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/export.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>
60 #include <linux/migrate.h>
61 #include <linux/string.h>
64 #include <asm/pgalloc.h>
65 #include <asm/uaccess.h>
67 #include <asm/tlbflush.h>
68 #include <asm/pgtable.h>
72 #ifdef LAST_NID_NOT_IN_PAGE_FLAGS
73 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_nid.
76 #ifndef CONFIG_NEED_MULTIPLE_NODES
77 /* use the per-pgdat data instead for discontigmem - mbligh */
78 unsigned long max_mapnr
;
81 EXPORT_SYMBOL(max_mapnr
);
82 EXPORT_SYMBOL(mem_map
);
85 unsigned long num_physpages
;
87 * A number of key systems in x86 including ioremap() rely on the assumption
88 * that high_memory defines the upper bound on direct map memory, then end
89 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
90 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
95 EXPORT_SYMBOL(num_physpages
);
96 EXPORT_SYMBOL(high_memory
);
99 * Randomize the address space (stacks, mmaps, brk, etc.).
101 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
102 * as ancient (libc5 based) binaries can segfault. )
104 int randomize_va_space __read_mostly
=
105 #ifdef CONFIG_COMPAT_BRK
111 static int __init
disable_randmaps(char *s
)
113 randomize_va_space
= 0;
116 __setup("norandmaps", disable_randmaps
);
118 unsigned long zero_pfn __read_mostly
;
119 unsigned long highest_memmap_pfn __read_mostly
;
122 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
124 static int __init
init_zero_pfn(void)
126 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
129 core_initcall(init_zero_pfn
);
132 #if defined(SPLIT_RSS_COUNTING)
134 void sync_mm_rss(struct mm_struct
*mm
)
138 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
139 if (current
->rss_stat
.count
[i
]) {
140 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
141 current
->rss_stat
.count
[i
] = 0;
144 current
->rss_stat
.events
= 0;
147 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
149 struct task_struct
*task
= current
;
151 if (likely(task
->mm
== mm
))
152 task
->rss_stat
.count
[member
] += val
;
154 add_mm_counter(mm
, member
, val
);
156 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
157 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
159 /* sync counter once per 64 page faults */
160 #define TASK_RSS_EVENTS_THRESH (64)
161 static void check_sync_rss_stat(struct task_struct
*task
)
163 if (unlikely(task
!= current
))
165 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
166 sync_mm_rss(task
->mm
);
168 #else /* SPLIT_RSS_COUNTING */
170 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
171 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
173 static void check_sync_rss_stat(struct task_struct
*task
)
177 #endif /* SPLIT_RSS_COUNTING */
179 #ifdef HAVE_GENERIC_MMU_GATHER
181 static int tlb_next_batch(struct mmu_gather
*tlb
)
183 struct mmu_gather_batch
*batch
;
187 tlb
->active
= batch
->next
;
191 if (tlb
->batch_count
== MAX_GATHER_BATCH_COUNT
)
194 batch
= (void *)__get_free_pages(GFP_NOWAIT
| __GFP_NOWARN
, 0);
201 batch
->max
= MAX_GATHER_BATCH
;
203 tlb
->active
->next
= batch
;
210 * Called to initialize an (on-stack) mmu_gather structure for page-table
211 * tear-down from @mm. The @fullmm argument is used when @mm is without
212 * users and we're going to destroy the full address space (exit/execve).
214 void tlb_gather_mmu(struct mmu_gather
*tlb
, struct mm_struct
*mm
, bool fullmm
)
218 tlb
->fullmm
= fullmm
;
219 tlb
->need_flush_all
= 0;
223 tlb
->local
.next
= NULL
;
225 tlb
->local
.max
= ARRAY_SIZE(tlb
->__pages
);
226 tlb
->active
= &tlb
->local
;
227 tlb
->batch_count
= 0;
229 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
234 void tlb_flush_mmu(struct mmu_gather
*tlb
)
236 struct mmu_gather_batch
*batch
;
238 if (!tlb
->need_flush
)
242 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
243 tlb_table_flush(tlb
);
246 for (batch
= &tlb
->local
; batch
; batch
= batch
->next
) {
247 free_pages_and_swap_cache(batch
->pages
, batch
->nr
);
250 tlb
->active
= &tlb
->local
;
254 * Called at the end of the shootdown operation to free up any resources
255 * that were required.
257 void tlb_finish_mmu(struct mmu_gather
*tlb
, unsigned long start
, unsigned long end
)
259 struct mmu_gather_batch
*batch
, *next
;
265 /* keep the page table cache within bounds */
268 for (batch
= tlb
->local
.next
; batch
; batch
= next
) {
270 free_pages((unsigned long)batch
, 0);
272 tlb
->local
.next
= NULL
;
276 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
277 * handling the additional races in SMP caused by other CPUs caching valid
278 * mappings in their TLBs. Returns the number of free page slots left.
279 * When out of page slots we must call tlb_flush_mmu().
281 int __tlb_remove_page(struct mmu_gather
*tlb
, struct page
*page
)
283 struct mmu_gather_batch
*batch
;
285 VM_BUG_ON(!tlb
->need_flush
);
288 batch
->pages
[batch
->nr
++] = page
;
289 if (batch
->nr
== batch
->max
) {
290 if (!tlb_next_batch(tlb
))
294 VM_BUG_ON(batch
->nr
> batch
->max
);
296 return batch
->max
- batch
->nr
;
299 #endif /* HAVE_GENERIC_MMU_GATHER */
301 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
304 * See the comment near struct mmu_table_batch.
307 static void tlb_remove_table_smp_sync(void *arg
)
309 /* Simply deliver the interrupt */
312 static void tlb_remove_table_one(void *table
)
315 * This isn't an RCU grace period and hence the page-tables cannot be
316 * assumed to be actually RCU-freed.
318 * It is however sufficient for software page-table walkers that rely on
319 * IRQ disabling. See the comment near struct mmu_table_batch.
321 smp_call_function(tlb_remove_table_smp_sync
, NULL
, 1);
322 __tlb_remove_table(table
);
325 static void tlb_remove_table_rcu(struct rcu_head
*head
)
327 struct mmu_table_batch
*batch
;
330 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
332 for (i
= 0; i
< batch
->nr
; i
++)
333 __tlb_remove_table(batch
->tables
[i
]);
335 free_page((unsigned long)batch
);
338 void tlb_table_flush(struct mmu_gather
*tlb
)
340 struct mmu_table_batch
**batch
= &tlb
->batch
;
343 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
348 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
350 struct mmu_table_batch
**batch
= &tlb
->batch
;
355 * When there's less then two users of this mm there cannot be a
356 * concurrent page-table walk.
358 if (atomic_read(&tlb
->mm
->mm_users
) < 2) {
359 __tlb_remove_table(table
);
363 if (*batch
== NULL
) {
364 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
365 if (*batch
== NULL
) {
366 tlb_remove_table_one(table
);
371 (*batch
)->tables
[(*batch
)->nr
++] = table
;
372 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
373 tlb_table_flush(tlb
);
376 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
379 * If a p?d_bad entry is found while walking page tables, report
380 * the error, before resetting entry to p?d_none. Usually (but
381 * very seldom) called out from the p?d_none_or_clear_bad macros.
384 void pgd_clear_bad(pgd_t
*pgd
)
390 void pud_clear_bad(pud_t
*pud
)
396 void pmd_clear_bad(pmd_t
*pmd
)
403 * Note: this doesn't free the actual pages themselves. That
404 * has been handled earlier when unmapping all the memory regions.
406 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
409 pgtable_t token
= pmd_pgtable(*pmd
);
411 pte_free_tlb(tlb
, token
, addr
);
415 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
416 unsigned long addr
, unsigned long end
,
417 unsigned long floor
, unsigned long ceiling
)
424 pmd
= pmd_offset(pud
, addr
);
426 next
= pmd_addr_end(addr
, end
);
427 if (pmd_none_or_clear_bad(pmd
))
429 free_pte_range(tlb
, pmd
, addr
);
430 } while (pmd
++, addr
= next
, addr
!= end
);
440 if (end
- 1 > ceiling
- 1)
443 pmd
= pmd_offset(pud
, start
);
445 pmd_free_tlb(tlb
, pmd
, start
);
448 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
449 unsigned long addr
, unsigned long end
,
450 unsigned long floor
, unsigned long ceiling
)
457 pud
= pud_offset(pgd
, addr
);
459 next
= pud_addr_end(addr
, end
);
460 if (pud_none_or_clear_bad(pud
))
462 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
463 } while (pud
++, addr
= next
, addr
!= end
);
469 ceiling
&= PGDIR_MASK
;
473 if (end
- 1 > ceiling
- 1)
476 pud
= pud_offset(pgd
, start
);
478 pud_free_tlb(tlb
, pud
, start
);
482 * This function frees user-level page tables of a process.
484 * Must be called with pagetable lock held.
486 void free_pgd_range(struct mmu_gather
*tlb
,
487 unsigned long addr
, unsigned long end
,
488 unsigned long floor
, unsigned long ceiling
)
494 * The next few lines have given us lots of grief...
496 * Why are we testing PMD* at this top level? Because often
497 * there will be no work to do at all, and we'd prefer not to
498 * go all the way down to the bottom just to discover that.
500 * Why all these "- 1"s? Because 0 represents both the bottom
501 * of the address space and the top of it (using -1 for the
502 * top wouldn't help much: the masks would do the wrong thing).
503 * The rule is that addr 0 and floor 0 refer to the bottom of
504 * the address space, but end 0 and ceiling 0 refer to the top
505 * Comparisons need to use "end - 1" and "ceiling - 1" (though
506 * that end 0 case should be mythical).
508 * Wherever addr is brought up or ceiling brought down, we must
509 * be careful to reject "the opposite 0" before it confuses the
510 * subsequent tests. But what about where end is brought down
511 * by PMD_SIZE below? no, end can't go down to 0 there.
513 * Whereas we round start (addr) and ceiling down, by different
514 * masks at different levels, in order to test whether a table
515 * now has no other vmas using it, so can be freed, we don't
516 * bother to round floor or end up - the tests don't need that.
530 if (end
- 1 > ceiling
- 1)
535 pgd
= pgd_offset(tlb
->mm
, addr
);
537 next
= pgd_addr_end(addr
, end
);
538 if (pgd_none_or_clear_bad(pgd
))
540 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
541 } while (pgd
++, addr
= next
, addr
!= end
);
544 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
545 unsigned long floor
, unsigned long ceiling
)
548 struct vm_area_struct
*next
= vma
->vm_next
;
549 unsigned long addr
= vma
->vm_start
;
552 * Hide vma from rmap and truncate_pagecache before freeing
555 unlink_anon_vmas(vma
);
556 unlink_file_vma(vma
);
558 if (is_vm_hugetlb_page(vma
)) {
559 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
560 floor
, next
? next
->vm_start
: ceiling
);
563 * Optimization: gather nearby vmas into one call down
565 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
566 && !is_vm_hugetlb_page(next
)) {
569 unlink_anon_vmas(vma
);
570 unlink_file_vma(vma
);
572 free_pgd_range(tlb
, addr
, vma
->vm_end
,
573 floor
, next
? next
->vm_start
: ceiling
);
579 int __pte_alloc(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
580 pmd_t
*pmd
, unsigned long address
)
582 pgtable_t
new = pte_alloc_one(mm
, address
);
583 int wait_split_huge_page
;
588 * Ensure all pte setup (eg. pte page lock and page clearing) are
589 * visible before the pte is made visible to other CPUs by being
590 * put into page tables.
592 * The other side of the story is the pointer chasing in the page
593 * table walking code (when walking the page table without locking;
594 * ie. most of the time). Fortunately, these data accesses consist
595 * of a chain of data-dependent loads, meaning most CPUs (alpha
596 * being the notable exception) will already guarantee loads are
597 * seen in-order. See the alpha page table accessors for the
598 * smp_read_barrier_depends() barriers in page table walking code.
600 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
602 spin_lock(&mm
->page_table_lock
);
603 wait_split_huge_page
= 0;
604 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
606 pmd_populate(mm
, pmd
, new);
608 } else if (unlikely(pmd_trans_splitting(*pmd
)))
609 wait_split_huge_page
= 1;
610 spin_unlock(&mm
->page_table_lock
);
613 if (wait_split_huge_page
)
614 wait_split_huge_page(vma
->anon_vma
, pmd
);
618 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
620 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
624 smp_wmb(); /* See comment in __pte_alloc */
626 spin_lock(&init_mm
.page_table_lock
);
627 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
628 pmd_populate_kernel(&init_mm
, pmd
, new);
631 VM_BUG_ON(pmd_trans_splitting(*pmd
));
632 spin_unlock(&init_mm
.page_table_lock
);
634 pte_free_kernel(&init_mm
, new);
638 static inline void init_rss_vec(int *rss
)
640 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
643 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
647 if (current
->mm
== mm
)
649 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
651 add_mm_counter(mm
, i
, rss
[i
]);
655 * This function is called to print an error when a bad pte
656 * is found. For example, we might have a PFN-mapped pte in
657 * a region that doesn't allow it.
659 * The calling function must still handle the error.
661 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
662 pte_t pte
, struct page
*page
)
664 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
665 pud_t
*pud
= pud_offset(pgd
, addr
);
666 pmd_t
*pmd
= pmd_offset(pud
, addr
);
667 struct address_space
*mapping
;
669 static unsigned long resume
;
670 static unsigned long nr_shown
;
671 static unsigned long nr_unshown
;
674 * Allow a burst of 60 reports, then keep quiet for that minute;
675 * or allow a steady drip of one report per second.
677 if (nr_shown
== 60) {
678 if (time_before(jiffies
, resume
)) {
684 "BUG: Bad page map: %lu messages suppressed\n",
691 resume
= jiffies
+ 60 * HZ
;
693 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
694 index
= linear_page_index(vma
, addr
);
697 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
699 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
703 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
704 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
706 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
709 printk(KERN_ALERT
"vma->vm_ops->fault: %pSR\n",
711 if (vma
->vm_file
&& vma
->vm_file
->f_op
)
712 printk(KERN_ALERT
"vma->vm_file->f_op->mmap: %pSR\n",
713 vma
->vm_file
->f_op
->mmap
);
715 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
718 static inline bool is_cow_mapping(vm_flags_t flags
)
720 return (flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
724 * vm_normal_page -- This function gets the "struct page" associated with a pte.
726 * "Special" mappings do not wish to be associated with a "struct page" (either
727 * it doesn't exist, or it exists but they don't want to touch it). In this
728 * case, NULL is returned here. "Normal" mappings do have a struct page.
730 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
731 * pte bit, in which case this function is trivial. Secondly, an architecture
732 * may not have a spare pte bit, which requires a more complicated scheme,
735 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
736 * special mapping (even if there are underlying and valid "struct pages").
737 * COWed pages of a VM_PFNMAP are always normal.
739 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
740 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
741 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
742 * mapping will always honor the rule
744 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
746 * And for normal mappings this is false.
748 * This restricts such mappings to be a linear translation from virtual address
749 * to pfn. To get around this restriction, we allow arbitrary mappings so long
750 * as the vma is not a COW mapping; in that case, we know that all ptes are
751 * special (because none can have been COWed).
754 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
756 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
757 * page" backing, however the difference is that _all_ pages with a struct
758 * page (that is, those where pfn_valid is true) are refcounted and considered
759 * normal pages by the VM. The disadvantage is that pages are refcounted
760 * (which can be slower and simply not an option for some PFNMAP users). The
761 * advantage is that we don't have to follow the strict linearity rule of
762 * PFNMAP mappings in order to support COWable mappings.
765 #ifdef __HAVE_ARCH_PTE_SPECIAL
766 # define HAVE_PTE_SPECIAL 1
768 # define HAVE_PTE_SPECIAL 0
770 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
773 unsigned long pfn
= pte_pfn(pte
);
775 if (HAVE_PTE_SPECIAL
) {
776 if (likely(!pte_special(pte
)))
778 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
780 if (!is_zero_pfn(pfn
))
781 print_bad_pte(vma
, addr
, pte
, NULL
);
785 /* !HAVE_PTE_SPECIAL case follows: */
787 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
788 if (vma
->vm_flags
& VM_MIXEDMAP
) {
794 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
795 if (pfn
== vma
->vm_pgoff
+ off
)
797 if (!is_cow_mapping(vma
->vm_flags
))
802 if (is_zero_pfn(pfn
))
805 if (unlikely(pfn
> highest_memmap_pfn
)) {
806 print_bad_pte(vma
, addr
, pte
, NULL
);
811 * NOTE! We still have PageReserved() pages in the page tables.
812 * eg. VDSO mappings can cause them to exist.
815 return pfn_to_page(pfn
);
819 * copy one vm_area from one task to the other. Assumes the page tables
820 * already present in the new task to be cleared in the whole range
821 * covered by this vma.
824 static inline unsigned long
825 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
826 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
827 unsigned long addr
, int *rss
)
829 unsigned long vm_flags
= vma
->vm_flags
;
830 pte_t pte
= *src_pte
;
833 /* pte contains position in swap or file, so copy. */
834 if (unlikely(!pte_present(pte
))) {
835 if (!pte_file(pte
)) {
836 swp_entry_t entry
= pte_to_swp_entry(pte
);
838 if (swap_duplicate(entry
) < 0)
841 /* make sure dst_mm is on swapoff's mmlist. */
842 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
843 spin_lock(&mmlist_lock
);
844 if (list_empty(&dst_mm
->mmlist
))
845 list_add(&dst_mm
->mmlist
,
847 spin_unlock(&mmlist_lock
);
849 if (likely(!non_swap_entry(entry
)))
851 else if (is_migration_entry(entry
)) {
852 page
= migration_entry_to_page(entry
);
859 if (is_write_migration_entry(entry
) &&
860 is_cow_mapping(vm_flags
)) {
862 * COW mappings require pages in both
863 * parent and child to be set to read.
865 make_migration_entry_read(&entry
);
866 pte
= swp_entry_to_pte(entry
);
867 set_pte_at(src_mm
, addr
, src_pte
, pte
);
875 * If it's a COW mapping, write protect it both
876 * in the parent and the child
878 if (is_cow_mapping(vm_flags
)) {
879 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
880 pte
= pte_wrprotect(pte
);
884 * If it's a shared mapping, mark it clean in
887 if (vm_flags
& VM_SHARED
)
888 pte
= pte_mkclean(pte
);
889 pte
= pte_mkold(pte
);
891 page
= vm_normal_page(vma
, addr
, pte
);
902 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
906 int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
907 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
908 unsigned long addr
, unsigned long end
)
910 pte_t
*orig_src_pte
, *orig_dst_pte
;
911 pte_t
*src_pte
, *dst_pte
;
912 spinlock_t
*src_ptl
, *dst_ptl
;
914 int rss
[NR_MM_COUNTERS
];
915 swp_entry_t entry
= (swp_entry_t
){0};
920 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
923 src_pte
= pte_offset_map(src_pmd
, addr
);
924 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
925 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
926 orig_src_pte
= src_pte
;
927 orig_dst_pte
= dst_pte
;
928 arch_enter_lazy_mmu_mode();
932 * We are holding two locks at this point - either of them
933 * could generate latencies in another task on another CPU.
935 if (progress
>= 32) {
937 if (need_resched() ||
938 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
941 if (pte_none(*src_pte
)) {
945 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
950 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
952 arch_leave_lazy_mmu_mode();
953 spin_unlock(src_ptl
);
954 pte_unmap(orig_src_pte
);
955 add_mm_rss_vec(dst_mm
, rss
);
956 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
960 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
969 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
970 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
971 unsigned long addr
, unsigned long end
)
973 pmd_t
*src_pmd
, *dst_pmd
;
976 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
979 src_pmd
= pmd_offset(src_pud
, addr
);
981 next
= pmd_addr_end(addr
, end
);
982 if (pmd_trans_huge(*src_pmd
)) {
984 VM_BUG_ON(next
-addr
!= HPAGE_PMD_SIZE
);
985 err
= copy_huge_pmd(dst_mm
, src_mm
,
986 dst_pmd
, src_pmd
, addr
, vma
);
993 if (pmd_none_or_clear_bad(src_pmd
))
995 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
998 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1002 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1003 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
1004 unsigned long addr
, unsigned long end
)
1006 pud_t
*src_pud
, *dst_pud
;
1009 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
1012 src_pud
= pud_offset(src_pgd
, addr
);
1014 next
= pud_addr_end(addr
, end
);
1015 if (pud_none_or_clear_bad(src_pud
))
1017 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
1020 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1024 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1025 struct vm_area_struct
*vma
)
1027 pgd_t
*src_pgd
, *dst_pgd
;
1029 unsigned long addr
= vma
->vm_start
;
1030 unsigned long end
= vma
->vm_end
;
1031 unsigned long mmun_start
; /* For mmu_notifiers */
1032 unsigned long mmun_end
; /* For mmu_notifiers */
1037 * Don't copy ptes where a page fault will fill them correctly.
1038 * Fork becomes much lighter when there are big shared or private
1039 * readonly mappings. The tradeoff is that copy_page_range is more
1040 * efficient than faulting.
1042 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_NONLINEAR
|
1043 VM_PFNMAP
| VM_MIXEDMAP
))) {
1048 if (is_vm_hugetlb_page(vma
))
1049 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1051 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
1053 * We do not free on error cases below as remove_vma
1054 * gets called on error from higher level routine
1056 ret
= track_pfn_copy(vma
);
1062 * We need to invalidate the secondary MMU mappings only when
1063 * there could be a permission downgrade on the ptes of the
1064 * parent mm. And a permission downgrade will only happen if
1065 * is_cow_mapping() returns true.
1067 is_cow
= is_cow_mapping(vma
->vm_flags
);
1071 mmu_notifier_invalidate_range_start(src_mm
, mmun_start
,
1075 dst_pgd
= pgd_offset(dst_mm
, addr
);
1076 src_pgd
= pgd_offset(src_mm
, addr
);
1078 next
= pgd_addr_end(addr
, end
);
1079 if (pgd_none_or_clear_bad(src_pgd
))
1081 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1082 vma
, addr
, next
))) {
1086 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1089 mmu_notifier_invalidate_range_end(src_mm
, mmun_start
, mmun_end
);
1093 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1094 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1095 unsigned long addr
, unsigned long end
,
1096 struct zap_details
*details
)
1098 struct mm_struct
*mm
= tlb
->mm
;
1099 int force_flush
= 0;
1100 int rss
[NR_MM_COUNTERS
];
1104 unsigned long range_start
= addr
;
1108 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1110 arch_enter_lazy_mmu_mode();
1113 if (pte_none(ptent
)) {
1117 if (pte_present(ptent
)) {
1120 page
= vm_normal_page(vma
, addr
, ptent
);
1121 if (unlikely(details
) && page
) {
1123 * unmap_shared_mapping_pages() wants to
1124 * invalidate cache without truncating:
1125 * unmap shared but keep private pages.
1127 if (details
->check_mapping
&&
1128 details
->check_mapping
!= page
->mapping
)
1131 * Each page->index must be checked when
1132 * invalidating or truncating nonlinear.
1134 if (details
->nonlinear_vma
&&
1135 (page
->index
< details
->first_index
||
1136 page
->index
> details
->last_index
))
1139 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1141 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1142 if (unlikely(!page
))
1144 if (unlikely(details
) && details
->nonlinear_vma
1145 && linear_page_index(details
->nonlinear_vma
,
1146 addr
) != page
->index
)
1147 set_pte_at(mm
, addr
, pte
,
1148 pgoff_to_pte(page
->index
));
1150 rss
[MM_ANONPAGES
]--;
1152 if (pte_dirty(ptent
))
1153 set_page_dirty(page
);
1154 if (pte_young(ptent
) &&
1155 likely(!VM_SequentialReadHint(vma
)))
1156 mark_page_accessed(page
);
1157 rss
[MM_FILEPAGES
]--;
1159 page_remove_rmap(page
);
1160 if (unlikely(page_mapcount(page
) < 0))
1161 print_bad_pte(vma
, addr
, ptent
, page
);
1162 force_flush
= !__tlb_remove_page(tlb
, page
);
1168 * If details->check_mapping, we leave swap entries;
1169 * if details->nonlinear_vma, we leave file entries.
1171 if (unlikely(details
))
1173 if (pte_file(ptent
)) {
1174 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
)))
1175 print_bad_pte(vma
, addr
, ptent
, NULL
);
1177 swp_entry_t entry
= pte_to_swp_entry(ptent
);
1179 if (!non_swap_entry(entry
))
1181 else if (is_migration_entry(entry
)) {
1184 page
= migration_entry_to_page(entry
);
1187 rss
[MM_ANONPAGES
]--;
1189 rss
[MM_FILEPAGES
]--;
1191 if (unlikely(!free_swap_and_cache(entry
)))
1192 print_bad_pte(vma
, addr
, ptent
, NULL
);
1194 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1195 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1197 add_mm_rss_vec(mm
, rss
);
1198 arch_leave_lazy_mmu_mode();
1199 pte_unmap_unlock(start_pte
, ptl
);
1202 * mmu_gather ran out of room to batch pages, we break out of
1203 * the PTE lock to avoid doing the potential expensive TLB invalidate
1204 * and page-free while holding it.
1209 #ifdef HAVE_GENERIC_MMU_GATHER
1210 tlb
->start
= range_start
;
1223 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1224 struct vm_area_struct
*vma
, pud_t
*pud
,
1225 unsigned long addr
, unsigned long end
,
1226 struct zap_details
*details
)
1231 pmd
= pmd_offset(pud
, addr
);
1233 next
= pmd_addr_end(addr
, end
);
1234 if (pmd_trans_huge(*pmd
)) {
1235 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1236 #ifdef CONFIG_DEBUG_VM
1237 if (!rwsem_is_locked(&tlb
->mm
->mmap_sem
)) {
1238 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1239 __func__
, addr
, end
,
1245 split_huge_page_pmd(vma
, addr
, pmd
);
1246 } else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1251 * Here there can be other concurrent MADV_DONTNEED or
1252 * trans huge page faults running, and if the pmd is
1253 * none or trans huge it can change under us. This is
1254 * because MADV_DONTNEED holds the mmap_sem in read
1257 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1259 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1262 } while (pmd
++, addr
= next
, addr
!= end
);
1267 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1268 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1269 unsigned long addr
, unsigned long end
,
1270 struct zap_details
*details
)
1275 pud
= pud_offset(pgd
, addr
);
1277 next
= pud_addr_end(addr
, end
);
1278 if (pud_none_or_clear_bad(pud
))
1280 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1281 } while (pud
++, addr
= next
, addr
!= end
);
1286 static void unmap_page_range(struct mmu_gather
*tlb
,
1287 struct vm_area_struct
*vma
,
1288 unsigned long addr
, unsigned long end
,
1289 struct zap_details
*details
)
1294 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
1297 BUG_ON(addr
>= end
);
1298 mem_cgroup_uncharge_start();
1299 tlb_start_vma(tlb
, vma
);
1300 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1302 next
= pgd_addr_end(addr
, end
);
1303 if (pgd_none_or_clear_bad(pgd
))
1305 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
1306 } while (pgd
++, addr
= next
, addr
!= end
);
1307 tlb_end_vma(tlb
, vma
);
1308 mem_cgroup_uncharge_end();
1312 static void unmap_single_vma(struct mmu_gather
*tlb
,
1313 struct vm_area_struct
*vma
, unsigned long start_addr
,
1314 unsigned long end_addr
,
1315 struct zap_details
*details
)
1317 unsigned long start
= max(vma
->vm_start
, start_addr
);
1320 if (start
>= vma
->vm_end
)
1322 end
= min(vma
->vm_end
, end_addr
);
1323 if (end
<= vma
->vm_start
)
1327 uprobe_munmap(vma
, start
, end
);
1329 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1330 untrack_pfn(vma
, 0, 0);
1333 if (unlikely(is_vm_hugetlb_page(vma
))) {
1335 * It is undesirable to test vma->vm_file as it
1336 * should be non-null for valid hugetlb area.
1337 * However, vm_file will be NULL in the error
1338 * cleanup path of do_mmap_pgoff. When
1339 * hugetlbfs ->mmap method fails,
1340 * do_mmap_pgoff() nullifies vma->vm_file
1341 * before calling this function to clean up.
1342 * Since no pte has actually been setup, it is
1343 * safe to do nothing in this case.
1346 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
1347 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1348 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
1351 unmap_page_range(tlb
, vma
, start
, end
, details
);
1356 * unmap_vmas - unmap a range of memory covered by a list of vma's
1357 * @tlb: address of the caller's struct mmu_gather
1358 * @vma: the starting vma
1359 * @start_addr: virtual address at which to start unmapping
1360 * @end_addr: virtual address at which to end unmapping
1362 * Unmap all pages in the vma list.
1364 * Only addresses between `start' and `end' will be unmapped.
1366 * The VMA list must be sorted in ascending virtual address order.
1368 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1369 * range after unmap_vmas() returns. So the only responsibility here is to
1370 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1371 * drops the lock and schedules.
1373 void unmap_vmas(struct mmu_gather
*tlb
,
1374 struct vm_area_struct
*vma
, unsigned long start_addr
,
1375 unsigned long end_addr
)
1377 struct mm_struct
*mm
= vma
->vm_mm
;
1379 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1380 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1381 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1382 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1386 * zap_page_range - remove user pages in a given range
1387 * @vma: vm_area_struct holding the applicable pages
1388 * @start: starting address of pages to zap
1389 * @size: number of bytes to zap
1390 * @details: details of nonlinear truncation or shared cache invalidation
1392 * Caller must protect the VMA list
1394 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1395 unsigned long size
, struct zap_details
*details
)
1397 struct mm_struct
*mm
= vma
->vm_mm
;
1398 struct mmu_gather tlb
;
1399 unsigned long end
= start
+ size
;
1402 tlb_gather_mmu(&tlb
, mm
, 0);
1403 update_hiwater_rss(mm
);
1404 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1405 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
)
1406 unmap_single_vma(&tlb
, vma
, start
, end
, details
);
1407 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1408 tlb_finish_mmu(&tlb
, start
, end
);
1412 * zap_page_range_single - remove user pages in a given range
1413 * @vma: vm_area_struct holding the applicable pages
1414 * @address: starting address of pages to zap
1415 * @size: number of bytes to zap
1416 * @details: details of nonlinear truncation or shared cache invalidation
1418 * The range must fit into one VMA.
1420 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1421 unsigned long size
, struct zap_details
*details
)
1423 struct mm_struct
*mm
= vma
->vm_mm
;
1424 struct mmu_gather tlb
;
1425 unsigned long end
= address
+ size
;
1428 tlb_gather_mmu(&tlb
, mm
, 0);
1429 update_hiwater_rss(mm
);
1430 mmu_notifier_invalidate_range_start(mm
, address
, end
);
1431 unmap_single_vma(&tlb
, vma
, address
, end
, details
);
1432 mmu_notifier_invalidate_range_end(mm
, address
, end
);
1433 tlb_finish_mmu(&tlb
, address
, end
);
1437 * zap_vma_ptes - remove ptes mapping the vma
1438 * @vma: vm_area_struct holding ptes to be zapped
1439 * @address: starting address of pages to zap
1440 * @size: number of bytes to zap
1442 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1444 * The entire address range must be fully contained within the vma.
1446 * Returns 0 if successful.
1448 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1451 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1452 !(vma
->vm_flags
& VM_PFNMAP
))
1454 zap_page_range_single(vma
, address
, size
, NULL
);
1457 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1460 * follow_page_mask - look up a page descriptor from a user-virtual address
1461 * @vma: vm_area_struct mapping @address
1462 * @address: virtual address to look up
1463 * @flags: flags modifying lookup behaviour
1464 * @page_mask: on output, *page_mask is set according to the size of the page
1466 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1468 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1469 * an error pointer if there is a mapping to something not represented
1470 * by a page descriptor (see also vm_normal_page()).
1472 struct page
*follow_page_mask(struct vm_area_struct
*vma
,
1473 unsigned long address
, unsigned int flags
,
1474 unsigned int *page_mask
)
1482 struct mm_struct
*mm
= vma
->vm_mm
;
1486 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
1487 if (!IS_ERR(page
)) {
1488 BUG_ON(flags
& FOLL_GET
);
1493 pgd
= pgd_offset(mm
, address
);
1494 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
1497 pud
= pud_offset(pgd
, address
);
1500 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
1501 BUG_ON(flags
& FOLL_GET
);
1502 page
= follow_huge_pud(mm
, address
, pud
, flags
& FOLL_WRITE
);
1505 if (unlikely(pud_bad(*pud
)))
1508 pmd
= pmd_offset(pud
, address
);
1511 if (pmd_huge(*pmd
) && vma
->vm_flags
& VM_HUGETLB
) {
1512 BUG_ON(flags
& FOLL_GET
);
1513 page
= follow_huge_pmd(mm
, address
, pmd
, flags
& FOLL_WRITE
);
1516 if ((flags
& FOLL_NUMA
) && pmd_numa(*pmd
))
1518 if (pmd_trans_huge(*pmd
)) {
1519 if (flags
& FOLL_SPLIT
) {
1520 split_huge_page_pmd(vma
, address
, pmd
);
1521 goto split_fallthrough
;
1523 spin_lock(&mm
->page_table_lock
);
1524 if (likely(pmd_trans_huge(*pmd
))) {
1525 if (unlikely(pmd_trans_splitting(*pmd
))) {
1526 spin_unlock(&mm
->page_table_lock
);
1527 wait_split_huge_page(vma
->anon_vma
, pmd
);
1529 page
= follow_trans_huge_pmd(vma
, address
,
1531 spin_unlock(&mm
->page_table_lock
);
1532 *page_mask
= HPAGE_PMD_NR
- 1;
1536 spin_unlock(&mm
->page_table_lock
);
1540 if (unlikely(pmd_bad(*pmd
)))
1543 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1546 if (!pte_present(pte
)) {
1549 * KSM's break_ksm() relies upon recognizing a ksm page
1550 * even while it is being migrated, so for that case we
1551 * need migration_entry_wait().
1553 if (likely(!(flags
& FOLL_MIGRATION
)))
1555 if (pte_none(pte
) || pte_file(pte
))
1557 entry
= pte_to_swp_entry(pte
);
1558 if (!is_migration_entry(entry
))
1560 pte_unmap_unlock(ptep
, ptl
);
1561 migration_entry_wait(mm
, pmd
, address
);
1562 goto split_fallthrough
;
1564 if ((flags
& FOLL_NUMA
) && pte_numa(pte
))
1566 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
1569 page
= vm_normal_page(vma
, address
, pte
);
1570 if (unlikely(!page
)) {
1571 if ((flags
& FOLL_DUMP
) ||
1572 !is_zero_pfn(pte_pfn(pte
)))
1574 page
= pte_page(pte
);
1577 if (flags
& FOLL_GET
)
1578 get_page_foll(page
);
1579 if (flags
& FOLL_TOUCH
) {
1580 if ((flags
& FOLL_WRITE
) &&
1581 !pte_dirty(pte
) && !PageDirty(page
))
1582 set_page_dirty(page
);
1584 * pte_mkyoung() would be more correct here, but atomic care
1585 * is needed to avoid losing the dirty bit: it is easier to use
1586 * mark_page_accessed().
1588 mark_page_accessed(page
);
1590 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1592 * The preliminary mapping check is mainly to avoid the
1593 * pointless overhead of lock_page on the ZERO_PAGE
1594 * which might bounce very badly if there is contention.
1596 * If the page is already locked, we don't need to
1597 * handle it now - vmscan will handle it later if and
1598 * when it attempts to reclaim the page.
1600 if (page
->mapping
&& trylock_page(page
)) {
1601 lru_add_drain(); /* push cached pages to LRU */
1603 * Because we lock page here, and migration is
1604 * blocked by the pte's page reference, and we
1605 * know the page is still mapped, we don't even
1606 * need to check for file-cache page truncation.
1608 mlock_vma_page(page
);
1613 pte_unmap_unlock(ptep
, ptl
);
1618 pte_unmap_unlock(ptep
, ptl
);
1619 return ERR_PTR(-EFAULT
);
1622 pte_unmap_unlock(ptep
, ptl
);
1628 * When core dumping an enormous anonymous area that nobody
1629 * has touched so far, we don't want to allocate unnecessary pages or
1630 * page tables. Return error instead of NULL to skip handle_mm_fault,
1631 * then get_dump_page() will return NULL to leave a hole in the dump.
1632 * But we can only make this optimization where a hole would surely
1633 * be zero-filled if handle_mm_fault() actually did handle it.
1635 if ((flags
& FOLL_DUMP
) &&
1636 (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
1637 return ERR_PTR(-EFAULT
);
1641 static inline int stack_guard_page(struct vm_area_struct
*vma
, unsigned long addr
)
1643 return stack_guard_page_start(vma
, addr
) ||
1644 stack_guard_page_end(vma
, addr
+PAGE_SIZE
);
1648 * __get_user_pages() - pin user pages in memory
1649 * @tsk: task_struct of target task
1650 * @mm: mm_struct of target mm
1651 * @start: starting user address
1652 * @nr_pages: number of pages from start to pin
1653 * @gup_flags: flags modifying pin behaviour
1654 * @pages: array that receives pointers to the pages pinned.
1655 * Should be at least nr_pages long. Or NULL, if caller
1656 * only intends to ensure the pages are faulted in.
1657 * @vmas: array of pointers to vmas corresponding to each page.
1658 * Or NULL if the caller does not require them.
1659 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1661 * Returns number of pages pinned. This may be fewer than the number
1662 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1663 * were pinned, returns -errno. Each page returned must be released
1664 * with a put_page() call when it is finished with. vmas will only
1665 * remain valid while mmap_sem is held.
1667 * Must be called with mmap_sem held for read or write.
1669 * __get_user_pages walks a process's page tables and takes a reference to
1670 * each struct page that each user address corresponds to at a given
1671 * instant. That is, it takes the page that would be accessed if a user
1672 * thread accesses the given user virtual address at that instant.
1674 * This does not guarantee that the page exists in the user mappings when
1675 * __get_user_pages returns, and there may even be a completely different
1676 * page there in some cases (eg. if mmapped pagecache has been invalidated
1677 * and subsequently re faulted). However it does guarantee that the page
1678 * won't be freed completely. And mostly callers simply care that the page
1679 * contains data that was valid *at some point in time*. Typically, an IO
1680 * or similar operation cannot guarantee anything stronger anyway because
1681 * locks can't be held over the syscall boundary.
1683 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1684 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1685 * appropriate) must be called after the page is finished with, and
1686 * before put_page is called.
1688 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1689 * or mmap_sem contention, and if waiting is needed to pin all pages,
1690 * *@nonblocking will be set to 0.
1692 * In most cases, get_user_pages or get_user_pages_fast should be used
1693 * instead of __get_user_pages. __get_user_pages should be used only if
1694 * you need some special @gup_flags.
1696 long __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1697 unsigned long start
, unsigned long nr_pages
,
1698 unsigned int gup_flags
, struct page
**pages
,
1699 struct vm_area_struct
**vmas
, int *nonblocking
)
1702 unsigned long vm_flags
;
1703 unsigned int page_mask
;
1708 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
1711 * Require read or write permissions.
1712 * If FOLL_FORCE is set, we only require the "MAY" flags.
1714 vm_flags
= (gup_flags
& FOLL_WRITE
) ?
1715 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1716 vm_flags
&= (gup_flags
& FOLL_FORCE
) ?
1717 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1720 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1721 * would be called on PROT_NONE ranges. We must never invoke
1722 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1723 * page faults would unprotect the PROT_NONE ranges if
1724 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1725 * bitflag. So to avoid that, don't set FOLL_NUMA if
1726 * FOLL_FORCE is set.
1728 if (!(gup_flags
& FOLL_FORCE
))
1729 gup_flags
|= FOLL_NUMA
;
1734 struct vm_area_struct
*vma
;
1736 vma
= find_extend_vma(mm
, start
);
1737 if (!vma
&& in_gate_area(mm
, start
)) {
1738 unsigned long pg
= start
& PAGE_MASK
;
1744 /* user gate pages are read-only */
1745 if (gup_flags
& FOLL_WRITE
)
1746 return i
? : -EFAULT
;
1748 pgd
= pgd_offset_k(pg
);
1750 pgd
= pgd_offset_gate(mm
, pg
);
1751 BUG_ON(pgd_none(*pgd
));
1752 pud
= pud_offset(pgd
, pg
);
1753 BUG_ON(pud_none(*pud
));
1754 pmd
= pmd_offset(pud
, pg
);
1756 return i
? : -EFAULT
;
1757 VM_BUG_ON(pmd_trans_huge(*pmd
));
1758 pte
= pte_offset_map(pmd
, pg
);
1759 if (pte_none(*pte
)) {
1761 return i
? : -EFAULT
;
1763 vma
= get_gate_vma(mm
);
1767 page
= vm_normal_page(vma
, start
, *pte
);
1769 if (!(gup_flags
& FOLL_DUMP
) &&
1770 is_zero_pfn(pte_pfn(*pte
)))
1771 page
= pte_page(*pte
);
1774 return i
? : -EFAULT
;
1786 (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1787 !(vm_flags
& vma
->vm_flags
))
1788 return i
? : -EFAULT
;
1790 if (is_vm_hugetlb_page(vma
)) {
1791 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1792 &start
, &nr_pages
, i
, gup_flags
);
1798 unsigned int foll_flags
= gup_flags
;
1799 unsigned int page_increm
;
1802 * If we have a pending SIGKILL, don't keep faulting
1803 * pages and potentially allocating memory.
1805 if (unlikely(fatal_signal_pending(current
)))
1806 return i
? i
: -ERESTARTSYS
;
1809 while (!(page
= follow_page_mask(vma
, start
,
1810 foll_flags
, &page_mask
))) {
1812 unsigned int fault_flags
= 0;
1814 /* For mlock, just skip the stack guard page. */
1815 if (foll_flags
& FOLL_MLOCK
) {
1816 if (stack_guard_page(vma
, start
))
1819 if (foll_flags
& FOLL_WRITE
)
1820 fault_flags
|= FAULT_FLAG_WRITE
;
1822 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
1823 if (foll_flags
& FOLL_NOWAIT
)
1824 fault_flags
|= (FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
);
1826 ret
= handle_mm_fault(mm
, vma
, start
,
1829 if (ret
& VM_FAULT_ERROR
) {
1830 if (ret
& VM_FAULT_OOM
)
1831 return i
? i
: -ENOMEM
;
1832 if (ret
& (VM_FAULT_HWPOISON
|
1833 VM_FAULT_HWPOISON_LARGE
)) {
1836 else if (gup_flags
& FOLL_HWPOISON
)
1841 if (ret
& VM_FAULT_SIGBUS
)
1842 return i
? i
: -EFAULT
;
1847 if (ret
& VM_FAULT_MAJOR
)
1853 if (ret
& VM_FAULT_RETRY
) {
1860 * The VM_FAULT_WRITE bit tells us that
1861 * do_wp_page has broken COW when necessary,
1862 * even if maybe_mkwrite decided not to set
1863 * pte_write. We can thus safely do subsequent
1864 * page lookups as if they were reads. But only
1865 * do so when looping for pte_write is futile:
1866 * in some cases userspace may also be wanting
1867 * to write to the gotten user page, which a
1868 * read fault here might prevent (a readonly
1869 * page might get reCOWed by userspace write).
1871 if ((ret
& VM_FAULT_WRITE
) &&
1872 !(vma
->vm_flags
& VM_WRITE
))
1873 foll_flags
&= ~FOLL_WRITE
;
1878 return i
? i
: PTR_ERR(page
);
1882 flush_anon_page(vma
, page
, start
);
1883 flush_dcache_page(page
);
1891 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & page_mask
);
1892 if (page_increm
> nr_pages
)
1893 page_increm
= nr_pages
;
1895 start
+= page_increm
* PAGE_SIZE
;
1896 nr_pages
-= page_increm
;
1897 } while (nr_pages
&& start
< vma
->vm_end
);
1901 EXPORT_SYMBOL(__get_user_pages
);
1904 * fixup_user_fault() - manually resolve a user page fault
1905 * @tsk: the task_struct to use for page fault accounting, or
1906 * NULL if faults are not to be recorded.
1907 * @mm: mm_struct of target mm
1908 * @address: user address
1909 * @fault_flags:flags to pass down to handle_mm_fault()
1911 * This is meant to be called in the specific scenario where for locking reasons
1912 * we try to access user memory in atomic context (within a pagefault_disable()
1913 * section), this returns -EFAULT, and we want to resolve the user fault before
1916 * Typically this is meant to be used by the futex code.
1918 * The main difference with get_user_pages() is that this function will
1919 * unconditionally call handle_mm_fault() which will in turn perform all the
1920 * necessary SW fixup of the dirty and young bits in the PTE, while
1921 * handle_mm_fault() only guarantees to update these in the struct page.
1923 * This is important for some architectures where those bits also gate the
1924 * access permission to the page because they are maintained in software. On
1925 * such architectures, gup() will not be enough to make a subsequent access
1928 * This should be called with the mm_sem held for read.
1930 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
1931 unsigned long address
, unsigned int fault_flags
)
1933 struct vm_area_struct
*vma
;
1936 vma
= find_extend_vma(mm
, address
);
1937 if (!vma
|| address
< vma
->vm_start
)
1940 ret
= handle_mm_fault(mm
, vma
, address
, fault_flags
);
1941 if (ret
& VM_FAULT_ERROR
) {
1942 if (ret
& VM_FAULT_OOM
)
1944 if (ret
& (VM_FAULT_HWPOISON
| VM_FAULT_HWPOISON_LARGE
))
1946 if (ret
& VM_FAULT_SIGBUS
)
1951 if (ret
& VM_FAULT_MAJOR
)
1960 * get_user_pages() - pin user pages in memory
1961 * @tsk: the task_struct to use for page fault accounting, or
1962 * NULL if faults are not to be recorded.
1963 * @mm: mm_struct of target mm
1964 * @start: starting user address
1965 * @nr_pages: number of pages from start to pin
1966 * @write: whether pages will be written to by the caller
1967 * @force: whether to force write access even if user mapping is
1968 * readonly. This will result in the page being COWed even
1969 * in MAP_SHARED mappings. You do not want this.
1970 * @pages: array that receives pointers to the pages pinned.
1971 * Should be at least nr_pages long. Or NULL, if caller
1972 * only intends to ensure the pages are faulted in.
1973 * @vmas: array of pointers to vmas corresponding to each page.
1974 * Or NULL if the caller does not require them.
1976 * Returns number of pages pinned. This may be fewer than the number
1977 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1978 * were pinned, returns -errno. Each page returned must be released
1979 * with a put_page() call when it is finished with. vmas will only
1980 * remain valid while mmap_sem is held.
1982 * Must be called with mmap_sem held for read or write.
1984 * get_user_pages walks a process's page tables and takes a reference to
1985 * each struct page that each user address corresponds to at a given
1986 * instant. That is, it takes the page that would be accessed if a user
1987 * thread accesses the given user virtual address at that instant.
1989 * This does not guarantee that the page exists in the user mappings when
1990 * get_user_pages returns, and there may even be a completely different
1991 * page there in some cases (eg. if mmapped pagecache has been invalidated
1992 * and subsequently re faulted). However it does guarantee that the page
1993 * won't be freed completely. And mostly callers simply care that the page
1994 * contains data that was valid *at some point in time*. Typically, an IO
1995 * or similar operation cannot guarantee anything stronger anyway because
1996 * locks can't be held over the syscall boundary.
1998 * If write=0, the page must not be written to. If the page is written to,
1999 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
2000 * after the page is finished with, and before put_page is called.
2002 * get_user_pages is typically used for fewer-copy IO operations, to get a
2003 * handle on the memory by some means other than accesses via the user virtual
2004 * addresses. The pages may be submitted for DMA to devices or accessed via
2005 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2006 * use the correct cache flushing APIs.
2008 * See also get_user_pages_fast, for performance critical applications.
2010 long get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
2011 unsigned long start
, unsigned long nr_pages
, int write
,
2012 int force
, struct page
**pages
, struct vm_area_struct
**vmas
)
2014 int flags
= FOLL_TOUCH
;
2019 flags
|= FOLL_WRITE
;
2021 flags
|= FOLL_FORCE
;
2023 return __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
, vmas
,
2026 EXPORT_SYMBOL(get_user_pages
);
2029 * get_dump_page() - pin user page in memory while writing it to core dump
2030 * @addr: user address
2032 * Returns struct page pointer of user page pinned for dump,
2033 * to be freed afterwards by page_cache_release() or put_page().
2035 * Returns NULL on any kind of failure - a hole must then be inserted into
2036 * the corefile, to preserve alignment with its headers; and also returns
2037 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2038 * allowing a hole to be left in the corefile to save diskspace.
2040 * Called without mmap_sem, but after all other threads have been killed.
2042 #ifdef CONFIG_ELF_CORE
2043 struct page
*get_dump_page(unsigned long addr
)
2045 struct vm_area_struct
*vma
;
2048 if (__get_user_pages(current
, current
->mm
, addr
, 1,
2049 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
2052 flush_cache_page(vma
, addr
, page_to_pfn(page
));
2055 #endif /* CONFIG_ELF_CORE */
2057 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
2060 pgd_t
* pgd
= pgd_offset(mm
, addr
);
2061 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
2063 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
2065 VM_BUG_ON(pmd_trans_huge(*pmd
));
2066 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
2073 * This is the old fallback for page remapping.
2075 * For historical reasons, it only allows reserved pages. Only
2076 * old drivers should use this, and they needed to mark their
2077 * pages reserved for the old functions anyway.
2079 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2080 struct page
*page
, pgprot_t prot
)
2082 struct mm_struct
*mm
= vma
->vm_mm
;
2091 flush_dcache_page(page
);
2092 pte
= get_locked_pte(mm
, addr
, &ptl
);
2096 if (!pte_none(*pte
))
2099 /* Ok, finally just insert the thing.. */
2101 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
2102 page_add_file_rmap(page
);
2103 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
2106 pte_unmap_unlock(pte
, ptl
);
2109 pte_unmap_unlock(pte
, ptl
);
2115 * vm_insert_page - insert single page into user vma
2116 * @vma: user vma to map to
2117 * @addr: target user address of this page
2118 * @page: source kernel page
2120 * This allows drivers to insert individual pages they've allocated
2123 * The page has to be a nice clean _individual_ kernel allocation.
2124 * If you allocate a compound page, you need to have marked it as
2125 * such (__GFP_COMP), or manually just split the page up yourself
2126 * (see split_page()).
2128 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2129 * took an arbitrary page protection parameter. This doesn't allow
2130 * that. Your vma protection will have to be set up correctly, which
2131 * means that if you want a shared writable mapping, you'd better
2132 * ask for a shared writable mapping!
2134 * The page does not need to be reserved.
2136 * Usually this function is called from f_op->mmap() handler
2137 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2138 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2139 * function from other places, for example from page-fault handler.
2141 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2144 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2146 if (!page_count(page
))
2148 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
2149 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
2150 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
2151 vma
->vm_flags
|= VM_MIXEDMAP
;
2153 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2155 EXPORT_SYMBOL(vm_insert_page
);
2157 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2158 unsigned long pfn
, pgprot_t prot
)
2160 struct mm_struct
*mm
= vma
->vm_mm
;
2166 pte
= get_locked_pte(mm
, addr
, &ptl
);
2170 if (!pte_none(*pte
))
2173 /* Ok, finally just insert the thing.. */
2174 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
2175 set_pte_at(mm
, addr
, pte
, entry
);
2176 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
2180 pte_unmap_unlock(pte
, ptl
);
2186 * vm_insert_pfn - insert single pfn into user vma
2187 * @vma: user vma to map to
2188 * @addr: target user address of this page
2189 * @pfn: source kernel pfn
2191 * Similar to vm_insert_page, this allows drivers to insert individual pages
2192 * they've allocated into a user vma. Same comments apply.
2194 * This function should only be called from a vm_ops->fault handler, and
2195 * in that case the handler should return NULL.
2197 * vma cannot be a COW mapping.
2199 * As this is called only for pages that do not currently exist, we
2200 * do not need to flush old virtual caches or the TLB.
2202 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2206 pgprot_t pgprot
= vma
->vm_page_prot
;
2208 * Technically, architectures with pte_special can avoid all these
2209 * restrictions (same for remap_pfn_range). However we would like
2210 * consistency in testing and feature parity among all, so we should
2211 * try to keep these invariants in place for everybody.
2213 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
2214 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
2215 (VM_PFNMAP
|VM_MIXEDMAP
));
2216 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
2217 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
2219 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2221 if (track_pfn_insert(vma
, &pgprot
, pfn
))
2224 ret
= insert_pfn(vma
, addr
, pfn
, pgprot
);
2228 EXPORT_SYMBOL(vm_insert_pfn
);
2230 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2233 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
2235 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2239 * If we don't have pte special, then we have to use the pfn_valid()
2240 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2241 * refcount the page if pfn_valid is true (hence insert_page rather
2242 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2243 * without pte special, it would there be refcounted as a normal page.
2245 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
2248 page
= pfn_to_page(pfn
);
2249 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2251 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
2253 EXPORT_SYMBOL(vm_insert_mixed
);
2256 * maps a range of physical memory into the requested pages. the old
2257 * mappings are removed. any references to nonexistent pages results
2258 * in null mappings (currently treated as "copy-on-access")
2260 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2261 unsigned long addr
, unsigned long end
,
2262 unsigned long pfn
, pgprot_t prot
)
2267 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2270 arch_enter_lazy_mmu_mode();
2272 BUG_ON(!pte_none(*pte
));
2273 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2275 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2276 arch_leave_lazy_mmu_mode();
2277 pte_unmap_unlock(pte
- 1, ptl
);
2281 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2282 unsigned long addr
, unsigned long end
,
2283 unsigned long pfn
, pgprot_t prot
)
2288 pfn
-= addr
>> PAGE_SHIFT
;
2289 pmd
= pmd_alloc(mm
, pud
, addr
);
2292 VM_BUG_ON(pmd_trans_huge(*pmd
));
2294 next
= pmd_addr_end(addr
, end
);
2295 if (remap_pte_range(mm
, pmd
, addr
, next
,
2296 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
2298 } while (pmd
++, addr
= next
, addr
!= end
);
2302 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2303 unsigned long addr
, unsigned long end
,
2304 unsigned long pfn
, pgprot_t prot
)
2309 pfn
-= addr
>> PAGE_SHIFT
;
2310 pud
= pud_alloc(mm
, pgd
, addr
);
2314 next
= pud_addr_end(addr
, end
);
2315 if (remap_pmd_range(mm
, pud
, addr
, next
,
2316 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
2318 } while (pud
++, addr
= next
, addr
!= end
);
2323 * remap_pfn_range - remap kernel memory to userspace
2324 * @vma: user vma to map to
2325 * @addr: target user address to start at
2326 * @pfn: physical address of kernel memory
2327 * @size: size of map area
2328 * @prot: page protection flags for this mapping
2330 * Note: this is only safe if the mm semaphore is held when called.
2332 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2333 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2337 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2338 struct mm_struct
*mm
= vma
->vm_mm
;
2342 * Physically remapped pages are special. Tell the
2343 * rest of the world about it:
2344 * VM_IO tells people not to look at these pages
2345 * (accesses can have side effects).
2346 * VM_PFNMAP tells the core MM that the base pages are just
2347 * raw PFN mappings, and do not have a "struct page" associated
2350 * Disable vma merging and expanding with mremap().
2352 * Omit vma from core dump, even when VM_IO turned off.
2354 * There's a horrible special case to handle copy-on-write
2355 * behaviour that some programs depend on. We mark the "original"
2356 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2357 * See vm_normal_page() for details.
2359 if (is_cow_mapping(vma
->vm_flags
)) {
2360 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2362 vma
->vm_pgoff
= pfn
;
2365 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
2369 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
2371 BUG_ON(addr
>= end
);
2372 pfn
-= addr
>> PAGE_SHIFT
;
2373 pgd
= pgd_offset(mm
, addr
);
2374 flush_cache_range(vma
, addr
, end
);
2376 next
= pgd_addr_end(addr
, end
);
2377 err
= remap_pud_range(mm
, pgd
, addr
, next
,
2378 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2381 } while (pgd
++, addr
= next
, addr
!= end
);
2384 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
));
2388 EXPORT_SYMBOL(remap_pfn_range
);
2391 * vm_iomap_memory - remap memory to userspace
2392 * @vma: user vma to map to
2393 * @start: start of area
2394 * @len: size of area
2396 * This is a simplified io_remap_pfn_range() for common driver use. The
2397 * driver just needs to give us the physical memory range to be mapped,
2398 * we'll figure out the rest from the vma information.
2400 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2401 * whatever write-combining details or similar.
2403 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2405 unsigned long vm_len
, pfn
, pages
;
2407 /* Check that the physical memory area passed in looks valid */
2408 if (start
+ len
< start
)
2411 * You *really* shouldn't map things that aren't page-aligned,
2412 * but we've historically allowed it because IO memory might
2413 * just have smaller alignment.
2415 len
+= start
& ~PAGE_MASK
;
2416 pfn
= start
>> PAGE_SHIFT
;
2417 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2418 if (pfn
+ pages
< pfn
)
2421 /* We start the mapping 'vm_pgoff' pages into the area */
2422 if (vma
->vm_pgoff
> pages
)
2424 pfn
+= vma
->vm_pgoff
;
2425 pages
-= vma
->vm_pgoff
;
2427 /* Can we fit all of the mapping? */
2428 vm_len
= vma
->vm_end
- vma
->vm_start
;
2429 if (vm_len
>> PAGE_SHIFT
> pages
)
2432 /* Ok, let it rip */
2433 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2435 EXPORT_SYMBOL(vm_iomap_memory
);
2437 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2438 unsigned long addr
, unsigned long end
,
2439 pte_fn_t fn
, void *data
)
2444 spinlock_t
*uninitialized_var(ptl
);
2446 pte
= (mm
== &init_mm
) ?
2447 pte_alloc_kernel(pmd
, addr
) :
2448 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2452 BUG_ON(pmd_huge(*pmd
));
2454 arch_enter_lazy_mmu_mode();
2456 token
= pmd_pgtable(*pmd
);
2459 err
= fn(pte
++, token
, addr
, data
);
2462 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2464 arch_leave_lazy_mmu_mode();
2467 pte_unmap_unlock(pte
-1, ptl
);
2471 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2472 unsigned long addr
, unsigned long end
,
2473 pte_fn_t fn
, void *data
)
2479 BUG_ON(pud_huge(*pud
));
2481 pmd
= pmd_alloc(mm
, pud
, addr
);
2485 next
= pmd_addr_end(addr
, end
);
2486 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
2489 } while (pmd
++, addr
= next
, addr
!= end
);
2493 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2494 unsigned long addr
, unsigned long end
,
2495 pte_fn_t fn
, void *data
)
2501 pud
= pud_alloc(mm
, pgd
, addr
);
2505 next
= pud_addr_end(addr
, end
);
2506 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
2509 } while (pud
++, addr
= next
, addr
!= end
);
2514 * Scan a region of virtual memory, filling in page tables as necessary
2515 * and calling a provided function on each leaf page table.
2517 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2518 unsigned long size
, pte_fn_t fn
, void *data
)
2522 unsigned long end
= addr
+ size
;
2525 BUG_ON(addr
>= end
);
2526 pgd
= pgd_offset(mm
, addr
);
2528 next
= pgd_addr_end(addr
, end
);
2529 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
2532 } while (pgd
++, addr
= next
, addr
!= end
);
2536 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2539 * handle_pte_fault chooses page fault handler according to an entry
2540 * which was read non-atomically. Before making any commitment, on
2541 * those architectures or configurations (e.g. i386 with PAE) which
2542 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2543 * must check under lock before unmapping the pte and proceeding
2544 * (but do_wp_page is only called after already making such a check;
2545 * and do_anonymous_page can safely check later on).
2547 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2548 pte_t
*page_table
, pte_t orig_pte
)
2551 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2552 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2553 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2555 same
= pte_same(*page_table
, orig_pte
);
2559 pte_unmap(page_table
);
2563 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
2566 * If the source page was a PFN mapping, we don't have
2567 * a "struct page" for it. We do a best-effort copy by
2568 * just copying from the original user address. If that
2569 * fails, we just zero-fill it. Live with it.
2571 if (unlikely(!src
)) {
2572 void *kaddr
= kmap_atomic(dst
);
2573 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
2576 * This really shouldn't fail, because the page is there
2577 * in the page tables. But it might just be unreadable,
2578 * in which case we just give up and fill the result with
2581 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
2583 kunmap_atomic(kaddr
);
2584 flush_dcache_page(dst
);
2586 copy_user_highpage(dst
, src
, va
, vma
);
2590 * This routine handles present pages, when users try to write
2591 * to a shared page. It is done by copying the page to a new address
2592 * and decrementing the shared-page counter for the old page.
2594 * Note that this routine assumes that the protection checks have been
2595 * done by the caller (the low-level page fault routine in most cases).
2596 * Thus we can safely just mark it writable once we've done any necessary
2599 * We also mark the page dirty at this point even though the page will
2600 * change only once the write actually happens. This avoids a few races,
2601 * and potentially makes it more efficient.
2603 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2604 * but allow concurrent faults), with pte both mapped and locked.
2605 * We return with mmap_sem still held, but pte unmapped and unlocked.
2607 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2608 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2609 spinlock_t
*ptl
, pte_t orig_pte
)
2612 struct page
*old_page
, *new_page
= NULL
;
2615 int page_mkwrite
= 0;
2616 struct page
*dirty_page
= NULL
;
2617 unsigned long mmun_start
= 0; /* For mmu_notifiers */
2618 unsigned long mmun_end
= 0; /* For mmu_notifiers */
2620 old_page
= vm_normal_page(vma
, address
, orig_pte
);
2623 * VM_MIXEDMAP !pfn_valid() case
2625 * We should not cow pages in a shared writeable mapping.
2626 * Just mark the pages writable as we can't do any dirty
2627 * accounting on raw pfn maps.
2629 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2630 (VM_WRITE
|VM_SHARED
))
2636 * Take out anonymous pages first, anonymous shared vmas are
2637 * not dirty accountable.
2639 if (PageAnon(old_page
) && !PageKsm(old_page
)) {
2640 if (!trylock_page(old_page
)) {
2641 page_cache_get(old_page
);
2642 pte_unmap_unlock(page_table
, ptl
);
2643 lock_page(old_page
);
2644 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2646 if (!pte_same(*page_table
, orig_pte
)) {
2647 unlock_page(old_page
);
2650 page_cache_release(old_page
);
2652 if (reuse_swap_page(old_page
)) {
2654 * The page is all ours. Move it to our anon_vma so
2655 * the rmap code will not search our parent or siblings.
2656 * Protected against the rmap code by the page lock.
2658 page_move_anon_rmap(old_page
, vma
, address
);
2659 unlock_page(old_page
);
2662 unlock_page(old_page
);
2663 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2664 (VM_WRITE
|VM_SHARED
))) {
2666 * Only catch write-faults on shared writable pages,
2667 * read-only shared pages can get COWed by
2668 * get_user_pages(.write=1, .force=1).
2670 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2671 struct vm_fault vmf
;
2674 vmf
.virtual_address
= (void __user
*)(address
&
2676 vmf
.pgoff
= old_page
->index
;
2677 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2678 vmf
.page
= old_page
;
2681 * Notify the address space that the page is about to
2682 * become writable so that it can prohibit this or wait
2683 * for the page to get into an appropriate state.
2685 * We do this without the lock held, so that it can
2686 * sleep if it needs to.
2688 page_cache_get(old_page
);
2689 pte_unmap_unlock(page_table
, ptl
);
2691 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
2693 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
2695 goto unwritable_page
;
2697 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
2698 lock_page(old_page
);
2699 if (!old_page
->mapping
) {
2700 ret
= 0; /* retry the fault */
2701 unlock_page(old_page
);
2702 goto unwritable_page
;
2705 VM_BUG_ON(!PageLocked(old_page
));
2708 * Since we dropped the lock we need to revalidate
2709 * the PTE as someone else may have changed it. If
2710 * they did, we just return, as we can count on the
2711 * MMU to tell us if they didn't also make it writable.
2713 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2715 if (!pte_same(*page_table
, orig_pte
)) {
2716 unlock_page(old_page
);
2722 dirty_page
= old_page
;
2723 get_page(dirty_page
);
2726 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2727 entry
= pte_mkyoung(orig_pte
);
2728 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2729 if (ptep_set_access_flags(vma
, address
, page_table
, entry
,1))
2730 update_mmu_cache(vma
, address
, page_table
);
2731 pte_unmap_unlock(page_table
, ptl
);
2732 ret
|= VM_FAULT_WRITE
;
2738 * Yes, Virginia, this is actually required to prevent a race
2739 * with clear_page_dirty_for_io() from clearing the page dirty
2740 * bit after it clear all dirty ptes, but before a racing
2741 * do_wp_page installs a dirty pte.
2743 * __do_fault is protected similarly.
2745 if (!page_mkwrite
) {
2746 wait_on_page_locked(dirty_page
);
2747 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2748 /* file_update_time outside page_lock */
2750 file_update_time(vma
->vm_file
);
2752 put_page(dirty_page
);
2754 struct address_space
*mapping
= dirty_page
->mapping
;
2756 set_page_dirty(dirty_page
);
2757 unlock_page(dirty_page
);
2758 page_cache_release(dirty_page
);
2761 * Some device drivers do not set page.mapping
2762 * but still dirty their pages
2764 balance_dirty_pages_ratelimited(mapping
);
2772 * Ok, we need to copy. Oh, well..
2774 page_cache_get(old_page
);
2776 pte_unmap_unlock(page_table
, ptl
);
2778 if (unlikely(anon_vma_prepare(vma
)))
2781 if (is_zero_pfn(pte_pfn(orig_pte
))) {
2782 new_page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2786 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2789 cow_user_page(new_page
, old_page
, address
, vma
);
2791 __SetPageUptodate(new_page
);
2793 if (mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
))
2796 mmun_start
= address
& PAGE_MASK
;
2797 mmun_end
= mmun_start
+ PAGE_SIZE
;
2798 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2801 * Re-check the pte - we dropped the lock
2803 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2804 if (likely(pte_same(*page_table
, orig_pte
))) {
2806 if (!PageAnon(old_page
)) {
2807 dec_mm_counter_fast(mm
, MM_FILEPAGES
);
2808 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2811 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2812 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2813 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2814 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2816 * Clear the pte entry and flush it first, before updating the
2817 * pte with the new entry. This will avoid a race condition
2818 * seen in the presence of one thread doing SMC and another
2821 ptep_clear_flush(vma
, address
, page_table
);
2822 page_add_new_anon_rmap(new_page
, vma
, address
);
2824 * We call the notify macro here because, when using secondary
2825 * mmu page tables (such as kvm shadow page tables), we want the
2826 * new page to be mapped directly into the secondary page table.
2828 set_pte_at_notify(mm
, address
, page_table
, entry
);
2829 update_mmu_cache(vma
, address
, page_table
);
2832 * Only after switching the pte to the new page may
2833 * we remove the mapcount here. Otherwise another
2834 * process may come and find the rmap count decremented
2835 * before the pte is switched to the new page, and
2836 * "reuse" the old page writing into it while our pte
2837 * here still points into it and can be read by other
2840 * The critical issue is to order this
2841 * page_remove_rmap with the ptp_clear_flush above.
2842 * Those stores are ordered by (if nothing else,)
2843 * the barrier present in the atomic_add_negative
2844 * in page_remove_rmap.
2846 * Then the TLB flush in ptep_clear_flush ensures that
2847 * no process can access the old page before the
2848 * decremented mapcount is visible. And the old page
2849 * cannot be reused until after the decremented
2850 * mapcount is visible. So transitively, TLBs to
2851 * old page will be flushed before it can be reused.
2853 page_remove_rmap(old_page
);
2856 /* Free the old page.. */
2857 new_page
= old_page
;
2858 ret
|= VM_FAULT_WRITE
;
2860 mem_cgroup_uncharge_page(new_page
);
2863 page_cache_release(new_page
);
2865 pte_unmap_unlock(page_table
, ptl
);
2866 if (mmun_end
> mmun_start
)
2867 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2870 * Don't let another task, with possibly unlocked vma,
2871 * keep the mlocked page.
2873 if ((ret
& VM_FAULT_WRITE
) && (vma
->vm_flags
& VM_LOCKED
)) {
2874 lock_page(old_page
); /* LRU manipulation */
2875 munlock_vma_page(old_page
);
2876 unlock_page(old_page
);
2878 page_cache_release(old_page
);
2882 page_cache_release(new_page
);
2885 page_cache_release(old_page
);
2886 return VM_FAULT_OOM
;
2889 page_cache_release(old_page
);
2893 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2894 unsigned long start_addr
, unsigned long end_addr
,
2895 struct zap_details
*details
)
2897 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2900 static inline void unmap_mapping_range_tree(struct rb_root
*root
,
2901 struct zap_details
*details
)
2903 struct vm_area_struct
*vma
;
2904 pgoff_t vba
, vea
, zba
, zea
;
2906 vma_interval_tree_foreach(vma
, root
,
2907 details
->first_index
, details
->last_index
) {
2909 vba
= vma
->vm_pgoff
;
2910 vea
= vba
+ vma_pages(vma
) - 1;
2911 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2912 zba
= details
->first_index
;
2915 zea
= details
->last_index
;
2919 unmap_mapping_range_vma(vma
,
2920 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2921 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2926 static inline void unmap_mapping_range_list(struct list_head
*head
,
2927 struct zap_details
*details
)
2929 struct vm_area_struct
*vma
;
2932 * In nonlinear VMAs there is no correspondence between virtual address
2933 * offset and file offset. So we must perform an exhaustive search
2934 * across *all* the pages in each nonlinear VMA, not just the pages
2935 * whose virtual address lies outside the file truncation point.
2937 list_for_each_entry(vma
, head
, shared
.nonlinear
) {
2938 details
->nonlinear_vma
= vma
;
2939 unmap_mapping_range_vma(vma
, vma
->vm_start
, vma
->vm_end
, details
);
2944 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2945 * @mapping: the address space containing mmaps to be unmapped.
2946 * @holebegin: byte in first page to unmap, relative to the start of
2947 * the underlying file. This will be rounded down to a PAGE_SIZE
2948 * boundary. Note that this is different from truncate_pagecache(), which
2949 * must keep the partial page. In contrast, we must get rid of
2951 * @holelen: size of prospective hole in bytes. This will be rounded
2952 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2954 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2955 * but 0 when invalidating pagecache, don't throw away private data.
2957 void unmap_mapping_range(struct address_space
*mapping
,
2958 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2960 struct zap_details details
;
2961 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2962 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2964 /* Check for overflow. */
2965 if (sizeof(holelen
) > sizeof(hlen
)) {
2967 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2968 if (holeend
& ~(long long)ULONG_MAX
)
2969 hlen
= ULONG_MAX
- hba
+ 1;
2972 details
.check_mapping
= even_cows
? NULL
: mapping
;
2973 details
.nonlinear_vma
= NULL
;
2974 details
.first_index
= hba
;
2975 details
.last_index
= hba
+ hlen
- 1;
2976 if (details
.last_index
< details
.first_index
)
2977 details
.last_index
= ULONG_MAX
;
2980 mutex_lock(&mapping
->i_mmap_mutex
);
2981 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
)))
2982 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2983 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
2984 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
2985 mutex_unlock(&mapping
->i_mmap_mutex
);
2987 EXPORT_SYMBOL(unmap_mapping_range
);
2990 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2991 * but allow concurrent faults), and pte mapped but not yet locked.
2992 * We return with mmap_sem still held, but pte unmapped and unlocked.
2994 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2995 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2996 unsigned int flags
, pte_t orig_pte
)
2999 struct page
*page
, *swapcache
;
3003 struct mem_cgroup
*ptr
;
3007 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
3010 entry
= pte_to_swp_entry(orig_pte
);
3011 if (unlikely(non_swap_entry(entry
))) {
3012 if (is_migration_entry(entry
)) {
3013 migration_entry_wait(mm
, pmd
, address
);
3014 } else if (is_hwpoison_entry(entry
)) {
3015 ret
= VM_FAULT_HWPOISON
;
3017 print_bad_pte(vma
, address
, orig_pte
, NULL
);
3018 ret
= VM_FAULT_SIGBUS
;
3022 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
3023 page
= lookup_swap_cache(entry
);
3025 page
= swapin_readahead(entry
,
3026 GFP_HIGHUSER_MOVABLE
, vma
, address
);
3029 * Back out if somebody else faulted in this pte
3030 * while we released the pte lock.
3032 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3033 if (likely(pte_same(*page_table
, orig_pte
)))
3035 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3039 /* Had to read the page from swap area: Major fault */
3040 ret
= VM_FAULT_MAJOR
;
3041 count_vm_event(PGMAJFAULT
);
3042 mem_cgroup_count_vm_event(mm
, PGMAJFAULT
);
3043 } else if (PageHWPoison(page
)) {
3045 * hwpoisoned dirty swapcache pages are kept for killing
3046 * owner processes (which may be unknown at hwpoison time)
3048 ret
= VM_FAULT_HWPOISON
;
3049 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3055 locked
= lock_page_or_retry(page
, mm
, flags
);
3057 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3059 ret
|= VM_FAULT_RETRY
;
3064 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3065 * release the swapcache from under us. The page pin, and pte_same
3066 * test below, are not enough to exclude that. Even if it is still
3067 * swapcache, we need to check that the page's swap has not changed.
3069 if (unlikely(!PageSwapCache(page
) || page_private(page
) != entry
.val
))
3072 page
= ksm_might_need_to_copy(page
, vma
, address
);
3073 if (unlikely(!page
)) {
3079 if (mem_cgroup_try_charge_swapin(mm
, page
, GFP_KERNEL
, &ptr
)) {
3085 * Back out if somebody else already faulted in this pte.
3087 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3088 if (unlikely(!pte_same(*page_table
, orig_pte
)))
3091 if (unlikely(!PageUptodate(page
))) {
3092 ret
= VM_FAULT_SIGBUS
;
3097 * The page isn't present yet, go ahead with the fault.
3099 * Be careful about the sequence of operations here.
3100 * To get its accounting right, reuse_swap_page() must be called
3101 * while the page is counted on swap but not yet in mapcount i.e.
3102 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3103 * must be called after the swap_free(), or it will never succeed.
3104 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3105 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3106 * in page->private. In this case, a record in swap_cgroup is silently
3107 * discarded at swap_free().
3110 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3111 dec_mm_counter_fast(mm
, MM_SWAPENTS
);
3112 pte
= mk_pte(page
, vma
->vm_page_prot
);
3113 if ((flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
)) {
3114 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3115 flags
&= ~FAULT_FLAG_WRITE
;
3116 ret
|= VM_FAULT_WRITE
;
3119 flush_icache_page(vma
, page
);
3120 set_pte_at(mm
, address
, page_table
, pte
);
3121 if (page
== swapcache
)
3122 do_page_add_anon_rmap(page
, vma
, address
, exclusive
);
3123 else /* ksm created a completely new copy */
3124 page_add_new_anon_rmap(page
, vma
, address
);
3125 /* It's better to call commit-charge after rmap is established */
3126 mem_cgroup_commit_charge_swapin(page
, ptr
);
3129 if (vm_swap_full() || (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
3130 try_to_free_swap(page
);
3132 if (page
!= swapcache
) {
3134 * Hold the lock to avoid the swap entry to be reused
3135 * until we take the PT lock for the pte_same() check
3136 * (to avoid false positives from pte_same). For
3137 * further safety release the lock after the swap_free
3138 * so that the swap count won't change under a
3139 * parallel locked swapcache.
3141 unlock_page(swapcache
);
3142 page_cache_release(swapcache
);
3145 if (flags
& FAULT_FLAG_WRITE
) {
3146 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
3147 if (ret
& VM_FAULT_ERROR
)
3148 ret
&= VM_FAULT_ERROR
;
3152 /* No need to invalidate - it was non-present before */
3153 update_mmu_cache(vma
, address
, page_table
);
3155 pte_unmap_unlock(page_table
, ptl
);
3159 mem_cgroup_cancel_charge_swapin(ptr
);
3160 pte_unmap_unlock(page_table
, ptl
);
3164 page_cache_release(page
);
3165 if (page
!= swapcache
) {
3166 unlock_page(swapcache
);
3167 page_cache_release(swapcache
);
3173 * This is like a special single-page "expand_{down|up}wards()",
3174 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3175 * doesn't hit another vma.
3177 static inline int check_stack_guard_page(struct vm_area_struct
*vma
, unsigned long address
)
3179 address
&= PAGE_MASK
;
3180 if ((vma
->vm_flags
& VM_GROWSDOWN
) && address
== vma
->vm_start
) {
3181 struct vm_area_struct
*prev
= vma
->vm_prev
;
3184 * Is there a mapping abutting this one below?
3186 * That's only ok if it's the same stack mapping
3187 * that has gotten split..
3189 if (prev
&& prev
->vm_end
== address
)
3190 return prev
->vm_flags
& VM_GROWSDOWN
? 0 : -ENOMEM
;
3192 expand_downwards(vma
, address
- PAGE_SIZE
);
3194 if ((vma
->vm_flags
& VM_GROWSUP
) && address
+ PAGE_SIZE
== vma
->vm_end
) {
3195 struct vm_area_struct
*next
= vma
->vm_next
;
3197 /* As VM_GROWSDOWN but s/below/above/ */
3198 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
)
3199 return next
->vm_flags
& VM_GROWSUP
? 0 : -ENOMEM
;
3201 expand_upwards(vma
, address
+ PAGE_SIZE
);
3207 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3208 * but allow concurrent faults), and pte mapped but not yet locked.
3209 * We return with mmap_sem still held, but pte unmapped and unlocked.
3211 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3212 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3219 pte_unmap(page_table
);
3221 /* Check if we need to add a guard page to the stack */
3222 if (check_stack_guard_page(vma
, address
) < 0)
3223 return VM_FAULT_SIGBUS
;
3225 /* Use the zero-page for reads */
3226 if (!(flags
& FAULT_FLAG_WRITE
)) {
3227 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(address
),
3228 vma
->vm_page_prot
));
3229 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3230 if (!pte_none(*page_table
))
3235 /* Allocate our own private page. */
3236 if (unlikely(anon_vma_prepare(vma
)))
3238 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
3242 * The memory barrier inside __SetPageUptodate makes sure that
3243 * preceeding stores to the page contents become visible before
3244 * the set_pte_at() write.
3246 __SetPageUptodate(page
);
3248 if (mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
))
3251 entry
= mk_pte(page
, vma
->vm_page_prot
);
3252 if (vma
->vm_flags
& VM_WRITE
)
3253 entry
= pte_mkwrite(pte_mkdirty(entry
));
3255 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3256 if (!pte_none(*page_table
))
3259 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3260 page_add_new_anon_rmap(page
, vma
, address
);
3262 set_pte_at(mm
, address
, page_table
, entry
);
3264 /* No need to invalidate - it was non-present before */
3265 update_mmu_cache(vma
, address
, page_table
);
3267 pte_unmap_unlock(page_table
, ptl
);
3270 mem_cgroup_uncharge_page(page
);
3271 page_cache_release(page
);
3274 page_cache_release(page
);
3276 return VM_FAULT_OOM
;
3280 * __do_fault() tries to create a new page mapping. It aggressively
3281 * tries to share with existing pages, but makes a separate copy if
3282 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3283 * the next page fault.
3285 * As this is called only for pages that do not currently exist, we
3286 * do not need to flush old virtual caches or the TLB.
3288 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3289 * but allow concurrent faults), and pte neither mapped nor locked.
3290 * We return with mmap_sem still held, but pte unmapped and unlocked.
3292 static int __do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3293 unsigned long address
, pmd_t
*pmd
,
3294 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3299 struct page
*cow_page
;
3302 struct page
*dirty_page
= NULL
;
3303 struct vm_fault vmf
;
3305 int page_mkwrite
= 0;
3308 * If we do COW later, allocate page befor taking lock_page()
3309 * on the file cache page. This will reduce lock holding time.
3311 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3313 if (unlikely(anon_vma_prepare(vma
)))
3314 return VM_FAULT_OOM
;
3316 cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
3318 return VM_FAULT_OOM
;
3320 if (mem_cgroup_newpage_charge(cow_page
, mm
, GFP_KERNEL
)) {
3321 page_cache_release(cow_page
);
3322 return VM_FAULT_OOM
;
3327 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
3332 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
3333 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
3337 if (unlikely(PageHWPoison(vmf
.page
))) {
3338 if (ret
& VM_FAULT_LOCKED
)
3339 unlock_page(vmf
.page
);
3340 ret
= VM_FAULT_HWPOISON
;
3345 * For consistency in subsequent calls, make the faulted page always
3348 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3349 lock_page(vmf
.page
);
3351 VM_BUG_ON(!PageLocked(vmf
.page
));
3354 * Should we do an early C-O-W break?
3357 if (flags
& FAULT_FLAG_WRITE
) {
3358 if (!(vma
->vm_flags
& VM_SHARED
)) {
3361 copy_user_highpage(page
, vmf
.page
, address
, vma
);
3362 __SetPageUptodate(page
);
3365 * If the page will be shareable, see if the backing
3366 * address space wants to know that the page is about
3367 * to become writable
3369 if (vma
->vm_ops
->page_mkwrite
) {
3373 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
3374 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
3376 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3378 goto unwritable_page
;
3380 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
3382 if (!page
->mapping
) {
3383 ret
= 0; /* retry the fault */
3385 goto unwritable_page
;
3388 VM_BUG_ON(!PageLocked(page
));
3395 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3398 * This silly early PAGE_DIRTY setting removes a race
3399 * due to the bad i386 page protection. But it's valid
3400 * for other architectures too.
3402 * Note that if FAULT_FLAG_WRITE is set, we either now have
3403 * an exclusive copy of the page, or this is a shared mapping,
3404 * so we can make it writable and dirty to avoid having to
3405 * handle that later.
3407 /* Only go through if we didn't race with anybody else... */
3408 if (likely(pte_same(*page_table
, orig_pte
))) {
3409 flush_icache_page(vma
, page
);
3410 entry
= mk_pte(page
, vma
->vm_page_prot
);
3411 if (flags
& FAULT_FLAG_WRITE
)
3412 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3414 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3415 page_add_new_anon_rmap(page
, vma
, address
);
3417 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
3418 page_add_file_rmap(page
);
3419 if (flags
& FAULT_FLAG_WRITE
) {
3421 get_page(dirty_page
);
3424 set_pte_at(mm
, address
, page_table
, entry
);
3426 /* no need to invalidate: a not-present page won't be cached */
3427 update_mmu_cache(vma
, address
, page_table
);
3430 mem_cgroup_uncharge_page(cow_page
);
3432 page_cache_release(page
);
3434 anon
= 1; /* no anon but release faulted_page */
3437 pte_unmap_unlock(page_table
, ptl
);
3440 struct address_space
*mapping
= page
->mapping
;
3443 if (set_page_dirty(dirty_page
))
3445 unlock_page(dirty_page
);
3446 put_page(dirty_page
);
3447 if ((dirtied
|| page_mkwrite
) && mapping
) {
3449 * Some device drivers do not set page.mapping but still
3452 balance_dirty_pages_ratelimited(mapping
);
3455 /* file_update_time outside page_lock */
3456 if (vma
->vm_file
&& !page_mkwrite
)
3457 file_update_time(vma
->vm_file
);
3459 unlock_page(vmf
.page
);
3461 page_cache_release(vmf
.page
);
3467 page_cache_release(page
);
3470 /* fs's fault handler get error */
3472 mem_cgroup_uncharge_page(cow_page
);
3473 page_cache_release(cow_page
);
3478 static int do_linear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3479 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3480 unsigned int flags
, pte_t orig_pte
)
3482 pgoff_t pgoff
= (((address
& PAGE_MASK
)
3483 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
3485 pte_unmap(page_table
);
3486 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3490 * Fault of a previously existing named mapping. Repopulate the pte
3491 * from the encoded file_pte if possible. This enables swappable
3494 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3495 * but allow concurrent faults), and pte mapped but not yet locked.
3496 * We return with mmap_sem still held, but pte unmapped and unlocked.
3498 static int do_nonlinear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3499 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3500 unsigned int flags
, pte_t orig_pte
)
3504 flags
|= FAULT_FLAG_NONLINEAR
;
3506 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
3509 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
))) {
3511 * Page table corrupted: show pte and kill process.
3513 print_bad_pte(vma
, address
, orig_pte
, NULL
);
3514 return VM_FAULT_SIGBUS
;
3517 pgoff
= pte_to_pgoff(orig_pte
);
3518 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3521 int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3522 unsigned long addr
, int current_nid
)
3526 count_vm_numa_event(NUMA_HINT_FAULTS
);
3527 if (current_nid
== numa_node_id())
3528 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3530 return mpol_misplaced(page
, vma
, addr
);
3533 int do_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3534 unsigned long addr
, pte_t pte
, pte_t
*ptep
, pmd_t
*pmd
)
3536 struct page
*page
= NULL
;
3538 int current_nid
= -1;
3540 bool migrated
= false;
3543 * The "pte" at this point cannot be used safely without
3544 * validation through pte_unmap_same(). It's of NUMA type but
3545 * the pfn may be screwed if the read is non atomic.
3547 * ptep_modify_prot_start is not called as this is clearing
3548 * the _PAGE_NUMA bit and it is not really expected that there
3549 * would be concurrent hardware modifications to the PTE.
3551 ptl
= pte_lockptr(mm
, pmd
);
3553 if (unlikely(!pte_same(*ptep
, pte
))) {
3554 pte_unmap_unlock(ptep
, ptl
);
3558 pte
= pte_mknonnuma(pte
);
3559 set_pte_at(mm
, addr
, ptep
, pte
);
3560 update_mmu_cache(vma
, addr
, ptep
);
3562 page
= vm_normal_page(vma
, addr
, pte
);
3564 pte_unmap_unlock(ptep
, ptl
);
3568 current_nid
= page_to_nid(page
);
3569 target_nid
= numa_migrate_prep(page
, vma
, addr
, current_nid
);
3570 pte_unmap_unlock(ptep
, ptl
);
3571 if (target_nid
== -1) {
3573 * Account for the fault against the current node if it not
3574 * being replaced regardless of where the page is located.
3576 current_nid
= numa_node_id();
3581 /* Migrate to the requested node */
3582 migrated
= migrate_misplaced_page(page
, target_nid
);
3584 current_nid
= target_nid
;
3587 if (current_nid
!= -1)
3588 task_numa_fault(current_nid
, 1, migrated
);
3592 /* NUMA hinting page fault entry point for regular pmds */
3593 #ifdef CONFIG_NUMA_BALANCING
3594 static int do_pmd_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3595 unsigned long addr
, pmd_t
*pmdp
)
3598 pte_t
*pte
, *orig_pte
;
3599 unsigned long _addr
= addr
& PMD_MASK
;
3600 unsigned long offset
;
3603 int local_nid
= numa_node_id();
3605 spin_lock(&mm
->page_table_lock
);
3607 if (pmd_numa(pmd
)) {
3608 set_pmd_at(mm
, _addr
, pmdp
, pmd_mknonnuma(pmd
));
3611 spin_unlock(&mm
->page_table_lock
);
3616 /* we're in a page fault so some vma must be in the range */
3618 BUG_ON(vma
->vm_start
>= _addr
+ PMD_SIZE
);
3619 offset
= max(_addr
, vma
->vm_start
) & ~PMD_MASK
;
3620 VM_BUG_ON(offset
>= PMD_SIZE
);
3621 orig_pte
= pte
= pte_offset_map_lock(mm
, pmdp
, _addr
, &ptl
);
3622 pte
+= offset
>> PAGE_SHIFT
;
3623 for (addr
= _addr
+ offset
; addr
< _addr
+ PMD_SIZE
; pte
++, addr
+= PAGE_SIZE
) {
3624 pte_t pteval
= *pte
;
3626 int curr_nid
= local_nid
;
3629 if (!pte_present(pteval
))
3631 if (!pte_numa(pteval
))
3633 if (addr
>= vma
->vm_end
) {
3634 vma
= find_vma(mm
, addr
);
3635 /* there's a pte present so there must be a vma */
3637 BUG_ON(addr
< vma
->vm_start
);
3639 if (pte_numa(pteval
)) {
3640 pteval
= pte_mknonnuma(pteval
);
3641 set_pte_at(mm
, addr
, pte
, pteval
);
3643 page
= vm_normal_page(vma
, addr
, pteval
);
3644 if (unlikely(!page
))
3646 /* only check non-shared pages */
3647 if (unlikely(page_mapcount(page
) != 1))
3651 * Note that the NUMA fault is later accounted to either
3652 * the node that is currently running or where the page is
3655 curr_nid
= local_nid
;
3656 target_nid
= numa_migrate_prep(page
, vma
, addr
,
3658 if (target_nid
== -1) {
3663 /* Migrate to the requested node */
3664 pte_unmap_unlock(pte
, ptl
);
3665 migrated
= migrate_misplaced_page(page
, target_nid
);
3667 curr_nid
= target_nid
;
3668 task_numa_fault(curr_nid
, 1, migrated
);
3670 pte
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
3672 pte_unmap_unlock(orig_pte
, ptl
);
3677 static int do_pmd_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3678 unsigned long addr
, pmd_t
*pmdp
)
3683 #endif /* CONFIG_NUMA_BALANCING */
3686 * These routines also need to handle stuff like marking pages dirty
3687 * and/or accessed for architectures that don't do it in hardware (most
3688 * RISC architectures). The early dirtying is also good on the i386.
3690 * There is also a hook called "update_mmu_cache()" that architectures
3691 * with external mmu caches can use to update those (ie the Sparc or
3692 * PowerPC hashed page tables that act as extended TLBs).
3694 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3695 * but allow concurrent faults), and pte mapped but not yet locked.
3696 * We return with mmap_sem still held, but pte unmapped and unlocked.
3698 int handle_pte_fault(struct mm_struct
*mm
,
3699 struct vm_area_struct
*vma
, unsigned long address
,
3700 pte_t
*pte
, pmd_t
*pmd
, unsigned int flags
)
3706 if (!pte_present(entry
)) {
3707 if (pte_none(entry
)) {
3709 if (likely(vma
->vm_ops
->fault
))
3710 return do_linear_fault(mm
, vma
, address
,
3711 pte
, pmd
, flags
, entry
);
3713 return do_anonymous_page(mm
, vma
, address
,
3716 if (pte_file(entry
))
3717 return do_nonlinear_fault(mm
, vma
, address
,
3718 pte
, pmd
, flags
, entry
);
3719 return do_swap_page(mm
, vma
, address
,
3720 pte
, pmd
, flags
, entry
);
3723 if (pte_numa(entry
))
3724 return do_numa_page(mm
, vma
, address
, entry
, pte
, pmd
);
3726 ptl
= pte_lockptr(mm
, pmd
);
3728 if (unlikely(!pte_same(*pte
, entry
)))
3730 if (flags
& FAULT_FLAG_WRITE
) {
3731 if (!pte_write(entry
))
3732 return do_wp_page(mm
, vma
, address
,
3733 pte
, pmd
, ptl
, entry
);
3734 entry
= pte_mkdirty(entry
);
3736 entry
= pte_mkyoung(entry
);
3737 if (ptep_set_access_flags(vma
, address
, pte
, entry
, flags
& FAULT_FLAG_WRITE
)) {
3738 update_mmu_cache(vma
, address
, pte
);
3741 * This is needed only for protection faults but the arch code
3742 * is not yet telling us if this is a protection fault or not.
3743 * This still avoids useless tlb flushes for .text page faults
3746 if (flags
& FAULT_FLAG_WRITE
)
3747 flush_tlb_fix_spurious_fault(vma
, address
);
3750 pte_unmap_unlock(pte
, ptl
);
3755 * By the time we get here, we already hold the mm semaphore
3757 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3758 unsigned long address
, unsigned int flags
)
3765 __set_current_state(TASK_RUNNING
);
3767 count_vm_event(PGFAULT
);
3768 mem_cgroup_count_vm_event(mm
, PGFAULT
);
3770 /* do counter updates before entering really critical section. */
3771 check_sync_rss_stat(current
);
3773 if (unlikely(is_vm_hugetlb_page(vma
)))
3774 return hugetlb_fault(mm
, vma
, address
, flags
);
3777 pgd
= pgd_offset(mm
, address
);
3778 pud
= pud_alloc(mm
, pgd
, address
);
3780 return VM_FAULT_OOM
;
3781 pmd
= pmd_alloc(mm
, pud
, address
);
3783 return VM_FAULT_OOM
;
3784 if (pmd_none(*pmd
) && transparent_hugepage_enabled(vma
)) {
3786 return do_huge_pmd_anonymous_page(mm
, vma
, address
,
3789 pmd_t orig_pmd
= *pmd
;
3793 if (pmd_trans_huge(orig_pmd
)) {
3794 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
3797 * If the pmd is splitting, return and retry the
3798 * the fault. Alternative: wait until the split
3799 * is done, and goto retry.
3801 if (pmd_trans_splitting(orig_pmd
))
3804 if (pmd_numa(orig_pmd
))
3805 return do_huge_pmd_numa_page(mm
, vma
, address
,
3808 if (dirty
&& !pmd_write(orig_pmd
)) {
3809 ret
= do_huge_pmd_wp_page(mm
, vma
, address
, pmd
,
3812 * If COW results in an oom, the huge pmd will
3813 * have been split, so retry the fault on the
3814 * pte for a smaller charge.
3816 if (unlikely(ret
& VM_FAULT_OOM
))
3820 huge_pmd_set_accessed(mm
, vma
, address
, pmd
,
3829 return do_pmd_numa_page(mm
, vma
, address
, pmd
);
3832 * Use __pte_alloc instead of pte_alloc_map, because we can't
3833 * run pte_offset_map on the pmd, if an huge pmd could
3834 * materialize from under us from a different thread.
3836 if (unlikely(pmd_none(*pmd
)) &&
3837 unlikely(__pte_alloc(mm
, vma
, pmd
, address
)))
3838 return VM_FAULT_OOM
;
3839 /* if an huge pmd materialized from under us just retry later */
3840 if (unlikely(pmd_trans_huge(*pmd
)))
3843 * A regular pmd is established and it can't morph into a huge pmd
3844 * from under us anymore at this point because we hold the mmap_sem
3845 * read mode and khugepaged takes it in write mode. So now it's
3846 * safe to run pte_offset_map().
3848 pte
= pte_offset_map(pmd
, address
);
3850 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
3853 #ifndef __PAGETABLE_PUD_FOLDED
3855 * Allocate page upper directory.
3856 * We've already handled the fast-path in-line.
3858 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
3860 pud_t
*new = pud_alloc_one(mm
, address
);
3864 smp_wmb(); /* See comment in __pte_alloc */
3866 spin_lock(&mm
->page_table_lock
);
3867 if (pgd_present(*pgd
)) /* Another has populated it */
3870 pgd_populate(mm
, pgd
, new);
3871 spin_unlock(&mm
->page_table_lock
);
3874 #endif /* __PAGETABLE_PUD_FOLDED */
3876 #ifndef __PAGETABLE_PMD_FOLDED
3878 * Allocate page middle directory.
3879 * We've already handled the fast-path in-line.
3881 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
3883 pmd_t
*new = pmd_alloc_one(mm
, address
);
3887 smp_wmb(); /* See comment in __pte_alloc */
3889 spin_lock(&mm
->page_table_lock
);
3890 #ifndef __ARCH_HAS_4LEVEL_HACK
3891 if (pud_present(*pud
)) /* Another has populated it */
3894 pud_populate(mm
, pud
, new);
3896 if (pgd_present(*pud
)) /* Another has populated it */
3899 pgd_populate(mm
, pud
, new);
3900 #endif /* __ARCH_HAS_4LEVEL_HACK */
3901 spin_unlock(&mm
->page_table_lock
);
3904 #endif /* __PAGETABLE_PMD_FOLDED */
3906 #if !defined(__HAVE_ARCH_GATE_AREA)
3908 #if defined(AT_SYSINFO_EHDR)
3909 static struct vm_area_struct gate_vma
;
3911 static int __init
gate_vma_init(void)
3913 gate_vma
.vm_mm
= NULL
;
3914 gate_vma
.vm_start
= FIXADDR_USER_START
;
3915 gate_vma
.vm_end
= FIXADDR_USER_END
;
3916 gate_vma
.vm_flags
= VM_READ
| VM_MAYREAD
| VM_EXEC
| VM_MAYEXEC
;
3917 gate_vma
.vm_page_prot
= __P101
;
3921 __initcall(gate_vma_init
);
3924 struct vm_area_struct
*get_gate_vma(struct mm_struct
*mm
)
3926 #ifdef AT_SYSINFO_EHDR
3933 int in_gate_area_no_mm(unsigned long addr
)
3935 #ifdef AT_SYSINFO_EHDR
3936 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
3942 #endif /* __HAVE_ARCH_GATE_AREA */
3944 static int __follow_pte(struct mm_struct
*mm
, unsigned long address
,
3945 pte_t
**ptepp
, spinlock_t
**ptlp
)
3952 pgd
= pgd_offset(mm
, address
);
3953 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
3956 pud
= pud_offset(pgd
, address
);
3957 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
3960 pmd
= pmd_offset(pud
, address
);
3961 VM_BUG_ON(pmd_trans_huge(*pmd
));
3962 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
3965 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3969 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
3972 if (!pte_present(*ptep
))
3977 pte_unmap_unlock(ptep
, *ptlp
);
3982 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
3983 pte_t
**ptepp
, spinlock_t
**ptlp
)
3987 /* (void) is needed to make gcc happy */
3988 (void) __cond_lock(*ptlp
,
3989 !(res
= __follow_pte(mm
, address
, ptepp
, ptlp
)));
3994 * follow_pfn - look up PFN at a user virtual address
3995 * @vma: memory mapping
3996 * @address: user virtual address
3997 * @pfn: location to store found PFN
3999 * Only IO mappings and raw PFN mappings are allowed.
4001 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4003 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
4010 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4013 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
4016 *pfn
= pte_pfn(*ptep
);
4017 pte_unmap_unlock(ptep
, ptl
);
4020 EXPORT_SYMBOL(follow_pfn
);
4022 #ifdef CONFIG_HAVE_IOREMAP_PROT
4023 int follow_phys(struct vm_area_struct
*vma
,
4024 unsigned long address
, unsigned int flags
,
4025 unsigned long *prot
, resource_size_t
*phys
)
4031 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4034 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
4038 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
4041 *prot
= pgprot_val(pte_pgprot(pte
));
4042 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
4046 pte_unmap_unlock(ptep
, ptl
);
4051 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
4052 void *buf
, int len
, int write
)
4054 resource_size_t phys_addr
;
4055 unsigned long prot
= 0;
4056 void __iomem
*maddr
;
4057 int offset
= addr
& (PAGE_SIZE
-1);
4059 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
4062 maddr
= ioremap_prot(phys_addr
, PAGE_SIZE
, prot
);
4064 memcpy_toio(maddr
+ offset
, buf
, len
);
4066 memcpy_fromio(buf
, maddr
+ offset
, len
);
4074 * Access another process' address space as given in mm. If non-NULL, use the
4075 * given task for page fault accounting.
4077 static int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
4078 unsigned long addr
, void *buf
, int len
, int write
)
4080 struct vm_area_struct
*vma
;
4081 void *old_buf
= buf
;
4083 down_read(&mm
->mmap_sem
);
4084 /* ignore errors, just check how much was successfully transferred */
4086 int bytes
, ret
, offset
;
4088 struct page
*page
= NULL
;
4090 ret
= get_user_pages(tsk
, mm
, addr
, 1,
4091 write
, 1, &page
, &vma
);
4094 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4095 * we can access using slightly different code.
4097 #ifdef CONFIG_HAVE_IOREMAP_PROT
4098 vma
= find_vma(mm
, addr
);
4099 if (!vma
|| vma
->vm_start
> addr
)
4101 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
4102 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
4110 offset
= addr
& (PAGE_SIZE
-1);
4111 if (bytes
> PAGE_SIZE
-offset
)
4112 bytes
= PAGE_SIZE
-offset
;
4116 copy_to_user_page(vma
, page
, addr
,
4117 maddr
+ offset
, buf
, bytes
);
4118 set_page_dirty_lock(page
);
4120 copy_from_user_page(vma
, page
, addr
,
4121 buf
, maddr
+ offset
, bytes
);
4124 page_cache_release(page
);
4130 up_read(&mm
->mmap_sem
);
4132 return buf
- old_buf
;
4136 * access_remote_vm - access another process' address space
4137 * @mm: the mm_struct of the target address space
4138 * @addr: start address to access
4139 * @buf: source or destination buffer
4140 * @len: number of bytes to transfer
4141 * @write: whether the access is a write
4143 * The caller must hold a reference on @mm.
4145 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
4146 void *buf
, int len
, int write
)
4148 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, write
);
4152 * Access another process' address space.
4153 * Source/target buffer must be kernel space,
4154 * Do not walk the page table directly, use get_user_pages
4156 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
4157 void *buf
, int len
, int write
)
4159 struct mm_struct
*mm
;
4162 mm
= get_task_mm(tsk
);
4166 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, write
);
4173 * Print the name of a VMA.
4175 void print_vma_addr(char *prefix
, unsigned long ip
)
4177 struct mm_struct
*mm
= current
->mm
;
4178 struct vm_area_struct
*vma
;
4181 * Do not print if we are in atomic
4182 * contexts (in exception stacks, etc.):
4184 if (preempt_count())
4187 down_read(&mm
->mmap_sem
);
4188 vma
= find_vma(mm
, ip
);
4189 if (vma
&& vma
->vm_file
) {
4190 struct file
*f
= vma
->vm_file
;
4191 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
4195 p
= d_path(&f
->f_path
, buf
, PAGE_SIZE
);
4198 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
4200 vma
->vm_end
- vma
->vm_start
);
4201 free_page((unsigned long)buf
);
4204 up_read(&mm
->mmap_sem
);
4207 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4208 void might_fault(void)
4211 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4212 * holding the mmap_sem, this is safe because kernel memory doesn't
4213 * get paged out, therefore we'll never actually fault, and the
4214 * below annotations will generate false positives.
4216 if (segment_eq(get_fs(), KERNEL_DS
))
4220 * it would be nicer only to annotate paths which are not under
4221 * pagefault_disable, however that requires a larger audit and
4222 * providing helpers like get_user_atomic.
4227 __might_sleep(__FILE__
, __LINE__
, 0);
4230 might_lock_read(¤t
->mm
->mmap_sem
);
4232 EXPORT_SYMBOL(might_fault
);
4235 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4236 static void clear_gigantic_page(struct page
*page
,
4238 unsigned int pages_per_huge_page
)
4241 struct page
*p
= page
;
4244 for (i
= 0; i
< pages_per_huge_page
;
4245 i
++, p
= mem_map_next(p
, page
, i
)) {
4247 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
4250 void clear_huge_page(struct page
*page
,
4251 unsigned long addr
, unsigned int pages_per_huge_page
)
4255 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4256 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
4261 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4263 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
4267 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
4269 struct vm_area_struct
*vma
,
4270 unsigned int pages_per_huge_page
)
4273 struct page
*dst_base
= dst
;
4274 struct page
*src_base
= src
;
4276 for (i
= 0; i
< pages_per_huge_page
; ) {
4278 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
4281 dst
= mem_map_next(dst
, dst_base
, i
);
4282 src
= mem_map_next(src
, src_base
, i
);
4286 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
4287 unsigned long addr
, struct vm_area_struct
*vma
,
4288 unsigned int pages_per_huge_page
)
4292 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4293 copy_user_gigantic_page(dst
, src
, addr
, vma
,
4294 pages_per_huge_page
);
4299 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4301 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
4304 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */