4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 #include <linux/kernel_stat.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/init.h>
52 #include <asm/pgalloc.h>
53 #include <asm/uaccess.h>
55 #include <asm/tlbflush.h>
56 #include <asm/pgtable.h>
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
61 #ifndef CONFIG_NEED_MULTIPLE_NODES
62 /* use the per-pgdat data instead for discontigmem - mbligh */
63 unsigned long max_mapnr
;
66 EXPORT_SYMBOL(max_mapnr
);
67 EXPORT_SYMBOL(mem_map
);
70 unsigned long num_physpages
;
72 * A number of key systems in x86 including ioremap() rely on the assumption
73 * that high_memory defines the upper bound on direct map memory, then end
74 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
75 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
79 unsigned long vmalloc_earlyreserve
;
81 EXPORT_SYMBOL(num_physpages
);
82 EXPORT_SYMBOL(high_memory
);
83 EXPORT_SYMBOL(vmalloc_earlyreserve
);
86 * If a p?d_bad entry is found while walking page tables, report
87 * the error, before resetting entry to p?d_none. Usually (but
88 * very seldom) called out from the p?d_none_or_clear_bad macros.
91 void pgd_clear_bad(pgd_t
*pgd
)
97 void pud_clear_bad(pud_t
*pud
)
103 void pmd_clear_bad(pmd_t
*pmd
)
110 * Note: this doesn't free the actual pages themselves. That
111 * has been handled earlier when unmapping all the memory regions.
113 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
)
115 struct page
*page
= pmd_page(*pmd
);
117 pte_free_tlb(tlb
, page
);
118 dec_page_state(nr_page_table_pages
);
122 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
123 unsigned long addr
, unsigned long end
,
124 unsigned long floor
, unsigned long ceiling
)
131 pmd
= pmd_offset(pud
, addr
);
133 next
= pmd_addr_end(addr
, end
);
134 if (pmd_none_or_clear_bad(pmd
))
136 free_pte_range(tlb
, pmd
);
137 } while (pmd
++, addr
= next
, addr
!= end
);
147 if (end
- 1 > ceiling
- 1)
150 pmd
= pmd_offset(pud
, start
);
152 pmd_free_tlb(tlb
, pmd
);
155 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
156 unsigned long addr
, unsigned long end
,
157 unsigned long floor
, unsigned long ceiling
)
164 pud
= pud_offset(pgd
, addr
);
166 next
= pud_addr_end(addr
, end
);
167 if (pud_none_or_clear_bad(pud
))
169 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
170 } while (pud
++, addr
= next
, addr
!= end
);
176 ceiling
&= PGDIR_MASK
;
180 if (end
- 1 > ceiling
- 1)
183 pud
= pud_offset(pgd
, start
);
185 pud_free_tlb(tlb
, pud
);
189 * This function frees user-level page tables of a process.
191 * Must be called with pagetable lock held.
193 void free_pgd_range(struct mmu_gather
**tlb
,
194 unsigned long addr
, unsigned long end
,
195 unsigned long floor
, unsigned long ceiling
)
202 * The next few lines have given us lots of grief...
204 * Why are we testing PMD* at this top level? Because often
205 * there will be no work to do at all, and we'd prefer not to
206 * go all the way down to the bottom just to discover that.
208 * Why all these "- 1"s? Because 0 represents both the bottom
209 * of the address space and the top of it (using -1 for the
210 * top wouldn't help much: the masks would do the wrong thing).
211 * The rule is that addr 0 and floor 0 refer to the bottom of
212 * the address space, but end 0 and ceiling 0 refer to the top
213 * Comparisons need to use "end - 1" and "ceiling - 1" (though
214 * that end 0 case should be mythical).
216 * Wherever addr is brought up or ceiling brought down, we must
217 * be careful to reject "the opposite 0" before it confuses the
218 * subsequent tests. But what about where end is brought down
219 * by PMD_SIZE below? no, end can't go down to 0 there.
221 * Whereas we round start (addr) and ceiling down, by different
222 * masks at different levels, in order to test whether a table
223 * now has no other vmas using it, so can be freed, we don't
224 * bother to round floor or end up - the tests don't need that.
238 if (end
- 1 > ceiling
- 1)
244 pgd
= pgd_offset((*tlb
)->mm
, addr
);
246 next
= pgd_addr_end(addr
, end
);
247 if (pgd_none_or_clear_bad(pgd
))
249 free_pud_range(*tlb
, pgd
, addr
, next
, floor
, ceiling
);
250 } while (pgd
++, addr
= next
, addr
!= end
);
253 flush_tlb_pgtables((*tlb
)->mm
, start
, end
);
256 void free_pgtables(struct mmu_gather
**tlb
, struct vm_area_struct
*vma
,
257 unsigned long floor
, unsigned long ceiling
)
260 struct vm_area_struct
*next
= vma
->vm_next
;
261 unsigned long addr
= vma
->vm_start
;
264 * Hide vma from rmap and vmtruncate before freeing pgtables
266 anon_vma_unlink(vma
);
267 unlink_file_vma(vma
);
269 if (is_hugepage_only_range(vma
->vm_mm
, addr
, HPAGE_SIZE
)) {
270 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
271 floor
, next
? next
->vm_start
: ceiling
);
274 * Optimization: gather nearby vmas into one call down
276 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
277 && !is_hugepage_only_range(vma
->vm_mm
, next
->vm_start
,
281 anon_vma_unlink(vma
);
282 unlink_file_vma(vma
);
284 free_pgd_range(tlb
, addr
, vma
->vm_end
,
285 floor
, next
? next
->vm_start
: ceiling
);
291 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
, unsigned long address
)
293 struct page
*new = pte_alloc_one(mm
, address
);
297 spin_lock(&mm
->page_table_lock
);
298 if (pmd_present(*pmd
)) /* Another has populated it */
302 inc_page_state(nr_page_table_pages
);
303 pmd_populate(mm
, pmd
, new);
305 spin_unlock(&mm
->page_table_lock
);
309 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
311 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
315 spin_lock(&init_mm
.page_table_lock
);
316 if (pmd_present(*pmd
)) /* Another has populated it */
317 pte_free_kernel(new);
319 pmd_populate_kernel(&init_mm
, pmd
, new);
320 spin_unlock(&init_mm
.page_table_lock
);
324 static inline void add_mm_rss(struct mm_struct
*mm
, int file_rss
, int anon_rss
)
327 add_mm_counter(mm
, file_rss
, file_rss
);
329 add_mm_counter(mm
, anon_rss
, anon_rss
);
333 * This function is called to print an error when a pte in a
334 * !VM_RESERVED region is found pointing to an invalid pfn (which
337 * The calling function must still handle the error.
339 void print_bad_pte(struct vm_area_struct
*vma
, pte_t pte
, unsigned long vaddr
)
341 printk(KERN_ERR
"Bad pte = %08llx, process = %s, "
342 "vm_flags = %lx, vaddr = %lx\n",
343 (long long)pte_val(pte
),
344 (vma
->vm_mm
== current
->mm
? current
->comm
: "???"),
345 vma
->vm_flags
, vaddr
);
350 * copy one vm_area from one task to the other. Assumes the page tables
351 * already present in the new task to be cleared in the whole range
352 * covered by this vma.
356 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
357 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
358 unsigned long addr
, int *rss
)
360 unsigned long vm_flags
= vma
->vm_flags
;
361 pte_t pte
= *src_pte
;
365 /* pte contains position in swap or file, so copy. */
366 if (unlikely(!pte_present(pte
))) {
367 if (!pte_file(pte
)) {
368 swap_duplicate(pte_to_swp_entry(pte
));
369 /* make sure dst_mm is on swapoff's mmlist. */
370 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
371 spin_lock(&mmlist_lock
);
372 list_add(&dst_mm
->mmlist
, &src_mm
->mmlist
);
373 spin_unlock(&mmlist_lock
);
379 /* If the region is VM_RESERVED, the mapping is not
380 * mapped via rmap - duplicate the pte as is.
382 if (vm_flags
& VM_RESERVED
)
386 /* If the pte points outside of valid memory but
387 * the region is not VM_RESERVED, we have a problem.
389 if (unlikely(!pfn_valid(pfn
))) {
390 print_bad_pte(vma
, pte
, addr
);
391 goto out_set_pte
; /* try to do something sane */
394 page
= pfn_to_page(pfn
);
397 * If it's a COW mapping, write protect it both
398 * in the parent and the child
400 if ((vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
) {
401 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
406 * If it's a shared mapping, mark it clean in
409 if (vm_flags
& VM_SHARED
)
410 pte
= pte_mkclean(pte
);
411 pte
= pte_mkold(pte
);
414 rss
[!!PageAnon(page
)]++;
417 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
420 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
421 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
422 unsigned long addr
, unsigned long end
)
424 pte_t
*src_pte
, *dst_pte
;
425 spinlock_t
*src_ptl
, *dst_ptl
;
431 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
434 src_pte
= pte_offset_map_nested(src_pmd
, addr
);
435 src_ptl
= &src_mm
->page_table_lock
;
440 * We are holding two locks at this point - either of them
441 * could generate latencies in another task on another CPU.
443 if (progress
>= 32) {
445 if (need_resched() ||
446 need_lockbreak(src_ptl
) ||
447 need_lockbreak(dst_ptl
))
450 if (pte_none(*src_pte
)) {
454 copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
, vma
, addr
, rss
);
456 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
458 spin_unlock(src_ptl
);
459 pte_unmap_nested(src_pte
- 1);
460 add_mm_rss(dst_mm
, rss
[0], rss
[1]);
461 pte_unmap_unlock(dst_pte
- 1, dst_ptl
);
468 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
469 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
470 unsigned long addr
, unsigned long end
)
472 pmd_t
*src_pmd
, *dst_pmd
;
475 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
478 src_pmd
= pmd_offset(src_pud
, addr
);
480 next
= pmd_addr_end(addr
, end
);
481 if (pmd_none_or_clear_bad(src_pmd
))
483 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
486 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
490 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
491 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
492 unsigned long addr
, unsigned long end
)
494 pud_t
*src_pud
, *dst_pud
;
497 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
500 src_pud
= pud_offset(src_pgd
, addr
);
502 next
= pud_addr_end(addr
, end
);
503 if (pud_none_or_clear_bad(src_pud
))
505 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
508 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
512 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
513 struct vm_area_struct
*vma
)
515 pgd_t
*src_pgd
, *dst_pgd
;
517 unsigned long addr
= vma
->vm_start
;
518 unsigned long end
= vma
->vm_end
;
521 * Don't copy ptes where a page fault will fill them correctly.
522 * Fork becomes much lighter when there are big shared or private
523 * readonly mappings. The tradeoff is that copy_page_range is more
524 * efficient than faulting.
526 if (!(vma
->vm_flags
& (VM_HUGETLB
|VM_NONLINEAR
|VM_RESERVED
))) {
531 if (is_vm_hugetlb_page(vma
))
532 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
534 dst_pgd
= pgd_offset(dst_mm
, addr
);
535 src_pgd
= pgd_offset(src_mm
, addr
);
537 next
= pgd_addr_end(addr
, end
);
538 if (pgd_none_or_clear_bad(src_pgd
))
540 if (copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
543 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
547 static void zap_pte_range(struct mmu_gather
*tlb
,
548 struct vm_area_struct
*vma
, pmd_t
*pmd
,
549 unsigned long addr
, unsigned long end
,
550 struct zap_details
*details
)
552 struct mm_struct
*mm
= tlb
->mm
;
558 pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
563 if (pte_present(ptent
)) {
564 struct page
*page
= NULL
;
565 if (!(vma
->vm_flags
& VM_RESERVED
)) {
566 unsigned long pfn
= pte_pfn(ptent
);
567 if (unlikely(!pfn_valid(pfn
)))
568 print_bad_pte(vma
, ptent
, addr
);
570 page
= pfn_to_page(pfn
);
572 if (unlikely(details
) && page
) {
574 * unmap_shared_mapping_pages() wants to
575 * invalidate cache without truncating:
576 * unmap shared but keep private pages.
578 if (details
->check_mapping
&&
579 details
->check_mapping
!= page
->mapping
)
582 * Each page->index must be checked when
583 * invalidating or truncating nonlinear.
585 if (details
->nonlinear_vma
&&
586 (page
->index
< details
->first_index
||
587 page
->index
> details
->last_index
))
590 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
592 tlb_remove_tlb_entry(tlb
, pte
, addr
);
595 if (unlikely(details
) && details
->nonlinear_vma
596 && linear_page_index(details
->nonlinear_vma
,
597 addr
) != page
->index
)
598 set_pte_at(mm
, addr
, pte
,
599 pgoff_to_pte(page
->index
));
603 if (pte_dirty(ptent
))
604 set_page_dirty(page
);
605 if (pte_young(ptent
))
606 mark_page_accessed(page
);
609 page_remove_rmap(page
);
610 tlb_remove_page(tlb
, page
);
614 * If details->check_mapping, we leave swap entries;
615 * if details->nonlinear_vma, we leave file entries.
617 if (unlikely(details
))
619 if (!pte_file(ptent
))
620 free_swap_and_cache(pte_to_swp_entry(ptent
));
621 pte_clear_full(mm
, addr
, pte
, tlb
->fullmm
);
622 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
624 add_mm_rss(mm
, file_rss
, anon_rss
);
625 pte_unmap_unlock(pte
- 1, ptl
);
628 static inline void zap_pmd_range(struct mmu_gather
*tlb
,
629 struct vm_area_struct
*vma
, pud_t
*pud
,
630 unsigned long addr
, unsigned long end
,
631 struct zap_details
*details
)
636 pmd
= pmd_offset(pud
, addr
);
638 next
= pmd_addr_end(addr
, end
);
639 if (pmd_none_or_clear_bad(pmd
))
641 zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
642 } while (pmd
++, addr
= next
, addr
!= end
);
645 static inline void zap_pud_range(struct mmu_gather
*tlb
,
646 struct vm_area_struct
*vma
, pgd_t
*pgd
,
647 unsigned long addr
, unsigned long end
,
648 struct zap_details
*details
)
653 pud
= pud_offset(pgd
, addr
);
655 next
= pud_addr_end(addr
, end
);
656 if (pud_none_or_clear_bad(pud
))
658 zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
659 } while (pud
++, addr
= next
, addr
!= end
);
662 static void unmap_page_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
663 unsigned long addr
, unsigned long end
,
664 struct zap_details
*details
)
669 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
673 tlb_start_vma(tlb
, vma
);
674 pgd
= pgd_offset(vma
->vm_mm
, addr
);
676 next
= pgd_addr_end(addr
, end
);
677 if (pgd_none_or_clear_bad(pgd
))
679 zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
680 } while (pgd
++, addr
= next
, addr
!= end
);
681 tlb_end_vma(tlb
, vma
);
684 #ifdef CONFIG_PREEMPT
685 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
687 /* No preempt: go for improved straight-line efficiency */
688 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
692 * unmap_vmas - unmap a range of memory covered by a list of vma's
693 * @tlbp: address of the caller's struct mmu_gather
694 * @vma: the starting vma
695 * @start_addr: virtual address at which to start unmapping
696 * @end_addr: virtual address at which to end unmapping
697 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
698 * @details: details of nonlinear truncation or shared cache invalidation
700 * Returns the end address of the unmapping (restart addr if interrupted).
702 * Unmap all pages in the vma list.
704 * We aim to not hold locks for too long (for scheduling latency reasons).
705 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
706 * return the ending mmu_gather to the caller.
708 * Only addresses between `start' and `end' will be unmapped.
710 * The VMA list must be sorted in ascending virtual address order.
712 * unmap_vmas() assumes that the caller will flush the whole unmapped address
713 * range after unmap_vmas() returns. So the only responsibility here is to
714 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
715 * drops the lock and schedules.
717 unsigned long unmap_vmas(struct mmu_gather
**tlbp
,
718 struct vm_area_struct
*vma
, unsigned long start_addr
,
719 unsigned long end_addr
, unsigned long *nr_accounted
,
720 struct zap_details
*details
)
722 unsigned long zap_bytes
= ZAP_BLOCK_SIZE
;
723 unsigned long tlb_start
= 0; /* For tlb_finish_mmu */
724 int tlb_start_valid
= 0;
725 unsigned long start
= start_addr
;
726 spinlock_t
*i_mmap_lock
= details
? details
->i_mmap_lock
: NULL
;
727 int fullmm
= (*tlbp
)->fullmm
;
729 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
) {
732 start
= max(vma
->vm_start
, start_addr
);
733 if (start
>= vma
->vm_end
)
735 end
= min(vma
->vm_end
, end_addr
);
736 if (end
<= vma
->vm_start
)
739 if (vma
->vm_flags
& VM_ACCOUNT
)
740 *nr_accounted
+= (end
- start
) >> PAGE_SHIFT
;
742 while (start
!= end
) {
745 if (!tlb_start_valid
) {
750 if (is_vm_hugetlb_page(vma
)) {
752 unmap_hugepage_range(vma
, start
, end
);
754 block
= min(zap_bytes
, end
- start
);
755 unmap_page_range(*tlbp
, vma
, start
,
756 start
+ block
, details
);
761 if ((long)zap_bytes
> 0)
764 tlb_finish_mmu(*tlbp
, tlb_start
, start
);
766 if (need_resched() ||
767 (i_mmap_lock
&& need_lockbreak(i_mmap_lock
))) {
775 *tlbp
= tlb_gather_mmu(vma
->vm_mm
, fullmm
);
777 zap_bytes
= ZAP_BLOCK_SIZE
;
781 return start
; /* which is now the end (or restart) address */
785 * zap_page_range - remove user pages in a given range
786 * @vma: vm_area_struct holding the applicable pages
787 * @address: starting address of pages to zap
788 * @size: number of bytes to zap
789 * @details: details of nonlinear truncation or shared cache invalidation
791 unsigned long zap_page_range(struct vm_area_struct
*vma
, unsigned long address
,
792 unsigned long size
, struct zap_details
*details
)
794 struct mm_struct
*mm
= vma
->vm_mm
;
795 struct mmu_gather
*tlb
;
796 unsigned long end
= address
+ size
;
797 unsigned long nr_accounted
= 0;
800 tlb
= tlb_gather_mmu(mm
, 0);
801 update_hiwater_rss(mm
);
802 end
= unmap_vmas(&tlb
, vma
, address
, end
, &nr_accounted
, details
);
804 tlb_finish_mmu(tlb
, address
, end
);
809 * Do a quick page-table lookup for a single page.
811 struct page
*follow_page(struct mm_struct
*mm
, unsigned long address
,
822 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
824 BUG_ON(flags
& FOLL_GET
);
829 pgd
= pgd_offset(mm
, address
);
830 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
833 pud
= pud_offset(pgd
, address
);
834 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
837 pmd
= pmd_offset(pud
, address
);
838 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
841 if (pmd_huge(*pmd
)) {
842 BUG_ON(flags
& FOLL_GET
);
843 page
= follow_huge_pmd(mm
, address
, pmd
, flags
& FOLL_WRITE
);
847 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
852 if (!pte_present(pte
))
854 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
860 page
= pfn_to_page(pfn
);
861 if (flags
& FOLL_GET
)
863 if (flags
& FOLL_TOUCH
) {
864 if ((flags
& FOLL_WRITE
) &&
865 !pte_dirty(pte
) && !PageDirty(page
))
866 set_page_dirty(page
);
867 mark_page_accessed(page
);
870 pte_unmap_unlock(ptep
, ptl
);
876 * When core dumping an enormous anonymous area that nobody
877 * has touched so far, we don't want to allocate page tables.
879 if (flags
& FOLL_ANON
) {
880 page
= ZERO_PAGE(address
);
881 if (flags
& FOLL_GET
)
883 BUG_ON(flags
& FOLL_WRITE
);
888 int get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
889 unsigned long start
, int len
, int write
, int force
,
890 struct page
**pages
, struct vm_area_struct
**vmas
)
893 unsigned int vm_flags
;
896 * Require read or write permissions.
897 * If 'force' is set, we only require the "MAY" flags.
899 vm_flags
= write
? (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
900 vm_flags
&= force
? (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
904 struct vm_area_struct
*vma
;
905 unsigned int foll_flags
;
907 vma
= find_extend_vma(mm
, start
);
908 if (!vma
&& in_gate_area(tsk
, start
)) {
909 unsigned long pg
= start
& PAGE_MASK
;
910 struct vm_area_struct
*gate_vma
= get_gate_vma(tsk
);
915 if (write
) /* user gate pages are read-only */
916 return i
? : -EFAULT
;
918 pgd
= pgd_offset_k(pg
);
920 pgd
= pgd_offset_gate(mm
, pg
);
921 BUG_ON(pgd_none(*pgd
));
922 pud
= pud_offset(pgd
, pg
);
923 BUG_ON(pud_none(*pud
));
924 pmd
= pmd_offset(pud
, pg
);
926 return i
? : -EFAULT
;
927 pte
= pte_offset_map(pmd
, pg
);
928 if (pte_none(*pte
)) {
930 return i
? : -EFAULT
;
933 pages
[i
] = pte_page(*pte
);
945 if (!vma
|| (vma
->vm_flags
& (VM_IO
| VM_RESERVED
))
946 || !(vm_flags
& vma
->vm_flags
))
947 return i
? : -EFAULT
;
949 if (is_vm_hugetlb_page(vma
)) {
950 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
955 foll_flags
= FOLL_TOUCH
;
957 foll_flags
|= FOLL_GET
;
958 if (!write
&& !(vma
->vm_flags
& VM_LOCKED
) &&
959 (!vma
->vm_ops
|| !vma
->vm_ops
->nopage
))
960 foll_flags
|= FOLL_ANON
;
966 foll_flags
|= FOLL_WRITE
;
969 while (!(page
= follow_page(mm
, start
, foll_flags
))) {
971 ret
= __handle_mm_fault(mm
, vma
, start
,
972 foll_flags
& FOLL_WRITE
);
974 * The VM_FAULT_WRITE bit tells us that do_wp_page has
975 * broken COW when necessary, even if maybe_mkwrite
976 * decided not to set pte_write. We can thus safely do
977 * subsequent page lookups as if they were reads.
979 if (ret
& VM_FAULT_WRITE
)
980 foll_flags
&= ~FOLL_WRITE
;
982 switch (ret
& ~VM_FAULT_WRITE
) {
989 case VM_FAULT_SIGBUS
:
990 return i
? i
: -EFAULT
;
992 return i
? i
: -ENOMEM
;
999 flush_dcache_page(page
);
1006 } while (len
&& start
< vma
->vm_end
);
1010 EXPORT_SYMBOL(get_user_pages
);
1012 static int zeromap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1013 unsigned long addr
, unsigned long end
, pgprot_t prot
)
1018 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1022 struct page
*page
= ZERO_PAGE(addr
);
1023 pte_t zero_pte
= pte_wrprotect(mk_pte(page
, prot
));
1024 page_cache_get(page
);
1025 page_add_file_rmap(page
);
1026 inc_mm_counter(mm
, file_rss
);
1027 BUG_ON(!pte_none(*pte
));
1028 set_pte_at(mm
, addr
, pte
, zero_pte
);
1029 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1030 pte_unmap_unlock(pte
- 1, ptl
);
1034 static inline int zeromap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1035 unsigned long addr
, unsigned long end
, pgprot_t prot
)
1040 pmd
= pmd_alloc(mm
, pud
, addr
);
1044 next
= pmd_addr_end(addr
, end
);
1045 if (zeromap_pte_range(mm
, pmd
, addr
, next
, prot
))
1047 } while (pmd
++, addr
= next
, addr
!= end
);
1051 static inline int zeromap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1052 unsigned long addr
, unsigned long end
, pgprot_t prot
)
1057 pud
= pud_alloc(mm
, pgd
, addr
);
1061 next
= pud_addr_end(addr
, end
);
1062 if (zeromap_pmd_range(mm
, pud
, addr
, next
, prot
))
1064 } while (pud
++, addr
= next
, addr
!= end
);
1068 int zeromap_page_range(struct vm_area_struct
*vma
,
1069 unsigned long addr
, unsigned long size
, pgprot_t prot
)
1073 unsigned long end
= addr
+ size
;
1074 struct mm_struct
*mm
= vma
->vm_mm
;
1077 BUG_ON(addr
>= end
);
1078 pgd
= pgd_offset(mm
, addr
);
1079 flush_cache_range(vma
, addr
, end
);
1081 next
= pgd_addr_end(addr
, end
);
1082 err
= zeromap_pud_range(mm
, pgd
, addr
, next
, prot
);
1085 } while (pgd
++, addr
= next
, addr
!= end
);
1090 * maps a range of physical memory into the requested pages. the old
1091 * mappings are removed. any references to nonexistent pages results
1092 * in null mappings (currently treated as "copy-on-access")
1094 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1095 unsigned long addr
, unsigned long end
,
1096 unsigned long pfn
, pgprot_t prot
)
1101 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1105 BUG_ON(!pte_none(*pte
));
1106 set_pte_at(mm
, addr
, pte
, pfn_pte(pfn
, prot
));
1108 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1109 pte_unmap_unlock(pte
- 1, ptl
);
1113 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1114 unsigned long addr
, unsigned long end
,
1115 unsigned long pfn
, pgprot_t prot
)
1120 pfn
-= addr
>> PAGE_SHIFT
;
1121 pmd
= pmd_alloc(mm
, pud
, addr
);
1125 next
= pmd_addr_end(addr
, end
);
1126 if (remap_pte_range(mm
, pmd
, addr
, next
,
1127 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1129 } while (pmd
++, addr
= next
, addr
!= end
);
1133 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1134 unsigned long addr
, unsigned long end
,
1135 unsigned long pfn
, pgprot_t prot
)
1140 pfn
-= addr
>> PAGE_SHIFT
;
1141 pud
= pud_alloc(mm
, pgd
, addr
);
1145 next
= pud_addr_end(addr
, end
);
1146 if (remap_pmd_range(mm
, pud
, addr
, next
,
1147 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1149 } while (pud
++, addr
= next
, addr
!= end
);
1153 /* Note: this is only safe if the mm semaphore is held when called. */
1154 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1155 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1159 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1160 struct mm_struct
*mm
= vma
->vm_mm
;
1164 * Physically remapped pages are special. Tell the
1165 * rest of the world about it:
1166 * VM_IO tells people not to look at these pages
1167 * (accesses can have side effects).
1168 * VM_RESERVED tells the core MM not to "manage" these pages
1169 * (e.g. refcount, mapcount, try to swap them out).
1171 vma
->vm_flags
|= VM_IO
| VM_RESERVED
;
1173 BUG_ON(addr
>= end
);
1174 pfn
-= addr
>> PAGE_SHIFT
;
1175 pgd
= pgd_offset(mm
, addr
);
1176 flush_cache_range(vma
, addr
, end
);
1178 next
= pgd_addr_end(addr
, end
);
1179 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1180 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1183 } while (pgd
++, addr
= next
, addr
!= end
);
1186 EXPORT_SYMBOL(remap_pfn_range
);
1189 * handle_pte_fault chooses page fault handler according to an entry
1190 * which was read non-atomically. Before making any commitment, on
1191 * those architectures or configurations (e.g. i386 with PAE) which
1192 * might give a mix of unmatched parts, do_swap_page and do_file_page
1193 * must check under lock before unmapping the pte and proceeding
1194 * (but do_wp_page is only called after already making such a check;
1195 * and do_anonymous_page and do_no_page can safely check later on).
1197 static inline int pte_unmap_same(struct mm_struct
*mm
,
1198 pte_t
*page_table
, pte_t orig_pte
)
1201 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1202 if (sizeof(pte_t
) > sizeof(unsigned long)) {
1203 spin_lock(&mm
->page_table_lock
);
1204 same
= pte_same(*page_table
, orig_pte
);
1205 spin_unlock(&mm
->page_table_lock
);
1208 pte_unmap(page_table
);
1213 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1214 * servicing faults for write access. In the normal case, do always want
1215 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1216 * that do not have writing enabled, when used by access_process_vm.
1218 static inline pte_t
maybe_mkwrite(pte_t pte
, struct vm_area_struct
*vma
)
1220 if (likely(vma
->vm_flags
& VM_WRITE
))
1221 pte
= pte_mkwrite(pte
);
1226 * This routine handles present pages, when users try to write
1227 * to a shared page. It is done by copying the page to a new address
1228 * and decrementing the shared-page counter for the old page.
1230 * Note that this routine assumes that the protection checks have been
1231 * done by the caller (the low-level page fault routine in most cases).
1232 * Thus we can safely just mark it writable once we've done any necessary
1235 * We also mark the page dirty at this point even though the page will
1236 * change only once the write actually happens. This avoids a few races,
1237 * and potentially makes it more efficient.
1239 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1240 * but allow concurrent faults), with pte both mapped and locked.
1241 * We return with mmap_sem still held, but pte unmapped and unlocked.
1243 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1244 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1245 spinlock_t
*ptl
, pte_t orig_pte
)
1247 struct page
*old_page
, *new_page
;
1248 unsigned long pfn
= pte_pfn(orig_pte
);
1250 int ret
= VM_FAULT_MINOR
;
1252 BUG_ON(vma
->vm_flags
& VM_RESERVED
);
1254 if (unlikely(!pfn_valid(pfn
))) {
1256 * Page table corrupted: show pte and kill process.
1258 print_bad_pte(vma
, orig_pte
, address
);
1262 old_page
= pfn_to_page(pfn
);
1264 if (PageAnon(old_page
) && !TestSetPageLocked(old_page
)) {
1265 int reuse
= can_share_swap_page(old_page
);
1266 unlock_page(old_page
);
1268 flush_cache_page(vma
, address
, pfn
);
1269 entry
= pte_mkyoung(orig_pte
);
1270 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1271 ptep_set_access_flags(vma
, address
, page_table
, entry
, 1);
1272 update_mmu_cache(vma
, address
, entry
);
1273 lazy_mmu_prot_update(entry
);
1274 ret
|= VM_FAULT_WRITE
;
1280 * Ok, we need to copy. Oh, well..
1282 page_cache_get(old_page
);
1283 pte_unmap_unlock(page_table
, ptl
);
1285 if (unlikely(anon_vma_prepare(vma
)))
1287 if (old_page
== ZERO_PAGE(address
)) {
1288 new_page
= alloc_zeroed_user_highpage(vma
, address
);
1292 new_page
= alloc_page_vma(GFP_HIGHUSER
, vma
, address
);
1295 copy_user_highpage(new_page
, old_page
, address
);
1299 * Re-check the pte - we dropped the lock
1301 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1302 if (likely(pte_same(*page_table
, orig_pte
))) {
1303 page_remove_rmap(old_page
);
1304 if (!PageAnon(old_page
)) {
1305 inc_mm_counter(mm
, anon_rss
);
1306 dec_mm_counter(mm
, file_rss
);
1308 flush_cache_page(vma
, address
, pfn
);
1309 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
1310 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1311 ptep_establish(vma
, address
, page_table
, entry
);
1312 update_mmu_cache(vma
, address
, entry
);
1313 lazy_mmu_prot_update(entry
);
1314 lru_cache_add_active(new_page
);
1315 page_add_anon_rmap(new_page
, vma
, address
);
1317 /* Free the old page.. */
1318 new_page
= old_page
;
1319 ret
|= VM_FAULT_WRITE
;
1321 page_cache_release(new_page
);
1322 page_cache_release(old_page
);
1324 pte_unmap_unlock(page_table
, ptl
);
1327 page_cache_release(old_page
);
1328 return VM_FAULT_OOM
;
1332 * Helper functions for unmap_mapping_range().
1334 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1336 * We have to restart searching the prio_tree whenever we drop the lock,
1337 * since the iterator is only valid while the lock is held, and anyway
1338 * a later vma might be split and reinserted earlier while lock dropped.
1340 * The list of nonlinear vmas could be handled more efficiently, using
1341 * a placeholder, but handle it in the same way until a need is shown.
1342 * It is important to search the prio_tree before nonlinear list: a vma
1343 * may become nonlinear and be shifted from prio_tree to nonlinear list
1344 * while the lock is dropped; but never shifted from list to prio_tree.
1346 * In order to make forward progress despite restarting the search,
1347 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1348 * quickly skip it next time around. Since the prio_tree search only
1349 * shows us those vmas affected by unmapping the range in question, we
1350 * can't efficiently keep all vmas in step with mapping->truncate_count:
1351 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1352 * mapping->truncate_count and vma->vm_truncate_count are protected by
1355 * In order to make forward progress despite repeatedly restarting some
1356 * large vma, note the restart_addr from unmap_vmas when it breaks out:
1357 * and restart from that address when we reach that vma again. It might
1358 * have been split or merged, shrunk or extended, but never shifted: so
1359 * restart_addr remains valid so long as it remains in the vma's range.
1360 * unmap_mapping_range forces truncate_count to leap over page-aligned
1361 * values so we can save vma's restart_addr in its truncate_count field.
1363 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1365 static void reset_vma_truncate_counts(struct address_space
*mapping
)
1367 struct vm_area_struct
*vma
;
1368 struct prio_tree_iter iter
;
1370 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, 0, ULONG_MAX
)
1371 vma
->vm_truncate_count
= 0;
1372 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
1373 vma
->vm_truncate_count
= 0;
1376 static int unmap_mapping_range_vma(struct vm_area_struct
*vma
,
1377 unsigned long start_addr
, unsigned long end_addr
,
1378 struct zap_details
*details
)
1380 unsigned long restart_addr
;
1384 restart_addr
= vma
->vm_truncate_count
;
1385 if (is_restart_addr(restart_addr
) && start_addr
< restart_addr
) {
1386 start_addr
= restart_addr
;
1387 if (start_addr
>= end_addr
) {
1388 /* Top of vma has been split off since last time */
1389 vma
->vm_truncate_count
= details
->truncate_count
;
1394 restart_addr
= zap_page_range(vma
, start_addr
,
1395 end_addr
- start_addr
, details
);
1396 need_break
= need_resched() ||
1397 need_lockbreak(details
->i_mmap_lock
);
1399 if (restart_addr
>= end_addr
) {
1400 /* We have now completed this vma: mark it so */
1401 vma
->vm_truncate_count
= details
->truncate_count
;
1405 /* Note restart_addr in vma's truncate_count field */
1406 vma
->vm_truncate_count
= restart_addr
;
1411 spin_unlock(details
->i_mmap_lock
);
1413 spin_lock(details
->i_mmap_lock
);
1417 static inline void unmap_mapping_range_tree(struct prio_tree_root
*root
,
1418 struct zap_details
*details
)
1420 struct vm_area_struct
*vma
;
1421 struct prio_tree_iter iter
;
1422 pgoff_t vba
, vea
, zba
, zea
;
1425 vma_prio_tree_foreach(vma
, &iter
, root
,
1426 details
->first_index
, details
->last_index
) {
1427 /* Skip quickly over those we have already dealt with */
1428 if (vma
->vm_truncate_count
== details
->truncate_count
)
1431 vba
= vma
->vm_pgoff
;
1432 vea
= vba
+ ((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) - 1;
1433 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1434 zba
= details
->first_index
;
1437 zea
= details
->last_index
;
1441 if (unmap_mapping_range_vma(vma
,
1442 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
1443 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
1449 static inline void unmap_mapping_range_list(struct list_head
*head
,
1450 struct zap_details
*details
)
1452 struct vm_area_struct
*vma
;
1455 * In nonlinear VMAs there is no correspondence between virtual address
1456 * offset and file offset. So we must perform an exhaustive search
1457 * across *all* the pages in each nonlinear VMA, not just the pages
1458 * whose virtual address lies outside the file truncation point.
1461 list_for_each_entry(vma
, head
, shared
.vm_set
.list
) {
1462 /* Skip quickly over those we have already dealt with */
1463 if (vma
->vm_truncate_count
== details
->truncate_count
)
1465 details
->nonlinear_vma
= vma
;
1466 if (unmap_mapping_range_vma(vma
, vma
->vm_start
,
1467 vma
->vm_end
, details
) < 0)
1473 * unmap_mapping_range - unmap the portion of all mmaps
1474 * in the specified address_space corresponding to the specified
1475 * page range in the underlying file.
1476 * @mapping: the address space containing mmaps to be unmapped.
1477 * @holebegin: byte in first page to unmap, relative to the start of
1478 * the underlying file. This will be rounded down to a PAGE_SIZE
1479 * boundary. Note that this is different from vmtruncate(), which
1480 * must keep the partial page. In contrast, we must get rid of
1482 * @holelen: size of prospective hole in bytes. This will be rounded
1483 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
1485 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1486 * but 0 when invalidating pagecache, don't throw away private data.
1488 void unmap_mapping_range(struct address_space
*mapping
,
1489 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
1491 struct zap_details details
;
1492 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
1493 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1495 /* Check for overflow. */
1496 if (sizeof(holelen
) > sizeof(hlen
)) {
1498 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1499 if (holeend
& ~(long long)ULONG_MAX
)
1500 hlen
= ULONG_MAX
- hba
+ 1;
1503 details
.check_mapping
= even_cows
? NULL
: mapping
;
1504 details
.nonlinear_vma
= NULL
;
1505 details
.first_index
= hba
;
1506 details
.last_index
= hba
+ hlen
- 1;
1507 if (details
.last_index
< details
.first_index
)
1508 details
.last_index
= ULONG_MAX
;
1509 details
.i_mmap_lock
= &mapping
->i_mmap_lock
;
1511 spin_lock(&mapping
->i_mmap_lock
);
1513 /* serialize i_size write against truncate_count write */
1515 /* Protect against page faults, and endless unmapping loops */
1516 mapping
->truncate_count
++;
1518 * For archs where spin_lock has inclusive semantics like ia64
1519 * this smp_mb() will prevent to read pagetable contents
1520 * before the truncate_count increment is visible to
1524 if (unlikely(is_restart_addr(mapping
->truncate_count
))) {
1525 if (mapping
->truncate_count
== 0)
1526 reset_vma_truncate_counts(mapping
);
1527 mapping
->truncate_count
++;
1529 details
.truncate_count
= mapping
->truncate_count
;
1531 if (unlikely(!prio_tree_empty(&mapping
->i_mmap
)))
1532 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
1533 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
1534 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
1535 spin_unlock(&mapping
->i_mmap_lock
);
1537 EXPORT_SYMBOL(unmap_mapping_range
);
1540 * Handle all mappings that got truncated by a "truncate()"
1543 * NOTE! We have to be ready to update the memory sharing
1544 * between the file and the memory map for a potential last
1545 * incomplete page. Ugly, but necessary.
1547 int vmtruncate(struct inode
* inode
, loff_t offset
)
1549 struct address_space
*mapping
= inode
->i_mapping
;
1550 unsigned long limit
;
1552 if (inode
->i_size
< offset
)
1555 * truncation of in-use swapfiles is disallowed - it would cause
1556 * subsequent swapout to scribble on the now-freed blocks.
1558 if (IS_SWAPFILE(inode
))
1560 i_size_write(inode
, offset
);
1561 unmap_mapping_range(mapping
, offset
+ PAGE_SIZE
- 1, 0, 1);
1562 truncate_inode_pages(mapping
, offset
);
1566 limit
= current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
;
1567 if (limit
!= RLIM_INFINITY
&& offset
> limit
)
1569 if (offset
> inode
->i_sb
->s_maxbytes
)
1571 i_size_write(inode
, offset
);
1574 if (inode
->i_op
&& inode
->i_op
->truncate
)
1575 inode
->i_op
->truncate(inode
);
1578 send_sig(SIGXFSZ
, current
, 0);
1585 EXPORT_SYMBOL(vmtruncate
);
1588 * Primitive swap readahead code. We simply read an aligned block of
1589 * (1 << page_cluster) entries in the swap area. This method is chosen
1590 * because it doesn't cost us any seek time. We also make sure to queue
1591 * the 'original' request together with the readahead ones...
1593 * This has been extended to use the NUMA policies from the mm triggering
1596 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1598 void swapin_readahead(swp_entry_t entry
, unsigned long addr
,struct vm_area_struct
*vma
)
1601 struct vm_area_struct
*next_vma
= vma
? vma
->vm_next
: NULL
;
1604 struct page
*new_page
;
1605 unsigned long offset
;
1608 * Get the number of handles we should do readahead io to.
1610 num
= valid_swaphandles(entry
, &offset
);
1611 for (i
= 0; i
< num
; offset
++, i
++) {
1612 /* Ok, do the async read-ahead now */
1613 new_page
= read_swap_cache_async(swp_entry(swp_type(entry
),
1614 offset
), vma
, addr
);
1617 page_cache_release(new_page
);
1620 * Find the next applicable VMA for the NUMA policy.
1626 if (addr
>= vma
->vm_end
) {
1628 next_vma
= vma
? vma
->vm_next
: NULL
;
1630 if (vma
&& addr
< vma
->vm_start
)
1633 if (next_vma
&& addr
>= next_vma
->vm_start
) {
1635 next_vma
= vma
->vm_next
;
1640 lru_add_drain(); /* Push any new pages onto the LRU now */
1644 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1645 * but allow concurrent faults), and pte mapped but not yet locked.
1646 * We return with mmap_sem still held, but pte unmapped and unlocked.
1648 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1649 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1650 int write_access
, pte_t orig_pte
)
1656 int ret
= VM_FAULT_MINOR
;
1658 if (!pte_unmap_same(mm
, page_table
, orig_pte
))
1661 entry
= pte_to_swp_entry(orig_pte
);
1662 page
= lookup_swap_cache(entry
);
1664 swapin_readahead(entry
, address
, vma
);
1665 page
= read_swap_cache_async(entry
, vma
, address
);
1668 * Back out if somebody else faulted in this pte
1669 * while we released the pte lock.
1671 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1672 if (likely(pte_same(*page_table
, orig_pte
)))
1677 /* Had to read the page from swap area: Major fault */
1678 ret
= VM_FAULT_MAJOR
;
1679 inc_page_state(pgmajfault
);
1683 mark_page_accessed(page
);
1687 * Back out if somebody else already faulted in this pte.
1689 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1690 if (unlikely(!pte_same(*page_table
, orig_pte
)))
1693 if (unlikely(!PageUptodate(page
))) {
1694 ret
= VM_FAULT_SIGBUS
;
1698 /* The page isn't present yet, go ahead with the fault. */
1700 inc_mm_counter(mm
, anon_rss
);
1701 pte
= mk_pte(page
, vma
->vm_page_prot
);
1702 if (write_access
&& can_share_swap_page(page
)) {
1703 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
1707 flush_icache_page(vma
, page
);
1708 set_pte_at(mm
, address
, page_table
, pte
);
1709 page_add_anon_rmap(page
, vma
, address
);
1713 remove_exclusive_swap_page(page
);
1717 if (do_wp_page(mm
, vma
, address
,
1718 page_table
, pmd
, ptl
, pte
) == VM_FAULT_OOM
)
1723 /* No need to invalidate - it was non-present before */
1724 update_mmu_cache(vma
, address
, pte
);
1725 lazy_mmu_prot_update(pte
);
1727 pte_unmap_unlock(page_table
, ptl
);
1731 pte_unmap_unlock(page_table
, ptl
);
1733 page_cache_release(page
);
1738 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1739 * but allow concurrent faults), and pte mapped but not yet locked.
1740 * We return with mmap_sem still held, but pte unmapped and unlocked.
1742 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1743 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1751 /* Allocate our own private page. */
1752 pte_unmap(page_table
);
1754 if (unlikely(anon_vma_prepare(vma
)))
1756 page
= alloc_zeroed_user_highpage(vma
, address
);
1760 entry
= mk_pte(page
, vma
->vm_page_prot
);
1761 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1763 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1764 if (!pte_none(*page_table
))
1766 inc_mm_counter(mm
, anon_rss
);
1767 lru_cache_add_active(page
);
1768 SetPageReferenced(page
);
1769 page_add_anon_rmap(page
, vma
, address
);
1771 /* Map the ZERO_PAGE - vm_page_prot is readonly */
1772 page
= ZERO_PAGE(address
);
1773 page_cache_get(page
);
1774 entry
= mk_pte(page
, vma
->vm_page_prot
);
1776 ptl
= &mm
->page_table_lock
;
1778 if (!pte_none(*page_table
))
1780 inc_mm_counter(mm
, file_rss
);
1781 page_add_file_rmap(page
);
1784 set_pte_at(mm
, address
, page_table
, entry
);
1786 /* No need to invalidate - it was non-present before */
1787 update_mmu_cache(vma
, address
, entry
);
1788 lazy_mmu_prot_update(entry
);
1790 pte_unmap_unlock(page_table
, ptl
);
1791 return VM_FAULT_MINOR
;
1793 page_cache_release(page
);
1796 return VM_FAULT_OOM
;
1800 * do_no_page() tries to create a new page mapping. It aggressively
1801 * tries to share with existing pages, but makes a separate copy if
1802 * the "write_access" parameter is true in order to avoid the next
1805 * As this is called only for pages that do not currently exist, we
1806 * do not need to flush old virtual caches or the TLB.
1808 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1809 * but allow concurrent faults), and pte mapped but not yet locked.
1810 * We return with mmap_sem still held, but pte unmapped and unlocked.
1812 static int do_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1813 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1817 struct page
*new_page
;
1818 struct address_space
*mapping
= NULL
;
1820 unsigned int sequence
= 0;
1821 int ret
= VM_FAULT_MINOR
;
1824 pte_unmap(page_table
);
1827 mapping
= vma
->vm_file
->f_mapping
;
1828 sequence
= mapping
->truncate_count
;
1829 smp_rmb(); /* serializes i_size against truncate_count */
1832 new_page
= vma
->vm_ops
->nopage(vma
, address
& PAGE_MASK
, &ret
);
1834 * No smp_rmb is needed here as long as there's a full
1835 * spin_lock/unlock sequence inside the ->nopage callback
1836 * (for the pagecache lookup) that acts as an implicit
1837 * smp_mb() and prevents the i_size read to happen
1838 * after the next truncate_count read.
1841 /* no page was available -- either SIGBUS or OOM */
1842 if (new_page
== NOPAGE_SIGBUS
)
1843 return VM_FAULT_SIGBUS
;
1844 if (new_page
== NOPAGE_OOM
)
1845 return VM_FAULT_OOM
;
1848 * Should we do an early C-O-W break?
1850 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
1853 if (unlikely(anon_vma_prepare(vma
)))
1855 page
= alloc_page_vma(GFP_HIGHUSER
, vma
, address
);
1858 copy_user_highpage(page
, new_page
, address
);
1859 page_cache_release(new_page
);
1864 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1866 * For a file-backed vma, someone could have truncated or otherwise
1867 * invalidated this page. If unmap_mapping_range got called,
1868 * retry getting the page.
1870 if (mapping
&& unlikely(sequence
!= mapping
->truncate_count
)) {
1871 pte_unmap_unlock(page_table
, ptl
);
1872 page_cache_release(new_page
);
1874 sequence
= mapping
->truncate_count
;
1880 * This silly early PAGE_DIRTY setting removes a race
1881 * due to the bad i386 page protection. But it's valid
1882 * for other architectures too.
1884 * Note that if write_access is true, we either now have
1885 * an exclusive copy of the page, or this is a shared mapping,
1886 * so we can make it writable and dirty to avoid having to
1887 * handle that later.
1889 /* Only go through if we didn't race with anybody else... */
1890 if (pte_none(*page_table
)) {
1891 flush_icache_page(vma
, new_page
);
1892 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
1894 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1895 set_pte_at(mm
, address
, page_table
, entry
);
1897 inc_mm_counter(mm
, anon_rss
);
1898 lru_cache_add_active(new_page
);
1899 page_add_anon_rmap(new_page
, vma
, address
);
1900 } else if (!(vma
->vm_flags
& VM_RESERVED
)) {
1901 inc_mm_counter(mm
, file_rss
);
1902 page_add_file_rmap(new_page
);
1905 /* One of our sibling threads was faster, back out. */
1906 page_cache_release(new_page
);
1910 /* no need to invalidate: a not-present page shouldn't be cached */
1911 update_mmu_cache(vma
, address
, entry
);
1912 lazy_mmu_prot_update(entry
);
1914 pte_unmap_unlock(page_table
, ptl
);
1917 page_cache_release(new_page
);
1918 return VM_FAULT_OOM
;
1922 * Fault of a previously existing named mapping. Repopulate the pte
1923 * from the encoded file_pte if possible. This enables swappable
1926 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1927 * but allow concurrent faults), and pte mapped but not yet locked.
1928 * We return with mmap_sem still held, but pte unmapped and unlocked.
1930 static int do_file_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1931 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1932 int write_access
, pte_t orig_pte
)
1937 if (!pte_unmap_same(mm
, page_table
, orig_pte
))
1938 return VM_FAULT_MINOR
;
1940 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
))) {
1942 * Page table corrupted: show pte and kill process.
1944 print_bad_pte(vma
, orig_pte
, address
);
1945 return VM_FAULT_OOM
;
1947 /* We can then assume vm->vm_ops && vma->vm_ops->populate */
1949 pgoff
= pte_to_pgoff(orig_pte
);
1950 err
= vma
->vm_ops
->populate(vma
, address
& PAGE_MASK
, PAGE_SIZE
,
1951 vma
->vm_page_prot
, pgoff
, 0);
1953 return VM_FAULT_OOM
;
1955 return VM_FAULT_SIGBUS
;
1956 return VM_FAULT_MAJOR
;
1960 * These routines also need to handle stuff like marking pages dirty
1961 * and/or accessed for architectures that don't do it in hardware (most
1962 * RISC architectures). The early dirtying is also good on the i386.
1964 * There is also a hook called "update_mmu_cache()" that architectures
1965 * with external mmu caches can use to update those (ie the Sparc or
1966 * PowerPC hashed page tables that act as extended TLBs).
1968 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1969 * but allow concurrent faults), and pte mapped but not yet locked.
1970 * We return with mmap_sem still held, but pte unmapped and unlocked.
1972 static inline int handle_pte_fault(struct mm_struct
*mm
,
1973 struct vm_area_struct
*vma
, unsigned long address
,
1974 pte_t
*pte
, pmd_t
*pmd
, int write_access
)
1980 if (!pte_present(entry
)) {
1981 if (pte_none(entry
)) {
1982 if (!vma
->vm_ops
|| !vma
->vm_ops
->nopage
)
1983 return do_anonymous_page(mm
, vma
, address
,
1984 pte
, pmd
, write_access
);
1985 return do_no_page(mm
, vma
, address
,
1986 pte
, pmd
, write_access
);
1988 if (pte_file(entry
))
1989 return do_file_page(mm
, vma
, address
,
1990 pte
, pmd
, write_access
, entry
);
1991 return do_swap_page(mm
, vma
, address
,
1992 pte
, pmd
, write_access
, entry
);
1995 ptl
= &mm
->page_table_lock
;
1997 if (unlikely(!pte_same(*pte
, entry
)))
2000 if (!pte_write(entry
))
2001 return do_wp_page(mm
, vma
, address
,
2002 pte
, pmd
, ptl
, entry
);
2003 entry
= pte_mkdirty(entry
);
2005 entry
= pte_mkyoung(entry
);
2006 ptep_set_access_flags(vma
, address
, pte
, entry
, write_access
);
2007 update_mmu_cache(vma
, address
, entry
);
2008 lazy_mmu_prot_update(entry
);
2010 pte_unmap_unlock(pte
, ptl
);
2011 return VM_FAULT_MINOR
;
2015 * By the time we get here, we already hold the mm semaphore
2017 int __handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2018 unsigned long address
, int write_access
)
2025 __set_current_state(TASK_RUNNING
);
2027 inc_page_state(pgfault
);
2029 if (unlikely(is_vm_hugetlb_page(vma
)))
2030 return hugetlb_fault(mm
, vma
, address
, write_access
);
2032 pgd
= pgd_offset(mm
, address
);
2033 pud
= pud_alloc(mm
, pgd
, address
);
2035 return VM_FAULT_OOM
;
2036 pmd
= pmd_alloc(mm
, pud
, address
);
2038 return VM_FAULT_OOM
;
2039 pte
= pte_alloc_map(mm
, pmd
, address
);
2041 return VM_FAULT_OOM
;
2043 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, write_access
);
2046 #ifndef __PAGETABLE_PUD_FOLDED
2048 * Allocate page upper directory.
2049 * We've already handled the fast-path in-line.
2051 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
2053 pud_t
*new = pud_alloc_one(mm
, address
);
2057 spin_lock(&mm
->page_table_lock
);
2058 if (pgd_present(*pgd
)) /* Another has populated it */
2061 pgd_populate(mm
, pgd
, new);
2062 spin_unlock(&mm
->page_table_lock
);
2065 #endif /* __PAGETABLE_PUD_FOLDED */
2067 #ifndef __PAGETABLE_PMD_FOLDED
2069 * Allocate page middle directory.
2070 * We've already handled the fast-path in-line.
2072 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
2074 pmd_t
*new = pmd_alloc_one(mm
, address
);
2078 spin_lock(&mm
->page_table_lock
);
2079 #ifndef __ARCH_HAS_4LEVEL_HACK
2080 if (pud_present(*pud
)) /* Another has populated it */
2083 pud_populate(mm
, pud
, new);
2085 if (pgd_present(*pud
)) /* Another has populated it */
2088 pgd_populate(mm
, pud
, new);
2089 #endif /* __ARCH_HAS_4LEVEL_HACK */
2090 spin_unlock(&mm
->page_table_lock
);
2093 #endif /* __PAGETABLE_PMD_FOLDED */
2095 int make_pages_present(unsigned long addr
, unsigned long end
)
2097 int ret
, len
, write
;
2098 struct vm_area_struct
* vma
;
2100 vma
= find_vma(current
->mm
, addr
);
2103 write
= (vma
->vm_flags
& VM_WRITE
) != 0;
2106 if (end
> vma
->vm_end
)
2108 len
= (end
+PAGE_SIZE
-1)/PAGE_SIZE
-addr
/PAGE_SIZE
;
2109 ret
= get_user_pages(current
, current
->mm
, addr
,
2110 len
, write
, 0, NULL
, NULL
);
2113 return ret
== len
? 0 : -1;
2117 * Map a vmalloc()-space virtual address to the physical page.
2119 struct page
* vmalloc_to_page(void * vmalloc_addr
)
2121 unsigned long addr
= (unsigned long) vmalloc_addr
;
2122 struct page
*page
= NULL
;
2123 pgd_t
*pgd
= pgd_offset_k(addr
);
2128 if (!pgd_none(*pgd
)) {
2129 pud
= pud_offset(pgd
, addr
);
2130 if (!pud_none(*pud
)) {
2131 pmd
= pmd_offset(pud
, addr
);
2132 if (!pmd_none(*pmd
)) {
2133 ptep
= pte_offset_map(pmd
, addr
);
2135 if (pte_present(pte
))
2136 page
= pte_page(pte
);
2144 EXPORT_SYMBOL(vmalloc_to_page
);
2147 * Map a vmalloc()-space virtual address to the physical page frame number.
2149 unsigned long vmalloc_to_pfn(void * vmalloc_addr
)
2151 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
2154 EXPORT_SYMBOL(vmalloc_to_pfn
);
2156 #if !defined(__HAVE_ARCH_GATE_AREA)
2158 #if defined(AT_SYSINFO_EHDR)
2159 static struct vm_area_struct gate_vma
;
2161 static int __init
gate_vma_init(void)
2163 gate_vma
.vm_mm
= NULL
;
2164 gate_vma
.vm_start
= FIXADDR_USER_START
;
2165 gate_vma
.vm_end
= FIXADDR_USER_END
;
2166 gate_vma
.vm_page_prot
= PAGE_READONLY
;
2167 gate_vma
.vm_flags
= VM_RESERVED
;
2170 __initcall(gate_vma_init
);
2173 struct vm_area_struct
*get_gate_vma(struct task_struct
*tsk
)
2175 #ifdef AT_SYSINFO_EHDR
2182 int in_gate_area_no_task(unsigned long addr
)
2184 #ifdef AT_SYSINFO_EHDR
2185 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
2191 #endif /* __HAVE_ARCH_GATE_AREA */
This page took 0.1925 seconds and 6 git commands to generate.