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 bad pte
337 * is found. For example, we might have a PFN-mapped pte in
338 * a region that doesn't allow it.
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 * This function gets the "struct page" associated with a pte.
355 * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
356 * will have each page table entry just pointing to a raw page frame
357 * number, and as far as the VM layer is concerned, those do not have
358 * pages associated with them - even if the PFN might point to memory
359 * that otherwise is perfectly fine and has a "struct page".
361 * The way we recognize those mappings is through the rules set up
362 * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
363 * and the vm_pgoff will point to the first PFN mapped: thus every
364 * page that is a raw mapping will always honor the rule
366 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
368 * and if that isn't true, the page has been COW'ed (in which case it
369 * _does_ have a "struct page" associated with it even if it is in a
372 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
, pte_t pte
)
374 unsigned long pfn
= pte_pfn(pte
);
376 if (vma
->vm_flags
& VM_PFNMAP
) {
377 unsigned long off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
378 if (pfn
== vma
->vm_pgoff
+ off
)
383 * Add some anal sanity checks for now. Eventually,
384 * we should just do "return pfn_to_page(pfn)", but
385 * in the meantime we check that we get a valid pfn,
386 * and that the resulting page looks ok.
388 * Remove this test eventually!
390 if (unlikely(!pfn_valid(pfn
))) {
391 print_bad_pte(vma
, pte
, addr
);
396 * NOTE! We still have PageReserved() pages in the page
399 * The PAGE_ZERO() pages and various VDSO mappings can
400 * cause them to exist.
402 return pfn_to_page(pfn
);
406 * copy one vm_area from one task to the other. Assumes the page tables
407 * already present in the new task to be cleared in the whole range
408 * covered by this vma.
412 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
413 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
414 unsigned long addr
, int *rss
)
416 unsigned long vm_flags
= vma
->vm_flags
;
417 pte_t pte
= *src_pte
;
420 /* pte contains position in swap or file, so copy. */
421 if (unlikely(!pte_present(pte
))) {
422 if (!pte_file(pte
)) {
423 swap_duplicate(pte_to_swp_entry(pte
));
424 /* make sure dst_mm is on swapoff's mmlist. */
425 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
426 spin_lock(&mmlist_lock
);
427 if (list_empty(&dst_mm
->mmlist
))
428 list_add(&dst_mm
->mmlist
,
430 spin_unlock(&mmlist_lock
);
437 * If it's a COW mapping, write protect it both
438 * in the parent and the child
440 if ((vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
) {
441 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
446 * If it's a shared mapping, mark it clean in
449 if (vm_flags
& VM_SHARED
)
450 pte
= pte_mkclean(pte
);
451 pte
= pte_mkold(pte
);
453 page
= vm_normal_page(vma
, addr
, pte
);
457 rss
[!!PageAnon(page
)]++;
461 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
464 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
465 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
466 unsigned long addr
, unsigned long end
)
468 pte_t
*src_pte
, *dst_pte
;
469 spinlock_t
*src_ptl
, *dst_ptl
;
475 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
478 src_pte
= pte_offset_map_nested(src_pmd
, addr
);
479 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
484 * We are holding two locks at this point - either of them
485 * could generate latencies in another task on another CPU.
487 if (progress
>= 32) {
489 if (need_resched() ||
490 need_lockbreak(src_ptl
) ||
491 need_lockbreak(dst_ptl
))
494 if (pte_none(*src_pte
)) {
498 copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
, vma
, addr
, rss
);
500 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
502 spin_unlock(src_ptl
);
503 pte_unmap_nested(src_pte
- 1);
504 add_mm_rss(dst_mm
, rss
[0], rss
[1]);
505 pte_unmap_unlock(dst_pte
- 1, dst_ptl
);
512 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
513 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
514 unsigned long addr
, unsigned long end
)
516 pmd_t
*src_pmd
, *dst_pmd
;
519 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
522 src_pmd
= pmd_offset(src_pud
, addr
);
524 next
= pmd_addr_end(addr
, end
);
525 if (pmd_none_or_clear_bad(src_pmd
))
527 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
530 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
534 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
535 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
536 unsigned long addr
, unsigned long end
)
538 pud_t
*src_pud
, *dst_pud
;
541 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
544 src_pud
= pud_offset(src_pgd
, addr
);
546 next
= pud_addr_end(addr
, end
);
547 if (pud_none_or_clear_bad(src_pud
))
549 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
552 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
556 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
557 struct vm_area_struct
*vma
)
559 pgd_t
*src_pgd
, *dst_pgd
;
561 unsigned long addr
= vma
->vm_start
;
562 unsigned long end
= vma
->vm_end
;
565 * Don't copy ptes where a page fault will fill them correctly.
566 * Fork becomes much lighter when there are big shared or private
567 * readonly mappings. The tradeoff is that copy_page_range is more
568 * efficient than faulting.
570 if (!(vma
->vm_flags
& (VM_HUGETLB
|VM_NONLINEAR
|VM_PFNMAP
))) {
575 if (is_vm_hugetlb_page(vma
))
576 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
578 dst_pgd
= pgd_offset(dst_mm
, addr
);
579 src_pgd
= pgd_offset(src_mm
, addr
);
581 next
= pgd_addr_end(addr
, end
);
582 if (pgd_none_or_clear_bad(src_pgd
))
584 if (copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
587 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
591 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
592 struct vm_area_struct
*vma
, pmd_t
*pmd
,
593 unsigned long addr
, unsigned long end
,
594 long *zap_work
, struct zap_details
*details
)
596 struct mm_struct
*mm
= tlb
->mm
;
602 pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
605 if (pte_none(ptent
)) {
609 if (pte_present(ptent
)) {
612 (*zap_work
) -= PAGE_SIZE
;
614 page
= vm_normal_page(vma
, addr
, ptent
);
615 if (unlikely(details
) && page
) {
617 * unmap_shared_mapping_pages() wants to
618 * invalidate cache without truncating:
619 * unmap shared but keep private pages.
621 if (details
->check_mapping
&&
622 details
->check_mapping
!= page
->mapping
)
625 * Each page->index must be checked when
626 * invalidating or truncating nonlinear.
628 if (details
->nonlinear_vma
&&
629 (page
->index
< details
->first_index
||
630 page
->index
> details
->last_index
))
633 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
635 tlb_remove_tlb_entry(tlb
, pte
, addr
);
638 if (unlikely(details
) && details
->nonlinear_vma
639 && linear_page_index(details
->nonlinear_vma
,
640 addr
) != page
->index
)
641 set_pte_at(mm
, addr
, pte
,
642 pgoff_to_pte(page
->index
));
646 if (pte_dirty(ptent
))
647 set_page_dirty(page
);
648 if (pte_young(ptent
))
649 mark_page_accessed(page
);
652 page_remove_rmap(page
);
653 tlb_remove_page(tlb
, page
);
657 * If details->check_mapping, we leave swap entries;
658 * if details->nonlinear_vma, we leave file entries.
660 if (unlikely(details
))
662 if (!pte_file(ptent
))
663 free_swap_and_cache(pte_to_swp_entry(ptent
));
664 pte_clear_full(mm
, addr
, pte
, tlb
->fullmm
);
665 } while (pte
++, addr
+= PAGE_SIZE
, (addr
!= end
&& *zap_work
> 0));
667 add_mm_rss(mm
, file_rss
, anon_rss
);
668 pte_unmap_unlock(pte
- 1, ptl
);
673 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
674 struct vm_area_struct
*vma
, pud_t
*pud
,
675 unsigned long addr
, unsigned long end
,
676 long *zap_work
, struct zap_details
*details
)
681 pmd
= pmd_offset(pud
, addr
);
683 next
= pmd_addr_end(addr
, end
);
684 if (pmd_none_or_clear_bad(pmd
)) {
688 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
,
690 } while (pmd
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
695 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
696 struct vm_area_struct
*vma
, pgd_t
*pgd
,
697 unsigned long addr
, unsigned long end
,
698 long *zap_work
, struct zap_details
*details
)
703 pud
= pud_offset(pgd
, addr
);
705 next
= pud_addr_end(addr
, end
);
706 if (pud_none_or_clear_bad(pud
)) {
710 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
,
712 } while (pud
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
717 static unsigned long unmap_page_range(struct mmu_gather
*tlb
,
718 struct vm_area_struct
*vma
,
719 unsigned long addr
, unsigned long end
,
720 long *zap_work
, struct zap_details
*details
)
725 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
729 tlb_start_vma(tlb
, vma
);
730 pgd
= pgd_offset(vma
->vm_mm
, addr
);
732 next
= pgd_addr_end(addr
, end
);
733 if (pgd_none_or_clear_bad(pgd
)) {
737 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
,
739 } while (pgd
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
740 tlb_end_vma(tlb
, vma
);
745 #ifdef CONFIG_PREEMPT
746 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
748 /* No preempt: go for improved straight-line efficiency */
749 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
753 * unmap_vmas - unmap a range of memory covered by a list of vma's
754 * @tlbp: address of the caller's struct mmu_gather
755 * @vma: the starting vma
756 * @start_addr: virtual address at which to start unmapping
757 * @end_addr: virtual address at which to end unmapping
758 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
759 * @details: details of nonlinear truncation or shared cache invalidation
761 * Returns the end address of the unmapping (restart addr if interrupted).
763 * Unmap all pages in the vma list.
765 * We aim to not hold locks for too long (for scheduling latency reasons).
766 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
767 * return the ending mmu_gather to the caller.
769 * Only addresses between `start' and `end' will be unmapped.
771 * The VMA list must be sorted in ascending virtual address order.
773 * unmap_vmas() assumes that the caller will flush the whole unmapped address
774 * range after unmap_vmas() returns. So the only responsibility here is to
775 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
776 * drops the lock and schedules.
778 unsigned long unmap_vmas(struct mmu_gather
**tlbp
,
779 struct vm_area_struct
*vma
, unsigned long start_addr
,
780 unsigned long end_addr
, unsigned long *nr_accounted
,
781 struct zap_details
*details
)
783 long zap_work
= ZAP_BLOCK_SIZE
;
784 unsigned long tlb_start
= 0; /* For tlb_finish_mmu */
785 int tlb_start_valid
= 0;
786 unsigned long start
= start_addr
;
787 spinlock_t
*i_mmap_lock
= details
? details
->i_mmap_lock
: NULL
;
788 int fullmm
= (*tlbp
)->fullmm
;
790 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
) {
793 start
= max(vma
->vm_start
, start_addr
);
794 if (start
>= vma
->vm_end
)
796 end
= min(vma
->vm_end
, end_addr
);
797 if (end
<= vma
->vm_start
)
800 if (vma
->vm_flags
& VM_ACCOUNT
)
801 *nr_accounted
+= (end
- start
) >> PAGE_SHIFT
;
803 while (start
!= end
) {
804 if (!tlb_start_valid
) {
809 if (unlikely(is_vm_hugetlb_page(vma
))) {
810 unmap_hugepage_range(vma
, start
, end
);
811 zap_work
-= (end
- start
) /
812 (HPAGE_SIZE
/ PAGE_SIZE
);
815 start
= unmap_page_range(*tlbp
, vma
,
816 start
, end
, &zap_work
, details
);
819 BUG_ON(start
!= end
);
823 tlb_finish_mmu(*tlbp
, tlb_start
, start
);
825 if (need_resched() ||
826 (i_mmap_lock
&& need_lockbreak(i_mmap_lock
))) {
834 *tlbp
= tlb_gather_mmu(vma
->vm_mm
, fullmm
);
836 zap_work
= ZAP_BLOCK_SIZE
;
840 return start
; /* which is now the end (or restart) address */
844 * zap_page_range - remove user pages in a given range
845 * @vma: vm_area_struct holding the applicable pages
846 * @address: starting address of pages to zap
847 * @size: number of bytes to zap
848 * @details: details of nonlinear truncation or shared cache invalidation
850 unsigned long zap_page_range(struct vm_area_struct
*vma
, unsigned long address
,
851 unsigned long size
, struct zap_details
*details
)
853 struct mm_struct
*mm
= vma
->vm_mm
;
854 struct mmu_gather
*tlb
;
855 unsigned long end
= address
+ size
;
856 unsigned long nr_accounted
= 0;
859 tlb
= tlb_gather_mmu(mm
, 0);
860 update_hiwater_rss(mm
);
861 end
= unmap_vmas(&tlb
, vma
, address
, end
, &nr_accounted
, details
);
863 tlb_finish_mmu(tlb
, address
, end
);
868 * Do a quick page-table lookup for a single page.
870 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
879 struct mm_struct
*mm
= vma
->vm_mm
;
881 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
883 BUG_ON(flags
& FOLL_GET
);
888 pgd
= pgd_offset(mm
, address
);
889 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
892 pud
= pud_offset(pgd
, address
);
893 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
896 pmd
= pmd_offset(pud
, address
);
897 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
900 if (pmd_huge(*pmd
)) {
901 BUG_ON(flags
& FOLL_GET
);
902 page
= follow_huge_pmd(mm
, address
, pmd
, flags
& FOLL_WRITE
);
906 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
911 if (!pte_present(pte
))
913 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
915 page
= vm_normal_page(vma
, address
, pte
);
919 if (flags
& FOLL_GET
)
921 if (flags
& FOLL_TOUCH
) {
922 if ((flags
& FOLL_WRITE
) &&
923 !pte_dirty(pte
) && !PageDirty(page
))
924 set_page_dirty(page
);
925 mark_page_accessed(page
);
928 pte_unmap_unlock(ptep
, ptl
);
934 * When core dumping an enormous anonymous area that nobody
935 * has touched so far, we don't want to allocate page tables.
937 if (flags
& FOLL_ANON
) {
938 page
= ZERO_PAGE(address
);
939 if (flags
& FOLL_GET
)
941 BUG_ON(flags
& FOLL_WRITE
);
946 int get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
947 unsigned long start
, int len
, int write
, int force
,
948 struct page
**pages
, struct vm_area_struct
**vmas
)
951 unsigned int vm_flags
;
954 * Require read or write permissions.
955 * If 'force' is set, we only require the "MAY" flags.
957 vm_flags
= write
? (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
958 vm_flags
&= force
? (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
962 struct vm_area_struct
*vma
;
963 unsigned int foll_flags
;
965 vma
= find_extend_vma(mm
, start
);
966 if (!vma
&& in_gate_area(tsk
, start
)) {
967 unsigned long pg
= start
& PAGE_MASK
;
968 struct vm_area_struct
*gate_vma
= get_gate_vma(tsk
);
973 if (write
) /* user gate pages are read-only */
974 return i
? : -EFAULT
;
976 pgd
= pgd_offset_k(pg
);
978 pgd
= pgd_offset_gate(mm
, pg
);
979 BUG_ON(pgd_none(*pgd
));
980 pud
= pud_offset(pgd
, pg
);
981 BUG_ON(pud_none(*pud
));
982 pmd
= pmd_offset(pud
, pg
);
984 return i
? : -EFAULT
;
985 pte
= pte_offset_map(pmd
, pg
);
986 if (pte_none(*pte
)) {
988 return i
? : -EFAULT
;
991 struct page
*page
= vm_normal_page(vma
, start
, *pte
);
1005 if (!vma
|| (vma
->vm_flags
& VM_IO
)
1006 || !(vm_flags
& vma
->vm_flags
))
1007 return i
? : -EFAULT
;
1009 if (is_vm_hugetlb_page(vma
)) {
1010 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1015 foll_flags
= FOLL_TOUCH
;
1017 foll_flags
|= FOLL_GET
;
1018 if (!write
&& !(vma
->vm_flags
& VM_LOCKED
) &&
1019 (!vma
->vm_ops
|| !vma
->vm_ops
->nopage
))
1020 foll_flags
|= FOLL_ANON
;
1026 foll_flags
|= FOLL_WRITE
;
1029 while (!(page
= follow_page(vma
, start
, foll_flags
))) {
1031 ret
= __handle_mm_fault(mm
, vma
, start
,
1032 foll_flags
& FOLL_WRITE
);
1034 * The VM_FAULT_WRITE bit tells us that do_wp_page has
1035 * broken COW when necessary, even if maybe_mkwrite
1036 * decided not to set pte_write. We can thus safely do
1037 * subsequent page lookups as if they were reads.
1039 if (ret
& VM_FAULT_WRITE
)
1040 foll_flags
&= ~FOLL_WRITE
;
1042 switch (ret
& ~VM_FAULT_WRITE
) {
1043 case VM_FAULT_MINOR
:
1046 case VM_FAULT_MAJOR
:
1049 case VM_FAULT_SIGBUS
:
1050 return i
? i
: -EFAULT
;
1052 return i
? i
: -ENOMEM
;
1059 flush_dcache_page(page
);
1066 } while (len
&& start
< vma
->vm_end
);
1070 EXPORT_SYMBOL(get_user_pages
);
1072 static int zeromap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1073 unsigned long addr
, unsigned long end
, pgprot_t prot
)
1078 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1082 struct page
*page
= ZERO_PAGE(addr
);
1083 pte_t zero_pte
= pte_wrprotect(mk_pte(page
, prot
));
1084 page_cache_get(page
);
1085 page_add_file_rmap(page
);
1086 inc_mm_counter(mm
, file_rss
);
1087 BUG_ON(!pte_none(*pte
));
1088 set_pte_at(mm
, addr
, pte
, zero_pte
);
1089 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1090 pte_unmap_unlock(pte
- 1, ptl
);
1094 static inline int zeromap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1095 unsigned long addr
, unsigned long end
, pgprot_t prot
)
1100 pmd
= pmd_alloc(mm
, pud
, addr
);
1104 next
= pmd_addr_end(addr
, end
);
1105 if (zeromap_pte_range(mm
, pmd
, addr
, next
, prot
))
1107 } while (pmd
++, addr
= next
, addr
!= end
);
1111 static inline int zeromap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1112 unsigned long addr
, unsigned long end
, pgprot_t prot
)
1117 pud
= pud_alloc(mm
, pgd
, addr
);
1121 next
= pud_addr_end(addr
, end
);
1122 if (zeromap_pmd_range(mm
, pud
, addr
, next
, prot
))
1124 } while (pud
++, addr
= next
, addr
!= end
);
1128 int zeromap_page_range(struct vm_area_struct
*vma
,
1129 unsigned long addr
, unsigned long size
, pgprot_t prot
)
1133 unsigned long end
= addr
+ size
;
1134 struct mm_struct
*mm
= vma
->vm_mm
;
1137 BUG_ON(addr
>= end
);
1138 pgd
= pgd_offset(mm
, addr
);
1139 flush_cache_range(vma
, addr
, end
);
1141 next
= pgd_addr_end(addr
, end
);
1142 err
= zeromap_pud_range(mm
, pgd
, addr
, next
, prot
);
1145 } while (pgd
++, addr
= next
, addr
!= end
);
1150 * maps a range of physical memory into the requested pages. the old
1151 * mappings are removed. any references to nonexistent pages results
1152 * in null mappings (currently treated as "copy-on-access")
1154 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1155 unsigned long addr
, unsigned long end
,
1156 unsigned long pfn
, pgprot_t prot
)
1161 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1165 BUG_ON(!pte_none(*pte
));
1166 set_pte_at(mm
, addr
, pte
, pfn_pte(pfn
, prot
));
1168 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1169 pte_unmap_unlock(pte
- 1, ptl
);
1173 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1174 unsigned long addr
, unsigned long end
,
1175 unsigned long pfn
, pgprot_t prot
)
1180 pfn
-= addr
>> PAGE_SHIFT
;
1181 pmd
= pmd_alloc(mm
, pud
, addr
);
1185 next
= pmd_addr_end(addr
, end
);
1186 if (remap_pte_range(mm
, pmd
, addr
, next
,
1187 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1189 } while (pmd
++, addr
= next
, addr
!= end
);
1193 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1194 unsigned long addr
, unsigned long end
,
1195 unsigned long pfn
, pgprot_t prot
)
1200 pfn
-= addr
>> PAGE_SHIFT
;
1201 pud
= pud_alloc(mm
, pgd
, addr
);
1205 next
= pud_addr_end(addr
, end
);
1206 if (remap_pmd_range(mm
, pud
, addr
, next
,
1207 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1209 } while (pud
++, addr
= next
, addr
!= end
);
1213 /* Note: this is only safe if the mm semaphore is held when called. */
1214 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1215 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1219 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1220 struct mm_struct
*mm
= vma
->vm_mm
;
1224 * Physically remapped pages are special. Tell the
1225 * rest of the world about it:
1226 * VM_IO tells people not to look at these pages
1227 * (accesses can have side effects).
1228 * VM_RESERVED is specified all over the place, because
1229 * in 2.4 it kept swapout's vma scan off this vma; but
1230 * in 2.6 the LRU scan won't even find its pages, so this
1231 * flag means no more than count its pages in reserved_vm,
1232 * and omit it from core dump, even when VM_IO turned off.
1233 * VM_PFNMAP tells the core MM that the base pages are just
1234 * raw PFN mappings, and do not have a "struct page" associated
1237 vma
->vm_flags
|= VM_IO
| VM_RESERVED
| VM_PFNMAP
;
1238 vma
->vm_pgoff
= pfn
;
1240 BUG_ON(addr
>= end
);
1241 pfn
-= addr
>> PAGE_SHIFT
;
1242 pgd
= pgd_offset(mm
, addr
);
1243 flush_cache_range(vma
, addr
, end
);
1245 next
= pgd_addr_end(addr
, end
);
1246 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1247 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1250 } while (pgd
++, addr
= next
, addr
!= end
);
1253 EXPORT_SYMBOL(remap_pfn_range
);
1256 * handle_pte_fault chooses page fault handler according to an entry
1257 * which was read non-atomically. Before making any commitment, on
1258 * those architectures or configurations (e.g. i386 with PAE) which
1259 * might give a mix of unmatched parts, do_swap_page and do_file_page
1260 * must check under lock before unmapping the pte and proceeding
1261 * (but do_wp_page is only called after already making such a check;
1262 * and do_anonymous_page and do_no_page can safely check later on).
1264 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
1265 pte_t
*page_table
, pte_t orig_pte
)
1268 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1269 if (sizeof(pte_t
) > sizeof(unsigned long)) {
1270 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
1272 same
= pte_same(*page_table
, orig_pte
);
1276 pte_unmap(page_table
);
1281 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1282 * servicing faults for write access. In the normal case, do always want
1283 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1284 * that do not have writing enabled, when used by access_process_vm.
1286 static inline pte_t
maybe_mkwrite(pte_t pte
, struct vm_area_struct
*vma
)
1288 if (likely(vma
->vm_flags
& VM_WRITE
))
1289 pte
= pte_mkwrite(pte
);
1293 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
)
1296 * If the source page was a PFN mapping, we don't have
1297 * a "struct page" for it. We do a best-effort copy by
1298 * just copying from the original user address. If that
1299 * fails, we just zero-fill it. Live with it.
1301 if (unlikely(!src
)) {
1302 void *kaddr
= kmap_atomic(dst
, KM_USER0
);
1303 unsigned long left
= __copy_from_user_inatomic(kaddr
, (void __user
*)va
, PAGE_SIZE
);
1305 memset(kaddr
, 0, PAGE_SIZE
);
1306 kunmap_atomic(kaddr
, KM_USER0
);
1310 copy_user_highpage(dst
, src
, va
);
1314 * This routine handles present pages, when users try to write
1315 * to a shared page. It is done by copying the page to a new address
1316 * and decrementing the shared-page counter for the old page.
1318 * Note that this routine assumes that the protection checks have been
1319 * done by the caller (the low-level page fault routine in most cases).
1320 * Thus we can safely just mark it writable once we've done any necessary
1323 * We also mark the page dirty at this point even though the page will
1324 * change only once the write actually happens. This avoids a few races,
1325 * and potentially makes it more efficient.
1327 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1328 * but allow concurrent faults), with pte both mapped and locked.
1329 * We return with mmap_sem still held, but pte unmapped and unlocked.
1331 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1332 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1333 spinlock_t
*ptl
, pte_t orig_pte
)
1335 struct page
*old_page
, *src_page
, *new_page
;
1337 int ret
= VM_FAULT_MINOR
;
1339 old_page
= vm_normal_page(vma
, address
, orig_pte
);
1340 src_page
= old_page
;
1344 if (PageAnon(old_page
) && !TestSetPageLocked(old_page
)) {
1345 int reuse
= can_share_swap_page(old_page
);
1346 unlock_page(old_page
);
1348 flush_cache_page(vma
, address
, pfn
);
1349 entry
= pte_mkyoung(orig_pte
);
1350 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1351 ptep_set_access_flags(vma
, address
, page_table
, entry
, 1);
1352 update_mmu_cache(vma
, address
, entry
);
1353 lazy_mmu_prot_update(entry
);
1354 ret
|= VM_FAULT_WRITE
;
1360 * Ok, we need to copy. Oh, well..
1362 page_cache_get(old_page
);
1364 pte_unmap_unlock(page_table
, ptl
);
1366 if (unlikely(anon_vma_prepare(vma
)))
1368 if (src_page
== ZERO_PAGE(address
)) {
1369 new_page
= alloc_zeroed_user_highpage(vma
, address
);
1373 new_page
= alloc_page_vma(GFP_HIGHUSER
, vma
, address
);
1376 cow_user_page(new_page
, src_page
, address
);
1380 * Re-check the pte - we dropped the lock
1382 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1383 if (likely(pte_same(*page_table
, orig_pte
))) {
1385 page_remove_rmap(old_page
);
1386 if (!PageAnon(old_page
)) {
1387 dec_mm_counter(mm
, file_rss
);
1388 inc_mm_counter(mm
, anon_rss
);
1391 inc_mm_counter(mm
, anon_rss
);
1392 flush_cache_page(vma
, address
, pfn
);
1393 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
1394 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1395 ptep_establish(vma
, address
, page_table
, entry
);
1396 update_mmu_cache(vma
, address
, entry
);
1397 lazy_mmu_prot_update(entry
);
1398 lru_cache_add_active(new_page
);
1399 page_add_anon_rmap(new_page
, vma
, address
);
1401 /* Free the old page.. */
1402 new_page
= old_page
;
1403 ret
|= VM_FAULT_WRITE
;
1406 page_cache_release(new_page
);
1408 page_cache_release(old_page
);
1410 pte_unmap_unlock(page_table
, ptl
);
1414 page_cache_release(old_page
);
1415 return VM_FAULT_OOM
;
1419 * Helper functions for unmap_mapping_range().
1421 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1423 * We have to restart searching the prio_tree whenever we drop the lock,
1424 * since the iterator is only valid while the lock is held, and anyway
1425 * a later vma might be split and reinserted earlier while lock dropped.
1427 * The list of nonlinear vmas could be handled more efficiently, using
1428 * a placeholder, but handle it in the same way until a need is shown.
1429 * It is important to search the prio_tree before nonlinear list: a vma
1430 * may become nonlinear and be shifted from prio_tree to nonlinear list
1431 * while the lock is dropped; but never shifted from list to prio_tree.
1433 * In order to make forward progress despite restarting the search,
1434 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1435 * quickly skip it next time around. Since the prio_tree search only
1436 * shows us those vmas affected by unmapping the range in question, we
1437 * can't efficiently keep all vmas in step with mapping->truncate_count:
1438 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1439 * mapping->truncate_count and vma->vm_truncate_count are protected by
1442 * In order to make forward progress despite repeatedly restarting some
1443 * large vma, note the restart_addr from unmap_vmas when it breaks out:
1444 * and restart from that address when we reach that vma again. It might
1445 * have been split or merged, shrunk or extended, but never shifted: so
1446 * restart_addr remains valid so long as it remains in the vma's range.
1447 * unmap_mapping_range forces truncate_count to leap over page-aligned
1448 * values so we can save vma's restart_addr in its truncate_count field.
1450 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1452 static void reset_vma_truncate_counts(struct address_space
*mapping
)
1454 struct vm_area_struct
*vma
;
1455 struct prio_tree_iter iter
;
1457 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, 0, ULONG_MAX
)
1458 vma
->vm_truncate_count
= 0;
1459 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
1460 vma
->vm_truncate_count
= 0;
1463 static int unmap_mapping_range_vma(struct vm_area_struct
*vma
,
1464 unsigned long start_addr
, unsigned long end_addr
,
1465 struct zap_details
*details
)
1467 unsigned long restart_addr
;
1471 restart_addr
= vma
->vm_truncate_count
;
1472 if (is_restart_addr(restart_addr
) && start_addr
< restart_addr
) {
1473 start_addr
= restart_addr
;
1474 if (start_addr
>= end_addr
) {
1475 /* Top of vma has been split off since last time */
1476 vma
->vm_truncate_count
= details
->truncate_count
;
1481 restart_addr
= zap_page_range(vma
, start_addr
,
1482 end_addr
- start_addr
, details
);
1483 need_break
= need_resched() ||
1484 need_lockbreak(details
->i_mmap_lock
);
1486 if (restart_addr
>= end_addr
) {
1487 /* We have now completed this vma: mark it so */
1488 vma
->vm_truncate_count
= details
->truncate_count
;
1492 /* Note restart_addr in vma's truncate_count field */
1493 vma
->vm_truncate_count
= restart_addr
;
1498 spin_unlock(details
->i_mmap_lock
);
1500 spin_lock(details
->i_mmap_lock
);
1504 static inline void unmap_mapping_range_tree(struct prio_tree_root
*root
,
1505 struct zap_details
*details
)
1507 struct vm_area_struct
*vma
;
1508 struct prio_tree_iter iter
;
1509 pgoff_t vba
, vea
, zba
, zea
;
1512 vma_prio_tree_foreach(vma
, &iter
, root
,
1513 details
->first_index
, details
->last_index
) {
1514 /* Skip quickly over those we have already dealt with */
1515 if (vma
->vm_truncate_count
== details
->truncate_count
)
1518 vba
= vma
->vm_pgoff
;
1519 vea
= vba
+ ((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) - 1;
1520 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1521 zba
= details
->first_index
;
1524 zea
= details
->last_index
;
1528 if (unmap_mapping_range_vma(vma
,
1529 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
1530 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
1536 static inline void unmap_mapping_range_list(struct list_head
*head
,
1537 struct zap_details
*details
)
1539 struct vm_area_struct
*vma
;
1542 * In nonlinear VMAs there is no correspondence between virtual address
1543 * offset and file offset. So we must perform an exhaustive search
1544 * across *all* the pages in each nonlinear VMA, not just the pages
1545 * whose virtual address lies outside the file truncation point.
1548 list_for_each_entry(vma
, head
, shared
.vm_set
.list
) {
1549 /* Skip quickly over those we have already dealt with */
1550 if (vma
->vm_truncate_count
== details
->truncate_count
)
1552 details
->nonlinear_vma
= vma
;
1553 if (unmap_mapping_range_vma(vma
, vma
->vm_start
,
1554 vma
->vm_end
, details
) < 0)
1560 * unmap_mapping_range - unmap the portion of all mmaps
1561 * in the specified address_space corresponding to the specified
1562 * page range in the underlying file.
1563 * @mapping: the address space containing mmaps to be unmapped.
1564 * @holebegin: byte in first page to unmap, relative to the start of
1565 * the underlying file. This will be rounded down to a PAGE_SIZE
1566 * boundary. Note that this is different from vmtruncate(), which
1567 * must keep the partial page. In contrast, we must get rid of
1569 * @holelen: size of prospective hole in bytes. This will be rounded
1570 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
1572 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1573 * but 0 when invalidating pagecache, don't throw away private data.
1575 void unmap_mapping_range(struct address_space
*mapping
,
1576 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
1578 struct zap_details details
;
1579 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
1580 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1582 /* Check for overflow. */
1583 if (sizeof(holelen
) > sizeof(hlen
)) {
1585 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1586 if (holeend
& ~(long long)ULONG_MAX
)
1587 hlen
= ULONG_MAX
- hba
+ 1;
1590 details
.check_mapping
= even_cows
? NULL
: mapping
;
1591 details
.nonlinear_vma
= NULL
;
1592 details
.first_index
= hba
;
1593 details
.last_index
= hba
+ hlen
- 1;
1594 if (details
.last_index
< details
.first_index
)
1595 details
.last_index
= ULONG_MAX
;
1596 details
.i_mmap_lock
= &mapping
->i_mmap_lock
;
1598 spin_lock(&mapping
->i_mmap_lock
);
1600 /* serialize i_size write against truncate_count write */
1602 /* Protect against page faults, and endless unmapping loops */
1603 mapping
->truncate_count
++;
1605 * For archs where spin_lock has inclusive semantics like ia64
1606 * this smp_mb() will prevent to read pagetable contents
1607 * before the truncate_count increment is visible to
1611 if (unlikely(is_restart_addr(mapping
->truncate_count
))) {
1612 if (mapping
->truncate_count
== 0)
1613 reset_vma_truncate_counts(mapping
);
1614 mapping
->truncate_count
++;
1616 details
.truncate_count
= mapping
->truncate_count
;
1618 if (unlikely(!prio_tree_empty(&mapping
->i_mmap
)))
1619 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
1620 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
1621 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
1622 spin_unlock(&mapping
->i_mmap_lock
);
1624 EXPORT_SYMBOL(unmap_mapping_range
);
1627 * Handle all mappings that got truncated by a "truncate()"
1630 * NOTE! We have to be ready to update the memory sharing
1631 * between the file and the memory map for a potential last
1632 * incomplete page. Ugly, but necessary.
1634 int vmtruncate(struct inode
* inode
, loff_t offset
)
1636 struct address_space
*mapping
= inode
->i_mapping
;
1637 unsigned long limit
;
1639 if (inode
->i_size
< offset
)
1642 * truncation of in-use swapfiles is disallowed - it would cause
1643 * subsequent swapout to scribble on the now-freed blocks.
1645 if (IS_SWAPFILE(inode
))
1647 i_size_write(inode
, offset
);
1648 unmap_mapping_range(mapping
, offset
+ PAGE_SIZE
- 1, 0, 1);
1649 truncate_inode_pages(mapping
, offset
);
1653 limit
= current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
;
1654 if (limit
!= RLIM_INFINITY
&& offset
> limit
)
1656 if (offset
> inode
->i_sb
->s_maxbytes
)
1658 i_size_write(inode
, offset
);
1661 if (inode
->i_op
&& inode
->i_op
->truncate
)
1662 inode
->i_op
->truncate(inode
);
1665 send_sig(SIGXFSZ
, current
, 0);
1672 EXPORT_SYMBOL(vmtruncate
);
1675 * Primitive swap readahead code. We simply read an aligned block of
1676 * (1 << page_cluster) entries in the swap area. This method is chosen
1677 * because it doesn't cost us any seek time. We also make sure to queue
1678 * the 'original' request together with the readahead ones...
1680 * This has been extended to use the NUMA policies from the mm triggering
1683 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1685 void swapin_readahead(swp_entry_t entry
, unsigned long addr
,struct vm_area_struct
*vma
)
1688 struct vm_area_struct
*next_vma
= vma
? vma
->vm_next
: NULL
;
1691 struct page
*new_page
;
1692 unsigned long offset
;
1695 * Get the number of handles we should do readahead io to.
1697 num
= valid_swaphandles(entry
, &offset
);
1698 for (i
= 0; i
< num
; offset
++, i
++) {
1699 /* Ok, do the async read-ahead now */
1700 new_page
= read_swap_cache_async(swp_entry(swp_type(entry
),
1701 offset
), vma
, addr
);
1704 page_cache_release(new_page
);
1707 * Find the next applicable VMA for the NUMA policy.
1713 if (addr
>= vma
->vm_end
) {
1715 next_vma
= vma
? vma
->vm_next
: NULL
;
1717 if (vma
&& addr
< vma
->vm_start
)
1720 if (next_vma
&& addr
>= next_vma
->vm_start
) {
1722 next_vma
= vma
->vm_next
;
1727 lru_add_drain(); /* Push any new pages onto the LRU now */
1731 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1732 * but allow concurrent faults), and pte mapped but not yet locked.
1733 * We return with mmap_sem still held, but pte unmapped and unlocked.
1735 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1736 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1737 int write_access
, pte_t orig_pte
)
1743 int ret
= VM_FAULT_MINOR
;
1745 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
1748 entry
= pte_to_swp_entry(orig_pte
);
1749 page
= lookup_swap_cache(entry
);
1751 swapin_readahead(entry
, address
, vma
);
1752 page
= read_swap_cache_async(entry
, vma
, address
);
1755 * Back out if somebody else faulted in this pte
1756 * while we released the pte lock.
1758 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1759 if (likely(pte_same(*page_table
, orig_pte
)))
1764 /* Had to read the page from swap area: Major fault */
1765 ret
= VM_FAULT_MAJOR
;
1766 inc_page_state(pgmajfault
);
1770 mark_page_accessed(page
);
1774 * Back out if somebody else already faulted in this pte.
1776 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1777 if (unlikely(!pte_same(*page_table
, orig_pte
)))
1780 if (unlikely(!PageUptodate(page
))) {
1781 ret
= VM_FAULT_SIGBUS
;
1785 /* The page isn't present yet, go ahead with the fault. */
1787 inc_mm_counter(mm
, anon_rss
);
1788 pte
= mk_pte(page
, vma
->vm_page_prot
);
1789 if (write_access
&& can_share_swap_page(page
)) {
1790 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
1794 flush_icache_page(vma
, page
);
1795 set_pte_at(mm
, address
, page_table
, pte
);
1796 page_add_anon_rmap(page
, vma
, address
);
1800 remove_exclusive_swap_page(page
);
1804 if (do_wp_page(mm
, vma
, address
,
1805 page_table
, pmd
, ptl
, pte
) == VM_FAULT_OOM
)
1810 /* No need to invalidate - it was non-present before */
1811 update_mmu_cache(vma
, address
, pte
);
1812 lazy_mmu_prot_update(pte
);
1814 pte_unmap_unlock(page_table
, ptl
);
1818 pte_unmap_unlock(page_table
, ptl
);
1820 page_cache_release(page
);
1825 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1826 * but allow concurrent faults), and pte mapped but not yet locked.
1827 * We return with mmap_sem still held, but pte unmapped and unlocked.
1829 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1830 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1838 /* Allocate our own private page. */
1839 pte_unmap(page_table
);
1841 if (unlikely(anon_vma_prepare(vma
)))
1843 page
= alloc_zeroed_user_highpage(vma
, address
);
1847 entry
= mk_pte(page
, vma
->vm_page_prot
);
1848 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1850 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1851 if (!pte_none(*page_table
))
1853 inc_mm_counter(mm
, anon_rss
);
1854 lru_cache_add_active(page
);
1855 SetPageReferenced(page
);
1856 page_add_anon_rmap(page
, vma
, address
);
1858 /* Map the ZERO_PAGE - vm_page_prot is readonly */
1859 page
= ZERO_PAGE(address
);
1860 page_cache_get(page
);
1861 entry
= mk_pte(page
, vma
->vm_page_prot
);
1863 ptl
= pte_lockptr(mm
, pmd
);
1865 if (!pte_none(*page_table
))
1867 inc_mm_counter(mm
, file_rss
);
1868 page_add_file_rmap(page
);
1871 set_pte_at(mm
, address
, page_table
, entry
);
1873 /* No need to invalidate - it was non-present before */
1874 update_mmu_cache(vma
, address
, entry
);
1875 lazy_mmu_prot_update(entry
);
1877 pte_unmap_unlock(page_table
, ptl
);
1878 return VM_FAULT_MINOR
;
1880 page_cache_release(page
);
1883 return VM_FAULT_OOM
;
1887 * do_no_page() tries to create a new page mapping. It aggressively
1888 * tries to share with existing pages, but makes a separate copy if
1889 * the "write_access" parameter is true in order to avoid the next
1892 * As this is called only for pages that do not currently exist, we
1893 * do not need to flush old virtual caches or the TLB.
1895 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1896 * but allow concurrent faults), and pte mapped but not yet locked.
1897 * We return with mmap_sem still held, but pte unmapped and unlocked.
1899 static int do_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1900 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1904 struct page
*new_page
;
1905 struct address_space
*mapping
= NULL
;
1907 unsigned int sequence
= 0;
1908 int ret
= VM_FAULT_MINOR
;
1911 pte_unmap(page_table
);
1913 mapping
= vma
->vm_file
->f_mapping
;
1914 sequence
= mapping
->truncate_count
;
1915 smp_rmb(); /* serializes i_size against truncate_count */
1918 new_page
= vma
->vm_ops
->nopage(vma
, address
& PAGE_MASK
, &ret
);
1920 * No smp_rmb is needed here as long as there's a full
1921 * spin_lock/unlock sequence inside the ->nopage callback
1922 * (for the pagecache lookup) that acts as an implicit
1923 * smp_mb() and prevents the i_size read to happen
1924 * after the next truncate_count read.
1927 /* no page was available -- either SIGBUS or OOM */
1928 if (new_page
== NOPAGE_SIGBUS
)
1929 return VM_FAULT_SIGBUS
;
1930 if (new_page
== NOPAGE_OOM
)
1931 return VM_FAULT_OOM
;
1934 * Should we do an early C-O-W break?
1936 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
1939 if (unlikely(anon_vma_prepare(vma
)))
1941 page
= alloc_page_vma(GFP_HIGHUSER
, vma
, address
);
1944 cow_user_page(page
, new_page
, address
);
1945 page_cache_release(new_page
);
1950 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1952 * For a file-backed vma, someone could have truncated or otherwise
1953 * invalidated this page. If unmap_mapping_range got called,
1954 * retry getting the page.
1956 if (mapping
&& unlikely(sequence
!= mapping
->truncate_count
)) {
1957 pte_unmap_unlock(page_table
, ptl
);
1958 page_cache_release(new_page
);
1960 sequence
= mapping
->truncate_count
;
1966 * This silly early PAGE_DIRTY setting removes a race
1967 * due to the bad i386 page protection. But it's valid
1968 * for other architectures too.
1970 * Note that if write_access is true, we either now have
1971 * an exclusive copy of the page, or this is a shared mapping,
1972 * so we can make it writable and dirty to avoid having to
1973 * handle that later.
1975 /* Only go through if we didn't race with anybody else... */
1976 if (pte_none(*page_table
)) {
1977 flush_icache_page(vma
, new_page
);
1978 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
1980 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1981 set_pte_at(mm
, address
, page_table
, entry
);
1983 inc_mm_counter(mm
, anon_rss
);
1984 lru_cache_add_active(new_page
);
1985 page_add_anon_rmap(new_page
, vma
, address
);
1987 inc_mm_counter(mm
, file_rss
);
1988 page_add_file_rmap(new_page
);
1991 /* One of our sibling threads was faster, back out. */
1992 page_cache_release(new_page
);
1996 /* no need to invalidate: a not-present page shouldn't be cached */
1997 update_mmu_cache(vma
, address
, entry
);
1998 lazy_mmu_prot_update(entry
);
2000 pte_unmap_unlock(page_table
, ptl
);
2003 page_cache_release(new_page
);
2004 return VM_FAULT_OOM
;
2008 * Fault of a previously existing named mapping. Repopulate the pte
2009 * from the encoded file_pte if possible. This enables swappable
2012 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2013 * but allow concurrent faults), and pte mapped but not yet locked.
2014 * We return with mmap_sem still held, but pte unmapped and unlocked.
2016 static int do_file_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2017 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2018 int write_access
, pte_t orig_pte
)
2023 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2024 return VM_FAULT_MINOR
;
2026 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
))) {
2028 * Page table corrupted: show pte and kill process.
2030 print_bad_pte(vma
, orig_pte
, address
);
2031 return VM_FAULT_OOM
;
2033 /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2035 pgoff
= pte_to_pgoff(orig_pte
);
2036 err
= vma
->vm_ops
->populate(vma
, address
& PAGE_MASK
, PAGE_SIZE
,
2037 vma
->vm_page_prot
, pgoff
, 0);
2039 return VM_FAULT_OOM
;
2041 return VM_FAULT_SIGBUS
;
2042 return VM_FAULT_MAJOR
;
2046 * These routines also need to handle stuff like marking pages dirty
2047 * and/or accessed for architectures that don't do it in hardware (most
2048 * RISC architectures). The early dirtying is also good on the i386.
2050 * There is also a hook called "update_mmu_cache()" that architectures
2051 * with external mmu caches can use to update those (ie the Sparc or
2052 * PowerPC hashed page tables that act as extended TLBs).
2054 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2055 * but allow concurrent faults), and pte mapped but not yet locked.
2056 * We return with mmap_sem still held, but pte unmapped and unlocked.
2058 static inline int handle_pte_fault(struct mm_struct
*mm
,
2059 struct vm_area_struct
*vma
, unsigned long address
,
2060 pte_t
*pte
, pmd_t
*pmd
, int write_access
)
2066 old_entry
= entry
= *pte
;
2067 if (!pte_present(entry
)) {
2068 if (pte_none(entry
)) {
2069 if (!vma
->vm_ops
|| !vma
->vm_ops
->nopage
)
2070 return do_anonymous_page(mm
, vma
, address
,
2071 pte
, pmd
, write_access
);
2072 return do_no_page(mm
, vma
, address
,
2073 pte
, pmd
, write_access
);
2075 if (pte_file(entry
))
2076 return do_file_page(mm
, vma
, address
,
2077 pte
, pmd
, write_access
, entry
);
2078 return do_swap_page(mm
, vma
, address
,
2079 pte
, pmd
, write_access
, entry
);
2082 ptl
= pte_lockptr(mm
, pmd
);
2084 if (unlikely(!pte_same(*pte
, entry
)))
2087 if (!pte_write(entry
))
2088 return do_wp_page(mm
, vma
, address
,
2089 pte
, pmd
, ptl
, entry
);
2090 entry
= pte_mkdirty(entry
);
2092 entry
= pte_mkyoung(entry
);
2093 if (!pte_same(old_entry
, entry
)) {
2094 ptep_set_access_flags(vma
, address
, pte
, entry
, write_access
);
2095 update_mmu_cache(vma
, address
, entry
);
2096 lazy_mmu_prot_update(entry
);
2099 * This is needed only for protection faults but the arch code
2100 * is not yet telling us if this is a protection fault or not.
2101 * This still avoids useless tlb flushes for .text page faults
2105 flush_tlb_page(vma
, address
);
2108 pte_unmap_unlock(pte
, ptl
);
2109 return VM_FAULT_MINOR
;
2113 * By the time we get here, we already hold the mm semaphore
2115 int __handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2116 unsigned long address
, int write_access
)
2123 __set_current_state(TASK_RUNNING
);
2125 inc_page_state(pgfault
);
2127 if (unlikely(is_vm_hugetlb_page(vma
)))
2128 return hugetlb_fault(mm
, vma
, address
, write_access
);
2130 pgd
= pgd_offset(mm
, address
);
2131 pud
= pud_alloc(mm
, pgd
, address
);
2133 return VM_FAULT_OOM
;
2134 pmd
= pmd_alloc(mm
, pud
, address
);
2136 return VM_FAULT_OOM
;
2137 pte
= pte_alloc_map(mm
, pmd
, address
);
2139 return VM_FAULT_OOM
;
2141 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, write_access
);
2144 #ifndef __PAGETABLE_PUD_FOLDED
2146 * Allocate page upper directory.
2147 * We've already handled the fast-path in-line.
2149 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
2151 pud_t
*new = pud_alloc_one(mm
, address
);
2155 spin_lock(&mm
->page_table_lock
);
2156 if (pgd_present(*pgd
)) /* Another has populated it */
2159 pgd_populate(mm
, pgd
, new);
2160 spin_unlock(&mm
->page_table_lock
);
2164 /* Workaround for gcc 2.96 */
2165 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
2169 #endif /* __PAGETABLE_PUD_FOLDED */
2171 #ifndef __PAGETABLE_PMD_FOLDED
2173 * Allocate page middle directory.
2174 * We've already handled the fast-path in-line.
2176 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
2178 pmd_t
*new = pmd_alloc_one(mm
, address
);
2182 spin_lock(&mm
->page_table_lock
);
2183 #ifndef __ARCH_HAS_4LEVEL_HACK
2184 if (pud_present(*pud
)) /* Another has populated it */
2187 pud_populate(mm
, pud
, new);
2189 if (pgd_present(*pud
)) /* Another has populated it */
2192 pgd_populate(mm
, pud
, new);
2193 #endif /* __ARCH_HAS_4LEVEL_HACK */
2194 spin_unlock(&mm
->page_table_lock
);
2198 /* Workaround for gcc 2.96 */
2199 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
2203 #endif /* __PAGETABLE_PMD_FOLDED */
2205 int make_pages_present(unsigned long addr
, unsigned long end
)
2207 int ret
, len
, write
;
2208 struct vm_area_struct
* vma
;
2210 vma
= find_vma(current
->mm
, addr
);
2213 write
= (vma
->vm_flags
& VM_WRITE
) != 0;
2216 if (end
> vma
->vm_end
)
2218 len
= (end
+PAGE_SIZE
-1)/PAGE_SIZE
-addr
/PAGE_SIZE
;
2219 ret
= get_user_pages(current
, current
->mm
, addr
,
2220 len
, write
, 0, NULL
, NULL
);
2223 return ret
== len
? 0 : -1;
2227 * Map a vmalloc()-space virtual address to the physical page.
2229 struct page
* vmalloc_to_page(void * vmalloc_addr
)
2231 unsigned long addr
= (unsigned long) vmalloc_addr
;
2232 struct page
*page
= NULL
;
2233 pgd_t
*pgd
= pgd_offset_k(addr
);
2238 if (!pgd_none(*pgd
)) {
2239 pud
= pud_offset(pgd
, addr
);
2240 if (!pud_none(*pud
)) {
2241 pmd
= pmd_offset(pud
, addr
);
2242 if (!pmd_none(*pmd
)) {
2243 ptep
= pte_offset_map(pmd
, addr
);
2245 if (pte_present(pte
))
2246 page
= pte_page(pte
);
2254 EXPORT_SYMBOL(vmalloc_to_page
);
2257 * Map a vmalloc()-space virtual address to the physical page frame number.
2259 unsigned long vmalloc_to_pfn(void * vmalloc_addr
)
2261 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
2264 EXPORT_SYMBOL(vmalloc_to_pfn
);
2266 #if !defined(__HAVE_ARCH_GATE_AREA)
2268 #if defined(AT_SYSINFO_EHDR)
2269 static struct vm_area_struct gate_vma
;
2271 static int __init
gate_vma_init(void)
2273 gate_vma
.vm_mm
= NULL
;
2274 gate_vma
.vm_start
= FIXADDR_USER_START
;
2275 gate_vma
.vm_end
= FIXADDR_USER_END
;
2276 gate_vma
.vm_page_prot
= PAGE_READONLY
;
2277 gate_vma
.vm_flags
= 0;
2280 __initcall(gate_vma_init
);
2283 struct vm_area_struct
*get_gate_vma(struct task_struct
*tsk
)
2285 #ifdef AT_SYSINFO_EHDR
2292 int in_gate_area_no_task(unsigned long addr
)
2294 #ifdef AT_SYSINFO_EHDR
2295 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
2301 #endif /* __HAVE_ARCH_GATE_AREA */