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
);
672 if (unlikely(is_pfn_mapping(vma
))) {
674 * We do not free on error cases below as remove_vma
675 * gets called on error from higher level routine
677 ret
= track_pfn_vma_copy(vma
);
683 * We need to invalidate the secondary MMU mappings only when
684 * there could be a permission downgrade on the ptes of the
685 * parent mm. And a permission downgrade will only happen if
686 * is_cow_mapping() returns true.
688 if (is_cow_mapping(vma
->vm_flags
))
689 mmu_notifier_invalidate_range_start(src_mm
, addr
, end
);
692 dst_pgd
= pgd_offset(dst_mm
, addr
);
693 src_pgd
= pgd_offset(src_mm
, addr
);
695 next
= pgd_addr_end(addr
, end
);
696 if (pgd_none_or_clear_bad(src_pgd
))
698 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
703 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
705 if (is_cow_mapping(vma
->vm_flags
))
706 mmu_notifier_invalidate_range_end(src_mm
,
711 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
712 struct vm_area_struct
*vma
, pmd_t
*pmd
,
713 unsigned long addr
, unsigned long end
,
714 long *zap_work
, struct zap_details
*details
)
716 struct mm_struct
*mm
= tlb
->mm
;
722 pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
723 arch_enter_lazy_mmu_mode();
726 if (pte_none(ptent
)) {
731 (*zap_work
) -= PAGE_SIZE
;
733 if (pte_present(ptent
)) {
736 page
= vm_normal_page(vma
, addr
, ptent
);
737 if (unlikely(details
) && page
) {
739 * unmap_shared_mapping_pages() wants to
740 * invalidate cache without truncating:
741 * unmap shared but keep private pages.
743 if (details
->check_mapping
&&
744 details
->check_mapping
!= page
->mapping
)
747 * Each page->index must be checked when
748 * invalidating or truncating nonlinear.
750 if (details
->nonlinear_vma
&&
751 (page
->index
< details
->first_index
||
752 page
->index
> details
->last_index
))
755 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
757 tlb_remove_tlb_entry(tlb
, pte
, addr
);
760 if (unlikely(details
) && details
->nonlinear_vma
761 && linear_page_index(details
->nonlinear_vma
,
762 addr
) != page
->index
)
763 set_pte_at(mm
, addr
, pte
,
764 pgoff_to_pte(page
->index
));
768 if (pte_dirty(ptent
))
769 set_page_dirty(page
);
770 if (pte_young(ptent
) &&
771 likely(!VM_SequentialReadHint(vma
)))
772 mark_page_accessed(page
);
775 page_remove_rmap(page
, vma
);
776 tlb_remove_page(tlb
, page
);
780 * If details->check_mapping, we leave swap entries;
781 * if details->nonlinear_vma, we leave file entries.
783 if (unlikely(details
))
785 if (!pte_file(ptent
))
786 free_swap_and_cache(pte_to_swp_entry(ptent
));
787 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
788 } while (pte
++, addr
+= PAGE_SIZE
, (addr
!= end
&& *zap_work
> 0));
790 add_mm_rss(mm
, file_rss
, anon_rss
);
791 arch_leave_lazy_mmu_mode();
792 pte_unmap_unlock(pte
- 1, ptl
);
797 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
798 struct vm_area_struct
*vma
, pud_t
*pud
,
799 unsigned long addr
, unsigned long end
,
800 long *zap_work
, struct zap_details
*details
)
805 pmd
= pmd_offset(pud
, addr
);
807 next
= pmd_addr_end(addr
, end
);
808 if (pmd_none_or_clear_bad(pmd
)) {
812 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
,
814 } while (pmd
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
819 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
820 struct vm_area_struct
*vma
, pgd_t
*pgd
,
821 unsigned long addr
, unsigned long end
,
822 long *zap_work
, struct zap_details
*details
)
827 pud
= pud_offset(pgd
, addr
);
829 next
= pud_addr_end(addr
, end
);
830 if (pud_none_or_clear_bad(pud
)) {
834 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
,
836 } while (pud
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
841 static unsigned long unmap_page_range(struct mmu_gather
*tlb
,
842 struct vm_area_struct
*vma
,
843 unsigned long addr
, unsigned long end
,
844 long *zap_work
, struct zap_details
*details
)
849 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
853 tlb_start_vma(tlb
, vma
);
854 pgd
= pgd_offset(vma
->vm_mm
, addr
);
856 next
= pgd_addr_end(addr
, end
);
857 if (pgd_none_or_clear_bad(pgd
)) {
861 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
,
863 } while (pgd
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
864 tlb_end_vma(tlb
, vma
);
869 #ifdef CONFIG_PREEMPT
870 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
872 /* No preempt: go for improved straight-line efficiency */
873 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
877 * unmap_vmas - unmap a range of memory covered by a list of vma's
878 * @tlbp: address of the caller's struct mmu_gather
879 * @vma: the starting vma
880 * @start_addr: virtual address at which to start unmapping
881 * @end_addr: virtual address at which to end unmapping
882 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
883 * @details: details of nonlinear truncation or shared cache invalidation
885 * Returns the end address of the unmapping (restart addr if interrupted).
887 * Unmap all pages in the vma list.
889 * We aim to not hold locks for too long (for scheduling latency reasons).
890 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
891 * return the ending mmu_gather to the caller.
893 * Only addresses between `start' and `end' will be unmapped.
895 * The VMA list must be sorted in ascending virtual address order.
897 * unmap_vmas() assumes that the caller will flush the whole unmapped address
898 * range after unmap_vmas() returns. So the only responsibility here is to
899 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
900 * drops the lock and schedules.
902 unsigned long unmap_vmas(struct mmu_gather
**tlbp
,
903 struct vm_area_struct
*vma
, unsigned long start_addr
,
904 unsigned long end_addr
, unsigned long *nr_accounted
,
905 struct zap_details
*details
)
907 long zap_work
= ZAP_BLOCK_SIZE
;
908 unsigned long tlb_start
= 0; /* For tlb_finish_mmu */
909 int tlb_start_valid
= 0;
910 unsigned long start
= start_addr
;
911 spinlock_t
*i_mmap_lock
= details
? details
->i_mmap_lock
: NULL
;
912 int fullmm
= (*tlbp
)->fullmm
;
913 struct mm_struct
*mm
= vma
->vm_mm
;
915 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
916 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
) {
919 start
= max(vma
->vm_start
, start_addr
);
920 if (start
>= vma
->vm_end
)
922 end
= min(vma
->vm_end
, end_addr
);
923 if (end
<= vma
->vm_start
)
926 if (vma
->vm_flags
& VM_ACCOUNT
)
927 *nr_accounted
+= (end
- start
) >> PAGE_SHIFT
;
929 if (unlikely(is_pfn_mapping(vma
)))
930 untrack_pfn_vma(vma
, 0, 0);
932 while (start
!= end
) {
933 if (!tlb_start_valid
) {
938 if (unlikely(is_vm_hugetlb_page(vma
))) {
940 * It is undesirable to test vma->vm_file as it
941 * should be non-null for valid hugetlb area.
942 * However, vm_file will be NULL in the error
943 * cleanup path of do_mmap_pgoff. When
944 * hugetlbfs ->mmap method fails,
945 * do_mmap_pgoff() nullifies vma->vm_file
946 * before calling this function to clean up.
947 * Since no pte has actually been setup, it is
948 * safe to do nothing in this case.
951 unmap_hugepage_range(vma
, start
, end
, NULL
);
952 zap_work
-= (end
- start
) /
953 pages_per_huge_page(hstate_vma(vma
));
958 start
= unmap_page_range(*tlbp
, vma
,
959 start
, end
, &zap_work
, details
);
962 BUG_ON(start
!= end
);
966 tlb_finish_mmu(*tlbp
, tlb_start
, start
);
968 if (need_resched() ||
969 (i_mmap_lock
&& spin_needbreak(i_mmap_lock
))) {
977 *tlbp
= tlb_gather_mmu(vma
->vm_mm
, fullmm
);
979 zap_work
= ZAP_BLOCK_SIZE
;
983 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
984 return start
; /* which is now the end (or restart) address */
988 * zap_page_range - remove user pages in a given range
989 * @vma: vm_area_struct holding the applicable pages
990 * @address: starting address of pages to zap
991 * @size: number of bytes to zap
992 * @details: details of nonlinear truncation or shared cache invalidation
994 unsigned long zap_page_range(struct vm_area_struct
*vma
, unsigned long address
,
995 unsigned long size
, struct zap_details
*details
)
997 struct mm_struct
*mm
= vma
->vm_mm
;
998 struct mmu_gather
*tlb
;
999 unsigned long end
= address
+ size
;
1000 unsigned long nr_accounted
= 0;
1003 tlb
= tlb_gather_mmu(mm
, 0);
1004 update_hiwater_rss(mm
);
1005 end
= unmap_vmas(&tlb
, vma
, address
, end
, &nr_accounted
, details
);
1007 tlb_finish_mmu(tlb
, address
, end
);
1012 * zap_vma_ptes - remove ptes mapping the vma
1013 * @vma: vm_area_struct holding ptes to be zapped
1014 * @address: starting address of pages to zap
1015 * @size: number of bytes to zap
1017 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1019 * The entire address range must be fully contained within the vma.
1021 * Returns 0 if successful.
1023 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1026 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1027 !(vma
->vm_flags
& VM_PFNMAP
))
1029 zap_page_range(vma
, address
, size
, NULL
);
1032 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1035 * Do a quick page-table lookup for a single page.
1037 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
1046 struct mm_struct
*mm
= vma
->vm_mm
;
1048 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
1049 if (!IS_ERR(page
)) {
1050 BUG_ON(flags
& FOLL_GET
);
1055 pgd
= pgd_offset(mm
, address
);
1056 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
1059 pud
= pud_offset(pgd
, address
);
1062 if (pud_huge(*pud
)) {
1063 BUG_ON(flags
& FOLL_GET
);
1064 page
= follow_huge_pud(mm
, address
, pud
, flags
& FOLL_WRITE
);
1067 if (unlikely(pud_bad(*pud
)))
1070 pmd
= pmd_offset(pud
, address
);
1073 if (pmd_huge(*pmd
)) {
1074 BUG_ON(flags
& FOLL_GET
);
1075 page
= follow_huge_pmd(mm
, address
, pmd
, flags
& FOLL_WRITE
);
1078 if (unlikely(pmd_bad(*pmd
)))
1081 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1084 if (!pte_present(pte
))
1086 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
1088 page
= vm_normal_page(vma
, address
, pte
);
1089 if (unlikely(!page
))
1092 if (flags
& FOLL_GET
)
1094 if (flags
& FOLL_TOUCH
) {
1095 if ((flags
& FOLL_WRITE
) &&
1096 !pte_dirty(pte
) && !PageDirty(page
))
1097 set_page_dirty(page
);
1098 mark_page_accessed(page
);
1101 pte_unmap_unlock(ptep
, ptl
);
1106 pte_unmap_unlock(ptep
, ptl
);
1107 return ERR_PTR(-EFAULT
);
1110 pte_unmap_unlock(ptep
, ptl
);
1113 /* Fall through to ZERO_PAGE handling */
1116 * When core dumping an enormous anonymous area that nobody
1117 * has touched so far, we don't want to allocate page tables.
1119 if (flags
& FOLL_ANON
) {
1120 page
= ZERO_PAGE(0);
1121 if (flags
& FOLL_GET
)
1123 BUG_ON(flags
& FOLL_WRITE
);
1128 /* Can we do the FOLL_ANON optimization? */
1129 static inline int use_zero_page(struct vm_area_struct
*vma
)
1132 * We don't want to optimize FOLL_ANON for make_pages_present()
1133 * when it tries to page in a VM_LOCKED region. As to VM_SHARED,
1134 * we want to get the page from the page tables to make sure
1135 * that we serialize and update with any other user of that
1138 if (vma
->vm_flags
& (VM_LOCKED
| VM_SHARED
))
1141 * And if we have a fault routine, it's not an anonymous region.
1143 return !vma
->vm_ops
|| !vma
->vm_ops
->fault
;
1148 int __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1149 unsigned long start
, int len
, int flags
,
1150 struct page
**pages
, struct vm_area_struct
**vmas
)
1153 unsigned int vm_flags
= 0;
1154 int write
= !!(flags
& GUP_FLAGS_WRITE
);
1155 int force
= !!(flags
& GUP_FLAGS_FORCE
);
1156 int ignore
= !!(flags
& GUP_FLAGS_IGNORE_VMA_PERMISSIONS
);
1161 * Require read or write permissions.
1162 * If 'force' is set, we only require the "MAY" flags.
1164 vm_flags
= write
? (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1165 vm_flags
&= force
? (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1169 struct vm_area_struct
*vma
;
1170 unsigned int foll_flags
;
1172 vma
= find_extend_vma(mm
, start
);
1173 if (!vma
&& in_gate_area(tsk
, start
)) {
1174 unsigned long pg
= start
& PAGE_MASK
;
1175 struct vm_area_struct
*gate_vma
= get_gate_vma(tsk
);
1181 /* user gate pages are read-only */
1182 if (!ignore
&& write
)
1183 return i
? : -EFAULT
;
1185 pgd
= pgd_offset_k(pg
);
1187 pgd
= pgd_offset_gate(mm
, pg
);
1188 BUG_ON(pgd_none(*pgd
));
1189 pud
= pud_offset(pgd
, pg
);
1190 BUG_ON(pud_none(*pud
));
1191 pmd
= pmd_offset(pud
, pg
);
1193 return i
? : -EFAULT
;
1194 pte
= pte_offset_map(pmd
, pg
);
1195 if (pte_none(*pte
)) {
1197 return i
? : -EFAULT
;
1200 struct page
*page
= vm_normal_page(gate_vma
, start
, *pte
);
1215 (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1216 (!ignore
&& !(vm_flags
& vma
->vm_flags
)))
1217 return i
? : -EFAULT
;
1219 if (is_vm_hugetlb_page(vma
)) {
1220 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1221 &start
, &len
, i
, write
);
1225 foll_flags
= FOLL_TOUCH
;
1227 foll_flags
|= FOLL_GET
;
1228 if (!write
&& use_zero_page(vma
))
1229 foll_flags
|= FOLL_ANON
;
1235 * If tsk is ooming, cut off its access to large memory
1236 * allocations. It has a pending SIGKILL, but it can't
1237 * be processed until returning to user space.
1239 if (unlikely(test_tsk_thread_flag(tsk
, TIF_MEMDIE
)))
1240 return i
? i
: -ENOMEM
;
1243 foll_flags
|= FOLL_WRITE
;
1246 while (!(page
= follow_page(vma
, start
, foll_flags
))) {
1248 ret
= handle_mm_fault(mm
, vma
, start
,
1249 foll_flags
& FOLL_WRITE
);
1250 if (ret
& VM_FAULT_ERROR
) {
1251 if (ret
& VM_FAULT_OOM
)
1252 return i
? i
: -ENOMEM
;
1253 else if (ret
& VM_FAULT_SIGBUS
)
1254 return i
? i
: -EFAULT
;
1257 if (ret
& VM_FAULT_MAJOR
)
1263 * The VM_FAULT_WRITE bit tells us that
1264 * do_wp_page has broken COW when necessary,
1265 * even if maybe_mkwrite decided not to set
1266 * pte_write. We can thus safely do subsequent
1267 * page lookups as if they were reads. But only
1268 * do so when looping for pte_write is futile:
1269 * in some cases userspace may also be wanting
1270 * to write to the gotten user page, which a
1271 * read fault here might prevent (a readonly
1272 * page might get reCOWed by userspace write).
1274 if ((ret
& VM_FAULT_WRITE
) &&
1275 !(vma
->vm_flags
& VM_WRITE
))
1276 foll_flags
&= ~FOLL_WRITE
;
1281 return i
? i
: PTR_ERR(page
);
1285 flush_anon_page(vma
, page
, start
);
1286 flush_dcache_page(page
);
1293 } while (len
&& start
< vma
->vm_end
);
1298 int get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1299 unsigned long start
, int len
, int write
, int force
,
1300 struct page
**pages
, struct vm_area_struct
**vmas
)
1305 flags
|= GUP_FLAGS_WRITE
;
1307 flags
|= GUP_FLAGS_FORCE
;
1309 return __get_user_pages(tsk
, mm
,
1314 EXPORT_SYMBOL(get_user_pages
);
1316 pte_t
*get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1319 pgd_t
* pgd
= pgd_offset(mm
, addr
);
1320 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
1322 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
1324 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1330 * This is the old fallback for page remapping.
1332 * For historical reasons, it only allows reserved pages. Only
1333 * old drivers should use this, and they needed to mark their
1334 * pages reserved for the old functions anyway.
1336 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1337 struct page
*page
, pgprot_t prot
)
1339 struct mm_struct
*mm
= vma
->vm_mm
;
1348 flush_dcache_page(page
);
1349 pte
= get_locked_pte(mm
, addr
, &ptl
);
1353 if (!pte_none(*pte
))
1356 /* Ok, finally just insert the thing.. */
1358 inc_mm_counter(mm
, file_rss
);
1359 page_add_file_rmap(page
);
1360 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1363 pte_unmap_unlock(pte
, ptl
);
1366 pte_unmap_unlock(pte
, ptl
);
1372 * vm_insert_page - insert single page into user vma
1373 * @vma: user vma to map to
1374 * @addr: target user address of this page
1375 * @page: source kernel page
1377 * This allows drivers to insert individual pages they've allocated
1380 * The page has to be a nice clean _individual_ kernel allocation.
1381 * If you allocate a compound page, you need to have marked it as
1382 * such (__GFP_COMP), or manually just split the page up yourself
1383 * (see split_page()).
1385 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1386 * took an arbitrary page protection parameter. This doesn't allow
1387 * that. Your vma protection will have to be set up correctly, which
1388 * means that if you want a shared writable mapping, you'd better
1389 * ask for a shared writable mapping!
1391 * The page does not need to be reserved.
1393 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1396 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1398 if (!page_count(page
))
1400 vma
->vm_flags
|= VM_INSERTPAGE
;
1401 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1403 EXPORT_SYMBOL(vm_insert_page
);
1405 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1406 unsigned long pfn
, pgprot_t prot
)
1408 struct mm_struct
*mm
= vma
->vm_mm
;
1414 pte
= get_locked_pte(mm
, addr
, &ptl
);
1418 if (!pte_none(*pte
))
1421 /* Ok, finally just insert the thing.. */
1422 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
1423 set_pte_at(mm
, addr
, pte
, entry
);
1424 update_mmu_cache(vma
, addr
, entry
); /* XXX: why not for insert_page? */
1428 pte_unmap_unlock(pte
, ptl
);
1434 * vm_insert_pfn - insert single pfn into user vma
1435 * @vma: user vma to map to
1436 * @addr: target user address of this page
1437 * @pfn: source kernel pfn
1439 * Similar to vm_inert_page, this allows drivers to insert individual pages
1440 * they've allocated into a user vma. Same comments apply.
1442 * This function should only be called from a vm_ops->fault handler, and
1443 * in that case the handler should return NULL.
1445 * vma cannot be a COW mapping.
1447 * As this is called only for pages that do not currently exist, we
1448 * do not need to flush old virtual caches or the TLB.
1450 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1455 * Technically, architectures with pte_special can avoid all these
1456 * restrictions (same for remap_pfn_range). However we would like
1457 * consistency in testing and feature parity among all, so we should
1458 * try to keep these invariants in place for everybody.
1460 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1461 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1462 (VM_PFNMAP
|VM_MIXEDMAP
));
1463 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1464 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1466 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1468 if (track_pfn_vma_new(vma
, vma
->vm_page_prot
, pfn
, PAGE_SIZE
))
1471 ret
= insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1474 untrack_pfn_vma(vma
, pfn
, PAGE_SIZE
);
1478 EXPORT_SYMBOL(vm_insert_pfn
);
1480 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1483 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
1485 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1489 * If we don't have pte special, then we have to use the pfn_valid()
1490 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1491 * refcount the page if pfn_valid is true (hence insert_page rather
1494 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
1497 page
= pfn_to_page(pfn
);
1498 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1500 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1502 EXPORT_SYMBOL(vm_insert_mixed
);
1505 * maps a range of physical memory into the requested pages. the old
1506 * mappings are removed. any references to nonexistent pages results
1507 * in null mappings (currently treated as "copy-on-access")
1509 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1510 unsigned long addr
, unsigned long end
,
1511 unsigned long pfn
, pgprot_t prot
)
1516 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1519 arch_enter_lazy_mmu_mode();
1521 BUG_ON(!pte_none(*pte
));
1522 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
1524 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1525 arch_leave_lazy_mmu_mode();
1526 pte_unmap_unlock(pte
- 1, ptl
);
1530 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1531 unsigned long addr
, unsigned long end
,
1532 unsigned long pfn
, pgprot_t prot
)
1537 pfn
-= addr
>> PAGE_SHIFT
;
1538 pmd
= pmd_alloc(mm
, pud
, addr
);
1542 next
= pmd_addr_end(addr
, end
);
1543 if (remap_pte_range(mm
, pmd
, addr
, next
,
1544 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1546 } while (pmd
++, addr
= next
, addr
!= end
);
1550 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1551 unsigned long addr
, unsigned long end
,
1552 unsigned long pfn
, pgprot_t prot
)
1557 pfn
-= addr
>> PAGE_SHIFT
;
1558 pud
= pud_alloc(mm
, pgd
, addr
);
1562 next
= pud_addr_end(addr
, end
);
1563 if (remap_pmd_range(mm
, pud
, addr
, next
,
1564 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1566 } while (pud
++, addr
= next
, addr
!= end
);
1571 * remap_pfn_range - remap kernel memory to userspace
1572 * @vma: user vma to map to
1573 * @addr: target user address to start at
1574 * @pfn: physical address of kernel memory
1575 * @size: size of map area
1576 * @prot: page protection flags for this mapping
1578 * Note: this is only safe if the mm semaphore is held when called.
1580 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1581 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1585 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1586 struct mm_struct
*mm
= vma
->vm_mm
;
1590 * Physically remapped pages are special. Tell the
1591 * rest of the world about it:
1592 * VM_IO tells people not to look at these pages
1593 * (accesses can have side effects).
1594 * VM_RESERVED is specified all over the place, because
1595 * in 2.4 it kept swapout's vma scan off this vma; but
1596 * in 2.6 the LRU scan won't even find its pages, so this
1597 * flag means no more than count its pages in reserved_vm,
1598 * and omit it from core dump, even when VM_IO turned off.
1599 * VM_PFNMAP tells the core MM that the base pages are just
1600 * raw PFN mappings, and do not have a "struct page" associated
1603 * There's a horrible special case to handle copy-on-write
1604 * behaviour that some programs depend on. We mark the "original"
1605 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1607 if (addr
== vma
->vm_start
&& end
== vma
->vm_end
)
1608 vma
->vm_pgoff
= pfn
;
1609 else if (is_cow_mapping(vma
->vm_flags
))
1612 vma
->vm_flags
|= VM_IO
| VM_RESERVED
| VM_PFNMAP
;
1614 err
= track_pfn_vma_new(vma
, prot
, pfn
, PAGE_ALIGN(size
));
1618 BUG_ON(addr
>= end
);
1619 pfn
-= addr
>> PAGE_SHIFT
;
1620 pgd
= pgd_offset(mm
, addr
);
1621 flush_cache_range(vma
, addr
, end
);
1623 next
= pgd_addr_end(addr
, end
);
1624 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1625 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1628 } while (pgd
++, addr
= next
, addr
!= end
);
1631 untrack_pfn_vma(vma
, pfn
, PAGE_ALIGN(size
));
1635 EXPORT_SYMBOL(remap_pfn_range
);
1637 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1638 unsigned long addr
, unsigned long end
,
1639 pte_fn_t fn
, void *data
)
1644 spinlock_t
*uninitialized_var(ptl
);
1646 pte
= (mm
== &init_mm
) ?
1647 pte_alloc_kernel(pmd
, addr
) :
1648 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1652 BUG_ON(pmd_huge(*pmd
));
1654 arch_enter_lazy_mmu_mode();
1656 token
= pmd_pgtable(*pmd
);
1659 err
= fn(pte
, token
, addr
, data
);
1662 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1664 arch_leave_lazy_mmu_mode();
1667 pte_unmap_unlock(pte
-1, ptl
);
1671 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1672 unsigned long addr
, unsigned long end
,
1673 pte_fn_t fn
, void *data
)
1679 BUG_ON(pud_huge(*pud
));
1681 pmd
= pmd_alloc(mm
, pud
, addr
);
1685 next
= pmd_addr_end(addr
, end
);
1686 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
1689 } while (pmd
++, addr
= next
, addr
!= end
);
1693 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1694 unsigned long addr
, unsigned long end
,
1695 pte_fn_t fn
, void *data
)
1701 pud
= pud_alloc(mm
, pgd
, addr
);
1705 next
= pud_addr_end(addr
, end
);
1706 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
1709 } while (pud
++, addr
= next
, addr
!= end
);
1714 * Scan a region of virtual memory, filling in page tables as necessary
1715 * and calling a provided function on each leaf page table.
1717 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
1718 unsigned long size
, pte_fn_t fn
, void *data
)
1722 unsigned long start
= addr
, end
= addr
+ size
;
1725 BUG_ON(addr
>= end
);
1726 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1727 pgd
= pgd_offset(mm
, addr
);
1729 next
= pgd_addr_end(addr
, end
);
1730 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
1733 } while (pgd
++, addr
= next
, addr
!= end
);
1734 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1737 EXPORT_SYMBOL_GPL(apply_to_page_range
);
1740 * handle_pte_fault chooses page fault handler according to an entry
1741 * which was read non-atomically. Before making any commitment, on
1742 * those architectures or configurations (e.g. i386 with PAE) which
1743 * might give a mix of unmatched parts, do_swap_page and do_file_page
1744 * must check under lock before unmapping the pte and proceeding
1745 * (but do_wp_page is only called after already making such a check;
1746 * and do_anonymous_page and do_no_page can safely check later on).
1748 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
1749 pte_t
*page_table
, pte_t orig_pte
)
1752 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1753 if (sizeof(pte_t
) > sizeof(unsigned long)) {
1754 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
1756 same
= pte_same(*page_table
, orig_pte
);
1760 pte_unmap(page_table
);
1765 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1766 * servicing faults for write access. In the normal case, do always want
1767 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1768 * that do not have writing enabled, when used by access_process_vm.
1770 static inline pte_t
maybe_mkwrite(pte_t pte
, struct vm_area_struct
*vma
)
1772 if (likely(vma
->vm_flags
& VM_WRITE
))
1773 pte
= pte_mkwrite(pte
);
1777 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
1780 * If the source page was a PFN mapping, we don't have
1781 * a "struct page" for it. We do a best-effort copy by
1782 * just copying from the original user address. If that
1783 * fails, we just zero-fill it. Live with it.
1785 if (unlikely(!src
)) {
1786 void *kaddr
= kmap_atomic(dst
, KM_USER0
);
1787 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
1790 * This really shouldn't fail, because the page is there
1791 * in the page tables. But it might just be unreadable,
1792 * in which case we just give up and fill the result with
1795 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
1796 memset(kaddr
, 0, PAGE_SIZE
);
1797 kunmap_atomic(kaddr
, KM_USER0
);
1798 flush_dcache_page(dst
);
1800 copy_user_highpage(dst
, src
, va
, vma
);
1804 * This routine handles present pages, when users try to write
1805 * to a shared page. It is done by copying the page to a new address
1806 * and decrementing the shared-page counter for the old page.
1808 * Note that this routine assumes that the protection checks have been
1809 * done by the caller (the low-level page fault routine in most cases).
1810 * Thus we can safely just mark it writable once we've done any necessary
1813 * We also mark the page dirty at this point even though the page will
1814 * change only once the write actually happens. This avoids a few races,
1815 * and potentially makes it more efficient.
1817 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1818 * but allow concurrent faults), with pte both mapped and locked.
1819 * We return with mmap_sem still held, but pte unmapped and unlocked.
1821 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1822 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1823 spinlock_t
*ptl
, pte_t orig_pte
)
1825 struct page
*old_page
, *new_page
;
1827 int reuse
= 0, ret
= 0;
1828 int page_mkwrite
= 0;
1829 struct page
*dirty_page
= NULL
;
1831 old_page
= vm_normal_page(vma
, address
, orig_pte
);
1834 * VM_MIXEDMAP !pfn_valid() case
1836 * We should not cow pages in a shared writeable mapping.
1837 * Just mark the pages writable as we can't do any dirty
1838 * accounting on raw pfn maps.
1840 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
1841 (VM_WRITE
|VM_SHARED
))
1847 * Take out anonymous pages first, anonymous shared vmas are
1848 * not dirty accountable.
1850 if (PageAnon(old_page
)) {
1851 if (!trylock_page(old_page
)) {
1852 page_cache_get(old_page
);
1853 pte_unmap_unlock(page_table
, ptl
);
1854 lock_page(old_page
);
1855 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
1857 if (!pte_same(*page_table
, orig_pte
)) {
1858 unlock_page(old_page
);
1859 page_cache_release(old_page
);
1862 page_cache_release(old_page
);
1864 reuse
= reuse_swap_page(old_page
);
1865 unlock_page(old_page
);
1866 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
1867 (VM_WRITE
|VM_SHARED
))) {
1869 * Only catch write-faults on shared writable pages,
1870 * read-only shared pages can get COWed by
1871 * get_user_pages(.write=1, .force=1).
1873 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
1875 * Notify the address space that the page is about to
1876 * become writable so that it can prohibit this or wait
1877 * for the page to get into an appropriate state.
1879 * We do this without the lock held, so that it can
1880 * sleep if it needs to.
1882 page_cache_get(old_page
);
1883 pte_unmap_unlock(page_table
, ptl
);
1885 if (vma
->vm_ops
->page_mkwrite(vma
, old_page
) < 0)
1886 goto unwritable_page
;
1889 * Since we dropped the lock we need to revalidate
1890 * the PTE as someone else may have changed it. If
1891 * they did, we just return, as we can count on the
1892 * MMU to tell us if they didn't also make it writable.
1894 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
1896 page_cache_release(old_page
);
1897 if (!pte_same(*page_table
, orig_pte
))
1902 dirty_page
= old_page
;
1903 get_page(dirty_page
);
1909 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
1910 entry
= pte_mkyoung(orig_pte
);
1911 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1912 if (ptep_set_access_flags(vma
, address
, page_table
, entry
,1))
1913 update_mmu_cache(vma
, address
, entry
);
1914 ret
|= VM_FAULT_WRITE
;
1919 * Ok, we need to copy. Oh, well..
1921 page_cache_get(old_page
);
1923 pte_unmap_unlock(page_table
, ptl
);
1925 if (unlikely(anon_vma_prepare(vma
)))
1927 VM_BUG_ON(old_page
== ZERO_PAGE(0));
1928 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
1932 * Don't let another task, with possibly unlocked vma,
1933 * keep the mlocked page.
1935 if (vma
->vm_flags
& VM_LOCKED
) {
1936 lock_page(old_page
); /* for LRU manipulation */
1937 clear_page_mlock(old_page
);
1938 unlock_page(old_page
);
1940 cow_user_page(new_page
, old_page
, address
, vma
);
1941 __SetPageUptodate(new_page
);
1943 if (mem_cgroup_charge(new_page
, mm
, GFP_KERNEL
))
1947 * Re-check the pte - we dropped the lock
1949 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1950 if (likely(pte_same(*page_table
, orig_pte
))) {
1952 if (!PageAnon(old_page
)) {
1953 dec_mm_counter(mm
, file_rss
);
1954 inc_mm_counter(mm
, anon_rss
);
1957 inc_mm_counter(mm
, anon_rss
);
1958 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
1959 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
1960 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1962 * Clear the pte entry and flush it first, before updating the
1963 * pte with the new entry. This will avoid a race condition
1964 * seen in the presence of one thread doing SMC and another
1967 ptep_clear_flush_notify(vma
, address
, page_table
);
1968 page_add_new_anon_rmap(new_page
, vma
, address
);
1969 set_pte_at(mm
, address
, page_table
, entry
);
1970 update_mmu_cache(vma
, address
, entry
);
1973 * Only after switching the pte to the new page may
1974 * we remove the mapcount here. Otherwise another
1975 * process may come and find the rmap count decremented
1976 * before the pte is switched to the new page, and
1977 * "reuse" the old page writing into it while our pte
1978 * here still points into it and can be read by other
1981 * The critical issue is to order this
1982 * page_remove_rmap with the ptp_clear_flush above.
1983 * Those stores are ordered by (if nothing else,)
1984 * the barrier present in the atomic_add_negative
1985 * in page_remove_rmap.
1987 * Then the TLB flush in ptep_clear_flush ensures that
1988 * no process can access the old page before the
1989 * decremented mapcount is visible. And the old page
1990 * cannot be reused until after the decremented
1991 * mapcount is visible. So transitively, TLBs to
1992 * old page will be flushed before it can be reused.
1994 page_remove_rmap(old_page
, vma
);
1997 /* Free the old page.. */
1998 new_page
= old_page
;
1999 ret
|= VM_FAULT_WRITE
;
2001 mem_cgroup_uncharge_page(new_page
);
2004 page_cache_release(new_page
);
2006 page_cache_release(old_page
);
2008 pte_unmap_unlock(page_table
, ptl
);
2011 file_update_time(vma
->vm_file
);
2014 * Yes, Virginia, this is actually required to prevent a race
2015 * with clear_page_dirty_for_io() from clearing the page dirty
2016 * bit after it clear all dirty ptes, but before a racing
2017 * do_wp_page installs a dirty pte.
2019 * do_no_page is protected similarly.
2021 wait_on_page_locked(dirty_page
);
2022 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2023 put_page(dirty_page
);
2027 page_cache_release(new_page
);
2030 page_cache_release(old_page
);
2031 return VM_FAULT_OOM
;
2034 page_cache_release(old_page
);
2035 return VM_FAULT_SIGBUS
;
2039 * Helper functions for unmap_mapping_range().
2041 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2043 * We have to restart searching the prio_tree whenever we drop the lock,
2044 * since the iterator is only valid while the lock is held, and anyway
2045 * a later vma might be split and reinserted earlier while lock dropped.
2047 * The list of nonlinear vmas could be handled more efficiently, using
2048 * a placeholder, but handle it in the same way until a need is shown.
2049 * It is important to search the prio_tree before nonlinear list: a vma
2050 * may become nonlinear and be shifted from prio_tree to nonlinear list
2051 * while the lock is dropped; but never shifted from list to prio_tree.
2053 * In order to make forward progress despite restarting the search,
2054 * vm_truncate_count is used to mark a vma as now dealt with, so we can
2055 * quickly skip it next time around. Since the prio_tree search only
2056 * shows us those vmas affected by unmapping the range in question, we
2057 * can't efficiently keep all vmas in step with mapping->truncate_count:
2058 * so instead reset them all whenever it wraps back to 0 (then go to 1).
2059 * mapping->truncate_count and vma->vm_truncate_count are protected by
2062 * In order to make forward progress despite repeatedly restarting some
2063 * large vma, note the restart_addr from unmap_vmas when it breaks out:
2064 * and restart from that address when we reach that vma again. It might
2065 * have been split or merged, shrunk or extended, but never shifted: so
2066 * restart_addr remains valid so long as it remains in the vma's range.
2067 * unmap_mapping_range forces truncate_count to leap over page-aligned
2068 * values so we can save vma's restart_addr in its truncate_count field.
2070 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2072 static void reset_vma_truncate_counts(struct address_space
*mapping
)
2074 struct vm_area_struct
*vma
;
2075 struct prio_tree_iter iter
;
2077 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, 0, ULONG_MAX
)
2078 vma
->vm_truncate_count
= 0;
2079 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
2080 vma
->vm_truncate_count
= 0;
2083 static int unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2084 unsigned long start_addr
, unsigned long end_addr
,
2085 struct zap_details
*details
)
2087 unsigned long restart_addr
;
2091 * files that support invalidating or truncating portions of the
2092 * file from under mmaped areas must have their ->fault function
2093 * return a locked page (and set VM_FAULT_LOCKED in the return).
2094 * This provides synchronisation against concurrent unmapping here.
2098 restart_addr
= vma
->vm_truncate_count
;
2099 if (is_restart_addr(restart_addr
) && start_addr
< restart_addr
) {
2100 start_addr
= restart_addr
;
2101 if (start_addr
>= end_addr
) {
2102 /* Top of vma has been split off since last time */
2103 vma
->vm_truncate_count
= details
->truncate_count
;
2108 restart_addr
= zap_page_range(vma
, start_addr
,
2109 end_addr
- start_addr
, details
);
2110 need_break
= need_resched() || spin_needbreak(details
->i_mmap_lock
);
2112 if (restart_addr
>= end_addr
) {
2113 /* We have now completed this vma: mark it so */
2114 vma
->vm_truncate_count
= details
->truncate_count
;
2118 /* Note restart_addr in vma's truncate_count field */
2119 vma
->vm_truncate_count
= restart_addr
;
2124 spin_unlock(details
->i_mmap_lock
);
2126 spin_lock(details
->i_mmap_lock
);
2130 static inline void unmap_mapping_range_tree(struct prio_tree_root
*root
,
2131 struct zap_details
*details
)
2133 struct vm_area_struct
*vma
;
2134 struct prio_tree_iter iter
;
2135 pgoff_t vba
, vea
, zba
, zea
;
2138 vma_prio_tree_foreach(vma
, &iter
, root
,
2139 details
->first_index
, details
->last_index
) {
2140 /* Skip quickly over those we have already dealt with */
2141 if (vma
->vm_truncate_count
== details
->truncate_count
)
2144 vba
= vma
->vm_pgoff
;
2145 vea
= vba
+ ((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) - 1;
2146 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2147 zba
= details
->first_index
;
2150 zea
= details
->last_index
;
2154 if (unmap_mapping_range_vma(vma
,
2155 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2156 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2162 static inline void unmap_mapping_range_list(struct list_head
*head
,
2163 struct zap_details
*details
)
2165 struct vm_area_struct
*vma
;
2168 * In nonlinear VMAs there is no correspondence between virtual address
2169 * offset and file offset. So we must perform an exhaustive search
2170 * across *all* the pages in each nonlinear VMA, not just the pages
2171 * whose virtual address lies outside the file truncation point.
2174 list_for_each_entry(vma
, head
, shared
.vm_set
.list
) {
2175 /* Skip quickly over those we have already dealt with */
2176 if (vma
->vm_truncate_count
== details
->truncate_count
)
2178 details
->nonlinear_vma
= vma
;
2179 if (unmap_mapping_range_vma(vma
, vma
->vm_start
,
2180 vma
->vm_end
, details
) < 0)
2186 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2187 * @mapping: the address space containing mmaps to be unmapped.
2188 * @holebegin: byte in first page to unmap, relative to the start of
2189 * the underlying file. This will be rounded down to a PAGE_SIZE
2190 * boundary. Note that this is different from vmtruncate(), which
2191 * must keep the partial page. In contrast, we must get rid of
2193 * @holelen: size of prospective hole in bytes. This will be rounded
2194 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2196 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2197 * but 0 when invalidating pagecache, don't throw away private data.
2199 void unmap_mapping_range(struct address_space
*mapping
,
2200 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2202 struct zap_details details
;
2203 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2204 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2206 /* Check for overflow. */
2207 if (sizeof(holelen
) > sizeof(hlen
)) {
2209 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2210 if (holeend
& ~(long long)ULONG_MAX
)
2211 hlen
= ULONG_MAX
- hba
+ 1;
2214 details
.check_mapping
= even_cows
? NULL
: mapping
;
2215 details
.nonlinear_vma
= NULL
;
2216 details
.first_index
= hba
;
2217 details
.last_index
= hba
+ hlen
- 1;
2218 if (details
.last_index
< details
.first_index
)
2219 details
.last_index
= ULONG_MAX
;
2220 details
.i_mmap_lock
= &mapping
->i_mmap_lock
;
2222 spin_lock(&mapping
->i_mmap_lock
);
2224 /* Protect against endless unmapping loops */
2225 mapping
->truncate_count
++;
2226 if (unlikely(is_restart_addr(mapping
->truncate_count
))) {
2227 if (mapping
->truncate_count
== 0)
2228 reset_vma_truncate_counts(mapping
);
2229 mapping
->truncate_count
++;
2231 details
.truncate_count
= mapping
->truncate_count
;
2233 if (unlikely(!prio_tree_empty(&mapping
->i_mmap
)))
2234 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2235 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
2236 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
2237 spin_unlock(&mapping
->i_mmap_lock
);
2239 EXPORT_SYMBOL(unmap_mapping_range
);
2242 * vmtruncate - unmap mappings "freed" by truncate() syscall
2243 * @inode: inode of the file used
2244 * @offset: file offset to start truncating
2246 * NOTE! We have to be ready to update the memory sharing
2247 * between the file and the memory map for a potential last
2248 * incomplete page. Ugly, but necessary.
2250 int vmtruncate(struct inode
* inode
, loff_t offset
)
2252 if (inode
->i_size
< offset
) {
2253 unsigned long limit
;
2255 limit
= current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
;
2256 if (limit
!= RLIM_INFINITY
&& offset
> limit
)
2258 if (offset
> inode
->i_sb
->s_maxbytes
)
2260 i_size_write(inode
, offset
);
2262 struct address_space
*mapping
= inode
->i_mapping
;
2265 * truncation of in-use swapfiles is disallowed - it would
2266 * cause subsequent swapout to scribble on the now-freed
2269 if (IS_SWAPFILE(inode
))
2271 i_size_write(inode
, offset
);
2274 * unmap_mapping_range is called twice, first simply for
2275 * efficiency so that truncate_inode_pages does fewer
2276 * single-page unmaps. However after this first call, and
2277 * before truncate_inode_pages finishes, it is possible for
2278 * private pages to be COWed, which remain after
2279 * truncate_inode_pages finishes, hence the second
2280 * unmap_mapping_range call must be made for correctness.
2282 unmap_mapping_range(mapping
, offset
+ PAGE_SIZE
- 1, 0, 1);
2283 truncate_inode_pages(mapping
, offset
);
2284 unmap_mapping_range(mapping
, offset
+ PAGE_SIZE
- 1, 0, 1);
2287 if (inode
->i_op
->truncate
)
2288 inode
->i_op
->truncate(inode
);
2292 send_sig(SIGXFSZ
, current
, 0);
2296 EXPORT_SYMBOL(vmtruncate
);
2298 int vmtruncate_range(struct inode
*inode
, loff_t offset
, loff_t end
)
2300 struct address_space
*mapping
= inode
->i_mapping
;
2303 * If the underlying filesystem is not going to provide
2304 * a way to truncate a range of blocks (punch a hole) -
2305 * we should return failure right now.
2307 if (!inode
->i_op
->truncate_range
)
2310 mutex_lock(&inode
->i_mutex
);
2311 down_write(&inode
->i_alloc_sem
);
2312 unmap_mapping_range(mapping
, offset
, (end
- offset
), 1);
2313 truncate_inode_pages_range(mapping
, offset
, end
);
2314 unmap_mapping_range(mapping
, offset
, (end
- offset
), 1);
2315 inode
->i_op
->truncate_range(inode
, offset
, end
);
2316 up_write(&inode
->i_alloc_sem
);
2317 mutex_unlock(&inode
->i_mutex
);
2323 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2324 * but allow concurrent faults), and pte mapped but not yet locked.
2325 * We return with mmap_sem still held, but pte unmapped and unlocked.
2327 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2328 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2329 int write_access
, pte_t orig_pte
)
2337 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2340 entry
= pte_to_swp_entry(orig_pte
);
2341 if (is_migration_entry(entry
)) {
2342 migration_entry_wait(mm
, pmd
, address
);
2345 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2346 page
= lookup_swap_cache(entry
);
2348 grab_swap_token(); /* Contend for token _before_ read-in */
2349 page
= swapin_readahead(entry
,
2350 GFP_HIGHUSER_MOVABLE
, vma
, address
);
2353 * Back out if somebody else faulted in this pte
2354 * while we released the pte lock.
2356 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2357 if (likely(pte_same(*page_table
, orig_pte
)))
2359 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2363 /* Had to read the page from swap area: Major fault */
2364 ret
= VM_FAULT_MAJOR
;
2365 count_vm_event(PGMAJFAULT
);
2368 mark_page_accessed(page
);
2371 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2373 if (mem_cgroup_charge(page
, mm
, GFP_KERNEL
)) {
2380 * Back out if somebody else already faulted in this pte.
2382 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2383 if (unlikely(!pte_same(*page_table
, orig_pte
)))
2386 if (unlikely(!PageUptodate(page
))) {
2387 ret
= VM_FAULT_SIGBUS
;
2391 /* The page isn't present yet, go ahead with the fault. */
2393 inc_mm_counter(mm
, anon_rss
);
2394 pte
= mk_pte(page
, vma
->vm_page_prot
);
2395 if (write_access
&& reuse_swap_page(page
)) {
2396 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
2400 flush_icache_page(vma
, page
);
2401 set_pte_at(mm
, address
, page_table
, pte
);
2402 page_add_anon_rmap(page
, vma
, address
);
2405 if (vm_swap_full() || (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
2406 try_to_free_swap(page
);
2410 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
2411 if (ret
& VM_FAULT_ERROR
)
2412 ret
&= VM_FAULT_ERROR
;
2416 /* No need to invalidate - it was non-present before */
2417 update_mmu_cache(vma
, address
, pte
);
2419 pte_unmap_unlock(page_table
, ptl
);
2423 mem_cgroup_uncharge_page(page
);
2424 pte_unmap_unlock(page_table
, ptl
);
2426 page_cache_release(page
);
2431 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2432 * but allow concurrent faults), and pte mapped but not yet locked.
2433 * We return with mmap_sem still held, but pte unmapped and unlocked.
2435 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2436 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2443 /* Allocate our own private page. */
2444 pte_unmap(page_table
);
2446 if (unlikely(anon_vma_prepare(vma
)))
2448 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2451 __SetPageUptodate(page
);
2453 if (mem_cgroup_charge(page
, mm
, GFP_KERNEL
))
2456 entry
= mk_pte(page
, vma
->vm_page_prot
);
2457 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2459 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2460 if (!pte_none(*page_table
))
2462 inc_mm_counter(mm
, anon_rss
);
2463 page_add_new_anon_rmap(page
, vma
, address
);
2464 set_pte_at(mm
, address
, page_table
, entry
);
2466 /* No need to invalidate - it was non-present before */
2467 update_mmu_cache(vma
, address
, entry
);
2469 pte_unmap_unlock(page_table
, ptl
);
2472 mem_cgroup_uncharge_page(page
);
2473 page_cache_release(page
);
2476 page_cache_release(page
);
2478 return VM_FAULT_OOM
;
2482 * __do_fault() tries to create a new page mapping. It aggressively
2483 * tries to share with existing pages, but makes a separate copy if
2484 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2485 * the next page fault.
2487 * As this is called only for pages that do not currently exist, we
2488 * do not need to flush old virtual caches or the TLB.
2490 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2491 * but allow concurrent faults), and pte neither mapped nor locked.
2492 * We return with mmap_sem still held, but pte unmapped and unlocked.
2494 static int __do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2495 unsigned long address
, pmd_t
*pmd
,
2496 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
2504 struct page
*dirty_page
= NULL
;
2505 struct vm_fault vmf
;
2507 int page_mkwrite
= 0;
2509 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
2514 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
2515 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2519 * For consistency in subsequent calls, make the faulted page always
2522 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
2523 lock_page(vmf
.page
);
2525 VM_BUG_ON(!PageLocked(vmf
.page
));
2528 * Should we do an early C-O-W break?
2531 if (flags
& FAULT_FLAG_WRITE
) {
2532 if (!(vma
->vm_flags
& VM_SHARED
)) {
2534 if (unlikely(anon_vma_prepare(vma
))) {
2538 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
,
2544 if (mem_cgroup_charge(page
, mm
, GFP_KERNEL
)) {
2546 page_cache_release(page
);
2551 * Don't let another task, with possibly unlocked vma,
2552 * keep the mlocked page.
2554 if (vma
->vm_flags
& VM_LOCKED
)
2555 clear_page_mlock(vmf
.page
);
2556 copy_user_highpage(page
, vmf
.page
, address
, vma
);
2557 __SetPageUptodate(page
);
2560 * If the page will be shareable, see if the backing
2561 * address space wants to know that the page is about
2562 * to become writable
2564 if (vma
->vm_ops
->page_mkwrite
) {
2566 if (vma
->vm_ops
->page_mkwrite(vma
, page
) < 0) {
2567 ret
= VM_FAULT_SIGBUS
;
2568 anon
= 1; /* no anon but release vmf.page */
2573 * XXX: this is not quite right (racy vs
2574 * invalidate) to unlock and relock the page
2575 * like this, however a better fix requires
2576 * reworking page_mkwrite locking API, which
2577 * is better done later.
2579 if (!page
->mapping
) {
2581 anon
= 1; /* no anon but release vmf.page */
2590 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2593 * This silly early PAGE_DIRTY setting removes a race
2594 * due to the bad i386 page protection. But it's valid
2595 * for other architectures too.
2597 * Note that if write_access is true, we either now have
2598 * an exclusive copy of the page, or this is a shared mapping,
2599 * so we can make it writable and dirty to avoid having to
2600 * handle that later.
2602 /* Only go through if we didn't race with anybody else... */
2603 if (likely(pte_same(*page_table
, orig_pte
))) {
2604 flush_icache_page(vma
, page
);
2605 entry
= mk_pte(page
, vma
->vm_page_prot
);
2606 if (flags
& FAULT_FLAG_WRITE
)
2607 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2609 inc_mm_counter(mm
, anon_rss
);
2610 page_add_new_anon_rmap(page
, vma
, address
);
2612 inc_mm_counter(mm
, file_rss
);
2613 page_add_file_rmap(page
);
2614 if (flags
& FAULT_FLAG_WRITE
) {
2616 get_page(dirty_page
);
2619 set_pte_at(mm
, address
, page_table
, entry
);
2621 /* no need to invalidate: a not-present page won't be cached */
2622 update_mmu_cache(vma
, address
, entry
);
2625 mem_cgroup_uncharge_page(page
);
2627 page_cache_release(page
);
2629 anon
= 1; /* no anon but release faulted_page */
2632 pte_unmap_unlock(page_table
, ptl
);
2635 unlock_page(vmf
.page
);
2638 page_cache_release(vmf
.page
);
2639 else if (dirty_page
) {
2641 file_update_time(vma
->vm_file
);
2643 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2644 put_page(dirty_page
);
2650 static int do_linear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2651 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2652 int write_access
, pte_t orig_pte
)
2654 pgoff_t pgoff
= (((address
& PAGE_MASK
)
2655 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
2656 unsigned int flags
= (write_access
? FAULT_FLAG_WRITE
: 0);
2658 pte_unmap(page_table
);
2659 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
2663 * Fault of a previously existing named mapping. Repopulate the pte
2664 * from the encoded file_pte if possible. This enables swappable
2667 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2668 * but allow concurrent faults), and pte mapped but not yet locked.
2669 * We return with mmap_sem still held, but pte unmapped and unlocked.
2671 static int do_nonlinear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2672 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2673 int write_access
, pte_t orig_pte
)
2675 unsigned int flags
= FAULT_FLAG_NONLINEAR
|
2676 (write_access
? FAULT_FLAG_WRITE
: 0);
2679 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2682 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
) ||
2683 !(vma
->vm_flags
& VM_CAN_NONLINEAR
))) {
2685 * Page table corrupted: show pte and kill process.
2687 print_bad_pte(vma
, orig_pte
, address
);
2688 return VM_FAULT_OOM
;
2691 pgoff
= pte_to_pgoff(orig_pte
);
2692 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
2696 * These routines also need to handle stuff like marking pages dirty
2697 * and/or accessed for architectures that don't do it in hardware (most
2698 * RISC architectures). The early dirtying is also good on the i386.
2700 * There is also a hook called "update_mmu_cache()" that architectures
2701 * with external mmu caches can use to update those (ie the Sparc or
2702 * PowerPC hashed page tables that act as extended TLBs).
2704 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2705 * but allow concurrent faults), and pte mapped but not yet locked.
2706 * We return with mmap_sem still held, but pte unmapped and unlocked.
2708 static inline int handle_pte_fault(struct mm_struct
*mm
,
2709 struct vm_area_struct
*vma
, unsigned long address
,
2710 pte_t
*pte
, pmd_t
*pmd
, int write_access
)
2716 if (!pte_present(entry
)) {
2717 if (pte_none(entry
)) {
2719 if (likely(vma
->vm_ops
->fault
))
2720 return do_linear_fault(mm
, vma
, address
,
2721 pte
, pmd
, write_access
, entry
);
2723 return do_anonymous_page(mm
, vma
, address
,
2724 pte
, pmd
, write_access
);
2726 if (pte_file(entry
))
2727 return do_nonlinear_fault(mm
, vma
, address
,
2728 pte
, pmd
, write_access
, entry
);
2729 return do_swap_page(mm
, vma
, address
,
2730 pte
, pmd
, write_access
, entry
);
2733 ptl
= pte_lockptr(mm
, pmd
);
2735 if (unlikely(!pte_same(*pte
, entry
)))
2738 if (!pte_write(entry
))
2739 return do_wp_page(mm
, vma
, address
,
2740 pte
, pmd
, ptl
, entry
);
2741 entry
= pte_mkdirty(entry
);
2743 entry
= pte_mkyoung(entry
);
2744 if (ptep_set_access_flags(vma
, address
, pte
, entry
, write_access
)) {
2745 update_mmu_cache(vma
, address
, entry
);
2748 * This is needed only for protection faults but the arch code
2749 * is not yet telling us if this is a protection fault or not.
2750 * This still avoids useless tlb flushes for .text page faults
2754 flush_tlb_page(vma
, address
);
2757 pte_unmap_unlock(pte
, ptl
);
2762 * By the time we get here, we already hold the mm semaphore
2764 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2765 unsigned long address
, int write_access
)
2772 __set_current_state(TASK_RUNNING
);
2774 count_vm_event(PGFAULT
);
2776 if (unlikely(is_vm_hugetlb_page(vma
)))
2777 return hugetlb_fault(mm
, vma
, address
, write_access
);
2779 pgd
= pgd_offset(mm
, address
);
2780 pud
= pud_alloc(mm
, pgd
, address
);
2782 return VM_FAULT_OOM
;
2783 pmd
= pmd_alloc(mm
, pud
, address
);
2785 return VM_FAULT_OOM
;
2786 pte
= pte_alloc_map(mm
, pmd
, address
);
2788 return VM_FAULT_OOM
;
2790 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, write_access
);
2793 #ifndef __PAGETABLE_PUD_FOLDED
2795 * Allocate page upper directory.
2796 * We've already handled the fast-path in-line.
2798 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
2800 pud_t
*new = pud_alloc_one(mm
, address
);
2804 smp_wmb(); /* See comment in __pte_alloc */
2806 spin_lock(&mm
->page_table_lock
);
2807 if (pgd_present(*pgd
)) /* Another has populated it */
2810 pgd_populate(mm
, pgd
, new);
2811 spin_unlock(&mm
->page_table_lock
);
2814 #endif /* __PAGETABLE_PUD_FOLDED */
2816 #ifndef __PAGETABLE_PMD_FOLDED
2818 * Allocate page middle directory.
2819 * We've already handled the fast-path in-line.
2821 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
2823 pmd_t
*new = pmd_alloc_one(mm
, address
);
2827 smp_wmb(); /* See comment in __pte_alloc */
2829 spin_lock(&mm
->page_table_lock
);
2830 #ifndef __ARCH_HAS_4LEVEL_HACK
2831 if (pud_present(*pud
)) /* Another has populated it */
2834 pud_populate(mm
, pud
, new);
2836 if (pgd_present(*pud
)) /* Another has populated it */
2839 pgd_populate(mm
, pud
, new);
2840 #endif /* __ARCH_HAS_4LEVEL_HACK */
2841 spin_unlock(&mm
->page_table_lock
);
2844 #endif /* __PAGETABLE_PMD_FOLDED */
2846 int make_pages_present(unsigned long addr
, unsigned long end
)
2848 int ret
, len
, write
;
2849 struct vm_area_struct
* vma
;
2851 vma
= find_vma(current
->mm
, addr
);
2854 write
= (vma
->vm_flags
& VM_WRITE
) != 0;
2855 BUG_ON(addr
>= end
);
2856 BUG_ON(end
> vma
->vm_end
);
2857 len
= DIV_ROUND_UP(end
, PAGE_SIZE
) - addr
/PAGE_SIZE
;
2858 ret
= get_user_pages(current
, current
->mm
, addr
,
2859 len
, write
, 0, NULL
, NULL
);
2862 return ret
== len
? 0 : -EFAULT
;
2865 #if !defined(__HAVE_ARCH_GATE_AREA)
2867 #if defined(AT_SYSINFO_EHDR)
2868 static struct vm_area_struct gate_vma
;
2870 static int __init
gate_vma_init(void)
2872 gate_vma
.vm_mm
= NULL
;
2873 gate_vma
.vm_start
= FIXADDR_USER_START
;
2874 gate_vma
.vm_end
= FIXADDR_USER_END
;
2875 gate_vma
.vm_flags
= VM_READ
| VM_MAYREAD
| VM_EXEC
| VM_MAYEXEC
;
2876 gate_vma
.vm_page_prot
= __P101
;
2878 * Make sure the vDSO gets into every core dump.
2879 * Dumping its contents makes post-mortem fully interpretable later
2880 * without matching up the same kernel and hardware config to see
2881 * what PC values meant.
2883 gate_vma
.vm_flags
|= VM_ALWAYSDUMP
;
2886 __initcall(gate_vma_init
);
2889 struct vm_area_struct
*get_gate_vma(struct task_struct
*tsk
)
2891 #ifdef AT_SYSINFO_EHDR
2898 int in_gate_area_no_task(unsigned long addr
)
2900 #ifdef AT_SYSINFO_EHDR
2901 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
2907 #endif /* __HAVE_ARCH_GATE_AREA */
2909 #ifdef CONFIG_HAVE_IOREMAP_PROT
2910 int follow_phys(struct vm_area_struct
*vma
,
2911 unsigned long address
, unsigned int flags
,
2912 unsigned long *prot
, resource_size_t
*phys
)
2919 resource_size_t phys_addr
= 0;
2920 struct mm_struct
*mm
= vma
->vm_mm
;
2923 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
2926 pgd
= pgd_offset(mm
, address
);
2927 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
2930 pud
= pud_offset(pgd
, address
);
2931 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
2934 pmd
= pmd_offset(pud
, address
);
2935 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
2938 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
2942 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2947 if (!pte_present(pte
))
2949 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
2951 phys_addr
= pte_pfn(pte
);
2952 phys_addr
<<= PAGE_SHIFT
; /* Shift here to avoid overflow on PAE */
2954 *prot
= pgprot_val(pte_pgprot(pte
));
2959 pte_unmap_unlock(ptep
, ptl
);
2964 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
2965 void *buf
, int len
, int write
)
2967 resource_size_t phys_addr
;
2968 unsigned long prot
= 0;
2969 void __iomem
*maddr
;
2970 int offset
= addr
& (PAGE_SIZE
-1);
2972 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
2975 maddr
= ioremap_prot(phys_addr
, PAGE_SIZE
, prot
);
2977 memcpy_toio(maddr
+ offset
, buf
, len
);
2979 memcpy_fromio(buf
, maddr
+ offset
, len
);
2987 * Access another process' address space.
2988 * Source/target buffer must be kernel space,
2989 * Do not walk the page table directly, use get_user_pages
2991 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
, void *buf
, int len
, int write
)
2993 struct mm_struct
*mm
;
2994 struct vm_area_struct
*vma
;
2995 void *old_buf
= buf
;
2997 mm
= get_task_mm(tsk
);
3001 down_read(&mm
->mmap_sem
);
3002 /* ignore errors, just check how much was successfully transferred */
3004 int bytes
, ret
, offset
;
3006 struct page
*page
= NULL
;
3008 ret
= get_user_pages(tsk
, mm
, addr
, 1,
3009 write
, 1, &page
, &vma
);
3012 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3013 * we can access using slightly different code.
3015 #ifdef CONFIG_HAVE_IOREMAP_PROT
3016 vma
= find_vma(mm
, addr
);
3019 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
3020 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
3028 offset
= addr
& (PAGE_SIZE
-1);
3029 if (bytes
> PAGE_SIZE
-offset
)
3030 bytes
= PAGE_SIZE
-offset
;
3034 copy_to_user_page(vma
, page
, addr
,
3035 maddr
+ offset
, buf
, bytes
);
3036 set_page_dirty_lock(page
);
3038 copy_from_user_page(vma
, page
, addr
,
3039 buf
, maddr
+ offset
, bytes
);
3042 page_cache_release(page
);
3048 up_read(&mm
->mmap_sem
);
3051 return buf
- old_buf
;
3055 * Print the name of a VMA.
3057 void print_vma_addr(char *prefix
, unsigned long ip
)
3059 struct mm_struct
*mm
= current
->mm
;
3060 struct vm_area_struct
*vma
;
3063 * Do not print if we are in atomic
3064 * contexts (in exception stacks, etc.):
3066 if (preempt_count())
3069 down_read(&mm
->mmap_sem
);
3070 vma
= find_vma(mm
, ip
);
3071 if (vma
&& vma
->vm_file
) {
3072 struct file
*f
= vma
->vm_file
;
3073 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
3077 p
= d_path(&f
->f_path
, buf
, PAGE_SIZE
);
3080 s
= strrchr(p
, '/');
3083 printk("%s%s[%lx+%lx]", prefix
, p
,
3085 vma
->vm_end
- vma
->vm_start
);
3086 free_page((unsigned long)buf
);
3089 up_read(¤t
->mm
->mmap_sem
);
3092 #ifdef CONFIG_PROVE_LOCKING
3093 void might_fault(void)
3097 * it would be nicer only to annotate paths which are not under
3098 * pagefault_disable, however that requires a larger audit and
3099 * providing helpers like get_user_atomic.
3101 if (!in_atomic() && current
->mm
)
3102 might_lock_read(¤t
->mm
->mmap_sem
);
3104 EXPORT_SYMBOL(might_fault
);