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>
55 #include <asm/pgalloc.h>
56 #include <asm/uaccess.h>
58 #include <asm/tlbflush.h>
59 #include <asm/pgtable.h>
61 #include <linux/swapops.h>
62 #include <linux/elf.h>
66 #ifndef CONFIG_NEED_MULTIPLE_NODES
67 /* use the per-pgdat data instead for discontigmem - mbligh */
68 unsigned long max_mapnr
;
71 EXPORT_SYMBOL(max_mapnr
);
72 EXPORT_SYMBOL(mem_map
);
75 unsigned long num_physpages
;
77 * A number of key systems in x86 including ioremap() rely on the assumption
78 * that high_memory defines the upper bound on direct map memory, then end
79 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
80 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
85 EXPORT_SYMBOL(num_physpages
);
86 EXPORT_SYMBOL(high_memory
);
89 * Randomize the address space (stacks, mmaps, brk, etc.).
91 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
92 * as ancient (libc5 based) binaries can segfault. )
94 int randomize_va_space __read_mostly
=
95 #ifdef CONFIG_COMPAT_BRK
101 static int __init
disable_randmaps(char *s
)
103 randomize_va_space
= 0;
106 __setup("norandmaps", disable_randmaps
);
110 * If a p?d_bad entry is found while walking page tables, report
111 * the error, before resetting entry to p?d_none. Usually (but
112 * very seldom) called out from the p?d_none_or_clear_bad macros.
115 void pgd_clear_bad(pgd_t
*pgd
)
121 void pud_clear_bad(pud_t
*pud
)
127 void pmd_clear_bad(pmd_t
*pmd
)
134 * Note: this doesn't free the actual pages themselves. That
135 * has been handled earlier when unmapping all the memory regions.
137 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
)
139 pgtable_t token
= pmd_pgtable(*pmd
);
141 pte_free_tlb(tlb
, token
);
145 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
146 unsigned long addr
, unsigned long end
,
147 unsigned long floor
, unsigned long ceiling
)
154 pmd
= pmd_offset(pud
, addr
);
156 next
= pmd_addr_end(addr
, end
);
157 if (pmd_none_or_clear_bad(pmd
))
159 free_pte_range(tlb
, pmd
);
160 } while (pmd
++, addr
= next
, addr
!= end
);
170 if (end
- 1 > ceiling
- 1)
173 pmd
= pmd_offset(pud
, start
);
175 pmd_free_tlb(tlb
, pmd
);
178 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
179 unsigned long addr
, unsigned long end
,
180 unsigned long floor
, unsigned long ceiling
)
187 pud
= pud_offset(pgd
, addr
);
189 next
= pud_addr_end(addr
, end
);
190 if (pud_none_or_clear_bad(pud
))
192 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
193 } while (pud
++, addr
= next
, addr
!= end
);
199 ceiling
&= PGDIR_MASK
;
203 if (end
- 1 > ceiling
- 1)
206 pud
= pud_offset(pgd
, start
);
208 pud_free_tlb(tlb
, pud
);
212 * This function frees user-level page tables of a process.
214 * Must be called with pagetable lock held.
216 void free_pgd_range(struct mmu_gather
*tlb
,
217 unsigned long addr
, unsigned long end
,
218 unsigned long floor
, unsigned long ceiling
)
225 * The next few lines have given us lots of grief...
227 * Why are we testing PMD* at this top level? Because often
228 * there will be no work to do at all, and we'd prefer not to
229 * go all the way down to the bottom just to discover that.
231 * Why all these "- 1"s? Because 0 represents both the bottom
232 * of the address space and the top of it (using -1 for the
233 * top wouldn't help much: the masks would do the wrong thing).
234 * The rule is that addr 0 and floor 0 refer to the bottom of
235 * the address space, but end 0 and ceiling 0 refer to the top
236 * Comparisons need to use "end - 1" and "ceiling - 1" (though
237 * that end 0 case should be mythical).
239 * Wherever addr is brought up or ceiling brought down, we must
240 * be careful to reject "the opposite 0" before it confuses the
241 * subsequent tests. But what about where end is brought down
242 * by PMD_SIZE below? no, end can't go down to 0 there.
244 * Whereas we round start (addr) and ceiling down, by different
245 * masks at different levels, in order to test whether a table
246 * now has no other vmas using it, so can be freed, we don't
247 * bother to round floor or end up - the tests don't need that.
261 if (end
- 1 > ceiling
- 1)
267 pgd
= pgd_offset(tlb
->mm
, addr
);
269 next
= pgd_addr_end(addr
, end
);
270 if (pgd_none_or_clear_bad(pgd
))
272 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
273 } while (pgd
++, addr
= next
, addr
!= end
);
276 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
277 unsigned long floor
, unsigned long ceiling
)
280 struct vm_area_struct
*next
= vma
->vm_next
;
281 unsigned long addr
= vma
->vm_start
;
284 * Hide vma from rmap and vmtruncate before freeing pgtables
286 anon_vma_unlink(vma
);
287 unlink_file_vma(vma
);
289 if (is_vm_hugetlb_page(vma
)) {
290 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
291 floor
, next
? next
->vm_start
: ceiling
);
294 * Optimization: gather nearby vmas into one call down
296 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
297 && !is_vm_hugetlb_page(next
)) {
300 anon_vma_unlink(vma
);
301 unlink_file_vma(vma
);
303 free_pgd_range(tlb
, addr
, vma
->vm_end
,
304 floor
, next
? next
->vm_start
: ceiling
);
310 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
, unsigned long address
)
312 pgtable_t
new = pte_alloc_one(mm
, address
);
317 * Ensure all pte setup (eg. pte page lock and page clearing) are
318 * visible before the pte is made visible to other CPUs by being
319 * put into page tables.
321 * The other side of the story is the pointer chasing in the page
322 * table walking code (when walking the page table without locking;
323 * ie. most of the time). Fortunately, these data accesses consist
324 * of a chain of data-dependent loads, meaning most CPUs (alpha
325 * being the notable exception) will already guarantee loads are
326 * seen in-order. See the alpha page table accessors for the
327 * smp_read_barrier_depends() barriers in page table walking code.
329 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
331 spin_lock(&mm
->page_table_lock
);
332 if (!pmd_present(*pmd
)) { /* Has another populated it ? */
334 pmd_populate(mm
, pmd
, new);
337 spin_unlock(&mm
->page_table_lock
);
343 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
345 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
349 smp_wmb(); /* See comment in __pte_alloc */
351 spin_lock(&init_mm
.page_table_lock
);
352 if (!pmd_present(*pmd
)) { /* Has another populated it ? */
353 pmd_populate_kernel(&init_mm
, pmd
, new);
356 spin_unlock(&init_mm
.page_table_lock
);
358 pte_free_kernel(&init_mm
, new);
362 static inline void add_mm_rss(struct mm_struct
*mm
, int file_rss
, int anon_rss
)
365 add_mm_counter(mm
, file_rss
, file_rss
);
367 add_mm_counter(mm
, anon_rss
, anon_rss
);
371 * This function is called to print an error when a bad pte
372 * is found. For example, we might have a PFN-mapped pte in
373 * a region that doesn't allow it.
375 * The calling function must still handle the error.
377 void print_bad_pte(struct vm_area_struct
*vma
, pte_t pte
, unsigned long vaddr
)
379 printk(KERN_ERR
"Bad pte = %08llx, process = %s, "
380 "vm_flags = %lx, vaddr = %lx\n",
381 (long long)pte_val(pte
),
382 (vma
->vm_mm
== current
->mm
? current
->comm
: "???"),
383 vma
->vm_flags
, vaddr
);
387 static inline int is_cow_mapping(unsigned int flags
)
389 return (flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
393 * vm_normal_page -- This function gets the "struct page" associated with a pte.
395 * "Special" mappings do not wish to be associated with a "struct page" (either
396 * it doesn't exist, or it exists but they don't want to touch it). In this
397 * case, NULL is returned here. "Normal" mappings do have a struct page.
399 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
400 * pte bit, in which case this function is trivial. Secondly, an architecture
401 * may not have a spare pte bit, which requires a more complicated scheme,
404 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
405 * special mapping (even if there are underlying and valid "struct pages").
406 * COWed pages of a VM_PFNMAP are always normal.
408 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
409 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
410 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
411 * mapping will always honor the rule
413 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
415 * And for normal mappings this is false.
417 * This restricts such mappings to be a linear translation from virtual address
418 * to pfn. To get around this restriction, we allow arbitrary mappings so long
419 * as the vma is not a COW mapping; in that case, we know that all ptes are
420 * special (because none can have been COWed).
423 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
425 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
426 * page" backing, however the difference is that _all_ pages with a struct
427 * page (that is, those where pfn_valid is true) are refcounted and considered
428 * normal pages by the VM. The disadvantage is that pages are refcounted
429 * (which can be slower and simply not an option for some PFNMAP users). The
430 * advantage is that we don't have to follow the strict linearity rule of
431 * PFNMAP mappings in order to support COWable mappings.
434 #ifdef __HAVE_ARCH_PTE_SPECIAL
435 # define HAVE_PTE_SPECIAL 1
437 # define HAVE_PTE_SPECIAL 0
439 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
444 if (HAVE_PTE_SPECIAL
) {
445 if (likely(!pte_special(pte
))) {
446 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
447 return pte_page(pte
);
449 VM_BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
)));
453 /* !HAVE_PTE_SPECIAL case follows: */
457 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
458 if (vma
->vm_flags
& VM_MIXEDMAP
) {
464 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
465 if (pfn
== vma
->vm_pgoff
+ off
)
467 if (!is_cow_mapping(vma
->vm_flags
))
472 VM_BUG_ON(!pfn_valid(pfn
));
475 * NOTE! We still have PageReserved() pages in the page tables.
477 * eg. VDSO mappings can cause them to exist.
480 return pfn_to_page(pfn
);
484 * copy one vm_area from one task to the other. Assumes the page tables
485 * already present in the new task to be cleared in the whole range
486 * covered by this vma.
490 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
491 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
492 unsigned long addr
, int *rss
)
494 unsigned long vm_flags
= vma
->vm_flags
;
495 pte_t pte
= *src_pte
;
498 /* pte contains position in swap or file, so copy. */
499 if (unlikely(!pte_present(pte
))) {
500 if (!pte_file(pte
)) {
501 swp_entry_t entry
= pte_to_swp_entry(pte
);
503 swap_duplicate(entry
);
504 /* make sure dst_mm is on swapoff's mmlist. */
505 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
506 spin_lock(&mmlist_lock
);
507 if (list_empty(&dst_mm
->mmlist
))
508 list_add(&dst_mm
->mmlist
,
510 spin_unlock(&mmlist_lock
);
512 if (is_write_migration_entry(entry
) &&
513 is_cow_mapping(vm_flags
)) {
515 * COW mappings require pages in both parent
516 * and child to be set to read.
518 make_migration_entry_read(&entry
);
519 pte
= swp_entry_to_pte(entry
);
520 set_pte_at(src_mm
, addr
, src_pte
, pte
);
527 * If it's a COW mapping, write protect it both
528 * in the parent and the child
530 if (is_cow_mapping(vm_flags
)) {
531 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
532 pte
= pte_wrprotect(pte
);
536 * If it's a shared mapping, mark it clean in
539 if (vm_flags
& VM_SHARED
)
540 pte
= pte_mkclean(pte
);
541 pte
= pte_mkold(pte
);
543 page
= vm_normal_page(vma
, addr
, pte
);
546 page_dup_rmap(page
, vma
, addr
);
547 rss
[!!PageAnon(page
)]++;
551 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
554 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
555 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
556 unsigned long addr
, unsigned long end
)
558 pte_t
*src_pte
, *dst_pte
;
559 spinlock_t
*src_ptl
, *dst_ptl
;
565 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
568 src_pte
= pte_offset_map_nested(src_pmd
, addr
);
569 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
570 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
571 arch_enter_lazy_mmu_mode();
575 * We are holding two locks at this point - either of them
576 * could generate latencies in another task on another CPU.
578 if (progress
>= 32) {
580 if (need_resched() ||
581 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
584 if (pte_none(*src_pte
)) {
588 copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
, vma
, addr
, rss
);
590 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
592 arch_leave_lazy_mmu_mode();
593 spin_unlock(src_ptl
);
594 pte_unmap_nested(src_pte
- 1);
595 add_mm_rss(dst_mm
, rss
[0], rss
[1]);
596 pte_unmap_unlock(dst_pte
- 1, dst_ptl
);
603 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
604 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
605 unsigned long addr
, unsigned long end
)
607 pmd_t
*src_pmd
, *dst_pmd
;
610 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
613 src_pmd
= pmd_offset(src_pud
, addr
);
615 next
= pmd_addr_end(addr
, end
);
616 if (pmd_none_or_clear_bad(src_pmd
))
618 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
621 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
625 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
626 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
627 unsigned long addr
, unsigned long end
)
629 pud_t
*src_pud
, *dst_pud
;
632 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
635 src_pud
= pud_offset(src_pgd
, addr
);
637 next
= pud_addr_end(addr
, end
);
638 if (pud_none_or_clear_bad(src_pud
))
640 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
643 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
647 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
648 struct vm_area_struct
*vma
)
650 pgd_t
*src_pgd
, *dst_pgd
;
652 unsigned long addr
= vma
->vm_start
;
653 unsigned long end
= vma
->vm_end
;
656 * Don't copy ptes where a page fault will fill them correctly.
657 * Fork becomes much lighter when there are big shared or private
658 * readonly mappings. The tradeoff is that copy_page_range is more
659 * efficient than faulting.
661 if (!(vma
->vm_flags
& (VM_HUGETLB
|VM_NONLINEAR
|VM_PFNMAP
|VM_INSERTPAGE
))) {
666 if (is_vm_hugetlb_page(vma
))
667 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
669 dst_pgd
= pgd_offset(dst_mm
, addr
);
670 src_pgd
= pgd_offset(src_mm
, addr
);
672 next
= pgd_addr_end(addr
, end
);
673 if (pgd_none_or_clear_bad(src_pgd
))
675 if (copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
678 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
682 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
683 struct vm_area_struct
*vma
, pmd_t
*pmd
,
684 unsigned long addr
, unsigned long end
,
685 long *zap_work
, struct zap_details
*details
)
687 struct mm_struct
*mm
= tlb
->mm
;
693 pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
694 arch_enter_lazy_mmu_mode();
697 if (pte_none(ptent
)) {
702 (*zap_work
) -= PAGE_SIZE
;
704 if (pte_present(ptent
)) {
707 page
= vm_normal_page(vma
, addr
, ptent
);
708 if (unlikely(details
) && page
) {
710 * unmap_shared_mapping_pages() wants to
711 * invalidate cache without truncating:
712 * unmap shared but keep private pages.
714 if (details
->check_mapping
&&
715 details
->check_mapping
!= page
->mapping
)
718 * Each page->index must be checked when
719 * invalidating or truncating nonlinear.
721 if (details
->nonlinear_vma
&&
722 (page
->index
< details
->first_index
||
723 page
->index
> details
->last_index
))
726 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
728 tlb_remove_tlb_entry(tlb
, pte
, addr
);
731 if (unlikely(details
) && details
->nonlinear_vma
732 && linear_page_index(details
->nonlinear_vma
,
733 addr
) != page
->index
)
734 set_pte_at(mm
, addr
, pte
,
735 pgoff_to_pte(page
->index
));
739 if (pte_dirty(ptent
))
740 set_page_dirty(page
);
741 if (pte_young(ptent
))
742 SetPageReferenced(page
);
745 page_remove_rmap(page
, vma
);
746 tlb_remove_page(tlb
, page
);
750 * If details->check_mapping, we leave swap entries;
751 * if details->nonlinear_vma, we leave file entries.
753 if (unlikely(details
))
755 if (!pte_file(ptent
))
756 free_swap_and_cache(pte_to_swp_entry(ptent
));
757 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
758 } while (pte
++, addr
+= PAGE_SIZE
, (addr
!= end
&& *zap_work
> 0));
760 add_mm_rss(mm
, file_rss
, anon_rss
);
761 arch_leave_lazy_mmu_mode();
762 pte_unmap_unlock(pte
- 1, ptl
);
767 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
768 struct vm_area_struct
*vma
, pud_t
*pud
,
769 unsigned long addr
, unsigned long end
,
770 long *zap_work
, struct zap_details
*details
)
775 pmd
= pmd_offset(pud
, addr
);
777 next
= pmd_addr_end(addr
, end
);
778 if (pmd_none_or_clear_bad(pmd
)) {
782 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
,
784 } while (pmd
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
789 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
790 struct vm_area_struct
*vma
, pgd_t
*pgd
,
791 unsigned long addr
, unsigned long end
,
792 long *zap_work
, struct zap_details
*details
)
797 pud
= pud_offset(pgd
, addr
);
799 next
= pud_addr_end(addr
, end
);
800 if (pud_none_or_clear_bad(pud
)) {
804 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
,
806 } while (pud
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
811 static unsigned long unmap_page_range(struct mmu_gather
*tlb
,
812 struct vm_area_struct
*vma
,
813 unsigned long addr
, unsigned long end
,
814 long *zap_work
, struct zap_details
*details
)
819 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
823 tlb_start_vma(tlb
, vma
);
824 pgd
= pgd_offset(vma
->vm_mm
, addr
);
826 next
= pgd_addr_end(addr
, end
);
827 if (pgd_none_or_clear_bad(pgd
)) {
831 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
,
833 } while (pgd
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
834 tlb_end_vma(tlb
, vma
);
839 #ifdef CONFIG_PREEMPT
840 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
842 /* No preempt: go for improved straight-line efficiency */
843 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
847 * unmap_vmas - unmap a range of memory covered by a list of vma's
848 * @tlbp: address of the caller's struct mmu_gather
849 * @vma: the starting vma
850 * @start_addr: virtual address at which to start unmapping
851 * @end_addr: virtual address at which to end unmapping
852 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
853 * @details: details of nonlinear truncation or shared cache invalidation
855 * Returns the end address of the unmapping (restart addr if interrupted).
857 * Unmap all pages in the vma list.
859 * We aim to not hold locks for too long (for scheduling latency reasons).
860 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
861 * return the ending mmu_gather to the caller.
863 * Only addresses between `start' and `end' will be unmapped.
865 * The VMA list must be sorted in ascending virtual address order.
867 * unmap_vmas() assumes that the caller will flush the whole unmapped address
868 * range after unmap_vmas() returns. So the only responsibility here is to
869 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
870 * drops the lock and schedules.
872 unsigned long unmap_vmas(struct mmu_gather
**tlbp
,
873 struct vm_area_struct
*vma
, unsigned long start_addr
,
874 unsigned long end_addr
, unsigned long *nr_accounted
,
875 struct zap_details
*details
)
877 long zap_work
= ZAP_BLOCK_SIZE
;
878 unsigned long tlb_start
= 0; /* For tlb_finish_mmu */
879 int tlb_start_valid
= 0;
880 unsigned long start
= start_addr
;
881 spinlock_t
*i_mmap_lock
= details
? details
->i_mmap_lock
: NULL
;
882 int fullmm
= (*tlbp
)->fullmm
;
884 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
) {
887 start
= max(vma
->vm_start
, start_addr
);
888 if (start
>= vma
->vm_end
)
890 end
= min(vma
->vm_end
, end_addr
);
891 if (end
<= vma
->vm_start
)
894 if (vma
->vm_flags
& VM_ACCOUNT
)
895 *nr_accounted
+= (end
- start
) >> PAGE_SHIFT
;
897 while (start
!= end
) {
898 if (!tlb_start_valid
) {
903 if (unlikely(is_vm_hugetlb_page(vma
))) {
904 unmap_hugepage_range(vma
, start
, end
, NULL
);
905 zap_work
-= (end
- start
) /
906 (HPAGE_SIZE
/ PAGE_SIZE
);
909 start
= unmap_page_range(*tlbp
, vma
,
910 start
, end
, &zap_work
, details
);
913 BUG_ON(start
!= end
);
917 tlb_finish_mmu(*tlbp
, tlb_start
, start
);
919 if (need_resched() ||
920 (i_mmap_lock
&& spin_needbreak(i_mmap_lock
))) {
928 *tlbp
= tlb_gather_mmu(vma
->vm_mm
, fullmm
);
930 zap_work
= ZAP_BLOCK_SIZE
;
934 return start
; /* which is now the end (or restart) address */
938 * zap_page_range - remove user pages in a given range
939 * @vma: vm_area_struct holding the applicable pages
940 * @address: starting address of pages to zap
941 * @size: number of bytes to zap
942 * @details: details of nonlinear truncation or shared cache invalidation
944 unsigned long zap_page_range(struct vm_area_struct
*vma
, unsigned long address
,
945 unsigned long size
, struct zap_details
*details
)
947 struct mm_struct
*mm
= vma
->vm_mm
;
948 struct mmu_gather
*tlb
;
949 unsigned long end
= address
+ size
;
950 unsigned long nr_accounted
= 0;
953 tlb
= tlb_gather_mmu(mm
, 0);
954 update_hiwater_rss(mm
);
955 end
= unmap_vmas(&tlb
, vma
, address
, end
, &nr_accounted
, details
);
957 tlb_finish_mmu(tlb
, address
, end
);
962 * Do a quick page-table lookup for a single page.
964 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
973 struct mm_struct
*mm
= vma
->vm_mm
;
975 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
977 BUG_ON(flags
& FOLL_GET
);
982 pgd
= pgd_offset(mm
, address
);
983 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
986 pud
= pud_offset(pgd
, address
);
987 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
990 pmd
= pmd_offset(pud
, address
);
994 if (pmd_huge(*pmd
)) {
995 BUG_ON(flags
& FOLL_GET
);
996 page
= follow_huge_pmd(mm
, address
, pmd
, flags
& FOLL_WRITE
);
1000 if (unlikely(pmd_bad(*pmd
)))
1003 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1006 if (!pte_present(pte
))
1008 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
1010 page
= vm_normal_page(vma
, address
, pte
);
1011 if (unlikely(!page
))
1014 if (flags
& FOLL_GET
)
1016 if (flags
& FOLL_TOUCH
) {
1017 if ((flags
& FOLL_WRITE
) &&
1018 !pte_dirty(pte
) && !PageDirty(page
))
1019 set_page_dirty(page
);
1020 mark_page_accessed(page
);
1023 pte_unmap_unlock(ptep
, ptl
);
1028 pte_unmap_unlock(ptep
, ptl
);
1029 return ERR_PTR(-EFAULT
);
1032 pte_unmap_unlock(ptep
, ptl
);
1035 /* Fall through to ZERO_PAGE handling */
1038 * When core dumping an enormous anonymous area that nobody
1039 * has touched so far, we don't want to allocate page tables.
1041 if (flags
& FOLL_ANON
) {
1042 page
= ZERO_PAGE(0);
1043 if (flags
& FOLL_GET
)
1045 BUG_ON(flags
& FOLL_WRITE
);
1050 /* Can we do the FOLL_ANON optimization? */
1051 static inline int use_zero_page(struct vm_area_struct
*vma
)
1054 * We don't want to optimize FOLL_ANON for make_pages_present()
1055 * when it tries to page in a VM_LOCKED region. As to VM_SHARED,
1056 * we want to get the page from the page tables to make sure
1057 * that we serialize and update with any other user of that
1060 if (vma
->vm_flags
& (VM_LOCKED
| VM_SHARED
))
1063 * And if we have a fault routine, it's not an anonymous region.
1065 return !vma
->vm_ops
|| !vma
->vm_ops
->fault
;
1068 int get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1069 unsigned long start
, int len
, int write
, int force
,
1070 struct page
**pages
, struct vm_area_struct
**vmas
)
1073 unsigned int vm_flags
;
1078 * Require read or write permissions.
1079 * If 'force' is set, we only require the "MAY" flags.
1081 vm_flags
= write
? (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1082 vm_flags
&= force
? (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1086 struct vm_area_struct
*vma
;
1087 unsigned int foll_flags
;
1089 vma
= find_extend_vma(mm
, start
);
1090 if (!vma
&& in_gate_area(tsk
, start
)) {
1091 unsigned long pg
= start
& PAGE_MASK
;
1092 struct vm_area_struct
*gate_vma
= get_gate_vma(tsk
);
1097 if (write
) /* user gate pages are read-only */
1098 return i
? : -EFAULT
;
1100 pgd
= pgd_offset_k(pg
);
1102 pgd
= pgd_offset_gate(mm
, pg
);
1103 BUG_ON(pgd_none(*pgd
));
1104 pud
= pud_offset(pgd
, pg
);
1105 BUG_ON(pud_none(*pud
));
1106 pmd
= pmd_offset(pud
, pg
);
1108 return i
? : -EFAULT
;
1109 pte
= pte_offset_map(pmd
, pg
);
1110 if (pte_none(*pte
)) {
1112 return i
? : -EFAULT
;
1115 struct page
*page
= vm_normal_page(gate_vma
, start
, *pte
);
1129 if (!vma
|| (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1130 || !(vm_flags
& vma
->vm_flags
))
1131 return i
? : -EFAULT
;
1133 if (is_vm_hugetlb_page(vma
)) {
1134 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1135 &start
, &len
, i
, write
);
1139 foll_flags
= FOLL_TOUCH
;
1141 foll_flags
|= FOLL_GET
;
1142 if (!write
&& use_zero_page(vma
))
1143 foll_flags
|= FOLL_ANON
;
1149 * If tsk is ooming, cut off its access to large memory
1150 * allocations. It has a pending SIGKILL, but it can't
1151 * be processed until returning to user space.
1153 if (unlikely(test_tsk_thread_flag(tsk
, TIF_MEMDIE
)))
1154 return i
? i
: -ENOMEM
;
1157 foll_flags
|= FOLL_WRITE
;
1160 while (!(page
= follow_page(vma
, start
, foll_flags
))) {
1162 ret
= handle_mm_fault(mm
, vma
, start
,
1163 foll_flags
& FOLL_WRITE
);
1164 if (ret
& VM_FAULT_ERROR
) {
1165 if (ret
& VM_FAULT_OOM
)
1166 return i
? i
: -ENOMEM
;
1167 else if (ret
& VM_FAULT_SIGBUS
)
1168 return i
? i
: -EFAULT
;
1171 if (ret
& VM_FAULT_MAJOR
)
1177 * The VM_FAULT_WRITE bit tells us that
1178 * do_wp_page has broken COW when necessary,
1179 * even if maybe_mkwrite decided not to set
1180 * pte_write. We can thus safely do subsequent
1181 * page lookups as if they were reads.
1183 if (ret
& VM_FAULT_WRITE
)
1184 foll_flags
&= ~FOLL_WRITE
;
1189 return i
? i
: PTR_ERR(page
);
1193 flush_anon_page(vma
, page
, start
);
1194 flush_dcache_page(page
);
1201 } while (len
&& start
< vma
->vm_end
);
1205 EXPORT_SYMBOL(get_user_pages
);
1207 pte_t
*get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1210 pgd_t
* pgd
= pgd_offset(mm
, addr
);
1211 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
1213 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
1215 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1221 * This is the old fallback for page remapping.
1223 * For historical reasons, it only allows reserved pages. Only
1224 * old drivers should use this, and they needed to mark their
1225 * pages reserved for the old functions anyway.
1227 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1228 struct page
*page
, pgprot_t prot
)
1230 struct mm_struct
*mm
= vma
->vm_mm
;
1235 retval
= mem_cgroup_charge(page
, mm
, GFP_KERNEL
);
1243 flush_dcache_page(page
);
1244 pte
= get_locked_pte(mm
, addr
, &ptl
);
1248 if (!pte_none(*pte
))
1251 /* Ok, finally just insert the thing.. */
1253 inc_mm_counter(mm
, file_rss
);
1254 page_add_file_rmap(page
);
1255 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1258 pte_unmap_unlock(pte
, ptl
);
1261 pte_unmap_unlock(pte
, ptl
);
1263 mem_cgroup_uncharge_page(page
);
1269 * vm_insert_page - insert single page into user vma
1270 * @vma: user vma to map to
1271 * @addr: target user address of this page
1272 * @page: source kernel page
1274 * This allows drivers to insert individual pages they've allocated
1277 * The page has to be a nice clean _individual_ kernel allocation.
1278 * If you allocate a compound page, you need to have marked it as
1279 * such (__GFP_COMP), or manually just split the page up yourself
1280 * (see split_page()).
1282 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1283 * took an arbitrary page protection parameter. This doesn't allow
1284 * that. Your vma protection will have to be set up correctly, which
1285 * means that if you want a shared writable mapping, you'd better
1286 * ask for a shared writable mapping!
1288 * The page does not need to be reserved.
1290 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1293 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1295 if (!page_count(page
))
1297 vma
->vm_flags
|= VM_INSERTPAGE
;
1298 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1300 EXPORT_SYMBOL(vm_insert_page
);
1302 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1303 unsigned long pfn
, pgprot_t prot
)
1305 struct mm_struct
*mm
= vma
->vm_mm
;
1311 pte
= get_locked_pte(mm
, addr
, &ptl
);
1315 if (!pte_none(*pte
))
1318 /* Ok, finally just insert the thing.. */
1319 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
1320 set_pte_at(mm
, addr
, pte
, entry
);
1321 update_mmu_cache(vma
, addr
, entry
); /* XXX: why not for insert_page? */
1325 pte_unmap_unlock(pte
, ptl
);
1331 * vm_insert_pfn - insert single pfn into user vma
1332 * @vma: user vma to map to
1333 * @addr: target user address of this page
1334 * @pfn: source kernel pfn
1336 * Similar to vm_inert_page, this allows drivers to insert individual pages
1337 * they've allocated into a user vma. Same comments apply.
1339 * This function should only be called from a vm_ops->fault handler, and
1340 * in that case the handler should return NULL.
1342 * vma cannot be a COW mapping.
1344 * As this is called only for pages that do not currently exist, we
1345 * do not need to flush old virtual caches or the TLB.
1347 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1351 * Technically, architectures with pte_special can avoid all these
1352 * restrictions (same for remap_pfn_range). However we would like
1353 * consistency in testing and feature parity among all, so we should
1354 * try to keep these invariants in place for everybody.
1356 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1357 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1358 (VM_PFNMAP
|VM_MIXEDMAP
));
1359 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1360 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1362 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1364 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1366 EXPORT_SYMBOL(vm_insert_pfn
);
1368 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1371 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
1373 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1377 * If we don't have pte special, then we have to use the pfn_valid()
1378 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1379 * refcount the page if pfn_valid is true (hence insert_page rather
1382 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
1385 page
= pfn_to_page(pfn
);
1386 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1388 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1390 EXPORT_SYMBOL(vm_insert_mixed
);
1393 * maps a range of physical memory into the requested pages. the old
1394 * mappings are removed. any references to nonexistent pages results
1395 * in null mappings (currently treated as "copy-on-access")
1397 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1398 unsigned long addr
, unsigned long end
,
1399 unsigned long pfn
, pgprot_t prot
)
1404 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1407 arch_enter_lazy_mmu_mode();
1409 BUG_ON(!pte_none(*pte
));
1410 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
1412 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1413 arch_leave_lazy_mmu_mode();
1414 pte_unmap_unlock(pte
- 1, ptl
);
1418 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1419 unsigned long addr
, unsigned long end
,
1420 unsigned long pfn
, pgprot_t prot
)
1425 pfn
-= addr
>> PAGE_SHIFT
;
1426 pmd
= pmd_alloc(mm
, pud
, addr
);
1430 next
= pmd_addr_end(addr
, end
);
1431 if (remap_pte_range(mm
, pmd
, addr
, next
,
1432 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1434 } while (pmd
++, addr
= next
, addr
!= end
);
1438 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1439 unsigned long addr
, unsigned long end
,
1440 unsigned long pfn
, pgprot_t prot
)
1445 pfn
-= addr
>> PAGE_SHIFT
;
1446 pud
= pud_alloc(mm
, pgd
, addr
);
1450 next
= pud_addr_end(addr
, end
);
1451 if (remap_pmd_range(mm
, pud
, addr
, next
,
1452 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1454 } while (pud
++, addr
= next
, addr
!= end
);
1459 * remap_pfn_range - remap kernel memory to userspace
1460 * @vma: user vma to map to
1461 * @addr: target user address to start at
1462 * @pfn: physical address of kernel memory
1463 * @size: size of map area
1464 * @prot: page protection flags for this mapping
1466 * Note: this is only safe if the mm semaphore is held when called.
1468 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1469 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1473 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1474 struct mm_struct
*mm
= vma
->vm_mm
;
1478 * Physically remapped pages are special. Tell the
1479 * rest of the world about it:
1480 * VM_IO tells people not to look at these pages
1481 * (accesses can have side effects).
1482 * VM_RESERVED is specified all over the place, because
1483 * in 2.4 it kept swapout's vma scan off this vma; but
1484 * in 2.6 the LRU scan won't even find its pages, so this
1485 * flag means no more than count its pages in reserved_vm,
1486 * and omit it from core dump, even when VM_IO turned off.
1487 * VM_PFNMAP tells the core MM that the base pages are just
1488 * raw PFN mappings, and do not have a "struct page" associated
1491 * There's a horrible special case to handle copy-on-write
1492 * behaviour that some programs depend on. We mark the "original"
1493 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1495 if (is_cow_mapping(vma
->vm_flags
)) {
1496 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
1498 vma
->vm_pgoff
= pfn
;
1501 vma
->vm_flags
|= VM_IO
| VM_RESERVED
| VM_PFNMAP
;
1503 BUG_ON(addr
>= end
);
1504 pfn
-= addr
>> PAGE_SHIFT
;
1505 pgd
= pgd_offset(mm
, addr
);
1506 flush_cache_range(vma
, addr
, end
);
1508 next
= pgd_addr_end(addr
, end
);
1509 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1510 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1513 } while (pgd
++, addr
= next
, addr
!= end
);
1516 EXPORT_SYMBOL(remap_pfn_range
);
1518 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1519 unsigned long addr
, unsigned long end
,
1520 pte_fn_t fn
, void *data
)
1525 spinlock_t
*uninitialized_var(ptl
);
1527 pte
= (mm
== &init_mm
) ?
1528 pte_alloc_kernel(pmd
, addr
) :
1529 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1533 BUG_ON(pmd_huge(*pmd
));
1535 token
= pmd_pgtable(*pmd
);
1538 err
= fn(pte
, token
, addr
, data
);
1541 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1544 pte_unmap_unlock(pte
-1, ptl
);
1548 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1549 unsigned long addr
, unsigned long end
,
1550 pte_fn_t fn
, void *data
)
1556 pmd
= pmd_alloc(mm
, pud
, addr
);
1560 next
= pmd_addr_end(addr
, end
);
1561 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
1564 } while (pmd
++, addr
= next
, addr
!= end
);
1568 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1569 unsigned long addr
, unsigned long end
,
1570 pte_fn_t fn
, void *data
)
1576 pud
= pud_alloc(mm
, pgd
, addr
);
1580 next
= pud_addr_end(addr
, end
);
1581 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
1584 } while (pud
++, addr
= next
, addr
!= end
);
1589 * Scan a region of virtual memory, filling in page tables as necessary
1590 * and calling a provided function on each leaf page table.
1592 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
1593 unsigned long size
, pte_fn_t fn
, void *data
)
1597 unsigned long end
= addr
+ size
;
1600 BUG_ON(addr
>= end
);
1601 pgd
= pgd_offset(mm
, addr
);
1603 next
= pgd_addr_end(addr
, end
);
1604 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
1607 } while (pgd
++, addr
= next
, addr
!= end
);
1610 EXPORT_SYMBOL_GPL(apply_to_page_range
);
1613 * handle_pte_fault chooses page fault handler according to an entry
1614 * which was read non-atomically. Before making any commitment, on
1615 * those architectures or configurations (e.g. i386 with PAE) which
1616 * might give a mix of unmatched parts, do_swap_page and do_file_page
1617 * must check under lock before unmapping the pte and proceeding
1618 * (but do_wp_page is only called after already making such a check;
1619 * and do_anonymous_page and do_no_page can safely check later on).
1621 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
1622 pte_t
*page_table
, pte_t orig_pte
)
1625 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1626 if (sizeof(pte_t
) > sizeof(unsigned long)) {
1627 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
1629 same
= pte_same(*page_table
, orig_pte
);
1633 pte_unmap(page_table
);
1638 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1639 * servicing faults for write access. In the normal case, do always want
1640 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1641 * that do not have writing enabled, when used by access_process_vm.
1643 static inline pte_t
maybe_mkwrite(pte_t pte
, struct vm_area_struct
*vma
)
1645 if (likely(vma
->vm_flags
& VM_WRITE
))
1646 pte
= pte_mkwrite(pte
);
1650 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
1653 * If the source page was a PFN mapping, we don't have
1654 * a "struct page" for it. We do a best-effort copy by
1655 * just copying from the original user address. If that
1656 * fails, we just zero-fill it. Live with it.
1658 if (unlikely(!src
)) {
1659 void *kaddr
= kmap_atomic(dst
, KM_USER0
);
1660 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
1663 * This really shouldn't fail, because the page is there
1664 * in the page tables. But it might just be unreadable,
1665 * in which case we just give up and fill the result with
1668 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
1669 memset(kaddr
, 0, PAGE_SIZE
);
1670 kunmap_atomic(kaddr
, KM_USER0
);
1671 flush_dcache_page(dst
);
1673 copy_user_highpage(dst
, src
, va
, vma
);
1677 * This routine handles present pages, when users try to write
1678 * to a shared page. It is done by copying the page to a new address
1679 * and decrementing the shared-page counter for the old page.
1681 * Note that this routine assumes that the protection checks have been
1682 * done by the caller (the low-level page fault routine in most cases).
1683 * Thus we can safely just mark it writable once we've done any necessary
1686 * We also mark the page dirty at this point even though the page will
1687 * change only once the write actually happens. This avoids a few races,
1688 * and potentially makes it more efficient.
1690 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1691 * but allow concurrent faults), with pte both mapped and locked.
1692 * We return with mmap_sem still held, but pte unmapped and unlocked.
1694 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1695 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
1696 spinlock_t
*ptl
, pte_t orig_pte
)
1698 struct page
*old_page
, *new_page
;
1700 int reuse
= 0, ret
= 0;
1701 int page_mkwrite
= 0;
1702 struct page
*dirty_page
= NULL
;
1704 old_page
= vm_normal_page(vma
, address
, orig_pte
);
1707 * VM_MIXEDMAP !pfn_valid() case
1709 * We should not cow pages in a shared writeable mapping.
1710 * Just mark the pages writable as we can't do any dirty
1711 * accounting on raw pfn maps.
1713 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
1714 (VM_WRITE
|VM_SHARED
))
1720 * Take out anonymous pages first, anonymous shared vmas are
1721 * not dirty accountable.
1723 if (PageAnon(old_page
)) {
1724 if (!TestSetPageLocked(old_page
)) {
1725 reuse
= can_share_swap_page(old_page
);
1726 unlock_page(old_page
);
1728 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
1729 (VM_WRITE
|VM_SHARED
))) {
1731 * Only catch write-faults on shared writable pages,
1732 * read-only shared pages can get COWed by
1733 * get_user_pages(.write=1, .force=1).
1735 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
1737 * Notify the address space that the page is about to
1738 * become writable so that it can prohibit this or wait
1739 * for the page to get into an appropriate state.
1741 * We do this without the lock held, so that it can
1742 * sleep if it needs to.
1744 page_cache_get(old_page
);
1745 pte_unmap_unlock(page_table
, ptl
);
1747 if (vma
->vm_ops
->page_mkwrite(vma
, old_page
) < 0)
1748 goto unwritable_page
;
1751 * Since we dropped the lock we need to revalidate
1752 * the PTE as someone else may have changed it. If
1753 * they did, we just return, as we can count on the
1754 * MMU to tell us if they didn't also make it writable.
1756 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
1758 page_cache_release(old_page
);
1759 if (!pte_same(*page_table
, orig_pte
))
1764 dirty_page
= old_page
;
1765 get_page(dirty_page
);
1771 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
1772 entry
= pte_mkyoung(orig_pte
);
1773 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1774 if (ptep_set_access_flags(vma
, address
, page_table
, entry
,1))
1775 update_mmu_cache(vma
, address
, entry
);
1776 ret
|= VM_FAULT_WRITE
;
1781 * Ok, we need to copy. Oh, well..
1783 page_cache_get(old_page
);
1785 pte_unmap_unlock(page_table
, ptl
);
1787 if (unlikely(anon_vma_prepare(vma
)))
1789 VM_BUG_ON(old_page
== ZERO_PAGE(0));
1790 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
1793 cow_user_page(new_page
, old_page
, address
, vma
);
1794 __SetPageUptodate(new_page
);
1796 if (mem_cgroup_charge(new_page
, mm
, GFP_KERNEL
))
1800 * Re-check the pte - we dropped the lock
1802 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1803 if (likely(pte_same(*page_table
, orig_pte
))) {
1805 if (!PageAnon(old_page
)) {
1806 dec_mm_counter(mm
, file_rss
);
1807 inc_mm_counter(mm
, anon_rss
);
1810 inc_mm_counter(mm
, anon_rss
);
1811 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
1812 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
1813 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1815 * Clear the pte entry and flush it first, before updating the
1816 * pte with the new entry. This will avoid a race condition
1817 * seen in the presence of one thread doing SMC and another
1820 ptep_clear_flush(vma
, address
, page_table
);
1821 set_pte_at(mm
, address
, page_table
, entry
);
1822 update_mmu_cache(vma
, address
, entry
);
1823 lru_cache_add_active(new_page
);
1824 page_add_new_anon_rmap(new_page
, vma
, address
);
1828 * Only after switching the pte to the new page may
1829 * we remove the mapcount here. Otherwise another
1830 * process may come and find the rmap count decremented
1831 * before the pte is switched to the new page, and
1832 * "reuse" the old page writing into it while our pte
1833 * here still points into it and can be read by other
1836 * The critical issue is to order this
1837 * page_remove_rmap with the ptp_clear_flush above.
1838 * Those stores are ordered by (if nothing else,)
1839 * the barrier present in the atomic_add_negative
1840 * in page_remove_rmap.
1842 * Then the TLB flush in ptep_clear_flush ensures that
1843 * no process can access the old page before the
1844 * decremented mapcount is visible. And the old page
1845 * cannot be reused until after the decremented
1846 * mapcount is visible. So transitively, TLBs to
1847 * old page will be flushed before it can be reused.
1849 page_remove_rmap(old_page
, vma
);
1852 /* Free the old page.. */
1853 new_page
= old_page
;
1854 ret
|= VM_FAULT_WRITE
;
1856 mem_cgroup_uncharge_page(new_page
);
1859 page_cache_release(new_page
);
1861 page_cache_release(old_page
);
1863 pte_unmap_unlock(page_table
, ptl
);
1866 file_update_time(vma
->vm_file
);
1869 * Yes, Virginia, this is actually required to prevent a race
1870 * with clear_page_dirty_for_io() from clearing the page dirty
1871 * bit after it clear all dirty ptes, but before a racing
1872 * do_wp_page installs a dirty pte.
1874 * do_no_page is protected similarly.
1876 wait_on_page_locked(dirty_page
);
1877 set_page_dirty_balance(dirty_page
, page_mkwrite
);
1878 put_page(dirty_page
);
1882 page_cache_release(new_page
);
1885 page_cache_release(old_page
);
1886 return VM_FAULT_OOM
;
1889 page_cache_release(old_page
);
1890 return VM_FAULT_SIGBUS
;
1894 * Helper functions for unmap_mapping_range().
1896 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1898 * We have to restart searching the prio_tree whenever we drop the lock,
1899 * since the iterator is only valid while the lock is held, and anyway
1900 * a later vma might be split and reinserted earlier while lock dropped.
1902 * The list of nonlinear vmas could be handled more efficiently, using
1903 * a placeholder, but handle it in the same way until a need is shown.
1904 * It is important to search the prio_tree before nonlinear list: a vma
1905 * may become nonlinear and be shifted from prio_tree to nonlinear list
1906 * while the lock is dropped; but never shifted from list to prio_tree.
1908 * In order to make forward progress despite restarting the search,
1909 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1910 * quickly skip it next time around. Since the prio_tree search only
1911 * shows us those vmas affected by unmapping the range in question, we
1912 * can't efficiently keep all vmas in step with mapping->truncate_count:
1913 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1914 * mapping->truncate_count and vma->vm_truncate_count are protected by
1917 * In order to make forward progress despite repeatedly restarting some
1918 * large vma, note the restart_addr from unmap_vmas when it breaks out:
1919 * and restart from that address when we reach that vma again. It might
1920 * have been split or merged, shrunk or extended, but never shifted: so
1921 * restart_addr remains valid so long as it remains in the vma's range.
1922 * unmap_mapping_range forces truncate_count to leap over page-aligned
1923 * values so we can save vma's restart_addr in its truncate_count field.
1925 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1927 static void reset_vma_truncate_counts(struct address_space
*mapping
)
1929 struct vm_area_struct
*vma
;
1930 struct prio_tree_iter iter
;
1932 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, 0, ULONG_MAX
)
1933 vma
->vm_truncate_count
= 0;
1934 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
1935 vma
->vm_truncate_count
= 0;
1938 static int unmap_mapping_range_vma(struct vm_area_struct
*vma
,
1939 unsigned long start_addr
, unsigned long end_addr
,
1940 struct zap_details
*details
)
1942 unsigned long restart_addr
;
1946 * files that support invalidating or truncating portions of the
1947 * file from under mmaped areas must have their ->fault function
1948 * return a locked page (and set VM_FAULT_LOCKED in the return).
1949 * This provides synchronisation against concurrent unmapping here.
1953 restart_addr
= vma
->vm_truncate_count
;
1954 if (is_restart_addr(restart_addr
) && start_addr
< restart_addr
) {
1955 start_addr
= restart_addr
;
1956 if (start_addr
>= end_addr
) {
1957 /* Top of vma has been split off since last time */
1958 vma
->vm_truncate_count
= details
->truncate_count
;
1963 restart_addr
= zap_page_range(vma
, start_addr
,
1964 end_addr
- start_addr
, details
);
1965 need_break
= need_resched() || spin_needbreak(details
->i_mmap_lock
);
1967 if (restart_addr
>= end_addr
) {
1968 /* We have now completed this vma: mark it so */
1969 vma
->vm_truncate_count
= details
->truncate_count
;
1973 /* Note restart_addr in vma's truncate_count field */
1974 vma
->vm_truncate_count
= restart_addr
;
1979 spin_unlock(details
->i_mmap_lock
);
1981 spin_lock(details
->i_mmap_lock
);
1985 static inline void unmap_mapping_range_tree(struct prio_tree_root
*root
,
1986 struct zap_details
*details
)
1988 struct vm_area_struct
*vma
;
1989 struct prio_tree_iter iter
;
1990 pgoff_t vba
, vea
, zba
, zea
;
1993 vma_prio_tree_foreach(vma
, &iter
, root
,
1994 details
->first_index
, details
->last_index
) {
1995 /* Skip quickly over those we have already dealt with */
1996 if (vma
->vm_truncate_count
== details
->truncate_count
)
1999 vba
= vma
->vm_pgoff
;
2000 vea
= vba
+ ((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) - 1;
2001 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2002 zba
= details
->first_index
;
2005 zea
= details
->last_index
;
2009 if (unmap_mapping_range_vma(vma
,
2010 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2011 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2017 static inline void unmap_mapping_range_list(struct list_head
*head
,
2018 struct zap_details
*details
)
2020 struct vm_area_struct
*vma
;
2023 * In nonlinear VMAs there is no correspondence between virtual address
2024 * offset and file offset. So we must perform an exhaustive search
2025 * across *all* the pages in each nonlinear VMA, not just the pages
2026 * whose virtual address lies outside the file truncation point.
2029 list_for_each_entry(vma
, head
, shared
.vm_set
.list
) {
2030 /* Skip quickly over those we have already dealt with */
2031 if (vma
->vm_truncate_count
== details
->truncate_count
)
2033 details
->nonlinear_vma
= vma
;
2034 if (unmap_mapping_range_vma(vma
, vma
->vm_start
,
2035 vma
->vm_end
, details
) < 0)
2041 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2042 * @mapping: the address space containing mmaps to be unmapped.
2043 * @holebegin: byte in first page to unmap, relative to the start of
2044 * the underlying file. This will be rounded down to a PAGE_SIZE
2045 * boundary. Note that this is different from vmtruncate(), which
2046 * must keep the partial page. In contrast, we must get rid of
2048 * @holelen: size of prospective hole in bytes. This will be rounded
2049 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2051 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2052 * but 0 when invalidating pagecache, don't throw away private data.
2054 void unmap_mapping_range(struct address_space
*mapping
,
2055 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2057 struct zap_details details
;
2058 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2059 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2061 /* Check for overflow. */
2062 if (sizeof(holelen
) > sizeof(hlen
)) {
2064 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2065 if (holeend
& ~(long long)ULONG_MAX
)
2066 hlen
= ULONG_MAX
- hba
+ 1;
2069 details
.check_mapping
= even_cows
? NULL
: mapping
;
2070 details
.nonlinear_vma
= NULL
;
2071 details
.first_index
= hba
;
2072 details
.last_index
= hba
+ hlen
- 1;
2073 if (details
.last_index
< details
.first_index
)
2074 details
.last_index
= ULONG_MAX
;
2075 details
.i_mmap_lock
= &mapping
->i_mmap_lock
;
2077 spin_lock(&mapping
->i_mmap_lock
);
2079 /* Protect against endless unmapping loops */
2080 mapping
->truncate_count
++;
2081 if (unlikely(is_restart_addr(mapping
->truncate_count
))) {
2082 if (mapping
->truncate_count
== 0)
2083 reset_vma_truncate_counts(mapping
);
2084 mapping
->truncate_count
++;
2086 details
.truncate_count
= mapping
->truncate_count
;
2088 if (unlikely(!prio_tree_empty(&mapping
->i_mmap
)))
2089 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2090 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
2091 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
2092 spin_unlock(&mapping
->i_mmap_lock
);
2094 EXPORT_SYMBOL(unmap_mapping_range
);
2097 * vmtruncate - unmap mappings "freed" by truncate() syscall
2098 * @inode: inode of the file used
2099 * @offset: file offset to start truncating
2101 * NOTE! We have to be ready to update the memory sharing
2102 * between the file and the memory map for a potential last
2103 * incomplete page. Ugly, but necessary.
2105 int vmtruncate(struct inode
* inode
, loff_t offset
)
2107 if (inode
->i_size
< offset
) {
2108 unsigned long limit
;
2110 limit
= current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
;
2111 if (limit
!= RLIM_INFINITY
&& offset
> limit
)
2113 if (offset
> inode
->i_sb
->s_maxbytes
)
2115 i_size_write(inode
, offset
);
2117 struct address_space
*mapping
= inode
->i_mapping
;
2120 * truncation of in-use swapfiles is disallowed - it would
2121 * cause subsequent swapout to scribble on the now-freed
2124 if (IS_SWAPFILE(inode
))
2126 i_size_write(inode
, offset
);
2129 * unmap_mapping_range is called twice, first simply for
2130 * efficiency so that truncate_inode_pages does fewer
2131 * single-page unmaps. However after this first call, and
2132 * before truncate_inode_pages finishes, it is possible for
2133 * private pages to be COWed, which remain after
2134 * truncate_inode_pages finishes, hence the second
2135 * unmap_mapping_range call must be made for correctness.
2137 unmap_mapping_range(mapping
, offset
+ PAGE_SIZE
- 1, 0, 1);
2138 truncate_inode_pages(mapping
, offset
);
2139 unmap_mapping_range(mapping
, offset
+ PAGE_SIZE
- 1, 0, 1);
2142 if (inode
->i_op
&& inode
->i_op
->truncate
)
2143 inode
->i_op
->truncate(inode
);
2147 send_sig(SIGXFSZ
, current
, 0);
2151 EXPORT_SYMBOL(vmtruncate
);
2153 int vmtruncate_range(struct inode
*inode
, loff_t offset
, loff_t end
)
2155 struct address_space
*mapping
= inode
->i_mapping
;
2158 * If the underlying filesystem is not going to provide
2159 * a way to truncate a range of blocks (punch a hole) -
2160 * we should return failure right now.
2162 if (!inode
->i_op
|| !inode
->i_op
->truncate_range
)
2165 mutex_lock(&inode
->i_mutex
);
2166 down_write(&inode
->i_alloc_sem
);
2167 unmap_mapping_range(mapping
, offset
, (end
- offset
), 1);
2168 truncate_inode_pages_range(mapping
, offset
, end
);
2169 unmap_mapping_range(mapping
, offset
, (end
- offset
), 1);
2170 inode
->i_op
->truncate_range(inode
, offset
, end
);
2171 up_write(&inode
->i_alloc_sem
);
2172 mutex_unlock(&inode
->i_mutex
);
2178 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2179 * but allow concurrent faults), and pte mapped but not yet locked.
2180 * We return with mmap_sem still held, but pte unmapped and unlocked.
2182 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2183 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2184 int write_access
, pte_t orig_pte
)
2192 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2195 entry
= pte_to_swp_entry(orig_pte
);
2196 if (is_migration_entry(entry
)) {
2197 migration_entry_wait(mm
, pmd
, address
);
2200 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2201 page
= lookup_swap_cache(entry
);
2203 grab_swap_token(); /* Contend for token _before_ read-in */
2204 page
= swapin_readahead(entry
,
2205 GFP_HIGHUSER_MOVABLE
, vma
, address
);
2208 * Back out if somebody else faulted in this pte
2209 * while we released the pte lock.
2211 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2212 if (likely(pte_same(*page_table
, orig_pte
)))
2214 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2218 /* Had to read the page from swap area: Major fault */
2219 ret
= VM_FAULT_MAJOR
;
2220 count_vm_event(PGMAJFAULT
);
2223 if (mem_cgroup_charge(page
, mm
, GFP_KERNEL
)) {
2224 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2229 mark_page_accessed(page
);
2231 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2234 * Back out if somebody else already faulted in this pte.
2236 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2237 if (unlikely(!pte_same(*page_table
, orig_pte
)))
2240 if (unlikely(!PageUptodate(page
))) {
2241 ret
= VM_FAULT_SIGBUS
;
2245 /* The page isn't present yet, go ahead with the fault. */
2247 inc_mm_counter(mm
, anon_rss
);
2248 pte
= mk_pte(page
, vma
->vm_page_prot
);
2249 if (write_access
&& can_share_swap_page(page
)) {
2250 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
2254 flush_icache_page(vma
, page
);
2255 set_pte_at(mm
, address
, page_table
, pte
);
2256 page_add_anon_rmap(page
, vma
, address
);
2260 remove_exclusive_swap_page(page
);
2264 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
2265 if (ret
& VM_FAULT_ERROR
)
2266 ret
&= VM_FAULT_ERROR
;
2270 /* No need to invalidate - it was non-present before */
2271 update_mmu_cache(vma
, address
, pte
);
2273 pte_unmap_unlock(page_table
, ptl
);
2277 mem_cgroup_uncharge_page(page
);
2278 pte_unmap_unlock(page_table
, ptl
);
2280 page_cache_release(page
);
2285 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2286 * but allow concurrent faults), and pte mapped but not yet locked.
2287 * We return with mmap_sem still held, but pte unmapped and unlocked.
2289 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2290 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2297 /* Allocate our own private page. */
2298 pte_unmap(page_table
);
2300 if (unlikely(anon_vma_prepare(vma
)))
2302 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2305 __SetPageUptodate(page
);
2307 if (mem_cgroup_charge(page
, mm
, GFP_KERNEL
))
2310 entry
= mk_pte(page
, vma
->vm_page_prot
);
2311 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2313 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2314 if (!pte_none(*page_table
))
2316 inc_mm_counter(mm
, anon_rss
);
2317 lru_cache_add_active(page
);
2318 page_add_new_anon_rmap(page
, vma
, address
);
2319 set_pte_at(mm
, address
, page_table
, entry
);
2321 /* No need to invalidate - it was non-present before */
2322 update_mmu_cache(vma
, address
, entry
);
2324 pte_unmap_unlock(page_table
, ptl
);
2327 mem_cgroup_uncharge_page(page
);
2328 page_cache_release(page
);
2331 page_cache_release(page
);
2333 return VM_FAULT_OOM
;
2337 * __do_fault() tries to create a new page mapping. It aggressively
2338 * tries to share with existing pages, but makes a separate copy if
2339 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2340 * the next page fault.
2342 * As this is called only for pages that do not currently exist, we
2343 * do not need to flush old virtual caches or the TLB.
2345 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2346 * but allow concurrent faults), and pte neither mapped nor locked.
2347 * We return with mmap_sem still held, but pte unmapped and unlocked.
2349 static int __do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2350 unsigned long address
, pmd_t
*pmd
,
2351 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
2358 struct page
*dirty_page
= NULL
;
2359 struct vm_fault vmf
;
2361 int page_mkwrite
= 0;
2363 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
2368 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
2369 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2373 * For consistency in subsequent calls, make the faulted page always
2376 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
2377 lock_page(vmf
.page
);
2379 VM_BUG_ON(!PageLocked(vmf
.page
));
2382 * Should we do an early C-O-W break?
2385 if (flags
& FAULT_FLAG_WRITE
) {
2386 if (!(vma
->vm_flags
& VM_SHARED
)) {
2388 if (unlikely(anon_vma_prepare(vma
))) {
2392 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
,
2398 copy_user_highpage(page
, vmf
.page
, address
, vma
);
2399 __SetPageUptodate(page
);
2402 * If the page will be shareable, see if the backing
2403 * address space wants to know that the page is about
2404 * to become writable
2406 if (vma
->vm_ops
->page_mkwrite
) {
2408 if (vma
->vm_ops
->page_mkwrite(vma
, page
) < 0) {
2409 ret
= VM_FAULT_SIGBUS
;
2410 anon
= 1; /* no anon but release vmf.page */
2415 * XXX: this is not quite right (racy vs
2416 * invalidate) to unlock and relock the page
2417 * like this, however a better fix requires
2418 * reworking page_mkwrite locking API, which
2419 * is better done later.
2421 if (!page
->mapping
) {
2423 anon
= 1; /* no anon but release vmf.page */
2432 if (mem_cgroup_charge(page
, mm
, GFP_KERNEL
)) {
2437 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2440 * This silly early PAGE_DIRTY setting removes a race
2441 * due to the bad i386 page protection. But it's valid
2442 * for other architectures too.
2444 * Note that if write_access is true, we either now have
2445 * an exclusive copy of the page, or this is a shared mapping,
2446 * so we can make it writable and dirty to avoid having to
2447 * handle that later.
2449 /* Only go through if we didn't race with anybody else... */
2450 if (likely(pte_same(*page_table
, orig_pte
))) {
2451 flush_icache_page(vma
, page
);
2452 entry
= mk_pte(page
, vma
->vm_page_prot
);
2453 if (flags
& FAULT_FLAG_WRITE
)
2454 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2455 set_pte_at(mm
, address
, page_table
, entry
);
2457 inc_mm_counter(mm
, anon_rss
);
2458 lru_cache_add_active(page
);
2459 page_add_new_anon_rmap(page
, vma
, address
);
2461 inc_mm_counter(mm
, file_rss
);
2462 page_add_file_rmap(page
);
2463 if (flags
& FAULT_FLAG_WRITE
) {
2465 get_page(dirty_page
);
2469 /* no need to invalidate: a not-present page won't be cached */
2470 update_mmu_cache(vma
, address
, entry
);
2472 mem_cgroup_uncharge_page(page
);
2474 page_cache_release(page
);
2476 anon
= 1; /* no anon but release faulted_page */
2479 pte_unmap_unlock(page_table
, ptl
);
2482 unlock_page(vmf
.page
);
2485 page_cache_release(vmf
.page
);
2486 else if (dirty_page
) {
2488 file_update_time(vma
->vm_file
);
2490 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2491 put_page(dirty_page
);
2497 static int do_linear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2498 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2499 int write_access
, pte_t orig_pte
)
2501 pgoff_t pgoff
= (((address
& PAGE_MASK
)
2502 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
2503 unsigned int flags
= (write_access
? FAULT_FLAG_WRITE
: 0);
2505 pte_unmap(page_table
);
2506 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
2510 * Fault of a previously existing named mapping. Repopulate the pte
2511 * from the encoded file_pte if possible. This enables swappable
2514 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2515 * but allow concurrent faults), and pte mapped but not yet locked.
2516 * We return with mmap_sem still held, but pte unmapped and unlocked.
2518 static int do_nonlinear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2519 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2520 int write_access
, pte_t orig_pte
)
2522 unsigned int flags
= FAULT_FLAG_NONLINEAR
|
2523 (write_access
? FAULT_FLAG_WRITE
: 0);
2526 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2529 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
) ||
2530 !(vma
->vm_flags
& VM_CAN_NONLINEAR
))) {
2532 * Page table corrupted: show pte and kill process.
2534 print_bad_pte(vma
, orig_pte
, address
);
2535 return VM_FAULT_OOM
;
2538 pgoff
= pte_to_pgoff(orig_pte
);
2539 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
2543 * These routines also need to handle stuff like marking pages dirty
2544 * and/or accessed for architectures that don't do it in hardware (most
2545 * RISC architectures). The early dirtying is also good on the i386.
2547 * There is also a hook called "update_mmu_cache()" that architectures
2548 * with external mmu caches can use to update those (ie the Sparc or
2549 * PowerPC hashed page tables that act as extended TLBs).
2551 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2552 * but allow concurrent faults), and pte mapped but not yet locked.
2553 * We return with mmap_sem still held, but pte unmapped and unlocked.
2555 static inline int handle_pte_fault(struct mm_struct
*mm
,
2556 struct vm_area_struct
*vma
, unsigned long address
,
2557 pte_t
*pte
, pmd_t
*pmd
, int write_access
)
2563 if (!pte_present(entry
)) {
2564 if (pte_none(entry
)) {
2566 if (likely(vma
->vm_ops
->fault
))
2567 return do_linear_fault(mm
, vma
, address
,
2568 pte
, pmd
, write_access
, entry
);
2570 return do_anonymous_page(mm
, vma
, address
,
2571 pte
, pmd
, write_access
);
2573 if (pte_file(entry
))
2574 return do_nonlinear_fault(mm
, vma
, address
,
2575 pte
, pmd
, write_access
, entry
);
2576 return do_swap_page(mm
, vma
, address
,
2577 pte
, pmd
, write_access
, entry
);
2580 ptl
= pte_lockptr(mm
, pmd
);
2582 if (unlikely(!pte_same(*pte
, entry
)))
2585 if (!pte_write(entry
))
2586 return do_wp_page(mm
, vma
, address
,
2587 pte
, pmd
, ptl
, entry
);
2588 entry
= pte_mkdirty(entry
);
2590 entry
= pte_mkyoung(entry
);
2591 if (ptep_set_access_flags(vma
, address
, pte
, entry
, write_access
)) {
2592 update_mmu_cache(vma
, address
, entry
);
2595 * This is needed only for protection faults but the arch code
2596 * is not yet telling us if this is a protection fault or not.
2597 * This still avoids useless tlb flushes for .text page faults
2601 flush_tlb_page(vma
, address
);
2604 pte_unmap_unlock(pte
, ptl
);
2609 * By the time we get here, we already hold the mm semaphore
2611 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2612 unsigned long address
, int write_access
)
2619 __set_current_state(TASK_RUNNING
);
2621 count_vm_event(PGFAULT
);
2623 if (unlikely(is_vm_hugetlb_page(vma
)))
2624 return hugetlb_fault(mm
, vma
, address
, write_access
);
2626 pgd
= pgd_offset(mm
, address
);
2627 pud
= pud_alloc(mm
, pgd
, address
);
2629 return VM_FAULT_OOM
;
2630 pmd
= pmd_alloc(mm
, pud
, address
);
2632 return VM_FAULT_OOM
;
2633 pte
= pte_alloc_map(mm
, pmd
, address
);
2635 return VM_FAULT_OOM
;
2637 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, write_access
);
2640 #ifndef __PAGETABLE_PUD_FOLDED
2642 * Allocate page upper directory.
2643 * We've already handled the fast-path in-line.
2645 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
2647 pud_t
*new = pud_alloc_one(mm
, address
);
2651 smp_wmb(); /* See comment in __pte_alloc */
2653 spin_lock(&mm
->page_table_lock
);
2654 if (pgd_present(*pgd
)) /* Another has populated it */
2657 pgd_populate(mm
, pgd
, new);
2658 spin_unlock(&mm
->page_table_lock
);
2661 #endif /* __PAGETABLE_PUD_FOLDED */
2663 #ifndef __PAGETABLE_PMD_FOLDED
2665 * Allocate page middle directory.
2666 * We've already handled the fast-path in-line.
2668 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
2670 pmd_t
*new = pmd_alloc_one(mm
, address
);
2674 smp_wmb(); /* See comment in __pte_alloc */
2676 spin_lock(&mm
->page_table_lock
);
2677 #ifndef __ARCH_HAS_4LEVEL_HACK
2678 if (pud_present(*pud
)) /* Another has populated it */
2681 pud_populate(mm
, pud
, new);
2683 if (pgd_present(*pud
)) /* Another has populated it */
2686 pgd_populate(mm
, pud
, new);
2687 #endif /* __ARCH_HAS_4LEVEL_HACK */
2688 spin_unlock(&mm
->page_table_lock
);
2691 #endif /* __PAGETABLE_PMD_FOLDED */
2693 int make_pages_present(unsigned long addr
, unsigned long end
)
2695 int ret
, len
, write
;
2696 struct vm_area_struct
* vma
;
2698 vma
= find_vma(current
->mm
, addr
);
2701 write
= (vma
->vm_flags
& VM_WRITE
) != 0;
2702 BUG_ON(addr
>= end
);
2703 BUG_ON(end
> vma
->vm_end
);
2704 len
= DIV_ROUND_UP(end
, PAGE_SIZE
) - addr
/PAGE_SIZE
;
2705 ret
= get_user_pages(current
, current
->mm
, addr
,
2706 len
, write
, 0, NULL
, NULL
);
2709 return ret
== len
? 0 : -1;
2712 #if !defined(__HAVE_ARCH_GATE_AREA)
2714 #if defined(AT_SYSINFO_EHDR)
2715 static struct vm_area_struct gate_vma
;
2717 static int __init
gate_vma_init(void)
2719 gate_vma
.vm_mm
= NULL
;
2720 gate_vma
.vm_start
= FIXADDR_USER_START
;
2721 gate_vma
.vm_end
= FIXADDR_USER_END
;
2722 gate_vma
.vm_flags
= VM_READ
| VM_MAYREAD
| VM_EXEC
| VM_MAYEXEC
;
2723 gate_vma
.vm_page_prot
= __P101
;
2725 * Make sure the vDSO gets into every core dump.
2726 * Dumping its contents makes post-mortem fully interpretable later
2727 * without matching up the same kernel and hardware config to see
2728 * what PC values meant.
2730 gate_vma
.vm_flags
|= VM_ALWAYSDUMP
;
2733 __initcall(gate_vma_init
);
2736 struct vm_area_struct
*get_gate_vma(struct task_struct
*tsk
)
2738 #ifdef AT_SYSINFO_EHDR
2745 int in_gate_area_no_task(unsigned long addr
)
2747 #ifdef AT_SYSINFO_EHDR
2748 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
2754 #endif /* __HAVE_ARCH_GATE_AREA */
2756 #ifdef CONFIG_HAVE_IOREMAP_PROT
2757 static resource_size_t
follow_phys(struct vm_area_struct
*vma
,
2758 unsigned long address
, unsigned int flags
,
2759 unsigned long *prot
)
2766 resource_size_t phys_addr
= 0;
2767 struct mm_struct
*mm
= vma
->vm_mm
;
2769 VM_BUG_ON(!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)));
2771 pgd
= pgd_offset(mm
, address
);
2772 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
2775 pud
= pud_offset(pgd
, address
);
2776 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
2779 pmd
= pmd_offset(pud
, address
);
2780 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
2783 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
2787 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2792 if (!pte_present(pte
))
2794 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
2796 phys_addr
= pte_pfn(pte
);
2797 phys_addr
<<= PAGE_SHIFT
; /* Shift here to avoid overflow on PAE */
2799 *prot
= pgprot_val(pte_pgprot(pte
));
2802 pte_unmap_unlock(ptep
, ptl
);
2809 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
2810 void *buf
, int len
, int write
)
2812 resource_size_t phys_addr
;
2813 unsigned long prot
= 0;
2815 int offset
= addr
& (PAGE_SIZE
-1);
2817 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
2820 phys_addr
= follow_phys(vma
, addr
, write
, &prot
);
2825 maddr
= ioremap_prot(phys_addr
, PAGE_SIZE
, prot
);
2827 memcpy_toio(maddr
+ offset
, buf
, len
);
2829 memcpy_fromio(buf
, maddr
+ offset
, len
);
2837 * Access another process' address space.
2838 * Source/target buffer must be kernel space,
2839 * Do not walk the page table directly, use get_user_pages
2841 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
, void *buf
, int len
, int write
)
2843 struct mm_struct
*mm
;
2844 struct vm_area_struct
*vma
;
2845 void *old_buf
= buf
;
2847 mm
= get_task_mm(tsk
);
2851 down_read(&mm
->mmap_sem
);
2852 /* ignore errors, just check how much was successfully transferred */
2854 int bytes
, ret
, offset
;
2856 struct page
*page
= NULL
;
2858 ret
= get_user_pages(tsk
, mm
, addr
, 1,
2859 write
, 1, &page
, &vma
);
2862 * Check if this is a VM_IO | VM_PFNMAP VMA, which
2863 * we can access using slightly different code.
2865 #ifdef CONFIG_HAVE_IOREMAP_PROT
2866 vma
= find_vma(mm
, addr
);
2869 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
2870 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
2878 offset
= addr
& (PAGE_SIZE
-1);
2879 if (bytes
> PAGE_SIZE
-offset
)
2880 bytes
= PAGE_SIZE
-offset
;
2884 copy_to_user_page(vma
, page
, addr
,
2885 maddr
+ offset
, buf
, bytes
);
2886 set_page_dirty_lock(page
);
2888 copy_from_user_page(vma
, page
, addr
,
2889 buf
, maddr
+ offset
, bytes
);
2892 page_cache_release(page
);
2898 up_read(&mm
->mmap_sem
);
2901 return buf
- old_buf
;
2905 * Print the name of a VMA.
2907 void print_vma_addr(char *prefix
, unsigned long ip
)
2909 struct mm_struct
*mm
= current
->mm
;
2910 struct vm_area_struct
*vma
;
2913 * Do not print if we are in atomic
2914 * contexts (in exception stacks, etc.):
2916 if (preempt_count())
2919 down_read(&mm
->mmap_sem
);
2920 vma
= find_vma(mm
, ip
);
2921 if (vma
&& vma
->vm_file
) {
2922 struct file
*f
= vma
->vm_file
;
2923 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
2927 p
= d_path(&f
->f_path
, buf
, PAGE_SIZE
);
2930 s
= strrchr(p
, '/');
2933 printk("%s%s[%lx+%lx]", prefix
, p
,
2935 vma
->vm_end
- vma
->vm_start
);
2936 free_page((unsigned long)buf
);
2939 up_read(¤t
->mm
->mmap_sem
);