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
)
196 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
199 * See the comment near struct mmu_table_batch.
202 static void tlb_remove_table_smp_sync(void *arg
)
204 /* Simply deliver the interrupt */
207 static void tlb_remove_table_one(void *table
)
210 * This isn't an RCU grace period and hence the page-tables cannot be
211 * assumed to be actually RCU-freed.
213 * It is however sufficient for software page-table walkers that rely on
214 * IRQ disabling. See the comment near struct mmu_table_batch.
216 smp_call_function(tlb_remove_table_smp_sync
, NULL
, 1);
217 __tlb_remove_table(table
);
220 static void tlb_remove_table_rcu(struct rcu_head
*head
)
222 struct mmu_table_batch
*batch
;
225 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
227 for (i
= 0; i
< batch
->nr
; i
++)
228 __tlb_remove_table(batch
->tables
[i
]);
230 free_page((unsigned long)batch
);
233 void tlb_table_flush(struct mmu_gather
*tlb
)
235 struct mmu_table_batch
**batch
= &tlb
->batch
;
238 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
243 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
245 struct mmu_table_batch
**batch
= &tlb
->batch
;
250 * When there's less then two users of this mm there cannot be a
251 * concurrent page-table walk.
253 if (atomic_read(&tlb
->mm
->mm_users
) < 2) {
254 __tlb_remove_table(table
);
258 if (*batch
== NULL
) {
259 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
260 if (*batch
== NULL
) {
261 tlb_remove_table_one(table
);
266 (*batch
)->tables
[(*batch
)->nr
++] = table
;
267 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
268 tlb_table_flush(tlb
);
274 * If a p?d_bad entry is found while walking page tables, report
275 * the error, before resetting entry to p?d_none. Usually (but
276 * very seldom) called out from the p?d_none_or_clear_bad macros.
279 void pgd_clear_bad(pgd_t
*pgd
)
285 void pud_clear_bad(pud_t
*pud
)
291 void pmd_clear_bad(pmd_t
*pmd
)
298 * Note: this doesn't free the actual pages themselves. That
299 * has been handled earlier when unmapping all the memory regions.
301 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
304 pgtable_t token
= pmd_pgtable(*pmd
);
306 pte_free_tlb(tlb
, token
, addr
);
310 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
311 unsigned long addr
, unsigned long end
,
312 unsigned long floor
, unsigned long ceiling
)
319 pmd
= pmd_offset(pud
, addr
);
321 next
= pmd_addr_end(addr
, end
);
322 if (pmd_none_or_clear_bad(pmd
))
324 free_pte_range(tlb
, pmd
, addr
);
325 } while (pmd
++, addr
= next
, addr
!= end
);
335 if (end
- 1 > ceiling
- 1)
338 pmd
= pmd_offset(pud
, start
);
340 pmd_free_tlb(tlb
, pmd
, start
);
343 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
344 unsigned long addr
, unsigned long end
,
345 unsigned long floor
, unsigned long ceiling
)
352 pud
= pud_offset(pgd
, addr
);
354 next
= pud_addr_end(addr
, end
);
355 if (pud_none_or_clear_bad(pud
))
357 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
358 } while (pud
++, addr
= next
, addr
!= end
);
364 ceiling
&= PGDIR_MASK
;
368 if (end
- 1 > ceiling
- 1)
371 pud
= pud_offset(pgd
, start
);
373 pud_free_tlb(tlb
, pud
, start
);
377 * This function frees user-level page tables of a process.
379 * Must be called with pagetable lock held.
381 void free_pgd_range(struct mmu_gather
*tlb
,
382 unsigned long addr
, unsigned long end
,
383 unsigned long floor
, unsigned long ceiling
)
389 * The next few lines have given us lots of grief...
391 * Why are we testing PMD* at this top level? Because often
392 * there will be no work to do at all, and we'd prefer not to
393 * go all the way down to the bottom just to discover that.
395 * Why all these "- 1"s? Because 0 represents both the bottom
396 * of the address space and the top of it (using -1 for the
397 * top wouldn't help much: the masks would do the wrong thing).
398 * The rule is that addr 0 and floor 0 refer to the bottom of
399 * the address space, but end 0 and ceiling 0 refer to the top
400 * Comparisons need to use "end - 1" and "ceiling - 1" (though
401 * that end 0 case should be mythical).
403 * Wherever addr is brought up or ceiling brought down, we must
404 * be careful to reject "the opposite 0" before it confuses the
405 * subsequent tests. But what about where end is brought down
406 * by PMD_SIZE below? no, end can't go down to 0 there.
408 * Whereas we round start (addr) and ceiling down, by different
409 * masks at different levels, in order to test whether a table
410 * now has no other vmas using it, so can be freed, we don't
411 * bother to round floor or end up - the tests don't need that.
425 if (end
- 1 > ceiling
- 1)
430 pgd
= pgd_offset(tlb
->mm
, addr
);
432 next
= pgd_addr_end(addr
, end
);
433 if (pgd_none_or_clear_bad(pgd
))
435 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
436 } while (pgd
++, addr
= next
, addr
!= end
);
439 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
440 unsigned long floor
, unsigned long ceiling
)
443 struct vm_area_struct
*next
= vma
->vm_next
;
444 unsigned long addr
= vma
->vm_start
;
447 * Hide vma from rmap and truncate_pagecache before freeing
450 unlink_anon_vmas(vma
);
451 unlink_file_vma(vma
);
453 if (is_vm_hugetlb_page(vma
)) {
454 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
455 floor
, next
? next
->vm_start
: ceiling
);
458 * Optimization: gather nearby vmas into one call down
460 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
461 && !is_vm_hugetlb_page(next
)) {
464 unlink_anon_vmas(vma
);
465 unlink_file_vma(vma
);
467 free_pgd_range(tlb
, addr
, vma
->vm_end
,
468 floor
, next
? next
->vm_start
: ceiling
);
474 int __pte_alloc(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
475 pmd_t
*pmd
, unsigned long address
)
477 pgtable_t
new = pte_alloc_one(mm
, address
);
478 int wait_split_huge_page
;
483 * Ensure all pte setup (eg. pte page lock and page clearing) are
484 * visible before the pte is made visible to other CPUs by being
485 * put into page tables.
487 * The other side of the story is the pointer chasing in the page
488 * table walking code (when walking the page table without locking;
489 * ie. most of the time). Fortunately, these data accesses consist
490 * of a chain of data-dependent loads, meaning most CPUs (alpha
491 * being the notable exception) will already guarantee loads are
492 * seen in-order. See the alpha page table accessors for the
493 * smp_read_barrier_depends() barriers in page table walking code.
495 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
497 spin_lock(&mm
->page_table_lock
);
498 wait_split_huge_page
= 0;
499 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
501 pmd_populate(mm
, pmd
, new);
503 } else if (unlikely(pmd_trans_splitting(*pmd
)))
504 wait_split_huge_page
= 1;
505 spin_unlock(&mm
->page_table_lock
);
508 if (wait_split_huge_page
)
509 wait_split_huge_page(vma
->anon_vma
, pmd
);
513 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
515 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
519 smp_wmb(); /* See comment in __pte_alloc */
521 spin_lock(&init_mm
.page_table_lock
);
522 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
523 pmd_populate_kernel(&init_mm
, pmd
, new);
526 VM_BUG_ON(pmd_trans_splitting(*pmd
));
527 spin_unlock(&init_mm
.page_table_lock
);
529 pte_free_kernel(&init_mm
, new);
533 static inline void init_rss_vec(int *rss
)
535 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
538 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
542 if (current
->mm
== mm
)
543 sync_mm_rss(current
, mm
);
544 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
546 add_mm_counter(mm
, i
, rss
[i
]);
550 * This function is called to print an error when a bad pte
551 * is found. For example, we might have a PFN-mapped pte in
552 * a region that doesn't allow it.
554 * The calling function must still handle the error.
556 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
557 pte_t pte
, struct page
*page
)
559 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
560 pud_t
*pud
= pud_offset(pgd
, addr
);
561 pmd_t
*pmd
= pmd_offset(pud
, addr
);
562 struct address_space
*mapping
;
564 static unsigned long resume
;
565 static unsigned long nr_shown
;
566 static unsigned long nr_unshown
;
569 * Allow a burst of 60 reports, then keep quiet for that minute;
570 * or allow a steady drip of one report per second.
572 if (nr_shown
== 60) {
573 if (time_before(jiffies
, resume
)) {
579 "BUG: Bad page map: %lu messages suppressed\n",
586 resume
= jiffies
+ 60 * HZ
;
588 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
589 index
= linear_page_index(vma
, addr
);
592 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
594 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
598 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
599 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
601 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
604 print_symbol(KERN_ALERT
"vma->vm_ops->fault: %s\n",
605 (unsigned long)vma
->vm_ops
->fault
);
606 if (vma
->vm_file
&& vma
->vm_file
->f_op
)
607 print_symbol(KERN_ALERT
"vma->vm_file->f_op->mmap: %s\n",
608 (unsigned long)vma
->vm_file
->f_op
->mmap
);
610 add_taint(TAINT_BAD_PAGE
);
613 static inline int is_cow_mapping(unsigned int flags
)
615 return (flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
619 static inline int is_zero_pfn(unsigned long pfn
)
621 return pfn
== zero_pfn
;
626 static inline unsigned long my_zero_pfn(unsigned long addr
)
633 * vm_normal_page -- This function gets the "struct page" associated with a pte.
635 * "Special" mappings do not wish to be associated with a "struct page" (either
636 * it doesn't exist, or it exists but they don't want to touch it). In this
637 * case, NULL is returned here. "Normal" mappings do have a struct page.
639 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
640 * pte bit, in which case this function is trivial. Secondly, an architecture
641 * may not have a spare pte bit, which requires a more complicated scheme,
644 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
645 * special mapping (even if there are underlying and valid "struct pages").
646 * COWed pages of a VM_PFNMAP are always normal.
648 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
649 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
650 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
651 * mapping will always honor the rule
653 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
655 * And for normal mappings this is false.
657 * This restricts such mappings to be a linear translation from virtual address
658 * to pfn. To get around this restriction, we allow arbitrary mappings so long
659 * as the vma is not a COW mapping; in that case, we know that all ptes are
660 * special (because none can have been COWed).
663 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
665 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
666 * page" backing, however the difference is that _all_ pages with a struct
667 * page (that is, those where pfn_valid is true) are refcounted and considered
668 * normal pages by the VM. The disadvantage is that pages are refcounted
669 * (which can be slower and simply not an option for some PFNMAP users). The
670 * advantage is that we don't have to follow the strict linearity rule of
671 * PFNMAP mappings in order to support COWable mappings.
674 #ifdef __HAVE_ARCH_PTE_SPECIAL
675 # define HAVE_PTE_SPECIAL 1
677 # define HAVE_PTE_SPECIAL 0
679 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
682 unsigned long pfn
= pte_pfn(pte
);
684 if (HAVE_PTE_SPECIAL
) {
685 if (likely(!pte_special(pte
)))
687 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
689 if (!is_zero_pfn(pfn
))
690 print_bad_pte(vma
, addr
, pte
, NULL
);
694 /* !HAVE_PTE_SPECIAL case follows: */
696 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
697 if (vma
->vm_flags
& VM_MIXEDMAP
) {
703 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
704 if (pfn
== vma
->vm_pgoff
+ off
)
706 if (!is_cow_mapping(vma
->vm_flags
))
711 if (is_zero_pfn(pfn
))
714 if (unlikely(pfn
> highest_memmap_pfn
)) {
715 print_bad_pte(vma
, addr
, pte
, NULL
);
720 * NOTE! We still have PageReserved() pages in the page tables.
721 * eg. VDSO mappings can cause them to exist.
724 return pfn_to_page(pfn
);
728 * copy one vm_area from one task to the other. Assumes the page tables
729 * already present in the new task to be cleared in the whole range
730 * covered by this vma.
733 static inline unsigned long
734 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
735 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
736 unsigned long addr
, int *rss
)
738 unsigned long vm_flags
= vma
->vm_flags
;
739 pte_t pte
= *src_pte
;
742 /* pte contains position in swap or file, so copy. */
743 if (unlikely(!pte_present(pte
))) {
744 if (!pte_file(pte
)) {
745 swp_entry_t entry
= pte_to_swp_entry(pte
);
747 if (swap_duplicate(entry
) < 0)
750 /* make sure dst_mm is on swapoff's mmlist. */
751 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
752 spin_lock(&mmlist_lock
);
753 if (list_empty(&dst_mm
->mmlist
))
754 list_add(&dst_mm
->mmlist
,
756 spin_unlock(&mmlist_lock
);
758 if (likely(!non_swap_entry(entry
)))
760 else if (is_write_migration_entry(entry
) &&
761 is_cow_mapping(vm_flags
)) {
763 * COW mappings require pages in both parent
764 * and child to be set to read.
766 make_migration_entry_read(&entry
);
767 pte
= swp_entry_to_pte(entry
);
768 set_pte_at(src_mm
, addr
, src_pte
, pte
);
775 * If it's a COW mapping, write protect it both
776 * in the parent and the child
778 if (is_cow_mapping(vm_flags
)) {
779 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
780 pte
= pte_wrprotect(pte
);
784 * If it's a shared mapping, mark it clean in
787 if (vm_flags
& VM_SHARED
)
788 pte
= pte_mkclean(pte
);
789 pte
= pte_mkold(pte
);
791 page
= vm_normal_page(vma
, addr
, pte
);
802 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
806 int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
807 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
808 unsigned long addr
, unsigned long end
)
810 pte_t
*orig_src_pte
, *orig_dst_pte
;
811 pte_t
*src_pte
, *dst_pte
;
812 spinlock_t
*src_ptl
, *dst_ptl
;
814 int rss
[NR_MM_COUNTERS
];
815 swp_entry_t entry
= (swp_entry_t
){0};
820 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
823 src_pte
= pte_offset_map(src_pmd
, addr
);
824 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
825 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
826 orig_src_pte
= src_pte
;
827 orig_dst_pte
= dst_pte
;
828 arch_enter_lazy_mmu_mode();
832 * We are holding two locks at this point - either of them
833 * could generate latencies in another task on another CPU.
835 if (progress
>= 32) {
837 if (need_resched() ||
838 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
841 if (pte_none(*src_pte
)) {
845 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
850 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
852 arch_leave_lazy_mmu_mode();
853 spin_unlock(src_ptl
);
854 pte_unmap(orig_src_pte
);
855 add_mm_rss_vec(dst_mm
, rss
);
856 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
860 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
869 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
870 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
871 unsigned long addr
, unsigned long end
)
873 pmd_t
*src_pmd
, *dst_pmd
;
876 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
879 src_pmd
= pmd_offset(src_pud
, addr
);
881 next
= pmd_addr_end(addr
, end
);
882 if (pmd_trans_huge(*src_pmd
)) {
884 VM_BUG_ON(next
-addr
!= HPAGE_PMD_SIZE
);
885 err
= copy_huge_pmd(dst_mm
, src_mm
,
886 dst_pmd
, src_pmd
, addr
, vma
);
893 if (pmd_none_or_clear_bad(src_pmd
))
895 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
898 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
902 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
903 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
904 unsigned long addr
, unsigned long end
)
906 pud_t
*src_pud
, *dst_pud
;
909 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
912 src_pud
= pud_offset(src_pgd
, addr
);
914 next
= pud_addr_end(addr
, end
);
915 if (pud_none_or_clear_bad(src_pud
))
917 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
920 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
924 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
925 struct vm_area_struct
*vma
)
927 pgd_t
*src_pgd
, *dst_pgd
;
929 unsigned long addr
= vma
->vm_start
;
930 unsigned long end
= vma
->vm_end
;
934 * Don't copy ptes where a page fault will fill them correctly.
935 * Fork becomes much lighter when there are big shared or private
936 * readonly mappings. The tradeoff is that copy_page_range is more
937 * efficient than faulting.
939 if (!(vma
->vm_flags
& (VM_HUGETLB
|VM_NONLINEAR
|VM_PFNMAP
|VM_INSERTPAGE
))) {
944 if (is_vm_hugetlb_page(vma
))
945 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
947 if (unlikely(is_pfn_mapping(vma
))) {
949 * We do not free on error cases below as remove_vma
950 * gets called on error from higher level routine
952 ret
= track_pfn_vma_copy(vma
);
958 * We need to invalidate the secondary MMU mappings only when
959 * there could be a permission downgrade on the ptes of the
960 * parent mm. And a permission downgrade will only happen if
961 * is_cow_mapping() returns true.
963 if (is_cow_mapping(vma
->vm_flags
))
964 mmu_notifier_invalidate_range_start(src_mm
, addr
, end
);
967 dst_pgd
= pgd_offset(dst_mm
, addr
);
968 src_pgd
= pgd_offset(src_mm
, addr
);
970 next
= pgd_addr_end(addr
, end
);
971 if (pgd_none_or_clear_bad(src_pgd
))
973 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
978 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
980 if (is_cow_mapping(vma
->vm_flags
))
981 mmu_notifier_invalidate_range_end(src_mm
,
986 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
987 struct vm_area_struct
*vma
, pmd_t
*pmd
,
988 unsigned long addr
, unsigned long end
,
989 struct zap_details
*details
)
991 struct mm_struct
*mm
= tlb
->mm
;
993 int rss
[NR_MM_COUNTERS
];
999 pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1000 arch_enter_lazy_mmu_mode();
1003 if (pte_none(ptent
)) {
1007 if (pte_present(ptent
)) {
1010 page
= vm_normal_page(vma
, addr
, ptent
);
1011 if (unlikely(details
) && page
) {
1013 * unmap_shared_mapping_pages() wants to
1014 * invalidate cache without truncating:
1015 * unmap shared but keep private pages.
1017 if (details
->check_mapping
&&
1018 details
->check_mapping
!= page
->mapping
)
1021 * Each page->index must be checked when
1022 * invalidating or truncating nonlinear.
1024 if (details
->nonlinear_vma
&&
1025 (page
->index
< details
->first_index
||
1026 page
->index
> details
->last_index
))
1029 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1031 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1032 if (unlikely(!page
))
1034 if (unlikely(details
) && details
->nonlinear_vma
1035 && linear_page_index(details
->nonlinear_vma
,
1036 addr
) != page
->index
)
1037 set_pte_at(mm
, addr
, pte
,
1038 pgoff_to_pte(page
->index
));
1040 rss
[MM_ANONPAGES
]--;
1042 if (pte_dirty(ptent
))
1043 set_page_dirty(page
);
1044 if (pte_young(ptent
) &&
1045 likely(!VM_SequentialReadHint(vma
)))
1046 mark_page_accessed(page
);
1047 rss
[MM_FILEPAGES
]--;
1049 page_remove_rmap(page
);
1050 if (unlikely(page_mapcount(page
) < 0))
1051 print_bad_pte(vma
, addr
, ptent
, page
);
1052 force_flush
= !__tlb_remove_page(tlb
, page
);
1058 * If details->check_mapping, we leave swap entries;
1059 * if details->nonlinear_vma, we leave file entries.
1061 if (unlikely(details
))
1063 if (pte_file(ptent
)) {
1064 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
)))
1065 print_bad_pte(vma
, addr
, ptent
, NULL
);
1067 swp_entry_t entry
= pte_to_swp_entry(ptent
);
1069 if (!non_swap_entry(entry
))
1071 if (unlikely(!free_swap_and_cache(entry
)))
1072 print_bad_pte(vma
, addr
, ptent
, NULL
);
1074 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1075 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1077 add_mm_rss_vec(mm
, rss
);
1078 arch_leave_lazy_mmu_mode();
1079 pte_unmap_unlock(pte
- 1, ptl
);
1082 * mmu_gather ran out of room to batch pages, we break out of
1083 * the PTE lock to avoid doing the potential expensive TLB invalidate
1084 * and page-free while holding it.
1096 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1097 struct vm_area_struct
*vma
, pud_t
*pud
,
1098 unsigned long addr
, unsigned long end
,
1099 struct zap_details
*details
)
1104 pmd
= pmd_offset(pud
, addr
);
1106 next
= pmd_addr_end(addr
, end
);
1107 if (pmd_trans_huge(*pmd
)) {
1108 if (next
-addr
!= HPAGE_PMD_SIZE
) {
1109 VM_BUG_ON(!rwsem_is_locked(&tlb
->mm
->mmap_sem
));
1110 split_huge_page_pmd(vma
->vm_mm
, pmd
);
1111 } else if (zap_huge_pmd(tlb
, vma
, pmd
))
1115 if (pmd_none_or_clear_bad(pmd
))
1117 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1119 } while (pmd
++, addr
= next
, addr
!= end
);
1124 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1125 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1126 unsigned long addr
, unsigned long end
,
1127 struct zap_details
*details
)
1132 pud
= pud_offset(pgd
, addr
);
1134 next
= pud_addr_end(addr
, end
);
1135 if (pud_none_or_clear_bad(pud
))
1137 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1138 } while (pud
++, addr
= next
, addr
!= end
);
1143 static unsigned long unmap_page_range(struct mmu_gather
*tlb
,
1144 struct vm_area_struct
*vma
,
1145 unsigned long addr
, unsigned long end
,
1146 struct zap_details
*details
)
1151 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
1154 BUG_ON(addr
>= end
);
1155 mem_cgroup_uncharge_start();
1156 tlb_start_vma(tlb
, vma
);
1157 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1159 next
= pgd_addr_end(addr
, end
);
1160 if (pgd_none_or_clear_bad(pgd
))
1162 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
1163 } while (pgd
++, addr
= next
, addr
!= end
);
1164 tlb_end_vma(tlb
, vma
);
1165 mem_cgroup_uncharge_end();
1170 #ifdef CONFIG_PREEMPT
1171 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
1173 /* No preempt: go for improved straight-line efficiency */
1174 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
1178 * unmap_vmas - unmap a range of memory covered by a list of vma's
1179 * @tlbp: address of the caller's struct mmu_gather
1180 * @vma: the starting vma
1181 * @start_addr: virtual address at which to start unmapping
1182 * @end_addr: virtual address at which to end unmapping
1183 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1184 * @details: details of nonlinear truncation or shared cache invalidation
1186 * Returns the end address of the unmapping (restart addr if interrupted).
1188 * Unmap all pages in the vma list.
1190 * We aim to not hold locks for too long (for scheduling latency reasons).
1191 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
1192 * return the ending mmu_gather to the caller.
1194 * Only addresses between `start' and `end' will be unmapped.
1196 * The VMA list must be sorted in ascending virtual address order.
1198 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1199 * range after unmap_vmas() returns. So the only responsibility here is to
1200 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1201 * drops the lock and schedules.
1203 unsigned long unmap_vmas(struct mmu_gather
*tlb
,
1204 struct vm_area_struct
*vma
, unsigned long start_addr
,
1205 unsigned long end_addr
, unsigned long *nr_accounted
,
1206 struct zap_details
*details
)
1208 unsigned long start
= start_addr
;
1209 struct mm_struct
*mm
= vma
->vm_mm
;
1211 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1212 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
) {
1215 start
= max(vma
->vm_start
, start_addr
);
1216 if (start
>= vma
->vm_end
)
1218 end
= min(vma
->vm_end
, end_addr
);
1219 if (end
<= vma
->vm_start
)
1222 if (vma
->vm_flags
& VM_ACCOUNT
)
1223 *nr_accounted
+= (end
- start
) >> PAGE_SHIFT
;
1225 if (unlikely(is_pfn_mapping(vma
)))
1226 untrack_pfn_vma(vma
, 0, 0);
1228 while (start
!= end
) {
1229 if (unlikely(is_vm_hugetlb_page(vma
))) {
1231 * It is undesirable to test vma->vm_file as it
1232 * should be non-null for valid hugetlb area.
1233 * However, vm_file will be NULL in the error
1234 * cleanup path of do_mmap_pgoff. When
1235 * hugetlbfs ->mmap method fails,
1236 * do_mmap_pgoff() nullifies vma->vm_file
1237 * before calling this function to clean up.
1238 * Since no pte has actually been setup, it is
1239 * safe to do nothing in this case.
1242 unmap_hugepage_range(vma
, start
, end
, NULL
);
1246 start
= unmap_page_range(tlb
, vma
, start
, end
, details
);
1250 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1251 return start
; /* which is now the end (or restart) address */
1255 * zap_page_range - remove user pages in a given range
1256 * @vma: vm_area_struct holding the applicable pages
1257 * @address: starting address of pages to zap
1258 * @size: number of bytes to zap
1259 * @details: details of nonlinear truncation or shared cache invalidation
1261 unsigned long zap_page_range(struct vm_area_struct
*vma
, unsigned long address
,
1262 unsigned long size
, struct zap_details
*details
)
1264 struct mm_struct
*mm
= vma
->vm_mm
;
1265 struct mmu_gather tlb
;
1266 unsigned long end
= address
+ size
;
1267 unsigned long nr_accounted
= 0;
1270 tlb_gather_mmu(&tlb
, mm
, 0);
1271 update_hiwater_rss(mm
);
1272 end
= unmap_vmas(&tlb
, vma
, address
, end
, &nr_accounted
, details
);
1273 tlb_finish_mmu(&tlb
, address
, end
);
1278 * zap_vma_ptes - remove ptes mapping the vma
1279 * @vma: vm_area_struct holding ptes to be zapped
1280 * @address: starting address of pages to zap
1281 * @size: number of bytes to zap
1283 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1285 * The entire address range must be fully contained within the vma.
1287 * Returns 0 if successful.
1289 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1292 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1293 !(vma
->vm_flags
& VM_PFNMAP
))
1295 zap_page_range(vma
, address
, size
, NULL
);
1298 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1301 * follow_page - look up a page descriptor from a user-virtual address
1302 * @vma: vm_area_struct mapping @address
1303 * @address: virtual address to look up
1304 * @flags: flags modifying lookup behaviour
1306 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1308 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1309 * an error pointer if there is a mapping to something not represented
1310 * by a page descriptor (see also vm_normal_page()).
1312 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
1321 struct mm_struct
*mm
= vma
->vm_mm
;
1323 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
1324 if (!IS_ERR(page
)) {
1325 BUG_ON(flags
& FOLL_GET
);
1330 pgd
= pgd_offset(mm
, address
);
1331 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
1334 pud
= pud_offset(pgd
, address
);
1337 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
1338 BUG_ON(flags
& FOLL_GET
);
1339 page
= follow_huge_pud(mm
, address
, pud
, flags
& FOLL_WRITE
);
1342 if (unlikely(pud_bad(*pud
)))
1345 pmd
= pmd_offset(pud
, address
);
1348 if (pmd_huge(*pmd
) && vma
->vm_flags
& VM_HUGETLB
) {
1349 BUG_ON(flags
& FOLL_GET
);
1350 page
= follow_huge_pmd(mm
, address
, pmd
, flags
& FOLL_WRITE
);
1353 if (pmd_trans_huge(*pmd
)) {
1354 if (flags
& FOLL_SPLIT
) {
1355 split_huge_page_pmd(mm
, pmd
);
1356 goto split_fallthrough
;
1358 spin_lock(&mm
->page_table_lock
);
1359 if (likely(pmd_trans_huge(*pmd
))) {
1360 if (unlikely(pmd_trans_splitting(*pmd
))) {
1361 spin_unlock(&mm
->page_table_lock
);
1362 wait_split_huge_page(vma
->anon_vma
, pmd
);
1364 page
= follow_trans_huge_pmd(mm
, address
,
1366 spin_unlock(&mm
->page_table_lock
);
1370 spin_unlock(&mm
->page_table_lock
);
1374 if (unlikely(pmd_bad(*pmd
)))
1377 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1380 if (!pte_present(pte
))
1382 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
1385 page
= vm_normal_page(vma
, address
, pte
);
1386 if (unlikely(!page
)) {
1387 if ((flags
& FOLL_DUMP
) ||
1388 !is_zero_pfn(pte_pfn(pte
)))
1390 page
= pte_page(pte
);
1393 if (flags
& FOLL_GET
)
1395 if (flags
& FOLL_TOUCH
) {
1396 if ((flags
& FOLL_WRITE
) &&
1397 !pte_dirty(pte
) && !PageDirty(page
))
1398 set_page_dirty(page
);
1400 * pte_mkyoung() would be more correct here, but atomic care
1401 * is needed to avoid losing the dirty bit: it is easier to use
1402 * mark_page_accessed().
1404 mark_page_accessed(page
);
1406 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1408 * The preliminary mapping check is mainly to avoid the
1409 * pointless overhead of lock_page on the ZERO_PAGE
1410 * which might bounce very badly if there is contention.
1412 * If the page is already locked, we don't need to
1413 * handle it now - vmscan will handle it later if and
1414 * when it attempts to reclaim the page.
1416 if (page
->mapping
&& trylock_page(page
)) {
1417 lru_add_drain(); /* push cached pages to LRU */
1419 * Because we lock page here and migration is
1420 * blocked by the pte's page reference, we need
1421 * only check for file-cache page truncation.
1424 mlock_vma_page(page
);
1429 pte_unmap_unlock(ptep
, ptl
);
1434 pte_unmap_unlock(ptep
, ptl
);
1435 return ERR_PTR(-EFAULT
);
1438 pte_unmap_unlock(ptep
, ptl
);
1444 * When core dumping an enormous anonymous area that nobody
1445 * has touched so far, we don't want to allocate unnecessary pages or
1446 * page tables. Return error instead of NULL to skip handle_mm_fault,
1447 * then get_dump_page() will return NULL to leave a hole in the dump.
1448 * But we can only make this optimization where a hole would surely
1449 * be zero-filled if handle_mm_fault() actually did handle it.
1451 if ((flags
& FOLL_DUMP
) &&
1452 (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
1453 return ERR_PTR(-EFAULT
);
1457 static inline int stack_guard_page(struct vm_area_struct
*vma
, unsigned long addr
)
1459 return stack_guard_page_start(vma
, addr
) ||
1460 stack_guard_page_end(vma
, addr
+PAGE_SIZE
);
1464 * __get_user_pages() - pin user pages in memory
1465 * @tsk: task_struct of target task
1466 * @mm: mm_struct of target mm
1467 * @start: starting user address
1468 * @nr_pages: number of pages from start to pin
1469 * @gup_flags: flags modifying pin behaviour
1470 * @pages: array that receives pointers to the pages pinned.
1471 * Should be at least nr_pages long. Or NULL, if caller
1472 * only intends to ensure the pages are faulted in.
1473 * @vmas: array of pointers to vmas corresponding to each page.
1474 * Or NULL if the caller does not require them.
1475 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1477 * Returns number of pages pinned. This may be fewer than the number
1478 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1479 * were pinned, returns -errno. Each page returned must be released
1480 * with a put_page() call when it is finished with. vmas will only
1481 * remain valid while mmap_sem is held.
1483 * Must be called with mmap_sem held for read or write.
1485 * __get_user_pages walks a process's page tables and takes a reference to
1486 * each struct page that each user address corresponds to at a given
1487 * instant. That is, it takes the page that would be accessed if a user
1488 * thread accesses the given user virtual address at that instant.
1490 * This does not guarantee that the page exists in the user mappings when
1491 * __get_user_pages returns, and there may even be a completely different
1492 * page there in some cases (eg. if mmapped pagecache has been invalidated
1493 * and subsequently re faulted). However it does guarantee that the page
1494 * won't be freed completely. And mostly callers simply care that the page
1495 * contains data that was valid *at some point in time*. Typically, an IO
1496 * or similar operation cannot guarantee anything stronger anyway because
1497 * locks can't be held over the syscall boundary.
1499 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1500 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1501 * appropriate) must be called after the page is finished with, and
1502 * before put_page is called.
1504 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1505 * or mmap_sem contention, and if waiting is needed to pin all pages,
1506 * *@nonblocking will be set to 0.
1508 * In most cases, get_user_pages or get_user_pages_fast should be used
1509 * instead of __get_user_pages. __get_user_pages should be used only if
1510 * you need some special @gup_flags.
1512 int __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1513 unsigned long start
, int nr_pages
, unsigned int gup_flags
,
1514 struct page
**pages
, struct vm_area_struct
**vmas
,
1518 unsigned long vm_flags
;
1523 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
1526 * Require read or write permissions.
1527 * If FOLL_FORCE is set, we only require the "MAY" flags.
1529 vm_flags
= (gup_flags
& FOLL_WRITE
) ?
1530 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1531 vm_flags
&= (gup_flags
& FOLL_FORCE
) ?
1532 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1536 struct vm_area_struct
*vma
;
1538 vma
= find_extend_vma(mm
, start
);
1539 if (!vma
&& in_gate_area(mm
, start
)) {
1540 unsigned long pg
= start
& PAGE_MASK
;
1546 /* user gate pages are read-only */
1547 if (gup_flags
& FOLL_WRITE
)
1548 return i
? : -EFAULT
;
1550 pgd
= pgd_offset_k(pg
);
1552 pgd
= pgd_offset_gate(mm
, pg
);
1553 BUG_ON(pgd_none(*pgd
));
1554 pud
= pud_offset(pgd
, pg
);
1555 BUG_ON(pud_none(*pud
));
1556 pmd
= pmd_offset(pud
, pg
);
1558 return i
? : -EFAULT
;
1559 VM_BUG_ON(pmd_trans_huge(*pmd
));
1560 pte
= pte_offset_map(pmd
, pg
);
1561 if (pte_none(*pte
)) {
1563 return i
? : -EFAULT
;
1565 vma
= get_gate_vma(mm
);
1569 page
= vm_normal_page(vma
, start
, *pte
);
1571 if (!(gup_flags
& FOLL_DUMP
) &&
1572 is_zero_pfn(pte_pfn(*pte
)))
1573 page
= pte_page(*pte
);
1576 return i
? : -EFAULT
;
1587 (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1588 !(vm_flags
& vma
->vm_flags
))
1589 return i
? : -EFAULT
;
1591 if (is_vm_hugetlb_page(vma
)) {
1592 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1593 &start
, &nr_pages
, i
, gup_flags
);
1599 unsigned int foll_flags
= gup_flags
;
1602 * If we have a pending SIGKILL, don't keep faulting
1603 * pages and potentially allocating memory.
1605 if (unlikely(fatal_signal_pending(current
)))
1606 return i
? i
: -ERESTARTSYS
;
1609 while (!(page
= follow_page(vma
, start
, foll_flags
))) {
1611 unsigned int fault_flags
= 0;
1613 /* For mlock, just skip the stack guard page. */
1614 if (foll_flags
& FOLL_MLOCK
) {
1615 if (stack_guard_page(vma
, start
))
1618 if (foll_flags
& FOLL_WRITE
)
1619 fault_flags
|= FAULT_FLAG_WRITE
;
1621 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
1622 if (foll_flags
& FOLL_NOWAIT
)
1623 fault_flags
|= (FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
);
1625 ret
= handle_mm_fault(mm
, vma
, start
,
1628 if (ret
& VM_FAULT_ERROR
) {
1629 if (ret
& VM_FAULT_OOM
)
1630 return i
? i
: -ENOMEM
;
1631 if (ret
& (VM_FAULT_HWPOISON
|
1632 VM_FAULT_HWPOISON_LARGE
)) {
1635 else if (gup_flags
& FOLL_HWPOISON
)
1640 if (ret
& VM_FAULT_SIGBUS
)
1641 return i
? i
: -EFAULT
;
1646 if (ret
& VM_FAULT_MAJOR
)
1652 if (ret
& VM_FAULT_RETRY
) {
1659 * The VM_FAULT_WRITE bit tells us that
1660 * do_wp_page has broken COW when necessary,
1661 * even if maybe_mkwrite decided not to set
1662 * pte_write. We can thus safely do subsequent
1663 * page lookups as if they were reads. But only
1664 * do so when looping for pte_write is futile:
1665 * in some cases userspace may also be wanting
1666 * to write to the gotten user page, which a
1667 * read fault here might prevent (a readonly
1668 * page might get reCOWed by userspace write).
1670 if ((ret
& VM_FAULT_WRITE
) &&
1671 !(vma
->vm_flags
& VM_WRITE
))
1672 foll_flags
&= ~FOLL_WRITE
;
1677 return i
? i
: PTR_ERR(page
);
1681 flush_anon_page(vma
, page
, start
);
1682 flush_dcache_page(page
);
1690 } while (nr_pages
&& start
< vma
->vm_end
);
1694 EXPORT_SYMBOL(__get_user_pages
);
1697 * get_user_pages() - pin user pages in memory
1698 * @tsk: the task_struct to use for page fault accounting, or
1699 * NULL if faults are not to be recorded.
1700 * @mm: mm_struct of target mm
1701 * @start: starting user address
1702 * @nr_pages: number of pages from start to pin
1703 * @write: whether pages will be written to by the caller
1704 * @force: whether to force write access even if user mapping is
1705 * readonly. This will result in the page being COWed even
1706 * in MAP_SHARED mappings. You do not want this.
1707 * @pages: array that receives pointers to the pages pinned.
1708 * Should be at least nr_pages long. Or NULL, if caller
1709 * only intends to ensure the pages are faulted in.
1710 * @vmas: array of pointers to vmas corresponding to each page.
1711 * Or NULL if the caller does not require them.
1713 * Returns number of pages pinned. This may be fewer than the number
1714 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1715 * were pinned, returns -errno. Each page returned must be released
1716 * with a put_page() call when it is finished with. vmas will only
1717 * remain valid while mmap_sem is held.
1719 * Must be called with mmap_sem held for read or write.
1721 * get_user_pages walks a process's page tables and takes a reference to
1722 * each struct page that each user address corresponds to at a given
1723 * instant. That is, it takes the page that would be accessed if a user
1724 * thread accesses the given user virtual address at that instant.
1726 * This does not guarantee that the page exists in the user mappings when
1727 * get_user_pages returns, and there may even be a completely different
1728 * page there in some cases (eg. if mmapped pagecache has been invalidated
1729 * and subsequently re faulted). However it does guarantee that the page
1730 * won't be freed completely. And mostly callers simply care that the page
1731 * contains data that was valid *at some point in time*. Typically, an IO
1732 * or similar operation cannot guarantee anything stronger anyway because
1733 * locks can't be held over the syscall boundary.
1735 * If write=0, the page must not be written to. If the page is written to,
1736 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1737 * after the page is finished with, and before put_page is called.
1739 * get_user_pages is typically used for fewer-copy IO operations, to get a
1740 * handle on the memory by some means other than accesses via the user virtual
1741 * addresses. The pages may be submitted for DMA to devices or accessed via
1742 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1743 * use the correct cache flushing APIs.
1745 * See also get_user_pages_fast, for performance critical applications.
1747 int get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1748 unsigned long start
, int nr_pages
, int write
, int force
,
1749 struct page
**pages
, struct vm_area_struct
**vmas
)
1751 int flags
= FOLL_TOUCH
;
1756 flags
|= FOLL_WRITE
;
1758 flags
|= FOLL_FORCE
;
1760 return __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
, vmas
,
1763 EXPORT_SYMBOL(get_user_pages
);
1766 * get_dump_page() - pin user page in memory while writing it to core dump
1767 * @addr: user address
1769 * Returns struct page pointer of user page pinned for dump,
1770 * to be freed afterwards by page_cache_release() or put_page().
1772 * Returns NULL on any kind of failure - a hole must then be inserted into
1773 * the corefile, to preserve alignment with its headers; and also returns
1774 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1775 * allowing a hole to be left in the corefile to save diskspace.
1777 * Called without mmap_sem, but after all other threads have been killed.
1779 #ifdef CONFIG_ELF_CORE
1780 struct page
*get_dump_page(unsigned long addr
)
1782 struct vm_area_struct
*vma
;
1785 if (__get_user_pages(current
, current
->mm
, addr
, 1,
1786 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
1789 flush_cache_page(vma
, addr
, page_to_pfn(page
));
1792 #endif /* CONFIG_ELF_CORE */
1794 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1797 pgd_t
* pgd
= pgd_offset(mm
, addr
);
1798 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
1800 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
1802 VM_BUG_ON(pmd_trans_huge(*pmd
));
1803 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1810 * This is the old fallback for page remapping.
1812 * For historical reasons, it only allows reserved pages. Only
1813 * old drivers should use this, and they needed to mark their
1814 * pages reserved for the old functions anyway.
1816 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1817 struct page
*page
, pgprot_t prot
)
1819 struct mm_struct
*mm
= vma
->vm_mm
;
1828 flush_dcache_page(page
);
1829 pte
= get_locked_pte(mm
, addr
, &ptl
);
1833 if (!pte_none(*pte
))
1836 /* Ok, finally just insert the thing.. */
1838 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
1839 page_add_file_rmap(page
);
1840 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1843 pte_unmap_unlock(pte
, ptl
);
1846 pte_unmap_unlock(pte
, ptl
);
1852 * vm_insert_page - insert single page into user vma
1853 * @vma: user vma to map to
1854 * @addr: target user address of this page
1855 * @page: source kernel page
1857 * This allows drivers to insert individual pages they've allocated
1860 * The page has to be a nice clean _individual_ kernel allocation.
1861 * If you allocate a compound page, you need to have marked it as
1862 * such (__GFP_COMP), or manually just split the page up yourself
1863 * (see split_page()).
1865 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1866 * took an arbitrary page protection parameter. This doesn't allow
1867 * that. Your vma protection will have to be set up correctly, which
1868 * means that if you want a shared writable mapping, you'd better
1869 * ask for a shared writable mapping!
1871 * The page does not need to be reserved.
1873 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1876 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1878 if (!page_count(page
))
1880 vma
->vm_flags
|= VM_INSERTPAGE
;
1881 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1883 EXPORT_SYMBOL(vm_insert_page
);
1885 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1886 unsigned long pfn
, pgprot_t prot
)
1888 struct mm_struct
*mm
= vma
->vm_mm
;
1894 pte
= get_locked_pte(mm
, addr
, &ptl
);
1898 if (!pte_none(*pte
))
1901 /* Ok, finally just insert the thing.. */
1902 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
1903 set_pte_at(mm
, addr
, pte
, entry
);
1904 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1908 pte_unmap_unlock(pte
, ptl
);
1914 * vm_insert_pfn - insert single pfn into user vma
1915 * @vma: user vma to map to
1916 * @addr: target user address of this page
1917 * @pfn: source kernel pfn
1919 * Similar to vm_inert_page, this allows drivers to insert individual pages
1920 * they've allocated into a user vma. Same comments apply.
1922 * This function should only be called from a vm_ops->fault handler, and
1923 * in that case the handler should return NULL.
1925 * vma cannot be a COW mapping.
1927 * As this is called only for pages that do not currently exist, we
1928 * do not need to flush old virtual caches or the TLB.
1930 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1934 pgprot_t pgprot
= vma
->vm_page_prot
;
1936 * Technically, architectures with pte_special can avoid all these
1937 * restrictions (same for remap_pfn_range). However we would like
1938 * consistency in testing and feature parity among all, so we should
1939 * try to keep these invariants in place for everybody.
1941 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1942 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1943 (VM_PFNMAP
|VM_MIXEDMAP
));
1944 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1945 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1947 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1949 if (track_pfn_vma_new(vma
, &pgprot
, pfn
, PAGE_SIZE
))
1952 ret
= insert_pfn(vma
, addr
, pfn
, pgprot
);
1955 untrack_pfn_vma(vma
, pfn
, PAGE_SIZE
);
1959 EXPORT_SYMBOL(vm_insert_pfn
);
1961 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1964 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
1966 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1970 * If we don't have pte special, then we have to use the pfn_valid()
1971 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1972 * refcount the page if pfn_valid is true (hence insert_page rather
1973 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1974 * without pte special, it would there be refcounted as a normal page.
1976 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
1979 page
= pfn_to_page(pfn
);
1980 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1982 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1984 EXPORT_SYMBOL(vm_insert_mixed
);
1987 * maps a range of physical memory into the requested pages. the old
1988 * mappings are removed. any references to nonexistent pages results
1989 * in null mappings (currently treated as "copy-on-access")
1991 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1992 unsigned long addr
, unsigned long end
,
1993 unsigned long pfn
, pgprot_t prot
)
1998 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2001 arch_enter_lazy_mmu_mode();
2003 BUG_ON(!pte_none(*pte
));
2004 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2006 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2007 arch_leave_lazy_mmu_mode();
2008 pte_unmap_unlock(pte
- 1, ptl
);
2012 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2013 unsigned long addr
, unsigned long end
,
2014 unsigned long pfn
, pgprot_t prot
)
2019 pfn
-= addr
>> PAGE_SHIFT
;
2020 pmd
= pmd_alloc(mm
, pud
, addr
);
2023 VM_BUG_ON(pmd_trans_huge(*pmd
));
2025 next
= pmd_addr_end(addr
, end
);
2026 if (remap_pte_range(mm
, pmd
, addr
, next
,
2027 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
2029 } while (pmd
++, addr
= next
, addr
!= end
);
2033 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2034 unsigned long addr
, unsigned long end
,
2035 unsigned long pfn
, pgprot_t prot
)
2040 pfn
-= addr
>> PAGE_SHIFT
;
2041 pud
= pud_alloc(mm
, pgd
, addr
);
2045 next
= pud_addr_end(addr
, end
);
2046 if (remap_pmd_range(mm
, pud
, addr
, next
,
2047 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
2049 } while (pud
++, addr
= next
, addr
!= end
);
2054 * remap_pfn_range - remap kernel memory to userspace
2055 * @vma: user vma to map to
2056 * @addr: target user address to start at
2057 * @pfn: physical address of kernel memory
2058 * @size: size of map area
2059 * @prot: page protection flags for this mapping
2061 * Note: this is only safe if the mm semaphore is held when called.
2063 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2064 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2068 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2069 struct mm_struct
*mm
= vma
->vm_mm
;
2073 * Physically remapped pages are special. Tell the
2074 * rest of the world about it:
2075 * VM_IO tells people not to look at these pages
2076 * (accesses can have side effects).
2077 * VM_RESERVED is specified all over the place, because
2078 * in 2.4 it kept swapout's vma scan off this vma; but
2079 * in 2.6 the LRU scan won't even find its pages, so this
2080 * flag means no more than count its pages in reserved_vm,
2081 * and omit it from core dump, even when VM_IO turned off.
2082 * VM_PFNMAP tells the core MM that the base pages are just
2083 * raw PFN mappings, and do not have a "struct page" associated
2086 * There's a horrible special case to handle copy-on-write
2087 * behaviour that some programs depend on. We mark the "original"
2088 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2090 if (addr
== vma
->vm_start
&& end
== vma
->vm_end
) {
2091 vma
->vm_pgoff
= pfn
;
2092 vma
->vm_flags
|= VM_PFN_AT_MMAP
;
2093 } else if (is_cow_mapping(vma
->vm_flags
))
2096 vma
->vm_flags
|= VM_IO
| VM_RESERVED
| VM_PFNMAP
;
2098 err
= track_pfn_vma_new(vma
, &prot
, pfn
, PAGE_ALIGN(size
));
2101 * To indicate that track_pfn related cleanup is not
2102 * needed from higher level routine calling unmap_vmas
2104 vma
->vm_flags
&= ~(VM_IO
| VM_RESERVED
| VM_PFNMAP
);
2105 vma
->vm_flags
&= ~VM_PFN_AT_MMAP
;
2109 BUG_ON(addr
>= end
);
2110 pfn
-= addr
>> PAGE_SHIFT
;
2111 pgd
= pgd_offset(mm
, addr
);
2112 flush_cache_range(vma
, addr
, end
);
2114 next
= pgd_addr_end(addr
, end
);
2115 err
= remap_pud_range(mm
, pgd
, addr
, next
,
2116 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2119 } while (pgd
++, addr
= next
, addr
!= end
);
2122 untrack_pfn_vma(vma
, pfn
, PAGE_ALIGN(size
));
2126 EXPORT_SYMBOL(remap_pfn_range
);
2128 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2129 unsigned long addr
, unsigned long end
,
2130 pte_fn_t fn
, void *data
)
2135 spinlock_t
*uninitialized_var(ptl
);
2137 pte
= (mm
== &init_mm
) ?
2138 pte_alloc_kernel(pmd
, addr
) :
2139 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2143 BUG_ON(pmd_huge(*pmd
));
2145 arch_enter_lazy_mmu_mode();
2147 token
= pmd_pgtable(*pmd
);
2150 err
= fn(pte
++, token
, addr
, data
);
2153 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2155 arch_leave_lazy_mmu_mode();
2158 pte_unmap_unlock(pte
-1, ptl
);
2162 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2163 unsigned long addr
, unsigned long end
,
2164 pte_fn_t fn
, void *data
)
2170 BUG_ON(pud_huge(*pud
));
2172 pmd
= pmd_alloc(mm
, pud
, addr
);
2176 next
= pmd_addr_end(addr
, end
);
2177 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
2180 } while (pmd
++, addr
= next
, addr
!= end
);
2184 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2185 unsigned long addr
, unsigned long end
,
2186 pte_fn_t fn
, void *data
)
2192 pud
= pud_alloc(mm
, pgd
, addr
);
2196 next
= pud_addr_end(addr
, end
);
2197 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
2200 } while (pud
++, addr
= next
, addr
!= end
);
2205 * Scan a region of virtual memory, filling in page tables as necessary
2206 * and calling a provided function on each leaf page table.
2208 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2209 unsigned long size
, pte_fn_t fn
, void *data
)
2213 unsigned long end
= addr
+ size
;
2216 BUG_ON(addr
>= end
);
2217 pgd
= pgd_offset(mm
, addr
);
2219 next
= pgd_addr_end(addr
, end
);
2220 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
2223 } while (pgd
++, addr
= next
, addr
!= end
);
2227 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2230 * handle_pte_fault chooses page fault handler according to an entry
2231 * which was read non-atomically. Before making any commitment, on
2232 * those architectures or configurations (e.g. i386 with PAE) which
2233 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2234 * must check under lock before unmapping the pte and proceeding
2235 * (but do_wp_page is only called after already making such a check;
2236 * and do_anonymous_page can safely check later on).
2238 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2239 pte_t
*page_table
, pte_t orig_pte
)
2242 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2243 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2244 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2246 same
= pte_same(*page_table
, orig_pte
);
2250 pte_unmap(page_table
);
2254 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
2257 * If the source page was a PFN mapping, we don't have
2258 * a "struct page" for it. We do a best-effort copy by
2259 * just copying from the original user address. If that
2260 * fails, we just zero-fill it. Live with it.
2262 if (unlikely(!src
)) {
2263 void *kaddr
= kmap_atomic(dst
, KM_USER0
);
2264 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
2267 * This really shouldn't fail, because the page is there
2268 * in the page tables. But it might just be unreadable,
2269 * in which case we just give up and fill the result with
2272 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
2274 kunmap_atomic(kaddr
, KM_USER0
);
2275 flush_dcache_page(dst
);
2277 copy_user_highpage(dst
, src
, va
, vma
);
2281 * This routine handles present pages, when users try to write
2282 * to a shared page. It is done by copying the page to a new address
2283 * and decrementing the shared-page counter for the old page.
2285 * Note that this routine assumes that the protection checks have been
2286 * done by the caller (the low-level page fault routine in most cases).
2287 * Thus we can safely just mark it writable once we've done any necessary
2290 * We also mark the page dirty at this point even though the page will
2291 * change only once the write actually happens. This avoids a few races,
2292 * and potentially makes it more efficient.
2294 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2295 * but allow concurrent faults), with pte both mapped and locked.
2296 * We return with mmap_sem still held, but pte unmapped and unlocked.
2298 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2299 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2300 spinlock_t
*ptl
, pte_t orig_pte
)
2303 struct page
*old_page
, *new_page
;
2306 int page_mkwrite
= 0;
2307 struct page
*dirty_page
= NULL
;
2309 old_page
= vm_normal_page(vma
, address
, orig_pte
);
2312 * VM_MIXEDMAP !pfn_valid() case
2314 * We should not cow pages in a shared writeable mapping.
2315 * Just mark the pages writable as we can't do any dirty
2316 * accounting on raw pfn maps.
2318 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2319 (VM_WRITE
|VM_SHARED
))
2325 * Take out anonymous pages first, anonymous shared vmas are
2326 * not dirty accountable.
2328 if (PageAnon(old_page
) && !PageKsm(old_page
)) {
2329 if (!trylock_page(old_page
)) {
2330 page_cache_get(old_page
);
2331 pte_unmap_unlock(page_table
, ptl
);
2332 lock_page(old_page
);
2333 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2335 if (!pte_same(*page_table
, orig_pte
)) {
2336 unlock_page(old_page
);
2339 page_cache_release(old_page
);
2341 if (reuse_swap_page(old_page
)) {
2343 * The page is all ours. Move it to our anon_vma so
2344 * the rmap code will not search our parent or siblings.
2345 * Protected against the rmap code by the page lock.
2347 page_move_anon_rmap(old_page
, vma
, address
);
2348 unlock_page(old_page
);
2351 unlock_page(old_page
);
2352 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2353 (VM_WRITE
|VM_SHARED
))) {
2355 * Only catch write-faults on shared writable pages,
2356 * read-only shared pages can get COWed by
2357 * get_user_pages(.write=1, .force=1).
2359 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2360 struct vm_fault vmf
;
2363 vmf
.virtual_address
= (void __user
*)(address
&
2365 vmf
.pgoff
= old_page
->index
;
2366 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2367 vmf
.page
= old_page
;
2370 * Notify the address space that the page is about to
2371 * become writable so that it can prohibit this or wait
2372 * for the page to get into an appropriate state.
2374 * We do this without the lock held, so that it can
2375 * sleep if it needs to.
2377 page_cache_get(old_page
);
2378 pte_unmap_unlock(page_table
, ptl
);
2380 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
2382 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
2384 goto unwritable_page
;
2386 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
2387 lock_page(old_page
);
2388 if (!old_page
->mapping
) {
2389 ret
= 0; /* retry the fault */
2390 unlock_page(old_page
);
2391 goto unwritable_page
;
2394 VM_BUG_ON(!PageLocked(old_page
));
2397 * Since we dropped the lock we need to revalidate
2398 * the PTE as someone else may have changed it. If
2399 * they did, we just return, as we can count on the
2400 * MMU to tell us if they didn't also make it writable.
2402 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2404 if (!pte_same(*page_table
, orig_pte
)) {
2405 unlock_page(old_page
);
2411 dirty_page
= old_page
;
2412 get_page(dirty_page
);
2415 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2416 entry
= pte_mkyoung(orig_pte
);
2417 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2418 if (ptep_set_access_flags(vma
, address
, page_table
, entry
,1))
2419 update_mmu_cache(vma
, address
, page_table
);
2420 pte_unmap_unlock(page_table
, ptl
);
2421 ret
|= VM_FAULT_WRITE
;
2427 * Yes, Virginia, this is actually required to prevent a race
2428 * with clear_page_dirty_for_io() from clearing the page dirty
2429 * bit after it clear all dirty ptes, but before a racing
2430 * do_wp_page installs a dirty pte.
2432 * __do_fault is protected similarly.
2434 if (!page_mkwrite
) {
2435 wait_on_page_locked(dirty_page
);
2436 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2438 put_page(dirty_page
);
2440 struct address_space
*mapping
= dirty_page
->mapping
;
2442 set_page_dirty(dirty_page
);
2443 unlock_page(dirty_page
);
2444 page_cache_release(dirty_page
);
2447 * Some device drivers do not set page.mapping
2448 * but still dirty their pages
2450 balance_dirty_pages_ratelimited(mapping
);
2454 /* file_update_time outside page_lock */
2456 file_update_time(vma
->vm_file
);
2462 * Ok, we need to copy. Oh, well..
2464 page_cache_get(old_page
);
2466 pte_unmap_unlock(page_table
, ptl
);
2468 if (unlikely(anon_vma_prepare(vma
)))
2471 if (is_zero_pfn(pte_pfn(orig_pte
))) {
2472 new_page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2476 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2479 cow_user_page(new_page
, old_page
, address
, vma
);
2481 __SetPageUptodate(new_page
);
2483 if (mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
))
2487 * Re-check the pte - we dropped the lock
2489 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2490 if (likely(pte_same(*page_table
, orig_pte
))) {
2492 if (!PageAnon(old_page
)) {
2493 dec_mm_counter_fast(mm
, MM_FILEPAGES
);
2494 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2497 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2498 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2499 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2500 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2502 * Clear the pte entry and flush it first, before updating the
2503 * pte with the new entry. This will avoid a race condition
2504 * seen in the presence of one thread doing SMC and another
2507 ptep_clear_flush(vma
, address
, page_table
);
2508 page_add_new_anon_rmap(new_page
, vma
, address
);
2510 * We call the notify macro here because, when using secondary
2511 * mmu page tables (such as kvm shadow page tables), we want the
2512 * new page to be mapped directly into the secondary page table.
2514 set_pte_at_notify(mm
, address
, page_table
, entry
);
2515 update_mmu_cache(vma
, address
, page_table
);
2518 * Only after switching the pte to the new page may
2519 * we remove the mapcount here. Otherwise another
2520 * process may come and find the rmap count decremented
2521 * before the pte is switched to the new page, and
2522 * "reuse" the old page writing into it while our pte
2523 * here still points into it and can be read by other
2526 * The critical issue is to order this
2527 * page_remove_rmap with the ptp_clear_flush above.
2528 * Those stores are ordered by (if nothing else,)
2529 * the barrier present in the atomic_add_negative
2530 * in page_remove_rmap.
2532 * Then the TLB flush in ptep_clear_flush ensures that
2533 * no process can access the old page before the
2534 * decremented mapcount is visible. And the old page
2535 * cannot be reused until after the decremented
2536 * mapcount is visible. So transitively, TLBs to
2537 * old page will be flushed before it can be reused.
2539 page_remove_rmap(old_page
);
2542 /* Free the old page.. */
2543 new_page
= old_page
;
2544 ret
|= VM_FAULT_WRITE
;
2546 mem_cgroup_uncharge_page(new_page
);
2549 page_cache_release(new_page
);
2551 pte_unmap_unlock(page_table
, ptl
);
2554 * Don't let another task, with possibly unlocked vma,
2555 * keep the mlocked page.
2557 if ((ret
& VM_FAULT_WRITE
) && (vma
->vm_flags
& VM_LOCKED
)) {
2558 lock_page(old_page
); /* LRU manipulation */
2559 munlock_vma_page(old_page
);
2560 unlock_page(old_page
);
2562 page_cache_release(old_page
);
2566 page_cache_release(new_page
);
2570 unlock_page(old_page
);
2571 page_cache_release(old_page
);
2573 page_cache_release(old_page
);
2575 return VM_FAULT_OOM
;
2578 page_cache_release(old_page
);
2582 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2583 unsigned long start_addr
, unsigned long end_addr
,
2584 struct zap_details
*details
)
2586 zap_page_range(vma
, start_addr
, end_addr
- start_addr
, details
);
2589 static inline void unmap_mapping_range_tree(struct prio_tree_root
*root
,
2590 struct zap_details
*details
)
2592 struct vm_area_struct
*vma
;
2593 struct prio_tree_iter iter
;
2594 pgoff_t vba
, vea
, zba
, zea
;
2596 vma_prio_tree_foreach(vma
, &iter
, root
,
2597 details
->first_index
, details
->last_index
) {
2599 vba
= vma
->vm_pgoff
;
2600 vea
= vba
+ ((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) - 1;
2601 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2602 zba
= details
->first_index
;
2605 zea
= details
->last_index
;
2609 unmap_mapping_range_vma(vma
,
2610 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2611 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2616 static inline void unmap_mapping_range_list(struct list_head
*head
,
2617 struct zap_details
*details
)
2619 struct vm_area_struct
*vma
;
2622 * In nonlinear VMAs there is no correspondence between virtual address
2623 * offset and file offset. So we must perform an exhaustive search
2624 * across *all* the pages in each nonlinear VMA, not just the pages
2625 * whose virtual address lies outside the file truncation point.
2627 list_for_each_entry(vma
, head
, shared
.vm_set
.list
) {
2628 details
->nonlinear_vma
= vma
;
2629 unmap_mapping_range_vma(vma
, vma
->vm_start
, vma
->vm_end
, details
);
2634 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2635 * @mapping: the address space containing mmaps to be unmapped.
2636 * @holebegin: byte in first page to unmap, relative to the start of
2637 * the underlying file. This will be rounded down to a PAGE_SIZE
2638 * boundary. Note that this is different from truncate_pagecache(), which
2639 * must keep the partial page. In contrast, we must get rid of
2641 * @holelen: size of prospective hole in bytes. This will be rounded
2642 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2644 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2645 * but 0 when invalidating pagecache, don't throw away private data.
2647 void unmap_mapping_range(struct address_space
*mapping
,
2648 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2650 struct zap_details details
;
2651 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2652 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2654 /* Check for overflow. */
2655 if (sizeof(holelen
) > sizeof(hlen
)) {
2657 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2658 if (holeend
& ~(long long)ULONG_MAX
)
2659 hlen
= ULONG_MAX
- hba
+ 1;
2662 details
.check_mapping
= even_cows
? NULL
: mapping
;
2663 details
.nonlinear_vma
= NULL
;
2664 details
.first_index
= hba
;
2665 details
.last_index
= hba
+ hlen
- 1;
2666 if (details
.last_index
< details
.first_index
)
2667 details
.last_index
= ULONG_MAX
;
2670 mutex_lock(&mapping
->i_mmap_mutex
);
2671 if (unlikely(!prio_tree_empty(&mapping
->i_mmap
)))
2672 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2673 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
2674 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
2675 mutex_unlock(&mapping
->i_mmap_mutex
);
2677 EXPORT_SYMBOL(unmap_mapping_range
);
2679 int vmtruncate_range(struct inode
*inode
, loff_t offset
, loff_t end
)
2681 struct address_space
*mapping
= inode
->i_mapping
;
2684 * If the underlying filesystem is not going to provide
2685 * a way to truncate a range of blocks (punch a hole) -
2686 * we should return failure right now.
2688 if (!inode
->i_op
->truncate_range
)
2691 mutex_lock(&inode
->i_mutex
);
2692 down_write(&inode
->i_alloc_sem
);
2693 unmap_mapping_range(mapping
, offset
, (end
- offset
), 1);
2694 truncate_inode_pages_range(mapping
, offset
, end
);
2695 unmap_mapping_range(mapping
, offset
, (end
- offset
), 1);
2696 inode
->i_op
->truncate_range(inode
, offset
, end
);
2697 up_write(&inode
->i_alloc_sem
);
2698 mutex_unlock(&inode
->i_mutex
);
2704 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2705 * but allow concurrent faults), and pte mapped but not yet locked.
2706 * We return with mmap_sem still held, but pte unmapped and unlocked.
2708 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2709 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2710 unsigned int flags
, pte_t orig_pte
)
2713 struct page
*page
, *swapcache
= NULL
;
2717 struct mem_cgroup
*ptr
;
2721 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2724 entry
= pte_to_swp_entry(orig_pte
);
2725 if (unlikely(non_swap_entry(entry
))) {
2726 if (is_migration_entry(entry
)) {
2727 migration_entry_wait(mm
, pmd
, address
);
2728 } else if (is_hwpoison_entry(entry
)) {
2729 ret
= VM_FAULT_HWPOISON
;
2731 print_bad_pte(vma
, address
, orig_pte
, NULL
);
2732 ret
= VM_FAULT_SIGBUS
;
2736 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2737 page
= lookup_swap_cache(entry
);
2739 grab_swap_token(mm
); /* Contend for token _before_ read-in */
2740 page
= swapin_readahead(entry
,
2741 GFP_HIGHUSER_MOVABLE
, vma
, address
);
2744 * Back out if somebody else faulted in this pte
2745 * while we released the pte lock.
2747 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2748 if (likely(pte_same(*page_table
, orig_pte
)))
2750 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2754 /* Had to read the page from swap area: Major fault */
2755 ret
= VM_FAULT_MAJOR
;
2756 count_vm_event(PGMAJFAULT
);
2757 } else if (PageHWPoison(page
)) {
2759 * hwpoisoned dirty swapcache pages are kept for killing
2760 * owner processes (which may be unknown at hwpoison time)
2762 ret
= VM_FAULT_HWPOISON
;
2763 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2767 locked
= lock_page_or_retry(page
, mm
, flags
);
2768 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2770 ret
|= VM_FAULT_RETRY
;
2775 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2776 * release the swapcache from under us. The page pin, and pte_same
2777 * test below, are not enough to exclude that. Even if it is still
2778 * swapcache, we need to check that the page's swap has not changed.
2780 if (unlikely(!PageSwapCache(page
) || page_private(page
) != entry
.val
))
2783 if (ksm_might_need_to_copy(page
, vma
, address
)) {
2785 page
= ksm_does_need_to_copy(page
, vma
, address
);
2787 if (unlikely(!page
)) {
2795 if (mem_cgroup_try_charge_swapin(mm
, page
, GFP_KERNEL
, &ptr
)) {
2801 * Back out if somebody else already faulted in this pte.
2803 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2804 if (unlikely(!pte_same(*page_table
, orig_pte
)))
2807 if (unlikely(!PageUptodate(page
))) {
2808 ret
= VM_FAULT_SIGBUS
;
2813 * The page isn't present yet, go ahead with the fault.
2815 * Be careful about the sequence of operations here.
2816 * To get its accounting right, reuse_swap_page() must be called
2817 * while the page is counted on swap but not yet in mapcount i.e.
2818 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2819 * must be called after the swap_free(), or it will never succeed.
2820 * Because delete_from_swap_page() may be called by reuse_swap_page(),
2821 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2822 * in page->private. In this case, a record in swap_cgroup is silently
2823 * discarded at swap_free().
2826 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2827 dec_mm_counter_fast(mm
, MM_SWAPENTS
);
2828 pte
= mk_pte(page
, vma
->vm_page_prot
);
2829 if ((flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
)) {
2830 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
2831 flags
&= ~FAULT_FLAG_WRITE
;
2832 ret
|= VM_FAULT_WRITE
;
2835 flush_icache_page(vma
, page
);
2836 set_pte_at(mm
, address
, page_table
, pte
);
2837 do_page_add_anon_rmap(page
, vma
, address
, exclusive
);
2838 /* It's better to call commit-charge after rmap is established */
2839 mem_cgroup_commit_charge_swapin(page
, ptr
);
2842 if (vm_swap_full() || (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
2843 try_to_free_swap(page
);
2847 * Hold the lock to avoid the swap entry to be reused
2848 * until we take the PT lock for the pte_same() check
2849 * (to avoid false positives from pte_same). For
2850 * further safety release the lock after the swap_free
2851 * so that the swap count won't change under a
2852 * parallel locked swapcache.
2854 unlock_page(swapcache
);
2855 page_cache_release(swapcache
);
2858 if (flags
& FAULT_FLAG_WRITE
) {
2859 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
2860 if (ret
& VM_FAULT_ERROR
)
2861 ret
&= VM_FAULT_ERROR
;
2865 /* No need to invalidate - it was non-present before */
2866 update_mmu_cache(vma
, address
, page_table
);
2868 pte_unmap_unlock(page_table
, ptl
);
2872 mem_cgroup_cancel_charge_swapin(ptr
);
2873 pte_unmap_unlock(page_table
, ptl
);
2877 page_cache_release(page
);
2879 unlock_page(swapcache
);
2880 page_cache_release(swapcache
);
2886 * This is like a special single-page "expand_{down|up}wards()",
2887 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2888 * doesn't hit another vma.
2890 static inline int check_stack_guard_page(struct vm_area_struct
*vma
, unsigned long address
)
2892 address
&= PAGE_MASK
;
2893 if ((vma
->vm_flags
& VM_GROWSDOWN
) && address
== vma
->vm_start
) {
2894 struct vm_area_struct
*prev
= vma
->vm_prev
;
2897 * Is there a mapping abutting this one below?
2899 * That's only ok if it's the same stack mapping
2900 * that has gotten split..
2902 if (prev
&& prev
->vm_end
== address
)
2903 return prev
->vm_flags
& VM_GROWSDOWN
? 0 : -ENOMEM
;
2905 expand_downwards(vma
, address
- PAGE_SIZE
);
2907 if ((vma
->vm_flags
& VM_GROWSUP
) && address
+ PAGE_SIZE
== vma
->vm_end
) {
2908 struct vm_area_struct
*next
= vma
->vm_next
;
2910 /* As VM_GROWSDOWN but s/below/above/ */
2911 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
)
2912 return next
->vm_flags
& VM_GROWSUP
? 0 : -ENOMEM
;
2914 expand_upwards(vma
, address
+ PAGE_SIZE
);
2920 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2921 * but allow concurrent faults), and pte mapped but not yet locked.
2922 * We return with mmap_sem still held, but pte unmapped and unlocked.
2924 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2925 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2932 pte_unmap(page_table
);
2934 /* Check if we need to add a guard page to the stack */
2935 if (check_stack_guard_page(vma
, address
) < 0)
2936 return VM_FAULT_SIGBUS
;
2938 /* Use the zero-page for reads */
2939 if (!(flags
& FAULT_FLAG_WRITE
)) {
2940 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(address
),
2941 vma
->vm_page_prot
));
2942 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2943 if (!pte_none(*page_table
))
2948 /* Allocate our own private page. */
2949 if (unlikely(anon_vma_prepare(vma
)))
2951 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2954 __SetPageUptodate(page
);
2956 if (mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
))
2959 entry
= mk_pte(page
, vma
->vm_page_prot
);
2960 if (vma
->vm_flags
& VM_WRITE
)
2961 entry
= pte_mkwrite(pte_mkdirty(entry
));
2963 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2964 if (!pte_none(*page_table
))
2967 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2968 page_add_new_anon_rmap(page
, vma
, address
);
2970 set_pte_at(mm
, address
, page_table
, entry
);
2972 /* No need to invalidate - it was non-present before */
2973 update_mmu_cache(vma
, address
, page_table
);
2975 pte_unmap_unlock(page_table
, ptl
);
2978 mem_cgroup_uncharge_page(page
);
2979 page_cache_release(page
);
2982 page_cache_release(page
);
2984 return VM_FAULT_OOM
;
2988 * __do_fault() tries to create a new page mapping. It aggressively
2989 * tries to share with existing pages, but makes a separate copy if
2990 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2991 * the next page fault.
2993 * As this is called only for pages that do not currently exist, we
2994 * do not need to flush old virtual caches or the TLB.
2996 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2997 * but allow concurrent faults), and pte neither mapped nor locked.
2998 * We return with mmap_sem still held, but pte unmapped and unlocked.
3000 static int __do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3001 unsigned long address
, pmd_t
*pmd
,
3002 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3010 struct page
*dirty_page
= NULL
;
3011 struct vm_fault vmf
;
3013 int page_mkwrite
= 0;
3015 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
3020 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
3021 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
3025 if (unlikely(PageHWPoison(vmf
.page
))) {
3026 if (ret
& VM_FAULT_LOCKED
)
3027 unlock_page(vmf
.page
);
3028 return VM_FAULT_HWPOISON
;
3032 * For consistency in subsequent calls, make the faulted page always
3035 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3036 lock_page(vmf
.page
);
3038 VM_BUG_ON(!PageLocked(vmf
.page
));
3041 * Should we do an early C-O-W break?
3044 if (flags
& FAULT_FLAG_WRITE
) {
3045 if (!(vma
->vm_flags
& VM_SHARED
)) {
3047 if (unlikely(anon_vma_prepare(vma
))) {
3051 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
,
3057 if (mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
)) {
3059 page_cache_release(page
);
3063 copy_user_highpage(page
, vmf
.page
, address
, vma
);
3064 __SetPageUptodate(page
);
3067 * If the page will be shareable, see if the backing
3068 * address space wants to know that the page is about
3069 * to become writable
3071 if (vma
->vm_ops
->page_mkwrite
) {
3075 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
3076 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
3078 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3080 goto unwritable_page
;
3082 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
3084 if (!page
->mapping
) {
3085 ret
= 0; /* retry the fault */
3087 goto unwritable_page
;
3090 VM_BUG_ON(!PageLocked(page
));
3097 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3100 * This silly early PAGE_DIRTY setting removes a race
3101 * due to the bad i386 page protection. But it's valid
3102 * for other architectures too.
3104 * Note that if FAULT_FLAG_WRITE is set, we either now have
3105 * an exclusive copy of the page, or this is a shared mapping,
3106 * so we can make it writable and dirty to avoid having to
3107 * handle that later.
3109 /* Only go through if we didn't race with anybody else... */
3110 if (likely(pte_same(*page_table
, orig_pte
))) {
3111 flush_icache_page(vma
, page
);
3112 entry
= mk_pte(page
, vma
->vm_page_prot
);
3113 if (flags
& FAULT_FLAG_WRITE
)
3114 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3116 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3117 page_add_new_anon_rmap(page
, vma
, address
);
3119 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
3120 page_add_file_rmap(page
);
3121 if (flags
& FAULT_FLAG_WRITE
) {
3123 get_page(dirty_page
);
3126 set_pte_at(mm
, address
, page_table
, entry
);
3128 /* no need to invalidate: a not-present page won't be cached */
3129 update_mmu_cache(vma
, address
, page_table
);
3132 mem_cgroup_uncharge_page(page
);
3134 page_cache_release(page
);
3136 anon
= 1; /* no anon but release faulted_page */
3139 pte_unmap_unlock(page_table
, ptl
);
3143 struct address_space
*mapping
= page
->mapping
;
3145 if (set_page_dirty(dirty_page
))
3147 unlock_page(dirty_page
);
3148 put_page(dirty_page
);
3149 if (page_mkwrite
&& mapping
) {
3151 * Some device drivers do not set page.mapping but still
3154 balance_dirty_pages_ratelimited(mapping
);
3157 /* file_update_time outside page_lock */
3159 file_update_time(vma
->vm_file
);
3161 unlock_page(vmf
.page
);
3163 page_cache_release(vmf
.page
);
3169 page_cache_release(page
);
3173 static int do_linear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3174 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3175 unsigned int flags
, pte_t orig_pte
)
3177 pgoff_t pgoff
= (((address
& PAGE_MASK
)
3178 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
3180 pte_unmap(page_table
);
3181 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3185 * Fault of a previously existing named mapping. Repopulate the pte
3186 * from the encoded file_pte if possible. This enables swappable
3189 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3190 * but allow concurrent faults), and pte mapped but not yet locked.
3191 * We return with mmap_sem still held, but pte unmapped and unlocked.
3193 static int do_nonlinear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3194 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3195 unsigned int flags
, pte_t orig_pte
)
3199 flags
|= FAULT_FLAG_NONLINEAR
;
3201 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
3204 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
))) {
3206 * Page table corrupted: show pte and kill process.
3208 print_bad_pte(vma
, address
, orig_pte
, NULL
);
3209 return VM_FAULT_SIGBUS
;
3212 pgoff
= pte_to_pgoff(orig_pte
);
3213 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3217 * These routines also need to handle stuff like marking pages dirty
3218 * and/or accessed for architectures that don't do it in hardware (most
3219 * RISC architectures). The early dirtying is also good on the i386.
3221 * There is also a hook called "update_mmu_cache()" that architectures
3222 * with external mmu caches can use to update those (ie the Sparc or
3223 * PowerPC hashed page tables that act as extended TLBs).
3225 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3226 * but allow concurrent faults), and pte mapped but not yet locked.
3227 * We return with mmap_sem still held, but pte unmapped and unlocked.
3229 int handle_pte_fault(struct mm_struct
*mm
,
3230 struct vm_area_struct
*vma
, unsigned long address
,
3231 pte_t
*pte
, pmd_t
*pmd
, unsigned int flags
)
3237 if (!pte_present(entry
)) {
3238 if (pte_none(entry
)) {
3240 if (likely(vma
->vm_ops
->fault
))
3241 return do_linear_fault(mm
, vma
, address
,
3242 pte
, pmd
, flags
, entry
);
3244 return do_anonymous_page(mm
, vma
, address
,
3247 if (pte_file(entry
))
3248 return do_nonlinear_fault(mm
, vma
, address
,
3249 pte
, pmd
, flags
, entry
);
3250 return do_swap_page(mm
, vma
, address
,
3251 pte
, pmd
, flags
, entry
);
3254 ptl
= pte_lockptr(mm
, pmd
);
3256 if (unlikely(!pte_same(*pte
, entry
)))
3258 if (flags
& FAULT_FLAG_WRITE
) {
3259 if (!pte_write(entry
))
3260 return do_wp_page(mm
, vma
, address
,
3261 pte
, pmd
, ptl
, entry
);
3262 entry
= pte_mkdirty(entry
);
3264 entry
= pte_mkyoung(entry
);
3265 if (ptep_set_access_flags(vma
, address
, pte
, entry
, flags
& FAULT_FLAG_WRITE
)) {
3266 update_mmu_cache(vma
, address
, pte
);
3269 * This is needed only for protection faults but the arch code
3270 * is not yet telling us if this is a protection fault or not.
3271 * This still avoids useless tlb flushes for .text page faults
3274 if (flags
& FAULT_FLAG_WRITE
)
3275 flush_tlb_fix_spurious_fault(vma
, address
);
3278 pte_unmap_unlock(pte
, ptl
);
3283 * By the time we get here, we already hold the mm semaphore
3285 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3286 unsigned long address
, unsigned int flags
)
3293 __set_current_state(TASK_RUNNING
);
3295 count_vm_event(PGFAULT
);
3297 /* do counter updates before entering really critical section. */
3298 check_sync_rss_stat(current
);
3300 if (unlikely(is_vm_hugetlb_page(vma
)))
3301 return hugetlb_fault(mm
, vma
, address
, flags
);
3303 pgd
= pgd_offset(mm
, address
);
3304 pud
= pud_alloc(mm
, pgd
, address
);
3306 return VM_FAULT_OOM
;
3307 pmd
= pmd_alloc(mm
, pud
, address
);
3309 return VM_FAULT_OOM
;
3310 if (pmd_none(*pmd
) && transparent_hugepage_enabled(vma
)) {
3312 return do_huge_pmd_anonymous_page(mm
, vma
, address
,
3315 pmd_t orig_pmd
= *pmd
;
3317 if (pmd_trans_huge(orig_pmd
)) {
3318 if (flags
& FAULT_FLAG_WRITE
&&
3319 !pmd_write(orig_pmd
) &&
3320 !pmd_trans_splitting(orig_pmd
))
3321 return do_huge_pmd_wp_page(mm
, vma
, address
,
3328 * Use __pte_alloc instead of pte_alloc_map, because we can't
3329 * run pte_offset_map on the pmd, if an huge pmd could
3330 * materialize from under us from a different thread.
3332 if (unlikely(pmd_none(*pmd
)) && __pte_alloc(mm
, vma
, pmd
, address
))
3333 return VM_FAULT_OOM
;
3334 /* if an huge pmd materialized from under us just retry later */
3335 if (unlikely(pmd_trans_huge(*pmd
)))
3338 * A regular pmd is established and it can't morph into a huge pmd
3339 * from under us anymore at this point because we hold the mmap_sem
3340 * read mode and khugepaged takes it in write mode. So now it's
3341 * safe to run pte_offset_map().
3343 pte
= pte_offset_map(pmd
, address
);
3345 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
3348 #ifndef __PAGETABLE_PUD_FOLDED
3350 * Allocate page upper directory.
3351 * We've already handled the fast-path in-line.
3353 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
3355 pud_t
*new = pud_alloc_one(mm
, address
);
3359 smp_wmb(); /* See comment in __pte_alloc */
3361 spin_lock(&mm
->page_table_lock
);
3362 if (pgd_present(*pgd
)) /* Another has populated it */
3365 pgd_populate(mm
, pgd
, new);
3366 spin_unlock(&mm
->page_table_lock
);
3369 #endif /* __PAGETABLE_PUD_FOLDED */
3371 #ifndef __PAGETABLE_PMD_FOLDED
3373 * Allocate page middle directory.
3374 * We've already handled the fast-path in-line.
3376 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
3378 pmd_t
*new = pmd_alloc_one(mm
, address
);
3382 smp_wmb(); /* See comment in __pte_alloc */
3384 spin_lock(&mm
->page_table_lock
);
3385 #ifndef __ARCH_HAS_4LEVEL_HACK
3386 if (pud_present(*pud
)) /* Another has populated it */
3389 pud_populate(mm
, pud
, new);
3391 if (pgd_present(*pud
)) /* Another has populated it */
3394 pgd_populate(mm
, pud
, new);
3395 #endif /* __ARCH_HAS_4LEVEL_HACK */
3396 spin_unlock(&mm
->page_table_lock
);
3399 #endif /* __PAGETABLE_PMD_FOLDED */
3401 int make_pages_present(unsigned long addr
, unsigned long end
)
3403 int ret
, len
, write
;
3404 struct vm_area_struct
* vma
;
3406 vma
= find_vma(current
->mm
, addr
);
3410 * We want to touch writable mappings with a write fault in order
3411 * to break COW, except for shared mappings because these don't COW
3412 * and we would not want to dirty them for nothing.
3414 write
= (vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
;
3415 BUG_ON(addr
>= end
);
3416 BUG_ON(end
> vma
->vm_end
);
3417 len
= DIV_ROUND_UP(end
, PAGE_SIZE
) - addr
/PAGE_SIZE
;
3418 ret
= get_user_pages(current
, current
->mm
, addr
,
3419 len
, write
, 0, NULL
, NULL
);
3422 return ret
== len
? 0 : -EFAULT
;
3425 #if !defined(__HAVE_ARCH_GATE_AREA)
3427 #if defined(AT_SYSINFO_EHDR)
3428 static struct vm_area_struct gate_vma
;
3430 static int __init
gate_vma_init(void)
3432 gate_vma
.vm_mm
= NULL
;
3433 gate_vma
.vm_start
= FIXADDR_USER_START
;
3434 gate_vma
.vm_end
= FIXADDR_USER_END
;
3435 gate_vma
.vm_flags
= VM_READ
| VM_MAYREAD
| VM_EXEC
| VM_MAYEXEC
;
3436 gate_vma
.vm_page_prot
= __P101
;
3438 * Make sure the vDSO gets into every core dump.
3439 * Dumping its contents makes post-mortem fully interpretable later
3440 * without matching up the same kernel and hardware config to see
3441 * what PC values meant.
3443 gate_vma
.vm_flags
|= VM_ALWAYSDUMP
;
3446 __initcall(gate_vma_init
);
3449 struct vm_area_struct
*get_gate_vma(struct mm_struct
*mm
)
3451 #ifdef AT_SYSINFO_EHDR
3458 int in_gate_area_no_mm(unsigned long addr
)
3460 #ifdef AT_SYSINFO_EHDR
3461 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
3467 #endif /* __HAVE_ARCH_GATE_AREA */
3469 static int __follow_pte(struct mm_struct
*mm
, unsigned long address
,
3470 pte_t
**ptepp
, spinlock_t
**ptlp
)
3477 pgd
= pgd_offset(mm
, address
);
3478 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
3481 pud
= pud_offset(pgd
, address
);
3482 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
3485 pmd
= pmd_offset(pud
, address
);
3486 VM_BUG_ON(pmd_trans_huge(*pmd
));
3487 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
3490 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3494 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
3497 if (!pte_present(*ptep
))
3502 pte_unmap_unlock(ptep
, *ptlp
);
3507 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
3508 pte_t
**ptepp
, spinlock_t
**ptlp
)
3512 /* (void) is needed to make gcc happy */
3513 (void) __cond_lock(*ptlp
,
3514 !(res
= __follow_pte(mm
, address
, ptepp
, ptlp
)));
3519 * follow_pfn - look up PFN at a user virtual address
3520 * @vma: memory mapping
3521 * @address: user virtual address
3522 * @pfn: location to store found PFN
3524 * Only IO mappings and raw PFN mappings are allowed.
3526 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3528 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
3535 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3538 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
3541 *pfn
= pte_pfn(*ptep
);
3542 pte_unmap_unlock(ptep
, ptl
);
3545 EXPORT_SYMBOL(follow_pfn
);
3547 #ifdef CONFIG_HAVE_IOREMAP_PROT
3548 int follow_phys(struct vm_area_struct
*vma
,
3549 unsigned long address
, unsigned int flags
,
3550 unsigned long *prot
, resource_size_t
*phys
)
3556 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3559 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
3563 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
3566 *prot
= pgprot_val(pte_pgprot(pte
));
3567 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
3571 pte_unmap_unlock(ptep
, ptl
);
3576 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
3577 void *buf
, int len
, int write
)
3579 resource_size_t phys_addr
;
3580 unsigned long prot
= 0;
3581 void __iomem
*maddr
;
3582 int offset
= addr
& (PAGE_SIZE
-1);
3584 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
3587 maddr
= ioremap_prot(phys_addr
, PAGE_SIZE
, prot
);
3589 memcpy_toio(maddr
+ offset
, buf
, len
);
3591 memcpy_fromio(buf
, maddr
+ offset
, len
);
3599 * Access another process' address space as given in mm. If non-NULL, use the
3600 * given task for page fault accounting.
3602 static int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
3603 unsigned long addr
, void *buf
, int len
, int write
)
3605 struct vm_area_struct
*vma
;
3606 void *old_buf
= buf
;
3608 down_read(&mm
->mmap_sem
);
3609 /* ignore errors, just check how much was successfully transferred */
3611 int bytes
, ret
, offset
;
3613 struct page
*page
= NULL
;
3615 ret
= get_user_pages(tsk
, mm
, addr
, 1,
3616 write
, 1, &page
, &vma
);
3619 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3620 * we can access using slightly different code.
3622 #ifdef CONFIG_HAVE_IOREMAP_PROT
3623 vma
= find_vma(mm
, addr
);
3624 if (!vma
|| vma
->vm_start
> addr
)
3626 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
3627 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
3635 offset
= addr
& (PAGE_SIZE
-1);
3636 if (bytes
> PAGE_SIZE
-offset
)
3637 bytes
= PAGE_SIZE
-offset
;
3641 copy_to_user_page(vma
, page
, addr
,
3642 maddr
+ offset
, buf
, bytes
);
3643 set_page_dirty_lock(page
);
3645 copy_from_user_page(vma
, page
, addr
,
3646 buf
, maddr
+ offset
, bytes
);
3649 page_cache_release(page
);
3655 up_read(&mm
->mmap_sem
);
3657 return buf
- old_buf
;
3661 * access_remote_vm - access another process' address space
3662 * @mm: the mm_struct of the target address space
3663 * @addr: start address to access
3664 * @buf: source or destination buffer
3665 * @len: number of bytes to transfer
3666 * @write: whether the access is a write
3668 * The caller must hold a reference on @mm.
3670 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
3671 void *buf
, int len
, int write
)
3673 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, write
);
3677 * Access another process' address space.
3678 * Source/target buffer must be kernel space,
3679 * Do not walk the page table directly, use get_user_pages
3681 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
3682 void *buf
, int len
, int write
)
3684 struct mm_struct
*mm
;
3687 mm
= get_task_mm(tsk
);
3691 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, write
);
3698 * Print the name of a VMA.
3700 void print_vma_addr(char *prefix
, unsigned long ip
)
3702 struct mm_struct
*mm
= current
->mm
;
3703 struct vm_area_struct
*vma
;
3706 * Do not print if we are in atomic
3707 * contexts (in exception stacks, etc.):
3709 if (preempt_count())
3712 down_read(&mm
->mmap_sem
);
3713 vma
= find_vma(mm
, ip
);
3714 if (vma
&& vma
->vm_file
) {
3715 struct file
*f
= vma
->vm_file
;
3716 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
3720 p
= d_path(&f
->f_path
, buf
, PAGE_SIZE
);
3723 s
= strrchr(p
, '/');
3726 printk("%s%s[%lx+%lx]", prefix
, p
,
3728 vma
->vm_end
- vma
->vm_start
);
3729 free_page((unsigned long)buf
);
3732 up_read(¤t
->mm
->mmap_sem
);
3735 #ifdef CONFIG_PROVE_LOCKING
3736 void might_fault(void)
3739 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3740 * holding the mmap_sem, this is safe because kernel memory doesn't
3741 * get paged out, therefore we'll never actually fault, and the
3742 * below annotations will generate false positives.
3744 if (segment_eq(get_fs(), KERNEL_DS
))
3749 * it would be nicer only to annotate paths which are not under
3750 * pagefault_disable, however that requires a larger audit and
3751 * providing helpers like get_user_atomic.
3753 if (!in_atomic() && current
->mm
)
3754 might_lock_read(¤t
->mm
->mmap_sem
);
3756 EXPORT_SYMBOL(might_fault
);
3759 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3760 static void clear_gigantic_page(struct page
*page
,
3762 unsigned int pages_per_huge_page
)
3765 struct page
*p
= page
;
3768 for (i
= 0; i
< pages_per_huge_page
;
3769 i
++, p
= mem_map_next(p
, page
, i
)) {
3771 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
3774 void clear_huge_page(struct page
*page
,
3775 unsigned long addr
, unsigned int pages_per_huge_page
)
3779 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3780 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
3785 for (i
= 0; i
< pages_per_huge_page
; i
++) {
3787 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
3791 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
3793 struct vm_area_struct
*vma
,
3794 unsigned int pages_per_huge_page
)
3797 struct page
*dst_base
= dst
;
3798 struct page
*src_base
= src
;
3800 for (i
= 0; i
< pages_per_huge_page
; ) {
3802 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
3805 dst
= mem_map_next(dst
, dst_base
, i
);
3806 src
= mem_map_next(src
, src_base
, i
);
3810 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
3811 unsigned long addr
, struct vm_area_struct
*vma
,
3812 unsigned int pages_per_huge_page
)
3816 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3817 copy_user_gigantic_page(dst
, src
, addr
, vma
,
3818 pages_per_huge_page
);
3823 for (i
= 0; i
< pages_per_huge_page
; i
++) {
3825 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
3828 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */