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;
226 tlb
->local
.next
= NULL
;
228 tlb
->local
.max
= ARRAY_SIZE(tlb
->__pages
);
229 tlb
->active
= &tlb
->local
;
230 tlb
->batch_count
= 0;
232 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
237 static void tlb_flush_mmu_tlbonly(struct mmu_gather
*tlb
)
241 mmu_notifier_invalidate_range(tlb
->mm
, tlb
->start
, tlb
->end
);
242 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
243 tlb_table_flush(tlb
);
247 static void tlb_flush_mmu_free(struct mmu_gather
*tlb
)
249 struct mmu_gather_batch
*batch
;
251 for (batch
= &tlb
->local
; batch
; batch
= batch
->next
) {
252 free_pages_and_swap_cache(batch
->pages
, batch
->nr
);
255 tlb
->active
= &tlb
->local
;
258 void tlb_flush_mmu(struct mmu_gather
*tlb
)
260 if (!tlb
->need_flush
)
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
->need_flush
);
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
;
366 * When there's less then two users of this mm there cannot be a
367 * concurrent page-table walk.
369 if (atomic_read(&tlb
->mm
->mm_users
) < 2) {
370 __tlb_remove_table(table
);
374 if (*batch
== NULL
) {
375 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
376 if (*batch
== NULL
) {
377 tlb_remove_table_one(table
);
382 (*batch
)->tables
[(*batch
)->nr
++] = table
;
383 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
384 tlb_table_flush(tlb
);
387 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
390 * Note: this doesn't free the actual pages themselves. That
391 * has been handled earlier when unmapping all the memory regions.
393 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
396 pgtable_t token
= pmd_pgtable(*pmd
);
398 pte_free_tlb(tlb
, token
, addr
);
399 atomic_long_dec(&tlb
->mm
->nr_ptes
);
402 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
403 unsigned long addr
, unsigned long end
,
404 unsigned long floor
, unsigned long ceiling
)
411 pmd
= pmd_offset(pud
, addr
);
413 next
= pmd_addr_end(addr
, end
);
414 if (pmd_none_or_clear_bad(pmd
))
416 free_pte_range(tlb
, pmd
, addr
);
417 } while (pmd
++, addr
= next
, addr
!= end
);
427 if (end
- 1 > ceiling
- 1)
430 pmd
= pmd_offset(pud
, start
);
432 pmd_free_tlb(tlb
, pmd
, start
);
435 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
436 unsigned long addr
, unsigned long end
,
437 unsigned long floor
, unsigned long ceiling
)
444 pud
= pud_offset(pgd
, addr
);
446 next
= pud_addr_end(addr
, end
);
447 if (pud_none_or_clear_bad(pud
))
449 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
450 } while (pud
++, addr
= next
, addr
!= end
);
456 ceiling
&= PGDIR_MASK
;
460 if (end
- 1 > ceiling
- 1)
463 pud
= pud_offset(pgd
, start
);
465 pud_free_tlb(tlb
, pud
, start
);
469 * This function frees user-level page tables of a process.
471 void free_pgd_range(struct mmu_gather
*tlb
,
472 unsigned long addr
, unsigned long end
,
473 unsigned long floor
, unsigned long ceiling
)
479 * The next few lines have given us lots of grief...
481 * Why are we testing PMD* at this top level? Because often
482 * there will be no work to do at all, and we'd prefer not to
483 * go all the way down to the bottom just to discover that.
485 * Why all these "- 1"s? Because 0 represents both the bottom
486 * of the address space and the top of it (using -1 for the
487 * top wouldn't help much: the masks would do the wrong thing).
488 * The rule is that addr 0 and floor 0 refer to the bottom of
489 * the address space, but end 0 and ceiling 0 refer to the top
490 * Comparisons need to use "end - 1" and "ceiling - 1" (though
491 * that end 0 case should be mythical).
493 * Wherever addr is brought up or ceiling brought down, we must
494 * be careful to reject "the opposite 0" before it confuses the
495 * subsequent tests. But what about where end is brought down
496 * by PMD_SIZE below? no, end can't go down to 0 there.
498 * Whereas we round start (addr) and ceiling down, by different
499 * masks at different levels, in order to test whether a table
500 * now has no other vmas using it, so can be freed, we don't
501 * bother to round floor or end up - the tests don't need that.
515 if (end
- 1 > ceiling
- 1)
520 pgd
= pgd_offset(tlb
->mm
, addr
);
522 next
= pgd_addr_end(addr
, end
);
523 if (pgd_none_or_clear_bad(pgd
))
525 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
526 } while (pgd
++, addr
= next
, addr
!= end
);
529 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
530 unsigned long floor
, unsigned long ceiling
)
533 struct vm_area_struct
*next
= vma
->vm_next
;
534 unsigned long addr
= vma
->vm_start
;
537 * Hide vma from rmap and truncate_pagecache before freeing
540 unlink_anon_vmas(vma
);
541 unlink_file_vma(vma
);
543 if (is_vm_hugetlb_page(vma
)) {
544 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
545 floor
, next
? next
->vm_start
: ceiling
);
548 * Optimization: gather nearby vmas into one call down
550 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
551 && !is_vm_hugetlb_page(next
)) {
554 unlink_anon_vmas(vma
);
555 unlink_file_vma(vma
);
557 free_pgd_range(tlb
, addr
, vma
->vm_end
,
558 floor
, next
? next
->vm_start
: ceiling
);
564 int __pte_alloc(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
565 pmd_t
*pmd
, unsigned long address
)
568 pgtable_t
new = pte_alloc_one(mm
, address
);
569 int wait_split_huge_page
;
574 * Ensure all pte setup (eg. pte page lock and page clearing) are
575 * visible before the pte is made visible to other CPUs by being
576 * put into page tables.
578 * The other side of the story is the pointer chasing in the page
579 * table walking code (when walking the page table without locking;
580 * ie. most of the time). Fortunately, these data accesses consist
581 * of a chain of data-dependent loads, meaning most CPUs (alpha
582 * being the notable exception) will already guarantee loads are
583 * seen in-order. See the alpha page table accessors for the
584 * smp_read_barrier_depends() barriers in page table walking code.
586 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
588 ptl
= pmd_lock(mm
, pmd
);
589 wait_split_huge_page
= 0;
590 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
591 atomic_long_inc(&mm
->nr_ptes
);
592 pmd_populate(mm
, pmd
, new);
594 } else if (unlikely(pmd_trans_splitting(*pmd
)))
595 wait_split_huge_page
= 1;
599 if (wait_split_huge_page
)
600 wait_split_huge_page(vma
->anon_vma
, pmd
);
604 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
606 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
610 smp_wmb(); /* See comment in __pte_alloc */
612 spin_lock(&init_mm
.page_table_lock
);
613 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
614 pmd_populate_kernel(&init_mm
, pmd
, new);
617 VM_BUG_ON(pmd_trans_splitting(*pmd
));
618 spin_unlock(&init_mm
.page_table_lock
);
620 pte_free_kernel(&init_mm
, new);
624 static inline void init_rss_vec(int *rss
)
626 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
629 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
633 if (current
->mm
== mm
)
635 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
637 add_mm_counter(mm
, i
, rss
[i
]);
641 * This function is called to print an error when a bad pte
642 * is found. For example, we might have a PFN-mapped pte in
643 * a region that doesn't allow it.
645 * The calling function must still handle the error.
647 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
648 pte_t pte
, struct page
*page
)
650 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
651 pud_t
*pud
= pud_offset(pgd
, addr
);
652 pmd_t
*pmd
= pmd_offset(pud
, addr
);
653 struct address_space
*mapping
;
655 static unsigned long resume
;
656 static unsigned long nr_shown
;
657 static unsigned long nr_unshown
;
660 * Allow a burst of 60 reports, then keep quiet for that minute;
661 * or allow a steady drip of one report per second.
663 if (nr_shown
== 60) {
664 if (time_before(jiffies
, resume
)) {
670 "BUG: Bad page map: %lu messages suppressed\n",
677 resume
= jiffies
+ 60 * HZ
;
679 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
680 index
= linear_page_index(vma
, addr
);
683 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
685 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
687 dump_page(page
, "bad pte");
689 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
690 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
692 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
695 printk(KERN_ALERT
"vma->vm_ops->fault: %pSR\n",
698 printk(KERN_ALERT
"vma->vm_file->f_op->mmap: %pSR\n",
699 vma
->vm_file
->f_op
->mmap
);
701 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
705 * vm_normal_page -- This function gets the "struct page" associated with a pte.
707 * "Special" mappings do not wish to be associated with a "struct page" (either
708 * it doesn't exist, or it exists but they don't want to touch it). In this
709 * case, NULL is returned here. "Normal" mappings do have a struct page.
711 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
712 * pte bit, in which case this function is trivial. Secondly, an architecture
713 * may not have a spare pte bit, which requires a more complicated scheme,
716 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
717 * special mapping (even if there are underlying and valid "struct pages").
718 * COWed pages of a VM_PFNMAP are always normal.
720 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
721 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
722 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
723 * mapping will always honor the rule
725 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
727 * And for normal mappings this is false.
729 * This restricts such mappings to be a linear translation from virtual address
730 * to pfn. To get around this restriction, we allow arbitrary mappings so long
731 * as the vma is not a COW mapping; in that case, we know that all ptes are
732 * special (because none can have been COWed).
735 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
737 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
738 * page" backing, however the difference is that _all_ pages with a struct
739 * page (that is, those where pfn_valid is true) are refcounted and considered
740 * normal pages by the VM. The disadvantage is that pages are refcounted
741 * (which can be slower and simply not an option for some PFNMAP users). The
742 * advantage is that we don't have to follow the strict linearity rule of
743 * PFNMAP mappings in order to support COWable mappings.
746 #ifdef __HAVE_ARCH_PTE_SPECIAL
747 # define HAVE_PTE_SPECIAL 1
749 # define HAVE_PTE_SPECIAL 0
751 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
754 unsigned long pfn
= pte_pfn(pte
);
756 if (HAVE_PTE_SPECIAL
) {
757 if (likely(!pte_special(pte
)))
759 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
761 if (!is_zero_pfn(pfn
))
762 print_bad_pte(vma
, addr
, pte
, NULL
);
766 /* !HAVE_PTE_SPECIAL case follows: */
768 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
769 if (vma
->vm_flags
& VM_MIXEDMAP
) {
775 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
776 if (pfn
== vma
->vm_pgoff
+ off
)
778 if (!is_cow_mapping(vma
->vm_flags
))
783 if (is_zero_pfn(pfn
))
786 if (unlikely(pfn
> highest_memmap_pfn
)) {
787 print_bad_pte(vma
, addr
, pte
, NULL
);
792 * NOTE! We still have PageReserved() pages in the page tables.
793 * eg. VDSO mappings can cause them to exist.
796 return pfn_to_page(pfn
);
800 * copy one vm_area from one task to the other. Assumes the page tables
801 * already present in the new task to be cleared in the whole range
802 * covered by this vma.
805 static inline unsigned long
806 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
807 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
808 unsigned long addr
, int *rss
)
810 unsigned long vm_flags
= vma
->vm_flags
;
811 pte_t pte
= *src_pte
;
814 /* pte contains position in swap or file, so copy. */
815 if (unlikely(!pte_present(pte
))) {
816 if (!pte_file(pte
)) {
817 swp_entry_t entry
= pte_to_swp_entry(pte
);
819 if (swap_duplicate(entry
) < 0)
822 /* make sure dst_mm is on swapoff's mmlist. */
823 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
824 spin_lock(&mmlist_lock
);
825 if (list_empty(&dst_mm
->mmlist
))
826 list_add(&dst_mm
->mmlist
,
828 spin_unlock(&mmlist_lock
);
830 if (likely(!non_swap_entry(entry
)))
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
);
858 * If it's a COW mapping, write protect it both
859 * in the parent and the child
861 if (is_cow_mapping(vm_flags
)) {
862 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
863 pte
= pte_wrprotect(pte
);
867 * If it's a shared mapping, mark it clean in
870 if (vm_flags
& VM_SHARED
)
871 pte
= pte_mkclean(pte
);
872 pte
= pte_mkold(pte
);
874 page
= vm_normal_page(vma
, addr
, pte
);
885 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
889 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
890 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
891 unsigned long addr
, unsigned long end
)
893 pte_t
*orig_src_pte
, *orig_dst_pte
;
894 pte_t
*src_pte
, *dst_pte
;
895 spinlock_t
*src_ptl
, *dst_ptl
;
897 int rss
[NR_MM_COUNTERS
];
898 swp_entry_t entry
= (swp_entry_t
){0};
903 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
906 src_pte
= pte_offset_map(src_pmd
, addr
);
907 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
908 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
909 orig_src_pte
= src_pte
;
910 orig_dst_pte
= dst_pte
;
911 arch_enter_lazy_mmu_mode();
915 * We are holding two locks at this point - either of them
916 * could generate latencies in another task on another CPU.
918 if (progress
>= 32) {
920 if (need_resched() ||
921 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
924 if (pte_none(*src_pte
)) {
928 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
933 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
935 arch_leave_lazy_mmu_mode();
936 spin_unlock(src_ptl
);
937 pte_unmap(orig_src_pte
);
938 add_mm_rss_vec(dst_mm
, rss
);
939 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
943 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
952 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
953 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
954 unsigned long addr
, unsigned long end
)
956 pmd_t
*src_pmd
, *dst_pmd
;
959 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
962 src_pmd
= pmd_offset(src_pud
, addr
);
964 next
= pmd_addr_end(addr
, end
);
965 if (pmd_trans_huge(*src_pmd
)) {
967 VM_BUG_ON(next
-addr
!= HPAGE_PMD_SIZE
);
968 err
= copy_huge_pmd(dst_mm
, src_mm
,
969 dst_pmd
, src_pmd
, addr
, vma
);
976 if (pmd_none_or_clear_bad(src_pmd
))
978 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
981 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
985 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
986 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
987 unsigned long addr
, unsigned long end
)
989 pud_t
*src_pud
, *dst_pud
;
992 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
995 src_pud
= pud_offset(src_pgd
, addr
);
997 next
= pud_addr_end(addr
, end
);
998 if (pud_none_or_clear_bad(src_pud
))
1000 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
1003 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1007 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1008 struct vm_area_struct
*vma
)
1010 pgd_t
*src_pgd
, *dst_pgd
;
1012 unsigned long addr
= vma
->vm_start
;
1013 unsigned long end
= vma
->vm_end
;
1014 unsigned long mmun_start
; /* For mmu_notifiers */
1015 unsigned long mmun_end
; /* For mmu_notifiers */
1020 * Don't copy ptes where a page fault will fill them correctly.
1021 * Fork becomes much lighter when there are big shared or private
1022 * readonly mappings. The tradeoff is that copy_page_range is more
1023 * efficient than faulting.
1025 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_NONLINEAR
|
1026 VM_PFNMAP
| VM_MIXEDMAP
))) {
1031 if (is_vm_hugetlb_page(vma
))
1032 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1034 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
1036 * We do not free on error cases below as remove_vma
1037 * gets called on error from higher level routine
1039 ret
= track_pfn_copy(vma
);
1045 * We need to invalidate the secondary MMU mappings only when
1046 * there could be a permission downgrade on the ptes of the
1047 * parent mm. And a permission downgrade will only happen if
1048 * is_cow_mapping() returns true.
1050 is_cow
= is_cow_mapping(vma
->vm_flags
);
1054 mmu_notifier_invalidate_range_start(src_mm
, mmun_start
,
1058 dst_pgd
= pgd_offset(dst_mm
, addr
);
1059 src_pgd
= pgd_offset(src_mm
, addr
);
1061 next
= pgd_addr_end(addr
, end
);
1062 if (pgd_none_or_clear_bad(src_pgd
))
1064 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1065 vma
, addr
, next
))) {
1069 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1072 mmu_notifier_invalidate_range_end(src_mm
, mmun_start
, mmun_end
);
1076 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1077 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1078 unsigned long addr
, unsigned long end
,
1079 struct zap_details
*details
)
1081 struct mm_struct
*mm
= tlb
->mm
;
1082 int force_flush
= 0;
1083 int rss
[NR_MM_COUNTERS
];
1090 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1092 arch_enter_lazy_mmu_mode();
1095 if (pte_none(ptent
)) {
1099 if (pte_present(ptent
)) {
1102 page
= vm_normal_page(vma
, addr
, ptent
);
1103 if (unlikely(details
) && page
) {
1105 * unmap_shared_mapping_pages() wants to
1106 * invalidate cache without truncating:
1107 * unmap shared but keep private pages.
1109 if (details
->check_mapping
&&
1110 details
->check_mapping
!= page
->mapping
)
1113 * Each page->index must be checked when
1114 * invalidating or truncating nonlinear.
1116 if (details
->nonlinear_vma
&&
1117 (page
->index
< details
->first_index
||
1118 page
->index
> details
->last_index
))
1121 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1123 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1124 if (unlikely(!page
))
1126 if (unlikely(details
) && details
->nonlinear_vma
1127 && linear_page_index(details
->nonlinear_vma
,
1128 addr
) != page
->index
) {
1129 pte_t ptfile
= pgoff_to_pte(page
->index
);
1130 if (pte_soft_dirty(ptent
))
1131 ptfile
= pte_file_mksoft_dirty(ptfile
);
1132 set_pte_at(mm
, addr
, pte
, ptfile
);
1135 rss
[MM_ANONPAGES
]--;
1137 if (pte_dirty(ptent
)) {
1139 set_page_dirty(page
);
1141 if (pte_young(ptent
) &&
1142 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1143 mark_page_accessed(page
);
1144 rss
[MM_FILEPAGES
]--;
1146 page_remove_rmap(page
);
1147 if (unlikely(page_mapcount(page
) < 0))
1148 print_bad_pte(vma
, addr
, ptent
, page
);
1149 if (unlikely(!__tlb_remove_page(tlb
, page
))) {
1157 * If details->check_mapping, we leave swap entries;
1158 * if details->nonlinear_vma, we leave file entries.
1160 if (unlikely(details
))
1162 if (pte_file(ptent
)) {
1163 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
)))
1164 print_bad_pte(vma
, addr
, ptent
, NULL
);
1166 swp_entry_t entry
= pte_to_swp_entry(ptent
);
1168 if (!non_swap_entry(entry
))
1170 else if (is_migration_entry(entry
)) {
1173 page
= migration_entry_to_page(entry
);
1176 rss
[MM_ANONPAGES
]--;
1178 rss
[MM_FILEPAGES
]--;
1180 if (unlikely(!free_swap_and_cache(entry
)))
1181 print_bad_pte(vma
, addr
, ptent
, NULL
);
1183 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1184 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1186 add_mm_rss_vec(mm
, rss
);
1187 arch_leave_lazy_mmu_mode();
1189 /* Do the actual TLB flush before dropping ptl */
1191 unsigned long old_end
;
1194 * Flush the TLB just for the previous segment,
1195 * then update the range to be the remaining
1200 tlb_flush_mmu_tlbonly(tlb
);
1204 pte_unmap_unlock(start_pte
, ptl
);
1207 * If we forced a TLB flush (either due to running out of
1208 * batch buffers or because we needed to flush dirty TLB
1209 * entries before releasing the ptl), free the batched
1210 * memory too. Restart if we didn't do everything.
1214 tlb_flush_mmu_free(tlb
);
1223 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1224 struct vm_area_struct
*vma
, pud_t
*pud
,
1225 unsigned long addr
, unsigned long end
,
1226 struct zap_details
*details
)
1231 pmd
= pmd_offset(pud
, addr
);
1233 next
= pmd_addr_end(addr
, end
);
1234 if (pmd_trans_huge(*pmd
)) {
1235 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1236 #ifdef CONFIG_DEBUG_VM
1237 if (!rwsem_is_locked(&tlb
->mm
->mmap_sem
)) {
1238 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1239 __func__
, addr
, end
,
1245 split_huge_page_pmd(vma
, addr
, pmd
);
1246 } else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1251 * Here there can be other concurrent MADV_DONTNEED or
1252 * trans huge page faults running, and if the pmd is
1253 * none or trans huge it can change under us. This is
1254 * because MADV_DONTNEED holds the mmap_sem in read
1257 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1259 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1262 } while (pmd
++, addr
= next
, addr
!= end
);
1267 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1268 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1269 unsigned long addr
, unsigned long end
,
1270 struct zap_details
*details
)
1275 pud
= pud_offset(pgd
, addr
);
1277 next
= pud_addr_end(addr
, end
);
1278 if (pud_none_or_clear_bad(pud
))
1280 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1281 } while (pud
++, addr
= next
, addr
!= end
);
1286 static void unmap_page_range(struct mmu_gather
*tlb
,
1287 struct vm_area_struct
*vma
,
1288 unsigned long addr
, unsigned long end
,
1289 struct zap_details
*details
)
1294 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
1297 BUG_ON(addr
>= end
);
1298 tlb_start_vma(tlb
, vma
);
1299 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1301 next
= pgd_addr_end(addr
, end
);
1302 if (pgd_none_or_clear_bad(pgd
))
1304 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
1305 } while (pgd
++, addr
= next
, addr
!= end
);
1306 tlb_end_vma(tlb
, vma
);
1310 static void unmap_single_vma(struct mmu_gather
*tlb
,
1311 struct vm_area_struct
*vma
, unsigned long start_addr
,
1312 unsigned long end_addr
,
1313 struct zap_details
*details
)
1315 unsigned long start
= max(vma
->vm_start
, start_addr
);
1318 if (start
>= vma
->vm_end
)
1320 end
= min(vma
->vm_end
, end_addr
);
1321 if (end
<= vma
->vm_start
)
1325 uprobe_munmap(vma
, start
, end
);
1327 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1328 untrack_pfn(vma
, 0, 0);
1331 if (unlikely(is_vm_hugetlb_page(vma
))) {
1333 * It is undesirable to test vma->vm_file as it
1334 * should be non-null for valid hugetlb area.
1335 * However, vm_file will be NULL in the error
1336 * cleanup path of mmap_region. When
1337 * hugetlbfs ->mmap method fails,
1338 * mmap_region() nullifies vma->vm_file
1339 * before calling this function to clean up.
1340 * Since no pte has actually been setup, it is
1341 * safe to do nothing in this case.
1344 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
1345 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1346 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
1349 unmap_page_range(tlb
, vma
, start
, end
, details
);
1354 * unmap_vmas - unmap a range of memory covered by a list of vma's
1355 * @tlb: address of the caller's struct mmu_gather
1356 * @vma: the starting vma
1357 * @start_addr: virtual address at which to start unmapping
1358 * @end_addr: virtual address at which to end unmapping
1360 * Unmap all pages in the vma list.
1362 * Only addresses between `start' and `end' will be unmapped.
1364 * The VMA list must be sorted in ascending virtual address order.
1366 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1367 * range after unmap_vmas() returns. So the only responsibility here is to
1368 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1369 * drops the lock and schedules.
1371 void unmap_vmas(struct mmu_gather
*tlb
,
1372 struct vm_area_struct
*vma
, unsigned long start_addr
,
1373 unsigned long end_addr
)
1375 struct mm_struct
*mm
= vma
->vm_mm
;
1377 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1378 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1379 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1380 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1384 * zap_page_range - remove user pages in a given range
1385 * @vma: vm_area_struct holding the applicable pages
1386 * @start: starting address of pages to zap
1387 * @size: number of bytes to zap
1388 * @details: details of nonlinear truncation or shared cache invalidation
1390 * Caller must protect the VMA list
1392 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1393 unsigned long size
, struct zap_details
*details
)
1395 struct mm_struct
*mm
= vma
->vm_mm
;
1396 struct mmu_gather tlb
;
1397 unsigned long end
= start
+ size
;
1400 tlb_gather_mmu(&tlb
, mm
, start
, end
);
1401 update_hiwater_rss(mm
);
1402 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1403 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
)
1404 unmap_single_vma(&tlb
, vma
, start
, end
, details
);
1405 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1406 tlb_finish_mmu(&tlb
, start
, end
);
1410 * zap_page_range_single - remove user pages in a given range
1411 * @vma: vm_area_struct holding the applicable pages
1412 * @address: starting address of pages to zap
1413 * @size: number of bytes to zap
1414 * @details: details of nonlinear truncation or shared cache invalidation
1416 * The range must fit into one VMA.
1418 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1419 unsigned long size
, struct zap_details
*details
)
1421 struct mm_struct
*mm
= vma
->vm_mm
;
1422 struct mmu_gather tlb
;
1423 unsigned long end
= address
+ size
;
1426 tlb_gather_mmu(&tlb
, mm
, address
, end
);
1427 update_hiwater_rss(mm
);
1428 mmu_notifier_invalidate_range_start(mm
, address
, end
);
1429 unmap_single_vma(&tlb
, vma
, address
, end
, details
);
1430 mmu_notifier_invalidate_range_end(mm
, address
, end
);
1431 tlb_finish_mmu(&tlb
, address
, end
);
1435 * zap_vma_ptes - remove ptes mapping the vma
1436 * @vma: vm_area_struct holding ptes to be zapped
1437 * @address: starting address of pages to zap
1438 * @size: number of bytes to zap
1440 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1442 * The entire address range must be fully contained within the vma.
1444 * Returns 0 if successful.
1446 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1449 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1450 !(vma
->vm_flags
& VM_PFNMAP
))
1452 zap_page_range_single(vma
, address
, size
, NULL
);
1455 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1457 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1460 pgd_t
* pgd
= pgd_offset(mm
, addr
);
1461 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
1463 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
1465 VM_BUG_ON(pmd_trans_huge(*pmd
));
1466 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1473 * This is the old fallback for page remapping.
1475 * For historical reasons, it only allows reserved pages. Only
1476 * old drivers should use this, and they needed to mark their
1477 * pages reserved for the old functions anyway.
1479 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1480 struct page
*page
, pgprot_t prot
)
1482 struct mm_struct
*mm
= vma
->vm_mm
;
1491 flush_dcache_page(page
);
1492 pte
= get_locked_pte(mm
, addr
, &ptl
);
1496 if (!pte_none(*pte
))
1499 /* Ok, finally just insert the thing.. */
1501 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
1502 page_add_file_rmap(page
);
1503 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1506 pte_unmap_unlock(pte
, ptl
);
1509 pte_unmap_unlock(pte
, ptl
);
1515 * vm_insert_page - insert single page into user vma
1516 * @vma: user vma to map to
1517 * @addr: target user address of this page
1518 * @page: source kernel page
1520 * This allows drivers to insert individual pages they've allocated
1523 * The page has to be a nice clean _individual_ kernel allocation.
1524 * If you allocate a compound page, you need to have marked it as
1525 * such (__GFP_COMP), or manually just split the page up yourself
1526 * (see split_page()).
1528 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1529 * took an arbitrary page protection parameter. This doesn't allow
1530 * that. Your vma protection will have to be set up correctly, which
1531 * means that if you want a shared writable mapping, you'd better
1532 * ask for a shared writable mapping!
1534 * The page does not need to be reserved.
1536 * Usually this function is called from f_op->mmap() handler
1537 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1538 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1539 * function from other places, for example from page-fault handler.
1541 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1544 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1546 if (!page_count(page
))
1548 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1549 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
1550 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1551 vma
->vm_flags
|= VM_MIXEDMAP
;
1553 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1555 EXPORT_SYMBOL(vm_insert_page
);
1557 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1558 unsigned long pfn
, pgprot_t prot
)
1560 struct mm_struct
*mm
= vma
->vm_mm
;
1566 pte
= get_locked_pte(mm
, addr
, &ptl
);
1570 if (!pte_none(*pte
))
1573 /* Ok, finally just insert the thing.. */
1574 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
1575 set_pte_at(mm
, addr
, pte
, entry
);
1576 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1580 pte_unmap_unlock(pte
, ptl
);
1586 * vm_insert_pfn - insert single pfn into user vma
1587 * @vma: user vma to map to
1588 * @addr: target user address of this page
1589 * @pfn: source kernel pfn
1591 * Similar to vm_insert_page, this allows drivers to insert individual pages
1592 * they've allocated into a user vma. Same comments apply.
1594 * This function should only be called from a vm_ops->fault handler, and
1595 * in that case the handler should return NULL.
1597 * vma cannot be a COW mapping.
1599 * As this is called only for pages that do not currently exist, we
1600 * do not need to flush old virtual caches or the TLB.
1602 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1606 pgprot_t pgprot
= vma
->vm_page_prot
;
1608 * Technically, architectures with pte_special can avoid all these
1609 * restrictions (same for remap_pfn_range). However we would like
1610 * consistency in testing and feature parity among all, so we should
1611 * try to keep these invariants in place for everybody.
1613 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1614 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1615 (VM_PFNMAP
|VM_MIXEDMAP
));
1616 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1617 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1619 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1621 if (track_pfn_insert(vma
, &pgprot
, pfn
))
1624 ret
= insert_pfn(vma
, addr
, pfn
, pgprot
);
1628 EXPORT_SYMBOL(vm_insert_pfn
);
1630 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1633 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
1635 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1639 * If we don't have pte special, then we have to use the pfn_valid()
1640 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1641 * refcount the page if pfn_valid is true (hence insert_page rather
1642 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1643 * without pte special, it would there be refcounted as a normal page.
1645 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
1648 page
= pfn_to_page(pfn
);
1649 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1651 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1653 EXPORT_SYMBOL(vm_insert_mixed
);
1656 * maps a range of physical memory into the requested pages. the old
1657 * mappings are removed. any references to nonexistent pages results
1658 * in null mappings (currently treated as "copy-on-access")
1660 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1661 unsigned long addr
, unsigned long end
,
1662 unsigned long pfn
, pgprot_t prot
)
1667 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1670 arch_enter_lazy_mmu_mode();
1672 BUG_ON(!pte_none(*pte
));
1673 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
1675 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1676 arch_leave_lazy_mmu_mode();
1677 pte_unmap_unlock(pte
- 1, ptl
);
1681 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1682 unsigned long addr
, unsigned long end
,
1683 unsigned long pfn
, pgprot_t prot
)
1688 pfn
-= addr
>> PAGE_SHIFT
;
1689 pmd
= pmd_alloc(mm
, pud
, addr
);
1692 VM_BUG_ON(pmd_trans_huge(*pmd
));
1694 next
= pmd_addr_end(addr
, end
);
1695 if (remap_pte_range(mm
, pmd
, addr
, next
,
1696 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1698 } while (pmd
++, addr
= next
, addr
!= end
);
1702 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1703 unsigned long addr
, unsigned long end
,
1704 unsigned long pfn
, pgprot_t prot
)
1709 pfn
-= addr
>> PAGE_SHIFT
;
1710 pud
= pud_alloc(mm
, pgd
, addr
);
1714 next
= pud_addr_end(addr
, end
);
1715 if (remap_pmd_range(mm
, pud
, addr
, next
,
1716 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1718 } while (pud
++, addr
= next
, addr
!= end
);
1723 * remap_pfn_range - remap kernel memory to userspace
1724 * @vma: user vma to map to
1725 * @addr: target user address to start at
1726 * @pfn: physical address of kernel memory
1727 * @size: size of map area
1728 * @prot: page protection flags for this mapping
1730 * Note: this is only safe if the mm semaphore is held when called.
1732 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1733 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1737 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1738 struct mm_struct
*mm
= vma
->vm_mm
;
1742 * Physically remapped pages are special. Tell the
1743 * rest of the world about it:
1744 * VM_IO tells people not to look at these pages
1745 * (accesses can have side effects).
1746 * VM_PFNMAP tells the core MM that the base pages are just
1747 * raw PFN mappings, and do not have a "struct page" associated
1750 * Disable vma merging and expanding with mremap().
1752 * Omit vma from core dump, even when VM_IO turned off.
1754 * There's a horrible special case to handle copy-on-write
1755 * behaviour that some programs depend on. We mark the "original"
1756 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1757 * See vm_normal_page() for details.
1759 if (is_cow_mapping(vma
->vm_flags
)) {
1760 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
1762 vma
->vm_pgoff
= pfn
;
1765 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
1769 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
1771 BUG_ON(addr
>= end
);
1772 pfn
-= addr
>> PAGE_SHIFT
;
1773 pgd
= pgd_offset(mm
, addr
);
1774 flush_cache_range(vma
, addr
, end
);
1776 next
= pgd_addr_end(addr
, end
);
1777 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1778 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1781 } while (pgd
++, addr
= next
, addr
!= end
);
1784 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
));
1788 EXPORT_SYMBOL(remap_pfn_range
);
1791 * vm_iomap_memory - remap memory to userspace
1792 * @vma: user vma to map to
1793 * @start: start of area
1794 * @len: size of area
1796 * This is a simplified io_remap_pfn_range() for common driver use. The
1797 * driver just needs to give us the physical memory range to be mapped,
1798 * we'll figure out the rest from the vma information.
1800 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1801 * whatever write-combining details or similar.
1803 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
1805 unsigned long vm_len
, pfn
, pages
;
1807 /* Check that the physical memory area passed in looks valid */
1808 if (start
+ len
< start
)
1811 * You *really* shouldn't map things that aren't page-aligned,
1812 * but we've historically allowed it because IO memory might
1813 * just have smaller alignment.
1815 len
+= start
& ~PAGE_MASK
;
1816 pfn
= start
>> PAGE_SHIFT
;
1817 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
1818 if (pfn
+ pages
< pfn
)
1821 /* We start the mapping 'vm_pgoff' pages into the area */
1822 if (vma
->vm_pgoff
> pages
)
1824 pfn
+= vma
->vm_pgoff
;
1825 pages
-= vma
->vm_pgoff
;
1827 /* Can we fit all of the mapping? */
1828 vm_len
= vma
->vm_end
- vma
->vm_start
;
1829 if (vm_len
>> PAGE_SHIFT
> pages
)
1832 /* Ok, let it rip */
1833 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
1835 EXPORT_SYMBOL(vm_iomap_memory
);
1837 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1838 unsigned long addr
, unsigned long end
,
1839 pte_fn_t fn
, void *data
)
1844 spinlock_t
*uninitialized_var(ptl
);
1846 pte
= (mm
== &init_mm
) ?
1847 pte_alloc_kernel(pmd
, addr
) :
1848 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1852 BUG_ON(pmd_huge(*pmd
));
1854 arch_enter_lazy_mmu_mode();
1856 token
= pmd_pgtable(*pmd
);
1859 err
= fn(pte
++, token
, addr
, data
);
1862 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1864 arch_leave_lazy_mmu_mode();
1867 pte_unmap_unlock(pte
-1, ptl
);
1871 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1872 unsigned long addr
, unsigned long end
,
1873 pte_fn_t fn
, void *data
)
1879 BUG_ON(pud_huge(*pud
));
1881 pmd
= pmd_alloc(mm
, pud
, addr
);
1885 next
= pmd_addr_end(addr
, end
);
1886 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
1889 } while (pmd
++, addr
= next
, addr
!= end
);
1893 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1894 unsigned long addr
, unsigned long end
,
1895 pte_fn_t fn
, void *data
)
1901 pud
= pud_alloc(mm
, pgd
, addr
);
1905 next
= pud_addr_end(addr
, end
);
1906 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
1909 } while (pud
++, addr
= next
, addr
!= end
);
1914 * Scan a region of virtual memory, filling in page tables as necessary
1915 * and calling a provided function on each leaf page table.
1917 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
1918 unsigned long size
, pte_fn_t fn
, void *data
)
1922 unsigned long end
= addr
+ size
;
1925 BUG_ON(addr
>= end
);
1926 pgd
= pgd_offset(mm
, addr
);
1928 next
= pgd_addr_end(addr
, end
);
1929 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
1932 } while (pgd
++, addr
= next
, addr
!= end
);
1936 EXPORT_SYMBOL_GPL(apply_to_page_range
);
1939 * handle_pte_fault chooses page fault handler according to an entry
1940 * which was read non-atomically. Before making any commitment, on
1941 * those architectures or configurations (e.g. i386 with PAE) which
1942 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
1943 * must check under lock before unmapping the pte and proceeding
1944 * (but do_wp_page is only called after already making such a check;
1945 * and do_anonymous_page can safely check later on).
1947 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
1948 pte_t
*page_table
, pte_t orig_pte
)
1951 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1952 if (sizeof(pte_t
) > sizeof(unsigned long)) {
1953 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
1955 same
= pte_same(*page_table
, orig_pte
);
1959 pte_unmap(page_table
);
1963 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
1965 debug_dma_assert_idle(src
);
1968 * If the source page was a PFN mapping, we don't have
1969 * a "struct page" for it. We do a best-effort copy by
1970 * just copying from the original user address. If that
1971 * fails, we just zero-fill it. Live with it.
1973 if (unlikely(!src
)) {
1974 void *kaddr
= kmap_atomic(dst
);
1975 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
1978 * This really shouldn't fail, because the page is there
1979 * in the page tables. But it might just be unreadable,
1980 * in which case we just give up and fill the result with
1983 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
1985 kunmap_atomic(kaddr
);
1986 flush_dcache_page(dst
);
1988 copy_user_highpage(dst
, src
, va
, vma
);
1992 * Notify the address space that the page is about to become writable so that
1993 * it can prohibit this or wait for the page to get into an appropriate state.
1995 * We do this without the lock held, so that it can sleep if it needs to.
1997 static int do_page_mkwrite(struct vm_area_struct
*vma
, struct page
*page
,
1998 unsigned long address
)
2000 struct vm_fault vmf
;
2003 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
2004 vmf
.pgoff
= page
->index
;
2005 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2008 ret
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
2009 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2011 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
2013 if (!page
->mapping
) {
2015 return 0; /* retry */
2017 ret
|= VM_FAULT_LOCKED
;
2019 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2024 * This routine handles present pages, when users try to write
2025 * to a shared page. It is done by copying the page to a new address
2026 * and decrementing the shared-page counter for the old page.
2028 * Note that this routine assumes that the protection checks have been
2029 * done by the caller (the low-level page fault routine in most cases).
2030 * Thus we can safely just mark it writable once we've done any necessary
2033 * We also mark the page dirty at this point even though the page will
2034 * change only once the write actually happens. This avoids a few races,
2035 * and potentially makes it more efficient.
2037 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2038 * but allow concurrent faults), with pte both mapped and locked.
2039 * We return with mmap_sem still held, but pte unmapped and unlocked.
2041 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2042 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2043 spinlock_t
*ptl
, pte_t orig_pte
)
2046 struct page
*old_page
, *new_page
= NULL
;
2049 int page_mkwrite
= 0;
2050 struct page
*dirty_page
= NULL
;
2051 unsigned long mmun_start
= 0; /* For mmu_notifiers */
2052 unsigned long mmun_end
= 0; /* For mmu_notifiers */
2053 struct mem_cgroup
*memcg
;
2055 old_page
= vm_normal_page(vma
, address
, orig_pte
);
2058 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2061 * We should not cow pages in a shared writeable mapping.
2062 * Just mark the pages writable as we can't do any dirty
2063 * accounting on raw pfn maps.
2065 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2066 (VM_WRITE
|VM_SHARED
))
2072 * Take out anonymous pages first, anonymous shared vmas are
2073 * not dirty accountable.
2075 if (PageAnon(old_page
) && !PageKsm(old_page
)) {
2076 if (!trylock_page(old_page
)) {
2077 page_cache_get(old_page
);
2078 pte_unmap_unlock(page_table
, ptl
);
2079 lock_page(old_page
);
2080 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2082 if (!pte_same(*page_table
, orig_pte
)) {
2083 unlock_page(old_page
);
2086 page_cache_release(old_page
);
2088 if (reuse_swap_page(old_page
)) {
2090 * The page is all ours. Move it to our anon_vma so
2091 * the rmap code will not search our parent or siblings.
2092 * Protected against the rmap code by the page lock.
2094 page_move_anon_rmap(old_page
, vma
, address
);
2095 unlock_page(old_page
);
2098 unlock_page(old_page
);
2099 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2100 (VM_WRITE
|VM_SHARED
))) {
2102 * Only catch write-faults on shared writable pages,
2103 * read-only shared pages can get COWed by
2104 * get_user_pages(.write=1, .force=1).
2106 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2108 page_cache_get(old_page
);
2109 pte_unmap_unlock(page_table
, ptl
);
2110 tmp
= do_page_mkwrite(vma
, old_page
, address
);
2111 if (unlikely(!tmp
|| (tmp
&
2112 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
2113 page_cache_release(old_page
);
2117 * Since we dropped the lock we need to revalidate
2118 * the PTE as someone else may have changed it. If
2119 * they did, we just return, as we can count on the
2120 * MMU to tell us if they didn't also make it writable.
2122 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2124 if (!pte_same(*page_table
, orig_pte
)) {
2125 unlock_page(old_page
);
2131 dirty_page
= old_page
;
2132 get_page(dirty_page
);
2136 * Clear the pages cpupid information as the existing
2137 * information potentially belongs to a now completely
2138 * unrelated process.
2141 page_cpupid_xchg_last(old_page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2143 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2144 entry
= pte_mkyoung(orig_pte
);
2145 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2146 if (ptep_set_access_flags(vma
, address
, page_table
, entry
,1))
2147 update_mmu_cache(vma
, address
, page_table
);
2148 pte_unmap_unlock(page_table
, ptl
);
2149 ret
|= VM_FAULT_WRITE
;
2155 * Yes, Virginia, this is actually required to prevent a race
2156 * with clear_page_dirty_for_io() from clearing the page dirty
2157 * bit after it clear all dirty ptes, but before a racing
2158 * do_wp_page installs a dirty pte.
2160 * do_shared_fault is protected similarly.
2162 if (!page_mkwrite
) {
2163 wait_on_page_locked(dirty_page
);
2164 set_page_dirty_balance(dirty_page
);
2165 /* file_update_time outside page_lock */
2167 file_update_time(vma
->vm_file
);
2169 put_page(dirty_page
);
2171 struct address_space
*mapping
= dirty_page
->mapping
;
2173 set_page_dirty(dirty_page
);
2174 unlock_page(dirty_page
);
2175 page_cache_release(dirty_page
);
2178 * Some device drivers do not set page.mapping
2179 * but still dirty their pages
2181 balance_dirty_pages_ratelimited(mapping
);
2189 * Ok, we need to copy. Oh, well..
2191 page_cache_get(old_page
);
2193 pte_unmap_unlock(page_table
, ptl
);
2195 if (unlikely(anon_vma_prepare(vma
)))
2198 if (is_zero_pfn(pte_pfn(orig_pte
))) {
2199 new_page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2203 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2206 cow_user_page(new_page
, old_page
, address
, vma
);
2208 __SetPageUptodate(new_page
);
2210 if (mem_cgroup_try_charge(new_page
, mm
, GFP_KERNEL
, &memcg
))
2213 mmun_start
= address
& PAGE_MASK
;
2214 mmun_end
= mmun_start
+ PAGE_SIZE
;
2215 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2218 * Re-check the pte - we dropped the lock
2220 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2221 if (likely(pte_same(*page_table
, orig_pte
))) {
2223 if (!PageAnon(old_page
)) {
2224 dec_mm_counter_fast(mm
, MM_FILEPAGES
);
2225 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2228 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2229 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2230 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2231 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2233 * Clear the pte entry and flush it first, before updating the
2234 * pte with the new entry. This will avoid a race condition
2235 * seen in the presence of one thread doing SMC and another
2238 ptep_clear_flush_notify(vma
, address
, page_table
);
2239 page_add_new_anon_rmap(new_page
, vma
, address
);
2240 mem_cgroup_commit_charge(new_page
, memcg
, false);
2241 lru_cache_add_active_or_unevictable(new_page
, vma
);
2243 * We call the notify macro here because, when using secondary
2244 * mmu page tables (such as kvm shadow page tables), we want the
2245 * new page to be mapped directly into the secondary page table.
2247 set_pte_at_notify(mm
, address
, page_table
, entry
);
2248 update_mmu_cache(vma
, address
, page_table
);
2251 * Only after switching the pte to the new page may
2252 * we remove the mapcount here. Otherwise another
2253 * process may come and find the rmap count decremented
2254 * before the pte is switched to the new page, and
2255 * "reuse" the old page writing into it while our pte
2256 * here still points into it and can be read by other
2259 * The critical issue is to order this
2260 * page_remove_rmap with the ptp_clear_flush above.
2261 * Those stores are ordered by (if nothing else,)
2262 * the barrier present in the atomic_add_negative
2263 * in page_remove_rmap.
2265 * Then the TLB flush in ptep_clear_flush ensures that
2266 * no process can access the old page before the
2267 * decremented mapcount is visible. And the old page
2268 * cannot be reused until after the decremented
2269 * mapcount is visible. So transitively, TLBs to
2270 * old page will be flushed before it can be reused.
2272 page_remove_rmap(old_page
);
2275 /* Free the old page.. */
2276 new_page
= old_page
;
2277 ret
|= VM_FAULT_WRITE
;
2279 mem_cgroup_cancel_charge(new_page
, memcg
);
2282 page_cache_release(new_page
);
2284 pte_unmap_unlock(page_table
, ptl
);
2285 if (mmun_end
> mmun_start
)
2286 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2289 * Don't let another task, with possibly unlocked vma,
2290 * keep the mlocked page.
2292 if ((ret
& VM_FAULT_WRITE
) && (vma
->vm_flags
& VM_LOCKED
)) {
2293 lock_page(old_page
); /* LRU manipulation */
2294 munlock_vma_page(old_page
);
2295 unlock_page(old_page
);
2297 page_cache_release(old_page
);
2301 page_cache_release(new_page
);
2304 page_cache_release(old_page
);
2305 return VM_FAULT_OOM
;
2308 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2309 unsigned long start_addr
, unsigned long end_addr
,
2310 struct zap_details
*details
)
2312 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2315 static inline void unmap_mapping_range_tree(struct rb_root
*root
,
2316 struct zap_details
*details
)
2318 struct vm_area_struct
*vma
;
2319 pgoff_t vba
, vea
, zba
, zea
;
2321 vma_interval_tree_foreach(vma
, root
,
2322 details
->first_index
, details
->last_index
) {
2324 vba
= vma
->vm_pgoff
;
2325 vea
= vba
+ vma_pages(vma
) - 1;
2326 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2327 zba
= details
->first_index
;
2330 zea
= details
->last_index
;
2334 unmap_mapping_range_vma(vma
,
2335 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2336 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2341 static inline void unmap_mapping_range_list(struct list_head
*head
,
2342 struct zap_details
*details
)
2344 struct vm_area_struct
*vma
;
2347 * In nonlinear VMAs there is no correspondence between virtual address
2348 * offset and file offset. So we must perform an exhaustive search
2349 * across *all* the pages in each nonlinear VMA, not just the pages
2350 * whose virtual address lies outside the file truncation point.
2352 list_for_each_entry(vma
, head
, shared
.nonlinear
) {
2353 details
->nonlinear_vma
= vma
;
2354 unmap_mapping_range_vma(vma
, vma
->vm_start
, vma
->vm_end
, details
);
2359 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
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
.nonlinear_vma
= NULL
;
2389 details
.first_index
= hba
;
2390 details
.last_index
= hba
+ hlen
- 1;
2391 if (details
.last_index
< details
.first_index
)
2392 details
.last_index
= ULONG_MAX
;
2395 mutex_lock(&mapping
->i_mmap_mutex
);
2396 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
)))
2397 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2398 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
2399 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
2400 mutex_unlock(&mapping
->i_mmap_mutex
);
2402 EXPORT_SYMBOL(unmap_mapping_range
);
2405 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2406 * but allow concurrent faults), and pte mapped but not yet locked.
2407 * We return with pte unmapped and unlocked.
2409 * We return with the mmap_sem locked or unlocked in the same cases
2410 * as does filemap_fault().
2412 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2413 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2414 unsigned int flags
, pte_t orig_pte
)
2417 struct page
*page
, *swapcache
;
2418 struct mem_cgroup
*memcg
;
2425 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2428 entry
= pte_to_swp_entry(orig_pte
);
2429 if (unlikely(non_swap_entry(entry
))) {
2430 if (is_migration_entry(entry
)) {
2431 migration_entry_wait(mm
, pmd
, address
);
2432 } else if (is_hwpoison_entry(entry
)) {
2433 ret
= VM_FAULT_HWPOISON
;
2435 print_bad_pte(vma
, address
, orig_pte
, NULL
);
2436 ret
= VM_FAULT_SIGBUS
;
2440 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2441 page
= lookup_swap_cache(entry
);
2443 page
= swapin_readahead(entry
,
2444 GFP_HIGHUSER_MOVABLE
, vma
, address
);
2447 * Back out if somebody else faulted in this pte
2448 * while we released the pte lock.
2450 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2451 if (likely(pte_same(*page_table
, orig_pte
)))
2453 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2457 /* Had to read the page from swap area: Major fault */
2458 ret
= VM_FAULT_MAJOR
;
2459 count_vm_event(PGMAJFAULT
);
2460 mem_cgroup_count_vm_event(mm
, PGMAJFAULT
);
2461 } else if (PageHWPoison(page
)) {
2463 * hwpoisoned dirty swapcache pages are kept for killing
2464 * owner processes (which may be unknown at hwpoison time)
2466 ret
= VM_FAULT_HWPOISON
;
2467 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2473 locked
= lock_page_or_retry(page
, mm
, flags
);
2475 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2477 ret
|= VM_FAULT_RETRY
;
2482 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2483 * release the swapcache from under us. The page pin, and pte_same
2484 * test below, are not enough to exclude that. Even if it is still
2485 * swapcache, we need to check that the page's swap has not changed.
2487 if (unlikely(!PageSwapCache(page
) || page_private(page
) != entry
.val
))
2490 page
= ksm_might_need_to_copy(page
, vma
, address
);
2491 if (unlikely(!page
)) {
2497 if (mem_cgroup_try_charge(page
, mm
, GFP_KERNEL
, &memcg
)) {
2503 * Back out if somebody else already faulted in this pte.
2505 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2506 if (unlikely(!pte_same(*page_table
, orig_pte
)))
2509 if (unlikely(!PageUptodate(page
))) {
2510 ret
= VM_FAULT_SIGBUS
;
2515 * The page isn't present yet, go ahead with the fault.
2517 * Be careful about the sequence of operations here.
2518 * To get its accounting right, reuse_swap_page() must be called
2519 * while the page is counted on swap but not yet in mapcount i.e.
2520 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2521 * must be called after the swap_free(), or it will never succeed.
2524 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2525 dec_mm_counter_fast(mm
, MM_SWAPENTS
);
2526 pte
= mk_pte(page
, vma
->vm_page_prot
);
2527 if ((flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
)) {
2528 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
2529 flags
&= ~FAULT_FLAG_WRITE
;
2530 ret
|= VM_FAULT_WRITE
;
2533 flush_icache_page(vma
, page
);
2534 if (pte_swp_soft_dirty(orig_pte
))
2535 pte
= pte_mksoft_dirty(pte
);
2536 set_pte_at(mm
, address
, page_table
, pte
);
2537 if (page
== swapcache
) {
2538 do_page_add_anon_rmap(page
, vma
, address
, exclusive
);
2539 mem_cgroup_commit_charge(page
, memcg
, true);
2540 } else { /* ksm created a completely new copy */
2541 page_add_new_anon_rmap(page
, vma
, address
);
2542 mem_cgroup_commit_charge(page
, memcg
, false);
2543 lru_cache_add_active_or_unevictable(page
, vma
);
2547 if (vm_swap_full() || (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
2548 try_to_free_swap(page
);
2550 if (page
!= swapcache
) {
2552 * Hold the lock to avoid the swap entry to be reused
2553 * until we take the PT lock for the pte_same() check
2554 * (to avoid false positives from pte_same). For
2555 * further safety release the lock after the swap_free
2556 * so that the swap count won't change under a
2557 * parallel locked swapcache.
2559 unlock_page(swapcache
);
2560 page_cache_release(swapcache
);
2563 if (flags
& FAULT_FLAG_WRITE
) {
2564 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
2565 if (ret
& VM_FAULT_ERROR
)
2566 ret
&= VM_FAULT_ERROR
;
2570 /* No need to invalidate - it was non-present before */
2571 update_mmu_cache(vma
, address
, page_table
);
2573 pte_unmap_unlock(page_table
, ptl
);
2577 mem_cgroup_cancel_charge(page
, memcg
);
2578 pte_unmap_unlock(page_table
, ptl
);
2582 page_cache_release(page
);
2583 if (page
!= swapcache
) {
2584 unlock_page(swapcache
);
2585 page_cache_release(swapcache
);
2591 * This is like a special single-page "expand_{down|up}wards()",
2592 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2593 * doesn't hit another vma.
2595 static inline int check_stack_guard_page(struct vm_area_struct
*vma
, unsigned long address
)
2597 address
&= PAGE_MASK
;
2598 if ((vma
->vm_flags
& VM_GROWSDOWN
) && address
== vma
->vm_start
) {
2599 struct vm_area_struct
*prev
= vma
->vm_prev
;
2602 * Is there a mapping abutting this one below?
2604 * That's only ok if it's the same stack mapping
2605 * that has gotten split..
2607 if (prev
&& prev
->vm_end
== address
)
2608 return prev
->vm_flags
& VM_GROWSDOWN
? 0 : -ENOMEM
;
2610 expand_downwards(vma
, address
- PAGE_SIZE
);
2612 if ((vma
->vm_flags
& VM_GROWSUP
) && address
+ PAGE_SIZE
== vma
->vm_end
) {
2613 struct vm_area_struct
*next
= vma
->vm_next
;
2615 /* As VM_GROWSDOWN but s/below/above/ */
2616 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
)
2617 return next
->vm_flags
& VM_GROWSUP
? 0 : -ENOMEM
;
2619 expand_upwards(vma
, address
+ PAGE_SIZE
);
2625 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2626 * but allow concurrent faults), and pte mapped but not yet locked.
2627 * We return with mmap_sem still held, but pte unmapped and unlocked.
2629 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2630 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2633 struct mem_cgroup
*memcg
;
2638 pte_unmap(page_table
);
2640 /* Check if we need to add a guard page to the stack */
2641 if (check_stack_guard_page(vma
, address
) < 0)
2642 return VM_FAULT_SIGBUS
;
2644 /* Use the zero-page for reads */
2645 if (!(flags
& FAULT_FLAG_WRITE
)) {
2646 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(address
),
2647 vma
->vm_page_prot
));
2648 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2649 if (!pte_none(*page_table
))
2654 /* Allocate our own private page. */
2655 if (unlikely(anon_vma_prepare(vma
)))
2657 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2661 * The memory barrier inside __SetPageUptodate makes sure that
2662 * preceeding stores to the page contents become visible before
2663 * the set_pte_at() write.
2665 __SetPageUptodate(page
);
2667 if (mem_cgroup_try_charge(page
, mm
, GFP_KERNEL
, &memcg
))
2670 entry
= mk_pte(page
, vma
->vm_page_prot
);
2671 if (vma
->vm_flags
& VM_WRITE
)
2672 entry
= pte_mkwrite(pte_mkdirty(entry
));
2674 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2675 if (!pte_none(*page_table
))
2678 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2679 page_add_new_anon_rmap(page
, vma
, address
);
2680 mem_cgroup_commit_charge(page
, memcg
, false);
2681 lru_cache_add_active_or_unevictable(page
, vma
);
2683 set_pte_at(mm
, address
, page_table
, entry
);
2685 /* No need to invalidate - it was non-present before */
2686 update_mmu_cache(vma
, address
, page_table
);
2688 pte_unmap_unlock(page_table
, ptl
);
2691 mem_cgroup_cancel_charge(page
, memcg
);
2692 page_cache_release(page
);
2695 page_cache_release(page
);
2697 return VM_FAULT_OOM
;
2701 * The mmap_sem must have been held on entry, and may have been
2702 * released depending on flags and vma->vm_ops->fault() return value.
2703 * See filemap_fault() and __lock_page_retry().
2705 static int __do_fault(struct vm_area_struct
*vma
, unsigned long address
,
2706 pgoff_t pgoff
, unsigned int flags
, struct page
**page
)
2708 struct vm_fault vmf
;
2711 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
2716 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
2717 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
2720 if (unlikely(PageHWPoison(vmf
.page
))) {
2721 if (ret
& VM_FAULT_LOCKED
)
2722 unlock_page(vmf
.page
);
2723 page_cache_release(vmf
.page
);
2724 return VM_FAULT_HWPOISON
;
2727 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
2728 lock_page(vmf
.page
);
2730 VM_BUG_ON_PAGE(!PageLocked(vmf
.page
), vmf
.page
);
2737 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2739 * @vma: virtual memory area
2740 * @address: user virtual address
2741 * @page: page to map
2742 * @pte: pointer to target page table entry
2743 * @write: true, if new entry is writable
2744 * @anon: true, if it's anonymous page
2746 * Caller must hold page table lock relevant for @pte.
2748 * Target users are page handler itself and implementations of
2749 * vm_ops->map_pages.
2751 void do_set_pte(struct vm_area_struct
*vma
, unsigned long address
,
2752 struct page
*page
, pte_t
*pte
, bool write
, bool anon
)
2756 flush_icache_page(vma
, page
);
2757 entry
= mk_pte(page
, vma
->vm_page_prot
);
2759 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2760 else if (pte_file(*pte
) && pte_file_soft_dirty(*pte
))
2761 entry
= pte_mksoft_dirty(entry
);
2763 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
2764 page_add_new_anon_rmap(page
, vma
, address
);
2766 inc_mm_counter_fast(vma
->vm_mm
, MM_FILEPAGES
);
2767 page_add_file_rmap(page
);
2769 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
2771 /* no need to invalidate: a not-present page won't be cached */
2772 update_mmu_cache(vma
, address
, pte
);
2775 static unsigned long fault_around_bytes __read_mostly
=
2776 rounddown_pow_of_two(65536);
2778 #ifdef CONFIG_DEBUG_FS
2779 static int fault_around_bytes_get(void *data
, u64
*val
)
2781 *val
= fault_around_bytes
;
2786 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2787 * rounded down to nearest page order. It's what do_fault_around() expects to
2790 static int fault_around_bytes_set(void *data
, u64 val
)
2792 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
2794 if (val
> PAGE_SIZE
)
2795 fault_around_bytes
= rounddown_pow_of_two(val
);
2797 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
2800 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops
,
2801 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
2803 static int __init
fault_around_debugfs(void)
2807 ret
= debugfs_create_file("fault_around_bytes", 0644, NULL
, NULL
,
2808 &fault_around_bytes_fops
);
2810 pr_warn("Failed to create fault_around_bytes in debugfs");
2813 late_initcall(fault_around_debugfs
);
2817 * do_fault_around() tries to map few pages around the fault address. The hope
2818 * is that the pages will be needed soon and this will lower the number of
2821 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2822 * not ready to be mapped: not up-to-date, locked, etc.
2824 * This function is called with the page table lock taken. In the split ptlock
2825 * case the page table lock only protects only those entries which belong to
2826 * the page table corresponding to the fault address.
2828 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2831 * fault_around_pages() defines how many pages we'll try to map.
2832 * do_fault_around() expects it to return a power of two less than or equal to
2835 * The virtual address of the area that we map is naturally aligned to the
2836 * fault_around_pages() value (and therefore to page order). This way it's
2837 * easier to guarantee that we don't cross page table boundaries.
2839 static void do_fault_around(struct vm_area_struct
*vma
, unsigned long address
,
2840 pte_t
*pte
, pgoff_t pgoff
, unsigned int flags
)
2842 unsigned long start_addr
, nr_pages
, mask
;
2844 struct vm_fault vmf
;
2847 nr_pages
= ACCESS_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
2848 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
2850 start_addr
= max(address
& mask
, vma
->vm_start
);
2851 off
= ((address
- start_addr
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
2856 * max_pgoff is either end of page table or end of vma
2857 * or fault_around_pages() from pgoff, depending what is nearest.
2859 max_pgoff
= pgoff
- ((start_addr
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
2861 max_pgoff
= min3(max_pgoff
, vma_pages(vma
) + vma
->vm_pgoff
- 1,
2862 pgoff
+ nr_pages
- 1);
2864 /* Check if it makes any sense to call ->map_pages */
2865 while (!pte_none(*pte
)) {
2866 if (++pgoff
> max_pgoff
)
2868 start_addr
+= PAGE_SIZE
;
2869 if (start_addr
>= vma
->vm_end
)
2874 vmf
.virtual_address
= (void __user
*) start_addr
;
2877 vmf
.max_pgoff
= max_pgoff
;
2879 vma
->vm_ops
->map_pages(vma
, &vmf
);
2882 static int do_read_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2883 unsigned long address
, pmd_t
*pmd
,
2884 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
2886 struct page
*fault_page
;
2892 * Let's call ->map_pages() first and use ->fault() as fallback
2893 * if page by the offset is not ready to be mapped (cold cache or
2896 if (vma
->vm_ops
->map_pages
&& !(flags
& FAULT_FLAG_NONLINEAR
) &&
2897 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
, &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
, &fault_page
);
2946 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
2949 copy_user_highpage(new_page
, fault_page
, address
, vma
);
2950 __SetPageUptodate(new_page
);
2952 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2953 if (unlikely(!pte_same(*pte
, orig_pte
))) {
2954 pte_unmap_unlock(pte
, ptl
);
2955 unlock_page(fault_page
);
2956 page_cache_release(fault_page
);
2959 do_set_pte(vma
, address
, new_page
, pte
, true, true);
2960 mem_cgroup_commit_charge(new_page
, memcg
, false);
2961 lru_cache_add_active_or_unevictable(new_page
, vma
);
2962 pte_unmap_unlock(pte
, ptl
);
2963 unlock_page(fault_page
);
2964 page_cache_release(fault_page
);
2967 mem_cgroup_cancel_charge(new_page
, memcg
);
2968 page_cache_release(new_page
);
2972 static int do_shared_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2973 unsigned long address
, pmd_t
*pmd
,
2974 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
2976 struct page
*fault_page
;
2977 struct address_space
*mapping
;
2983 ret
= __do_fault(vma
, address
, pgoff
, flags
, &fault_page
);
2984 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
2988 * Check if the backing address space wants to know that the page is
2989 * about to become writable
2991 if (vma
->vm_ops
->page_mkwrite
) {
2992 unlock_page(fault_page
);
2993 tmp
= do_page_mkwrite(vma
, fault_page
, address
);
2994 if (unlikely(!tmp
||
2995 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
2996 page_cache_release(fault_page
);
3001 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3002 if (unlikely(!pte_same(*pte
, orig_pte
))) {
3003 pte_unmap_unlock(pte
, ptl
);
3004 unlock_page(fault_page
);
3005 page_cache_release(fault_page
);
3008 do_set_pte(vma
, address
, fault_page
, pte
, true, false);
3009 pte_unmap_unlock(pte
, ptl
);
3011 if (set_page_dirty(fault_page
))
3013 mapping
= fault_page
->mapping
;
3014 unlock_page(fault_page
);
3015 if ((dirtied
|| vma
->vm_ops
->page_mkwrite
) && mapping
) {
3017 * Some device drivers do not set page.mapping but still
3020 balance_dirty_pages_ratelimited(mapping
);
3023 /* file_update_time outside page_lock */
3024 if (vma
->vm_file
&& !vma
->vm_ops
->page_mkwrite
)
3025 file_update_time(vma
->vm_file
);
3031 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3032 * but allow concurrent faults).
3033 * The mmap_sem may have been released depending on flags and our
3034 * return value. See filemap_fault() and __lock_page_or_retry().
3036 static int do_linear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3037 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3038 unsigned int flags
, pte_t orig_pte
)
3040 pgoff_t pgoff
= (((address
& PAGE_MASK
)
3041 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
3043 pte_unmap(page_table
);
3044 if (!(flags
& FAULT_FLAG_WRITE
))
3045 return do_read_fault(mm
, vma
, address
, pmd
, pgoff
, flags
,
3047 if (!(vma
->vm_flags
& VM_SHARED
))
3048 return do_cow_fault(mm
, vma
, address
, pmd
, pgoff
, flags
,
3050 return do_shared_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3054 * Fault of a previously existing named mapping. Repopulate the pte
3055 * from the encoded file_pte if possible. This enables swappable
3058 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3059 * but allow concurrent faults), and pte mapped but not yet locked.
3060 * We return with pte unmapped and unlocked.
3061 * The mmap_sem may have been released depending on flags and our
3062 * return value. See filemap_fault() and __lock_page_or_retry().
3064 static int do_nonlinear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3065 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3066 unsigned int flags
, pte_t orig_pte
)
3070 flags
|= FAULT_FLAG_NONLINEAR
;
3072 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
3075 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
))) {
3077 * Page table corrupted: show pte and kill process.
3079 print_bad_pte(vma
, address
, orig_pte
, NULL
);
3080 return VM_FAULT_SIGBUS
;
3083 pgoff
= pte_to_pgoff(orig_pte
);
3084 if (!(flags
& FAULT_FLAG_WRITE
))
3085 return do_read_fault(mm
, vma
, address
, pmd
, pgoff
, flags
,
3087 if (!(vma
->vm_flags
& VM_SHARED
))
3088 return do_cow_fault(mm
, vma
, address
, pmd
, pgoff
, flags
,
3090 return do_shared_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3093 static int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3094 unsigned long addr
, int page_nid
,
3099 count_vm_numa_event(NUMA_HINT_FAULTS
);
3100 if (page_nid
== numa_node_id()) {
3101 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3102 *flags
|= TNF_FAULT_LOCAL
;
3105 return mpol_misplaced(page
, vma
, addr
);
3108 static int do_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3109 unsigned long addr
, pte_t pte
, pte_t
*ptep
, pmd_t
*pmd
)
3111 struct page
*page
= NULL
;
3116 bool migrated
= false;
3120 * The "pte" at this point cannot be used safely without
3121 * validation through pte_unmap_same(). It's of NUMA type but
3122 * the pfn may be screwed if the read is non atomic.
3124 * ptep_modify_prot_start is not called as this is clearing
3125 * the _PAGE_NUMA bit and it is not really expected that there
3126 * would be concurrent hardware modifications to the PTE.
3128 ptl
= pte_lockptr(mm
, pmd
);
3130 if (unlikely(!pte_same(*ptep
, pte
))) {
3131 pte_unmap_unlock(ptep
, ptl
);
3135 pte
= pte_mknonnuma(pte
);
3136 set_pte_at(mm
, addr
, ptep
, pte
);
3137 update_mmu_cache(vma
, addr
, ptep
);
3139 page
= vm_normal_page(vma
, addr
, pte
);
3141 pte_unmap_unlock(ptep
, ptl
);
3144 BUG_ON(is_zero_pfn(page_to_pfn(page
)));
3147 * Avoid grouping on DSO/COW pages in specific and RO pages
3148 * in general, RO pages shouldn't hurt as much anyway since
3149 * they can be in shared cache state.
3151 if (!pte_write(pte
))
3152 flags
|= TNF_NO_GROUP
;
3155 * Flag if the page is shared between multiple address spaces. This
3156 * is later used when determining whether to group tasks together
3158 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
3159 flags
|= TNF_SHARED
;
3161 last_cpupid
= page_cpupid_last(page
);
3162 page_nid
= page_to_nid(page
);
3163 target_nid
= numa_migrate_prep(page
, vma
, addr
, page_nid
, &flags
);
3164 pte_unmap_unlock(ptep
, ptl
);
3165 if (target_nid
== -1) {
3170 /* Migrate to the requested node */
3171 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
3173 page_nid
= target_nid
;
3174 flags
|= TNF_MIGRATED
;
3179 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
3184 * These routines also need to handle stuff like marking pages dirty
3185 * and/or accessed for architectures that don't do it in hardware (most
3186 * RISC architectures). The early dirtying is also good on the i386.
3188 * There is also a hook called "update_mmu_cache()" that architectures
3189 * with external mmu caches can use to update those (ie the Sparc or
3190 * PowerPC hashed page tables that act as extended TLBs).
3192 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3193 * but allow concurrent faults), and pte mapped but not yet locked.
3194 * We return with pte unmapped and unlocked.
3196 * The mmap_sem may have been released depending on flags and our
3197 * return value. See filemap_fault() and __lock_page_or_retry().
3199 static int handle_pte_fault(struct mm_struct
*mm
,
3200 struct vm_area_struct
*vma
, unsigned long address
,
3201 pte_t
*pte
, pmd_t
*pmd
, unsigned int flags
)
3206 entry
= ACCESS_ONCE(*pte
);
3207 if (!pte_present(entry
)) {
3208 if (pte_none(entry
)) {
3210 if (likely(vma
->vm_ops
->fault
))
3211 return do_linear_fault(mm
, vma
, address
,
3212 pte
, pmd
, flags
, entry
);
3214 return do_anonymous_page(mm
, vma
, address
,
3217 if (pte_file(entry
))
3218 return do_nonlinear_fault(mm
, vma
, address
,
3219 pte
, pmd
, flags
, entry
);
3220 return do_swap_page(mm
, vma
, address
,
3221 pte
, pmd
, flags
, entry
);
3224 if (pte_numa(entry
))
3225 return do_numa_page(mm
, vma
, address
, entry
, pte
, pmd
);
3227 ptl
= pte_lockptr(mm
, pmd
);
3229 if (unlikely(!pte_same(*pte
, entry
)))
3231 if (flags
& FAULT_FLAG_WRITE
) {
3232 if (!pte_write(entry
))
3233 return do_wp_page(mm
, vma
, address
,
3234 pte
, pmd
, ptl
, entry
);
3235 entry
= pte_mkdirty(entry
);
3237 entry
= pte_mkyoung(entry
);
3238 if (ptep_set_access_flags(vma
, address
, pte
, entry
, flags
& FAULT_FLAG_WRITE
)) {
3239 update_mmu_cache(vma
, address
, pte
);
3242 * This is needed only for protection faults but the arch code
3243 * is not yet telling us if this is a protection fault or not.
3244 * This still avoids useless tlb flushes for .text page faults
3247 if (flags
& FAULT_FLAG_WRITE
)
3248 flush_tlb_fix_spurious_fault(vma
, address
);
3251 pte_unmap_unlock(pte
, ptl
);
3256 * By the time we get here, we already hold the mm semaphore
3258 * The mmap_sem may have been released depending on flags and our
3259 * return value. See filemap_fault() and __lock_page_or_retry().
3261 static int __handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3262 unsigned long address
, unsigned int flags
)
3269 if (unlikely(is_vm_hugetlb_page(vma
)))
3270 return hugetlb_fault(mm
, vma
, address
, flags
);
3272 pgd
= pgd_offset(mm
, address
);
3273 pud
= pud_alloc(mm
, pgd
, address
);
3275 return VM_FAULT_OOM
;
3276 pmd
= pmd_alloc(mm
, pud
, address
);
3278 return VM_FAULT_OOM
;
3279 if (pmd_none(*pmd
) && transparent_hugepage_enabled(vma
)) {
3280 int ret
= VM_FAULT_FALLBACK
;
3282 ret
= do_huge_pmd_anonymous_page(mm
, vma
, address
,
3284 if (!(ret
& VM_FAULT_FALLBACK
))
3287 pmd_t orig_pmd
= *pmd
;
3291 if (pmd_trans_huge(orig_pmd
)) {
3292 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
3295 * If the pmd is splitting, return and retry the
3296 * the fault. Alternative: wait until the split
3297 * is done, and goto retry.
3299 if (pmd_trans_splitting(orig_pmd
))
3302 if (pmd_numa(orig_pmd
))
3303 return do_huge_pmd_numa_page(mm
, vma
, address
,
3306 if (dirty
&& !pmd_write(orig_pmd
)) {
3307 ret
= do_huge_pmd_wp_page(mm
, vma
, address
, pmd
,
3309 if (!(ret
& VM_FAULT_FALLBACK
))
3312 huge_pmd_set_accessed(mm
, vma
, address
, pmd
,
3320 * Use __pte_alloc instead of pte_alloc_map, because we can't
3321 * run pte_offset_map on the pmd, if an huge pmd could
3322 * materialize from under us from a different thread.
3324 if (unlikely(pmd_none(*pmd
)) &&
3325 unlikely(__pte_alloc(mm
, vma
, pmd
, address
)))
3326 return VM_FAULT_OOM
;
3327 /* if an huge pmd materialized from under us just retry later */
3328 if (unlikely(pmd_trans_huge(*pmd
)))
3331 * A regular pmd is established and it can't morph into a huge pmd
3332 * from under us anymore at this point because we hold the mmap_sem
3333 * read mode and khugepaged takes it in write mode. So now it's
3334 * safe to run pte_offset_map().
3336 pte
= pte_offset_map(pmd
, address
);
3338 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
3342 * By the time we get here, we already hold the mm semaphore
3344 * The mmap_sem may have been released depending on flags and our
3345 * return value. See filemap_fault() and __lock_page_or_retry().
3347 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3348 unsigned long address
, unsigned int flags
)
3352 __set_current_state(TASK_RUNNING
);
3354 count_vm_event(PGFAULT
);
3355 mem_cgroup_count_vm_event(mm
, PGFAULT
);
3357 /* do counter updates before entering really critical section. */
3358 check_sync_rss_stat(current
);
3361 * Enable the memcg OOM handling for faults triggered in user
3362 * space. Kernel faults are handled more gracefully.
3364 if (flags
& FAULT_FLAG_USER
)
3365 mem_cgroup_oom_enable();
3367 ret
= __handle_mm_fault(mm
, vma
, address
, flags
);
3369 if (flags
& FAULT_FLAG_USER
) {
3370 mem_cgroup_oom_disable();
3372 * The task may have entered a memcg OOM situation but
3373 * if the allocation error was handled gracefully (no
3374 * VM_FAULT_OOM), there is no need to kill anything.
3375 * Just clean up the OOM state peacefully.
3377 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
3378 mem_cgroup_oom_synchronize(false);
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
)) /* Another has populated it */
3425 pud_populate(mm
, pud
, new);
3427 if (pgd_present(*pud
)) /* Another has populated it */
3430 pgd_populate(mm
, pud
, new);
3431 #endif /* __ARCH_HAS_4LEVEL_HACK */
3432 spin_unlock(&mm
->page_table_lock
);
3435 #endif /* __PAGETABLE_PMD_FOLDED */
3437 static int __follow_pte(struct mm_struct
*mm
, unsigned long address
,
3438 pte_t
**ptepp
, spinlock_t
**ptlp
)
3445 pgd
= pgd_offset(mm
, address
);
3446 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
3449 pud
= pud_offset(pgd
, address
);
3450 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
3453 pmd
= pmd_offset(pud
, address
);
3454 VM_BUG_ON(pmd_trans_huge(*pmd
));
3455 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
3458 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3462 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
3465 if (!pte_present(*ptep
))
3470 pte_unmap_unlock(ptep
, *ptlp
);
3475 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
3476 pte_t
**ptepp
, spinlock_t
**ptlp
)
3480 /* (void) is needed to make gcc happy */
3481 (void) __cond_lock(*ptlp
,
3482 !(res
= __follow_pte(mm
, address
, ptepp
, ptlp
)));
3487 * follow_pfn - look up PFN at a user virtual address
3488 * @vma: memory mapping
3489 * @address: user virtual address
3490 * @pfn: location to store found PFN
3492 * Only IO mappings and raw PFN mappings are allowed.
3494 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3496 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
3503 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3506 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
3509 *pfn
= pte_pfn(*ptep
);
3510 pte_unmap_unlock(ptep
, ptl
);
3513 EXPORT_SYMBOL(follow_pfn
);
3515 #ifdef CONFIG_HAVE_IOREMAP_PROT
3516 int follow_phys(struct vm_area_struct
*vma
,
3517 unsigned long address
, unsigned int flags
,
3518 unsigned long *prot
, resource_size_t
*phys
)
3524 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3527 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
3531 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
3534 *prot
= pgprot_val(pte_pgprot(pte
));
3535 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
3539 pte_unmap_unlock(ptep
, ptl
);
3544 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
3545 void *buf
, int len
, int write
)
3547 resource_size_t phys_addr
;
3548 unsigned long prot
= 0;
3549 void __iomem
*maddr
;
3550 int offset
= addr
& (PAGE_SIZE
-1);
3552 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
3555 maddr
= ioremap_prot(phys_addr
, PAGE_SIZE
, prot
);
3557 memcpy_toio(maddr
+ offset
, buf
, len
);
3559 memcpy_fromio(buf
, maddr
+ offset
, len
);
3564 EXPORT_SYMBOL_GPL(generic_access_phys
);
3568 * Access another process' address space as given in mm. If non-NULL, use the
3569 * given task for page fault accounting.
3571 static int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
3572 unsigned long addr
, void *buf
, int len
, int write
)
3574 struct vm_area_struct
*vma
;
3575 void *old_buf
= buf
;
3577 down_read(&mm
->mmap_sem
);
3578 /* ignore errors, just check how much was successfully transferred */
3580 int bytes
, ret
, offset
;
3582 struct page
*page
= NULL
;
3584 ret
= get_user_pages(tsk
, mm
, addr
, 1,
3585 write
, 1, &page
, &vma
);
3587 #ifndef CONFIG_HAVE_IOREMAP_PROT
3591 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3592 * we can access using slightly different code.
3594 vma
= find_vma(mm
, addr
);
3595 if (!vma
|| vma
->vm_start
> addr
)
3597 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
3598 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
3606 offset
= addr
& (PAGE_SIZE
-1);
3607 if (bytes
> PAGE_SIZE
-offset
)
3608 bytes
= PAGE_SIZE
-offset
;
3612 copy_to_user_page(vma
, page
, addr
,
3613 maddr
+ offset
, buf
, bytes
);
3614 set_page_dirty_lock(page
);
3616 copy_from_user_page(vma
, page
, addr
,
3617 buf
, maddr
+ offset
, bytes
);
3620 page_cache_release(page
);
3626 up_read(&mm
->mmap_sem
);
3628 return buf
- old_buf
;
3632 * access_remote_vm - access another process' address space
3633 * @mm: the mm_struct of the target address space
3634 * @addr: start address to access
3635 * @buf: source or destination buffer
3636 * @len: number of bytes to transfer
3637 * @write: whether the access is a write
3639 * The caller must hold a reference on @mm.
3641 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
3642 void *buf
, int len
, int write
)
3644 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, write
);
3648 * Access another process' address space.
3649 * Source/target buffer must be kernel space,
3650 * Do not walk the page table directly, use get_user_pages
3652 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
3653 void *buf
, int len
, int write
)
3655 struct mm_struct
*mm
;
3658 mm
= get_task_mm(tsk
);
3662 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, write
);
3669 * Print the name of a VMA.
3671 void print_vma_addr(char *prefix
, unsigned long ip
)
3673 struct mm_struct
*mm
= current
->mm
;
3674 struct vm_area_struct
*vma
;
3677 * Do not print if we are in atomic
3678 * contexts (in exception stacks, etc.):
3680 if (preempt_count())
3683 down_read(&mm
->mmap_sem
);
3684 vma
= find_vma(mm
, ip
);
3685 if (vma
&& vma
->vm_file
) {
3686 struct file
*f
= vma
->vm_file
;
3687 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
3691 p
= d_path(&f
->f_path
, buf
, PAGE_SIZE
);
3694 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
3696 vma
->vm_end
- vma
->vm_start
);
3697 free_page((unsigned long)buf
);
3700 up_read(&mm
->mmap_sem
);
3703 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3704 void might_fault(void)
3707 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3708 * holding the mmap_sem, this is safe because kernel memory doesn't
3709 * get paged out, therefore we'll never actually fault, and the
3710 * below annotations will generate false positives.
3712 if (segment_eq(get_fs(), KERNEL_DS
))
3716 * it would be nicer only to annotate paths which are not under
3717 * pagefault_disable, however that requires a larger audit and
3718 * providing helpers like get_user_atomic.
3723 __might_sleep(__FILE__
, __LINE__
, 0);
3726 might_lock_read(¤t
->mm
->mmap_sem
);
3728 EXPORT_SYMBOL(might_fault
);
3731 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3732 static void clear_gigantic_page(struct page
*page
,
3734 unsigned int pages_per_huge_page
)
3737 struct page
*p
= page
;
3740 for (i
= 0; i
< pages_per_huge_page
;
3741 i
++, p
= mem_map_next(p
, page
, i
)) {
3743 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
3746 void clear_huge_page(struct page
*page
,
3747 unsigned long addr
, unsigned int pages_per_huge_page
)
3751 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3752 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
3757 for (i
= 0; i
< pages_per_huge_page
; i
++) {
3759 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
3763 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
3765 struct vm_area_struct
*vma
,
3766 unsigned int pages_per_huge_page
)
3769 struct page
*dst_base
= dst
;
3770 struct page
*src_base
= src
;
3772 for (i
= 0; i
< pages_per_huge_page
; ) {
3774 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
3777 dst
= mem_map_next(dst
, dst_base
, i
);
3778 src
= mem_map_next(src
, src_base
, i
);
3782 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
3783 unsigned long addr
, struct vm_area_struct
*vma
,
3784 unsigned int pages_per_huge_page
)
3788 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3789 copy_user_gigantic_page(dst
, src
, addr
, vma
,
3790 pages_per_huge_page
);
3795 for (i
= 0; i
< pages_per_huge_page
; i
++) {
3797 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
3800 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3802 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3804 static struct kmem_cache
*page_ptl_cachep
;
3806 void __init
ptlock_cache_init(void)
3808 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
3812 bool ptlock_alloc(struct page
*page
)
3816 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
3823 void ptlock_free(struct page
*page
)
3825 kmem_cache_free(page_ptl_cachep
, page
->ptl
);