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>
62 #include <linux/dma-debug.h>
63 #include <linux/debugfs.h>
66 #include <asm/pgalloc.h>
67 #include <asm/uaccess.h>
69 #include <asm/tlbflush.h>
70 #include <asm/pgtable.h>
74 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
75 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
78 #ifndef CONFIG_NEED_MULTIPLE_NODES
79 /* use the per-pgdat data instead for discontigmem - mbligh */
80 unsigned long max_mapnr
;
83 EXPORT_SYMBOL(max_mapnr
);
84 EXPORT_SYMBOL(mem_map
);
88 * A number of key systems in x86 including ioremap() rely on the assumption
89 * that high_memory defines the upper bound on direct map memory, then end
90 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
91 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
96 EXPORT_SYMBOL(high_memory
);
99 * Randomize the address space (stacks, mmaps, brk, etc.).
101 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
102 * as ancient (libc5 based) binaries can segfault. )
104 int randomize_va_space __read_mostly
=
105 #ifdef CONFIG_COMPAT_BRK
111 static int __init
disable_randmaps(char *s
)
113 randomize_va_space
= 0;
116 __setup("norandmaps", disable_randmaps
);
118 unsigned long zero_pfn __read_mostly
;
119 unsigned long highest_memmap_pfn __read_mostly
;
121 EXPORT_SYMBOL(zero_pfn
);
124 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
126 static int __init
init_zero_pfn(void)
128 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
131 core_initcall(init_zero_pfn
);
134 #if defined(SPLIT_RSS_COUNTING)
136 void sync_mm_rss(struct mm_struct
*mm
)
140 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
141 if (current
->rss_stat
.count
[i
]) {
142 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
143 current
->rss_stat
.count
[i
] = 0;
146 current
->rss_stat
.events
= 0;
149 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
151 struct task_struct
*task
= current
;
153 if (likely(task
->mm
== mm
))
154 task
->rss_stat
.count
[member
] += val
;
156 add_mm_counter(mm
, member
, val
);
158 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
159 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
161 /* sync counter once per 64 page faults */
162 #define TASK_RSS_EVENTS_THRESH (64)
163 static void check_sync_rss_stat(struct task_struct
*task
)
165 if (unlikely(task
!= current
))
167 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
168 sync_mm_rss(task
->mm
);
170 #else /* SPLIT_RSS_COUNTING */
172 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
173 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
175 static void check_sync_rss_stat(struct task_struct
*task
)
179 #endif /* SPLIT_RSS_COUNTING */
181 #ifdef HAVE_GENERIC_MMU_GATHER
183 static int tlb_next_batch(struct mmu_gather
*tlb
)
185 struct mmu_gather_batch
*batch
;
189 tlb
->active
= batch
->next
;
193 if (tlb
->batch_count
== MAX_GATHER_BATCH_COUNT
)
196 batch
= (void *)__get_free_pages(GFP_NOWAIT
| __GFP_NOWARN
, 0);
203 batch
->max
= MAX_GATHER_BATCH
;
205 tlb
->active
->next
= batch
;
212 * Called to initialize an (on-stack) mmu_gather structure for page-table
213 * tear-down from @mm. The @fullmm argument is used when @mm is without
214 * users and we're going to destroy the full address space (exit/execve).
216 void tlb_gather_mmu(struct mmu_gather
*tlb
, struct mm_struct
*mm
, unsigned long start
, unsigned long end
)
220 /* Is it from 0 to ~0? */
221 tlb
->fullmm
= !(start
| (end
+1));
222 tlb
->need_flush_all
= 0;
223 tlb
->local
.next
= NULL
;
225 tlb
->local
.max
= ARRAY_SIZE(tlb
->__pages
);
226 tlb
->active
= &tlb
->local
;
227 tlb
->batch_count
= 0;
229 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
233 __tlb_reset_range(tlb
);
236 static void tlb_flush_mmu_tlbonly(struct mmu_gather
*tlb
)
242 mmu_notifier_invalidate_range(tlb
->mm
, tlb
->start
, tlb
->end
);
243 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
244 tlb_table_flush(tlb
);
246 __tlb_reset_range(tlb
);
249 static void tlb_flush_mmu_free(struct mmu_gather
*tlb
)
251 struct mmu_gather_batch
*batch
;
253 for (batch
= &tlb
->local
; batch
&& batch
->nr
; batch
= batch
->next
) {
254 free_pages_and_swap_cache(batch
->pages
, batch
->nr
);
257 tlb
->active
= &tlb
->local
;
260 void tlb_flush_mmu(struct mmu_gather
*tlb
)
262 tlb_flush_mmu_tlbonly(tlb
);
263 tlb_flush_mmu_free(tlb
);
267 * Called at the end of the shootdown operation to free up any resources
268 * that were required.
270 void tlb_finish_mmu(struct mmu_gather
*tlb
, unsigned long start
, unsigned long end
)
272 struct mmu_gather_batch
*batch
, *next
;
276 /* keep the page table cache within bounds */
279 for (batch
= tlb
->local
.next
; batch
; batch
= next
) {
281 free_pages((unsigned long)batch
, 0);
283 tlb
->local
.next
= NULL
;
287 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
288 * handling the additional races in SMP caused by other CPUs caching valid
289 * mappings in their TLBs. Returns the number of free page slots left.
290 * When out of page slots we must call tlb_flush_mmu().
292 int __tlb_remove_page(struct mmu_gather
*tlb
, struct page
*page
)
294 struct mmu_gather_batch
*batch
;
296 VM_BUG_ON(!tlb
->end
);
299 batch
->pages
[batch
->nr
++] = page
;
300 if (batch
->nr
== batch
->max
) {
301 if (!tlb_next_batch(tlb
))
305 VM_BUG_ON_PAGE(batch
->nr
> batch
->max
, page
);
307 return batch
->max
- batch
->nr
;
310 #endif /* HAVE_GENERIC_MMU_GATHER */
312 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
315 * See the comment near struct mmu_table_batch.
318 static void tlb_remove_table_smp_sync(void *arg
)
320 /* Simply deliver the interrupt */
323 static void tlb_remove_table_one(void *table
)
326 * This isn't an RCU grace period and hence the page-tables cannot be
327 * assumed to be actually RCU-freed.
329 * It is however sufficient for software page-table walkers that rely on
330 * IRQ disabling. See the comment near struct mmu_table_batch.
332 smp_call_function(tlb_remove_table_smp_sync
, NULL
, 1);
333 __tlb_remove_table(table
);
336 static void tlb_remove_table_rcu(struct rcu_head
*head
)
338 struct mmu_table_batch
*batch
;
341 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
343 for (i
= 0; i
< batch
->nr
; i
++)
344 __tlb_remove_table(batch
->tables
[i
]);
346 free_page((unsigned long)batch
);
349 void tlb_table_flush(struct mmu_gather
*tlb
)
351 struct mmu_table_batch
**batch
= &tlb
->batch
;
354 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
359 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
361 struct mmu_table_batch
**batch
= &tlb
->batch
;
364 * When there's less then two users of this mm there cannot be a
365 * concurrent page-table walk.
367 if (atomic_read(&tlb
->mm
->mm_users
) < 2) {
368 __tlb_remove_table(table
);
372 if (*batch
== NULL
) {
373 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
374 if (*batch
== NULL
) {
375 tlb_remove_table_one(table
);
380 (*batch
)->tables
[(*batch
)->nr
++] = table
;
381 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
382 tlb_table_flush(tlb
);
385 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
388 * Note: this doesn't free the actual pages themselves. That
389 * has been handled earlier when unmapping all the memory regions.
391 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
394 pgtable_t token
= pmd_pgtable(*pmd
);
396 pte_free_tlb(tlb
, token
, addr
);
397 atomic_long_dec(&tlb
->mm
->nr_ptes
);
400 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
401 unsigned long addr
, unsigned long end
,
402 unsigned long floor
, unsigned long ceiling
)
409 pmd
= pmd_offset(pud
, addr
);
411 next
= pmd_addr_end(addr
, end
);
412 if (pmd_none_or_clear_bad(pmd
))
414 free_pte_range(tlb
, pmd
, addr
);
415 } while (pmd
++, addr
= next
, addr
!= end
);
425 if (end
- 1 > ceiling
- 1)
428 pmd
= pmd_offset(pud
, start
);
430 pmd_free_tlb(tlb
, pmd
, start
);
431 mm_dec_nr_pmds(tlb
->mm
);
434 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
435 unsigned long addr
, unsigned long end
,
436 unsigned long floor
, unsigned long ceiling
)
443 pud
= pud_offset(pgd
, addr
);
445 next
= pud_addr_end(addr
, end
);
446 if (pud_none_or_clear_bad(pud
))
448 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
449 } while (pud
++, addr
= next
, addr
!= end
);
455 ceiling
&= PGDIR_MASK
;
459 if (end
- 1 > ceiling
- 1)
462 pud
= pud_offset(pgd
, start
);
464 pud_free_tlb(tlb
, pud
, start
);
468 * This function frees user-level page tables of a process.
470 void free_pgd_range(struct mmu_gather
*tlb
,
471 unsigned long addr
, unsigned long end
,
472 unsigned long floor
, unsigned long ceiling
)
478 * The next few lines have given us lots of grief...
480 * Why are we testing PMD* at this top level? Because often
481 * there will be no work to do at all, and we'd prefer not to
482 * go all the way down to the bottom just to discover that.
484 * Why all these "- 1"s? Because 0 represents both the bottom
485 * of the address space and the top of it (using -1 for the
486 * top wouldn't help much: the masks would do the wrong thing).
487 * The rule is that addr 0 and floor 0 refer to the bottom of
488 * the address space, but end 0 and ceiling 0 refer to the top
489 * Comparisons need to use "end - 1" and "ceiling - 1" (though
490 * that end 0 case should be mythical).
492 * Wherever addr is brought up or ceiling brought down, we must
493 * be careful to reject "the opposite 0" before it confuses the
494 * subsequent tests. But what about where end is brought down
495 * by PMD_SIZE below? no, end can't go down to 0 there.
497 * Whereas we round start (addr) and ceiling down, by different
498 * masks at different levels, in order to test whether a table
499 * now has no other vmas using it, so can be freed, we don't
500 * bother to round floor or end up - the tests don't need that.
514 if (end
- 1 > ceiling
- 1)
519 pgd
= pgd_offset(tlb
->mm
, addr
);
521 next
= pgd_addr_end(addr
, end
);
522 if (pgd_none_or_clear_bad(pgd
))
524 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
525 } while (pgd
++, addr
= next
, addr
!= end
);
528 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
529 unsigned long floor
, unsigned long ceiling
)
532 struct vm_area_struct
*next
= vma
->vm_next
;
533 unsigned long addr
= vma
->vm_start
;
536 * Hide vma from rmap and truncate_pagecache before freeing
539 unlink_anon_vmas(vma
);
540 unlink_file_vma(vma
);
542 if (is_vm_hugetlb_page(vma
)) {
543 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
544 floor
, next
? next
->vm_start
: ceiling
);
547 * Optimization: gather nearby vmas into one call down
549 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
550 && !is_vm_hugetlb_page(next
)) {
553 unlink_anon_vmas(vma
);
554 unlink_file_vma(vma
);
556 free_pgd_range(tlb
, addr
, vma
->vm_end
,
557 floor
, next
? next
->vm_start
: ceiling
);
563 int __pte_alloc(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
564 pmd_t
*pmd
, unsigned long address
)
567 pgtable_t
new = pte_alloc_one(mm
, address
);
568 int wait_split_huge_page
;
573 * Ensure all pte setup (eg. pte page lock and page clearing) are
574 * visible before the pte is made visible to other CPUs by being
575 * put into page tables.
577 * The other side of the story is the pointer chasing in the page
578 * table walking code (when walking the page table without locking;
579 * ie. most of the time). Fortunately, these data accesses consist
580 * of a chain of data-dependent loads, meaning most CPUs (alpha
581 * being the notable exception) will already guarantee loads are
582 * seen in-order. See the alpha page table accessors for the
583 * smp_read_barrier_depends() barriers in page table walking code.
585 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
587 ptl
= pmd_lock(mm
, pmd
);
588 wait_split_huge_page
= 0;
589 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
590 atomic_long_inc(&mm
->nr_ptes
);
591 pmd_populate(mm
, pmd
, new);
593 } else if (unlikely(pmd_trans_splitting(*pmd
)))
594 wait_split_huge_page
= 1;
598 if (wait_split_huge_page
)
599 wait_split_huge_page(vma
->anon_vma
, pmd
);
603 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
605 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
609 smp_wmb(); /* See comment in __pte_alloc */
611 spin_lock(&init_mm
.page_table_lock
);
612 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
613 pmd_populate_kernel(&init_mm
, pmd
, new);
616 VM_BUG_ON(pmd_trans_splitting(*pmd
));
617 spin_unlock(&init_mm
.page_table_lock
);
619 pte_free_kernel(&init_mm
, new);
623 static inline void init_rss_vec(int *rss
)
625 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
628 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
632 if (current
->mm
== mm
)
634 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
636 add_mm_counter(mm
, i
, rss
[i
]);
640 * This function is called to print an error when a bad pte
641 * is found. For example, we might have a PFN-mapped pte in
642 * a region that doesn't allow it.
644 * The calling function must still handle the error.
646 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
647 pte_t pte
, struct page
*page
)
649 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
650 pud_t
*pud
= pud_offset(pgd
, addr
);
651 pmd_t
*pmd
= pmd_offset(pud
, addr
);
652 struct address_space
*mapping
;
654 static unsigned long resume
;
655 static unsigned long nr_shown
;
656 static unsigned long nr_unshown
;
659 * Allow a burst of 60 reports, then keep quiet for that minute;
660 * or allow a steady drip of one report per second.
662 if (nr_shown
== 60) {
663 if (time_before(jiffies
, resume
)) {
669 "BUG: Bad page map: %lu messages suppressed\n",
676 resume
= jiffies
+ 60 * HZ
;
678 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
679 index
= linear_page_index(vma
, addr
);
682 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
684 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
686 dump_page(page
, "bad pte");
688 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
689 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
691 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
694 printk(KERN_ALERT
"vma->vm_ops->fault: %pSR\n",
697 printk(KERN_ALERT
"vma->vm_file->f_op->mmap: %pSR\n",
698 vma
->vm_file
->f_op
->mmap
);
700 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
704 * vm_normal_page -- This function gets the "struct page" associated with a pte.
706 * "Special" mappings do not wish to be associated with a "struct page" (either
707 * it doesn't exist, or it exists but they don't want to touch it). In this
708 * case, NULL is returned here. "Normal" mappings do have a struct page.
710 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
711 * pte bit, in which case this function is trivial. Secondly, an architecture
712 * may not have a spare pte bit, which requires a more complicated scheme,
715 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
716 * special mapping (even if there are underlying and valid "struct pages").
717 * COWed pages of a VM_PFNMAP are always normal.
719 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
720 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
721 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
722 * mapping will always honor the rule
724 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
726 * And for normal mappings this is false.
728 * This restricts such mappings to be a linear translation from virtual address
729 * to pfn. To get around this restriction, we allow arbitrary mappings so long
730 * as the vma is not a COW mapping; in that case, we know that all ptes are
731 * special (because none can have been COWed).
734 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
736 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
737 * page" backing, however the difference is that _all_ pages with a struct
738 * page (that is, those where pfn_valid is true) are refcounted and considered
739 * normal pages by the VM. The disadvantage is that pages are refcounted
740 * (which can be slower and simply not an option for some PFNMAP users). The
741 * advantage is that we don't have to follow the strict linearity rule of
742 * PFNMAP mappings in order to support COWable mappings.
745 #ifdef __HAVE_ARCH_PTE_SPECIAL
746 # define HAVE_PTE_SPECIAL 1
748 # define HAVE_PTE_SPECIAL 0
750 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
753 unsigned long pfn
= pte_pfn(pte
);
755 if (HAVE_PTE_SPECIAL
) {
756 if (likely(!pte_special(pte
)))
758 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
759 return vma
->vm_ops
->find_special_page(vma
, addr
);
760 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
762 if (!is_zero_pfn(pfn
))
763 print_bad_pte(vma
, addr
, pte
, NULL
);
767 /* !HAVE_PTE_SPECIAL case follows: */
769 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
770 if (vma
->vm_flags
& VM_MIXEDMAP
) {
776 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
777 if (pfn
== vma
->vm_pgoff
+ off
)
779 if (!is_cow_mapping(vma
->vm_flags
))
784 if (is_zero_pfn(pfn
))
787 if (unlikely(pfn
> highest_memmap_pfn
)) {
788 print_bad_pte(vma
, addr
, pte
, NULL
);
793 * NOTE! We still have PageReserved() pages in the page tables.
794 * eg. VDSO mappings can cause them to exist.
797 return pfn_to_page(pfn
);
801 * copy one vm_area from one task to the other. Assumes the page tables
802 * already present in the new task to be cleared in the whole range
803 * covered by this vma.
806 static inline unsigned long
807 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
808 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
809 unsigned long addr
, int *rss
)
811 unsigned long vm_flags
= vma
->vm_flags
;
812 pte_t pte
= *src_pte
;
815 /* pte contains position in swap or file, so copy. */
816 if (unlikely(!pte_present(pte
))) {
817 swp_entry_t entry
= pte_to_swp_entry(pte
);
819 if (likely(!non_swap_entry(entry
))) {
820 if (swap_duplicate(entry
) < 0)
823 /* make sure dst_mm is on swapoff's mmlist. */
824 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
825 spin_lock(&mmlist_lock
);
826 if (list_empty(&dst_mm
->mmlist
))
827 list_add(&dst_mm
->mmlist
,
829 spin_unlock(&mmlist_lock
);
832 } else if (is_migration_entry(entry
)) {
833 page
= migration_entry_to_page(entry
);
840 if (is_write_migration_entry(entry
) &&
841 is_cow_mapping(vm_flags
)) {
843 * COW mappings require pages in both
844 * parent and child to be set to read.
846 make_migration_entry_read(&entry
);
847 pte
= swp_entry_to_pte(entry
);
848 if (pte_swp_soft_dirty(*src_pte
))
849 pte
= pte_swp_mksoft_dirty(pte
);
850 set_pte_at(src_mm
, addr
, src_pte
, pte
);
857 * If it's a COW mapping, write protect it both
858 * in the parent and the child
860 if (is_cow_mapping(vm_flags
)) {
861 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
862 pte
= pte_wrprotect(pte
);
866 * If it's a shared mapping, mark it clean in
869 if (vm_flags
& VM_SHARED
)
870 pte
= pte_mkclean(pte
);
871 pte
= pte_mkold(pte
);
873 page
= vm_normal_page(vma
, addr
, pte
);
884 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
888 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
889 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
890 unsigned long addr
, unsigned long end
)
892 pte_t
*orig_src_pte
, *orig_dst_pte
;
893 pte_t
*src_pte
, *dst_pte
;
894 spinlock_t
*src_ptl
, *dst_ptl
;
896 int rss
[NR_MM_COUNTERS
];
897 swp_entry_t entry
= (swp_entry_t
){0};
902 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
905 src_pte
= pte_offset_map(src_pmd
, addr
);
906 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
907 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
908 orig_src_pte
= src_pte
;
909 orig_dst_pte
= dst_pte
;
910 arch_enter_lazy_mmu_mode();
914 * We are holding two locks at this point - either of them
915 * could generate latencies in another task on another CPU.
917 if (progress
>= 32) {
919 if (need_resched() ||
920 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
923 if (pte_none(*src_pte
)) {
927 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
932 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
934 arch_leave_lazy_mmu_mode();
935 spin_unlock(src_ptl
);
936 pte_unmap(orig_src_pte
);
937 add_mm_rss_vec(dst_mm
, rss
);
938 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
942 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
951 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
952 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
953 unsigned long addr
, unsigned long end
)
955 pmd_t
*src_pmd
, *dst_pmd
;
958 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
961 src_pmd
= pmd_offset(src_pud
, addr
);
963 next
= pmd_addr_end(addr
, end
);
964 if (pmd_trans_huge(*src_pmd
)) {
966 VM_BUG_ON(next
-addr
!= HPAGE_PMD_SIZE
);
967 err
= copy_huge_pmd(dst_mm
, src_mm
,
968 dst_pmd
, src_pmd
, addr
, vma
);
975 if (pmd_none_or_clear_bad(src_pmd
))
977 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
980 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
984 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
985 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
986 unsigned long addr
, unsigned long end
)
988 pud_t
*src_pud
, *dst_pud
;
991 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
994 src_pud
= pud_offset(src_pgd
, addr
);
996 next
= pud_addr_end(addr
, end
);
997 if (pud_none_or_clear_bad(src_pud
))
999 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
1002 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1006 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1007 struct vm_area_struct
*vma
)
1009 pgd_t
*src_pgd
, *dst_pgd
;
1011 unsigned long addr
= vma
->vm_start
;
1012 unsigned long end
= vma
->vm_end
;
1013 unsigned long mmun_start
; /* For mmu_notifiers */
1014 unsigned long mmun_end
; /* For mmu_notifiers */
1019 * Don't copy ptes where a page fault will fill them correctly.
1020 * Fork becomes much lighter when there are big shared or private
1021 * readonly mappings. The tradeoff is that copy_page_range is more
1022 * efficient than faulting.
1024 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_PFNMAP
| VM_MIXEDMAP
)) &&
1028 if (is_vm_hugetlb_page(vma
))
1029 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1031 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
1033 * We do not free on error cases below as remove_vma
1034 * gets called on error from higher level routine
1036 ret
= track_pfn_copy(vma
);
1042 * We need to invalidate the secondary MMU mappings only when
1043 * there could be a permission downgrade on the ptes of the
1044 * parent mm. And a permission downgrade will only happen if
1045 * is_cow_mapping() returns true.
1047 is_cow
= is_cow_mapping(vma
->vm_flags
);
1051 mmu_notifier_invalidate_range_start(src_mm
, mmun_start
,
1055 dst_pgd
= pgd_offset(dst_mm
, addr
);
1056 src_pgd
= pgd_offset(src_mm
, addr
);
1058 next
= pgd_addr_end(addr
, end
);
1059 if (pgd_none_or_clear_bad(src_pgd
))
1061 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1062 vma
, addr
, next
))) {
1066 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1069 mmu_notifier_invalidate_range_end(src_mm
, mmun_start
, mmun_end
);
1073 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1074 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1075 unsigned long addr
, unsigned long end
,
1076 struct zap_details
*details
)
1078 struct mm_struct
*mm
= tlb
->mm
;
1079 int force_flush
= 0;
1080 int rss
[NR_MM_COUNTERS
];
1088 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1090 arch_enter_lazy_mmu_mode();
1093 if (pte_none(ptent
)) {
1097 if (pte_present(ptent
)) {
1100 page
= vm_normal_page(vma
, addr
, ptent
);
1101 if (unlikely(details
) && page
) {
1103 * unmap_shared_mapping_pages() wants to
1104 * invalidate cache without truncating:
1105 * unmap shared but keep private pages.
1107 if (details
->check_mapping
&&
1108 details
->check_mapping
!= page
->mapping
)
1111 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1113 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1114 if (unlikely(!page
))
1117 rss
[MM_ANONPAGES
]--;
1119 if (pte_dirty(ptent
)) {
1121 set_page_dirty(page
);
1123 if (pte_young(ptent
) &&
1124 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1125 mark_page_accessed(page
);
1126 rss
[MM_FILEPAGES
]--;
1128 page_remove_rmap(page
);
1129 if (unlikely(page_mapcount(page
) < 0))
1130 print_bad_pte(vma
, addr
, ptent
, page
);
1131 if (unlikely(!__tlb_remove_page(tlb
, page
))) {
1138 /* If details->check_mapping, we leave swap entries. */
1139 if (unlikely(details
))
1142 entry
= pte_to_swp_entry(ptent
);
1143 if (!non_swap_entry(entry
))
1145 else if (is_migration_entry(entry
)) {
1148 page
= migration_entry_to_page(entry
);
1151 rss
[MM_ANONPAGES
]--;
1153 rss
[MM_FILEPAGES
]--;
1155 if (unlikely(!free_swap_and_cache(entry
)))
1156 print_bad_pte(vma
, addr
, ptent
, NULL
);
1157 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1158 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1160 add_mm_rss_vec(mm
, rss
);
1161 arch_leave_lazy_mmu_mode();
1163 /* Do the actual TLB flush before dropping ptl */
1165 tlb_flush_mmu_tlbonly(tlb
);
1166 pte_unmap_unlock(start_pte
, ptl
);
1169 * If we forced a TLB flush (either due to running out of
1170 * batch buffers or because we needed to flush dirty TLB
1171 * entries before releasing the ptl), free the batched
1172 * memory too. Restart if we didn't do everything.
1176 tlb_flush_mmu_free(tlb
);
1185 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1186 struct vm_area_struct
*vma
, pud_t
*pud
,
1187 unsigned long addr
, unsigned long end
,
1188 struct zap_details
*details
)
1193 pmd
= pmd_offset(pud
, addr
);
1195 next
= pmd_addr_end(addr
, end
);
1196 if (pmd_trans_huge(*pmd
)) {
1197 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1198 #ifdef CONFIG_DEBUG_VM
1199 if (!rwsem_is_locked(&tlb
->mm
->mmap_sem
)) {
1200 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1201 __func__
, addr
, end
,
1207 split_huge_page_pmd(vma
, addr
, pmd
);
1208 } else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1213 * Here there can be other concurrent MADV_DONTNEED or
1214 * trans huge page faults running, and if the pmd is
1215 * none or trans huge it can change under us. This is
1216 * because MADV_DONTNEED holds the mmap_sem in read
1219 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1221 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1224 } while (pmd
++, addr
= next
, addr
!= end
);
1229 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1230 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1231 unsigned long addr
, unsigned long end
,
1232 struct zap_details
*details
)
1237 pud
= pud_offset(pgd
, addr
);
1239 next
= pud_addr_end(addr
, end
);
1240 if (pud_none_or_clear_bad(pud
))
1242 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1243 } while (pud
++, addr
= next
, addr
!= end
);
1248 static void unmap_page_range(struct mmu_gather
*tlb
,
1249 struct vm_area_struct
*vma
,
1250 unsigned long addr
, unsigned long end
,
1251 struct zap_details
*details
)
1256 if (details
&& !details
->check_mapping
)
1259 BUG_ON(addr
>= end
);
1260 tlb_start_vma(tlb
, vma
);
1261 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1263 next
= pgd_addr_end(addr
, end
);
1264 if (pgd_none_or_clear_bad(pgd
))
1266 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
1267 } while (pgd
++, addr
= next
, addr
!= end
);
1268 tlb_end_vma(tlb
, vma
);
1272 static void unmap_single_vma(struct mmu_gather
*tlb
,
1273 struct vm_area_struct
*vma
, unsigned long start_addr
,
1274 unsigned long end_addr
,
1275 struct zap_details
*details
)
1277 unsigned long start
= max(vma
->vm_start
, start_addr
);
1280 if (start
>= vma
->vm_end
)
1282 end
= min(vma
->vm_end
, end_addr
);
1283 if (end
<= vma
->vm_start
)
1287 uprobe_munmap(vma
, start
, end
);
1289 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1290 untrack_pfn(vma
, 0, 0);
1293 if (unlikely(is_vm_hugetlb_page(vma
))) {
1295 * It is undesirable to test vma->vm_file as it
1296 * should be non-null for valid hugetlb area.
1297 * However, vm_file will be NULL in the error
1298 * cleanup path of mmap_region. When
1299 * hugetlbfs ->mmap method fails,
1300 * mmap_region() nullifies vma->vm_file
1301 * before calling this function to clean up.
1302 * Since no pte has actually been setup, it is
1303 * safe to do nothing in this case.
1306 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
1307 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1308 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
1311 unmap_page_range(tlb
, vma
, start
, end
, details
);
1316 * unmap_vmas - unmap a range of memory covered by a list of vma's
1317 * @tlb: address of the caller's struct mmu_gather
1318 * @vma: the starting vma
1319 * @start_addr: virtual address at which to start unmapping
1320 * @end_addr: virtual address at which to end unmapping
1322 * Unmap all pages in the vma list.
1324 * Only addresses between `start' and `end' will be unmapped.
1326 * The VMA list must be sorted in ascending virtual address order.
1328 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1329 * range after unmap_vmas() returns. So the only responsibility here is to
1330 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1331 * drops the lock and schedules.
1333 void unmap_vmas(struct mmu_gather
*tlb
,
1334 struct vm_area_struct
*vma
, unsigned long start_addr
,
1335 unsigned long end_addr
)
1337 struct mm_struct
*mm
= vma
->vm_mm
;
1339 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1340 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1341 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1342 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1346 * zap_page_range - remove user pages in a given range
1347 * @vma: vm_area_struct holding the applicable pages
1348 * @start: starting address of pages to zap
1349 * @size: number of bytes to zap
1350 * @details: details of shared cache invalidation
1352 * Caller must protect the VMA list
1354 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1355 unsigned long size
, struct zap_details
*details
)
1357 struct mm_struct
*mm
= vma
->vm_mm
;
1358 struct mmu_gather tlb
;
1359 unsigned long end
= start
+ size
;
1362 tlb_gather_mmu(&tlb
, mm
, start
, end
);
1363 update_hiwater_rss(mm
);
1364 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1365 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
)
1366 unmap_single_vma(&tlb
, vma
, start
, end
, details
);
1367 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1368 tlb_finish_mmu(&tlb
, start
, end
);
1372 * zap_page_range_single - remove user pages in a given range
1373 * @vma: vm_area_struct holding the applicable pages
1374 * @address: starting address of pages to zap
1375 * @size: number of bytes to zap
1376 * @details: details of shared cache invalidation
1378 * The range must fit into one VMA.
1380 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1381 unsigned long size
, struct zap_details
*details
)
1383 struct mm_struct
*mm
= vma
->vm_mm
;
1384 struct mmu_gather tlb
;
1385 unsigned long end
= address
+ size
;
1388 tlb_gather_mmu(&tlb
, mm
, address
, end
);
1389 update_hiwater_rss(mm
);
1390 mmu_notifier_invalidate_range_start(mm
, address
, end
);
1391 unmap_single_vma(&tlb
, vma
, address
, end
, details
);
1392 mmu_notifier_invalidate_range_end(mm
, address
, end
);
1393 tlb_finish_mmu(&tlb
, address
, end
);
1397 * zap_vma_ptes - remove ptes mapping the vma
1398 * @vma: vm_area_struct holding ptes to be zapped
1399 * @address: starting address of pages to zap
1400 * @size: number of bytes to zap
1402 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1404 * The entire address range must be fully contained within the vma.
1406 * Returns 0 if successful.
1408 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1411 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1412 !(vma
->vm_flags
& VM_PFNMAP
))
1414 zap_page_range_single(vma
, address
, size
, NULL
);
1417 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1419 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1422 pgd_t
* pgd
= pgd_offset(mm
, addr
);
1423 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
1425 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
1427 VM_BUG_ON(pmd_trans_huge(*pmd
));
1428 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1435 * This is the old fallback for page remapping.
1437 * For historical reasons, it only allows reserved pages. Only
1438 * old drivers should use this, and they needed to mark their
1439 * pages reserved for the old functions anyway.
1441 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1442 struct page
*page
, pgprot_t prot
)
1444 struct mm_struct
*mm
= vma
->vm_mm
;
1453 flush_dcache_page(page
);
1454 pte
= get_locked_pte(mm
, addr
, &ptl
);
1458 if (!pte_none(*pte
))
1461 /* Ok, finally just insert the thing.. */
1463 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
1464 page_add_file_rmap(page
);
1465 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1468 pte_unmap_unlock(pte
, ptl
);
1471 pte_unmap_unlock(pte
, ptl
);
1477 * vm_insert_page - insert single page into user vma
1478 * @vma: user vma to map to
1479 * @addr: target user address of this page
1480 * @page: source kernel page
1482 * This allows drivers to insert individual pages they've allocated
1485 * The page has to be a nice clean _individual_ kernel allocation.
1486 * If you allocate a compound page, you need to have marked it as
1487 * such (__GFP_COMP), or manually just split the page up yourself
1488 * (see split_page()).
1490 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1491 * took an arbitrary page protection parameter. This doesn't allow
1492 * that. Your vma protection will have to be set up correctly, which
1493 * means that if you want a shared writable mapping, you'd better
1494 * ask for a shared writable mapping!
1496 * The page does not need to be reserved.
1498 * Usually this function is called from f_op->mmap() handler
1499 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1500 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1501 * function from other places, for example from page-fault handler.
1503 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1506 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1508 if (!page_count(page
))
1510 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1511 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
1512 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1513 vma
->vm_flags
|= VM_MIXEDMAP
;
1515 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1517 EXPORT_SYMBOL(vm_insert_page
);
1519 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1520 unsigned long pfn
, pgprot_t prot
)
1522 struct mm_struct
*mm
= vma
->vm_mm
;
1528 pte
= get_locked_pte(mm
, addr
, &ptl
);
1532 if (!pte_none(*pte
))
1535 /* Ok, finally just insert the thing.. */
1536 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
1537 set_pte_at(mm
, addr
, pte
, entry
);
1538 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1542 pte_unmap_unlock(pte
, ptl
);
1548 * vm_insert_pfn - insert single pfn into user vma
1549 * @vma: user vma to map to
1550 * @addr: target user address of this page
1551 * @pfn: source kernel pfn
1553 * Similar to vm_insert_page, this allows drivers to insert individual pages
1554 * they've allocated into a user vma. Same comments apply.
1556 * This function should only be called from a vm_ops->fault handler, and
1557 * in that case the handler should return NULL.
1559 * vma cannot be a COW mapping.
1561 * As this is called only for pages that do not currently exist, we
1562 * do not need to flush old virtual caches or the TLB.
1564 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1568 pgprot_t pgprot
= vma
->vm_page_prot
;
1570 * Technically, architectures with pte_special can avoid all these
1571 * restrictions (same for remap_pfn_range). However we would like
1572 * consistency in testing and feature parity among all, so we should
1573 * try to keep these invariants in place for everybody.
1575 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1576 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1577 (VM_PFNMAP
|VM_MIXEDMAP
));
1578 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1579 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1581 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1583 if (track_pfn_insert(vma
, &pgprot
, pfn
))
1586 ret
= insert_pfn(vma
, addr
, pfn
, pgprot
);
1590 EXPORT_SYMBOL(vm_insert_pfn
);
1592 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1595 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
1597 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1601 * If we don't have pte special, then we have to use the pfn_valid()
1602 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1603 * refcount the page if pfn_valid is true (hence insert_page rather
1604 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1605 * without pte special, it would there be refcounted as a normal page.
1607 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
1610 page
= pfn_to_page(pfn
);
1611 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1613 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1615 EXPORT_SYMBOL(vm_insert_mixed
);
1618 * maps a range of physical memory into the requested pages. the old
1619 * mappings are removed. any references to nonexistent pages results
1620 * in null mappings (currently treated as "copy-on-access")
1622 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1623 unsigned long addr
, unsigned long end
,
1624 unsigned long pfn
, pgprot_t prot
)
1629 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1632 arch_enter_lazy_mmu_mode();
1634 BUG_ON(!pte_none(*pte
));
1635 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
1637 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1638 arch_leave_lazy_mmu_mode();
1639 pte_unmap_unlock(pte
- 1, ptl
);
1643 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1644 unsigned long addr
, unsigned long end
,
1645 unsigned long pfn
, pgprot_t prot
)
1650 pfn
-= addr
>> PAGE_SHIFT
;
1651 pmd
= pmd_alloc(mm
, pud
, addr
);
1654 VM_BUG_ON(pmd_trans_huge(*pmd
));
1656 next
= pmd_addr_end(addr
, end
);
1657 if (remap_pte_range(mm
, pmd
, addr
, next
,
1658 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1660 } while (pmd
++, addr
= next
, addr
!= end
);
1664 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1665 unsigned long addr
, unsigned long end
,
1666 unsigned long pfn
, pgprot_t prot
)
1671 pfn
-= addr
>> PAGE_SHIFT
;
1672 pud
= pud_alloc(mm
, pgd
, addr
);
1676 next
= pud_addr_end(addr
, end
);
1677 if (remap_pmd_range(mm
, pud
, addr
, next
,
1678 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1680 } while (pud
++, addr
= next
, addr
!= end
);
1685 * remap_pfn_range - remap kernel memory to userspace
1686 * @vma: user vma to map to
1687 * @addr: target user address to start at
1688 * @pfn: physical address of kernel memory
1689 * @size: size of map area
1690 * @prot: page protection flags for this mapping
1692 * Note: this is only safe if the mm semaphore is held when called.
1694 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1695 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1699 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1700 struct mm_struct
*mm
= vma
->vm_mm
;
1704 * Physically remapped pages are special. Tell the
1705 * rest of the world about it:
1706 * VM_IO tells people not to look at these pages
1707 * (accesses can have side effects).
1708 * VM_PFNMAP tells the core MM that the base pages are just
1709 * raw PFN mappings, and do not have a "struct page" associated
1712 * Disable vma merging and expanding with mremap().
1714 * Omit vma from core dump, even when VM_IO turned off.
1716 * There's a horrible special case to handle copy-on-write
1717 * behaviour that some programs depend on. We mark the "original"
1718 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1719 * See vm_normal_page() for details.
1721 if (is_cow_mapping(vma
->vm_flags
)) {
1722 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
1724 vma
->vm_pgoff
= pfn
;
1727 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
1731 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
1733 BUG_ON(addr
>= end
);
1734 pfn
-= addr
>> PAGE_SHIFT
;
1735 pgd
= pgd_offset(mm
, addr
);
1736 flush_cache_range(vma
, addr
, end
);
1738 next
= pgd_addr_end(addr
, end
);
1739 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1740 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1743 } while (pgd
++, addr
= next
, addr
!= end
);
1746 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
));
1750 EXPORT_SYMBOL(remap_pfn_range
);
1753 * vm_iomap_memory - remap memory to userspace
1754 * @vma: user vma to map to
1755 * @start: start of area
1756 * @len: size of area
1758 * This is a simplified io_remap_pfn_range() for common driver use. The
1759 * driver just needs to give us the physical memory range to be mapped,
1760 * we'll figure out the rest from the vma information.
1762 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1763 * whatever write-combining details or similar.
1765 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
1767 unsigned long vm_len
, pfn
, pages
;
1769 /* Check that the physical memory area passed in looks valid */
1770 if (start
+ len
< start
)
1773 * You *really* shouldn't map things that aren't page-aligned,
1774 * but we've historically allowed it because IO memory might
1775 * just have smaller alignment.
1777 len
+= start
& ~PAGE_MASK
;
1778 pfn
= start
>> PAGE_SHIFT
;
1779 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
1780 if (pfn
+ pages
< pfn
)
1783 /* We start the mapping 'vm_pgoff' pages into the area */
1784 if (vma
->vm_pgoff
> pages
)
1786 pfn
+= vma
->vm_pgoff
;
1787 pages
-= vma
->vm_pgoff
;
1789 /* Can we fit all of the mapping? */
1790 vm_len
= vma
->vm_end
- vma
->vm_start
;
1791 if (vm_len
>> PAGE_SHIFT
> pages
)
1794 /* Ok, let it rip */
1795 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
1797 EXPORT_SYMBOL(vm_iomap_memory
);
1799 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1800 unsigned long addr
, unsigned long end
,
1801 pte_fn_t fn
, void *data
)
1806 spinlock_t
*uninitialized_var(ptl
);
1808 pte
= (mm
== &init_mm
) ?
1809 pte_alloc_kernel(pmd
, addr
) :
1810 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1814 BUG_ON(pmd_huge(*pmd
));
1816 arch_enter_lazy_mmu_mode();
1818 token
= pmd_pgtable(*pmd
);
1821 err
= fn(pte
++, token
, addr
, data
);
1824 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1826 arch_leave_lazy_mmu_mode();
1829 pte_unmap_unlock(pte
-1, ptl
);
1833 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1834 unsigned long addr
, unsigned long end
,
1835 pte_fn_t fn
, void *data
)
1841 BUG_ON(pud_huge(*pud
));
1843 pmd
= pmd_alloc(mm
, pud
, addr
);
1847 next
= pmd_addr_end(addr
, end
);
1848 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
1851 } while (pmd
++, addr
= next
, addr
!= end
);
1855 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1856 unsigned long addr
, unsigned long end
,
1857 pte_fn_t fn
, void *data
)
1863 pud
= pud_alloc(mm
, pgd
, addr
);
1867 next
= pud_addr_end(addr
, end
);
1868 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
1871 } while (pud
++, addr
= next
, addr
!= end
);
1876 * Scan a region of virtual memory, filling in page tables as necessary
1877 * and calling a provided function on each leaf page table.
1879 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
1880 unsigned long size
, pte_fn_t fn
, void *data
)
1884 unsigned long end
= addr
+ size
;
1887 BUG_ON(addr
>= end
);
1888 pgd
= pgd_offset(mm
, addr
);
1890 next
= pgd_addr_end(addr
, end
);
1891 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
1894 } while (pgd
++, addr
= next
, addr
!= end
);
1898 EXPORT_SYMBOL_GPL(apply_to_page_range
);
1901 * handle_pte_fault chooses page fault handler according to an entry which was
1902 * read non-atomically. Before making any commitment, on those architectures
1903 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1904 * parts, do_swap_page must check under lock before unmapping the pte and
1905 * proceeding (but do_wp_page is only called after already making such a check;
1906 * and do_anonymous_page can safely check later on).
1908 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
1909 pte_t
*page_table
, pte_t orig_pte
)
1912 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1913 if (sizeof(pte_t
) > sizeof(unsigned long)) {
1914 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
1916 same
= pte_same(*page_table
, orig_pte
);
1920 pte_unmap(page_table
);
1924 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
1926 debug_dma_assert_idle(src
);
1929 * If the source page was a PFN mapping, we don't have
1930 * a "struct page" for it. We do a best-effort copy by
1931 * just copying from the original user address. If that
1932 * fails, we just zero-fill it. Live with it.
1934 if (unlikely(!src
)) {
1935 void *kaddr
= kmap_atomic(dst
);
1936 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
1939 * This really shouldn't fail, because the page is there
1940 * in the page tables. But it might just be unreadable,
1941 * in which case we just give up and fill the result with
1944 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
1946 kunmap_atomic(kaddr
);
1947 flush_dcache_page(dst
);
1949 copy_user_highpage(dst
, src
, va
, vma
);
1953 * Notify the address space that the page is about to become writable so that
1954 * it can prohibit this or wait for the page to get into an appropriate state.
1956 * We do this without the lock held, so that it can sleep if it needs to.
1958 static int do_page_mkwrite(struct vm_area_struct
*vma
, struct page
*page
,
1959 unsigned long address
)
1961 struct vm_fault vmf
;
1964 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
1965 vmf
.pgoff
= page
->index
;
1966 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
1968 vmf
.cow_page
= NULL
;
1970 ret
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
1971 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
1973 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
1975 if (!page
->mapping
) {
1977 return 0; /* retry */
1979 ret
|= VM_FAULT_LOCKED
;
1981 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1986 * Handle write page faults for pages that can be reused in the current vma
1988 * This can happen either due to the mapping being with the VM_SHARED flag,
1989 * or due to us being the last reference standing to the page. In either
1990 * case, all we need to do here is to mark the page as writable and update
1991 * any related book-keeping.
1993 static inline int wp_page_reuse(struct mm_struct
*mm
,
1994 struct vm_area_struct
*vma
, unsigned long address
,
1995 pte_t
*page_table
, spinlock_t
*ptl
, pte_t orig_pte
,
1996 struct page
*page
, int page_mkwrite
,
2002 * Clear the pages cpupid information as the existing
2003 * information potentially belongs to a now completely
2004 * unrelated process.
2007 page_cpupid_xchg_last(page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2009 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2010 entry
= pte_mkyoung(orig_pte
);
2011 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2012 if (ptep_set_access_flags(vma
, address
, page_table
, entry
, 1))
2013 update_mmu_cache(vma
, address
, page_table
);
2014 pte_unmap_unlock(page_table
, ptl
);
2017 struct address_space
*mapping
;
2023 dirtied
= set_page_dirty(page
);
2024 VM_BUG_ON_PAGE(PageAnon(page
), page
);
2025 mapping
= page
->mapping
;
2027 page_cache_release(page
);
2029 if ((dirtied
|| page_mkwrite
) && mapping
) {
2031 * Some device drivers do not set page.mapping
2032 * but still dirty their pages
2034 balance_dirty_pages_ratelimited(mapping
);
2038 file_update_time(vma
->vm_file
);
2041 return VM_FAULT_WRITE
;
2045 * Handle the case of a page which we actually need to copy to a new page.
2047 * Called with mmap_sem locked and the old page referenced, but
2048 * without the ptl held.
2050 * High level logic flow:
2052 * - Allocate a page, copy the content of the old page to the new one.
2053 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2054 * - Take the PTL. If the pte changed, bail out and release the allocated page
2055 * - If the pte is still the way we remember it, update the page table and all
2056 * relevant references. This includes dropping the reference the page-table
2057 * held to the old page, as well as updating the rmap.
2058 * - In any case, unlock the PTL and drop the reference we took to the old page.
2060 static int wp_page_copy(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2061 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2062 pte_t orig_pte
, struct page
*old_page
)
2064 struct page
*new_page
= NULL
;
2065 spinlock_t
*ptl
= NULL
;
2067 int page_copied
= 0;
2068 const unsigned long mmun_start
= address
& PAGE_MASK
; /* For mmu_notifiers */
2069 const unsigned long mmun_end
= mmun_start
+ PAGE_SIZE
; /* For mmu_notifiers */
2070 struct mem_cgroup
*memcg
;
2072 if (unlikely(anon_vma_prepare(vma
)))
2075 if (is_zero_pfn(pte_pfn(orig_pte
))) {
2076 new_page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2080 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2083 cow_user_page(new_page
, old_page
, address
, vma
);
2085 __SetPageUptodate(new_page
);
2087 if (mem_cgroup_try_charge(new_page
, mm
, GFP_KERNEL
, &memcg
))
2090 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2093 * Re-check the pte - we dropped the lock
2095 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2096 if (likely(pte_same(*page_table
, orig_pte
))) {
2098 if (!PageAnon(old_page
)) {
2099 dec_mm_counter_fast(mm
, MM_FILEPAGES
);
2100 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2103 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2105 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2106 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2107 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2109 * Clear the pte entry and flush it first, before updating the
2110 * pte with the new entry. This will avoid a race condition
2111 * seen in the presence of one thread doing SMC and another
2114 ptep_clear_flush_notify(vma
, address
, page_table
);
2115 page_add_new_anon_rmap(new_page
, vma
, address
);
2116 mem_cgroup_commit_charge(new_page
, memcg
, false);
2117 lru_cache_add_active_or_unevictable(new_page
, vma
);
2119 * We call the notify macro here because, when using secondary
2120 * mmu page tables (such as kvm shadow page tables), we want the
2121 * new page to be mapped directly into the secondary page table.
2123 set_pte_at_notify(mm
, address
, page_table
, entry
);
2124 update_mmu_cache(vma
, address
, page_table
);
2127 * Only after switching the pte to the new page may
2128 * we remove the mapcount here. Otherwise another
2129 * process may come and find the rmap count decremented
2130 * before the pte is switched to the new page, and
2131 * "reuse" the old page writing into it while our pte
2132 * here still points into it and can be read by other
2135 * The critical issue is to order this
2136 * page_remove_rmap with the ptp_clear_flush above.
2137 * Those stores are ordered by (if nothing else,)
2138 * the barrier present in the atomic_add_negative
2139 * in page_remove_rmap.
2141 * Then the TLB flush in ptep_clear_flush ensures that
2142 * no process can access the old page before the
2143 * decremented mapcount is visible. And the old page
2144 * cannot be reused until after the decremented
2145 * mapcount is visible. So transitively, TLBs to
2146 * old page will be flushed before it can be reused.
2148 page_remove_rmap(old_page
);
2151 /* Free the old page.. */
2152 new_page
= old_page
;
2155 mem_cgroup_cancel_charge(new_page
, memcg
);
2159 page_cache_release(new_page
);
2161 pte_unmap_unlock(page_table
, ptl
);
2162 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2165 * Don't let another task, with possibly unlocked vma,
2166 * keep the mlocked page.
2168 if (page_copied
&& (vma
->vm_flags
& VM_LOCKED
)) {
2169 lock_page(old_page
); /* LRU manipulation */
2170 munlock_vma_page(old_page
);
2171 unlock_page(old_page
);
2173 page_cache_release(old_page
);
2175 return page_copied
? VM_FAULT_WRITE
: 0;
2177 page_cache_release(new_page
);
2180 page_cache_release(old_page
);
2181 return VM_FAULT_OOM
;
2184 static int wp_page_shared(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2185 unsigned long address
, pte_t
*page_table
,
2186 pmd_t
*pmd
, spinlock_t
*ptl
, pte_t orig_pte
,
2187 struct page
*old_page
)
2190 int page_mkwrite
= 0;
2192 page_cache_get(old_page
);
2195 * Only catch write-faults on shared writable pages,
2196 * read-only shared pages can get COWed by
2197 * get_user_pages(.write=1, .force=1).
2199 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2202 pte_unmap_unlock(page_table
, ptl
);
2203 tmp
= do_page_mkwrite(vma
, old_page
, address
);
2204 if (unlikely(!tmp
|| (tmp
&
2205 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
2206 page_cache_release(old_page
);
2210 * Since we dropped the lock we need to revalidate
2211 * the PTE as someone else may have changed it. If
2212 * they did, we just return, as we can count on the
2213 * MMU to tell us if they didn't also make it writable.
2215 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2217 if (!pte_same(*page_table
, orig_pte
)) {
2218 unlock_page(old_page
);
2219 pte_unmap_unlock(page_table
, ptl
);
2220 page_cache_release(old_page
);
2226 return wp_page_reuse(mm
, vma
, address
, page_table
, ptl
,
2227 orig_pte
, old_page
, page_mkwrite
, 1);
2231 * This routine handles present pages, when users try to write
2232 * to a shared page. It is done by copying the page to a new address
2233 * and decrementing the shared-page counter for the old page.
2235 * Note that this routine assumes that the protection checks have been
2236 * done by the caller (the low-level page fault routine in most cases).
2237 * Thus we can safely just mark it writable once we've done any necessary
2240 * We also mark the page dirty at this point even though the page will
2241 * change only once the write actually happens. This avoids a few races,
2242 * and potentially makes it more efficient.
2244 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2245 * but allow concurrent faults), with pte both mapped and locked.
2246 * We return with mmap_sem still held, but pte unmapped and unlocked.
2248 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2249 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2250 spinlock_t
*ptl
, pte_t orig_pte
)
2253 struct page
*old_page
;
2255 old_page
= vm_normal_page(vma
, address
, orig_pte
);
2258 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2261 * We should not cow pages in a shared writeable mapping.
2262 * Just mark the pages writable as we can't do any dirty
2263 * accounting on raw pfn maps.
2265 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2266 (VM_WRITE
|VM_SHARED
))
2267 return wp_page_reuse(mm
, vma
, address
, page_table
, ptl
,
2268 orig_pte
, old_page
, 0, 0);
2270 pte_unmap_unlock(page_table
, ptl
);
2271 return wp_page_copy(mm
, vma
, address
, page_table
, pmd
,
2272 orig_pte
, old_page
);
2276 * Take out anonymous pages first, anonymous shared vmas are
2277 * not dirty accountable.
2279 if (PageAnon(old_page
) && !PageKsm(old_page
)) {
2280 if (!trylock_page(old_page
)) {
2281 page_cache_get(old_page
);
2282 pte_unmap_unlock(page_table
, ptl
);
2283 lock_page(old_page
);
2284 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2286 if (!pte_same(*page_table
, orig_pte
)) {
2287 unlock_page(old_page
);
2288 pte_unmap_unlock(page_table
, ptl
);
2289 page_cache_release(old_page
);
2292 page_cache_release(old_page
);
2294 if (reuse_swap_page(old_page
)) {
2296 * The page is all ours. Move it to our anon_vma so
2297 * the rmap code will not search our parent or siblings.
2298 * Protected against the rmap code by the page lock.
2300 page_move_anon_rmap(old_page
, vma
, address
);
2301 unlock_page(old_page
);
2302 return wp_page_reuse(mm
, vma
, address
, page_table
, ptl
,
2303 orig_pte
, old_page
, 0, 0);
2305 unlock_page(old_page
);
2306 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2307 (VM_WRITE
|VM_SHARED
))) {
2308 return wp_page_shared(mm
, vma
, address
, page_table
, pmd
,
2309 ptl
, orig_pte
, old_page
);
2313 * Ok, we need to copy. Oh, well..
2315 page_cache_get(old_page
);
2317 pte_unmap_unlock(page_table
, ptl
);
2318 return wp_page_copy(mm
, vma
, address
, page_table
, pmd
,
2319 orig_pte
, old_page
);
2322 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2323 unsigned long start_addr
, unsigned long end_addr
,
2324 struct zap_details
*details
)
2326 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2329 static inline void unmap_mapping_range_tree(struct rb_root
*root
,
2330 struct zap_details
*details
)
2332 struct vm_area_struct
*vma
;
2333 pgoff_t vba
, vea
, zba
, zea
;
2335 vma_interval_tree_foreach(vma
, root
,
2336 details
->first_index
, details
->last_index
) {
2338 vba
= vma
->vm_pgoff
;
2339 vea
= vba
+ vma_pages(vma
) - 1;
2340 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2341 zba
= details
->first_index
;
2344 zea
= details
->last_index
;
2348 unmap_mapping_range_vma(vma
,
2349 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2350 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2356 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2357 * address_space corresponding to the specified page range in the underlying
2360 * @mapping: the address space containing mmaps to be unmapped.
2361 * @holebegin: byte in first page to unmap, relative to the start of
2362 * the underlying file. This will be rounded down to a PAGE_SIZE
2363 * boundary. Note that this is different from truncate_pagecache(), which
2364 * must keep the partial page. In contrast, we must get rid of
2366 * @holelen: size of prospective hole in bytes. This will be rounded
2367 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2369 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2370 * but 0 when invalidating pagecache, don't throw away private data.
2372 void unmap_mapping_range(struct address_space
*mapping
,
2373 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2375 struct zap_details details
;
2376 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2377 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2379 /* Check for overflow. */
2380 if (sizeof(holelen
) > sizeof(hlen
)) {
2382 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2383 if (holeend
& ~(long long)ULONG_MAX
)
2384 hlen
= ULONG_MAX
- hba
+ 1;
2387 details
.check_mapping
= even_cows
? NULL
: mapping
;
2388 details
.first_index
= hba
;
2389 details
.last_index
= hba
+ hlen
- 1;
2390 if (details
.last_index
< details
.first_index
)
2391 details
.last_index
= ULONG_MAX
;
2394 /* DAX uses i_mmap_lock to serialise file truncate vs page fault */
2395 i_mmap_lock_write(mapping
);
2396 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
)))
2397 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2398 i_mmap_unlock_write(mapping
);
2400 EXPORT_SYMBOL(unmap_mapping_range
);
2403 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2404 * but allow concurrent faults), and pte mapped but not yet locked.
2405 * We return with pte unmapped and unlocked.
2407 * We return with the mmap_sem locked or unlocked in the same cases
2408 * as does filemap_fault().
2410 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2411 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2412 unsigned int flags
, pte_t orig_pte
)
2415 struct page
*page
, *swapcache
;
2416 struct mem_cgroup
*memcg
;
2423 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2426 entry
= pte_to_swp_entry(orig_pte
);
2427 if (unlikely(non_swap_entry(entry
))) {
2428 if (is_migration_entry(entry
)) {
2429 migration_entry_wait(mm
, pmd
, address
);
2430 } else if (is_hwpoison_entry(entry
)) {
2431 ret
= VM_FAULT_HWPOISON
;
2433 print_bad_pte(vma
, address
, orig_pte
, NULL
);
2434 ret
= VM_FAULT_SIGBUS
;
2438 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2439 page
= lookup_swap_cache(entry
);
2441 page
= swapin_readahead(entry
,
2442 GFP_HIGHUSER_MOVABLE
, vma
, address
);
2445 * Back out if somebody else faulted in this pte
2446 * while we released the pte lock.
2448 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2449 if (likely(pte_same(*page_table
, orig_pte
)))
2451 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2455 /* Had to read the page from swap area: Major fault */
2456 ret
= VM_FAULT_MAJOR
;
2457 count_vm_event(PGMAJFAULT
);
2458 mem_cgroup_count_vm_event(mm
, PGMAJFAULT
);
2459 } else if (PageHWPoison(page
)) {
2461 * hwpoisoned dirty swapcache pages are kept for killing
2462 * owner processes (which may be unknown at hwpoison time)
2464 ret
= VM_FAULT_HWPOISON
;
2465 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2471 locked
= lock_page_or_retry(page
, mm
, flags
);
2473 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2475 ret
|= VM_FAULT_RETRY
;
2480 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2481 * release the swapcache from under us. The page pin, and pte_same
2482 * test below, are not enough to exclude that. Even if it is still
2483 * swapcache, we need to check that the page's swap has not changed.
2485 if (unlikely(!PageSwapCache(page
) || page_private(page
) != entry
.val
))
2488 page
= ksm_might_need_to_copy(page
, vma
, address
);
2489 if (unlikely(!page
)) {
2495 if (mem_cgroup_try_charge(page
, mm
, GFP_KERNEL
, &memcg
)) {
2501 * Back out if somebody else already faulted in this pte.
2503 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2504 if (unlikely(!pte_same(*page_table
, orig_pte
)))
2507 if (unlikely(!PageUptodate(page
))) {
2508 ret
= VM_FAULT_SIGBUS
;
2513 * The page isn't present yet, go ahead with the fault.
2515 * Be careful about the sequence of operations here.
2516 * To get its accounting right, reuse_swap_page() must be called
2517 * while the page is counted on swap but not yet in mapcount i.e.
2518 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2519 * must be called after the swap_free(), or it will never succeed.
2522 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2523 dec_mm_counter_fast(mm
, MM_SWAPENTS
);
2524 pte
= mk_pte(page
, vma
->vm_page_prot
);
2525 if ((flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
)) {
2526 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
2527 flags
&= ~FAULT_FLAG_WRITE
;
2528 ret
|= VM_FAULT_WRITE
;
2531 flush_icache_page(vma
, page
);
2532 if (pte_swp_soft_dirty(orig_pte
))
2533 pte
= pte_mksoft_dirty(pte
);
2534 set_pte_at(mm
, address
, page_table
, pte
);
2535 if (page
== swapcache
) {
2536 do_page_add_anon_rmap(page
, vma
, address
, exclusive
);
2537 mem_cgroup_commit_charge(page
, memcg
, true);
2538 } else { /* ksm created a completely new copy */
2539 page_add_new_anon_rmap(page
, vma
, address
);
2540 mem_cgroup_commit_charge(page
, memcg
, false);
2541 lru_cache_add_active_or_unevictable(page
, vma
);
2545 if (vm_swap_full() || (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
2546 try_to_free_swap(page
);
2548 if (page
!= swapcache
) {
2550 * Hold the lock to avoid the swap entry to be reused
2551 * until we take the PT lock for the pte_same() check
2552 * (to avoid false positives from pte_same). For
2553 * further safety release the lock after the swap_free
2554 * so that the swap count won't change under a
2555 * parallel locked swapcache.
2557 unlock_page(swapcache
);
2558 page_cache_release(swapcache
);
2561 if (flags
& FAULT_FLAG_WRITE
) {
2562 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
2563 if (ret
& VM_FAULT_ERROR
)
2564 ret
&= VM_FAULT_ERROR
;
2568 /* No need to invalidate - it was non-present before */
2569 update_mmu_cache(vma
, address
, page_table
);
2571 pte_unmap_unlock(page_table
, ptl
);
2575 mem_cgroup_cancel_charge(page
, memcg
);
2576 pte_unmap_unlock(page_table
, ptl
);
2580 page_cache_release(page
);
2581 if (page
!= swapcache
) {
2582 unlock_page(swapcache
);
2583 page_cache_release(swapcache
);
2589 * This is like a special single-page "expand_{down|up}wards()",
2590 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2591 * doesn't hit another vma.
2593 static inline int check_stack_guard_page(struct vm_area_struct
*vma
, unsigned long address
)
2595 address
&= PAGE_MASK
;
2596 if ((vma
->vm_flags
& VM_GROWSDOWN
) && address
== vma
->vm_start
) {
2597 struct vm_area_struct
*prev
= vma
->vm_prev
;
2600 * Is there a mapping abutting this one below?
2602 * That's only ok if it's the same stack mapping
2603 * that has gotten split..
2605 if (prev
&& prev
->vm_end
== address
)
2606 return prev
->vm_flags
& VM_GROWSDOWN
? 0 : -ENOMEM
;
2608 return expand_downwards(vma
, address
- PAGE_SIZE
);
2610 if ((vma
->vm_flags
& VM_GROWSUP
) && address
+ PAGE_SIZE
== vma
->vm_end
) {
2611 struct vm_area_struct
*next
= vma
->vm_next
;
2613 /* As VM_GROWSDOWN but s/below/above/ */
2614 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
)
2615 return next
->vm_flags
& VM_GROWSUP
? 0 : -ENOMEM
;
2617 return expand_upwards(vma
, address
+ PAGE_SIZE
);
2623 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2624 * but allow concurrent faults), and pte mapped but not yet locked.
2625 * We return with mmap_sem still held, but pte unmapped and unlocked.
2627 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2628 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2631 struct mem_cgroup
*memcg
;
2636 pte_unmap(page_table
);
2638 /* Check if we need to add a guard page to the stack */
2639 if (check_stack_guard_page(vma
, address
) < 0)
2640 return VM_FAULT_SIGSEGV
;
2642 /* Use the zero-page for reads */
2643 if (!(flags
& FAULT_FLAG_WRITE
) && !mm_forbids_zeropage(mm
)) {
2644 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(address
),
2645 vma
->vm_page_prot
));
2646 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2647 if (!pte_none(*page_table
))
2652 /* Allocate our own private page. */
2653 if (unlikely(anon_vma_prepare(vma
)))
2655 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2659 * The memory barrier inside __SetPageUptodate makes sure that
2660 * preceeding stores to the page contents become visible before
2661 * the set_pte_at() write.
2663 __SetPageUptodate(page
);
2665 if (mem_cgroup_try_charge(page
, mm
, GFP_KERNEL
, &memcg
))
2668 entry
= mk_pte(page
, vma
->vm_page_prot
);
2669 if (vma
->vm_flags
& VM_WRITE
)
2670 entry
= pte_mkwrite(pte_mkdirty(entry
));
2672 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2673 if (!pte_none(*page_table
))
2676 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2677 page_add_new_anon_rmap(page
, vma
, address
);
2678 mem_cgroup_commit_charge(page
, memcg
, false);
2679 lru_cache_add_active_or_unevictable(page
, vma
);
2681 set_pte_at(mm
, address
, page_table
, entry
);
2683 /* No need to invalidate - it was non-present before */
2684 update_mmu_cache(vma
, address
, page_table
);
2686 pte_unmap_unlock(page_table
, ptl
);
2689 mem_cgroup_cancel_charge(page
, memcg
);
2690 page_cache_release(page
);
2693 page_cache_release(page
);
2695 return VM_FAULT_OOM
;
2699 * The mmap_sem must have been held on entry, and may have been
2700 * released depending on flags and vma->vm_ops->fault() return value.
2701 * See filemap_fault() and __lock_page_retry().
2703 static int __do_fault(struct vm_area_struct
*vma
, unsigned long address
,
2704 pgoff_t pgoff
, unsigned int flags
,
2705 struct page
*cow_page
, struct page
**page
)
2707 struct vm_fault vmf
;
2710 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
2714 vmf
.cow_page
= cow_page
;
2716 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
2717 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
2722 if (unlikely(PageHWPoison(vmf
.page
))) {
2723 if (ret
& VM_FAULT_LOCKED
)
2724 unlock_page(vmf
.page
);
2725 page_cache_release(vmf
.page
);
2726 return VM_FAULT_HWPOISON
;
2729 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
2730 lock_page(vmf
.page
);
2732 VM_BUG_ON_PAGE(!PageLocked(vmf
.page
), vmf
.page
);
2740 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2742 * @vma: virtual memory area
2743 * @address: user virtual address
2744 * @page: page to map
2745 * @pte: pointer to target page table entry
2746 * @write: true, if new entry is writable
2747 * @anon: true, if it's anonymous page
2749 * Caller must hold page table lock relevant for @pte.
2751 * Target users are page handler itself and implementations of
2752 * vm_ops->map_pages.
2754 void do_set_pte(struct vm_area_struct
*vma
, unsigned long address
,
2755 struct page
*page
, pte_t
*pte
, bool write
, bool anon
)
2759 flush_icache_page(vma
, page
);
2760 entry
= mk_pte(page
, vma
->vm_page_prot
);
2762 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2764 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
2765 page_add_new_anon_rmap(page
, vma
, address
);
2767 inc_mm_counter_fast(vma
->vm_mm
, MM_FILEPAGES
);
2768 page_add_file_rmap(page
);
2770 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
2772 /* no need to invalidate: a not-present page won't be cached */
2773 update_mmu_cache(vma
, address
, pte
);
2776 static unsigned long fault_around_bytes __read_mostly
=
2777 rounddown_pow_of_two(65536);
2779 #ifdef CONFIG_DEBUG_FS
2780 static int fault_around_bytes_get(void *data
, u64
*val
)
2782 *val
= fault_around_bytes
;
2787 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2788 * rounded down to nearest page order. It's what do_fault_around() expects to
2791 static int fault_around_bytes_set(void *data
, u64 val
)
2793 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
2795 if (val
> PAGE_SIZE
)
2796 fault_around_bytes
= rounddown_pow_of_two(val
);
2798 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
2801 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops
,
2802 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
2804 static int __init
fault_around_debugfs(void)
2808 ret
= debugfs_create_file("fault_around_bytes", 0644, NULL
, NULL
,
2809 &fault_around_bytes_fops
);
2811 pr_warn("Failed to create fault_around_bytes in debugfs");
2814 late_initcall(fault_around_debugfs
);
2818 * do_fault_around() tries to map few pages around the fault address. The hope
2819 * is that the pages will be needed soon and this will lower the number of
2822 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2823 * not ready to be mapped: not up-to-date, locked, etc.
2825 * This function is called with the page table lock taken. In the split ptlock
2826 * case the page table lock only protects only those entries which belong to
2827 * the page table corresponding to the fault address.
2829 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2832 * fault_around_pages() defines how many pages we'll try to map.
2833 * do_fault_around() expects it to return a power of two less than or equal to
2836 * The virtual address of the area that we map is naturally aligned to the
2837 * fault_around_pages() value (and therefore to page order). This way it's
2838 * easier to guarantee that we don't cross page table boundaries.
2840 static void do_fault_around(struct vm_area_struct
*vma
, unsigned long address
,
2841 pte_t
*pte
, pgoff_t pgoff
, unsigned int flags
)
2843 unsigned long start_addr
, nr_pages
, mask
;
2845 struct vm_fault vmf
;
2848 nr_pages
= ACCESS_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
2849 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
2851 start_addr
= max(address
& mask
, vma
->vm_start
);
2852 off
= ((address
- start_addr
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
2857 * max_pgoff is either end of page table or end of vma
2858 * or fault_around_pages() from pgoff, depending what is nearest.
2860 max_pgoff
= pgoff
- ((start_addr
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
2862 max_pgoff
= min3(max_pgoff
, vma_pages(vma
) + vma
->vm_pgoff
- 1,
2863 pgoff
+ nr_pages
- 1);
2865 /* Check if it makes any sense to call ->map_pages */
2866 while (!pte_none(*pte
)) {
2867 if (++pgoff
> max_pgoff
)
2869 start_addr
+= PAGE_SIZE
;
2870 if (start_addr
>= vma
->vm_end
)
2875 vmf
.virtual_address
= (void __user
*) start_addr
;
2878 vmf
.max_pgoff
= max_pgoff
;
2880 vma
->vm_ops
->map_pages(vma
, &vmf
);
2883 static int do_read_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2884 unsigned long address
, pmd_t
*pmd
,
2885 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
2887 struct page
*fault_page
;
2893 * Let's call ->map_pages() first and use ->fault() as fallback
2894 * if page by the offset is not ready to be mapped (cold cache or
2897 if (vma
->vm_ops
->map_pages
&& fault_around_bytes
>> PAGE_SHIFT
> 1) {
2898 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2899 do_fault_around(vma
, address
, pte
, pgoff
, flags
);
2900 if (!pte_same(*pte
, orig_pte
))
2902 pte_unmap_unlock(pte
, ptl
);
2905 ret
= __do_fault(vma
, address
, pgoff
, flags
, NULL
, &fault_page
);
2906 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
2909 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2910 if (unlikely(!pte_same(*pte
, orig_pte
))) {
2911 pte_unmap_unlock(pte
, ptl
);
2912 unlock_page(fault_page
);
2913 page_cache_release(fault_page
);
2916 do_set_pte(vma
, address
, fault_page
, pte
, false, false);
2917 unlock_page(fault_page
);
2919 pte_unmap_unlock(pte
, ptl
);
2923 static int do_cow_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2924 unsigned long address
, pmd_t
*pmd
,
2925 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
2927 struct page
*fault_page
, *new_page
;
2928 struct mem_cgroup
*memcg
;
2933 if (unlikely(anon_vma_prepare(vma
)))
2934 return VM_FAULT_OOM
;
2936 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2938 return VM_FAULT_OOM
;
2940 if (mem_cgroup_try_charge(new_page
, mm
, GFP_KERNEL
, &memcg
)) {
2941 page_cache_release(new_page
);
2942 return VM_FAULT_OOM
;
2945 ret
= __do_fault(vma
, address
, pgoff
, flags
, new_page
, &fault_page
);
2946 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
2950 copy_user_highpage(new_page
, fault_page
, address
, vma
);
2951 __SetPageUptodate(new_page
);
2953 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2954 if (unlikely(!pte_same(*pte
, orig_pte
))) {
2955 pte_unmap_unlock(pte
, ptl
);
2957 unlock_page(fault_page
);
2958 page_cache_release(fault_page
);
2961 * The fault handler has no page to lock, so it holds
2962 * i_mmap_lock for read to protect against truncate.
2964 i_mmap_unlock_read(vma
->vm_file
->f_mapping
);
2968 do_set_pte(vma
, address
, new_page
, pte
, true, true);
2969 mem_cgroup_commit_charge(new_page
, memcg
, false);
2970 lru_cache_add_active_or_unevictable(new_page
, vma
);
2971 pte_unmap_unlock(pte
, ptl
);
2973 unlock_page(fault_page
);
2974 page_cache_release(fault_page
);
2977 * The fault handler has no page to lock, so it holds
2978 * i_mmap_lock for read to protect against truncate.
2980 i_mmap_unlock_read(vma
->vm_file
->f_mapping
);
2984 mem_cgroup_cancel_charge(new_page
, memcg
);
2985 page_cache_release(new_page
);
2989 static int do_shared_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2990 unsigned long address
, pmd_t
*pmd
,
2991 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
2993 struct page
*fault_page
;
2994 struct address_space
*mapping
;
3000 ret
= __do_fault(vma
, address
, pgoff
, flags
, NULL
, &fault_page
);
3001 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3005 * Check if the backing address space wants to know that the page is
3006 * about to become writable
3008 if (vma
->vm_ops
->page_mkwrite
) {
3009 unlock_page(fault_page
);
3010 tmp
= do_page_mkwrite(vma
, fault_page
, address
);
3011 if (unlikely(!tmp
||
3012 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3013 page_cache_release(fault_page
);
3018 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3019 if (unlikely(!pte_same(*pte
, orig_pte
))) {
3020 pte_unmap_unlock(pte
, ptl
);
3021 unlock_page(fault_page
);
3022 page_cache_release(fault_page
);
3025 do_set_pte(vma
, address
, fault_page
, pte
, true, false);
3026 pte_unmap_unlock(pte
, ptl
);
3028 if (set_page_dirty(fault_page
))
3031 * Take a local copy of the address_space - page.mapping may be zeroed
3032 * by truncate after unlock_page(). The address_space itself remains
3033 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3034 * release semantics to prevent the compiler from undoing this copying.
3036 mapping
= fault_page
->mapping
;
3037 unlock_page(fault_page
);
3038 if ((dirtied
|| vma
->vm_ops
->page_mkwrite
) && mapping
) {
3040 * Some device drivers do not set page.mapping but still
3043 balance_dirty_pages_ratelimited(mapping
);
3046 if (!vma
->vm_ops
->page_mkwrite
)
3047 file_update_time(vma
->vm_file
);
3053 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3054 * but allow concurrent faults).
3055 * The mmap_sem may have been released depending on flags and our
3056 * return value. See filemap_fault() and __lock_page_or_retry().
3058 static int do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3059 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3060 unsigned int flags
, pte_t orig_pte
)
3062 pgoff_t pgoff
= (((address
& PAGE_MASK
)
3063 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
3065 pte_unmap(page_table
);
3066 if (!(flags
& FAULT_FLAG_WRITE
))
3067 return do_read_fault(mm
, vma
, address
, pmd
, pgoff
, flags
,
3069 if (!(vma
->vm_flags
& VM_SHARED
))
3070 return do_cow_fault(mm
, vma
, address
, pmd
, pgoff
, flags
,
3072 return do_shared_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3075 static int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3076 unsigned long addr
, int page_nid
,
3081 count_vm_numa_event(NUMA_HINT_FAULTS
);
3082 if (page_nid
== numa_node_id()) {
3083 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3084 *flags
|= TNF_FAULT_LOCAL
;
3087 return mpol_misplaced(page
, vma
, addr
);
3090 static int do_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3091 unsigned long addr
, pte_t pte
, pte_t
*ptep
, pmd_t
*pmd
)
3093 struct page
*page
= NULL
;
3098 bool migrated
= false;
3099 bool was_writable
= pte_write(pte
);
3102 /* A PROT_NONE fault should not end up here */
3103 BUG_ON(!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)));
3106 * The "pte" at this point cannot be used safely without
3107 * validation through pte_unmap_same(). It's of NUMA type but
3108 * the pfn may be screwed if the read is non atomic.
3110 * We can safely just do a "set_pte_at()", because the old
3111 * page table entry is not accessible, so there would be no
3112 * concurrent hardware modifications to the PTE.
3114 ptl
= pte_lockptr(mm
, pmd
);
3116 if (unlikely(!pte_same(*ptep
, pte
))) {
3117 pte_unmap_unlock(ptep
, ptl
);
3121 /* Make it present again */
3122 pte
= pte_modify(pte
, vma
->vm_page_prot
);
3123 pte
= pte_mkyoung(pte
);
3125 pte
= pte_mkwrite(pte
);
3126 set_pte_at(mm
, addr
, ptep
, pte
);
3127 update_mmu_cache(vma
, addr
, ptep
);
3129 page
= vm_normal_page(vma
, addr
, pte
);
3131 pte_unmap_unlock(ptep
, ptl
);
3136 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3137 * much anyway since they can be in shared cache state. This misses
3138 * the case where a mapping is writable but the process never writes
3139 * to it but pte_write gets cleared during protection updates and
3140 * pte_dirty has unpredictable behaviour between PTE scan updates,
3141 * background writeback, dirty balancing and application behaviour.
3143 if (!(vma
->vm_flags
& VM_WRITE
))
3144 flags
|= TNF_NO_GROUP
;
3147 * Flag if the page is shared between multiple address spaces. This
3148 * is later used when determining whether to group tasks together
3150 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
3151 flags
|= TNF_SHARED
;
3153 last_cpupid
= page_cpupid_last(page
);
3154 page_nid
= page_to_nid(page
);
3155 target_nid
= numa_migrate_prep(page
, vma
, addr
, page_nid
, &flags
);
3156 pte_unmap_unlock(ptep
, ptl
);
3157 if (target_nid
== -1) {
3162 /* Migrate to the requested node */
3163 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
3165 page_nid
= target_nid
;
3166 flags
|= TNF_MIGRATED
;
3168 flags
|= TNF_MIGRATE_FAIL
;
3172 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
3177 * These routines also need to handle stuff like marking pages dirty
3178 * and/or accessed for architectures that don't do it in hardware (most
3179 * RISC architectures). The early dirtying is also good on the i386.
3181 * There is also a hook called "update_mmu_cache()" that architectures
3182 * with external mmu caches can use to update those (ie the Sparc or
3183 * PowerPC hashed page tables that act as extended TLBs).
3185 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3186 * but allow concurrent faults), and pte mapped but not yet locked.
3187 * We return with pte unmapped and unlocked.
3189 * The mmap_sem may have been released depending on flags and our
3190 * return value. See filemap_fault() and __lock_page_or_retry().
3192 static int handle_pte_fault(struct mm_struct
*mm
,
3193 struct vm_area_struct
*vma
, unsigned long address
,
3194 pte_t
*pte
, pmd_t
*pmd
, unsigned int flags
)
3200 * some architectures can have larger ptes than wordsize,
3201 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3202 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3203 * The code below just needs a consistent view for the ifs and
3204 * we later double check anyway with the ptl lock held. So here
3205 * a barrier will do.
3209 if (!pte_present(entry
)) {
3210 if (pte_none(entry
)) {
3212 if (likely(vma
->vm_ops
->fault
))
3213 return do_fault(mm
, vma
, address
, pte
,
3216 return do_anonymous_page(mm
, vma
, address
,
3219 return do_swap_page(mm
, vma
, address
,
3220 pte
, pmd
, flags
, entry
);
3223 if (pte_protnone(entry
))
3224 return do_numa_page(mm
, vma
, address
, entry
, pte
, pmd
);
3226 ptl
= pte_lockptr(mm
, pmd
);
3228 if (unlikely(!pte_same(*pte
, entry
)))
3230 if (flags
& FAULT_FLAG_WRITE
) {
3231 if (!pte_write(entry
))
3232 return do_wp_page(mm
, vma
, address
,
3233 pte
, pmd
, ptl
, entry
);
3234 entry
= pte_mkdirty(entry
);
3236 entry
= pte_mkyoung(entry
);
3237 if (ptep_set_access_flags(vma
, address
, pte
, entry
, flags
& FAULT_FLAG_WRITE
)) {
3238 update_mmu_cache(vma
, address
, pte
);
3241 * This is needed only for protection faults but the arch code
3242 * is not yet telling us if this is a protection fault or not.
3243 * This still avoids useless tlb flushes for .text page faults
3246 if (flags
& FAULT_FLAG_WRITE
)
3247 flush_tlb_fix_spurious_fault(vma
, address
);
3250 pte_unmap_unlock(pte
, ptl
);
3255 * By the time we get here, we already hold the mm semaphore
3257 * The mmap_sem may have been released depending on flags and our
3258 * return value. See filemap_fault() and __lock_page_or_retry().
3260 static int __handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3261 unsigned long address
, unsigned int flags
)
3268 if (unlikely(is_vm_hugetlb_page(vma
)))
3269 return hugetlb_fault(mm
, vma
, address
, flags
);
3271 pgd
= pgd_offset(mm
, address
);
3272 pud
= pud_alloc(mm
, pgd
, address
);
3274 return VM_FAULT_OOM
;
3275 pmd
= pmd_alloc(mm
, pud
, address
);
3277 return VM_FAULT_OOM
;
3278 if (pmd_none(*pmd
) && transparent_hugepage_enabled(vma
)) {
3279 int ret
= VM_FAULT_FALLBACK
;
3281 ret
= do_huge_pmd_anonymous_page(mm
, vma
, address
,
3283 if (!(ret
& VM_FAULT_FALLBACK
))
3286 pmd_t orig_pmd
= *pmd
;
3290 if (pmd_trans_huge(orig_pmd
)) {
3291 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
3294 * If the pmd is splitting, return and retry the
3295 * the fault. Alternative: wait until the split
3296 * is done, and goto retry.
3298 if (pmd_trans_splitting(orig_pmd
))
3301 if (pmd_protnone(orig_pmd
))
3302 return do_huge_pmd_numa_page(mm
, vma
, address
,
3305 if (dirty
&& !pmd_write(orig_pmd
)) {
3306 ret
= do_huge_pmd_wp_page(mm
, vma
, address
, pmd
,
3308 if (!(ret
& VM_FAULT_FALLBACK
))
3311 huge_pmd_set_accessed(mm
, vma
, address
, pmd
,
3319 * Use __pte_alloc instead of pte_alloc_map, because we can't
3320 * run pte_offset_map on the pmd, if an huge pmd could
3321 * materialize from under us from a different thread.
3323 if (unlikely(pmd_none(*pmd
)) &&
3324 unlikely(__pte_alloc(mm
, vma
, pmd
, address
)))
3325 return VM_FAULT_OOM
;
3326 /* if an huge pmd materialized from under us just retry later */
3327 if (unlikely(pmd_trans_huge(*pmd
)))
3330 * A regular pmd is established and it can't morph into a huge pmd
3331 * from under us anymore at this point because we hold the mmap_sem
3332 * read mode and khugepaged takes it in write mode. So now it's
3333 * safe to run pte_offset_map().
3335 pte
= pte_offset_map(pmd
, address
);
3337 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
3341 * By the time we get here, we already hold the mm semaphore
3343 * The mmap_sem may have been released depending on flags and our
3344 * return value. See filemap_fault() and __lock_page_or_retry().
3346 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3347 unsigned long address
, unsigned int flags
)
3351 __set_current_state(TASK_RUNNING
);
3353 count_vm_event(PGFAULT
);
3354 mem_cgroup_count_vm_event(mm
, PGFAULT
);
3356 /* do counter updates before entering really critical section. */
3357 check_sync_rss_stat(current
);
3360 * Enable the memcg OOM handling for faults triggered in user
3361 * space. Kernel faults are handled more gracefully.
3363 if (flags
& FAULT_FLAG_USER
)
3364 mem_cgroup_oom_enable();
3366 ret
= __handle_mm_fault(mm
, vma
, address
, flags
);
3368 if (flags
& FAULT_FLAG_USER
) {
3369 mem_cgroup_oom_disable();
3371 * The task may have entered a memcg OOM situation but
3372 * if the allocation error was handled gracefully (no
3373 * VM_FAULT_OOM), there is no need to kill anything.
3374 * Just clean up the OOM state peacefully.
3376 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
3377 mem_cgroup_oom_synchronize(false);
3382 EXPORT_SYMBOL_GPL(handle_mm_fault
);
3384 #ifndef __PAGETABLE_PUD_FOLDED
3386 * Allocate page upper directory.
3387 * We've already handled the fast-path in-line.
3389 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
3391 pud_t
*new = pud_alloc_one(mm
, address
);
3395 smp_wmb(); /* See comment in __pte_alloc */
3397 spin_lock(&mm
->page_table_lock
);
3398 if (pgd_present(*pgd
)) /* Another has populated it */
3401 pgd_populate(mm
, pgd
, new);
3402 spin_unlock(&mm
->page_table_lock
);
3405 #endif /* __PAGETABLE_PUD_FOLDED */
3407 #ifndef __PAGETABLE_PMD_FOLDED
3409 * Allocate page middle directory.
3410 * We've already handled the fast-path in-line.
3412 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
3414 pmd_t
*new = pmd_alloc_one(mm
, address
);
3418 smp_wmb(); /* See comment in __pte_alloc */
3420 spin_lock(&mm
->page_table_lock
);
3421 #ifndef __ARCH_HAS_4LEVEL_HACK
3422 if (!pud_present(*pud
)) {
3424 pud_populate(mm
, pud
, new);
3425 } else /* Another has populated it */
3428 if (!pgd_present(*pud
)) {
3430 pgd_populate(mm
, pud
, new);
3431 } else /* Another has populated it */
3433 #endif /* __ARCH_HAS_4LEVEL_HACK */
3434 spin_unlock(&mm
->page_table_lock
);
3437 #endif /* __PAGETABLE_PMD_FOLDED */
3439 static int __follow_pte(struct mm_struct
*mm
, unsigned long address
,
3440 pte_t
**ptepp
, spinlock_t
**ptlp
)
3447 pgd
= pgd_offset(mm
, address
);
3448 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
3451 pud
= pud_offset(pgd
, address
);
3452 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
3455 pmd
= pmd_offset(pud
, address
);
3456 VM_BUG_ON(pmd_trans_huge(*pmd
));
3457 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
3460 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3464 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
3467 if (!pte_present(*ptep
))
3472 pte_unmap_unlock(ptep
, *ptlp
);
3477 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
3478 pte_t
**ptepp
, spinlock_t
**ptlp
)
3482 /* (void) is needed to make gcc happy */
3483 (void) __cond_lock(*ptlp
,
3484 !(res
= __follow_pte(mm
, address
, ptepp
, ptlp
)));
3489 * follow_pfn - look up PFN at a user virtual address
3490 * @vma: memory mapping
3491 * @address: user virtual address
3492 * @pfn: location to store found PFN
3494 * Only IO mappings and raw PFN mappings are allowed.
3496 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3498 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
3505 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3508 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
3511 *pfn
= pte_pfn(*ptep
);
3512 pte_unmap_unlock(ptep
, ptl
);
3515 EXPORT_SYMBOL(follow_pfn
);
3517 #ifdef CONFIG_HAVE_IOREMAP_PROT
3518 int follow_phys(struct vm_area_struct
*vma
,
3519 unsigned long address
, unsigned int flags
,
3520 unsigned long *prot
, resource_size_t
*phys
)
3526 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3529 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
3533 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
3536 *prot
= pgprot_val(pte_pgprot(pte
));
3537 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
3541 pte_unmap_unlock(ptep
, ptl
);
3546 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
3547 void *buf
, int len
, int write
)
3549 resource_size_t phys_addr
;
3550 unsigned long prot
= 0;
3551 void __iomem
*maddr
;
3552 int offset
= addr
& (PAGE_SIZE
-1);
3554 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
3557 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
3559 memcpy_toio(maddr
+ offset
, buf
, len
);
3561 memcpy_fromio(buf
, maddr
+ offset
, len
);
3566 EXPORT_SYMBOL_GPL(generic_access_phys
);
3570 * Access another process' address space as given in mm. If non-NULL, use the
3571 * given task for page fault accounting.
3573 static int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
3574 unsigned long addr
, void *buf
, int len
, int write
)
3576 struct vm_area_struct
*vma
;
3577 void *old_buf
= buf
;
3579 down_read(&mm
->mmap_sem
);
3580 /* ignore errors, just check how much was successfully transferred */
3582 int bytes
, ret
, offset
;
3584 struct page
*page
= NULL
;
3586 ret
= get_user_pages(tsk
, mm
, addr
, 1,
3587 write
, 1, &page
, &vma
);
3589 #ifndef CONFIG_HAVE_IOREMAP_PROT
3593 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3594 * we can access using slightly different code.
3596 vma
= find_vma(mm
, addr
);
3597 if (!vma
|| vma
->vm_start
> addr
)
3599 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
3600 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
3608 offset
= addr
& (PAGE_SIZE
-1);
3609 if (bytes
> PAGE_SIZE
-offset
)
3610 bytes
= PAGE_SIZE
-offset
;
3614 copy_to_user_page(vma
, page
, addr
,
3615 maddr
+ offset
, buf
, bytes
);
3616 set_page_dirty_lock(page
);
3618 copy_from_user_page(vma
, page
, addr
,
3619 buf
, maddr
+ offset
, bytes
);
3622 page_cache_release(page
);
3628 up_read(&mm
->mmap_sem
);
3630 return buf
- old_buf
;
3634 * access_remote_vm - access another process' address space
3635 * @mm: the mm_struct of the target address space
3636 * @addr: start address to access
3637 * @buf: source or destination buffer
3638 * @len: number of bytes to transfer
3639 * @write: whether the access is a write
3641 * The caller must hold a reference on @mm.
3643 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
3644 void *buf
, int len
, int write
)
3646 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, write
);
3650 * Access another process' address space.
3651 * Source/target buffer must be kernel space,
3652 * Do not walk the page table directly, use get_user_pages
3654 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
3655 void *buf
, int len
, int write
)
3657 struct mm_struct
*mm
;
3660 mm
= get_task_mm(tsk
);
3664 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, write
);
3671 * Print the name of a VMA.
3673 void print_vma_addr(char *prefix
, unsigned long ip
)
3675 struct mm_struct
*mm
= current
->mm
;
3676 struct vm_area_struct
*vma
;
3679 * Do not print if we are in atomic
3680 * contexts (in exception stacks, etc.):
3682 if (preempt_count())
3685 down_read(&mm
->mmap_sem
);
3686 vma
= find_vma(mm
, ip
);
3687 if (vma
&& vma
->vm_file
) {
3688 struct file
*f
= vma
->vm_file
;
3689 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
3693 p
= d_path(&f
->f_path
, buf
, PAGE_SIZE
);
3696 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
3698 vma
->vm_end
- vma
->vm_start
);
3699 free_page((unsigned long)buf
);
3702 up_read(&mm
->mmap_sem
);
3705 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3706 void might_fault(void)
3709 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3710 * holding the mmap_sem, this is safe because kernel memory doesn't
3711 * get paged out, therefore we'll never actually fault, and the
3712 * below annotations will generate false positives.
3714 if (segment_eq(get_fs(), KERNEL_DS
))
3718 * it would be nicer only to annotate paths which are not under
3719 * pagefault_disable, however that requires a larger audit and
3720 * providing helpers like get_user_atomic.
3725 __might_sleep(__FILE__
, __LINE__
, 0);
3728 might_lock_read(¤t
->mm
->mmap_sem
);
3730 EXPORT_SYMBOL(might_fault
);
3733 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3734 static void clear_gigantic_page(struct page
*page
,
3736 unsigned int pages_per_huge_page
)
3739 struct page
*p
= page
;
3742 for (i
= 0; i
< pages_per_huge_page
;
3743 i
++, p
= mem_map_next(p
, page
, i
)) {
3745 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
3748 void clear_huge_page(struct page
*page
,
3749 unsigned long addr
, unsigned int pages_per_huge_page
)
3753 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3754 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
3759 for (i
= 0; i
< pages_per_huge_page
; i
++) {
3761 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
3765 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
3767 struct vm_area_struct
*vma
,
3768 unsigned int pages_per_huge_page
)
3771 struct page
*dst_base
= dst
;
3772 struct page
*src_base
= src
;
3774 for (i
= 0; i
< pages_per_huge_page
; ) {
3776 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
3779 dst
= mem_map_next(dst
, dst_base
, i
);
3780 src
= mem_map_next(src
, src_base
, i
);
3784 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
3785 unsigned long addr
, struct vm_area_struct
*vma
,
3786 unsigned int pages_per_huge_page
)
3790 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3791 copy_user_gigantic_page(dst
, src
, addr
, vma
,
3792 pages_per_huge_page
);
3797 for (i
= 0; i
< pages_per_huge_page
; i
++) {
3799 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
3802 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3804 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3806 static struct kmem_cache
*page_ptl_cachep
;
3808 void __init
ptlock_cache_init(void)
3810 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
3814 bool ptlock_alloc(struct page
*page
)
3818 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
3825 void ptlock_free(struct page
*page
)
3827 kmem_cache_free(page_ptl_cachep
, page
->ptl
);