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.
1269 if (ret
& VM_FAULT_WRITE
)
1270 foll_flags
&= ~FOLL_WRITE
;
1275 return i
? i
: PTR_ERR(page
);
1279 flush_anon_page(vma
, page
, start
);
1280 flush_dcache_page(page
);
1287 } while (len
&& start
< vma
->vm_end
);
1292 int get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1293 unsigned long start
, int len
, int write
, int force
,
1294 struct page
**pages
, struct vm_area_struct
**vmas
)
1299 flags
|= GUP_FLAGS_WRITE
;
1301 flags
|= GUP_FLAGS_FORCE
;
1303 return __get_user_pages(tsk
, mm
,
1308 EXPORT_SYMBOL(get_user_pages
);
1310 pte_t
*get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1313 pgd_t
* pgd
= pgd_offset(mm
, addr
);
1314 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
1316 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
1318 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1324 * This is the old fallback for page remapping.
1326 * For historical reasons, it only allows reserved pages. Only
1327 * old drivers should use this, and they needed to mark their
1328 * pages reserved for the old functions anyway.
1330 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1331 struct page
*page
, pgprot_t prot
)
1333 struct mm_struct
*mm
= vma
->vm_mm
;
1342 flush_dcache_page(page
);
1343 pte
= get_locked_pte(mm
, addr
, &ptl
);
1347 if (!pte_none(*pte
))
1350 /* Ok, finally just insert the thing.. */
1352 inc_mm_counter(mm
, file_rss
);
1353 page_add_file_rmap(page
);
1354 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1357 pte_unmap_unlock(pte
, ptl
);
1360 pte_unmap_unlock(pte
, ptl
);
1366 * vm_insert_page - insert single page into user vma
1367 * @vma: user vma to map to
1368 * @addr: target user address of this page
1369 * @page: source kernel page
1371 * This allows drivers to insert individual pages they've allocated
1374 * The page has to be a nice clean _individual_ kernel allocation.
1375 * If you allocate a compound page, you need to have marked it as
1376 * such (__GFP_COMP), or manually just split the page up yourself
1377 * (see split_page()).
1379 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1380 * took an arbitrary page protection parameter. This doesn't allow
1381 * that. Your vma protection will have to be set up correctly, which
1382 * means that if you want a shared writable mapping, you'd better
1383 * ask for a shared writable mapping!
1385 * The page does not need to be reserved.
1387 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1390 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1392 if (!page_count(page
))
1394 vma
->vm_flags
|= VM_INSERTPAGE
;
1395 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1397 EXPORT_SYMBOL(vm_insert_page
);
1399 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1400 unsigned long pfn
, pgprot_t prot
)
1402 struct mm_struct
*mm
= vma
->vm_mm
;
1408 pte
= get_locked_pte(mm
, addr
, &ptl
);
1412 if (!pte_none(*pte
))
1415 /* Ok, finally just insert the thing.. */
1416 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
1417 set_pte_at(mm
, addr
, pte
, entry
);
1418 update_mmu_cache(vma
, addr
, entry
); /* XXX: why not for insert_page? */
1422 pte_unmap_unlock(pte
, ptl
);
1428 * vm_insert_pfn - insert single pfn into user vma
1429 * @vma: user vma to map to
1430 * @addr: target user address of this page
1431 * @pfn: source kernel pfn
1433 * Similar to vm_inert_page, this allows drivers to insert individual pages
1434 * they've allocated into a user vma. Same comments apply.
1436 * This function should only be called from a vm_ops->fault handler, and
1437 * in that case the handler should return NULL.
1439 * vma cannot be a COW mapping.
1441 * As this is called only for pages that do not currently exist, we
1442 * do not need to flush old virtual caches or the TLB.
1444 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1449 * Technically, architectures with pte_special can avoid all these
1450 * restrictions (same for remap_pfn_range). However we would like
1451 * consistency in testing and feature parity among all, so we should
1452 * try to keep these invariants in place for everybody.
1454 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1455 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1456 (VM_PFNMAP
|VM_MIXEDMAP
));
1457 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1458 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1460 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1462 if (track_pfn_vma_new(vma
, vma
->vm_page_prot
, pfn
, PAGE_SIZE
))
1465 ret
= insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1468 untrack_pfn_vma(vma
, pfn
, PAGE_SIZE
);
1472 EXPORT_SYMBOL(vm_insert_pfn
);
1474 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1477 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
1479 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1483 * If we don't have pte special, then we have to use the pfn_valid()
1484 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1485 * refcount the page if pfn_valid is true (hence insert_page rather
1488 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
1491 page
= pfn_to_page(pfn
);
1492 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1494 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1496 EXPORT_SYMBOL(vm_insert_mixed
);
1499 * maps a range of physical memory into the requested pages. the old
1500 * mappings are removed. any references to nonexistent pages results
1501 * in null mappings (currently treated as "copy-on-access")
1503 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1504 unsigned long addr
, unsigned long end
,
1505 unsigned long pfn
, pgprot_t prot
)
1510 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1513 arch_enter_lazy_mmu_mode();
1515 BUG_ON(!pte_none(*pte
));
1516 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
1518 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1519 arch_leave_lazy_mmu_mode();
1520 pte_unmap_unlock(pte
- 1, ptl
);
1524 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1525 unsigned long addr
, unsigned long end
,
1526 unsigned long pfn
, pgprot_t prot
)
1531 pfn
-= addr
>> PAGE_SHIFT
;
1532 pmd
= pmd_alloc(mm
, pud
, addr
);
1536 next
= pmd_addr_end(addr
, end
);
1537 if (remap_pte_range(mm
, pmd
, addr
, next
,
1538 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1540 } while (pmd
++, addr
= next
, addr
!= end
);
1544 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1545 unsigned long addr
, unsigned long end
,
1546 unsigned long pfn
, pgprot_t prot
)
1551 pfn
-= addr
>> PAGE_SHIFT
;
1552 pud
= pud_alloc(mm
, pgd
, addr
);
1556 next
= pud_addr_end(addr
, end
);
1557 if (remap_pmd_range(mm
, pud
, addr
, next
,
1558 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1560 } while (pud
++, addr
= next
, addr
!= end
);
1565 * remap_pfn_range - remap kernel memory to userspace
1566 * @vma: user vma to map to
1567 * @addr: target user address to start at
1568 * @pfn: physical address of kernel memory
1569 * @size: size of map area
1570 * @prot: page protection flags for this mapping
1572 * Note: this is only safe if the mm semaphore is held when called.
1574 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1575 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1579 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1580 struct mm_struct
*mm
= vma
->vm_mm
;
1584 * Physically remapped pages are special. Tell the
1585 * rest of the world about it:
1586 * VM_IO tells people not to look at these pages
1587 * (accesses can have side effects).
1588 * VM_RESERVED is specified all over the place, because
1589 * in 2.4 it kept swapout's vma scan off this vma; but
1590 * in 2.6 the LRU scan won't even find its pages, so this
1591 * flag means no more than count its pages in reserved_vm,
1592 * and omit it from core dump, even when VM_IO turned off.
1593 * VM_PFNMAP tells the core MM that the base pages are just
1594 * raw PFN mappings, and do not have a "struct page" associated
1597 * There's a horrible special case to handle copy-on-write
1598 * behaviour that some programs depend on. We mark the "original"
1599 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1601 if (addr
== vma
->vm_start
&& end
== vma
->vm_end
)
1602 vma
->vm_pgoff
= pfn
;
1603 else if (is_cow_mapping(vma
->vm_flags
))
1606 vma
->vm_flags
|= VM_IO
| VM_RESERVED
| VM_PFNMAP
;
1608 err
= track_pfn_vma_new(vma
, prot
, pfn
, PAGE_ALIGN(size
));
1612 BUG_ON(addr
>= end
);
1613 pfn
-= addr
>> PAGE_SHIFT
;
1614 pgd
= pgd_offset(mm
, addr
);
1615 flush_cache_range(vma
, addr
, end
);
1617 next
= pgd_addr_end(addr
, end
);
1618 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1619 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1622 } while (pgd
++, addr
= next
, addr
!= end
);
1625 untrack_pfn_vma(vma
, pfn
, PAGE_ALIGN(size
));
1629 EXPORT_SYMBOL(remap_pfn_range
);
1631 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1632 unsigned long addr
, unsigned long end
,
1633 pte_fn_t fn
, void *data
)
1638 spinlock_t
*uninitialized_var(ptl
);
1640 pte
= (mm
== &init_mm
) ?
1641 pte_alloc_kernel(pmd
, addr
) :
1642 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1646 BUG_ON(pmd_huge(*pmd
));
1648 arch_enter_lazy_mmu_mode();
1650 token
= pmd_pgtable(*pmd
);
1653 err
= fn(pte
, token
, addr
, data
);
1656 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1658 arch_leave_lazy_mmu_mode();
1661 pte_unmap_unlock(pte
-1, ptl
);
1665 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1666 unsigned long addr
, unsigned long end
,
1667 pte_fn_t fn
, void *data
)
1673 BUG_ON(pud_huge(*pud
));
1675 pmd
= pmd_alloc(mm
, pud
, addr
);
1679 next
= pmd_addr_end(addr
, end
);
1680 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
1683 } while (pmd
++, addr
= next
, addr
!= end
);
1687 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1688 unsigned long addr
, unsigned long end
,
1689 pte_fn_t fn
, void *data
)
1695 pud
= pud_alloc(mm
, pgd
, addr
);
1699 next
= pud_addr_end(addr
, end
);
1700 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
1703 } while (pud
++, addr
= next
, addr
!= end
);
1708 * Scan a region of virtual memory, filling in page tables as necessary
1709 * and calling a provided function on each leaf page table.
1711 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
1712 unsigned long size
, pte_fn_t fn
, void *data
)
1716 unsigned long start
= addr
, end
= addr
+ size
;
1719 BUG_ON(addr
>= end
);
1720 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1721 pgd
= pgd_offset(mm
, addr
);
1723 next
= pgd_addr_end(addr
, end
);
1724 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
1727 } while (pgd
++, addr
= next
, addr
!= end
);
1728 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1731 EXPORT_SYMBOL_GPL(apply_to_page_range
);
1734 * handle_pte_fault chooses page fault handler according to an entry
1735 * which was read non-atomically. Before making any commitment, on
1736 * those architectures or configurations (e.g. i386 with PAE) which
1737 * might give a mix of unmatched parts, do_swap_page and do_file_page
1738 * must check under lock before unmapping the pte and proceeding
1739 * (but do_wp_page is only called after already making such a check;
1740 * and do_anonymous_page and do_no_page can safely check later on).
1742 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
1743 pte_t
*page_table
, pte_t orig_pte
)
1746 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1747 if (sizeof(pte_t
) > sizeof(unsigned long)) {
1748 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
1750 same
= pte_same(*page_table
, orig_pte
);
1754 pte_unmap(page_table
);
1759 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1760 * servicing faults for write access. In the normal case, do always want
1761 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1762 * that do not have writing enabled, when used by access_process_vm.
1764 static inline pte_t
maybe_mkwrite(pte_t pte
, struct vm_area_struct
*vma
)
1766 if (likely(vma
->vm_flags
& VM_WRITE
))
1767 pte
= pte_mkwrite(pte
);
1771 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
1774 * If the source page was a PFN mapping, we don't have
1775 * a "struct page" for it. We do a best-effort copy by
1776 * just copying from the original user address. If that
1777 * fails, we just zero-fill it. Live with it.
1779 if (unlikely(!src
)) {
1780 void *kaddr
= kmap_atomic(dst
, KM_USER0
);
1781 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
1784 * This really shouldn't fail, because the page is there
1785 * in the page tables. But it might just be unreadable,
1786 * in which case we just give up and fill the result with
1789 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
1790 memset(kaddr
, 0, PAGE_SIZE
);
1791 kunmap_atomic(kaddr
, KM_USER0
);
1792 flush_dcache_page(dst
);
1794 copy_user_highpage(dst
, src
, va
, vma
);
1798 * This routine handles present pages, when users try to write
1799 * to a shared page. It is done by copying the page to a new address
1800 * and decrementing the shared-page counter for the old page.
1802 * Note that this routine assumes that the protection checks have been
1803 * done by the caller (the low-level page fault routine in most cases).
1804 * Thus we can safely just mark it writable once we've done any necessary
1807 * We also mark the page dirty at this point even though the page will
1808 * change only once the write actually happens. This avoids a few races,
1809 * and potentially makes it more efficient.
1811 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1812 * but allow concurrent faults), with pte both mapped and locked.
1813 * We return with mmap_sem still held, but pte unmapped and unlocked.
1815 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1816 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1817 spinlock_t
*ptl
, pte_t orig_pte
)
1819 struct page
*old_page
, *new_page
;
1821 int reuse
= 0, ret
= 0;
1822 int page_mkwrite
= 0;
1823 struct page
*dirty_page
= NULL
;
1825 old_page
= vm_normal_page(vma
, address
, orig_pte
);
1828 * VM_MIXEDMAP !pfn_valid() case
1830 * We should not cow pages in a shared writeable mapping.
1831 * Just mark the pages writable as we can't do any dirty
1832 * accounting on raw pfn maps.
1834 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
1835 (VM_WRITE
|VM_SHARED
))
1841 * Take out anonymous pages first, anonymous shared vmas are
1842 * not dirty accountable.
1844 if (PageAnon(old_page
)) {
1845 if (trylock_page(old_page
)) {
1846 reuse
= can_share_swap_page(old_page
);
1847 unlock_page(old_page
);
1849 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
1850 (VM_WRITE
|VM_SHARED
))) {
1852 * Only catch write-faults on shared writable pages,
1853 * read-only shared pages can get COWed by
1854 * get_user_pages(.write=1, .force=1).
1856 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
1858 * Notify the address space that the page is about to
1859 * become writable so that it can prohibit this or wait
1860 * for the page to get into an appropriate state.
1862 * We do this without the lock held, so that it can
1863 * sleep if it needs to.
1865 page_cache_get(old_page
);
1866 pte_unmap_unlock(page_table
, ptl
);
1868 if (vma
->vm_ops
->page_mkwrite(vma
, old_page
) < 0)
1869 goto unwritable_page
;
1872 * Since we dropped the lock we need to revalidate
1873 * the PTE as someone else may have changed it. If
1874 * they did, we just return, as we can count on the
1875 * MMU to tell us if they didn't also make it writable.
1877 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
1879 page_cache_release(old_page
);
1880 if (!pte_same(*page_table
, orig_pte
))
1885 dirty_page
= old_page
;
1886 get_page(dirty_page
);
1892 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
1893 entry
= pte_mkyoung(orig_pte
);
1894 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1895 if (ptep_set_access_flags(vma
, address
, page_table
, entry
,1))
1896 update_mmu_cache(vma
, address
, entry
);
1897 ret
|= VM_FAULT_WRITE
;
1902 * Ok, we need to copy. Oh, well..
1904 page_cache_get(old_page
);
1906 pte_unmap_unlock(page_table
, ptl
);
1908 if (unlikely(anon_vma_prepare(vma
)))
1910 VM_BUG_ON(old_page
== ZERO_PAGE(0));
1911 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
1915 * Don't let another task, with possibly unlocked vma,
1916 * keep the mlocked page.
1918 if (vma
->vm_flags
& VM_LOCKED
) {
1919 lock_page(old_page
); /* for LRU manipulation */
1920 clear_page_mlock(old_page
);
1921 unlock_page(old_page
);
1923 cow_user_page(new_page
, old_page
, address
, vma
);
1924 __SetPageUptodate(new_page
);
1926 if (mem_cgroup_charge(new_page
, mm
, GFP_KERNEL
))
1930 * Re-check the pte - we dropped the lock
1932 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1933 if (likely(pte_same(*page_table
, orig_pte
))) {
1935 if (!PageAnon(old_page
)) {
1936 dec_mm_counter(mm
, file_rss
);
1937 inc_mm_counter(mm
, anon_rss
);
1940 inc_mm_counter(mm
, anon_rss
);
1941 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
1942 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
1943 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1945 * Clear the pte entry and flush it first, before updating the
1946 * pte with the new entry. This will avoid a race condition
1947 * seen in the presence of one thread doing SMC and another
1950 ptep_clear_flush_notify(vma
, address
, page_table
);
1951 page_add_new_anon_rmap(new_page
, vma
, address
);
1952 set_pte_at(mm
, address
, page_table
, entry
);
1953 update_mmu_cache(vma
, address
, entry
);
1956 * Only after switching the pte to the new page may
1957 * we remove the mapcount here. Otherwise another
1958 * process may come and find the rmap count decremented
1959 * before the pte is switched to the new page, and
1960 * "reuse" the old page writing into it while our pte
1961 * here still points into it and can be read by other
1964 * The critical issue is to order this
1965 * page_remove_rmap with the ptp_clear_flush above.
1966 * Those stores are ordered by (if nothing else,)
1967 * the barrier present in the atomic_add_negative
1968 * in page_remove_rmap.
1970 * Then the TLB flush in ptep_clear_flush ensures that
1971 * no process can access the old page before the
1972 * decremented mapcount is visible. And the old page
1973 * cannot be reused until after the decremented
1974 * mapcount is visible. So transitively, TLBs to
1975 * old page will be flushed before it can be reused.
1977 page_remove_rmap(old_page
, vma
);
1980 /* Free the old page.. */
1981 new_page
= old_page
;
1982 ret
|= VM_FAULT_WRITE
;
1984 mem_cgroup_uncharge_page(new_page
);
1987 page_cache_release(new_page
);
1989 page_cache_release(old_page
);
1991 pte_unmap_unlock(page_table
, ptl
);
1994 file_update_time(vma
->vm_file
);
1997 * Yes, Virginia, this is actually required to prevent a race
1998 * with clear_page_dirty_for_io() from clearing the page dirty
1999 * bit after it clear all dirty ptes, but before a racing
2000 * do_wp_page installs a dirty pte.
2002 * do_no_page is protected similarly.
2004 wait_on_page_locked(dirty_page
);
2005 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2006 put_page(dirty_page
);
2010 page_cache_release(new_page
);
2013 page_cache_release(old_page
);
2014 return VM_FAULT_OOM
;
2017 page_cache_release(old_page
);
2018 return VM_FAULT_SIGBUS
;
2022 * Helper functions for unmap_mapping_range().
2024 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2026 * We have to restart searching the prio_tree whenever we drop the lock,
2027 * since the iterator is only valid while the lock is held, and anyway
2028 * a later vma might be split and reinserted earlier while lock dropped.
2030 * The list of nonlinear vmas could be handled more efficiently, using
2031 * a placeholder, but handle it in the same way until a need is shown.
2032 * It is important to search the prio_tree before nonlinear list: a vma
2033 * may become nonlinear and be shifted from prio_tree to nonlinear list
2034 * while the lock is dropped; but never shifted from list to prio_tree.
2036 * In order to make forward progress despite restarting the search,
2037 * vm_truncate_count is used to mark a vma as now dealt with, so we can
2038 * quickly skip it next time around. Since the prio_tree search only
2039 * shows us those vmas affected by unmapping the range in question, we
2040 * can't efficiently keep all vmas in step with mapping->truncate_count:
2041 * so instead reset them all whenever it wraps back to 0 (then go to 1).
2042 * mapping->truncate_count and vma->vm_truncate_count are protected by
2045 * In order to make forward progress despite repeatedly restarting some
2046 * large vma, note the restart_addr from unmap_vmas when it breaks out:
2047 * and restart from that address when we reach that vma again. It might
2048 * have been split or merged, shrunk or extended, but never shifted: so
2049 * restart_addr remains valid so long as it remains in the vma's range.
2050 * unmap_mapping_range forces truncate_count to leap over page-aligned
2051 * values so we can save vma's restart_addr in its truncate_count field.
2053 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2055 static void reset_vma_truncate_counts(struct address_space
*mapping
)
2057 struct vm_area_struct
*vma
;
2058 struct prio_tree_iter iter
;
2060 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, 0, ULONG_MAX
)
2061 vma
->vm_truncate_count
= 0;
2062 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
2063 vma
->vm_truncate_count
= 0;
2066 static int unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2067 unsigned long start_addr
, unsigned long end_addr
,
2068 struct zap_details
*details
)
2070 unsigned long restart_addr
;
2074 * files that support invalidating or truncating portions of the
2075 * file from under mmaped areas must have their ->fault function
2076 * return a locked page (and set VM_FAULT_LOCKED in the return).
2077 * This provides synchronisation against concurrent unmapping here.
2081 restart_addr
= vma
->vm_truncate_count
;
2082 if (is_restart_addr(restart_addr
) && start_addr
< restart_addr
) {
2083 start_addr
= restart_addr
;
2084 if (start_addr
>= end_addr
) {
2085 /* Top of vma has been split off since last time */
2086 vma
->vm_truncate_count
= details
->truncate_count
;
2091 restart_addr
= zap_page_range(vma
, start_addr
,
2092 end_addr
- start_addr
, details
);
2093 need_break
= need_resched() || spin_needbreak(details
->i_mmap_lock
);
2095 if (restart_addr
>= end_addr
) {
2096 /* We have now completed this vma: mark it so */
2097 vma
->vm_truncate_count
= details
->truncate_count
;
2101 /* Note restart_addr in vma's truncate_count field */
2102 vma
->vm_truncate_count
= restart_addr
;
2107 spin_unlock(details
->i_mmap_lock
);
2109 spin_lock(details
->i_mmap_lock
);
2113 static inline void unmap_mapping_range_tree(struct prio_tree_root
*root
,
2114 struct zap_details
*details
)
2116 struct vm_area_struct
*vma
;
2117 struct prio_tree_iter iter
;
2118 pgoff_t vba
, vea
, zba
, zea
;
2121 vma_prio_tree_foreach(vma
, &iter
, root
,
2122 details
->first_index
, details
->last_index
) {
2123 /* Skip quickly over those we have already dealt with */
2124 if (vma
->vm_truncate_count
== details
->truncate_count
)
2127 vba
= vma
->vm_pgoff
;
2128 vea
= vba
+ ((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) - 1;
2129 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2130 zba
= details
->first_index
;
2133 zea
= details
->last_index
;
2137 if (unmap_mapping_range_vma(vma
,
2138 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2139 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2145 static inline void unmap_mapping_range_list(struct list_head
*head
,
2146 struct zap_details
*details
)
2148 struct vm_area_struct
*vma
;
2151 * In nonlinear VMAs there is no correspondence between virtual address
2152 * offset and file offset. So we must perform an exhaustive search
2153 * across *all* the pages in each nonlinear VMA, not just the pages
2154 * whose virtual address lies outside the file truncation point.
2157 list_for_each_entry(vma
, head
, shared
.vm_set
.list
) {
2158 /* Skip quickly over those we have already dealt with */
2159 if (vma
->vm_truncate_count
== details
->truncate_count
)
2161 details
->nonlinear_vma
= vma
;
2162 if (unmap_mapping_range_vma(vma
, vma
->vm_start
,
2163 vma
->vm_end
, details
) < 0)
2169 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2170 * @mapping: the address space containing mmaps to be unmapped.
2171 * @holebegin: byte in first page to unmap, relative to the start of
2172 * the underlying file. This will be rounded down to a PAGE_SIZE
2173 * boundary. Note that this is different from vmtruncate(), which
2174 * must keep the partial page. In contrast, we must get rid of
2176 * @holelen: size of prospective hole in bytes. This will be rounded
2177 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2179 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2180 * but 0 when invalidating pagecache, don't throw away private data.
2182 void unmap_mapping_range(struct address_space
*mapping
,
2183 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2185 struct zap_details details
;
2186 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2187 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2189 /* Check for overflow. */
2190 if (sizeof(holelen
) > sizeof(hlen
)) {
2192 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2193 if (holeend
& ~(long long)ULONG_MAX
)
2194 hlen
= ULONG_MAX
- hba
+ 1;
2197 details
.check_mapping
= even_cows
? NULL
: mapping
;
2198 details
.nonlinear_vma
= NULL
;
2199 details
.first_index
= hba
;
2200 details
.last_index
= hba
+ hlen
- 1;
2201 if (details
.last_index
< details
.first_index
)
2202 details
.last_index
= ULONG_MAX
;
2203 details
.i_mmap_lock
= &mapping
->i_mmap_lock
;
2205 spin_lock(&mapping
->i_mmap_lock
);
2207 /* Protect against endless unmapping loops */
2208 mapping
->truncate_count
++;
2209 if (unlikely(is_restart_addr(mapping
->truncate_count
))) {
2210 if (mapping
->truncate_count
== 0)
2211 reset_vma_truncate_counts(mapping
);
2212 mapping
->truncate_count
++;
2214 details
.truncate_count
= mapping
->truncate_count
;
2216 if (unlikely(!prio_tree_empty(&mapping
->i_mmap
)))
2217 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2218 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
2219 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
2220 spin_unlock(&mapping
->i_mmap_lock
);
2222 EXPORT_SYMBOL(unmap_mapping_range
);
2225 * vmtruncate - unmap mappings "freed" by truncate() syscall
2226 * @inode: inode of the file used
2227 * @offset: file offset to start truncating
2229 * NOTE! We have to be ready to update the memory sharing
2230 * between the file and the memory map for a potential last
2231 * incomplete page. Ugly, but necessary.
2233 int vmtruncate(struct inode
* inode
, loff_t offset
)
2235 if (inode
->i_size
< offset
) {
2236 unsigned long limit
;
2238 limit
= current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
;
2239 if (limit
!= RLIM_INFINITY
&& offset
> limit
)
2241 if (offset
> inode
->i_sb
->s_maxbytes
)
2243 i_size_write(inode
, offset
);
2245 struct address_space
*mapping
= inode
->i_mapping
;
2248 * truncation of in-use swapfiles is disallowed - it would
2249 * cause subsequent swapout to scribble on the now-freed
2252 if (IS_SWAPFILE(inode
))
2254 i_size_write(inode
, offset
);
2257 * unmap_mapping_range is called twice, first simply for
2258 * efficiency so that truncate_inode_pages does fewer
2259 * single-page unmaps. However after this first call, and
2260 * before truncate_inode_pages finishes, it is possible for
2261 * private pages to be COWed, which remain after
2262 * truncate_inode_pages finishes, hence the second
2263 * unmap_mapping_range call must be made for correctness.
2265 unmap_mapping_range(mapping
, offset
+ PAGE_SIZE
- 1, 0, 1);
2266 truncate_inode_pages(mapping
, offset
);
2267 unmap_mapping_range(mapping
, offset
+ PAGE_SIZE
- 1, 0, 1);
2270 if (inode
->i_op
->truncate
)
2271 inode
->i_op
->truncate(inode
);
2275 send_sig(SIGXFSZ
, current
, 0);
2279 EXPORT_SYMBOL(vmtruncate
);
2281 int vmtruncate_range(struct inode
*inode
, loff_t offset
, loff_t end
)
2283 struct address_space
*mapping
= inode
->i_mapping
;
2286 * If the underlying filesystem is not going to provide
2287 * a way to truncate a range of blocks (punch a hole) -
2288 * we should return failure right now.
2290 if (!inode
->i_op
->truncate_range
)
2293 mutex_lock(&inode
->i_mutex
);
2294 down_write(&inode
->i_alloc_sem
);
2295 unmap_mapping_range(mapping
, offset
, (end
- offset
), 1);
2296 truncate_inode_pages_range(mapping
, offset
, end
);
2297 unmap_mapping_range(mapping
, offset
, (end
- offset
), 1);
2298 inode
->i_op
->truncate_range(inode
, offset
, end
);
2299 up_write(&inode
->i_alloc_sem
);
2300 mutex_unlock(&inode
->i_mutex
);
2306 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2307 * but allow concurrent faults), and pte mapped but not yet locked.
2308 * We return with mmap_sem still held, but pte unmapped and unlocked.
2310 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2311 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2312 int write_access
, pte_t orig_pte
)
2320 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2323 entry
= pte_to_swp_entry(orig_pte
);
2324 if (is_migration_entry(entry
)) {
2325 migration_entry_wait(mm
, pmd
, address
);
2328 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2329 page
= lookup_swap_cache(entry
);
2331 grab_swap_token(); /* Contend for token _before_ read-in */
2332 page
= swapin_readahead(entry
,
2333 GFP_HIGHUSER_MOVABLE
, vma
, address
);
2336 * Back out if somebody else faulted in this pte
2337 * while we released the pte lock.
2339 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2340 if (likely(pte_same(*page_table
, orig_pte
)))
2342 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2346 /* Had to read the page from swap area: Major fault */
2347 ret
= VM_FAULT_MAJOR
;
2348 count_vm_event(PGMAJFAULT
);
2351 mark_page_accessed(page
);
2354 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2356 if (mem_cgroup_charge(page
, mm
, GFP_KERNEL
)) {
2363 * Back out if somebody else already faulted in this pte.
2365 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2366 if (unlikely(!pte_same(*page_table
, orig_pte
)))
2369 if (unlikely(!PageUptodate(page
))) {
2370 ret
= VM_FAULT_SIGBUS
;
2374 /* The page isn't present yet, go ahead with the fault. */
2376 inc_mm_counter(mm
, anon_rss
);
2377 pte
= mk_pte(page
, vma
->vm_page_prot
);
2378 if (write_access
&& can_share_swap_page(page
)) {
2379 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
2383 flush_icache_page(vma
, page
);
2384 set_pte_at(mm
, address
, page_table
, pte
);
2385 page_add_anon_rmap(page
, vma
, address
);
2388 if (vm_swap_full() || (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
2389 remove_exclusive_swap_page(page
);
2393 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
2394 if (ret
& VM_FAULT_ERROR
)
2395 ret
&= VM_FAULT_ERROR
;
2399 /* No need to invalidate - it was non-present before */
2400 update_mmu_cache(vma
, address
, pte
);
2402 pte_unmap_unlock(page_table
, ptl
);
2406 mem_cgroup_uncharge_page(page
);
2407 pte_unmap_unlock(page_table
, ptl
);
2409 page_cache_release(page
);
2414 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2415 * but allow concurrent faults), and pte mapped but not yet locked.
2416 * We return with mmap_sem still held, but pte unmapped and unlocked.
2418 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2419 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2426 /* Allocate our own private page. */
2427 pte_unmap(page_table
);
2429 if (unlikely(anon_vma_prepare(vma
)))
2431 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2434 __SetPageUptodate(page
);
2436 if (mem_cgroup_charge(page
, mm
, GFP_KERNEL
))
2439 entry
= mk_pte(page
, vma
->vm_page_prot
);
2440 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2442 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2443 if (!pte_none(*page_table
))
2445 inc_mm_counter(mm
, anon_rss
);
2446 page_add_new_anon_rmap(page
, vma
, address
);
2447 set_pte_at(mm
, address
, page_table
, entry
);
2449 /* No need to invalidate - it was non-present before */
2450 update_mmu_cache(vma
, address
, entry
);
2452 pte_unmap_unlock(page_table
, ptl
);
2455 mem_cgroup_uncharge_page(page
);
2456 page_cache_release(page
);
2459 page_cache_release(page
);
2461 return VM_FAULT_OOM
;
2465 * __do_fault() tries to create a new page mapping. It aggressively
2466 * tries to share with existing pages, but makes a separate copy if
2467 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2468 * the next page fault.
2470 * As this is called only for pages that do not currently exist, we
2471 * do not need to flush old virtual caches or the TLB.
2473 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2474 * but allow concurrent faults), and pte neither mapped nor locked.
2475 * We return with mmap_sem still held, but pte unmapped and unlocked.
2477 static int __do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2478 unsigned long address
, pmd_t
*pmd
,
2479 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
2487 struct page
*dirty_page
= NULL
;
2488 struct vm_fault vmf
;
2490 int page_mkwrite
= 0;
2492 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
2497 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
2498 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2502 * For consistency in subsequent calls, make the faulted page always
2505 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
2506 lock_page(vmf
.page
);
2508 VM_BUG_ON(!PageLocked(vmf
.page
));
2511 * Should we do an early C-O-W break?
2514 if (flags
& FAULT_FLAG_WRITE
) {
2515 if (!(vma
->vm_flags
& VM_SHARED
)) {
2517 if (unlikely(anon_vma_prepare(vma
))) {
2521 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
,
2527 if (mem_cgroup_charge(page
, mm
, GFP_KERNEL
)) {
2529 page_cache_release(page
);
2534 * Don't let another task, with possibly unlocked vma,
2535 * keep the mlocked page.
2537 if (vma
->vm_flags
& VM_LOCKED
)
2538 clear_page_mlock(vmf
.page
);
2539 copy_user_highpage(page
, vmf
.page
, address
, vma
);
2540 __SetPageUptodate(page
);
2543 * If the page will be shareable, see if the backing
2544 * address space wants to know that the page is about
2545 * to become writable
2547 if (vma
->vm_ops
->page_mkwrite
) {
2549 if (vma
->vm_ops
->page_mkwrite(vma
, page
) < 0) {
2550 ret
= VM_FAULT_SIGBUS
;
2551 anon
= 1; /* no anon but release vmf.page */
2556 * XXX: this is not quite right (racy vs
2557 * invalidate) to unlock and relock the page
2558 * like this, however a better fix requires
2559 * reworking page_mkwrite locking API, which
2560 * is better done later.
2562 if (!page
->mapping
) {
2564 anon
= 1; /* no anon but release vmf.page */
2573 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2576 * This silly early PAGE_DIRTY setting removes a race
2577 * due to the bad i386 page protection. But it's valid
2578 * for other architectures too.
2580 * Note that if write_access is true, we either now have
2581 * an exclusive copy of the page, or this is a shared mapping,
2582 * so we can make it writable and dirty to avoid having to
2583 * handle that later.
2585 /* Only go through if we didn't race with anybody else... */
2586 if (likely(pte_same(*page_table
, orig_pte
))) {
2587 flush_icache_page(vma
, page
);
2588 entry
= mk_pte(page
, vma
->vm_page_prot
);
2589 if (flags
& FAULT_FLAG_WRITE
)
2590 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2592 inc_mm_counter(mm
, anon_rss
);
2593 page_add_new_anon_rmap(page
, vma
, address
);
2595 inc_mm_counter(mm
, file_rss
);
2596 page_add_file_rmap(page
);
2597 if (flags
& FAULT_FLAG_WRITE
) {
2599 get_page(dirty_page
);
2602 set_pte_at(mm
, address
, page_table
, entry
);
2604 /* no need to invalidate: a not-present page won't be cached */
2605 update_mmu_cache(vma
, address
, entry
);
2608 mem_cgroup_uncharge_page(page
);
2610 page_cache_release(page
);
2612 anon
= 1; /* no anon but release faulted_page */
2615 pte_unmap_unlock(page_table
, ptl
);
2618 unlock_page(vmf
.page
);
2621 page_cache_release(vmf
.page
);
2622 else if (dirty_page
) {
2624 file_update_time(vma
->vm_file
);
2626 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2627 put_page(dirty_page
);
2633 static int do_linear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2634 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2635 int write_access
, pte_t orig_pte
)
2637 pgoff_t pgoff
= (((address
& PAGE_MASK
)
2638 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
2639 unsigned int flags
= (write_access
? FAULT_FLAG_WRITE
: 0);
2641 pte_unmap(page_table
);
2642 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
2646 * Fault of a previously existing named mapping. Repopulate the pte
2647 * from the encoded file_pte if possible. This enables swappable
2650 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2651 * but allow concurrent faults), and pte mapped but not yet locked.
2652 * We return with mmap_sem still held, but pte unmapped and unlocked.
2654 static int do_nonlinear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2655 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2656 int write_access
, pte_t orig_pte
)
2658 unsigned int flags
= FAULT_FLAG_NONLINEAR
|
2659 (write_access
? FAULT_FLAG_WRITE
: 0);
2662 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2665 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
) ||
2666 !(vma
->vm_flags
& VM_CAN_NONLINEAR
))) {
2668 * Page table corrupted: show pte and kill process.
2670 print_bad_pte(vma
, orig_pte
, address
);
2671 return VM_FAULT_OOM
;
2674 pgoff
= pte_to_pgoff(orig_pte
);
2675 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
2679 * These routines also need to handle stuff like marking pages dirty
2680 * and/or accessed for architectures that don't do it in hardware (most
2681 * RISC architectures). The early dirtying is also good on the i386.
2683 * There is also a hook called "update_mmu_cache()" that architectures
2684 * with external mmu caches can use to update those (ie the Sparc or
2685 * PowerPC hashed page tables that act as extended TLBs).
2687 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2688 * but allow concurrent faults), and pte mapped but not yet locked.
2689 * We return with mmap_sem still held, but pte unmapped and unlocked.
2691 static inline int handle_pte_fault(struct mm_struct
*mm
,
2692 struct vm_area_struct
*vma
, unsigned long address
,
2693 pte_t
*pte
, pmd_t
*pmd
, int write_access
)
2699 if (!pte_present(entry
)) {
2700 if (pte_none(entry
)) {
2702 if (likely(vma
->vm_ops
->fault
))
2703 return do_linear_fault(mm
, vma
, address
,
2704 pte
, pmd
, write_access
, entry
);
2706 return do_anonymous_page(mm
, vma
, address
,
2707 pte
, pmd
, write_access
);
2709 if (pte_file(entry
))
2710 return do_nonlinear_fault(mm
, vma
, address
,
2711 pte
, pmd
, write_access
, entry
);
2712 return do_swap_page(mm
, vma
, address
,
2713 pte
, pmd
, write_access
, entry
);
2716 ptl
= pte_lockptr(mm
, pmd
);
2718 if (unlikely(!pte_same(*pte
, entry
)))
2721 if (!pte_write(entry
))
2722 return do_wp_page(mm
, vma
, address
,
2723 pte
, pmd
, ptl
, entry
);
2724 entry
= pte_mkdirty(entry
);
2726 entry
= pte_mkyoung(entry
);
2727 if (ptep_set_access_flags(vma
, address
, pte
, entry
, write_access
)) {
2728 update_mmu_cache(vma
, address
, entry
);
2731 * This is needed only for protection faults but the arch code
2732 * is not yet telling us if this is a protection fault or not.
2733 * This still avoids useless tlb flushes for .text page faults
2737 flush_tlb_page(vma
, address
);
2740 pte_unmap_unlock(pte
, ptl
);
2745 * By the time we get here, we already hold the mm semaphore
2747 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2748 unsigned long address
, int write_access
)
2755 __set_current_state(TASK_RUNNING
);
2757 count_vm_event(PGFAULT
);
2759 if (unlikely(is_vm_hugetlb_page(vma
)))
2760 return hugetlb_fault(mm
, vma
, address
, write_access
);
2762 pgd
= pgd_offset(mm
, address
);
2763 pud
= pud_alloc(mm
, pgd
, address
);
2765 return VM_FAULT_OOM
;
2766 pmd
= pmd_alloc(mm
, pud
, address
);
2768 return VM_FAULT_OOM
;
2769 pte
= pte_alloc_map(mm
, pmd
, address
);
2771 return VM_FAULT_OOM
;
2773 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, write_access
);
2776 #ifndef __PAGETABLE_PUD_FOLDED
2778 * Allocate page upper directory.
2779 * We've already handled the fast-path in-line.
2781 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
2783 pud_t
*new = pud_alloc_one(mm
, address
);
2787 smp_wmb(); /* See comment in __pte_alloc */
2789 spin_lock(&mm
->page_table_lock
);
2790 if (pgd_present(*pgd
)) /* Another has populated it */
2793 pgd_populate(mm
, pgd
, new);
2794 spin_unlock(&mm
->page_table_lock
);
2797 #endif /* __PAGETABLE_PUD_FOLDED */
2799 #ifndef __PAGETABLE_PMD_FOLDED
2801 * Allocate page middle directory.
2802 * We've already handled the fast-path in-line.
2804 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
2806 pmd_t
*new = pmd_alloc_one(mm
, address
);
2810 smp_wmb(); /* See comment in __pte_alloc */
2812 spin_lock(&mm
->page_table_lock
);
2813 #ifndef __ARCH_HAS_4LEVEL_HACK
2814 if (pud_present(*pud
)) /* Another has populated it */
2817 pud_populate(mm
, pud
, new);
2819 if (pgd_present(*pud
)) /* Another has populated it */
2822 pgd_populate(mm
, pud
, new);
2823 #endif /* __ARCH_HAS_4LEVEL_HACK */
2824 spin_unlock(&mm
->page_table_lock
);
2827 #endif /* __PAGETABLE_PMD_FOLDED */
2829 int make_pages_present(unsigned long addr
, unsigned long end
)
2831 int ret
, len
, write
;
2832 struct vm_area_struct
* vma
;
2834 vma
= find_vma(current
->mm
, addr
);
2837 write
= (vma
->vm_flags
& VM_WRITE
) != 0;
2838 BUG_ON(addr
>= end
);
2839 BUG_ON(end
> vma
->vm_end
);
2840 len
= DIV_ROUND_UP(end
, PAGE_SIZE
) - addr
/PAGE_SIZE
;
2841 ret
= get_user_pages(current
, current
->mm
, addr
,
2842 len
, write
, 0, NULL
, NULL
);
2845 return ret
== len
? 0 : -EFAULT
;
2848 #if !defined(__HAVE_ARCH_GATE_AREA)
2850 #if defined(AT_SYSINFO_EHDR)
2851 static struct vm_area_struct gate_vma
;
2853 static int __init
gate_vma_init(void)
2855 gate_vma
.vm_mm
= NULL
;
2856 gate_vma
.vm_start
= FIXADDR_USER_START
;
2857 gate_vma
.vm_end
= FIXADDR_USER_END
;
2858 gate_vma
.vm_flags
= VM_READ
| VM_MAYREAD
| VM_EXEC
| VM_MAYEXEC
;
2859 gate_vma
.vm_page_prot
= __P101
;
2861 * Make sure the vDSO gets into every core dump.
2862 * Dumping its contents makes post-mortem fully interpretable later
2863 * without matching up the same kernel and hardware config to see
2864 * what PC values meant.
2866 gate_vma
.vm_flags
|= VM_ALWAYSDUMP
;
2869 __initcall(gate_vma_init
);
2872 struct vm_area_struct
*get_gate_vma(struct task_struct
*tsk
)
2874 #ifdef AT_SYSINFO_EHDR
2881 int in_gate_area_no_task(unsigned long addr
)
2883 #ifdef AT_SYSINFO_EHDR
2884 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
2890 #endif /* __HAVE_ARCH_GATE_AREA */
2892 #ifdef CONFIG_HAVE_IOREMAP_PROT
2893 int follow_phys(struct vm_area_struct
*vma
,
2894 unsigned long address
, unsigned int flags
,
2895 unsigned long *prot
, resource_size_t
*phys
)
2902 resource_size_t phys_addr
= 0;
2903 struct mm_struct
*mm
= vma
->vm_mm
;
2906 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
2909 pgd
= pgd_offset(mm
, address
);
2910 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
2913 pud
= pud_offset(pgd
, address
);
2914 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
2917 pmd
= pmd_offset(pud
, address
);
2918 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
2921 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
2925 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2930 if (!pte_present(pte
))
2932 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
2934 phys_addr
= pte_pfn(pte
);
2935 phys_addr
<<= PAGE_SHIFT
; /* Shift here to avoid overflow on PAE */
2937 *prot
= pgprot_val(pte_pgprot(pte
));
2942 pte_unmap_unlock(ptep
, ptl
);
2947 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
2948 void *buf
, int len
, int write
)
2950 resource_size_t phys_addr
;
2951 unsigned long prot
= 0;
2953 int offset
= addr
& (PAGE_SIZE
-1);
2955 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
2958 maddr
= ioremap_prot(phys_addr
, PAGE_SIZE
, prot
);
2960 memcpy_toio(maddr
+ offset
, buf
, len
);
2962 memcpy_fromio(buf
, maddr
+ offset
, len
);
2970 * Access another process' address space.
2971 * Source/target buffer must be kernel space,
2972 * Do not walk the page table directly, use get_user_pages
2974 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
, void *buf
, int len
, int write
)
2976 struct mm_struct
*mm
;
2977 struct vm_area_struct
*vma
;
2978 void *old_buf
= buf
;
2980 mm
= get_task_mm(tsk
);
2984 down_read(&mm
->mmap_sem
);
2985 /* ignore errors, just check how much was successfully transferred */
2987 int bytes
, ret
, offset
;
2989 struct page
*page
= NULL
;
2991 ret
= get_user_pages(tsk
, mm
, addr
, 1,
2992 write
, 1, &page
, &vma
);
2995 * Check if this is a VM_IO | VM_PFNMAP VMA, which
2996 * we can access using slightly different code.
2998 #ifdef CONFIG_HAVE_IOREMAP_PROT
2999 vma
= find_vma(mm
, addr
);
3002 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
3003 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
3011 offset
= addr
& (PAGE_SIZE
-1);
3012 if (bytes
> PAGE_SIZE
-offset
)
3013 bytes
= PAGE_SIZE
-offset
;
3017 copy_to_user_page(vma
, page
, addr
,
3018 maddr
+ offset
, buf
, bytes
);
3019 set_page_dirty_lock(page
);
3021 copy_from_user_page(vma
, page
, addr
,
3022 buf
, maddr
+ offset
, bytes
);
3025 page_cache_release(page
);
3031 up_read(&mm
->mmap_sem
);
3034 return buf
- old_buf
;
3038 * Print the name of a VMA.
3040 void print_vma_addr(char *prefix
, unsigned long ip
)
3042 struct mm_struct
*mm
= current
->mm
;
3043 struct vm_area_struct
*vma
;
3046 * Do not print if we are in atomic
3047 * contexts (in exception stacks, etc.):
3049 if (preempt_count())
3052 down_read(&mm
->mmap_sem
);
3053 vma
= find_vma(mm
, ip
);
3054 if (vma
&& vma
->vm_file
) {
3055 struct file
*f
= vma
->vm_file
;
3056 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
3060 p
= d_path(&f
->f_path
, buf
, PAGE_SIZE
);
3063 s
= strrchr(p
, '/');
3066 printk("%s%s[%lx+%lx]", prefix
, p
,
3068 vma
->vm_end
- vma
->vm_start
);
3069 free_page((unsigned long)buf
);
3072 up_read(¤t
->mm
->mmap_sem
);
3075 #ifdef CONFIG_PROVE_LOCKING
3076 void might_fault(void)
3080 * it would be nicer only to annotate paths which are not under
3081 * pagefault_disable, however that requires a larger audit and
3082 * providing helpers like get_user_atomic.
3084 if (!in_atomic() && current
->mm
)
3085 might_lock_read(¤t
->mm
->mmap_sem
);
3087 EXPORT_SYMBOL(might_fault
);