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>
64 #include <linux/userfaultfd_k.h>
67 #include <asm/pgalloc.h>
68 #include <asm/uaccess.h>
70 #include <asm/tlbflush.h>
71 #include <asm/pgtable.h>
75 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
76 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
79 #ifndef CONFIG_NEED_MULTIPLE_NODES
80 /* use the per-pgdat data instead for discontigmem - mbligh */
81 unsigned long max_mapnr
;
84 EXPORT_SYMBOL(max_mapnr
);
85 EXPORT_SYMBOL(mem_map
);
89 * A number of key systems in x86 including ioremap() rely on the assumption
90 * that high_memory defines the upper bound on direct map memory, then end
91 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
92 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
97 EXPORT_SYMBOL(high_memory
);
100 * Randomize the address space (stacks, mmaps, brk, etc.).
102 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
103 * as ancient (libc5 based) binaries can segfault. )
105 int randomize_va_space __read_mostly
=
106 #ifdef CONFIG_COMPAT_BRK
112 static int __init
disable_randmaps(char *s
)
114 randomize_va_space
= 0;
117 __setup("norandmaps", disable_randmaps
);
119 unsigned long zero_pfn __read_mostly
;
120 unsigned long highest_memmap_pfn __read_mostly
;
122 EXPORT_SYMBOL(zero_pfn
);
125 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
127 static int __init
init_zero_pfn(void)
129 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
132 core_initcall(init_zero_pfn
);
135 #if defined(SPLIT_RSS_COUNTING)
137 void sync_mm_rss(struct mm_struct
*mm
)
141 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
142 if (current
->rss_stat
.count
[i
]) {
143 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
144 current
->rss_stat
.count
[i
] = 0;
147 current
->rss_stat
.events
= 0;
150 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
152 struct task_struct
*task
= current
;
154 if (likely(task
->mm
== mm
))
155 task
->rss_stat
.count
[member
] += val
;
157 add_mm_counter(mm
, member
, val
);
159 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
160 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
162 /* sync counter once per 64 page faults */
163 #define TASK_RSS_EVENTS_THRESH (64)
164 static void check_sync_rss_stat(struct task_struct
*task
)
166 if (unlikely(task
!= current
))
168 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
169 sync_mm_rss(task
->mm
);
171 #else /* SPLIT_RSS_COUNTING */
173 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
174 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
176 static void check_sync_rss_stat(struct task_struct
*task
)
180 #endif /* SPLIT_RSS_COUNTING */
182 #ifdef HAVE_GENERIC_MMU_GATHER
184 static bool tlb_next_batch(struct mmu_gather
*tlb
)
186 struct mmu_gather_batch
*batch
;
190 tlb
->active
= batch
->next
;
194 if (tlb
->batch_count
== MAX_GATHER_BATCH_COUNT
)
197 batch
= (void *)__get_free_pages(GFP_NOWAIT
| __GFP_NOWARN
, 0);
204 batch
->max
= MAX_GATHER_BATCH
;
206 tlb
->active
->next
= batch
;
213 * Called to initialize an (on-stack) mmu_gather structure for page-table
214 * tear-down from @mm. The @fullmm argument is used when @mm is without
215 * users and we're going to destroy the full address space (exit/execve).
217 void tlb_gather_mmu(struct mmu_gather
*tlb
, struct mm_struct
*mm
, unsigned long start
, unsigned long end
)
221 /* Is it from 0 to ~0? */
222 tlb
->fullmm
= !(start
| (end
+1));
223 tlb
->need_flush_all
= 0;
224 tlb
->local
.next
= NULL
;
226 tlb
->local
.max
= ARRAY_SIZE(tlb
->__pages
);
227 tlb
->active
= &tlb
->local
;
228 tlb
->batch_count
= 0;
230 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
234 __tlb_reset_range(tlb
);
237 static void tlb_flush_mmu_tlbonly(struct mmu_gather
*tlb
)
243 mmu_notifier_invalidate_range(tlb
->mm
, tlb
->start
, tlb
->end
);
244 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
245 tlb_table_flush(tlb
);
247 __tlb_reset_range(tlb
);
250 static void tlb_flush_mmu_free(struct mmu_gather
*tlb
)
252 struct mmu_gather_batch
*batch
;
254 for (batch
= &tlb
->local
; batch
&& batch
->nr
; batch
= batch
->next
) {
255 free_pages_and_swap_cache(batch
->pages
, batch
->nr
);
258 tlb
->active
= &tlb
->local
;
261 void tlb_flush_mmu(struct mmu_gather
*tlb
)
263 tlb_flush_mmu_tlbonly(tlb
);
264 tlb_flush_mmu_free(tlb
);
268 * Called at the end of the shootdown operation to free up any resources
269 * that were required.
271 void tlb_finish_mmu(struct mmu_gather
*tlb
, unsigned long start
, unsigned long end
)
273 struct mmu_gather_batch
*batch
, *next
;
277 /* keep the page table cache within bounds */
280 for (batch
= tlb
->local
.next
; batch
; batch
= next
) {
282 free_pages((unsigned long)batch
, 0);
284 tlb
->local
.next
= NULL
;
288 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
289 * handling the additional races in SMP caused by other CPUs caching valid
290 * mappings in their TLBs. Returns the number of free page slots left.
291 * When out of page slots we must call tlb_flush_mmu().
293 int __tlb_remove_page(struct mmu_gather
*tlb
, struct page
*page
)
295 struct mmu_gather_batch
*batch
;
297 VM_BUG_ON(!tlb
->end
);
300 batch
->pages
[batch
->nr
++] = page
;
301 if (batch
->nr
== batch
->max
) {
302 if (!tlb_next_batch(tlb
))
306 VM_BUG_ON_PAGE(batch
->nr
> batch
->max
, page
);
308 return batch
->max
- batch
->nr
;
311 #endif /* HAVE_GENERIC_MMU_GATHER */
313 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
316 * See the comment near struct mmu_table_batch.
319 static void tlb_remove_table_smp_sync(void *arg
)
321 /* Simply deliver the interrupt */
324 static void tlb_remove_table_one(void *table
)
327 * This isn't an RCU grace period and hence the page-tables cannot be
328 * assumed to be actually RCU-freed.
330 * It is however sufficient for software page-table walkers that rely on
331 * IRQ disabling. See the comment near struct mmu_table_batch.
333 smp_call_function(tlb_remove_table_smp_sync
, NULL
, 1);
334 __tlb_remove_table(table
);
337 static void tlb_remove_table_rcu(struct rcu_head
*head
)
339 struct mmu_table_batch
*batch
;
342 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
344 for (i
= 0; i
< batch
->nr
; i
++)
345 __tlb_remove_table(batch
->tables
[i
]);
347 free_page((unsigned long)batch
);
350 void tlb_table_flush(struct mmu_gather
*tlb
)
352 struct mmu_table_batch
**batch
= &tlb
->batch
;
355 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
360 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
362 struct mmu_table_batch
**batch
= &tlb
->batch
;
365 * When there's less then two users of this mm there cannot be a
366 * concurrent page-table walk.
368 if (atomic_read(&tlb
->mm
->mm_users
) < 2) {
369 __tlb_remove_table(table
);
373 if (*batch
== NULL
) {
374 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
375 if (*batch
== NULL
) {
376 tlb_remove_table_one(table
);
381 (*batch
)->tables
[(*batch
)->nr
++] = table
;
382 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
383 tlb_table_flush(tlb
);
386 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
389 * Note: this doesn't free the actual pages themselves. That
390 * has been handled earlier when unmapping all the memory regions.
392 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
395 pgtable_t token
= pmd_pgtable(*pmd
);
397 pte_free_tlb(tlb
, token
, addr
);
398 atomic_long_dec(&tlb
->mm
->nr_ptes
);
401 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
402 unsigned long addr
, unsigned long end
,
403 unsigned long floor
, unsigned long ceiling
)
410 pmd
= pmd_offset(pud
, addr
);
412 next
= pmd_addr_end(addr
, end
);
413 if (pmd_none_or_clear_bad(pmd
))
415 free_pte_range(tlb
, pmd
, addr
);
416 } while (pmd
++, addr
= next
, addr
!= end
);
426 if (end
- 1 > ceiling
- 1)
429 pmd
= pmd_offset(pud
, start
);
431 pmd_free_tlb(tlb
, pmd
, start
);
432 mm_dec_nr_pmds(tlb
->mm
);
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
694 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
696 vma
->vm_ops
? vma
->vm_ops
->fault
: NULL
,
697 vma
->vm_file
? vma
->vm_file
->f_op
->mmap
: NULL
,
698 mapping
? mapping
->a_ops
->readpage
: NULL
);
700 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
704 * vm_normal_page -- This function gets the "struct page" associated with a pte.
706 * "Special" mappings do not wish to be associated with a "struct page" (either
707 * it doesn't exist, or it exists but they don't want to touch it). In this
708 * case, NULL is returned here. "Normal" mappings do have a struct page.
710 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
711 * pte bit, in which case this function is trivial. Secondly, an architecture
712 * may not have a spare pte bit, which requires a more complicated scheme,
715 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
716 * special mapping (even if there are underlying and valid "struct pages").
717 * COWed pages of a VM_PFNMAP are always normal.
719 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
720 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
721 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
722 * mapping will always honor the rule
724 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
726 * And for normal mappings this is false.
728 * This restricts such mappings to be a linear translation from virtual address
729 * to pfn. To get around this restriction, we allow arbitrary mappings so long
730 * as the vma is not a COW mapping; in that case, we know that all ptes are
731 * special (because none can have been COWed).
734 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
736 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
737 * page" backing, however the difference is that _all_ pages with a struct
738 * page (that is, those where pfn_valid is true) are refcounted and considered
739 * normal pages by the VM. The disadvantage is that pages are refcounted
740 * (which can be slower and simply not an option for some PFNMAP users). The
741 * advantage is that we don't have to follow the strict linearity rule of
742 * PFNMAP mappings in order to support COWable mappings.
745 #ifdef __HAVE_ARCH_PTE_SPECIAL
746 # define HAVE_PTE_SPECIAL 1
748 # define HAVE_PTE_SPECIAL 0
750 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
753 unsigned long pfn
= pte_pfn(pte
);
755 if (HAVE_PTE_SPECIAL
) {
756 if (likely(!pte_special(pte
)))
758 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
759 return vma
->vm_ops
->find_special_page(vma
, addr
);
760 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
762 if (!is_zero_pfn(pfn
))
763 print_bad_pte(vma
, addr
, pte
, NULL
);
767 /* !HAVE_PTE_SPECIAL case follows: */
769 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
770 if (vma
->vm_flags
& VM_MIXEDMAP
) {
776 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
777 if (pfn
== vma
->vm_pgoff
+ off
)
779 if (!is_cow_mapping(vma
->vm_flags
))
784 if (is_zero_pfn(pfn
))
787 if (unlikely(pfn
> highest_memmap_pfn
)) {
788 print_bad_pte(vma
, addr
, pte
, NULL
);
793 * NOTE! We still have PageReserved() pages in the page tables.
794 * eg. VDSO mappings can cause them to exist.
797 return pfn_to_page(pfn
);
801 * copy one vm_area from one task to the other. Assumes the page tables
802 * already present in the new task to be cleared in the whole range
803 * covered by this vma.
806 static inline unsigned long
807 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
808 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
809 unsigned long addr
, int *rss
)
811 unsigned long vm_flags
= vma
->vm_flags
;
812 pte_t pte
= *src_pte
;
815 /* pte contains position in swap or file, so copy. */
816 if (unlikely(!pte_present(pte
))) {
817 swp_entry_t entry
= pte_to_swp_entry(pte
);
819 if (likely(!non_swap_entry(entry
))) {
820 if (swap_duplicate(entry
) < 0)
823 /* make sure dst_mm is on swapoff's mmlist. */
824 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
825 spin_lock(&mmlist_lock
);
826 if (list_empty(&dst_mm
->mmlist
))
827 list_add(&dst_mm
->mmlist
,
829 spin_unlock(&mmlist_lock
);
832 } else if (is_migration_entry(entry
)) {
833 page
= migration_entry_to_page(entry
);
835 rss
[mm_counter(page
)]++;
837 if (is_write_migration_entry(entry
) &&
838 is_cow_mapping(vm_flags
)) {
840 * COW mappings require pages in both
841 * parent and child to be set to read.
843 make_migration_entry_read(&entry
);
844 pte
= swp_entry_to_pte(entry
);
845 if (pte_swp_soft_dirty(*src_pte
))
846 pte
= pte_swp_mksoft_dirty(pte
);
847 set_pte_at(src_mm
, addr
, src_pte
, pte
);
854 * If it's a COW mapping, write protect it both
855 * in the parent and the child
857 if (is_cow_mapping(vm_flags
)) {
858 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
859 pte
= pte_wrprotect(pte
);
863 * If it's a shared mapping, mark it clean in
866 if (vm_flags
& VM_SHARED
)
867 pte
= pte_mkclean(pte
);
868 pte
= pte_mkold(pte
);
870 page
= vm_normal_page(vma
, addr
, pte
);
874 rss
[mm_counter(page
)]++;
878 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
882 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
883 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
884 unsigned long addr
, unsigned long end
)
886 pte_t
*orig_src_pte
, *orig_dst_pte
;
887 pte_t
*src_pte
, *dst_pte
;
888 spinlock_t
*src_ptl
, *dst_ptl
;
890 int rss
[NR_MM_COUNTERS
];
891 swp_entry_t entry
= (swp_entry_t
){0};
896 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
899 src_pte
= pte_offset_map(src_pmd
, addr
);
900 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
901 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
902 orig_src_pte
= src_pte
;
903 orig_dst_pte
= dst_pte
;
904 arch_enter_lazy_mmu_mode();
908 * We are holding two locks at this point - either of them
909 * could generate latencies in another task on another CPU.
911 if (progress
>= 32) {
913 if (need_resched() ||
914 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
917 if (pte_none(*src_pte
)) {
921 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
926 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
928 arch_leave_lazy_mmu_mode();
929 spin_unlock(src_ptl
);
930 pte_unmap(orig_src_pte
);
931 add_mm_rss_vec(dst_mm
, rss
);
932 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
936 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
945 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
946 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
947 unsigned long addr
, unsigned long end
)
949 pmd_t
*src_pmd
, *dst_pmd
;
952 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
955 src_pmd
= pmd_offset(src_pud
, addr
);
957 next
= pmd_addr_end(addr
, end
);
958 if (pmd_trans_huge(*src_pmd
)) {
960 VM_BUG_ON(next
-addr
!= HPAGE_PMD_SIZE
);
961 err
= copy_huge_pmd(dst_mm
, src_mm
,
962 dst_pmd
, src_pmd
, addr
, vma
);
969 if (pmd_none_or_clear_bad(src_pmd
))
971 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
974 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
978 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
979 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
980 unsigned long addr
, unsigned long end
)
982 pud_t
*src_pud
, *dst_pud
;
985 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
988 src_pud
= pud_offset(src_pgd
, addr
);
990 next
= pud_addr_end(addr
, end
);
991 if (pud_none_or_clear_bad(src_pud
))
993 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
996 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1000 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1001 struct vm_area_struct
*vma
)
1003 pgd_t
*src_pgd
, *dst_pgd
;
1005 unsigned long addr
= vma
->vm_start
;
1006 unsigned long end
= vma
->vm_end
;
1007 unsigned long mmun_start
; /* For mmu_notifiers */
1008 unsigned long mmun_end
; /* For mmu_notifiers */
1013 * Don't copy ptes where a page fault will fill them correctly.
1014 * Fork becomes much lighter when there are big shared or private
1015 * readonly mappings. The tradeoff is that copy_page_range is more
1016 * efficient than faulting.
1018 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_PFNMAP
| VM_MIXEDMAP
)) &&
1022 if (is_vm_hugetlb_page(vma
))
1023 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1025 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
1027 * We do not free on error cases below as remove_vma
1028 * gets called on error from higher level routine
1030 ret
= track_pfn_copy(vma
);
1036 * We need to invalidate the secondary MMU mappings only when
1037 * there could be a permission downgrade on the ptes of the
1038 * parent mm. And a permission downgrade will only happen if
1039 * is_cow_mapping() returns true.
1041 is_cow
= is_cow_mapping(vma
->vm_flags
);
1045 mmu_notifier_invalidate_range_start(src_mm
, mmun_start
,
1049 dst_pgd
= pgd_offset(dst_mm
, addr
);
1050 src_pgd
= pgd_offset(src_mm
, addr
);
1052 next
= pgd_addr_end(addr
, end
);
1053 if (pgd_none_or_clear_bad(src_pgd
))
1055 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1056 vma
, addr
, next
))) {
1060 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1063 mmu_notifier_invalidate_range_end(src_mm
, mmun_start
, mmun_end
);
1067 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1068 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1069 unsigned long addr
, unsigned long end
,
1070 struct zap_details
*details
)
1072 struct mm_struct
*mm
= tlb
->mm
;
1073 int force_flush
= 0;
1074 int rss
[NR_MM_COUNTERS
];
1082 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1084 arch_enter_lazy_mmu_mode();
1087 if (pte_none(ptent
)) {
1091 if (pte_present(ptent
)) {
1094 page
= vm_normal_page(vma
, addr
, ptent
);
1095 if (unlikely(details
) && page
) {
1097 * unmap_shared_mapping_pages() wants to
1098 * invalidate cache without truncating:
1099 * unmap shared but keep private pages.
1101 if (details
->check_mapping
&&
1102 details
->check_mapping
!= page
->mapping
)
1105 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1107 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1108 if (unlikely(!page
))
1111 if (!PageAnon(page
)) {
1112 if (pte_dirty(ptent
)) {
1114 set_page_dirty(page
);
1116 if (pte_young(ptent
) &&
1117 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1118 mark_page_accessed(page
);
1120 rss
[mm_counter(page
)]--;
1121 page_remove_rmap(page
);
1122 if (unlikely(page_mapcount(page
) < 0))
1123 print_bad_pte(vma
, addr
, ptent
, page
);
1124 if (unlikely(!__tlb_remove_page(tlb
, page
))) {
1131 /* If details->check_mapping, we leave swap entries. */
1132 if (unlikely(details
))
1135 entry
= pte_to_swp_entry(ptent
);
1136 if (!non_swap_entry(entry
))
1138 else if (is_migration_entry(entry
)) {
1141 page
= migration_entry_to_page(entry
);
1142 rss
[mm_counter(page
)]--;
1144 if (unlikely(!free_swap_and_cache(entry
)))
1145 print_bad_pte(vma
, addr
, ptent
, NULL
);
1146 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1147 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1149 add_mm_rss_vec(mm
, rss
);
1150 arch_leave_lazy_mmu_mode();
1152 /* Do the actual TLB flush before dropping ptl */
1154 tlb_flush_mmu_tlbonly(tlb
);
1155 pte_unmap_unlock(start_pte
, ptl
);
1158 * If we forced a TLB flush (either due to running out of
1159 * batch buffers or because we needed to flush dirty TLB
1160 * entries before releasing the ptl), free the batched
1161 * memory too. Restart if we didn't do everything.
1165 tlb_flush_mmu_free(tlb
);
1174 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1175 struct vm_area_struct
*vma
, pud_t
*pud
,
1176 unsigned long addr
, unsigned long end
,
1177 struct zap_details
*details
)
1182 pmd
= pmd_offset(pud
, addr
);
1184 next
= pmd_addr_end(addr
, end
);
1185 if (pmd_trans_huge(*pmd
)) {
1186 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1187 #ifdef CONFIG_DEBUG_VM
1188 if (!rwsem_is_locked(&tlb
->mm
->mmap_sem
)) {
1189 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1190 __func__
, addr
, end
,
1196 split_huge_page_pmd(vma
, addr
, pmd
);
1197 } else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1202 * Here there can be other concurrent MADV_DONTNEED or
1203 * trans huge page faults running, and if the pmd is
1204 * none or trans huge it can change under us. This is
1205 * because MADV_DONTNEED holds the mmap_sem in read
1208 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1210 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1213 } while (pmd
++, addr
= next
, addr
!= end
);
1218 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1219 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1220 unsigned long addr
, unsigned long end
,
1221 struct zap_details
*details
)
1226 pud
= pud_offset(pgd
, addr
);
1228 next
= pud_addr_end(addr
, end
);
1229 if (pud_none_or_clear_bad(pud
))
1231 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1232 } while (pud
++, addr
= next
, addr
!= end
);
1237 static void unmap_page_range(struct mmu_gather
*tlb
,
1238 struct vm_area_struct
*vma
,
1239 unsigned long addr
, unsigned long end
,
1240 struct zap_details
*details
)
1245 if (details
&& !details
->check_mapping
)
1248 BUG_ON(addr
>= end
);
1249 tlb_start_vma(tlb
, vma
);
1250 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1252 next
= pgd_addr_end(addr
, end
);
1253 if (pgd_none_or_clear_bad(pgd
))
1255 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
1256 } while (pgd
++, addr
= next
, addr
!= end
);
1257 tlb_end_vma(tlb
, vma
);
1261 static void unmap_single_vma(struct mmu_gather
*tlb
,
1262 struct vm_area_struct
*vma
, unsigned long start_addr
,
1263 unsigned long end_addr
,
1264 struct zap_details
*details
)
1266 unsigned long start
= max(vma
->vm_start
, start_addr
);
1269 if (start
>= vma
->vm_end
)
1271 end
= min(vma
->vm_end
, end_addr
);
1272 if (end
<= vma
->vm_start
)
1276 uprobe_munmap(vma
, start
, end
);
1278 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1279 untrack_pfn(vma
, 0, 0);
1282 if (unlikely(is_vm_hugetlb_page(vma
))) {
1284 * It is undesirable to test vma->vm_file as it
1285 * should be non-null for valid hugetlb area.
1286 * However, vm_file will be NULL in the error
1287 * cleanup path of mmap_region. When
1288 * hugetlbfs ->mmap method fails,
1289 * mmap_region() nullifies vma->vm_file
1290 * before calling this function to clean up.
1291 * Since no pte has actually been setup, it is
1292 * safe to do nothing in this case.
1295 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
1296 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1297 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
1300 unmap_page_range(tlb
, vma
, start
, end
, details
);
1305 * unmap_vmas - unmap a range of memory covered by a list of vma's
1306 * @tlb: address of the caller's struct mmu_gather
1307 * @vma: the starting vma
1308 * @start_addr: virtual address at which to start unmapping
1309 * @end_addr: virtual address at which to end unmapping
1311 * Unmap all pages in the vma list.
1313 * Only addresses between `start' and `end' will be unmapped.
1315 * The VMA list must be sorted in ascending virtual address order.
1317 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1318 * range after unmap_vmas() returns. So the only responsibility here is to
1319 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1320 * drops the lock and schedules.
1322 void unmap_vmas(struct mmu_gather
*tlb
,
1323 struct vm_area_struct
*vma
, unsigned long start_addr
,
1324 unsigned long end_addr
)
1326 struct mm_struct
*mm
= vma
->vm_mm
;
1328 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1329 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1330 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1331 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1335 * zap_page_range - remove user pages in a given range
1336 * @vma: vm_area_struct holding the applicable pages
1337 * @start: starting address of pages to zap
1338 * @size: number of bytes to zap
1339 * @details: details of shared cache invalidation
1341 * Caller must protect the VMA list
1343 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1344 unsigned long size
, struct zap_details
*details
)
1346 struct mm_struct
*mm
= vma
->vm_mm
;
1347 struct mmu_gather tlb
;
1348 unsigned long end
= start
+ size
;
1351 tlb_gather_mmu(&tlb
, mm
, start
, end
);
1352 update_hiwater_rss(mm
);
1353 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1354 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
)
1355 unmap_single_vma(&tlb
, vma
, start
, end
, details
);
1356 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1357 tlb_finish_mmu(&tlb
, start
, end
);
1361 * zap_page_range_single - remove user pages in a given range
1362 * @vma: vm_area_struct holding the applicable pages
1363 * @address: starting address of pages to zap
1364 * @size: number of bytes to zap
1365 * @details: details of shared cache invalidation
1367 * The range must fit into one VMA.
1369 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1370 unsigned long size
, struct zap_details
*details
)
1372 struct mm_struct
*mm
= vma
->vm_mm
;
1373 struct mmu_gather tlb
;
1374 unsigned long end
= address
+ size
;
1377 tlb_gather_mmu(&tlb
, mm
, address
, end
);
1378 update_hiwater_rss(mm
);
1379 mmu_notifier_invalidate_range_start(mm
, address
, end
);
1380 unmap_single_vma(&tlb
, vma
, address
, end
, details
);
1381 mmu_notifier_invalidate_range_end(mm
, address
, end
);
1382 tlb_finish_mmu(&tlb
, address
, end
);
1386 * zap_vma_ptes - remove ptes mapping the vma
1387 * @vma: vm_area_struct holding ptes to be zapped
1388 * @address: starting address of pages to zap
1389 * @size: number of bytes to zap
1391 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1393 * The entire address range must be fully contained within the vma.
1395 * Returns 0 if successful.
1397 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1400 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1401 !(vma
->vm_flags
& VM_PFNMAP
))
1403 zap_page_range_single(vma
, address
, size
, NULL
);
1406 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1408 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1411 pgd_t
* pgd
= pgd_offset(mm
, addr
);
1412 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
1414 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
1416 VM_BUG_ON(pmd_trans_huge(*pmd
));
1417 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1424 * This is the old fallback for page remapping.
1426 * For historical reasons, it only allows reserved pages. Only
1427 * old drivers should use this, and they needed to mark their
1428 * pages reserved for the old functions anyway.
1430 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1431 struct page
*page
, pgprot_t prot
)
1433 struct mm_struct
*mm
= vma
->vm_mm
;
1442 flush_dcache_page(page
);
1443 pte
= get_locked_pte(mm
, addr
, &ptl
);
1447 if (!pte_none(*pte
))
1450 /* Ok, finally just insert the thing.. */
1452 inc_mm_counter_fast(mm
, mm_counter_file(page
));
1453 page_add_file_rmap(page
);
1454 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1457 pte_unmap_unlock(pte
, ptl
);
1460 pte_unmap_unlock(pte
, ptl
);
1466 * vm_insert_page - insert single page into user vma
1467 * @vma: user vma to map to
1468 * @addr: target user address of this page
1469 * @page: source kernel page
1471 * This allows drivers to insert individual pages they've allocated
1474 * The page has to be a nice clean _individual_ kernel allocation.
1475 * If you allocate a compound page, you need to have marked it as
1476 * such (__GFP_COMP), or manually just split the page up yourself
1477 * (see split_page()).
1479 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1480 * took an arbitrary page protection parameter. This doesn't allow
1481 * that. Your vma protection will have to be set up correctly, which
1482 * means that if you want a shared writable mapping, you'd better
1483 * ask for a shared writable mapping!
1485 * The page does not need to be reserved.
1487 * Usually this function is called from f_op->mmap() handler
1488 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1489 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1490 * function from other places, for example from page-fault handler.
1492 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1495 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1497 if (!page_count(page
))
1499 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1500 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
1501 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1502 vma
->vm_flags
|= VM_MIXEDMAP
;
1504 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1506 EXPORT_SYMBOL(vm_insert_page
);
1508 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1509 unsigned long pfn
, pgprot_t prot
)
1511 struct mm_struct
*mm
= vma
->vm_mm
;
1517 pte
= get_locked_pte(mm
, addr
, &ptl
);
1521 if (!pte_none(*pte
))
1524 /* Ok, finally just insert the thing.. */
1525 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
1526 set_pte_at(mm
, addr
, pte
, entry
);
1527 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1531 pte_unmap_unlock(pte
, ptl
);
1537 * vm_insert_pfn - insert single pfn into user vma
1538 * @vma: user vma to map to
1539 * @addr: target user address of this page
1540 * @pfn: source kernel pfn
1542 * Similar to vm_insert_page, this allows drivers to insert individual pages
1543 * they've allocated into a user vma. Same comments apply.
1545 * This function should only be called from a vm_ops->fault handler, and
1546 * in that case the handler should return NULL.
1548 * vma cannot be a COW mapping.
1550 * As this is called only for pages that do not currently exist, we
1551 * do not need to flush old virtual caches or the TLB.
1553 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1557 pgprot_t pgprot
= vma
->vm_page_prot
;
1559 * Technically, architectures with pte_special can avoid all these
1560 * restrictions (same for remap_pfn_range). However we would like
1561 * consistency in testing and feature parity among all, so we should
1562 * try to keep these invariants in place for everybody.
1564 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1565 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1566 (VM_PFNMAP
|VM_MIXEDMAP
));
1567 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1568 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1570 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1572 if (track_pfn_insert(vma
, &pgprot
, pfn
))
1575 ret
= insert_pfn(vma
, addr
, pfn
, pgprot
);
1579 EXPORT_SYMBOL(vm_insert_pfn
);
1581 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1584 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
1586 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1590 * If we don't have pte special, then we have to use the pfn_valid()
1591 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1592 * refcount the page if pfn_valid is true (hence insert_page rather
1593 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1594 * without pte special, it would there be refcounted as a normal page.
1596 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
1599 page
= pfn_to_page(pfn
);
1600 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1602 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1604 EXPORT_SYMBOL(vm_insert_mixed
);
1607 * maps a range of physical memory into the requested pages. the old
1608 * mappings are removed. any references to nonexistent pages results
1609 * in null mappings (currently treated as "copy-on-access")
1611 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1612 unsigned long addr
, unsigned long end
,
1613 unsigned long pfn
, pgprot_t prot
)
1618 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1621 arch_enter_lazy_mmu_mode();
1623 BUG_ON(!pte_none(*pte
));
1624 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
1626 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1627 arch_leave_lazy_mmu_mode();
1628 pte_unmap_unlock(pte
- 1, ptl
);
1632 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1633 unsigned long addr
, unsigned long end
,
1634 unsigned long pfn
, pgprot_t prot
)
1639 pfn
-= addr
>> PAGE_SHIFT
;
1640 pmd
= pmd_alloc(mm
, pud
, addr
);
1643 VM_BUG_ON(pmd_trans_huge(*pmd
));
1645 next
= pmd_addr_end(addr
, end
);
1646 if (remap_pte_range(mm
, pmd
, addr
, next
,
1647 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1649 } while (pmd
++, addr
= next
, addr
!= end
);
1653 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1654 unsigned long addr
, unsigned long end
,
1655 unsigned long pfn
, pgprot_t prot
)
1660 pfn
-= addr
>> PAGE_SHIFT
;
1661 pud
= pud_alloc(mm
, pgd
, addr
);
1665 next
= pud_addr_end(addr
, end
);
1666 if (remap_pmd_range(mm
, pud
, addr
, next
,
1667 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1669 } while (pud
++, addr
= next
, addr
!= end
);
1674 * remap_pfn_range - remap kernel memory to userspace
1675 * @vma: user vma to map to
1676 * @addr: target user address to start at
1677 * @pfn: physical address of kernel memory
1678 * @size: size of map area
1679 * @prot: page protection flags for this mapping
1681 * Note: this is only safe if the mm semaphore is held when called.
1683 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1684 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1688 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1689 struct mm_struct
*mm
= vma
->vm_mm
;
1693 * Physically remapped pages are special. Tell the
1694 * rest of the world about it:
1695 * VM_IO tells people not to look at these pages
1696 * (accesses can have side effects).
1697 * VM_PFNMAP tells the core MM that the base pages are just
1698 * raw PFN mappings, and do not have a "struct page" associated
1701 * Disable vma merging and expanding with mremap().
1703 * Omit vma from core dump, even when VM_IO turned off.
1705 * There's a horrible special case to handle copy-on-write
1706 * behaviour that some programs depend on. We mark the "original"
1707 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1708 * See vm_normal_page() for details.
1710 if (is_cow_mapping(vma
->vm_flags
)) {
1711 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
1713 vma
->vm_pgoff
= pfn
;
1716 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
1720 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
1722 BUG_ON(addr
>= end
);
1723 pfn
-= addr
>> PAGE_SHIFT
;
1724 pgd
= pgd_offset(mm
, addr
);
1725 flush_cache_range(vma
, addr
, end
);
1727 next
= pgd_addr_end(addr
, end
);
1728 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1729 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1732 } while (pgd
++, addr
= next
, addr
!= end
);
1735 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
));
1739 EXPORT_SYMBOL(remap_pfn_range
);
1742 * vm_iomap_memory - remap memory to userspace
1743 * @vma: user vma to map to
1744 * @start: start of area
1745 * @len: size of area
1747 * This is a simplified io_remap_pfn_range() for common driver use. The
1748 * driver just needs to give us the physical memory range to be mapped,
1749 * we'll figure out the rest from the vma information.
1751 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1752 * whatever write-combining details or similar.
1754 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
1756 unsigned long vm_len
, pfn
, pages
;
1758 /* Check that the physical memory area passed in looks valid */
1759 if (start
+ len
< start
)
1762 * You *really* shouldn't map things that aren't page-aligned,
1763 * but we've historically allowed it because IO memory might
1764 * just have smaller alignment.
1766 len
+= start
& ~PAGE_MASK
;
1767 pfn
= start
>> PAGE_SHIFT
;
1768 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
1769 if (pfn
+ pages
< pfn
)
1772 /* We start the mapping 'vm_pgoff' pages into the area */
1773 if (vma
->vm_pgoff
> pages
)
1775 pfn
+= vma
->vm_pgoff
;
1776 pages
-= vma
->vm_pgoff
;
1778 /* Can we fit all of the mapping? */
1779 vm_len
= vma
->vm_end
- vma
->vm_start
;
1780 if (vm_len
>> PAGE_SHIFT
> pages
)
1783 /* Ok, let it rip */
1784 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
1786 EXPORT_SYMBOL(vm_iomap_memory
);
1788 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1789 unsigned long addr
, unsigned long end
,
1790 pte_fn_t fn
, void *data
)
1795 spinlock_t
*uninitialized_var(ptl
);
1797 pte
= (mm
== &init_mm
) ?
1798 pte_alloc_kernel(pmd
, addr
) :
1799 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1803 BUG_ON(pmd_huge(*pmd
));
1805 arch_enter_lazy_mmu_mode();
1807 token
= pmd_pgtable(*pmd
);
1810 err
= fn(pte
++, token
, addr
, data
);
1813 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1815 arch_leave_lazy_mmu_mode();
1818 pte_unmap_unlock(pte
-1, ptl
);
1822 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1823 unsigned long addr
, unsigned long end
,
1824 pte_fn_t fn
, void *data
)
1830 BUG_ON(pud_huge(*pud
));
1832 pmd
= pmd_alloc(mm
, pud
, addr
);
1836 next
= pmd_addr_end(addr
, end
);
1837 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
1840 } while (pmd
++, addr
= next
, addr
!= end
);
1844 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1845 unsigned long addr
, unsigned long end
,
1846 pte_fn_t fn
, void *data
)
1852 pud
= pud_alloc(mm
, pgd
, addr
);
1856 next
= pud_addr_end(addr
, end
);
1857 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
1860 } while (pud
++, addr
= next
, addr
!= end
);
1865 * Scan a region of virtual memory, filling in page tables as necessary
1866 * and calling a provided function on each leaf page table.
1868 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
1869 unsigned long size
, pte_fn_t fn
, void *data
)
1873 unsigned long end
= addr
+ size
;
1876 BUG_ON(addr
>= end
);
1877 pgd
= pgd_offset(mm
, addr
);
1879 next
= pgd_addr_end(addr
, end
);
1880 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
1883 } while (pgd
++, addr
= next
, addr
!= end
);
1887 EXPORT_SYMBOL_GPL(apply_to_page_range
);
1890 * handle_pte_fault chooses page fault handler according to an entry which was
1891 * read non-atomically. Before making any commitment, on those architectures
1892 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1893 * parts, do_swap_page must check under lock before unmapping the pte and
1894 * proceeding (but do_wp_page is only called after already making such a check;
1895 * and do_anonymous_page can safely check later on).
1897 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
1898 pte_t
*page_table
, pte_t orig_pte
)
1901 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1902 if (sizeof(pte_t
) > sizeof(unsigned long)) {
1903 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
1905 same
= pte_same(*page_table
, orig_pte
);
1909 pte_unmap(page_table
);
1913 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
1915 debug_dma_assert_idle(src
);
1918 * If the source page was a PFN mapping, we don't have
1919 * a "struct page" for it. We do a best-effort copy by
1920 * just copying from the original user address. If that
1921 * fails, we just zero-fill it. Live with it.
1923 if (unlikely(!src
)) {
1924 void *kaddr
= kmap_atomic(dst
);
1925 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
1928 * This really shouldn't fail, because the page is there
1929 * in the page tables. But it might just be unreadable,
1930 * in which case we just give up and fill the result with
1933 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
1935 kunmap_atomic(kaddr
);
1936 flush_dcache_page(dst
);
1938 copy_user_highpage(dst
, src
, va
, vma
);
1941 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
1943 struct file
*vm_file
= vma
->vm_file
;
1946 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
1949 * Special mappings (e.g. VDSO) do not have any file so fake
1950 * a default GFP_KERNEL for them.
1956 * Notify the address space that the page is about to become writable so that
1957 * it can prohibit this or wait for the page to get into an appropriate state.
1959 * We do this without the lock held, so that it can sleep if it needs to.
1961 static int do_page_mkwrite(struct vm_area_struct
*vma
, struct page
*page
,
1962 unsigned long address
)
1964 struct vm_fault vmf
;
1967 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
1968 vmf
.pgoff
= page
->index
;
1969 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
1970 vmf
.gfp_mask
= __get_fault_gfp_mask(vma
);
1972 vmf
.cow_page
= NULL
;
1974 ret
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
1975 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
1977 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
1979 if (!page
->mapping
) {
1981 return 0; /* retry */
1983 ret
|= VM_FAULT_LOCKED
;
1985 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1990 * Handle write page faults for pages that can be reused in the current vma
1992 * This can happen either due to the mapping being with the VM_SHARED flag,
1993 * or due to us being the last reference standing to the page. In either
1994 * case, all we need to do here is to mark the page as writable and update
1995 * any related book-keeping.
1997 static inline int wp_page_reuse(struct mm_struct
*mm
,
1998 struct vm_area_struct
*vma
, unsigned long address
,
1999 pte_t
*page_table
, spinlock_t
*ptl
, pte_t orig_pte
,
2000 struct page
*page
, int page_mkwrite
,
2006 * Clear the pages cpupid information as the existing
2007 * information potentially belongs to a now completely
2008 * unrelated process.
2011 page_cpupid_xchg_last(page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2013 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2014 entry
= pte_mkyoung(orig_pte
);
2015 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2016 if (ptep_set_access_flags(vma
, address
, page_table
, entry
, 1))
2017 update_mmu_cache(vma
, address
, page_table
);
2018 pte_unmap_unlock(page_table
, ptl
);
2021 struct address_space
*mapping
;
2027 dirtied
= set_page_dirty(page
);
2028 VM_BUG_ON_PAGE(PageAnon(page
), page
);
2029 mapping
= page
->mapping
;
2031 page_cache_release(page
);
2033 if ((dirtied
|| page_mkwrite
) && mapping
) {
2035 * Some device drivers do not set page.mapping
2036 * but still dirty their pages
2038 balance_dirty_pages_ratelimited(mapping
);
2042 file_update_time(vma
->vm_file
);
2045 return VM_FAULT_WRITE
;
2049 * Handle the case of a page which we actually need to copy to a new page.
2051 * Called with mmap_sem locked and the old page referenced, but
2052 * without the ptl held.
2054 * High level logic flow:
2056 * - Allocate a page, copy the content of the old page to the new one.
2057 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2058 * - Take the PTL. If the pte changed, bail out and release the allocated page
2059 * - If the pte is still the way we remember it, update the page table and all
2060 * relevant references. This includes dropping the reference the page-table
2061 * held to the old page, as well as updating the rmap.
2062 * - In any case, unlock the PTL and drop the reference we took to the old page.
2064 static int wp_page_copy(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2065 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2066 pte_t orig_pte
, struct page
*old_page
)
2068 struct page
*new_page
= NULL
;
2069 spinlock_t
*ptl
= NULL
;
2071 int page_copied
= 0;
2072 const unsigned long mmun_start
= address
& PAGE_MASK
; /* For mmu_notifiers */
2073 const unsigned long mmun_end
= mmun_start
+ PAGE_SIZE
; /* For mmu_notifiers */
2074 struct mem_cgroup
*memcg
;
2076 if (unlikely(anon_vma_prepare(vma
)))
2079 if (is_zero_pfn(pte_pfn(orig_pte
))) {
2080 new_page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2084 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2087 cow_user_page(new_page
, old_page
, address
, vma
);
2090 if (mem_cgroup_try_charge(new_page
, mm
, GFP_KERNEL
, &memcg
))
2093 __SetPageUptodate(new_page
);
2095 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2098 * Re-check the pte - we dropped the lock
2100 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2101 if (likely(pte_same(*page_table
, orig_pte
))) {
2103 if (!PageAnon(old_page
)) {
2104 dec_mm_counter_fast(mm
,
2105 mm_counter_file(old_page
));
2106 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2109 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2111 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2112 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2113 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2115 * Clear the pte entry and flush it first, before updating the
2116 * pte with the new entry. This will avoid a race condition
2117 * seen in the presence of one thread doing SMC and another
2120 ptep_clear_flush_notify(vma
, address
, page_table
);
2121 page_add_new_anon_rmap(new_page
, vma
, address
);
2122 mem_cgroup_commit_charge(new_page
, memcg
, false);
2123 lru_cache_add_active_or_unevictable(new_page
, vma
);
2125 * We call the notify macro here because, when using secondary
2126 * mmu page tables (such as kvm shadow page tables), we want the
2127 * new page to be mapped directly into the secondary page table.
2129 set_pte_at_notify(mm
, address
, page_table
, entry
);
2130 update_mmu_cache(vma
, address
, page_table
);
2133 * Only after switching the pte to the new page may
2134 * we remove the mapcount here. Otherwise another
2135 * process may come and find the rmap count decremented
2136 * before the pte is switched to the new page, and
2137 * "reuse" the old page writing into it while our pte
2138 * here still points into it and can be read by other
2141 * The critical issue is to order this
2142 * page_remove_rmap with the ptp_clear_flush above.
2143 * Those stores are ordered by (if nothing else,)
2144 * the barrier present in the atomic_add_negative
2145 * in page_remove_rmap.
2147 * Then the TLB flush in ptep_clear_flush ensures that
2148 * no process can access the old page before the
2149 * decremented mapcount is visible. And the old page
2150 * cannot be reused until after the decremented
2151 * mapcount is visible. So transitively, TLBs to
2152 * old page will be flushed before it can be reused.
2154 page_remove_rmap(old_page
);
2157 /* Free the old page.. */
2158 new_page
= old_page
;
2161 mem_cgroup_cancel_charge(new_page
, memcg
);
2165 page_cache_release(new_page
);
2167 pte_unmap_unlock(page_table
, ptl
);
2168 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2171 * Don't let another task, with possibly unlocked vma,
2172 * keep the mlocked page.
2174 if (page_copied
&& (vma
->vm_flags
& VM_LOCKED
)) {
2175 lock_page(old_page
); /* LRU manipulation */
2176 munlock_vma_page(old_page
);
2177 unlock_page(old_page
);
2179 page_cache_release(old_page
);
2181 return page_copied
? VM_FAULT_WRITE
: 0;
2183 page_cache_release(new_page
);
2186 page_cache_release(old_page
);
2187 return VM_FAULT_OOM
;
2191 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2194 static int wp_pfn_shared(struct mm_struct
*mm
,
2195 struct vm_area_struct
*vma
, unsigned long address
,
2196 pte_t
*page_table
, spinlock_t
*ptl
, pte_t orig_pte
,
2199 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
2200 struct vm_fault vmf
= {
2202 .pgoff
= linear_page_index(vma
, address
),
2203 .virtual_address
= (void __user
*)(address
& PAGE_MASK
),
2204 .flags
= FAULT_FLAG_WRITE
| FAULT_FLAG_MKWRITE
,
2208 pte_unmap_unlock(page_table
, ptl
);
2209 ret
= vma
->vm_ops
->pfn_mkwrite(vma
, &vmf
);
2210 if (ret
& VM_FAULT_ERROR
)
2212 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2214 * We might have raced with another page fault while we
2215 * released the pte_offset_map_lock.
2217 if (!pte_same(*page_table
, orig_pte
)) {
2218 pte_unmap_unlock(page_table
, ptl
);
2222 return wp_page_reuse(mm
, vma
, address
, page_table
, ptl
, orig_pte
,
2226 static int wp_page_shared(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2227 unsigned long address
, pte_t
*page_table
,
2228 pmd_t
*pmd
, spinlock_t
*ptl
, pte_t orig_pte
,
2229 struct page
*old_page
)
2232 int page_mkwrite
= 0;
2234 page_cache_get(old_page
);
2237 * Only catch write-faults on shared writable pages,
2238 * read-only shared pages can get COWed by
2239 * get_user_pages(.write=1, .force=1).
2241 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2244 pte_unmap_unlock(page_table
, ptl
);
2245 tmp
= do_page_mkwrite(vma
, old_page
, address
);
2246 if (unlikely(!tmp
|| (tmp
&
2247 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
2248 page_cache_release(old_page
);
2252 * Since we dropped the lock we need to revalidate
2253 * the PTE as someone else may have changed it. If
2254 * they did, we just return, as we can count on the
2255 * MMU to tell us if they didn't also make it writable.
2257 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2259 if (!pte_same(*page_table
, orig_pte
)) {
2260 unlock_page(old_page
);
2261 pte_unmap_unlock(page_table
, ptl
);
2262 page_cache_release(old_page
);
2268 return wp_page_reuse(mm
, vma
, address
, page_table
, ptl
,
2269 orig_pte
, old_page
, page_mkwrite
, 1);
2273 * This routine handles present pages, when users try to write
2274 * to a shared page. It is done by copying the page to a new address
2275 * and decrementing the shared-page counter for the old page.
2277 * Note that this routine assumes that the protection checks have been
2278 * done by the caller (the low-level page fault routine in most cases).
2279 * Thus we can safely just mark it writable once we've done any necessary
2282 * We also mark the page dirty at this point even though the page will
2283 * change only once the write actually happens. This avoids a few races,
2284 * and potentially makes it more efficient.
2286 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2287 * but allow concurrent faults), with pte both mapped and locked.
2288 * We return with mmap_sem still held, but pte unmapped and unlocked.
2290 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2291 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2292 spinlock_t
*ptl
, pte_t orig_pte
)
2295 struct page
*old_page
;
2297 old_page
= vm_normal_page(vma
, address
, orig_pte
);
2300 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2303 * We should not cow pages in a shared writeable mapping.
2304 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2306 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2307 (VM_WRITE
|VM_SHARED
))
2308 return wp_pfn_shared(mm
, vma
, address
, page_table
, ptl
,
2311 pte_unmap_unlock(page_table
, ptl
);
2312 return wp_page_copy(mm
, vma
, address
, page_table
, pmd
,
2313 orig_pte
, old_page
);
2317 * Take out anonymous pages first, anonymous shared vmas are
2318 * not dirty accountable.
2320 if (PageAnon(old_page
) && !PageKsm(old_page
)) {
2321 if (!trylock_page(old_page
)) {
2322 page_cache_get(old_page
);
2323 pte_unmap_unlock(page_table
, ptl
);
2324 lock_page(old_page
);
2325 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2327 if (!pte_same(*page_table
, orig_pte
)) {
2328 unlock_page(old_page
);
2329 pte_unmap_unlock(page_table
, ptl
);
2330 page_cache_release(old_page
);
2333 page_cache_release(old_page
);
2335 if (reuse_swap_page(old_page
)) {
2337 * The page is all ours. Move it to our anon_vma so
2338 * the rmap code will not search our parent or siblings.
2339 * Protected against the rmap code by the page lock.
2341 page_move_anon_rmap(old_page
, vma
, address
);
2342 unlock_page(old_page
);
2343 return wp_page_reuse(mm
, vma
, address
, page_table
, ptl
,
2344 orig_pte
, old_page
, 0, 0);
2346 unlock_page(old_page
);
2347 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2348 (VM_WRITE
|VM_SHARED
))) {
2349 return wp_page_shared(mm
, vma
, address
, page_table
, pmd
,
2350 ptl
, orig_pte
, old_page
);
2354 * Ok, we need to copy. Oh, well..
2356 page_cache_get(old_page
);
2358 pte_unmap_unlock(page_table
, ptl
);
2359 return wp_page_copy(mm
, vma
, address
, page_table
, pmd
,
2360 orig_pte
, old_page
);
2363 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2364 unsigned long start_addr
, unsigned long end_addr
,
2365 struct zap_details
*details
)
2367 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2370 static inline void unmap_mapping_range_tree(struct rb_root
*root
,
2371 struct zap_details
*details
)
2373 struct vm_area_struct
*vma
;
2374 pgoff_t vba
, vea
, zba
, zea
;
2376 vma_interval_tree_foreach(vma
, root
,
2377 details
->first_index
, details
->last_index
) {
2379 vba
= vma
->vm_pgoff
;
2380 vea
= vba
+ vma_pages(vma
) - 1;
2381 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2382 zba
= details
->first_index
;
2385 zea
= details
->last_index
;
2389 unmap_mapping_range_vma(vma
,
2390 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2391 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2397 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2398 * address_space corresponding to the specified page range in the underlying
2401 * @mapping: the address space containing mmaps to be unmapped.
2402 * @holebegin: byte in first page to unmap, relative to the start of
2403 * the underlying file. This will be rounded down to a PAGE_SIZE
2404 * boundary. Note that this is different from truncate_pagecache(), which
2405 * must keep the partial page. In contrast, we must get rid of
2407 * @holelen: size of prospective hole in bytes. This will be rounded
2408 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2410 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2411 * but 0 when invalidating pagecache, don't throw away private data.
2413 void unmap_mapping_range(struct address_space
*mapping
,
2414 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2416 struct zap_details details
;
2417 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2418 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2420 /* Check for overflow. */
2421 if (sizeof(holelen
) > sizeof(hlen
)) {
2423 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2424 if (holeend
& ~(long long)ULONG_MAX
)
2425 hlen
= ULONG_MAX
- hba
+ 1;
2428 details
.check_mapping
= even_cows
? NULL
: mapping
;
2429 details
.first_index
= hba
;
2430 details
.last_index
= hba
+ hlen
- 1;
2431 if (details
.last_index
< details
.first_index
)
2432 details
.last_index
= ULONG_MAX
;
2435 /* DAX uses i_mmap_lock to serialise file truncate vs page fault */
2436 i_mmap_lock_write(mapping
);
2437 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
)))
2438 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2439 i_mmap_unlock_write(mapping
);
2441 EXPORT_SYMBOL(unmap_mapping_range
);
2444 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2445 * but allow concurrent faults), and pte mapped but not yet locked.
2446 * We return with pte unmapped and unlocked.
2448 * We return with the mmap_sem locked or unlocked in the same cases
2449 * as does filemap_fault().
2451 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2452 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2453 unsigned int flags
, pte_t orig_pte
)
2456 struct page
*page
, *swapcache
;
2457 struct mem_cgroup
*memcg
;
2464 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2467 entry
= pte_to_swp_entry(orig_pte
);
2468 if (unlikely(non_swap_entry(entry
))) {
2469 if (is_migration_entry(entry
)) {
2470 migration_entry_wait(mm
, pmd
, address
);
2471 } else if (is_hwpoison_entry(entry
)) {
2472 ret
= VM_FAULT_HWPOISON
;
2474 print_bad_pte(vma
, address
, orig_pte
, NULL
);
2475 ret
= VM_FAULT_SIGBUS
;
2479 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2480 page
= lookup_swap_cache(entry
);
2482 page
= swapin_readahead(entry
,
2483 GFP_HIGHUSER_MOVABLE
, vma
, address
);
2486 * Back out if somebody else faulted in this pte
2487 * while we released the pte lock.
2489 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2490 if (likely(pte_same(*page_table
, orig_pte
)))
2492 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2496 /* Had to read the page from swap area: Major fault */
2497 ret
= VM_FAULT_MAJOR
;
2498 count_vm_event(PGMAJFAULT
);
2499 mem_cgroup_count_vm_event(mm
, PGMAJFAULT
);
2500 } else if (PageHWPoison(page
)) {
2502 * hwpoisoned dirty swapcache pages are kept for killing
2503 * owner processes (which may be unknown at hwpoison time)
2505 ret
= VM_FAULT_HWPOISON
;
2506 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2512 locked
= lock_page_or_retry(page
, mm
, flags
);
2514 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2516 ret
|= VM_FAULT_RETRY
;
2521 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2522 * release the swapcache from under us. The page pin, and pte_same
2523 * test below, are not enough to exclude that. Even if it is still
2524 * swapcache, we need to check that the page's swap has not changed.
2526 if (unlikely(!PageSwapCache(page
) || page_private(page
) != entry
.val
))
2529 page
= ksm_might_need_to_copy(page
, vma
, address
);
2530 if (unlikely(!page
)) {
2536 if (mem_cgroup_try_charge(page
, mm
, GFP_KERNEL
, &memcg
)) {
2542 * Back out if somebody else already faulted in this pte.
2544 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2545 if (unlikely(!pte_same(*page_table
, orig_pte
)))
2548 if (unlikely(!PageUptodate(page
))) {
2549 ret
= VM_FAULT_SIGBUS
;
2554 * The page isn't present yet, go ahead with the fault.
2556 * Be careful about the sequence of operations here.
2557 * To get its accounting right, reuse_swap_page() must be called
2558 * while the page is counted on swap but not yet in mapcount i.e.
2559 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2560 * must be called after the swap_free(), or it will never succeed.
2563 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2564 dec_mm_counter_fast(mm
, MM_SWAPENTS
);
2565 pte
= mk_pte(page
, vma
->vm_page_prot
);
2566 if ((flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
)) {
2567 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
2568 flags
&= ~FAULT_FLAG_WRITE
;
2569 ret
|= VM_FAULT_WRITE
;
2572 flush_icache_page(vma
, page
);
2573 if (pte_swp_soft_dirty(orig_pte
))
2574 pte
= pte_mksoft_dirty(pte
);
2575 set_pte_at(mm
, address
, page_table
, pte
);
2576 if (page
== swapcache
) {
2577 do_page_add_anon_rmap(page
, vma
, address
, exclusive
);
2578 mem_cgroup_commit_charge(page
, memcg
, true);
2579 } else { /* ksm created a completely new copy */
2580 page_add_new_anon_rmap(page
, vma
, address
);
2581 mem_cgroup_commit_charge(page
, memcg
, false);
2582 lru_cache_add_active_or_unevictable(page
, vma
);
2586 if (vm_swap_full() || (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
2587 try_to_free_swap(page
);
2589 if (page
!= swapcache
) {
2591 * Hold the lock to avoid the swap entry to be reused
2592 * until we take the PT lock for the pte_same() check
2593 * (to avoid false positives from pte_same). For
2594 * further safety release the lock after the swap_free
2595 * so that the swap count won't change under a
2596 * parallel locked swapcache.
2598 unlock_page(swapcache
);
2599 page_cache_release(swapcache
);
2602 if (flags
& FAULT_FLAG_WRITE
) {
2603 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
2604 if (ret
& VM_FAULT_ERROR
)
2605 ret
&= VM_FAULT_ERROR
;
2609 /* No need to invalidate - it was non-present before */
2610 update_mmu_cache(vma
, address
, page_table
);
2612 pte_unmap_unlock(page_table
, ptl
);
2616 mem_cgroup_cancel_charge(page
, memcg
);
2617 pte_unmap_unlock(page_table
, ptl
);
2621 page_cache_release(page
);
2622 if (page
!= swapcache
) {
2623 unlock_page(swapcache
);
2624 page_cache_release(swapcache
);
2630 * This is like a special single-page "expand_{down|up}wards()",
2631 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2632 * doesn't hit another vma.
2634 static inline int check_stack_guard_page(struct vm_area_struct
*vma
, unsigned long address
)
2636 address
&= PAGE_MASK
;
2637 if ((vma
->vm_flags
& VM_GROWSDOWN
) && address
== vma
->vm_start
) {
2638 struct vm_area_struct
*prev
= vma
->vm_prev
;
2641 * Is there a mapping abutting this one below?
2643 * That's only ok if it's the same stack mapping
2644 * that has gotten split..
2646 if (prev
&& prev
->vm_end
== address
)
2647 return prev
->vm_flags
& VM_GROWSDOWN
? 0 : -ENOMEM
;
2649 return expand_downwards(vma
, address
- PAGE_SIZE
);
2651 if ((vma
->vm_flags
& VM_GROWSUP
) && address
+ PAGE_SIZE
== vma
->vm_end
) {
2652 struct vm_area_struct
*next
= vma
->vm_next
;
2654 /* As VM_GROWSDOWN but s/below/above/ */
2655 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
)
2656 return next
->vm_flags
& VM_GROWSUP
? 0 : -ENOMEM
;
2658 return expand_upwards(vma
, address
+ PAGE_SIZE
);
2664 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2665 * but allow concurrent faults), and pte mapped but not yet locked.
2666 * We return with mmap_sem still held, but pte unmapped and unlocked.
2668 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2669 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2672 struct mem_cgroup
*memcg
;
2677 pte_unmap(page_table
);
2679 /* File mapping without ->vm_ops ? */
2680 if (vma
->vm_flags
& VM_SHARED
)
2681 return VM_FAULT_SIGBUS
;
2683 /* Check if we need to add a guard page to the stack */
2684 if (check_stack_guard_page(vma
, address
) < 0)
2685 return VM_FAULT_SIGSEGV
;
2687 /* Use the zero-page for reads */
2688 if (!(flags
& FAULT_FLAG_WRITE
) && !mm_forbids_zeropage(mm
)) {
2689 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(address
),
2690 vma
->vm_page_prot
));
2691 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2692 if (!pte_none(*page_table
))
2694 /* Deliver the page fault to userland, check inside PT lock */
2695 if (userfaultfd_missing(vma
)) {
2696 pte_unmap_unlock(page_table
, ptl
);
2697 return handle_userfault(vma
, address
, flags
,
2703 /* Allocate our own private page. */
2704 if (unlikely(anon_vma_prepare(vma
)))
2706 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2710 if (mem_cgroup_try_charge(page
, mm
, GFP_KERNEL
, &memcg
))
2714 * The memory barrier inside __SetPageUptodate makes sure that
2715 * preceeding stores to the page contents become visible before
2716 * the set_pte_at() write.
2718 __SetPageUptodate(page
);
2720 entry
= mk_pte(page
, vma
->vm_page_prot
);
2721 if (vma
->vm_flags
& VM_WRITE
)
2722 entry
= pte_mkwrite(pte_mkdirty(entry
));
2724 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2725 if (!pte_none(*page_table
))
2728 /* Deliver the page fault to userland, check inside PT lock */
2729 if (userfaultfd_missing(vma
)) {
2730 pte_unmap_unlock(page_table
, ptl
);
2731 mem_cgroup_cancel_charge(page
, memcg
);
2732 page_cache_release(page
);
2733 return handle_userfault(vma
, address
, flags
,
2737 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2738 page_add_new_anon_rmap(page
, vma
, address
);
2739 mem_cgroup_commit_charge(page
, memcg
, false);
2740 lru_cache_add_active_or_unevictable(page
, vma
);
2742 set_pte_at(mm
, address
, page_table
, entry
);
2744 /* No need to invalidate - it was non-present before */
2745 update_mmu_cache(vma
, address
, page_table
);
2747 pte_unmap_unlock(page_table
, ptl
);
2750 mem_cgroup_cancel_charge(page
, memcg
);
2751 page_cache_release(page
);
2754 page_cache_release(page
);
2756 return VM_FAULT_OOM
;
2760 * The mmap_sem must have been held on entry, and may have been
2761 * released depending on flags and vma->vm_ops->fault() return value.
2762 * See filemap_fault() and __lock_page_retry().
2764 static int __do_fault(struct vm_area_struct
*vma
, unsigned long address
,
2765 pgoff_t pgoff
, unsigned int flags
,
2766 struct page
*cow_page
, struct page
**page
)
2768 struct vm_fault vmf
;
2771 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
2775 vmf
.gfp_mask
= __get_fault_gfp_mask(vma
);
2776 vmf
.cow_page
= cow_page
;
2778 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
2779 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
2784 if (unlikely(PageHWPoison(vmf
.page
))) {
2785 if (ret
& VM_FAULT_LOCKED
)
2786 unlock_page(vmf
.page
);
2787 page_cache_release(vmf
.page
);
2788 return VM_FAULT_HWPOISON
;
2791 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
2792 lock_page(vmf
.page
);
2794 VM_BUG_ON_PAGE(!PageLocked(vmf
.page
), vmf
.page
);
2802 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2804 * @vma: virtual memory area
2805 * @address: user virtual address
2806 * @page: page to map
2807 * @pte: pointer to target page table entry
2808 * @write: true, if new entry is writable
2809 * @anon: true, if it's anonymous page
2811 * Caller must hold page table lock relevant for @pte.
2813 * Target users are page handler itself and implementations of
2814 * vm_ops->map_pages.
2816 void do_set_pte(struct vm_area_struct
*vma
, unsigned long address
,
2817 struct page
*page
, pte_t
*pte
, bool write
, bool anon
)
2821 flush_icache_page(vma
, page
);
2822 entry
= mk_pte(page
, vma
->vm_page_prot
);
2824 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2826 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
2827 page_add_new_anon_rmap(page
, vma
, address
);
2829 inc_mm_counter_fast(vma
->vm_mm
, mm_counter_file(page
));
2830 page_add_file_rmap(page
);
2832 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
2834 /* no need to invalidate: a not-present page won't be cached */
2835 update_mmu_cache(vma
, address
, pte
);
2838 static unsigned long fault_around_bytes __read_mostly
=
2839 rounddown_pow_of_two(65536);
2841 #ifdef CONFIG_DEBUG_FS
2842 static int fault_around_bytes_get(void *data
, u64
*val
)
2844 *val
= fault_around_bytes
;
2849 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2850 * rounded down to nearest page order. It's what do_fault_around() expects to
2853 static int fault_around_bytes_set(void *data
, u64 val
)
2855 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
2857 if (val
> PAGE_SIZE
)
2858 fault_around_bytes
= rounddown_pow_of_two(val
);
2860 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
2863 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops
,
2864 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
2866 static int __init
fault_around_debugfs(void)
2870 ret
= debugfs_create_file("fault_around_bytes", 0644, NULL
, NULL
,
2871 &fault_around_bytes_fops
);
2873 pr_warn("Failed to create fault_around_bytes in debugfs");
2876 late_initcall(fault_around_debugfs
);
2880 * do_fault_around() tries to map few pages around the fault address. The hope
2881 * is that the pages will be needed soon and this will lower the number of
2884 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2885 * not ready to be mapped: not up-to-date, locked, etc.
2887 * This function is called with the page table lock taken. In the split ptlock
2888 * case the page table lock only protects only those entries which belong to
2889 * the page table corresponding to the fault address.
2891 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2894 * fault_around_pages() defines how many pages we'll try to map.
2895 * do_fault_around() expects it to return a power of two less than or equal to
2898 * The virtual address of the area that we map is naturally aligned to the
2899 * fault_around_pages() value (and therefore to page order). This way it's
2900 * easier to guarantee that we don't cross page table boundaries.
2902 static void do_fault_around(struct vm_area_struct
*vma
, unsigned long address
,
2903 pte_t
*pte
, pgoff_t pgoff
, unsigned int flags
)
2905 unsigned long start_addr
, nr_pages
, mask
;
2907 struct vm_fault vmf
;
2910 nr_pages
= READ_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
2911 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
2913 start_addr
= max(address
& mask
, vma
->vm_start
);
2914 off
= ((address
- start_addr
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
2919 * max_pgoff is either end of page table or end of vma
2920 * or fault_around_pages() from pgoff, depending what is nearest.
2922 max_pgoff
= pgoff
- ((start_addr
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
2924 max_pgoff
= min3(max_pgoff
, vma_pages(vma
) + vma
->vm_pgoff
- 1,
2925 pgoff
+ nr_pages
- 1);
2927 /* Check if it makes any sense to call ->map_pages */
2928 while (!pte_none(*pte
)) {
2929 if (++pgoff
> max_pgoff
)
2931 start_addr
+= PAGE_SIZE
;
2932 if (start_addr
>= vma
->vm_end
)
2937 vmf
.virtual_address
= (void __user
*) start_addr
;
2940 vmf
.max_pgoff
= max_pgoff
;
2942 vmf
.gfp_mask
= __get_fault_gfp_mask(vma
);
2943 vma
->vm_ops
->map_pages(vma
, &vmf
);
2946 static int do_read_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2947 unsigned long address
, pmd_t
*pmd
,
2948 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
2950 struct page
*fault_page
;
2956 * Let's call ->map_pages() first and use ->fault() as fallback
2957 * if page by the offset is not ready to be mapped (cold cache or
2960 if (vma
->vm_ops
->map_pages
&& fault_around_bytes
>> PAGE_SHIFT
> 1) {
2961 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2962 do_fault_around(vma
, address
, pte
, pgoff
, flags
);
2963 if (!pte_same(*pte
, orig_pte
))
2965 pte_unmap_unlock(pte
, ptl
);
2968 ret
= __do_fault(vma
, address
, pgoff
, flags
, NULL
, &fault_page
);
2969 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
2972 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2973 if (unlikely(!pte_same(*pte
, orig_pte
))) {
2974 pte_unmap_unlock(pte
, ptl
);
2975 unlock_page(fault_page
);
2976 page_cache_release(fault_page
);
2979 do_set_pte(vma
, address
, fault_page
, pte
, false, false);
2980 unlock_page(fault_page
);
2982 pte_unmap_unlock(pte
, ptl
);
2986 static int do_cow_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2987 unsigned long address
, pmd_t
*pmd
,
2988 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
2990 struct page
*fault_page
, *new_page
;
2991 struct mem_cgroup
*memcg
;
2996 if (unlikely(anon_vma_prepare(vma
)))
2997 return VM_FAULT_OOM
;
2999 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
3001 return VM_FAULT_OOM
;
3003 if (mem_cgroup_try_charge(new_page
, mm
, GFP_KERNEL
, &memcg
)) {
3004 page_cache_release(new_page
);
3005 return VM_FAULT_OOM
;
3008 ret
= __do_fault(vma
, address
, pgoff
, flags
, new_page
, &fault_page
);
3009 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3013 copy_user_highpage(new_page
, fault_page
, address
, vma
);
3014 __SetPageUptodate(new_page
);
3016 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3017 if (unlikely(!pte_same(*pte
, orig_pte
))) {
3018 pte_unmap_unlock(pte
, ptl
);
3020 unlock_page(fault_page
);
3021 page_cache_release(fault_page
);
3024 * The fault handler has no page to lock, so it holds
3025 * i_mmap_lock for read to protect against truncate.
3027 i_mmap_unlock_read(vma
->vm_file
->f_mapping
);
3031 do_set_pte(vma
, address
, new_page
, pte
, true, true);
3032 mem_cgroup_commit_charge(new_page
, memcg
, false);
3033 lru_cache_add_active_or_unevictable(new_page
, vma
);
3034 pte_unmap_unlock(pte
, ptl
);
3036 unlock_page(fault_page
);
3037 page_cache_release(fault_page
);
3040 * The fault handler has no page to lock, so it holds
3041 * i_mmap_lock for read to protect against truncate.
3043 i_mmap_unlock_read(vma
->vm_file
->f_mapping
);
3047 mem_cgroup_cancel_charge(new_page
, memcg
);
3048 page_cache_release(new_page
);
3052 static int do_shared_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3053 unsigned long address
, pmd_t
*pmd
,
3054 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3056 struct page
*fault_page
;
3057 struct address_space
*mapping
;
3063 ret
= __do_fault(vma
, address
, pgoff
, flags
, NULL
, &fault_page
);
3064 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3068 * Check if the backing address space wants to know that the page is
3069 * about to become writable
3071 if (vma
->vm_ops
->page_mkwrite
) {
3072 unlock_page(fault_page
);
3073 tmp
= do_page_mkwrite(vma
, fault_page
, address
);
3074 if (unlikely(!tmp
||
3075 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3076 page_cache_release(fault_page
);
3081 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3082 if (unlikely(!pte_same(*pte
, orig_pte
))) {
3083 pte_unmap_unlock(pte
, ptl
);
3084 unlock_page(fault_page
);
3085 page_cache_release(fault_page
);
3088 do_set_pte(vma
, address
, fault_page
, pte
, true, false);
3089 pte_unmap_unlock(pte
, ptl
);
3091 if (set_page_dirty(fault_page
))
3094 * Take a local copy of the address_space - page.mapping may be zeroed
3095 * by truncate after unlock_page(). The address_space itself remains
3096 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3097 * release semantics to prevent the compiler from undoing this copying.
3099 mapping
= fault_page
->mapping
;
3100 unlock_page(fault_page
);
3101 if ((dirtied
|| vma
->vm_ops
->page_mkwrite
) && mapping
) {
3103 * Some device drivers do not set page.mapping but still
3106 balance_dirty_pages_ratelimited(mapping
);
3109 if (!vma
->vm_ops
->page_mkwrite
)
3110 file_update_time(vma
->vm_file
);
3116 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3117 * but allow concurrent faults).
3118 * The mmap_sem may have been released depending on flags and our
3119 * return value. See filemap_fault() and __lock_page_or_retry().
3121 static int do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3122 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3123 unsigned int flags
, pte_t orig_pte
)
3125 pgoff_t pgoff
= (((address
& PAGE_MASK
)
3126 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
3128 pte_unmap(page_table
);
3129 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3130 if (!vma
->vm_ops
->fault
)
3131 return VM_FAULT_SIGBUS
;
3132 if (!(flags
& FAULT_FLAG_WRITE
))
3133 return do_read_fault(mm
, vma
, address
, pmd
, pgoff
, flags
,
3135 if (!(vma
->vm_flags
& VM_SHARED
))
3136 return do_cow_fault(mm
, vma
, address
, pmd
, pgoff
, flags
,
3138 return do_shared_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3141 static int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3142 unsigned long addr
, int page_nid
,
3147 count_vm_numa_event(NUMA_HINT_FAULTS
);
3148 if (page_nid
== numa_node_id()) {
3149 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3150 *flags
|= TNF_FAULT_LOCAL
;
3153 return mpol_misplaced(page
, vma
, addr
);
3156 static int do_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3157 unsigned long addr
, pte_t pte
, pte_t
*ptep
, pmd_t
*pmd
)
3159 struct page
*page
= NULL
;
3164 bool migrated
= false;
3165 bool was_writable
= pte_write(pte
);
3168 /* A PROT_NONE fault should not end up here */
3169 BUG_ON(!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)));
3172 * The "pte" at this point cannot be used safely without
3173 * validation through pte_unmap_same(). It's of NUMA type but
3174 * the pfn may be screwed if the read is non atomic.
3176 * We can safely just do a "set_pte_at()", because the old
3177 * page table entry is not accessible, so there would be no
3178 * concurrent hardware modifications to the PTE.
3180 ptl
= pte_lockptr(mm
, pmd
);
3182 if (unlikely(!pte_same(*ptep
, pte
))) {
3183 pte_unmap_unlock(ptep
, ptl
);
3187 /* Make it present again */
3188 pte
= pte_modify(pte
, vma
->vm_page_prot
);
3189 pte
= pte_mkyoung(pte
);
3191 pte
= pte_mkwrite(pte
);
3192 set_pte_at(mm
, addr
, ptep
, pte
);
3193 update_mmu_cache(vma
, addr
, ptep
);
3195 page
= vm_normal_page(vma
, addr
, pte
);
3197 pte_unmap_unlock(ptep
, ptl
);
3202 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3203 * much anyway since they can be in shared cache state. This misses
3204 * the case where a mapping is writable but the process never writes
3205 * to it but pte_write gets cleared during protection updates and
3206 * pte_dirty has unpredictable behaviour between PTE scan updates,
3207 * background writeback, dirty balancing and application behaviour.
3209 if (!(vma
->vm_flags
& VM_WRITE
))
3210 flags
|= TNF_NO_GROUP
;
3213 * Flag if the page is shared between multiple address spaces. This
3214 * is later used when determining whether to group tasks together
3216 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
3217 flags
|= TNF_SHARED
;
3219 last_cpupid
= page_cpupid_last(page
);
3220 page_nid
= page_to_nid(page
);
3221 target_nid
= numa_migrate_prep(page
, vma
, addr
, page_nid
, &flags
);
3222 pte_unmap_unlock(ptep
, ptl
);
3223 if (target_nid
== -1) {
3228 /* Migrate to the requested node */
3229 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
3231 page_nid
= target_nid
;
3232 flags
|= TNF_MIGRATED
;
3234 flags
|= TNF_MIGRATE_FAIL
;
3238 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
3242 static int create_huge_pmd(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3243 unsigned long address
, pmd_t
*pmd
, unsigned int flags
)
3245 if (vma_is_anonymous(vma
))
3246 return do_huge_pmd_anonymous_page(mm
, vma
, address
, pmd
, flags
);
3247 if (vma
->vm_ops
->pmd_fault
)
3248 return vma
->vm_ops
->pmd_fault(vma
, address
, pmd
, flags
);
3249 return VM_FAULT_FALLBACK
;
3252 static int wp_huge_pmd(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3253 unsigned long address
, pmd_t
*pmd
, pmd_t orig_pmd
,
3256 if (vma_is_anonymous(vma
))
3257 return do_huge_pmd_wp_page(mm
, vma
, address
, pmd
, orig_pmd
);
3258 if (vma
->vm_ops
->pmd_fault
)
3259 return vma
->vm_ops
->pmd_fault(vma
, address
, pmd
, flags
);
3260 return VM_FAULT_FALLBACK
;
3264 * These routines also need to handle stuff like marking pages dirty
3265 * and/or accessed for architectures that don't do it in hardware (most
3266 * RISC architectures). The early dirtying is also good on the i386.
3268 * There is also a hook called "update_mmu_cache()" that architectures
3269 * with external mmu caches can use to update those (ie the Sparc or
3270 * PowerPC hashed page tables that act as extended TLBs).
3272 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3273 * but allow concurrent faults), and pte mapped but not yet locked.
3274 * We return with pte unmapped and unlocked.
3276 * The mmap_sem may have been released depending on flags and our
3277 * return value. See filemap_fault() and __lock_page_or_retry().
3279 static int handle_pte_fault(struct mm_struct
*mm
,
3280 struct vm_area_struct
*vma
, unsigned long address
,
3281 pte_t
*pte
, pmd_t
*pmd
, unsigned int flags
)
3287 * some architectures can have larger ptes than wordsize,
3288 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3289 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3290 * The code below just needs a consistent view for the ifs and
3291 * we later double check anyway with the ptl lock held. So here
3292 * a barrier will do.
3296 if (!pte_present(entry
)) {
3297 if (pte_none(entry
)) {
3298 if (vma_is_anonymous(vma
))
3299 return do_anonymous_page(mm
, vma
, address
,
3302 return do_fault(mm
, vma
, address
, pte
, pmd
,
3305 return do_swap_page(mm
, vma
, address
,
3306 pte
, pmd
, flags
, entry
);
3309 if (pte_protnone(entry
))
3310 return do_numa_page(mm
, vma
, address
, entry
, pte
, pmd
);
3312 ptl
= pte_lockptr(mm
, pmd
);
3314 if (unlikely(!pte_same(*pte
, entry
)))
3316 if (flags
& FAULT_FLAG_WRITE
) {
3317 if (!pte_write(entry
))
3318 return do_wp_page(mm
, vma
, address
,
3319 pte
, pmd
, ptl
, entry
);
3320 entry
= pte_mkdirty(entry
);
3322 entry
= pte_mkyoung(entry
);
3323 if (ptep_set_access_flags(vma
, address
, pte
, entry
, flags
& FAULT_FLAG_WRITE
)) {
3324 update_mmu_cache(vma
, address
, pte
);
3327 * This is needed only for protection faults but the arch code
3328 * is not yet telling us if this is a protection fault or not.
3329 * This still avoids useless tlb flushes for .text page faults
3332 if (flags
& FAULT_FLAG_WRITE
)
3333 flush_tlb_fix_spurious_fault(vma
, address
);
3336 pte_unmap_unlock(pte
, ptl
);
3341 * By the time we get here, we already hold the mm semaphore
3343 * The mmap_sem may have been released depending on flags and our
3344 * return value. See filemap_fault() and __lock_page_or_retry().
3346 static int __handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3347 unsigned long address
, unsigned int flags
)
3354 if (unlikely(is_vm_hugetlb_page(vma
)))
3355 return hugetlb_fault(mm
, vma
, address
, flags
);
3357 pgd
= pgd_offset(mm
, address
);
3358 pud
= pud_alloc(mm
, pgd
, address
);
3360 return VM_FAULT_OOM
;
3361 pmd
= pmd_alloc(mm
, pud
, address
);
3363 return VM_FAULT_OOM
;
3364 if (pmd_none(*pmd
) && transparent_hugepage_enabled(vma
)) {
3365 int ret
= create_huge_pmd(mm
, vma
, address
, pmd
, flags
);
3366 if (!(ret
& VM_FAULT_FALLBACK
))
3369 pmd_t orig_pmd
= *pmd
;
3373 if (pmd_trans_huge(orig_pmd
)) {
3374 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
3377 * If the pmd is splitting, return and retry the
3378 * the fault. Alternative: wait until the split
3379 * is done, and goto retry.
3381 if (pmd_trans_splitting(orig_pmd
))
3384 if (pmd_protnone(orig_pmd
))
3385 return do_huge_pmd_numa_page(mm
, vma
, address
,
3388 if (dirty
&& !pmd_write(orig_pmd
)) {
3389 ret
= wp_huge_pmd(mm
, vma
, address
, pmd
,
3391 if (!(ret
& VM_FAULT_FALLBACK
))
3394 huge_pmd_set_accessed(mm
, vma
, address
, pmd
,
3402 * Use __pte_alloc instead of pte_alloc_map, because we can't
3403 * run pte_offset_map on the pmd, if an huge pmd could
3404 * materialize from under us from a different thread.
3406 if (unlikely(pmd_none(*pmd
)) &&
3407 unlikely(__pte_alloc(mm
, vma
, pmd
, address
)))
3408 return VM_FAULT_OOM
;
3409 /* if an huge pmd materialized from under us just retry later */
3410 if (unlikely(pmd_trans_huge(*pmd
)))
3413 * A regular pmd is established and it can't morph into a huge pmd
3414 * from under us anymore at this point because we hold the mmap_sem
3415 * read mode and khugepaged takes it in write mode. So now it's
3416 * safe to run pte_offset_map().
3418 pte
= pte_offset_map(pmd
, address
);
3420 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
3424 * By the time we get here, we already hold the mm semaphore
3426 * The mmap_sem may have been released depending on flags and our
3427 * return value. See filemap_fault() and __lock_page_or_retry().
3429 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3430 unsigned long address
, unsigned int flags
)
3434 __set_current_state(TASK_RUNNING
);
3436 count_vm_event(PGFAULT
);
3437 mem_cgroup_count_vm_event(mm
, PGFAULT
);
3439 /* do counter updates before entering really critical section. */
3440 check_sync_rss_stat(current
);
3443 * Enable the memcg OOM handling for faults triggered in user
3444 * space. Kernel faults are handled more gracefully.
3446 if (flags
& FAULT_FLAG_USER
)
3447 mem_cgroup_oom_enable();
3449 ret
= __handle_mm_fault(mm
, vma
, address
, flags
);
3451 if (flags
& FAULT_FLAG_USER
) {
3452 mem_cgroup_oom_disable();
3454 * The task may have entered a memcg OOM situation but
3455 * if the allocation error was handled gracefully (no
3456 * VM_FAULT_OOM), there is no need to kill anything.
3457 * Just clean up the OOM state peacefully.
3459 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
3460 mem_cgroup_oom_synchronize(false);
3465 EXPORT_SYMBOL_GPL(handle_mm_fault
);
3467 #ifndef __PAGETABLE_PUD_FOLDED
3469 * Allocate page upper directory.
3470 * We've already handled the fast-path in-line.
3472 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
3474 pud_t
*new = pud_alloc_one(mm
, address
);
3478 smp_wmb(); /* See comment in __pte_alloc */
3480 spin_lock(&mm
->page_table_lock
);
3481 if (pgd_present(*pgd
)) /* Another has populated it */
3484 pgd_populate(mm
, pgd
, new);
3485 spin_unlock(&mm
->page_table_lock
);
3488 #endif /* __PAGETABLE_PUD_FOLDED */
3490 #ifndef __PAGETABLE_PMD_FOLDED
3492 * Allocate page middle directory.
3493 * We've already handled the fast-path in-line.
3495 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
3497 pmd_t
*new = pmd_alloc_one(mm
, address
);
3501 smp_wmb(); /* See comment in __pte_alloc */
3503 spin_lock(&mm
->page_table_lock
);
3504 #ifndef __ARCH_HAS_4LEVEL_HACK
3505 if (!pud_present(*pud
)) {
3507 pud_populate(mm
, pud
, new);
3508 } else /* Another has populated it */
3511 if (!pgd_present(*pud
)) {
3513 pgd_populate(mm
, pud
, new);
3514 } else /* Another has populated it */
3516 #endif /* __ARCH_HAS_4LEVEL_HACK */
3517 spin_unlock(&mm
->page_table_lock
);
3520 #endif /* __PAGETABLE_PMD_FOLDED */
3522 static int __follow_pte(struct mm_struct
*mm
, unsigned long address
,
3523 pte_t
**ptepp
, spinlock_t
**ptlp
)
3530 pgd
= pgd_offset(mm
, address
);
3531 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
3534 pud
= pud_offset(pgd
, address
);
3535 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
3538 pmd
= pmd_offset(pud
, address
);
3539 VM_BUG_ON(pmd_trans_huge(*pmd
));
3540 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
3543 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3547 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
3550 if (!pte_present(*ptep
))
3555 pte_unmap_unlock(ptep
, *ptlp
);
3560 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
3561 pte_t
**ptepp
, spinlock_t
**ptlp
)
3565 /* (void) is needed to make gcc happy */
3566 (void) __cond_lock(*ptlp
,
3567 !(res
= __follow_pte(mm
, address
, ptepp
, ptlp
)));
3572 * follow_pfn - look up PFN at a user virtual address
3573 * @vma: memory mapping
3574 * @address: user virtual address
3575 * @pfn: location to store found PFN
3577 * Only IO mappings and raw PFN mappings are allowed.
3579 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3581 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
3588 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3591 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
3594 *pfn
= pte_pfn(*ptep
);
3595 pte_unmap_unlock(ptep
, ptl
);
3598 EXPORT_SYMBOL(follow_pfn
);
3600 #ifdef CONFIG_HAVE_IOREMAP_PROT
3601 int follow_phys(struct vm_area_struct
*vma
,
3602 unsigned long address
, unsigned int flags
,
3603 unsigned long *prot
, resource_size_t
*phys
)
3609 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3612 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
3616 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
3619 *prot
= pgprot_val(pte_pgprot(pte
));
3620 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
3624 pte_unmap_unlock(ptep
, ptl
);
3629 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
3630 void *buf
, int len
, int write
)
3632 resource_size_t phys_addr
;
3633 unsigned long prot
= 0;
3634 void __iomem
*maddr
;
3635 int offset
= addr
& (PAGE_SIZE
-1);
3637 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
3640 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
3642 memcpy_toio(maddr
+ offset
, buf
, len
);
3644 memcpy_fromio(buf
, maddr
+ offset
, len
);
3649 EXPORT_SYMBOL_GPL(generic_access_phys
);
3653 * Access another process' address space as given in mm. If non-NULL, use the
3654 * given task for page fault accounting.
3656 static int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
3657 unsigned long addr
, void *buf
, int len
, int write
)
3659 struct vm_area_struct
*vma
;
3660 void *old_buf
= buf
;
3662 down_read(&mm
->mmap_sem
);
3663 /* ignore errors, just check how much was successfully transferred */
3665 int bytes
, ret
, offset
;
3667 struct page
*page
= NULL
;
3669 ret
= get_user_pages(tsk
, mm
, addr
, 1,
3670 write
, 1, &page
, &vma
);
3672 #ifndef CONFIG_HAVE_IOREMAP_PROT
3676 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3677 * we can access using slightly different code.
3679 vma
= find_vma(mm
, addr
);
3680 if (!vma
|| vma
->vm_start
> addr
)
3682 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
3683 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
3691 offset
= addr
& (PAGE_SIZE
-1);
3692 if (bytes
> PAGE_SIZE
-offset
)
3693 bytes
= PAGE_SIZE
-offset
;
3697 copy_to_user_page(vma
, page
, addr
,
3698 maddr
+ offset
, buf
, bytes
);
3699 set_page_dirty_lock(page
);
3701 copy_from_user_page(vma
, page
, addr
,
3702 buf
, maddr
+ offset
, bytes
);
3705 page_cache_release(page
);
3711 up_read(&mm
->mmap_sem
);
3713 return buf
- old_buf
;
3717 * access_remote_vm - access another process' address space
3718 * @mm: the mm_struct of the target address space
3719 * @addr: start address to access
3720 * @buf: source or destination buffer
3721 * @len: number of bytes to transfer
3722 * @write: whether the access is a write
3724 * The caller must hold a reference on @mm.
3726 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
3727 void *buf
, int len
, int write
)
3729 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, write
);
3733 * Access another process' address space.
3734 * Source/target buffer must be kernel space,
3735 * Do not walk the page table directly, use get_user_pages
3737 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
3738 void *buf
, int len
, int write
)
3740 struct mm_struct
*mm
;
3743 mm
= get_task_mm(tsk
);
3747 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, write
);
3754 * Print the name of a VMA.
3756 void print_vma_addr(char *prefix
, unsigned long ip
)
3758 struct mm_struct
*mm
= current
->mm
;
3759 struct vm_area_struct
*vma
;
3762 * Do not print if we are in atomic
3763 * contexts (in exception stacks, etc.):
3765 if (preempt_count())
3768 down_read(&mm
->mmap_sem
);
3769 vma
= find_vma(mm
, ip
);
3770 if (vma
&& vma
->vm_file
) {
3771 struct file
*f
= vma
->vm_file
;
3772 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
3776 p
= file_path(f
, buf
, PAGE_SIZE
);
3779 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
3781 vma
->vm_end
- vma
->vm_start
);
3782 free_page((unsigned long)buf
);
3785 up_read(&mm
->mmap_sem
);
3788 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3789 void __might_fault(const char *file
, int line
)
3792 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3793 * holding the mmap_sem, this is safe because kernel memory doesn't
3794 * get paged out, therefore we'll never actually fault, and the
3795 * below annotations will generate false positives.
3797 if (segment_eq(get_fs(), KERNEL_DS
))
3799 if (pagefault_disabled())
3801 __might_sleep(file
, line
, 0);
3802 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3804 might_lock_read(¤t
->mm
->mmap_sem
);
3807 EXPORT_SYMBOL(__might_fault
);
3810 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3811 static void clear_gigantic_page(struct page
*page
,
3813 unsigned int pages_per_huge_page
)
3816 struct page
*p
= page
;
3819 for (i
= 0; i
< pages_per_huge_page
;
3820 i
++, p
= mem_map_next(p
, page
, i
)) {
3822 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
3825 void clear_huge_page(struct page
*page
,
3826 unsigned long addr
, unsigned int pages_per_huge_page
)
3830 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3831 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
3836 for (i
= 0; i
< pages_per_huge_page
; i
++) {
3838 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
3842 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
3844 struct vm_area_struct
*vma
,
3845 unsigned int pages_per_huge_page
)
3848 struct page
*dst_base
= dst
;
3849 struct page
*src_base
= src
;
3851 for (i
= 0; i
< pages_per_huge_page
; ) {
3853 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
3856 dst
= mem_map_next(dst
, dst_base
, i
);
3857 src
= mem_map_next(src
, src_base
, i
);
3861 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
3862 unsigned long addr
, struct vm_area_struct
*vma
,
3863 unsigned int pages_per_huge_page
)
3867 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3868 copy_user_gigantic_page(dst
, src
, addr
, vma
,
3869 pages_per_huge_page
);
3874 for (i
= 0; i
< pages_per_huge_page
; i
++) {
3876 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
3879 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3881 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3883 static struct kmem_cache
*page_ptl_cachep
;
3885 void __init
ptlock_cache_init(void)
3887 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
3891 bool ptlock_alloc(struct page
*page
)
3895 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
3902 void ptlock_free(struct page
*page
)
3904 kmem_cache_free(page_ptl_cachep
, page
->ptl
);