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_lock_deinit(page
);
118 pte_free_tlb(tlb
, page
);
119 dec_page_state(nr_page_table_pages
);
123 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
124 unsigned long addr
, unsigned long end
,
125 unsigned long floor
, unsigned long ceiling
)
132 pmd
= pmd_offset(pud
, addr
);
134 next
= pmd_addr_end(addr
, end
);
135 if (pmd_none_or_clear_bad(pmd
))
137 free_pte_range(tlb
, pmd
);
138 } while (pmd
++, addr
= next
, addr
!= end
);
148 if (end
- 1 > ceiling
- 1)
151 pmd
= pmd_offset(pud
, start
);
153 pmd_free_tlb(tlb
, pmd
);
156 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
157 unsigned long addr
, unsigned long end
,
158 unsigned long floor
, unsigned long ceiling
)
165 pud
= pud_offset(pgd
, addr
);
167 next
= pud_addr_end(addr
, end
);
168 if (pud_none_or_clear_bad(pud
))
170 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
171 } while (pud
++, addr
= next
, addr
!= end
);
177 ceiling
&= PGDIR_MASK
;
181 if (end
- 1 > ceiling
- 1)
184 pud
= pud_offset(pgd
, start
);
186 pud_free_tlb(tlb
, pud
);
190 * This function frees user-level page tables of a process.
192 * Must be called with pagetable lock held.
194 void free_pgd_range(struct mmu_gather
**tlb
,
195 unsigned long addr
, unsigned long end
,
196 unsigned long floor
, unsigned long ceiling
)
203 * The next few lines have given us lots of grief...
205 * Why are we testing PMD* at this top level? Because often
206 * there will be no work to do at all, and we'd prefer not to
207 * go all the way down to the bottom just to discover that.
209 * Why all these "- 1"s? Because 0 represents both the bottom
210 * of the address space and the top of it (using -1 for the
211 * top wouldn't help much: the masks would do the wrong thing).
212 * The rule is that addr 0 and floor 0 refer to the bottom of
213 * the address space, but end 0 and ceiling 0 refer to the top
214 * Comparisons need to use "end - 1" and "ceiling - 1" (though
215 * that end 0 case should be mythical).
217 * Wherever addr is brought up or ceiling brought down, we must
218 * be careful to reject "the opposite 0" before it confuses the
219 * subsequent tests. But what about where end is brought down
220 * by PMD_SIZE below? no, end can't go down to 0 there.
222 * Whereas we round start (addr) and ceiling down, by different
223 * masks at different levels, in order to test whether a table
224 * now has no other vmas using it, so can be freed, we don't
225 * bother to round floor or end up - the tests don't need that.
239 if (end
- 1 > ceiling
- 1)
245 pgd
= pgd_offset((*tlb
)->mm
, addr
);
247 next
= pgd_addr_end(addr
, end
);
248 if (pgd_none_or_clear_bad(pgd
))
250 free_pud_range(*tlb
, pgd
, addr
, next
, floor
, ceiling
);
251 } while (pgd
++, addr
= next
, addr
!= end
);
254 flush_tlb_pgtables((*tlb
)->mm
, start
, end
);
257 void free_pgtables(struct mmu_gather
**tlb
, struct vm_area_struct
*vma
,
258 unsigned long floor
, unsigned long ceiling
)
261 struct vm_area_struct
*next
= vma
->vm_next
;
262 unsigned long addr
= vma
->vm_start
;
265 * Hide vma from rmap and vmtruncate before freeing pgtables
267 anon_vma_unlink(vma
);
268 unlink_file_vma(vma
);
270 if (is_hugepage_only_range(vma
->vm_mm
, addr
, HPAGE_SIZE
)) {
271 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
272 floor
, next
? next
->vm_start
: ceiling
);
275 * Optimization: gather nearby vmas into one call down
277 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
278 && !is_hugepage_only_range(vma
->vm_mm
, next
->vm_start
,
282 anon_vma_unlink(vma
);
283 unlink_file_vma(vma
);
285 free_pgd_range(tlb
, addr
, vma
->vm_end
,
286 floor
, next
? next
->vm_start
: ceiling
);
292 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
, unsigned long address
)
294 struct page
*new = pte_alloc_one(mm
, address
);
299 spin_lock(&mm
->page_table_lock
);
300 if (pmd_present(*pmd
)) { /* Another has populated it */
301 pte_lock_deinit(new);
305 inc_page_state(nr_page_table_pages
);
306 pmd_populate(mm
, pmd
, new);
308 spin_unlock(&mm
->page_table_lock
);
312 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
314 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
318 spin_lock(&init_mm
.page_table_lock
);
319 if (pmd_present(*pmd
)) /* Another has populated it */
320 pte_free_kernel(new);
322 pmd_populate_kernel(&init_mm
, pmd
, new);
323 spin_unlock(&init_mm
.page_table_lock
);
327 static inline void add_mm_rss(struct mm_struct
*mm
, int file_rss
, int anon_rss
)
330 add_mm_counter(mm
, file_rss
, file_rss
);
332 add_mm_counter(mm
, anon_rss
, anon_rss
);
336 * This function is called to print an error when a pte in a
337 * !VM_UNPAGED region is found pointing to an invalid pfn (which
340 * The calling function must still handle the error.
342 void print_bad_pte(struct vm_area_struct
*vma
, pte_t pte
, unsigned long vaddr
)
344 printk(KERN_ERR
"Bad pte = %08llx, process = %s, "
345 "vm_flags = %lx, vaddr = %lx\n",
346 (long long)pte_val(pte
),
347 (vma
->vm_mm
== current
->mm
? current
->comm
: "???"),
348 vma
->vm_flags
, vaddr
);
353 * page_is_anon applies strict checks for an anonymous page belonging to
354 * this vma at this address. It is used on VM_UNPAGED vmas, which are
355 * usually populated with shared originals (which must not be counted),
356 * but occasionally contain private COWed copies (when !VM_SHARED, or
357 * perhaps via ptrace when VM_SHARED). An mmap of /dev/mem might window
358 * free pages, pages from other processes, or from other parts of this:
359 * it's tricky, but try not to be deceived by foreign anonymous pages.
361 static inline int page_is_anon(struct page
*page
,
362 struct vm_area_struct
*vma
, unsigned long addr
)
364 return page
&& PageAnon(page
) && page_mapped(page
) &&
365 page_address_in_vma(page
, vma
) == addr
;
369 * copy one vm_area from one task to the other. Assumes the page tables
370 * already present in the new task to be cleared in the whole range
371 * covered by this vma.
375 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
376 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
377 unsigned long addr
, int *rss
)
379 unsigned long vm_flags
= vma
->vm_flags
;
380 pte_t pte
= *src_pte
;
384 /* pte contains position in swap or file, so copy. */
385 if (unlikely(!pte_present(pte
))) {
386 if (!pte_file(pte
)) {
387 swap_duplicate(pte_to_swp_entry(pte
));
388 /* make sure dst_mm is on swapoff's mmlist. */
389 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
390 spin_lock(&mmlist_lock
);
391 if (list_empty(&dst_mm
->mmlist
))
392 list_add(&dst_mm
->mmlist
,
394 spin_unlock(&mmlist_lock
);
401 page
= pfn_valid(pfn
)? pfn_to_page(pfn
): NULL
;
403 if (unlikely(vm_flags
& VM_UNPAGED
))
404 if (!page_is_anon(page
, vma
, addr
))
408 * If the pte points outside of valid memory but
409 * the region is not VM_UNPAGED, we have a problem.
411 if (unlikely(!page
)) {
412 print_bad_pte(vma
, pte
, addr
);
413 goto out_set_pte
; /* try to do something sane */
417 * If it's a COW mapping, write protect it both
418 * in the parent and the child
420 if ((vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
) {
421 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
426 * If it's a shared mapping, mark it clean in
429 if (vm_flags
& VM_SHARED
)
430 pte
= pte_mkclean(pte
);
431 pte
= pte_mkold(pte
);
434 rss
[!!PageAnon(page
)]++;
437 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
440 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
441 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
442 unsigned long addr
, unsigned long end
)
444 pte_t
*src_pte
, *dst_pte
;
445 spinlock_t
*src_ptl
, *dst_ptl
;
451 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
454 src_pte
= pte_offset_map_nested(src_pmd
, addr
);
455 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
460 * We are holding two locks at this point - either of them
461 * could generate latencies in another task on another CPU.
463 if (progress
>= 32) {
465 if (need_resched() ||
466 need_lockbreak(src_ptl
) ||
467 need_lockbreak(dst_ptl
))
470 if (pte_none(*src_pte
)) {
474 copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
, vma
, addr
, rss
);
476 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
478 spin_unlock(src_ptl
);
479 pte_unmap_nested(src_pte
- 1);
480 add_mm_rss(dst_mm
, rss
[0], rss
[1]);
481 pte_unmap_unlock(dst_pte
- 1, dst_ptl
);
488 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
489 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
490 unsigned long addr
, unsigned long end
)
492 pmd_t
*src_pmd
, *dst_pmd
;
495 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
498 src_pmd
= pmd_offset(src_pud
, addr
);
500 next
= pmd_addr_end(addr
, end
);
501 if (pmd_none_or_clear_bad(src_pmd
))
503 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
506 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
510 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
511 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
512 unsigned long addr
, unsigned long end
)
514 pud_t
*src_pud
, *dst_pud
;
517 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
520 src_pud
= pud_offset(src_pgd
, addr
);
522 next
= pud_addr_end(addr
, end
);
523 if (pud_none_or_clear_bad(src_pud
))
525 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
528 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
532 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
533 struct vm_area_struct
*vma
)
535 pgd_t
*src_pgd
, *dst_pgd
;
537 unsigned long addr
= vma
->vm_start
;
538 unsigned long end
= vma
->vm_end
;
541 * Don't copy ptes where a page fault will fill them correctly.
542 * Fork becomes much lighter when there are big shared or private
543 * readonly mappings. The tradeoff is that copy_page_range is more
544 * efficient than faulting.
546 if (!(vma
->vm_flags
& (VM_HUGETLB
|VM_NONLINEAR
|VM_UNPAGED
))) {
551 if (is_vm_hugetlb_page(vma
))
552 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
554 dst_pgd
= pgd_offset(dst_mm
, addr
);
555 src_pgd
= pgd_offset(src_mm
, addr
);
557 next
= pgd_addr_end(addr
, end
);
558 if (pgd_none_or_clear_bad(src_pgd
))
560 if (copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
563 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
567 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
568 struct vm_area_struct
*vma
, pmd_t
*pmd
,
569 unsigned long addr
, unsigned long end
,
570 long *zap_work
, struct zap_details
*details
)
572 struct mm_struct
*mm
= tlb
->mm
;
578 pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
581 if (pte_none(ptent
)) {
585 if (pte_present(ptent
)) {
589 (*zap_work
) -= PAGE_SIZE
;
591 pfn
= pte_pfn(ptent
);
592 page
= pfn_valid(pfn
)? pfn_to_page(pfn
): NULL
;
594 if (unlikely(vma
->vm_flags
& VM_UNPAGED
)) {
595 if (!page_is_anon(page
, vma
, addr
))
597 } else if (unlikely(!page
))
598 print_bad_pte(vma
, ptent
, addr
);
600 if (unlikely(details
) && page
) {
602 * unmap_shared_mapping_pages() wants to
603 * invalidate cache without truncating:
604 * unmap shared but keep private pages.
606 if (details
->check_mapping
&&
607 details
->check_mapping
!= page
->mapping
)
610 * Each page->index must be checked when
611 * invalidating or truncating nonlinear.
613 if (details
->nonlinear_vma
&&
614 (page
->index
< details
->first_index
||
615 page
->index
> details
->last_index
))
618 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
620 tlb_remove_tlb_entry(tlb
, pte
, addr
);
623 if (unlikely(details
) && details
->nonlinear_vma
624 && linear_page_index(details
->nonlinear_vma
,
625 addr
) != page
->index
)
626 set_pte_at(mm
, addr
, pte
,
627 pgoff_to_pte(page
->index
));
631 if (pte_dirty(ptent
))
632 set_page_dirty(page
);
633 if (pte_young(ptent
))
634 mark_page_accessed(page
);
637 page_remove_rmap(page
);
638 tlb_remove_page(tlb
, page
);
642 * If details->check_mapping, we leave swap entries;
643 * if details->nonlinear_vma, we leave file entries.
645 if (unlikely(details
))
647 if (!pte_file(ptent
))
648 free_swap_and_cache(pte_to_swp_entry(ptent
));
649 pte_clear_full(mm
, addr
, pte
, tlb
->fullmm
);
650 } while (pte
++, addr
+= PAGE_SIZE
, (addr
!= end
&& *zap_work
> 0));
652 add_mm_rss(mm
, file_rss
, anon_rss
);
653 pte_unmap_unlock(pte
- 1, ptl
);
658 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
659 struct vm_area_struct
*vma
, pud_t
*pud
,
660 unsigned long addr
, unsigned long end
,
661 long *zap_work
, struct zap_details
*details
)
666 pmd
= pmd_offset(pud
, addr
);
668 next
= pmd_addr_end(addr
, end
);
669 if (pmd_none_or_clear_bad(pmd
)) {
673 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
,
675 } while (pmd
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
680 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
681 struct vm_area_struct
*vma
, pgd_t
*pgd
,
682 unsigned long addr
, unsigned long end
,
683 long *zap_work
, struct zap_details
*details
)
688 pud
= pud_offset(pgd
, addr
);
690 next
= pud_addr_end(addr
, end
);
691 if (pud_none_or_clear_bad(pud
)) {
695 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
,
697 } while (pud
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
702 static unsigned long unmap_page_range(struct mmu_gather
*tlb
,
703 struct vm_area_struct
*vma
,
704 unsigned long addr
, unsigned long end
,
705 long *zap_work
, struct zap_details
*details
)
710 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
714 tlb_start_vma(tlb
, vma
);
715 pgd
= pgd_offset(vma
->vm_mm
, addr
);
717 next
= pgd_addr_end(addr
, end
);
718 if (pgd_none_or_clear_bad(pgd
)) {
722 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
,
724 } while (pgd
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
725 tlb_end_vma(tlb
, vma
);
730 #ifdef CONFIG_PREEMPT
731 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
733 /* No preempt: go for improved straight-line efficiency */
734 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
738 * unmap_vmas - unmap a range of memory covered by a list of vma's
739 * @tlbp: address of the caller's struct mmu_gather
740 * @vma: the starting vma
741 * @start_addr: virtual address at which to start unmapping
742 * @end_addr: virtual address at which to end unmapping
743 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
744 * @details: details of nonlinear truncation or shared cache invalidation
746 * Returns the end address of the unmapping (restart addr if interrupted).
748 * Unmap all pages in the vma list.
750 * We aim to not hold locks for too long (for scheduling latency reasons).
751 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
752 * return the ending mmu_gather to the caller.
754 * Only addresses between `start' and `end' will be unmapped.
756 * The VMA list must be sorted in ascending virtual address order.
758 * unmap_vmas() assumes that the caller will flush the whole unmapped address
759 * range after unmap_vmas() returns. So the only responsibility here is to
760 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
761 * drops the lock and schedules.
763 unsigned long unmap_vmas(struct mmu_gather
**tlbp
,
764 struct vm_area_struct
*vma
, unsigned long start_addr
,
765 unsigned long end_addr
, unsigned long *nr_accounted
,
766 struct zap_details
*details
)
768 long zap_work
= ZAP_BLOCK_SIZE
;
769 unsigned long tlb_start
= 0; /* For tlb_finish_mmu */
770 int tlb_start_valid
= 0;
771 unsigned long start
= start_addr
;
772 spinlock_t
*i_mmap_lock
= details
? details
->i_mmap_lock
: NULL
;
773 int fullmm
= (*tlbp
)->fullmm
;
775 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
) {
778 start
= max(vma
->vm_start
, start_addr
);
779 if (start
>= vma
->vm_end
)
781 end
= min(vma
->vm_end
, end_addr
);
782 if (end
<= vma
->vm_start
)
785 if (vma
->vm_flags
& VM_ACCOUNT
)
786 *nr_accounted
+= (end
- start
) >> PAGE_SHIFT
;
788 while (start
!= end
) {
789 if (!tlb_start_valid
) {
794 if (unlikely(is_vm_hugetlb_page(vma
))) {
795 unmap_hugepage_range(vma
, start
, end
);
796 zap_work
-= (end
- start
) /
797 (HPAGE_SIZE
/ PAGE_SIZE
);
800 start
= unmap_page_range(*tlbp
, vma
,
801 start
, end
, &zap_work
, details
);
804 BUG_ON(start
!= end
);
808 tlb_finish_mmu(*tlbp
, tlb_start
, start
);
810 if (need_resched() ||
811 (i_mmap_lock
&& need_lockbreak(i_mmap_lock
))) {
819 *tlbp
= tlb_gather_mmu(vma
->vm_mm
, fullmm
);
821 zap_work
= ZAP_BLOCK_SIZE
;
825 return start
; /* which is now the end (or restart) address */
829 * zap_page_range - remove user pages in a given range
830 * @vma: vm_area_struct holding the applicable pages
831 * @address: starting address of pages to zap
832 * @size: number of bytes to zap
833 * @details: details of nonlinear truncation or shared cache invalidation
835 unsigned long zap_page_range(struct vm_area_struct
*vma
, unsigned long address
,
836 unsigned long size
, struct zap_details
*details
)
838 struct mm_struct
*mm
= vma
->vm_mm
;
839 struct mmu_gather
*tlb
;
840 unsigned long end
= address
+ size
;
841 unsigned long nr_accounted
= 0;
844 tlb
= tlb_gather_mmu(mm
, 0);
845 update_hiwater_rss(mm
);
846 end
= unmap_vmas(&tlb
, vma
, address
, end
, &nr_accounted
, details
);
848 tlb_finish_mmu(tlb
, address
, end
);
853 * Do a quick page-table lookup for a single page.
855 struct page
*follow_page(struct mm_struct
*mm
, unsigned long address
,
866 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
868 BUG_ON(flags
& FOLL_GET
);
873 pgd
= pgd_offset(mm
, address
);
874 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
877 pud
= pud_offset(pgd
, address
);
878 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
881 pmd
= pmd_offset(pud
, address
);
882 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
885 if (pmd_huge(*pmd
)) {
886 BUG_ON(flags
& FOLL_GET
);
887 page
= follow_huge_pmd(mm
, address
, pmd
, flags
& FOLL_WRITE
);
891 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
896 if (!pte_present(pte
))
898 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
904 page
= pfn_to_page(pfn
);
905 if (flags
& FOLL_GET
)
907 if (flags
& FOLL_TOUCH
) {
908 if ((flags
& FOLL_WRITE
) &&
909 !pte_dirty(pte
) && !PageDirty(page
))
910 set_page_dirty(page
);
911 mark_page_accessed(page
);
914 pte_unmap_unlock(ptep
, ptl
);
920 * When core dumping an enormous anonymous area that nobody
921 * has touched so far, we don't want to allocate page tables.
923 if (flags
& FOLL_ANON
) {
924 page
= ZERO_PAGE(address
);
925 if (flags
& FOLL_GET
)
927 BUG_ON(flags
& FOLL_WRITE
);
932 int get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
933 unsigned long start
, int len
, int write
, int force
,
934 struct page
**pages
, struct vm_area_struct
**vmas
)
937 unsigned int vm_flags
;
940 * Require read or write permissions.
941 * If 'force' is set, we only require the "MAY" flags.
943 vm_flags
= write
? (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
944 vm_flags
&= force
? (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
948 struct vm_area_struct
*vma
;
949 unsigned int foll_flags
;
951 vma
= find_extend_vma(mm
, start
);
952 if (!vma
&& in_gate_area(tsk
, start
)) {
953 unsigned long pg
= start
& PAGE_MASK
;
954 struct vm_area_struct
*gate_vma
= get_gate_vma(tsk
);
959 if (write
) /* user gate pages are read-only */
960 return i
? : -EFAULT
;
962 pgd
= pgd_offset_k(pg
);
964 pgd
= pgd_offset_gate(mm
, pg
);
965 BUG_ON(pgd_none(*pgd
));
966 pud
= pud_offset(pgd
, pg
);
967 BUG_ON(pud_none(*pud
));
968 pmd
= pmd_offset(pud
, pg
);
970 return i
? : -EFAULT
;
971 pte
= pte_offset_map(pmd
, pg
);
972 if (pte_none(*pte
)) {
974 return i
? : -EFAULT
;
977 pages
[i
] = pte_page(*pte
);
989 if (!vma
|| (vma
->vm_flags
& VM_IO
)
990 || !(vm_flags
& vma
->vm_flags
))
991 return i
? : -EFAULT
;
993 if (is_vm_hugetlb_page(vma
)) {
994 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
999 foll_flags
= FOLL_TOUCH
;
1001 foll_flags
|= FOLL_GET
;
1002 if (!write
&& !(vma
->vm_flags
& VM_LOCKED
) &&
1003 (!vma
->vm_ops
|| !vma
->vm_ops
->nopage
))
1004 foll_flags
|= FOLL_ANON
;
1010 foll_flags
|= FOLL_WRITE
;
1013 while (!(page
= follow_page(mm
, start
, foll_flags
))) {
1015 ret
= __handle_mm_fault(mm
, vma
, start
,
1016 foll_flags
& FOLL_WRITE
);
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
)
1024 foll_flags
&= ~FOLL_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
;
1043 flush_dcache_page(page
);
1050 } while (len
&& start
< vma
->vm_end
);
1054 EXPORT_SYMBOL(get_user_pages
);
1056 static int zeromap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1057 unsigned long addr
, unsigned long end
, pgprot_t prot
)
1062 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1066 struct page
*page
= ZERO_PAGE(addr
);
1067 pte_t zero_pte
= pte_wrprotect(mk_pte(page
, prot
));
1068 page_cache_get(page
);
1069 page_add_file_rmap(page
);
1070 inc_mm_counter(mm
, file_rss
);
1071 BUG_ON(!pte_none(*pte
));
1072 set_pte_at(mm
, addr
, pte
, zero_pte
);
1073 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1074 pte_unmap_unlock(pte
- 1, ptl
);
1078 static inline int zeromap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1079 unsigned long addr
, unsigned long end
, pgprot_t prot
)
1084 pmd
= pmd_alloc(mm
, pud
, addr
);
1088 next
= pmd_addr_end(addr
, end
);
1089 if (zeromap_pte_range(mm
, pmd
, addr
, next
, prot
))
1091 } while (pmd
++, addr
= next
, addr
!= end
);
1095 static inline int zeromap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1096 unsigned long addr
, unsigned long end
, pgprot_t prot
)
1101 pud
= pud_alloc(mm
, pgd
, addr
);
1105 next
= pud_addr_end(addr
, end
);
1106 if (zeromap_pmd_range(mm
, pud
, addr
, next
, prot
))
1108 } while (pud
++, addr
= next
, addr
!= end
);
1112 int zeromap_page_range(struct vm_area_struct
*vma
,
1113 unsigned long addr
, unsigned long size
, pgprot_t prot
)
1117 unsigned long end
= addr
+ size
;
1118 struct mm_struct
*mm
= vma
->vm_mm
;
1121 BUG_ON(addr
>= end
);
1122 pgd
= pgd_offset(mm
, addr
);
1123 flush_cache_range(vma
, addr
, end
);
1125 next
= pgd_addr_end(addr
, end
);
1126 err
= zeromap_pud_range(mm
, pgd
, addr
, next
, prot
);
1129 } while (pgd
++, addr
= next
, addr
!= end
);
1134 * maps a range of physical memory into the requested pages. the old
1135 * mappings are removed. any references to nonexistent pages results
1136 * in null mappings (currently treated as "copy-on-access")
1138 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1139 unsigned long addr
, unsigned long end
,
1140 unsigned long pfn
, pgprot_t prot
)
1145 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1149 BUG_ON(!pte_none(*pte
));
1150 set_pte_at(mm
, addr
, pte
, pfn_pte(pfn
, prot
));
1152 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1153 pte_unmap_unlock(pte
- 1, ptl
);
1157 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1158 unsigned long addr
, unsigned long end
,
1159 unsigned long pfn
, pgprot_t prot
)
1164 pfn
-= addr
>> PAGE_SHIFT
;
1165 pmd
= pmd_alloc(mm
, pud
, addr
);
1169 next
= pmd_addr_end(addr
, end
);
1170 if (remap_pte_range(mm
, pmd
, addr
, next
,
1171 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1173 } while (pmd
++, addr
= next
, addr
!= end
);
1177 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1178 unsigned long addr
, unsigned long end
,
1179 unsigned long pfn
, pgprot_t prot
)
1184 pfn
-= addr
>> PAGE_SHIFT
;
1185 pud
= pud_alloc(mm
, pgd
, addr
);
1189 next
= pud_addr_end(addr
, end
);
1190 if (remap_pmd_range(mm
, pud
, addr
, next
,
1191 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1193 } while (pud
++, addr
= next
, addr
!= end
);
1197 /* Note: this is only safe if the mm semaphore is held when called. */
1198 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1199 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1203 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1204 struct mm_struct
*mm
= vma
->vm_mm
;
1208 * Physically remapped pages are special. Tell the
1209 * rest of the world about it:
1210 * VM_IO tells people not to look at these pages
1211 * (accesses can have side effects).
1212 * VM_RESERVED is specified all over the place, because
1213 * in 2.4 it kept swapout's vma scan off this vma; but
1214 * in 2.6 the LRU scan won't even find its pages, so this
1215 * flag means no more than count its pages in reserved_vm,
1216 * and omit it from core dump, even when VM_IO turned off.
1217 * VM_UNPAGED tells the core MM not to "manage" these pages
1218 * (e.g. refcount, mapcount, try to swap them out): in
1219 * particular, zap_pte_range does not try to free them.
1221 vma
->vm_flags
|= VM_IO
| VM_RESERVED
| VM_UNPAGED
;
1223 BUG_ON(addr
>= end
);
1224 pfn
-= addr
>> PAGE_SHIFT
;
1225 pgd
= pgd_offset(mm
, addr
);
1226 flush_cache_range(vma
, addr
, end
);
1228 next
= pgd_addr_end(addr
, end
);
1229 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1230 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1233 } while (pgd
++, addr
= next
, addr
!= end
);
1236 EXPORT_SYMBOL(remap_pfn_range
);
1239 * handle_pte_fault chooses page fault handler according to an entry
1240 * which was read non-atomically. Before making any commitment, on
1241 * those architectures or configurations (e.g. i386 with PAE) which
1242 * might give a mix of unmatched parts, do_swap_page and do_file_page
1243 * must check under lock before unmapping the pte and proceeding
1244 * (but do_wp_page is only called after already making such a check;
1245 * and do_anonymous_page and do_no_page can safely check later on).
1247 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
1248 pte_t
*page_table
, pte_t orig_pte
)
1251 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1252 if (sizeof(pte_t
) > sizeof(unsigned long)) {
1253 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
1255 same
= pte_same(*page_table
, orig_pte
);
1259 pte_unmap(page_table
);
1264 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1265 * servicing faults for write access. In the normal case, do always want
1266 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1267 * that do not have writing enabled, when used by access_process_vm.
1269 static inline pte_t
maybe_mkwrite(pte_t pte
, struct vm_area_struct
*vma
)
1271 if (likely(vma
->vm_flags
& VM_WRITE
))
1272 pte
= pte_mkwrite(pte
);
1277 * This routine handles present pages, when users try to write
1278 * to a shared page. It is done by copying the page to a new address
1279 * and decrementing the shared-page counter for the old page.
1281 * Note that this routine assumes that the protection checks have been
1282 * done by the caller (the low-level page fault routine in most cases).
1283 * Thus we can safely just mark it writable once we've done any necessary
1286 * We also mark the page dirty at this point even though the page will
1287 * change only once the write actually happens. This avoids a few races,
1288 * and potentially makes it more efficient.
1290 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1291 * but allow concurrent faults), with pte both mapped and locked.
1292 * We return with mmap_sem still held, but pte unmapped and unlocked.
1294 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1295 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1296 spinlock_t
*ptl
, pte_t orig_pte
)
1298 struct page
*old_page
, *src_page
, *new_page
;
1299 unsigned long pfn
= pte_pfn(orig_pte
);
1301 int ret
= VM_FAULT_MINOR
;
1303 if (unlikely(!pfn_valid(pfn
))) {
1305 * Page table corrupted: show pte and kill process.
1306 * Or it's an attempt to COW an out-of-map VM_UNPAGED
1307 * entry, which copy_user_highpage does not support.
1309 print_bad_pte(vma
, orig_pte
, address
);
1313 old_page
= pfn_to_page(pfn
);
1314 src_page
= old_page
;
1316 if (unlikely(vma
->vm_flags
& VM_UNPAGED
))
1317 if (!page_is_anon(old_page
, vma
, address
)) {
1322 if (PageAnon(old_page
) && !TestSetPageLocked(old_page
)) {
1323 int reuse
= can_share_swap_page(old_page
);
1324 unlock_page(old_page
);
1326 flush_cache_page(vma
, address
, pfn
);
1327 entry
= pte_mkyoung(orig_pte
);
1328 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1329 ptep_set_access_flags(vma
, address
, page_table
, entry
, 1);
1330 update_mmu_cache(vma
, address
, entry
);
1331 lazy_mmu_prot_update(entry
);
1332 ret
|= VM_FAULT_WRITE
;
1338 * Ok, we need to copy. Oh, well..
1340 page_cache_get(old_page
);
1342 pte_unmap_unlock(page_table
, ptl
);
1344 if (unlikely(anon_vma_prepare(vma
)))
1346 if (src_page
== ZERO_PAGE(address
)) {
1347 new_page
= alloc_zeroed_user_highpage(vma
, address
);
1351 new_page
= alloc_page_vma(GFP_HIGHUSER
, vma
, address
);
1354 copy_user_highpage(new_page
, src_page
, address
);
1358 * Re-check the pte - we dropped the lock
1360 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1361 if (likely(pte_same(*page_table
, orig_pte
))) {
1363 page_remove_rmap(old_page
);
1364 if (!PageAnon(old_page
)) {
1365 dec_mm_counter(mm
, file_rss
);
1366 inc_mm_counter(mm
, anon_rss
);
1369 inc_mm_counter(mm
, anon_rss
);
1370 flush_cache_page(vma
, address
, pfn
);
1371 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
1372 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1373 ptep_establish(vma
, address
, page_table
, entry
);
1374 update_mmu_cache(vma
, address
, entry
);
1375 lazy_mmu_prot_update(entry
);
1376 lru_cache_add_active(new_page
);
1377 page_add_anon_rmap(new_page
, vma
, address
);
1379 /* Free the old page.. */
1380 new_page
= old_page
;
1381 ret
|= VM_FAULT_WRITE
;
1384 page_cache_release(new_page
);
1386 page_cache_release(old_page
);
1388 pte_unmap_unlock(page_table
, ptl
);
1392 page_cache_release(old_page
);
1393 return VM_FAULT_OOM
;
1397 * Helper functions for unmap_mapping_range().
1399 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1401 * We have to restart searching the prio_tree whenever we drop the lock,
1402 * since the iterator is only valid while the lock is held, and anyway
1403 * a later vma might be split and reinserted earlier while lock dropped.
1405 * The list of nonlinear vmas could be handled more efficiently, using
1406 * a placeholder, but handle it in the same way until a need is shown.
1407 * It is important to search the prio_tree before nonlinear list: a vma
1408 * may become nonlinear and be shifted from prio_tree to nonlinear list
1409 * while the lock is dropped; but never shifted from list to prio_tree.
1411 * In order to make forward progress despite restarting the search,
1412 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1413 * quickly skip it next time around. Since the prio_tree search only
1414 * shows us those vmas affected by unmapping the range in question, we
1415 * can't efficiently keep all vmas in step with mapping->truncate_count:
1416 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1417 * mapping->truncate_count and vma->vm_truncate_count are protected by
1420 * In order to make forward progress despite repeatedly restarting some
1421 * large vma, note the restart_addr from unmap_vmas when it breaks out:
1422 * and restart from that address when we reach that vma again. It might
1423 * have been split or merged, shrunk or extended, but never shifted: so
1424 * restart_addr remains valid so long as it remains in the vma's range.
1425 * unmap_mapping_range forces truncate_count to leap over page-aligned
1426 * values so we can save vma's restart_addr in its truncate_count field.
1428 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1430 static void reset_vma_truncate_counts(struct address_space
*mapping
)
1432 struct vm_area_struct
*vma
;
1433 struct prio_tree_iter iter
;
1435 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, 0, ULONG_MAX
)
1436 vma
->vm_truncate_count
= 0;
1437 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
1438 vma
->vm_truncate_count
= 0;
1441 static int unmap_mapping_range_vma(struct vm_area_struct
*vma
,
1442 unsigned long start_addr
, unsigned long end_addr
,
1443 struct zap_details
*details
)
1445 unsigned long restart_addr
;
1449 restart_addr
= vma
->vm_truncate_count
;
1450 if (is_restart_addr(restart_addr
) && start_addr
< restart_addr
) {
1451 start_addr
= restart_addr
;
1452 if (start_addr
>= end_addr
) {
1453 /* Top of vma has been split off since last time */
1454 vma
->vm_truncate_count
= details
->truncate_count
;
1459 restart_addr
= zap_page_range(vma
, start_addr
,
1460 end_addr
- start_addr
, details
);
1461 need_break
= need_resched() ||
1462 need_lockbreak(details
->i_mmap_lock
);
1464 if (restart_addr
>= end_addr
) {
1465 /* We have now completed this vma: mark it so */
1466 vma
->vm_truncate_count
= details
->truncate_count
;
1470 /* Note restart_addr in vma's truncate_count field */
1471 vma
->vm_truncate_count
= restart_addr
;
1476 spin_unlock(details
->i_mmap_lock
);
1478 spin_lock(details
->i_mmap_lock
);
1482 static inline void unmap_mapping_range_tree(struct prio_tree_root
*root
,
1483 struct zap_details
*details
)
1485 struct vm_area_struct
*vma
;
1486 struct prio_tree_iter iter
;
1487 pgoff_t vba
, vea
, zba
, zea
;
1490 vma_prio_tree_foreach(vma
, &iter
, root
,
1491 details
->first_index
, details
->last_index
) {
1492 /* Skip quickly over those we have already dealt with */
1493 if (vma
->vm_truncate_count
== details
->truncate_count
)
1496 vba
= vma
->vm_pgoff
;
1497 vea
= vba
+ ((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) - 1;
1498 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1499 zba
= details
->first_index
;
1502 zea
= details
->last_index
;
1506 if (unmap_mapping_range_vma(vma
,
1507 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
1508 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
1514 static inline void unmap_mapping_range_list(struct list_head
*head
,
1515 struct zap_details
*details
)
1517 struct vm_area_struct
*vma
;
1520 * In nonlinear VMAs there is no correspondence between virtual address
1521 * offset and file offset. So we must perform an exhaustive search
1522 * across *all* the pages in each nonlinear VMA, not just the pages
1523 * whose virtual address lies outside the file truncation point.
1526 list_for_each_entry(vma
, head
, shared
.vm_set
.list
) {
1527 /* Skip quickly over those we have already dealt with */
1528 if (vma
->vm_truncate_count
== details
->truncate_count
)
1530 details
->nonlinear_vma
= vma
;
1531 if (unmap_mapping_range_vma(vma
, vma
->vm_start
,
1532 vma
->vm_end
, details
) < 0)
1538 * unmap_mapping_range - unmap the portion of all mmaps
1539 * in the specified address_space corresponding to the specified
1540 * page range in the underlying file.
1541 * @mapping: the address space containing mmaps to be unmapped.
1542 * @holebegin: byte in first page to unmap, relative to the start of
1543 * the underlying file. This will be rounded down to a PAGE_SIZE
1544 * boundary. Note that this is different from vmtruncate(), which
1545 * must keep the partial page. In contrast, we must get rid of
1547 * @holelen: size of prospective hole in bytes. This will be rounded
1548 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
1550 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1551 * but 0 when invalidating pagecache, don't throw away private data.
1553 void unmap_mapping_range(struct address_space
*mapping
,
1554 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
1556 struct zap_details details
;
1557 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
1558 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1560 /* Check for overflow. */
1561 if (sizeof(holelen
) > sizeof(hlen
)) {
1563 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1564 if (holeend
& ~(long long)ULONG_MAX
)
1565 hlen
= ULONG_MAX
- hba
+ 1;
1568 details
.check_mapping
= even_cows
? NULL
: mapping
;
1569 details
.nonlinear_vma
= NULL
;
1570 details
.first_index
= hba
;
1571 details
.last_index
= hba
+ hlen
- 1;
1572 if (details
.last_index
< details
.first_index
)
1573 details
.last_index
= ULONG_MAX
;
1574 details
.i_mmap_lock
= &mapping
->i_mmap_lock
;
1576 spin_lock(&mapping
->i_mmap_lock
);
1578 /* serialize i_size write against truncate_count write */
1580 /* Protect against page faults, and endless unmapping loops */
1581 mapping
->truncate_count
++;
1583 * For archs where spin_lock has inclusive semantics like ia64
1584 * this smp_mb() will prevent to read pagetable contents
1585 * before the truncate_count increment is visible to
1589 if (unlikely(is_restart_addr(mapping
->truncate_count
))) {
1590 if (mapping
->truncate_count
== 0)
1591 reset_vma_truncate_counts(mapping
);
1592 mapping
->truncate_count
++;
1594 details
.truncate_count
= mapping
->truncate_count
;
1596 if (unlikely(!prio_tree_empty(&mapping
->i_mmap
)))
1597 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
1598 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
1599 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
1600 spin_unlock(&mapping
->i_mmap_lock
);
1602 EXPORT_SYMBOL(unmap_mapping_range
);
1605 * Handle all mappings that got truncated by a "truncate()"
1608 * NOTE! We have to be ready to update the memory sharing
1609 * between the file and the memory map for a potential last
1610 * incomplete page. Ugly, but necessary.
1612 int vmtruncate(struct inode
* inode
, loff_t offset
)
1614 struct address_space
*mapping
= inode
->i_mapping
;
1615 unsigned long limit
;
1617 if (inode
->i_size
< offset
)
1620 * truncation of in-use swapfiles is disallowed - it would cause
1621 * subsequent swapout to scribble on the now-freed blocks.
1623 if (IS_SWAPFILE(inode
))
1625 i_size_write(inode
, offset
);
1626 unmap_mapping_range(mapping
, offset
+ PAGE_SIZE
- 1, 0, 1);
1627 truncate_inode_pages(mapping
, offset
);
1631 limit
= current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
;
1632 if (limit
!= RLIM_INFINITY
&& offset
> limit
)
1634 if (offset
> inode
->i_sb
->s_maxbytes
)
1636 i_size_write(inode
, offset
);
1639 if (inode
->i_op
&& inode
->i_op
->truncate
)
1640 inode
->i_op
->truncate(inode
);
1643 send_sig(SIGXFSZ
, current
, 0);
1650 EXPORT_SYMBOL(vmtruncate
);
1653 * Primitive swap readahead code. We simply read an aligned block of
1654 * (1 << page_cluster) entries in the swap area. This method is chosen
1655 * because it doesn't cost us any seek time. We also make sure to queue
1656 * the 'original' request together with the readahead ones...
1658 * This has been extended to use the NUMA policies from the mm triggering
1661 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1663 void swapin_readahead(swp_entry_t entry
, unsigned long addr
,struct vm_area_struct
*vma
)
1666 struct vm_area_struct
*next_vma
= vma
? vma
->vm_next
: NULL
;
1669 struct page
*new_page
;
1670 unsigned long offset
;
1673 * Get the number of handles we should do readahead io to.
1675 num
= valid_swaphandles(entry
, &offset
);
1676 for (i
= 0; i
< num
; offset
++, i
++) {
1677 /* Ok, do the async read-ahead now */
1678 new_page
= read_swap_cache_async(swp_entry(swp_type(entry
),
1679 offset
), vma
, addr
);
1682 page_cache_release(new_page
);
1685 * Find the next applicable VMA for the NUMA policy.
1691 if (addr
>= vma
->vm_end
) {
1693 next_vma
= vma
? vma
->vm_next
: NULL
;
1695 if (vma
&& addr
< vma
->vm_start
)
1698 if (next_vma
&& addr
>= next_vma
->vm_start
) {
1700 next_vma
= vma
->vm_next
;
1705 lru_add_drain(); /* Push any new pages onto the LRU now */
1709 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1710 * but allow concurrent faults), and pte mapped but not yet locked.
1711 * We return with mmap_sem still held, but pte unmapped and unlocked.
1713 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1714 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1715 int write_access
, pte_t orig_pte
)
1721 int ret
= VM_FAULT_MINOR
;
1723 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
1726 entry
= pte_to_swp_entry(orig_pte
);
1727 page
= lookup_swap_cache(entry
);
1729 swapin_readahead(entry
, address
, vma
);
1730 page
= read_swap_cache_async(entry
, vma
, address
);
1733 * Back out if somebody else faulted in this pte
1734 * while we released the pte lock.
1736 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1737 if (likely(pte_same(*page_table
, orig_pte
)))
1742 /* Had to read the page from swap area: Major fault */
1743 ret
= VM_FAULT_MAJOR
;
1744 inc_page_state(pgmajfault
);
1748 mark_page_accessed(page
);
1752 * Back out if somebody else already faulted in this pte.
1754 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1755 if (unlikely(!pte_same(*page_table
, orig_pte
)))
1758 if (unlikely(!PageUptodate(page
))) {
1759 ret
= VM_FAULT_SIGBUS
;
1763 /* The page isn't present yet, go ahead with the fault. */
1765 inc_mm_counter(mm
, anon_rss
);
1766 pte
= mk_pte(page
, vma
->vm_page_prot
);
1767 if (write_access
&& can_share_swap_page(page
)) {
1768 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
1772 flush_icache_page(vma
, page
);
1773 set_pte_at(mm
, address
, page_table
, pte
);
1774 page_add_anon_rmap(page
, vma
, address
);
1778 remove_exclusive_swap_page(page
);
1782 if (do_wp_page(mm
, vma
, address
,
1783 page_table
, pmd
, ptl
, pte
) == VM_FAULT_OOM
)
1788 /* No need to invalidate - it was non-present before */
1789 update_mmu_cache(vma
, address
, pte
);
1790 lazy_mmu_prot_update(pte
);
1792 pte_unmap_unlock(page_table
, ptl
);
1796 pte_unmap_unlock(page_table
, ptl
);
1798 page_cache_release(page
);
1803 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1804 * but allow concurrent faults), and pte mapped but not yet locked.
1805 * We return with mmap_sem still held, but pte unmapped and unlocked.
1807 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1808 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1816 * A VM_UNPAGED vma will normally be filled with present ptes
1817 * by remap_pfn_range, and never arrive here; but it might have
1818 * holes, or if !VM_DONTEXPAND, mremap might have expanded it.
1819 * It's weird enough handling anon pages in unpaged vmas, we do
1820 * not want to worry about ZERO_PAGEs too (it may or may not
1821 * matter if their counts wrap): just give them anon pages.
1824 if (write_access
|| (vma
->vm_flags
& VM_UNPAGED
)) {
1825 /* Allocate our own private page. */
1826 pte_unmap(page_table
);
1828 if (unlikely(anon_vma_prepare(vma
)))
1830 page
= alloc_zeroed_user_highpage(vma
, address
);
1834 entry
= mk_pte(page
, vma
->vm_page_prot
);
1835 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1837 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1838 if (!pte_none(*page_table
))
1840 inc_mm_counter(mm
, anon_rss
);
1841 lru_cache_add_active(page
);
1842 SetPageReferenced(page
);
1843 page_add_anon_rmap(page
, vma
, address
);
1845 /* Map the ZERO_PAGE - vm_page_prot is readonly */
1846 page
= ZERO_PAGE(address
);
1847 page_cache_get(page
);
1848 entry
= mk_pte(page
, vma
->vm_page_prot
);
1850 ptl
= pte_lockptr(mm
, pmd
);
1852 if (!pte_none(*page_table
))
1854 inc_mm_counter(mm
, file_rss
);
1855 page_add_file_rmap(page
);
1858 set_pte_at(mm
, address
, page_table
, entry
);
1860 /* No need to invalidate - it was non-present before */
1861 update_mmu_cache(vma
, address
, entry
);
1862 lazy_mmu_prot_update(entry
);
1864 pte_unmap_unlock(page_table
, ptl
);
1865 return VM_FAULT_MINOR
;
1867 page_cache_release(page
);
1870 return VM_FAULT_OOM
;
1874 * do_no_page() tries to create a new page mapping. It aggressively
1875 * tries to share with existing pages, but makes a separate copy if
1876 * the "write_access" parameter is true in order to avoid the next
1879 * As this is called only for pages that do not currently exist, we
1880 * do not need to flush old virtual caches or the TLB.
1882 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1883 * but allow concurrent faults), and pte mapped but not yet locked.
1884 * We return with mmap_sem still held, but pte unmapped and unlocked.
1886 static int do_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1887 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1891 struct page
*new_page
;
1892 struct address_space
*mapping
= NULL
;
1894 unsigned int sequence
= 0;
1895 int ret
= VM_FAULT_MINOR
;
1898 pte_unmap(page_table
);
1899 BUG_ON(vma
->vm_flags
& VM_UNPAGED
);
1902 mapping
= vma
->vm_file
->f_mapping
;
1903 sequence
= mapping
->truncate_count
;
1904 smp_rmb(); /* serializes i_size against truncate_count */
1907 new_page
= vma
->vm_ops
->nopage(vma
, address
& PAGE_MASK
, &ret
);
1909 * No smp_rmb is needed here as long as there's a full
1910 * spin_lock/unlock sequence inside the ->nopage callback
1911 * (for the pagecache lookup) that acts as an implicit
1912 * smp_mb() and prevents the i_size read to happen
1913 * after the next truncate_count read.
1916 /* no page was available -- either SIGBUS or OOM */
1917 if (new_page
== NOPAGE_SIGBUS
)
1918 return VM_FAULT_SIGBUS
;
1919 if (new_page
== NOPAGE_OOM
)
1920 return VM_FAULT_OOM
;
1923 * Should we do an early C-O-W break?
1925 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
1928 if (unlikely(anon_vma_prepare(vma
)))
1930 page
= alloc_page_vma(GFP_HIGHUSER
, vma
, address
);
1933 copy_user_highpage(page
, new_page
, address
);
1934 page_cache_release(new_page
);
1939 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1941 * For a file-backed vma, someone could have truncated or otherwise
1942 * invalidated this page. If unmap_mapping_range got called,
1943 * retry getting the page.
1945 if (mapping
&& unlikely(sequence
!= mapping
->truncate_count
)) {
1946 pte_unmap_unlock(page_table
, ptl
);
1947 page_cache_release(new_page
);
1949 sequence
= mapping
->truncate_count
;
1955 * This silly early PAGE_DIRTY setting removes a race
1956 * due to the bad i386 page protection. But it's valid
1957 * for other architectures too.
1959 * Note that if write_access is true, we either now have
1960 * an exclusive copy of the page, or this is a shared mapping,
1961 * so we can make it writable and dirty to avoid having to
1962 * handle that later.
1964 /* Only go through if we didn't race with anybody else... */
1965 if (pte_none(*page_table
)) {
1966 flush_icache_page(vma
, new_page
);
1967 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
1969 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1970 set_pte_at(mm
, address
, page_table
, entry
);
1972 inc_mm_counter(mm
, anon_rss
);
1973 lru_cache_add_active(new_page
);
1974 page_add_anon_rmap(new_page
, vma
, address
);
1976 inc_mm_counter(mm
, file_rss
);
1977 page_add_file_rmap(new_page
);
1980 /* One of our sibling threads was faster, back out. */
1981 page_cache_release(new_page
);
1985 /* no need to invalidate: a not-present page shouldn't be cached */
1986 update_mmu_cache(vma
, address
, entry
);
1987 lazy_mmu_prot_update(entry
);
1989 pte_unmap_unlock(page_table
, ptl
);
1992 page_cache_release(new_page
);
1993 return VM_FAULT_OOM
;
1997 * Fault of a previously existing named mapping. Repopulate the pte
1998 * from the encoded file_pte if possible. This enables swappable
2001 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2002 * but allow concurrent faults), and pte mapped but not yet locked.
2003 * We return with mmap_sem still held, but pte unmapped and unlocked.
2005 static int do_file_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2006 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2007 int write_access
, pte_t orig_pte
)
2012 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2013 return VM_FAULT_MINOR
;
2015 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
))) {
2017 * Page table corrupted: show pte and kill process.
2019 print_bad_pte(vma
, orig_pte
, address
);
2020 return VM_FAULT_OOM
;
2022 /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2024 pgoff
= pte_to_pgoff(orig_pte
);
2025 err
= vma
->vm_ops
->populate(vma
, address
& PAGE_MASK
, PAGE_SIZE
,
2026 vma
->vm_page_prot
, pgoff
, 0);
2028 return VM_FAULT_OOM
;
2030 return VM_FAULT_SIGBUS
;
2031 return VM_FAULT_MAJOR
;
2035 * These routines also need to handle stuff like marking pages dirty
2036 * and/or accessed for architectures that don't do it in hardware (most
2037 * RISC architectures). The early dirtying is also good on the i386.
2039 * There is also a hook called "update_mmu_cache()" that architectures
2040 * with external mmu caches can use to update those (ie the Sparc or
2041 * PowerPC hashed page tables that act as extended TLBs).
2043 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2044 * but allow concurrent faults), and pte mapped but not yet locked.
2045 * We return with mmap_sem still held, but pte unmapped and unlocked.
2047 static inline int handle_pte_fault(struct mm_struct
*mm
,
2048 struct vm_area_struct
*vma
, unsigned long address
,
2049 pte_t
*pte
, pmd_t
*pmd
, int write_access
)
2055 old_entry
= entry
= *pte
;
2056 if (!pte_present(entry
)) {
2057 if (pte_none(entry
)) {
2058 if (!vma
->vm_ops
|| !vma
->vm_ops
->nopage
)
2059 return do_anonymous_page(mm
, vma
, address
,
2060 pte
, pmd
, write_access
);
2061 return do_no_page(mm
, vma
, address
,
2062 pte
, pmd
, write_access
);
2064 if (pte_file(entry
))
2065 return do_file_page(mm
, vma
, address
,
2066 pte
, pmd
, write_access
, entry
);
2067 return do_swap_page(mm
, vma
, address
,
2068 pte
, pmd
, write_access
, entry
);
2071 ptl
= pte_lockptr(mm
, pmd
);
2073 if (unlikely(!pte_same(*pte
, entry
)))
2076 if (!pte_write(entry
))
2077 return do_wp_page(mm
, vma
, address
,
2078 pte
, pmd
, ptl
, entry
);
2079 entry
= pte_mkdirty(entry
);
2081 entry
= pte_mkyoung(entry
);
2082 if (!pte_same(old_entry
, entry
)) {
2083 ptep_set_access_flags(vma
, address
, pte
, entry
, write_access
);
2084 update_mmu_cache(vma
, address
, entry
);
2085 lazy_mmu_prot_update(entry
);
2088 * This is needed only for protection faults but the arch code
2089 * is not yet telling us if this is a protection fault or not.
2090 * This still avoids useless tlb flushes for .text page faults
2094 flush_tlb_page(vma
, address
);
2097 pte_unmap_unlock(pte
, ptl
);
2098 return VM_FAULT_MINOR
;
2102 * By the time we get here, we already hold the mm semaphore
2104 int __handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2105 unsigned long address
, int write_access
)
2112 __set_current_state(TASK_RUNNING
);
2114 inc_page_state(pgfault
);
2116 if (unlikely(is_vm_hugetlb_page(vma
)))
2117 return hugetlb_fault(mm
, vma
, address
, write_access
);
2119 pgd
= pgd_offset(mm
, address
);
2120 pud
= pud_alloc(mm
, pgd
, address
);
2122 return VM_FAULT_OOM
;
2123 pmd
= pmd_alloc(mm
, pud
, address
);
2125 return VM_FAULT_OOM
;
2126 pte
= pte_alloc_map(mm
, pmd
, address
);
2128 return VM_FAULT_OOM
;
2130 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, write_access
);
2133 #ifndef __PAGETABLE_PUD_FOLDED
2135 * Allocate page upper directory.
2136 * We've already handled the fast-path in-line.
2138 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
2140 pud_t
*new = pud_alloc_one(mm
, address
);
2144 spin_lock(&mm
->page_table_lock
);
2145 if (pgd_present(*pgd
)) /* Another has populated it */
2148 pgd_populate(mm
, pgd
, new);
2149 spin_unlock(&mm
->page_table_lock
);
2152 #endif /* __PAGETABLE_PUD_FOLDED */
2154 #ifndef __PAGETABLE_PMD_FOLDED
2156 * Allocate page middle directory.
2157 * We've already handled the fast-path in-line.
2159 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
2161 pmd_t
*new = pmd_alloc_one(mm
, address
);
2165 spin_lock(&mm
->page_table_lock
);
2166 #ifndef __ARCH_HAS_4LEVEL_HACK
2167 if (pud_present(*pud
)) /* Another has populated it */
2170 pud_populate(mm
, pud
, new);
2172 if (pgd_present(*pud
)) /* Another has populated it */
2175 pgd_populate(mm
, pud
, new);
2176 #endif /* __ARCH_HAS_4LEVEL_HACK */
2177 spin_unlock(&mm
->page_table_lock
);
2180 #endif /* __PAGETABLE_PMD_FOLDED */
2182 int make_pages_present(unsigned long addr
, unsigned long end
)
2184 int ret
, len
, write
;
2185 struct vm_area_struct
* vma
;
2187 vma
= find_vma(current
->mm
, addr
);
2190 write
= (vma
->vm_flags
& VM_WRITE
) != 0;
2193 if (end
> vma
->vm_end
)
2195 len
= (end
+PAGE_SIZE
-1)/PAGE_SIZE
-addr
/PAGE_SIZE
;
2196 ret
= get_user_pages(current
, current
->mm
, addr
,
2197 len
, write
, 0, NULL
, NULL
);
2200 return ret
== len
? 0 : -1;
2204 * Map a vmalloc()-space virtual address to the physical page.
2206 struct page
* vmalloc_to_page(void * vmalloc_addr
)
2208 unsigned long addr
= (unsigned long) vmalloc_addr
;
2209 struct page
*page
= NULL
;
2210 pgd_t
*pgd
= pgd_offset_k(addr
);
2215 if (!pgd_none(*pgd
)) {
2216 pud
= pud_offset(pgd
, addr
);
2217 if (!pud_none(*pud
)) {
2218 pmd
= pmd_offset(pud
, addr
);
2219 if (!pmd_none(*pmd
)) {
2220 ptep
= pte_offset_map(pmd
, addr
);
2222 if (pte_present(pte
))
2223 page
= pte_page(pte
);
2231 EXPORT_SYMBOL(vmalloc_to_page
);
2234 * Map a vmalloc()-space virtual address to the physical page frame number.
2236 unsigned long vmalloc_to_pfn(void * vmalloc_addr
)
2238 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
2241 EXPORT_SYMBOL(vmalloc_to_pfn
);
2243 #if !defined(__HAVE_ARCH_GATE_AREA)
2245 #if defined(AT_SYSINFO_EHDR)
2246 static struct vm_area_struct gate_vma
;
2248 static int __init
gate_vma_init(void)
2250 gate_vma
.vm_mm
= NULL
;
2251 gate_vma
.vm_start
= FIXADDR_USER_START
;
2252 gate_vma
.vm_end
= FIXADDR_USER_END
;
2253 gate_vma
.vm_page_prot
= PAGE_READONLY
;
2254 gate_vma
.vm_flags
= 0;
2257 __initcall(gate_vma_init
);
2260 struct vm_area_struct
*get_gate_vma(struct task_struct
*tsk
)
2262 #ifdef AT_SYSINFO_EHDR
2269 int in_gate_area_no_task(unsigned long addr
)
2271 #ifdef AT_SYSINFO_EHDR
2272 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
2278 #endif /* __HAVE_ARCH_GATE_AREA */