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
;
263 if (is_hugepage_only_range(vma
->vm_mm
, addr
, HPAGE_SIZE
)) {
264 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
265 floor
, next
? next
->vm_start
: ceiling
);
268 * Optimization: gather nearby vmas into one call down
270 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
271 && !is_hugepage_only_range(vma
->vm_mm
, next
->vm_start
,
276 free_pgd_range(tlb
, addr
, vma
->vm_end
,
277 floor
, next
? next
->vm_start
: ceiling
);
283 pte_t fastcall
*pte_alloc_map(struct mm_struct
*mm
, pmd_t
*pmd
,
284 unsigned long address
)
286 if (!pmd_present(*pmd
)) {
289 spin_unlock(&mm
->page_table_lock
);
290 new = pte_alloc_one(mm
, address
);
291 spin_lock(&mm
->page_table_lock
);
295 * Because we dropped the lock, we should re-check the
296 * entry, as somebody else could have populated it..
298 if (pmd_present(*pmd
)) {
303 inc_page_state(nr_page_table_pages
);
304 pmd_populate(mm
, pmd
, new);
307 return pte_offset_map(pmd
, address
);
310 pte_t fastcall
* pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
312 if (!pmd_present(*pmd
)) {
315 new = pte_alloc_one_kernel(&init_mm
, address
);
319 spin_lock(&init_mm
.page_table_lock
);
320 if (pmd_present(*pmd
))
321 pte_free_kernel(new);
323 pmd_populate_kernel(&init_mm
, pmd
, new);
324 spin_unlock(&init_mm
.page_table_lock
);
326 return pte_offset_kernel(pmd
, address
);
329 static inline void add_mm_rss(struct mm_struct
*mm
, int file_rss
, int anon_rss
)
332 add_mm_counter(mm
, file_rss
, file_rss
);
334 add_mm_counter(mm
, anon_rss
, anon_rss
);
338 * This function is called to print an error when a pte in a
339 * !VM_RESERVED region is found pointing to an invalid pfn (which
342 * The calling function must still handle the error.
344 void print_bad_pte(struct vm_area_struct
*vma
, pte_t pte
, unsigned long vaddr
)
346 printk(KERN_ERR
"Bad pte = %08llx, process = %s, "
347 "vm_flags = %lx, vaddr = %lx\n",
348 (long long)pte_val(pte
),
349 (vma
->vm_mm
== current
->mm
? current
->comm
: "???"),
350 vma
->vm_flags
, vaddr
);
355 * copy one vm_area from one task to the other. Assumes the page tables
356 * already present in the new task to be cleared in the whole range
357 * covered by this vma.
359 * dst->page_table_lock is held on entry and exit,
360 * but may be dropped within p[mg]d_alloc() and pte_alloc_map().
364 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
365 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
366 unsigned long addr
, int *rss
)
368 unsigned long vm_flags
= vma
->vm_flags
;
369 pte_t pte
= *src_pte
;
373 /* pte contains position in swap or file, so copy. */
374 if (unlikely(!pte_present(pte
))) {
375 if (!pte_file(pte
)) {
376 swap_duplicate(pte_to_swp_entry(pte
));
377 /* make sure dst_mm is on swapoff's mmlist. */
378 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
379 spin_lock(&mmlist_lock
);
380 list_add(&dst_mm
->mmlist
, &src_mm
->mmlist
);
381 spin_unlock(&mmlist_lock
);
387 /* If the region is VM_RESERVED, the mapping is not
388 * mapped via rmap - duplicate the pte as is.
390 if (vm_flags
& VM_RESERVED
)
394 /* If the pte points outside of valid memory but
395 * the region is not VM_RESERVED, we have a problem.
397 if (unlikely(!pfn_valid(pfn
))) {
398 print_bad_pte(vma
, pte
, addr
);
399 goto out_set_pte
; /* try to do something sane */
402 page
= pfn_to_page(pfn
);
405 * If it's a COW mapping, write protect it both
406 * in the parent and the child
408 if ((vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
) {
409 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
414 * If it's a shared mapping, mark it clean in
417 if (vm_flags
& VM_SHARED
)
418 pte
= pte_mkclean(pte
);
419 pte
= pte_mkold(pte
);
422 rss
[!!PageAnon(page
)]++;
425 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
428 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
429 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
430 unsigned long addr
, unsigned long end
)
432 pte_t
*src_pte
, *dst_pte
;
438 dst_pte
= pte_alloc_map(dst_mm
, dst_pmd
, addr
);
441 src_pte
= pte_offset_map_nested(src_pmd
, addr
);
443 spin_lock(&src_mm
->page_table_lock
);
446 * We are holding two locks at this point - either of them
447 * could generate latencies in another task on another CPU.
449 if (progress
>= 32) {
451 if (need_resched() ||
452 need_lockbreak(&src_mm
->page_table_lock
) ||
453 need_lockbreak(&dst_mm
->page_table_lock
))
456 if (pte_none(*src_pte
)) {
460 copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
, vma
, addr
, rss
);
462 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
463 spin_unlock(&src_mm
->page_table_lock
);
465 pte_unmap_nested(src_pte
- 1);
466 pte_unmap(dst_pte
- 1);
467 add_mm_rss(dst_mm
, rss
[0], rss
[1]);
468 cond_resched_lock(&dst_mm
->page_table_lock
);
474 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
475 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
476 unsigned long addr
, unsigned long end
)
478 pmd_t
*src_pmd
, *dst_pmd
;
481 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
484 src_pmd
= pmd_offset(src_pud
, addr
);
486 next
= pmd_addr_end(addr
, end
);
487 if (pmd_none_or_clear_bad(src_pmd
))
489 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
492 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
496 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
497 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
498 unsigned long addr
, unsigned long end
)
500 pud_t
*src_pud
, *dst_pud
;
503 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
506 src_pud
= pud_offset(src_pgd
, addr
);
508 next
= pud_addr_end(addr
, end
);
509 if (pud_none_or_clear_bad(src_pud
))
511 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
514 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
518 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
519 struct vm_area_struct
*vma
)
521 pgd_t
*src_pgd
, *dst_pgd
;
523 unsigned long addr
= vma
->vm_start
;
524 unsigned long end
= vma
->vm_end
;
527 * Don't copy ptes where a page fault will fill them correctly.
528 * Fork becomes much lighter when there are big shared or private
529 * readonly mappings. The tradeoff is that copy_page_range is more
530 * efficient than faulting.
532 if (!(vma
->vm_flags
& (VM_HUGETLB
|VM_NONLINEAR
|VM_RESERVED
))) {
537 if (is_vm_hugetlb_page(vma
))
538 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
540 dst_pgd
= pgd_offset(dst_mm
, addr
);
541 src_pgd
= pgd_offset(src_mm
, addr
);
543 next
= pgd_addr_end(addr
, end
);
544 if (pgd_none_or_clear_bad(src_pgd
))
546 if (copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
549 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
553 static void zap_pte_range(struct mmu_gather
*tlb
,
554 struct vm_area_struct
*vma
, pmd_t
*pmd
,
555 unsigned long addr
, unsigned long end
,
556 struct zap_details
*details
)
558 struct mm_struct
*mm
= tlb
->mm
;
563 pte
= pte_offset_map(pmd
, addr
);
568 if (pte_present(ptent
)) {
569 struct page
*page
= NULL
;
570 if (!(vma
->vm_flags
& VM_RESERVED
)) {
571 unsigned long pfn
= pte_pfn(ptent
);
572 if (unlikely(!pfn_valid(pfn
)))
573 print_bad_pte(vma
, ptent
, addr
);
575 page
= pfn_to_page(pfn
);
577 if (unlikely(details
) && page
) {
579 * unmap_shared_mapping_pages() wants to
580 * invalidate cache without truncating:
581 * unmap shared but keep private pages.
583 if (details
->check_mapping
&&
584 details
->check_mapping
!= page
->mapping
)
587 * Each page->index must be checked when
588 * invalidating or truncating nonlinear.
590 if (details
->nonlinear_vma
&&
591 (page
->index
< details
->first_index
||
592 page
->index
> details
->last_index
))
595 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
597 tlb_remove_tlb_entry(tlb
, pte
, addr
);
600 if (unlikely(details
) && details
->nonlinear_vma
601 && linear_page_index(details
->nonlinear_vma
,
602 addr
) != page
->index
)
603 set_pte_at(mm
, addr
, pte
,
604 pgoff_to_pte(page
->index
));
608 if (pte_dirty(ptent
))
609 set_page_dirty(page
);
610 if (pte_young(ptent
))
611 mark_page_accessed(page
);
614 page_remove_rmap(page
);
615 tlb_remove_page(tlb
, page
);
619 * If details->check_mapping, we leave swap entries;
620 * if details->nonlinear_vma, we leave file entries.
622 if (unlikely(details
))
624 if (!pte_file(ptent
))
625 free_swap_and_cache(pte_to_swp_entry(ptent
));
626 pte_clear_full(mm
, addr
, pte
, tlb
->fullmm
);
627 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
629 add_mm_rss(mm
, file_rss
, anon_rss
);
633 static inline void zap_pmd_range(struct mmu_gather
*tlb
,
634 struct vm_area_struct
*vma
, pud_t
*pud
,
635 unsigned long addr
, unsigned long end
,
636 struct zap_details
*details
)
641 pmd
= pmd_offset(pud
, addr
);
643 next
= pmd_addr_end(addr
, end
);
644 if (pmd_none_or_clear_bad(pmd
))
646 zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
647 } while (pmd
++, addr
= next
, addr
!= end
);
650 static inline void zap_pud_range(struct mmu_gather
*tlb
,
651 struct vm_area_struct
*vma
, pgd_t
*pgd
,
652 unsigned long addr
, unsigned long end
,
653 struct zap_details
*details
)
658 pud
= pud_offset(pgd
, addr
);
660 next
= pud_addr_end(addr
, end
);
661 if (pud_none_or_clear_bad(pud
))
663 zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
664 } while (pud
++, addr
= next
, addr
!= end
);
667 static void unmap_page_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
668 unsigned long addr
, unsigned long end
,
669 struct zap_details
*details
)
674 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
678 tlb_start_vma(tlb
, vma
);
679 pgd
= pgd_offset(vma
->vm_mm
, addr
);
681 next
= pgd_addr_end(addr
, end
);
682 if (pgd_none_or_clear_bad(pgd
))
684 zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
685 } while (pgd
++, addr
= next
, addr
!= end
);
686 tlb_end_vma(tlb
, vma
);
689 #ifdef CONFIG_PREEMPT
690 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
692 /* No preempt: go for improved straight-line efficiency */
693 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
697 * unmap_vmas - unmap a range of memory covered by a list of vma's
698 * @tlbp: address of the caller's struct mmu_gather
699 * @mm: the controlling mm_struct
700 * @vma: the starting vma
701 * @start_addr: virtual address at which to start unmapping
702 * @end_addr: virtual address at which to end unmapping
703 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
704 * @details: details of nonlinear truncation or shared cache invalidation
706 * Returns the end address of the unmapping (restart addr if interrupted).
708 * Unmap all pages in the vma list. Called under page_table_lock.
710 * We aim to not hold page_table_lock for too long (for scheduling latency
711 * reasons). So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
712 * return the ending mmu_gather to the caller.
714 * Only addresses between `start' and `end' will be unmapped.
716 * The VMA list must be sorted in ascending virtual address order.
718 * unmap_vmas() assumes that the caller will flush the whole unmapped address
719 * range after unmap_vmas() returns. So the only responsibility here is to
720 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
721 * drops the lock and schedules.
723 unsigned long unmap_vmas(struct mmu_gather
**tlbp
, struct mm_struct
*mm
,
724 struct vm_area_struct
*vma
, unsigned long start_addr
,
725 unsigned long end_addr
, unsigned long *nr_accounted
,
726 struct zap_details
*details
)
728 unsigned long zap_bytes
= ZAP_BLOCK_SIZE
;
729 unsigned long tlb_start
= 0; /* For tlb_finish_mmu */
730 int tlb_start_valid
= 0;
731 unsigned long start
= start_addr
;
732 spinlock_t
*i_mmap_lock
= details
? details
->i_mmap_lock
: NULL
;
733 int fullmm
= (*tlbp
)->fullmm
;
735 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
) {
738 start
= max(vma
->vm_start
, start_addr
);
739 if (start
>= vma
->vm_end
)
741 end
= min(vma
->vm_end
, end_addr
);
742 if (end
<= vma
->vm_start
)
745 if (vma
->vm_flags
& VM_ACCOUNT
)
746 *nr_accounted
+= (end
- start
) >> PAGE_SHIFT
;
748 while (start
!= end
) {
751 if (!tlb_start_valid
) {
756 if (is_vm_hugetlb_page(vma
)) {
758 unmap_hugepage_range(vma
, start
, end
);
760 block
= min(zap_bytes
, end
- start
);
761 unmap_page_range(*tlbp
, vma
, start
,
762 start
+ block
, details
);
767 if ((long)zap_bytes
> 0)
770 tlb_finish_mmu(*tlbp
, tlb_start
, start
);
772 if (need_resched() ||
773 need_lockbreak(&mm
->page_table_lock
) ||
774 (i_mmap_lock
&& need_lockbreak(i_mmap_lock
))) {
776 /* must reset count of rss freed */
777 *tlbp
= tlb_gather_mmu(mm
, fullmm
);
780 spin_unlock(&mm
->page_table_lock
);
782 spin_lock(&mm
->page_table_lock
);
785 *tlbp
= tlb_gather_mmu(mm
, fullmm
);
787 zap_bytes
= ZAP_BLOCK_SIZE
;
791 return start
; /* which is now the end (or restart) address */
795 * zap_page_range - remove user pages in a given range
796 * @vma: vm_area_struct holding the applicable pages
797 * @address: starting address of pages to zap
798 * @size: number of bytes to zap
799 * @details: details of nonlinear truncation or shared cache invalidation
801 unsigned long zap_page_range(struct vm_area_struct
*vma
, unsigned long address
,
802 unsigned long size
, struct zap_details
*details
)
804 struct mm_struct
*mm
= vma
->vm_mm
;
805 struct mmu_gather
*tlb
;
806 unsigned long end
= address
+ size
;
807 unsigned long nr_accounted
= 0;
809 if (is_vm_hugetlb_page(vma
)) {
810 zap_hugepage_range(vma
, address
, size
);
815 spin_lock(&mm
->page_table_lock
);
816 tlb
= tlb_gather_mmu(mm
, 0);
817 update_hiwater_rss(mm
);
818 end
= unmap_vmas(&tlb
, mm
, vma
, address
, end
, &nr_accounted
, details
);
819 tlb_finish_mmu(tlb
, address
, end
);
820 spin_unlock(&mm
->page_table_lock
);
825 * Do a quick page-table lookup for a single page.
826 * mm->page_table_lock must be held.
828 static struct page
*__follow_page(struct mm_struct
*mm
, unsigned long address
,
829 int read
, int write
, int accessed
)
838 page
= follow_huge_addr(mm
, address
, write
);
842 pgd
= pgd_offset(mm
, address
);
843 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
846 pud
= pud_offset(pgd
, address
);
847 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
850 pmd
= pmd_offset(pud
, address
);
851 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
854 return follow_huge_pmd(mm
, address
, pmd
, write
);
856 ptep
= pte_offset_map(pmd
, address
);
862 if (pte_present(pte
)) {
863 if (write
&& !pte_write(pte
))
865 if (read
&& !pte_read(pte
))
868 if (pfn_valid(pfn
)) {
869 page
= pfn_to_page(pfn
);
871 if (write
&& !pte_dirty(pte
) &&!PageDirty(page
))
872 set_page_dirty(page
);
873 mark_page_accessed(page
);
884 follow_page(struct mm_struct
*mm
, unsigned long address
, int write
)
886 return __follow_page(mm
, address
, 0, write
, 1);
890 * check_user_page_readable() can be called frm niterrupt context by oprofile,
891 * so we need to avoid taking any non-irq-safe locks
893 int check_user_page_readable(struct mm_struct
*mm
, unsigned long address
)
895 return __follow_page(mm
, address
, 1, 0, 0) != NULL
;
897 EXPORT_SYMBOL(check_user_page_readable
);
900 untouched_anonymous_page(struct mm_struct
* mm
, struct vm_area_struct
*vma
,
901 unsigned long address
)
907 /* Check if the vma is for an anonymous mapping. */
908 if (vma
->vm_ops
&& vma
->vm_ops
->nopage
)
911 /* Check if page directory entry exists. */
912 pgd
= pgd_offset(mm
, address
);
913 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
916 pud
= pud_offset(pgd
, address
);
917 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
920 /* Check if page middle directory entry exists. */
921 pmd
= pmd_offset(pud
, address
);
922 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
925 /* There is a pte slot for 'address' in 'mm'. */
929 int get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
930 unsigned long start
, int len
, int write
, int force
,
931 struct page
**pages
, struct vm_area_struct
**vmas
)
937 * Require read or write permissions.
938 * If 'force' is set, we only require the "MAY" flags.
940 flags
= write
? (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
941 flags
&= force
? (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
945 struct vm_area_struct
* vma
;
947 vma
= find_extend_vma(mm
, start
);
948 if (!vma
&& in_gate_area(tsk
, start
)) {
949 unsigned long pg
= start
& PAGE_MASK
;
950 struct vm_area_struct
*gate_vma
= get_gate_vma(tsk
);
955 if (write
) /* user gate pages are read-only */
956 return i
? : -EFAULT
;
958 pgd
= pgd_offset_k(pg
);
960 pgd
= pgd_offset_gate(mm
, pg
);
961 BUG_ON(pgd_none(*pgd
));
962 pud
= pud_offset(pgd
, pg
);
963 BUG_ON(pud_none(*pud
));
964 pmd
= pmd_offset(pud
, pg
);
966 return i
? : -EFAULT
;
967 pte
= pte_offset_map(pmd
, pg
);
968 if (pte_none(*pte
)) {
970 return i
? : -EFAULT
;
973 pages
[i
] = pte_page(*pte
);
985 if (!vma
|| (vma
->vm_flags
& (VM_IO
| VM_RESERVED
))
986 || !(flags
& vma
->vm_flags
))
987 return i
? : -EFAULT
;
989 if (is_vm_hugetlb_page(vma
)) {
990 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
994 spin_lock(&mm
->page_table_lock
);
996 int write_access
= write
;
999 cond_resched_lock(&mm
->page_table_lock
);
1000 while (!(page
= follow_page(mm
, start
, write_access
))) {
1004 * Shortcut for anonymous pages. We don't want
1005 * to force the creation of pages tables for
1006 * insanely big anonymously mapped areas that
1007 * nobody touched so far. This is important
1008 * for doing a core dump for these mappings.
1010 if (!write
&& untouched_anonymous_page(mm
,vma
,start
)) {
1011 page
= ZERO_PAGE(start
);
1014 spin_unlock(&mm
->page_table_lock
);
1015 ret
= __handle_mm_fault(mm
, vma
, start
, write_access
);
1018 * The VM_FAULT_WRITE bit tells us that do_wp_page has
1019 * broken COW when necessary, even if maybe_mkwrite
1020 * decided not to set pte_write. We can thus safely do
1021 * subsequent page lookups as if they were reads.
1023 if (ret
& VM_FAULT_WRITE
)
1026 switch (ret
& ~VM_FAULT_WRITE
) {
1027 case VM_FAULT_MINOR
:
1030 case VM_FAULT_MAJOR
:
1033 case VM_FAULT_SIGBUS
:
1034 return i
? i
: -EFAULT
;
1036 return i
? i
: -ENOMEM
;
1040 spin_lock(&mm
->page_table_lock
);
1044 flush_dcache_page(page
);
1045 page_cache_get(page
);
1052 } while (len
&& start
< vma
->vm_end
);
1053 spin_unlock(&mm
->page_table_lock
);
1057 EXPORT_SYMBOL(get_user_pages
);
1059 static int zeromap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1060 unsigned long addr
, unsigned long end
, pgprot_t prot
)
1064 pte
= pte_alloc_map(mm
, pmd
, addr
);
1068 struct page
*page
= ZERO_PAGE(addr
);
1069 pte_t zero_pte
= pte_wrprotect(mk_pte(page
, prot
));
1070 page_cache_get(page
);
1071 page_add_file_rmap(page
);
1072 inc_mm_counter(mm
, file_rss
);
1073 BUG_ON(!pte_none(*pte
));
1074 set_pte_at(mm
, addr
, pte
, zero_pte
);
1075 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1080 static inline int zeromap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1081 unsigned long addr
, unsigned long end
, pgprot_t prot
)
1086 pmd
= pmd_alloc(mm
, pud
, addr
);
1090 next
= pmd_addr_end(addr
, end
);
1091 if (zeromap_pte_range(mm
, pmd
, addr
, next
, prot
))
1093 } while (pmd
++, addr
= next
, addr
!= end
);
1097 static inline int zeromap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1098 unsigned long addr
, unsigned long end
, pgprot_t prot
)
1103 pud
= pud_alloc(mm
, pgd
, addr
);
1107 next
= pud_addr_end(addr
, end
);
1108 if (zeromap_pmd_range(mm
, pud
, addr
, next
, prot
))
1110 } while (pud
++, addr
= next
, addr
!= end
);
1114 int zeromap_page_range(struct vm_area_struct
*vma
,
1115 unsigned long addr
, unsigned long size
, pgprot_t prot
)
1119 unsigned long end
= addr
+ size
;
1120 struct mm_struct
*mm
= vma
->vm_mm
;
1123 BUG_ON(addr
>= end
);
1124 pgd
= pgd_offset(mm
, addr
);
1125 flush_cache_range(vma
, addr
, end
);
1126 spin_lock(&mm
->page_table_lock
);
1128 next
= pgd_addr_end(addr
, end
);
1129 err
= zeromap_pud_range(mm
, pgd
, addr
, next
, prot
);
1132 } while (pgd
++, addr
= next
, addr
!= end
);
1133 spin_unlock(&mm
->page_table_lock
);
1138 * maps a range of physical memory into the requested pages. the old
1139 * mappings are removed. any references to nonexistent pages results
1140 * in null mappings (currently treated as "copy-on-access")
1142 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1143 unsigned long addr
, unsigned long end
,
1144 unsigned long pfn
, pgprot_t prot
)
1148 pte
= pte_alloc_map(mm
, pmd
, addr
);
1152 BUG_ON(!pte_none(*pte
));
1153 set_pte_at(mm
, addr
, pte
, pfn_pte(pfn
, prot
));
1155 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1160 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1161 unsigned long addr
, unsigned long end
,
1162 unsigned long pfn
, pgprot_t prot
)
1167 pfn
-= addr
>> PAGE_SHIFT
;
1168 pmd
= pmd_alloc(mm
, pud
, addr
);
1172 next
= pmd_addr_end(addr
, end
);
1173 if (remap_pte_range(mm
, pmd
, addr
, next
,
1174 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1176 } while (pmd
++, addr
= next
, addr
!= end
);
1180 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1181 unsigned long addr
, unsigned long end
,
1182 unsigned long pfn
, pgprot_t prot
)
1187 pfn
-= addr
>> PAGE_SHIFT
;
1188 pud
= pud_alloc(mm
, pgd
, addr
);
1192 next
= pud_addr_end(addr
, end
);
1193 if (remap_pmd_range(mm
, pud
, addr
, next
,
1194 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1196 } while (pud
++, addr
= next
, addr
!= end
);
1200 /* Note: this is only safe if the mm semaphore is held when called. */
1201 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1202 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1206 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1207 struct mm_struct
*mm
= vma
->vm_mm
;
1211 * Physically remapped pages are special. Tell the
1212 * rest of the world about it:
1213 * VM_IO tells people not to look at these pages
1214 * (accesses can have side effects).
1215 * VM_RESERVED tells the core MM not to "manage" these pages
1216 * (e.g. refcount, mapcount, try to swap them out).
1218 vma
->vm_flags
|= VM_IO
| VM_RESERVED
;
1220 BUG_ON(addr
>= end
);
1221 pfn
-= addr
>> PAGE_SHIFT
;
1222 pgd
= pgd_offset(mm
, addr
);
1223 flush_cache_range(vma
, addr
, end
);
1224 spin_lock(&mm
->page_table_lock
);
1226 next
= pgd_addr_end(addr
, end
);
1227 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1228 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1231 } while (pgd
++, addr
= next
, addr
!= end
);
1232 spin_unlock(&mm
->page_table_lock
);
1235 EXPORT_SYMBOL(remap_pfn_range
);
1238 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1239 * servicing faults for write access. In the normal case, do always want
1240 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1241 * that do not have writing enabled, when used by access_process_vm.
1243 static inline pte_t
maybe_mkwrite(pte_t pte
, struct vm_area_struct
*vma
)
1245 if (likely(vma
->vm_flags
& VM_WRITE
))
1246 pte
= pte_mkwrite(pte
);
1251 * This routine handles present pages, when users try to write
1252 * to a shared page. It is done by copying the page to a new address
1253 * and decrementing the shared-page counter for the old page.
1255 * Note that this routine assumes that the protection checks have been
1256 * done by the caller (the low-level page fault routine in most cases).
1257 * Thus we can safely just mark it writable once we've done any necessary
1260 * We also mark the page dirty at this point even though the page will
1261 * change only once the write actually happens. This avoids a few races,
1262 * and potentially makes it more efficient.
1264 * We hold the mm semaphore and the page_table_lock on entry and exit
1265 * with the page_table_lock released.
1267 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1268 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1271 struct page
*old_page
, *new_page
;
1272 unsigned long pfn
= pte_pfn(orig_pte
);
1274 int ret
= VM_FAULT_MINOR
;
1276 BUG_ON(vma
->vm_flags
& VM_RESERVED
);
1278 if (unlikely(!pfn_valid(pfn
))) {
1280 * Page table corrupted: show pte and kill process.
1282 print_bad_pte(vma
, orig_pte
, address
);
1286 old_page
= pfn_to_page(pfn
);
1288 if (PageAnon(old_page
) && !TestSetPageLocked(old_page
)) {
1289 int reuse
= can_share_swap_page(old_page
);
1290 unlock_page(old_page
);
1292 flush_cache_page(vma
, address
, pfn
);
1293 entry
= pte_mkyoung(orig_pte
);
1294 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1295 ptep_set_access_flags(vma
, address
, page_table
, entry
, 1);
1296 update_mmu_cache(vma
, address
, entry
);
1297 lazy_mmu_prot_update(entry
);
1298 ret
|= VM_FAULT_WRITE
;
1304 * Ok, we need to copy. Oh, well..
1306 page_cache_get(old_page
);
1307 pte_unmap(page_table
);
1308 spin_unlock(&mm
->page_table_lock
);
1310 if (unlikely(anon_vma_prepare(vma
)))
1312 if (old_page
== ZERO_PAGE(address
)) {
1313 new_page
= alloc_zeroed_user_highpage(vma
, address
);
1317 new_page
= alloc_page_vma(GFP_HIGHUSER
, vma
, address
);
1320 copy_user_highpage(new_page
, old_page
, address
);
1324 * Re-check the pte - we dropped the lock
1326 spin_lock(&mm
->page_table_lock
);
1327 page_table
= pte_offset_map(pmd
, address
);
1328 if (likely(pte_same(*page_table
, orig_pte
))) {
1329 page_remove_rmap(old_page
);
1330 if (!PageAnon(old_page
)) {
1331 inc_mm_counter(mm
, anon_rss
);
1332 dec_mm_counter(mm
, file_rss
);
1334 flush_cache_page(vma
, address
, pfn
);
1335 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
1336 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1337 ptep_establish(vma
, address
, page_table
, entry
);
1338 update_mmu_cache(vma
, address
, entry
);
1339 lazy_mmu_prot_update(entry
);
1341 lru_cache_add_active(new_page
);
1342 page_add_anon_rmap(new_page
, vma
, address
);
1344 /* Free the old page.. */
1345 new_page
= old_page
;
1346 ret
|= VM_FAULT_WRITE
;
1348 page_cache_release(new_page
);
1349 page_cache_release(old_page
);
1351 pte_unmap(page_table
);
1352 spin_unlock(&mm
->page_table_lock
);
1355 page_cache_release(old_page
);
1356 return VM_FAULT_OOM
;
1360 * Helper functions for unmap_mapping_range().
1362 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1364 * We have to restart searching the prio_tree whenever we drop the lock,
1365 * since the iterator is only valid while the lock is held, and anyway
1366 * a later vma might be split and reinserted earlier while lock dropped.
1368 * The list of nonlinear vmas could be handled more efficiently, using
1369 * a placeholder, but handle it in the same way until a need is shown.
1370 * It is important to search the prio_tree before nonlinear list: a vma
1371 * may become nonlinear and be shifted from prio_tree to nonlinear list
1372 * while the lock is dropped; but never shifted from list to prio_tree.
1374 * In order to make forward progress despite restarting the search,
1375 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1376 * quickly skip it next time around. Since the prio_tree search only
1377 * shows us those vmas affected by unmapping the range in question, we
1378 * can't efficiently keep all vmas in step with mapping->truncate_count:
1379 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1380 * mapping->truncate_count and vma->vm_truncate_count are protected by
1383 * In order to make forward progress despite repeatedly restarting some
1384 * large vma, note the restart_addr from unmap_vmas when it breaks out:
1385 * and restart from that address when we reach that vma again. It might
1386 * have been split or merged, shrunk or extended, but never shifted: so
1387 * restart_addr remains valid so long as it remains in the vma's range.
1388 * unmap_mapping_range forces truncate_count to leap over page-aligned
1389 * values so we can save vma's restart_addr in its truncate_count field.
1391 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1393 static void reset_vma_truncate_counts(struct address_space
*mapping
)
1395 struct vm_area_struct
*vma
;
1396 struct prio_tree_iter iter
;
1398 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, 0, ULONG_MAX
)
1399 vma
->vm_truncate_count
= 0;
1400 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
1401 vma
->vm_truncate_count
= 0;
1404 static int unmap_mapping_range_vma(struct vm_area_struct
*vma
,
1405 unsigned long start_addr
, unsigned long end_addr
,
1406 struct zap_details
*details
)
1408 unsigned long restart_addr
;
1412 restart_addr
= vma
->vm_truncate_count
;
1413 if (is_restart_addr(restart_addr
) && start_addr
< restart_addr
) {
1414 start_addr
= restart_addr
;
1415 if (start_addr
>= end_addr
) {
1416 /* Top of vma has been split off since last time */
1417 vma
->vm_truncate_count
= details
->truncate_count
;
1422 restart_addr
= zap_page_range(vma
, start_addr
,
1423 end_addr
- start_addr
, details
);
1426 * We cannot rely on the break test in unmap_vmas:
1427 * on the one hand, we don't want to restart our loop
1428 * just because that broke out for the page_table_lock;
1429 * on the other hand, it does no test when vma is small.
1431 need_break
= need_resched() ||
1432 need_lockbreak(details
->i_mmap_lock
);
1434 if (restart_addr
>= end_addr
) {
1435 /* We have now completed this vma: mark it so */
1436 vma
->vm_truncate_count
= details
->truncate_count
;
1440 /* Note restart_addr in vma's truncate_count field */
1441 vma
->vm_truncate_count
= restart_addr
;
1446 spin_unlock(details
->i_mmap_lock
);
1448 spin_lock(details
->i_mmap_lock
);
1452 static inline void unmap_mapping_range_tree(struct prio_tree_root
*root
,
1453 struct zap_details
*details
)
1455 struct vm_area_struct
*vma
;
1456 struct prio_tree_iter iter
;
1457 pgoff_t vba
, vea
, zba
, zea
;
1460 vma_prio_tree_foreach(vma
, &iter
, root
,
1461 details
->first_index
, details
->last_index
) {
1462 /* Skip quickly over those we have already dealt with */
1463 if (vma
->vm_truncate_count
== details
->truncate_count
)
1466 vba
= vma
->vm_pgoff
;
1467 vea
= vba
+ ((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) - 1;
1468 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1469 zba
= details
->first_index
;
1472 zea
= details
->last_index
;
1476 if (unmap_mapping_range_vma(vma
,
1477 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
1478 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
1484 static inline void unmap_mapping_range_list(struct list_head
*head
,
1485 struct zap_details
*details
)
1487 struct vm_area_struct
*vma
;
1490 * In nonlinear VMAs there is no correspondence between virtual address
1491 * offset and file offset. So we must perform an exhaustive search
1492 * across *all* the pages in each nonlinear VMA, not just the pages
1493 * whose virtual address lies outside the file truncation point.
1496 list_for_each_entry(vma
, head
, shared
.vm_set
.list
) {
1497 /* Skip quickly over those we have already dealt with */
1498 if (vma
->vm_truncate_count
== details
->truncate_count
)
1500 details
->nonlinear_vma
= vma
;
1501 if (unmap_mapping_range_vma(vma
, vma
->vm_start
,
1502 vma
->vm_end
, details
) < 0)
1508 * unmap_mapping_range - unmap the portion of all mmaps
1509 * in the specified address_space corresponding to the specified
1510 * page range in the underlying file.
1511 * @mapping: the address space containing mmaps to be unmapped.
1512 * @holebegin: byte in first page to unmap, relative to the start of
1513 * the underlying file. This will be rounded down to a PAGE_SIZE
1514 * boundary. Note that this is different from vmtruncate(), which
1515 * must keep the partial page. In contrast, we must get rid of
1517 * @holelen: size of prospective hole in bytes. This will be rounded
1518 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
1520 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1521 * but 0 when invalidating pagecache, don't throw away private data.
1523 void unmap_mapping_range(struct address_space
*mapping
,
1524 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
1526 struct zap_details details
;
1527 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
1528 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1530 /* Check for overflow. */
1531 if (sizeof(holelen
) > sizeof(hlen
)) {
1533 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1534 if (holeend
& ~(long long)ULONG_MAX
)
1535 hlen
= ULONG_MAX
- hba
+ 1;
1538 details
.check_mapping
= even_cows
? NULL
: mapping
;
1539 details
.nonlinear_vma
= NULL
;
1540 details
.first_index
= hba
;
1541 details
.last_index
= hba
+ hlen
- 1;
1542 if (details
.last_index
< details
.first_index
)
1543 details
.last_index
= ULONG_MAX
;
1544 details
.i_mmap_lock
= &mapping
->i_mmap_lock
;
1546 spin_lock(&mapping
->i_mmap_lock
);
1548 /* serialize i_size write against truncate_count write */
1550 /* Protect against page faults, and endless unmapping loops */
1551 mapping
->truncate_count
++;
1553 * For archs where spin_lock has inclusive semantics like ia64
1554 * this smp_mb() will prevent to read pagetable contents
1555 * before the truncate_count increment is visible to
1559 if (unlikely(is_restart_addr(mapping
->truncate_count
))) {
1560 if (mapping
->truncate_count
== 0)
1561 reset_vma_truncate_counts(mapping
);
1562 mapping
->truncate_count
++;
1564 details
.truncate_count
= mapping
->truncate_count
;
1566 if (unlikely(!prio_tree_empty(&mapping
->i_mmap
)))
1567 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
1568 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
1569 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
1570 spin_unlock(&mapping
->i_mmap_lock
);
1572 EXPORT_SYMBOL(unmap_mapping_range
);
1575 * Handle all mappings that got truncated by a "truncate()"
1578 * NOTE! We have to be ready to update the memory sharing
1579 * between the file and the memory map for a potential last
1580 * incomplete page. Ugly, but necessary.
1582 int vmtruncate(struct inode
* inode
, loff_t offset
)
1584 struct address_space
*mapping
= inode
->i_mapping
;
1585 unsigned long limit
;
1587 if (inode
->i_size
< offset
)
1590 * truncation of in-use swapfiles is disallowed - it would cause
1591 * subsequent swapout to scribble on the now-freed blocks.
1593 if (IS_SWAPFILE(inode
))
1595 i_size_write(inode
, offset
);
1596 unmap_mapping_range(mapping
, offset
+ PAGE_SIZE
- 1, 0, 1);
1597 truncate_inode_pages(mapping
, offset
);
1601 limit
= current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
;
1602 if (limit
!= RLIM_INFINITY
&& offset
> limit
)
1604 if (offset
> inode
->i_sb
->s_maxbytes
)
1606 i_size_write(inode
, offset
);
1609 if (inode
->i_op
&& inode
->i_op
->truncate
)
1610 inode
->i_op
->truncate(inode
);
1613 send_sig(SIGXFSZ
, current
, 0);
1620 EXPORT_SYMBOL(vmtruncate
);
1623 * Primitive swap readahead code. We simply read an aligned block of
1624 * (1 << page_cluster) entries in the swap area. This method is chosen
1625 * because it doesn't cost us any seek time. We also make sure to queue
1626 * the 'original' request together with the readahead ones...
1628 * This has been extended to use the NUMA policies from the mm triggering
1631 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1633 void swapin_readahead(swp_entry_t entry
, unsigned long addr
,struct vm_area_struct
*vma
)
1636 struct vm_area_struct
*next_vma
= vma
? vma
->vm_next
: NULL
;
1639 struct page
*new_page
;
1640 unsigned long offset
;
1643 * Get the number of handles we should do readahead io to.
1645 num
= valid_swaphandles(entry
, &offset
);
1646 for (i
= 0; i
< num
; offset
++, i
++) {
1647 /* Ok, do the async read-ahead now */
1648 new_page
= read_swap_cache_async(swp_entry(swp_type(entry
),
1649 offset
), vma
, addr
);
1652 page_cache_release(new_page
);
1655 * Find the next applicable VMA for the NUMA policy.
1661 if (addr
>= vma
->vm_end
) {
1663 next_vma
= vma
? vma
->vm_next
: NULL
;
1665 if (vma
&& addr
< vma
->vm_start
)
1668 if (next_vma
&& addr
>= next_vma
->vm_start
) {
1670 next_vma
= vma
->vm_next
;
1675 lru_add_drain(); /* Push any new pages onto the LRU now */
1679 * We hold the mm semaphore and the page_table_lock on entry and
1680 * should release the pagetable lock on exit..
1682 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1683 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1684 int write_access
, pte_t orig_pte
)
1689 int ret
= VM_FAULT_MINOR
;
1691 pte_unmap(page_table
);
1692 spin_unlock(&mm
->page_table_lock
);
1694 entry
= pte_to_swp_entry(orig_pte
);
1695 page
= lookup_swap_cache(entry
);
1697 swapin_readahead(entry
, address
, vma
);
1698 page
= read_swap_cache_async(entry
, vma
, address
);
1701 * Back out if somebody else faulted in this pte while
1702 * we released the page table lock.
1704 spin_lock(&mm
->page_table_lock
);
1705 page_table
= pte_offset_map(pmd
, address
);
1706 if (likely(pte_same(*page_table
, orig_pte
)))
1711 /* Had to read the page from swap area: Major fault */
1712 ret
= VM_FAULT_MAJOR
;
1713 inc_page_state(pgmajfault
);
1717 mark_page_accessed(page
);
1721 * Back out if somebody else faulted in this pte while we
1722 * released the page table lock.
1724 spin_lock(&mm
->page_table_lock
);
1725 page_table
= pte_offset_map(pmd
, address
);
1726 if (unlikely(!pte_same(*page_table
, orig_pte
)))
1729 if (unlikely(!PageUptodate(page
))) {
1730 ret
= VM_FAULT_SIGBUS
;
1734 /* The page isn't present yet, go ahead with the fault. */
1736 inc_mm_counter(mm
, anon_rss
);
1737 pte
= mk_pte(page
, vma
->vm_page_prot
);
1738 if (write_access
&& can_share_swap_page(page
)) {
1739 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
1743 flush_icache_page(vma
, page
);
1744 set_pte_at(mm
, address
, page_table
, pte
);
1745 page_add_anon_rmap(page
, vma
, address
);
1749 remove_exclusive_swap_page(page
);
1753 if (do_wp_page(mm
, vma
, address
,
1754 page_table
, pmd
, pte
) == VM_FAULT_OOM
)
1759 /* No need to invalidate - it was non-present before */
1760 update_mmu_cache(vma
, address
, pte
);
1761 lazy_mmu_prot_update(pte
);
1763 pte_unmap(page_table
);
1764 spin_unlock(&mm
->page_table_lock
);
1768 pte_unmap(page_table
);
1769 spin_unlock(&mm
->page_table_lock
);
1771 page_cache_release(page
);
1776 * We are called with the MM semaphore and page_table_lock
1777 * spinlock held to protect against concurrent faults in
1778 * multithreaded programs.
1780 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1781 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1784 struct page
*page
= ZERO_PAGE(addr
);
1787 /* Mapping of ZERO_PAGE - vm_page_prot is readonly */
1788 entry
= mk_pte(page
, vma
->vm_page_prot
);
1791 /* Allocate our own private page. */
1792 pte_unmap(page_table
);
1793 spin_unlock(&mm
->page_table_lock
);
1795 if (unlikely(anon_vma_prepare(vma
)))
1797 page
= alloc_zeroed_user_highpage(vma
, address
);
1801 spin_lock(&mm
->page_table_lock
);
1802 page_table
= pte_offset_map(pmd
, address
);
1804 if (!pte_none(*page_table
)) {
1805 page_cache_release(page
);
1808 inc_mm_counter(mm
, anon_rss
);
1809 entry
= mk_pte(page
, vma
->vm_page_prot
);
1810 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1811 lru_cache_add_active(page
);
1812 SetPageReferenced(page
);
1813 page_add_anon_rmap(page
, vma
, address
);
1815 inc_mm_counter(mm
, file_rss
);
1816 page_add_file_rmap(page
);
1817 page_cache_get(page
);
1820 set_pte_at(mm
, address
, page_table
, entry
);
1822 /* No need to invalidate - it was non-present before */
1823 update_mmu_cache(vma
, address
, entry
);
1824 lazy_mmu_prot_update(entry
);
1826 pte_unmap(page_table
);
1827 spin_unlock(&mm
->page_table_lock
);
1828 return VM_FAULT_MINOR
;
1830 return VM_FAULT_OOM
;
1834 * do_no_page() tries to create a new page mapping. It aggressively
1835 * tries to share with existing pages, but makes a separate copy if
1836 * the "write_access" parameter is true in order to avoid the next
1839 * As this is called only for pages that do not currently exist, we
1840 * do not need to flush old virtual caches or the TLB.
1842 * This is called with the MM semaphore held and the page table
1843 * spinlock held. Exit with the spinlock released.
1845 static int do_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1846 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1849 struct page
*new_page
;
1850 struct address_space
*mapping
= NULL
;
1852 unsigned int sequence
= 0;
1853 int ret
= VM_FAULT_MINOR
;
1856 pte_unmap(page_table
);
1857 spin_unlock(&mm
->page_table_lock
);
1860 mapping
= vma
->vm_file
->f_mapping
;
1861 sequence
= mapping
->truncate_count
;
1862 smp_rmb(); /* serializes i_size against truncate_count */
1865 new_page
= vma
->vm_ops
->nopage(vma
, address
& PAGE_MASK
, &ret
);
1867 * No smp_rmb is needed here as long as there's a full
1868 * spin_lock/unlock sequence inside the ->nopage callback
1869 * (for the pagecache lookup) that acts as an implicit
1870 * smp_mb() and prevents the i_size read to happen
1871 * after the next truncate_count read.
1874 /* no page was available -- either SIGBUS or OOM */
1875 if (new_page
== NOPAGE_SIGBUS
)
1876 return VM_FAULT_SIGBUS
;
1877 if (new_page
== NOPAGE_OOM
)
1878 return VM_FAULT_OOM
;
1881 * Should we do an early C-O-W break?
1883 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
1886 if (unlikely(anon_vma_prepare(vma
)))
1888 page
= alloc_page_vma(GFP_HIGHUSER
, vma
, address
);
1891 copy_user_highpage(page
, new_page
, address
);
1892 page_cache_release(new_page
);
1897 spin_lock(&mm
->page_table_lock
);
1899 * For a file-backed vma, someone could have truncated or otherwise
1900 * invalidated this page. If unmap_mapping_range got called,
1901 * retry getting the page.
1903 if (mapping
&& unlikely(sequence
!= mapping
->truncate_count
)) {
1904 spin_unlock(&mm
->page_table_lock
);
1905 page_cache_release(new_page
);
1907 sequence
= mapping
->truncate_count
;
1911 page_table
= pte_offset_map(pmd
, address
);
1914 * This silly early PAGE_DIRTY setting removes a race
1915 * due to the bad i386 page protection. But it's valid
1916 * for other architectures too.
1918 * Note that if write_access is true, we either now have
1919 * an exclusive copy of the page, or this is a shared mapping,
1920 * so we can make it writable and dirty to avoid having to
1921 * handle that later.
1923 /* Only go through if we didn't race with anybody else... */
1924 if (pte_none(*page_table
)) {
1925 flush_icache_page(vma
, new_page
);
1926 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
1928 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1929 set_pte_at(mm
, address
, page_table
, entry
);
1931 inc_mm_counter(mm
, anon_rss
);
1932 lru_cache_add_active(new_page
);
1933 page_add_anon_rmap(new_page
, vma
, address
);
1934 } else if (!(vma
->vm_flags
& VM_RESERVED
)) {
1935 inc_mm_counter(mm
, file_rss
);
1936 page_add_file_rmap(new_page
);
1939 /* One of our sibling threads was faster, back out. */
1940 page_cache_release(new_page
);
1944 /* no need to invalidate: a not-present page shouldn't be cached */
1945 update_mmu_cache(vma
, address
, entry
);
1946 lazy_mmu_prot_update(entry
);
1948 pte_unmap(page_table
);
1949 spin_unlock(&mm
->page_table_lock
);
1952 page_cache_release(new_page
);
1953 return VM_FAULT_OOM
;
1957 * Fault of a previously existing named mapping. Repopulate the pte
1958 * from the encoded file_pte if possible. This enables swappable
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 pte_unmap(page_table
);
1969 spin_unlock(&mm
->page_table_lock
);
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 * Note the "page_table_lock". It is to protect against kswapd removing
2000 * pages from under us. Note that kswapd only ever _removes_ pages, never
2001 * adds them. As such, once we have noticed that the page is not present,
2002 * we can drop the lock early.
2004 * The adding of pages is protected by the MM semaphore (which we hold),
2005 * so we don't need to worry about a page being suddenly been added into
2008 * We enter with the pagetable spinlock held, we are supposed to
2009 * release it when done.
2011 static inline int handle_pte_fault(struct mm_struct
*mm
,
2012 struct vm_area_struct
*vma
, unsigned long address
,
2013 pte_t
*pte
, pmd_t
*pmd
, int write_access
)
2018 if (!pte_present(entry
)) {
2019 if (pte_none(entry
)) {
2020 if (!vma
->vm_ops
|| !vma
->vm_ops
->nopage
)
2021 return do_anonymous_page(mm
, vma
, address
,
2022 pte
, pmd
, write_access
);
2023 return do_no_page(mm
, vma
, address
,
2024 pte
, pmd
, write_access
);
2026 if (pte_file(entry
))
2027 return do_file_page(mm
, vma
, address
,
2028 pte
, pmd
, write_access
, entry
);
2029 return do_swap_page(mm
, vma
, address
,
2030 pte
, pmd
, write_access
, entry
);
2034 if (!pte_write(entry
))
2035 return do_wp_page(mm
, vma
, address
, pte
, pmd
, entry
);
2036 entry
= pte_mkdirty(entry
);
2038 entry
= pte_mkyoung(entry
);
2039 ptep_set_access_flags(vma
, address
, pte
, entry
, write_access
);
2040 update_mmu_cache(vma
, address
, entry
);
2041 lazy_mmu_prot_update(entry
);
2043 spin_unlock(&mm
->page_table_lock
);
2044 return VM_FAULT_MINOR
;
2048 * By the time we get here, we already hold the mm semaphore
2050 int __handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2051 unsigned long address
, int write_access
)
2058 __set_current_state(TASK_RUNNING
);
2060 inc_page_state(pgfault
);
2062 if (unlikely(is_vm_hugetlb_page(vma
)))
2063 return hugetlb_fault(mm
, vma
, address
, write_access
);
2066 * We need the page table lock to synchronize with kswapd
2067 * and the SMP-safe atomic PTE updates.
2069 pgd
= pgd_offset(mm
, address
);
2070 spin_lock(&mm
->page_table_lock
);
2072 pud
= pud_alloc(mm
, pgd
, address
);
2076 pmd
= pmd_alloc(mm
, pud
, address
);
2080 pte
= pte_alloc_map(mm
, pmd
, address
);
2084 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, write_access
);
2087 spin_unlock(&mm
->page_table_lock
);
2088 return VM_FAULT_OOM
;
2091 #ifndef __PAGETABLE_PUD_FOLDED
2093 * Allocate page upper directory.
2094 * We've already handled the fast-path in-line.
2096 pud_t fastcall
*__pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
2100 if (mm
!= &init_mm
) /* Temporary bridging hack */
2101 spin_unlock(&mm
->page_table_lock
);
2102 new = pud_alloc_one(mm
, address
);
2104 if (mm
!= &init_mm
) /* Temporary bridging hack */
2105 spin_lock(&mm
->page_table_lock
);
2109 spin_lock(&mm
->page_table_lock
);
2110 if (pgd_present(*pgd
)) {
2114 pgd_populate(mm
, pgd
, new);
2116 if (mm
== &init_mm
) /* Temporary bridging hack */
2117 spin_unlock(&mm
->page_table_lock
);
2118 return pud_offset(pgd
, address
);
2120 #endif /* __PAGETABLE_PUD_FOLDED */
2122 #ifndef __PAGETABLE_PMD_FOLDED
2124 * Allocate page middle directory.
2125 * We've already handled the fast-path in-line.
2127 pmd_t fastcall
*__pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
2131 if (mm
!= &init_mm
) /* Temporary bridging hack */
2132 spin_unlock(&mm
->page_table_lock
);
2133 new = pmd_alloc_one(mm
, address
);
2135 if (mm
!= &init_mm
) /* Temporary bridging hack */
2136 spin_lock(&mm
->page_table_lock
);
2140 spin_lock(&mm
->page_table_lock
);
2141 #ifndef __ARCH_HAS_4LEVEL_HACK
2142 if (pud_present(*pud
)) {
2146 pud_populate(mm
, pud
, new);
2148 if (pgd_present(*pud
)) {
2152 pgd_populate(mm
, pud
, new);
2153 #endif /* __ARCH_HAS_4LEVEL_HACK */
2156 if (mm
== &init_mm
) /* Temporary bridging hack */
2157 spin_unlock(&mm
->page_table_lock
);
2158 return pmd_offset(pud
, address
);
2160 #endif /* __PAGETABLE_PMD_FOLDED */
2162 int make_pages_present(unsigned long addr
, unsigned long end
)
2164 int ret
, len
, write
;
2165 struct vm_area_struct
* vma
;
2167 vma
= find_vma(current
->mm
, addr
);
2170 write
= (vma
->vm_flags
& VM_WRITE
) != 0;
2173 if (end
> vma
->vm_end
)
2175 len
= (end
+PAGE_SIZE
-1)/PAGE_SIZE
-addr
/PAGE_SIZE
;
2176 ret
= get_user_pages(current
, current
->mm
, addr
,
2177 len
, write
, 0, NULL
, NULL
);
2180 return ret
== len
? 0 : -1;
2184 * Map a vmalloc()-space virtual address to the physical page.
2186 struct page
* vmalloc_to_page(void * vmalloc_addr
)
2188 unsigned long addr
= (unsigned long) vmalloc_addr
;
2189 struct page
*page
= NULL
;
2190 pgd_t
*pgd
= pgd_offset_k(addr
);
2195 if (!pgd_none(*pgd
)) {
2196 pud
= pud_offset(pgd
, addr
);
2197 if (!pud_none(*pud
)) {
2198 pmd
= pmd_offset(pud
, addr
);
2199 if (!pmd_none(*pmd
)) {
2200 ptep
= pte_offset_map(pmd
, addr
);
2202 if (pte_present(pte
))
2203 page
= pte_page(pte
);
2211 EXPORT_SYMBOL(vmalloc_to_page
);
2214 * Map a vmalloc()-space virtual address to the physical page frame number.
2216 unsigned long vmalloc_to_pfn(void * vmalloc_addr
)
2218 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
2221 EXPORT_SYMBOL(vmalloc_to_pfn
);
2223 #if !defined(__HAVE_ARCH_GATE_AREA)
2225 #if defined(AT_SYSINFO_EHDR)
2226 static struct vm_area_struct gate_vma
;
2228 static int __init
gate_vma_init(void)
2230 gate_vma
.vm_mm
= NULL
;
2231 gate_vma
.vm_start
= FIXADDR_USER_START
;
2232 gate_vma
.vm_end
= FIXADDR_USER_END
;
2233 gate_vma
.vm_page_prot
= PAGE_READONLY
;
2234 gate_vma
.vm_flags
= VM_RESERVED
;
2237 __initcall(gate_vma_init
);
2240 struct vm_area_struct
*get_gate_vma(struct task_struct
*tsk
)
2242 #ifdef AT_SYSINFO_EHDR
2249 int in_gate_area_no_task(unsigned long addr
)
2251 #ifdef AT_SYSINFO_EHDR
2252 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
2258 #endif /* __HAVE_ARCH_GATE_AREA */