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
;
222 tlb
->fast_mode
= (num_possible_cpus() == 1);
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 if (tlb_fast_mode(tlb
))
249 for (batch
= &tlb
->local
; batch
; batch
= batch
->next
) {
250 free_pages_and_swap_cache(batch
->pages
, batch
->nr
);
253 tlb
->active
= &tlb
->local
;
257 * Called at the end of the shootdown operation to free up any resources
258 * that were required.
260 void tlb_finish_mmu(struct mmu_gather
*tlb
, unsigned long start
, unsigned long end
)
262 struct mmu_gather_batch
*batch
, *next
;
268 /* keep the page table cache within bounds */
271 for (batch
= tlb
->local
.next
; batch
; batch
= next
) {
273 free_pages((unsigned long)batch
, 0);
275 tlb
->local
.next
= NULL
;
279 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
280 * handling the additional races in SMP caused by other CPUs caching valid
281 * mappings in their TLBs. Returns the number of free page slots left.
282 * When out of page slots we must call tlb_flush_mmu().
284 int __tlb_remove_page(struct mmu_gather
*tlb
, struct page
*page
)
286 struct mmu_gather_batch
*batch
;
288 VM_BUG_ON(!tlb
->need_flush
);
290 if (tlb_fast_mode(tlb
)) {
291 free_page_and_swap_cache(page
);
292 return 1; /* avoid calling tlb_flush_mmu() */
296 batch
->pages
[batch
->nr
++] = page
;
297 if (batch
->nr
== batch
->max
) {
298 if (!tlb_next_batch(tlb
))
302 VM_BUG_ON(batch
->nr
> batch
->max
);
304 return batch
->max
- batch
->nr
;
307 #endif /* HAVE_GENERIC_MMU_GATHER */
309 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
312 * See the comment near struct mmu_table_batch.
315 static void tlb_remove_table_smp_sync(void *arg
)
317 /* Simply deliver the interrupt */
320 static void tlb_remove_table_one(void *table
)
323 * This isn't an RCU grace period and hence the page-tables cannot be
324 * assumed to be actually RCU-freed.
326 * It is however sufficient for software page-table walkers that rely on
327 * IRQ disabling. See the comment near struct mmu_table_batch.
329 smp_call_function(tlb_remove_table_smp_sync
, NULL
, 1);
330 __tlb_remove_table(table
);
333 static void tlb_remove_table_rcu(struct rcu_head
*head
)
335 struct mmu_table_batch
*batch
;
338 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
340 for (i
= 0; i
< batch
->nr
; i
++)
341 __tlb_remove_table(batch
->tables
[i
]);
343 free_page((unsigned long)batch
);
346 void tlb_table_flush(struct mmu_gather
*tlb
)
348 struct mmu_table_batch
**batch
= &tlb
->batch
;
351 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
356 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
358 struct mmu_table_batch
**batch
= &tlb
->batch
;
363 * When there's less then two users of this mm there cannot be a
364 * concurrent page-table walk.
366 if (atomic_read(&tlb
->mm
->mm_users
) < 2) {
367 __tlb_remove_table(table
);
371 if (*batch
== NULL
) {
372 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
373 if (*batch
== NULL
) {
374 tlb_remove_table_one(table
);
379 (*batch
)->tables
[(*batch
)->nr
++] = table
;
380 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
381 tlb_table_flush(tlb
);
384 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
387 * If a p?d_bad entry is found while walking page tables, report
388 * the error, before resetting entry to p?d_none. Usually (but
389 * very seldom) called out from the p?d_none_or_clear_bad macros.
392 void pgd_clear_bad(pgd_t
*pgd
)
398 void pud_clear_bad(pud_t
*pud
)
404 void pmd_clear_bad(pmd_t
*pmd
)
411 * Note: this doesn't free the actual pages themselves. That
412 * has been handled earlier when unmapping all the memory regions.
414 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
417 pgtable_t token
= pmd_pgtable(*pmd
);
419 pte_free_tlb(tlb
, token
, addr
);
423 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
424 unsigned long addr
, unsigned long end
,
425 unsigned long floor
, unsigned long ceiling
)
432 pmd
= pmd_offset(pud
, addr
);
434 next
= pmd_addr_end(addr
, end
);
435 if (pmd_none_or_clear_bad(pmd
))
437 free_pte_range(tlb
, pmd
, addr
);
438 } while (pmd
++, addr
= next
, addr
!= end
);
448 if (end
- 1 > ceiling
- 1)
451 pmd
= pmd_offset(pud
, start
);
453 pmd_free_tlb(tlb
, pmd
, start
);
456 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
457 unsigned long addr
, unsigned long end
,
458 unsigned long floor
, unsigned long ceiling
)
465 pud
= pud_offset(pgd
, addr
);
467 next
= pud_addr_end(addr
, end
);
468 if (pud_none_or_clear_bad(pud
))
470 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
471 } while (pud
++, addr
= next
, addr
!= end
);
477 ceiling
&= PGDIR_MASK
;
481 if (end
- 1 > ceiling
- 1)
484 pud
= pud_offset(pgd
, start
);
486 pud_free_tlb(tlb
, pud
, start
);
490 * This function frees user-level page tables of a process.
492 * Must be called with pagetable lock held.
494 void free_pgd_range(struct mmu_gather
*tlb
,
495 unsigned long addr
, unsigned long end
,
496 unsigned long floor
, unsigned long ceiling
)
502 * The next few lines have given us lots of grief...
504 * Why are we testing PMD* at this top level? Because often
505 * there will be no work to do at all, and we'd prefer not to
506 * go all the way down to the bottom just to discover that.
508 * Why all these "- 1"s? Because 0 represents both the bottom
509 * of the address space and the top of it (using -1 for the
510 * top wouldn't help much: the masks would do the wrong thing).
511 * The rule is that addr 0 and floor 0 refer to the bottom of
512 * the address space, but end 0 and ceiling 0 refer to the top
513 * Comparisons need to use "end - 1" and "ceiling - 1" (though
514 * that end 0 case should be mythical).
516 * Wherever addr is brought up or ceiling brought down, we must
517 * be careful to reject "the opposite 0" before it confuses the
518 * subsequent tests. But what about where end is brought down
519 * by PMD_SIZE below? no, end can't go down to 0 there.
521 * Whereas we round start (addr) and ceiling down, by different
522 * masks at different levels, in order to test whether a table
523 * now has no other vmas using it, so can be freed, we don't
524 * bother to round floor or end up - the tests don't need that.
538 if (end
- 1 > ceiling
- 1)
543 pgd
= pgd_offset(tlb
->mm
, addr
);
545 next
= pgd_addr_end(addr
, end
);
546 if (pgd_none_or_clear_bad(pgd
))
548 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
549 } while (pgd
++, addr
= next
, addr
!= end
);
552 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
553 unsigned long floor
, unsigned long ceiling
)
556 struct vm_area_struct
*next
= vma
->vm_next
;
557 unsigned long addr
= vma
->vm_start
;
560 * Hide vma from rmap and truncate_pagecache before freeing
563 unlink_anon_vmas(vma
);
564 unlink_file_vma(vma
);
566 if (is_vm_hugetlb_page(vma
)) {
567 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
568 floor
, next
? next
->vm_start
: ceiling
);
571 * Optimization: gather nearby vmas into one call down
573 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
574 && !is_vm_hugetlb_page(next
)) {
577 unlink_anon_vmas(vma
);
578 unlink_file_vma(vma
);
580 free_pgd_range(tlb
, addr
, vma
->vm_end
,
581 floor
, next
? next
->vm_start
: ceiling
);
587 int __pte_alloc(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
588 pmd_t
*pmd
, unsigned long address
)
590 pgtable_t
new = pte_alloc_one(mm
, address
);
591 int wait_split_huge_page
;
596 * Ensure all pte setup (eg. pte page lock and page clearing) are
597 * visible before the pte is made visible to other CPUs by being
598 * put into page tables.
600 * The other side of the story is the pointer chasing in the page
601 * table walking code (when walking the page table without locking;
602 * ie. most of the time). Fortunately, these data accesses consist
603 * of a chain of data-dependent loads, meaning most CPUs (alpha
604 * being the notable exception) will already guarantee loads are
605 * seen in-order. See the alpha page table accessors for the
606 * smp_read_barrier_depends() barriers in page table walking code.
608 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
610 spin_lock(&mm
->page_table_lock
);
611 wait_split_huge_page
= 0;
612 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
614 pmd_populate(mm
, pmd
, new);
616 } else if (unlikely(pmd_trans_splitting(*pmd
)))
617 wait_split_huge_page
= 1;
618 spin_unlock(&mm
->page_table_lock
);
621 if (wait_split_huge_page
)
622 wait_split_huge_page(vma
->anon_vma
, pmd
);
626 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
628 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
632 smp_wmb(); /* See comment in __pte_alloc */
634 spin_lock(&init_mm
.page_table_lock
);
635 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
636 pmd_populate_kernel(&init_mm
, pmd
, new);
639 VM_BUG_ON(pmd_trans_splitting(*pmd
));
640 spin_unlock(&init_mm
.page_table_lock
);
642 pte_free_kernel(&init_mm
, new);
646 static inline void init_rss_vec(int *rss
)
648 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
651 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
655 if (current
->mm
== mm
)
657 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
659 add_mm_counter(mm
, i
, rss
[i
]);
663 * This function is called to print an error when a bad pte
664 * is found. For example, we might have a PFN-mapped pte in
665 * a region that doesn't allow it.
667 * The calling function must still handle the error.
669 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
670 pte_t pte
, struct page
*page
)
672 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
673 pud_t
*pud
= pud_offset(pgd
, addr
);
674 pmd_t
*pmd
= pmd_offset(pud
, addr
);
675 struct address_space
*mapping
;
677 static unsigned long resume
;
678 static unsigned long nr_shown
;
679 static unsigned long nr_unshown
;
682 * Allow a burst of 60 reports, then keep quiet for that minute;
683 * or allow a steady drip of one report per second.
685 if (nr_shown
== 60) {
686 if (time_before(jiffies
, resume
)) {
692 "BUG: Bad page map: %lu messages suppressed\n",
699 resume
= jiffies
+ 60 * HZ
;
701 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
702 index
= linear_page_index(vma
, addr
);
705 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
707 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
711 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
712 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
714 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
717 print_symbol(KERN_ALERT
"vma->vm_ops->fault: %s\n",
718 (unsigned long)vma
->vm_ops
->fault
);
719 if (vma
->vm_file
&& vma
->vm_file
->f_op
)
720 print_symbol(KERN_ALERT
"vma->vm_file->f_op->mmap: %s\n",
721 (unsigned long)vma
->vm_file
->f_op
->mmap
);
723 add_taint(TAINT_BAD_PAGE
);
726 static inline bool is_cow_mapping(vm_flags_t flags
)
728 return (flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
732 * vm_normal_page -- This function gets the "struct page" associated with a pte.
734 * "Special" mappings do not wish to be associated with a "struct page" (either
735 * it doesn't exist, or it exists but they don't want to touch it). In this
736 * case, NULL is returned here. "Normal" mappings do have a struct page.
738 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
739 * pte bit, in which case this function is trivial. Secondly, an architecture
740 * may not have a spare pte bit, which requires a more complicated scheme,
743 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
744 * special mapping (even if there are underlying and valid "struct pages").
745 * COWed pages of a VM_PFNMAP are always normal.
747 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
748 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
749 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
750 * mapping will always honor the rule
752 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
754 * And for normal mappings this is false.
756 * This restricts such mappings to be a linear translation from virtual address
757 * to pfn. To get around this restriction, we allow arbitrary mappings so long
758 * as the vma is not a COW mapping; in that case, we know that all ptes are
759 * special (because none can have been COWed).
762 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
764 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
765 * page" backing, however the difference is that _all_ pages with a struct
766 * page (that is, those where pfn_valid is true) are refcounted and considered
767 * normal pages by the VM. The disadvantage is that pages are refcounted
768 * (which can be slower and simply not an option for some PFNMAP users). The
769 * advantage is that we don't have to follow the strict linearity rule of
770 * PFNMAP mappings in order to support COWable mappings.
773 #ifdef __HAVE_ARCH_PTE_SPECIAL
774 # define HAVE_PTE_SPECIAL 1
776 # define HAVE_PTE_SPECIAL 0
778 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
781 unsigned long pfn
= pte_pfn(pte
);
783 if (HAVE_PTE_SPECIAL
) {
784 if (likely(!pte_special(pte
)))
786 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
788 if (!is_zero_pfn(pfn
))
789 print_bad_pte(vma
, addr
, pte
, NULL
);
793 /* !HAVE_PTE_SPECIAL case follows: */
795 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
796 if (vma
->vm_flags
& VM_MIXEDMAP
) {
802 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
803 if (pfn
== vma
->vm_pgoff
+ off
)
805 if (!is_cow_mapping(vma
->vm_flags
))
810 if (is_zero_pfn(pfn
))
813 if (unlikely(pfn
> highest_memmap_pfn
)) {
814 print_bad_pte(vma
, addr
, pte
, NULL
);
819 * NOTE! We still have PageReserved() pages in the page tables.
820 * eg. VDSO mappings can cause them to exist.
823 return pfn_to_page(pfn
);
827 * copy one vm_area from one task to the other. Assumes the page tables
828 * already present in the new task to be cleared in the whole range
829 * covered by this vma.
832 static inline unsigned long
833 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
834 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
835 unsigned long addr
, int *rss
)
837 unsigned long vm_flags
= vma
->vm_flags
;
838 pte_t pte
= *src_pte
;
841 /* pte contains position in swap or file, so copy. */
842 if (unlikely(!pte_present(pte
))) {
843 if (!pte_file(pte
)) {
844 swp_entry_t entry
= pte_to_swp_entry(pte
);
846 if (swap_duplicate(entry
) < 0)
849 /* make sure dst_mm is on swapoff's mmlist. */
850 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
851 spin_lock(&mmlist_lock
);
852 if (list_empty(&dst_mm
->mmlist
))
853 list_add(&dst_mm
->mmlist
,
855 spin_unlock(&mmlist_lock
);
857 if (likely(!non_swap_entry(entry
)))
859 else if (is_migration_entry(entry
)) {
860 page
= migration_entry_to_page(entry
);
867 if (is_write_migration_entry(entry
) &&
868 is_cow_mapping(vm_flags
)) {
870 * COW mappings require pages in both
871 * parent and child to be set to read.
873 make_migration_entry_read(&entry
);
874 pte
= swp_entry_to_pte(entry
);
875 set_pte_at(src_mm
, addr
, src_pte
, pte
);
883 * If it's a COW mapping, write protect it both
884 * in the parent and the child
886 if (is_cow_mapping(vm_flags
)) {
887 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
888 pte
= pte_wrprotect(pte
);
892 * If it's a shared mapping, mark it clean in
895 if (vm_flags
& VM_SHARED
)
896 pte
= pte_mkclean(pte
);
897 pte
= pte_mkold(pte
);
899 page
= vm_normal_page(vma
, addr
, pte
);
910 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
914 int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
915 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
916 unsigned long addr
, unsigned long end
)
918 pte_t
*orig_src_pte
, *orig_dst_pte
;
919 pte_t
*src_pte
, *dst_pte
;
920 spinlock_t
*src_ptl
, *dst_ptl
;
922 int rss
[NR_MM_COUNTERS
];
923 swp_entry_t entry
= (swp_entry_t
){0};
928 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
931 src_pte
= pte_offset_map(src_pmd
, addr
);
932 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
933 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
934 orig_src_pte
= src_pte
;
935 orig_dst_pte
= dst_pte
;
936 arch_enter_lazy_mmu_mode();
940 * We are holding two locks at this point - either of them
941 * could generate latencies in another task on another CPU.
943 if (progress
>= 32) {
945 if (need_resched() ||
946 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
949 if (pte_none(*src_pte
)) {
953 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
958 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
960 arch_leave_lazy_mmu_mode();
961 spin_unlock(src_ptl
);
962 pte_unmap(orig_src_pte
);
963 add_mm_rss_vec(dst_mm
, rss
);
964 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
968 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
977 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
978 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
979 unsigned long addr
, unsigned long end
)
981 pmd_t
*src_pmd
, *dst_pmd
;
984 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
987 src_pmd
= pmd_offset(src_pud
, addr
);
989 next
= pmd_addr_end(addr
, end
);
990 if (pmd_trans_huge(*src_pmd
)) {
992 VM_BUG_ON(next
-addr
!= HPAGE_PMD_SIZE
);
993 err
= copy_huge_pmd(dst_mm
, src_mm
,
994 dst_pmd
, src_pmd
, addr
, vma
);
1001 if (pmd_none_or_clear_bad(src_pmd
))
1003 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
1006 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1010 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1011 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
1012 unsigned long addr
, unsigned long end
)
1014 pud_t
*src_pud
, *dst_pud
;
1017 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
1020 src_pud
= pud_offset(src_pgd
, addr
);
1022 next
= pud_addr_end(addr
, end
);
1023 if (pud_none_or_clear_bad(src_pud
))
1025 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
1028 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1032 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1033 struct vm_area_struct
*vma
)
1035 pgd_t
*src_pgd
, *dst_pgd
;
1037 unsigned long addr
= vma
->vm_start
;
1038 unsigned long end
= vma
->vm_end
;
1039 unsigned long mmun_start
; /* For mmu_notifiers */
1040 unsigned long mmun_end
; /* For mmu_notifiers */
1045 * Don't copy ptes where a page fault will fill them correctly.
1046 * Fork becomes much lighter when there are big shared or private
1047 * readonly mappings. The tradeoff is that copy_page_range is more
1048 * efficient than faulting.
1050 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_NONLINEAR
|
1051 VM_PFNMAP
| VM_MIXEDMAP
))) {
1056 if (is_vm_hugetlb_page(vma
))
1057 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1059 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
1061 * We do not free on error cases below as remove_vma
1062 * gets called on error from higher level routine
1064 ret
= track_pfn_copy(vma
);
1070 * We need to invalidate the secondary MMU mappings only when
1071 * there could be a permission downgrade on the ptes of the
1072 * parent mm. And a permission downgrade will only happen if
1073 * is_cow_mapping() returns true.
1075 is_cow
= is_cow_mapping(vma
->vm_flags
);
1079 mmu_notifier_invalidate_range_start(src_mm
, mmun_start
,
1083 dst_pgd
= pgd_offset(dst_mm
, addr
);
1084 src_pgd
= pgd_offset(src_mm
, addr
);
1086 next
= pgd_addr_end(addr
, end
);
1087 if (pgd_none_or_clear_bad(src_pgd
))
1089 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1090 vma
, addr
, next
))) {
1094 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1097 mmu_notifier_invalidate_range_end(src_mm
, mmun_start
, mmun_end
);
1101 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1102 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1103 unsigned long addr
, unsigned long end
,
1104 struct zap_details
*details
)
1106 struct mm_struct
*mm
= tlb
->mm
;
1107 int force_flush
= 0;
1108 int rss
[NR_MM_COUNTERS
];
1115 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1117 arch_enter_lazy_mmu_mode();
1120 if (pte_none(ptent
)) {
1124 if (pte_present(ptent
)) {
1127 page
= vm_normal_page(vma
, addr
, ptent
);
1128 if (unlikely(details
) && page
) {
1130 * unmap_shared_mapping_pages() wants to
1131 * invalidate cache without truncating:
1132 * unmap shared but keep private pages.
1134 if (details
->check_mapping
&&
1135 details
->check_mapping
!= page
->mapping
)
1138 * Each page->index must be checked when
1139 * invalidating or truncating nonlinear.
1141 if (details
->nonlinear_vma
&&
1142 (page
->index
< details
->first_index
||
1143 page
->index
> details
->last_index
))
1146 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1148 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1149 if (unlikely(!page
))
1151 if (unlikely(details
) && details
->nonlinear_vma
1152 && linear_page_index(details
->nonlinear_vma
,
1153 addr
) != page
->index
)
1154 set_pte_at(mm
, addr
, pte
,
1155 pgoff_to_pte(page
->index
));
1157 rss
[MM_ANONPAGES
]--;
1159 if (pte_dirty(ptent
))
1160 set_page_dirty(page
);
1161 if (pte_young(ptent
) &&
1162 likely(!VM_SequentialReadHint(vma
)))
1163 mark_page_accessed(page
);
1164 rss
[MM_FILEPAGES
]--;
1166 page_remove_rmap(page
);
1167 if (unlikely(page_mapcount(page
) < 0))
1168 print_bad_pte(vma
, addr
, ptent
, page
);
1169 force_flush
= !__tlb_remove_page(tlb
, page
);
1175 * If details->check_mapping, we leave swap entries;
1176 * if details->nonlinear_vma, we leave file entries.
1178 if (unlikely(details
))
1180 if (pte_file(ptent
)) {
1181 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
)))
1182 print_bad_pte(vma
, addr
, ptent
, NULL
);
1184 swp_entry_t entry
= pte_to_swp_entry(ptent
);
1186 if (!non_swap_entry(entry
))
1188 else if (is_migration_entry(entry
)) {
1191 page
= migration_entry_to_page(entry
);
1194 rss
[MM_ANONPAGES
]--;
1196 rss
[MM_FILEPAGES
]--;
1198 if (unlikely(!free_swap_and_cache(entry
)))
1199 print_bad_pte(vma
, addr
, ptent
, NULL
);
1201 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1202 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1204 add_mm_rss_vec(mm
, rss
);
1205 arch_leave_lazy_mmu_mode();
1206 pte_unmap_unlock(start_pte
, ptl
);
1209 * mmu_gather ran out of room to batch pages, we break out of
1210 * the PTE lock to avoid doing the potential expensive TLB invalidate
1211 * and page-free while holding it.
1216 #ifdef HAVE_GENERIC_MMU_GATHER
1228 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1229 struct vm_area_struct
*vma
, pud_t
*pud
,
1230 unsigned long addr
, unsigned long end
,
1231 struct zap_details
*details
)
1236 pmd
= pmd_offset(pud
, addr
);
1238 next
= pmd_addr_end(addr
, end
);
1239 if (pmd_trans_huge(*pmd
)) {
1240 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1241 #ifdef CONFIG_DEBUG_VM
1242 if (!rwsem_is_locked(&tlb
->mm
->mmap_sem
)) {
1243 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1244 __func__
, addr
, end
,
1250 split_huge_page_pmd(vma
, addr
, pmd
);
1251 } else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1256 * Here there can be other concurrent MADV_DONTNEED or
1257 * trans huge page faults running, and if the pmd is
1258 * none or trans huge it can change under us. This is
1259 * because MADV_DONTNEED holds the mmap_sem in read
1262 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1264 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1267 } while (pmd
++, addr
= next
, addr
!= end
);
1272 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1273 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1274 unsigned long addr
, unsigned long end
,
1275 struct zap_details
*details
)
1280 pud
= pud_offset(pgd
, addr
);
1282 next
= pud_addr_end(addr
, end
);
1283 if (pud_none_or_clear_bad(pud
))
1285 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1286 } while (pud
++, addr
= next
, addr
!= end
);
1291 static void unmap_page_range(struct mmu_gather
*tlb
,
1292 struct vm_area_struct
*vma
,
1293 unsigned long addr
, unsigned long end
,
1294 struct zap_details
*details
)
1299 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
1302 BUG_ON(addr
>= end
);
1303 mem_cgroup_uncharge_start();
1304 tlb_start_vma(tlb
, vma
);
1305 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1307 next
= pgd_addr_end(addr
, end
);
1308 if (pgd_none_or_clear_bad(pgd
))
1310 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
1311 } while (pgd
++, addr
= next
, addr
!= end
);
1312 tlb_end_vma(tlb
, vma
);
1313 mem_cgroup_uncharge_end();
1317 static void unmap_single_vma(struct mmu_gather
*tlb
,
1318 struct vm_area_struct
*vma
, unsigned long start_addr
,
1319 unsigned long end_addr
,
1320 struct zap_details
*details
)
1322 unsigned long start
= max(vma
->vm_start
, start_addr
);
1325 if (start
>= vma
->vm_end
)
1327 end
= min(vma
->vm_end
, end_addr
);
1328 if (end
<= vma
->vm_start
)
1332 uprobe_munmap(vma
, start
, end
);
1334 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1335 untrack_pfn(vma
, 0, 0);
1338 if (unlikely(is_vm_hugetlb_page(vma
))) {
1340 * It is undesirable to test vma->vm_file as it
1341 * should be non-null for valid hugetlb area.
1342 * However, vm_file will be NULL in the error
1343 * cleanup path of do_mmap_pgoff. When
1344 * hugetlbfs ->mmap method fails,
1345 * do_mmap_pgoff() nullifies vma->vm_file
1346 * before calling this function to clean up.
1347 * Since no pte has actually been setup, it is
1348 * safe to do nothing in this case.
1351 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
1352 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1353 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
1356 unmap_page_range(tlb
, vma
, start
, end
, details
);
1361 * unmap_vmas - unmap a range of memory covered by a list of vma's
1362 * @tlb: address of the caller's struct mmu_gather
1363 * @vma: the starting vma
1364 * @start_addr: virtual address at which to start unmapping
1365 * @end_addr: virtual address at which to end unmapping
1367 * Unmap all pages in the vma list.
1369 * Only addresses between `start' and `end' will be unmapped.
1371 * The VMA list must be sorted in ascending virtual address order.
1373 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1374 * range after unmap_vmas() returns. So the only responsibility here is to
1375 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1376 * drops the lock and schedules.
1378 void unmap_vmas(struct mmu_gather
*tlb
,
1379 struct vm_area_struct
*vma
, unsigned long start_addr
,
1380 unsigned long end_addr
)
1382 struct mm_struct
*mm
= vma
->vm_mm
;
1384 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1385 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1386 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1387 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1391 * zap_page_range - remove user pages in a given range
1392 * @vma: vm_area_struct holding the applicable pages
1393 * @start: starting address of pages to zap
1394 * @size: number of bytes to zap
1395 * @details: details of nonlinear truncation or shared cache invalidation
1397 * Caller must protect the VMA list
1399 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1400 unsigned long size
, struct zap_details
*details
)
1402 struct mm_struct
*mm
= vma
->vm_mm
;
1403 struct mmu_gather tlb
;
1404 unsigned long end
= start
+ size
;
1407 tlb_gather_mmu(&tlb
, mm
, 0);
1408 update_hiwater_rss(mm
);
1409 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1410 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
)
1411 unmap_single_vma(&tlb
, vma
, start
, end
, details
);
1412 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1413 tlb_finish_mmu(&tlb
, start
, end
);
1417 * zap_page_range_single - remove user pages in a given range
1418 * @vma: vm_area_struct holding the applicable pages
1419 * @address: starting address of pages to zap
1420 * @size: number of bytes to zap
1421 * @details: details of nonlinear truncation or shared cache invalidation
1423 * The range must fit into one VMA.
1425 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1426 unsigned long size
, struct zap_details
*details
)
1428 struct mm_struct
*mm
= vma
->vm_mm
;
1429 struct mmu_gather tlb
;
1430 unsigned long end
= address
+ size
;
1433 tlb_gather_mmu(&tlb
, mm
, 0);
1434 update_hiwater_rss(mm
);
1435 mmu_notifier_invalidate_range_start(mm
, address
, end
);
1436 unmap_single_vma(&tlb
, vma
, address
, end
, details
);
1437 mmu_notifier_invalidate_range_end(mm
, address
, end
);
1438 tlb_finish_mmu(&tlb
, address
, end
);
1442 * zap_vma_ptes - remove ptes mapping the vma
1443 * @vma: vm_area_struct holding ptes to be zapped
1444 * @address: starting address of pages to zap
1445 * @size: number of bytes to zap
1447 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1449 * The entire address range must be fully contained within the vma.
1451 * Returns 0 if successful.
1453 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1456 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1457 !(vma
->vm_flags
& VM_PFNMAP
))
1459 zap_page_range_single(vma
, address
, size
, NULL
);
1462 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1465 * follow_page - look up a page descriptor from a user-virtual address
1466 * @vma: vm_area_struct mapping @address
1467 * @address: virtual address to look up
1468 * @flags: flags modifying lookup behaviour
1470 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1472 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1473 * an error pointer if there is a mapping to something not represented
1474 * by a page descriptor (see also vm_normal_page()).
1476 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
1485 struct mm_struct
*mm
= vma
->vm_mm
;
1487 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
1488 if (!IS_ERR(page
)) {
1489 BUG_ON(flags
& FOLL_GET
);
1494 pgd
= pgd_offset(mm
, address
);
1495 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
1498 pud
= pud_offset(pgd
, address
);
1501 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
1502 BUG_ON(flags
& FOLL_GET
);
1503 page
= follow_huge_pud(mm
, address
, pud
, flags
& FOLL_WRITE
);
1506 if (unlikely(pud_bad(*pud
)))
1509 pmd
= pmd_offset(pud
, address
);
1512 if (pmd_huge(*pmd
) && vma
->vm_flags
& VM_HUGETLB
) {
1513 BUG_ON(flags
& FOLL_GET
);
1514 page
= follow_huge_pmd(mm
, address
, pmd
, flags
& FOLL_WRITE
);
1517 if ((flags
& FOLL_NUMA
) && pmd_numa(*pmd
))
1519 if (pmd_trans_huge(*pmd
)) {
1520 if (flags
& FOLL_SPLIT
) {
1521 split_huge_page_pmd(vma
, address
, pmd
);
1522 goto split_fallthrough
;
1524 spin_lock(&mm
->page_table_lock
);
1525 if (likely(pmd_trans_huge(*pmd
))) {
1526 if (unlikely(pmd_trans_splitting(*pmd
))) {
1527 spin_unlock(&mm
->page_table_lock
);
1528 wait_split_huge_page(vma
->anon_vma
, pmd
);
1530 page
= follow_trans_huge_pmd(vma
, address
,
1532 spin_unlock(&mm
->page_table_lock
);
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
))
1548 if ((flags
& FOLL_NUMA
) && pte_numa(pte
))
1550 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
1553 page
= vm_normal_page(vma
, address
, pte
);
1554 if (unlikely(!page
)) {
1555 if ((flags
& FOLL_DUMP
) ||
1556 !is_zero_pfn(pte_pfn(pte
)))
1558 page
= pte_page(pte
);
1561 if (flags
& FOLL_GET
)
1562 get_page_foll(page
);
1563 if (flags
& FOLL_TOUCH
) {
1564 if ((flags
& FOLL_WRITE
) &&
1565 !pte_dirty(pte
) && !PageDirty(page
))
1566 set_page_dirty(page
);
1568 * pte_mkyoung() would be more correct here, but atomic care
1569 * is needed to avoid losing the dirty bit: it is easier to use
1570 * mark_page_accessed().
1572 mark_page_accessed(page
);
1574 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1576 * The preliminary mapping check is mainly to avoid the
1577 * pointless overhead of lock_page on the ZERO_PAGE
1578 * which might bounce very badly if there is contention.
1580 * If the page is already locked, we don't need to
1581 * handle it now - vmscan will handle it later if and
1582 * when it attempts to reclaim the page.
1584 if (page
->mapping
&& trylock_page(page
)) {
1585 lru_add_drain(); /* push cached pages to LRU */
1587 * Because we lock page here, and migration is
1588 * blocked by the pte's page reference, and we
1589 * know the page is still mapped, we don't even
1590 * need to check for file-cache page truncation.
1592 mlock_vma_page(page
);
1597 pte_unmap_unlock(ptep
, ptl
);
1602 pte_unmap_unlock(ptep
, ptl
);
1603 return ERR_PTR(-EFAULT
);
1606 pte_unmap_unlock(ptep
, ptl
);
1612 * When core dumping an enormous anonymous area that nobody
1613 * has touched so far, we don't want to allocate unnecessary pages or
1614 * page tables. Return error instead of NULL to skip handle_mm_fault,
1615 * then get_dump_page() will return NULL to leave a hole in the dump.
1616 * But we can only make this optimization where a hole would surely
1617 * be zero-filled if handle_mm_fault() actually did handle it.
1619 if ((flags
& FOLL_DUMP
) &&
1620 (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
1621 return ERR_PTR(-EFAULT
);
1625 static inline int stack_guard_page(struct vm_area_struct
*vma
, unsigned long addr
)
1627 return stack_guard_page_start(vma
, addr
) ||
1628 stack_guard_page_end(vma
, addr
+PAGE_SIZE
);
1632 * __get_user_pages() - pin user pages in memory
1633 * @tsk: task_struct of target task
1634 * @mm: mm_struct of target mm
1635 * @start: starting user address
1636 * @nr_pages: number of pages from start to pin
1637 * @gup_flags: flags modifying pin behaviour
1638 * @pages: array that receives pointers to the pages pinned.
1639 * Should be at least nr_pages long. Or NULL, if caller
1640 * only intends to ensure the pages are faulted in.
1641 * @vmas: array of pointers to vmas corresponding to each page.
1642 * Or NULL if the caller does not require them.
1643 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1645 * Returns number of pages pinned. This may be fewer than the number
1646 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1647 * were pinned, returns -errno. Each page returned must be released
1648 * with a put_page() call when it is finished with. vmas will only
1649 * remain valid while mmap_sem is held.
1651 * Must be called with mmap_sem held for read or write.
1653 * __get_user_pages walks a process's page tables and takes a reference to
1654 * each struct page that each user address corresponds to at a given
1655 * instant. That is, it takes the page that would be accessed if a user
1656 * thread accesses the given user virtual address at that instant.
1658 * This does not guarantee that the page exists in the user mappings when
1659 * __get_user_pages returns, and there may even be a completely different
1660 * page there in some cases (eg. if mmapped pagecache has been invalidated
1661 * and subsequently re faulted). However it does guarantee that the page
1662 * won't be freed completely. And mostly callers simply care that the page
1663 * contains data that was valid *at some point in time*. Typically, an IO
1664 * or similar operation cannot guarantee anything stronger anyway because
1665 * locks can't be held over the syscall boundary.
1667 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1668 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1669 * appropriate) must be called after the page is finished with, and
1670 * before put_page is called.
1672 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1673 * or mmap_sem contention, and if waiting is needed to pin all pages,
1674 * *@nonblocking will be set to 0.
1676 * In most cases, get_user_pages or get_user_pages_fast should be used
1677 * instead of __get_user_pages. __get_user_pages should be used only if
1678 * you need some special @gup_flags.
1680 int __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1681 unsigned long start
, int nr_pages
, unsigned int gup_flags
,
1682 struct page
**pages
, struct vm_area_struct
**vmas
,
1686 unsigned long vm_flags
;
1691 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
1694 * Require read or write permissions.
1695 * If FOLL_FORCE is set, we only require the "MAY" flags.
1697 vm_flags
= (gup_flags
& FOLL_WRITE
) ?
1698 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1699 vm_flags
&= (gup_flags
& FOLL_FORCE
) ?
1700 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1703 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1704 * would be called on PROT_NONE ranges. We must never invoke
1705 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1706 * page faults would unprotect the PROT_NONE ranges if
1707 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1708 * bitflag. So to avoid that, don't set FOLL_NUMA if
1709 * FOLL_FORCE is set.
1711 if (!(gup_flags
& FOLL_FORCE
))
1712 gup_flags
|= FOLL_NUMA
;
1717 struct vm_area_struct
*vma
;
1719 vma
= find_extend_vma(mm
, start
);
1720 if (!vma
&& in_gate_area(mm
, start
)) {
1721 unsigned long pg
= start
& PAGE_MASK
;
1727 /* user gate pages are read-only */
1728 if (gup_flags
& FOLL_WRITE
)
1729 return i
? : -EFAULT
;
1731 pgd
= pgd_offset_k(pg
);
1733 pgd
= pgd_offset_gate(mm
, pg
);
1734 BUG_ON(pgd_none(*pgd
));
1735 pud
= pud_offset(pgd
, pg
);
1736 BUG_ON(pud_none(*pud
));
1737 pmd
= pmd_offset(pud
, pg
);
1739 return i
? : -EFAULT
;
1740 VM_BUG_ON(pmd_trans_huge(*pmd
));
1741 pte
= pte_offset_map(pmd
, pg
);
1742 if (pte_none(*pte
)) {
1744 return i
? : -EFAULT
;
1746 vma
= get_gate_vma(mm
);
1750 page
= vm_normal_page(vma
, start
, *pte
);
1752 if (!(gup_flags
& FOLL_DUMP
) &&
1753 is_zero_pfn(pte_pfn(*pte
)))
1754 page
= pte_page(*pte
);
1757 return i
? : -EFAULT
;
1768 (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1769 !(vm_flags
& vma
->vm_flags
))
1770 return i
? : -EFAULT
;
1772 if (is_vm_hugetlb_page(vma
)) {
1773 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1774 &start
, &nr_pages
, i
, gup_flags
);
1780 unsigned int foll_flags
= gup_flags
;
1783 * If we have a pending SIGKILL, don't keep faulting
1784 * pages and potentially allocating memory.
1786 if (unlikely(fatal_signal_pending(current
)))
1787 return i
? i
: -ERESTARTSYS
;
1790 while (!(page
= follow_page(vma
, start
, foll_flags
))) {
1792 unsigned int fault_flags
= 0;
1794 /* For mlock, just skip the stack guard page. */
1795 if (foll_flags
& FOLL_MLOCK
) {
1796 if (stack_guard_page(vma
, start
))
1799 if (foll_flags
& FOLL_WRITE
)
1800 fault_flags
|= FAULT_FLAG_WRITE
;
1802 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
1803 if (foll_flags
& FOLL_NOWAIT
)
1804 fault_flags
|= (FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
);
1806 ret
= handle_mm_fault(mm
, vma
, start
,
1809 if (ret
& VM_FAULT_ERROR
) {
1810 if (ret
& VM_FAULT_OOM
)
1811 return i
? i
: -ENOMEM
;
1812 if (ret
& (VM_FAULT_HWPOISON
|
1813 VM_FAULT_HWPOISON_LARGE
)) {
1816 else if (gup_flags
& FOLL_HWPOISON
)
1821 if (ret
& VM_FAULT_SIGBUS
)
1822 return i
? i
: -EFAULT
;
1827 if (ret
& VM_FAULT_MAJOR
)
1833 if (ret
& VM_FAULT_RETRY
) {
1840 * The VM_FAULT_WRITE bit tells us that
1841 * do_wp_page has broken COW when necessary,
1842 * even if maybe_mkwrite decided not to set
1843 * pte_write. We can thus safely do subsequent
1844 * page lookups as if they were reads. But only
1845 * do so when looping for pte_write is futile:
1846 * in some cases userspace may also be wanting
1847 * to write to the gotten user page, which a
1848 * read fault here might prevent (a readonly
1849 * page might get reCOWed by userspace write).
1851 if ((ret
& VM_FAULT_WRITE
) &&
1852 !(vma
->vm_flags
& VM_WRITE
))
1853 foll_flags
&= ~FOLL_WRITE
;
1858 return i
? i
: PTR_ERR(page
);
1862 flush_anon_page(vma
, page
, start
);
1863 flush_dcache_page(page
);
1871 } while (nr_pages
&& start
< vma
->vm_end
);
1875 EXPORT_SYMBOL(__get_user_pages
);
1878 * fixup_user_fault() - manually resolve a user page fault
1879 * @tsk: the task_struct to use for page fault accounting, or
1880 * NULL if faults are not to be recorded.
1881 * @mm: mm_struct of target mm
1882 * @address: user address
1883 * @fault_flags:flags to pass down to handle_mm_fault()
1885 * This is meant to be called in the specific scenario where for locking reasons
1886 * we try to access user memory in atomic context (within a pagefault_disable()
1887 * section), this returns -EFAULT, and we want to resolve the user fault before
1890 * Typically this is meant to be used by the futex code.
1892 * The main difference with get_user_pages() is that this function will
1893 * unconditionally call handle_mm_fault() which will in turn perform all the
1894 * necessary SW fixup of the dirty and young bits in the PTE, while
1895 * handle_mm_fault() only guarantees to update these in the struct page.
1897 * This is important for some architectures where those bits also gate the
1898 * access permission to the page because they are maintained in software. On
1899 * such architectures, gup() will not be enough to make a subsequent access
1902 * This should be called with the mm_sem held for read.
1904 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
1905 unsigned long address
, unsigned int fault_flags
)
1907 struct vm_area_struct
*vma
;
1910 vma
= find_extend_vma(mm
, address
);
1911 if (!vma
|| address
< vma
->vm_start
)
1914 ret
= handle_mm_fault(mm
, vma
, address
, fault_flags
);
1915 if (ret
& VM_FAULT_ERROR
) {
1916 if (ret
& VM_FAULT_OOM
)
1918 if (ret
& (VM_FAULT_HWPOISON
| VM_FAULT_HWPOISON_LARGE
))
1920 if (ret
& VM_FAULT_SIGBUS
)
1925 if (ret
& VM_FAULT_MAJOR
)
1934 * get_user_pages() - pin user pages in memory
1935 * @tsk: the task_struct to use for page fault accounting, or
1936 * NULL if faults are not to be recorded.
1937 * @mm: mm_struct of target mm
1938 * @start: starting user address
1939 * @nr_pages: number of pages from start to pin
1940 * @write: whether pages will be written to by the caller
1941 * @force: whether to force write access even if user mapping is
1942 * readonly. This will result in the page being COWed even
1943 * in MAP_SHARED mappings. You do not want this.
1944 * @pages: array that receives pointers to the pages pinned.
1945 * Should be at least nr_pages long. Or NULL, if caller
1946 * only intends to ensure the pages are faulted in.
1947 * @vmas: array of pointers to vmas corresponding to each page.
1948 * Or NULL if the caller does not require them.
1950 * Returns number of pages pinned. This may be fewer than the number
1951 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1952 * were pinned, returns -errno. Each page returned must be released
1953 * with a put_page() call when it is finished with. vmas will only
1954 * remain valid while mmap_sem is held.
1956 * Must be called with mmap_sem held for read or write.
1958 * get_user_pages walks a process's page tables and takes a reference to
1959 * each struct page that each user address corresponds to at a given
1960 * instant. That is, it takes the page that would be accessed if a user
1961 * thread accesses the given user virtual address at that instant.
1963 * This does not guarantee that the page exists in the user mappings when
1964 * get_user_pages returns, and there may even be a completely different
1965 * page there in some cases (eg. if mmapped pagecache has been invalidated
1966 * and subsequently re faulted). However it does guarantee that the page
1967 * won't be freed completely. And mostly callers simply care that the page
1968 * contains data that was valid *at some point in time*. Typically, an IO
1969 * or similar operation cannot guarantee anything stronger anyway because
1970 * locks can't be held over the syscall boundary.
1972 * If write=0, the page must not be written to. If the page is written to,
1973 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1974 * after the page is finished with, and before put_page is called.
1976 * get_user_pages is typically used for fewer-copy IO operations, to get a
1977 * handle on the memory by some means other than accesses via the user virtual
1978 * addresses. The pages may be submitted for DMA to devices or accessed via
1979 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1980 * use the correct cache flushing APIs.
1982 * See also get_user_pages_fast, for performance critical applications.
1984 int get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1985 unsigned long start
, int nr_pages
, int write
, int force
,
1986 struct page
**pages
, struct vm_area_struct
**vmas
)
1988 int flags
= FOLL_TOUCH
;
1993 flags
|= FOLL_WRITE
;
1995 flags
|= FOLL_FORCE
;
1997 return __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
, vmas
,
2000 EXPORT_SYMBOL(get_user_pages
);
2003 * get_dump_page() - pin user page in memory while writing it to core dump
2004 * @addr: user address
2006 * Returns struct page pointer of user page pinned for dump,
2007 * to be freed afterwards by page_cache_release() or put_page().
2009 * Returns NULL on any kind of failure - a hole must then be inserted into
2010 * the corefile, to preserve alignment with its headers; and also returns
2011 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2012 * allowing a hole to be left in the corefile to save diskspace.
2014 * Called without mmap_sem, but after all other threads have been killed.
2016 #ifdef CONFIG_ELF_CORE
2017 struct page
*get_dump_page(unsigned long addr
)
2019 struct vm_area_struct
*vma
;
2022 if (__get_user_pages(current
, current
->mm
, addr
, 1,
2023 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
2026 flush_cache_page(vma
, addr
, page_to_pfn(page
));
2029 #endif /* CONFIG_ELF_CORE */
2031 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
2034 pgd_t
* pgd
= pgd_offset(mm
, addr
);
2035 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
2037 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
2039 VM_BUG_ON(pmd_trans_huge(*pmd
));
2040 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
2047 * This is the old fallback for page remapping.
2049 * For historical reasons, it only allows reserved pages. Only
2050 * old drivers should use this, and they needed to mark their
2051 * pages reserved for the old functions anyway.
2053 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2054 struct page
*page
, pgprot_t prot
)
2056 struct mm_struct
*mm
= vma
->vm_mm
;
2065 flush_dcache_page(page
);
2066 pte
= get_locked_pte(mm
, addr
, &ptl
);
2070 if (!pte_none(*pte
))
2073 /* Ok, finally just insert the thing.. */
2075 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
2076 page_add_file_rmap(page
);
2077 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
2080 pte_unmap_unlock(pte
, ptl
);
2083 pte_unmap_unlock(pte
, ptl
);
2089 * vm_insert_page - insert single page into user vma
2090 * @vma: user vma to map to
2091 * @addr: target user address of this page
2092 * @page: source kernel page
2094 * This allows drivers to insert individual pages they've allocated
2097 * The page has to be a nice clean _individual_ kernel allocation.
2098 * If you allocate a compound page, you need to have marked it as
2099 * such (__GFP_COMP), or manually just split the page up yourself
2100 * (see split_page()).
2102 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2103 * took an arbitrary page protection parameter. This doesn't allow
2104 * that. Your vma protection will have to be set up correctly, which
2105 * means that if you want a shared writable mapping, you'd better
2106 * ask for a shared writable mapping!
2108 * The page does not need to be reserved.
2110 * Usually this function is called from f_op->mmap() handler
2111 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2112 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2113 * function from other places, for example from page-fault handler.
2115 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2118 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2120 if (!page_count(page
))
2122 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
2123 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
2124 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
2125 vma
->vm_flags
|= VM_MIXEDMAP
;
2127 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2129 EXPORT_SYMBOL(vm_insert_page
);
2131 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2132 unsigned long pfn
, pgprot_t prot
)
2134 struct mm_struct
*mm
= vma
->vm_mm
;
2140 pte
= get_locked_pte(mm
, addr
, &ptl
);
2144 if (!pte_none(*pte
))
2147 /* Ok, finally just insert the thing.. */
2148 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
2149 set_pte_at(mm
, addr
, pte
, entry
);
2150 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
2154 pte_unmap_unlock(pte
, ptl
);
2160 * vm_insert_pfn - insert single pfn into user vma
2161 * @vma: user vma to map to
2162 * @addr: target user address of this page
2163 * @pfn: source kernel pfn
2165 * Similar to vm_insert_page, this allows drivers to insert individual pages
2166 * they've allocated into a user vma. Same comments apply.
2168 * This function should only be called from a vm_ops->fault handler, and
2169 * in that case the handler should return NULL.
2171 * vma cannot be a COW mapping.
2173 * As this is called only for pages that do not currently exist, we
2174 * do not need to flush old virtual caches or the TLB.
2176 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2180 pgprot_t pgprot
= vma
->vm_page_prot
;
2182 * Technically, architectures with pte_special can avoid all these
2183 * restrictions (same for remap_pfn_range). However we would like
2184 * consistency in testing and feature parity among all, so we should
2185 * try to keep these invariants in place for everybody.
2187 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
2188 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
2189 (VM_PFNMAP
|VM_MIXEDMAP
));
2190 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
2191 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
2193 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2195 if (track_pfn_insert(vma
, &pgprot
, pfn
))
2198 ret
= insert_pfn(vma
, addr
, pfn
, pgprot
);
2202 EXPORT_SYMBOL(vm_insert_pfn
);
2204 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2207 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
2209 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2213 * If we don't have pte special, then we have to use the pfn_valid()
2214 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2215 * refcount the page if pfn_valid is true (hence insert_page rather
2216 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2217 * without pte special, it would there be refcounted as a normal page.
2219 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
2222 page
= pfn_to_page(pfn
);
2223 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2225 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
2227 EXPORT_SYMBOL(vm_insert_mixed
);
2230 * maps a range of physical memory into the requested pages. the old
2231 * mappings are removed. any references to nonexistent pages results
2232 * in null mappings (currently treated as "copy-on-access")
2234 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2235 unsigned long addr
, unsigned long end
,
2236 unsigned long pfn
, pgprot_t prot
)
2241 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2244 arch_enter_lazy_mmu_mode();
2246 BUG_ON(!pte_none(*pte
));
2247 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2249 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2250 arch_leave_lazy_mmu_mode();
2251 pte_unmap_unlock(pte
- 1, ptl
);
2255 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2256 unsigned long addr
, unsigned long end
,
2257 unsigned long pfn
, pgprot_t prot
)
2262 pfn
-= addr
>> PAGE_SHIFT
;
2263 pmd
= pmd_alloc(mm
, pud
, addr
);
2266 VM_BUG_ON(pmd_trans_huge(*pmd
));
2268 next
= pmd_addr_end(addr
, end
);
2269 if (remap_pte_range(mm
, pmd
, addr
, next
,
2270 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
2272 } while (pmd
++, addr
= next
, addr
!= end
);
2276 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2277 unsigned long addr
, unsigned long end
,
2278 unsigned long pfn
, pgprot_t prot
)
2283 pfn
-= addr
>> PAGE_SHIFT
;
2284 pud
= pud_alloc(mm
, pgd
, addr
);
2288 next
= pud_addr_end(addr
, end
);
2289 if (remap_pmd_range(mm
, pud
, addr
, next
,
2290 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
2292 } while (pud
++, addr
= next
, addr
!= end
);
2297 * remap_pfn_range - remap kernel memory to userspace
2298 * @vma: user vma to map to
2299 * @addr: target user address to start at
2300 * @pfn: physical address of kernel memory
2301 * @size: size of map area
2302 * @prot: page protection flags for this mapping
2304 * Note: this is only safe if the mm semaphore is held when called.
2306 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2307 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2311 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2312 struct mm_struct
*mm
= vma
->vm_mm
;
2316 * Physically remapped pages are special. Tell the
2317 * rest of the world about it:
2318 * VM_IO tells people not to look at these pages
2319 * (accesses can have side effects).
2320 * VM_PFNMAP tells the core MM that the base pages are just
2321 * raw PFN mappings, and do not have a "struct page" associated
2324 * Disable vma merging and expanding with mremap().
2326 * Omit vma from core dump, even when VM_IO turned off.
2328 * There's a horrible special case to handle copy-on-write
2329 * behaviour that some programs depend on. We mark the "original"
2330 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2331 * See vm_normal_page() for details.
2333 if (is_cow_mapping(vma
->vm_flags
)) {
2334 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2336 vma
->vm_pgoff
= pfn
;
2339 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
2343 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
2345 BUG_ON(addr
>= end
);
2346 pfn
-= addr
>> PAGE_SHIFT
;
2347 pgd
= pgd_offset(mm
, addr
);
2348 flush_cache_range(vma
, addr
, end
);
2350 next
= pgd_addr_end(addr
, end
);
2351 err
= remap_pud_range(mm
, pgd
, addr
, next
,
2352 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2355 } while (pgd
++, addr
= next
, addr
!= end
);
2358 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
));
2362 EXPORT_SYMBOL(remap_pfn_range
);
2364 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2365 unsigned long addr
, unsigned long end
,
2366 pte_fn_t fn
, void *data
)
2371 spinlock_t
*uninitialized_var(ptl
);
2373 pte
= (mm
== &init_mm
) ?
2374 pte_alloc_kernel(pmd
, addr
) :
2375 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2379 BUG_ON(pmd_huge(*pmd
));
2381 arch_enter_lazy_mmu_mode();
2383 token
= pmd_pgtable(*pmd
);
2386 err
= fn(pte
++, token
, addr
, data
);
2389 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2391 arch_leave_lazy_mmu_mode();
2394 pte_unmap_unlock(pte
-1, ptl
);
2398 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2399 unsigned long addr
, unsigned long end
,
2400 pte_fn_t fn
, void *data
)
2406 BUG_ON(pud_huge(*pud
));
2408 pmd
= pmd_alloc(mm
, pud
, addr
);
2412 next
= pmd_addr_end(addr
, end
);
2413 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
2416 } while (pmd
++, addr
= next
, addr
!= end
);
2420 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2421 unsigned long addr
, unsigned long end
,
2422 pte_fn_t fn
, void *data
)
2428 pud
= pud_alloc(mm
, pgd
, addr
);
2432 next
= pud_addr_end(addr
, end
);
2433 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
2436 } while (pud
++, addr
= next
, addr
!= end
);
2441 * Scan a region of virtual memory, filling in page tables as necessary
2442 * and calling a provided function on each leaf page table.
2444 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2445 unsigned long size
, pte_fn_t fn
, void *data
)
2449 unsigned long end
= addr
+ size
;
2452 BUG_ON(addr
>= end
);
2453 pgd
= pgd_offset(mm
, addr
);
2455 next
= pgd_addr_end(addr
, end
);
2456 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
2459 } while (pgd
++, addr
= next
, addr
!= end
);
2463 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2466 * handle_pte_fault chooses page fault handler according to an entry
2467 * which was read non-atomically. Before making any commitment, on
2468 * those architectures or configurations (e.g. i386 with PAE) which
2469 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2470 * must check under lock before unmapping the pte and proceeding
2471 * (but do_wp_page is only called after already making such a check;
2472 * and do_anonymous_page can safely check later on).
2474 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2475 pte_t
*page_table
, pte_t orig_pte
)
2478 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2479 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2480 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2482 same
= pte_same(*page_table
, orig_pte
);
2486 pte_unmap(page_table
);
2490 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
2493 * If the source page was a PFN mapping, we don't have
2494 * a "struct page" for it. We do a best-effort copy by
2495 * just copying from the original user address. If that
2496 * fails, we just zero-fill it. Live with it.
2498 if (unlikely(!src
)) {
2499 void *kaddr
= kmap_atomic(dst
);
2500 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
2503 * This really shouldn't fail, because the page is there
2504 * in the page tables. But it might just be unreadable,
2505 * in which case we just give up and fill the result with
2508 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
2510 kunmap_atomic(kaddr
);
2511 flush_dcache_page(dst
);
2513 copy_user_highpage(dst
, src
, va
, vma
);
2517 * This routine handles present pages, when users try to write
2518 * to a shared page. It is done by copying the page to a new address
2519 * and decrementing the shared-page counter for the old page.
2521 * Note that this routine assumes that the protection checks have been
2522 * done by the caller (the low-level page fault routine in most cases).
2523 * Thus we can safely just mark it writable once we've done any necessary
2526 * We also mark the page dirty at this point even though the page will
2527 * change only once the write actually happens. This avoids a few races,
2528 * and potentially makes it more efficient.
2530 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2531 * but allow concurrent faults), with pte both mapped and locked.
2532 * We return with mmap_sem still held, but pte unmapped and unlocked.
2534 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2535 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2536 spinlock_t
*ptl
, pte_t orig_pte
)
2539 struct page
*old_page
, *new_page
= NULL
;
2542 int page_mkwrite
= 0;
2543 struct page
*dirty_page
= NULL
;
2544 unsigned long mmun_start
= 0; /* For mmu_notifiers */
2545 unsigned long mmun_end
= 0; /* For mmu_notifiers */
2547 old_page
= vm_normal_page(vma
, address
, orig_pte
);
2550 * VM_MIXEDMAP !pfn_valid() case
2552 * We should not cow pages in a shared writeable mapping.
2553 * Just mark the pages writable as we can't do any dirty
2554 * accounting on raw pfn maps.
2556 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2557 (VM_WRITE
|VM_SHARED
))
2563 * Take out anonymous pages first, anonymous shared vmas are
2564 * not dirty accountable.
2566 if (PageAnon(old_page
) && !PageKsm(old_page
)) {
2567 if (!trylock_page(old_page
)) {
2568 page_cache_get(old_page
);
2569 pte_unmap_unlock(page_table
, ptl
);
2570 lock_page(old_page
);
2571 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2573 if (!pte_same(*page_table
, orig_pte
)) {
2574 unlock_page(old_page
);
2577 page_cache_release(old_page
);
2579 if (reuse_swap_page(old_page
)) {
2581 * The page is all ours. Move it to our anon_vma so
2582 * the rmap code will not search our parent or siblings.
2583 * Protected against the rmap code by the page lock.
2585 page_move_anon_rmap(old_page
, vma
, address
);
2586 unlock_page(old_page
);
2589 unlock_page(old_page
);
2590 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2591 (VM_WRITE
|VM_SHARED
))) {
2593 * Only catch write-faults on shared writable pages,
2594 * read-only shared pages can get COWed by
2595 * get_user_pages(.write=1, .force=1).
2597 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2598 struct vm_fault vmf
;
2601 vmf
.virtual_address
= (void __user
*)(address
&
2603 vmf
.pgoff
= old_page
->index
;
2604 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2605 vmf
.page
= old_page
;
2608 * Notify the address space that the page is about to
2609 * become writable so that it can prohibit this or wait
2610 * for the page to get into an appropriate state.
2612 * We do this without the lock held, so that it can
2613 * sleep if it needs to.
2615 page_cache_get(old_page
);
2616 pte_unmap_unlock(page_table
, ptl
);
2618 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
2620 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
2622 goto unwritable_page
;
2624 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
2625 lock_page(old_page
);
2626 if (!old_page
->mapping
) {
2627 ret
= 0; /* retry the fault */
2628 unlock_page(old_page
);
2629 goto unwritable_page
;
2632 VM_BUG_ON(!PageLocked(old_page
));
2635 * Since we dropped the lock we need to revalidate
2636 * the PTE as someone else may have changed it. If
2637 * they did, we just return, as we can count on the
2638 * MMU to tell us if they didn't also make it writable.
2640 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2642 if (!pte_same(*page_table
, orig_pte
)) {
2643 unlock_page(old_page
);
2649 dirty_page
= old_page
;
2650 get_page(dirty_page
);
2653 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2654 entry
= pte_mkyoung(orig_pte
);
2655 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2656 if (ptep_set_access_flags(vma
, address
, page_table
, entry
,1))
2657 update_mmu_cache(vma
, address
, page_table
);
2658 pte_unmap_unlock(page_table
, ptl
);
2659 ret
|= VM_FAULT_WRITE
;
2665 * Yes, Virginia, this is actually required to prevent a race
2666 * with clear_page_dirty_for_io() from clearing the page dirty
2667 * bit after it clear all dirty ptes, but before a racing
2668 * do_wp_page installs a dirty pte.
2670 * __do_fault is protected similarly.
2672 if (!page_mkwrite
) {
2673 wait_on_page_locked(dirty_page
);
2674 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2675 /* file_update_time outside page_lock */
2677 file_update_time(vma
->vm_file
);
2679 put_page(dirty_page
);
2681 struct address_space
*mapping
= dirty_page
->mapping
;
2683 set_page_dirty(dirty_page
);
2684 unlock_page(dirty_page
);
2685 page_cache_release(dirty_page
);
2688 * Some device drivers do not set page.mapping
2689 * but still dirty their pages
2691 balance_dirty_pages_ratelimited(mapping
);
2699 * Ok, we need to copy. Oh, well..
2701 page_cache_get(old_page
);
2703 pte_unmap_unlock(page_table
, ptl
);
2705 if (unlikely(anon_vma_prepare(vma
)))
2708 if (is_zero_pfn(pte_pfn(orig_pte
))) {
2709 new_page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2713 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2716 cow_user_page(new_page
, old_page
, address
, vma
);
2718 __SetPageUptodate(new_page
);
2720 if (mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
))
2723 mmun_start
= address
& PAGE_MASK
;
2724 mmun_end
= mmun_start
+ PAGE_SIZE
;
2725 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2728 * Re-check the pte - we dropped the lock
2730 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2731 if (likely(pte_same(*page_table
, orig_pte
))) {
2733 if (!PageAnon(old_page
)) {
2734 dec_mm_counter_fast(mm
, MM_FILEPAGES
);
2735 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2738 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2739 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2740 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2741 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2743 * Clear the pte entry and flush it first, before updating the
2744 * pte with the new entry. This will avoid a race condition
2745 * seen in the presence of one thread doing SMC and another
2748 ptep_clear_flush(vma
, address
, page_table
);
2749 page_add_new_anon_rmap(new_page
, vma
, address
);
2751 * We call the notify macro here because, when using secondary
2752 * mmu page tables (such as kvm shadow page tables), we want the
2753 * new page to be mapped directly into the secondary page table.
2755 set_pte_at_notify(mm
, address
, page_table
, entry
);
2756 update_mmu_cache(vma
, address
, page_table
);
2759 * Only after switching the pte to the new page may
2760 * we remove the mapcount here. Otherwise another
2761 * process may come and find the rmap count decremented
2762 * before the pte is switched to the new page, and
2763 * "reuse" the old page writing into it while our pte
2764 * here still points into it and can be read by other
2767 * The critical issue is to order this
2768 * page_remove_rmap with the ptp_clear_flush above.
2769 * Those stores are ordered by (if nothing else,)
2770 * the barrier present in the atomic_add_negative
2771 * in page_remove_rmap.
2773 * Then the TLB flush in ptep_clear_flush ensures that
2774 * no process can access the old page before the
2775 * decremented mapcount is visible. And the old page
2776 * cannot be reused until after the decremented
2777 * mapcount is visible. So transitively, TLBs to
2778 * old page will be flushed before it can be reused.
2780 page_remove_rmap(old_page
);
2783 /* Free the old page.. */
2784 new_page
= old_page
;
2785 ret
|= VM_FAULT_WRITE
;
2787 mem_cgroup_uncharge_page(new_page
);
2790 page_cache_release(new_page
);
2792 pte_unmap_unlock(page_table
, ptl
);
2793 if (mmun_end
> mmun_start
)
2794 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2797 * Don't let another task, with possibly unlocked vma,
2798 * keep the mlocked page.
2800 if ((ret
& VM_FAULT_WRITE
) && (vma
->vm_flags
& VM_LOCKED
)) {
2801 lock_page(old_page
); /* LRU manipulation */
2802 munlock_vma_page(old_page
);
2803 unlock_page(old_page
);
2805 page_cache_release(old_page
);
2809 page_cache_release(new_page
);
2812 page_cache_release(old_page
);
2813 return VM_FAULT_OOM
;
2816 page_cache_release(old_page
);
2820 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2821 unsigned long start_addr
, unsigned long end_addr
,
2822 struct zap_details
*details
)
2824 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2827 static inline void unmap_mapping_range_tree(struct rb_root
*root
,
2828 struct zap_details
*details
)
2830 struct vm_area_struct
*vma
;
2831 pgoff_t vba
, vea
, zba
, zea
;
2833 vma_interval_tree_foreach(vma
, root
,
2834 details
->first_index
, details
->last_index
) {
2836 vba
= vma
->vm_pgoff
;
2837 vea
= vba
+ ((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) - 1;
2838 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2839 zba
= details
->first_index
;
2842 zea
= details
->last_index
;
2846 unmap_mapping_range_vma(vma
,
2847 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2848 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2853 static inline void unmap_mapping_range_list(struct list_head
*head
,
2854 struct zap_details
*details
)
2856 struct vm_area_struct
*vma
;
2859 * In nonlinear VMAs there is no correspondence between virtual address
2860 * offset and file offset. So we must perform an exhaustive search
2861 * across *all* the pages in each nonlinear VMA, not just the pages
2862 * whose virtual address lies outside the file truncation point.
2864 list_for_each_entry(vma
, head
, shared
.nonlinear
) {
2865 details
->nonlinear_vma
= vma
;
2866 unmap_mapping_range_vma(vma
, vma
->vm_start
, vma
->vm_end
, details
);
2871 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2872 * @mapping: the address space containing mmaps to be unmapped.
2873 * @holebegin: byte in first page to unmap, relative to the start of
2874 * the underlying file. This will be rounded down to a PAGE_SIZE
2875 * boundary. Note that this is different from truncate_pagecache(), which
2876 * must keep the partial page. In contrast, we must get rid of
2878 * @holelen: size of prospective hole in bytes. This will be rounded
2879 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2881 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2882 * but 0 when invalidating pagecache, don't throw away private data.
2884 void unmap_mapping_range(struct address_space
*mapping
,
2885 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2887 struct zap_details details
;
2888 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2889 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2891 /* Check for overflow. */
2892 if (sizeof(holelen
) > sizeof(hlen
)) {
2894 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2895 if (holeend
& ~(long long)ULONG_MAX
)
2896 hlen
= ULONG_MAX
- hba
+ 1;
2899 details
.check_mapping
= even_cows
? NULL
: mapping
;
2900 details
.nonlinear_vma
= NULL
;
2901 details
.first_index
= hba
;
2902 details
.last_index
= hba
+ hlen
- 1;
2903 if (details
.last_index
< details
.first_index
)
2904 details
.last_index
= ULONG_MAX
;
2907 mutex_lock(&mapping
->i_mmap_mutex
);
2908 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
)))
2909 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2910 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
2911 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
2912 mutex_unlock(&mapping
->i_mmap_mutex
);
2914 EXPORT_SYMBOL(unmap_mapping_range
);
2917 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2918 * but allow concurrent faults), and pte mapped but not yet locked.
2919 * We return with mmap_sem still held, but pte unmapped and unlocked.
2921 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2922 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2923 unsigned int flags
, pte_t orig_pte
)
2926 struct page
*page
, *swapcache
= NULL
;
2930 struct mem_cgroup
*ptr
;
2934 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2937 entry
= pte_to_swp_entry(orig_pte
);
2938 if (unlikely(non_swap_entry(entry
))) {
2939 if (is_migration_entry(entry
)) {
2940 migration_entry_wait(mm
, pmd
, address
);
2941 } else if (is_hwpoison_entry(entry
)) {
2942 ret
= VM_FAULT_HWPOISON
;
2944 print_bad_pte(vma
, address
, orig_pte
, NULL
);
2945 ret
= VM_FAULT_SIGBUS
;
2949 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2950 page
= lookup_swap_cache(entry
);
2952 page
= swapin_readahead(entry
,
2953 GFP_HIGHUSER_MOVABLE
, vma
, address
);
2956 * Back out if somebody else faulted in this pte
2957 * while we released the pte lock.
2959 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2960 if (likely(pte_same(*page_table
, orig_pte
)))
2962 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2966 /* Had to read the page from swap area: Major fault */
2967 ret
= VM_FAULT_MAJOR
;
2968 count_vm_event(PGMAJFAULT
);
2969 mem_cgroup_count_vm_event(mm
, PGMAJFAULT
);
2970 } else if (PageHWPoison(page
)) {
2972 * hwpoisoned dirty swapcache pages are kept for killing
2973 * owner processes (which may be unknown at hwpoison time)
2975 ret
= VM_FAULT_HWPOISON
;
2976 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2980 locked
= lock_page_or_retry(page
, mm
, flags
);
2982 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2984 ret
|= VM_FAULT_RETRY
;
2989 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2990 * release the swapcache from under us. The page pin, and pte_same
2991 * test below, are not enough to exclude that. Even if it is still
2992 * swapcache, we need to check that the page's swap has not changed.
2994 if (unlikely(!PageSwapCache(page
) || page_private(page
) != entry
.val
))
2998 page
= ksm_might_need_to_copy(page
, vma
, address
);
2999 if (unlikely(!page
)) {
3005 if (page
== swapcache
)
3008 if (mem_cgroup_try_charge_swapin(mm
, page
, GFP_KERNEL
, &ptr
)) {
3014 * Back out if somebody else already faulted in this pte.
3016 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3017 if (unlikely(!pte_same(*page_table
, orig_pte
)))
3020 if (unlikely(!PageUptodate(page
))) {
3021 ret
= VM_FAULT_SIGBUS
;
3026 * The page isn't present yet, go ahead with the fault.
3028 * Be careful about the sequence of operations here.
3029 * To get its accounting right, reuse_swap_page() must be called
3030 * while the page is counted on swap but not yet in mapcount i.e.
3031 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3032 * must be called after the swap_free(), or it will never succeed.
3033 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3034 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3035 * in page->private. In this case, a record in swap_cgroup is silently
3036 * discarded at swap_free().
3039 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3040 dec_mm_counter_fast(mm
, MM_SWAPENTS
);
3041 pte
= mk_pte(page
, vma
->vm_page_prot
);
3042 if ((flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
)) {
3043 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3044 flags
&= ~FAULT_FLAG_WRITE
;
3045 ret
|= VM_FAULT_WRITE
;
3048 flush_icache_page(vma
, page
);
3049 set_pte_at(mm
, address
, page_table
, pte
);
3050 if (swapcache
) /* ksm created a completely new copy */
3051 page_add_new_anon_rmap(page
, vma
, address
);
3053 do_page_add_anon_rmap(page
, vma
, address
, exclusive
);
3054 /* It's better to call commit-charge after rmap is established */
3055 mem_cgroup_commit_charge_swapin(page
, ptr
);
3058 if (vm_swap_full() || (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
3059 try_to_free_swap(page
);
3063 * Hold the lock to avoid the swap entry to be reused
3064 * until we take the PT lock for the pte_same() check
3065 * (to avoid false positives from pte_same). For
3066 * further safety release the lock after the swap_free
3067 * so that the swap count won't change under a
3068 * parallel locked swapcache.
3070 unlock_page(swapcache
);
3071 page_cache_release(swapcache
);
3074 if (flags
& FAULT_FLAG_WRITE
) {
3075 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
3076 if (ret
& VM_FAULT_ERROR
)
3077 ret
&= VM_FAULT_ERROR
;
3081 /* No need to invalidate - it was non-present before */
3082 update_mmu_cache(vma
, address
, page_table
);
3084 pte_unmap_unlock(page_table
, ptl
);
3088 mem_cgroup_cancel_charge_swapin(ptr
);
3089 pte_unmap_unlock(page_table
, ptl
);
3093 page_cache_release(page
);
3095 unlock_page(swapcache
);
3096 page_cache_release(swapcache
);
3102 * This is like a special single-page "expand_{down|up}wards()",
3103 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3104 * doesn't hit another vma.
3106 static inline int check_stack_guard_page(struct vm_area_struct
*vma
, unsigned long address
)
3108 address
&= PAGE_MASK
;
3109 if ((vma
->vm_flags
& VM_GROWSDOWN
) && address
== vma
->vm_start
) {
3110 struct vm_area_struct
*prev
= vma
->vm_prev
;
3113 * Is there a mapping abutting this one below?
3115 * That's only ok if it's the same stack mapping
3116 * that has gotten split..
3118 if (prev
&& prev
->vm_end
== address
)
3119 return prev
->vm_flags
& VM_GROWSDOWN
? 0 : -ENOMEM
;
3121 expand_downwards(vma
, address
- PAGE_SIZE
);
3123 if ((vma
->vm_flags
& VM_GROWSUP
) && address
+ PAGE_SIZE
== vma
->vm_end
) {
3124 struct vm_area_struct
*next
= vma
->vm_next
;
3126 /* As VM_GROWSDOWN but s/below/above/ */
3127 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
)
3128 return next
->vm_flags
& VM_GROWSUP
? 0 : -ENOMEM
;
3130 expand_upwards(vma
, address
+ PAGE_SIZE
);
3136 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3137 * but allow concurrent faults), and pte mapped but not yet locked.
3138 * We return with mmap_sem still held, but pte unmapped and unlocked.
3140 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3141 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3148 pte_unmap(page_table
);
3150 /* Check if we need to add a guard page to the stack */
3151 if (check_stack_guard_page(vma
, address
) < 0)
3152 return VM_FAULT_SIGBUS
;
3154 /* Use the zero-page for reads */
3155 if (!(flags
& FAULT_FLAG_WRITE
)) {
3156 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(address
),
3157 vma
->vm_page_prot
));
3158 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3159 if (!pte_none(*page_table
))
3164 /* Allocate our own private page. */
3165 if (unlikely(anon_vma_prepare(vma
)))
3167 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
3170 __SetPageUptodate(page
);
3172 if (mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
))
3175 entry
= mk_pte(page
, vma
->vm_page_prot
);
3176 if (vma
->vm_flags
& VM_WRITE
)
3177 entry
= pte_mkwrite(pte_mkdirty(entry
));
3179 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3180 if (!pte_none(*page_table
))
3183 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3184 page_add_new_anon_rmap(page
, vma
, address
);
3186 set_pte_at(mm
, address
, page_table
, entry
);
3188 /* No need to invalidate - it was non-present before */
3189 update_mmu_cache(vma
, address
, page_table
);
3191 pte_unmap_unlock(page_table
, ptl
);
3194 mem_cgroup_uncharge_page(page
);
3195 page_cache_release(page
);
3198 page_cache_release(page
);
3200 return VM_FAULT_OOM
;
3204 * __do_fault() tries to create a new page mapping. It aggressively
3205 * tries to share with existing pages, but makes a separate copy if
3206 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3207 * the next page fault.
3209 * As this is called only for pages that do not currently exist, we
3210 * do not need to flush old virtual caches or the TLB.
3212 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3213 * but allow concurrent faults), and pte neither mapped nor locked.
3214 * We return with mmap_sem still held, but pte unmapped and unlocked.
3216 static int __do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3217 unsigned long address
, pmd_t
*pmd
,
3218 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3223 struct page
*cow_page
;
3226 struct page
*dirty_page
= NULL
;
3227 struct vm_fault vmf
;
3229 int page_mkwrite
= 0;
3232 * If we do COW later, allocate page befor taking lock_page()
3233 * on the file cache page. This will reduce lock holding time.
3235 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3237 if (unlikely(anon_vma_prepare(vma
)))
3238 return VM_FAULT_OOM
;
3240 cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
3242 return VM_FAULT_OOM
;
3244 if (mem_cgroup_newpage_charge(cow_page
, mm
, GFP_KERNEL
)) {
3245 page_cache_release(cow_page
);
3246 return VM_FAULT_OOM
;
3251 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
3256 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
3257 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
3261 if (unlikely(PageHWPoison(vmf
.page
))) {
3262 if (ret
& VM_FAULT_LOCKED
)
3263 unlock_page(vmf
.page
);
3264 ret
= VM_FAULT_HWPOISON
;
3269 * For consistency in subsequent calls, make the faulted page always
3272 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3273 lock_page(vmf
.page
);
3275 VM_BUG_ON(!PageLocked(vmf
.page
));
3278 * Should we do an early C-O-W break?
3281 if (flags
& FAULT_FLAG_WRITE
) {
3282 if (!(vma
->vm_flags
& VM_SHARED
)) {
3285 copy_user_highpage(page
, vmf
.page
, address
, vma
);
3286 __SetPageUptodate(page
);
3289 * If the page will be shareable, see if the backing
3290 * address space wants to know that the page is about
3291 * to become writable
3293 if (vma
->vm_ops
->page_mkwrite
) {
3297 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
3298 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
3300 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3302 goto unwritable_page
;
3304 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
3306 if (!page
->mapping
) {
3307 ret
= 0; /* retry the fault */
3309 goto unwritable_page
;
3312 VM_BUG_ON(!PageLocked(page
));
3319 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3322 * This silly early PAGE_DIRTY setting removes a race
3323 * due to the bad i386 page protection. But it's valid
3324 * for other architectures too.
3326 * Note that if FAULT_FLAG_WRITE is set, we either now have
3327 * an exclusive copy of the page, or this is a shared mapping,
3328 * so we can make it writable and dirty to avoid having to
3329 * handle that later.
3331 /* Only go through if we didn't race with anybody else... */
3332 if (likely(pte_same(*page_table
, orig_pte
))) {
3333 flush_icache_page(vma
, page
);
3334 entry
= mk_pte(page
, vma
->vm_page_prot
);
3335 if (flags
& FAULT_FLAG_WRITE
)
3336 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3338 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3339 page_add_new_anon_rmap(page
, vma
, address
);
3341 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
3342 page_add_file_rmap(page
);
3343 if (flags
& FAULT_FLAG_WRITE
) {
3345 get_page(dirty_page
);
3348 set_pte_at(mm
, address
, page_table
, entry
);
3350 /* no need to invalidate: a not-present page won't be cached */
3351 update_mmu_cache(vma
, address
, page_table
);
3354 mem_cgroup_uncharge_page(cow_page
);
3356 page_cache_release(page
);
3358 anon
= 1; /* no anon but release faulted_page */
3361 pte_unmap_unlock(page_table
, ptl
);
3364 struct address_space
*mapping
= page
->mapping
;
3367 if (set_page_dirty(dirty_page
))
3369 unlock_page(dirty_page
);
3370 put_page(dirty_page
);
3371 if ((dirtied
|| page_mkwrite
) && mapping
) {
3373 * Some device drivers do not set page.mapping but still
3376 balance_dirty_pages_ratelimited(mapping
);
3379 /* file_update_time outside page_lock */
3380 if (vma
->vm_file
&& !page_mkwrite
)
3381 file_update_time(vma
->vm_file
);
3383 unlock_page(vmf
.page
);
3385 page_cache_release(vmf
.page
);
3391 page_cache_release(page
);
3394 /* fs's fault handler get error */
3396 mem_cgroup_uncharge_page(cow_page
);
3397 page_cache_release(cow_page
);
3402 static int do_linear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3403 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3404 unsigned int flags
, pte_t orig_pte
)
3406 pgoff_t pgoff
= (((address
& PAGE_MASK
)
3407 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
3409 pte_unmap(page_table
);
3410 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3414 * Fault of a previously existing named mapping. Repopulate the pte
3415 * from the encoded file_pte if possible. This enables swappable
3418 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3419 * but allow concurrent faults), and pte mapped but not yet locked.
3420 * We return with mmap_sem still held, but pte unmapped and unlocked.
3422 static int do_nonlinear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3423 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3424 unsigned int flags
, pte_t orig_pte
)
3428 flags
|= FAULT_FLAG_NONLINEAR
;
3430 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
3433 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
))) {
3435 * Page table corrupted: show pte and kill process.
3437 print_bad_pte(vma
, address
, orig_pte
, NULL
);
3438 return VM_FAULT_SIGBUS
;
3441 pgoff
= pte_to_pgoff(orig_pte
);
3442 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3445 int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3446 unsigned long addr
, int current_nid
)
3450 count_vm_numa_event(NUMA_HINT_FAULTS
);
3451 if (current_nid
== numa_node_id())
3452 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3454 return mpol_misplaced(page
, vma
, addr
);
3457 int do_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3458 unsigned long addr
, pte_t pte
, pte_t
*ptep
, pmd_t
*pmd
)
3460 struct page
*page
= NULL
;
3462 int current_nid
= -1;
3464 bool migrated
= false;
3467 * The "pte" at this point cannot be used safely without
3468 * validation through pte_unmap_same(). It's of NUMA type but
3469 * the pfn may be screwed if the read is non atomic.
3471 * ptep_modify_prot_start is not called as this is clearing
3472 * the _PAGE_NUMA bit and it is not really expected that there
3473 * would be concurrent hardware modifications to the PTE.
3475 ptl
= pte_lockptr(mm
, pmd
);
3477 if (unlikely(!pte_same(*ptep
, pte
))) {
3478 pte_unmap_unlock(ptep
, ptl
);
3482 pte
= pte_mknonnuma(pte
);
3483 set_pte_at(mm
, addr
, ptep
, pte
);
3484 update_mmu_cache(vma
, addr
, ptep
);
3486 page
= vm_normal_page(vma
, addr
, pte
);
3488 pte_unmap_unlock(ptep
, ptl
);
3492 current_nid
= page_to_nid(page
);
3493 target_nid
= numa_migrate_prep(page
, vma
, addr
, current_nid
);
3494 pte_unmap_unlock(ptep
, ptl
);
3495 if (target_nid
== -1) {
3497 * Account for the fault against the current node if it not
3498 * being replaced regardless of where the page is located.
3500 current_nid
= numa_node_id();
3505 /* Migrate to the requested node */
3506 migrated
= migrate_misplaced_page(page
, target_nid
);
3508 current_nid
= target_nid
;
3511 if (current_nid
!= -1)
3512 task_numa_fault(current_nid
, 1, migrated
);
3516 /* NUMA hinting page fault entry point for regular pmds */
3517 #ifdef CONFIG_NUMA_BALANCING
3518 static int do_pmd_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3519 unsigned long addr
, pmd_t
*pmdp
)
3522 pte_t
*pte
, *orig_pte
;
3523 unsigned long _addr
= addr
& PMD_MASK
;
3524 unsigned long offset
;
3527 int local_nid
= numa_node_id();
3529 spin_lock(&mm
->page_table_lock
);
3531 if (pmd_numa(pmd
)) {
3532 set_pmd_at(mm
, _addr
, pmdp
, pmd_mknonnuma(pmd
));
3535 spin_unlock(&mm
->page_table_lock
);
3540 /* we're in a page fault so some vma must be in the range */
3542 BUG_ON(vma
->vm_start
>= _addr
+ PMD_SIZE
);
3543 offset
= max(_addr
, vma
->vm_start
) & ~PMD_MASK
;
3544 VM_BUG_ON(offset
>= PMD_SIZE
);
3545 orig_pte
= pte
= pte_offset_map_lock(mm
, pmdp
, _addr
, &ptl
);
3546 pte
+= offset
>> PAGE_SHIFT
;
3547 for (addr
= _addr
+ offset
; addr
< _addr
+ PMD_SIZE
; pte
++, addr
+= PAGE_SIZE
) {
3548 pte_t pteval
= *pte
;
3550 int curr_nid
= local_nid
;
3553 if (!pte_present(pteval
))
3555 if (!pte_numa(pteval
))
3557 if (addr
>= vma
->vm_end
) {
3558 vma
= find_vma(mm
, addr
);
3559 /* there's a pte present so there must be a vma */
3561 BUG_ON(addr
< vma
->vm_start
);
3563 if (pte_numa(pteval
)) {
3564 pteval
= pte_mknonnuma(pteval
);
3565 set_pte_at(mm
, addr
, pte
, pteval
);
3567 page
= vm_normal_page(vma
, addr
, pteval
);
3568 if (unlikely(!page
))
3570 /* only check non-shared pages */
3571 if (unlikely(page_mapcount(page
) != 1))
3575 * Note that the NUMA fault is later accounted to either
3576 * the node that is currently running or where the page is
3579 curr_nid
= local_nid
;
3580 target_nid
= numa_migrate_prep(page
, vma
, addr
,
3582 if (target_nid
== -1) {
3587 /* Migrate to the requested node */
3588 pte_unmap_unlock(pte
, ptl
);
3589 migrated
= migrate_misplaced_page(page
, target_nid
);
3591 curr_nid
= target_nid
;
3592 task_numa_fault(curr_nid
, 1, migrated
);
3594 pte
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
3596 pte_unmap_unlock(orig_pte
, ptl
);
3601 static int do_pmd_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3602 unsigned long addr
, pmd_t
*pmdp
)
3607 #endif /* CONFIG_NUMA_BALANCING */
3610 * These routines also need to handle stuff like marking pages dirty
3611 * and/or accessed for architectures that don't do it in hardware (most
3612 * RISC architectures). The early dirtying is also good on the i386.
3614 * There is also a hook called "update_mmu_cache()" that architectures
3615 * with external mmu caches can use to update those (ie the Sparc or
3616 * PowerPC hashed page tables that act as extended TLBs).
3618 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3619 * but allow concurrent faults), and pte mapped but not yet locked.
3620 * We return with mmap_sem still held, but pte unmapped and unlocked.
3622 int handle_pte_fault(struct mm_struct
*mm
,
3623 struct vm_area_struct
*vma
, unsigned long address
,
3624 pte_t
*pte
, pmd_t
*pmd
, unsigned int flags
)
3630 if (!pte_present(entry
)) {
3631 if (pte_none(entry
)) {
3633 if (likely(vma
->vm_ops
->fault
))
3634 return do_linear_fault(mm
, vma
, address
,
3635 pte
, pmd
, flags
, entry
);
3637 return do_anonymous_page(mm
, vma
, address
,
3640 if (pte_file(entry
))
3641 return do_nonlinear_fault(mm
, vma
, address
,
3642 pte
, pmd
, flags
, entry
);
3643 return do_swap_page(mm
, vma
, address
,
3644 pte
, pmd
, flags
, entry
);
3647 if (pte_numa(entry
))
3648 return do_numa_page(mm
, vma
, address
, entry
, pte
, pmd
);
3650 ptl
= pte_lockptr(mm
, pmd
);
3652 if (unlikely(!pte_same(*pte
, entry
)))
3654 if (flags
& FAULT_FLAG_WRITE
) {
3655 if (!pte_write(entry
))
3656 return do_wp_page(mm
, vma
, address
,
3657 pte
, pmd
, ptl
, entry
);
3658 entry
= pte_mkdirty(entry
);
3660 entry
= pte_mkyoung(entry
);
3661 if (ptep_set_access_flags(vma
, address
, pte
, entry
, flags
& FAULT_FLAG_WRITE
)) {
3662 update_mmu_cache(vma
, address
, pte
);
3665 * This is needed only for protection faults but the arch code
3666 * is not yet telling us if this is a protection fault or not.
3667 * This still avoids useless tlb flushes for .text page faults
3670 if (flags
& FAULT_FLAG_WRITE
)
3671 flush_tlb_fix_spurious_fault(vma
, address
);
3674 pte_unmap_unlock(pte
, ptl
);
3679 * By the time we get here, we already hold the mm semaphore
3681 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3682 unsigned long address
, unsigned int flags
)
3689 __set_current_state(TASK_RUNNING
);
3691 count_vm_event(PGFAULT
);
3692 mem_cgroup_count_vm_event(mm
, PGFAULT
);
3694 /* do counter updates before entering really critical section. */
3695 check_sync_rss_stat(current
);
3697 if (unlikely(is_vm_hugetlb_page(vma
)))
3698 return hugetlb_fault(mm
, vma
, address
, flags
);
3701 pgd
= pgd_offset(mm
, address
);
3702 pud
= pud_alloc(mm
, pgd
, address
);
3704 return VM_FAULT_OOM
;
3705 pmd
= pmd_alloc(mm
, pud
, address
);
3707 return VM_FAULT_OOM
;
3708 if (pmd_none(*pmd
) && transparent_hugepage_enabled(vma
)) {
3710 return do_huge_pmd_anonymous_page(mm
, vma
, address
,
3713 pmd_t orig_pmd
= *pmd
;
3717 if (pmd_trans_huge(orig_pmd
)) {
3718 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
3721 * If the pmd is splitting, return and retry the
3722 * the fault. Alternative: wait until the split
3723 * is done, and goto retry.
3725 if (pmd_trans_splitting(orig_pmd
))
3728 if (pmd_numa(orig_pmd
))
3729 return do_huge_pmd_numa_page(mm
, vma
, address
,
3732 if (dirty
&& !pmd_write(orig_pmd
)) {
3733 ret
= do_huge_pmd_wp_page(mm
, vma
, address
, pmd
,
3736 * If COW results in an oom, the huge pmd will
3737 * have been split, so retry the fault on the
3738 * pte for a smaller charge.
3740 if (unlikely(ret
& VM_FAULT_OOM
))
3744 huge_pmd_set_accessed(mm
, vma
, address
, pmd
,
3753 return do_pmd_numa_page(mm
, vma
, address
, pmd
);
3756 * Use __pte_alloc instead of pte_alloc_map, because we can't
3757 * run pte_offset_map on the pmd, if an huge pmd could
3758 * materialize from under us from a different thread.
3760 if (unlikely(pmd_none(*pmd
)) &&
3761 unlikely(__pte_alloc(mm
, vma
, pmd
, address
)))
3762 return VM_FAULT_OOM
;
3763 /* if an huge pmd materialized from under us just retry later */
3764 if (unlikely(pmd_trans_huge(*pmd
)))
3767 * A regular pmd is established and it can't morph into a huge pmd
3768 * from under us anymore at this point because we hold the mmap_sem
3769 * read mode and khugepaged takes it in write mode. So now it's
3770 * safe to run pte_offset_map().
3772 pte
= pte_offset_map(pmd
, address
);
3774 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
3777 #ifndef __PAGETABLE_PUD_FOLDED
3779 * Allocate page upper directory.
3780 * We've already handled the fast-path in-line.
3782 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
3784 pud_t
*new = pud_alloc_one(mm
, address
);
3788 smp_wmb(); /* See comment in __pte_alloc */
3790 spin_lock(&mm
->page_table_lock
);
3791 if (pgd_present(*pgd
)) /* Another has populated it */
3794 pgd_populate(mm
, pgd
, new);
3795 spin_unlock(&mm
->page_table_lock
);
3798 #endif /* __PAGETABLE_PUD_FOLDED */
3800 #ifndef __PAGETABLE_PMD_FOLDED
3802 * Allocate page middle directory.
3803 * We've already handled the fast-path in-line.
3805 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
3807 pmd_t
*new = pmd_alloc_one(mm
, address
);
3811 smp_wmb(); /* See comment in __pte_alloc */
3813 spin_lock(&mm
->page_table_lock
);
3814 #ifndef __ARCH_HAS_4LEVEL_HACK
3815 if (pud_present(*pud
)) /* Another has populated it */
3818 pud_populate(mm
, pud
, new);
3820 if (pgd_present(*pud
)) /* Another has populated it */
3823 pgd_populate(mm
, pud
, new);
3824 #endif /* __ARCH_HAS_4LEVEL_HACK */
3825 spin_unlock(&mm
->page_table_lock
);
3828 #endif /* __PAGETABLE_PMD_FOLDED */
3830 #if !defined(__HAVE_ARCH_GATE_AREA)
3832 #if defined(AT_SYSINFO_EHDR)
3833 static struct vm_area_struct gate_vma
;
3835 static int __init
gate_vma_init(void)
3837 gate_vma
.vm_mm
= NULL
;
3838 gate_vma
.vm_start
= FIXADDR_USER_START
;
3839 gate_vma
.vm_end
= FIXADDR_USER_END
;
3840 gate_vma
.vm_flags
= VM_READ
| VM_MAYREAD
| VM_EXEC
| VM_MAYEXEC
;
3841 gate_vma
.vm_page_prot
= __P101
;
3845 __initcall(gate_vma_init
);
3848 struct vm_area_struct
*get_gate_vma(struct mm_struct
*mm
)
3850 #ifdef AT_SYSINFO_EHDR
3857 int in_gate_area_no_mm(unsigned long addr
)
3859 #ifdef AT_SYSINFO_EHDR
3860 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
3866 #endif /* __HAVE_ARCH_GATE_AREA */
3868 static int __follow_pte(struct mm_struct
*mm
, unsigned long address
,
3869 pte_t
**ptepp
, spinlock_t
**ptlp
)
3876 pgd
= pgd_offset(mm
, address
);
3877 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
3880 pud
= pud_offset(pgd
, address
);
3881 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
3884 pmd
= pmd_offset(pud
, address
);
3885 VM_BUG_ON(pmd_trans_huge(*pmd
));
3886 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
3889 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3893 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
3896 if (!pte_present(*ptep
))
3901 pte_unmap_unlock(ptep
, *ptlp
);
3906 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
3907 pte_t
**ptepp
, spinlock_t
**ptlp
)
3911 /* (void) is needed to make gcc happy */
3912 (void) __cond_lock(*ptlp
,
3913 !(res
= __follow_pte(mm
, address
, ptepp
, ptlp
)));
3918 * follow_pfn - look up PFN at a user virtual address
3919 * @vma: memory mapping
3920 * @address: user virtual address
3921 * @pfn: location to store found PFN
3923 * Only IO mappings and raw PFN mappings are allowed.
3925 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3927 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
3934 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3937 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
3940 *pfn
= pte_pfn(*ptep
);
3941 pte_unmap_unlock(ptep
, ptl
);
3944 EXPORT_SYMBOL(follow_pfn
);
3946 #ifdef CONFIG_HAVE_IOREMAP_PROT
3947 int follow_phys(struct vm_area_struct
*vma
,
3948 unsigned long address
, unsigned int flags
,
3949 unsigned long *prot
, resource_size_t
*phys
)
3955 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3958 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
3962 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
3965 *prot
= pgprot_val(pte_pgprot(pte
));
3966 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
3970 pte_unmap_unlock(ptep
, ptl
);
3975 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
3976 void *buf
, int len
, int write
)
3978 resource_size_t phys_addr
;
3979 unsigned long prot
= 0;
3980 void __iomem
*maddr
;
3981 int offset
= addr
& (PAGE_SIZE
-1);
3983 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
3986 maddr
= ioremap_prot(phys_addr
, PAGE_SIZE
, prot
);
3988 memcpy_toio(maddr
+ offset
, buf
, len
);
3990 memcpy_fromio(buf
, maddr
+ offset
, len
);
3998 * Access another process' address space as given in mm. If non-NULL, use the
3999 * given task for page fault accounting.
4001 static int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
4002 unsigned long addr
, void *buf
, int len
, int write
)
4004 struct vm_area_struct
*vma
;
4005 void *old_buf
= buf
;
4007 down_read(&mm
->mmap_sem
);
4008 /* ignore errors, just check how much was successfully transferred */
4010 int bytes
, ret
, offset
;
4012 struct page
*page
= NULL
;
4014 ret
= get_user_pages(tsk
, mm
, addr
, 1,
4015 write
, 1, &page
, &vma
);
4018 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4019 * we can access using slightly different code.
4021 #ifdef CONFIG_HAVE_IOREMAP_PROT
4022 vma
= find_vma(mm
, addr
);
4023 if (!vma
|| vma
->vm_start
> addr
)
4025 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
4026 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
4034 offset
= addr
& (PAGE_SIZE
-1);
4035 if (bytes
> PAGE_SIZE
-offset
)
4036 bytes
= PAGE_SIZE
-offset
;
4040 copy_to_user_page(vma
, page
, addr
,
4041 maddr
+ offset
, buf
, bytes
);
4042 set_page_dirty_lock(page
);
4044 copy_from_user_page(vma
, page
, addr
,
4045 buf
, maddr
+ offset
, bytes
);
4048 page_cache_release(page
);
4054 up_read(&mm
->mmap_sem
);
4056 return buf
- old_buf
;
4060 * access_remote_vm - access another process' address space
4061 * @mm: the mm_struct of the target address space
4062 * @addr: start address to access
4063 * @buf: source or destination buffer
4064 * @len: number of bytes to transfer
4065 * @write: whether the access is a write
4067 * The caller must hold a reference on @mm.
4069 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
4070 void *buf
, int len
, int write
)
4072 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, write
);
4076 * Access another process' address space.
4077 * Source/target buffer must be kernel space,
4078 * Do not walk the page table directly, use get_user_pages
4080 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
4081 void *buf
, int len
, int write
)
4083 struct mm_struct
*mm
;
4086 mm
= get_task_mm(tsk
);
4090 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, write
);
4097 * Print the name of a VMA.
4099 void print_vma_addr(char *prefix
, unsigned long ip
)
4101 struct mm_struct
*mm
= current
->mm
;
4102 struct vm_area_struct
*vma
;
4105 * Do not print if we are in atomic
4106 * contexts (in exception stacks, etc.):
4108 if (preempt_count())
4111 down_read(&mm
->mmap_sem
);
4112 vma
= find_vma(mm
, ip
);
4113 if (vma
&& vma
->vm_file
) {
4114 struct file
*f
= vma
->vm_file
;
4115 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
4119 p
= d_path(&f
->f_path
, buf
, PAGE_SIZE
);
4122 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
4124 vma
->vm_end
- vma
->vm_start
);
4125 free_page((unsigned long)buf
);
4128 up_read(&mm
->mmap_sem
);
4131 #ifdef CONFIG_PROVE_LOCKING
4132 void might_fault(void)
4135 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4136 * holding the mmap_sem, this is safe because kernel memory doesn't
4137 * get paged out, therefore we'll never actually fault, and the
4138 * below annotations will generate false positives.
4140 if (segment_eq(get_fs(), KERNEL_DS
))
4145 * it would be nicer only to annotate paths which are not under
4146 * pagefault_disable, however that requires a larger audit and
4147 * providing helpers like get_user_atomic.
4149 if (!in_atomic() && current
->mm
)
4150 might_lock_read(¤t
->mm
->mmap_sem
);
4152 EXPORT_SYMBOL(might_fault
);
4155 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4156 static void clear_gigantic_page(struct page
*page
,
4158 unsigned int pages_per_huge_page
)
4161 struct page
*p
= page
;
4164 for (i
= 0; i
< pages_per_huge_page
;
4165 i
++, p
= mem_map_next(p
, page
, i
)) {
4167 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
4170 void clear_huge_page(struct page
*page
,
4171 unsigned long addr
, unsigned int pages_per_huge_page
)
4175 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4176 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
4181 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4183 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
4187 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
4189 struct vm_area_struct
*vma
,
4190 unsigned int pages_per_huge_page
)
4193 struct page
*dst_base
= dst
;
4194 struct page
*src_base
= src
;
4196 for (i
= 0; i
< pages_per_huge_page
; ) {
4198 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
4201 dst
= mem_map_next(dst
, dst_base
, i
);
4202 src
= mem_map_next(src
, src_base
, i
);
4206 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
4207 unsigned long addr
, struct vm_area_struct
*vma
,
4208 unsigned int pages_per_huge_page
)
4212 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4213 copy_user_gigantic_page(dst
, src
, addr
, vma
,
4214 pages_per_huge_page
);
4219 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4221 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
4224 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */