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/rmap.h>
49 #include <linux/module.h>
50 #include <linux/init.h>
52 #include <asm/pgalloc.h>
53 #include <asm/uaccess.h>
55 #include <asm/tlbflush.h>
56 #include <asm/pgtable.h>
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
61 #ifndef CONFIG_NEED_MULTIPLE_NODES
62 /* use the per-pgdat data instead for discontigmem - mbligh */
63 unsigned long max_mapnr
;
66 EXPORT_SYMBOL(max_mapnr
);
67 EXPORT_SYMBOL(mem_map
);
70 unsigned long num_physpages
;
72 * A number of key systems in x86 including ioremap() rely on the assumption
73 * that high_memory defines the upper bound on direct map memory, then end
74 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
75 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
79 unsigned long vmalloc_earlyreserve
;
81 EXPORT_SYMBOL(num_physpages
);
82 EXPORT_SYMBOL(high_memory
);
83 EXPORT_SYMBOL(vmalloc_earlyreserve
);
86 * If a p?d_bad entry is found while walking page tables, report
87 * the error, before resetting entry to p?d_none. Usually (but
88 * very seldom) called out from the p?d_none_or_clear_bad macros.
91 void pgd_clear_bad(pgd_t
*pgd
)
97 void pud_clear_bad(pud_t
*pud
)
103 void pmd_clear_bad(pmd_t
*pmd
)
110 * Note: this doesn't free the actual pages themselves. That
111 * has been handled earlier when unmapping all the memory regions.
113 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
)
115 struct page
*page
= pmd_page(*pmd
);
117 pte_free_tlb(tlb
, page
);
118 dec_page_state(nr_page_table_pages
);
122 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
123 unsigned long addr
, unsigned long end
,
124 unsigned long floor
, unsigned long ceiling
)
131 pmd
= pmd_offset(pud
, addr
);
133 next
= pmd_addr_end(addr
, end
);
134 if (pmd_none_or_clear_bad(pmd
))
136 free_pte_range(tlb
, pmd
);
137 } while (pmd
++, addr
= next
, addr
!= end
);
147 if (end
- 1 > ceiling
- 1)
150 pmd
= pmd_offset(pud
, start
);
152 pmd_free_tlb(tlb
, pmd
);
155 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
156 unsigned long addr
, unsigned long end
,
157 unsigned long floor
, unsigned long ceiling
)
164 pud
= pud_offset(pgd
, addr
);
166 next
= pud_addr_end(addr
, end
);
167 if (pud_none_or_clear_bad(pud
))
169 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
170 } while (pud
++, addr
= next
, addr
!= end
);
176 ceiling
&= PGDIR_MASK
;
180 if (end
- 1 > ceiling
- 1)
183 pud
= pud_offset(pgd
, start
);
185 pud_free_tlb(tlb
, pud
);
189 * This function frees user-level page tables of a process.
191 * Must be called with pagetable lock held.
193 void free_pgd_range(struct mmu_gather
**tlb
,
194 unsigned long addr
, unsigned long end
,
195 unsigned long floor
, unsigned long ceiling
)
202 * The next few lines have given us lots of grief...
204 * Why are we testing PMD* at this top level? Because often
205 * there will be no work to do at all, and we'd prefer not to
206 * go all the way down to the bottom just to discover that.
208 * Why all these "- 1"s? Because 0 represents both the bottom
209 * of the address space and the top of it (using -1 for the
210 * top wouldn't help much: the masks would do the wrong thing).
211 * The rule is that addr 0 and floor 0 refer to the bottom of
212 * the address space, but end 0 and ceiling 0 refer to the top
213 * Comparisons need to use "end - 1" and "ceiling - 1" (though
214 * that end 0 case should be mythical).
216 * Wherever addr is brought up or ceiling brought down, we must
217 * be careful to reject "the opposite 0" before it confuses the
218 * subsequent tests. But what about where end is brought down
219 * by PMD_SIZE below? no, end can't go down to 0 there.
221 * Whereas we round start (addr) and ceiling down, by different
222 * masks at different levels, in order to test whether a table
223 * now has no other vmas using it, so can be freed, we don't
224 * bother to round floor or end up - the tests don't need that.
238 if (end
- 1 > ceiling
- 1)
244 pgd
= pgd_offset((*tlb
)->mm
, addr
);
246 next
= pgd_addr_end(addr
, end
);
247 if (pgd_none_or_clear_bad(pgd
))
249 free_pud_range(*tlb
, pgd
, addr
, next
, floor
, ceiling
);
250 } while (pgd
++, addr
= next
, addr
!= end
);
253 flush_tlb_pgtables((*tlb
)->mm
, start
, end
);
256 void free_pgtables(struct mmu_gather
**tlb
, struct vm_area_struct
*vma
,
257 unsigned long floor
, unsigned long ceiling
)
260 struct vm_area_struct
*next
= vma
->vm_next
;
261 unsigned long addr
= vma
->vm_start
;
264 * Hide vma from rmap and vmtruncate before freeing pgtables
266 anon_vma_unlink(vma
);
267 unlink_file_vma(vma
);
269 if (is_hugepage_only_range(vma
->vm_mm
, addr
, HPAGE_SIZE
)) {
270 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
271 floor
, next
? next
->vm_start
: ceiling
);
274 * Optimization: gather nearby vmas into one call down
276 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
277 && !is_hugepage_only_range(vma
->vm_mm
, next
->vm_start
,
281 anon_vma_unlink(vma
);
282 unlink_file_vma(vma
);
284 free_pgd_range(tlb
, addr
, vma
->vm_end
,
285 floor
, next
? next
->vm_start
: ceiling
);
291 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
, unsigned long address
)
293 struct page
*new = pte_alloc_one(mm
, address
);
297 spin_lock(&mm
->page_table_lock
);
298 if (pmd_present(*pmd
)) /* Another has populated it */
302 inc_page_state(nr_page_table_pages
);
303 pmd_populate(mm
, pmd
, new);
305 spin_unlock(&mm
->page_table_lock
);
309 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
311 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
315 spin_lock(&init_mm
.page_table_lock
);
316 if (pmd_present(*pmd
)) /* Another has populated it */
317 pte_free_kernel(new);
319 pmd_populate_kernel(&init_mm
, pmd
, new);
320 spin_unlock(&init_mm
.page_table_lock
);
324 static inline void add_mm_rss(struct mm_struct
*mm
, int file_rss
, int anon_rss
)
327 add_mm_counter(mm
, file_rss
, file_rss
);
329 add_mm_counter(mm
, anon_rss
, anon_rss
);
333 * This function is called to print an error when a pte in a
334 * !VM_RESERVED region is found pointing to an invalid pfn (which
337 * The calling function must still handle the error.
339 void print_bad_pte(struct vm_area_struct
*vma
, pte_t pte
, unsigned long vaddr
)
341 printk(KERN_ERR
"Bad pte = %08llx, process = %s, "
342 "vm_flags = %lx, vaddr = %lx\n",
343 (long long)pte_val(pte
),
344 (vma
->vm_mm
== current
->mm
? current
->comm
: "???"),
345 vma
->vm_flags
, vaddr
);
350 * copy one vm_area from one task to the other. Assumes the page tables
351 * already present in the new task to be cleared in the whole range
352 * covered by this vma.
356 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
357 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
358 unsigned long addr
, int *rss
)
360 unsigned long vm_flags
= vma
->vm_flags
;
361 pte_t pte
= *src_pte
;
365 /* pte contains position in swap or file, so copy. */
366 if (unlikely(!pte_present(pte
))) {
367 if (!pte_file(pte
)) {
368 swap_duplicate(pte_to_swp_entry(pte
));
369 /* make sure dst_mm is on swapoff's mmlist. */
370 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
371 spin_lock(&mmlist_lock
);
372 list_add(&dst_mm
->mmlist
, &src_mm
->mmlist
);
373 spin_unlock(&mmlist_lock
);
379 /* If the region is VM_RESERVED, the mapping is not
380 * mapped via rmap - duplicate the pte as is.
382 if (vm_flags
& VM_RESERVED
)
386 /* If the pte points outside of valid memory but
387 * the region is not VM_RESERVED, we have a problem.
389 if (unlikely(!pfn_valid(pfn
))) {
390 print_bad_pte(vma
, pte
, addr
);
391 goto out_set_pte
; /* try to do something sane */
394 page
= pfn_to_page(pfn
);
397 * If it's a COW mapping, write protect it both
398 * in the parent and the child
400 if ((vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
) {
401 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
406 * If it's a shared mapping, mark it clean in
409 if (vm_flags
& VM_SHARED
)
410 pte
= pte_mkclean(pte
);
411 pte
= pte_mkold(pte
);
414 rss
[!!PageAnon(page
)]++;
417 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
420 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
421 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
422 unsigned long addr
, unsigned long end
)
424 pte_t
*src_pte
, *dst_pte
;
425 spinlock_t
*src_ptl
, *dst_ptl
;
431 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
434 src_pte
= pte_offset_map_nested(src_pmd
, addr
);
435 src_ptl
= &src_mm
->page_table_lock
;
440 * We are holding two locks at this point - either of them
441 * could generate latencies in another task on another CPU.
443 if (progress
>= 32) {
445 if (need_resched() ||
446 need_lockbreak(src_ptl
) ||
447 need_lockbreak(dst_ptl
))
450 if (pte_none(*src_pte
)) {
454 copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
, vma
, addr
, rss
);
456 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
458 spin_unlock(src_ptl
);
459 pte_unmap_nested(src_pte
- 1);
460 add_mm_rss(dst_mm
, rss
[0], rss
[1]);
461 pte_unmap_unlock(dst_pte
- 1, dst_ptl
);
468 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
469 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
470 unsigned long addr
, unsigned long end
)
472 pmd_t
*src_pmd
, *dst_pmd
;
475 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
478 src_pmd
= pmd_offset(src_pud
, addr
);
480 next
= pmd_addr_end(addr
, end
);
481 if (pmd_none_or_clear_bad(src_pmd
))
483 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
486 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
490 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
491 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
492 unsigned long addr
, unsigned long end
)
494 pud_t
*src_pud
, *dst_pud
;
497 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
500 src_pud
= pud_offset(src_pgd
, addr
);
502 next
= pud_addr_end(addr
, end
);
503 if (pud_none_or_clear_bad(src_pud
))
505 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
508 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
512 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
513 struct vm_area_struct
*vma
)
515 pgd_t
*src_pgd
, *dst_pgd
;
517 unsigned long addr
= vma
->vm_start
;
518 unsigned long end
= vma
->vm_end
;
521 * Don't copy ptes where a page fault will fill them correctly.
522 * Fork becomes much lighter when there are big shared or private
523 * readonly mappings. The tradeoff is that copy_page_range is more
524 * efficient than faulting.
526 if (!(vma
->vm_flags
& (VM_HUGETLB
|VM_NONLINEAR
|VM_RESERVED
))) {
531 if (is_vm_hugetlb_page(vma
))
532 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
534 dst_pgd
= pgd_offset(dst_mm
, addr
);
535 src_pgd
= pgd_offset(src_mm
, addr
);
537 next
= pgd_addr_end(addr
, end
);
538 if (pgd_none_or_clear_bad(src_pgd
))
540 if (copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
543 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
547 static void zap_pte_range(struct mmu_gather
*tlb
,
548 struct vm_area_struct
*vma
, pmd_t
*pmd
,
549 unsigned long addr
, unsigned long end
,
550 struct zap_details
*details
)
552 struct mm_struct
*mm
= tlb
->mm
;
558 pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
563 if (pte_present(ptent
)) {
564 struct page
*page
= NULL
;
565 if (!(vma
->vm_flags
& VM_RESERVED
)) {
566 unsigned long pfn
= pte_pfn(ptent
);
567 if (unlikely(!pfn_valid(pfn
)))
568 print_bad_pte(vma
, ptent
, addr
);
570 page
= pfn_to_page(pfn
);
572 if (unlikely(details
) && page
) {
574 * unmap_shared_mapping_pages() wants to
575 * invalidate cache without truncating:
576 * unmap shared but keep private pages.
578 if (details
->check_mapping
&&
579 details
->check_mapping
!= page
->mapping
)
582 * Each page->index must be checked when
583 * invalidating or truncating nonlinear.
585 if (details
->nonlinear_vma
&&
586 (page
->index
< details
->first_index
||
587 page
->index
> details
->last_index
))
590 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
592 tlb_remove_tlb_entry(tlb
, pte
, addr
);
595 if (unlikely(details
) && details
->nonlinear_vma
596 && linear_page_index(details
->nonlinear_vma
,
597 addr
) != page
->index
)
598 set_pte_at(mm
, addr
, pte
,
599 pgoff_to_pte(page
->index
));
603 if (pte_dirty(ptent
))
604 set_page_dirty(page
);
605 if (pte_young(ptent
))
606 mark_page_accessed(page
);
609 page_remove_rmap(page
);
610 tlb_remove_page(tlb
, page
);
614 * If details->check_mapping, we leave swap entries;
615 * if details->nonlinear_vma, we leave file entries.
617 if (unlikely(details
))
619 if (!pte_file(ptent
))
620 free_swap_and_cache(pte_to_swp_entry(ptent
));
621 pte_clear_full(mm
, addr
, pte
, tlb
->fullmm
);
622 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
624 add_mm_rss(mm
, file_rss
, anon_rss
);
625 pte_unmap_unlock(pte
- 1, ptl
);
628 static inline void zap_pmd_range(struct mmu_gather
*tlb
,
629 struct vm_area_struct
*vma
, pud_t
*pud
,
630 unsigned long addr
, unsigned long end
,
631 struct zap_details
*details
)
636 pmd
= pmd_offset(pud
, addr
);
638 next
= pmd_addr_end(addr
, end
);
639 if (pmd_none_or_clear_bad(pmd
))
641 zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
642 } while (pmd
++, addr
= next
, addr
!= end
);
645 static inline void zap_pud_range(struct mmu_gather
*tlb
,
646 struct vm_area_struct
*vma
, pgd_t
*pgd
,
647 unsigned long addr
, unsigned long end
,
648 struct zap_details
*details
)
653 pud
= pud_offset(pgd
, addr
);
655 next
= pud_addr_end(addr
, end
);
656 if (pud_none_or_clear_bad(pud
))
658 zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
659 } while (pud
++, addr
= next
, addr
!= end
);
662 static void unmap_page_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
663 unsigned long addr
, unsigned long end
,
664 struct zap_details
*details
)
669 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
673 tlb_start_vma(tlb
, vma
);
674 pgd
= pgd_offset(vma
->vm_mm
, addr
);
676 next
= pgd_addr_end(addr
, end
);
677 if (pgd_none_or_clear_bad(pgd
))
679 zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
680 } while (pgd
++, addr
= next
, addr
!= end
);
681 tlb_end_vma(tlb
, vma
);
684 #ifdef CONFIG_PREEMPT
685 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
687 /* No preempt: go for improved straight-line efficiency */
688 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
692 * unmap_vmas - unmap a range of memory covered by a list of vma's
693 * @tlbp: address of the caller's struct mmu_gather
694 * @vma: the starting vma
695 * @start_addr: virtual address at which to start unmapping
696 * @end_addr: virtual address at which to end unmapping
697 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
698 * @details: details of nonlinear truncation or shared cache invalidation
700 * Returns the end address of the unmapping (restart addr if interrupted).
702 * Unmap all pages in the vma list.
704 * We aim to not hold locks for too long (for scheduling latency reasons).
705 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
706 * return the ending mmu_gather to the caller.
708 * Only addresses between `start' and `end' will be unmapped.
710 * The VMA list must be sorted in ascending virtual address order.
712 * unmap_vmas() assumes that the caller will flush the whole unmapped address
713 * range after unmap_vmas() returns. So the only responsibility here is to
714 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
715 * drops the lock and schedules.
717 unsigned long unmap_vmas(struct mmu_gather
**tlbp
,
718 struct vm_area_struct
*vma
, unsigned long start_addr
,
719 unsigned long end_addr
, unsigned long *nr_accounted
,
720 struct zap_details
*details
)
722 unsigned long zap_bytes
= ZAP_BLOCK_SIZE
;
723 unsigned long tlb_start
= 0; /* For tlb_finish_mmu */
724 int tlb_start_valid
= 0;
725 unsigned long start
= start_addr
;
726 spinlock_t
*i_mmap_lock
= details
? details
->i_mmap_lock
: NULL
;
727 int fullmm
= (*tlbp
)->fullmm
;
729 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
) {
732 start
= max(vma
->vm_start
, start_addr
);
733 if (start
>= vma
->vm_end
)
735 end
= min(vma
->vm_end
, end_addr
);
736 if (end
<= vma
->vm_start
)
739 if (vma
->vm_flags
& VM_ACCOUNT
)
740 *nr_accounted
+= (end
- start
) >> PAGE_SHIFT
;
742 while (start
!= end
) {
745 if (!tlb_start_valid
) {
750 if (is_vm_hugetlb_page(vma
)) {
752 unmap_hugepage_range(vma
, start
, end
);
754 block
= min(zap_bytes
, end
- start
);
755 unmap_page_range(*tlbp
, vma
, start
,
756 start
+ block
, details
);
761 if ((long)zap_bytes
> 0)
764 tlb_finish_mmu(*tlbp
, tlb_start
, start
);
766 if (need_resched() ||
767 (i_mmap_lock
&& need_lockbreak(i_mmap_lock
))) {
775 *tlbp
= tlb_gather_mmu(vma
->vm_mm
, fullmm
);
777 zap_bytes
= ZAP_BLOCK_SIZE
;
781 return start
; /* which is now the end (or restart) address */
785 * zap_page_range - remove user pages in a given range
786 * @vma: vm_area_struct holding the applicable pages
787 * @address: starting address of pages to zap
788 * @size: number of bytes to zap
789 * @details: details of nonlinear truncation or shared cache invalidation
791 unsigned long zap_page_range(struct vm_area_struct
*vma
, unsigned long address
,
792 unsigned long size
, struct zap_details
*details
)
794 struct mm_struct
*mm
= vma
->vm_mm
;
795 struct mmu_gather
*tlb
;
796 unsigned long end
= address
+ size
;
797 unsigned long nr_accounted
= 0;
800 tlb
= tlb_gather_mmu(mm
, 0);
801 update_hiwater_rss(mm
);
802 end
= unmap_vmas(&tlb
, vma
, address
, end
, &nr_accounted
, details
);
804 tlb_finish_mmu(tlb
, address
, end
);
809 * Do a quick page-table lookup for a single page.
810 * mm->page_table_lock must be held.
812 static struct page
*__follow_page(struct mm_struct
*mm
, unsigned long address
,
813 int read
, int write
, int accessed
)
822 page
= follow_huge_addr(mm
, address
, write
);
826 pgd
= pgd_offset(mm
, address
);
827 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
830 pud
= pud_offset(pgd
, address
);
831 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
834 pmd
= pmd_offset(pud
, address
);
835 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
838 return follow_huge_pmd(mm
, address
, pmd
, write
);
840 ptep
= pte_offset_map(pmd
, address
);
846 if (pte_present(pte
)) {
847 if (write
&& !pte_write(pte
))
849 if (read
&& !pte_read(pte
))
852 if (pfn_valid(pfn
)) {
853 page
= pfn_to_page(pfn
);
855 if (write
&& !pte_dirty(pte
) &&!PageDirty(page
))
856 set_page_dirty(page
);
857 mark_page_accessed(page
);
868 follow_page(struct mm_struct
*mm
, unsigned long address
, int write
)
870 return __follow_page(mm
, address
, 0, write
, 1);
874 * check_user_page_readable() can be called frm niterrupt context by oprofile,
875 * so we need to avoid taking any non-irq-safe locks
877 int check_user_page_readable(struct mm_struct
*mm
, unsigned long address
)
879 return __follow_page(mm
, address
, 1, 0, 0) != NULL
;
881 EXPORT_SYMBOL(check_user_page_readable
);
884 untouched_anonymous_page(struct mm_struct
* mm
, struct vm_area_struct
*vma
,
885 unsigned long address
)
891 /* Check if the vma is for an anonymous mapping. */
892 if (vma
->vm_ops
&& vma
->vm_ops
->nopage
)
895 /* Check if page directory entry exists. */
896 pgd
= pgd_offset(mm
, address
);
897 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
900 pud
= pud_offset(pgd
, address
);
901 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
904 /* Check if page middle directory entry exists. */
905 pmd
= pmd_offset(pud
, address
);
906 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
909 /* There is a pte slot for 'address' in 'mm'. */
913 int get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
914 unsigned long start
, int len
, int write
, int force
,
915 struct page
**pages
, struct vm_area_struct
**vmas
)
921 * Require read or write permissions.
922 * If 'force' is set, we only require the "MAY" flags.
924 flags
= write
? (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
925 flags
&= force
? (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
929 struct vm_area_struct
* vma
;
931 vma
= find_extend_vma(mm
, start
);
932 if (!vma
&& in_gate_area(tsk
, start
)) {
933 unsigned long pg
= start
& PAGE_MASK
;
934 struct vm_area_struct
*gate_vma
= get_gate_vma(tsk
);
939 if (write
) /* user gate pages are read-only */
940 return i
? : -EFAULT
;
942 pgd
= pgd_offset_k(pg
);
944 pgd
= pgd_offset_gate(mm
, pg
);
945 BUG_ON(pgd_none(*pgd
));
946 pud
= pud_offset(pgd
, pg
);
947 BUG_ON(pud_none(*pud
));
948 pmd
= pmd_offset(pud
, pg
);
950 return i
? : -EFAULT
;
951 pte
= pte_offset_map(pmd
, pg
);
952 if (pte_none(*pte
)) {
954 return i
? : -EFAULT
;
957 pages
[i
] = pte_page(*pte
);
969 if (!vma
|| (vma
->vm_flags
& (VM_IO
| VM_RESERVED
))
970 || !(flags
& vma
->vm_flags
))
971 return i
? : -EFAULT
;
973 if (is_vm_hugetlb_page(vma
)) {
974 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
978 spin_lock(&mm
->page_table_lock
);
980 int write_access
= write
;
983 cond_resched_lock(&mm
->page_table_lock
);
984 while (!(page
= follow_page(mm
, start
, write_access
))) {
988 * Shortcut for anonymous pages. We don't want
989 * to force the creation of pages tables for
990 * insanely big anonymously mapped areas that
991 * nobody touched so far. This is important
992 * for doing a core dump for these mappings.
994 if (!write
&& untouched_anonymous_page(mm
,vma
,start
)) {
995 page
= ZERO_PAGE(start
);
998 spin_unlock(&mm
->page_table_lock
);
999 ret
= __handle_mm_fault(mm
, vma
, start
, write_access
);
1002 * The VM_FAULT_WRITE bit tells us that do_wp_page has
1003 * broken COW when necessary, even if maybe_mkwrite
1004 * decided not to set pte_write. We can thus safely do
1005 * subsequent page lookups as if they were reads.
1007 if (ret
& VM_FAULT_WRITE
)
1010 switch (ret
& ~VM_FAULT_WRITE
) {
1011 case VM_FAULT_MINOR
:
1014 case VM_FAULT_MAJOR
:
1017 case VM_FAULT_SIGBUS
:
1018 return i
? i
: -EFAULT
;
1020 return i
? i
: -ENOMEM
;
1024 spin_lock(&mm
->page_table_lock
);
1028 flush_dcache_page(page
);
1029 page_cache_get(page
);
1036 } while (len
&& start
< vma
->vm_end
);
1037 spin_unlock(&mm
->page_table_lock
);
1041 EXPORT_SYMBOL(get_user_pages
);
1043 static int zeromap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1044 unsigned long addr
, unsigned long end
, pgprot_t prot
)
1049 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1053 struct page
*page
= ZERO_PAGE(addr
);
1054 pte_t zero_pte
= pte_wrprotect(mk_pte(page
, prot
));
1055 page_cache_get(page
);
1056 page_add_file_rmap(page
);
1057 inc_mm_counter(mm
, file_rss
);
1058 BUG_ON(!pte_none(*pte
));
1059 set_pte_at(mm
, addr
, pte
, zero_pte
);
1060 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1061 pte_unmap_unlock(pte
- 1, ptl
);
1065 static inline int zeromap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1066 unsigned long addr
, unsigned long end
, pgprot_t prot
)
1071 pmd
= pmd_alloc(mm
, pud
, addr
);
1075 next
= pmd_addr_end(addr
, end
);
1076 if (zeromap_pte_range(mm
, pmd
, addr
, next
, prot
))
1078 } while (pmd
++, addr
= next
, addr
!= end
);
1082 static inline int zeromap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1083 unsigned long addr
, unsigned long end
, pgprot_t prot
)
1088 pud
= pud_alloc(mm
, pgd
, addr
);
1092 next
= pud_addr_end(addr
, end
);
1093 if (zeromap_pmd_range(mm
, pud
, addr
, next
, prot
))
1095 } while (pud
++, addr
= next
, addr
!= end
);
1099 int zeromap_page_range(struct vm_area_struct
*vma
,
1100 unsigned long addr
, unsigned long size
, pgprot_t prot
)
1104 unsigned long end
= addr
+ size
;
1105 struct mm_struct
*mm
= vma
->vm_mm
;
1108 BUG_ON(addr
>= end
);
1109 pgd
= pgd_offset(mm
, addr
);
1110 flush_cache_range(vma
, addr
, end
);
1112 next
= pgd_addr_end(addr
, end
);
1113 err
= zeromap_pud_range(mm
, pgd
, addr
, next
, prot
);
1116 } while (pgd
++, addr
= next
, addr
!= end
);
1121 * maps a range of physical memory into the requested pages. the old
1122 * mappings are removed. any references to nonexistent pages results
1123 * in null mappings (currently treated as "copy-on-access")
1125 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1126 unsigned long addr
, unsigned long end
,
1127 unsigned long pfn
, pgprot_t prot
)
1132 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1136 BUG_ON(!pte_none(*pte
));
1137 set_pte_at(mm
, addr
, pte
, pfn_pte(pfn
, prot
));
1139 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1140 pte_unmap_unlock(pte
- 1, ptl
);
1144 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1145 unsigned long addr
, unsigned long end
,
1146 unsigned long pfn
, pgprot_t prot
)
1151 pfn
-= addr
>> PAGE_SHIFT
;
1152 pmd
= pmd_alloc(mm
, pud
, addr
);
1156 next
= pmd_addr_end(addr
, end
);
1157 if (remap_pte_range(mm
, pmd
, addr
, next
,
1158 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1160 } while (pmd
++, addr
= next
, addr
!= end
);
1164 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1165 unsigned long addr
, unsigned long end
,
1166 unsigned long pfn
, pgprot_t prot
)
1171 pfn
-= addr
>> PAGE_SHIFT
;
1172 pud
= pud_alloc(mm
, pgd
, addr
);
1176 next
= pud_addr_end(addr
, end
);
1177 if (remap_pmd_range(mm
, pud
, addr
, next
,
1178 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1180 } while (pud
++, addr
= next
, addr
!= end
);
1184 /* Note: this is only safe if the mm semaphore is held when called. */
1185 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1186 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1190 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1191 struct mm_struct
*mm
= vma
->vm_mm
;
1195 * Physically remapped pages are special. Tell the
1196 * rest of the world about it:
1197 * VM_IO tells people not to look at these pages
1198 * (accesses can have side effects).
1199 * VM_RESERVED tells the core MM not to "manage" these pages
1200 * (e.g. refcount, mapcount, try to swap them out).
1202 vma
->vm_flags
|= VM_IO
| VM_RESERVED
;
1204 BUG_ON(addr
>= end
);
1205 pfn
-= addr
>> PAGE_SHIFT
;
1206 pgd
= pgd_offset(mm
, addr
);
1207 flush_cache_range(vma
, addr
, end
);
1209 next
= pgd_addr_end(addr
, end
);
1210 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1211 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1214 } while (pgd
++, addr
= next
, addr
!= end
);
1217 EXPORT_SYMBOL(remap_pfn_range
);
1220 * handle_pte_fault chooses page fault handler according to an entry
1221 * which was read non-atomically. Before making any commitment, on
1222 * those architectures or configurations (e.g. i386 with PAE) which
1223 * might give a mix of unmatched parts, do_swap_page and do_file_page
1224 * must check under lock before unmapping the pte and proceeding
1225 * (but do_wp_page is only called after already making such a check;
1226 * and do_anonymous_page and do_no_page can safely check later on).
1228 static inline int pte_unmap_same(struct mm_struct
*mm
,
1229 pte_t
*page_table
, pte_t orig_pte
)
1232 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1233 if (sizeof(pte_t
) > sizeof(unsigned long)) {
1234 spin_lock(&mm
->page_table_lock
);
1235 same
= pte_same(*page_table
, orig_pte
);
1236 spin_unlock(&mm
->page_table_lock
);
1239 pte_unmap(page_table
);
1244 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1245 * servicing faults for write access. In the normal case, do always want
1246 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1247 * that do not have writing enabled, when used by access_process_vm.
1249 static inline pte_t
maybe_mkwrite(pte_t pte
, struct vm_area_struct
*vma
)
1251 if (likely(vma
->vm_flags
& VM_WRITE
))
1252 pte
= pte_mkwrite(pte
);
1257 * This routine handles present pages, when users try to write
1258 * to a shared page. It is done by copying the page to a new address
1259 * and decrementing the shared-page counter for the old page.
1261 * Note that this routine assumes that the protection checks have been
1262 * done by the caller (the low-level page fault routine in most cases).
1263 * Thus we can safely just mark it writable once we've done any necessary
1266 * We also mark the page dirty at this point even though the page will
1267 * change only once the write actually happens. This avoids a few races,
1268 * and potentially makes it more efficient.
1270 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1271 * but allow concurrent faults), with pte both mapped and locked.
1272 * We return with mmap_sem still held, but pte unmapped and unlocked.
1274 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1275 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1276 spinlock_t
*ptl
, pte_t orig_pte
)
1278 struct page
*old_page
, *new_page
;
1279 unsigned long pfn
= pte_pfn(orig_pte
);
1281 int ret
= VM_FAULT_MINOR
;
1283 BUG_ON(vma
->vm_flags
& VM_RESERVED
);
1285 if (unlikely(!pfn_valid(pfn
))) {
1287 * Page table corrupted: show pte and kill process.
1289 print_bad_pte(vma
, orig_pte
, address
);
1293 old_page
= pfn_to_page(pfn
);
1295 if (PageAnon(old_page
) && !TestSetPageLocked(old_page
)) {
1296 int reuse
= can_share_swap_page(old_page
);
1297 unlock_page(old_page
);
1299 flush_cache_page(vma
, address
, pfn
);
1300 entry
= pte_mkyoung(orig_pte
);
1301 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1302 ptep_set_access_flags(vma
, address
, page_table
, entry
, 1);
1303 update_mmu_cache(vma
, address
, entry
);
1304 lazy_mmu_prot_update(entry
);
1305 ret
|= VM_FAULT_WRITE
;
1311 * Ok, we need to copy. Oh, well..
1313 page_cache_get(old_page
);
1314 pte_unmap_unlock(page_table
, ptl
);
1316 if (unlikely(anon_vma_prepare(vma
)))
1318 if (old_page
== ZERO_PAGE(address
)) {
1319 new_page
= alloc_zeroed_user_highpage(vma
, address
);
1323 new_page
= alloc_page_vma(GFP_HIGHUSER
, vma
, address
);
1326 copy_user_highpage(new_page
, old_page
, address
);
1330 * Re-check the pte - we dropped the lock
1332 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1333 if (likely(pte_same(*page_table
, orig_pte
))) {
1334 page_remove_rmap(old_page
);
1335 if (!PageAnon(old_page
)) {
1336 inc_mm_counter(mm
, anon_rss
);
1337 dec_mm_counter(mm
, file_rss
);
1339 flush_cache_page(vma
, address
, pfn
);
1340 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
1341 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1342 ptep_establish(vma
, address
, page_table
, entry
);
1343 update_mmu_cache(vma
, address
, entry
);
1344 lazy_mmu_prot_update(entry
);
1345 lru_cache_add_active(new_page
);
1346 page_add_anon_rmap(new_page
, vma
, address
);
1348 /* Free the old page.. */
1349 new_page
= old_page
;
1350 ret
|= VM_FAULT_WRITE
;
1352 page_cache_release(new_page
);
1353 page_cache_release(old_page
);
1355 pte_unmap_unlock(page_table
, ptl
);
1358 page_cache_release(old_page
);
1359 return VM_FAULT_OOM
;
1363 * Helper functions for unmap_mapping_range().
1365 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1367 * We have to restart searching the prio_tree whenever we drop the lock,
1368 * since the iterator is only valid while the lock is held, and anyway
1369 * a later vma might be split and reinserted earlier while lock dropped.
1371 * The list of nonlinear vmas could be handled more efficiently, using
1372 * a placeholder, but handle it in the same way until a need is shown.
1373 * It is important to search the prio_tree before nonlinear list: a vma
1374 * may become nonlinear and be shifted from prio_tree to nonlinear list
1375 * while the lock is dropped; but never shifted from list to prio_tree.
1377 * In order to make forward progress despite restarting the search,
1378 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1379 * quickly skip it next time around. Since the prio_tree search only
1380 * shows us those vmas affected by unmapping the range in question, we
1381 * can't efficiently keep all vmas in step with mapping->truncate_count:
1382 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1383 * mapping->truncate_count and vma->vm_truncate_count are protected by
1386 * In order to make forward progress despite repeatedly restarting some
1387 * large vma, note the restart_addr from unmap_vmas when it breaks out:
1388 * and restart from that address when we reach that vma again. It might
1389 * have been split or merged, shrunk or extended, but never shifted: so
1390 * restart_addr remains valid so long as it remains in the vma's range.
1391 * unmap_mapping_range forces truncate_count to leap over page-aligned
1392 * values so we can save vma's restart_addr in its truncate_count field.
1394 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1396 static void reset_vma_truncate_counts(struct address_space
*mapping
)
1398 struct vm_area_struct
*vma
;
1399 struct prio_tree_iter iter
;
1401 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, 0, ULONG_MAX
)
1402 vma
->vm_truncate_count
= 0;
1403 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
1404 vma
->vm_truncate_count
= 0;
1407 static int unmap_mapping_range_vma(struct vm_area_struct
*vma
,
1408 unsigned long start_addr
, unsigned long end_addr
,
1409 struct zap_details
*details
)
1411 unsigned long restart_addr
;
1415 restart_addr
= vma
->vm_truncate_count
;
1416 if (is_restart_addr(restart_addr
) && start_addr
< restart_addr
) {
1417 start_addr
= restart_addr
;
1418 if (start_addr
>= end_addr
) {
1419 /* Top of vma has been split off since last time */
1420 vma
->vm_truncate_count
= details
->truncate_count
;
1425 restart_addr
= zap_page_range(vma
, start_addr
,
1426 end_addr
- start_addr
, details
);
1427 need_break
= need_resched() ||
1428 need_lockbreak(details
->i_mmap_lock
);
1430 if (restart_addr
>= end_addr
) {
1431 /* We have now completed this vma: mark it so */
1432 vma
->vm_truncate_count
= details
->truncate_count
;
1436 /* Note restart_addr in vma's truncate_count field */
1437 vma
->vm_truncate_count
= restart_addr
;
1442 spin_unlock(details
->i_mmap_lock
);
1444 spin_lock(details
->i_mmap_lock
);
1448 static inline void unmap_mapping_range_tree(struct prio_tree_root
*root
,
1449 struct zap_details
*details
)
1451 struct vm_area_struct
*vma
;
1452 struct prio_tree_iter iter
;
1453 pgoff_t vba
, vea
, zba
, zea
;
1456 vma_prio_tree_foreach(vma
, &iter
, root
,
1457 details
->first_index
, details
->last_index
) {
1458 /* Skip quickly over those we have already dealt with */
1459 if (vma
->vm_truncate_count
== details
->truncate_count
)
1462 vba
= vma
->vm_pgoff
;
1463 vea
= vba
+ ((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) - 1;
1464 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1465 zba
= details
->first_index
;
1468 zea
= details
->last_index
;
1472 if (unmap_mapping_range_vma(vma
,
1473 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
1474 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
1480 static inline void unmap_mapping_range_list(struct list_head
*head
,
1481 struct zap_details
*details
)
1483 struct vm_area_struct
*vma
;
1486 * In nonlinear VMAs there is no correspondence between virtual address
1487 * offset and file offset. So we must perform an exhaustive search
1488 * across *all* the pages in each nonlinear VMA, not just the pages
1489 * whose virtual address lies outside the file truncation point.
1492 list_for_each_entry(vma
, head
, shared
.vm_set
.list
) {
1493 /* Skip quickly over those we have already dealt with */
1494 if (vma
->vm_truncate_count
== details
->truncate_count
)
1496 details
->nonlinear_vma
= vma
;
1497 if (unmap_mapping_range_vma(vma
, vma
->vm_start
,
1498 vma
->vm_end
, details
) < 0)
1504 * unmap_mapping_range - unmap the portion of all mmaps
1505 * in the specified address_space corresponding to the specified
1506 * page range in the underlying file.
1507 * @mapping: the address space containing mmaps to be unmapped.
1508 * @holebegin: byte in first page to unmap, relative to the start of
1509 * the underlying file. This will be rounded down to a PAGE_SIZE
1510 * boundary. Note that this is different from vmtruncate(), which
1511 * must keep the partial page. In contrast, we must get rid of
1513 * @holelen: size of prospective hole in bytes. This will be rounded
1514 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
1516 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1517 * but 0 when invalidating pagecache, don't throw away private data.
1519 void unmap_mapping_range(struct address_space
*mapping
,
1520 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
1522 struct zap_details details
;
1523 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
1524 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1526 /* Check for overflow. */
1527 if (sizeof(holelen
) > sizeof(hlen
)) {
1529 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1530 if (holeend
& ~(long long)ULONG_MAX
)
1531 hlen
= ULONG_MAX
- hba
+ 1;
1534 details
.check_mapping
= even_cows
? NULL
: mapping
;
1535 details
.nonlinear_vma
= NULL
;
1536 details
.first_index
= hba
;
1537 details
.last_index
= hba
+ hlen
- 1;
1538 if (details
.last_index
< details
.first_index
)
1539 details
.last_index
= ULONG_MAX
;
1540 details
.i_mmap_lock
= &mapping
->i_mmap_lock
;
1542 spin_lock(&mapping
->i_mmap_lock
);
1544 /* serialize i_size write against truncate_count write */
1546 /* Protect against page faults, and endless unmapping loops */
1547 mapping
->truncate_count
++;
1549 * For archs where spin_lock has inclusive semantics like ia64
1550 * this smp_mb() will prevent to read pagetable contents
1551 * before the truncate_count increment is visible to
1555 if (unlikely(is_restart_addr(mapping
->truncate_count
))) {
1556 if (mapping
->truncate_count
== 0)
1557 reset_vma_truncate_counts(mapping
);
1558 mapping
->truncate_count
++;
1560 details
.truncate_count
= mapping
->truncate_count
;
1562 if (unlikely(!prio_tree_empty(&mapping
->i_mmap
)))
1563 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
1564 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
1565 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
1566 spin_unlock(&mapping
->i_mmap_lock
);
1568 EXPORT_SYMBOL(unmap_mapping_range
);
1571 * Handle all mappings that got truncated by a "truncate()"
1574 * NOTE! We have to be ready to update the memory sharing
1575 * between the file and the memory map for a potential last
1576 * incomplete page. Ugly, but necessary.
1578 int vmtruncate(struct inode
* inode
, loff_t offset
)
1580 struct address_space
*mapping
= inode
->i_mapping
;
1581 unsigned long limit
;
1583 if (inode
->i_size
< offset
)
1586 * truncation of in-use swapfiles is disallowed - it would cause
1587 * subsequent swapout to scribble on the now-freed blocks.
1589 if (IS_SWAPFILE(inode
))
1591 i_size_write(inode
, offset
);
1592 unmap_mapping_range(mapping
, offset
+ PAGE_SIZE
- 1, 0, 1);
1593 truncate_inode_pages(mapping
, offset
);
1597 limit
= current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
;
1598 if (limit
!= RLIM_INFINITY
&& offset
> limit
)
1600 if (offset
> inode
->i_sb
->s_maxbytes
)
1602 i_size_write(inode
, offset
);
1605 if (inode
->i_op
&& inode
->i_op
->truncate
)
1606 inode
->i_op
->truncate(inode
);
1609 send_sig(SIGXFSZ
, current
, 0);
1616 EXPORT_SYMBOL(vmtruncate
);
1619 * Primitive swap readahead code. We simply read an aligned block of
1620 * (1 << page_cluster) entries in the swap area. This method is chosen
1621 * because it doesn't cost us any seek time. We also make sure to queue
1622 * the 'original' request together with the readahead ones...
1624 * This has been extended to use the NUMA policies from the mm triggering
1627 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1629 void swapin_readahead(swp_entry_t entry
, unsigned long addr
,struct vm_area_struct
*vma
)
1632 struct vm_area_struct
*next_vma
= vma
? vma
->vm_next
: NULL
;
1635 struct page
*new_page
;
1636 unsigned long offset
;
1639 * Get the number of handles we should do readahead io to.
1641 num
= valid_swaphandles(entry
, &offset
);
1642 for (i
= 0; i
< num
; offset
++, i
++) {
1643 /* Ok, do the async read-ahead now */
1644 new_page
= read_swap_cache_async(swp_entry(swp_type(entry
),
1645 offset
), vma
, addr
);
1648 page_cache_release(new_page
);
1651 * Find the next applicable VMA for the NUMA policy.
1657 if (addr
>= vma
->vm_end
) {
1659 next_vma
= vma
? vma
->vm_next
: NULL
;
1661 if (vma
&& addr
< vma
->vm_start
)
1664 if (next_vma
&& addr
>= next_vma
->vm_start
) {
1666 next_vma
= vma
->vm_next
;
1671 lru_add_drain(); /* Push any new pages onto the LRU now */
1675 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1676 * but allow concurrent faults), and pte mapped but not yet locked.
1677 * We return with mmap_sem still held, but pte unmapped and unlocked.
1679 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1680 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1681 int write_access
, pte_t orig_pte
)
1687 int ret
= VM_FAULT_MINOR
;
1689 if (!pte_unmap_same(mm
, page_table
, orig_pte
))
1692 entry
= pte_to_swp_entry(orig_pte
);
1693 page
= lookup_swap_cache(entry
);
1695 swapin_readahead(entry
, address
, vma
);
1696 page
= read_swap_cache_async(entry
, vma
, address
);
1699 * Back out if somebody else faulted in this pte
1700 * while we released the pte lock.
1702 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1703 if (likely(pte_same(*page_table
, orig_pte
)))
1708 /* Had to read the page from swap area: Major fault */
1709 ret
= VM_FAULT_MAJOR
;
1710 inc_page_state(pgmajfault
);
1714 mark_page_accessed(page
);
1718 * Back out if somebody else already faulted in this pte.
1720 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1721 if (unlikely(!pte_same(*page_table
, orig_pte
)))
1724 if (unlikely(!PageUptodate(page
))) {
1725 ret
= VM_FAULT_SIGBUS
;
1729 /* The page isn't present yet, go ahead with the fault. */
1731 inc_mm_counter(mm
, anon_rss
);
1732 pte
= mk_pte(page
, vma
->vm_page_prot
);
1733 if (write_access
&& can_share_swap_page(page
)) {
1734 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
1738 flush_icache_page(vma
, page
);
1739 set_pte_at(mm
, address
, page_table
, pte
);
1740 page_add_anon_rmap(page
, vma
, address
);
1744 remove_exclusive_swap_page(page
);
1748 if (do_wp_page(mm
, vma
, address
,
1749 page_table
, pmd
, ptl
, pte
) == VM_FAULT_OOM
)
1754 /* No need to invalidate - it was non-present before */
1755 update_mmu_cache(vma
, address
, pte
);
1756 lazy_mmu_prot_update(pte
);
1758 pte_unmap_unlock(page_table
, ptl
);
1762 pte_unmap_unlock(page_table
, ptl
);
1764 page_cache_release(page
);
1769 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1770 * but allow concurrent faults), and pte mapped but not yet locked.
1771 * We return with mmap_sem still held, but pte unmapped and unlocked.
1773 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1774 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1782 /* Allocate our own private page. */
1783 pte_unmap(page_table
);
1785 if (unlikely(anon_vma_prepare(vma
)))
1787 page
= alloc_zeroed_user_highpage(vma
, address
);
1791 entry
= mk_pte(page
, vma
->vm_page_prot
);
1792 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1794 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1795 if (!pte_none(*page_table
))
1797 inc_mm_counter(mm
, anon_rss
);
1798 lru_cache_add_active(page
);
1799 SetPageReferenced(page
);
1800 page_add_anon_rmap(page
, vma
, address
);
1802 /* Map the ZERO_PAGE - vm_page_prot is readonly */
1803 page
= ZERO_PAGE(address
);
1804 page_cache_get(page
);
1805 entry
= mk_pte(page
, vma
->vm_page_prot
);
1807 ptl
= &mm
->page_table_lock
;
1809 if (!pte_none(*page_table
))
1811 inc_mm_counter(mm
, file_rss
);
1812 page_add_file_rmap(page
);
1815 set_pte_at(mm
, address
, page_table
, entry
);
1817 /* No need to invalidate - it was non-present before */
1818 update_mmu_cache(vma
, address
, entry
);
1819 lazy_mmu_prot_update(entry
);
1821 pte_unmap_unlock(page_table
, ptl
);
1822 return VM_FAULT_MINOR
;
1824 page_cache_release(page
);
1827 return VM_FAULT_OOM
;
1831 * do_no_page() tries to create a new page mapping. It aggressively
1832 * tries to share with existing pages, but makes a separate copy if
1833 * the "write_access" parameter is true in order to avoid the next
1836 * As this is called only for pages that do not currently exist, we
1837 * do not need to flush old virtual caches or the TLB.
1839 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1840 * but allow concurrent faults), and pte mapped but not yet locked.
1841 * We return with mmap_sem still held, but pte unmapped and unlocked.
1843 static int do_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1844 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1848 struct page
*new_page
;
1849 struct address_space
*mapping
= NULL
;
1851 unsigned int sequence
= 0;
1852 int ret
= VM_FAULT_MINOR
;
1855 pte_unmap(page_table
);
1858 mapping
= vma
->vm_file
->f_mapping
;
1859 sequence
= mapping
->truncate_count
;
1860 smp_rmb(); /* serializes i_size against truncate_count */
1863 new_page
= vma
->vm_ops
->nopage(vma
, address
& PAGE_MASK
, &ret
);
1865 * No smp_rmb is needed here as long as there's a full
1866 * spin_lock/unlock sequence inside the ->nopage callback
1867 * (for the pagecache lookup) that acts as an implicit
1868 * smp_mb() and prevents the i_size read to happen
1869 * after the next truncate_count read.
1872 /* no page was available -- either SIGBUS or OOM */
1873 if (new_page
== NOPAGE_SIGBUS
)
1874 return VM_FAULT_SIGBUS
;
1875 if (new_page
== NOPAGE_OOM
)
1876 return VM_FAULT_OOM
;
1879 * Should we do an early C-O-W break?
1881 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
1884 if (unlikely(anon_vma_prepare(vma
)))
1886 page
= alloc_page_vma(GFP_HIGHUSER
, vma
, address
);
1889 copy_user_highpage(page
, new_page
, address
);
1890 page_cache_release(new_page
);
1895 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1897 * For a file-backed vma, someone could have truncated or otherwise
1898 * invalidated this page. If unmap_mapping_range got called,
1899 * retry getting the page.
1901 if (mapping
&& unlikely(sequence
!= mapping
->truncate_count
)) {
1902 pte_unmap_unlock(page_table
, ptl
);
1903 page_cache_release(new_page
);
1905 sequence
= mapping
->truncate_count
;
1911 * This silly early PAGE_DIRTY setting removes a race
1912 * due to the bad i386 page protection. But it's valid
1913 * for other architectures too.
1915 * Note that if write_access is true, we either now have
1916 * an exclusive copy of the page, or this is a shared mapping,
1917 * so we can make it writable and dirty to avoid having to
1918 * handle that later.
1920 /* Only go through if we didn't race with anybody else... */
1921 if (pte_none(*page_table
)) {
1922 flush_icache_page(vma
, new_page
);
1923 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
1925 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1926 set_pte_at(mm
, address
, page_table
, entry
);
1928 inc_mm_counter(mm
, anon_rss
);
1929 lru_cache_add_active(new_page
);
1930 page_add_anon_rmap(new_page
, vma
, address
);
1931 } else if (!(vma
->vm_flags
& VM_RESERVED
)) {
1932 inc_mm_counter(mm
, file_rss
);
1933 page_add_file_rmap(new_page
);
1936 /* One of our sibling threads was faster, back out. */
1937 page_cache_release(new_page
);
1941 /* no need to invalidate: a not-present page shouldn't be cached */
1942 update_mmu_cache(vma
, address
, entry
);
1943 lazy_mmu_prot_update(entry
);
1945 pte_unmap_unlock(page_table
, ptl
);
1948 page_cache_release(new_page
);
1949 return VM_FAULT_OOM
;
1953 * Fault of a previously existing named mapping. Repopulate the pte
1954 * from the encoded file_pte if possible. This enables swappable
1957 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1958 * but allow concurrent faults), and pte mapped but not yet locked.
1959 * We return with mmap_sem still held, but pte unmapped and unlocked.
1961 static int do_file_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1962 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1963 int write_access
, pte_t orig_pte
)
1968 if (!pte_unmap_same(mm
, page_table
, orig_pte
))
1969 return VM_FAULT_MINOR
;
1971 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
))) {
1973 * Page table corrupted: show pte and kill process.
1975 print_bad_pte(vma
, orig_pte
, address
);
1976 return VM_FAULT_OOM
;
1978 /* We can then assume vm->vm_ops && vma->vm_ops->populate */
1980 pgoff
= pte_to_pgoff(orig_pte
);
1981 err
= vma
->vm_ops
->populate(vma
, address
& PAGE_MASK
, PAGE_SIZE
,
1982 vma
->vm_page_prot
, pgoff
, 0);
1984 return VM_FAULT_OOM
;
1986 return VM_FAULT_SIGBUS
;
1987 return VM_FAULT_MAJOR
;
1991 * These routines also need to handle stuff like marking pages dirty
1992 * and/or accessed for architectures that don't do it in hardware (most
1993 * RISC architectures). The early dirtying is also good on the i386.
1995 * There is also a hook called "update_mmu_cache()" that architectures
1996 * with external mmu caches can use to update those (ie the Sparc or
1997 * PowerPC hashed page tables that act as extended TLBs).
1999 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2000 * but allow concurrent faults), and pte mapped but not yet locked.
2001 * We return with mmap_sem still held, but pte unmapped and unlocked.
2003 static inline int handle_pte_fault(struct mm_struct
*mm
,
2004 struct vm_area_struct
*vma
, unsigned long address
,
2005 pte_t
*pte
, pmd_t
*pmd
, int write_access
)
2011 if (!pte_present(entry
)) {
2012 if (pte_none(entry
)) {
2013 if (!vma
->vm_ops
|| !vma
->vm_ops
->nopage
)
2014 return do_anonymous_page(mm
, vma
, address
,
2015 pte
, pmd
, write_access
);
2016 return do_no_page(mm
, vma
, address
,
2017 pte
, pmd
, write_access
);
2019 if (pte_file(entry
))
2020 return do_file_page(mm
, vma
, address
,
2021 pte
, pmd
, write_access
, entry
);
2022 return do_swap_page(mm
, vma
, address
,
2023 pte
, pmd
, write_access
, entry
);
2026 ptl
= &mm
->page_table_lock
;
2028 if (unlikely(!pte_same(*pte
, entry
)))
2031 if (!pte_write(entry
))
2032 return do_wp_page(mm
, vma
, address
,
2033 pte
, pmd
, ptl
, entry
);
2034 entry
= pte_mkdirty(entry
);
2036 entry
= pte_mkyoung(entry
);
2037 ptep_set_access_flags(vma
, address
, pte
, entry
, write_access
);
2038 update_mmu_cache(vma
, address
, entry
);
2039 lazy_mmu_prot_update(entry
);
2041 pte_unmap_unlock(pte
, ptl
);
2042 return VM_FAULT_MINOR
;
2046 * By the time we get here, we already hold the mm semaphore
2048 int __handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2049 unsigned long address
, int write_access
)
2056 __set_current_state(TASK_RUNNING
);
2058 inc_page_state(pgfault
);
2060 if (unlikely(is_vm_hugetlb_page(vma
)))
2061 return hugetlb_fault(mm
, vma
, address
, write_access
);
2063 pgd
= pgd_offset(mm
, address
);
2064 pud
= pud_alloc(mm
, pgd
, address
);
2066 return VM_FAULT_OOM
;
2067 pmd
= pmd_alloc(mm
, pud
, address
);
2069 return VM_FAULT_OOM
;
2070 pte
= pte_alloc_map(mm
, pmd
, address
);
2072 return VM_FAULT_OOM
;
2074 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, write_access
);
2077 #ifndef __PAGETABLE_PUD_FOLDED
2079 * Allocate page upper directory.
2080 * We've already handled the fast-path in-line.
2082 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
2084 pud_t
*new = pud_alloc_one(mm
, address
);
2088 spin_lock(&mm
->page_table_lock
);
2089 if (pgd_present(*pgd
)) /* Another has populated it */
2092 pgd_populate(mm
, pgd
, new);
2093 spin_unlock(&mm
->page_table_lock
);
2096 #endif /* __PAGETABLE_PUD_FOLDED */
2098 #ifndef __PAGETABLE_PMD_FOLDED
2100 * Allocate page middle directory.
2101 * We've already handled the fast-path in-line.
2103 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
2105 pmd_t
*new = pmd_alloc_one(mm
, address
);
2109 spin_lock(&mm
->page_table_lock
);
2110 #ifndef __ARCH_HAS_4LEVEL_HACK
2111 if (pud_present(*pud
)) /* Another has populated it */
2114 pud_populate(mm
, pud
, new);
2116 if (pgd_present(*pud
)) /* Another has populated it */
2119 pgd_populate(mm
, pud
, new);
2120 #endif /* __ARCH_HAS_4LEVEL_HACK */
2121 spin_unlock(&mm
->page_table_lock
);
2124 #endif /* __PAGETABLE_PMD_FOLDED */
2126 int make_pages_present(unsigned long addr
, unsigned long end
)
2128 int ret
, len
, write
;
2129 struct vm_area_struct
* vma
;
2131 vma
= find_vma(current
->mm
, addr
);
2134 write
= (vma
->vm_flags
& VM_WRITE
) != 0;
2137 if (end
> vma
->vm_end
)
2139 len
= (end
+PAGE_SIZE
-1)/PAGE_SIZE
-addr
/PAGE_SIZE
;
2140 ret
= get_user_pages(current
, current
->mm
, addr
,
2141 len
, write
, 0, NULL
, NULL
);
2144 return ret
== len
? 0 : -1;
2148 * Map a vmalloc()-space virtual address to the physical page.
2150 struct page
* vmalloc_to_page(void * vmalloc_addr
)
2152 unsigned long addr
= (unsigned long) vmalloc_addr
;
2153 struct page
*page
= NULL
;
2154 pgd_t
*pgd
= pgd_offset_k(addr
);
2159 if (!pgd_none(*pgd
)) {
2160 pud
= pud_offset(pgd
, addr
);
2161 if (!pud_none(*pud
)) {
2162 pmd
= pmd_offset(pud
, addr
);
2163 if (!pmd_none(*pmd
)) {
2164 ptep
= pte_offset_map(pmd
, addr
);
2166 if (pte_present(pte
))
2167 page
= pte_page(pte
);
2175 EXPORT_SYMBOL(vmalloc_to_page
);
2178 * Map a vmalloc()-space virtual address to the physical page frame number.
2180 unsigned long vmalloc_to_pfn(void * vmalloc_addr
)
2182 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
2185 EXPORT_SYMBOL(vmalloc_to_pfn
);
2187 #if !defined(__HAVE_ARCH_GATE_AREA)
2189 #if defined(AT_SYSINFO_EHDR)
2190 static struct vm_area_struct gate_vma
;
2192 static int __init
gate_vma_init(void)
2194 gate_vma
.vm_mm
= NULL
;
2195 gate_vma
.vm_start
= FIXADDR_USER_START
;
2196 gate_vma
.vm_end
= FIXADDR_USER_END
;
2197 gate_vma
.vm_page_prot
= PAGE_READONLY
;
2198 gate_vma
.vm_flags
= VM_RESERVED
;
2201 __initcall(gate_vma_init
);
2204 struct vm_area_struct
*get_gate_vma(struct task_struct
*tsk
)
2206 #ifdef AT_SYSINFO_EHDR
2213 int in_gate_area_no_task(unsigned long addr
)
2215 #ifdef AT_SYSINFO_EHDR
2216 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
2222 #endif /* __HAVE_ARCH_GATE_AREA */