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/delayacct.h>
51 #include <linux/init.h>
52 #include <linux/writeback.h>
53 #include <linux/memcontrol.h>
54 #include <linux/mmu_notifier.h>
56 #include <asm/pgalloc.h>
57 #include <asm/uaccess.h>
59 #include <asm/tlbflush.h>
60 #include <asm/pgtable.h>
62 #include <linux/swapops.h>
63 #include <linux/elf.h>
67 #ifndef CONFIG_NEED_MULTIPLE_NODES
68 /* use the per-pgdat data instead for discontigmem - mbligh */
69 unsigned long max_mapnr
;
72 EXPORT_SYMBOL(max_mapnr
);
73 EXPORT_SYMBOL(mem_map
);
76 unsigned long num_physpages
;
78 * A number of key systems in x86 including ioremap() rely on the assumption
79 * that high_memory defines the upper bound on direct map memory, then end
80 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
81 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
86 EXPORT_SYMBOL(num_physpages
);
87 EXPORT_SYMBOL(high_memory
);
90 * Randomize the address space (stacks, mmaps, brk, etc.).
92 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
93 * as ancient (libc5 based) binaries can segfault. )
95 int randomize_va_space __read_mostly
=
96 #ifdef CONFIG_COMPAT_BRK
102 static int __init
disable_randmaps(char *s
)
104 randomize_va_space
= 0;
107 __setup("norandmaps", disable_randmaps
);
111 * If a p?d_bad entry is found while walking page tables, report
112 * the error, before resetting entry to p?d_none. Usually (but
113 * very seldom) called out from the p?d_none_or_clear_bad macros.
116 void pgd_clear_bad(pgd_t
*pgd
)
122 void pud_clear_bad(pud_t
*pud
)
128 void pmd_clear_bad(pmd_t
*pmd
)
135 * Note: this doesn't free the actual pages themselves. That
136 * has been handled earlier when unmapping all the memory regions.
138 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
)
140 pgtable_t token
= pmd_pgtable(*pmd
);
142 pte_free_tlb(tlb
, token
);
146 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
147 unsigned long addr
, unsigned long end
,
148 unsigned long floor
, unsigned long ceiling
)
155 pmd
= pmd_offset(pud
, addr
);
157 next
= pmd_addr_end(addr
, end
);
158 if (pmd_none_or_clear_bad(pmd
))
160 free_pte_range(tlb
, pmd
);
161 } while (pmd
++, addr
= next
, addr
!= end
);
171 if (end
- 1 > ceiling
- 1)
174 pmd
= pmd_offset(pud
, start
);
176 pmd_free_tlb(tlb
, pmd
);
179 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
180 unsigned long addr
, unsigned long end
,
181 unsigned long floor
, unsigned long ceiling
)
188 pud
= pud_offset(pgd
, addr
);
190 next
= pud_addr_end(addr
, end
);
191 if (pud_none_or_clear_bad(pud
))
193 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
194 } while (pud
++, addr
= next
, addr
!= end
);
200 ceiling
&= PGDIR_MASK
;
204 if (end
- 1 > ceiling
- 1)
207 pud
= pud_offset(pgd
, start
);
209 pud_free_tlb(tlb
, pud
);
213 * This function frees user-level page tables of a process.
215 * Must be called with pagetable lock held.
217 void free_pgd_range(struct mmu_gather
*tlb
,
218 unsigned long addr
, unsigned long end
,
219 unsigned long floor
, unsigned long ceiling
)
226 * The next few lines have given us lots of grief...
228 * Why are we testing PMD* at this top level? Because often
229 * there will be no work to do at all, and we'd prefer not to
230 * go all the way down to the bottom just to discover that.
232 * Why all these "- 1"s? Because 0 represents both the bottom
233 * of the address space and the top of it (using -1 for the
234 * top wouldn't help much: the masks would do the wrong thing).
235 * The rule is that addr 0 and floor 0 refer to the bottom of
236 * the address space, but end 0 and ceiling 0 refer to the top
237 * Comparisons need to use "end - 1" and "ceiling - 1" (though
238 * that end 0 case should be mythical).
240 * Wherever addr is brought up or ceiling brought down, we must
241 * be careful to reject "the opposite 0" before it confuses the
242 * subsequent tests. But what about where end is brought down
243 * by PMD_SIZE below? no, end can't go down to 0 there.
245 * Whereas we round start (addr) and ceiling down, by different
246 * masks at different levels, in order to test whether a table
247 * now has no other vmas using it, so can be freed, we don't
248 * bother to round floor or end up - the tests don't need that.
262 if (end
- 1 > ceiling
- 1)
268 pgd
= pgd_offset(tlb
->mm
, addr
);
270 next
= pgd_addr_end(addr
, end
);
271 if (pgd_none_or_clear_bad(pgd
))
273 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
274 } while (pgd
++, addr
= next
, addr
!= end
);
277 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
278 unsigned long floor
, unsigned long ceiling
)
281 struct vm_area_struct
*next
= vma
->vm_next
;
282 unsigned long addr
= vma
->vm_start
;
285 * Hide vma from rmap and vmtruncate before freeing pgtables
287 anon_vma_unlink(vma
);
288 unlink_file_vma(vma
);
290 if (is_vm_hugetlb_page(vma
)) {
291 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
292 floor
, next
? next
->vm_start
: ceiling
);
295 * Optimization: gather nearby vmas into one call down
297 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
298 && !is_vm_hugetlb_page(next
)) {
301 anon_vma_unlink(vma
);
302 unlink_file_vma(vma
);
304 free_pgd_range(tlb
, addr
, vma
->vm_end
,
305 floor
, next
? next
->vm_start
: ceiling
);
311 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
, unsigned long address
)
313 pgtable_t
new = pte_alloc_one(mm
, address
);
318 * Ensure all pte setup (eg. pte page lock and page clearing) are
319 * visible before the pte is made visible to other CPUs by being
320 * put into page tables.
322 * The other side of the story is the pointer chasing in the page
323 * table walking code (when walking the page table without locking;
324 * ie. most of the time). Fortunately, these data accesses consist
325 * of a chain of data-dependent loads, meaning most CPUs (alpha
326 * being the notable exception) will already guarantee loads are
327 * seen in-order. See the alpha page table accessors for the
328 * smp_read_barrier_depends() barriers in page table walking code.
330 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
332 spin_lock(&mm
->page_table_lock
);
333 if (!pmd_present(*pmd
)) { /* Has another populated it ? */
335 pmd_populate(mm
, pmd
, new);
338 spin_unlock(&mm
->page_table_lock
);
344 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
346 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
350 smp_wmb(); /* See comment in __pte_alloc */
352 spin_lock(&init_mm
.page_table_lock
);
353 if (!pmd_present(*pmd
)) { /* Has another populated it ? */
354 pmd_populate_kernel(&init_mm
, pmd
, new);
357 spin_unlock(&init_mm
.page_table_lock
);
359 pte_free_kernel(&init_mm
, new);
363 static inline void add_mm_rss(struct mm_struct
*mm
, int file_rss
, int anon_rss
)
366 add_mm_counter(mm
, file_rss
, file_rss
);
368 add_mm_counter(mm
, anon_rss
, anon_rss
);
372 * This function is called to print an error when a bad pte
373 * is found. For example, we might have a PFN-mapped pte in
374 * a region that doesn't allow it.
376 * The calling function must still handle the error.
378 static void print_bad_pte(struct vm_area_struct
*vma
, pte_t pte
,
381 printk(KERN_ERR
"Bad pte = %08llx, process = %s, "
382 "vm_flags = %lx, vaddr = %lx\n",
383 (long long)pte_val(pte
),
384 (vma
->vm_mm
== current
->mm
? current
->comm
: "???"),
385 vma
->vm_flags
, vaddr
);
389 static inline int is_cow_mapping(unsigned int flags
)
391 return (flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
395 * vm_normal_page -- This function gets the "struct page" associated with a pte.
397 * "Special" mappings do not wish to be associated with a "struct page" (either
398 * it doesn't exist, or it exists but they don't want to touch it). In this
399 * case, NULL is returned here. "Normal" mappings do have a struct page.
401 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
402 * pte bit, in which case this function is trivial. Secondly, an architecture
403 * may not have a spare pte bit, which requires a more complicated scheme,
406 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
407 * special mapping (even if there are underlying and valid "struct pages").
408 * COWed pages of a VM_PFNMAP are always normal.
410 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
411 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
412 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
413 * mapping will always honor the rule
415 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
417 * And for normal mappings this is false.
419 * This restricts such mappings to be a linear translation from virtual address
420 * to pfn. To get around this restriction, we allow arbitrary mappings so long
421 * as the vma is not a COW mapping; in that case, we know that all ptes are
422 * special (because none can have been COWed).
425 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
427 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
428 * page" backing, however the difference is that _all_ pages with a struct
429 * page (that is, those where pfn_valid is true) are refcounted and considered
430 * normal pages by the VM. The disadvantage is that pages are refcounted
431 * (which can be slower and simply not an option for some PFNMAP users). The
432 * advantage is that we don't have to follow the strict linearity rule of
433 * PFNMAP mappings in order to support COWable mappings.
436 #ifdef __HAVE_ARCH_PTE_SPECIAL
437 # define HAVE_PTE_SPECIAL 1
439 # define HAVE_PTE_SPECIAL 0
441 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
446 if (HAVE_PTE_SPECIAL
) {
447 if (likely(!pte_special(pte
))) {
448 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
449 return pte_page(pte
);
451 VM_BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
)));
455 /* !HAVE_PTE_SPECIAL case follows: */
459 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
460 if (vma
->vm_flags
& VM_MIXEDMAP
) {
466 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
467 if (pfn
== vma
->vm_pgoff
+ off
)
469 if (!is_cow_mapping(vma
->vm_flags
))
474 VM_BUG_ON(!pfn_valid(pfn
));
477 * NOTE! We still have PageReserved() pages in the page tables.
479 * eg. VDSO mappings can cause them to exist.
482 return pfn_to_page(pfn
);
486 * copy one vm_area from one task to the other. Assumes the page tables
487 * already present in the new task to be cleared in the whole range
488 * covered by this vma.
492 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
493 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
494 unsigned long addr
, int *rss
)
496 unsigned long vm_flags
= vma
->vm_flags
;
497 pte_t pte
= *src_pte
;
500 /* pte contains position in swap or file, so copy. */
501 if (unlikely(!pte_present(pte
))) {
502 if (!pte_file(pte
)) {
503 swp_entry_t entry
= pte_to_swp_entry(pte
);
505 swap_duplicate(entry
);
506 /* make sure dst_mm is on swapoff's mmlist. */
507 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
508 spin_lock(&mmlist_lock
);
509 if (list_empty(&dst_mm
->mmlist
))
510 list_add(&dst_mm
->mmlist
,
512 spin_unlock(&mmlist_lock
);
514 if (is_write_migration_entry(entry
) &&
515 is_cow_mapping(vm_flags
)) {
517 * COW mappings require pages in both parent
518 * and child to be set to read.
520 make_migration_entry_read(&entry
);
521 pte
= swp_entry_to_pte(entry
);
522 set_pte_at(src_mm
, addr
, src_pte
, pte
);
529 * If it's a COW mapping, write protect it both
530 * in the parent and the child
532 if (is_cow_mapping(vm_flags
)) {
533 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
534 pte
= pte_wrprotect(pte
);
538 * If it's a shared mapping, mark it clean in
541 if (vm_flags
& VM_SHARED
)
542 pte
= pte_mkclean(pte
);
543 pte
= pte_mkold(pte
);
545 page
= vm_normal_page(vma
, addr
, pte
);
548 page_dup_rmap(page
, vma
, addr
);
549 rss
[!!PageAnon(page
)]++;
553 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
556 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
557 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
558 unsigned long addr
, unsigned long end
)
560 pte_t
*src_pte
, *dst_pte
;
561 spinlock_t
*src_ptl
, *dst_ptl
;
567 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
570 src_pte
= pte_offset_map_nested(src_pmd
, addr
);
571 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
572 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
573 arch_enter_lazy_mmu_mode();
577 * We are holding two locks at this point - either of them
578 * could generate latencies in another task on another CPU.
580 if (progress
>= 32) {
582 if (need_resched() ||
583 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
586 if (pte_none(*src_pte
)) {
590 copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
, vma
, addr
, rss
);
592 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
594 arch_leave_lazy_mmu_mode();
595 spin_unlock(src_ptl
);
596 pte_unmap_nested(src_pte
- 1);
597 add_mm_rss(dst_mm
, rss
[0], rss
[1]);
598 pte_unmap_unlock(dst_pte
- 1, dst_ptl
);
605 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
606 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
607 unsigned long addr
, unsigned long end
)
609 pmd_t
*src_pmd
, *dst_pmd
;
612 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
615 src_pmd
= pmd_offset(src_pud
, addr
);
617 next
= pmd_addr_end(addr
, end
);
618 if (pmd_none_or_clear_bad(src_pmd
))
620 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
623 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
627 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
628 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
629 unsigned long addr
, unsigned long end
)
631 pud_t
*src_pud
, *dst_pud
;
634 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
637 src_pud
= pud_offset(src_pgd
, addr
);
639 next
= pud_addr_end(addr
, end
);
640 if (pud_none_or_clear_bad(src_pud
))
642 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
645 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
649 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
650 struct vm_area_struct
*vma
)
652 pgd_t
*src_pgd
, *dst_pgd
;
654 unsigned long addr
= vma
->vm_start
;
655 unsigned long end
= vma
->vm_end
;
659 * Don't copy ptes where a page fault will fill them correctly.
660 * Fork becomes much lighter when there are big shared or private
661 * readonly mappings. The tradeoff is that copy_page_range is more
662 * efficient than faulting.
664 if (!(vma
->vm_flags
& (VM_HUGETLB
|VM_NONLINEAR
|VM_PFNMAP
|VM_INSERTPAGE
))) {
669 if (is_vm_hugetlb_page(vma
))
670 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
673 * We need to invalidate the secondary MMU mappings only when
674 * there could be a permission downgrade on the ptes of the
675 * parent mm. And a permission downgrade will only happen if
676 * is_cow_mapping() returns true.
678 if (is_cow_mapping(vma
->vm_flags
))
679 mmu_notifier_invalidate_range_start(src_mm
, addr
, end
);
682 dst_pgd
= pgd_offset(dst_mm
, addr
);
683 src_pgd
= pgd_offset(src_mm
, addr
);
685 next
= pgd_addr_end(addr
, end
);
686 if (pgd_none_or_clear_bad(src_pgd
))
688 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
693 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
695 if (is_cow_mapping(vma
->vm_flags
))
696 mmu_notifier_invalidate_range_end(src_mm
,
701 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
702 struct vm_area_struct
*vma
, pmd_t
*pmd
,
703 unsigned long addr
, unsigned long end
,
704 long *zap_work
, struct zap_details
*details
)
706 struct mm_struct
*mm
= tlb
->mm
;
712 pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
713 arch_enter_lazy_mmu_mode();
716 if (pte_none(ptent
)) {
721 (*zap_work
) -= PAGE_SIZE
;
723 if (pte_present(ptent
)) {
726 page
= vm_normal_page(vma
, addr
, ptent
);
727 if (unlikely(details
) && page
) {
729 * unmap_shared_mapping_pages() wants to
730 * invalidate cache without truncating:
731 * unmap shared but keep private pages.
733 if (details
->check_mapping
&&
734 details
->check_mapping
!= page
->mapping
)
737 * Each page->index must be checked when
738 * invalidating or truncating nonlinear.
740 if (details
->nonlinear_vma
&&
741 (page
->index
< details
->first_index
||
742 page
->index
> details
->last_index
))
745 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
747 tlb_remove_tlb_entry(tlb
, pte
, addr
);
750 if (unlikely(details
) && details
->nonlinear_vma
751 && linear_page_index(details
->nonlinear_vma
,
752 addr
) != page
->index
)
753 set_pte_at(mm
, addr
, pte
,
754 pgoff_to_pte(page
->index
));
758 if (pte_dirty(ptent
))
759 set_page_dirty(page
);
760 if (pte_young(ptent
))
761 SetPageReferenced(page
);
764 page_remove_rmap(page
, vma
);
765 tlb_remove_page(tlb
, page
);
769 * If details->check_mapping, we leave swap entries;
770 * if details->nonlinear_vma, we leave file entries.
772 if (unlikely(details
))
774 if (!pte_file(ptent
))
775 free_swap_and_cache(pte_to_swp_entry(ptent
));
776 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
777 } while (pte
++, addr
+= PAGE_SIZE
, (addr
!= end
&& *zap_work
> 0));
779 add_mm_rss(mm
, file_rss
, anon_rss
);
780 arch_leave_lazy_mmu_mode();
781 pte_unmap_unlock(pte
- 1, ptl
);
786 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
787 struct vm_area_struct
*vma
, pud_t
*pud
,
788 unsigned long addr
, unsigned long end
,
789 long *zap_work
, struct zap_details
*details
)
794 pmd
= pmd_offset(pud
, addr
);
796 next
= pmd_addr_end(addr
, end
);
797 if (pmd_none_or_clear_bad(pmd
)) {
801 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
,
803 } while (pmd
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
808 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
809 struct vm_area_struct
*vma
, pgd_t
*pgd
,
810 unsigned long addr
, unsigned long end
,
811 long *zap_work
, struct zap_details
*details
)
816 pud
= pud_offset(pgd
, addr
);
818 next
= pud_addr_end(addr
, end
);
819 if (pud_none_or_clear_bad(pud
)) {
823 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
,
825 } while (pud
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
830 static unsigned long unmap_page_range(struct mmu_gather
*tlb
,
831 struct vm_area_struct
*vma
,
832 unsigned long addr
, unsigned long end
,
833 long *zap_work
, struct zap_details
*details
)
838 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
842 tlb_start_vma(tlb
, vma
);
843 pgd
= pgd_offset(vma
->vm_mm
, addr
);
845 next
= pgd_addr_end(addr
, end
);
846 if (pgd_none_or_clear_bad(pgd
)) {
850 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
,
852 } while (pgd
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
853 tlb_end_vma(tlb
, vma
);
858 #ifdef CONFIG_PREEMPT
859 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
861 /* No preempt: go for improved straight-line efficiency */
862 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
866 * unmap_vmas - unmap a range of memory covered by a list of vma's
867 * @tlbp: address of the caller's struct mmu_gather
868 * @vma: the starting vma
869 * @start_addr: virtual address at which to start unmapping
870 * @end_addr: virtual address at which to end unmapping
871 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
872 * @details: details of nonlinear truncation or shared cache invalidation
874 * Returns the end address of the unmapping (restart addr if interrupted).
876 * Unmap all pages in the vma list.
878 * We aim to not hold locks for too long (for scheduling latency reasons).
879 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
880 * return the ending mmu_gather to the caller.
882 * Only addresses between `start' and `end' will be unmapped.
884 * The VMA list must be sorted in ascending virtual address order.
886 * unmap_vmas() assumes that the caller will flush the whole unmapped address
887 * range after unmap_vmas() returns. So the only responsibility here is to
888 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
889 * drops the lock and schedules.
891 unsigned long unmap_vmas(struct mmu_gather
**tlbp
,
892 struct vm_area_struct
*vma
, unsigned long start_addr
,
893 unsigned long end_addr
, unsigned long *nr_accounted
,
894 struct zap_details
*details
)
896 long zap_work
= ZAP_BLOCK_SIZE
;
897 unsigned long tlb_start
= 0; /* For tlb_finish_mmu */
898 int tlb_start_valid
= 0;
899 unsigned long start
= start_addr
;
900 spinlock_t
*i_mmap_lock
= details
? details
->i_mmap_lock
: NULL
;
901 int fullmm
= (*tlbp
)->fullmm
;
902 struct mm_struct
*mm
= vma
->vm_mm
;
904 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
905 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
) {
908 start
= max(vma
->vm_start
, start_addr
);
909 if (start
>= vma
->vm_end
)
911 end
= min(vma
->vm_end
, end_addr
);
912 if (end
<= vma
->vm_start
)
915 if (vma
->vm_flags
& VM_ACCOUNT
)
916 *nr_accounted
+= (end
- start
) >> PAGE_SHIFT
;
918 while (start
!= end
) {
919 if (!tlb_start_valid
) {
924 if (unlikely(is_vm_hugetlb_page(vma
))) {
926 * It is undesirable to test vma->vm_file as it
927 * should be non-null for valid hugetlb area.
928 * However, vm_file will be NULL in the error
929 * cleanup path of do_mmap_pgoff. When
930 * hugetlbfs ->mmap method fails,
931 * do_mmap_pgoff() nullifies vma->vm_file
932 * before calling this function to clean up.
933 * Since no pte has actually been setup, it is
934 * safe to do nothing in this case.
937 unmap_hugepage_range(vma
, start
, end
, NULL
);
938 zap_work
-= (end
- start
) /
939 pages_per_huge_page(hstate_vma(vma
));
944 start
= unmap_page_range(*tlbp
, vma
,
945 start
, end
, &zap_work
, details
);
948 BUG_ON(start
!= end
);
952 tlb_finish_mmu(*tlbp
, tlb_start
, start
);
954 if (need_resched() ||
955 (i_mmap_lock
&& spin_needbreak(i_mmap_lock
))) {
963 *tlbp
= tlb_gather_mmu(vma
->vm_mm
, fullmm
);
965 zap_work
= ZAP_BLOCK_SIZE
;
969 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
970 return start
; /* which is now the end (or restart) address */
974 * zap_page_range - remove user pages in a given range
975 * @vma: vm_area_struct holding the applicable pages
976 * @address: starting address of pages to zap
977 * @size: number of bytes to zap
978 * @details: details of nonlinear truncation or shared cache invalidation
980 unsigned long zap_page_range(struct vm_area_struct
*vma
, unsigned long address
,
981 unsigned long size
, struct zap_details
*details
)
983 struct mm_struct
*mm
= vma
->vm_mm
;
984 struct mmu_gather
*tlb
;
985 unsigned long end
= address
+ size
;
986 unsigned long nr_accounted
= 0;
989 tlb
= tlb_gather_mmu(mm
, 0);
990 update_hiwater_rss(mm
);
991 end
= unmap_vmas(&tlb
, vma
, address
, end
, &nr_accounted
, details
);
993 tlb_finish_mmu(tlb
, address
, end
);
996 EXPORT_SYMBOL_GPL(zap_page_range
);
999 * zap_vma_ptes - remove ptes mapping the vma
1000 * @vma: vm_area_struct holding ptes to be zapped
1001 * @address: starting address of pages to zap
1002 * @size: number of bytes to zap
1004 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1006 * The entire address range must be fully contained within the vma.
1008 * Returns 0 if successful.
1010 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1013 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1014 !(vma
->vm_flags
& VM_PFNMAP
))
1016 zap_page_range(vma
, address
, size
, NULL
);
1019 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1022 * Do a quick page-table lookup for a single page.
1024 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
1033 struct mm_struct
*mm
= vma
->vm_mm
;
1035 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
1036 if (!IS_ERR(page
)) {
1037 BUG_ON(flags
& FOLL_GET
);
1042 pgd
= pgd_offset(mm
, address
);
1043 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
1046 pud
= pud_offset(pgd
, address
);
1049 if (pud_huge(*pud
)) {
1050 BUG_ON(flags
& FOLL_GET
);
1051 page
= follow_huge_pud(mm
, address
, pud
, flags
& FOLL_WRITE
);
1054 if (unlikely(pud_bad(*pud
)))
1057 pmd
= pmd_offset(pud
, address
);
1060 if (pmd_huge(*pmd
)) {
1061 BUG_ON(flags
& FOLL_GET
);
1062 page
= follow_huge_pmd(mm
, address
, pmd
, flags
& FOLL_WRITE
);
1065 if (unlikely(pmd_bad(*pmd
)))
1068 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1071 if (!pte_present(pte
))
1073 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
1075 page
= vm_normal_page(vma
, address
, pte
);
1076 if (unlikely(!page
))
1079 if (flags
& FOLL_GET
)
1081 if (flags
& FOLL_TOUCH
) {
1082 if ((flags
& FOLL_WRITE
) &&
1083 !pte_dirty(pte
) && !PageDirty(page
))
1084 set_page_dirty(page
);
1085 mark_page_accessed(page
);
1088 pte_unmap_unlock(ptep
, ptl
);
1093 pte_unmap_unlock(ptep
, ptl
);
1094 return ERR_PTR(-EFAULT
);
1097 pte_unmap_unlock(ptep
, ptl
);
1100 /* Fall through to ZERO_PAGE handling */
1103 * When core dumping an enormous anonymous area that nobody
1104 * has touched so far, we don't want to allocate page tables.
1106 if (flags
& FOLL_ANON
) {
1107 page
= ZERO_PAGE(0);
1108 if (flags
& FOLL_GET
)
1110 BUG_ON(flags
& FOLL_WRITE
);
1114 EXPORT_SYMBOL_GPL(follow_page
);
1116 /* Can we do the FOLL_ANON optimization? */
1117 static inline int use_zero_page(struct vm_area_struct
*vma
)
1120 * We don't want to optimize FOLL_ANON for make_pages_present()
1121 * when it tries to page in a VM_LOCKED region. As to VM_SHARED,
1122 * we want to get the page from the page tables to make sure
1123 * that we serialize and update with any other user of that
1126 if (vma
->vm_flags
& (VM_LOCKED
| VM_SHARED
))
1129 * And if we have a fault routine, it's not an anonymous region.
1131 return !vma
->vm_ops
|| !vma
->vm_ops
->fault
;
1134 int get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1135 unsigned long start
, int len
, int write
, int force
,
1136 struct page
**pages
, struct vm_area_struct
**vmas
)
1139 unsigned int vm_flags
;
1144 * Require read or write permissions.
1145 * If 'force' is set, we only require the "MAY" flags.
1147 vm_flags
= write
? (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1148 vm_flags
&= force
? (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1152 struct vm_area_struct
*vma
;
1153 unsigned int foll_flags
;
1155 vma
= find_extend_vma(mm
, start
);
1156 if (!vma
&& in_gate_area(tsk
, start
)) {
1157 unsigned long pg
= start
& PAGE_MASK
;
1158 struct vm_area_struct
*gate_vma
= get_gate_vma(tsk
);
1163 if (write
) /* user gate pages are read-only */
1164 return i
? : -EFAULT
;
1166 pgd
= pgd_offset_k(pg
);
1168 pgd
= pgd_offset_gate(mm
, pg
);
1169 BUG_ON(pgd_none(*pgd
));
1170 pud
= pud_offset(pgd
, pg
);
1171 BUG_ON(pud_none(*pud
));
1172 pmd
= pmd_offset(pud
, pg
);
1174 return i
? : -EFAULT
;
1175 pte
= pte_offset_map(pmd
, pg
);
1176 if (pte_none(*pte
)) {
1178 return i
? : -EFAULT
;
1181 struct page
*page
= vm_normal_page(gate_vma
, start
, *pte
);
1195 if (!vma
|| (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1196 || !(vm_flags
& vma
->vm_flags
))
1197 return i
? : -EFAULT
;
1199 if (is_vm_hugetlb_page(vma
)) {
1200 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1201 &start
, &len
, i
, write
);
1205 foll_flags
= FOLL_TOUCH
;
1207 foll_flags
|= FOLL_GET
;
1208 if (!write
&& use_zero_page(vma
))
1209 foll_flags
|= FOLL_ANON
;
1215 * If tsk is ooming, cut off its access to large memory
1216 * allocations. It has a pending SIGKILL, but it can't
1217 * be processed until returning to user space.
1219 if (unlikely(test_tsk_thread_flag(tsk
, TIF_MEMDIE
)))
1220 return i
? i
: -ENOMEM
;
1223 foll_flags
|= FOLL_WRITE
;
1226 while (!(page
= follow_page(vma
, start
, foll_flags
))) {
1228 ret
= handle_mm_fault(mm
, vma
, start
,
1229 foll_flags
& FOLL_WRITE
);
1230 if (ret
& VM_FAULT_ERROR
) {
1231 if (ret
& VM_FAULT_OOM
)
1232 return i
? i
: -ENOMEM
;
1233 else if (ret
& VM_FAULT_SIGBUS
)
1234 return i
? i
: -EFAULT
;
1237 if (ret
& VM_FAULT_MAJOR
)
1243 * The VM_FAULT_WRITE bit tells us that
1244 * do_wp_page has broken COW when necessary,
1245 * even if maybe_mkwrite decided not to set
1246 * pte_write. We can thus safely do subsequent
1247 * page lookups as if they were reads.
1249 if (ret
& VM_FAULT_WRITE
)
1250 foll_flags
&= ~FOLL_WRITE
;
1255 return i
? i
: PTR_ERR(page
);
1259 flush_anon_page(vma
, page
, start
);
1260 flush_dcache_page(page
);
1267 } while (len
&& start
< vma
->vm_end
);
1271 EXPORT_SYMBOL(get_user_pages
);
1273 pte_t
*get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1276 pgd_t
* pgd
= pgd_offset(mm
, addr
);
1277 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
1279 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
1281 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1287 * This is the old fallback for page remapping.
1289 * For historical reasons, it only allows reserved pages. Only
1290 * old drivers should use this, and they needed to mark their
1291 * pages reserved for the old functions anyway.
1293 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1294 struct page
*page
, pgprot_t prot
)
1296 struct mm_struct
*mm
= vma
->vm_mm
;
1301 retval
= mem_cgroup_charge(page
, mm
, GFP_KERNEL
);
1309 flush_dcache_page(page
);
1310 pte
= get_locked_pte(mm
, addr
, &ptl
);
1314 if (!pte_none(*pte
))
1317 /* Ok, finally just insert the thing.. */
1319 inc_mm_counter(mm
, file_rss
);
1320 page_add_file_rmap(page
);
1321 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1324 pte_unmap_unlock(pte
, ptl
);
1327 pte_unmap_unlock(pte
, ptl
);
1329 mem_cgroup_uncharge_page(page
);
1335 * vm_insert_page - insert single page into user vma
1336 * @vma: user vma to map to
1337 * @addr: target user address of this page
1338 * @page: source kernel page
1340 * This allows drivers to insert individual pages they've allocated
1343 * The page has to be a nice clean _individual_ kernel allocation.
1344 * If you allocate a compound page, you need to have marked it as
1345 * such (__GFP_COMP), or manually just split the page up yourself
1346 * (see split_page()).
1348 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1349 * took an arbitrary page protection parameter. This doesn't allow
1350 * that. Your vma protection will have to be set up correctly, which
1351 * means that if you want a shared writable mapping, you'd better
1352 * ask for a shared writable mapping!
1354 * The page does not need to be reserved.
1356 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1359 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1361 if (!page_count(page
))
1363 vma
->vm_flags
|= VM_INSERTPAGE
;
1364 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1366 EXPORT_SYMBOL(vm_insert_page
);
1368 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1369 unsigned long pfn
, pgprot_t prot
)
1371 struct mm_struct
*mm
= vma
->vm_mm
;
1377 pte
= get_locked_pte(mm
, addr
, &ptl
);
1381 if (!pte_none(*pte
))
1384 /* Ok, finally just insert the thing.. */
1385 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
1386 set_pte_at(mm
, addr
, pte
, entry
);
1387 update_mmu_cache(vma
, addr
, entry
); /* XXX: why not for insert_page? */
1391 pte_unmap_unlock(pte
, ptl
);
1397 * vm_insert_pfn - insert single pfn into user vma
1398 * @vma: user vma to map to
1399 * @addr: target user address of this page
1400 * @pfn: source kernel pfn
1402 * Similar to vm_inert_page, this allows drivers to insert individual pages
1403 * they've allocated into a user vma. Same comments apply.
1405 * This function should only be called from a vm_ops->fault handler, and
1406 * in that case the handler should return NULL.
1408 * vma cannot be a COW mapping.
1410 * As this is called only for pages that do not currently exist, we
1411 * do not need to flush old virtual caches or the TLB.
1413 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1417 * Technically, architectures with pte_special can avoid all these
1418 * restrictions (same for remap_pfn_range). However we would like
1419 * consistency in testing and feature parity among all, so we should
1420 * try to keep these invariants in place for everybody.
1422 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1423 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1424 (VM_PFNMAP
|VM_MIXEDMAP
));
1425 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1426 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1428 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1430 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1432 EXPORT_SYMBOL(vm_insert_pfn
);
1434 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1437 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
1439 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1443 * If we don't have pte special, then we have to use the pfn_valid()
1444 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1445 * refcount the page if pfn_valid is true (hence insert_page rather
1448 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
1451 page
= pfn_to_page(pfn
);
1452 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1454 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1456 EXPORT_SYMBOL(vm_insert_mixed
);
1459 * maps a range of physical memory into the requested pages. the old
1460 * mappings are removed. any references to nonexistent pages results
1461 * in null mappings (currently treated as "copy-on-access")
1463 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1464 unsigned long addr
, unsigned long end
,
1465 unsigned long pfn
, pgprot_t prot
)
1470 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1473 arch_enter_lazy_mmu_mode();
1475 BUG_ON(!pte_none(*pte
));
1476 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
1478 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1479 arch_leave_lazy_mmu_mode();
1480 pte_unmap_unlock(pte
- 1, ptl
);
1484 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1485 unsigned long addr
, unsigned long end
,
1486 unsigned long pfn
, pgprot_t prot
)
1491 pfn
-= addr
>> PAGE_SHIFT
;
1492 pmd
= pmd_alloc(mm
, pud
, addr
);
1496 next
= pmd_addr_end(addr
, end
);
1497 if (remap_pte_range(mm
, pmd
, addr
, next
,
1498 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1500 } while (pmd
++, addr
= next
, addr
!= end
);
1504 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1505 unsigned long addr
, unsigned long end
,
1506 unsigned long pfn
, pgprot_t prot
)
1511 pfn
-= addr
>> PAGE_SHIFT
;
1512 pud
= pud_alloc(mm
, pgd
, addr
);
1516 next
= pud_addr_end(addr
, end
);
1517 if (remap_pmd_range(mm
, pud
, addr
, next
,
1518 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1520 } while (pud
++, addr
= next
, addr
!= end
);
1525 * remap_pfn_range - remap kernel memory to userspace
1526 * @vma: user vma to map to
1527 * @addr: target user address to start at
1528 * @pfn: physical address of kernel memory
1529 * @size: size of map area
1530 * @prot: page protection flags for this mapping
1532 * Note: this is only safe if the mm semaphore is held when called.
1534 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1535 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1539 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1540 struct mm_struct
*mm
= vma
->vm_mm
;
1544 * Physically remapped pages are special. Tell the
1545 * rest of the world about it:
1546 * VM_IO tells people not to look at these pages
1547 * (accesses can have side effects).
1548 * VM_RESERVED is specified all over the place, because
1549 * in 2.4 it kept swapout's vma scan off this vma; but
1550 * in 2.6 the LRU scan won't even find its pages, so this
1551 * flag means no more than count its pages in reserved_vm,
1552 * and omit it from core dump, even when VM_IO turned off.
1553 * VM_PFNMAP tells the core MM that the base pages are just
1554 * raw PFN mappings, and do not have a "struct page" associated
1557 * There's a horrible special case to handle copy-on-write
1558 * behaviour that some programs depend on. We mark the "original"
1559 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1561 if (is_cow_mapping(vma
->vm_flags
)) {
1562 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
1564 vma
->vm_pgoff
= pfn
;
1567 vma
->vm_flags
|= VM_IO
| VM_RESERVED
| VM_PFNMAP
;
1569 BUG_ON(addr
>= end
);
1570 pfn
-= addr
>> PAGE_SHIFT
;
1571 pgd
= pgd_offset(mm
, addr
);
1572 flush_cache_range(vma
, addr
, end
);
1574 next
= pgd_addr_end(addr
, end
);
1575 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1576 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1579 } while (pgd
++, addr
= next
, addr
!= end
);
1582 EXPORT_SYMBOL(remap_pfn_range
);
1584 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1585 unsigned long addr
, unsigned long end
,
1586 pte_fn_t fn
, void *data
)
1591 spinlock_t
*uninitialized_var(ptl
);
1593 pte
= (mm
== &init_mm
) ?
1594 pte_alloc_kernel(pmd
, addr
) :
1595 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1599 BUG_ON(pmd_huge(*pmd
));
1601 token
= pmd_pgtable(*pmd
);
1604 err
= fn(pte
, token
, addr
, data
);
1607 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1610 pte_unmap_unlock(pte
-1, ptl
);
1614 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1615 unsigned long addr
, unsigned long end
,
1616 pte_fn_t fn
, void *data
)
1622 BUG_ON(pud_huge(*pud
));
1624 pmd
= pmd_alloc(mm
, pud
, addr
);
1628 next
= pmd_addr_end(addr
, end
);
1629 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
1632 } while (pmd
++, addr
= next
, addr
!= end
);
1636 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1637 unsigned long addr
, unsigned long end
,
1638 pte_fn_t fn
, void *data
)
1644 pud
= pud_alloc(mm
, pgd
, addr
);
1648 next
= pud_addr_end(addr
, end
);
1649 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
1652 } while (pud
++, addr
= next
, addr
!= end
);
1657 * Scan a region of virtual memory, filling in page tables as necessary
1658 * and calling a provided function on each leaf page table.
1660 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
1661 unsigned long size
, pte_fn_t fn
, void *data
)
1665 unsigned long start
= addr
, end
= addr
+ size
;
1668 BUG_ON(addr
>= end
);
1669 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1670 pgd
= pgd_offset(mm
, addr
);
1672 next
= pgd_addr_end(addr
, end
);
1673 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
1676 } while (pgd
++, addr
= next
, addr
!= end
);
1677 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1680 EXPORT_SYMBOL_GPL(apply_to_page_range
);
1683 * handle_pte_fault chooses page fault handler according to an entry
1684 * which was read non-atomically. Before making any commitment, on
1685 * those architectures or configurations (e.g. i386 with PAE) which
1686 * might give a mix of unmatched parts, do_swap_page and do_file_page
1687 * must check under lock before unmapping the pte and proceeding
1688 * (but do_wp_page is only called after already making such a check;
1689 * and do_anonymous_page and do_no_page can safely check later on).
1691 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
1692 pte_t
*page_table
, pte_t orig_pte
)
1695 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1696 if (sizeof(pte_t
) > sizeof(unsigned long)) {
1697 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
1699 same
= pte_same(*page_table
, orig_pte
);
1703 pte_unmap(page_table
);
1708 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1709 * servicing faults for write access. In the normal case, do always want
1710 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1711 * that do not have writing enabled, when used by access_process_vm.
1713 static inline pte_t
maybe_mkwrite(pte_t pte
, struct vm_area_struct
*vma
)
1715 if (likely(vma
->vm_flags
& VM_WRITE
))
1716 pte
= pte_mkwrite(pte
);
1720 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
1723 * If the source page was a PFN mapping, we don't have
1724 * a "struct page" for it. We do a best-effort copy by
1725 * just copying from the original user address. If that
1726 * fails, we just zero-fill it. Live with it.
1728 if (unlikely(!src
)) {
1729 void *kaddr
= kmap_atomic(dst
, KM_USER0
);
1730 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
1733 * This really shouldn't fail, because the page is there
1734 * in the page tables. But it might just be unreadable,
1735 * in which case we just give up and fill the result with
1738 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
1739 memset(kaddr
, 0, PAGE_SIZE
);
1740 kunmap_atomic(kaddr
, KM_USER0
);
1741 flush_dcache_page(dst
);
1743 copy_user_highpage(dst
, src
, va
, vma
);
1747 * This routine handles present pages, when users try to write
1748 * to a shared page. It is done by copying the page to a new address
1749 * and decrementing the shared-page counter for the old page.
1751 * Note that this routine assumes that the protection checks have been
1752 * done by the caller (the low-level page fault routine in most cases).
1753 * Thus we can safely just mark it writable once we've done any necessary
1756 * We also mark the page dirty at this point even though the page will
1757 * change only once the write actually happens. This avoids a few races,
1758 * and potentially makes it more efficient.
1760 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1761 * but allow concurrent faults), with pte both mapped and locked.
1762 * We return with mmap_sem still held, but pte unmapped and unlocked.
1764 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1765 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1766 spinlock_t
*ptl
, pte_t orig_pte
)
1768 struct page
*old_page
, *new_page
;
1770 int reuse
= 0, ret
= 0;
1771 int page_mkwrite
= 0;
1772 struct page
*dirty_page
= NULL
;
1774 old_page
= vm_normal_page(vma
, address
, orig_pte
);
1777 * VM_MIXEDMAP !pfn_valid() case
1779 * We should not cow pages in a shared writeable mapping.
1780 * Just mark the pages writable as we can't do any dirty
1781 * accounting on raw pfn maps.
1783 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
1784 (VM_WRITE
|VM_SHARED
))
1790 * Take out anonymous pages first, anonymous shared vmas are
1791 * not dirty accountable.
1793 if (PageAnon(old_page
)) {
1794 if (!TestSetPageLocked(old_page
)) {
1795 reuse
= can_share_swap_page(old_page
);
1796 unlock_page(old_page
);
1798 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
1799 (VM_WRITE
|VM_SHARED
))) {
1801 * Only catch write-faults on shared writable pages,
1802 * read-only shared pages can get COWed by
1803 * get_user_pages(.write=1, .force=1).
1805 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
1807 * Notify the address space that the page is about to
1808 * become writable so that it can prohibit this or wait
1809 * for the page to get into an appropriate state.
1811 * We do this without the lock held, so that it can
1812 * sleep if it needs to.
1814 page_cache_get(old_page
);
1815 pte_unmap_unlock(page_table
, ptl
);
1817 if (vma
->vm_ops
->page_mkwrite(vma
, old_page
) < 0)
1818 goto unwritable_page
;
1821 * Since we dropped the lock we need to revalidate
1822 * the PTE as someone else may have changed it. If
1823 * they did, we just return, as we can count on the
1824 * MMU to tell us if they didn't also make it writable.
1826 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
1828 page_cache_release(old_page
);
1829 if (!pte_same(*page_table
, orig_pte
))
1834 dirty_page
= old_page
;
1835 get_page(dirty_page
);
1841 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
1842 entry
= pte_mkyoung(orig_pte
);
1843 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1844 if (ptep_set_access_flags(vma
, address
, page_table
, entry
,1))
1845 update_mmu_cache(vma
, address
, entry
);
1846 ret
|= VM_FAULT_WRITE
;
1851 * Ok, we need to copy. Oh, well..
1853 page_cache_get(old_page
);
1855 pte_unmap_unlock(page_table
, ptl
);
1857 if (unlikely(anon_vma_prepare(vma
)))
1859 VM_BUG_ON(old_page
== ZERO_PAGE(0));
1860 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
1863 cow_user_page(new_page
, old_page
, address
, vma
);
1864 __SetPageUptodate(new_page
);
1866 if (mem_cgroup_charge(new_page
, mm
, GFP_KERNEL
))
1870 * Re-check the pte - we dropped the lock
1872 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1873 if (likely(pte_same(*page_table
, orig_pte
))) {
1875 if (!PageAnon(old_page
)) {
1876 dec_mm_counter(mm
, file_rss
);
1877 inc_mm_counter(mm
, anon_rss
);
1880 inc_mm_counter(mm
, anon_rss
);
1881 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
1882 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
1883 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1885 * Clear the pte entry and flush it first, before updating the
1886 * pte with the new entry. This will avoid a race condition
1887 * seen in the presence of one thread doing SMC and another
1890 ptep_clear_flush_notify(vma
, address
, page_table
);
1891 set_pte_at(mm
, address
, page_table
, entry
);
1892 update_mmu_cache(vma
, address
, entry
);
1893 lru_cache_add_active(new_page
);
1894 page_add_new_anon_rmap(new_page
, vma
, address
);
1898 * Only after switching the pte to the new page may
1899 * we remove the mapcount here. Otherwise another
1900 * process may come and find the rmap count decremented
1901 * before the pte is switched to the new page, and
1902 * "reuse" the old page writing into it while our pte
1903 * here still points into it and can be read by other
1906 * The critical issue is to order this
1907 * page_remove_rmap with the ptp_clear_flush above.
1908 * Those stores are ordered by (if nothing else,)
1909 * the barrier present in the atomic_add_negative
1910 * in page_remove_rmap.
1912 * Then the TLB flush in ptep_clear_flush ensures that
1913 * no process can access the old page before the
1914 * decremented mapcount is visible. And the old page
1915 * cannot be reused until after the decremented
1916 * mapcount is visible. So transitively, TLBs to
1917 * old page will be flushed before it can be reused.
1919 page_remove_rmap(old_page
, vma
);
1922 /* Free the old page.. */
1923 new_page
= old_page
;
1924 ret
|= VM_FAULT_WRITE
;
1926 mem_cgroup_uncharge_page(new_page
);
1929 page_cache_release(new_page
);
1931 page_cache_release(old_page
);
1933 pte_unmap_unlock(page_table
, ptl
);
1936 file_update_time(vma
->vm_file
);
1939 * Yes, Virginia, this is actually required to prevent a race
1940 * with clear_page_dirty_for_io() from clearing the page dirty
1941 * bit after it clear all dirty ptes, but before a racing
1942 * do_wp_page installs a dirty pte.
1944 * do_no_page is protected similarly.
1946 wait_on_page_locked(dirty_page
);
1947 set_page_dirty_balance(dirty_page
, page_mkwrite
);
1948 put_page(dirty_page
);
1952 page_cache_release(new_page
);
1955 page_cache_release(old_page
);
1956 return VM_FAULT_OOM
;
1959 page_cache_release(old_page
);
1960 return VM_FAULT_SIGBUS
;
1964 * Helper functions for unmap_mapping_range().
1966 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1968 * We have to restart searching the prio_tree whenever we drop the lock,
1969 * since the iterator is only valid while the lock is held, and anyway
1970 * a later vma might be split and reinserted earlier while lock dropped.
1972 * The list of nonlinear vmas could be handled more efficiently, using
1973 * a placeholder, but handle it in the same way until a need is shown.
1974 * It is important to search the prio_tree before nonlinear list: a vma
1975 * may become nonlinear and be shifted from prio_tree to nonlinear list
1976 * while the lock is dropped; but never shifted from list to prio_tree.
1978 * In order to make forward progress despite restarting the search,
1979 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1980 * quickly skip it next time around. Since the prio_tree search only
1981 * shows us those vmas affected by unmapping the range in question, we
1982 * can't efficiently keep all vmas in step with mapping->truncate_count:
1983 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1984 * mapping->truncate_count and vma->vm_truncate_count are protected by
1987 * In order to make forward progress despite repeatedly restarting some
1988 * large vma, note the restart_addr from unmap_vmas when it breaks out:
1989 * and restart from that address when we reach that vma again. It might
1990 * have been split or merged, shrunk or extended, but never shifted: so
1991 * restart_addr remains valid so long as it remains in the vma's range.
1992 * unmap_mapping_range forces truncate_count to leap over page-aligned
1993 * values so we can save vma's restart_addr in its truncate_count field.
1995 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1997 static void reset_vma_truncate_counts(struct address_space
*mapping
)
1999 struct vm_area_struct
*vma
;
2000 struct prio_tree_iter iter
;
2002 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, 0, ULONG_MAX
)
2003 vma
->vm_truncate_count
= 0;
2004 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
2005 vma
->vm_truncate_count
= 0;
2008 static int unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2009 unsigned long start_addr
, unsigned long end_addr
,
2010 struct zap_details
*details
)
2012 unsigned long restart_addr
;
2016 * files that support invalidating or truncating portions of the
2017 * file from under mmaped areas must have their ->fault function
2018 * return a locked page (and set VM_FAULT_LOCKED in the return).
2019 * This provides synchronisation against concurrent unmapping here.
2023 restart_addr
= vma
->vm_truncate_count
;
2024 if (is_restart_addr(restart_addr
) && start_addr
< restart_addr
) {
2025 start_addr
= restart_addr
;
2026 if (start_addr
>= end_addr
) {
2027 /* Top of vma has been split off since last time */
2028 vma
->vm_truncate_count
= details
->truncate_count
;
2033 restart_addr
= zap_page_range(vma
, start_addr
,
2034 end_addr
- start_addr
, details
);
2035 need_break
= need_resched() || spin_needbreak(details
->i_mmap_lock
);
2037 if (restart_addr
>= end_addr
) {
2038 /* We have now completed this vma: mark it so */
2039 vma
->vm_truncate_count
= details
->truncate_count
;
2043 /* Note restart_addr in vma's truncate_count field */
2044 vma
->vm_truncate_count
= restart_addr
;
2049 spin_unlock(details
->i_mmap_lock
);
2051 spin_lock(details
->i_mmap_lock
);
2055 static inline void unmap_mapping_range_tree(struct prio_tree_root
*root
,
2056 struct zap_details
*details
)
2058 struct vm_area_struct
*vma
;
2059 struct prio_tree_iter iter
;
2060 pgoff_t vba
, vea
, zba
, zea
;
2063 vma_prio_tree_foreach(vma
, &iter
, root
,
2064 details
->first_index
, details
->last_index
) {
2065 /* Skip quickly over those we have already dealt with */
2066 if (vma
->vm_truncate_count
== details
->truncate_count
)
2069 vba
= vma
->vm_pgoff
;
2070 vea
= vba
+ ((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) - 1;
2071 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2072 zba
= details
->first_index
;
2075 zea
= details
->last_index
;
2079 if (unmap_mapping_range_vma(vma
,
2080 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2081 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2087 static inline void unmap_mapping_range_list(struct list_head
*head
,
2088 struct zap_details
*details
)
2090 struct vm_area_struct
*vma
;
2093 * In nonlinear VMAs there is no correspondence between virtual address
2094 * offset and file offset. So we must perform an exhaustive search
2095 * across *all* the pages in each nonlinear VMA, not just the pages
2096 * whose virtual address lies outside the file truncation point.
2099 list_for_each_entry(vma
, head
, shared
.vm_set
.list
) {
2100 /* Skip quickly over those we have already dealt with */
2101 if (vma
->vm_truncate_count
== details
->truncate_count
)
2103 details
->nonlinear_vma
= vma
;
2104 if (unmap_mapping_range_vma(vma
, vma
->vm_start
,
2105 vma
->vm_end
, details
) < 0)
2111 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2112 * @mapping: the address space containing mmaps to be unmapped.
2113 * @holebegin: byte in first page to unmap, relative to the start of
2114 * the underlying file. This will be rounded down to a PAGE_SIZE
2115 * boundary. Note that this is different from vmtruncate(), which
2116 * must keep the partial page. In contrast, we must get rid of
2118 * @holelen: size of prospective hole in bytes. This will be rounded
2119 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2121 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2122 * but 0 when invalidating pagecache, don't throw away private data.
2124 void unmap_mapping_range(struct address_space
*mapping
,
2125 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2127 struct zap_details details
;
2128 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2129 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2131 /* Check for overflow. */
2132 if (sizeof(holelen
) > sizeof(hlen
)) {
2134 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2135 if (holeend
& ~(long long)ULONG_MAX
)
2136 hlen
= ULONG_MAX
- hba
+ 1;
2139 details
.check_mapping
= even_cows
? NULL
: mapping
;
2140 details
.nonlinear_vma
= NULL
;
2141 details
.first_index
= hba
;
2142 details
.last_index
= hba
+ hlen
- 1;
2143 if (details
.last_index
< details
.first_index
)
2144 details
.last_index
= ULONG_MAX
;
2145 details
.i_mmap_lock
= &mapping
->i_mmap_lock
;
2147 spin_lock(&mapping
->i_mmap_lock
);
2149 /* Protect against endless unmapping loops */
2150 mapping
->truncate_count
++;
2151 if (unlikely(is_restart_addr(mapping
->truncate_count
))) {
2152 if (mapping
->truncate_count
== 0)
2153 reset_vma_truncate_counts(mapping
);
2154 mapping
->truncate_count
++;
2156 details
.truncate_count
= mapping
->truncate_count
;
2158 if (unlikely(!prio_tree_empty(&mapping
->i_mmap
)))
2159 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2160 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
2161 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
2162 spin_unlock(&mapping
->i_mmap_lock
);
2164 EXPORT_SYMBOL(unmap_mapping_range
);
2167 * vmtruncate - unmap mappings "freed" by truncate() syscall
2168 * @inode: inode of the file used
2169 * @offset: file offset to start truncating
2171 * NOTE! We have to be ready to update the memory sharing
2172 * between the file and the memory map for a potential last
2173 * incomplete page. Ugly, but necessary.
2175 int vmtruncate(struct inode
* inode
, loff_t offset
)
2177 if (inode
->i_size
< offset
) {
2178 unsigned long limit
;
2180 limit
= current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
;
2181 if (limit
!= RLIM_INFINITY
&& offset
> limit
)
2183 if (offset
> inode
->i_sb
->s_maxbytes
)
2185 i_size_write(inode
, offset
);
2187 struct address_space
*mapping
= inode
->i_mapping
;
2190 * truncation of in-use swapfiles is disallowed - it would
2191 * cause subsequent swapout to scribble on the now-freed
2194 if (IS_SWAPFILE(inode
))
2196 i_size_write(inode
, offset
);
2199 * unmap_mapping_range is called twice, first simply for
2200 * efficiency so that truncate_inode_pages does fewer
2201 * single-page unmaps. However after this first call, and
2202 * before truncate_inode_pages finishes, it is possible for
2203 * private pages to be COWed, which remain after
2204 * truncate_inode_pages finishes, hence the second
2205 * unmap_mapping_range call must be made for correctness.
2207 unmap_mapping_range(mapping
, offset
+ PAGE_SIZE
- 1, 0, 1);
2208 truncate_inode_pages(mapping
, offset
);
2209 unmap_mapping_range(mapping
, offset
+ PAGE_SIZE
- 1, 0, 1);
2212 if (inode
->i_op
&& inode
->i_op
->truncate
)
2213 inode
->i_op
->truncate(inode
);
2217 send_sig(SIGXFSZ
, current
, 0);
2221 EXPORT_SYMBOL(vmtruncate
);
2223 int vmtruncate_range(struct inode
*inode
, loff_t offset
, loff_t end
)
2225 struct address_space
*mapping
= inode
->i_mapping
;
2228 * If the underlying filesystem is not going to provide
2229 * a way to truncate a range of blocks (punch a hole) -
2230 * we should return failure right now.
2232 if (!inode
->i_op
|| !inode
->i_op
->truncate_range
)
2235 mutex_lock(&inode
->i_mutex
);
2236 down_write(&inode
->i_alloc_sem
);
2237 unmap_mapping_range(mapping
, offset
, (end
- offset
), 1);
2238 truncate_inode_pages_range(mapping
, offset
, end
);
2239 unmap_mapping_range(mapping
, offset
, (end
- offset
), 1);
2240 inode
->i_op
->truncate_range(inode
, offset
, end
);
2241 up_write(&inode
->i_alloc_sem
);
2242 mutex_unlock(&inode
->i_mutex
);
2248 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2249 * but allow concurrent faults), and pte mapped but not yet locked.
2250 * We return with mmap_sem still held, but pte unmapped and unlocked.
2252 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2253 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2254 int write_access
, pte_t orig_pte
)
2262 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2265 entry
= pte_to_swp_entry(orig_pte
);
2266 if (is_migration_entry(entry
)) {
2267 migration_entry_wait(mm
, pmd
, address
);
2270 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2271 page
= lookup_swap_cache(entry
);
2273 grab_swap_token(); /* Contend for token _before_ read-in */
2274 page
= swapin_readahead(entry
,
2275 GFP_HIGHUSER_MOVABLE
, vma
, address
);
2278 * Back out if somebody else faulted in this pte
2279 * while we released the pte lock.
2281 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2282 if (likely(pte_same(*page_table
, orig_pte
)))
2284 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2288 /* Had to read the page from swap area: Major fault */
2289 ret
= VM_FAULT_MAJOR
;
2290 count_vm_event(PGMAJFAULT
);
2293 if (mem_cgroup_charge(page
, mm
, GFP_KERNEL
)) {
2294 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2299 mark_page_accessed(page
);
2301 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2304 * Back out if somebody else already faulted in this pte.
2306 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2307 if (unlikely(!pte_same(*page_table
, orig_pte
)))
2310 if (unlikely(!PageUptodate(page
))) {
2311 ret
= VM_FAULT_SIGBUS
;
2315 /* The page isn't present yet, go ahead with the fault. */
2317 inc_mm_counter(mm
, anon_rss
);
2318 pte
= mk_pte(page
, vma
->vm_page_prot
);
2319 if (write_access
&& can_share_swap_page(page
)) {
2320 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
2324 flush_icache_page(vma
, page
);
2325 set_pte_at(mm
, address
, page_table
, pte
);
2326 page_add_anon_rmap(page
, vma
, address
);
2330 remove_exclusive_swap_page(page
);
2334 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
2335 if (ret
& VM_FAULT_ERROR
)
2336 ret
&= VM_FAULT_ERROR
;
2340 /* No need to invalidate - it was non-present before */
2341 update_mmu_cache(vma
, address
, pte
);
2343 pte_unmap_unlock(page_table
, ptl
);
2347 mem_cgroup_uncharge_page(page
);
2348 pte_unmap_unlock(page_table
, ptl
);
2350 page_cache_release(page
);
2355 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2356 * but allow concurrent faults), and pte mapped but not yet locked.
2357 * We return with mmap_sem still held, but pte unmapped and unlocked.
2359 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2360 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2367 /* Allocate our own private page. */
2368 pte_unmap(page_table
);
2370 if (unlikely(anon_vma_prepare(vma
)))
2372 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2375 __SetPageUptodate(page
);
2377 if (mem_cgroup_charge(page
, mm
, GFP_KERNEL
))
2380 entry
= mk_pte(page
, vma
->vm_page_prot
);
2381 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2383 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2384 if (!pte_none(*page_table
))
2386 inc_mm_counter(mm
, anon_rss
);
2387 lru_cache_add_active(page
);
2388 page_add_new_anon_rmap(page
, vma
, address
);
2389 set_pte_at(mm
, address
, page_table
, entry
);
2391 /* No need to invalidate - it was non-present before */
2392 update_mmu_cache(vma
, address
, entry
);
2394 pte_unmap_unlock(page_table
, ptl
);
2397 mem_cgroup_uncharge_page(page
);
2398 page_cache_release(page
);
2401 page_cache_release(page
);
2403 return VM_FAULT_OOM
;
2407 * __do_fault() tries to create a new page mapping. It aggressively
2408 * tries to share with existing pages, but makes a separate copy if
2409 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2410 * the next page fault.
2412 * As this is called only for pages that do not currently exist, we
2413 * do not need to flush old virtual caches or the TLB.
2415 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2416 * but allow concurrent faults), and pte neither mapped nor locked.
2417 * We return with mmap_sem still held, but pte unmapped and unlocked.
2419 static int __do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2420 unsigned long address
, pmd_t
*pmd
,
2421 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
2428 struct page
*dirty_page
= NULL
;
2429 struct vm_fault vmf
;
2431 int page_mkwrite
= 0;
2433 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
2438 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
2439 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2443 * For consistency in subsequent calls, make the faulted page always
2446 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
2447 lock_page(vmf
.page
);
2449 VM_BUG_ON(!PageLocked(vmf
.page
));
2452 * Should we do an early C-O-W break?
2455 if (flags
& FAULT_FLAG_WRITE
) {
2456 if (!(vma
->vm_flags
& VM_SHARED
)) {
2458 if (unlikely(anon_vma_prepare(vma
))) {
2462 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
,
2468 copy_user_highpage(page
, vmf
.page
, address
, vma
);
2469 __SetPageUptodate(page
);
2472 * If the page will be shareable, see if the backing
2473 * address space wants to know that the page is about
2474 * to become writable
2476 if (vma
->vm_ops
->page_mkwrite
) {
2478 if (vma
->vm_ops
->page_mkwrite(vma
, page
) < 0) {
2479 ret
= VM_FAULT_SIGBUS
;
2480 anon
= 1; /* no anon but release vmf.page */
2485 * XXX: this is not quite right (racy vs
2486 * invalidate) to unlock and relock the page
2487 * like this, however a better fix requires
2488 * reworking page_mkwrite locking API, which
2489 * is better done later.
2491 if (!page
->mapping
) {
2493 anon
= 1; /* no anon but release vmf.page */
2502 if (mem_cgroup_charge(page
, mm
, GFP_KERNEL
)) {
2507 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2510 * This silly early PAGE_DIRTY setting removes a race
2511 * due to the bad i386 page protection. But it's valid
2512 * for other architectures too.
2514 * Note that if write_access is true, we either now have
2515 * an exclusive copy of the page, or this is a shared mapping,
2516 * so we can make it writable and dirty to avoid having to
2517 * handle that later.
2519 /* Only go through if we didn't race with anybody else... */
2520 if (likely(pte_same(*page_table
, orig_pte
))) {
2521 flush_icache_page(vma
, page
);
2522 entry
= mk_pte(page
, vma
->vm_page_prot
);
2523 if (flags
& FAULT_FLAG_WRITE
)
2524 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2525 set_pte_at(mm
, address
, page_table
, entry
);
2527 inc_mm_counter(mm
, anon_rss
);
2528 lru_cache_add_active(page
);
2529 page_add_new_anon_rmap(page
, vma
, address
);
2531 inc_mm_counter(mm
, file_rss
);
2532 page_add_file_rmap(page
);
2533 if (flags
& FAULT_FLAG_WRITE
) {
2535 get_page(dirty_page
);
2539 /* no need to invalidate: a not-present page won't be cached */
2540 update_mmu_cache(vma
, address
, entry
);
2542 mem_cgroup_uncharge_page(page
);
2544 page_cache_release(page
);
2546 anon
= 1; /* no anon but release faulted_page */
2549 pte_unmap_unlock(page_table
, ptl
);
2552 unlock_page(vmf
.page
);
2555 page_cache_release(vmf
.page
);
2556 else if (dirty_page
) {
2558 file_update_time(vma
->vm_file
);
2560 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2561 put_page(dirty_page
);
2567 static int do_linear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2568 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2569 int write_access
, pte_t orig_pte
)
2571 pgoff_t pgoff
= (((address
& PAGE_MASK
)
2572 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
2573 unsigned int flags
= (write_access
? FAULT_FLAG_WRITE
: 0);
2575 pte_unmap(page_table
);
2576 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
2580 * Fault of a previously existing named mapping. Repopulate the pte
2581 * from the encoded file_pte if possible. This enables swappable
2584 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2585 * but allow concurrent faults), and pte mapped but not yet locked.
2586 * We return with mmap_sem still held, but pte unmapped and unlocked.
2588 static int do_nonlinear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2589 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2590 int write_access
, pte_t orig_pte
)
2592 unsigned int flags
= FAULT_FLAG_NONLINEAR
|
2593 (write_access
? FAULT_FLAG_WRITE
: 0);
2596 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2599 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
) ||
2600 !(vma
->vm_flags
& VM_CAN_NONLINEAR
))) {
2602 * Page table corrupted: show pte and kill process.
2604 print_bad_pte(vma
, orig_pte
, address
);
2605 return VM_FAULT_OOM
;
2608 pgoff
= pte_to_pgoff(orig_pte
);
2609 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
2613 * These routines also need to handle stuff like marking pages dirty
2614 * and/or accessed for architectures that don't do it in hardware (most
2615 * RISC architectures). The early dirtying is also good on the i386.
2617 * There is also a hook called "update_mmu_cache()" that architectures
2618 * with external mmu caches can use to update those (ie the Sparc or
2619 * PowerPC hashed page tables that act as extended TLBs).
2621 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2622 * but allow concurrent faults), and pte mapped but not yet locked.
2623 * We return with mmap_sem still held, but pte unmapped and unlocked.
2625 static inline int handle_pte_fault(struct mm_struct
*mm
,
2626 struct vm_area_struct
*vma
, unsigned long address
,
2627 pte_t
*pte
, pmd_t
*pmd
, int write_access
)
2633 if (!pte_present(entry
)) {
2634 if (pte_none(entry
)) {
2636 if (likely(vma
->vm_ops
->fault
))
2637 return do_linear_fault(mm
, vma
, address
,
2638 pte
, pmd
, write_access
, entry
);
2640 return do_anonymous_page(mm
, vma
, address
,
2641 pte
, pmd
, write_access
);
2643 if (pte_file(entry
))
2644 return do_nonlinear_fault(mm
, vma
, address
,
2645 pte
, pmd
, write_access
, entry
);
2646 return do_swap_page(mm
, vma
, address
,
2647 pte
, pmd
, write_access
, entry
);
2650 ptl
= pte_lockptr(mm
, pmd
);
2652 if (unlikely(!pte_same(*pte
, entry
)))
2655 if (!pte_write(entry
))
2656 return do_wp_page(mm
, vma
, address
,
2657 pte
, pmd
, ptl
, entry
);
2658 entry
= pte_mkdirty(entry
);
2660 entry
= pte_mkyoung(entry
);
2661 if (ptep_set_access_flags(vma
, address
, pte
, entry
, write_access
)) {
2662 update_mmu_cache(vma
, address
, entry
);
2665 * This is needed only for protection faults but the arch code
2666 * is not yet telling us if this is a protection fault or not.
2667 * This still avoids useless tlb flushes for .text page faults
2671 flush_tlb_page(vma
, address
);
2674 pte_unmap_unlock(pte
, ptl
);
2679 * By the time we get here, we already hold the mm semaphore
2681 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2682 unsigned long address
, int write_access
)
2689 __set_current_state(TASK_RUNNING
);
2691 count_vm_event(PGFAULT
);
2693 if (unlikely(is_vm_hugetlb_page(vma
)))
2694 return hugetlb_fault(mm
, vma
, address
, write_access
);
2696 pgd
= pgd_offset(mm
, address
);
2697 pud
= pud_alloc(mm
, pgd
, address
);
2699 return VM_FAULT_OOM
;
2700 pmd
= pmd_alloc(mm
, pud
, address
);
2702 return VM_FAULT_OOM
;
2703 pte
= pte_alloc_map(mm
, pmd
, address
);
2705 return VM_FAULT_OOM
;
2707 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, write_access
);
2710 #ifndef __PAGETABLE_PUD_FOLDED
2712 * Allocate page upper directory.
2713 * We've already handled the fast-path in-line.
2715 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
2717 pud_t
*new = pud_alloc_one(mm
, address
);
2721 smp_wmb(); /* See comment in __pte_alloc */
2723 spin_lock(&mm
->page_table_lock
);
2724 if (pgd_present(*pgd
)) /* Another has populated it */
2727 pgd_populate(mm
, pgd
, new);
2728 spin_unlock(&mm
->page_table_lock
);
2731 #endif /* __PAGETABLE_PUD_FOLDED */
2733 #ifndef __PAGETABLE_PMD_FOLDED
2735 * Allocate page middle directory.
2736 * We've already handled the fast-path in-line.
2738 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
2740 pmd_t
*new = pmd_alloc_one(mm
, address
);
2744 smp_wmb(); /* See comment in __pte_alloc */
2746 spin_lock(&mm
->page_table_lock
);
2747 #ifndef __ARCH_HAS_4LEVEL_HACK
2748 if (pud_present(*pud
)) /* Another has populated it */
2751 pud_populate(mm
, pud
, new);
2753 if (pgd_present(*pud
)) /* Another has populated it */
2756 pgd_populate(mm
, pud
, new);
2757 #endif /* __ARCH_HAS_4LEVEL_HACK */
2758 spin_unlock(&mm
->page_table_lock
);
2761 #endif /* __PAGETABLE_PMD_FOLDED */
2763 int make_pages_present(unsigned long addr
, unsigned long end
)
2765 int ret
, len
, write
;
2766 struct vm_area_struct
* vma
;
2768 vma
= find_vma(current
->mm
, addr
);
2771 write
= (vma
->vm_flags
& VM_WRITE
) != 0;
2772 BUG_ON(addr
>= end
);
2773 BUG_ON(end
> vma
->vm_end
);
2774 len
= DIV_ROUND_UP(end
, PAGE_SIZE
) - addr
/PAGE_SIZE
;
2775 ret
= get_user_pages(current
, current
->mm
, addr
,
2776 len
, write
, 0, NULL
, NULL
);
2779 return ret
== len
? 0 : -1;
2782 #if !defined(__HAVE_ARCH_GATE_AREA)
2784 #if defined(AT_SYSINFO_EHDR)
2785 static struct vm_area_struct gate_vma
;
2787 static int __init
gate_vma_init(void)
2789 gate_vma
.vm_mm
= NULL
;
2790 gate_vma
.vm_start
= FIXADDR_USER_START
;
2791 gate_vma
.vm_end
= FIXADDR_USER_END
;
2792 gate_vma
.vm_flags
= VM_READ
| VM_MAYREAD
| VM_EXEC
| VM_MAYEXEC
;
2793 gate_vma
.vm_page_prot
= __P101
;
2795 * Make sure the vDSO gets into every core dump.
2796 * Dumping its contents makes post-mortem fully interpretable later
2797 * without matching up the same kernel and hardware config to see
2798 * what PC values meant.
2800 gate_vma
.vm_flags
|= VM_ALWAYSDUMP
;
2803 __initcall(gate_vma_init
);
2806 struct vm_area_struct
*get_gate_vma(struct task_struct
*tsk
)
2808 #ifdef AT_SYSINFO_EHDR
2815 int in_gate_area_no_task(unsigned long addr
)
2817 #ifdef AT_SYSINFO_EHDR
2818 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
2824 #endif /* __HAVE_ARCH_GATE_AREA */
2826 #ifdef CONFIG_HAVE_IOREMAP_PROT
2827 static resource_size_t
follow_phys(struct vm_area_struct
*vma
,
2828 unsigned long address
, unsigned int flags
,
2829 unsigned long *prot
)
2836 resource_size_t phys_addr
= 0;
2837 struct mm_struct
*mm
= vma
->vm_mm
;
2839 VM_BUG_ON(!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)));
2841 pgd
= pgd_offset(mm
, address
);
2842 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
2845 pud
= pud_offset(pgd
, address
);
2846 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
2849 pmd
= pmd_offset(pud
, address
);
2850 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
2853 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
2857 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2862 if (!pte_present(pte
))
2864 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
2866 phys_addr
= pte_pfn(pte
);
2867 phys_addr
<<= PAGE_SHIFT
; /* Shift here to avoid overflow on PAE */
2869 *prot
= pgprot_val(pte_pgprot(pte
));
2872 pte_unmap_unlock(ptep
, ptl
);
2879 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
2880 void *buf
, int len
, int write
)
2882 resource_size_t phys_addr
;
2883 unsigned long prot
= 0;
2885 int offset
= addr
& (PAGE_SIZE
-1);
2887 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
2890 phys_addr
= follow_phys(vma
, addr
, write
, &prot
);
2895 maddr
= ioremap_prot(phys_addr
, PAGE_SIZE
, prot
);
2897 memcpy_toio(maddr
+ offset
, buf
, len
);
2899 memcpy_fromio(buf
, maddr
+ offset
, len
);
2907 * Access another process' address space.
2908 * Source/target buffer must be kernel space,
2909 * Do not walk the page table directly, use get_user_pages
2911 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
, void *buf
, int len
, int write
)
2913 struct mm_struct
*mm
;
2914 struct vm_area_struct
*vma
;
2915 void *old_buf
= buf
;
2917 mm
= get_task_mm(tsk
);
2921 down_read(&mm
->mmap_sem
);
2922 /* ignore errors, just check how much was successfully transferred */
2924 int bytes
, ret
, offset
;
2926 struct page
*page
= NULL
;
2928 ret
= get_user_pages(tsk
, mm
, addr
, 1,
2929 write
, 1, &page
, &vma
);
2932 * Check if this is a VM_IO | VM_PFNMAP VMA, which
2933 * we can access using slightly different code.
2935 #ifdef CONFIG_HAVE_IOREMAP_PROT
2936 vma
= find_vma(mm
, addr
);
2939 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
2940 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
2948 offset
= addr
& (PAGE_SIZE
-1);
2949 if (bytes
> PAGE_SIZE
-offset
)
2950 bytes
= PAGE_SIZE
-offset
;
2954 copy_to_user_page(vma
, page
, addr
,
2955 maddr
+ offset
, buf
, bytes
);
2956 set_page_dirty_lock(page
);
2958 copy_from_user_page(vma
, page
, addr
,
2959 buf
, maddr
+ offset
, bytes
);
2962 page_cache_release(page
);
2968 up_read(&mm
->mmap_sem
);
2971 return buf
- old_buf
;
2975 * Print the name of a VMA.
2977 void print_vma_addr(char *prefix
, unsigned long ip
)
2979 struct mm_struct
*mm
= current
->mm
;
2980 struct vm_area_struct
*vma
;
2983 * Do not print if we are in atomic
2984 * contexts (in exception stacks, etc.):
2986 if (preempt_count())
2989 down_read(&mm
->mmap_sem
);
2990 vma
= find_vma(mm
, ip
);
2991 if (vma
&& vma
->vm_file
) {
2992 struct file
*f
= vma
->vm_file
;
2993 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
2997 p
= d_path(&f
->f_path
, buf
, PAGE_SIZE
);
3000 s
= strrchr(p
, '/');
3003 printk("%s%s[%lx+%lx]", prefix
, p
,
3005 vma
->vm_end
- vma
->vm_start
);
3006 free_page((unsigned long)buf
);
3009 up_read(¤t
->mm
->mmap_sem
);