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 #ifndef track_pfn_vma_new
104 * Interface that can be used by architecture code to keep track of
105 * memory type of pfn mappings (remap_pfn_range, vm_insert_pfn)
107 * track_pfn_vma_new is called when a _new_ pfn mapping is being established
108 * for physical range indicated by pfn and size.
110 int track_pfn_vma_new(struct vm_area_struct
*vma
, pgprot_t prot
,
111 unsigned long pfn
, unsigned long size
)
117 #ifndef track_pfn_vma_copy
119 * Interface that can be used by architecture code to keep track of
120 * memory type of pfn mappings (remap_pfn_range, vm_insert_pfn)
122 * track_pfn_vma_copy is called when vma that is covering the pfnmap gets
123 * copied through copy_page_range().
125 int track_pfn_vma_copy(struct vm_area_struct
*vma
)
131 #ifndef untrack_pfn_vma
133 * Interface that can be used by architecture code to keep track of
134 * memory type of pfn mappings (remap_pfn_range, vm_insert_pfn)
136 * untrack_pfn_vma is called while unmapping a pfnmap for a region.
137 * untrack can be called for a specific region indicated by pfn and size or
138 * can be for the entire vma (in which case size can be zero).
140 void untrack_pfn_vma(struct vm_area_struct
*vma
, unsigned long pfn
,
146 static int __init
disable_randmaps(char *s
)
148 randomize_va_space
= 0;
151 __setup("norandmaps", disable_randmaps
);
155 * If a p?d_bad entry is found while walking page tables, report
156 * the error, before resetting entry to p?d_none. Usually (but
157 * very seldom) called out from the p?d_none_or_clear_bad macros.
160 void pgd_clear_bad(pgd_t
*pgd
)
166 void pud_clear_bad(pud_t
*pud
)
172 void pmd_clear_bad(pmd_t
*pmd
)
179 * Note: this doesn't free the actual pages themselves. That
180 * has been handled earlier when unmapping all the memory regions.
182 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
)
184 pgtable_t token
= pmd_pgtable(*pmd
);
186 pte_free_tlb(tlb
, token
);
190 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
191 unsigned long addr
, unsigned long end
,
192 unsigned long floor
, unsigned long ceiling
)
199 pmd
= pmd_offset(pud
, addr
);
201 next
= pmd_addr_end(addr
, end
);
202 if (pmd_none_or_clear_bad(pmd
))
204 free_pte_range(tlb
, pmd
);
205 } while (pmd
++, addr
= next
, addr
!= end
);
215 if (end
- 1 > ceiling
- 1)
218 pmd
= pmd_offset(pud
, start
);
220 pmd_free_tlb(tlb
, pmd
);
223 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
224 unsigned long addr
, unsigned long end
,
225 unsigned long floor
, unsigned long ceiling
)
232 pud
= pud_offset(pgd
, addr
);
234 next
= pud_addr_end(addr
, end
);
235 if (pud_none_or_clear_bad(pud
))
237 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
238 } while (pud
++, addr
= next
, addr
!= end
);
244 ceiling
&= PGDIR_MASK
;
248 if (end
- 1 > ceiling
- 1)
251 pud
= pud_offset(pgd
, start
);
253 pud_free_tlb(tlb
, pud
);
257 * This function frees user-level page tables of a process.
259 * Must be called with pagetable lock held.
261 void free_pgd_range(struct mmu_gather
*tlb
,
262 unsigned long addr
, unsigned long end
,
263 unsigned long floor
, unsigned long ceiling
)
270 * The next few lines have given us lots of grief...
272 * Why are we testing PMD* at this top level? Because often
273 * there will be no work to do at all, and we'd prefer not to
274 * go all the way down to the bottom just to discover that.
276 * Why all these "- 1"s? Because 0 represents both the bottom
277 * of the address space and the top of it (using -1 for the
278 * top wouldn't help much: the masks would do the wrong thing).
279 * The rule is that addr 0 and floor 0 refer to the bottom of
280 * the address space, but end 0 and ceiling 0 refer to the top
281 * Comparisons need to use "end - 1" and "ceiling - 1" (though
282 * that end 0 case should be mythical).
284 * Wherever addr is brought up or ceiling brought down, we must
285 * be careful to reject "the opposite 0" before it confuses the
286 * subsequent tests. But what about where end is brought down
287 * by PMD_SIZE below? no, end can't go down to 0 there.
289 * Whereas we round start (addr) and ceiling down, by different
290 * masks at different levels, in order to test whether a table
291 * now has no other vmas using it, so can be freed, we don't
292 * bother to round floor or end up - the tests don't need that.
306 if (end
- 1 > ceiling
- 1)
312 pgd
= pgd_offset(tlb
->mm
, addr
);
314 next
= pgd_addr_end(addr
, end
);
315 if (pgd_none_or_clear_bad(pgd
))
317 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
318 } while (pgd
++, addr
= next
, addr
!= end
);
321 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
322 unsigned long floor
, unsigned long ceiling
)
325 struct vm_area_struct
*next
= vma
->vm_next
;
326 unsigned long addr
= vma
->vm_start
;
329 * Hide vma from rmap and vmtruncate before freeing pgtables
331 anon_vma_unlink(vma
);
332 unlink_file_vma(vma
);
334 if (is_vm_hugetlb_page(vma
)) {
335 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
336 floor
, next
? next
->vm_start
: ceiling
);
339 * Optimization: gather nearby vmas into one call down
341 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
342 && !is_vm_hugetlb_page(next
)) {
345 anon_vma_unlink(vma
);
346 unlink_file_vma(vma
);
348 free_pgd_range(tlb
, addr
, vma
->vm_end
,
349 floor
, next
? next
->vm_start
: ceiling
);
355 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
, unsigned long address
)
357 pgtable_t
new = pte_alloc_one(mm
, address
);
362 * Ensure all pte setup (eg. pte page lock and page clearing) are
363 * visible before the pte is made visible to other CPUs by being
364 * put into page tables.
366 * The other side of the story is the pointer chasing in the page
367 * table walking code (when walking the page table without locking;
368 * ie. most of the time). Fortunately, these data accesses consist
369 * of a chain of data-dependent loads, meaning most CPUs (alpha
370 * being the notable exception) will already guarantee loads are
371 * seen in-order. See the alpha page table accessors for the
372 * smp_read_barrier_depends() barriers in page table walking code.
374 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
376 spin_lock(&mm
->page_table_lock
);
377 if (!pmd_present(*pmd
)) { /* Has another populated it ? */
379 pmd_populate(mm
, pmd
, new);
382 spin_unlock(&mm
->page_table_lock
);
388 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
390 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
394 smp_wmb(); /* See comment in __pte_alloc */
396 spin_lock(&init_mm
.page_table_lock
);
397 if (!pmd_present(*pmd
)) { /* Has another populated it ? */
398 pmd_populate_kernel(&init_mm
, pmd
, new);
401 spin_unlock(&init_mm
.page_table_lock
);
403 pte_free_kernel(&init_mm
, new);
407 static inline void add_mm_rss(struct mm_struct
*mm
, int file_rss
, int anon_rss
)
410 add_mm_counter(mm
, file_rss
, file_rss
);
412 add_mm_counter(mm
, anon_rss
, anon_rss
);
416 * This function is called to print an error when a bad pte
417 * is found. For example, we might have a PFN-mapped pte in
418 * a region that doesn't allow it.
420 * The calling function must still handle the error.
422 static void print_bad_pte(struct vm_area_struct
*vma
, pte_t pte
,
425 printk(KERN_ERR
"Bad pte = %08llx, process = %s, "
426 "vm_flags = %lx, vaddr = %lx\n",
427 (long long)pte_val(pte
),
428 (vma
->vm_mm
== current
->mm
? current
->comm
: "???"),
429 vma
->vm_flags
, vaddr
);
433 static inline int is_cow_mapping(unsigned int flags
)
435 return (flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
439 * vm_normal_page -- This function gets the "struct page" associated with a pte.
441 * "Special" mappings do not wish to be associated with a "struct page" (either
442 * it doesn't exist, or it exists but they don't want to touch it). In this
443 * case, NULL is returned here. "Normal" mappings do have a struct page.
445 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
446 * pte bit, in which case this function is trivial. Secondly, an architecture
447 * may not have a spare pte bit, which requires a more complicated scheme,
450 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
451 * special mapping (even if there are underlying and valid "struct pages").
452 * COWed pages of a VM_PFNMAP are always normal.
454 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
455 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
456 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
457 * mapping will always honor the rule
459 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
461 * And for normal mappings this is false.
463 * This restricts such mappings to be a linear translation from virtual address
464 * to pfn. To get around this restriction, we allow arbitrary mappings so long
465 * as the vma is not a COW mapping; in that case, we know that all ptes are
466 * special (because none can have been COWed).
469 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
471 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
472 * page" backing, however the difference is that _all_ pages with a struct
473 * page (that is, those where pfn_valid is true) are refcounted and considered
474 * normal pages by the VM. The disadvantage is that pages are refcounted
475 * (which can be slower and simply not an option for some PFNMAP users). The
476 * advantage is that we don't have to follow the strict linearity rule of
477 * PFNMAP mappings in order to support COWable mappings.
480 #ifdef __HAVE_ARCH_PTE_SPECIAL
481 # define HAVE_PTE_SPECIAL 1
483 # define HAVE_PTE_SPECIAL 0
485 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
490 if (HAVE_PTE_SPECIAL
) {
491 if (likely(!pte_special(pte
))) {
492 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
493 return pte_page(pte
);
495 VM_BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
)));
499 /* !HAVE_PTE_SPECIAL case follows: */
503 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
504 if (vma
->vm_flags
& VM_MIXEDMAP
) {
510 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
511 if (pfn
== vma
->vm_pgoff
+ off
)
513 if (!is_cow_mapping(vma
->vm_flags
))
518 VM_BUG_ON(!pfn_valid(pfn
));
521 * NOTE! We still have PageReserved() pages in the page tables.
523 * eg. VDSO mappings can cause them to exist.
526 return pfn_to_page(pfn
);
530 * copy one vm_area from one task to the other. Assumes the page tables
531 * already present in the new task to be cleared in the whole range
532 * covered by this vma.
536 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
537 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
538 unsigned long addr
, int *rss
)
540 unsigned long vm_flags
= vma
->vm_flags
;
541 pte_t pte
= *src_pte
;
544 /* pte contains position in swap or file, so copy. */
545 if (unlikely(!pte_present(pte
))) {
546 if (!pte_file(pte
)) {
547 swp_entry_t entry
= pte_to_swp_entry(pte
);
549 swap_duplicate(entry
);
550 /* make sure dst_mm is on swapoff's mmlist. */
551 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
552 spin_lock(&mmlist_lock
);
553 if (list_empty(&dst_mm
->mmlist
))
554 list_add(&dst_mm
->mmlist
,
556 spin_unlock(&mmlist_lock
);
558 if (is_write_migration_entry(entry
) &&
559 is_cow_mapping(vm_flags
)) {
561 * COW mappings require pages in both parent
562 * and child to be set to read.
564 make_migration_entry_read(&entry
);
565 pte
= swp_entry_to_pte(entry
);
566 set_pte_at(src_mm
, addr
, src_pte
, pte
);
573 * If it's a COW mapping, write protect it both
574 * in the parent and the child
576 if (is_cow_mapping(vm_flags
)) {
577 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
578 pte
= pte_wrprotect(pte
);
582 * If it's a shared mapping, mark it clean in
585 if (vm_flags
& VM_SHARED
)
586 pte
= pte_mkclean(pte
);
587 pte
= pte_mkold(pte
);
589 page
= vm_normal_page(vma
, addr
, pte
);
592 page_dup_rmap(page
, vma
, addr
);
593 rss
[!!PageAnon(page
)]++;
597 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
600 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
601 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
602 unsigned long addr
, unsigned long end
)
604 pte_t
*src_pte
, *dst_pte
;
605 spinlock_t
*src_ptl
, *dst_ptl
;
611 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
614 src_pte
= pte_offset_map_nested(src_pmd
, addr
);
615 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
616 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
617 arch_enter_lazy_mmu_mode();
621 * We are holding two locks at this point - either of them
622 * could generate latencies in another task on another CPU.
624 if (progress
>= 32) {
626 if (need_resched() ||
627 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
630 if (pte_none(*src_pte
)) {
634 copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
, vma
, addr
, rss
);
636 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
638 arch_leave_lazy_mmu_mode();
639 spin_unlock(src_ptl
);
640 pte_unmap_nested(src_pte
- 1);
641 add_mm_rss(dst_mm
, rss
[0], rss
[1]);
642 pte_unmap_unlock(dst_pte
- 1, dst_ptl
);
649 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
650 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
651 unsigned long addr
, unsigned long end
)
653 pmd_t
*src_pmd
, *dst_pmd
;
656 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
659 src_pmd
= pmd_offset(src_pud
, addr
);
661 next
= pmd_addr_end(addr
, end
);
662 if (pmd_none_or_clear_bad(src_pmd
))
664 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
667 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
671 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
672 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
673 unsigned long addr
, unsigned long end
)
675 pud_t
*src_pud
, *dst_pud
;
678 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
681 src_pud
= pud_offset(src_pgd
, addr
);
683 next
= pud_addr_end(addr
, end
);
684 if (pud_none_or_clear_bad(src_pud
))
686 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
689 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
693 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
694 struct vm_area_struct
*vma
)
696 pgd_t
*src_pgd
, *dst_pgd
;
698 unsigned long addr
= vma
->vm_start
;
699 unsigned long end
= vma
->vm_end
;
703 * Don't copy ptes where a page fault will fill them correctly.
704 * Fork becomes much lighter when there are big shared or private
705 * readonly mappings. The tradeoff is that copy_page_range is more
706 * efficient than faulting.
708 if (!(vma
->vm_flags
& (VM_HUGETLB
|VM_NONLINEAR
|VM_PFNMAP
|VM_INSERTPAGE
))) {
713 if (is_vm_hugetlb_page(vma
))
714 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
716 if (is_pfn_mapping(vma
)) {
718 * We do not free on error cases below as remove_vma
719 * gets called on error from higher level routine
721 ret
= track_pfn_vma_copy(vma
);
727 * We need to invalidate the secondary MMU mappings only when
728 * there could be a permission downgrade on the ptes of the
729 * parent mm. And a permission downgrade will only happen if
730 * is_cow_mapping() returns true.
732 if (is_cow_mapping(vma
->vm_flags
))
733 mmu_notifier_invalidate_range_start(src_mm
, addr
, end
);
736 dst_pgd
= pgd_offset(dst_mm
, addr
);
737 src_pgd
= pgd_offset(src_mm
, addr
);
739 next
= pgd_addr_end(addr
, end
);
740 if (pgd_none_or_clear_bad(src_pgd
))
742 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
747 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
749 if (is_cow_mapping(vma
->vm_flags
))
750 mmu_notifier_invalidate_range_end(src_mm
,
755 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
756 struct vm_area_struct
*vma
, pmd_t
*pmd
,
757 unsigned long addr
, unsigned long end
,
758 long *zap_work
, struct zap_details
*details
)
760 struct mm_struct
*mm
= tlb
->mm
;
766 pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
767 arch_enter_lazy_mmu_mode();
770 if (pte_none(ptent
)) {
775 (*zap_work
) -= PAGE_SIZE
;
777 if (pte_present(ptent
)) {
780 page
= vm_normal_page(vma
, addr
, ptent
);
781 if (unlikely(details
) && page
) {
783 * unmap_shared_mapping_pages() wants to
784 * invalidate cache without truncating:
785 * unmap shared but keep private pages.
787 if (details
->check_mapping
&&
788 details
->check_mapping
!= page
->mapping
)
791 * Each page->index must be checked when
792 * invalidating or truncating nonlinear.
794 if (details
->nonlinear_vma
&&
795 (page
->index
< details
->first_index
||
796 page
->index
> details
->last_index
))
799 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
801 tlb_remove_tlb_entry(tlb
, pte
, addr
);
804 if (unlikely(details
) && details
->nonlinear_vma
805 && linear_page_index(details
->nonlinear_vma
,
806 addr
) != page
->index
)
807 set_pte_at(mm
, addr
, pte
,
808 pgoff_to_pte(page
->index
));
812 if (pte_dirty(ptent
))
813 set_page_dirty(page
);
814 if (pte_young(ptent
))
815 SetPageReferenced(page
);
818 page_remove_rmap(page
, vma
);
819 tlb_remove_page(tlb
, page
);
823 * If details->check_mapping, we leave swap entries;
824 * if details->nonlinear_vma, we leave file entries.
826 if (unlikely(details
))
828 if (!pte_file(ptent
))
829 free_swap_and_cache(pte_to_swp_entry(ptent
));
830 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
831 } while (pte
++, addr
+= PAGE_SIZE
, (addr
!= end
&& *zap_work
> 0));
833 add_mm_rss(mm
, file_rss
, anon_rss
);
834 arch_leave_lazy_mmu_mode();
835 pte_unmap_unlock(pte
- 1, ptl
);
840 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
841 struct vm_area_struct
*vma
, pud_t
*pud
,
842 unsigned long addr
, unsigned long end
,
843 long *zap_work
, struct zap_details
*details
)
848 pmd
= pmd_offset(pud
, addr
);
850 next
= pmd_addr_end(addr
, end
);
851 if (pmd_none_or_clear_bad(pmd
)) {
855 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
,
857 } while (pmd
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
862 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
863 struct vm_area_struct
*vma
, pgd_t
*pgd
,
864 unsigned long addr
, unsigned long end
,
865 long *zap_work
, struct zap_details
*details
)
870 pud
= pud_offset(pgd
, addr
);
872 next
= pud_addr_end(addr
, end
);
873 if (pud_none_or_clear_bad(pud
)) {
877 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
,
879 } while (pud
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
884 static unsigned long unmap_page_range(struct mmu_gather
*tlb
,
885 struct vm_area_struct
*vma
,
886 unsigned long addr
, unsigned long end
,
887 long *zap_work
, struct zap_details
*details
)
892 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
896 tlb_start_vma(tlb
, vma
);
897 pgd
= pgd_offset(vma
->vm_mm
, addr
);
899 next
= pgd_addr_end(addr
, end
);
900 if (pgd_none_or_clear_bad(pgd
)) {
904 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
,
906 } while (pgd
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
907 tlb_end_vma(tlb
, vma
);
912 #ifdef CONFIG_PREEMPT
913 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
915 /* No preempt: go for improved straight-line efficiency */
916 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
920 * unmap_vmas - unmap a range of memory covered by a list of vma's
921 * @tlbp: address of the caller's struct mmu_gather
922 * @vma: the starting vma
923 * @start_addr: virtual address at which to start unmapping
924 * @end_addr: virtual address at which to end unmapping
925 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
926 * @details: details of nonlinear truncation or shared cache invalidation
928 * Returns the end address of the unmapping (restart addr if interrupted).
930 * Unmap all pages in the vma list.
932 * We aim to not hold locks for too long (for scheduling latency reasons).
933 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
934 * return the ending mmu_gather to the caller.
936 * Only addresses between `start' and `end' will be unmapped.
938 * The VMA list must be sorted in ascending virtual address order.
940 * unmap_vmas() assumes that the caller will flush the whole unmapped address
941 * range after unmap_vmas() returns. So the only responsibility here is to
942 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
943 * drops the lock and schedules.
945 unsigned long unmap_vmas(struct mmu_gather
**tlbp
,
946 struct vm_area_struct
*vma
, unsigned long start_addr
,
947 unsigned long end_addr
, unsigned long *nr_accounted
,
948 struct zap_details
*details
)
950 long zap_work
= ZAP_BLOCK_SIZE
;
951 unsigned long tlb_start
= 0; /* For tlb_finish_mmu */
952 int tlb_start_valid
= 0;
953 unsigned long start
= start_addr
;
954 spinlock_t
*i_mmap_lock
= details
? details
->i_mmap_lock
: NULL
;
955 int fullmm
= (*tlbp
)->fullmm
;
956 struct mm_struct
*mm
= vma
->vm_mm
;
958 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
959 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
) {
962 start
= max(vma
->vm_start
, start_addr
);
963 if (start
>= vma
->vm_end
)
965 end
= min(vma
->vm_end
, end_addr
);
966 if (end
<= vma
->vm_start
)
969 if (vma
->vm_flags
& VM_ACCOUNT
)
970 *nr_accounted
+= (end
- start
) >> PAGE_SHIFT
;
972 if (is_pfn_mapping(vma
))
973 untrack_pfn_vma(vma
, 0, 0);
975 while (start
!= end
) {
976 if (!tlb_start_valid
) {
981 if (unlikely(is_vm_hugetlb_page(vma
))) {
983 * It is undesirable to test vma->vm_file as it
984 * should be non-null for valid hugetlb area.
985 * However, vm_file will be NULL in the error
986 * cleanup path of do_mmap_pgoff. When
987 * hugetlbfs ->mmap method fails,
988 * do_mmap_pgoff() nullifies vma->vm_file
989 * before calling this function to clean up.
990 * Since no pte has actually been setup, it is
991 * safe to do nothing in this case.
994 unmap_hugepage_range(vma
, start
, end
, NULL
);
995 zap_work
-= (end
- start
) /
996 pages_per_huge_page(hstate_vma(vma
));
1001 start
= unmap_page_range(*tlbp
, vma
,
1002 start
, end
, &zap_work
, details
);
1005 BUG_ON(start
!= end
);
1009 tlb_finish_mmu(*tlbp
, tlb_start
, start
);
1011 if (need_resched() ||
1012 (i_mmap_lock
&& spin_needbreak(i_mmap_lock
))) {
1020 *tlbp
= tlb_gather_mmu(vma
->vm_mm
, fullmm
);
1021 tlb_start_valid
= 0;
1022 zap_work
= ZAP_BLOCK_SIZE
;
1026 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1027 return start
; /* which is now the end (or restart) address */
1031 * zap_page_range - remove user pages in a given range
1032 * @vma: vm_area_struct holding the applicable pages
1033 * @address: starting address of pages to zap
1034 * @size: number of bytes to zap
1035 * @details: details of nonlinear truncation or shared cache invalidation
1037 unsigned long zap_page_range(struct vm_area_struct
*vma
, unsigned long address
,
1038 unsigned long size
, struct zap_details
*details
)
1040 struct mm_struct
*mm
= vma
->vm_mm
;
1041 struct mmu_gather
*tlb
;
1042 unsigned long end
= address
+ size
;
1043 unsigned long nr_accounted
= 0;
1046 tlb
= tlb_gather_mmu(mm
, 0);
1047 update_hiwater_rss(mm
);
1048 end
= unmap_vmas(&tlb
, vma
, address
, end
, &nr_accounted
, details
);
1050 tlb_finish_mmu(tlb
, address
, end
);
1055 * zap_vma_ptes - remove ptes mapping the vma
1056 * @vma: vm_area_struct holding ptes to be zapped
1057 * @address: starting address of pages to zap
1058 * @size: number of bytes to zap
1060 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1062 * The entire address range must be fully contained within the vma.
1064 * Returns 0 if successful.
1066 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1069 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1070 !(vma
->vm_flags
& VM_PFNMAP
))
1072 zap_page_range(vma
, address
, size
, NULL
);
1075 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1078 * Do a quick page-table lookup for a single page.
1080 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
1089 struct mm_struct
*mm
= vma
->vm_mm
;
1091 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
1092 if (!IS_ERR(page
)) {
1093 BUG_ON(flags
& FOLL_GET
);
1098 pgd
= pgd_offset(mm
, address
);
1099 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
1102 pud
= pud_offset(pgd
, address
);
1105 if (pud_huge(*pud
)) {
1106 BUG_ON(flags
& FOLL_GET
);
1107 page
= follow_huge_pud(mm
, address
, pud
, flags
& FOLL_WRITE
);
1110 if (unlikely(pud_bad(*pud
)))
1113 pmd
= pmd_offset(pud
, address
);
1116 if (pmd_huge(*pmd
)) {
1117 BUG_ON(flags
& FOLL_GET
);
1118 page
= follow_huge_pmd(mm
, address
, pmd
, flags
& FOLL_WRITE
);
1121 if (unlikely(pmd_bad(*pmd
)))
1124 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1127 if (!pte_present(pte
))
1129 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
1131 page
= vm_normal_page(vma
, address
, pte
);
1132 if (unlikely(!page
))
1135 if (flags
& FOLL_GET
)
1137 if (flags
& FOLL_TOUCH
) {
1138 if ((flags
& FOLL_WRITE
) &&
1139 !pte_dirty(pte
) && !PageDirty(page
))
1140 set_page_dirty(page
);
1141 mark_page_accessed(page
);
1144 pte_unmap_unlock(ptep
, ptl
);
1149 pte_unmap_unlock(ptep
, ptl
);
1150 return ERR_PTR(-EFAULT
);
1153 pte_unmap_unlock(ptep
, ptl
);
1156 /* Fall through to ZERO_PAGE handling */
1159 * When core dumping an enormous anonymous area that nobody
1160 * has touched so far, we don't want to allocate page tables.
1162 if (flags
& FOLL_ANON
) {
1163 page
= ZERO_PAGE(0);
1164 if (flags
& FOLL_GET
)
1166 BUG_ON(flags
& FOLL_WRITE
);
1171 /* Can we do the FOLL_ANON optimization? */
1172 static inline int use_zero_page(struct vm_area_struct
*vma
)
1175 * We don't want to optimize FOLL_ANON for make_pages_present()
1176 * when it tries to page in a VM_LOCKED region. As to VM_SHARED,
1177 * we want to get the page from the page tables to make sure
1178 * that we serialize and update with any other user of that
1181 if (vma
->vm_flags
& (VM_LOCKED
| VM_SHARED
))
1184 * And if we have a fault routine, it's not an anonymous region.
1186 return !vma
->vm_ops
|| !vma
->vm_ops
->fault
;
1191 int __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1192 unsigned long start
, int len
, int flags
,
1193 struct page
**pages
, struct vm_area_struct
**vmas
)
1196 unsigned int vm_flags
= 0;
1197 int write
= !!(flags
& GUP_FLAGS_WRITE
);
1198 int force
= !!(flags
& GUP_FLAGS_FORCE
);
1199 int ignore
= !!(flags
& GUP_FLAGS_IGNORE_VMA_PERMISSIONS
);
1204 * Require read or write permissions.
1205 * If 'force' is set, we only require the "MAY" flags.
1207 vm_flags
= write
? (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1208 vm_flags
&= force
? (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1212 struct vm_area_struct
*vma
;
1213 unsigned int foll_flags
;
1215 vma
= find_extend_vma(mm
, start
);
1216 if (!vma
&& in_gate_area(tsk
, start
)) {
1217 unsigned long pg
= start
& PAGE_MASK
;
1218 struct vm_area_struct
*gate_vma
= get_gate_vma(tsk
);
1224 /* user gate pages are read-only */
1225 if (!ignore
&& write
)
1226 return i
? : -EFAULT
;
1228 pgd
= pgd_offset_k(pg
);
1230 pgd
= pgd_offset_gate(mm
, pg
);
1231 BUG_ON(pgd_none(*pgd
));
1232 pud
= pud_offset(pgd
, pg
);
1233 BUG_ON(pud_none(*pud
));
1234 pmd
= pmd_offset(pud
, pg
);
1236 return i
? : -EFAULT
;
1237 pte
= pte_offset_map(pmd
, pg
);
1238 if (pte_none(*pte
)) {
1240 return i
? : -EFAULT
;
1243 struct page
*page
= vm_normal_page(gate_vma
, start
, *pte
);
1258 (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1259 (!ignore
&& !(vm_flags
& vma
->vm_flags
)))
1260 return i
? : -EFAULT
;
1262 if (is_vm_hugetlb_page(vma
)) {
1263 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1264 &start
, &len
, i
, write
);
1268 foll_flags
= FOLL_TOUCH
;
1270 foll_flags
|= FOLL_GET
;
1271 if (!write
&& use_zero_page(vma
))
1272 foll_flags
|= FOLL_ANON
;
1278 * If tsk is ooming, cut off its access to large memory
1279 * allocations. It has a pending SIGKILL, but it can't
1280 * be processed until returning to user space.
1282 if (unlikely(test_tsk_thread_flag(tsk
, TIF_MEMDIE
)))
1283 return i
? i
: -ENOMEM
;
1286 foll_flags
|= FOLL_WRITE
;
1289 while (!(page
= follow_page(vma
, start
, foll_flags
))) {
1291 ret
= handle_mm_fault(mm
, vma
, start
,
1292 foll_flags
& FOLL_WRITE
);
1293 if (ret
& VM_FAULT_ERROR
) {
1294 if (ret
& VM_FAULT_OOM
)
1295 return i
? i
: -ENOMEM
;
1296 else if (ret
& VM_FAULT_SIGBUS
)
1297 return i
? i
: -EFAULT
;
1300 if (ret
& VM_FAULT_MAJOR
)
1306 * The VM_FAULT_WRITE bit tells us that
1307 * do_wp_page has broken COW when necessary,
1308 * even if maybe_mkwrite decided not to set
1309 * pte_write. We can thus safely do subsequent
1310 * page lookups as if they were reads.
1312 if (ret
& VM_FAULT_WRITE
)
1313 foll_flags
&= ~FOLL_WRITE
;
1318 return i
? i
: PTR_ERR(page
);
1322 flush_anon_page(vma
, page
, start
);
1323 flush_dcache_page(page
);
1330 } while (len
&& start
< vma
->vm_end
);
1335 int get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1336 unsigned long start
, int len
, int write
, int force
,
1337 struct page
**pages
, struct vm_area_struct
**vmas
)
1342 flags
|= GUP_FLAGS_WRITE
;
1344 flags
|= GUP_FLAGS_FORCE
;
1346 return __get_user_pages(tsk
, mm
,
1351 EXPORT_SYMBOL(get_user_pages
);
1353 pte_t
*get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1356 pgd_t
* pgd
= pgd_offset(mm
, addr
);
1357 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
1359 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
1361 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1367 * This is the old fallback for page remapping.
1369 * For historical reasons, it only allows reserved pages. Only
1370 * old drivers should use this, and they needed to mark their
1371 * pages reserved for the old functions anyway.
1373 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1374 struct page
*page
, pgprot_t prot
)
1376 struct mm_struct
*mm
= vma
->vm_mm
;
1385 flush_dcache_page(page
);
1386 pte
= get_locked_pte(mm
, addr
, &ptl
);
1390 if (!pte_none(*pte
))
1393 /* Ok, finally just insert the thing.. */
1395 inc_mm_counter(mm
, file_rss
);
1396 page_add_file_rmap(page
);
1397 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1400 pte_unmap_unlock(pte
, ptl
);
1403 pte_unmap_unlock(pte
, ptl
);
1409 * vm_insert_page - insert single page into user vma
1410 * @vma: user vma to map to
1411 * @addr: target user address of this page
1412 * @page: source kernel page
1414 * This allows drivers to insert individual pages they've allocated
1417 * The page has to be a nice clean _individual_ kernel allocation.
1418 * If you allocate a compound page, you need to have marked it as
1419 * such (__GFP_COMP), or manually just split the page up yourself
1420 * (see split_page()).
1422 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1423 * took an arbitrary page protection parameter. This doesn't allow
1424 * that. Your vma protection will have to be set up correctly, which
1425 * means that if you want a shared writable mapping, you'd better
1426 * ask for a shared writable mapping!
1428 * The page does not need to be reserved.
1430 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1433 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1435 if (!page_count(page
))
1437 vma
->vm_flags
|= VM_INSERTPAGE
;
1438 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1440 EXPORT_SYMBOL(vm_insert_page
);
1442 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1443 unsigned long pfn
, pgprot_t prot
)
1445 struct mm_struct
*mm
= vma
->vm_mm
;
1451 pte
= get_locked_pte(mm
, addr
, &ptl
);
1455 if (!pte_none(*pte
))
1458 /* Ok, finally just insert the thing.. */
1459 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
1460 set_pte_at(mm
, addr
, pte
, entry
);
1461 update_mmu_cache(vma
, addr
, entry
); /* XXX: why not for insert_page? */
1465 pte_unmap_unlock(pte
, ptl
);
1471 * vm_insert_pfn - insert single pfn into user vma
1472 * @vma: user vma to map to
1473 * @addr: target user address of this page
1474 * @pfn: source kernel pfn
1476 * Similar to vm_inert_page, this allows drivers to insert individual pages
1477 * they've allocated into a user vma. Same comments apply.
1479 * This function should only be called from a vm_ops->fault handler, and
1480 * in that case the handler should return NULL.
1482 * vma cannot be a COW mapping.
1484 * As this is called only for pages that do not currently exist, we
1485 * do not need to flush old virtual caches or the TLB.
1487 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1492 * Technically, architectures with pte_special can avoid all these
1493 * restrictions (same for remap_pfn_range). However we would like
1494 * consistency in testing and feature parity among all, so we should
1495 * try to keep these invariants in place for everybody.
1497 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1498 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1499 (VM_PFNMAP
|VM_MIXEDMAP
));
1500 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1501 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1503 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1505 if (track_pfn_vma_new(vma
, vma
->vm_page_prot
, pfn
, PAGE_SIZE
))
1508 ret
= insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1511 untrack_pfn_vma(vma
, pfn
, PAGE_SIZE
);
1515 EXPORT_SYMBOL(vm_insert_pfn
);
1517 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1520 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
1522 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1526 * If we don't have pte special, then we have to use the pfn_valid()
1527 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1528 * refcount the page if pfn_valid is true (hence insert_page rather
1531 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
1534 page
= pfn_to_page(pfn
);
1535 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1537 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1539 EXPORT_SYMBOL(vm_insert_mixed
);
1542 * maps a range of physical memory into the requested pages. the old
1543 * mappings are removed. any references to nonexistent pages results
1544 * in null mappings (currently treated as "copy-on-access")
1546 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1547 unsigned long addr
, unsigned long end
,
1548 unsigned long pfn
, pgprot_t prot
)
1553 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1556 arch_enter_lazy_mmu_mode();
1558 BUG_ON(!pte_none(*pte
));
1559 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
1561 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1562 arch_leave_lazy_mmu_mode();
1563 pte_unmap_unlock(pte
- 1, ptl
);
1567 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1568 unsigned long addr
, unsigned long end
,
1569 unsigned long pfn
, pgprot_t prot
)
1574 pfn
-= addr
>> PAGE_SHIFT
;
1575 pmd
= pmd_alloc(mm
, pud
, addr
);
1579 next
= pmd_addr_end(addr
, end
);
1580 if (remap_pte_range(mm
, pmd
, addr
, next
,
1581 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1583 } while (pmd
++, addr
= next
, addr
!= end
);
1587 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1588 unsigned long addr
, unsigned long end
,
1589 unsigned long pfn
, pgprot_t prot
)
1594 pfn
-= addr
>> PAGE_SHIFT
;
1595 pud
= pud_alloc(mm
, pgd
, addr
);
1599 next
= pud_addr_end(addr
, end
);
1600 if (remap_pmd_range(mm
, pud
, addr
, next
,
1601 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1603 } while (pud
++, addr
= next
, addr
!= end
);
1608 * remap_pfn_range - remap kernel memory to userspace
1609 * @vma: user vma to map to
1610 * @addr: target user address to start at
1611 * @pfn: physical address of kernel memory
1612 * @size: size of map area
1613 * @prot: page protection flags for this mapping
1615 * Note: this is only safe if the mm semaphore is held when called.
1617 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1618 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1622 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1623 struct mm_struct
*mm
= vma
->vm_mm
;
1627 * Physically remapped pages are special. Tell the
1628 * rest of the world about it:
1629 * VM_IO tells people not to look at these pages
1630 * (accesses can have side effects).
1631 * VM_RESERVED is specified all over the place, because
1632 * in 2.4 it kept swapout's vma scan off this vma; but
1633 * in 2.6 the LRU scan won't even find its pages, so this
1634 * flag means no more than count its pages in reserved_vm,
1635 * and omit it from core dump, even when VM_IO turned off.
1636 * VM_PFNMAP tells the core MM that the base pages are just
1637 * raw PFN mappings, and do not have a "struct page" associated
1640 * There's a horrible special case to handle copy-on-write
1641 * behaviour that some programs depend on. We mark the "original"
1642 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1644 if (addr
== vma
->vm_start
&& end
== vma
->vm_end
)
1645 vma
->vm_pgoff
= pfn
;
1646 else if (is_cow_mapping(vma
->vm_flags
))
1649 vma
->vm_flags
|= VM_IO
| VM_RESERVED
| VM_PFNMAP
;
1651 err
= track_pfn_vma_new(vma
, prot
, pfn
, PAGE_ALIGN(size
));
1655 BUG_ON(addr
>= end
);
1656 pfn
-= addr
>> PAGE_SHIFT
;
1657 pgd
= pgd_offset(mm
, addr
);
1658 flush_cache_range(vma
, addr
, end
);
1660 next
= pgd_addr_end(addr
, end
);
1661 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1662 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1665 } while (pgd
++, addr
= next
, addr
!= end
);
1668 untrack_pfn_vma(vma
, pfn
, PAGE_ALIGN(size
));
1672 EXPORT_SYMBOL(remap_pfn_range
);
1674 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1675 unsigned long addr
, unsigned long end
,
1676 pte_fn_t fn
, void *data
)
1681 spinlock_t
*uninitialized_var(ptl
);
1683 pte
= (mm
== &init_mm
) ?
1684 pte_alloc_kernel(pmd
, addr
) :
1685 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1689 BUG_ON(pmd_huge(*pmd
));
1691 token
= pmd_pgtable(*pmd
);
1694 err
= fn(pte
, token
, addr
, data
);
1697 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1700 pte_unmap_unlock(pte
-1, ptl
);
1704 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1705 unsigned long addr
, unsigned long end
,
1706 pte_fn_t fn
, void *data
)
1712 BUG_ON(pud_huge(*pud
));
1714 pmd
= pmd_alloc(mm
, pud
, addr
);
1718 next
= pmd_addr_end(addr
, end
);
1719 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
1722 } while (pmd
++, addr
= next
, addr
!= end
);
1726 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1727 unsigned long addr
, unsigned long end
,
1728 pte_fn_t fn
, void *data
)
1734 pud
= pud_alloc(mm
, pgd
, addr
);
1738 next
= pud_addr_end(addr
, end
);
1739 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
1742 } while (pud
++, addr
= next
, addr
!= end
);
1747 * Scan a region of virtual memory, filling in page tables as necessary
1748 * and calling a provided function on each leaf page table.
1750 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
1751 unsigned long size
, pte_fn_t fn
, void *data
)
1755 unsigned long start
= addr
, end
= addr
+ size
;
1758 BUG_ON(addr
>= end
);
1759 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1760 pgd
= pgd_offset(mm
, addr
);
1762 next
= pgd_addr_end(addr
, end
);
1763 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
1766 } while (pgd
++, addr
= next
, addr
!= end
);
1767 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1770 EXPORT_SYMBOL_GPL(apply_to_page_range
);
1773 * handle_pte_fault chooses page fault handler according to an entry
1774 * which was read non-atomically. Before making any commitment, on
1775 * those architectures or configurations (e.g. i386 with PAE) which
1776 * might give a mix of unmatched parts, do_swap_page and do_file_page
1777 * must check under lock before unmapping the pte and proceeding
1778 * (but do_wp_page is only called after already making such a check;
1779 * and do_anonymous_page and do_no_page can safely check later on).
1781 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
1782 pte_t
*page_table
, pte_t orig_pte
)
1785 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1786 if (sizeof(pte_t
) > sizeof(unsigned long)) {
1787 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
1789 same
= pte_same(*page_table
, orig_pte
);
1793 pte_unmap(page_table
);
1798 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1799 * servicing faults for write access. In the normal case, do always want
1800 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1801 * that do not have writing enabled, when used by access_process_vm.
1803 static inline pte_t
maybe_mkwrite(pte_t pte
, struct vm_area_struct
*vma
)
1805 if (likely(vma
->vm_flags
& VM_WRITE
))
1806 pte
= pte_mkwrite(pte
);
1810 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
1813 * If the source page was a PFN mapping, we don't have
1814 * a "struct page" for it. We do a best-effort copy by
1815 * just copying from the original user address. If that
1816 * fails, we just zero-fill it. Live with it.
1818 if (unlikely(!src
)) {
1819 void *kaddr
= kmap_atomic(dst
, KM_USER0
);
1820 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
1823 * This really shouldn't fail, because the page is there
1824 * in the page tables. But it might just be unreadable,
1825 * in which case we just give up and fill the result with
1828 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
1829 memset(kaddr
, 0, PAGE_SIZE
);
1830 kunmap_atomic(kaddr
, KM_USER0
);
1831 flush_dcache_page(dst
);
1833 copy_user_highpage(dst
, src
, va
, vma
);
1837 * This routine handles present pages, when users try to write
1838 * to a shared page. It is done by copying the page to a new address
1839 * and decrementing the shared-page counter for the old page.
1841 * Note that this routine assumes that the protection checks have been
1842 * done by the caller (the low-level page fault routine in most cases).
1843 * Thus we can safely just mark it writable once we've done any necessary
1846 * We also mark the page dirty at this point even though the page will
1847 * change only once the write actually happens. This avoids a few races,
1848 * and potentially makes it more efficient.
1850 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1851 * but allow concurrent faults), with pte both mapped and locked.
1852 * We return with mmap_sem still held, but pte unmapped and unlocked.
1854 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1855 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1856 spinlock_t
*ptl
, pte_t orig_pte
)
1858 struct page
*old_page
, *new_page
;
1860 int reuse
= 0, ret
= 0;
1861 int page_mkwrite
= 0;
1862 struct page
*dirty_page
= NULL
;
1864 old_page
= vm_normal_page(vma
, address
, orig_pte
);
1867 * VM_MIXEDMAP !pfn_valid() case
1869 * We should not cow pages in a shared writeable mapping.
1870 * Just mark the pages writable as we can't do any dirty
1871 * accounting on raw pfn maps.
1873 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
1874 (VM_WRITE
|VM_SHARED
))
1880 * Take out anonymous pages first, anonymous shared vmas are
1881 * not dirty accountable.
1883 if (PageAnon(old_page
)) {
1884 if (trylock_page(old_page
)) {
1885 reuse
= can_share_swap_page(old_page
);
1886 unlock_page(old_page
);
1888 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
1889 (VM_WRITE
|VM_SHARED
))) {
1891 * Only catch write-faults on shared writable pages,
1892 * read-only shared pages can get COWed by
1893 * get_user_pages(.write=1, .force=1).
1895 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
1897 * Notify the address space that the page is about to
1898 * become writable so that it can prohibit this or wait
1899 * for the page to get into an appropriate state.
1901 * We do this without the lock held, so that it can
1902 * sleep if it needs to.
1904 page_cache_get(old_page
);
1905 pte_unmap_unlock(page_table
, ptl
);
1907 if (vma
->vm_ops
->page_mkwrite(vma
, old_page
) < 0)
1908 goto unwritable_page
;
1911 * Since we dropped the lock we need to revalidate
1912 * the PTE as someone else may have changed it. If
1913 * they did, we just return, as we can count on the
1914 * MMU to tell us if they didn't also make it writable.
1916 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
1918 page_cache_release(old_page
);
1919 if (!pte_same(*page_table
, orig_pte
))
1924 dirty_page
= old_page
;
1925 get_page(dirty_page
);
1931 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
1932 entry
= pte_mkyoung(orig_pte
);
1933 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1934 if (ptep_set_access_flags(vma
, address
, page_table
, entry
,1))
1935 update_mmu_cache(vma
, address
, entry
);
1936 ret
|= VM_FAULT_WRITE
;
1941 * Ok, we need to copy. Oh, well..
1943 page_cache_get(old_page
);
1945 pte_unmap_unlock(page_table
, ptl
);
1947 if (unlikely(anon_vma_prepare(vma
)))
1949 VM_BUG_ON(old_page
== ZERO_PAGE(0));
1950 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
1954 * Don't let another task, with possibly unlocked vma,
1955 * keep the mlocked page.
1957 if (vma
->vm_flags
& VM_LOCKED
) {
1958 lock_page(old_page
); /* for LRU manipulation */
1959 clear_page_mlock(old_page
);
1960 unlock_page(old_page
);
1962 cow_user_page(new_page
, old_page
, address
, vma
);
1963 __SetPageUptodate(new_page
);
1965 if (mem_cgroup_charge(new_page
, mm
, GFP_KERNEL
))
1969 * Re-check the pte - we dropped the lock
1971 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1972 if (likely(pte_same(*page_table
, orig_pte
))) {
1974 if (!PageAnon(old_page
)) {
1975 dec_mm_counter(mm
, file_rss
);
1976 inc_mm_counter(mm
, anon_rss
);
1979 inc_mm_counter(mm
, anon_rss
);
1980 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
1981 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
1982 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1984 * Clear the pte entry and flush it first, before updating the
1985 * pte with the new entry. This will avoid a race condition
1986 * seen in the presence of one thread doing SMC and another
1989 ptep_clear_flush_notify(vma
, address
, page_table
);
1990 SetPageSwapBacked(new_page
);
1991 lru_cache_add_active_or_unevictable(new_page
, vma
);
1992 page_add_new_anon_rmap(new_page
, vma
, address
);
1994 //TODO: is this safe? do_anonymous_page() does it this way.
1995 set_pte_at(mm
, address
, page_table
, entry
);
1996 update_mmu_cache(vma
, address
, entry
);
1999 * Only after switching the pte to the new page may
2000 * we remove the mapcount here. Otherwise another
2001 * process may come and find the rmap count decremented
2002 * before the pte is switched to the new page, and
2003 * "reuse" the old page writing into it while our pte
2004 * here still points into it and can be read by other
2007 * The critical issue is to order this
2008 * page_remove_rmap with the ptp_clear_flush above.
2009 * Those stores are ordered by (if nothing else,)
2010 * the barrier present in the atomic_add_negative
2011 * in page_remove_rmap.
2013 * Then the TLB flush in ptep_clear_flush ensures that
2014 * no process can access the old page before the
2015 * decremented mapcount is visible. And the old page
2016 * cannot be reused until after the decremented
2017 * mapcount is visible. So transitively, TLBs to
2018 * old page will be flushed before it can be reused.
2020 page_remove_rmap(old_page
, vma
);
2023 /* Free the old page.. */
2024 new_page
= old_page
;
2025 ret
|= VM_FAULT_WRITE
;
2027 mem_cgroup_uncharge_page(new_page
);
2030 page_cache_release(new_page
);
2032 page_cache_release(old_page
);
2034 pte_unmap_unlock(page_table
, ptl
);
2037 file_update_time(vma
->vm_file
);
2040 * Yes, Virginia, this is actually required to prevent a race
2041 * with clear_page_dirty_for_io() from clearing the page dirty
2042 * bit after it clear all dirty ptes, but before a racing
2043 * do_wp_page installs a dirty pte.
2045 * do_no_page is protected similarly.
2047 wait_on_page_locked(dirty_page
);
2048 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2049 put_page(dirty_page
);
2053 page_cache_release(new_page
);
2056 page_cache_release(old_page
);
2057 return VM_FAULT_OOM
;
2060 page_cache_release(old_page
);
2061 return VM_FAULT_SIGBUS
;
2065 * Helper functions for unmap_mapping_range().
2067 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2069 * We have to restart searching the prio_tree whenever we drop the lock,
2070 * since the iterator is only valid while the lock is held, and anyway
2071 * a later vma might be split and reinserted earlier while lock dropped.
2073 * The list of nonlinear vmas could be handled more efficiently, using
2074 * a placeholder, but handle it in the same way until a need is shown.
2075 * It is important to search the prio_tree before nonlinear list: a vma
2076 * may become nonlinear and be shifted from prio_tree to nonlinear list
2077 * while the lock is dropped; but never shifted from list to prio_tree.
2079 * In order to make forward progress despite restarting the search,
2080 * vm_truncate_count is used to mark a vma as now dealt with, so we can
2081 * quickly skip it next time around. Since the prio_tree search only
2082 * shows us those vmas affected by unmapping the range in question, we
2083 * can't efficiently keep all vmas in step with mapping->truncate_count:
2084 * so instead reset them all whenever it wraps back to 0 (then go to 1).
2085 * mapping->truncate_count and vma->vm_truncate_count are protected by
2088 * In order to make forward progress despite repeatedly restarting some
2089 * large vma, note the restart_addr from unmap_vmas when it breaks out:
2090 * and restart from that address when we reach that vma again. It might
2091 * have been split or merged, shrunk or extended, but never shifted: so
2092 * restart_addr remains valid so long as it remains in the vma's range.
2093 * unmap_mapping_range forces truncate_count to leap over page-aligned
2094 * values so we can save vma's restart_addr in its truncate_count field.
2096 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2098 static void reset_vma_truncate_counts(struct address_space
*mapping
)
2100 struct vm_area_struct
*vma
;
2101 struct prio_tree_iter iter
;
2103 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, 0, ULONG_MAX
)
2104 vma
->vm_truncate_count
= 0;
2105 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
2106 vma
->vm_truncate_count
= 0;
2109 static int unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2110 unsigned long start_addr
, unsigned long end_addr
,
2111 struct zap_details
*details
)
2113 unsigned long restart_addr
;
2117 * files that support invalidating or truncating portions of the
2118 * file from under mmaped areas must have their ->fault function
2119 * return a locked page (and set VM_FAULT_LOCKED in the return).
2120 * This provides synchronisation against concurrent unmapping here.
2124 restart_addr
= vma
->vm_truncate_count
;
2125 if (is_restart_addr(restart_addr
) && start_addr
< restart_addr
) {
2126 start_addr
= restart_addr
;
2127 if (start_addr
>= end_addr
) {
2128 /* Top of vma has been split off since last time */
2129 vma
->vm_truncate_count
= details
->truncate_count
;
2134 restart_addr
= zap_page_range(vma
, start_addr
,
2135 end_addr
- start_addr
, details
);
2136 need_break
= need_resched() || spin_needbreak(details
->i_mmap_lock
);
2138 if (restart_addr
>= end_addr
) {
2139 /* We have now completed this vma: mark it so */
2140 vma
->vm_truncate_count
= details
->truncate_count
;
2144 /* Note restart_addr in vma's truncate_count field */
2145 vma
->vm_truncate_count
= restart_addr
;
2150 spin_unlock(details
->i_mmap_lock
);
2152 spin_lock(details
->i_mmap_lock
);
2156 static inline void unmap_mapping_range_tree(struct prio_tree_root
*root
,
2157 struct zap_details
*details
)
2159 struct vm_area_struct
*vma
;
2160 struct prio_tree_iter iter
;
2161 pgoff_t vba
, vea
, zba
, zea
;
2164 vma_prio_tree_foreach(vma
, &iter
, root
,
2165 details
->first_index
, details
->last_index
) {
2166 /* Skip quickly over those we have already dealt with */
2167 if (vma
->vm_truncate_count
== details
->truncate_count
)
2170 vba
= vma
->vm_pgoff
;
2171 vea
= vba
+ ((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) - 1;
2172 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2173 zba
= details
->first_index
;
2176 zea
= details
->last_index
;
2180 if (unmap_mapping_range_vma(vma
,
2181 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2182 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2188 static inline void unmap_mapping_range_list(struct list_head
*head
,
2189 struct zap_details
*details
)
2191 struct vm_area_struct
*vma
;
2194 * In nonlinear VMAs there is no correspondence between virtual address
2195 * offset and file offset. So we must perform an exhaustive search
2196 * across *all* the pages in each nonlinear VMA, not just the pages
2197 * whose virtual address lies outside the file truncation point.
2200 list_for_each_entry(vma
, head
, shared
.vm_set
.list
) {
2201 /* Skip quickly over those we have already dealt with */
2202 if (vma
->vm_truncate_count
== details
->truncate_count
)
2204 details
->nonlinear_vma
= vma
;
2205 if (unmap_mapping_range_vma(vma
, vma
->vm_start
,
2206 vma
->vm_end
, details
) < 0)
2212 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2213 * @mapping: the address space containing mmaps to be unmapped.
2214 * @holebegin: byte in first page to unmap, relative to the start of
2215 * the underlying file. This will be rounded down to a PAGE_SIZE
2216 * boundary. Note that this is different from vmtruncate(), which
2217 * must keep the partial page. In contrast, we must get rid of
2219 * @holelen: size of prospective hole in bytes. This will be rounded
2220 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2222 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2223 * but 0 when invalidating pagecache, don't throw away private data.
2225 void unmap_mapping_range(struct address_space
*mapping
,
2226 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2228 struct zap_details details
;
2229 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2230 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2232 /* Check for overflow. */
2233 if (sizeof(holelen
) > sizeof(hlen
)) {
2235 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2236 if (holeend
& ~(long long)ULONG_MAX
)
2237 hlen
= ULONG_MAX
- hba
+ 1;
2240 details
.check_mapping
= even_cows
? NULL
: mapping
;
2241 details
.nonlinear_vma
= NULL
;
2242 details
.first_index
= hba
;
2243 details
.last_index
= hba
+ hlen
- 1;
2244 if (details
.last_index
< details
.first_index
)
2245 details
.last_index
= ULONG_MAX
;
2246 details
.i_mmap_lock
= &mapping
->i_mmap_lock
;
2248 spin_lock(&mapping
->i_mmap_lock
);
2250 /* Protect against endless unmapping loops */
2251 mapping
->truncate_count
++;
2252 if (unlikely(is_restart_addr(mapping
->truncate_count
))) {
2253 if (mapping
->truncate_count
== 0)
2254 reset_vma_truncate_counts(mapping
);
2255 mapping
->truncate_count
++;
2257 details
.truncate_count
= mapping
->truncate_count
;
2259 if (unlikely(!prio_tree_empty(&mapping
->i_mmap
)))
2260 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2261 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
2262 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
2263 spin_unlock(&mapping
->i_mmap_lock
);
2265 EXPORT_SYMBOL(unmap_mapping_range
);
2268 * vmtruncate - unmap mappings "freed" by truncate() syscall
2269 * @inode: inode of the file used
2270 * @offset: file offset to start truncating
2272 * NOTE! We have to be ready to update the memory sharing
2273 * between the file and the memory map for a potential last
2274 * incomplete page. Ugly, but necessary.
2276 int vmtruncate(struct inode
* inode
, loff_t offset
)
2278 if (inode
->i_size
< offset
) {
2279 unsigned long limit
;
2281 limit
= current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
;
2282 if (limit
!= RLIM_INFINITY
&& offset
> limit
)
2284 if (offset
> inode
->i_sb
->s_maxbytes
)
2286 i_size_write(inode
, offset
);
2288 struct address_space
*mapping
= inode
->i_mapping
;
2291 * truncation of in-use swapfiles is disallowed - it would
2292 * cause subsequent swapout to scribble on the now-freed
2295 if (IS_SWAPFILE(inode
))
2297 i_size_write(inode
, offset
);
2300 * unmap_mapping_range is called twice, first simply for
2301 * efficiency so that truncate_inode_pages does fewer
2302 * single-page unmaps. However after this first call, and
2303 * before truncate_inode_pages finishes, it is possible for
2304 * private pages to be COWed, which remain after
2305 * truncate_inode_pages finishes, hence the second
2306 * unmap_mapping_range call must be made for correctness.
2308 unmap_mapping_range(mapping
, offset
+ PAGE_SIZE
- 1, 0, 1);
2309 truncate_inode_pages(mapping
, offset
);
2310 unmap_mapping_range(mapping
, offset
+ PAGE_SIZE
- 1, 0, 1);
2313 if (inode
->i_op
&& inode
->i_op
->truncate
)
2314 inode
->i_op
->truncate(inode
);
2318 send_sig(SIGXFSZ
, current
, 0);
2322 EXPORT_SYMBOL(vmtruncate
);
2324 int vmtruncate_range(struct inode
*inode
, loff_t offset
, loff_t end
)
2326 struct address_space
*mapping
= inode
->i_mapping
;
2329 * If the underlying filesystem is not going to provide
2330 * a way to truncate a range of blocks (punch a hole) -
2331 * we should return failure right now.
2333 if (!inode
->i_op
|| !inode
->i_op
->truncate_range
)
2336 mutex_lock(&inode
->i_mutex
);
2337 down_write(&inode
->i_alloc_sem
);
2338 unmap_mapping_range(mapping
, offset
, (end
- offset
), 1);
2339 truncate_inode_pages_range(mapping
, offset
, end
);
2340 unmap_mapping_range(mapping
, offset
, (end
- offset
), 1);
2341 inode
->i_op
->truncate_range(inode
, offset
, end
);
2342 up_write(&inode
->i_alloc_sem
);
2343 mutex_unlock(&inode
->i_mutex
);
2349 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2350 * but allow concurrent faults), and pte mapped but not yet locked.
2351 * We return with mmap_sem still held, but pte unmapped and unlocked.
2353 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2354 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2355 int write_access
, pte_t orig_pte
)
2363 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2366 entry
= pte_to_swp_entry(orig_pte
);
2367 if (is_migration_entry(entry
)) {
2368 migration_entry_wait(mm
, pmd
, address
);
2371 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2372 page
= lookup_swap_cache(entry
);
2374 grab_swap_token(); /* Contend for token _before_ read-in */
2375 page
= swapin_readahead(entry
,
2376 GFP_HIGHUSER_MOVABLE
, vma
, address
);
2379 * Back out if somebody else faulted in this pte
2380 * while we released the pte lock.
2382 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2383 if (likely(pte_same(*page_table
, orig_pte
)))
2385 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2389 /* Had to read the page from swap area: Major fault */
2390 ret
= VM_FAULT_MAJOR
;
2391 count_vm_event(PGMAJFAULT
);
2394 mark_page_accessed(page
);
2397 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2399 if (mem_cgroup_charge(page
, mm
, GFP_KERNEL
)) {
2406 * Back out if somebody else already faulted in this pte.
2408 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2409 if (unlikely(!pte_same(*page_table
, orig_pte
)))
2412 if (unlikely(!PageUptodate(page
))) {
2413 ret
= VM_FAULT_SIGBUS
;
2417 /* The page isn't present yet, go ahead with the fault. */
2419 inc_mm_counter(mm
, anon_rss
);
2420 pte
= mk_pte(page
, vma
->vm_page_prot
);
2421 if (write_access
&& can_share_swap_page(page
)) {
2422 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
2426 flush_icache_page(vma
, page
);
2427 set_pte_at(mm
, address
, page_table
, pte
);
2428 page_add_anon_rmap(page
, vma
, address
);
2431 if (vm_swap_full() || (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
2432 remove_exclusive_swap_page(page
);
2436 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
2437 if (ret
& VM_FAULT_ERROR
)
2438 ret
&= VM_FAULT_ERROR
;
2442 /* No need to invalidate - it was non-present before */
2443 update_mmu_cache(vma
, address
, pte
);
2445 pte_unmap_unlock(page_table
, ptl
);
2449 mem_cgroup_uncharge_page(page
);
2450 pte_unmap_unlock(page_table
, ptl
);
2452 page_cache_release(page
);
2457 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2458 * but allow concurrent faults), and pte mapped but not yet locked.
2459 * We return with mmap_sem still held, but pte unmapped and unlocked.
2461 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2462 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2469 /* Allocate our own private page. */
2470 pte_unmap(page_table
);
2472 if (unlikely(anon_vma_prepare(vma
)))
2474 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2477 __SetPageUptodate(page
);
2479 if (mem_cgroup_charge(page
, mm
, GFP_KERNEL
))
2482 entry
= mk_pte(page
, vma
->vm_page_prot
);
2483 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2485 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2486 if (!pte_none(*page_table
))
2488 inc_mm_counter(mm
, anon_rss
);
2489 SetPageSwapBacked(page
);
2490 lru_cache_add_active_or_unevictable(page
, vma
);
2491 page_add_new_anon_rmap(page
, vma
, address
);
2492 set_pte_at(mm
, address
, page_table
, entry
);
2494 /* No need to invalidate - it was non-present before */
2495 update_mmu_cache(vma
, address
, entry
);
2497 pte_unmap_unlock(page_table
, ptl
);
2500 mem_cgroup_uncharge_page(page
);
2501 page_cache_release(page
);
2504 page_cache_release(page
);
2506 return VM_FAULT_OOM
;
2510 * __do_fault() tries to create a new page mapping. It aggressively
2511 * tries to share with existing pages, but makes a separate copy if
2512 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2513 * the next page fault.
2515 * As this is called only for pages that do not currently exist, we
2516 * do not need to flush old virtual caches or the TLB.
2518 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2519 * but allow concurrent faults), and pte neither mapped nor locked.
2520 * We return with mmap_sem still held, but pte unmapped and unlocked.
2522 static int __do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2523 unsigned long address
, pmd_t
*pmd
,
2524 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
2532 struct page
*dirty_page
= NULL
;
2533 struct vm_fault vmf
;
2535 int page_mkwrite
= 0;
2537 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
2542 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
2543 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2547 * For consistency in subsequent calls, make the faulted page always
2550 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
2551 lock_page(vmf
.page
);
2553 VM_BUG_ON(!PageLocked(vmf
.page
));
2556 * Should we do an early C-O-W break?
2559 if (flags
& FAULT_FLAG_WRITE
) {
2560 if (!(vma
->vm_flags
& VM_SHARED
)) {
2562 if (unlikely(anon_vma_prepare(vma
))) {
2566 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
,
2572 if (mem_cgroup_charge(page
, mm
, GFP_KERNEL
)) {
2574 page_cache_release(page
);
2579 * Don't let another task, with possibly unlocked vma,
2580 * keep the mlocked page.
2582 if (vma
->vm_flags
& VM_LOCKED
)
2583 clear_page_mlock(vmf
.page
);
2584 copy_user_highpage(page
, vmf
.page
, address
, vma
);
2585 __SetPageUptodate(page
);
2588 * If the page will be shareable, see if the backing
2589 * address space wants to know that the page is about
2590 * to become writable
2592 if (vma
->vm_ops
->page_mkwrite
) {
2594 if (vma
->vm_ops
->page_mkwrite(vma
, page
) < 0) {
2595 ret
= VM_FAULT_SIGBUS
;
2596 anon
= 1; /* no anon but release vmf.page */
2601 * XXX: this is not quite right (racy vs
2602 * invalidate) to unlock and relock the page
2603 * like this, however a better fix requires
2604 * reworking page_mkwrite locking API, which
2605 * is better done later.
2607 if (!page
->mapping
) {
2609 anon
= 1; /* no anon but release vmf.page */
2618 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2621 * This silly early PAGE_DIRTY setting removes a race
2622 * due to the bad i386 page protection. But it's valid
2623 * for other architectures too.
2625 * Note that if write_access is true, we either now have
2626 * an exclusive copy of the page, or this is a shared mapping,
2627 * so we can make it writable and dirty to avoid having to
2628 * handle that later.
2630 /* Only go through if we didn't race with anybody else... */
2631 if (likely(pte_same(*page_table
, orig_pte
))) {
2632 flush_icache_page(vma
, page
);
2633 entry
= mk_pte(page
, vma
->vm_page_prot
);
2634 if (flags
& FAULT_FLAG_WRITE
)
2635 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2637 inc_mm_counter(mm
, anon_rss
);
2638 SetPageSwapBacked(page
);
2639 lru_cache_add_active_or_unevictable(page
, vma
);
2640 page_add_new_anon_rmap(page
, vma
, address
);
2642 inc_mm_counter(mm
, file_rss
);
2643 page_add_file_rmap(page
);
2644 if (flags
& FAULT_FLAG_WRITE
) {
2646 get_page(dirty_page
);
2649 //TODO: is this safe? do_anonymous_page() does it this way.
2650 set_pte_at(mm
, address
, page_table
, entry
);
2652 /* no need to invalidate: a not-present page won't be cached */
2653 update_mmu_cache(vma
, address
, entry
);
2656 mem_cgroup_uncharge_page(page
);
2658 page_cache_release(page
);
2660 anon
= 1; /* no anon but release faulted_page */
2663 pte_unmap_unlock(page_table
, ptl
);
2666 unlock_page(vmf
.page
);
2669 page_cache_release(vmf
.page
);
2670 else if (dirty_page
) {
2672 file_update_time(vma
->vm_file
);
2674 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2675 put_page(dirty_page
);
2681 static int do_linear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2682 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2683 int write_access
, pte_t orig_pte
)
2685 pgoff_t pgoff
= (((address
& PAGE_MASK
)
2686 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
2687 unsigned int flags
= (write_access
? FAULT_FLAG_WRITE
: 0);
2689 pte_unmap(page_table
);
2690 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
2694 * Fault of a previously existing named mapping. Repopulate the pte
2695 * from the encoded file_pte if possible. This enables swappable
2698 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2699 * but allow concurrent faults), and pte mapped but not yet locked.
2700 * We return with mmap_sem still held, but pte unmapped and unlocked.
2702 static int do_nonlinear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2703 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2704 int write_access
, pte_t orig_pte
)
2706 unsigned int flags
= FAULT_FLAG_NONLINEAR
|
2707 (write_access
? FAULT_FLAG_WRITE
: 0);
2710 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2713 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
) ||
2714 !(vma
->vm_flags
& VM_CAN_NONLINEAR
))) {
2716 * Page table corrupted: show pte and kill process.
2718 print_bad_pte(vma
, orig_pte
, address
);
2719 return VM_FAULT_OOM
;
2722 pgoff
= pte_to_pgoff(orig_pte
);
2723 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
2727 * These routines also need to handle stuff like marking pages dirty
2728 * and/or accessed for architectures that don't do it in hardware (most
2729 * RISC architectures). The early dirtying is also good on the i386.
2731 * There is also a hook called "update_mmu_cache()" that architectures
2732 * with external mmu caches can use to update those (ie the Sparc or
2733 * PowerPC hashed page tables that act as extended TLBs).
2735 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2736 * but allow concurrent faults), and pte mapped but not yet locked.
2737 * We return with mmap_sem still held, but pte unmapped and unlocked.
2739 static inline int handle_pte_fault(struct mm_struct
*mm
,
2740 struct vm_area_struct
*vma
, unsigned long address
,
2741 pte_t
*pte
, pmd_t
*pmd
, int write_access
)
2747 if (!pte_present(entry
)) {
2748 if (pte_none(entry
)) {
2750 if (likely(vma
->vm_ops
->fault
))
2751 return do_linear_fault(mm
, vma
, address
,
2752 pte
, pmd
, write_access
, entry
);
2754 return do_anonymous_page(mm
, vma
, address
,
2755 pte
, pmd
, write_access
);
2757 if (pte_file(entry
))
2758 return do_nonlinear_fault(mm
, vma
, address
,
2759 pte
, pmd
, write_access
, entry
);
2760 return do_swap_page(mm
, vma
, address
,
2761 pte
, pmd
, write_access
, entry
);
2764 ptl
= pte_lockptr(mm
, pmd
);
2766 if (unlikely(!pte_same(*pte
, entry
)))
2769 if (!pte_write(entry
))
2770 return do_wp_page(mm
, vma
, address
,
2771 pte
, pmd
, ptl
, entry
);
2772 entry
= pte_mkdirty(entry
);
2774 entry
= pte_mkyoung(entry
);
2775 if (ptep_set_access_flags(vma
, address
, pte
, entry
, write_access
)) {
2776 update_mmu_cache(vma
, address
, entry
);
2779 * This is needed only for protection faults but the arch code
2780 * is not yet telling us if this is a protection fault or not.
2781 * This still avoids useless tlb flushes for .text page faults
2785 flush_tlb_page(vma
, address
);
2788 pte_unmap_unlock(pte
, ptl
);
2793 * By the time we get here, we already hold the mm semaphore
2795 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2796 unsigned long address
, int write_access
)
2803 __set_current_state(TASK_RUNNING
);
2805 count_vm_event(PGFAULT
);
2807 if (unlikely(is_vm_hugetlb_page(vma
)))
2808 return hugetlb_fault(mm
, vma
, address
, write_access
);
2810 pgd
= pgd_offset(mm
, address
);
2811 pud
= pud_alloc(mm
, pgd
, address
);
2813 return VM_FAULT_OOM
;
2814 pmd
= pmd_alloc(mm
, pud
, address
);
2816 return VM_FAULT_OOM
;
2817 pte
= pte_alloc_map(mm
, pmd
, address
);
2819 return VM_FAULT_OOM
;
2821 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, write_access
);
2824 #ifndef __PAGETABLE_PUD_FOLDED
2826 * Allocate page upper directory.
2827 * We've already handled the fast-path in-line.
2829 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
2831 pud_t
*new = pud_alloc_one(mm
, address
);
2835 smp_wmb(); /* See comment in __pte_alloc */
2837 spin_lock(&mm
->page_table_lock
);
2838 if (pgd_present(*pgd
)) /* Another has populated it */
2841 pgd_populate(mm
, pgd
, new);
2842 spin_unlock(&mm
->page_table_lock
);
2845 #endif /* __PAGETABLE_PUD_FOLDED */
2847 #ifndef __PAGETABLE_PMD_FOLDED
2849 * Allocate page middle directory.
2850 * We've already handled the fast-path in-line.
2852 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
2854 pmd_t
*new = pmd_alloc_one(mm
, address
);
2858 smp_wmb(); /* See comment in __pte_alloc */
2860 spin_lock(&mm
->page_table_lock
);
2861 #ifndef __ARCH_HAS_4LEVEL_HACK
2862 if (pud_present(*pud
)) /* Another has populated it */
2865 pud_populate(mm
, pud
, new);
2867 if (pgd_present(*pud
)) /* Another has populated it */
2870 pgd_populate(mm
, pud
, new);
2871 #endif /* __ARCH_HAS_4LEVEL_HACK */
2872 spin_unlock(&mm
->page_table_lock
);
2875 #endif /* __PAGETABLE_PMD_FOLDED */
2877 int make_pages_present(unsigned long addr
, unsigned long end
)
2879 int ret
, len
, write
;
2880 struct vm_area_struct
* vma
;
2882 vma
= find_vma(current
->mm
, addr
);
2885 write
= (vma
->vm_flags
& VM_WRITE
) != 0;
2886 BUG_ON(addr
>= end
);
2887 BUG_ON(end
> vma
->vm_end
);
2888 len
= DIV_ROUND_UP(end
, PAGE_SIZE
) - addr
/PAGE_SIZE
;
2889 ret
= get_user_pages(current
, current
->mm
, addr
,
2890 len
, write
, 0, NULL
, NULL
);
2893 return ret
== len
? 0 : -EFAULT
;
2896 #if !defined(__HAVE_ARCH_GATE_AREA)
2898 #if defined(AT_SYSINFO_EHDR)
2899 static struct vm_area_struct gate_vma
;
2901 static int __init
gate_vma_init(void)
2903 gate_vma
.vm_mm
= NULL
;
2904 gate_vma
.vm_start
= FIXADDR_USER_START
;
2905 gate_vma
.vm_end
= FIXADDR_USER_END
;
2906 gate_vma
.vm_flags
= VM_READ
| VM_MAYREAD
| VM_EXEC
| VM_MAYEXEC
;
2907 gate_vma
.vm_page_prot
= __P101
;
2909 * Make sure the vDSO gets into every core dump.
2910 * Dumping its contents makes post-mortem fully interpretable later
2911 * without matching up the same kernel and hardware config to see
2912 * what PC values meant.
2914 gate_vma
.vm_flags
|= VM_ALWAYSDUMP
;
2917 __initcall(gate_vma_init
);
2920 struct vm_area_struct
*get_gate_vma(struct task_struct
*tsk
)
2922 #ifdef AT_SYSINFO_EHDR
2929 int in_gate_area_no_task(unsigned long addr
)
2931 #ifdef AT_SYSINFO_EHDR
2932 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
2938 #endif /* __HAVE_ARCH_GATE_AREA */
2940 #ifdef CONFIG_HAVE_IOREMAP_PROT
2941 int follow_phys(struct vm_area_struct
*vma
,
2942 unsigned long address
, unsigned int flags
,
2943 unsigned long *prot
, resource_size_t
*phys
)
2950 resource_size_t phys_addr
= 0;
2951 struct mm_struct
*mm
= vma
->vm_mm
;
2954 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
2957 pgd
= pgd_offset(mm
, address
);
2958 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
2961 pud
= pud_offset(pgd
, address
);
2962 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
2965 pmd
= pmd_offset(pud
, address
);
2966 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
2969 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
2973 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2978 if (!pte_present(pte
))
2980 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
2982 phys_addr
= pte_pfn(pte
);
2983 phys_addr
<<= PAGE_SHIFT
; /* Shift here to avoid overflow on PAE */
2985 *prot
= pgprot_val(pte_pgprot(pte
));
2990 pte_unmap_unlock(ptep
, ptl
);
2995 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
2996 void *buf
, int len
, int write
)
2998 resource_size_t phys_addr
;
2999 unsigned long prot
= 0;
3001 int offset
= addr
& (PAGE_SIZE
-1);
3003 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
3006 maddr
= ioremap_prot(phys_addr
, PAGE_SIZE
, prot
);
3008 memcpy_toio(maddr
+ offset
, buf
, len
);
3010 memcpy_fromio(buf
, maddr
+ offset
, len
);
3018 * Access another process' address space.
3019 * Source/target buffer must be kernel space,
3020 * Do not walk the page table directly, use get_user_pages
3022 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
, void *buf
, int len
, int write
)
3024 struct mm_struct
*mm
;
3025 struct vm_area_struct
*vma
;
3026 void *old_buf
= buf
;
3028 mm
= get_task_mm(tsk
);
3032 down_read(&mm
->mmap_sem
);
3033 /* ignore errors, just check how much was successfully transferred */
3035 int bytes
, ret
, offset
;
3037 struct page
*page
= NULL
;
3039 ret
= get_user_pages(tsk
, mm
, addr
, 1,
3040 write
, 1, &page
, &vma
);
3043 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3044 * we can access using slightly different code.
3046 #ifdef CONFIG_HAVE_IOREMAP_PROT
3047 vma
= find_vma(mm
, addr
);
3050 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
3051 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
3059 offset
= addr
& (PAGE_SIZE
-1);
3060 if (bytes
> PAGE_SIZE
-offset
)
3061 bytes
= PAGE_SIZE
-offset
;
3065 copy_to_user_page(vma
, page
, addr
,
3066 maddr
+ offset
, buf
, bytes
);
3067 set_page_dirty_lock(page
);
3069 copy_from_user_page(vma
, page
, addr
,
3070 buf
, maddr
+ offset
, bytes
);
3073 page_cache_release(page
);
3079 up_read(&mm
->mmap_sem
);
3082 return buf
- old_buf
;
3086 * Print the name of a VMA.
3088 void print_vma_addr(char *prefix
, unsigned long ip
)
3090 struct mm_struct
*mm
= current
->mm
;
3091 struct vm_area_struct
*vma
;
3094 * Do not print if we are in atomic
3095 * contexts (in exception stacks, etc.):
3097 if (preempt_count())
3100 down_read(&mm
->mmap_sem
);
3101 vma
= find_vma(mm
, ip
);
3102 if (vma
&& vma
->vm_file
) {
3103 struct file
*f
= vma
->vm_file
;
3104 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
3108 p
= d_path(&f
->f_path
, buf
, PAGE_SIZE
);
3111 s
= strrchr(p
, '/');
3114 printk("%s%s[%lx+%lx]", prefix
, p
,
3116 vma
->vm_end
- vma
->vm_start
);
3117 free_page((unsigned long)buf
);
3120 up_read(¤t
->mm
->mmap_sem
);