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/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/module.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
62 #include <asm/pgalloc.h>
63 #include <asm/uaccess.h>
65 #include <asm/tlbflush.h>
66 #include <asm/pgtable.h>
70 #ifndef CONFIG_NEED_MULTIPLE_NODES
71 /* use the per-pgdat data instead for discontigmem - mbligh */
72 unsigned long max_mapnr
;
75 EXPORT_SYMBOL(max_mapnr
);
76 EXPORT_SYMBOL(mem_map
);
79 unsigned long num_physpages
;
81 * A number of key systems in x86 including ioremap() rely on the assumption
82 * that high_memory defines the upper bound on direct map memory, then end
83 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
84 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
89 EXPORT_SYMBOL(num_physpages
);
90 EXPORT_SYMBOL(high_memory
);
93 * Randomize the address space (stacks, mmaps, brk, etc.).
95 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96 * as ancient (libc5 based) binaries can segfault. )
98 int randomize_va_space __read_mostly
=
99 #ifdef CONFIG_COMPAT_BRK
105 static int __init
disable_randmaps(char *s
)
107 randomize_va_space
= 0;
110 __setup("norandmaps", disable_randmaps
);
112 unsigned long zero_pfn __read_mostly
;
113 unsigned long highest_memmap_pfn __read_mostly
;
116 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
118 static int __init
init_zero_pfn(void)
120 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
123 core_initcall(init_zero_pfn
);
126 #if defined(SPLIT_RSS_COUNTING)
128 static void __sync_task_rss_stat(struct task_struct
*task
, struct mm_struct
*mm
)
132 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
133 if (task
->rss_stat
.count
[i
]) {
134 add_mm_counter(mm
, i
, task
->rss_stat
.count
[i
]);
135 task
->rss_stat
.count
[i
] = 0;
138 task
->rss_stat
.events
= 0;
141 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
143 struct task_struct
*task
= current
;
145 if (likely(task
->mm
== mm
))
146 task
->rss_stat
.count
[member
] += val
;
148 add_mm_counter(mm
, member
, val
);
150 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
153 /* sync counter once per 64 page faults */
154 #define TASK_RSS_EVENTS_THRESH (64)
155 static void check_sync_rss_stat(struct task_struct
*task
)
157 if (unlikely(task
!= current
))
159 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
160 __sync_task_rss_stat(task
, task
->mm
);
163 unsigned long get_mm_counter(struct mm_struct
*mm
, int member
)
168 * Don't use task->mm here...for avoiding to use task_get_mm()..
169 * The caller must guarantee task->mm is not invalid.
171 val
= atomic_long_read(&mm
->rss_stat
.count
[member
]);
173 * counter is updated in asynchronous manner and may go to minus.
174 * But it's never be expected number for users.
178 return (unsigned long)val
;
181 void sync_mm_rss(struct task_struct
*task
, struct mm_struct
*mm
)
183 __sync_task_rss_stat(task
, mm
);
187 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
188 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
190 static void check_sync_rss_stat(struct task_struct
*task
)
197 * If a p?d_bad entry is found while walking page tables, report
198 * the error, before resetting entry to p?d_none. Usually (but
199 * very seldom) called out from the p?d_none_or_clear_bad macros.
202 void pgd_clear_bad(pgd_t
*pgd
)
208 void pud_clear_bad(pud_t
*pud
)
214 void pmd_clear_bad(pmd_t
*pmd
)
221 * Note: this doesn't free the actual pages themselves. That
222 * has been handled earlier when unmapping all the memory regions.
224 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
227 pgtable_t token
= pmd_pgtable(*pmd
);
229 pte_free_tlb(tlb
, token
, addr
);
233 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
234 unsigned long addr
, unsigned long end
,
235 unsigned long floor
, unsigned long ceiling
)
242 pmd
= pmd_offset(pud
, addr
);
244 next
= pmd_addr_end(addr
, end
);
245 if (pmd_none_or_clear_bad(pmd
))
247 free_pte_range(tlb
, pmd
, addr
);
248 } while (pmd
++, addr
= next
, addr
!= end
);
258 if (end
- 1 > ceiling
- 1)
261 pmd
= pmd_offset(pud
, start
);
263 pmd_free_tlb(tlb
, pmd
, start
);
266 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
267 unsigned long addr
, unsigned long end
,
268 unsigned long floor
, unsigned long ceiling
)
275 pud
= pud_offset(pgd
, addr
);
277 next
= pud_addr_end(addr
, end
);
278 if (pud_none_or_clear_bad(pud
))
280 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
281 } while (pud
++, addr
= next
, addr
!= end
);
287 ceiling
&= PGDIR_MASK
;
291 if (end
- 1 > ceiling
- 1)
294 pud
= pud_offset(pgd
, start
);
296 pud_free_tlb(tlb
, pud
, start
);
300 * This function frees user-level page tables of a process.
302 * Must be called with pagetable lock held.
304 void free_pgd_range(struct mmu_gather
*tlb
,
305 unsigned long addr
, unsigned long end
,
306 unsigned long floor
, unsigned long ceiling
)
312 * The next few lines have given us lots of grief...
314 * Why are we testing PMD* at this top level? Because often
315 * there will be no work to do at all, and we'd prefer not to
316 * go all the way down to the bottom just to discover that.
318 * Why all these "- 1"s? Because 0 represents both the bottom
319 * of the address space and the top of it (using -1 for the
320 * top wouldn't help much: the masks would do the wrong thing).
321 * The rule is that addr 0 and floor 0 refer to the bottom of
322 * the address space, but end 0 and ceiling 0 refer to the top
323 * Comparisons need to use "end - 1" and "ceiling - 1" (though
324 * that end 0 case should be mythical).
326 * Wherever addr is brought up or ceiling brought down, we must
327 * be careful to reject "the opposite 0" before it confuses the
328 * subsequent tests. But what about where end is brought down
329 * by PMD_SIZE below? no, end can't go down to 0 there.
331 * Whereas we round start (addr) and ceiling down, by different
332 * masks at different levels, in order to test whether a table
333 * now has no other vmas using it, so can be freed, we don't
334 * bother to round floor or end up - the tests don't need that.
348 if (end
- 1 > ceiling
- 1)
353 pgd
= pgd_offset(tlb
->mm
, addr
);
355 next
= pgd_addr_end(addr
, end
);
356 if (pgd_none_or_clear_bad(pgd
))
358 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
359 } while (pgd
++, addr
= next
, addr
!= end
);
362 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
363 unsigned long floor
, unsigned long ceiling
)
366 struct vm_area_struct
*next
= vma
->vm_next
;
367 unsigned long addr
= vma
->vm_start
;
370 * Hide vma from rmap and truncate_pagecache before freeing
373 unlink_anon_vmas(vma
);
374 unlink_file_vma(vma
);
376 if (is_vm_hugetlb_page(vma
)) {
377 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
378 floor
, next
? next
->vm_start
: ceiling
);
381 * Optimization: gather nearby vmas into one call down
383 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
384 && !is_vm_hugetlb_page(next
)) {
387 unlink_anon_vmas(vma
);
388 unlink_file_vma(vma
);
390 free_pgd_range(tlb
, addr
, vma
->vm_end
,
391 floor
, next
? next
->vm_start
: ceiling
);
397 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
, unsigned long address
)
399 pgtable_t
new = pte_alloc_one(mm
, address
);
404 * Ensure all pte setup (eg. pte page lock and page clearing) are
405 * visible before the pte is made visible to other CPUs by being
406 * put into page tables.
408 * The other side of the story is the pointer chasing in the page
409 * table walking code (when walking the page table without locking;
410 * ie. most of the time). Fortunately, these data accesses consist
411 * of a chain of data-dependent loads, meaning most CPUs (alpha
412 * being the notable exception) will already guarantee loads are
413 * seen in-order. See the alpha page table accessors for the
414 * smp_read_barrier_depends() barriers in page table walking code.
416 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
418 spin_lock(&mm
->page_table_lock
);
419 if (!pmd_present(*pmd
)) { /* Has another populated it ? */
421 pmd_populate(mm
, pmd
, new);
424 spin_unlock(&mm
->page_table_lock
);
430 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
432 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
436 smp_wmb(); /* See comment in __pte_alloc */
438 spin_lock(&init_mm
.page_table_lock
);
439 if (!pmd_present(*pmd
)) { /* Has another populated it ? */
440 pmd_populate_kernel(&init_mm
, pmd
, new);
443 spin_unlock(&init_mm
.page_table_lock
);
445 pte_free_kernel(&init_mm
, new);
449 static inline void init_rss_vec(int *rss
)
451 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
454 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
458 if (current
->mm
== mm
)
459 sync_mm_rss(current
, mm
);
460 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
462 add_mm_counter(mm
, i
, rss
[i
]);
466 * This function is called to print an error when a bad pte
467 * is found. For example, we might have a PFN-mapped pte in
468 * a region that doesn't allow it.
470 * The calling function must still handle the error.
472 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
473 pte_t pte
, struct page
*page
)
475 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
476 pud_t
*pud
= pud_offset(pgd
, addr
);
477 pmd_t
*pmd
= pmd_offset(pud
, addr
);
478 struct address_space
*mapping
;
480 static unsigned long resume
;
481 static unsigned long nr_shown
;
482 static unsigned long nr_unshown
;
485 * Allow a burst of 60 reports, then keep quiet for that minute;
486 * or allow a steady drip of one report per second.
488 if (nr_shown
== 60) {
489 if (time_before(jiffies
, resume
)) {
495 "BUG: Bad page map: %lu messages suppressed\n",
502 resume
= jiffies
+ 60 * HZ
;
504 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
505 index
= linear_page_index(vma
, addr
);
508 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
510 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
514 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
515 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
517 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
520 print_symbol(KERN_ALERT
"vma->vm_ops->fault: %s\n",
521 (unsigned long)vma
->vm_ops
->fault
);
522 if (vma
->vm_file
&& vma
->vm_file
->f_op
)
523 print_symbol(KERN_ALERT
"vma->vm_file->f_op->mmap: %s\n",
524 (unsigned long)vma
->vm_file
->f_op
->mmap
);
526 add_taint(TAINT_BAD_PAGE
);
529 static inline int is_cow_mapping(unsigned int flags
)
531 return (flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
535 static inline int is_zero_pfn(unsigned long pfn
)
537 return pfn
== zero_pfn
;
542 static inline unsigned long my_zero_pfn(unsigned long addr
)
549 * vm_normal_page -- This function gets the "struct page" associated with a pte.
551 * "Special" mappings do not wish to be associated with a "struct page" (either
552 * it doesn't exist, or it exists but they don't want to touch it). In this
553 * case, NULL is returned here. "Normal" mappings do have a struct page.
555 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
556 * pte bit, in which case this function is trivial. Secondly, an architecture
557 * may not have a spare pte bit, which requires a more complicated scheme,
560 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
561 * special mapping (even if there are underlying and valid "struct pages").
562 * COWed pages of a VM_PFNMAP are always normal.
564 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
565 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
566 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
567 * mapping will always honor the rule
569 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
571 * And for normal mappings this is false.
573 * This restricts such mappings to be a linear translation from virtual address
574 * to pfn. To get around this restriction, we allow arbitrary mappings so long
575 * as the vma is not a COW mapping; in that case, we know that all ptes are
576 * special (because none can have been COWed).
579 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
581 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
582 * page" backing, however the difference is that _all_ pages with a struct
583 * page (that is, those where pfn_valid is true) are refcounted and considered
584 * normal pages by the VM. The disadvantage is that pages are refcounted
585 * (which can be slower and simply not an option for some PFNMAP users). The
586 * advantage is that we don't have to follow the strict linearity rule of
587 * PFNMAP mappings in order to support COWable mappings.
590 #ifdef __HAVE_ARCH_PTE_SPECIAL
591 # define HAVE_PTE_SPECIAL 1
593 # define HAVE_PTE_SPECIAL 0
595 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
598 unsigned long pfn
= pte_pfn(pte
);
600 if (HAVE_PTE_SPECIAL
) {
601 if (likely(!pte_special(pte
)))
603 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
605 if (!is_zero_pfn(pfn
))
606 print_bad_pte(vma
, addr
, pte
, NULL
);
610 /* !HAVE_PTE_SPECIAL case follows: */
612 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
613 if (vma
->vm_flags
& VM_MIXEDMAP
) {
619 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
620 if (pfn
== vma
->vm_pgoff
+ off
)
622 if (!is_cow_mapping(vma
->vm_flags
))
627 if (is_zero_pfn(pfn
))
630 if (unlikely(pfn
> highest_memmap_pfn
)) {
631 print_bad_pte(vma
, addr
, pte
, NULL
);
636 * NOTE! We still have PageReserved() pages in the page tables.
637 * eg. VDSO mappings can cause them to exist.
640 return pfn_to_page(pfn
);
644 * copy one vm_area from one task to the other. Assumes the page tables
645 * already present in the new task to be cleared in the whole range
646 * covered by this vma.
649 static inline unsigned long
650 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
651 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
652 unsigned long addr
, int *rss
)
654 unsigned long vm_flags
= vma
->vm_flags
;
655 pte_t pte
= *src_pte
;
658 /* pte contains position in swap or file, so copy. */
659 if (unlikely(!pte_present(pte
))) {
660 if (!pte_file(pte
)) {
661 swp_entry_t entry
= pte_to_swp_entry(pte
);
663 if (swap_duplicate(entry
) < 0)
666 /* make sure dst_mm is on swapoff's mmlist. */
667 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
668 spin_lock(&mmlist_lock
);
669 if (list_empty(&dst_mm
->mmlist
))
670 list_add(&dst_mm
->mmlist
,
672 spin_unlock(&mmlist_lock
);
674 if (likely(!non_swap_entry(entry
)))
676 else if (is_write_migration_entry(entry
) &&
677 is_cow_mapping(vm_flags
)) {
679 * COW mappings require pages in both parent
680 * and child to be set to read.
682 make_migration_entry_read(&entry
);
683 pte
= swp_entry_to_pte(entry
);
684 set_pte_at(src_mm
, addr
, src_pte
, pte
);
691 * If it's a COW mapping, write protect it both
692 * in the parent and the child
694 if (is_cow_mapping(vm_flags
)) {
695 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
696 pte
= pte_wrprotect(pte
);
700 * If it's a shared mapping, mark it clean in
703 if (vm_flags
& VM_SHARED
)
704 pte
= pte_mkclean(pte
);
705 pte
= pte_mkold(pte
);
707 page
= vm_normal_page(vma
, addr
, pte
);
718 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
722 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
723 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
724 unsigned long addr
, unsigned long end
)
726 pte_t
*orig_src_pte
, *orig_dst_pte
;
727 pte_t
*src_pte
, *dst_pte
;
728 spinlock_t
*src_ptl
, *dst_ptl
;
730 int rss
[NR_MM_COUNTERS
];
731 swp_entry_t entry
= (swp_entry_t
){0};
736 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
739 src_pte
= pte_offset_map(src_pmd
, addr
);
740 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
741 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
742 orig_src_pte
= src_pte
;
743 orig_dst_pte
= dst_pte
;
744 arch_enter_lazy_mmu_mode();
748 * We are holding two locks at this point - either of them
749 * could generate latencies in another task on another CPU.
751 if (progress
>= 32) {
753 if (need_resched() ||
754 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
757 if (pte_none(*src_pte
)) {
761 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
766 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
768 arch_leave_lazy_mmu_mode();
769 spin_unlock(src_ptl
);
770 pte_unmap(orig_src_pte
);
771 add_mm_rss_vec(dst_mm
, rss
);
772 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
776 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
785 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
786 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
787 unsigned long addr
, unsigned long end
)
789 pmd_t
*src_pmd
, *dst_pmd
;
792 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
795 src_pmd
= pmd_offset(src_pud
, addr
);
797 next
= pmd_addr_end(addr
, end
);
798 if (pmd_none_or_clear_bad(src_pmd
))
800 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
803 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
807 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
808 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
809 unsigned long addr
, unsigned long end
)
811 pud_t
*src_pud
, *dst_pud
;
814 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
817 src_pud
= pud_offset(src_pgd
, addr
);
819 next
= pud_addr_end(addr
, end
);
820 if (pud_none_or_clear_bad(src_pud
))
822 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
825 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
829 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
830 struct vm_area_struct
*vma
)
832 pgd_t
*src_pgd
, *dst_pgd
;
834 unsigned long addr
= vma
->vm_start
;
835 unsigned long end
= vma
->vm_end
;
839 * Don't copy ptes where a page fault will fill them correctly.
840 * Fork becomes much lighter when there are big shared or private
841 * readonly mappings. The tradeoff is that copy_page_range is more
842 * efficient than faulting.
844 if (!(vma
->vm_flags
& (VM_HUGETLB
|VM_NONLINEAR
|VM_PFNMAP
|VM_INSERTPAGE
))) {
849 if (is_vm_hugetlb_page(vma
))
850 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
852 if (unlikely(is_pfn_mapping(vma
))) {
854 * We do not free on error cases below as remove_vma
855 * gets called on error from higher level routine
857 ret
= track_pfn_vma_copy(vma
);
863 * We need to invalidate the secondary MMU mappings only when
864 * there could be a permission downgrade on the ptes of the
865 * parent mm. And a permission downgrade will only happen if
866 * is_cow_mapping() returns true.
868 if (is_cow_mapping(vma
->vm_flags
))
869 mmu_notifier_invalidate_range_start(src_mm
, addr
, end
);
872 dst_pgd
= pgd_offset(dst_mm
, addr
);
873 src_pgd
= pgd_offset(src_mm
, addr
);
875 next
= pgd_addr_end(addr
, end
);
876 if (pgd_none_or_clear_bad(src_pgd
))
878 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
883 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
885 if (is_cow_mapping(vma
->vm_flags
))
886 mmu_notifier_invalidate_range_end(src_mm
,
891 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
892 struct vm_area_struct
*vma
, pmd_t
*pmd
,
893 unsigned long addr
, unsigned long end
,
894 long *zap_work
, struct zap_details
*details
)
896 struct mm_struct
*mm
= tlb
->mm
;
899 int rss
[NR_MM_COUNTERS
];
903 pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
904 arch_enter_lazy_mmu_mode();
907 if (pte_none(ptent
)) {
912 (*zap_work
) -= PAGE_SIZE
;
914 if (pte_present(ptent
)) {
917 page
= vm_normal_page(vma
, addr
, ptent
);
918 if (unlikely(details
) && page
) {
920 * unmap_shared_mapping_pages() wants to
921 * invalidate cache without truncating:
922 * unmap shared but keep private pages.
924 if (details
->check_mapping
&&
925 details
->check_mapping
!= page
->mapping
)
928 * Each page->index must be checked when
929 * invalidating or truncating nonlinear.
931 if (details
->nonlinear_vma
&&
932 (page
->index
< details
->first_index
||
933 page
->index
> details
->last_index
))
936 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
938 tlb_remove_tlb_entry(tlb
, pte
, addr
);
941 if (unlikely(details
) && details
->nonlinear_vma
942 && linear_page_index(details
->nonlinear_vma
,
943 addr
) != page
->index
)
944 set_pte_at(mm
, addr
, pte
,
945 pgoff_to_pte(page
->index
));
949 if (pte_dirty(ptent
))
950 set_page_dirty(page
);
951 if (pte_young(ptent
) &&
952 likely(!VM_SequentialReadHint(vma
)))
953 mark_page_accessed(page
);
956 page_remove_rmap(page
);
957 if (unlikely(page_mapcount(page
) < 0))
958 print_bad_pte(vma
, addr
, ptent
, page
);
959 tlb_remove_page(tlb
, page
);
963 * If details->check_mapping, we leave swap entries;
964 * if details->nonlinear_vma, we leave file entries.
966 if (unlikely(details
))
968 if (pte_file(ptent
)) {
969 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
)))
970 print_bad_pte(vma
, addr
, ptent
, NULL
);
972 swp_entry_t entry
= pte_to_swp_entry(ptent
);
974 if (!non_swap_entry(entry
))
976 if (unlikely(!free_swap_and_cache(entry
)))
977 print_bad_pte(vma
, addr
, ptent
, NULL
);
979 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
980 } while (pte
++, addr
+= PAGE_SIZE
, (addr
!= end
&& *zap_work
> 0));
982 add_mm_rss_vec(mm
, rss
);
983 arch_leave_lazy_mmu_mode();
984 pte_unmap_unlock(pte
- 1, ptl
);
989 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
990 struct vm_area_struct
*vma
, pud_t
*pud
,
991 unsigned long addr
, unsigned long end
,
992 long *zap_work
, struct zap_details
*details
)
997 pmd
= pmd_offset(pud
, addr
);
999 next
= pmd_addr_end(addr
, end
);
1000 if (pmd_none_or_clear_bad(pmd
)) {
1004 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
,
1006 } while (pmd
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
1011 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1012 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1013 unsigned long addr
, unsigned long end
,
1014 long *zap_work
, struct zap_details
*details
)
1019 pud
= pud_offset(pgd
, addr
);
1021 next
= pud_addr_end(addr
, end
);
1022 if (pud_none_or_clear_bad(pud
)) {
1026 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
,
1028 } while (pud
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
1033 static unsigned long unmap_page_range(struct mmu_gather
*tlb
,
1034 struct vm_area_struct
*vma
,
1035 unsigned long addr
, unsigned long end
,
1036 long *zap_work
, struct zap_details
*details
)
1041 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
1044 BUG_ON(addr
>= end
);
1045 mem_cgroup_uncharge_start();
1046 tlb_start_vma(tlb
, vma
);
1047 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1049 next
= pgd_addr_end(addr
, end
);
1050 if (pgd_none_or_clear_bad(pgd
)) {
1054 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
,
1056 } while (pgd
++, addr
= next
, (addr
!= end
&& *zap_work
> 0));
1057 tlb_end_vma(tlb
, vma
);
1058 mem_cgroup_uncharge_end();
1063 #ifdef CONFIG_PREEMPT
1064 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
1066 /* No preempt: go for improved straight-line efficiency */
1067 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
1071 * unmap_vmas - unmap a range of memory covered by a list of vma's
1072 * @tlbp: address of the caller's struct mmu_gather
1073 * @vma: the starting vma
1074 * @start_addr: virtual address at which to start unmapping
1075 * @end_addr: virtual address at which to end unmapping
1076 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1077 * @details: details of nonlinear truncation or shared cache invalidation
1079 * Returns the end address of the unmapping (restart addr if interrupted).
1081 * Unmap all pages in the vma list.
1083 * We aim to not hold locks for too long (for scheduling latency reasons).
1084 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
1085 * return the ending mmu_gather to the caller.
1087 * Only addresses between `start' and `end' will be unmapped.
1089 * The VMA list must be sorted in ascending virtual address order.
1091 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1092 * range after unmap_vmas() returns. So the only responsibility here is to
1093 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1094 * drops the lock and schedules.
1096 unsigned long unmap_vmas(struct mmu_gather
**tlbp
,
1097 struct vm_area_struct
*vma
, unsigned long start_addr
,
1098 unsigned long end_addr
, unsigned long *nr_accounted
,
1099 struct zap_details
*details
)
1101 long zap_work
= ZAP_BLOCK_SIZE
;
1102 unsigned long tlb_start
= 0; /* For tlb_finish_mmu */
1103 int tlb_start_valid
= 0;
1104 unsigned long start
= start_addr
;
1105 spinlock_t
*i_mmap_lock
= details
? details
->i_mmap_lock
: NULL
;
1106 int fullmm
= (*tlbp
)->fullmm
;
1107 struct mm_struct
*mm
= vma
->vm_mm
;
1109 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1110 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
) {
1113 start
= max(vma
->vm_start
, start_addr
);
1114 if (start
>= vma
->vm_end
)
1116 end
= min(vma
->vm_end
, end_addr
);
1117 if (end
<= vma
->vm_start
)
1120 if (vma
->vm_flags
& VM_ACCOUNT
)
1121 *nr_accounted
+= (end
- start
) >> PAGE_SHIFT
;
1123 if (unlikely(is_pfn_mapping(vma
)))
1124 untrack_pfn_vma(vma
, 0, 0);
1126 while (start
!= end
) {
1127 if (!tlb_start_valid
) {
1129 tlb_start_valid
= 1;
1132 if (unlikely(is_vm_hugetlb_page(vma
))) {
1134 * It is undesirable to test vma->vm_file as it
1135 * should be non-null for valid hugetlb area.
1136 * However, vm_file will be NULL in the error
1137 * cleanup path of do_mmap_pgoff. When
1138 * hugetlbfs ->mmap method fails,
1139 * do_mmap_pgoff() nullifies vma->vm_file
1140 * before calling this function to clean up.
1141 * Since no pte has actually been setup, it is
1142 * safe to do nothing in this case.
1145 unmap_hugepage_range(vma
, start
, end
, NULL
);
1146 zap_work
-= (end
- start
) /
1147 pages_per_huge_page(hstate_vma(vma
));
1152 start
= unmap_page_range(*tlbp
, vma
,
1153 start
, end
, &zap_work
, details
);
1156 BUG_ON(start
!= end
);
1160 tlb_finish_mmu(*tlbp
, tlb_start
, start
);
1162 if (need_resched() ||
1163 (i_mmap_lock
&& spin_needbreak(i_mmap_lock
))) {
1171 *tlbp
= tlb_gather_mmu(vma
->vm_mm
, fullmm
);
1172 tlb_start_valid
= 0;
1173 zap_work
= ZAP_BLOCK_SIZE
;
1177 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1178 return start
; /* which is now the end (or restart) address */
1182 * zap_page_range - remove user pages in a given range
1183 * @vma: vm_area_struct holding the applicable pages
1184 * @address: starting address of pages to zap
1185 * @size: number of bytes to zap
1186 * @details: details of nonlinear truncation or shared cache invalidation
1188 unsigned long zap_page_range(struct vm_area_struct
*vma
, unsigned long address
,
1189 unsigned long size
, struct zap_details
*details
)
1191 struct mm_struct
*mm
= vma
->vm_mm
;
1192 struct mmu_gather
*tlb
;
1193 unsigned long end
= address
+ size
;
1194 unsigned long nr_accounted
= 0;
1197 tlb
= tlb_gather_mmu(mm
, 0);
1198 update_hiwater_rss(mm
);
1199 end
= unmap_vmas(&tlb
, vma
, address
, end
, &nr_accounted
, details
);
1201 tlb_finish_mmu(tlb
, address
, end
);
1206 * zap_vma_ptes - remove ptes mapping the vma
1207 * @vma: vm_area_struct holding ptes to be zapped
1208 * @address: starting address of pages to zap
1209 * @size: number of bytes to zap
1211 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1213 * The entire address range must be fully contained within the vma.
1215 * Returns 0 if successful.
1217 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1220 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1221 !(vma
->vm_flags
& VM_PFNMAP
))
1223 zap_page_range(vma
, address
, size
, NULL
);
1226 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1229 * follow_page - look up a page descriptor from a user-virtual address
1230 * @vma: vm_area_struct mapping @address
1231 * @address: virtual address to look up
1232 * @flags: flags modifying lookup behaviour
1234 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1236 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1237 * an error pointer if there is a mapping to something not represented
1238 * by a page descriptor (see also vm_normal_page()).
1240 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
1249 struct mm_struct
*mm
= vma
->vm_mm
;
1251 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
1252 if (!IS_ERR(page
)) {
1253 BUG_ON(flags
& FOLL_GET
);
1258 pgd
= pgd_offset(mm
, address
);
1259 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
1262 pud
= pud_offset(pgd
, address
);
1265 if (pud_huge(*pud
)) {
1266 BUG_ON(flags
& FOLL_GET
);
1267 page
= follow_huge_pud(mm
, address
, pud
, flags
& FOLL_WRITE
);
1270 if (unlikely(pud_bad(*pud
)))
1273 pmd
= pmd_offset(pud
, address
);
1276 if (pmd_huge(*pmd
)) {
1277 BUG_ON(flags
& FOLL_GET
);
1278 page
= follow_huge_pmd(mm
, address
, pmd
, flags
& FOLL_WRITE
);
1281 if (unlikely(pmd_bad(*pmd
)))
1284 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1287 if (!pte_present(pte
))
1289 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
1292 page
= vm_normal_page(vma
, address
, pte
);
1293 if (unlikely(!page
)) {
1294 if ((flags
& FOLL_DUMP
) ||
1295 !is_zero_pfn(pte_pfn(pte
)))
1297 page
= pte_page(pte
);
1300 if (flags
& FOLL_GET
)
1302 if (flags
& FOLL_TOUCH
) {
1303 if ((flags
& FOLL_WRITE
) &&
1304 !pte_dirty(pte
) && !PageDirty(page
))
1305 set_page_dirty(page
);
1307 * pte_mkyoung() would be more correct here, but atomic care
1308 * is needed to avoid losing the dirty bit: it is easier to use
1309 * mark_page_accessed().
1311 mark_page_accessed(page
);
1313 if (flags
& FOLL_MLOCK
) {
1315 * The preliminary mapping check is mainly to avoid the
1316 * pointless overhead of lock_page on the ZERO_PAGE
1317 * which might bounce very badly if there is contention.
1319 * If the page is already locked, we don't need to
1320 * handle it now - vmscan will handle it later if and
1321 * when it attempts to reclaim the page.
1323 if (page
->mapping
&& trylock_page(page
)) {
1324 lru_add_drain(); /* push cached pages to LRU */
1326 * Because we lock page here and migration is
1327 * blocked by the pte's page reference, we need
1328 * only check for file-cache page truncation.
1331 mlock_vma_page(page
);
1336 pte_unmap_unlock(ptep
, ptl
);
1341 pte_unmap_unlock(ptep
, ptl
);
1342 return ERR_PTR(-EFAULT
);
1345 pte_unmap_unlock(ptep
, ptl
);
1351 * When core dumping an enormous anonymous area that nobody
1352 * has touched so far, we don't want to allocate unnecessary pages or
1353 * page tables. Return error instead of NULL to skip handle_mm_fault,
1354 * then get_dump_page() will return NULL to leave a hole in the dump.
1355 * But we can only make this optimization where a hole would surely
1356 * be zero-filled if handle_mm_fault() actually did handle it.
1358 if ((flags
& FOLL_DUMP
) &&
1359 (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
1360 return ERR_PTR(-EFAULT
);
1364 int __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1365 unsigned long start
, int nr_pages
, unsigned int gup_flags
,
1366 struct page
**pages
, struct vm_area_struct
**vmas
)
1369 unsigned long vm_flags
;
1374 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
1377 * Require read or write permissions.
1378 * If FOLL_FORCE is set, we only require the "MAY" flags.
1380 vm_flags
= (gup_flags
& FOLL_WRITE
) ?
1381 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1382 vm_flags
&= (gup_flags
& FOLL_FORCE
) ?
1383 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1387 struct vm_area_struct
*vma
;
1389 vma
= find_extend_vma(mm
, start
);
1390 if (!vma
&& in_gate_area(tsk
, start
)) {
1391 unsigned long pg
= start
& PAGE_MASK
;
1392 struct vm_area_struct
*gate_vma
= get_gate_vma(tsk
);
1398 /* user gate pages are read-only */
1399 if (gup_flags
& FOLL_WRITE
)
1400 return i
? : -EFAULT
;
1402 pgd
= pgd_offset_k(pg
);
1404 pgd
= pgd_offset_gate(mm
, pg
);
1405 BUG_ON(pgd_none(*pgd
));
1406 pud
= pud_offset(pgd
, pg
);
1407 BUG_ON(pud_none(*pud
));
1408 pmd
= pmd_offset(pud
, pg
);
1410 return i
? : -EFAULT
;
1411 pte
= pte_offset_map(pmd
, pg
);
1412 if (pte_none(*pte
)) {
1414 return i
? : -EFAULT
;
1419 page
= vm_normal_page(gate_vma
, start
, *pte
);
1421 if (!(gup_flags
& FOLL_DUMP
) &&
1422 is_zero_pfn(pte_pfn(*pte
)))
1423 page
= pte_page(*pte
);
1426 return i
? : -EFAULT
;
1442 (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1443 !(vm_flags
& vma
->vm_flags
))
1444 return i
? : -EFAULT
;
1446 if (is_vm_hugetlb_page(vma
)) {
1447 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1448 &start
, &nr_pages
, i
, gup_flags
);
1454 unsigned int foll_flags
= gup_flags
;
1457 * If we have a pending SIGKILL, don't keep faulting
1458 * pages and potentially allocating memory.
1460 if (unlikely(fatal_signal_pending(current
)))
1461 return i
? i
: -ERESTARTSYS
;
1464 while (!(page
= follow_page(vma
, start
, foll_flags
))) {
1467 ret
= handle_mm_fault(mm
, vma
, start
,
1468 (foll_flags
& FOLL_WRITE
) ?
1469 FAULT_FLAG_WRITE
: 0);
1471 if (ret
& VM_FAULT_ERROR
) {
1472 if (ret
& VM_FAULT_OOM
)
1473 return i
? i
: -ENOMEM
;
1475 (VM_FAULT_HWPOISON
|VM_FAULT_HWPOISON_LARGE
|
1477 return i
? i
: -EFAULT
;
1480 if (ret
& VM_FAULT_MAJOR
)
1486 * The VM_FAULT_WRITE bit tells us that
1487 * do_wp_page has broken COW when necessary,
1488 * even if maybe_mkwrite decided not to set
1489 * pte_write. We can thus safely do subsequent
1490 * page lookups as if they were reads. But only
1491 * do so when looping for pte_write is futile:
1492 * in some cases userspace may also be wanting
1493 * to write to the gotten user page, which a
1494 * read fault here might prevent (a readonly
1495 * page might get reCOWed by userspace write).
1497 if ((ret
& VM_FAULT_WRITE
) &&
1498 !(vma
->vm_flags
& VM_WRITE
))
1499 foll_flags
&= ~FOLL_WRITE
;
1504 return i
? i
: PTR_ERR(page
);
1508 flush_anon_page(vma
, page
, start
);
1509 flush_dcache_page(page
);
1516 } while (nr_pages
&& start
< vma
->vm_end
);
1522 * get_user_pages() - pin user pages in memory
1523 * @tsk: task_struct of target task
1524 * @mm: mm_struct of target mm
1525 * @start: starting user address
1526 * @nr_pages: number of pages from start to pin
1527 * @write: whether pages will be written to by the caller
1528 * @force: whether to force write access even if user mapping is
1529 * readonly. This will result in the page being COWed even
1530 * in MAP_SHARED mappings. You do not want this.
1531 * @pages: array that receives pointers to the pages pinned.
1532 * Should be at least nr_pages long. Or NULL, if caller
1533 * only intends to ensure the pages are faulted in.
1534 * @vmas: array of pointers to vmas corresponding to each page.
1535 * Or NULL if the caller does not require them.
1537 * Returns number of pages pinned. This may be fewer than the number
1538 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1539 * were pinned, returns -errno. Each page returned must be released
1540 * with a put_page() call when it is finished with. vmas will only
1541 * remain valid while mmap_sem is held.
1543 * Must be called with mmap_sem held for read or write.
1545 * get_user_pages walks a process's page tables and takes a reference to
1546 * each struct page that each user address corresponds to at a given
1547 * instant. That is, it takes the page that would be accessed if a user
1548 * thread accesses the given user virtual address at that instant.
1550 * This does not guarantee that the page exists in the user mappings when
1551 * get_user_pages returns, and there may even be a completely different
1552 * page there in some cases (eg. if mmapped pagecache has been invalidated
1553 * and subsequently re faulted). However it does guarantee that the page
1554 * won't be freed completely. And mostly callers simply care that the page
1555 * contains data that was valid *at some point in time*. Typically, an IO
1556 * or similar operation cannot guarantee anything stronger anyway because
1557 * locks can't be held over the syscall boundary.
1559 * If write=0, the page must not be written to. If the page is written to,
1560 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1561 * after the page is finished with, and before put_page is called.
1563 * get_user_pages is typically used for fewer-copy IO operations, to get a
1564 * handle on the memory by some means other than accesses via the user virtual
1565 * addresses. The pages may be submitted for DMA to devices or accessed via
1566 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1567 * use the correct cache flushing APIs.
1569 * See also get_user_pages_fast, for performance critical applications.
1571 int get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1572 unsigned long start
, int nr_pages
, int write
, int force
,
1573 struct page
**pages
, struct vm_area_struct
**vmas
)
1575 int flags
= FOLL_TOUCH
;
1580 flags
|= FOLL_WRITE
;
1582 flags
|= FOLL_FORCE
;
1584 return __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
, vmas
);
1586 EXPORT_SYMBOL(get_user_pages
);
1589 * get_dump_page() - pin user page in memory while writing it to core dump
1590 * @addr: user address
1592 * Returns struct page pointer of user page pinned for dump,
1593 * to be freed afterwards by page_cache_release() or put_page().
1595 * Returns NULL on any kind of failure - a hole must then be inserted into
1596 * the corefile, to preserve alignment with its headers; and also returns
1597 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1598 * allowing a hole to be left in the corefile to save diskspace.
1600 * Called without mmap_sem, but after all other threads have been killed.
1602 #ifdef CONFIG_ELF_CORE
1603 struct page
*get_dump_page(unsigned long addr
)
1605 struct vm_area_struct
*vma
;
1608 if (__get_user_pages(current
, current
->mm
, addr
, 1,
1609 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
) < 1)
1611 flush_cache_page(vma
, addr
, page_to_pfn(page
));
1614 #endif /* CONFIG_ELF_CORE */
1616 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1619 pgd_t
* pgd
= pgd_offset(mm
, addr
);
1620 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
1622 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
1624 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1630 * This is the old fallback for page remapping.
1632 * For historical reasons, it only allows reserved pages. Only
1633 * old drivers should use this, and they needed to mark their
1634 * pages reserved for the old functions anyway.
1636 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1637 struct page
*page
, pgprot_t prot
)
1639 struct mm_struct
*mm
= vma
->vm_mm
;
1648 flush_dcache_page(page
);
1649 pte
= get_locked_pte(mm
, addr
, &ptl
);
1653 if (!pte_none(*pte
))
1656 /* Ok, finally just insert the thing.. */
1658 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
1659 page_add_file_rmap(page
);
1660 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1663 pte_unmap_unlock(pte
, ptl
);
1666 pte_unmap_unlock(pte
, ptl
);
1672 * vm_insert_page - insert single page into user vma
1673 * @vma: user vma to map to
1674 * @addr: target user address of this page
1675 * @page: source kernel page
1677 * This allows drivers to insert individual pages they've allocated
1680 * The page has to be a nice clean _individual_ kernel allocation.
1681 * If you allocate a compound page, you need to have marked it as
1682 * such (__GFP_COMP), or manually just split the page up yourself
1683 * (see split_page()).
1685 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1686 * took an arbitrary page protection parameter. This doesn't allow
1687 * that. Your vma protection will have to be set up correctly, which
1688 * means that if you want a shared writable mapping, you'd better
1689 * ask for a shared writable mapping!
1691 * The page does not need to be reserved.
1693 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1696 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1698 if (!page_count(page
))
1700 vma
->vm_flags
|= VM_INSERTPAGE
;
1701 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1703 EXPORT_SYMBOL(vm_insert_page
);
1705 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1706 unsigned long pfn
, pgprot_t prot
)
1708 struct mm_struct
*mm
= vma
->vm_mm
;
1714 pte
= get_locked_pte(mm
, addr
, &ptl
);
1718 if (!pte_none(*pte
))
1721 /* Ok, finally just insert the thing.. */
1722 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
1723 set_pte_at(mm
, addr
, pte
, entry
);
1724 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1728 pte_unmap_unlock(pte
, ptl
);
1734 * vm_insert_pfn - insert single pfn into user vma
1735 * @vma: user vma to map to
1736 * @addr: target user address of this page
1737 * @pfn: source kernel pfn
1739 * Similar to vm_inert_page, this allows drivers to insert individual pages
1740 * they've allocated into a user vma. Same comments apply.
1742 * This function should only be called from a vm_ops->fault handler, and
1743 * in that case the handler should return NULL.
1745 * vma cannot be a COW mapping.
1747 * As this is called only for pages that do not currently exist, we
1748 * do not need to flush old virtual caches or the TLB.
1750 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1754 pgprot_t pgprot
= vma
->vm_page_prot
;
1756 * Technically, architectures with pte_special can avoid all these
1757 * restrictions (same for remap_pfn_range). However we would like
1758 * consistency in testing and feature parity among all, so we should
1759 * try to keep these invariants in place for everybody.
1761 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1762 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1763 (VM_PFNMAP
|VM_MIXEDMAP
));
1764 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1765 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1767 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1769 if (track_pfn_vma_new(vma
, &pgprot
, pfn
, PAGE_SIZE
))
1772 ret
= insert_pfn(vma
, addr
, pfn
, pgprot
);
1775 untrack_pfn_vma(vma
, pfn
, PAGE_SIZE
);
1779 EXPORT_SYMBOL(vm_insert_pfn
);
1781 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1784 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
1786 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1790 * If we don't have pte special, then we have to use the pfn_valid()
1791 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1792 * refcount the page if pfn_valid is true (hence insert_page rather
1793 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1794 * without pte special, it would there be refcounted as a normal page.
1796 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
1799 page
= pfn_to_page(pfn
);
1800 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1802 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1804 EXPORT_SYMBOL(vm_insert_mixed
);
1807 * maps a range of physical memory into the requested pages. the old
1808 * mappings are removed. any references to nonexistent pages results
1809 * in null mappings (currently treated as "copy-on-access")
1811 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1812 unsigned long addr
, unsigned long end
,
1813 unsigned long pfn
, pgprot_t prot
)
1818 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1821 arch_enter_lazy_mmu_mode();
1823 BUG_ON(!pte_none(*pte
));
1824 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
1826 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1827 arch_leave_lazy_mmu_mode();
1828 pte_unmap_unlock(pte
- 1, ptl
);
1832 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1833 unsigned long addr
, unsigned long end
,
1834 unsigned long pfn
, pgprot_t prot
)
1839 pfn
-= addr
>> PAGE_SHIFT
;
1840 pmd
= pmd_alloc(mm
, pud
, addr
);
1844 next
= pmd_addr_end(addr
, end
);
1845 if (remap_pte_range(mm
, pmd
, addr
, next
,
1846 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1848 } while (pmd
++, addr
= next
, addr
!= end
);
1852 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1853 unsigned long addr
, unsigned long end
,
1854 unsigned long pfn
, pgprot_t prot
)
1859 pfn
-= addr
>> PAGE_SHIFT
;
1860 pud
= pud_alloc(mm
, pgd
, addr
);
1864 next
= pud_addr_end(addr
, end
);
1865 if (remap_pmd_range(mm
, pud
, addr
, next
,
1866 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1868 } while (pud
++, addr
= next
, addr
!= end
);
1873 * remap_pfn_range - remap kernel memory to userspace
1874 * @vma: user vma to map to
1875 * @addr: target user address to start at
1876 * @pfn: physical address of kernel memory
1877 * @size: size of map area
1878 * @prot: page protection flags for this mapping
1880 * Note: this is only safe if the mm semaphore is held when called.
1882 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1883 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1887 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1888 struct mm_struct
*mm
= vma
->vm_mm
;
1892 * Physically remapped pages are special. Tell the
1893 * rest of the world about it:
1894 * VM_IO tells people not to look at these pages
1895 * (accesses can have side effects).
1896 * VM_RESERVED is specified all over the place, because
1897 * in 2.4 it kept swapout's vma scan off this vma; but
1898 * in 2.6 the LRU scan won't even find its pages, so this
1899 * flag means no more than count its pages in reserved_vm,
1900 * and omit it from core dump, even when VM_IO turned off.
1901 * VM_PFNMAP tells the core MM that the base pages are just
1902 * raw PFN mappings, and do not have a "struct page" associated
1905 * There's a horrible special case to handle copy-on-write
1906 * behaviour that some programs depend on. We mark the "original"
1907 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1909 if (addr
== vma
->vm_start
&& end
== vma
->vm_end
) {
1910 vma
->vm_pgoff
= pfn
;
1911 vma
->vm_flags
|= VM_PFN_AT_MMAP
;
1912 } else if (is_cow_mapping(vma
->vm_flags
))
1915 vma
->vm_flags
|= VM_IO
| VM_RESERVED
| VM_PFNMAP
;
1917 err
= track_pfn_vma_new(vma
, &prot
, pfn
, PAGE_ALIGN(size
));
1920 * To indicate that track_pfn related cleanup is not
1921 * needed from higher level routine calling unmap_vmas
1923 vma
->vm_flags
&= ~(VM_IO
| VM_RESERVED
| VM_PFNMAP
);
1924 vma
->vm_flags
&= ~VM_PFN_AT_MMAP
;
1928 BUG_ON(addr
>= end
);
1929 pfn
-= addr
>> PAGE_SHIFT
;
1930 pgd
= pgd_offset(mm
, addr
);
1931 flush_cache_range(vma
, addr
, end
);
1933 next
= pgd_addr_end(addr
, end
);
1934 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1935 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1938 } while (pgd
++, addr
= next
, addr
!= end
);
1941 untrack_pfn_vma(vma
, pfn
, PAGE_ALIGN(size
));
1945 EXPORT_SYMBOL(remap_pfn_range
);
1947 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1948 unsigned long addr
, unsigned long end
,
1949 pte_fn_t fn
, void *data
)
1954 spinlock_t
*uninitialized_var(ptl
);
1956 pte
= (mm
== &init_mm
) ?
1957 pte_alloc_kernel(pmd
, addr
) :
1958 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1962 BUG_ON(pmd_huge(*pmd
));
1964 arch_enter_lazy_mmu_mode();
1966 token
= pmd_pgtable(*pmd
);
1969 err
= fn(pte
++, token
, addr
, data
);
1972 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1974 arch_leave_lazy_mmu_mode();
1977 pte_unmap_unlock(pte
-1, ptl
);
1981 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1982 unsigned long addr
, unsigned long end
,
1983 pte_fn_t fn
, void *data
)
1989 BUG_ON(pud_huge(*pud
));
1991 pmd
= pmd_alloc(mm
, pud
, addr
);
1995 next
= pmd_addr_end(addr
, end
);
1996 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
1999 } while (pmd
++, addr
= next
, addr
!= end
);
2003 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2004 unsigned long addr
, unsigned long end
,
2005 pte_fn_t fn
, void *data
)
2011 pud
= pud_alloc(mm
, pgd
, addr
);
2015 next
= pud_addr_end(addr
, end
);
2016 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
2019 } while (pud
++, addr
= next
, addr
!= end
);
2024 * Scan a region of virtual memory, filling in page tables as necessary
2025 * and calling a provided function on each leaf page table.
2027 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2028 unsigned long size
, pte_fn_t fn
, void *data
)
2032 unsigned long end
= addr
+ size
;
2035 BUG_ON(addr
>= end
);
2036 pgd
= pgd_offset(mm
, addr
);
2038 next
= pgd_addr_end(addr
, end
);
2039 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
2042 } while (pgd
++, addr
= next
, addr
!= end
);
2046 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2049 * handle_pte_fault chooses page fault handler according to an entry
2050 * which was read non-atomically. Before making any commitment, on
2051 * those architectures or configurations (e.g. i386 with PAE) which
2052 * might give a mix of unmatched parts, do_swap_page and do_file_page
2053 * must check under lock before unmapping the pte and proceeding
2054 * (but do_wp_page is only called after already making such a check;
2055 * and do_anonymous_page and do_no_page can safely check later on).
2057 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2058 pte_t
*page_table
, pte_t orig_pte
)
2061 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2062 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2063 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2065 same
= pte_same(*page_table
, orig_pte
);
2069 pte_unmap(page_table
);
2074 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
2075 * servicing faults for write access. In the normal case, do always want
2076 * pte_mkwrite. But get_user_pages can cause write faults for mappings
2077 * that do not have writing enabled, when used by access_process_vm.
2079 static inline pte_t
maybe_mkwrite(pte_t pte
, struct vm_area_struct
*vma
)
2081 if (likely(vma
->vm_flags
& VM_WRITE
))
2082 pte
= pte_mkwrite(pte
);
2086 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
2089 * If the source page was a PFN mapping, we don't have
2090 * a "struct page" for it. We do a best-effort copy by
2091 * just copying from the original user address. If that
2092 * fails, we just zero-fill it. Live with it.
2094 if (unlikely(!src
)) {
2095 void *kaddr
= kmap_atomic(dst
, KM_USER0
);
2096 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
2099 * This really shouldn't fail, because the page is there
2100 * in the page tables. But it might just be unreadable,
2101 * in which case we just give up and fill the result with
2104 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
2106 kunmap_atomic(kaddr
, KM_USER0
);
2107 flush_dcache_page(dst
);
2109 copy_user_highpage(dst
, src
, va
, vma
);
2113 * This routine handles present pages, when users try to write
2114 * to a shared page. It is done by copying the page to a new address
2115 * and decrementing the shared-page counter for the old page.
2117 * Note that this routine assumes that the protection checks have been
2118 * done by the caller (the low-level page fault routine in most cases).
2119 * Thus we can safely just mark it writable once we've done any necessary
2122 * We also mark the page dirty at this point even though the page will
2123 * change only once the write actually happens. This avoids a few races,
2124 * and potentially makes it more efficient.
2126 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2127 * but allow concurrent faults), with pte both mapped and locked.
2128 * We return with mmap_sem still held, but pte unmapped and unlocked.
2130 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2131 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2132 spinlock_t
*ptl
, pte_t orig_pte
)
2135 struct page
*old_page
, *new_page
;
2138 int page_mkwrite
= 0;
2139 struct page
*dirty_page
= NULL
;
2141 old_page
= vm_normal_page(vma
, address
, orig_pte
);
2144 * VM_MIXEDMAP !pfn_valid() case
2146 * We should not cow pages in a shared writeable mapping.
2147 * Just mark the pages writable as we can't do any dirty
2148 * accounting on raw pfn maps.
2150 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2151 (VM_WRITE
|VM_SHARED
))
2157 * Take out anonymous pages first, anonymous shared vmas are
2158 * not dirty accountable.
2160 if (PageAnon(old_page
) && !PageKsm(old_page
)) {
2161 if (!trylock_page(old_page
)) {
2162 page_cache_get(old_page
);
2163 pte_unmap_unlock(page_table
, ptl
);
2164 lock_page(old_page
);
2165 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2167 if (!pte_same(*page_table
, orig_pte
)) {
2168 unlock_page(old_page
);
2169 page_cache_release(old_page
);
2172 page_cache_release(old_page
);
2174 if (reuse_swap_page(old_page
)) {
2176 * The page is all ours. Move it to our anon_vma so
2177 * the rmap code will not search our parent or siblings.
2178 * Protected against the rmap code by the page lock.
2180 page_move_anon_rmap(old_page
, vma
, address
);
2181 unlock_page(old_page
);
2184 unlock_page(old_page
);
2185 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2186 (VM_WRITE
|VM_SHARED
))) {
2188 * Only catch write-faults on shared writable pages,
2189 * read-only shared pages can get COWed by
2190 * get_user_pages(.write=1, .force=1).
2192 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2193 struct vm_fault vmf
;
2196 vmf
.virtual_address
= (void __user
*)(address
&
2198 vmf
.pgoff
= old_page
->index
;
2199 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2200 vmf
.page
= old_page
;
2203 * Notify the address space that the page is about to
2204 * become writable so that it can prohibit this or wait
2205 * for the page to get into an appropriate state.
2207 * We do this without the lock held, so that it can
2208 * sleep if it needs to.
2210 page_cache_get(old_page
);
2211 pte_unmap_unlock(page_table
, ptl
);
2213 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
2215 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
2217 goto unwritable_page
;
2219 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
2220 lock_page(old_page
);
2221 if (!old_page
->mapping
) {
2222 ret
= 0; /* retry the fault */
2223 unlock_page(old_page
);
2224 goto unwritable_page
;
2227 VM_BUG_ON(!PageLocked(old_page
));
2230 * Since we dropped the lock we need to revalidate
2231 * the PTE as someone else may have changed it. If
2232 * they did, we just return, as we can count on the
2233 * MMU to tell us if they didn't also make it writable.
2235 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2237 if (!pte_same(*page_table
, orig_pte
)) {
2238 unlock_page(old_page
);
2239 page_cache_release(old_page
);
2245 dirty_page
= old_page
;
2246 get_page(dirty_page
);
2249 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2250 entry
= pte_mkyoung(orig_pte
);
2251 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2252 if (ptep_set_access_flags(vma
, address
, page_table
, entry
,1))
2253 update_mmu_cache(vma
, address
, page_table
);
2254 pte_unmap_unlock(page_table
, ptl
);
2255 ret
|= VM_FAULT_WRITE
;
2261 * Yes, Virginia, this is actually required to prevent a race
2262 * with clear_page_dirty_for_io() from clearing the page dirty
2263 * bit after it clear all dirty ptes, but before a racing
2264 * do_wp_page installs a dirty pte.
2266 * do_no_page is protected similarly.
2268 if (!page_mkwrite
) {
2269 wait_on_page_locked(dirty_page
);
2270 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2272 put_page(dirty_page
);
2274 struct address_space
*mapping
= dirty_page
->mapping
;
2276 set_page_dirty(dirty_page
);
2277 unlock_page(dirty_page
);
2278 page_cache_release(dirty_page
);
2281 * Some device drivers do not set page.mapping
2282 * but still dirty their pages
2284 balance_dirty_pages_ratelimited(mapping
);
2288 /* file_update_time outside page_lock */
2290 file_update_time(vma
->vm_file
);
2296 * Ok, we need to copy. Oh, well..
2298 page_cache_get(old_page
);
2300 pte_unmap_unlock(page_table
, ptl
);
2302 if (unlikely(anon_vma_prepare(vma
)))
2305 if (is_zero_pfn(pte_pfn(orig_pte
))) {
2306 new_page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2310 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2313 cow_user_page(new_page
, old_page
, address
, vma
);
2315 __SetPageUptodate(new_page
);
2318 * Don't let another task, with possibly unlocked vma,
2319 * keep the mlocked page.
2321 if ((vma
->vm_flags
& VM_LOCKED
) && old_page
) {
2322 lock_page(old_page
); /* for LRU manipulation */
2323 clear_page_mlock(old_page
);
2324 unlock_page(old_page
);
2327 if (mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
))
2331 * Re-check the pte - we dropped the lock
2333 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2334 if (likely(pte_same(*page_table
, orig_pte
))) {
2336 if (!PageAnon(old_page
)) {
2337 dec_mm_counter_fast(mm
, MM_FILEPAGES
);
2338 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2341 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2342 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2343 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2344 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2346 * Clear the pte entry and flush it first, before updating the
2347 * pte with the new entry. This will avoid a race condition
2348 * seen in the presence of one thread doing SMC and another
2351 ptep_clear_flush(vma
, address
, page_table
);
2352 page_add_new_anon_rmap(new_page
, vma
, address
);
2354 * We call the notify macro here because, when using secondary
2355 * mmu page tables (such as kvm shadow page tables), we want the
2356 * new page to be mapped directly into the secondary page table.
2358 set_pte_at_notify(mm
, address
, page_table
, entry
);
2359 update_mmu_cache(vma
, address
, page_table
);
2362 * Only after switching the pte to the new page may
2363 * we remove the mapcount here. Otherwise another
2364 * process may come and find the rmap count decremented
2365 * before the pte is switched to the new page, and
2366 * "reuse" the old page writing into it while our pte
2367 * here still points into it and can be read by other
2370 * The critical issue is to order this
2371 * page_remove_rmap with the ptp_clear_flush above.
2372 * Those stores are ordered by (if nothing else,)
2373 * the barrier present in the atomic_add_negative
2374 * in page_remove_rmap.
2376 * Then the TLB flush in ptep_clear_flush ensures that
2377 * no process can access the old page before the
2378 * decremented mapcount is visible. And the old page
2379 * cannot be reused until after the decremented
2380 * mapcount is visible. So transitively, TLBs to
2381 * old page will be flushed before it can be reused.
2383 page_remove_rmap(old_page
);
2386 /* Free the old page.. */
2387 new_page
= old_page
;
2388 ret
|= VM_FAULT_WRITE
;
2390 mem_cgroup_uncharge_page(new_page
);
2393 page_cache_release(new_page
);
2395 page_cache_release(old_page
);
2397 pte_unmap_unlock(page_table
, ptl
);
2400 page_cache_release(new_page
);
2404 unlock_page(old_page
);
2405 page_cache_release(old_page
);
2407 page_cache_release(old_page
);
2409 return VM_FAULT_OOM
;
2412 page_cache_release(old_page
);
2417 * Helper functions for unmap_mapping_range().
2419 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2421 * We have to restart searching the prio_tree whenever we drop the lock,
2422 * since the iterator is only valid while the lock is held, and anyway
2423 * a later vma might be split and reinserted earlier while lock dropped.
2425 * The list of nonlinear vmas could be handled more efficiently, using
2426 * a placeholder, but handle it in the same way until a need is shown.
2427 * It is important to search the prio_tree before nonlinear list: a vma
2428 * may become nonlinear and be shifted from prio_tree to nonlinear list
2429 * while the lock is dropped; but never shifted from list to prio_tree.
2431 * In order to make forward progress despite restarting the search,
2432 * vm_truncate_count is used to mark a vma as now dealt with, so we can
2433 * quickly skip it next time around. Since the prio_tree search only
2434 * shows us those vmas affected by unmapping the range in question, we
2435 * can't efficiently keep all vmas in step with mapping->truncate_count:
2436 * so instead reset them all whenever it wraps back to 0 (then go to 1).
2437 * mapping->truncate_count and vma->vm_truncate_count are protected by
2440 * In order to make forward progress despite repeatedly restarting some
2441 * large vma, note the restart_addr from unmap_vmas when it breaks out:
2442 * and restart from that address when we reach that vma again. It might
2443 * have been split or merged, shrunk or extended, but never shifted: so
2444 * restart_addr remains valid so long as it remains in the vma's range.
2445 * unmap_mapping_range forces truncate_count to leap over page-aligned
2446 * values so we can save vma's restart_addr in its truncate_count field.
2448 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2450 static void reset_vma_truncate_counts(struct address_space
*mapping
)
2452 struct vm_area_struct
*vma
;
2453 struct prio_tree_iter iter
;
2455 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, 0, ULONG_MAX
)
2456 vma
->vm_truncate_count
= 0;
2457 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
2458 vma
->vm_truncate_count
= 0;
2461 static int unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2462 unsigned long start_addr
, unsigned long end_addr
,
2463 struct zap_details
*details
)
2465 unsigned long restart_addr
;
2469 * files that support invalidating or truncating portions of the
2470 * file from under mmaped areas must have their ->fault function
2471 * return a locked page (and set VM_FAULT_LOCKED in the return).
2472 * This provides synchronisation against concurrent unmapping here.
2476 restart_addr
= vma
->vm_truncate_count
;
2477 if (is_restart_addr(restart_addr
) && start_addr
< restart_addr
) {
2478 start_addr
= restart_addr
;
2479 if (start_addr
>= end_addr
) {
2480 /* Top of vma has been split off since last time */
2481 vma
->vm_truncate_count
= details
->truncate_count
;
2486 restart_addr
= zap_page_range(vma
, start_addr
,
2487 end_addr
- start_addr
, details
);
2488 need_break
= need_resched() || spin_needbreak(details
->i_mmap_lock
);
2490 if (restart_addr
>= end_addr
) {
2491 /* We have now completed this vma: mark it so */
2492 vma
->vm_truncate_count
= details
->truncate_count
;
2496 /* Note restart_addr in vma's truncate_count field */
2497 vma
->vm_truncate_count
= restart_addr
;
2502 spin_unlock(details
->i_mmap_lock
);
2504 spin_lock(details
->i_mmap_lock
);
2508 static inline void unmap_mapping_range_tree(struct prio_tree_root
*root
,
2509 struct zap_details
*details
)
2511 struct vm_area_struct
*vma
;
2512 struct prio_tree_iter iter
;
2513 pgoff_t vba
, vea
, zba
, zea
;
2516 vma_prio_tree_foreach(vma
, &iter
, root
,
2517 details
->first_index
, details
->last_index
) {
2518 /* Skip quickly over those we have already dealt with */
2519 if (vma
->vm_truncate_count
== details
->truncate_count
)
2522 vba
= vma
->vm_pgoff
;
2523 vea
= vba
+ ((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) - 1;
2524 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2525 zba
= details
->first_index
;
2528 zea
= details
->last_index
;
2532 if (unmap_mapping_range_vma(vma
,
2533 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2534 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2540 static inline void unmap_mapping_range_list(struct list_head
*head
,
2541 struct zap_details
*details
)
2543 struct vm_area_struct
*vma
;
2546 * In nonlinear VMAs there is no correspondence between virtual address
2547 * offset and file offset. So we must perform an exhaustive search
2548 * across *all* the pages in each nonlinear VMA, not just the pages
2549 * whose virtual address lies outside the file truncation point.
2552 list_for_each_entry(vma
, head
, shared
.vm_set
.list
) {
2553 /* Skip quickly over those we have already dealt with */
2554 if (vma
->vm_truncate_count
== details
->truncate_count
)
2556 details
->nonlinear_vma
= vma
;
2557 if (unmap_mapping_range_vma(vma
, vma
->vm_start
,
2558 vma
->vm_end
, details
) < 0)
2564 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2565 * @mapping: the address space containing mmaps to be unmapped.
2566 * @holebegin: byte in first page to unmap, relative to the start of
2567 * the underlying file. This will be rounded down to a PAGE_SIZE
2568 * boundary. Note that this is different from truncate_pagecache(), which
2569 * must keep the partial page. In contrast, we must get rid of
2571 * @holelen: size of prospective hole in bytes. This will be rounded
2572 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2574 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2575 * but 0 when invalidating pagecache, don't throw away private data.
2577 void unmap_mapping_range(struct address_space
*mapping
,
2578 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2580 struct zap_details details
;
2581 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2582 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2584 /* Check for overflow. */
2585 if (sizeof(holelen
) > sizeof(hlen
)) {
2587 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2588 if (holeend
& ~(long long)ULONG_MAX
)
2589 hlen
= ULONG_MAX
- hba
+ 1;
2592 details
.check_mapping
= even_cows
? NULL
: mapping
;
2593 details
.nonlinear_vma
= NULL
;
2594 details
.first_index
= hba
;
2595 details
.last_index
= hba
+ hlen
- 1;
2596 if (details
.last_index
< details
.first_index
)
2597 details
.last_index
= ULONG_MAX
;
2598 details
.i_mmap_lock
= &mapping
->i_mmap_lock
;
2600 spin_lock(&mapping
->i_mmap_lock
);
2602 /* Protect against endless unmapping loops */
2603 mapping
->truncate_count
++;
2604 if (unlikely(is_restart_addr(mapping
->truncate_count
))) {
2605 if (mapping
->truncate_count
== 0)
2606 reset_vma_truncate_counts(mapping
);
2607 mapping
->truncate_count
++;
2609 details
.truncate_count
= mapping
->truncate_count
;
2611 if (unlikely(!prio_tree_empty(&mapping
->i_mmap
)))
2612 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2613 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
2614 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
2615 spin_unlock(&mapping
->i_mmap_lock
);
2617 EXPORT_SYMBOL(unmap_mapping_range
);
2619 int vmtruncate_range(struct inode
*inode
, loff_t offset
, loff_t end
)
2621 struct address_space
*mapping
= inode
->i_mapping
;
2624 * If the underlying filesystem is not going to provide
2625 * a way to truncate a range of blocks (punch a hole) -
2626 * we should return failure right now.
2628 if (!inode
->i_op
->truncate_range
)
2631 mutex_lock(&inode
->i_mutex
);
2632 down_write(&inode
->i_alloc_sem
);
2633 unmap_mapping_range(mapping
, offset
, (end
- offset
), 1);
2634 truncate_inode_pages_range(mapping
, offset
, end
);
2635 unmap_mapping_range(mapping
, offset
, (end
- offset
), 1);
2636 inode
->i_op
->truncate_range(inode
, offset
, end
);
2637 up_write(&inode
->i_alloc_sem
);
2638 mutex_unlock(&inode
->i_mutex
);
2644 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2645 * but allow concurrent faults), and pte mapped but not yet locked.
2646 * We return with mmap_sem still held, but pte unmapped and unlocked.
2648 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2649 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2650 unsigned int flags
, pte_t orig_pte
)
2653 struct page
*page
, *swapcache
= NULL
;
2657 struct mem_cgroup
*ptr
= NULL
;
2661 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2664 entry
= pte_to_swp_entry(orig_pte
);
2665 if (unlikely(non_swap_entry(entry
))) {
2666 if (is_migration_entry(entry
)) {
2667 migration_entry_wait(mm
, pmd
, address
);
2668 } else if (is_hwpoison_entry(entry
)) {
2669 ret
= VM_FAULT_HWPOISON
;
2671 print_bad_pte(vma
, address
, orig_pte
, NULL
);
2672 ret
= VM_FAULT_SIGBUS
;
2676 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2677 page
= lookup_swap_cache(entry
);
2679 grab_swap_token(mm
); /* Contend for token _before_ read-in */
2680 page
= swapin_readahead(entry
,
2681 GFP_HIGHUSER_MOVABLE
, vma
, address
);
2684 * Back out if somebody else faulted in this pte
2685 * while we released the pte lock.
2687 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2688 if (likely(pte_same(*page_table
, orig_pte
)))
2690 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2694 /* Had to read the page from swap area: Major fault */
2695 ret
= VM_FAULT_MAJOR
;
2696 count_vm_event(PGMAJFAULT
);
2697 } else if (PageHWPoison(page
)) {
2699 * hwpoisoned dirty swapcache pages are kept for killing
2700 * owner processes (which may be unknown at hwpoison time)
2702 ret
= VM_FAULT_HWPOISON
;
2703 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2707 locked
= lock_page_or_retry(page
, mm
, flags
);
2708 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2710 ret
|= VM_FAULT_RETRY
;
2715 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2716 * release the swapcache from under us. The page pin, and pte_same
2717 * test below, are not enough to exclude that. Even if it is still
2718 * swapcache, we need to check that the page's swap has not changed.
2720 if (unlikely(!PageSwapCache(page
) || page_private(page
) != entry
.val
))
2723 if (ksm_might_need_to_copy(page
, vma
, address
)) {
2725 page
= ksm_does_need_to_copy(page
, vma
, address
);
2727 if (unlikely(!page
)) {
2735 if (mem_cgroup_try_charge_swapin(mm
, page
, GFP_KERNEL
, &ptr
)) {
2741 * Back out if somebody else already faulted in this pte.
2743 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2744 if (unlikely(!pte_same(*page_table
, orig_pte
)))
2747 if (unlikely(!PageUptodate(page
))) {
2748 ret
= VM_FAULT_SIGBUS
;
2753 * The page isn't present yet, go ahead with the fault.
2755 * Be careful about the sequence of operations here.
2756 * To get its accounting right, reuse_swap_page() must be called
2757 * while the page is counted on swap but not yet in mapcount i.e.
2758 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2759 * must be called after the swap_free(), or it will never succeed.
2760 * Because delete_from_swap_page() may be called by reuse_swap_page(),
2761 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2762 * in page->private. In this case, a record in swap_cgroup is silently
2763 * discarded at swap_free().
2766 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2767 dec_mm_counter_fast(mm
, MM_SWAPENTS
);
2768 pte
= mk_pte(page
, vma
->vm_page_prot
);
2769 if ((flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
)) {
2770 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
2771 flags
&= ~FAULT_FLAG_WRITE
;
2772 ret
|= VM_FAULT_WRITE
;
2775 flush_icache_page(vma
, page
);
2776 set_pte_at(mm
, address
, page_table
, pte
);
2777 do_page_add_anon_rmap(page
, vma
, address
, exclusive
);
2778 /* It's better to call commit-charge after rmap is established */
2779 mem_cgroup_commit_charge_swapin(page
, ptr
);
2782 if (vm_swap_full() || (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
2783 try_to_free_swap(page
);
2787 * Hold the lock to avoid the swap entry to be reused
2788 * until we take the PT lock for the pte_same() check
2789 * (to avoid false positives from pte_same). For
2790 * further safety release the lock after the swap_free
2791 * so that the swap count won't change under a
2792 * parallel locked swapcache.
2794 unlock_page(swapcache
);
2795 page_cache_release(swapcache
);
2798 if (flags
& FAULT_FLAG_WRITE
) {
2799 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
2800 if (ret
& VM_FAULT_ERROR
)
2801 ret
&= VM_FAULT_ERROR
;
2805 /* No need to invalidate - it was non-present before */
2806 update_mmu_cache(vma
, address
, page_table
);
2808 pte_unmap_unlock(page_table
, ptl
);
2812 mem_cgroup_cancel_charge_swapin(ptr
);
2813 pte_unmap_unlock(page_table
, ptl
);
2817 page_cache_release(page
);
2819 unlock_page(swapcache
);
2820 page_cache_release(swapcache
);
2826 * This is like a special single-page "expand_{down|up}wards()",
2827 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2828 * doesn't hit another vma.
2830 static inline int check_stack_guard_page(struct vm_area_struct
*vma
, unsigned long address
)
2832 address
&= PAGE_MASK
;
2833 if ((vma
->vm_flags
& VM_GROWSDOWN
) && address
== vma
->vm_start
) {
2834 struct vm_area_struct
*prev
= vma
->vm_prev
;
2837 * Is there a mapping abutting this one below?
2839 * That's only ok if it's the same stack mapping
2840 * that has gotten split..
2842 if (prev
&& prev
->vm_end
== address
)
2843 return prev
->vm_flags
& VM_GROWSDOWN
? 0 : -ENOMEM
;
2845 expand_stack(vma
, address
- PAGE_SIZE
);
2847 if ((vma
->vm_flags
& VM_GROWSUP
) && address
+ PAGE_SIZE
== vma
->vm_end
) {
2848 struct vm_area_struct
*next
= vma
->vm_next
;
2850 /* As VM_GROWSDOWN but s/below/above/ */
2851 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
)
2852 return next
->vm_flags
& VM_GROWSUP
? 0 : -ENOMEM
;
2854 expand_upwards(vma
, address
+ PAGE_SIZE
);
2860 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2861 * but allow concurrent faults), and pte mapped but not yet locked.
2862 * We return with mmap_sem still held, but pte unmapped and unlocked.
2864 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2865 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2872 pte_unmap(page_table
);
2874 /* Check if we need to add a guard page to the stack */
2875 if (check_stack_guard_page(vma
, address
) < 0)
2876 return VM_FAULT_SIGBUS
;
2878 /* Use the zero-page for reads */
2879 if (!(flags
& FAULT_FLAG_WRITE
)) {
2880 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(address
),
2881 vma
->vm_page_prot
));
2882 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2883 if (!pte_none(*page_table
))
2888 /* Allocate our own private page. */
2889 if (unlikely(anon_vma_prepare(vma
)))
2891 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2894 __SetPageUptodate(page
);
2896 if (mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
))
2899 entry
= mk_pte(page
, vma
->vm_page_prot
);
2900 if (vma
->vm_flags
& VM_WRITE
)
2901 entry
= pte_mkwrite(pte_mkdirty(entry
));
2903 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2904 if (!pte_none(*page_table
))
2907 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2908 page_add_new_anon_rmap(page
, vma
, address
);
2910 set_pte_at(mm
, address
, page_table
, entry
);
2912 /* No need to invalidate - it was non-present before */
2913 update_mmu_cache(vma
, address
, page_table
);
2915 pte_unmap_unlock(page_table
, ptl
);
2918 mem_cgroup_uncharge_page(page
);
2919 page_cache_release(page
);
2922 page_cache_release(page
);
2924 return VM_FAULT_OOM
;
2928 * __do_fault() tries to create a new page mapping. It aggressively
2929 * tries to share with existing pages, but makes a separate copy if
2930 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2931 * the next page fault.
2933 * As this is called only for pages that do not currently exist, we
2934 * do not need to flush old virtual caches or the TLB.
2936 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2937 * but allow concurrent faults), and pte neither mapped nor locked.
2938 * We return with mmap_sem still held, but pte unmapped and unlocked.
2940 static int __do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2941 unsigned long address
, pmd_t
*pmd
,
2942 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
2950 struct page
*dirty_page
= NULL
;
2951 struct vm_fault vmf
;
2953 int page_mkwrite
= 0;
2955 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
2960 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
2961 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
2965 if (unlikely(PageHWPoison(vmf
.page
))) {
2966 if (ret
& VM_FAULT_LOCKED
)
2967 unlock_page(vmf
.page
);
2968 return VM_FAULT_HWPOISON
;
2972 * For consistency in subsequent calls, make the faulted page always
2975 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
2976 lock_page(vmf
.page
);
2978 VM_BUG_ON(!PageLocked(vmf
.page
));
2981 * Should we do an early C-O-W break?
2984 if (flags
& FAULT_FLAG_WRITE
) {
2985 if (!(vma
->vm_flags
& VM_SHARED
)) {
2987 if (unlikely(anon_vma_prepare(vma
))) {
2991 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
,
2997 if (mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
)) {
2999 page_cache_release(page
);
3004 * Don't let another task, with possibly unlocked vma,
3005 * keep the mlocked page.
3007 if (vma
->vm_flags
& VM_LOCKED
)
3008 clear_page_mlock(vmf
.page
);
3009 copy_user_highpage(page
, vmf
.page
, address
, vma
);
3010 __SetPageUptodate(page
);
3013 * If the page will be shareable, see if the backing
3014 * address space wants to know that the page is about
3015 * to become writable
3017 if (vma
->vm_ops
->page_mkwrite
) {
3021 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
3022 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
3024 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3026 goto unwritable_page
;
3028 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
3030 if (!page
->mapping
) {
3031 ret
= 0; /* retry the fault */
3033 goto unwritable_page
;
3036 VM_BUG_ON(!PageLocked(page
));
3043 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3046 * This silly early PAGE_DIRTY setting removes a race
3047 * due to the bad i386 page protection. But it's valid
3048 * for other architectures too.
3050 * Note that if FAULT_FLAG_WRITE is set, we either now have
3051 * an exclusive copy of the page, or this is a shared mapping,
3052 * so we can make it writable and dirty to avoid having to
3053 * handle that later.
3055 /* Only go through if we didn't race with anybody else... */
3056 if (likely(pte_same(*page_table
, orig_pte
))) {
3057 flush_icache_page(vma
, page
);
3058 entry
= mk_pte(page
, vma
->vm_page_prot
);
3059 if (flags
& FAULT_FLAG_WRITE
)
3060 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3062 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3063 page_add_new_anon_rmap(page
, vma
, address
);
3065 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
3066 page_add_file_rmap(page
);
3067 if (flags
& FAULT_FLAG_WRITE
) {
3069 get_page(dirty_page
);
3072 set_pte_at(mm
, address
, page_table
, entry
);
3074 /* no need to invalidate: a not-present page won't be cached */
3075 update_mmu_cache(vma
, address
, page_table
);
3078 mem_cgroup_uncharge_page(page
);
3080 page_cache_release(page
);
3082 anon
= 1; /* no anon but release faulted_page */
3085 pte_unmap_unlock(page_table
, ptl
);
3089 struct address_space
*mapping
= page
->mapping
;
3091 if (set_page_dirty(dirty_page
))
3093 unlock_page(dirty_page
);
3094 put_page(dirty_page
);
3095 if (page_mkwrite
&& mapping
) {
3097 * Some device drivers do not set page.mapping but still
3100 balance_dirty_pages_ratelimited(mapping
);
3103 /* file_update_time outside page_lock */
3105 file_update_time(vma
->vm_file
);
3107 unlock_page(vmf
.page
);
3109 page_cache_release(vmf
.page
);
3115 page_cache_release(page
);
3119 static int do_linear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3120 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3121 unsigned int flags
, pte_t orig_pte
)
3123 pgoff_t pgoff
= (((address
& PAGE_MASK
)
3124 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
3126 pte_unmap(page_table
);
3127 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3131 * Fault of a previously existing named mapping. Repopulate the pte
3132 * from the encoded file_pte if possible. This enables swappable
3135 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3136 * but allow concurrent faults), and pte mapped but not yet locked.
3137 * We return with mmap_sem still held, but pte unmapped and unlocked.
3139 static int do_nonlinear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3140 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3141 unsigned int flags
, pte_t orig_pte
)
3145 flags
|= FAULT_FLAG_NONLINEAR
;
3147 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
3150 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
))) {
3152 * Page table corrupted: show pte and kill process.
3154 print_bad_pte(vma
, address
, orig_pte
, NULL
);
3155 return VM_FAULT_SIGBUS
;
3158 pgoff
= pte_to_pgoff(orig_pte
);
3159 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3163 * These routines also need to handle stuff like marking pages dirty
3164 * and/or accessed for architectures that don't do it in hardware (most
3165 * RISC architectures). The early dirtying is also good on the i386.
3167 * There is also a hook called "update_mmu_cache()" that architectures
3168 * with external mmu caches can use to update those (ie the Sparc or
3169 * PowerPC hashed page tables that act as extended TLBs).
3171 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3172 * but allow concurrent faults), and pte mapped but not yet locked.
3173 * We return with mmap_sem still held, but pte unmapped and unlocked.
3175 static inline int handle_pte_fault(struct mm_struct
*mm
,
3176 struct vm_area_struct
*vma
, unsigned long address
,
3177 pte_t
*pte
, pmd_t
*pmd
, unsigned int flags
)
3183 if (!pte_present(entry
)) {
3184 if (pte_none(entry
)) {
3186 if (likely(vma
->vm_ops
->fault
))
3187 return do_linear_fault(mm
, vma
, address
,
3188 pte
, pmd
, flags
, entry
);
3190 return do_anonymous_page(mm
, vma
, address
,
3193 if (pte_file(entry
))
3194 return do_nonlinear_fault(mm
, vma
, address
,
3195 pte
, pmd
, flags
, entry
);
3196 return do_swap_page(mm
, vma
, address
,
3197 pte
, pmd
, flags
, entry
);
3200 ptl
= pte_lockptr(mm
, pmd
);
3202 if (unlikely(!pte_same(*pte
, entry
)))
3204 if (flags
& FAULT_FLAG_WRITE
) {
3205 if (!pte_write(entry
))
3206 return do_wp_page(mm
, vma
, address
,
3207 pte
, pmd
, ptl
, entry
);
3208 entry
= pte_mkdirty(entry
);
3210 entry
= pte_mkyoung(entry
);
3211 if (ptep_set_access_flags(vma
, address
, pte
, entry
, flags
& FAULT_FLAG_WRITE
)) {
3212 update_mmu_cache(vma
, address
, pte
);
3215 * This is needed only for protection faults but the arch code
3216 * is not yet telling us if this is a protection fault or not.
3217 * This still avoids useless tlb flushes for .text page faults
3220 if (flags
& FAULT_FLAG_WRITE
)
3221 flush_tlb_fix_spurious_fault(vma
, address
);
3224 pte_unmap_unlock(pte
, ptl
);
3229 * By the time we get here, we already hold the mm semaphore
3231 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3232 unsigned long address
, unsigned int flags
)
3239 __set_current_state(TASK_RUNNING
);
3241 count_vm_event(PGFAULT
);
3243 /* do counter updates before entering really critical section. */
3244 check_sync_rss_stat(current
);
3246 if (unlikely(is_vm_hugetlb_page(vma
)))
3247 return hugetlb_fault(mm
, vma
, address
, flags
);
3249 pgd
= pgd_offset(mm
, address
);
3250 pud
= pud_alloc(mm
, pgd
, address
);
3252 return VM_FAULT_OOM
;
3253 pmd
= pmd_alloc(mm
, pud
, address
);
3255 return VM_FAULT_OOM
;
3256 pte
= pte_alloc_map(mm
, pmd
, address
);
3258 return VM_FAULT_OOM
;
3260 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
3263 #ifndef __PAGETABLE_PUD_FOLDED
3265 * Allocate page upper directory.
3266 * We've already handled the fast-path in-line.
3268 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
3270 pud_t
*new = pud_alloc_one(mm
, address
);
3274 smp_wmb(); /* See comment in __pte_alloc */
3276 spin_lock(&mm
->page_table_lock
);
3277 if (pgd_present(*pgd
)) /* Another has populated it */
3280 pgd_populate(mm
, pgd
, new);
3281 spin_unlock(&mm
->page_table_lock
);
3284 #endif /* __PAGETABLE_PUD_FOLDED */
3286 #ifndef __PAGETABLE_PMD_FOLDED
3288 * Allocate page middle directory.
3289 * We've already handled the fast-path in-line.
3291 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
3293 pmd_t
*new = pmd_alloc_one(mm
, address
);
3297 smp_wmb(); /* See comment in __pte_alloc */
3299 spin_lock(&mm
->page_table_lock
);
3300 #ifndef __ARCH_HAS_4LEVEL_HACK
3301 if (pud_present(*pud
)) /* Another has populated it */
3304 pud_populate(mm
, pud
, new);
3306 if (pgd_present(*pud
)) /* Another has populated it */
3309 pgd_populate(mm
, pud
, new);
3310 #endif /* __ARCH_HAS_4LEVEL_HACK */
3311 spin_unlock(&mm
->page_table_lock
);
3314 #endif /* __PAGETABLE_PMD_FOLDED */
3316 int make_pages_present(unsigned long addr
, unsigned long end
)
3318 int ret
, len
, write
;
3319 struct vm_area_struct
* vma
;
3321 vma
= find_vma(current
->mm
, addr
);
3325 * We want to touch writable mappings with a write fault in order
3326 * to break COW, except for shared mappings because these don't COW
3327 * and we would not want to dirty them for nothing.
3329 write
= (vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
;
3330 BUG_ON(addr
>= end
);
3331 BUG_ON(end
> vma
->vm_end
);
3332 len
= DIV_ROUND_UP(end
, PAGE_SIZE
) - addr
/PAGE_SIZE
;
3333 ret
= get_user_pages(current
, current
->mm
, addr
,
3334 len
, write
, 0, NULL
, NULL
);
3337 return ret
== len
? 0 : -EFAULT
;
3340 #if !defined(__HAVE_ARCH_GATE_AREA)
3342 #if defined(AT_SYSINFO_EHDR)
3343 static struct vm_area_struct gate_vma
;
3345 static int __init
gate_vma_init(void)
3347 gate_vma
.vm_mm
= NULL
;
3348 gate_vma
.vm_start
= FIXADDR_USER_START
;
3349 gate_vma
.vm_end
= FIXADDR_USER_END
;
3350 gate_vma
.vm_flags
= VM_READ
| VM_MAYREAD
| VM_EXEC
| VM_MAYEXEC
;
3351 gate_vma
.vm_page_prot
= __P101
;
3353 * Make sure the vDSO gets into every core dump.
3354 * Dumping its contents makes post-mortem fully interpretable later
3355 * without matching up the same kernel and hardware config to see
3356 * what PC values meant.
3358 gate_vma
.vm_flags
|= VM_ALWAYSDUMP
;
3361 __initcall(gate_vma_init
);
3364 struct vm_area_struct
*get_gate_vma(struct task_struct
*tsk
)
3366 #ifdef AT_SYSINFO_EHDR
3373 int in_gate_area_no_task(unsigned long addr
)
3375 #ifdef AT_SYSINFO_EHDR
3376 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
3382 #endif /* __HAVE_ARCH_GATE_AREA */
3384 static int __follow_pte(struct mm_struct
*mm
, unsigned long address
,
3385 pte_t
**ptepp
, spinlock_t
**ptlp
)
3392 pgd
= pgd_offset(mm
, address
);
3393 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
3396 pud
= pud_offset(pgd
, address
);
3397 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
3400 pmd
= pmd_offset(pud
, address
);
3401 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
3404 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3408 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
3411 if (!pte_present(*ptep
))
3416 pte_unmap_unlock(ptep
, *ptlp
);
3421 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
3422 pte_t
**ptepp
, spinlock_t
**ptlp
)
3426 /* (void) is needed to make gcc happy */
3427 (void) __cond_lock(*ptlp
,
3428 !(res
= __follow_pte(mm
, address
, ptepp
, ptlp
)));
3433 * follow_pfn - look up PFN at a user virtual address
3434 * @vma: memory mapping
3435 * @address: user virtual address
3436 * @pfn: location to store found PFN
3438 * Only IO mappings and raw PFN mappings are allowed.
3440 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3442 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
3449 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3452 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
3455 *pfn
= pte_pfn(*ptep
);
3456 pte_unmap_unlock(ptep
, ptl
);
3459 EXPORT_SYMBOL(follow_pfn
);
3461 #ifdef CONFIG_HAVE_IOREMAP_PROT
3462 int follow_phys(struct vm_area_struct
*vma
,
3463 unsigned long address
, unsigned int flags
,
3464 unsigned long *prot
, resource_size_t
*phys
)
3470 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3473 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
3477 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
3480 *prot
= pgprot_val(pte_pgprot(pte
));
3481 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
3485 pte_unmap_unlock(ptep
, ptl
);
3490 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
3491 void *buf
, int len
, int write
)
3493 resource_size_t phys_addr
;
3494 unsigned long prot
= 0;
3495 void __iomem
*maddr
;
3496 int offset
= addr
& (PAGE_SIZE
-1);
3498 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
3501 maddr
= ioremap_prot(phys_addr
, PAGE_SIZE
, prot
);
3503 memcpy_toio(maddr
+ offset
, buf
, len
);
3505 memcpy_fromio(buf
, maddr
+ offset
, len
);
3513 * Access another process' address space.
3514 * Source/target buffer must be kernel space,
3515 * Do not walk the page table directly, use get_user_pages
3517 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
, void *buf
, int len
, int write
)
3519 struct mm_struct
*mm
;
3520 struct vm_area_struct
*vma
;
3521 void *old_buf
= buf
;
3523 mm
= get_task_mm(tsk
);
3527 down_read(&mm
->mmap_sem
);
3528 /* ignore errors, just check how much was successfully transferred */
3530 int bytes
, ret
, offset
;
3532 struct page
*page
= NULL
;
3534 ret
= get_user_pages(tsk
, mm
, addr
, 1,
3535 write
, 1, &page
, &vma
);
3538 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3539 * we can access using slightly different code.
3541 #ifdef CONFIG_HAVE_IOREMAP_PROT
3542 vma
= find_vma(mm
, addr
);
3545 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
3546 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
3554 offset
= addr
& (PAGE_SIZE
-1);
3555 if (bytes
> PAGE_SIZE
-offset
)
3556 bytes
= PAGE_SIZE
-offset
;
3560 copy_to_user_page(vma
, page
, addr
,
3561 maddr
+ offset
, buf
, bytes
);
3562 set_page_dirty_lock(page
);
3564 copy_from_user_page(vma
, page
, addr
,
3565 buf
, maddr
+ offset
, bytes
);
3568 page_cache_release(page
);
3574 up_read(&mm
->mmap_sem
);
3577 return buf
- old_buf
;
3581 * Print the name of a VMA.
3583 void print_vma_addr(char *prefix
, unsigned long ip
)
3585 struct mm_struct
*mm
= current
->mm
;
3586 struct vm_area_struct
*vma
;
3589 * Do not print if we are in atomic
3590 * contexts (in exception stacks, etc.):
3592 if (preempt_count())
3595 down_read(&mm
->mmap_sem
);
3596 vma
= find_vma(mm
, ip
);
3597 if (vma
&& vma
->vm_file
) {
3598 struct file
*f
= vma
->vm_file
;
3599 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
3603 p
= d_path(&f
->f_path
, buf
, PAGE_SIZE
);
3606 s
= strrchr(p
, '/');
3609 printk("%s%s[%lx+%lx]", prefix
, p
,
3611 vma
->vm_end
- vma
->vm_start
);
3612 free_page((unsigned long)buf
);
3615 up_read(¤t
->mm
->mmap_sem
);
3618 #ifdef CONFIG_PROVE_LOCKING
3619 void might_fault(void)
3622 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3623 * holding the mmap_sem, this is safe because kernel memory doesn't
3624 * get paged out, therefore we'll never actually fault, and the
3625 * below annotations will generate false positives.
3627 if (segment_eq(get_fs(), KERNEL_DS
))
3632 * it would be nicer only to annotate paths which are not under
3633 * pagefault_disable, however that requires a larger audit and
3634 * providing helpers like get_user_atomic.
3636 if (!in_atomic() && current
->mm
)
3637 might_lock_read(¤t
->mm
->mmap_sem
);
3639 EXPORT_SYMBOL(might_fault
);