent page-table walk.
355 if (atomic_read(&tlb
->mm
->mm_users
) < 2) {
356 __tlb_remove_table(table
);
360 if (*batch
== NULL
) {
361 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
362 if (*batch
== NULL
) {
363 tlb_remove_table_one(table
);
368 (*batch
)->tables
[(*batch
)->nr
++] = table
;
369 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
370 tlb_table_flush(tlb
);
373 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
376 * If a p?d_bad entry is found while walking page tables, report
377 * the error, before resetting entry to p?d_none. Usually (but
378 * very seldom) called out from the p?d_none_or_clear_bad macros.
381 void pgd_clear_bad(pgd_t
*pgd
)
387 void pud_clear_bad(pud_t
*pud
)
393 void pmd_clear_bad(pmd_t
*pmd
)
400 * Note: this doesn't free the actual pages themselves. That
401 * has been handled earlier when unmapping all the memory regions.
403 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
406 pgtable_t token
= pmd_pgtable(*pmd
);
408 pte_free_tlb(tlb
, token
, addr
);
412 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
413 unsigned long addr
, unsigned long end
,
414 unsigned long floor
, unsigned long ceiling
)
421 pmd
= pmd_offset(pud
, addr
);
423 next
= pmd_addr_end(addr
, end
);
424 if (pmd_none_or_clear_bad(pmd
))
426 free_pte_range(tlb
, pmd
, addr
);
427 } while (pmd
++, addr
= next
, addr
!= end
);
437 if (end
- 1 > ceiling
- 1)
440 pmd
= pmd_offset(pud
, start
);
442 pmd_free_tlb(tlb
, pmd
, start
);
445 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
446 unsigned long addr
, unsigned long end
,
447 unsigned long floor
, unsigned long ceiling
)
454 pud
= pud_offset(pgd
, addr
);
456 next
= pud_addr_end(addr
, end
);
457 if (pud_none_or_clear_bad(pud
))
459 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
460 } while (pud
++, addr
= next
, addr
!= end
);
466 ceiling
&= PGDIR_MASK
;
470 if (end
- 1 > ceiling
- 1)
473 pud
= pud_offset(pgd
, start
);
475 pud_free_tlb(tlb
, pud
, start
);
479 * This function frees user-level page tables of a process.
481 * Must be called with pagetable lock held.
483 void free_pgd_range(struct mmu_gather
*tlb
,
484 unsigned long addr
, unsigned long end
,
485 unsigned long floor
, unsigned long ceiling
)
491 * The next few lines have given us lots of grief...
493 * Why are we testing PMD* at this top level? Because often
494 * there will be no work to do at all, and we'd prefer not to
495 * go all the way down to the bottom just to discover that.
497 * Why all these "- 1"s? Because 0 represents both the bottom
498 * of the address space and the top of it (using -1 for the
499 * top wouldn't help much: the masks would do the wrong thing).
500 * The rule is that addr 0 and floor 0 refer to the bottom of
501 * the address space, but end 0 and ceiling 0 refer to the top
502 * Comparisons need to use "end - 1" and "ceiling - 1" (though
503 * that end 0 case should be mythical).
505 * Wherever addr is brought up or ceiling brought down, we must
506 * be careful to reject "the opposite 0" before it confuses the
507 * subsequent tests. But what about where end is brought down
508 * by PMD_SIZE below? no, end can't go down to 0 there.
510 * Whereas we round start (addr) and ceiling down, by different
511 * masks at different levels, in order to test whether a table
512 * now has no other vmas using it, so can be freed, we don't
513 * bother to round floor or end up - the tests don't need that.
527 if (end
- 1 > ceiling
- 1)
532 pgd
= pgd_offset(tlb
->mm
, addr
);
534 next
= pgd_addr_end(addr
, end
);
535 if (pgd_none_or_clear_bad(pgd
))
537 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
538 } while (pgd
++, addr
= next
, addr
!= end
);
541 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
542 unsigned long floor
, unsigned long ceiling
)
545 struct vm_area_struct
*next
= vma
->vm_next
;
546 unsigned long addr
= vma
->vm_start
;
549 * Hide vma from rmap and truncate_pagecache before freeing
552 unlink_anon_vmas(vma
);
553 unlink_file_vma(vma
);
555 if (is_vm_hugetlb_page(vma
)) {
556 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
557 floor
, next
? next
->vm_start
: ceiling
);
560 * Optimization: gather nearby vmas into one call down
562 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
563 && !is_vm_hugetlb_page(next
)) {
566 unlink_anon_vmas(vma
);
567 unlink_file_vma(vma
);
569 free_pgd_range(tlb
, addr
, vma
->vm_end
,
570 floor
, next
? next
->vm_start
: ceiling
);
576 int __pte_alloc(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
577 pmd_t
*pmd
, unsigned long address
)
579 pgtable_t
new = pte_alloc_one(mm
, address
);
580 int wait_split_huge_page
;
585 * Ensure all pte setup (eg. pte page lock and page clearing) are
586 * visible before the pte is made visible to other CPUs by being
587 * put into page tables.
589 * The other side of the story is the pointer chasing in the page
590 * table walking code (when walking the page table without locking;
591 * ie. most of the time). Fortunately, these data accesses consist
592 * of a chain of data-dependent loads, meaning most CPUs (alpha
593 * being the notable exception) will already guarantee loads are
594 * seen in-order. See the alpha page table accessors for the
595 * smp_read_barrier_depends() barriers in page table walking code.
597 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
599 spin_lock(&mm
->page_table_lock
);
600 wait_split_huge_page
= 0;
601 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
603 pmd_populate(mm
, pmd
, new);
605 } else if (unlikely(pmd_trans_splitting(*pmd
)))
606 wait_split_huge_page
= 1;
607 spin_unlock(&mm
->page_table_lock
);
610 if (wait_split_huge_page
)
611 wait_split_huge_page(vma
->anon_vma
, pmd
);
615 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
617 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
621 smp_wmb(); /* See comment in __pte_alloc */
623 spin_lock(&init_mm
.page_table_lock
);
624 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
625 pmd_populate_kernel(&init_mm
, pmd
, new);
628 VM_BUG_ON(pmd_trans_splitting(*pmd
));
629 spin_unlock(&init_mm
.page_table_lock
);
631 pte_free_kernel(&init_mm
, new);
635 static inline void init_rss_vec(int *rss
)
637 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
640 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
644 if (current
->mm
== mm
)
646 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
648 add_mm_counter(mm
, i
, rss
[i
]);
652 * This function is called to print an error when a bad pte
653 * is found. For example, we might have a PFN-mapped pte in
654 * a region that doesn't allow it.
656 * The calling function must still handle the error.
658 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
659 pte_t pte
, struct page
*page
)
661 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
662 pud_t
*pud
= pud_offset(pgd
, addr
);
663 pmd_t
*pmd
= pmd_offset(pud
, addr
);
664 struct address_space
*mapping
;
666 static unsigned long resume
;
667 static unsigned long nr_shown
;
668 static unsigned long nr_unshown
;
671 * Allow a burst of 60 reports, then keep quiet for that minute;
672 * or allow a steady drip of one report per second.
674 if (nr_shown
== 60) {
675 if (time_before(jiffies
, resume
)) {
681 "BUG: Bad page map: %lu messages suppressed\n",
688 resume
= jiffies
+ 60 * HZ
;
690 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
691 index
= linear_page_index(vma
, addr
);
694 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
696 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
700 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
701 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
703 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
706 print_symbol(KERN_ALERT
"vma->vm_ops->fault: %s\n",
707 (unsigned long)vma
->vm_ops
->fault
);
708 if (vma
->vm_file
&& vma
->vm_file
->f_op
)
709 print_symbol(KERN_ALERT
"vma->vm_file->f_op->mmap: %s\n",
710 (unsigned long)vma
->vm_file
->f_op
->mmap
);
712 add_taint(TAINT_BAD_PAGE
);
715 static inline int is_cow_mapping(vm_flags_t flags
)
717 return (flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
721 static inline int is_zero_pfn(unsigned long pfn
)
723 return pfn
== zero_pfn
;
728 static inline unsigned long my_zero_pfn(unsigned long addr
)
735 * vm_normal_page -- This function gets the "struct page" associated with a pte.
737 * "Special" mappings do not wish to be associated with a "struct page" (either
738 * it doesn't exist, or it exists but they don't want to touch it). In this
739 * case, NULL is returned here. "Normal" mappings do have a struct page.
741 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
742 * pte bit, in which case this function is trivial. Secondly, an architecture
743 * may not have a spare pte bit, which requires a more complicated scheme,
746 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
747 * special mapping (even if there are underlying and valid "struct pages").
748 * COWed pages of a VM_PFNMAP are always normal.
750 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
751 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
752 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
753 * mapping will always honor the rule
755 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
757 * And for normal mappings this is false.
759 * This restricts such mappings to be a linear translation from virtual address
760 * to pfn. To get around this restriction, we allow arbitrary mappings so long
761 * as the vma is not a COW mapping; in that case, we know that all ptes are
762 * special (because none can have been COWed).
765 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
767 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
768 * page" backing, however the difference is that _all_ pages with a struct
769 * page (that is, those where pfn_valid is true) are refcounted and considered
770 * normal pages by the VM. The disadvantage is that pages are refcounted
771 * (which can be slower and simply not an option for some PFNMAP users). The
772 * advantage is that we don't have to follow the strict linearity rule of
773 * PFNMAP mappings in order to support COWable mappings.
776 #ifdef __HAVE_ARCH_PTE_SPECIAL
777 # define HAVE_PTE_SPECIAL 1
779 # define HAVE_PTE_SPECIAL 0
781 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
784 unsigned long pfn
= pte_pfn(pte
);
786 if (HAVE_PTE_SPECIAL
) {
787 if (likely(!pte_special(pte
)))
789 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
791 if (!is_zero_pfn(pfn
))
792 print_bad_pte(vma
, addr
, pte
, NULL
);
796 /* !HAVE_PTE_SPECIAL case follows: */
798 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
799 if (vma
->vm_flags
& VM_MIXEDMAP
) {
805 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
806 if (pfn
== vma
->vm_pgoff
+ off
)
808 if (!is_cow_mapping(vma
->vm_flags
))
813 if (is_zero_pfn(pfn
))
816 if (unlikely(pfn
> highest_memmap_pfn
)) {
817 print_bad_pte(vma
, addr
, pte
, NULL
);
822 * NOTE! We still have PageReserved() pages in the page tables.
823 * eg. VDSO mappings can cause them to exist.
826 return pfn_to_page(pfn
);
830 * copy one vm_area from one task to the other. Assumes the page tables
831 * already present in the new task to be cleared in the whole range
832 * covered by this vma.
835 static inline unsigned long
836 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
837 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
838 unsigned long addr
, int *rss
)
840 unsigned long vm_flags
= vma
->vm_flags
;
841 pte_t pte
= *src_pte
;
844 /* pte contains position in swap or file, so copy. */
845 if (unlikely(!pte_present(pte
))) {
846 if (!pte_file(pte
)) {
847 swp_entry_t entry
= pte_to_swp_entry(pte
);
849 if (swap_duplicate(entry
) < 0)
852 /* make sure dst_mm is on swapoff's mmlist. */
853 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
854 spin_lock(&mmlist_lock
);
855 if (list_empty(&dst_mm
->mmlist
))
856 list_add(&dst_mm
->mmlist
,
858 spin_unlock(&mmlist_lock
);
860 if (likely(!non_swap_entry(entry
)))
862 else if (is_migration_entry(entry
)) {
863 page
= migration_entry_to_page(entry
);
870 if (is_write_migration_entry(entry
) &&
871 is_cow_mapping(vm_flags
)) {
873 * COW mappings require pages in both
874 * parent and child to be set to read.
876 make_migration_entry_read(&entry
);
877 pte
= swp_entry_to_pte(entry
);
878 set_pte_at(src_mm
, addr
, src_pte
, pte
);
886 * If it's a COW mapping, write protect it both
887 * in the parent and the child
889 if (is_cow_mapping(vm_flags
)) {
890 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
891 pte
= pte_wrprotect(pte
);
895 * If it's a shared mapping, mark it clean in
898 if (vm_flags
& VM_SHARED
)
899 pte
= pte_mkclean(pte
);
900 pte
= pte_mkold(pte
);
902 page
= vm_normal_page(vma
, addr
, pte
);
913 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
917 int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
918 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
919 unsigned long addr
, unsigned long end
)
921 pte_t
*orig_src_pte
, *orig_dst_pte
;
922 pte_t
*src_pte
, *dst_pte
;
923 spinlock_t
*src_ptl
, *dst_ptl
;
925 int rss
[NR_MM_COUNTERS
];
926 swp_entry_t entry
= (swp_entry_t
){0};
931 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
934 src_pte
= pte_offset_map(src_pmd
, addr
);
935 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
936 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
937 orig_src_pte
= src_pte
;
938 orig_dst_pte
= dst_pte
;
939 arch_enter_lazy_mmu_mode();
943 * We are holding two locks at this point - either of them
944 * could generate latencies in another task on another CPU.
946 if (progress
>= 32) {
948 if (need_resched() ||
949 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
952 if (pte_none(*src_pte
)) {
956 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
961 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
963 arch_leave_lazy_mmu_mode();
964 spin_unlock(src_ptl
);
965 pte_unmap(orig_src_pte
);
966 add_mm_rss_vec(dst_mm
, rss
);
967 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
971 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
980 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
981 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
982 unsigned long addr
, unsigned long end
)
984 pmd_t
*src_pmd
, *dst_pmd
;
987 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
990 src_pmd
= pmd_offset(src_pud
, addr
);
992 next
= pmd_addr_end(addr
, end
);
993 if (pmd_trans_huge(*src_pmd
)) {
995 VM_BUG_ON(next
-addr
!= HPAGE_PMD_SIZE
);
996 err
= copy_huge_pmd(dst_mm
, src_mm
,
997 dst_pmd
, src_pmd
, addr
, vma
);
1004 if (pmd_none_or_clear_bad(src_pmd
))
1006 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
1009 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1013 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1014 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
1015 unsigned long addr
, unsigned long end
)
1017 pud_t
*src_pud
, *dst_pud
;
1020 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
1023 src_pud
= pud_offset(src_pgd
, addr
);
1025 next
= pud_addr_end(addr
, end
);
1026 if (pud_none_or_clear_bad(src_pud
))
1028 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
1031 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1035 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1036 struct vm_area_struct
*vma
)
1038 pgd_t
*src_pgd
, *dst_pgd
;
1040 unsigned long addr
= vma
->vm_start
;
1041 unsigned long end
= vma
->vm_end
;
1045 * Don't copy ptes where a page fault will fill them correctly.
1046 * Fork becomes much lighter when there are big shared or private
1047 * readonly mappings. The tradeoff is that copy_page_range is more
1048 * efficient than faulting.
1050 if (!(vma
->vm_flags
& (VM_HUGETLB
|VM_NONLINEAR
|VM_PFNMAP
|VM_INSERTPAGE
))) {
1055 if (is_vm_hugetlb_page(vma
))
1056 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1058 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
1060 * We do not free on error cases below as remove_vma
1061 * gets called on error from higher level routine
1063 ret
= track_pfn_copy(vma
);
1069 * We need to invalidate the secondary MMU mappings only when
1070 * there could be a permission downgrade on the ptes of the
1071 * parent mm. And a permission downgrade will only happen if
1072 * is_cow_mapping() returns true.
1074 if (is_cow_mapping(vma
->vm_flags
))
1075 mmu_notifier_invalidate_range_start(src_mm
, addr
, end
);
1078 dst_pgd
= pgd_offset(dst_mm
, addr
);
1079 src_pgd
= pgd_offset(src_mm
, addr
);
1081 next
= pgd_addr_end(addr
, end
);
1082 if (pgd_none_or_clear_bad(src_pgd
))
1084 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1085 vma
, addr
, next
))) {
1089 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1091 if (is_cow_mapping(vma
->vm_flags
))
1092 mmu_notifier_invalidate_range_end(src_mm
,
1093 vma
->vm_start
, end
);
1097 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1098 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1099 unsigned long addr
, unsigned long end
,
1100 struct zap_details
*details
)
1102 struct mm_struct
*mm
= tlb
->mm
;
1103 int force_flush
= 0;
1104 int rss
[NR_MM_COUNTERS
];
1111 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1113 arch_enter_lazy_mmu_mode();
1116 if (pte_none(ptent
)) {
1120 if (pte_present(ptent
)) {
1123 page
= vm_normal_page(vma
, addr
, ptent
);
1124 if (unlikely(details
) && page
) {
1126 * unmap_shared_mapping_pages() wants to
1127 * invalidate cache without truncating:
1128 * unmap shared but keep private pages.
1130 if (details
->check_mapping
&&
1131 details
->check_mapping
!= page
->mapping
)
1134 * Each page->index must be checked when
1135 * invalidating or truncating nonlinear.
1137 if (details
->nonlinear_vma
&&
1138 (page
->index
< details
->first_index
||
1139 page
->index
> details
->last_index
))
1142 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1144 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1145 if (unlikely(!page
))
1147 if (unlikely(details
) && details
->nonlinear_vma
1148 && linear_page_index(details
->nonlinear_vma
,
1149 addr
) != page
->index
)
1150 set_pte_at(mm
, addr
, pte
,
1151 pgoff_to_pte(page
->index
));
1153 rss
[MM_ANONPAGES
]--;
1155 if (pte_dirty(ptent
))
1156 set_page_dirty(page
);
1157 if (pte_young(ptent
) &&
1158 likely(!VM_SequentialReadHint(vma
)))
1159 mark_page_accessed(page
);
1160 rss
[MM_FILEPAGES
]--;
1162 page_remove_rmap(page
);
1163 if (unlikely(page_mapcount(page
) < 0))
1164 print_bad_pte(vma
, addr
, ptent
, page
);
1165 force_flush
= !__tlb_remove_page(tlb
, page
);
1171 * If details->check_mapping, we leave swap entries;
1172 * if details->nonlinear_vma, we leave file entries.
1174 if (unlikely(details
))
1176 if (pte_file(ptent
)) {
1177 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
)))
1178 print_bad_pte(vma
, addr
, ptent
, NULL
);
1180 swp_entry_t entry
= pte_to_swp_entry(ptent
);
1182 if (!non_swap_entry(entry
))
1184 else if (is_migration_entry(entry
)) {
1187 page
= migration_entry_to_page(entry
);
1190 rss
[MM_ANONPAGES
]--;
1192 rss
[MM_FILEPAGES
]--;
1194 if (unlikely(!free_swap_and_cache(entry
)))
1195 print_bad_pte(vma
, addr
, ptent
, NULL
);
1197 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1198 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1200 add_mm_rss_vec(mm
, rss
);
1201 arch_leave_lazy_mmu_mode();
1202 pte_unmap_unlock(start_pte
, ptl
);
1205 * mmu_gather ran out of room to batch pages, we break out of
1206 * the PTE lock to avoid doing the potential expensive TLB invalidate
1207 * and page-free while holding it.
1212 #ifdef HAVE_GENERIC_MMU_GATHER
1224 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1225 struct vm_area_struct
*vma
, pud_t
*pud
,
1226 unsigned long addr
, unsigned long end
,
1227 struct zap_details
*details
)
1232 pmd
= pmd_offset(pud
, addr
);
1234 next
= pmd_addr_end(addr
, end
);
1235 if (pmd_trans_huge(*pmd
)) {
1236 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1237 #ifdef CONFIG_DEBUG_VM
1238 if (!rwsem_is_locked(&tlb
->mm
->mmap_sem
)) {
1239 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1240 __func__
, addr
, end
,
1246 split_huge_page_pmd(vma
->vm_mm
, pmd
);
1247 } else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1252 * Here there can be other concurrent MADV_DONTNEED or
1253 * trans huge page faults running, and if the pmd is
1254 * none or trans huge it can change under us. This is
1255 * because MADV_DONTNEED holds the mmap_sem in read
1258 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1260 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1263 } while (pmd
++, addr
= next
, addr
!= end
);
1268 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1269 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1270 unsigned long addr
, unsigned long end
,
1271 struct zap_details
*details
)
1276 pud
= pud_offset(pgd
, addr
);
1278 next
= pud_addr_end(addr
, end
);
1279 if (pud_none_or_clear_bad(pud
))
1281 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1282 } while (pud
++, addr
= next
, addr
!= end
);
1287 static void unmap_page_range(struct mmu_gather
*tlb
,
1288 struct vm_area_struct
*vma
,
1289 unsigned long addr
, unsigned long end
,
1290 struct zap_details
*details
)
1295 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
1298 BUG_ON(addr
>= end
);
1299 mem_cgroup_uncharge_start();
1300 tlb_start_vma(tlb
, vma
);
1301 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1303 next
= pgd_addr_end(addr
, end
);
1304 if (pgd_none_or_clear_bad(pgd
))
1306 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
1307 } while (pgd
++, addr
= next
, addr
!= end
);
1308 tlb_end_vma(tlb
, vma
);
1309 mem_cgroup_uncharge_end();
1313 static void unmap_single_vma(struct mmu_gather
*tlb
,
1314 struct vm_area_struct
*vma
, unsigned long start_addr
,
1315 unsigned long end_addr
,
1316 struct zap_details
*details
)
1318 unsigned long start
= max(vma
->vm_start
, start_addr
);
1321 if (start
>= vma
->vm_end
)
1323 end
= min(vma
->vm_end
, end_addr
);
1324 if (end
<= vma
->vm_start
)
1328 uprobe_munmap(vma
, start
, end
);
1330 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1331 untrack_pfn(vma
, 0, 0);
1334 if (unlikely(is_vm_hugetlb_page(vma
))) {
1336 * It is undesirable to test vma->vm_file as it
1337 * should be non-null for valid hugetlb area.
1338 * However, vm_file will be NULL in the error
1339 * cleanup path of do_mmap_pgoff. When
1340 * hugetlbfs ->mmap method fails,
1341 * do_mmap_pgoff() nullifies vma->vm_file
1342 * before calling this function to clean up.
1343 * Since no pte has actually been setup, it is
1344 * safe to do nothing in this case.
1347 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
1348 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1349 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
1352 unmap_page_range(tlb
, vma
, start
, end
, details
);
1357 * unmap_vmas - unmap a range of memory covered by a list of vma's
1358 * @tlb: address of the caller's struct mmu_gather
1359 * @vma: the starting vma
1360 * @start_addr: virtual address at which to start unmapping
1361 * @end_addr: virtual address at which to end unmapping
1363 * Unmap all pages in the vma list.
1365 * Only addresses between `start' and `end' will be unmapped.
1367 * The VMA list must be sorted in ascending virtual address order.
1369 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1370 * range after unmap_vmas() returns. So the only responsibility here is to
1371 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1372 * drops the lock and schedules.
1374 void unmap_vmas(struct mmu_gather
*tlb
,
1375 struct vm_area_struct
*vma
, unsigned long start_addr
,
1376 unsigned long end_addr
)
1378 struct mm_struct
*mm
= vma
->vm_mm
;
1380 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1381 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1382 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1383 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1387 * zap_page_range - remove user pages in a given range
1388 * @vma: vm_area_struct holding the applicable pages
1389 * @start: starting address of pages to zap
1390 * @size: number of bytes to zap
1391 * @details: details of nonlinear truncation or shared cache invalidation
1393 * Caller must protect the VMA list
1395 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1396 unsigned long size
, struct zap_details
*details
)
1398 struct mm_struct
*mm
= vma
->vm_mm
;
1399 struct mmu_gather tlb
;
1400 unsigned long end
= start
+ size
;
1403 tlb_gather_mmu(&tlb
, mm
, 0);
1404 update_hiwater_rss(mm
);
1405 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1406 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
)
1407 unmap_single_vma(&tlb
, vma
, start
, end
, details
);
1408 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1409 tlb_finish_mmu(&tlb
, start
, end
);
1413 * zap_page_range_single - remove user pages in a given range
1414 * @vma: vm_area_struct holding the applicable pages
1415 * @address: starting address of pages to zap
1416 * @size: number of bytes to zap
1417 * @details: details of nonlinear truncation or shared cache invalidation
1419 * The range must fit into one VMA.
1421 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1422 unsigned long size
, struct zap_details
*details
)
1424 struct mm_struct
*mm
= vma
->vm_mm
;
1425 struct mmu_gather tlb
;
1426 unsigned long end
= address
+ size
;
1429 tlb_gather_mmu(&tlb
, mm
, 0);
1430 update_hiwater_rss(mm
);
1431 mmu_notifier_invalidate_range_start(mm
, address
, end
);
1432 unmap_single_vma(&tlb
, vma
, address
, end
, details
);
1433 mmu_notifier_invalidate_range_end(mm
, address
, end
);
1434 tlb_finish_mmu(&tlb
, address
, end
);
1438 * zap_vma_ptes - remove ptes mapping the vma
1439 * @vma: vm_area_struct holding ptes to be zapped
1440 * @address: starting address of pages to zap
1441 * @size: number of bytes to zap
1443 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1445 * The entire address range must be fully contained within the vma.
1447 * Returns 0 if successful.
1449 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1452 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1453 !(vma
->vm_flags
& VM_PFNMAP
))
1455 zap_page_range_single(vma
, address
, size
, NULL
);
1458 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1461 * follow_page - look up a page descriptor from a user-virtual address
1462 * @vma: vm_area_struct mapping @address
1463 * @address: virtual address to look up
1464 * @flags: flags modifying lookup behaviour
1466 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1468 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1469 * an error pointer if there is a mapping to something not represented
1470 * by a page descriptor (see also vm_normal_page()).
1472 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
1481 struct mm_struct
*mm
= vma
->vm_mm
;
1483 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
1484 if (!IS_ERR(page
)) {
1485 BUG_ON(flags
& FOLL_GET
);
1490 pgd
= pgd_offset(mm
, address
);
1491 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
1494 pud
= pud_offset(pgd
, address
);
1497 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
1498 BUG_ON(flags
& FOLL_GET
);
1499 page
= follow_huge_pud(mm
, address
, pud
, flags
& FOLL_WRITE
);
1502 if (unlikely(pud_bad(*pud
)))
1505 pmd
= pmd_offset(pud
, address
);
1508 if (pmd_huge(*pmd
) && vma
->vm_flags
& VM_HUGETLB
) {
1509 BUG_ON(flags
& FOLL_GET
);
1510 page
= follow_huge_pmd(mm
, address
, pmd
, flags
& FOLL_WRITE
);
1513 if (pmd_trans_huge(*pmd
)) {
1514 if (flags
& FOLL_SPLIT
) {
1515 split_huge_page_pmd(mm
, pmd
);
1516 goto split_fallthrough
;
1518 spin_lock(&mm
->page_table_lock
);
1519 if (likely(pmd_trans_huge(*pmd
))) {
1520 if (unlikely(pmd_trans_splitting(*pmd
))) {
1521 spin_unlock(&mm
->page_table_lock
);
1522 wait_split_huge_page(vma
->anon_vma
, pmd
);
1524 page
= follow_trans_huge_pmd(mm
, address
,
1526 spin_unlock(&mm
->page_table_lock
);
1530 spin_unlock(&mm
->page_table_lock
);
1534 if (unlikely(pmd_bad(*pmd
)))
1537 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1540 if (!pte_present(pte
))
1542 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
1545 page
= vm_normal_page(vma
, address
, pte
);
1546 if (unlikely(!page
)) {
1547 if ((flags
& FOLL_DUMP
) ||
1548 !is_zero_pfn(pte_pfn(pte
)))
1550 page
= pte_page(pte
);
1553 if (flags
& FOLL_GET
)
1554 get_page_foll(page
);
1555 if (flags
& FOLL_TOUCH
) {
1556 if ((flags
& FOLL_WRITE
) &&
1557 !pte_dirty(pte
) && !PageDirty(page
))
1558 set_page_dirty(page
);
1560 * pte_mkyoung() would be more correct here, but atomic care
1561 * is needed to avoid losing the dirty bit: it is easier to use
1562 * mark_page_accessed().
1564 mark_page_accessed(page
);
1566 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1568 * The preliminary mapping check is mainly to avoid the
1569 * pointless overhead of lock_page on the ZERO_PAGE
1570 * which might bounce very badly if there is contention.
1572 * If the page is already locked, we don't need to
1573 * handle it now - vmscan will handle it later if and
1574 * when it attempts to reclaim the page.
1576 if (page
->mapping
&& trylock_page(page
)) {
1577 lru_add_drain(); /* push cached pages to LRU */
1579 * Because we lock page here and migration is
1580 * blocked by the pte's page reference, we need
1581 * only check for file-cache page truncation.
1584 mlock_vma_page(page
);
1589 pte_unmap_unlock(ptep
, ptl
);
1594 pte_unmap_unlock(ptep
, ptl
);
1595 return ERR_PTR(-EFAULT
);
1598 pte_unmap_unlock(ptep
, ptl
);
1604 * When core dumping an enormous anonymous area that nobody
1605 * has touched so far, we don't want to allocate unnecessary pages or
1606 * page tables. Return error instead of NULL to skip handle_mm_fault,
1607 * then get_dump_page() will return NULL to leave a hole in the dump.
1608 * But we can only make this optimization where a hole would surely
1609 * be zero-filled if handle_mm_fault() actually did handle it.
1611 if ((flags
& FOLL_DUMP
) &&
1612 (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
1613 return ERR_PTR(-EFAULT
);
1617 static inline int stack_guard_page(struct vm_area_struct
*vma
, unsigned long addr
)
1619 return stack_guard_page_start(vma
, addr
) ||
1620 stack_guard_page_end(vma
, addr
+PAGE_SIZE
);
1624 * __get_user_pages() - pin user pages in memory
1625 * @tsk: task_struct of target task
1626 * @mm: mm_struct of target mm
1627 * @start: starting user address
1628 * @nr_pages: number of pages from start to pin
1629 * @gup_flags: flags modifying pin behaviour
1630 * @pages: array that receives pointers to the pages pinned.
1631 * Should be at least nr_pages long. Or NULL, if caller
1632 * only intends to ensure the pages are faulted in.
1633 * @vmas: array of pointers to vmas corresponding to each page.
1634 * Or NULL if the caller does not require them.
1635 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1637 * Returns number of pages pinned. This may be fewer than the number
1638 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1639 * were pinned, returns -errno. Each page returned must be released
1640 * with a put_page() call when it is finished with. vmas will only
1641 * remain valid while mmap_sem is held.
1643 * Must be called with mmap_sem held for read or write.
1645 * __get_user_pages walks a process's page tables and takes a reference to
1646 * each struct page that each user address corresponds to at a given
1647 * instant. That is, it takes the page that would be accessed if a user
1648 * thread accesses the given user virtual address at that instant.
1650 * This does not guarantee that the page exists in the user mappings when
1651 * __get_user_pages returns, and there may even be a completely different
1652 * page there in some cases (eg. if mmapped pagecache has been invalidated
1653 * and subsequently re faulted). However it does guarantee that the page
1654 * won't be freed completely. And mostly callers simply care that the page
1655 * contains data that was valid *at some point in time*. Typically, an IO
1656 * or similar operation cannot guarantee anything stronger anyway because
1657 * locks can't be held over the syscall boundary.
1659 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1660 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1661 * appropriate) must be called after the page is finished with, and
1662 * before put_page is called.
1664 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1665 * or mmap_sem contention, and if waiting is needed to pin all pages,
1666 * *@nonblocking will be set to 0.
1668 * In most cases, get_user_pages or get_user_pages_fast should be used
1669 * instead of __get_user_pages. __get_user_pages should be used only if
1670 * you need some special @gup_flags.
1672 int __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1673 unsigned long start
, int nr_pages
, unsigned int gup_flags
,
1674 struct page
**pages
, struct vm_area_struct
**vmas
,
1678 unsigned long vm_flags
;
1683 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
1686 * Require read or write permissions.
1687 * If FOLL_FORCE is set, we only require the "MAY" flags.
1689 vm_flags
= (gup_flags
& FOLL_WRITE
) ?
1690 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1691 vm_flags
&= (gup_flags
& FOLL_FORCE
) ?
1692 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1696 struct vm_area_struct
*vma
;
1698 vma
= find_extend_vma(mm
, start
);
1699 if (!vma
&& in_gate_area(mm
, start
)) {
1700 unsigned long pg
= start
& PAGE_MASK
;
1706 /* user gate pages are read-only */
1707 if (gup_flags
& FOLL_WRITE
)
1708 return i
? : -EFAULT
;
1710 pgd
= pgd_offset_k(pg
);
1712 pgd
= pgd_offset_gate(mm
, pg
);
1713 BUG_ON(pgd_none(*pgd
));
1714 pud
= pud_offset(pgd
, pg
);
1715 BUG_ON(pud_none(*pud
));
1716 pmd
= pmd_offset(pud
, pg
);
1718 return i
? : -EFAULT
;
1719 VM_BUG_ON(pmd_trans_huge(*pmd
));
1720 pte
= pte_offset_map(pmd
, pg
);
1721 if (pte_none(*pte
)) {
1723 return i
? : -EFAULT
;
1725 vma
= get_gate_vma(mm
);
1729 page
= vm_normal_page(vma
, start
, *pte
);
1731 if (!(gup_flags
& FOLL_DUMP
) &&
1732 is_zero_pfn(pte_pfn(*pte
)))
1733 page
= pte_page(*pte
);
1736 return i
? : -EFAULT
;
1747 (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1748 !(vm_flags
& vma
->vm_flags
))
1749 return i
? : -EFAULT
;
1751 if (is_vm_hugetlb_page(vma
)) {
1752 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1753 &start
, &nr_pages
, i
, gup_flags
);
1759 unsigned int foll_flags
= gup_flags
;
1762 * If we have a pending SIGKILL, don't keep faulting
1763 * pages and potentially allocating memory.
1765 if (unlikely(fatal_signal_pending(current
)))
1766 return i
? i
: -ERESTARTSYS
;
1769 while (!(page
= follow_page(vma
, start
, foll_flags
))) {
1771 unsigned int fault_flags
= 0;
1773 /* For mlock, just skip the stack guard page. */
1774 if (foll_flags
& FOLL_MLOCK
) {
1775 if (stack_guard_page(vma
, start
))
1778 if (foll_flags
& FOLL_WRITE
)
1779 fault_flags
|= FAULT_FLAG_WRITE
;
1781 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
1782 if (foll_flags
& FOLL_NOWAIT
)
1783 fault_flags
|= (FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
);
1785 ret
= handle_mm_fault(mm
, vma
, start
,
1788 if (ret
& VM_FAULT_ERROR
) {
1789 if (ret
& VM_FAULT_OOM
)
1790 return i
? i
: -ENOMEM
;
1791 if (ret
& (VM_FAULT_HWPOISON
|
1792 VM_FAULT_HWPOISON_LARGE
)) {
1795 else if (gup_flags
& FOLL_HWPOISON
)
1800 if (ret
& VM_FAULT_SIGBUS
)
1801 return i
? i
: -EFAULT
;
1806 if (ret
& VM_FAULT_MAJOR
)
1812 if (ret
& VM_FAULT_RETRY
) {
1819 * The VM_FAULT_WRITE bit tells us that
1820 * do_wp_page has broken COW when necessary,
1821 * even if maybe_mkwrite decided not to set
1822 * pte_write. We can thus safely do subsequent
1823 * page lookups as if they were reads. But only
1824 * do so when looping for pte_write is futile:
1825 * in some cases userspace may also be wanting
1826 * to write to the gotten user page, which a
1827 * read fault here might prevent (a readonly
1828 * page might get reCOWed by userspace write).
1830 if ((ret
& VM_FAULT_WRITE
) &&
1831 !(vma
->vm_flags
& VM_WRITE
))
1832 foll_flags
&= ~FOLL_WRITE
;
1837 return i
? i
: PTR_ERR(page
);
1841 flush_anon_page(vma
, page
, start
);
1842 flush_dcache_page(page
);
1850 } while (nr_pages
&& start
< vma
->vm_end
);
1854 EXPORT_SYMBOL(__get_user_pages
);
1857 * fixup_user_fault() - manually resolve a user page fault
1858 * @tsk: the task_struct to use for page fault accounting, or
1859 * NULL if faults are not to be recorded.
1860 * @mm: mm_struct of target mm
1861 * @address: user address
1862 * @fault_flags:flags to pass down to handle_mm_fault()
1864 * This is meant to be called in the specific scenario where for locking reasons
1865 * we try to access user memory in atomic context (within a pagefault_disable()
1866 * section), this returns -EFAULT, and we want to resolve the user fault before
1869 * Typically this is meant to be used by the futex code.
1871 * The main difference with get_user_pages() is that this function will
1872 * unconditionally call handle_mm_fault() which will in turn perform all the
1873 * necessary SW fixup of the dirty and young bits in the PTE, while
1874 * handle_mm_fault() only guarantees to update these in the struct page.
1876 * This is important for some architectures where those bits also gate the
1877 * access permission to the page because they are maintained in software. On
1878 * such architectures, gup() will not be enough to make a subsequent access
1881 * This should be called with the mm_sem held for read.
1883 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
1884 unsigned long address
, unsigned int fault_flags
)
1886 struct vm_area_struct
*vma
;
1889 vma
= find_extend_vma(mm
, address
);
1890 if (!vma
|| address
< vma
->vm_start
)
1893 ret
= handle_mm_fault(mm
, vma
, address
, fault_flags
);
1894 if (ret
& VM_FAULT_ERROR
) {
1895 if (ret
& VM_FAULT_OOM
)
1897 if (ret
& (VM_FAULT_HWPOISON
| VM_FAULT_HWPOISON_LARGE
))
1899 if (ret
& VM_FAULT_SIGBUS
)
1904 if (ret
& VM_FAULT_MAJOR
)
1913 * get_user_pages() - pin user pages in memory
1914 * @tsk: the task_struct to use for page fault accounting, or
1915 * NULL if faults are not to be recorded.
1916 * @mm: mm_struct of target mm
1917 * @start: starting user address
1918 * @nr_pages: number of pages from start to pin
1919 * @write: whether pages will be written to by the caller
1920 * @force: whether to force write access even if user mapping is
1921 * readonly. This will result in the page being COWed even
1922 * in MAP_SHARED mappings. You do not want this.
1923 * @pages: array that receives pointers to the pages pinned.
1924 * Should be at least nr_pages long. Or NULL, if caller
1925 * only intends to ensure the pages are faulted in.
1926 * @vmas: array of pointers to vmas corresponding to each page.
1927 * Or NULL if the caller does not require them.
1929 * Returns number of pages pinned. This may be fewer than the number
1930 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1931 * were pinned, returns -errno. Each page returned must be released
1932 * with a put_page() call when it is finished with. vmas will only
1933 * remain valid while mmap_sem is held.
1935 * Must be called with mmap_sem held for read or write.
1937 * get_user_pages walks a process's page tables and takes a reference to
1938 * each struct page that each user address corresponds to at a given
1939 * instant. That is, it takes the page that would be accessed if a user
1940 * thread accesses the given user virtual address at that instant.
1942 * This does not guarantee that the page exists in the user mappings when
1943 * get_user_pages returns, and there may even be a completely different
1944 * page there in some cases (eg. if mmapped pagecache has been invalidated
1945 * and subsequently re faulted). However it does guarantee that the page
1946 * won't be freed completely. And mostly callers simply care that the page
1947 * contains data that was valid *at some point in time*. Typically, an IO
1948 * or similar operation cannot guarantee anything stronger anyway because
1949 * locks can't be held over the syscall boundary.
1951 * If write=0, the page must not be written to. If the page is written to,
1952 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1953 * after the page is finished with, and before put_page is called.
1955 * get_user_pages is typically used for fewer-copy IO operations, to get a
1956 * handle on the memory by some means other than accesses via the user virtual
1957 * addresses. The pages may be submitted for DMA to devices or accessed via
1958 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1959 * use the correct cache flushing APIs.
1961 * See also get_user_pages_fast, for performance critical applications.
1963 int get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1964 unsigned long start
, int nr_pages
, int write
, int force
,
1965 struct page
**pages
, struct vm_area_struct
**vmas
)
1967 int flags
= FOLL_TOUCH
;
1972 flags
|= FOLL_WRITE
;
1974 flags
|= FOLL_FORCE
;
1976 return __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
, vmas
,
1979 EXPORT_SYMBOL(get_user_pages
);
1982 * get_dump_page() - pin user page in memory while writing it to core dump
1983 * @addr: user address
1985 * Returns struct page pointer of user page pinned for dump,
1986 * to be freed afterwards by page_cache_release() or put_page().
1988 * Returns NULL on any kind of failure - a hole must then be inserted into
1989 * the corefile, to preserve alignment with its headers; and also returns
1990 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1991 * allowing a hole to be left in the corefile to save diskspace.
1993 * Called without mmap_sem, but after all other threads have been killed.
1995 #ifdef CONFIG_ELF_CORE
1996 struct page
*get_dump_page(unsigned long addr
)
1998 struct vm_area_struct
*vma
;
2001 if (__get_user_pages(current
, current
->mm
, addr
, 1,
2002 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
2005 flush_cache_page(vma
, addr
, page_to_pfn(page
));
2008 #endif /* CONFIG_ELF_CORE */
2010 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
2013 pgd_t
* pgd
= pgd_offset(mm
, addr
);
2014 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
2016 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
2018 VM_BUG_ON(pmd_trans_huge(*pmd
));
2019 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
2026 * This is the old fallback for page remapping.
2028 * For historical reasons, it only allows reserved pages. Only
2029 * old drivers should use this, and they needed to mark their
2030 * pages reserved for the old functions anyway.
2032 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2033 struct page
*page
, pgprot_t prot
)
2035 struct mm_struct
*mm
= vma
->vm_mm
;
2044 flush_dcache_page(page
);
2045 pte
= get_locked_pte(mm
, addr
, &ptl
);
2049 if (!pte_none(*pte
))
2052 /* Ok, finally just insert the thing.. */
2054 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
2055 page_add_file_rmap(page
);
2056 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
2059 pte_unmap_unlock(pte
, ptl
);
2062 pte_unmap_unlock(pte
, ptl
);
2068 * vm_insert_page - insert single page into user vma
2069 * @vma: user vma to map to
2070 * @addr: target user address of this page
2071 * @page: source kernel page
2073 * This allows drivers to insert individual pages they've allocated
2076 * The page has to be a nice clean _individual_ kernel allocation.
2077 * If you allocate a compound page, you need to have marked it as
2078 * such (__GFP_COMP), or manually just split the page up yourself
2079 * (see split_page()).
2081 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2082 * took an arbitrary page protection parameter. This doesn't allow
2083 * that. Your vma protection will have to be set up correctly, which
2084 * means that if you want a shared writable mapping, you'd better
2085 * ask for a shared writable mapping!
2087 * The page does not need to be reserved.
2089 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2092 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2094 if (!page_count(page
))
2096 vma
->vm_flags
|= VM_INSERTPAGE
;
2097 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2099 EXPORT_SYMBOL(vm_insert_page
);
2101 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2102 unsigned long pfn
, pgprot_t prot
)
2104 struct mm_struct
*mm
= vma
->vm_mm
;
2110 pte
= get_locked_pte(mm
, addr
, &ptl
);
2114 if (!pte_none(*pte
))
2117 /* Ok, finally just insert the thing.. */
2118 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
2119 set_pte_at(mm
, addr
, pte
, entry
);
2120 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
2124 pte_unmap_unlock(pte
, ptl
);
2130 * vm_insert_pfn - insert single pfn into user vma
2131 * @vma: user vma to map to
2132 * @addr: target user address of this page
2133 * @pfn: source kernel pfn
2135 * Similar to vm_inert_page, this allows drivers to insert individual pages
2136 * they've allocated into a user vma. Same comments apply.
2138 * This function should only be called from a vm_ops->fault handler, and
2139 * in that case the handler should return NULL.
2141 * vma cannot be a COW mapping.
2143 * As this is called only for pages that do not currently exist, we
2144 * do not need to flush old virtual caches or the TLB.
2146 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2150 pgprot_t pgprot
= vma
->vm_page_prot
;
2152 * Technically, architectures with pte_special can avoid all these
2153 * restrictions (same for remap_pfn_range). However we would like
2154 * consistency in testing and feature parity among all, so we should
2155 * try to keep these invariants in place for everybody.
2157 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
2158 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
2159 (VM_PFNMAP
|VM_MIXEDMAP
));
2160 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
2161 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
2163 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2165 if (track_pfn_insert(vma
, &pgprot
, pfn
))
2168 ret
= insert_pfn(vma
, addr
, pfn
, pgprot
);
2172 EXPORT_SYMBOL(vm_insert_pfn
);
2174 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2177 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
2179 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2183 * If we don't have pte special, then we have to use the pfn_valid()
2184 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2185 * refcount the page if pfn_valid is true (hence insert_page rather
2186 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2187 * without pte special, it would there be refcounted as a normal page.
2189 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
2192 page
= pfn_to_page(pfn
);
2193 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2195 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
2197 EXPORT_SYMBOL(vm_insert_mixed
);
2200 * maps a range of physical memory into the requested pages. the old
2201 * mappings are removed. any references to nonexistent pages results
2202 * in null mappings (currently treated as "copy-on-access")
2204 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2205 unsigned long addr
, unsigned long end
,
2206 unsigned long pfn
, pgprot_t prot
)
2211 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2214 arch_enter_lazy_mmu_mode();
2216 BUG_ON(!pte_none(*pte
));
2217 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2219 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2220 arch_leave_lazy_mmu_mode();
2221 pte_unmap_unlock(pte
- 1, ptl
);
2225 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2226 unsigned long addr
, unsigned long end
,
2227 unsigned long pfn
, pgprot_t prot
)
2232 pfn
-= addr
>> PAGE_SHIFT
;
2233 pmd
= pmd_alloc(mm
, pud
, addr
);
2236 VM_BUG_ON(pmd_trans_huge(*pmd
));
2238 next
= pmd_addr_end(addr
, end
);
2239 if (remap_pte_range(mm
, pmd
, addr
, next
,
2240 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
2242 } while (pmd
++, addr
= next
, addr
!= end
);
2246 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2247 unsigned long addr
, unsigned long end
,
2248 unsigned long pfn
, pgprot_t prot
)
2253 pfn
-= addr
>> PAGE_SHIFT
;
2254 pud
= pud_alloc(mm
, pgd
, addr
);
2258 next
= pud_addr_end(addr
, end
);
2259 if (remap_pmd_range(mm
, pud
, addr
, next
,
2260 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
2262 } while (pud
++, addr
= next
, addr
!= end
);
2267 * remap_pfn_range - remap kernel memory to userspace
2268 * @vma: user vma to map to
2269 * @addr: target user address to start at
2270 * @pfn: physical address of kernel memory
2271 * @size: size of map area
2272 * @prot: page protection flags for this mapping
2274 * Note: this is only safe if the mm semaphore is held when called.
2276 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2277 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2281 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2282 struct mm_struct
*mm
= vma
->vm_mm
;
2286 * Physically remapped pages are special. Tell the
2287 * rest of the world about it:
2288 * VM_IO tells people not to look at these pages
2289 * (accesses can have side effects).
2290 * VM_RESERVED is specified all over the place, because
2291 * in 2.4 it kept swapout's vma scan off this vma; but
2292 * in 2.6 the LRU scan won't even find its pages, so this
2293 * flag means no more than count its pages in reserved_vm,
2294 * and omit it from core dump, even when VM_IO turned off.
2295 * VM_PFNMAP tells the core MM that the base pages are just
2296 * raw PFN mappings, and do not have a "struct page" associated
2299 * There's a horrible special case to handle copy-on-write
2300 * behaviour that some programs depend on. We mark the "original"
2301 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2302 * See vm_normal_page() for details.
2304 if (is_cow_mapping(vma
->vm_flags
)) {
2305 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2307 vma
->vm_pgoff
= pfn
;
2310 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
2314 vma
->vm_flags
|= VM_IO
| VM_RESERVED
| VM_PFNMAP
;
2316 BUG_ON(addr
>= end
);
2317 pfn
-= addr
>> PAGE_SHIFT
;
2318 pgd
= pgd_offset(mm
, addr
);
2319 flush_cache_range(vma
, addr
, end
);
2321 next
= pgd_addr_end(addr
, end
);
2322 err
= remap_pud_range(mm
, pgd
, addr
, next
,
2323 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2326 } while (pgd
++, addr
= next
, addr
!= end
);
2329 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
));
2333 EXPORT_SYMBOL(remap_pfn_range
);
2335 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2336 unsigned long addr
, unsigned long end
,
2337 pte_fn_t fn
, void *data
)
2342 spinlock_t
*uninitialized_var(ptl
);
2344 pte
= (mm
== &init_mm
) ?
2345 pte_alloc_kernel(pmd
, addr
) :
2346 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2350 BUG_ON(pmd_huge(*pmd
));
2352 arch_enter_lazy_mmu_mode();
2354 token
= pmd_pgtable(*pmd
);
2357 err
= fn(pte
++, token
, addr
, data
);
2360 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2362 arch_leave_lazy_mmu_mode();
2365 pte_unmap_unlock(pte
-1, ptl
);
2369 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2370 unsigned long addr
, unsigned long end
,
2371 pte_fn_t fn
, void *data
)
2377 BUG_ON(pud_huge(*pud
));
2379 pmd
= pmd_alloc(mm
, pud
, addr
);
2383 next
= pmd_addr_end(addr
, end
);
2384 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
2387 } while (pmd
++, addr
= next
, addr
!= end
);
2391 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2392 unsigned long addr
, unsigned long end
,
2393 pte_fn_t fn
, void *data
)
2399 pud
= pud_alloc(mm
, pgd
, addr
);
2403 next
= pud_addr_end(addr
, end
);
2404 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
2407 } while (pud
++, addr
= next
, addr
!= end
);
2412 * Scan a region of virtual memory, filling in page tables as necessary
2413 * and calling a provided function on each leaf page table.
2415 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2416 unsigned long size
, pte_fn_t fn
, void *data
)
2420 unsigned long end
= addr
+ size
;
2423 BUG_ON(addr
>= end
);
2424 pgd
= pgd_offset(mm
, addr
);
2426 next
= pgd_addr_end(addr
, end
);
2427 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
2430 } while (pgd
++, addr
= next
, addr
!= end
);
2434 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2437 * handle_pte_fault chooses page fault handler according to an entry
2438 * which was read non-atomically. Before making any commitment, on
2439 * those architectures or configurations (e.g. i386 with PAE) which
2440 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2441 * must check under lock before unmapping the pte and proceeding
2442 * (but do_wp_page is only called after already making such a check;
2443 * and do_anonymous_page can safely check later on).
2445 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2446 pte_t
*page_table
, pte_t orig_pte
)
2449 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2450 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2451 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2453 same
= pte_same(*page_table
, orig_pte
);
2457 pte_unmap(page_table
);
2461 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
2464 * If the source page was a PFN mapping, we don't have
2465 * a "struct page" for it. We do a best-effort copy by
2466 * just copying from the original user address. If that
2467 * fails, we just zero-fill it. Live with it.
2469 if (unlikely(!src
)) {
2470 void *kaddr
= kmap_atomic(dst
);
2471 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
2474 * This really shouldn't fail, because the page is there
2475 * in the page tables. But it might just be unreadable,
2476 * in which case we just give up and fill the result with
2479 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
2481 kunmap_atomic(kaddr
);
2482 flush_dcache_page(dst
);
2484 copy_user_highpage(dst
, src
, va
, vma
);
2488 * This routine handles present pages, when users try to write
2489 * to a shared page. It is done by copying the page to a new address
2490 * and decrementing the shared-page counter for the old page.
2492 * Note that this routine assumes that the protection checks have been
2493 * done by the caller (the low-level page fault routine in most cases).
2494 * Thus we can safely just mark it writable once we've done any necessary
2497 * We also mark the page dirty at this point even though the page will
2498 * change only once the write actually happens. This avoids a few races,
2499 * and potentially makes it more efficient.
2501 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2502 * but allow concurrent faults), with pte both mapped and locked.
2503 * We return with mmap_sem still held, but pte unmapped and unlocked.
2505 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2506 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2507 spinlock_t
*ptl
, pte_t orig_pte
)
2510 struct page
*old_page
, *new_page
;
2513 int page_mkwrite
= 0;
2514 struct page
*dirty_page
= NULL
;
2516 old_page
= vm_normal_page(vma
, address
, orig_pte
);
2519 * VM_MIXEDMAP !pfn_valid() case
2521 * We should not cow pages in a shared writeable mapping.
2522 * Just mark the pages writable as we can't do any dirty
2523 * accounting on raw pfn maps.
2525 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2526 (VM_WRITE
|VM_SHARED
))
2532 * Take out anonymous pages first, anonymous shared vmas are
2533 * not dirty accountable.
2535 if (PageAnon(old_page
) && !PageKsm(old_page
)) {
2536 if (!trylock_page(old_page
)) {
2537 page_cache_get(old_page
);
2538 pte_unmap_unlock(page_table
, ptl
);
2539 lock_page(old_page
);
2540 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2542 if (!pte_same(*page_table
, orig_pte
)) {
2543 unlock_page(old_page
);
2546 page_cache_release(old_page
);
2548 if (reuse_swap_page(old_page
)) {
2550 * The page is all ours. Move it to our anon_vma so
2551 * the rmap code will not search our parent or siblings.
2552 * Protected against the rmap code by the page lock.
2554 page_move_anon_rmap(old_page
, vma
, address
);
2555 unlock_page(old_page
);
2558 unlock_page(old_page
);
2559 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2560 (VM_WRITE
|VM_SHARED
))) {
2562 * Only catch write-faults on shared writable pages,
2563 * read-only shared pages can get COWed by
2564 * get_user_pages(.write=1, .force=1).
2566 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2567 struct vm_fault vmf
;
2570 vmf
.virtual_address
= (void __user
*)(address
&
2572 vmf
.pgoff
= old_page
->index
;
2573 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2574 vmf
.page
= old_page
;
2577 * Notify the address space that the page is about to
2578 * become writable so that it can prohibit this or wait
2579 * for the page to get into an appropriate state.
2581 * We do this without the lock held, so that it can
2582 * sleep if it needs to.
2584 page_cache_get(old_page
);
2585 pte_unmap_unlock(page_table
, ptl
);
2587 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
2589 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
2591 goto unwritable_page
;
2593 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
2594 lock_page(old_page
);
2595 if (!old_page
->mapping
) {
2596 ret
= 0; /* retry the fault */
2597 unlock_page(old_page
);
2598 goto unwritable_page
;
2601 VM_BUG_ON(!PageLocked(old_page
));
2604 * Since we dropped the lock we need to revalidate
2605 * the PTE as someone else may have changed it. If
2606 * they did, we just return, as we can count on the
2607 * MMU to tell us if they didn't also make it writable.
2609 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2611 if (!pte_same(*page_table
, orig_pte
)) {
2612 unlock_page(old_page
);
2618 dirty_page
= old_page
;
2619 get_page(dirty_page
);
2622 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2623 entry
= pte_mkyoung(orig_pte
);
2624 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2625 if (ptep_set_access_flags(vma
, address
, page_table
, entry
,1))
2626 update_mmu_cache(vma
, address
, page_table
);
2627 pte_unmap_unlock(page_table
, ptl
);
2628 ret
|= VM_FAULT_WRITE
;
2634 * Yes, Virginia, this is actually required to prevent a race
2635 * with clear_page_dirty_for_io() from clearing the page dirty
2636 * bit after it clear all dirty ptes, but before a racing
2637 * do_wp_page installs a dirty pte.
2639 * __do_fault is protected similarly.
2641 if (!page_mkwrite
) {
2642 wait_on_page_locked(dirty_page
);
2643 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2644 /* file_update_time outside page_lock */
2646 file_update_time(vma
->vm_file
);
2648 put_page(dirty_page
);
2650 struct address_space
*mapping
= dirty_page
->mapping
;
2652 set_page_dirty(dirty_page
);
2653 unlock_page(dirty_page
);
2654 page_cache_release(dirty_page
);
2657 * Some device drivers do not set page.mapping
2658 * but still dirty their pages
2660 balance_dirty_pages_ratelimited(mapping
);
2668 * Ok, we need to copy. Oh, well..
2670 page_cache_get(old_page
);
2672 pte_unmap_unlock(page_table
, ptl
);
2674 if (unlikely(anon_vma_prepare(vma
)))
2677 if (is_zero_pfn(pte_pfn(orig_pte
))) {
2678 new_page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2682 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2685 cow_user_page(new_page
, old_page
, address
, vma
);
2687 __SetPageUptodate(new_page
);
2689 if (mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
))
2693 * Re-check the pte - we dropped the lock
2695 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2696 if (likely(pte_same(*page_table
, orig_pte
))) {
2698 if (!PageAnon(old_page
)) {
2699 dec_mm_counter_fast(mm
, MM_FILEPAGES
);
2700 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2703 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2704 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2705 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2706 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2708 * Clear the pte entry and flush it first, before updating the
2709 * pte with the new entry. This will avoid a race condition
2710 * seen in the presence of one thread doing SMC and another
2713 ptep_clear_flush(vma
, address
, page_table
);
2714 page_add_new_anon_rmap(new_page
, vma
, address
);
2716 * We call the notify macro here because, when using secondary
2717 * mmu page tables (such as kvm shadow page tables), we want the
2718 * new page to be mapped directly into the secondary page table.
2720 set_pte_at_notify(mm
, address
, page_table
, entry
);
2721 update_mmu_cache(vma
, address
, page_table
);
2724 * Only after switching the pte to the new page may
2725 * we remove the mapcount here. Otherwise another
2726 * process may come and find the rmap count decremented
2727 * before the pte is switched to the new page, and
2728 * "reuse" the old page writing into it while our pte
2729 * here still points into it and can be read by other
2732 * The critical issue is to order this
2733 * page_remove_rmap with the ptp_clear_flush above.
2734 * Those stores are ordered by (if nothing else,)
2735 * the barrier present in the atomic_add_negative
2736 * in page_remove_rmap.
2738 * Then the TLB flush in ptep_clear_flush ensures that
2739 * no process can access the old page before the
2740 * decremented mapcount is visible. And the old page
2741 * cannot be reused until after the decremented
2742 * mapcount is visible. So transitively, TLBs to
2743 * old page will be flushed before it can be reused.
2745 page_remove_rmap(old_page
);
2748 /* Free the old page.. */
2749 new_page
= old_page
;
2750 ret
|= VM_FAULT_WRITE
;
2752 mem_cgroup_uncharge_page(new_page
);
2755 page_cache_release(new_page
);
2757 pte_unmap_unlock(page_table
, ptl
);
2760 * Don't let another task, with possibly unlocked vma,
2761 * keep the mlocked page.
2763 if ((ret
& VM_FAULT_WRITE
) && (vma
->vm_flags
& VM_LOCKED
)) {
2764 lock_page(old_page
); /* LRU manipulation */
2765 munlock_vma_page(old_page
);
2766 unlock_page(old_page
);
2768 page_cache_release(old_page
);
2772 page_cache_release(new_page
);
2776 unlock_page(old_page
);
2777 page_cache_release(old_page
);
2779 page_cache_release(old_page
);
2781 return VM_FAULT_OOM
;
2784 page_cache_release(old_page
);
2788 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2789 unsigned long start_addr
, unsigned long end_addr
,
2790 struct zap_details
*details
)
2792 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2795 static inline void unmap_mapping_range_tree(struct prio_tree_root
*root
,
2796 struct zap_details
*details
)
2798 struct vm_area_struct
*vma
;
2799 struct prio_tree_iter iter
;
2800 pgoff_t vba
, vea
, zba
, zea
;
2802 vma_prio_tree_foreach(vma
, &iter
, root
,
2803 details
->first_index
, details
->last_index
) {
2805 vba
= vma
->vm_pgoff
;
2806 vea
= vba
+ ((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) - 1;
2807 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2808 zba
= details
->first_index
;
2811 zea
= details
->last_index
;
2815 unmap_mapping_range_vma(vma
,
2816 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2817 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2822 static inline void unmap_mapping_range_list(struct list_head
*head
,
2823 struct zap_details
*details
)
2825 struct vm_area_struct
*vma
;
2828 * In nonlinear VMAs there is no correspondence between virtual address
2829 * offset and file offset. So we must perform an exhaustive search
2830 * across *all* the pages in each nonlinear VMA, not just the pages
2831 * whose virtual address lies outside the file truncation point.
2833 list_for_each_entry(vma
, head
, shared
.vm_set
.list
) {
2834 details
->nonlinear_vma
= vma
;
2835 unmap_mapping_range_vma(vma
, vma
->vm_start
, vma
->vm_end
, details
);
2840 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2841 * @mapping: the address space containing mmaps to be unmapped.
2842 * @holebegin: byte in first page to unmap, relative to the start of
2843 * the underlying file. This will be rounded down to a PAGE_SIZE
2844 * boundary. Note that this is different from truncate_pagecache(), which
2845 * must keep the partial page. In contrast, we must get rid of
2847 * @holelen: size of prospective hole in bytes. This will be rounded
2848 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2850 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2851 * but 0 when invalidating pagecache, don't throw away private data.
2853 void unmap_mapping_range(struct address_space
*mapping
,
2854 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2856 struct zap_details details
;
2857 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2858 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2860 /* Check for overflow. */
2861 if (sizeof(holelen
) > sizeof(hlen
)) {
2863 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2864 if (holeend
& ~(long long)ULONG_MAX
)
2865 hlen
= ULONG_MAX
- hba
+ 1;
2868 details
.check_mapping
= even_cows
? NULL
: mapping
;
2869 details
.nonlinear_vma
= NULL
;
2870 details
.first_index
= hba
;
2871 details
.last_index
= hba
+ hlen
- 1;
2872 if (details
.last_index
< details
.first_index
)
2873 details
.last_index
= ULONG_MAX
;
2876 mutex_lock(&mapping
->i_mmap_mutex
);
2877 if (unlikely(!prio_tree_empty(&mapping
->i_mmap
)))
2878 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2879 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
2880 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
2881 mutex_unlock(&mapping
->i_mmap_mutex
);
2883 EXPORT_SYMBOL(unmap_mapping_range
);
2886 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2887 * but allow concurrent faults), and pte mapped but not yet locked.
2888 * We return with mmap_sem still held, but pte unmapped and unlocked.
2890 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2891 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2892 unsigned int flags
, pte_t orig_pte
)
2895 struct page
*page
, *swapcache
= NULL
;
2899 struct mem_cgroup
*ptr
;
2903 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2906 entry
= pte_to_swp_entry(orig_pte
);
2907 if (unlikely(non_swap_entry(entry
))) {
2908 if (is_migration_entry(entry
)) {
2909 migration_entry_wait(mm
, pmd
, address
);
2910 } else if (is_hwpoison_entry(entry
)) {
2911 ret
= VM_FAULT_HWPOISON
;
2913 print_bad_pte(vma
, address
, orig_pte
, NULL
);
2914 ret
= VM_FAULT_SIGBUS
;
2918 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2919 page
= lookup_swap_cache(entry
);
2921 page
= swapin_readahead(entry
,
2922 GFP_HIGHUSER_MOVABLE
, vma
, address
);
2925 * Back out if somebody else faulted in this pte
2926 * while we released the pte lock.
2928 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2929 if (likely(pte_same(*page_table
, orig_pte
)))
2931 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2935 /* Had to read the page from swap area: Major fault */
2936 ret
= VM_FAULT_MAJOR
;
2937 count_vm_event(PGMAJFAULT
);
2938 mem_cgroup_count_vm_event(mm
, PGMAJFAULT
);
2939 } else if (PageHWPoison(page
)) {
2941 * hwpoisoned dirty swapcache pages are kept for killing
2942 * owner processes (which may be unknown at hwpoison time)
2944 ret
= VM_FAULT_HWPOISON
;
2945 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2949 locked
= lock_page_or_retry(page
, mm
, flags
);
2951 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2953 ret
|= VM_FAULT_RETRY
;
2958 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2959 * release the swapcache from under us. The page pin, and pte_same
2960 * test below, are not enough to exclude that. Even if it is still
2961 * swapcache, we need to check that the page's swap has not changed.
2963 if (unlikely(!PageSwapCache(page
) || page_private(page
) != entry
.val
))
2966 if (ksm_might_need_to_copy(page
, vma
, address
)) {
2968 page
= ksm_does_need_to_copy(page
, vma
, address
);
2970 if (unlikely(!page
)) {
2978 if (mem_cgroup_try_charge_swapin(mm
, page
, GFP_KERNEL
, &ptr
)) {
2984 * Back out if somebody else already faulted in this pte.
2986 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2987 if (unlikely(!pte_same(*page_table
, orig_pte
)))
2990 if (unlikely(!PageUptodate(page
))) {
2991 ret
= VM_FAULT_SIGBUS
;
2996 * The page isn't present yet, go ahead with the fault.
2998 * Be careful about the sequence of operations here.
2999 * To get its accounting right, reuse_swap_page() must be called
3000 * while the page is counted on swap but not yet in mapcount i.e.
3001 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3002 * must be called after the swap_free(), or it will never succeed.
3003 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3004 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3005 * in page->private. In this case, a record in swap_cgroup is silently
3006 * discarded at swap_free().
3009 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3010 dec_mm_counter_fast(mm
, MM_SWAPENTS
);
3011 pte
= mk_pte(page
, vma
->vm_page_prot
);
3012 if ((flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
)) {
3013 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3014 flags
&= ~FAULT_FLAG_WRITE
;
3015 ret
|= VM_FAULT_WRITE
;
3018 flush_icache_page(vma
, page
);
3019 set_pte_at(mm
, address
, page_table
, pte
);
3020 do_page_add_anon_rmap(page
, vma
, address
, exclusive
);
3021 /* It's better to call commit-charge after rmap is established */
3022 mem_cgroup_commit_charge_swapin(page
, ptr
);
3025 if (vm_swap_full() || (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
3026 try_to_free_swap(page
);
3030 * Hold the lock to avoid the swap entry to be reused
3031 * until we take the PT lock for the pte_same() check
3032 * (to avoid false positives from pte_same). For
3033 * further safety release the lock after the swap_free
3034 * so that the swap count won't change under a
3035 * parallel locked swapcache.
3037 unlock_page(swapcache
);
3038 page_cache_release(swapcache
);
3041 if (flags
& FAULT_FLAG_WRITE
) {
3042 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
3043 if (ret
& VM_FAULT_ERROR
)
3044 ret
&= VM_FAULT_ERROR
;
3048 /* No need to invalidate - it was non-present before */
3049 update_mmu_cache(vma
, address
, page_table
);
3051 pte_unmap_unlock(page_table
, ptl
);
3055 mem_cgroup_cancel_charge_swapin(ptr
);
3056 pte_unmap_unlock(page_table
, ptl
);
3060 page_cache_release(page
);
3062 unlock_page(swapcache
);
3063 page_cache_release(swapcache
);
3069 * This is like a special single-page "expand_{down|up}wards()",
3070 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3071 * doesn't hit another vma.
3073 static inline int check_stack_guard_page(struct vm_area_struct
*vma
, unsigned long address
)
3075 address
&= PAGE_MASK
;
3076 if ((vma
->vm_flags
& VM_GROWSDOWN
) && address
== vma
->vm_start
) {
3077 struct vm_area_struct
*prev
= vma
->vm_prev
;
3080 * Is there a mapping abutting this one below?
3082 * That's only ok if it's the same stack mapping
3083 * that has gotten split..
3085 if (prev
&& prev
->vm_end
== address
)
3086 return prev
->vm_flags
& VM_GROWSDOWN
? 0 : -ENOMEM
;
3088 expand_downwards(vma
, address
- PAGE_SIZE
);
3090 if ((vma
->vm_flags
& VM_GROWSUP
) && address
+ PAGE_SIZE
== vma
->vm_end
) {
3091 struct vm_area_struct
*next
= vma
->vm_next
;
3093 /* As VM_GROWSDOWN but s/below/above/ */
3094 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
)
3095 return next
->vm_flags
& VM_GROWSUP
? 0 : -ENOMEM
;
3097 expand_upwards(vma
, address
+ PAGE_SIZE
);
3103 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3104 * but allow concurrent faults), and pte mapped but not yet locked.
3105 * We return with mmap_sem still held, but pte unmapped and unlocked.
3107 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3108 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3115 pte_unmap(page_table
);
3117 /* Check if we need to add a guard page to the stack */
3118 if (check_stack_guard_page(vma
, address
) < 0)
3119 return VM_FAULT_SIGBUS
;
3121 /* Use the zero-page for reads */
3122 if (!(flags
& FAULT_FLAG_WRITE
)) {
3123 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(address
),
3124 vma
->vm_page_prot
));
3125 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3126 if (!pte_none(*page_table
))
3131 /* Allocate our own private page. */
3132 if (unlikely(anon_vma_prepare(vma
)))
3134 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
3137 __SetPageUptodate(page
);
3139 if (mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
))
3142 entry
= mk_pte(page
, vma
->vm_page_prot
);
3143 if (vma
->vm_flags
& VM_WRITE
)
3144 entry
= pte_mkwrite(pte_mkdirty(entry
));
3146 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3147 if (!pte_none(*page_table
))
3150 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3151 page_add_new_anon_rmap(page
, vma
, address
);
3153 set_pte_at(mm
, address
, page_table
, entry
);
3155 /* No need to invalidate - it was non-present before */
3156 update_mmu_cache(vma
, address
, page_table
);
3158 pte_unmap_unlock(page_table
, ptl
);
3161 mem_cgroup_uncharge_page(page
);
3162 page_cache_release(page
);
3165 page_cache_release(page
);
3167 return VM_FAULT_OOM
;
3171 * __do_fault() tries to create a new page mapping. It aggressively
3172 * tries to share with existing pages, but makes a separate copy if
3173 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3174 * the next page fault.
3176 * As this is called only for pages that do not currently exist, we
3177 * do not need to flush old virtual caches or the TLB.
3179 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3180 * but allow concurrent faults), and pte neither mapped nor locked.
3181 * We return with mmap_sem still held, but pte unmapped and unlocked.
3183 static int __do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3184 unsigned long address
, pmd_t
*pmd
,
3185 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3190 struct page
*cow_page
;
3193 struct page
*dirty_page
= NULL
;
3194 struct vm_fault vmf
;
3196 int page_mkwrite
= 0;
3199 * If we do COW later, allocate page befor taking lock_page()
3200 * on the file cache page. This will reduce lock holding time.
3202 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3204 if (unlikely(anon_vma_prepare(vma
)))
3205 return VM_FAULT_OOM
;
3207 cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
3209 return VM_FAULT_OOM
;
3211 if (mem_cgroup_newpage_charge(cow_page
, mm
, GFP_KERNEL
)) {
3212 page_cache_release(cow_page
);
3213 return VM_FAULT_OOM
;
3218 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
3223 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
3224 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
3228 if (unlikely(PageHWPoison(vmf
.page
))) {
3229 if (ret
& VM_FAULT_LOCKED
)
3230 unlock_page(vmf
.page
);
3231 ret
= VM_FAULT_HWPOISON
;
3236 * For consistency in subsequent calls, make the faulted page always
3239 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3240 lock_page(vmf
.page
);
3242 VM_BUG_ON(!PageLocked(vmf
.page
));
3245 * Should we do an early C-O-W break?
3248 if (flags
& FAULT_FLAG_WRITE
) {
3249 if (!(vma
->vm_flags
& VM_SHARED
)) {
3252 copy_user_highpage(page
, vmf
.page
, address
, vma
);
3253 __SetPageUptodate(page
);
3256 * If the page will be shareable, see if the backing
3257 * address space wants to know that the page is about
3258 * to become writable
3260 if (vma
->vm_ops
->page_mkwrite
) {
3264 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
3265 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
3267 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3269 goto unwritable_page
;
3271 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
3273 if (!page
->mapping
) {
3274 ret
= 0; /* retry the fault */
3276 goto unwritable_page
;
3279 VM_BUG_ON(!PageLocked(page
));
3286 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3289 * This silly early PAGE_DIRTY setting removes a race
3290 * due to the bad i386 page protection. But it's valid
3291 * for other architectures too.
3293 * Note that if FAULT_FLAG_WRITE is set, we either now have
3294 * an exclusive copy of the page, or this is a shared mapping,
3295 * so we can make it writable and dirty to avoid having to
3296 * handle that later.
3298 /* Only go through if we didn't race with anybody else... */
3299 if (likely(pte_same(*page_table
, orig_pte
))) {
3300 flush_icache_page(vma
, page
);
3301 entry
= mk_pte(page
, vma
->vm_page_prot
);
3302 if (flags
& FAULT_FLAG_WRITE
)
3303 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3305 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3306 page_add_new_anon_rmap(page
, vma
, address
);
3308 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
3309 page_add_file_rmap(page
);
3310 if (flags
& FAULT_FLAG_WRITE
) {
3312 get_page(dirty_page
);
3315 set_pte_at(mm
, address
, page_table
, entry
);
3317 /* no need to invalidate: a not-present page won't be cached */
3318 update_mmu_cache(vma
, address
, page_table
);
3321 mem_cgroup_uncharge_page(cow_page
);
3323 page_cache_release(page
);
3325 anon
= 1; /* no anon but release faulted_page */
3328 pte_unmap_unlock(page_table
, ptl
);
3331 struct address_space
*mapping
= page
->mapping
;
3334 if (set_page_dirty(dirty_page
))
3336 unlock_page(dirty_page
);
3337 put_page(dirty_page
);
3338 if ((dirtied
|| page_mkwrite
) && mapping
) {
3340 * Some device drivers do not set page.mapping but still
3343 balance_dirty_pages_ratelimited(mapping
);
3346 /* file_update_time outside page_lock */
3347 if (vma
->vm_file
&& !page_mkwrite
)
3348 file_update_time(vma
->vm_file
);
3350 unlock_page(vmf
.page
);
3352 page_cache_release(vmf
.page
);
3358 page_cache_release(page
);
3361 /* fs's fault handler get error */
3363 mem_cgroup_uncharge_page(cow_page
);
3364 page_cache_release(cow_page
);
3369 static int do_linear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3370 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3371 unsigned int flags
, pte_t orig_pte
)
3373 pgoff_t pgoff
= (((address
& PAGE_MASK
)
3374 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
3376 pte_unmap(page_table
);
3377 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3381 * Fault of a previously existing named mapping. Repopulate the pte
3382 * from the encoded file_pte if possible. This enables swappable
3385 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3386 * but allow concurrent faults), and pte mapped but not yet locked.
3387 * We return with mmap_sem still held, but pte unmapped and unlocked.
3389 static int do_nonlinear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3390 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3391 unsigned int flags
, pte_t orig_pte
)
3395 flags
|= FAULT_FLAG_NONLINEAR
;
3397 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
3400 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
))) {
3402 * Page table corrupted: show pte and kill process.
3404 print_bad_pte(vma
, address
, orig_pte
, NULL
);
3405 return VM_FAULT_SIGBUS
;
3408 pgoff
= pte_to_pgoff(orig_pte
);
3409 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3413 * These routines also need to handle stuff like marking pages dirty
3414 * and/or accessed for architectures that don't do it in hardware (most
3415 * RISC architectures). The early dirtying is also good on the i386.
3417 * There is also a hook called "update_mmu_cache()" that architectures
3418 * with external mmu caches can use to update those (ie the Sparc or
3419 * PowerPC hashed page tables that act as extended TLBs).
3421 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3422 * but allow concurrent faults), and pte mapped but not yet locked.
3423 * We return with mmap_sem still held, but pte unmapped and unlocked.
3425 int handle_pte_fault(struct mm_struct
*mm
,
3426 struct vm_area_struct
*vma
, unsigned long address
,
3427 pte_t
*pte
, pmd_t
*pmd
, unsigned int flags
)
3433 if (!pte_present(entry
)) {
3434 if (pte_none(entry
)) {
3436 if (likely(vma
->vm_ops
->fault
))
3437 return do_linear_fault(mm
, vma
, address
,
3438 pte
, pmd
, flags
, entry
);
3440 return do_anonymous_page(mm
, vma
, address
,
3443 if (pte_file(entry
))
3444 return do_nonlinear_fault(mm
, vma
, address
,
3445 pte
, pmd
, flags
, entry
);
3446 return do_swap_page(mm
, vma
, address
,
3447 pte
, pmd
, flags
, entry
);
3450 ptl
= pte_lockptr(mm
, pmd
);
3452 if (unlikely(!pte_same(*pte
, entry
)))
3454 if (flags
& FAULT_FLAG_WRITE
) {
3455 if (!pte_write(entry
))
3456 return do_wp_page(mm
, vma
, address
,
3457 pte
, pmd
, ptl
, entry
);
3458 entry
= pte_mkdirty(entry
);
3460 entry
= pte_mkyoung(entry
);
3461 if (ptep_set_access_flags(vma
, address
, pte
, entry
, flags
& FAULT_FLAG_WRITE
)) {
3462 update_mmu_cache(vma
, address
, pte
);
3465 * This is needed only for protection faults but the arch code
3466 * is not yet telling us if this is a protection fault or not.
3467 * This still avoids useless tlb flushes for .text page faults
3470 if (flags
& FAULT_FLAG_WRITE
)
3471 flush_tlb_fix_spurious_fault(vma
, address
);
3474 pte_unmap_unlock(pte
, ptl
);
3479 * By the time we get here, we already hold the mm semaphore
3481 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3482 unsigned long address
, unsigned int flags
)
3489 __set_current_state(TASK_RUNNING
);
3491 count_vm_event(PGFAULT
);
3492 mem_cgroup_count_vm_event(mm
, PGFAULT
);
3494 /* do counter updates before entering really critical section. */
3495 check_sync_rss_stat(current
);
3497 if (unlikely(is_vm_hugetlb_page(vma
)))
3498 return hugetlb_fault(mm
, vma
, address
, flags
);
3501 pgd
= pgd_offset(mm
, address
);
3502 pud
= pud_alloc(mm
, pgd
, address
);
3504 return VM_FAULT_OOM
;
3505 pmd
= pmd_alloc(mm
, pud
, address
);
3507 return VM_FAULT_OOM
;
3508 if (pmd_none(*pmd
) && transparent_hugepage_enabled(vma
)) {
3510 return do_huge_pmd_anonymous_page(mm
, vma
, address
,
3513 pmd_t orig_pmd
= *pmd
;
3517 if (pmd_trans_huge(orig_pmd
)) {
3518 if (flags
& FAULT_FLAG_WRITE
&&
3519 !pmd_write(orig_pmd
) &&
3520 !pmd_trans_splitting(orig_pmd
)) {
3521 ret
= do_huge_pmd_wp_page(mm
, vma
, address
, pmd
,
3524 * If COW results in an oom, the huge pmd will
3525 * have been split, so retry the fault on the
3526 * pte for a smaller charge.
3528 if (unlikely(ret
& VM_FAULT_OOM
))
3537 * Use __pte_alloc instead of pte_alloc_map, because we can't
3538 * run pte_offset_map on the pmd, if an huge pmd could
3539 * materialize from under us from a different thread.
3541 if (unlikely(pmd_none(*pmd
)) && __pte_alloc(mm
, vma
, pmd
, address
))
3542 return VM_FAULT_OOM
;
3543 /* if an huge pmd materialized from under us just retry later */
3544 if (unlikely(pmd_trans_huge(*pmd
)))
3547 * A regular pmd is established and it can't morph into a huge pmd
3548 * from under us anymore at this point because we hold the mmap_sem
3549 * read mode and khugepaged takes it in write mode. So now it's
3550 * safe to run pte_offset_map().
3552 pte
= pte_offset_map(pmd
, address
);
3554 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
3557 #ifndef __PAGETABLE_PUD_FOLDED
3559 * Allocate page upper directory.
3560 * We've already handled the fast-path in-line.
3562 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
3564 pud_t
*new = pud_alloc_one(mm
, address
);
3568 smp_wmb(); /* See comment in __pte_alloc */
3570 spin_lock(&mm
->page_table_lock
);
3571 if (pgd_present(*pgd
)) /* Another has populated it */
3574 pgd_populate(mm
, pgd
, new);
3575 spin_unlock(&mm
->page_table_lock
);
3578 #endif /* __PAGETABLE_PUD_FOLDED */
3580 #ifndef __PAGETABLE_PMD_FOLDED
3582 * Allocate page middle directory.
3583 * We've already handled the fast-path in-line.
3585 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
3587 pmd_t
*new = pmd_alloc_one(mm
, address
);
3591 smp_wmb(); /* See comment in __pte_alloc */
3593 spin_lock(&mm
->page_table_lock
);
3594 #ifndef __ARCH_HAS_4LEVEL_HACK
3595 if (pud_present(*pud
)) /* Another has populated it */
3598 pud_populate(mm
, pud
, new);
3600 if (pgd_present(*pud
)) /* Another has populated it */
3603 pgd_populate(mm
, pud
, new);
3604 #endif /* __ARCH_HAS_4LEVEL_HACK */
3605 spin_unlock(&mm
->page_table_lock
);
3608 #endif /* __PAGETABLE_PMD_FOLDED */
3610 int make_pages_present(unsigned long addr
, unsigned long end
)
3612 int ret
, len
, write
;
3613 struct vm_area_struct
* vma
;
3615 vma
= find_vma(current
->mm
, addr
);
3619 * We want to touch writable mappings with a write fault in order
3620 * to break COW, except for shared mappings because these don't COW
3621 * and we would not want to dirty them for nothing.
3623 write
= (vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
;
3624 BUG_ON(addr
>= end
);
3625 BUG_ON(end
> vma
->vm_end
);
3626 len
= DIV_ROUND_UP(end
, PAGE_SIZE
) - addr
/PAGE_SIZE
;
3627 ret
= get_user_pages(current
, current
->mm
, addr
,
3628 len
, write
, 0, NULL
, NULL
);
3631 return ret
== len
? 0 : -EFAULT
;
3634 #if !defined(__HAVE_ARCH_GATE_AREA)
3636 #if defined(AT_SYSINFO_EHDR)
3637 static struct vm_area_struct gate_vma
;
3639 static int __init
gate_vma_init(void)
3641 gate_vma
.vm_mm
= NULL
;
3642 gate_vma
.vm_start
= FIXADDR_USER_START
;
3643 gate_vma
.vm_end
= FIXADDR_USER_END
;
3644 gate_vma
.vm_flags
= VM_READ
| VM_MAYREAD
| VM_EXEC
| VM_MAYEXEC
;
3645 gate_vma
.vm_page_prot
= __P101
;
3649 __initcall(gate_vma_init
);
3652 struct vm_area_struct
*get_gate_vma(struct mm_struct
*mm
)
3654 #ifdef AT_SYSINFO_EHDR
3661 int in_gate_area_no_mm(unsigned long addr
)
3663 #ifdef AT_SYSINFO_EHDR
3664 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
3670 #endif /* __HAVE_ARCH_GATE_AREA */
3672 static int __follow_pte(struct mm_struct
*mm
, unsigned long address
,
3673 pte_t
**ptepp
, spinlock_t
**ptlp
)
3680 pgd
= pgd_offset(mm
, address
);
3681 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
3684 pud
= pud_offset(pgd
, address
);
3685 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
3688 pmd
= pmd_offset(pud
, address
);
3689 VM_BUG_ON(pmd_trans_huge(*pmd
));
3690 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
3693 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3697 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
3700 if (!pte_present(*ptep
))
3705 pte_unmap_unlock(ptep
, *ptlp
);
3710 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
3711 pte_t
**ptepp
, spinlock_t
**ptlp
)
3715 /* (void) is needed to make gcc happy */
3716 (void) __cond_lock(*ptlp
,
3717 !(res
= __follow_pte(mm
, address
, ptepp
, ptlp
)));
3722 * follow_pfn - look up PFN at a user virtual address
3723 * @vma: memory mapping
3724 * @address: user virtual address
3725 * @pfn: location to store found PFN
3727 * Only IO mappings and raw PFN mappings are allowed.
3729 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3731 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
3738 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3741 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
3744 *pfn
= pte_pfn(*ptep
);
3745 pte_unmap_unlock(ptep
, ptl
);
3748 EXPORT_SYMBOL(follow_pfn
);
3750 #ifdef CONFIG_HAVE_IOREMAP_PROT
3751 int follow_phys(struct vm_area_struct
*vma
,
3752 unsigned long address
, unsigned int flags
,
3753 unsigned long *prot
, resource_size_t
*phys
)
3759 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3762 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
3766 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
3769 *prot
= pgprot_val(pte_pgprot(pte
));
3770 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
3774 pte_unmap_unlock(ptep
, ptl
);
3779 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
3780 void *buf
, int len
, int write
)
3782 resource_size_t phys_addr
;
3783 unsigned long prot
= 0;
3784 void __iomem
*maddr
;
3785 int offset
= addr
& (PAGE_SIZE
-1);
3787 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
3790 maddr
= ioremap_prot(phys_addr
, PAGE_SIZE
, prot
);
3792 memcpy_toio(maddr
+ offset
, buf
, len
);
3794 memcpy_fromio(buf
, maddr
+ offset
, len
);
3802 * Access another process' address space as given in mm. If non-NULL, use the
3803 * given task for page fault accounting.
3805 static int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
3806 unsigned long addr
, void *buf
, int len
, int write
)
3808 struct vm_area_struct
*vma
;
3809 void *old_buf
= buf
;
3811 down_read(&mm
->mmap_sem
);
3812 /* ignore errors, just check how much was successfully transferred */
3814 int bytes
, ret
, offset
;
3816 struct page
*page
= NULL
;
3818 ret
= get_user_pages(tsk
, mm
, addr
, 1,
3819 write
, 1, &page
, &vma
);
3822 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3823 * we can access using slightly different code.
3825 #ifdef CONFIG_HAVE_IOREMAP_PROT
3826 vma
= find_vma(mm
, addr
);
3827 if (!vma
|| vma
->vm_start
> addr
)
3829 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
3830 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
3838 offset
= addr
& (PAGE_SIZE
-1);
3839 if (bytes
> PAGE_SIZE
-offset
)
3840 bytes
= PAGE_SIZE
-offset
;
3844 copy_to_user_page(vma
, page
, addr
,
3845 maddr
+ offset
, buf
, bytes
);
3846 set_page_dirty_lock(page
);
3848 copy_from_user_page(vma
, page
, addr
,
3849 buf
, maddr
+ offset
, bytes
);
3852 page_cache_release(page
);
3858 up_read(&mm
->mmap_sem
);
3860 return buf
- old_buf
;
3864 * access_remote_vm - access another process' address space
3865 * @mm: the mm_struct of the target address space
3866 * @addr: start address to access
3867 * @buf: source or destination buffer
3868 * @len: number of bytes to transfer
3869 * @write: whether the access is a write
3871 * The caller must hold a reference on @mm.
3873 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
3874 void *buf
, int len
, int write
)
3876 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, write
);
3880 * Access another process' address space.
3881 * Source/target buffer must be kernel space,
3882 * Do not walk the page table directly, use get_user_pages
3884 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
3885 void *buf
, int len
, int write
)
3887 struct mm_struct
*mm
;
3890 mm
= get_task_mm(tsk
);
3894 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, write
);
3901 * Print the name of a VMA.
3903 void print_vma_addr(char *prefix
, unsigned long ip
)
3905 struct mm_struct
*mm
= current
->mm
;
3906 struct vm_area_struct
*vma
;
3909 * Do not print if we are in atomic
3910 * contexts (in exception stacks, etc.):
3912 if (preempt_count())
3915 down_read(&mm
->mmap_sem
);
3916 vma
= find_vma(mm
, ip
);
3917 if (vma
&& vma
->vm_file
) {
3918 struct file
*f
= vma
->vm_file
;
3919 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
3923 p
= d_path(&f
->f_path
, buf
, PAGE_SIZE
);
3926 s
= strrchr(p
, '/');
3929 printk("%s%s[%lx+%lx]", prefix
, p
,
3931 vma
->vm_end
- vma
->vm_start
);
3932 free_page((unsigned long)buf
);
3935 up_read(&mm
->mmap_sem
);
3938 #ifdef CONFIG_PROVE_LOCKING
3939 void might_fault(void)
3942 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3943 * holding the mmap_sem, this is safe because kernel memory doesn't
3944 * get paged out, therefore we'll never actually fault, and the
3945 * below annotations will generate false positives.
3947 if (segment_eq(get_fs(), KERNEL_DS
))
3952 * it would be nicer only to annotate paths which are not under
3953 * pagefault_disable, however that requires a larger audit and
3954 * providing helpers like get_user_atomic.
3956 if (!in_atomic() && current
->mm
)
3957 might_lock_read(¤t
->mm
->mmap_sem
);
3959 EXPORT_SYMBOL(might_fault
);
3962 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3963 static void clear_gigantic_page(struct page
*page
,
3965 unsigned int pages_per_huge_page
)
3968 struct page
*p
= page
;
3971 for (i
= 0; i
< pages_per_huge_page
;
3972 i
++, p
= mem_map_next(p
, page
, i
)) {
3974 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
3977 void clear_huge_page(struct page
*page
,
3978 unsigned long addr
, unsigned int pages_per_huge_page
)
3982 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3983 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
3988 for (i
= 0; i
< pages_per_huge_page
; i
++) {
3990 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
3994 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
3996 struct vm_area_struct
*vma
,
3997 unsigned int pages_per_huge_page
)
4000 struct page
*dst_base
= dst
;
4001 struct page
*src_base
= src
;
4003 for (i
= 0; i
< pages_per_huge_page
; ) {
4005 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
4008 dst
= mem_map_next(dst
, dst_base
, i
);
4009 src
= mem_map_next(src
, src_base
, i
);
4013 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
4014 unsigned long addr
, struct vm_area_struct
*vma
,
4015 unsigned int pages_per_huge_page
)
4019 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4020 copy_user_gigantic_page(dst
, src
, addr
, vma
,
4021 pages_per_huge_page
);
4026 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4028 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
4031 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */