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_CPUPID_NOT_IN_PAGE_FLAGS
73 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
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
);
86 * A number of key systems in x86 including ioremap() rely on the assumption
87 * that high_memory defines the upper bound on direct map memory, then end
88 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
89 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
94 EXPORT_SYMBOL(high_memory
);
97 * Randomize the address space (stacks, mmaps, brk, etc.).
99 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
100 * as ancient (libc5 based) binaries can segfault. )
102 int randomize_va_space __read_mostly
=
103 #ifdef CONFIG_COMPAT_BRK
109 static int __init
disable_randmaps(char *s
)
111 randomize_va_space
= 0;
114 __setup("norandmaps", disable_randmaps
);
116 unsigned long zero_pfn __read_mostly
;
117 unsigned long highest_memmap_pfn __read_mostly
;
120 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
122 static int __init
init_zero_pfn(void)
124 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
127 core_initcall(init_zero_pfn
);
130 #if defined(SPLIT_RSS_COUNTING)
132 void sync_mm_rss(struct mm_struct
*mm
)
136 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
137 if (current
->rss_stat
.count
[i
]) {
138 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
139 current
->rss_stat
.count
[i
] = 0;
142 current
->rss_stat
.events
= 0;
145 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
147 struct task_struct
*task
= current
;
149 if (likely(task
->mm
== mm
))
150 task
->rss_stat
.count
[member
] += val
;
152 add_mm_counter(mm
, member
, val
);
154 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
155 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
157 /* sync counter once per 64 page faults */
158 #define TASK_RSS_EVENTS_THRESH (64)
159 static void check_sync_rss_stat(struct task_struct
*task
)
161 if (unlikely(task
!= current
))
163 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
164 sync_mm_rss(task
->mm
);
166 #else /* SPLIT_RSS_COUNTING */
168 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
169 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
171 static void check_sync_rss_stat(struct task_struct
*task
)
175 #endif /* SPLIT_RSS_COUNTING */
177 #ifdef HAVE_GENERIC_MMU_GATHER
179 static int tlb_next_batch(struct mmu_gather
*tlb
)
181 struct mmu_gather_batch
*batch
;
185 tlb
->active
= batch
->next
;
189 if (tlb
->batch_count
== MAX_GATHER_BATCH_COUNT
)
192 batch
= (void *)__get_free_pages(GFP_NOWAIT
| __GFP_NOWARN
, 0);
199 batch
->max
= MAX_GATHER_BATCH
;
201 tlb
->active
->next
= batch
;
208 * Called to initialize an (on-stack) mmu_gather structure for page-table
209 * tear-down from @mm. The @fullmm argument is used when @mm is without
210 * users and we're going to destroy the full address space (exit/execve).
212 void tlb_gather_mmu(struct mmu_gather
*tlb
, struct mm_struct
*mm
, unsigned long start
, unsigned long end
)
216 /* Is it from 0 to ~0? */
217 tlb
->fullmm
= !(start
| (end
+1));
218 tlb
->need_flush_all
= 0;
222 tlb
->local
.next
= NULL
;
224 tlb
->local
.max
= ARRAY_SIZE(tlb
->__pages
);
225 tlb
->active
= &tlb
->local
;
226 tlb
->batch_count
= 0;
228 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
233 void tlb_flush_mmu(struct mmu_gather
*tlb
)
235 struct mmu_gather_batch
*batch
;
237 if (!tlb
->need_flush
)
241 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
242 tlb_table_flush(tlb
);
245 for (batch
= &tlb
->local
; batch
; batch
= batch
->next
) {
246 free_pages_and_swap_cache(batch
->pages
, batch
->nr
);
249 tlb
->active
= &tlb
->local
;
253 * Called at the end of the shootdown operation to free up any resources
254 * that were required.
256 void tlb_finish_mmu(struct mmu_gather
*tlb
, unsigned long start
, unsigned long end
)
258 struct mmu_gather_batch
*batch
, *next
;
262 /* keep the page table cache within bounds */
265 for (batch
= tlb
->local
.next
; batch
; batch
= next
) {
267 free_pages((unsigned long)batch
, 0);
269 tlb
->local
.next
= NULL
;
273 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
274 * handling the additional races in SMP caused by other CPUs caching valid
275 * mappings in their TLBs. Returns the number of free page slots left.
276 * When out of page slots we must call tlb_flush_mmu().
278 int __tlb_remove_page(struct mmu_gather
*tlb
, struct page
*page
)
280 struct mmu_gather_batch
*batch
;
282 VM_BUG_ON(!tlb
->need_flush
);
285 batch
->pages
[batch
->nr
++] = page
;
286 if (batch
->nr
== batch
->max
) {
287 if (!tlb_next_batch(tlb
))
291 VM_BUG_ON(batch
->nr
> batch
->max
);
293 return batch
->max
- batch
->nr
;
296 #endif /* HAVE_GENERIC_MMU_GATHER */
298 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
301 * See the comment near struct mmu_table_batch.
304 static void tlb_remove_table_smp_sync(void *arg
)
306 /* Simply deliver the interrupt */
309 static void tlb_remove_table_one(void *table
)
312 * This isn't an RCU grace period and hence the page-tables cannot be
313 * assumed to be actually RCU-freed.
315 * It is however sufficient for software page-table walkers that rely on
316 * IRQ disabling. See the comment near struct mmu_table_batch.
318 smp_call_function(tlb_remove_table_smp_sync
, NULL
, 1);
319 __tlb_remove_table(table
);
322 static void tlb_remove_table_rcu(struct rcu_head
*head
)
324 struct mmu_table_batch
*batch
;
327 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
329 for (i
= 0; i
< batch
->nr
; i
++)
330 __tlb_remove_table(batch
->tables
[i
]);
332 free_page((unsigned long)batch
);
335 void tlb_table_flush(struct mmu_gather
*tlb
)
337 struct mmu_table_batch
**batch
= &tlb
->batch
;
340 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
345 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
347 struct mmu_table_batch
**batch
= &tlb
->batch
;
352 * When there's less then two users of this mm there cannot be a
353 * concurrent page-table walk.
355 if (atomic_read(&tlb
->mm
->mm_users
) < 2) {
356 __tlb_remove_table(table
);
360 if (*batch
== NULL
) {
361 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
362 if (*batch
== NULL
) {
363 tlb_remove_table_one(table
);
368 (*batch
)->tables
[(*batch
)->nr
++] = table
;
369 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
370 tlb_table_flush(tlb
);
373 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
376 * Note: this doesn't free the actual pages themselves. That
377 * has been handled earlier when unmapping all the memory regions.
379 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
382 pgtable_t token
= pmd_pgtable(*pmd
);
384 pte_free_tlb(tlb
, token
, addr
);
388 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
389 unsigned long addr
, unsigned long end
,
390 unsigned long floor
, unsigned long ceiling
)
397 pmd
= pmd_offset(pud
, addr
);
399 next
= pmd_addr_end(addr
, end
);
400 if (pmd_none_or_clear_bad(pmd
))
402 free_pte_range(tlb
, pmd
, addr
);
403 } while (pmd
++, addr
= next
, addr
!= end
);
413 if (end
- 1 > ceiling
- 1)
416 pmd
= pmd_offset(pud
, start
);
418 pmd_free_tlb(tlb
, pmd
, start
);
421 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
422 unsigned long addr
, unsigned long end
,
423 unsigned long floor
, unsigned long ceiling
)
430 pud
= pud_offset(pgd
, addr
);
432 next
= pud_addr_end(addr
, end
);
433 if (pud_none_or_clear_bad(pud
))
435 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
436 } while (pud
++, addr
= next
, addr
!= end
);
442 ceiling
&= PGDIR_MASK
;
446 if (end
- 1 > ceiling
- 1)
449 pud
= pud_offset(pgd
, start
);
451 pud_free_tlb(tlb
, pud
, start
);
455 * This function frees user-level page tables of a process.
457 void free_pgd_range(struct mmu_gather
*tlb
,
458 unsigned long addr
, unsigned long end
,
459 unsigned long floor
, unsigned long ceiling
)
465 * The next few lines have given us lots of grief...
467 * Why are we testing PMD* at this top level? Because often
468 * there will be no work to do at all, and we'd prefer not to
469 * go all the way down to the bottom just to discover that.
471 * Why all these "- 1"s? Because 0 represents both the bottom
472 * of the address space and the top of it (using -1 for the
473 * top wouldn't help much: the masks would do the wrong thing).
474 * The rule is that addr 0 and floor 0 refer to the bottom of
475 * the address space, but end 0 and ceiling 0 refer to the top
476 * Comparisons need to use "end - 1" and "ceiling - 1" (though
477 * that end 0 case should be mythical).
479 * Wherever addr is brought up or ceiling brought down, we must
480 * be careful to reject "the opposite 0" before it confuses the
481 * subsequent tests. But what about where end is brought down
482 * by PMD_SIZE below? no, end can't go down to 0 there.
484 * Whereas we round start (addr) and ceiling down, by different
485 * masks at different levels, in order to test whether a table
486 * now has no other vmas using it, so can be freed, we don't
487 * bother to round floor or end up - the tests don't need that.
501 if (end
- 1 > ceiling
- 1)
506 pgd
= pgd_offset(tlb
->mm
, addr
);
508 next
= pgd_addr_end(addr
, end
);
509 if (pgd_none_or_clear_bad(pgd
))
511 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
512 } while (pgd
++, addr
= next
, addr
!= end
);
515 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
516 unsigned long floor
, unsigned long ceiling
)
519 struct vm_area_struct
*next
= vma
->vm_next
;
520 unsigned long addr
= vma
->vm_start
;
523 * Hide vma from rmap and truncate_pagecache before freeing
526 unlink_anon_vmas(vma
);
527 unlink_file_vma(vma
);
529 if (is_vm_hugetlb_page(vma
)) {
530 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
531 floor
, next
? next
->vm_start
: ceiling
);
534 * Optimization: gather nearby vmas into one call down
536 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
537 && !is_vm_hugetlb_page(next
)) {
540 unlink_anon_vmas(vma
);
541 unlink_file_vma(vma
);
543 free_pgd_range(tlb
, addr
, vma
->vm_end
,
544 floor
, next
? next
->vm_start
: ceiling
);
550 int __pte_alloc(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
551 pmd_t
*pmd
, unsigned long address
)
553 pgtable_t
new = pte_alloc_one(mm
, address
);
554 int wait_split_huge_page
;
559 * Ensure all pte setup (eg. pte page lock and page clearing) are
560 * visible before the pte is made visible to other CPUs by being
561 * put into page tables.
563 * The other side of the story is the pointer chasing in the page
564 * table walking code (when walking the page table without locking;
565 * ie. most of the time). Fortunately, these data accesses consist
566 * of a chain of data-dependent loads, meaning most CPUs (alpha
567 * being the notable exception) will already guarantee loads are
568 * seen in-order. See the alpha page table accessors for the
569 * smp_read_barrier_depends() barriers in page table walking code.
571 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
573 spin_lock(&mm
->page_table_lock
);
574 wait_split_huge_page
= 0;
575 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
577 pmd_populate(mm
, pmd
, new);
579 } else if (unlikely(pmd_trans_splitting(*pmd
)))
580 wait_split_huge_page
= 1;
581 spin_unlock(&mm
->page_table_lock
);
584 if (wait_split_huge_page
)
585 wait_split_huge_page(vma
->anon_vma
, pmd
);
589 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
591 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
595 smp_wmb(); /* See comment in __pte_alloc */
597 spin_lock(&init_mm
.page_table_lock
);
598 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
599 pmd_populate_kernel(&init_mm
, pmd
, new);
602 VM_BUG_ON(pmd_trans_splitting(*pmd
));
603 spin_unlock(&init_mm
.page_table_lock
);
605 pte_free_kernel(&init_mm
, new);
609 static inline void init_rss_vec(int *rss
)
611 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
614 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
618 if (current
->mm
== mm
)
620 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
622 add_mm_counter(mm
, i
, rss
[i
]);
626 * This function is called to print an error when a bad pte
627 * is found. For example, we might have a PFN-mapped pte in
628 * a region that doesn't allow it.
630 * The calling function must still handle the error.
632 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
633 pte_t pte
, struct page
*page
)
635 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
636 pud_t
*pud
= pud_offset(pgd
, addr
);
637 pmd_t
*pmd
= pmd_offset(pud
, addr
);
638 struct address_space
*mapping
;
640 static unsigned long resume
;
641 static unsigned long nr_shown
;
642 static unsigned long nr_unshown
;
645 * Allow a burst of 60 reports, then keep quiet for that minute;
646 * or allow a steady drip of one report per second.
648 if (nr_shown
== 60) {
649 if (time_before(jiffies
, resume
)) {
655 "BUG: Bad page map: %lu messages suppressed\n",
662 resume
= jiffies
+ 60 * HZ
;
664 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
665 index
= linear_page_index(vma
, addr
);
668 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
670 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
674 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
675 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
677 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
680 printk(KERN_ALERT
"vma->vm_ops->fault: %pSR\n",
682 if (vma
->vm_file
&& vma
->vm_file
->f_op
)
683 printk(KERN_ALERT
"vma->vm_file->f_op->mmap: %pSR\n",
684 vma
->vm_file
->f_op
->mmap
);
686 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
689 static inline bool is_cow_mapping(vm_flags_t flags
)
691 return (flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
695 * vm_normal_page -- This function gets the "struct page" associated with a pte.
697 * "Special" mappings do not wish to be associated with a "struct page" (either
698 * it doesn't exist, or it exists but they don't want to touch it). In this
699 * case, NULL is returned here. "Normal" mappings do have a struct page.
701 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
702 * pte bit, in which case this function is trivial. Secondly, an architecture
703 * may not have a spare pte bit, which requires a more complicated scheme,
706 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
707 * special mapping (even if there are underlying and valid "struct pages").
708 * COWed pages of a VM_PFNMAP are always normal.
710 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
711 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
712 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
713 * mapping will always honor the rule
715 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
717 * And for normal mappings this is false.
719 * This restricts such mappings to be a linear translation from virtual address
720 * to pfn. To get around this restriction, we allow arbitrary mappings so long
721 * as the vma is not a COW mapping; in that case, we know that all ptes are
722 * special (because none can have been COWed).
725 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
727 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
728 * page" backing, however the difference is that _all_ pages with a struct
729 * page (that is, those where pfn_valid is true) are refcounted and considered
730 * normal pages by the VM. The disadvantage is that pages are refcounted
731 * (which can be slower and simply not an option for some PFNMAP users). The
732 * advantage is that we don't have to follow the strict linearity rule of
733 * PFNMAP mappings in order to support COWable mappings.
736 #ifdef __HAVE_ARCH_PTE_SPECIAL
737 # define HAVE_PTE_SPECIAL 1
739 # define HAVE_PTE_SPECIAL 0
741 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
744 unsigned long pfn
= pte_pfn(pte
);
746 if (HAVE_PTE_SPECIAL
) {
747 if (likely(!pte_special(pte
)))
749 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
751 if (!is_zero_pfn(pfn
))
752 print_bad_pte(vma
, addr
, pte
, NULL
);
756 /* !HAVE_PTE_SPECIAL case follows: */
758 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
759 if (vma
->vm_flags
& VM_MIXEDMAP
) {
765 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
766 if (pfn
== vma
->vm_pgoff
+ off
)
768 if (!is_cow_mapping(vma
->vm_flags
))
773 if (is_zero_pfn(pfn
))
776 if (unlikely(pfn
> highest_memmap_pfn
)) {
777 print_bad_pte(vma
, addr
, pte
, NULL
);
782 * NOTE! We still have PageReserved() pages in the page tables.
783 * eg. VDSO mappings can cause them to exist.
786 return pfn_to_page(pfn
);
790 * copy one vm_area from one task to the other. Assumes the page tables
791 * already present in the new task to be cleared in the whole range
792 * covered by this vma.
795 static inline unsigned long
796 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
797 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
798 unsigned long addr
, int *rss
)
800 unsigned long vm_flags
= vma
->vm_flags
;
801 pte_t pte
= *src_pte
;
804 /* pte contains position in swap or file, so copy. */
805 if (unlikely(!pte_present(pte
))) {
806 if (!pte_file(pte
)) {
807 swp_entry_t entry
= pte_to_swp_entry(pte
);
809 if (swap_duplicate(entry
) < 0)
812 /* make sure dst_mm is on swapoff's mmlist. */
813 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
814 spin_lock(&mmlist_lock
);
815 if (list_empty(&dst_mm
->mmlist
))
816 list_add(&dst_mm
->mmlist
,
818 spin_unlock(&mmlist_lock
);
820 if (likely(!non_swap_entry(entry
)))
822 else if (is_migration_entry(entry
)) {
823 page
= migration_entry_to_page(entry
);
830 if (is_write_migration_entry(entry
) &&
831 is_cow_mapping(vm_flags
)) {
833 * COW mappings require pages in both
834 * parent and child to be set to read.
836 make_migration_entry_read(&entry
);
837 pte
= swp_entry_to_pte(entry
);
838 if (pte_swp_soft_dirty(*src_pte
))
839 pte
= pte_swp_mksoft_dirty(pte
);
840 set_pte_at(src_mm
, addr
, src_pte
, pte
);
848 * If it's a COW mapping, write protect it both
849 * in the parent and the child
851 if (is_cow_mapping(vm_flags
)) {
852 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
853 pte
= pte_wrprotect(pte
);
857 * If it's a shared mapping, mark it clean in
860 if (vm_flags
& VM_SHARED
)
861 pte
= pte_mkclean(pte
);
862 pte
= pte_mkold(pte
);
864 page
= vm_normal_page(vma
, addr
, pte
);
875 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
879 int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
880 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
881 unsigned long addr
, unsigned long end
)
883 pte_t
*orig_src_pte
, *orig_dst_pte
;
884 pte_t
*src_pte
, *dst_pte
;
885 spinlock_t
*src_ptl
, *dst_ptl
;
887 int rss
[NR_MM_COUNTERS
];
888 swp_entry_t entry
= (swp_entry_t
){0};
893 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
896 src_pte
= pte_offset_map(src_pmd
, addr
);
897 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
898 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
899 orig_src_pte
= src_pte
;
900 orig_dst_pte
= dst_pte
;
901 arch_enter_lazy_mmu_mode();
905 * We are holding two locks at this point - either of them
906 * could generate latencies in another task on another CPU.
908 if (progress
>= 32) {
910 if (need_resched() ||
911 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
914 if (pte_none(*src_pte
)) {
918 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
923 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
925 arch_leave_lazy_mmu_mode();
926 spin_unlock(src_ptl
);
927 pte_unmap(orig_src_pte
);
928 add_mm_rss_vec(dst_mm
, rss
);
929 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
933 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
942 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
943 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
944 unsigned long addr
, unsigned long end
)
946 pmd_t
*src_pmd
, *dst_pmd
;
949 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
952 src_pmd
= pmd_offset(src_pud
, addr
);
954 next
= pmd_addr_end(addr
, end
);
955 if (pmd_trans_huge(*src_pmd
)) {
957 VM_BUG_ON(next
-addr
!= HPAGE_PMD_SIZE
);
958 err
= copy_huge_pmd(dst_mm
, src_mm
,
959 dst_pmd
, src_pmd
, addr
, vma
);
966 if (pmd_none_or_clear_bad(src_pmd
))
968 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
971 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
975 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
976 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
977 unsigned long addr
, unsigned long end
)
979 pud_t
*src_pud
, *dst_pud
;
982 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
985 src_pud
= pud_offset(src_pgd
, addr
);
987 next
= pud_addr_end(addr
, end
);
988 if (pud_none_or_clear_bad(src_pud
))
990 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
993 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
997 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
998 struct vm_area_struct
*vma
)
1000 pgd_t
*src_pgd
, *dst_pgd
;
1002 unsigned long addr
= vma
->vm_start
;
1003 unsigned long end
= vma
->vm_end
;
1004 unsigned long mmun_start
; /* For mmu_notifiers */
1005 unsigned long mmun_end
; /* For mmu_notifiers */
1010 * Don't copy ptes where a page fault will fill them correctly.
1011 * Fork becomes much lighter when there are big shared or private
1012 * readonly mappings. The tradeoff is that copy_page_range is more
1013 * efficient than faulting.
1015 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_NONLINEAR
|
1016 VM_PFNMAP
| VM_MIXEDMAP
))) {
1021 if (is_vm_hugetlb_page(vma
))
1022 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1024 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
1026 * We do not free on error cases below as remove_vma
1027 * gets called on error from higher level routine
1029 ret
= track_pfn_copy(vma
);
1035 * We need to invalidate the secondary MMU mappings only when
1036 * there could be a permission downgrade on the ptes of the
1037 * parent mm. And a permission downgrade will only happen if
1038 * is_cow_mapping() returns true.
1040 is_cow
= is_cow_mapping(vma
->vm_flags
);
1044 mmu_notifier_invalidate_range_start(src_mm
, mmun_start
,
1048 dst_pgd
= pgd_offset(dst_mm
, addr
);
1049 src_pgd
= pgd_offset(src_mm
, addr
);
1051 next
= pgd_addr_end(addr
, end
);
1052 if (pgd_none_or_clear_bad(src_pgd
))
1054 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1055 vma
, addr
, next
))) {
1059 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1062 mmu_notifier_invalidate_range_end(src_mm
, mmun_start
, mmun_end
);
1066 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1067 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1068 unsigned long addr
, unsigned long end
,
1069 struct zap_details
*details
)
1071 struct mm_struct
*mm
= tlb
->mm
;
1072 int force_flush
= 0;
1073 int rss
[NR_MM_COUNTERS
];
1080 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1082 arch_enter_lazy_mmu_mode();
1085 if (pte_none(ptent
)) {
1089 if (pte_present(ptent
)) {
1092 page
= vm_normal_page(vma
, addr
, ptent
);
1093 if (unlikely(details
) && page
) {
1095 * unmap_shared_mapping_pages() wants to
1096 * invalidate cache without truncating:
1097 * unmap shared but keep private pages.
1099 if (details
->check_mapping
&&
1100 details
->check_mapping
!= page
->mapping
)
1103 * Each page->index must be checked when
1104 * invalidating or truncating nonlinear.
1106 if (details
->nonlinear_vma
&&
1107 (page
->index
< details
->first_index
||
1108 page
->index
> details
->last_index
))
1111 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1113 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1114 if (unlikely(!page
))
1116 if (unlikely(details
) && details
->nonlinear_vma
1117 && linear_page_index(details
->nonlinear_vma
,
1118 addr
) != page
->index
) {
1119 pte_t ptfile
= pgoff_to_pte(page
->index
);
1120 if (pte_soft_dirty(ptent
))
1121 pte_file_mksoft_dirty(ptfile
);
1122 set_pte_at(mm
, addr
, pte
, ptfile
);
1125 rss
[MM_ANONPAGES
]--;
1127 if (pte_dirty(ptent
))
1128 set_page_dirty(page
);
1129 if (pte_young(ptent
) &&
1130 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1131 mark_page_accessed(page
);
1132 rss
[MM_FILEPAGES
]--;
1134 page_remove_rmap(page
);
1135 if (unlikely(page_mapcount(page
) < 0))
1136 print_bad_pte(vma
, addr
, ptent
, page
);
1137 force_flush
= !__tlb_remove_page(tlb
, page
);
1143 * If details->check_mapping, we leave swap entries;
1144 * if details->nonlinear_vma, we leave file entries.
1146 if (unlikely(details
))
1148 if (pte_file(ptent
)) {
1149 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
)))
1150 print_bad_pte(vma
, addr
, ptent
, NULL
);
1152 swp_entry_t entry
= pte_to_swp_entry(ptent
);
1154 if (!non_swap_entry(entry
))
1156 else if (is_migration_entry(entry
)) {
1159 page
= migration_entry_to_page(entry
);
1162 rss
[MM_ANONPAGES
]--;
1164 rss
[MM_FILEPAGES
]--;
1166 if (unlikely(!free_swap_and_cache(entry
)))
1167 print_bad_pte(vma
, addr
, ptent
, NULL
);
1169 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1170 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1172 add_mm_rss_vec(mm
, rss
);
1173 arch_leave_lazy_mmu_mode();
1174 pte_unmap_unlock(start_pte
, ptl
);
1177 * mmu_gather ran out of room to batch pages, we break out of
1178 * the PTE lock to avoid doing the potential expensive TLB invalidate
1179 * and page-free while holding it.
1182 unsigned long old_end
;
1187 * Flush the TLB just for the previous segment,
1188 * then update the range to be the remaining
1206 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1207 struct vm_area_struct
*vma
, pud_t
*pud
,
1208 unsigned long addr
, unsigned long end
,
1209 struct zap_details
*details
)
1214 pmd
= pmd_offset(pud
, addr
);
1216 next
= pmd_addr_end(addr
, end
);
1217 if (pmd_trans_huge(*pmd
)) {
1218 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1219 #ifdef CONFIG_DEBUG_VM
1220 if (!rwsem_is_locked(&tlb
->mm
->mmap_sem
)) {
1221 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1222 __func__
, addr
, end
,
1228 split_huge_page_pmd(vma
, addr
, pmd
);
1229 } else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1234 * Here there can be other concurrent MADV_DONTNEED or
1235 * trans huge page faults running, and if the pmd is
1236 * none or trans huge it can change under us. This is
1237 * because MADV_DONTNEED holds the mmap_sem in read
1240 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1242 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1245 } while (pmd
++, addr
= next
, addr
!= end
);
1250 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1251 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1252 unsigned long addr
, unsigned long end
,
1253 struct zap_details
*details
)
1258 pud
= pud_offset(pgd
, addr
);
1260 next
= pud_addr_end(addr
, end
);
1261 if (pud_none_or_clear_bad(pud
))
1263 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1264 } while (pud
++, addr
= next
, addr
!= end
);
1269 static void unmap_page_range(struct mmu_gather
*tlb
,
1270 struct vm_area_struct
*vma
,
1271 unsigned long addr
, unsigned long end
,
1272 struct zap_details
*details
)
1277 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
1280 BUG_ON(addr
>= end
);
1281 mem_cgroup_uncharge_start();
1282 tlb_start_vma(tlb
, vma
);
1283 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1285 next
= pgd_addr_end(addr
, end
);
1286 if (pgd_none_or_clear_bad(pgd
))
1288 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
1289 } while (pgd
++, addr
= next
, addr
!= end
);
1290 tlb_end_vma(tlb
, vma
);
1291 mem_cgroup_uncharge_end();
1295 static void unmap_single_vma(struct mmu_gather
*tlb
,
1296 struct vm_area_struct
*vma
, unsigned long start_addr
,
1297 unsigned long end_addr
,
1298 struct zap_details
*details
)
1300 unsigned long start
= max(vma
->vm_start
, start_addr
);
1303 if (start
>= vma
->vm_end
)
1305 end
= min(vma
->vm_end
, end_addr
);
1306 if (end
<= vma
->vm_start
)
1310 uprobe_munmap(vma
, start
, end
);
1312 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1313 untrack_pfn(vma
, 0, 0);
1316 if (unlikely(is_vm_hugetlb_page(vma
))) {
1318 * It is undesirable to test vma->vm_file as it
1319 * should be non-null for valid hugetlb area.
1320 * However, vm_file will be NULL in the error
1321 * cleanup path of do_mmap_pgoff. When
1322 * hugetlbfs ->mmap method fails,
1323 * do_mmap_pgoff() nullifies vma->vm_file
1324 * before calling this function to clean up.
1325 * Since no pte has actually been setup, it is
1326 * safe to do nothing in this case.
1329 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
1330 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1331 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
1334 unmap_page_range(tlb
, vma
, start
, end
, details
);
1339 * unmap_vmas - unmap a range of memory covered by a list of vma's
1340 * @tlb: address of the caller's struct mmu_gather
1341 * @vma: the starting vma
1342 * @start_addr: virtual address at which to start unmapping
1343 * @end_addr: virtual address at which to end unmapping
1345 * Unmap all pages in the vma list.
1347 * Only addresses between `start' and `end' will be unmapped.
1349 * The VMA list must be sorted in ascending virtual address order.
1351 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1352 * range after unmap_vmas() returns. So the only responsibility here is to
1353 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1354 * drops the lock and schedules.
1356 void unmap_vmas(struct mmu_gather
*tlb
,
1357 struct vm_area_struct
*vma
, unsigned long start_addr
,
1358 unsigned long end_addr
)
1360 struct mm_struct
*mm
= vma
->vm_mm
;
1362 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1363 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1364 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1365 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1369 * zap_page_range - remove user pages in a given range
1370 * @vma: vm_area_struct holding the applicable pages
1371 * @start: starting address of pages to zap
1372 * @size: number of bytes to zap
1373 * @details: details of nonlinear truncation or shared cache invalidation
1375 * Caller must protect the VMA list
1377 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1378 unsigned long size
, struct zap_details
*details
)
1380 struct mm_struct
*mm
= vma
->vm_mm
;
1381 struct mmu_gather tlb
;
1382 unsigned long end
= start
+ size
;
1385 tlb_gather_mmu(&tlb
, mm
, start
, end
);
1386 update_hiwater_rss(mm
);
1387 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1388 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
)
1389 unmap_single_vma(&tlb
, vma
, start
, end
, details
);
1390 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1391 tlb_finish_mmu(&tlb
, start
, end
);
1395 * zap_page_range_single - remove user pages in a given range
1396 * @vma: vm_area_struct holding the applicable pages
1397 * @address: starting address of pages to zap
1398 * @size: number of bytes to zap
1399 * @details: details of nonlinear truncation or shared cache invalidation
1401 * The range must fit into one VMA.
1403 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1404 unsigned long size
, struct zap_details
*details
)
1406 struct mm_struct
*mm
= vma
->vm_mm
;
1407 struct mmu_gather tlb
;
1408 unsigned long end
= address
+ size
;
1411 tlb_gather_mmu(&tlb
, mm
, address
, end
);
1412 update_hiwater_rss(mm
);
1413 mmu_notifier_invalidate_range_start(mm
, address
, end
);
1414 unmap_single_vma(&tlb
, vma
, address
, end
, details
);
1415 mmu_notifier_invalidate_range_end(mm
, address
, end
);
1416 tlb_finish_mmu(&tlb
, address
, end
);
1420 * zap_vma_ptes - remove ptes mapping the vma
1421 * @vma: vm_area_struct holding ptes to be zapped
1422 * @address: starting address of pages to zap
1423 * @size: number of bytes to zap
1425 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1427 * The entire address range must be fully contained within the vma.
1429 * Returns 0 if successful.
1431 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1434 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1435 !(vma
->vm_flags
& VM_PFNMAP
))
1437 zap_page_range_single(vma
, address
, size
, NULL
);
1440 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1443 * follow_page_mask - look up a page descriptor from a user-virtual address
1444 * @vma: vm_area_struct mapping @address
1445 * @address: virtual address to look up
1446 * @flags: flags modifying lookup behaviour
1447 * @page_mask: on output, *page_mask is set according to the size of the page
1449 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1451 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1452 * an error pointer if there is a mapping to something not represented
1453 * by a page descriptor (see also vm_normal_page()).
1455 struct page
*follow_page_mask(struct vm_area_struct
*vma
,
1456 unsigned long address
, unsigned int flags
,
1457 unsigned int *page_mask
)
1465 struct mm_struct
*mm
= vma
->vm_mm
;
1469 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
1470 if (!IS_ERR(page
)) {
1471 BUG_ON(flags
& FOLL_GET
);
1476 pgd
= pgd_offset(mm
, address
);
1477 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
1480 pud
= pud_offset(pgd
, address
);
1483 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
1484 if (flags
& FOLL_GET
)
1486 page
= follow_huge_pud(mm
, address
, pud
, flags
& FOLL_WRITE
);
1489 if (unlikely(pud_bad(*pud
)))
1492 pmd
= pmd_offset(pud
, address
);
1495 if (pmd_huge(*pmd
) && vma
->vm_flags
& VM_HUGETLB
) {
1496 page
= follow_huge_pmd(mm
, address
, pmd
, flags
& FOLL_WRITE
);
1497 if (flags
& FOLL_GET
) {
1499 * Refcount on tail pages are not well-defined and
1500 * shouldn't be taken. The caller should handle a NULL
1501 * return when trying to follow tail pages.
1512 if ((flags
& FOLL_NUMA
) && pmd_numa(*pmd
))
1514 if (pmd_trans_huge(*pmd
)) {
1515 if (flags
& FOLL_SPLIT
) {
1516 split_huge_page_pmd(vma
, address
, pmd
);
1517 goto split_fallthrough
;
1519 spin_lock(&mm
->page_table_lock
);
1520 if (likely(pmd_trans_huge(*pmd
))) {
1521 if (unlikely(pmd_trans_splitting(*pmd
))) {
1522 spin_unlock(&mm
->page_table_lock
);
1523 wait_split_huge_page(vma
->anon_vma
, pmd
);
1525 page
= follow_trans_huge_pmd(vma
, address
,
1527 spin_unlock(&mm
->page_table_lock
);
1528 *page_mask
= HPAGE_PMD_NR
- 1;
1532 spin_unlock(&mm
->page_table_lock
);
1536 if (unlikely(pmd_bad(*pmd
)))
1539 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1542 if (!pte_present(pte
)) {
1545 * KSM's break_ksm() relies upon recognizing a ksm page
1546 * even while it is being migrated, so for that case we
1547 * need migration_entry_wait().
1549 if (likely(!(flags
& FOLL_MIGRATION
)))
1551 if (pte_none(pte
) || pte_file(pte
))
1553 entry
= pte_to_swp_entry(pte
);
1554 if (!is_migration_entry(entry
))
1556 pte_unmap_unlock(ptep
, ptl
);
1557 migration_entry_wait(mm
, pmd
, address
);
1558 goto split_fallthrough
;
1560 if ((flags
& FOLL_NUMA
) && pte_numa(pte
))
1562 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
1565 page
= vm_normal_page(vma
, address
, pte
);
1566 if (unlikely(!page
)) {
1567 if ((flags
& FOLL_DUMP
) ||
1568 !is_zero_pfn(pte_pfn(pte
)))
1570 page
= pte_page(pte
);
1573 if (flags
& FOLL_GET
)
1574 get_page_foll(page
);
1575 if (flags
& FOLL_TOUCH
) {
1576 if ((flags
& FOLL_WRITE
) &&
1577 !pte_dirty(pte
) && !PageDirty(page
))
1578 set_page_dirty(page
);
1580 * pte_mkyoung() would be more correct here, but atomic care
1581 * is needed to avoid losing the dirty bit: it is easier to use
1582 * mark_page_accessed().
1584 mark_page_accessed(page
);
1586 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1588 * The preliminary mapping check is mainly to avoid the
1589 * pointless overhead of lock_page on the ZERO_PAGE
1590 * which might bounce very badly if there is contention.
1592 * If the page is already locked, we don't need to
1593 * handle it now - vmscan will handle it later if and
1594 * when it attempts to reclaim the page.
1596 if (page
->mapping
&& trylock_page(page
)) {
1597 lru_add_drain(); /* push cached pages to LRU */
1599 * Because we lock page here, and migration is
1600 * blocked by the pte's page reference, and we
1601 * know the page is still mapped, we don't even
1602 * need to check for file-cache page truncation.
1604 mlock_vma_page(page
);
1609 pte_unmap_unlock(ptep
, ptl
);
1614 pte_unmap_unlock(ptep
, ptl
);
1615 return ERR_PTR(-EFAULT
);
1618 pte_unmap_unlock(ptep
, ptl
);
1624 * When core dumping an enormous anonymous area that nobody
1625 * has touched so far, we don't want to allocate unnecessary pages or
1626 * page tables. Return error instead of NULL to skip handle_mm_fault,
1627 * then get_dump_page() will return NULL to leave a hole in the dump.
1628 * But we can only make this optimization where a hole would surely
1629 * be zero-filled if handle_mm_fault() actually did handle it.
1631 if ((flags
& FOLL_DUMP
) &&
1632 (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
1633 return ERR_PTR(-EFAULT
);
1637 static inline int stack_guard_page(struct vm_area_struct
*vma
, unsigned long addr
)
1639 return stack_guard_page_start(vma
, addr
) ||
1640 stack_guard_page_end(vma
, addr
+PAGE_SIZE
);
1644 * __get_user_pages() - pin user pages in memory
1645 * @tsk: task_struct of target task
1646 * @mm: mm_struct of target mm
1647 * @start: starting user address
1648 * @nr_pages: number of pages from start to pin
1649 * @gup_flags: flags modifying pin behaviour
1650 * @pages: array that receives pointers to the pages pinned.
1651 * Should be at least nr_pages long. Or NULL, if caller
1652 * only intends to ensure the pages are faulted in.
1653 * @vmas: array of pointers to vmas corresponding to each page.
1654 * Or NULL if the caller does not require them.
1655 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1657 * Returns number of pages pinned. This may be fewer than the number
1658 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1659 * were pinned, returns -errno. Each page returned must be released
1660 * with a put_page() call when it is finished with. vmas will only
1661 * remain valid while mmap_sem is held.
1663 * Must be called with mmap_sem held for read or write.
1665 * __get_user_pages walks a process's page tables and takes a reference to
1666 * each struct page that each user address corresponds to at a given
1667 * instant. That is, it takes the page that would be accessed if a user
1668 * thread accesses the given user virtual address at that instant.
1670 * This does not guarantee that the page exists in the user mappings when
1671 * __get_user_pages returns, and there may even be a completely different
1672 * page there in some cases (eg. if mmapped pagecache has been invalidated
1673 * and subsequently re faulted). However it does guarantee that the page
1674 * won't be freed completely. And mostly callers simply care that the page
1675 * contains data that was valid *at some point in time*. Typically, an IO
1676 * or similar operation cannot guarantee anything stronger anyway because
1677 * locks can't be held over the syscall boundary.
1679 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1680 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1681 * appropriate) must be called after the page is finished with, and
1682 * before put_page is called.
1684 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1685 * or mmap_sem contention, and if waiting is needed to pin all pages,
1686 * *@nonblocking will be set to 0.
1688 * In most cases, get_user_pages or get_user_pages_fast should be used
1689 * instead of __get_user_pages. __get_user_pages should be used only if
1690 * you need some special @gup_flags.
1692 long __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1693 unsigned long start
, unsigned long nr_pages
,
1694 unsigned int gup_flags
, struct page
**pages
,
1695 struct vm_area_struct
**vmas
, int *nonblocking
)
1698 unsigned long vm_flags
;
1699 unsigned int page_mask
;
1704 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
1707 * Require read or write permissions.
1708 * If FOLL_FORCE is set, we only require the "MAY" flags.
1710 vm_flags
= (gup_flags
& FOLL_WRITE
) ?
1711 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1712 vm_flags
&= (gup_flags
& FOLL_FORCE
) ?
1713 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1716 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1717 * would be called on PROT_NONE ranges. We must never invoke
1718 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1719 * page faults would unprotect the PROT_NONE ranges if
1720 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1721 * bitflag. So to avoid that, don't set FOLL_NUMA if
1722 * FOLL_FORCE is set.
1724 if (!(gup_flags
& FOLL_FORCE
))
1725 gup_flags
|= FOLL_NUMA
;
1730 struct vm_area_struct
*vma
;
1732 vma
= find_extend_vma(mm
, start
);
1733 if (!vma
&& in_gate_area(mm
, start
)) {
1734 unsigned long pg
= start
& PAGE_MASK
;
1740 /* user gate pages are read-only */
1741 if (gup_flags
& FOLL_WRITE
)
1742 return i
? : -EFAULT
;
1744 pgd
= pgd_offset_k(pg
);
1746 pgd
= pgd_offset_gate(mm
, pg
);
1747 BUG_ON(pgd_none(*pgd
));
1748 pud
= pud_offset(pgd
, pg
);
1749 BUG_ON(pud_none(*pud
));
1750 pmd
= pmd_offset(pud
, pg
);
1752 return i
? : -EFAULT
;
1753 VM_BUG_ON(pmd_trans_huge(*pmd
));
1754 pte
= pte_offset_map(pmd
, pg
);
1755 if (pte_none(*pte
)) {
1757 return i
? : -EFAULT
;
1759 vma
= get_gate_vma(mm
);
1763 page
= vm_normal_page(vma
, start
, *pte
);
1765 if (!(gup_flags
& FOLL_DUMP
) &&
1766 is_zero_pfn(pte_pfn(*pte
)))
1767 page
= pte_page(*pte
);
1770 return i
? : -EFAULT
;
1782 (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1783 !(vm_flags
& vma
->vm_flags
))
1784 return i
? : -EFAULT
;
1786 if (is_vm_hugetlb_page(vma
)) {
1787 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1788 &start
, &nr_pages
, i
, gup_flags
);
1794 unsigned int foll_flags
= gup_flags
;
1795 unsigned int page_increm
;
1798 * If we have a pending SIGKILL, don't keep faulting
1799 * pages and potentially allocating memory.
1801 if (unlikely(fatal_signal_pending(current
)))
1802 return i
? i
: -ERESTARTSYS
;
1805 while (!(page
= follow_page_mask(vma
, start
,
1806 foll_flags
, &page_mask
))) {
1808 unsigned int fault_flags
= 0;
1810 /* For mlock, just skip the stack guard page. */
1811 if (foll_flags
& FOLL_MLOCK
) {
1812 if (stack_guard_page(vma
, start
))
1815 if (foll_flags
& FOLL_WRITE
)
1816 fault_flags
|= FAULT_FLAG_WRITE
;
1818 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
1819 if (foll_flags
& FOLL_NOWAIT
)
1820 fault_flags
|= (FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
);
1822 ret
= handle_mm_fault(mm
, vma
, start
,
1825 if (ret
& VM_FAULT_ERROR
) {
1826 if (ret
& VM_FAULT_OOM
)
1827 return i
? i
: -ENOMEM
;
1828 if (ret
& (VM_FAULT_HWPOISON
|
1829 VM_FAULT_HWPOISON_LARGE
)) {
1832 else if (gup_flags
& FOLL_HWPOISON
)
1837 if (ret
& VM_FAULT_SIGBUS
)
1838 return i
? i
: -EFAULT
;
1843 if (ret
& VM_FAULT_MAJOR
)
1849 if (ret
& VM_FAULT_RETRY
) {
1856 * The VM_FAULT_WRITE bit tells us that
1857 * do_wp_page has broken COW when necessary,
1858 * even if maybe_mkwrite decided not to set
1859 * pte_write. We can thus safely do subsequent
1860 * page lookups as if they were reads. But only
1861 * do so when looping for pte_write is futile:
1862 * in some cases userspace may also be wanting
1863 * to write to the gotten user page, which a
1864 * read fault here might prevent (a readonly
1865 * page might get reCOWed by userspace write).
1867 if ((ret
& VM_FAULT_WRITE
) &&
1868 !(vma
->vm_flags
& VM_WRITE
))
1869 foll_flags
&= ~FOLL_WRITE
;
1874 return i
? i
: PTR_ERR(page
);
1878 flush_anon_page(vma
, page
, start
);
1879 flush_dcache_page(page
);
1887 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & page_mask
);
1888 if (page_increm
> nr_pages
)
1889 page_increm
= nr_pages
;
1891 start
+= page_increm
* PAGE_SIZE
;
1892 nr_pages
-= page_increm
;
1893 } while (nr_pages
&& start
< vma
->vm_end
);
1897 EXPORT_SYMBOL(__get_user_pages
);
1900 * fixup_user_fault() - manually resolve a user page fault
1901 * @tsk: the task_struct to use for page fault accounting, or
1902 * NULL if faults are not to be recorded.
1903 * @mm: mm_struct of target mm
1904 * @address: user address
1905 * @fault_flags:flags to pass down to handle_mm_fault()
1907 * This is meant to be called in the specific scenario where for locking reasons
1908 * we try to access user memory in atomic context (within a pagefault_disable()
1909 * section), this returns -EFAULT, and we want to resolve the user fault before
1912 * Typically this is meant to be used by the futex code.
1914 * The main difference with get_user_pages() is that this function will
1915 * unconditionally call handle_mm_fault() which will in turn perform all the
1916 * necessary SW fixup of the dirty and young bits in the PTE, while
1917 * handle_mm_fault() only guarantees to update these in the struct page.
1919 * This is important for some architectures where those bits also gate the
1920 * access permission to the page because they are maintained in software. On
1921 * such architectures, gup() will not be enough to make a subsequent access
1924 * This should be called with the mm_sem held for read.
1926 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
1927 unsigned long address
, unsigned int fault_flags
)
1929 struct vm_area_struct
*vma
;
1932 vma
= find_extend_vma(mm
, address
);
1933 if (!vma
|| address
< vma
->vm_start
)
1936 ret
= handle_mm_fault(mm
, vma
, address
, fault_flags
);
1937 if (ret
& VM_FAULT_ERROR
) {
1938 if (ret
& VM_FAULT_OOM
)
1940 if (ret
& (VM_FAULT_HWPOISON
| VM_FAULT_HWPOISON_LARGE
))
1942 if (ret
& VM_FAULT_SIGBUS
)
1947 if (ret
& VM_FAULT_MAJOR
)
1956 * get_user_pages() - pin user pages in memory
1957 * @tsk: the task_struct to use for page fault accounting, or
1958 * NULL if faults are not to be recorded.
1959 * @mm: mm_struct of target mm
1960 * @start: starting user address
1961 * @nr_pages: number of pages from start to pin
1962 * @write: whether pages will be written to by the caller
1963 * @force: whether to force write access even if user mapping is
1964 * readonly. This will result in the page being COWed even
1965 * in MAP_SHARED mappings. You do not want this.
1966 * @pages: array that receives pointers to the pages pinned.
1967 * Should be at least nr_pages long. Or NULL, if caller
1968 * only intends to ensure the pages are faulted in.
1969 * @vmas: array of pointers to vmas corresponding to each page.
1970 * Or NULL if the caller does not require them.
1972 * Returns number of pages pinned. This may be fewer than the number
1973 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1974 * were pinned, returns -errno. Each page returned must be released
1975 * with a put_page() call when it is finished with. vmas will only
1976 * remain valid while mmap_sem is held.
1978 * Must be called with mmap_sem held for read or write.
1980 * get_user_pages walks a process's page tables and takes a reference to
1981 * each struct page that each user address corresponds to at a given
1982 * instant. That is, it takes the page that would be accessed if a user
1983 * thread accesses the given user virtual address at that instant.
1985 * This does not guarantee that the page exists in the user mappings when
1986 * get_user_pages returns, and there may even be a completely different
1987 * page there in some cases (eg. if mmapped pagecache has been invalidated
1988 * and subsequently re faulted). However it does guarantee that the page
1989 * won't be freed completely. And mostly callers simply care that the page
1990 * contains data that was valid *at some point in time*. Typically, an IO
1991 * or similar operation cannot guarantee anything stronger anyway because
1992 * locks can't be held over the syscall boundary.
1994 * If write=0, the page must not be written to. If the page is written to,
1995 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1996 * after the page is finished with, and before put_page is called.
1998 * get_user_pages is typically used for fewer-copy IO operations, to get a
1999 * handle on the memory by some means other than accesses via the user virtual
2000 * addresses. The pages may be submitted for DMA to devices or accessed via
2001 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2002 * use the correct cache flushing APIs.
2004 * See also get_user_pages_fast, for performance critical applications.
2006 long get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
2007 unsigned long start
, unsigned long nr_pages
, int write
,
2008 int force
, struct page
**pages
, struct vm_area_struct
**vmas
)
2010 int flags
= FOLL_TOUCH
;
2015 flags
|= FOLL_WRITE
;
2017 flags
|= FOLL_FORCE
;
2019 return __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
, vmas
,
2022 EXPORT_SYMBOL(get_user_pages
);
2025 * get_dump_page() - pin user page in memory while writing it to core dump
2026 * @addr: user address
2028 * Returns struct page pointer of user page pinned for dump,
2029 * to be freed afterwards by page_cache_release() or put_page().
2031 * Returns NULL on any kind of failure - a hole must then be inserted into
2032 * the corefile, to preserve alignment with its headers; and also returns
2033 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2034 * allowing a hole to be left in the corefile to save diskspace.
2036 * Called without mmap_sem, but after all other threads have been killed.
2038 #ifdef CONFIG_ELF_CORE
2039 struct page
*get_dump_page(unsigned long addr
)
2041 struct vm_area_struct
*vma
;
2044 if (__get_user_pages(current
, current
->mm
, addr
, 1,
2045 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
2048 flush_cache_page(vma
, addr
, page_to_pfn(page
));
2051 #endif /* CONFIG_ELF_CORE */
2053 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
2056 pgd_t
* pgd
= pgd_offset(mm
, addr
);
2057 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
2059 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
2061 VM_BUG_ON(pmd_trans_huge(*pmd
));
2062 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
2069 * This is the old fallback for page remapping.
2071 * For historical reasons, it only allows reserved pages. Only
2072 * old drivers should use this, and they needed to mark their
2073 * pages reserved for the old functions anyway.
2075 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2076 struct page
*page
, pgprot_t prot
)
2078 struct mm_struct
*mm
= vma
->vm_mm
;
2087 flush_dcache_page(page
);
2088 pte
= get_locked_pte(mm
, addr
, &ptl
);
2092 if (!pte_none(*pte
))
2095 /* Ok, finally just insert the thing.. */
2097 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
2098 page_add_file_rmap(page
);
2099 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
2102 pte_unmap_unlock(pte
, ptl
);
2105 pte_unmap_unlock(pte
, ptl
);
2111 * vm_insert_page - insert single page into user vma
2112 * @vma: user vma to map to
2113 * @addr: target user address of this page
2114 * @page: source kernel page
2116 * This allows drivers to insert individual pages they've allocated
2119 * The page has to be a nice clean _individual_ kernel allocation.
2120 * If you allocate a compound page, you need to have marked it as
2121 * such (__GFP_COMP), or manually just split the page up yourself
2122 * (see split_page()).
2124 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2125 * took an arbitrary page protection parameter. This doesn't allow
2126 * that. Your vma protection will have to be set up correctly, which
2127 * means that if you want a shared writable mapping, you'd better
2128 * ask for a shared writable mapping!
2130 * The page does not need to be reserved.
2132 * Usually this function is called from f_op->mmap() handler
2133 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2134 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2135 * function from other places, for example from page-fault handler.
2137 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2140 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2142 if (!page_count(page
))
2144 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
2145 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
2146 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
2147 vma
->vm_flags
|= VM_MIXEDMAP
;
2149 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2151 EXPORT_SYMBOL(vm_insert_page
);
2153 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2154 unsigned long pfn
, pgprot_t prot
)
2156 struct mm_struct
*mm
= vma
->vm_mm
;
2162 pte
= get_locked_pte(mm
, addr
, &ptl
);
2166 if (!pte_none(*pte
))
2169 /* Ok, finally just insert the thing.. */
2170 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
2171 set_pte_at(mm
, addr
, pte
, entry
);
2172 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
2176 pte_unmap_unlock(pte
, ptl
);
2182 * vm_insert_pfn - insert single pfn into user vma
2183 * @vma: user vma to map to
2184 * @addr: target user address of this page
2185 * @pfn: source kernel pfn
2187 * Similar to vm_insert_page, this allows drivers to insert individual pages
2188 * they've allocated into a user vma. Same comments apply.
2190 * This function should only be called from a vm_ops->fault handler, and
2191 * in that case the handler should return NULL.
2193 * vma cannot be a COW mapping.
2195 * As this is called only for pages that do not currently exist, we
2196 * do not need to flush old virtual caches or the TLB.
2198 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2202 pgprot_t pgprot
= vma
->vm_page_prot
;
2204 * Technically, architectures with pte_special can avoid all these
2205 * restrictions (same for remap_pfn_range). However we would like
2206 * consistency in testing and feature parity among all, so we should
2207 * try to keep these invariants in place for everybody.
2209 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
2210 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
2211 (VM_PFNMAP
|VM_MIXEDMAP
));
2212 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
2213 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
2215 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2217 if (track_pfn_insert(vma
, &pgprot
, pfn
))
2220 ret
= insert_pfn(vma
, addr
, pfn
, pgprot
);
2224 EXPORT_SYMBOL(vm_insert_pfn
);
2226 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2229 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
2231 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2235 * If we don't have pte special, then we have to use the pfn_valid()
2236 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2237 * refcount the page if pfn_valid is true (hence insert_page rather
2238 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2239 * without pte special, it would there be refcounted as a normal page.
2241 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
2244 page
= pfn_to_page(pfn
);
2245 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2247 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
2249 EXPORT_SYMBOL(vm_insert_mixed
);
2252 * maps a range of physical memory into the requested pages. the old
2253 * mappings are removed. any references to nonexistent pages results
2254 * in null mappings (currently treated as "copy-on-access")
2256 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2257 unsigned long addr
, unsigned long end
,
2258 unsigned long pfn
, pgprot_t prot
)
2263 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2266 arch_enter_lazy_mmu_mode();
2268 BUG_ON(!pte_none(*pte
));
2269 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2271 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2272 arch_leave_lazy_mmu_mode();
2273 pte_unmap_unlock(pte
- 1, ptl
);
2277 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2278 unsigned long addr
, unsigned long end
,
2279 unsigned long pfn
, pgprot_t prot
)
2284 pfn
-= addr
>> PAGE_SHIFT
;
2285 pmd
= pmd_alloc(mm
, pud
, addr
);
2288 VM_BUG_ON(pmd_trans_huge(*pmd
));
2290 next
= pmd_addr_end(addr
, end
);
2291 if (remap_pte_range(mm
, pmd
, addr
, next
,
2292 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
2294 } while (pmd
++, addr
= next
, addr
!= end
);
2298 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2299 unsigned long addr
, unsigned long end
,
2300 unsigned long pfn
, pgprot_t prot
)
2305 pfn
-= addr
>> PAGE_SHIFT
;
2306 pud
= pud_alloc(mm
, pgd
, addr
);
2310 next
= pud_addr_end(addr
, end
);
2311 if (remap_pmd_range(mm
, pud
, addr
, next
,
2312 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
2314 } while (pud
++, addr
= next
, addr
!= end
);
2319 * remap_pfn_range - remap kernel memory to userspace
2320 * @vma: user vma to map to
2321 * @addr: target user address to start at
2322 * @pfn: physical address of kernel memory
2323 * @size: size of map area
2324 * @prot: page protection flags for this mapping
2326 * Note: this is only safe if the mm semaphore is held when called.
2328 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2329 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2333 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2334 struct mm_struct
*mm
= vma
->vm_mm
;
2338 * Physically remapped pages are special. Tell the
2339 * rest of the world about it:
2340 * VM_IO tells people not to look at these pages
2341 * (accesses can have side effects).
2342 * VM_PFNMAP tells the core MM that the base pages are just
2343 * raw PFN mappings, and do not have a "struct page" associated
2346 * Disable vma merging and expanding with mremap().
2348 * Omit vma from core dump, even when VM_IO turned off.
2350 * There's a horrible special case to handle copy-on-write
2351 * behaviour that some programs depend on. We mark the "original"
2352 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2353 * See vm_normal_page() for details.
2355 if (is_cow_mapping(vma
->vm_flags
)) {
2356 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2358 vma
->vm_pgoff
= pfn
;
2361 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
2365 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
2367 BUG_ON(addr
>= end
);
2368 pfn
-= addr
>> PAGE_SHIFT
;
2369 pgd
= pgd_offset(mm
, addr
);
2370 flush_cache_range(vma
, addr
, end
);
2372 next
= pgd_addr_end(addr
, end
);
2373 err
= remap_pud_range(mm
, pgd
, addr
, next
,
2374 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2377 } while (pgd
++, addr
= next
, addr
!= end
);
2380 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
));
2384 EXPORT_SYMBOL(remap_pfn_range
);
2387 * vm_iomap_memory - remap memory to userspace
2388 * @vma: user vma to map to
2389 * @start: start of area
2390 * @len: size of area
2392 * This is a simplified io_remap_pfn_range() for common driver use. The
2393 * driver just needs to give us the physical memory range to be mapped,
2394 * we'll figure out the rest from the vma information.
2396 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2397 * whatever write-combining details or similar.
2399 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2401 unsigned long vm_len
, pfn
, pages
;
2403 /* Check that the physical memory area passed in looks valid */
2404 if (start
+ len
< start
)
2407 * You *really* shouldn't map things that aren't page-aligned,
2408 * but we've historically allowed it because IO memory might
2409 * just have smaller alignment.
2411 len
+= start
& ~PAGE_MASK
;
2412 pfn
= start
>> PAGE_SHIFT
;
2413 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2414 if (pfn
+ pages
< pfn
)
2417 /* We start the mapping 'vm_pgoff' pages into the area */
2418 if (vma
->vm_pgoff
> pages
)
2420 pfn
+= vma
->vm_pgoff
;
2421 pages
-= vma
->vm_pgoff
;
2423 /* Can we fit all of the mapping? */
2424 vm_len
= vma
->vm_end
- vma
->vm_start
;
2425 if (vm_len
>> PAGE_SHIFT
> pages
)
2428 /* Ok, let it rip */
2429 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2431 EXPORT_SYMBOL(vm_iomap_memory
);
2433 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2434 unsigned long addr
, unsigned long end
,
2435 pte_fn_t fn
, void *data
)
2440 spinlock_t
*uninitialized_var(ptl
);
2442 pte
= (mm
== &init_mm
) ?
2443 pte_alloc_kernel(pmd
, addr
) :
2444 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2448 BUG_ON(pmd_huge(*pmd
));
2450 arch_enter_lazy_mmu_mode();
2452 token
= pmd_pgtable(*pmd
);
2455 err
= fn(pte
++, token
, addr
, data
);
2458 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2460 arch_leave_lazy_mmu_mode();
2463 pte_unmap_unlock(pte
-1, ptl
);
2467 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2468 unsigned long addr
, unsigned long end
,
2469 pte_fn_t fn
, void *data
)
2475 BUG_ON(pud_huge(*pud
));
2477 pmd
= pmd_alloc(mm
, pud
, addr
);
2481 next
= pmd_addr_end(addr
, end
);
2482 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
2485 } while (pmd
++, addr
= next
, addr
!= end
);
2489 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2490 unsigned long addr
, unsigned long end
,
2491 pte_fn_t fn
, void *data
)
2497 pud
= pud_alloc(mm
, pgd
, addr
);
2501 next
= pud_addr_end(addr
, end
);
2502 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
2505 } while (pud
++, addr
= next
, addr
!= end
);
2510 * Scan a region of virtual memory, filling in page tables as necessary
2511 * and calling a provided function on each leaf page table.
2513 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2514 unsigned long size
, pte_fn_t fn
, void *data
)
2518 unsigned long end
= addr
+ size
;
2521 BUG_ON(addr
>= end
);
2522 pgd
= pgd_offset(mm
, addr
);
2524 next
= pgd_addr_end(addr
, end
);
2525 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
2528 } while (pgd
++, addr
= next
, addr
!= end
);
2532 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2535 * handle_pte_fault chooses page fault handler according to an entry
2536 * which was read non-atomically. Before making any commitment, on
2537 * those architectures or configurations (e.g. i386 with PAE) which
2538 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2539 * must check under lock before unmapping the pte and proceeding
2540 * (but do_wp_page is only called after already making such a check;
2541 * and do_anonymous_page can safely check later on).
2543 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2544 pte_t
*page_table
, pte_t orig_pte
)
2547 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2548 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2549 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2551 same
= pte_same(*page_table
, orig_pte
);
2555 pte_unmap(page_table
);
2559 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
2562 * If the source page was a PFN mapping, we don't have
2563 * a "struct page" for it. We do a best-effort copy by
2564 * just copying from the original user address. If that
2565 * fails, we just zero-fill it. Live with it.
2567 if (unlikely(!src
)) {
2568 void *kaddr
= kmap_atomic(dst
);
2569 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
2572 * This really shouldn't fail, because the page is there
2573 * in the page tables. But it might just be unreadable,
2574 * in which case we just give up and fill the result with
2577 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
2579 kunmap_atomic(kaddr
);
2580 flush_dcache_page(dst
);
2582 copy_user_highpage(dst
, src
, va
, vma
);
2586 * This routine handles present pages, when users try to write
2587 * to a shared page. It is done by copying the page to a new address
2588 * and decrementing the shared-page counter for the old page.
2590 * Note that this routine assumes that the protection checks have been
2591 * done by the caller (the low-level page fault routine in most cases).
2592 * Thus we can safely just mark it writable once we've done any necessary
2595 * We also mark the page dirty at this point even though the page will
2596 * change only once the write actually happens. This avoids a few races,
2597 * and potentially makes it more efficient.
2599 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2600 * but allow concurrent faults), with pte both mapped and locked.
2601 * We return with mmap_sem still held, but pte unmapped and unlocked.
2603 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2604 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2605 spinlock_t
*ptl
, pte_t orig_pte
)
2608 struct page
*old_page
, *new_page
= NULL
;
2611 int page_mkwrite
= 0;
2612 struct page
*dirty_page
= NULL
;
2613 unsigned long mmun_start
= 0; /* For mmu_notifiers */
2614 unsigned long mmun_end
= 0; /* For mmu_notifiers */
2616 old_page
= vm_normal_page(vma
, address
, orig_pte
);
2619 * VM_MIXEDMAP !pfn_valid() case
2621 * We should not cow pages in a shared writeable mapping.
2622 * Just mark the pages writable as we can't do any dirty
2623 * accounting on raw pfn maps.
2625 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2626 (VM_WRITE
|VM_SHARED
))
2632 * Take out anonymous pages first, anonymous shared vmas are
2633 * not dirty accountable.
2635 if (PageAnon(old_page
) && !PageKsm(old_page
)) {
2636 if (!trylock_page(old_page
)) {
2637 page_cache_get(old_page
);
2638 pte_unmap_unlock(page_table
, ptl
);
2639 lock_page(old_page
);
2640 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2642 if (!pte_same(*page_table
, orig_pte
)) {
2643 unlock_page(old_page
);
2646 page_cache_release(old_page
);
2648 if (reuse_swap_page(old_page
)) {
2650 * The page is all ours. Move it to our anon_vma so
2651 * the rmap code will not search our parent or siblings.
2652 * Protected against the rmap code by the page lock.
2654 page_move_anon_rmap(old_page
, vma
, address
);
2655 unlock_page(old_page
);
2658 unlock_page(old_page
);
2659 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2660 (VM_WRITE
|VM_SHARED
))) {
2662 * Only catch write-faults on shared writable pages,
2663 * read-only shared pages can get COWed by
2664 * get_user_pages(.write=1, .force=1).
2666 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2667 struct vm_fault vmf
;
2670 vmf
.virtual_address
= (void __user
*)(address
&
2672 vmf
.pgoff
= old_page
->index
;
2673 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2674 vmf
.page
= old_page
;
2677 * Notify the address space that the page is about to
2678 * become writable so that it can prohibit this or wait
2679 * for the page to get into an appropriate state.
2681 * We do this without the lock held, so that it can
2682 * sleep if it needs to.
2684 page_cache_get(old_page
);
2685 pte_unmap_unlock(page_table
, ptl
);
2687 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
2689 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
2691 goto unwritable_page
;
2693 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
2694 lock_page(old_page
);
2695 if (!old_page
->mapping
) {
2696 ret
= 0; /* retry the fault */
2697 unlock_page(old_page
);
2698 goto unwritable_page
;
2701 VM_BUG_ON(!PageLocked(old_page
));
2704 * Since we dropped the lock we need to revalidate
2705 * the PTE as someone else may have changed it. If
2706 * they did, we just return, as we can count on the
2707 * MMU to tell us if they didn't also make it writable.
2709 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2711 if (!pte_same(*page_table
, orig_pte
)) {
2712 unlock_page(old_page
);
2718 dirty_page
= old_page
;
2719 get_page(dirty_page
);
2723 * Clear the pages cpupid information as the existing
2724 * information potentially belongs to a now completely
2725 * unrelated process.
2728 page_cpupid_xchg_last(old_page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2730 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2731 entry
= pte_mkyoung(orig_pte
);
2732 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2733 if (ptep_set_access_flags(vma
, address
, page_table
, entry
,1))
2734 update_mmu_cache(vma
, address
, page_table
);
2735 pte_unmap_unlock(page_table
, ptl
);
2736 ret
|= VM_FAULT_WRITE
;
2742 * Yes, Virginia, this is actually required to prevent a race
2743 * with clear_page_dirty_for_io() from clearing the page dirty
2744 * bit after it clear all dirty ptes, but before a racing
2745 * do_wp_page installs a dirty pte.
2747 * __do_fault is protected similarly.
2749 if (!page_mkwrite
) {
2750 wait_on_page_locked(dirty_page
);
2751 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2752 /* file_update_time outside page_lock */
2754 file_update_time(vma
->vm_file
);
2756 put_page(dirty_page
);
2758 struct address_space
*mapping
= dirty_page
->mapping
;
2760 set_page_dirty(dirty_page
);
2761 unlock_page(dirty_page
);
2762 page_cache_release(dirty_page
);
2765 * Some device drivers do not set page.mapping
2766 * but still dirty their pages
2768 balance_dirty_pages_ratelimited(mapping
);
2776 * Ok, we need to copy. Oh, well..
2778 page_cache_get(old_page
);
2780 pte_unmap_unlock(page_table
, ptl
);
2782 if (unlikely(anon_vma_prepare(vma
)))
2785 if (is_zero_pfn(pte_pfn(orig_pte
))) {
2786 new_page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2790 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2793 cow_user_page(new_page
, old_page
, address
, vma
);
2795 __SetPageUptodate(new_page
);
2797 if (mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
))
2800 mmun_start
= address
& PAGE_MASK
;
2801 mmun_end
= mmun_start
+ PAGE_SIZE
;
2802 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2805 * Re-check the pte - we dropped the lock
2807 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2808 if (likely(pte_same(*page_table
, orig_pte
))) {
2810 if (!PageAnon(old_page
)) {
2811 dec_mm_counter_fast(mm
, MM_FILEPAGES
);
2812 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2815 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2816 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2817 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2818 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2820 * Clear the pte entry and flush it first, before updating the
2821 * pte with the new entry. This will avoid a race condition
2822 * seen in the presence of one thread doing SMC and another
2825 ptep_clear_flush(vma
, address
, page_table
);
2826 page_add_new_anon_rmap(new_page
, vma
, address
);
2828 * We call the notify macro here because, when using secondary
2829 * mmu page tables (such as kvm shadow page tables), we want the
2830 * new page to be mapped directly into the secondary page table.
2832 set_pte_at_notify(mm
, address
, page_table
, entry
);
2833 update_mmu_cache(vma
, address
, page_table
);
2836 * Only after switching the pte to the new page may
2837 * we remove the mapcount here. Otherwise another
2838 * process may come and find the rmap count decremented
2839 * before the pte is switched to the new page, and
2840 * "reuse" the old page writing into it while our pte
2841 * here still points into it and can be read by other
2844 * The critical issue is to order this
2845 * page_remove_rmap with the ptp_clear_flush above.
2846 * Those stores are ordered by (if nothing else,)
2847 * the barrier present in the atomic_add_negative
2848 * in page_remove_rmap.
2850 * Then the TLB flush in ptep_clear_flush ensures that
2851 * no process can access the old page before the
2852 * decremented mapcount is visible. And the old page
2853 * cannot be reused until after the decremented
2854 * mapcount is visible. So transitively, TLBs to
2855 * old page will be flushed before it can be reused.
2857 page_remove_rmap(old_page
);
2860 /* Free the old page.. */
2861 new_page
= old_page
;
2862 ret
|= VM_FAULT_WRITE
;
2864 mem_cgroup_uncharge_page(new_page
);
2867 page_cache_release(new_page
);
2869 pte_unmap_unlock(page_table
, ptl
);
2870 if (mmun_end
> mmun_start
)
2871 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2874 * Don't let another task, with possibly unlocked vma,
2875 * keep the mlocked page.
2877 if ((ret
& VM_FAULT_WRITE
) && (vma
->vm_flags
& VM_LOCKED
)) {
2878 lock_page(old_page
); /* LRU manipulation */
2879 munlock_vma_page(old_page
);
2880 unlock_page(old_page
);
2882 page_cache_release(old_page
);
2886 page_cache_release(new_page
);
2889 page_cache_release(old_page
);
2890 return VM_FAULT_OOM
;
2893 page_cache_release(old_page
);
2897 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2898 unsigned long start_addr
, unsigned long end_addr
,
2899 struct zap_details
*details
)
2901 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2904 static inline void unmap_mapping_range_tree(struct rb_root
*root
,
2905 struct zap_details
*details
)
2907 struct vm_area_struct
*vma
;
2908 pgoff_t vba
, vea
, zba
, zea
;
2910 vma_interval_tree_foreach(vma
, root
,
2911 details
->first_index
, details
->last_index
) {
2913 vba
= vma
->vm_pgoff
;
2914 vea
= vba
+ vma_pages(vma
) - 1;
2915 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2916 zba
= details
->first_index
;
2919 zea
= details
->last_index
;
2923 unmap_mapping_range_vma(vma
,
2924 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2925 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2930 static inline void unmap_mapping_range_list(struct list_head
*head
,
2931 struct zap_details
*details
)
2933 struct vm_area_struct
*vma
;
2936 * In nonlinear VMAs there is no correspondence between virtual address
2937 * offset and file offset. So we must perform an exhaustive search
2938 * across *all* the pages in each nonlinear VMA, not just the pages
2939 * whose virtual address lies outside the file truncation point.
2941 list_for_each_entry(vma
, head
, shared
.nonlinear
) {
2942 details
->nonlinear_vma
= vma
;
2943 unmap_mapping_range_vma(vma
, vma
->vm_start
, vma
->vm_end
, details
);
2948 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2949 * @mapping: the address space containing mmaps to be unmapped.
2950 * @holebegin: byte in first page to unmap, relative to the start of
2951 * the underlying file. This will be rounded down to a PAGE_SIZE
2952 * boundary. Note that this is different from truncate_pagecache(), which
2953 * must keep the partial page. In contrast, we must get rid of
2955 * @holelen: size of prospective hole in bytes. This will be rounded
2956 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2958 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2959 * but 0 when invalidating pagecache, don't throw away private data.
2961 void unmap_mapping_range(struct address_space
*mapping
,
2962 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2964 struct zap_details details
;
2965 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2966 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2968 /* Check for overflow. */
2969 if (sizeof(holelen
) > sizeof(hlen
)) {
2971 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2972 if (holeend
& ~(long long)ULONG_MAX
)
2973 hlen
= ULONG_MAX
- hba
+ 1;
2976 details
.check_mapping
= even_cows
? NULL
: mapping
;
2977 details
.nonlinear_vma
= NULL
;
2978 details
.first_index
= hba
;
2979 details
.last_index
= hba
+ hlen
- 1;
2980 if (details
.last_index
< details
.first_index
)
2981 details
.last_index
= ULONG_MAX
;
2984 mutex_lock(&mapping
->i_mmap_mutex
);
2985 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
)))
2986 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2987 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
2988 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
2989 mutex_unlock(&mapping
->i_mmap_mutex
);
2991 EXPORT_SYMBOL(unmap_mapping_range
);
2994 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2995 * but allow concurrent faults), and pte mapped but not yet locked.
2996 * We return with mmap_sem still held, but pte unmapped and unlocked.
2998 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2999 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3000 unsigned int flags
, pte_t orig_pte
)
3003 struct page
*page
, *swapcache
;
3007 struct mem_cgroup
*ptr
;
3011 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
3014 entry
= pte_to_swp_entry(orig_pte
);
3015 if (unlikely(non_swap_entry(entry
))) {
3016 if (is_migration_entry(entry
)) {
3017 migration_entry_wait(mm
, pmd
, address
);
3018 } else if (is_hwpoison_entry(entry
)) {
3019 ret
= VM_FAULT_HWPOISON
;
3021 print_bad_pte(vma
, address
, orig_pte
, NULL
);
3022 ret
= VM_FAULT_SIGBUS
;
3026 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
3027 page
= lookup_swap_cache(entry
);
3029 page
= swapin_readahead(entry
,
3030 GFP_HIGHUSER_MOVABLE
, vma
, address
);
3033 * Back out if somebody else faulted in this pte
3034 * while we released the pte lock.
3036 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3037 if (likely(pte_same(*page_table
, orig_pte
)))
3039 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3043 /* Had to read the page from swap area: Major fault */
3044 ret
= VM_FAULT_MAJOR
;
3045 count_vm_event(PGMAJFAULT
);
3046 mem_cgroup_count_vm_event(mm
, PGMAJFAULT
);
3047 } else if (PageHWPoison(page
)) {
3049 * hwpoisoned dirty swapcache pages are kept for killing
3050 * owner processes (which may be unknown at hwpoison time)
3052 ret
= VM_FAULT_HWPOISON
;
3053 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3059 locked
= lock_page_or_retry(page
, mm
, flags
);
3061 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3063 ret
|= VM_FAULT_RETRY
;
3068 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3069 * release the swapcache from under us. The page pin, and pte_same
3070 * test below, are not enough to exclude that. Even if it is still
3071 * swapcache, we need to check that the page's swap has not changed.
3073 if (unlikely(!PageSwapCache(page
) || page_private(page
) != entry
.val
))
3076 page
= ksm_might_need_to_copy(page
, vma
, address
);
3077 if (unlikely(!page
)) {
3083 if (mem_cgroup_try_charge_swapin(mm
, page
, GFP_KERNEL
, &ptr
)) {
3089 * Back out if somebody else already faulted in this pte.
3091 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3092 if (unlikely(!pte_same(*page_table
, orig_pte
)))
3095 if (unlikely(!PageUptodate(page
))) {
3096 ret
= VM_FAULT_SIGBUS
;
3101 * The page isn't present yet, go ahead with the fault.
3103 * Be careful about the sequence of operations here.
3104 * To get its accounting right, reuse_swap_page() must be called
3105 * while the page is counted on swap but not yet in mapcount i.e.
3106 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3107 * must be called after the swap_free(), or it will never succeed.
3108 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3109 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3110 * in page->private. In this case, a record in swap_cgroup is silently
3111 * discarded at swap_free().
3114 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3115 dec_mm_counter_fast(mm
, MM_SWAPENTS
);
3116 pte
= mk_pte(page
, vma
->vm_page_prot
);
3117 if ((flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
)) {
3118 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3119 flags
&= ~FAULT_FLAG_WRITE
;
3120 ret
|= VM_FAULT_WRITE
;
3123 flush_icache_page(vma
, page
);
3124 if (pte_swp_soft_dirty(orig_pte
))
3125 pte
= pte_mksoft_dirty(pte
);
3126 set_pte_at(mm
, address
, page_table
, pte
);
3127 if (page
== swapcache
)
3128 do_page_add_anon_rmap(page
, vma
, address
, exclusive
);
3129 else /* ksm created a completely new copy */
3130 page_add_new_anon_rmap(page
, vma
, address
);
3131 /* It's better to call commit-charge after rmap is established */
3132 mem_cgroup_commit_charge_swapin(page
, ptr
);
3135 if (vm_swap_full() || (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
3136 try_to_free_swap(page
);
3138 if (page
!= swapcache
) {
3140 * Hold the lock to avoid the swap entry to be reused
3141 * until we take the PT lock for the pte_same() check
3142 * (to avoid false positives from pte_same). For
3143 * further safety release the lock after the swap_free
3144 * so that the swap count won't change under a
3145 * parallel locked swapcache.
3147 unlock_page(swapcache
);
3148 page_cache_release(swapcache
);
3151 if (flags
& FAULT_FLAG_WRITE
) {
3152 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
3153 if (ret
& VM_FAULT_ERROR
)
3154 ret
&= VM_FAULT_ERROR
;
3158 /* No need to invalidate - it was non-present before */
3159 update_mmu_cache(vma
, address
, page_table
);
3161 pte_unmap_unlock(page_table
, ptl
);
3165 mem_cgroup_cancel_charge_swapin(ptr
);
3166 pte_unmap_unlock(page_table
, ptl
);
3170 page_cache_release(page
);
3171 if (page
!= swapcache
) {
3172 unlock_page(swapcache
);
3173 page_cache_release(swapcache
);
3179 * This is like a special single-page "expand_{down|up}wards()",
3180 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3181 * doesn't hit another vma.
3183 static inline int check_stack_guard_page(struct vm_area_struct
*vma
, unsigned long address
)
3185 address
&= PAGE_MASK
;
3186 if ((vma
->vm_flags
& VM_GROWSDOWN
) && address
== vma
->vm_start
) {
3187 struct vm_area_struct
*prev
= vma
->vm_prev
;
3190 * Is there a mapping abutting this one below?
3192 * That's only ok if it's the same stack mapping
3193 * that has gotten split..
3195 if (prev
&& prev
->vm_end
== address
)
3196 return prev
->vm_flags
& VM_GROWSDOWN
? 0 : -ENOMEM
;
3198 expand_downwards(vma
, address
- PAGE_SIZE
);
3200 if ((vma
->vm_flags
& VM_GROWSUP
) && address
+ PAGE_SIZE
== vma
->vm_end
) {
3201 struct vm_area_struct
*next
= vma
->vm_next
;
3203 /* As VM_GROWSDOWN but s/below/above/ */
3204 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
)
3205 return next
->vm_flags
& VM_GROWSUP
? 0 : -ENOMEM
;
3207 expand_upwards(vma
, address
+ PAGE_SIZE
);
3213 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3214 * but allow concurrent faults), and pte mapped but not yet locked.
3215 * We return with mmap_sem still held, but pte unmapped and unlocked.
3217 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3218 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3225 pte_unmap(page_table
);
3227 /* Check if we need to add a guard page to the stack */
3228 if (check_stack_guard_page(vma
, address
) < 0)
3229 return VM_FAULT_SIGBUS
;
3231 /* Use the zero-page for reads */
3232 if (!(flags
& FAULT_FLAG_WRITE
)) {
3233 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(address
),
3234 vma
->vm_page_prot
));
3235 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3236 if (!pte_none(*page_table
))
3241 /* Allocate our own private page. */
3242 if (unlikely(anon_vma_prepare(vma
)))
3244 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
3248 * The memory barrier inside __SetPageUptodate makes sure that
3249 * preceeding stores to the page contents become visible before
3250 * the set_pte_at() write.
3252 __SetPageUptodate(page
);
3254 if (mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
))
3257 entry
= mk_pte(page
, vma
->vm_page_prot
);
3258 if (vma
->vm_flags
& VM_WRITE
)
3259 entry
= pte_mkwrite(pte_mkdirty(entry
));
3261 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3262 if (!pte_none(*page_table
))
3265 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3266 page_add_new_anon_rmap(page
, vma
, address
);
3268 set_pte_at(mm
, address
, page_table
, entry
);
3270 /* No need to invalidate - it was non-present before */
3271 update_mmu_cache(vma
, address
, page_table
);
3273 pte_unmap_unlock(page_table
, ptl
);
3276 mem_cgroup_uncharge_page(page
);
3277 page_cache_release(page
);
3280 page_cache_release(page
);
3282 return VM_FAULT_OOM
;
3286 * __do_fault() tries to create a new page mapping. It aggressively
3287 * tries to share with existing pages, but makes a separate copy if
3288 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3289 * the next page fault.
3291 * As this is called only for pages that do not currently exist, we
3292 * do not need to flush old virtual caches or the TLB.
3294 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3295 * but allow concurrent faults), and pte neither mapped nor locked.
3296 * We return with mmap_sem still held, but pte unmapped and unlocked.
3298 static int __do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3299 unsigned long address
, pmd_t
*pmd
,
3300 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3305 struct page
*cow_page
;
3308 struct page
*dirty_page
= NULL
;
3309 struct vm_fault vmf
;
3311 int page_mkwrite
= 0;
3314 * If we do COW later, allocate page befor taking lock_page()
3315 * on the file cache page. This will reduce lock holding time.
3317 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3319 if (unlikely(anon_vma_prepare(vma
)))
3320 return VM_FAULT_OOM
;
3322 cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
3324 return VM_FAULT_OOM
;
3326 if (mem_cgroup_newpage_charge(cow_page
, mm
, GFP_KERNEL
)) {
3327 page_cache_release(cow_page
);
3328 return VM_FAULT_OOM
;
3333 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
3338 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
3339 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
3343 if (unlikely(PageHWPoison(vmf
.page
))) {
3344 if (ret
& VM_FAULT_LOCKED
)
3345 unlock_page(vmf
.page
);
3346 ret
= VM_FAULT_HWPOISON
;
3351 * For consistency in subsequent calls, make the faulted page always
3354 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3355 lock_page(vmf
.page
);
3357 VM_BUG_ON(!PageLocked(vmf
.page
));
3360 * Should we do an early C-O-W break?
3363 if (flags
& FAULT_FLAG_WRITE
) {
3364 if (!(vma
->vm_flags
& VM_SHARED
)) {
3367 copy_user_highpage(page
, vmf
.page
, address
, vma
);
3368 __SetPageUptodate(page
);
3371 * If the page will be shareable, see if the backing
3372 * address space wants to know that the page is about
3373 * to become writable
3375 if (vma
->vm_ops
->page_mkwrite
) {
3379 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
3380 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
3382 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3384 goto unwritable_page
;
3386 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
3388 if (!page
->mapping
) {
3389 ret
= 0; /* retry the fault */
3391 goto unwritable_page
;
3394 VM_BUG_ON(!PageLocked(page
));
3401 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3404 * This silly early PAGE_DIRTY setting removes a race
3405 * due to the bad i386 page protection. But it's valid
3406 * for other architectures too.
3408 * Note that if FAULT_FLAG_WRITE is set, we either now have
3409 * an exclusive copy of the page, or this is a shared mapping,
3410 * so we can make it writable and dirty to avoid having to
3411 * handle that later.
3413 /* Only go through if we didn't race with anybody else... */
3414 if (likely(pte_same(*page_table
, orig_pte
))) {
3415 flush_icache_page(vma
, page
);
3416 entry
= mk_pte(page
, vma
->vm_page_prot
);
3417 if (flags
& FAULT_FLAG_WRITE
)
3418 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3419 else if (pte_file(orig_pte
) && pte_file_soft_dirty(orig_pte
))
3420 pte_mksoft_dirty(entry
);
3422 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3423 page_add_new_anon_rmap(page
, vma
, address
);
3425 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
3426 page_add_file_rmap(page
);
3427 if (flags
& FAULT_FLAG_WRITE
) {
3429 get_page(dirty_page
);
3432 set_pte_at(mm
, address
, page_table
, entry
);
3434 /* no need to invalidate: a not-present page won't be cached */
3435 update_mmu_cache(vma
, address
, page_table
);
3438 mem_cgroup_uncharge_page(cow_page
);
3440 page_cache_release(page
);
3442 anon
= 1; /* no anon but release faulted_page */
3445 pte_unmap_unlock(page_table
, ptl
);
3448 struct address_space
*mapping
= page
->mapping
;
3451 if (set_page_dirty(dirty_page
))
3453 unlock_page(dirty_page
);
3454 put_page(dirty_page
);
3455 if ((dirtied
|| page_mkwrite
) && mapping
) {
3457 * Some device drivers do not set page.mapping but still
3460 balance_dirty_pages_ratelimited(mapping
);
3463 /* file_update_time outside page_lock */
3464 if (vma
->vm_file
&& !page_mkwrite
)
3465 file_update_time(vma
->vm_file
);
3467 unlock_page(vmf
.page
);
3469 page_cache_release(vmf
.page
);
3475 page_cache_release(page
);
3478 /* fs's fault handler get error */
3480 mem_cgroup_uncharge_page(cow_page
);
3481 page_cache_release(cow_page
);
3486 static int do_linear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3487 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3488 unsigned int flags
, pte_t orig_pte
)
3490 pgoff_t pgoff
= (((address
& PAGE_MASK
)
3491 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
3493 pte_unmap(page_table
);
3494 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3498 * Fault of a previously existing named mapping. Repopulate the pte
3499 * from the encoded file_pte if possible. This enables swappable
3502 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3503 * but allow concurrent faults), and pte mapped but not yet locked.
3504 * We return with mmap_sem still held, but pte unmapped and unlocked.
3506 static int do_nonlinear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3507 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3508 unsigned int flags
, pte_t orig_pte
)
3512 flags
|= FAULT_FLAG_NONLINEAR
;
3514 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
3517 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
))) {
3519 * Page table corrupted: show pte and kill process.
3521 print_bad_pte(vma
, address
, orig_pte
, NULL
);
3522 return VM_FAULT_SIGBUS
;
3525 pgoff
= pte_to_pgoff(orig_pte
);
3526 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3529 int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3530 unsigned long addr
, int page_nid
,
3535 count_vm_numa_event(NUMA_HINT_FAULTS
);
3536 if (page_nid
== numa_node_id()) {
3537 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3538 *flags
|= TNF_FAULT_LOCAL
;
3541 return mpol_misplaced(page
, vma
, addr
);
3544 int do_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3545 unsigned long addr
, pte_t pte
, pte_t
*ptep
, pmd_t
*pmd
)
3547 struct page
*page
= NULL
;
3552 bool migrated
= false;
3556 * The "pte" at this point cannot be used safely without
3557 * validation through pte_unmap_same(). It's of NUMA type but
3558 * the pfn may be screwed if the read is non atomic.
3560 * ptep_modify_prot_start is not called as this is clearing
3561 * the _PAGE_NUMA bit and it is not really expected that there
3562 * would be concurrent hardware modifications to the PTE.
3564 ptl
= pte_lockptr(mm
, pmd
);
3566 if (unlikely(!pte_same(*ptep
, pte
))) {
3567 pte_unmap_unlock(ptep
, ptl
);
3571 pte
= pte_mknonnuma(pte
);
3572 set_pte_at(mm
, addr
, ptep
, pte
);
3573 update_mmu_cache(vma
, addr
, ptep
);
3575 page
= vm_normal_page(vma
, addr
, pte
);
3577 pte_unmap_unlock(ptep
, ptl
);
3580 BUG_ON(is_zero_pfn(page_to_pfn(page
)));
3583 * Avoid grouping on DSO/COW pages in specific and RO pages
3584 * in general, RO pages shouldn't hurt as much anyway since
3585 * they can be in shared cache state.
3587 if (!pte_write(pte
))
3588 flags
|= TNF_NO_GROUP
;
3591 * Flag if the page is shared between multiple address spaces. This
3592 * is later used when determining whether to group tasks together
3594 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
3595 flags
|= TNF_SHARED
;
3597 last_cpupid
= page_cpupid_last(page
);
3598 page_nid
= page_to_nid(page
);
3599 target_nid
= numa_migrate_prep(page
, vma
, addr
, page_nid
, &flags
);
3600 pte_unmap_unlock(ptep
, ptl
);
3601 if (target_nid
== -1) {
3606 /* Migrate to the requested node */
3607 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
3609 page_nid
= target_nid
;
3610 flags
|= TNF_MIGRATED
;
3615 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
3620 * These routines also need to handle stuff like marking pages dirty
3621 * and/or accessed for architectures that don't do it in hardware (most
3622 * RISC architectures). The early dirtying is also good on the i386.
3624 * There is also a hook called "update_mmu_cache()" that architectures
3625 * with external mmu caches can use to update those (ie the Sparc or
3626 * PowerPC hashed page tables that act as extended TLBs).
3628 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3629 * but allow concurrent faults), and pte mapped but not yet locked.
3630 * We return with mmap_sem still held, but pte unmapped and unlocked.
3632 static int handle_pte_fault(struct mm_struct
*mm
,
3633 struct vm_area_struct
*vma
, unsigned long address
,
3634 pte_t
*pte
, pmd_t
*pmd
, unsigned int flags
)
3640 if (!pte_present(entry
)) {
3641 if (pte_none(entry
)) {
3643 if (likely(vma
->vm_ops
->fault
))
3644 return do_linear_fault(mm
, vma
, address
,
3645 pte
, pmd
, flags
, entry
);
3647 return do_anonymous_page(mm
, vma
, address
,
3650 if (pte_file(entry
))
3651 return do_nonlinear_fault(mm
, vma
, address
,
3652 pte
, pmd
, flags
, entry
);
3653 return do_swap_page(mm
, vma
, address
,
3654 pte
, pmd
, flags
, entry
);
3657 if (pte_numa(entry
))
3658 return do_numa_page(mm
, vma
, address
, entry
, pte
, pmd
);
3660 ptl
= pte_lockptr(mm
, pmd
);
3662 if (unlikely(!pte_same(*pte
, entry
)))
3664 if (flags
& FAULT_FLAG_WRITE
) {
3665 if (!pte_write(entry
))
3666 return do_wp_page(mm
, vma
, address
,
3667 pte
, pmd
, ptl
, entry
);
3668 entry
= pte_mkdirty(entry
);
3670 entry
= pte_mkyoung(entry
);
3671 if (ptep_set_access_flags(vma
, address
, pte
, entry
, flags
& FAULT_FLAG_WRITE
)) {
3672 update_mmu_cache(vma
, address
, pte
);
3675 * This is needed only for protection faults but the arch code
3676 * is not yet telling us if this is a protection fault or not.
3677 * This still avoids useless tlb flushes for .text page faults
3680 if (flags
& FAULT_FLAG_WRITE
)
3681 flush_tlb_fix_spurious_fault(vma
, address
);
3684 pte_unmap_unlock(pte
, ptl
);
3689 * By the time we get here, we already hold the mm semaphore
3691 static int __handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3692 unsigned long address
, unsigned int flags
)
3699 if (unlikely(is_vm_hugetlb_page(vma
)))
3700 return hugetlb_fault(mm
, vma
, address
, flags
);
3703 pgd
= pgd_offset(mm
, address
);
3704 pud
= pud_alloc(mm
, pgd
, address
);
3706 return VM_FAULT_OOM
;
3707 pmd
= pmd_alloc(mm
, pud
, address
);
3709 return VM_FAULT_OOM
;
3710 if (pmd_none(*pmd
) && transparent_hugepage_enabled(vma
)) {
3711 int ret
= VM_FAULT_FALLBACK
;
3713 ret
= do_huge_pmd_anonymous_page(mm
, vma
, address
,
3715 if (!(ret
& VM_FAULT_FALLBACK
))
3718 pmd_t orig_pmd
= *pmd
;
3722 if (pmd_trans_huge(orig_pmd
)) {
3723 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
3726 * If the pmd is splitting, return and retry the
3727 * the fault. Alternative: wait until the split
3728 * is done, and goto retry.
3730 if (pmd_trans_splitting(orig_pmd
))
3733 if (pmd_numa(orig_pmd
))
3734 return do_huge_pmd_numa_page(mm
, vma
, address
,
3737 if (dirty
&& !pmd_write(orig_pmd
)) {
3738 ret
= do_huge_pmd_wp_page(mm
, vma
, address
, pmd
,
3741 * If COW results in an oom, the huge pmd will
3742 * have been split, so retry the fault on the
3743 * pte for a smaller charge.
3745 if (unlikely(ret
& VM_FAULT_OOM
))
3749 huge_pmd_set_accessed(mm
, vma
, address
, pmd
,
3757 /* THP should already have been handled */
3758 BUG_ON(pmd_numa(*pmd
));
3761 * Use __pte_alloc instead of pte_alloc_map, because we can't
3762 * run pte_offset_map on the pmd, if an huge pmd could
3763 * materialize from under us from a different thread.
3765 if (unlikely(pmd_none(*pmd
)) &&
3766 unlikely(__pte_alloc(mm
, vma
, pmd
, address
)))
3767 return VM_FAULT_OOM
;
3768 /* if an huge pmd materialized from under us just retry later */
3769 if (unlikely(pmd_trans_huge(*pmd
)))
3772 * A regular pmd is established and it can't morph into a huge pmd
3773 * from under us anymore at this point because we hold the mmap_sem
3774 * read mode and khugepaged takes it in write mode. So now it's
3775 * safe to run pte_offset_map().
3777 pte
= pte_offset_map(pmd
, address
);
3779 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
3782 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3783 unsigned long address
, unsigned int flags
)
3787 __set_current_state(TASK_RUNNING
);
3789 count_vm_event(PGFAULT
);
3790 mem_cgroup_count_vm_event(mm
, PGFAULT
);
3792 /* do counter updates before entering really critical section. */
3793 check_sync_rss_stat(current
);
3796 * Enable the memcg OOM handling for faults triggered in user
3797 * space. Kernel faults are handled more gracefully.
3799 if (flags
& FAULT_FLAG_USER
)
3800 mem_cgroup_oom_enable();
3802 ret
= __handle_mm_fault(mm
, vma
, address
, flags
);
3804 if (flags
& FAULT_FLAG_USER
) {
3805 mem_cgroup_oom_disable();
3807 * The task may have entered a memcg OOM situation but
3808 * if the allocation error was handled gracefully (no
3809 * VM_FAULT_OOM), there is no need to kill anything.
3810 * Just clean up the OOM state peacefully.
3812 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
3813 mem_cgroup_oom_synchronize(false);
3819 #ifndef __PAGETABLE_PUD_FOLDED
3821 * Allocate page upper directory.
3822 * We've already handled the fast-path in-line.
3824 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
3826 pud_t
*new = pud_alloc_one(mm
, address
);
3830 smp_wmb(); /* See comment in __pte_alloc */
3832 spin_lock(&mm
->page_table_lock
);
3833 if (pgd_present(*pgd
)) /* Another has populated it */
3836 pgd_populate(mm
, pgd
, new);
3837 spin_unlock(&mm
->page_table_lock
);
3840 #endif /* __PAGETABLE_PUD_FOLDED */
3842 #ifndef __PAGETABLE_PMD_FOLDED
3844 * Allocate page middle directory.
3845 * We've already handled the fast-path in-line.
3847 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
3849 pmd_t
*new = pmd_alloc_one(mm
, address
);
3853 smp_wmb(); /* See comment in __pte_alloc */
3855 spin_lock(&mm
->page_table_lock
);
3856 #ifndef __ARCH_HAS_4LEVEL_HACK
3857 if (pud_present(*pud
)) /* Another has populated it */
3860 pud_populate(mm
, pud
, new);
3862 if (pgd_present(*pud
)) /* Another has populated it */
3865 pgd_populate(mm
, pud
, new);
3866 #endif /* __ARCH_HAS_4LEVEL_HACK */
3867 spin_unlock(&mm
->page_table_lock
);
3870 #endif /* __PAGETABLE_PMD_FOLDED */
3872 #if !defined(__HAVE_ARCH_GATE_AREA)
3874 #if defined(AT_SYSINFO_EHDR)
3875 static struct vm_area_struct gate_vma
;
3877 static int __init
gate_vma_init(void)
3879 gate_vma
.vm_mm
= NULL
;
3880 gate_vma
.vm_start
= FIXADDR_USER_START
;
3881 gate_vma
.vm_end
= FIXADDR_USER_END
;
3882 gate_vma
.vm_flags
= VM_READ
| VM_MAYREAD
| VM_EXEC
| VM_MAYEXEC
;
3883 gate_vma
.vm_page_prot
= __P101
;
3887 __initcall(gate_vma_init
);
3890 struct vm_area_struct
*get_gate_vma(struct mm_struct
*mm
)
3892 #ifdef AT_SYSINFO_EHDR
3899 int in_gate_area_no_mm(unsigned long addr
)
3901 #ifdef AT_SYSINFO_EHDR
3902 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
3908 #endif /* __HAVE_ARCH_GATE_AREA */
3910 static int __follow_pte(struct mm_struct
*mm
, unsigned long address
,
3911 pte_t
**ptepp
, spinlock_t
**ptlp
)
3918 pgd
= pgd_offset(mm
, address
);
3919 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
3922 pud
= pud_offset(pgd
, address
);
3923 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
3926 pmd
= pmd_offset(pud
, address
);
3927 VM_BUG_ON(pmd_trans_huge(*pmd
));
3928 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
3931 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3935 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
3938 if (!pte_present(*ptep
))
3943 pte_unmap_unlock(ptep
, *ptlp
);
3948 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
3949 pte_t
**ptepp
, spinlock_t
**ptlp
)
3953 /* (void) is needed to make gcc happy */
3954 (void) __cond_lock(*ptlp
,
3955 !(res
= __follow_pte(mm
, address
, ptepp
, ptlp
)));
3960 * follow_pfn - look up PFN at a user virtual address
3961 * @vma: memory mapping
3962 * @address: user virtual address
3963 * @pfn: location to store found PFN
3965 * Only IO mappings and raw PFN mappings are allowed.
3967 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3969 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
3976 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3979 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
3982 *pfn
= pte_pfn(*ptep
);
3983 pte_unmap_unlock(ptep
, ptl
);
3986 EXPORT_SYMBOL(follow_pfn
);
3988 #ifdef CONFIG_HAVE_IOREMAP_PROT
3989 int follow_phys(struct vm_area_struct
*vma
,
3990 unsigned long address
, unsigned int flags
,
3991 unsigned long *prot
, resource_size_t
*phys
)
3997 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4000 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
4004 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
4007 *prot
= pgprot_val(pte_pgprot(pte
));
4008 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
4012 pte_unmap_unlock(ptep
, ptl
);
4017 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
4018 void *buf
, int len
, int write
)
4020 resource_size_t phys_addr
;
4021 unsigned long prot
= 0;
4022 void __iomem
*maddr
;
4023 int offset
= addr
& (PAGE_SIZE
-1);
4025 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
4028 maddr
= ioremap_prot(phys_addr
, PAGE_SIZE
, prot
);
4030 memcpy_toio(maddr
+ offset
, buf
, len
);
4032 memcpy_fromio(buf
, maddr
+ offset
, len
);
4037 EXPORT_SYMBOL_GPL(generic_access_phys
);
4041 * Access another process' address space as given in mm. If non-NULL, use the
4042 * given task for page fault accounting.
4044 static int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
4045 unsigned long addr
, void *buf
, int len
, int write
)
4047 struct vm_area_struct
*vma
;
4048 void *old_buf
= buf
;
4050 down_read(&mm
->mmap_sem
);
4051 /* ignore errors, just check how much was successfully transferred */
4053 int bytes
, ret
, offset
;
4055 struct page
*page
= NULL
;
4057 ret
= get_user_pages(tsk
, mm
, addr
, 1,
4058 write
, 1, &page
, &vma
);
4061 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4062 * we can access using slightly different code.
4064 #ifdef CONFIG_HAVE_IOREMAP_PROT
4065 vma
= find_vma(mm
, addr
);
4066 if (!vma
|| vma
->vm_start
> addr
)
4068 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
4069 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
4077 offset
= addr
& (PAGE_SIZE
-1);
4078 if (bytes
> PAGE_SIZE
-offset
)
4079 bytes
= PAGE_SIZE
-offset
;
4083 copy_to_user_page(vma
, page
, addr
,
4084 maddr
+ offset
, buf
, bytes
);
4085 set_page_dirty_lock(page
);
4087 copy_from_user_page(vma
, page
, addr
,
4088 buf
, maddr
+ offset
, bytes
);
4091 page_cache_release(page
);
4097 up_read(&mm
->mmap_sem
);
4099 return buf
- old_buf
;
4103 * access_remote_vm - access another process' address space
4104 * @mm: the mm_struct of the target address space
4105 * @addr: start address to access
4106 * @buf: source or destination buffer
4107 * @len: number of bytes to transfer
4108 * @write: whether the access is a write
4110 * The caller must hold a reference on @mm.
4112 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
4113 void *buf
, int len
, int write
)
4115 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, write
);
4119 * Access another process' address space.
4120 * Source/target buffer must be kernel space,
4121 * Do not walk the page table directly, use get_user_pages
4123 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
4124 void *buf
, int len
, int write
)
4126 struct mm_struct
*mm
;
4129 mm
= get_task_mm(tsk
);
4133 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, write
);
4140 * Print the name of a VMA.
4142 void print_vma_addr(char *prefix
, unsigned long ip
)
4144 struct mm_struct
*mm
= current
->mm
;
4145 struct vm_area_struct
*vma
;
4148 * Do not print if we are in atomic
4149 * contexts (in exception stacks, etc.):
4151 if (preempt_count())
4154 down_read(&mm
->mmap_sem
);
4155 vma
= find_vma(mm
, ip
);
4156 if (vma
&& vma
->vm_file
) {
4157 struct file
*f
= vma
->vm_file
;
4158 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
4162 p
= d_path(&f
->f_path
, buf
, PAGE_SIZE
);
4165 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
4167 vma
->vm_end
- vma
->vm_start
);
4168 free_page((unsigned long)buf
);
4171 up_read(&mm
->mmap_sem
);
4174 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4175 void might_fault(void)
4178 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4179 * holding the mmap_sem, this is safe because kernel memory doesn't
4180 * get paged out, therefore we'll never actually fault, and the
4181 * below annotations will generate false positives.
4183 if (segment_eq(get_fs(), KERNEL_DS
))
4187 * it would be nicer only to annotate paths which are not under
4188 * pagefault_disable, however that requires a larger audit and
4189 * providing helpers like get_user_atomic.
4194 __might_sleep(__FILE__
, __LINE__
, 0);
4197 might_lock_read(¤t
->mm
->mmap_sem
);
4199 EXPORT_SYMBOL(might_fault
);
4202 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4203 static void clear_gigantic_page(struct page
*page
,
4205 unsigned int pages_per_huge_page
)
4208 struct page
*p
= page
;
4211 for (i
= 0; i
< pages_per_huge_page
;
4212 i
++, p
= mem_map_next(p
, page
, i
)) {
4214 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
4217 void clear_huge_page(struct page
*page
,
4218 unsigned long addr
, unsigned int pages_per_huge_page
)
4222 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4223 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
4228 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4230 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
4234 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
4236 struct vm_area_struct
*vma
,
4237 unsigned int pages_per_huge_page
)
4240 struct page
*dst_base
= dst
;
4241 struct page
*src_base
= src
;
4243 for (i
= 0; i
< pages_per_huge_page
; ) {
4245 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
4248 dst
= mem_map_next(dst
, dst_base
, i
);
4249 src
= mem_map_next(src
, src_base
, i
);
4253 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
4254 unsigned long addr
, struct vm_area_struct
*vma
,
4255 unsigned int pages_per_huge_page
)
4259 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4260 copy_user_gigantic_page(dst
, src
, addr
, vma
,
4261 pages_per_huge_page
);
4266 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4268 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
4271 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */