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/pfn_t.h>
54 #include <linux/writeback.h>
55 #include <linux/memcontrol.h>
56 #include <linux/mmu_notifier.h>
57 #include <linux/kallsyms.h>
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
60 #include <linux/gfp.h>
61 #include <linux/migrate.h>
62 #include <linux/string.h>
63 #include <linux/dma-debug.h>
64 #include <linux/debugfs.h>
65 #include <linux/userfaultfd_k.h>
66 #include <linux/dax.h>
69 #include <asm/mmu_context.h>
70 #include <asm/pgalloc.h>
71 #include <asm/uaccess.h>
73 #include <asm/tlbflush.h>
74 #include <asm/pgtable.h>
78 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
79 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
82 #ifndef CONFIG_NEED_MULTIPLE_NODES
83 /* use the per-pgdat data instead for discontigmem - mbligh */
84 unsigned long max_mapnr
;
87 EXPORT_SYMBOL(max_mapnr
);
88 EXPORT_SYMBOL(mem_map
);
92 * A number of key systems in x86 including ioremap() rely on the assumption
93 * that high_memory defines the upper bound on direct map memory, then end
94 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
95 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
100 EXPORT_SYMBOL(high_memory
);
103 * Randomize the address space (stacks, mmaps, brk, etc.).
105 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
106 * as ancient (libc5 based) binaries can segfault. )
108 int randomize_va_space __read_mostly
=
109 #ifdef CONFIG_COMPAT_BRK
115 static int __init
disable_randmaps(char *s
)
117 randomize_va_space
= 0;
120 __setup("norandmaps", disable_randmaps
);
122 unsigned long zero_pfn __read_mostly
;
123 unsigned long highest_memmap_pfn __read_mostly
;
125 EXPORT_SYMBOL(zero_pfn
);
128 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
130 static int __init
init_zero_pfn(void)
132 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
135 core_initcall(init_zero_pfn
);
138 #if defined(SPLIT_RSS_COUNTING)
140 void sync_mm_rss(struct mm_struct
*mm
)
144 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
145 if (current
->rss_stat
.count
[i
]) {
146 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
147 current
->rss_stat
.count
[i
] = 0;
150 current
->rss_stat
.events
= 0;
153 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
155 struct task_struct
*task
= current
;
157 if (likely(task
->mm
== mm
))
158 task
->rss_stat
.count
[member
] += val
;
160 add_mm_counter(mm
, member
, val
);
162 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
163 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
165 /* sync counter once per 64 page faults */
166 #define TASK_RSS_EVENTS_THRESH (64)
167 static void check_sync_rss_stat(struct task_struct
*task
)
169 if (unlikely(task
!= current
))
171 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
172 sync_mm_rss(task
->mm
);
174 #else /* SPLIT_RSS_COUNTING */
176 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
177 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
179 static void check_sync_rss_stat(struct task_struct
*task
)
183 #endif /* SPLIT_RSS_COUNTING */
185 #ifdef HAVE_GENERIC_MMU_GATHER
187 static bool tlb_next_batch(struct mmu_gather
*tlb
)
189 struct mmu_gather_batch
*batch
;
193 tlb
->active
= batch
->next
;
197 if (tlb
->batch_count
== MAX_GATHER_BATCH_COUNT
)
200 batch
= (void *)__get_free_pages(GFP_NOWAIT
| __GFP_NOWARN
, 0);
207 batch
->max
= MAX_GATHER_BATCH
;
209 tlb
->active
->next
= batch
;
216 * Called to initialize an (on-stack) mmu_gather structure for page-table
217 * tear-down from @mm. The @fullmm argument is used when @mm is without
218 * users and we're going to destroy the full address space (exit/execve).
220 void tlb_gather_mmu(struct mmu_gather
*tlb
, struct mm_struct
*mm
, unsigned long start
, unsigned long end
)
224 /* Is it from 0 to ~0? */
225 tlb
->fullmm
= !(start
| (end
+1));
226 tlb
->need_flush_all
= 0;
227 tlb
->local
.next
= NULL
;
229 tlb
->local
.max
= ARRAY_SIZE(tlb
->__pages
);
230 tlb
->active
= &tlb
->local
;
231 tlb
->batch_count
= 0;
233 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
237 __tlb_reset_range(tlb
);
240 static void tlb_flush_mmu_tlbonly(struct mmu_gather
*tlb
)
246 mmu_notifier_invalidate_range(tlb
->mm
, tlb
->start
, tlb
->end
);
247 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
248 tlb_table_flush(tlb
);
250 __tlb_reset_range(tlb
);
253 static void tlb_flush_mmu_free(struct mmu_gather
*tlb
)
255 struct mmu_gather_batch
*batch
;
257 for (batch
= &tlb
->local
; batch
&& batch
->nr
; batch
= batch
->next
) {
258 free_pages_and_swap_cache(batch
->pages
, batch
->nr
);
261 tlb
->active
= &tlb
->local
;
264 void tlb_flush_mmu(struct mmu_gather
*tlb
)
266 tlb_flush_mmu_tlbonly(tlb
);
267 tlb_flush_mmu_free(tlb
);
271 * Called at the end of the shootdown operation to free up any resources
272 * that were required.
274 void tlb_finish_mmu(struct mmu_gather
*tlb
, unsigned long start
, unsigned long end
)
276 struct mmu_gather_batch
*batch
, *next
;
280 /* keep the page table cache within bounds */
283 for (batch
= tlb
->local
.next
; batch
; batch
= next
) {
285 free_pages((unsigned long)batch
, 0);
287 tlb
->local
.next
= NULL
;
291 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
292 * handling the additional races in SMP caused by other CPUs caching valid
293 * mappings in their TLBs. Returns the number of free page slots left.
294 * When out of page slots we must call tlb_flush_mmu().
296 int __tlb_remove_page(struct mmu_gather
*tlb
, struct page
*page
)
298 struct mmu_gather_batch
*batch
;
300 VM_BUG_ON(!tlb
->end
);
303 batch
->pages
[batch
->nr
++] = page
;
304 if (batch
->nr
== batch
->max
) {
305 if (!tlb_next_batch(tlb
))
309 VM_BUG_ON_PAGE(batch
->nr
> batch
->max
, page
);
311 return batch
->max
- batch
->nr
;
314 #endif /* HAVE_GENERIC_MMU_GATHER */
316 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
319 * See the comment near struct mmu_table_batch.
322 static void tlb_remove_table_smp_sync(void *arg
)
324 /* Simply deliver the interrupt */
327 static void tlb_remove_table_one(void *table
)
330 * This isn't an RCU grace period and hence the page-tables cannot be
331 * assumed to be actually RCU-freed.
333 * It is however sufficient for software page-table walkers that rely on
334 * IRQ disabling. See the comment near struct mmu_table_batch.
336 smp_call_function(tlb_remove_table_smp_sync
, NULL
, 1);
337 __tlb_remove_table(table
);
340 static void tlb_remove_table_rcu(struct rcu_head
*head
)
342 struct mmu_table_batch
*batch
;
345 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
347 for (i
= 0; i
< batch
->nr
; i
++)
348 __tlb_remove_table(batch
->tables
[i
]);
350 free_page((unsigned long)batch
);
353 void tlb_table_flush(struct mmu_gather
*tlb
)
355 struct mmu_table_batch
**batch
= &tlb
->batch
;
358 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
363 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
365 struct mmu_table_batch
**batch
= &tlb
->batch
;
368 * When there's less then two users of this mm there cannot be a
369 * concurrent page-table walk.
371 if (atomic_read(&tlb
->mm
->mm_users
) < 2) {
372 __tlb_remove_table(table
);
376 if (*batch
== NULL
) {
377 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
378 if (*batch
== NULL
) {
379 tlb_remove_table_one(table
);
384 (*batch
)->tables
[(*batch
)->nr
++] = table
;
385 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
386 tlb_table_flush(tlb
);
389 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
392 * Note: this doesn't free the actual pages themselves. That
393 * has been handled earlier when unmapping all the memory regions.
395 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
398 pgtable_t token
= pmd_pgtable(*pmd
);
400 pte_free_tlb(tlb
, token
, addr
);
401 atomic_long_dec(&tlb
->mm
->nr_ptes
);
404 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
405 unsigned long addr
, unsigned long end
,
406 unsigned long floor
, unsigned long ceiling
)
413 pmd
= pmd_offset(pud
, addr
);
415 next
= pmd_addr_end(addr
, end
);
416 if (pmd_none_or_clear_bad(pmd
))
418 free_pte_range(tlb
, pmd
, addr
);
419 } while (pmd
++, addr
= next
, addr
!= end
);
429 if (end
- 1 > ceiling
- 1)
432 pmd
= pmd_offset(pud
, start
);
434 pmd_free_tlb(tlb
, pmd
, start
);
435 mm_dec_nr_pmds(tlb
->mm
);
438 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
439 unsigned long addr
, unsigned long end
,
440 unsigned long floor
, unsigned long ceiling
)
447 pud
= pud_offset(pgd
, addr
);
449 next
= pud_addr_end(addr
, end
);
450 if (pud_none_or_clear_bad(pud
))
452 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
453 } while (pud
++, addr
= next
, addr
!= end
);
459 ceiling
&= PGDIR_MASK
;
463 if (end
- 1 > ceiling
- 1)
466 pud
= pud_offset(pgd
, start
);
468 pud_free_tlb(tlb
, pud
, start
);
472 * This function frees user-level page tables of a process.
474 void free_pgd_range(struct mmu_gather
*tlb
,
475 unsigned long addr
, unsigned long end
,
476 unsigned long floor
, unsigned long ceiling
)
482 * The next few lines have given us lots of grief...
484 * Why are we testing PMD* at this top level? Because often
485 * there will be no work to do at all, and we'd prefer not to
486 * go all the way down to the bottom just to discover that.
488 * Why all these "- 1"s? Because 0 represents both the bottom
489 * of the address space and the top of it (using -1 for the
490 * top wouldn't help much: the masks would do the wrong thing).
491 * The rule is that addr 0 and floor 0 refer to the bottom of
492 * the address space, but end 0 and ceiling 0 refer to the top
493 * Comparisons need to use "end - 1" and "ceiling - 1" (though
494 * that end 0 case should be mythical).
496 * Wherever addr is brought up or ceiling brought down, we must
497 * be careful to reject "the opposite 0" before it confuses the
498 * subsequent tests. But what about where end is brought down
499 * by PMD_SIZE below? no, end can't go down to 0 there.
501 * Whereas we round start (addr) and ceiling down, by different
502 * masks at different levels, in order to test whether a table
503 * now has no other vmas using it, so can be freed, we don't
504 * bother to round floor or end up - the tests don't need that.
518 if (end
- 1 > ceiling
- 1)
523 pgd
= pgd_offset(tlb
->mm
, addr
);
525 next
= pgd_addr_end(addr
, end
);
526 if (pgd_none_or_clear_bad(pgd
))
528 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
529 } while (pgd
++, addr
= next
, addr
!= end
);
532 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
533 unsigned long floor
, unsigned long ceiling
)
536 struct vm_area_struct
*next
= vma
->vm_next
;
537 unsigned long addr
= vma
->vm_start
;
540 * Hide vma from rmap and truncate_pagecache before freeing
543 unlink_anon_vmas(vma
);
544 unlink_file_vma(vma
);
546 if (is_vm_hugetlb_page(vma
)) {
547 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
548 floor
, next
? next
->vm_start
: ceiling
);
551 * Optimization: gather nearby vmas into one call down
553 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
554 && !is_vm_hugetlb_page(next
)) {
557 unlink_anon_vmas(vma
);
558 unlink_file_vma(vma
);
560 free_pgd_range(tlb
, addr
, vma
->vm_end
,
561 floor
, next
? next
->vm_start
: ceiling
);
567 int __pte_alloc(struct mm_struct
*mm
, pmd_t
*pmd
, unsigned long address
)
570 pgtable_t
new = pte_alloc_one(mm
, address
);
575 * Ensure all pte setup (eg. pte page lock and page clearing) are
576 * visible before the pte is made visible to other CPUs by being
577 * put into page tables.
579 * The other side of the story is the pointer chasing in the page
580 * table walking code (when walking the page table without locking;
581 * ie. most of the time). Fortunately, these data accesses consist
582 * of a chain of data-dependent loads, meaning most CPUs (alpha
583 * being the notable exception) will already guarantee loads are
584 * seen in-order. See the alpha page table accessors for the
585 * smp_read_barrier_depends() barriers in page table walking code.
587 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
589 ptl
= pmd_lock(mm
, pmd
);
590 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
591 atomic_long_inc(&mm
->nr_ptes
);
592 pmd_populate(mm
, pmd
, new);
601 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
603 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
607 smp_wmb(); /* See comment in __pte_alloc */
609 spin_lock(&init_mm
.page_table_lock
);
610 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
611 pmd_populate_kernel(&init_mm
, pmd
, new);
614 spin_unlock(&init_mm
.page_table_lock
);
616 pte_free_kernel(&init_mm
, new);
620 static inline void init_rss_vec(int *rss
)
622 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
625 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
629 if (current
->mm
== mm
)
631 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
633 add_mm_counter(mm
, i
, rss
[i
]);
637 * This function is called to print an error when a bad pte
638 * is found. For example, we might have a PFN-mapped pte in
639 * a region that doesn't allow it.
641 * The calling function must still handle the error.
643 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
644 pte_t pte
, struct page
*page
)
646 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
647 pud_t
*pud
= pud_offset(pgd
, addr
);
648 pmd_t
*pmd
= pmd_offset(pud
, addr
);
649 struct address_space
*mapping
;
651 static unsigned long resume
;
652 static unsigned long nr_shown
;
653 static unsigned long nr_unshown
;
656 * Allow a burst of 60 reports, then keep quiet for that minute;
657 * or allow a steady drip of one report per second.
659 if (nr_shown
== 60) {
660 if (time_before(jiffies
, resume
)) {
665 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
672 resume
= jiffies
+ 60 * HZ
;
674 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
675 index
= linear_page_index(vma
, addr
);
677 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
679 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
681 dump_page(page
, "bad pte");
682 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
683 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
685 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
687 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
689 vma
->vm_ops
? vma
->vm_ops
->fault
: NULL
,
690 vma
->vm_file
? vma
->vm_file
->f_op
->mmap
: NULL
,
691 mapping
? mapping
->a_ops
->readpage
: NULL
);
693 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
697 * vm_normal_page -- This function gets the "struct page" associated with a pte.
699 * "Special" mappings do not wish to be associated with a "struct page" (either
700 * it doesn't exist, or it exists but they don't want to touch it). In this
701 * case, NULL is returned here. "Normal" mappings do have a struct page.
703 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
704 * pte bit, in which case this function is trivial. Secondly, an architecture
705 * may not have a spare pte bit, which requires a more complicated scheme,
708 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
709 * special mapping (even if there are underlying and valid "struct pages").
710 * COWed pages of a VM_PFNMAP are always normal.
712 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
713 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
714 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
715 * mapping will always honor the rule
717 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
719 * And for normal mappings this is false.
721 * This restricts such mappings to be a linear translation from virtual address
722 * to pfn. To get around this restriction, we allow arbitrary mappings so long
723 * as the vma is not a COW mapping; in that case, we know that all ptes are
724 * special (because none can have been COWed).
727 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
729 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
730 * page" backing, however the difference is that _all_ pages with a struct
731 * page (that is, those where pfn_valid is true) are refcounted and considered
732 * normal pages by the VM. The disadvantage is that pages are refcounted
733 * (which can be slower and simply not an option for some PFNMAP users). The
734 * advantage is that we don't have to follow the strict linearity rule of
735 * PFNMAP mappings in order to support COWable mappings.
738 #ifdef __HAVE_ARCH_PTE_SPECIAL
739 # define HAVE_PTE_SPECIAL 1
741 # define HAVE_PTE_SPECIAL 0
743 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
746 unsigned long pfn
= pte_pfn(pte
);
748 if (HAVE_PTE_SPECIAL
) {
749 if (likely(!pte_special(pte
)))
751 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
752 return vma
->vm_ops
->find_special_page(vma
, addr
);
753 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
755 if (!is_zero_pfn(pfn
))
756 print_bad_pte(vma
, addr
, pte
, NULL
);
760 /* !HAVE_PTE_SPECIAL case follows: */
762 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
763 if (vma
->vm_flags
& VM_MIXEDMAP
) {
769 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
770 if (pfn
== vma
->vm_pgoff
+ off
)
772 if (!is_cow_mapping(vma
->vm_flags
))
777 if (is_zero_pfn(pfn
))
780 if (unlikely(pfn
> highest_memmap_pfn
)) {
781 print_bad_pte(vma
, addr
, pte
, NULL
);
786 * NOTE! We still have PageReserved() pages in the page tables.
787 * eg. VDSO mappings can cause them to exist.
790 return pfn_to_page(pfn
);
793 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
794 struct page
*vm_normal_page_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
797 unsigned long pfn
= pmd_pfn(pmd
);
800 * There is no pmd_special() but there may be special pmds, e.g.
801 * in a direct-access (dax) mapping, so let's just replicate the
802 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
804 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
805 if (vma
->vm_flags
& VM_MIXEDMAP
) {
811 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
812 if (pfn
== vma
->vm_pgoff
+ off
)
814 if (!is_cow_mapping(vma
->vm_flags
))
819 if (is_zero_pfn(pfn
))
821 if (unlikely(pfn
> highest_memmap_pfn
))
825 * NOTE! We still have PageReserved() pages in the page tables.
826 * eg. VDSO mappings can cause them to exist.
829 return pfn_to_page(pfn
);
834 * copy one vm_area from one task to the other. Assumes the page tables
835 * already present in the new task to be cleared in the whole range
836 * covered by this vma.
839 static inline unsigned long
840 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
841 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
842 unsigned long addr
, int *rss
)
844 unsigned long vm_flags
= vma
->vm_flags
;
845 pte_t pte
= *src_pte
;
848 /* pte contains position in swap or file, so copy. */
849 if (unlikely(!pte_present(pte
))) {
850 swp_entry_t entry
= pte_to_swp_entry(pte
);
852 if (likely(!non_swap_entry(entry
))) {
853 if (swap_duplicate(entry
) < 0)
856 /* make sure dst_mm is on swapoff's mmlist. */
857 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
858 spin_lock(&mmlist_lock
);
859 if (list_empty(&dst_mm
->mmlist
))
860 list_add(&dst_mm
->mmlist
,
862 spin_unlock(&mmlist_lock
);
865 } else if (is_migration_entry(entry
)) {
866 page
= migration_entry_to_page(entry
);
868 rss
[mm_counter(page
)]++;
870 if (is_write_migration_entry(entry
) &&
871 is_cow_mapping(vm_flags
)) {
873 * COW mappings require pages in both
874 * parent and child to be set to read.
876 make_migration_entry_read(&entry
);
877 pte
= swp_entry_to_pte(entry
);
878 if (pte_swp_soft_dirty(*src_pte
))
879 pte
= pte_swp_mksoft_dirty(pte
);
880 set_pte_at(src_mm
, addr
, src_pte
, pte
);
887 * If it's a COW mapping, write protect it both
888 * in the parent and the child
890 if (is_cow_mapping(vm_flags
)) {
891 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
892 pte
= pte_wrprotect(pte
);
896 * If it's a shared mapping, mark it clean in
899 if (vm_flags
& VM_SHARED
)
900 pte
= pte_mkclean(pte
);
901 pte
= pte_mkold(pte
);
903 page
= vm_normal_page(vma
, addr
, pte
);
906 page_dup_rmap(page
, false);
907 rss
[mm_counter(page
)]++;
911 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
915 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
916 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
917 unsigned long addr
, unsigned long end
)
919 pte_t
*orig_src_pte
, *orig_dst_pte
;
920 pte_t
*src_pte
, *dst_pte
;
921 spinlock_t
*src_ptl
, *dst_ptl
;
923 int rss
[NR_MM_COUNTERS
];
924 swp_entry_t entry
= (swp_entry_t
){0};
929 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
932 src_pte
= pte_offset_map(src_pmd
, addr
);
933 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
934 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
935 orig_src_pte
= src_pte
;
936 orig_dst_pte
= dst_pte
;
937 arch_enter_lazy_mmu_mode();
941 * We are holding two locks at this point - either of them
942 * could generate latencies in another task on another CPU.
944 if (progress
>= 32) {
946 if (need_resched() ||
947 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
950 if (pte_none(*src_pte
)) {
954 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
959 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
961 arch_leave_lazy_mmu_mode();
962 spin_unlock(src_ptl
);
963 pte_unmap(orig_src_pte
);
964 add_mm_rss_vec(dst_mm
, rss
);
965 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
969 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
978 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
979 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
980 unsigned long addr
, unsigned long end
)
982 pmd_t
*src_pmd
, *dst_pmd
;
985 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
988 src_pmd
= pmd_offset(src_pud
, addr
);
990 next
= pmd_addr_end(addr
, end
);
991 if (pmd_trans_huge(*src_pmd
) || pmd_devmap(*src_pmd
)) {
993 VM_BUG_ON(next
-addr
!= HPAGE_PMD_SIZE
);
994 err
= copy_huge_pmd(dst_mm
, src_mm
,
995 dst_pmd
, src_pmd
, addr
, vma
);
1002 if (pmd_none_or_clear_bad(src_pmd
))
1004 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
1007 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1011 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1012 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
1013 unsigned long addr
, unsigned long end
)
1015 pud_t
*src_pud
, *dst_pud
;
1018 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
1021 src_pud
= pud_offset(src_pgd
, addr
);
1023 next
= pud_addr_end(addr
, end
);
1024 if (pud_none_or_clear_bad(src_pud
))
1026 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
1029 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1033 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1034 struct vm_area_struct
*vma
)
1036 pgd_t
*src_pgd
, *dst_pgd
;
1038 unsigned long addr
= vma
->vm_start
;
1039 unsigned long end
= vma
->vm_end
;
1040 unsigned long mmun_start
; /* For mmu_notifiers */
1041 unsigned long mmun_end
; /* For mmu_notifiers */
1046 * Don't copy ptes where a page fault will fill them correctly.
1047 * Fork becomes much lighter when there are big shared or private
1048 * readonly mappings. The tradeoff is that copy_page_range is more
1049 * efficient than faulting.
1051 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_PFNMAP
| VM_MIXEDMAP
)) &&
1055 if (is_vm_hugetlb_page(vma
))
1056 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1058 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
1060 * We do not free on error cases below as remove_vma
1061 * gets called on error from higher level routine
1063 ret
= track_pfn_copy(vma
);
1069 * We need to invalidate the secondary MMU mappings only when
1070 * there could be a permission downgrade on the ptes of the
1071 * parent mm. And a permission downgrade will only happen if
1072 * is_cow_mapping() returns true.
1074 is_cow
= is_cow_mapping(vma
->vm_flags
);
1078 mmu_notifier_invalidate_range_start(src_mm
, mmun_start
,
1082 dst_pgd
= pgd_offset(dst_mm
, addr
);
1083 src_pgd
= pgd_offset(src_mm
, addr
);
1085 next
= pgd_addr_end(addr
, end
);
1086 if (pgd_none_or_clear_bad(src_pgd
))
1088 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1089 vma
, addr
, next
))) {
1093 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1096 mmu_notifier_invalidate_range_end(src_mm
, mmun_start
, mmun_end
);
1100 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1101 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1102 unsigned long addr
, unsigned long end
,
1103 struct zap_details
*details
)
1105 struct mm_struct
*mm
= tlb
->mm
;
1106 int force_flush
= 0;
1107 int rss
[NR_MM_COUNTERS
];
1115 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1117 arch_enter_lazy_mmu_mode();
1120 if (pte_none(ptent
)) {
1124 if (pte_present(ptent
)) {
1127 page
= vm_normal_page(vma
, addr
, ptent
);
1128 if (unlikely(details
) && page
) {
1130 * unmap_shared_mapping_pages() wants to
1131 * invalidate cache without truncating:
1132 * unmap shared but keep private pages.
1134 if (details
->check_mapping
&&
1135 details
->check_mapping
!= page
->mapping
)
1138 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1140 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1141 if (unlikely(!page
))
1144 if (!PageAnon(page
)) {
1145 if (pte_dirty(ptent
)) {
1147 * oom_reaper cannot tear down dirty
1150 if (unlikely(details
&& details
->ignore_dirty
))
1153 set_page_dirty(page
);
1155 if (pte_young(ptent
) &&
1156 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1157 mark_page_accessed(page
);
1159 rss
[mm_counter(page
)]--;
1160 page_remove_rmap(page
, false);
1161 if (unlikely(page_mapcount(page
) < 0))
1162 print_bad_pte(vma
, addr
, ptent
, page
);
1163 if (unlikely(!__tlb_remove_page(tlb
, page
))) {
1170 /* only check swap_entries if explicitly asked for in details */
1171 if (unlikely(details
&& !details
->check_swap_entries
))
1174 entry
= pte_to_swp_entry(ptent
);
1175 if (!non_swap_entry(entry
))
1177 else if (is_migration_entry(entry
)) {
1180 page
= migration_entry_to_page(entry
);
1181 rss
[mm_counter(page
)]--;
1183 if (unlikely(!free_swap_and_cache(entry
)))
1184 print_bad_pte(vma
, addr
, ptent
, NULL
);
1185 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1186 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1188 add_mm_rss_vec(mm
, rss
);
1189 arch_leave_lazy_mmu_mode();
1191 /* Do the actual TLB flush before dropping ptl */
1193 tlb_flush_mmu_tlbonly(tlb
);
1194 pte_unmap_unlock(start_pte
, ptl
);
1197 * If we forced a TLB flush (either due to running out of
1198 * batch buffers or because we needed to flush dirty TLB
1199 * entries before releasing the ptl), free the batched
1200 * memory too. Restart if we didn't do everything.
1204 tlb_flush_mmu_free(tlb
);
1213 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1214 struct vm_area_struct
*vma
, pud_t
*pud
,
1215 unsigned long addr
, unsigned long end
,
1216 struct zap_details
*details
)
1221 pmd
= pmd_offset(pud
, addr
);
1223 next
= pmd_addr_end(addr
, end
);
1224 if (pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1225 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1226 VM_BUG_ON_VMA(vma_is_anonymous(vma
) &&
1227 !rwsem_is_locked(&tlb
->mm
->mmap_sem
), vma
);
1228 split_huge_pmd(vma
, pmd
, addr
);
1229 } else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1234 * Here there can be other concurrent MADV_DONTNEED or
1235 * trans huge page faults running, and if the pmd is
1236 * none or trans huge it can change under us. This is
1237 * because MADV_DONTNEED holds the mmap_sem in read
1240 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1242 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1245 } while (pmd
++, addr
= next
, addr
!= end
);
1250 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1251 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1252 unsigned long addr
, unsigned long end
,
1253 struct zap_details
*details
)
1258 pud
= pud_offset(pgd
, addr
);
1260 next
= pud_addr_end(addr
, end
);
1261 if (pud_none_or_clear_bad(pud
))
1263 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1264 } while (pud
++, addr
= next
, addr
!= end
);
1269 void unmap_page_range(struct mmu_gather
*tlb
,
1270 struct vm_area_struct
*vma
,
1271 unsigned long addr
, unsigned long end
,
1272 struct zap_details
*details
)
1277 BUG_ON(addr
>= end
);
1278 tlb_start_vma(tlb
, vma
);
1279 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1281 next
= pgd_addr_end(addr
, end
);
1282 if (pgd_none_or_clear_bad(pgd
))
1284 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
1285 } while (pgd
++, addr
= next
, addr
!= end
);
1286 tlb_end_vma(tlb
, vma
);
1290 static void unmap_single_vma(struct mmu_gather
*tlb
,
1291 struct vm_area_struct
*vma
, unsigned long start_addr
,
1292 unsigned long end_addr
,
1293 struct zap_details
*details
)
1295 unsigned long start
= max(vma
->vm_start
, start_addr
);
1298 if (start
>= vma
->vm_end
)
1300 end
= min(vma
->vm_end
, end_addr
);
1301 if (end
<= vma
->vm_start
)
1305 uprobe_munmap(vma
, start
, end
);
1307 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1308 untrack_pfn(vma
, 0, 0);
1311 if (unlikely(is_vm_hugetlb_page(vma
))) {
1313 * It is undesirable to test vma->vm_file as it
1314 * should be non-null for valid hugetlb area.
1315 * However, vm_file will be NULL in the error
1316 * cleanup path of mmap_region. When
1317 * hugetlbfs ->mmap method fails,
1318 * mmap_region() nullifies vma->vm_file
1319 * before calling this function to clean up.
1320 * Since no pte has actually been setup, it is
1321 * safe to do nothing in this case.
1324 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
1325 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1326 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
1329 unmap_page_range(tlb
, vma
, start
, end
, details
);
1334 * unmap_vmas - unmap a range of memory covered by a list of vma's
1335 * @tlb: address of the caller's struct mmu_gather
1336 * @vma: the starting vma
1337 * @start_addr: virtual address at which to start unmapping
1338 * @end_addr: virtual address at which to end unmapping
1340 * Unmap all pages in the vma list.
1342 * Only addresses between `start' and `end' will be unmapped.
1344 * The VMA list must be sorted in ascending virtual address order.
1346 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1347 * range after unmap_vmas() returns. So the only responsibility here is to
1348 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1349 * drops the lock and schedules.
1351 void unmap_vmas(struct mmu_gather
*tlb
,
1352 struct vm_area_struct
*vma
, unsigned long start_addr
,
1353 unsigned long end_addr
)
1355 struct mm_struct
*mm
= vma
->vm_mm
;
1357 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1358 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1359 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1360 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1364 * zap_page_range - remove user pages in a given range
1365 * @vma: vm_area_struct holding the applicable pages
1366 * @start: starting address of pages to zap
1367 * @size: number of bytes to zap
1368 * @details: details of shared cache invalidation
1370 * Caller must protect the VMA list
1372 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1373 unsigned long size
, struct zap_details
*details
)
1375 struct mm_struct
*mm
= vma
->vm_mm
;
1376 struct mmu_gather tlb
;
1377 unsigned long end
= start
+ size
;
1380 tlb_gather_mmu(&tlb
, mm
, start
, end
);
1381 update_hiwater_rss(mm
);
1382 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1383 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
)
1384 unmap_single_vma(&tlb
, vma
, start
, end
, details
);
1385 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1386 tlb_finish_mmu(&tlb
, start
, end
);
1390 * zap_page_range_single - remove user pages in a given range
1391 * @vma: vm_area_struct holding the applicable pages
1392 * @address: starting address of pages to zap
1393 * @size: number of bytes to zap
1394 * @details: details of shared cache invalidation
1396 * The range must fit into one VMA.
1398 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1399 unsigned long size
, struct zap_details
*details
)
1401 struct mm_struct
*mm
= vma
->vm_mm
;
1402 struct mmu_gather tlb
;
1403 unsigned long end
= address
+ size
;
1406 tlb_gather_mmu(&tlb
, mm
, address
, end
);
1407 update_hiwater_rss(mm
);
1408 mmu_notifier_invalidate_range_start(mm
, address
, end
);
1409 unmap_single_vma(&tlb
, vma
, address
, end
, details
);
1410 mmu_notifier_invalidate_range_end(mm
, address
, end
);
1411 tlb_finish_mmu(&tlb
, address
, end
);
1415 * zap_vma_ptes - remove ptes mapping the vma
1416 * @vma: vm_area_struct holding ptes to be zapped
1417 * @address: starting address of pages to zap
1418 * @size: number of bytes to zap
1420 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1422 * The entire address range must be fully contained within the vma.
1424 * Returns 0 if successful.
1426 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1429 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1430 !(vma
->vm_flags
& VM_PFNMAP
))
1432 zap_page_range_single(vma
, address
, size
, NULL
);
1435 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1437 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1440 pgd_t
* pgd
= pgd_offset(mm
, addr
);
1441 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
1443 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
1445 VM_BUG_ON(pmd_trans_huge(*pmd
));
1446 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1453 * This is the old fallback for page remapping.
1455 * For historical reasons, it only allows reserved pages. Only
1456 * old drivers should use this, and they needed to mark their
1457 * pages reserved for the old functions anyway.
1459 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1460 struct page
*page
, pgprot_t prot
)
1462 struct mm_struct
*mm
= vma
->vm_mm
;
1471 flush_dcache_page(page
);
1472 pte
= get_locked_pte(mm
, addr
, &ptl
);
1476 if (!pte_none(*pte
))
1479 /* Ok, finally just insert the thing.. */
1481 inc_mm_counter_fast(mm
, mm_counter_file(page
));
1482 page_add_file_rmap(page
);
1483 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1486 pte_unmap_unlock(pte
, ptl
);
1489 pte_unmap_unlock(pte
, ptl
);
1495 * vm_insert_page - insert single page into user vma
1496 * @vma: user vma to map to
1497 * @addr: target user address of this page
1498 * @page: source kernel page
1500 * This allows drivers to insert individual pages they've allocated
1503 * The page has to be a nice clean _individual_ kernel allocation.
1504 * If you allocate a compound page, you need to have marked it as
1505 * such (__GFP_COMP), or manually just split the page up yourself
1506 * (see split_page()).
1508 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1509 * took an arbitrary page protection parameter. This doesn't allow
1510 * that. Your vma protection will have to be set up correctly, which
1511 * means that if you want a shared writable mapping, you'd better
1512 * ask for a shared writable mapping!
1514 * The page does not need to be reserved.
1516 * Usually this function is called from f_op->mmap() handler
1517 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1518 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1519 * function from other places, for example from page-fault handler.
1521 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1524 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1526 if (!page_count(page
))
1528 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1529 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
1530 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1531 vma
->vm_flags
|= VM_MIXEDMAP
;
1533 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1535 EXPORT_SYMBOL(vm_insert_page
);
1537 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1538 pfn_t pfn
, pgprot_t prot
)
1540 struct mm_struct
*mm
= vma
->vm_mm
;
1546 pte
= get_locked_pte(mm
, addr
, &ptl
);
1550 if (!pte_none(*pte
))
1553 /* Ok, finally just insert the thing.. */
1554 if (pfn_t_devmap(pfn
))
1555 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
1557 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
1558 set_pte_at(mm
, addr
, pte
, entry
);
1559 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1563 pte_unmap_unlock(pte
, ptl
);
1569 * vm_insert_pfn - insert single pfn into user vma
1570 * @vma: user vma to map to
1571 * @addr: target user address of this page
1572 * @pfn: source kernel pfn
1574 * Similar to vm_insert_page, this allows drivers to insert individual pages
1575 * they've allocated into a user vma. Same comments apply.
1577 * This function should only be called from a vm_ops->fault handler, and
1578 * in that case the handler should return NULL.
1580 * vma cannot be a COW mapping.
1582 * As this is called only for pages that do not currently exist, we
1583 * do not need to flush old virtual caches or the TLB.
1585 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1588 return vm_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
1590 EXPORT_SYMBOL(vm_insert_pfn
);
1593 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1594 * @vma: user vma to map to
1595 * @addr: target user address of this page
1596 * @pfn: source kernel pfn
1597 * @pgprot: pgprot flags for the inserted page
1599 * This is exactly like vm_insert_pfn, except that it allows drivers to
1600 * to override pgprot on a per-page basis.
1602 * This only makes sense for IO mappings, and it makes no sense for
1603 * cow mappings. In general, using multiple vmas is preferable;
1604 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1607 int vm_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
1608 unsigned long pfn
, pgprot_t pgprot
)
1612 * Technically, architectures with pte_special can avoid all these
1613 * restrictions (same for remap_pfn_range). However we would like
1614 * consistency in testing and feature parity among all, so we should
1615 * try to keep these invariants in place for everybody.
1617 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1618 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1619 (VM_PFNMAP
|VM_MIXEDMAP
));
1620 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1621 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1623 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1625 if (track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
)))
1628 ret
= insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
);
1632 EXPORT_SYMBOL(vm_insert_pfn_prot
);
1634 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1637 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
1639 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1643 * If we don't have pte special, then we have to use the pfn_valid()
1644 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1645 * refcount the page if pfn_valid is true (hence insert_page rather
1646 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1647 * without pte special, it would there be refcounted as a normal page.
1649 if (!HAVE_PTE_SPECIAL
&& !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
1653 * At this point we are committed to insert_page()
1654 * regardless of whether the caller specified flags that
1655 * result in pfn_t_has_page() == false.
1657 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
1658 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1660 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1662 EXPORT_SYMBOL(vm_insert_mixed
);
1665 * maps a range of physical memory into the requested pages. the old
1666 * mappings are removed. any references to nonexistent pages results
1667 * in null mappings (currently treated as "copy-on-access")
1669 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1670 unsigned long addr
, unsigned long end
,
1671 unsigned long pfn
, pgprot_t prot
)
1676 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1679 arch_enter_lazy_mmu_mode();
1681 BUG_ON(!pte_none(*pte
));
1682 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
1684 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1685 arch_leave_lazy_mmu_mode();
1686 pte_unmap_unlock(pte
- 1, ptl
);
1690 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1691 unsigned long addr
, unsigned long end
,
1692 unsigned long pfn
, pgprot_t prot
)
1697 pfn
-= addr
>> PAGE_SHIFT
;
1698 pmd
= pmd_alloc(mm
, pud
, addr
);
1701 VM_BUG_ON(pmd_trans_huge(*pmd
));
1703 next
= pmd_addr_end(addr
, end
);
1704 if (remap_pte_range(mm
, pmd
, addr
, next
,
1705 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1707 } while (pmd
++, addr
= next
, addr
!= end
);
1711 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1712 unsigned long addr
, unsigned long end
,
1713 unsigned long pfn
, pgprot_t prot
)
1718 pfn
-= addr
>> PAGE_SHIFT
;
1719 pud
= pud_alloc(mm
, pgd
, addr
);
1723 next
= pud_addr_end(addr
, end
);
1724 if (remap_pmd_range(mm
, pud
, addr
, next
,
1725 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1727 } while (pud
++, addr
= next
, addr
!= end
);
1732 * remap_pfn_range - remap kernel memory to userspace
1733 * @vma: user vma to map to
1734 * @addr: target user address to start at
1735 * @pfn: physical address of kernel memory
1736 * @size: size of map area
1737 * @prot: page protection flags for this mapping
1739 * Note: this is only safe if the mm semaphore is held when called.
1741 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1742 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1746 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1747 struct mm_struct
*mm
= vma
->vm_mm
;
1748 unsigned long remap_pfn
= pfn
;
1752 * Physically remapped pages are special. Tell the
1753 * rest of the world about it:
1754 * VM_IO tells people not to look at these pages
1755 * (accesses can have side effects).
1756 * VM_PFNMAP tells the core MM that the base pages are just
1757 * raw PFN mappings, and do not have a "struct page" associated
1760 * Disable vma merging and expanding with mremap().
1762 * Omit vma from core dump, even when VM_IO turned off.
1764 * There's a horrible special case to handle copy-on-write
1765 * behaviour that some programs depend on. We mark the "original"
1766 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1767 * See vm_normal_page() for details.
1769 if (is_cow_mapping(vma
->vm_flags
)) {
1770 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
1772 vma
->vm_pgoff
= pfn
;
1775 err
= track_pfn_remap(vma
, &prot
, remap_pfn
, addr
, PAGE_ALIGN(size
));
1779 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
1781 BUG_ON(addr
>= end
);
1782 pfn
-= addr
>> PAGE_SHIFT
;
1783 pgd
= pgd_offset(mm
, addr
);
1784 flush_cache_range(vma
, addr
, end
);
1786 next
= pgd_addr_end(addr
, end
);
1787 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1788 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1791 } while (pgd
++, addr
= next
, addr
!= end
);
1794 untrack_pfn(vma
, remap_pfn
, PAGE_ALIGN(size
));
1798 EXPORT_SYMBOL(remap_pfn_range
);
1801 * vm_iomap_memory - remap memory to userspace
1802 * @vma: user vma to map to
1803 * @start: start of area
1804 * @len: size of area
1806 * This is a simplified io_remap_pfn_range() for common driver use. The
1807 * driver just needs to give us the physical memory range to be mapped,
1808 * we'll figure out the rest from the vma information.
1810 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1811 * whatever write-combining details or similar.
1813 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
1815 unsigned long vm_len
, pfn
, pages
;
1817 /* Check that the physical memory area passed in looks valid */
1818 if (start
+ len
< start
)
1821 * You *really* shouldn't map things that aren't page-aligned,
1822 * but we've historically allowed it because IO memory might
1823 * just have smaller alignment.
1825 len
+= start
& ~PAGE_MASK
;
1826 pfn
= start
>> PAGE_SHIFT
;
1827 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
1828 if (pfn
+ pages
< pfn
)
1831 /* We start the mapping 'vm_pgoff' pages into the area */
1832 if (vma
->vm_pgoff
> pages
)
1834 pfn
+= vma
->vm_pgoff
;
1835 pages
-= vma
->vm_pgoff
;
1837 /* Can we fit all of the mapping? */
1838 vm_len
= vma
->vm_end
- vma
->vm_start
;
1839 if (vm_len
>> PAGE_SHIFT
> pages
)
1842 /* Ok, let it rip */
1843 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
1845 EXPORT_SYMBOL(vm_iomap_memory
);
1847 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1848 unsigned long addr
, unsigned long end
,
1849 pte_fn_t fn
, void *data
)
1854 spinlock_t
*uninitialized_var(ptl
);
1856 pte
= (mm
== &init_mm
) ?
1857 pte_alloc_kernel(pmd
, addr
) :
1858 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1862 BUG_ON(pmd_huge(*pmd
));
1864 arch_enter_lazy_mmu_mode();
1866 token
= pmd_pgtable(*pmd
);
1869 err
= fn(pte
++, token
, addr
, data
);
1872 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1874 arch_leave_lazy_mmu_mode();
1877 pte_unmap_unlock(pte
-1, ptl
);
1881 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1882 unsigned long addr
, unsigned long end
,
1883 pte_fn_t fn
, void *data
)
1889 BUG_ON(pud_huge(*pud
));
1891 pmd
= pmd_alloc(mm
, pud
, addr
);
1895 next
= pmd_addr_end(addr
, end
);
1896 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
1899 } while (pmd
++, addr
= next
, addr
!= end
);
1903 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1904 unsigned long addr
, unsigned long end
,
1905 pte_fn_t fn
, void *data
)
1911 pud
= pud_alloc(mm
, pgd
, addr
);
1915 next
= pud_addr_end(addr
, end
);
1916 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
1919 } while (pud
++, addr
= next
, addr
!= end
);
1924 * Scan a region of virtual memory, filling in page tables as necessary
1925 * and calling a provided function on each leaf page table.
1927 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
1928 unsigned long size
, pte_fn_t fn
, void *data
)
1932 unsigned long end
= addr
+ size
;
1935 if (WARN_ON(addr
>= end
))
1938 pgd
= pgd_offset(mm
, addr
);
1940 next
= pgd_addr_end(addr
, end
);
1941 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
1944 } while (pgd
++, addr
= next
, addr
!= end
);
1948 EXPORT_SYMBOL_GPL(apply_to_page_range
);
1951 * handle_pte_fault chooses page fault handler according to an entry which was
1952 * read non-atomically. Before making any commitment, on those architectures
1953 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1954 * parts, do_swap_page must check under lock before unmapping the pte and
1955 * proceeding (but do_wp_page is only called after already making such a check;
1956 * and do_anonymous_page can safely check later on).
1958 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
1959 pte_t
*page_table
, pte_t orig_pte
)
1962 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1963 if (sizeof(pte_t
) > sizeof(unsigned long)) {
1964 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
1966 same
= pte_same(*page_table
, orig_pte
);
1970 pte_unmap(page_table
);
1974 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
1976 debug_dma_assert_idle(src
);
1979 * If the source page was a PFN mapping, we don't have
1980 * a "struct page" for it. We do a best-effort copy by
1981 * just copying from the original user address. If that
1982 * fails, we just zero-fill it. Live with it.
1984 if (unlikely(!src
)) {
1985 void *kaddr
= kmap_atomic(dst
);
1986 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
1989 * This really shouldn't fail, because the page is there
1990 * in the page tables. But it might just be unreadable,
1991 * in which case we just give up and fill the result with
1994 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
1996 kunmap_atomic(kaddr
);
1997 flush_dcache_page(dst
);
1999 copy_user_highpage(dst
, src
, va
, vma
);
2002 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
2004 struct file
*vm_file
= vma
->vm_file
;
2007 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
2010 * Special mappings (e.g. VDSO) do not have any file so fake
2011 * a default GFP_KERNEL for them.
2017 * Notify the address space that the page is about to become writable so that
2018 * it can prohibit this or wait for the page to get into an appropriate state.
2020 * We do this without the lock held, so that it can sleep if it needs to.
2022 static int do_page_mkwrite(struct vm_area_struct
*vma
, struct page
*page
,
2023 unsigned long address
)
2025 struct vm_fault vmf
;
2028 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
2029 vmf
.pgoff
= page
->index
;
2030 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2031 vmf
.gfp_mask
= __get_fault_gfp_mask(vma
);
2033 vmf
.cow_page
= NULL
;
2035 ret
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
2036 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2038 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
2040 if (!page
->mapping
) {
2042 return 0; /* retry */
2044 ret
|= VM_FAULT_LOCKED
;
2046 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2051 * Handle write page faults for pages that can be reused in the current vma
2053 * This can happen either due to the mapping being with the VM_SHARED flag,
2054 * or due to us being the last reference standing to the page. In either
2055 * case, all we need to do here is to mark the page as writable and update
2056 * any related book-keeping.
2058 static inline int wp_page_reuse(struct mm_struct
*mm
,
2059 struct vm_area_struct
*vma
, unsigned long address
,
2060 pte_t
*page_table
, spinlock_t
*ptl
, pte_t orig_pte
,
2061 struct page
*page
, int page_mkwrite
,
2067 * Clear the pages cpupid information as the existing
2068 * information potentially belongs to a now completely
2069 * unrelated process.
2072 page_cpupid_xchg_last(page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2074 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2075 entry
= pte_mkyoung(orig_pte
);
2076 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2077 if (ptep_set_access_flags(vma
, address
, page_table
, entry
, 1))
2078 update_mmu_cache(vma
, address
, page_table
);
2079 pte_unmap_unlock(page_table
, ptl
);
2082 struct address_space
*mapping
;
2088 dirtied
= set_page_dirty(page
);
2089 VM_BUG_ON_PAGE(PageAnon(page
), page
);
2090 mapping
= page
->mapping
;
2094 if ((dirtied
|| page_mkwrite
) && mapping
) {
2096 * Some device drivers do not set page.mapping
2097 * but still dirty their pages
2099 balance_dirty_pages_ratelimited(mapping
);
2103 file_update_time(vma
->vm_file
);
2106 return VM_FAULT_WRITE
;
2110 * Handle the case of a page which we actually need to copy to a new page.
2112 * Called with mmap_sem locked and the old page referenced, but
2113 * without the ptl held.
2115 * High level logic flow:
2117 * - Allocate a page, copy the content of the old page to the new one.
2118 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2119 * - Take the PTL. If the pte changed, bail out and release the allocated page
2120 * - If the pte is still the way we remember it, update the page table and all
2121 * relevant references. This includes dropping the reference the page-table
2122 * held to the old page, as well as updating the rmap.
2123 * - In any case, unlock the PTL and drop the reference we took to the old page.
2125 static int wp_page_copy(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2126 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2127 pte_t orig_pte
, struct page
*old_page
)
2129 struct page
*new_page
= NULL
;
2130 spinlock_t
*ptl
= NULL
;
2132 int page_copied
= 0;
2133 const unsigned long mmun_start
= address
& PAGE_MASK
; /* For mmu_notifiers */
2134 const unsigned long mmun_end
= mmun_start
+ PAGE_SIZE
; /* For mmu_notifiers */
2135 struct mem_cgroup
*memcg
;
2137 if (unlikely(anon_vma_prepare(vma
)))
2140 if (is_zero_pfn(pte_pfn(orig_pte
))) {
2141 new_page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2145 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2148 cow_user_page(new_page
, old_page
, address
, vma
);
2151 if (mem_cgroup_try_charge(new_page
, mm
, GFP_KERNEL
, &memcg
, false))
2154 __SetPageUptodate(new_page
);
2156 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2159 * Re-check the pte - we dropped the lock
2161 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2162 if (likely(pte_same(*page_table
, orig_pte
))) {
2164 if (!PageAnon(old_page
)) {
2165 dec_mm_counter_fast(mm
,
2166 mm_counter_file(old_page
));
2167 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2170 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2172 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2173 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2174 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2176 * Clear the pte entry and flush it first, before updating the
2177 * pte with the new entry. This will avoid a race condition
2178 * seen in the presence of one thread doing SMC and another
2181 ptep_clear_flush_notify(vma
, address
, page_table
);
2182 page_add_new_anon_rmap(new_page
, vma
, address
, false);
2183 mem_cgroup_commit_charge(new_page
, memcg
, false, false);
2184 lru_cache_add_active_or_unevictable(new_page
, vma
);
2186 * We call the notify macro here because, when using secondary
2187 * mmu page tables (such as kvm shadow page tables), we want the
2188 * new page to be mapped directly into the secondary page table.
2190 set_pte_at_notify(mm
, address
, page_table
, entry
);
2191 update_mmu_cache(vma
, address
, page_table
);
2194 * Only after switching the pte to the new page may
2195 * we remove the mapcount here. Otherwise another
2196 * process may come and find the rmap count decremented
2197 * before the pte is switched to the new page, and
2198 * "reuse" the old page writing into it while our pte
2199 * here still points into it and can be read by other
2202 * The critical issue is to order this
2203 * page_remove_rmap with the ptp_clear_flush above.
2204 * Those stores are ordered by (if nothing else,)
2205 * the barrier present in the atomic_add_negative
2206 * in page_remove_rmap.
2208 * Then the TLB flush in ptep_clear_flush ensures that
2209 * no process can access the old page before the
2210 * decremented mapcount is visible. And the old page
2211 * cannot be reused until after the decremented
2212 * mapcount is visible. So transitively, TLBs to
2213 * old page will be flushed before it can be reused.
2215 page_remove_rmap(old_page
, false);
2218 /* Free the old page.. */
2219 new_page
= old_page
;
2222 mem_cgroup_cancel_charge(new_page
, memcg
, false);
2228 pte_unmap_unlock(page_table
, ptl
);
2229 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2232 * Don't let another task, with possibly unlocked vma,
2233 * keep the mlocked page.
2235 if (page_copied
&& (vma
->vm_flags
& VM_LOCKED
)) {
2236 lock_page(old_page
); /* LRU manipulation */
2237 if (PageMlocked(old_page
))
2238 munlock_vma_page(old_page
);
2239 unlock_page(old_page
);
2243 return page_copied
? VM_FAULT_WRITE
: 0;
2249 return VM_FAULT_OOM
;
2253 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2256 static int wp_pfn_shared(struct mm_struct
*mm
,
2257 struct vm_area_struct
*vma
, unsigned long address
,
2258 pte_t
*page_table
, spinlock_t
*ptl
, pte_t orig_pte
,
2261 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
2262 struct vm_fault vmf
= {
2264 .pgoff
= linear_page_index(vma
, address
),
2265 .virtual_address
= (void __user
*)(address
& PAGE_MASK
),
2266 .flags
= FAULT_FLAG_WRITE
| FAULT_FLAG_MKWRITE
,
2270 pte_unmap_unlock(page_table
, ptl
);
2271 ret
= vma
->vm_ops
->pfn_mkwrite(vma
, &vmf
);
2272 if (ret
& VM_FAULT_ERROR
)
2274 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2276 * We might have raced with another page fault while we
2277 * released the pte_offset_map_lock.
2279 if (!pte_same(*page_table
, orig_pte
)) {
2280 pte_unmap_unlock(page_table
, ptl
);
2284 return wp_page_reuse(mm
, vma
, address
, page_table
, ptl
, orig_pte
,
2288 static int wp_page_shared(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2289 unsigned long address
, pte_t
*page_table
,
2290 pmd_t
*pmd
, spinlock_t
*ptl
, pte_t orig_pte
,
2291 struct page
*old_page
)
2294 int page_mkwrite
= 0;
2298 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2301 pte_unmap_unlock(page_table
, ptl
);
2302 tmp
= do_page_mkwrite(vma
, old_page
, address
);
2303 if (unlikely(!tmp
|| (tmp
&
2304 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
2309 * Since we dropped the lock we need to revalidate
2310 * the PTE as someone else may have changed it. If
2311 * they did, we just return, as we can count on the
2312 * MMU to tell us if they didn't also make it writable.
2314 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2316 if (!pte_same(*page_table
, orig_pte
)) {
2317 unlock_page(old_page
);
2318 pte_unmap_unlock(page_table
, ptl
);
2325 return wp_page_reuse(mm
, vma
, address
, page_table
, ptl
,
2326 orig_pte
, old_page
, page_mkwrite
, 1);
2330 * This routine handles present pages, when users try to write
2331 * to a shared page. It is done by copying the page to a new address
2332 * and decrementing the shared-page counter for the old page.
2334 * Note that this routine assumes that the protection checks have been
2335 * done by the caller (the low-level page fault routine in most cases).
2336 * Thus we can safely just mark it writable once we've done any necessary
2339 * We also mark the page dirty at this point even though the page will
2340 * change only once the write actually happens. This avoids a few races,
2341 * and potentially makes it more efficient.
2343 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2344 * but allow concurrent faults), with pte both mapped and locked.
2345 * We return with mmap_sem still held, but pte unmapped and unlocked.
2347 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2348 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2349 spinlock_t
*ptl
, pte_t orig_pte
)
2352 struct page
*old_page
;
2354 old_page
= vm_normal_page(vma
, address
, orig_pte
);
2357 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2360 * We should not cow pages in a shared writeable mapping.
2361 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2363 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2364 (VM_WRITE
|VM_SHARED
))
2365 return wp_pfn_shared(mm
, vma
, address
, page_table
, ptl
,
2368 pte_unmap_unlock(page_table
, ptl
);
2369 return wp_page_copy(mm
, vma
, address
, page_table
, pmd
,
2370 orig_pte
, old_page
);
2374 * Take out anonymous pages first, anonymous shared vmas are
2375 * not dirty accountable.
2377 if (PageAnon(old_page
) && !PageKsm(old_page
)) {
2379 if (!trylock_page(old_page
)) {
2381 pte_unmap_unlock(page_table
, ptl
);
2382 lock_page(old_page
);
2383 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2385 if (!pte_same(*page_table
, orig_pte
)) {
2386 unlock_page(old_page
);
2387 pte_unmap_unlock(page_table
, ptl
);
2393 if (reuse_swap_page(old_page
, &total_mapcount
)) {
2394 if (total_mapcount
== 1) {
2396 * The page is all ours. Move it to
2397 * our anon_vma so the rmap code will
2398 * not search our parent or siblings.
2399 * Protected against the rmap code by
2402 page_move_anon_rmap(compound_head(old_page
),
2405 unlock_page(old_page
);
2406 return wp_page_reuse(mm
, vma
, address
, page_table
, ptl
,
2407 orig_pte
, old_page
, 0, 0);
2409 unlock_page(old_page
);
2410 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2411 (VM_WRITE
|VM_SHARED
))) {
2412 return wp_page_shared(mm
, vma
, address
, page_table
, pmd
,
2413 ptl
, orig_pte
, old_page
);
2417 * Ok, we need to copy. Oh, well..
2421 pte_unmap_unlock(page_table
, ptl
);
2422 return wp_page_copy(mm
, vma
, address
, page_table
, pmd
,
2423 orig_pte
, old_page
);
2426 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2427 unsigned long start_addr
, unsigned long end_addr
,
2428 struct zap_details
*details
)
2430 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2433 static inline void unmap_mapping_range_tree(struct rb_root
*root
,
2434 struct zap_details
*details
)
2436 struct vm_area_struct
*vma
;
2437 pgoff_t vba
, vea
, zba
, zea
;
2439 vma_interval_tree_foreach(vma
, root
,
2440 details
->first_index
, details
->last_index
) {
2442 vba
= vma
->vm_pgoff
;
2443 vea
= vba
+ vma_pages(vma
) - 1;
2444 zba
= details
->first_index
;
2447 zea
= details
->last_index
;
2451 unmap_mapping_range_vma(vma
,
2452 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2453 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2459 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2460 * address_space corresponding to the specified page range in the underlying
2463 * @mapping: the address space containing mmaps to be unmapped.
2464 * @holebegin: byte in first page to unmap, relative to the start of
2465 * the underlying file. This will be rounded down to a PAGE_SIZE
2466 * boundary. Note that this is different from truncate_pagecache(), which
2467 * must keep the partial page. In contrast, we must get rid of
2469 * @holelen: size of prospective hole in bytes. This will be rounded
2470 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2472 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2473 * but 0 when invalidating pagecache, don't throw away private data.
2475 void unmap_mapping_range(struct address_space
*mapping
,
2476 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2478 struct zap_details details
= { };
2479 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2480 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2482 /* Check for overflow. */
2483 if (sizeof(holelen
) > sizeof(hlen
)) {
2485 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2486 if (holeend
& ~(long long)ULONG_MAX
)
2487 hlen
= ULONG_MAX
- hba
+ 1;
2490 details
.check_mapping
= even_cows
? NULL
: mapping
;
2491 details
.first_index
= hba
;
2492 details
.last_index
= hba
+ hlen
- 1;
2493 if (details
.last_index
< details
.first_index
)
2494 details
.last_index
= ULONG_MAX
;
2496 i_mmap_lock_write(mapping
);
2497 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
)))
2498 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2499 i_mmap_unlock_write(mapping
);
2501 EXPORT_SYMBOL(unmap_mapping_range
);
2504 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2505 * but allow concurrent faults), and pte mapped but not yet locked.
2506 * We return with pte unmapped and unlocked.
2508 * We return with the mmap_sem locked or unlocked in the same cases
2509 * as does filemap_fault().
2511 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2512 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2513 unsigned int flags
, pte_t orig_pte
)
2516 struct page
*page
, *swapcache
;
2517 struct mem_cgroup
*memcg
;
2524 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2527 entry
= pte_to_swp_entry(orig_pte
);
2528 if (unlikely(non_swap_entry(entry
))) {
2529 if (is_migration_entry(entry
)) {
2530 migration_entry_wait(mm
, pmd
, address
);
2531 } else if (is_hwpoison_entry(entry
)) {
2532 ret
= VM_FAULT_HWPOISON
;
2534 print_bad_pte(vma
, address
, orig_pte
, NULL
);
2535 ret
= VM_FAULT_SIGBUS
;
2539 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2540 page
= lookup_swap_cache(entry
);
2542 page
= swapin_readahead(entry
,
2543 GFP_HIGHUSER_MOVABLE
, vma
, address
);
2546 * Back out if somebody else faulted in this pte
2547 * while we released the pte lock.
2549 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2550 if (likely(pte_same(*page_table
, orig_pte
)))
2552 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2556 /* Had to read the page from swap area: Major fault */
2557 ret
= VM_FAULT_MAJOR
;
2558 count_vm_event(PGMAJFAULT
);
2559 mem_cgroup_count_vm_event(mm
, PGMAJFAULT
);
2560 } else if (PageHWPoison(page
)) {
2562 * hwpoisoned dirty swapcache pages are kept for killing
2563 * owner processes (which may be unknown at hwpoison time)
2565 ret
= VM_FAULT_HWPOISON
;
2566 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2572 locked
= lock_page_or_retry(page
, mm
, flags
);
2574 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2576 ret
|= VM_FAULT_RETRY
;
2581 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2582 * release the swapcache from under us. The page pin, and pte_same
2583 * test below, are not enough to exclude that. Even if it is still
2584 * swapcache, we need to check that the page's swap has not changed.
2586 if (unlikely(!PageSwapCache(page
) || page_private(page
) != entry
.val
))
2589 page
= ksm_might_need_to_copy(page
, vma
, address
);
2590 if (unlikely(!page
)) {
2596 if (mem_cgroup_try_charge(page
, mm
, GFP_KERNEL
, &memcg
, false)) {
2602 * Back out if somebody else already faulted in this pte.
2604 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2605 if (unlikely(!pte_same(*page_table
, orig_pte
)))
2608 if (unlikely(!PageUptodate(page
))) {
2609 ret
= VM_FAULT_SIGBUS
;
2614 * The page isn't present yet, go ahead with the fault.
2616 * Be careful about the sequence of operations here.
2617 * To get its accounting right, reuse_swap_page() must be called
2618 * while the page is counted on swap but not yet in mapcount i.e.
2619 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2620 * must be called after the swap_free(), or it will never succeed.
2623 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2624 dec_mm_counter_fast(mm
, MM_SWAPENTS
);
2625 pte
= mk_pte(page
, vma
->vm_page_prot
);
2626 if ((flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
, NULL
)) {
2627 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
2628 flags
&= ~FAULT_FLAG_WRITE
;
2629 ret
|= VM_FAULT_WRITE
;
2630 exclusive
= RMAP_EXCLUSIVE
;
2632 flush_icache_page(vma
, page
);
2633 if (pte_swp_soft_dirty(orig_pte
))
2634 pte
= pte_mksoft_dirty(pte
);
2635 set_pte_at(mm
, address
, page_table
, pte
);
2636 if (page
== swapcache
) {
2637 do_page_add_anon_rmap(page
, vma
, address
, exclusive
);
2638 mem_cgroup_commit_charge(page
, memcg
, true, false);
2639 } else { /* ksm created a completely new copy */
2640 page_add_new_anon_rmap(page
, vma
, address
, false);
2641 mem_cgroup_commit_charge(page
, memcg
, false, false);
2642 lru_cache_add_active_or_unevictable(page
, vma
);
2646 if (mem_cgroup_swap_full(page
) ||
2647 (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
2648 try_to_free_swap(page
);
2650 if (page
!= swapcache
) {
2652 * Hold the lock to avoid the swap entry to be reused
2653 * until we take the PT lock for the pte_same() check
2654 * (to avoid false positives from pte_same). For
2655 * further safety release the lock after the swap_free
2656 * so that the swap count won't change under a
2657 * parallel locked swapcache.
2659 unlock_page(swapcache
);
2660 put_page(swapcache
);
2663 if (flags
& FAULT_FLAG_WRITE
) {
2664 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
2665 if (ret
& VM_FAULT_ERROR
)
2666 ret
&= VM_FAULT_ERROR
;
2670 /* No need to invalidate - it was non-present before */
2671 update_mmu_cache(vma
, address
, page_table
);
2673 pte_unmap_unlock(page_table
, ptl
);
2677 mem_cgroup_cancel_charge(page
, memcg
, false);
2678 pte_unmap_unlock(page_table
, ptl
);
2683 if (page
!= swapcache
) {
2684 unlock_page(swapcache
);
2685 put_page(swapcache
);
2691 * This is like a special single-page "expand_{down|up}wards()",
2692 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2693 * doesn't hit another vma.
2695 static inline int check_stack_guard_page(struct vm_area_struct
*vma
, unsigned long address
)
2697 address
&= PAGE_MASK
;
2698 if ((vma
->vm_flags
& VM_GROWSDOWN
) && address
== vma
->vm_start
) {
2699 struct vm_area_struct
*prev
= vma
->vm_prev
;
2702 * Is there a mapping abutting this one below?
2704 * That's only ok if it's the same stack mapping
2705 * that has gotten split..
2707 if (prev
&& prev
->vm_end
== address
)
2708 return prev
->vm_flags
& VM_GROWSDOWN
? 0 : -ENOMEM
;
2710 return expand_downwards(vma
, address
- PAGE_SIZE
);
2712 if ((vma
->vm_flags
& VM_GROWSUP
) && address
+ PAGE_SIZE
== vma
->vm_end
) {
2713 struct vm_area_struct
*next
= vma
->vm_next
;
2715 /* As VM_GROWSDOWN but s/below/above/ */
2716 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
)
2717 return next
->vm_flags
& VM_GROWSUP
? 0 : -ENOMEM
;
2719 return expand_upwards(vma
, address
+ PAGE_SIZE
);
2725 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2726 * but allow concurrent faults), and pte mapped but not yet locked.
2727 * We return with mmap_sem still held, but pte unmapped and unlocked.
2729 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2730 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2733 struct mem_cgroup
*memcg
;
2738 pte_unmap(page_table
);
2740 /* File mapping without ->vm_ops ? */
2741 if (vma
->vm_flags
& VM_SHARED
)
2742 return VM_FAULT_SIGBUS
;
2744 /* Check if we need to add a guard page to the stack */
2745 if (check_stack_guard_page(vma
, address
) < 0)
2746 return VM_FAULT_SIGSEGV
;
2748 /* Use the zero-page for reads */
2749 if (!(flags
& FAULT_FLAG_WRITE
) && !mm_forbids_zeropage(mm
)) {
2750 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(address
),
2751 vma
->vm_page_prot
));
2752 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2753 if (!pte_none(*page_table
))
2755 /* Deliver the page fault to userland, check inside PT lock */
2756 if (userfaultfd_missing(vma
)) {
2757 pte_unmap_unlock(page_table
, ptl
);
2758 return handle_userfault(vma
, address
, flags
,
2764 /* Allocate our own private page. */
2765 if (unlikely(anon_vma_prepare(vma
)))
2767 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2771 if (mem_cgroup_try_charge(page
, mm
, GFP_KERNEL
, &memcg
, false))
2775 * The memory barrier inside __SetPageUptodate makes sure that
2776 * preceeding stores to the page contents become visible before
2777 * the set_pte_at() write.
2779 __SetPageUptodate(page
);
2781 entry
= mk_pte(page
, vma
->vm_page_prot
);
2782 if (vma
->vm_flags
& VM_WRITE
)
2783 entry
= pte_mkwrite(pte_mkdirty(entry
));
2785 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2786 if (!pte_none(*page_table
))
2789 /* Deliver the page fault to userland, check inside PT lock */
2790 if (userfaultfd_missing(vma
)) {
2791 pte_unmap_unlock(page_table
, ptl
);
2792 mem_cgroup_cancel_charge(page
, memcg
, false);
2794 return handle_userfault(vma
, address
, flags
,
2798 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2799 page_add_new_anon_rmap(page
, vma
, address
, false);
2800 mem_cgroup_commit_charge(page
, memcg
, false, false);
2801 lru_cache_add_active_or_unevictable(page
, vma
);
2803 set_pte_at(mm
, address
, page_table
, entry
);
2805 /* No need to invalidate - it was non-present before */
2806 update_mmu_cache(vma
, address
, page_table
);
2808 pte_unmap_unlock(page_table
, ptl
);
2811 mem_cgroup_cancel_charge(page
, memcg
, false);
2817 return VM_FAULT_OOM
;
2821 * The mmap_sem must have been held on entry, and may have been
2822 * released depending on flags and vma->vm_ops->fault() return value.
2823 * See filemap_fault() and __lock_page_retry().
2825 static int __do_fault(struct vm_area_struct
*vma
, unsigned long address
,
2826 pgoff_t pgoff
, unsigned int flags
,
2827 struct page
*cow_page
, struct page
**page
,
2830 struct vm_fault vmf
;
2833 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
2837 vmf
.gfp_mask
= __get_fault_gfp_mask(vma
);
2838 vmf
.cow_page
= cow_page
;
2840 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
2841 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
2843 if (ret
& VM_FAULT_DAX_LOCKED
) {
2848 if (unlikely(PageHWPoison(vmf
.page
))) {
2849 if (ret
& VM_FAULT_LOCKED
)
2850 unlock_page(vmf
.page
);
2852 return VM_FAULT_HWPOISON
;
2855 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
2856 lock_page(vmf
.page
);
2858 VM_BUG_ON_PAGE(!PageLocked(vmf
.page
), vmf
.page
);
2865 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2867 * @vma: virtual memory area
2868 * @address: user virtual address
2869 * @page: page to map
2870 * @pte: pointer to target page table entry
2871 * @write: true, if new entry is writable
2872 * @anon: true, if it's anonymous page
2874 * Caller must hold page table lock relevant for @pte.
2876 * Target users are page handler itself and implementations of
2877 * vm_ops->map_pages.
2879 void do_set_pte(struct vm_area_struct
*vma
, unsigned long address
,
2880 struct page
*page
, pte_t
*pte
, bool write
, bool anon
, bool old
)
2884 flush_icache_page(vma
, page
);
2885 entry
= mk_pte(page
, vma
->vm_page_prot
);
2887 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2889 entry
= pte_mkold(entry
);
2891 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
2892 page_add_new_anon_rmap(page
, vma
, address
, false);
2894 inc_mm_counter_fast(vma
->vm_mm
, mm_counter_file(page
));
2895 page_add_file_rmap(page
);
2897 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
2899 /* no need to invalidate: a not-present page won't be cached */
2900 update_mmu_cache(vma
, address
, pte
);
2904 * If architecture emulates "accessed" or "young" bit without HW support,
2905 * there is no much gain with fault_around.
2907 static unsigned long fault_around_bytes __read_mostly
=
2908 #ifndef __HAVE_ARCH_PTEP_SET_ACCESS_FLAGS
2911 rounddown_pow_of_two(65536);
2914 #ifdef CONFIG_DEBUG_FS
2915 static int fault_around_bytes_get(void *data
, u64
*val
)
2917 *val
= fault_around_bytes
;
2922 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2923 * rounded down to nearest page order. It's what do_fault_around() expects to
2926 static int fault_around_bytes_set(void *data
, u64 val
)
2928 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
2930 if (val
> PAGE_SIZE
)
2931 fault_around_bytes
= rounddown_pow_of_two(val
);
2933 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
2936 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops
,
2937 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
2939 static int __init
fault_around_debugfs(void)
2943 ret
= debugfs_create_file("fault_around_bytes", 0644, NULL
, NULL
,
2944 &fault_around_bytes_fops
);
2946 pr_warn("Failed to create fault_around_bytes in debugfs");
2949 late_initcall(fault_around_debugfs
);
2953 * do_fault_around() tries to map few pages around the fault address. The hope
2954 * is that the pages will be needed soon and this will lower the number of
2957 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2958 * not ready to be mapped: not up-to-date, locked, etc.
2960 * This function is called with the page table lock taken. In the split ptlock
2961 * case the page table lock only protects only those entries which belong to
2962 * the page table corresponding to the fault address.
2964 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2967 * fault_around_pages() defines how many pages we'll try to map.
2968 * do_fault_around() expects it to return a power of two less than or equal to
2971 * The virtual address of the area that we map is naturally aligned to the
2972 * fault_around_pages() value (and therefore to page order). This way it's
2973 * easier to guarantee that we don't cross page table boundaries.
2975 static void do_fault_around(struct vm_area_struct
*vma
, unsigned long address
,
2976 pte_t
*pte
, pgoff_t pgoff
, unsigned int flags
)
2978 unsigned long start_addr
, nr_pages
, mask
;
2980 struct vm_fault vmf
;
2983 nr_pages
= READ_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
2984 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
2986 start_addr
= max(address
& mask
, vma
->vm_start
);
2987 off
= ((address
- start_addr
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
2992 * max_pgoff is either end of page table or end of vma
2993 * or fault_around_pages() from pgoff, depending what is nearest.
2995 max_pgoff
= pgoff
- ((start_addr
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
2997 max_pgoff
= min3(max_pgoff
, vma_pages(vma
) + vma
->vm_pgoff
- 1,
2998 pgoff
+ nr_pages
- 1);
3000 /* Check if it makes any sense to call ->map_pages */
3001 while (!pte_none(*pte
)) {
3002 if (++pgoff
> max_pgoff
)
3004 start_addr
+= PAGE_SIZE
;
3005 if (start_addr
>= vma
->vm_end
)
3010 vmf
.virtual_address
= (void __user
*) start_addr
;
3013 vmf
.max_pgoff
= max_pgoff
;
3015 vmf
.gfp_mask
= __get_fault_gfp_mask(vma
);
3016 vma
->vm_ops
->map_pages(vma
, &vmf
);
3019 static int do_read_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3020 unsigned long address
, pmd_t
*pmd
,
3021 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3023 struct page
*fault_page
;
3029 * Let's call ->map_pages() first and use ->fault() as fallback
3030 * if page by the offset is not ready to be mapped (cold cache or
3033 if (vma
->vm_ops
->map_pages
&& fault_around_bytes
>> PAGE_SHIFT
> 1) {
3034 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3035 if (!pte_same(*pte
, orig_pte
))
3037 do_fault_around(vma
, address
, pte
, pgoff
, flags
);
3038 /* Check if the fault is handled by faultaround */
3039 if (!pte_same(*pte
, orig_pte
)) {
3041 * Faultaround produce old pte, but the pte we've
3042 * handler fault for should be young.
3044 pte_t entry
= pte_mkyoung(*pte
);
3045 if (ptep_set_access_flags(vma
, address
, pte
, entry
, 0))
3046 update_mmu_cache(vma
, address
, pte
);
3049 pte_unmap_unlock(pte
, ptl
);
3052 ret
= __do_fault(vma
, address
, pgoff
, flags
, NULL
, &fault_page
, NULL
);
3053 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3056 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3057 if (unlikely(!pte_same(*pte
, orig_pte
))) {
3058 pte_unmap_unlock(pte
, ptl
);
3059 unlock_page(fault_page
);
3060 put_page(fault_page
);
3063 do_set_pte(vma
, address
, fault_page
, pte
, false, false, false);
3064 unlock_page(fault_page
);
3066 pte_unmap_unlock(pte
, ptl
);
3070 static int do_cow_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3071 unsigned long address
, pmd_t
*pmd
,
3072 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3074 struct page
*fault_page
, *new_page
;
3076 struct mem_cgroup
*memcg
;
3081 if (unlikely(anon_vma_prepare(vma
)))
3082 return VM_FAULT_OOM
;
3084 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
3086 return VM_FAULT_OOM
;
3088 if (mem_cgroup_try_charge(new_page
, mm
, GFP_KERNEL
, &memcg
, false)) {
3090 return VM_FAULT_OOM
;
3093 ret
= __do_fault(vma
, address
, pgoff
, flags
, new_page
, &fault_page
,
3095 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3098 if (!(ret
& VM_FAULT_DAX_LOCKED
))
3099 copy_user_highpage(new_page
, fault_page
, address
, vma
);
3100 __SetPageUptodate(new_page
);
3102 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3103 if (unlikely(!pte_same(*pte
, orig_pte
))) {
3104 pte_unmap_unlock(pte
, ptl
);
3105 if (!(ret
& VM_FAULT_DAX_LOCKED
)) {
3106 unlock_page(fault_page
);
3107 put_page(fault_page
);
3109 dax_unlock_mapping_entry(vma
->vm_file
->f_mapping
,
3114 do_set_pte(vma
, address
, new_page
, pte
, true, true, false);
3115 mem_cgroup_commit_charge(new_page
, memcg
, false, false);
3116 lru_cache_add_active_or_unevictable(new_page
, vma
);
3117 pte_unmap_unlock(pte
, ptl
);
3118 if (!(ret
& VM_FAULT_DAX_LOCKED
)) {
3119 unlock_page(fault_page
);
3120 put_page(fault_page
);
3122 dax_unlock_mapping_entry(vma
->vm_file
->f_mapping
, pgoff
);
3126 mem_cgroup_cancel_charge(new_page
, memcg
, false);
3131 static int do_shared_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3132 unsigned long address
, pmd_t
*pmd
,
3133 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3135 struct page
*fault_page
;
3136 struct address_space
*mapping
;
3142 ret
= __do_fault(vma
, address
, pgoff
, flags
, NULL
, &fault_page
, NULL
);
3143 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3147 * Check if the backing address space wants to know that the page is
3148 * about to become writable
3150 if (vma
->vm_ops
->page_mkwrite
) {
3151 unlock_page(fault_page
);
3152 tmp
= do_page_mkwrite(vma
, fault_page
, address
);
3153 if (unlikely(!tmp
||
3154 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3155 put_page(fault_page
);
3160 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3161 if (unlikely(!pte_same(*pte
, orig_pte
))) {
3162 pte_unmap_unlock(pte
, ptl
);
3163 unlock_page(fault_page
);
3164 put_page(fault_page
);
3167 do_set_pte(vma
, address
, fault_page
, pte
, true, false, false);
3168 pte_unmap_unlock(pte
, ptl
);
3170 if (set_page_dirty(fault_page
))
3173 * Take a local copy of the address_space - page.mapping may be zeroed
3174 * by truncate after unlock_page(). The address_space itself remains
3175 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3176 * release semantics to prevent the compiler from undoing this copying.
3178 mapping
= page_rmapping(fault_page
);
3179 unlock_page(fault_page
);
3180 if ((dirtied
|| vma
->vm_ops
->page_mkwrite
) && mapping
) {
3182 * Some device drivers do not set page.mapping but still
3185 balance_dirty_pages_ratelimited(mapping
);
3188 if (!vma
->vm_ops
->page_mkwrite
)
3189 file_update_time(vma
->vm_file
);
3195 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3196 * but allow concurrent faults).
3197 * The mmap_sem may have been released depending on flags and our
3198 * return value. See filemap_fault() and __lock_page_or_retry().
3200 static int do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3201 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3202 unsigned int flags
, pte_t orig_pte
)
3204 pgoff_t pgoff
= linear_page_index(vma
, address
);
3206 pte_unmap(page_table
);
3207 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3208 if (!vma
->vm_ops
->fault
)
3209 return VM_FAULT_SIGBUS
;
3210 if (!(flags
& FAULT_FLAG_WRITE
))
3211 return do_read_fault(mm
, vma
, address
, pmd
, pgoff
, flags
,
3213 if (!(vma
->vm_flags
& VM_SHARED
))
3214 return do_cow_fault(mm
, vma
, address
, pmd
, pgoff
, flags
,
3216 return do_shared_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3219 static int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3220 unsigned long addr
, int page_nid
,
3225 count_vm_numa_event(NUMA_HINT_FAULTS
);
3226 if (page_nid
== numa_node_id()) {
3227 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3228 *flags
|= TNF_FAULT_LOCAL
;
3231 return mpol_misplaced(page
, vma
, addr
);
3234 static int do_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3235 unsigned long addr
, pte_t pte
, pte_t
*ptep
, pmd_t
*pmd
)
3237 struct page
*page
= NULL
;
3242 bool migrated
= false;
3243 bool was_writable
= pte_write(pte
);
3246 /* A PROT_NONE fault should not end up here */
3247 BUG_ON(!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)));
3250 * The "pte" at this point cannot be used safely without
3251 * validation through pte_unmap_same(). It's of NUMA type but
3252 * the pfn may be screwed if the read is non atomic.
3254 * We can safely just do a "set_pte_at()", because the old
3255 * page table entry is not accessible, so there would be no
3256 * concurrent hardware modifications to the PTE.
3258 ptl
= pte_lockptr(mm
, pmd
);
3260 if (unlikely(!pte_same(*ptep
, pte
))) {
3261 pte_unmap_unlock(ptep
, ptl
);
3265 /* Make it present again */
3266 pte
= pte_modify(pte
, vma
->vm_page_prot
);
3267 pte
= pte_mkyoung(pte
);
3269 pte
= pte_mkwrite(pte
);
3270 set_pte_at(mm
, addr
, ptep
, pte
);
3271 update_mmu_cache(vma
, addr
, ptep
);
3273 page
= vm_normal_page(vma
, addr
, pte
);
3275 pte_unmap_unlock(ptep
, ptl
);
3279 /* TODO: handle PTE-mapped THP */
3280 if (PageCompound(page
)) {
3281 pte_unmap_unlock(ptep
, ptl
);
3286 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3287 * much anyway since they can be in shared cache state. This misses
3288 * the case where a mapping is writable but the process never writes
3289 * to it but pte_write gets cleared during protection updates and
3290 * pte_dirty has unpredictable behaviour between PTE scan updates,
3291 * background writeback, dirty balancing and application behaviour.
3293 if (!(vma
->vm_flags
& VM_WRITE
))
3294 flags
|= TNF_NO_GROUP
;
3297 * Flag if the page is shared between multiple address spaces. This
3298 * is later used when determining whether to group tasks together
3300 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
3301 flags
|= TNF_SHARED
;
3303 last_cpupid
= page_cpupid_last(page
);
3304 page_nid
= page_to_nid(page
);
3305 target_nid
= numa_migrate_prep(page
, vma
, addr
, page_nid
, &flags
);
3306 pte_unmap_unlock(ptep
, ptl
);
3307 if (target_nid
== -1) {
3312 /* Migrate to the requested node */
3313 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
3315 page_nid
= target_nid
;
3316 flags
|= TNF_MIGRATED
;
3318 flags
|= TNF_MIGRATE_FAIL
;
3322 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
3326 static int create_huge_pmd(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3327 unsigned long address
, pmd_t
*pmd
, unsigned int flags
)
3329 if (vma_is_anonymous(vma
))
3330 return do_huge_pmd_anonymous_page(mm
, vma
, address
, pmd
, flags
);
3331 if (vma
->vm_ops
->pmd_fault
)
3332 return vma
->vm_ops
->pmd_fault(vma
, address
, pmd
, flags
);
3333 return VM_FAULT_FALLBACK
;
3336 static int wp_huge_pmd(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3337 unsigned long address
, pmd_t
*pmd
, pmd_t orig_pmd
,
3340 if (vma_is_anonymous(vma
))
3341 return do_huge_pmd_wp_page(mm
, vma
, address
, pmd
, orig_pmd
);
3342 if (vma
->vm_ops
->pmd_fault
)
3343 return vma
->vm_ops
->pmd_fault(vma
, address
, pmd
, flags
);
3344 return VM_FAULT_FALLBACK
;
3348 * These routines also need to handle stuff like marking pages dirty
3349 * and/or accessed for architectures that don't do it in hardware (most
3350 * RISC architectures). The early dirtying is also good on the i386.
3352 * There is also a hook called "update_mmu_cache()" that architectures
3353 * with external mmu caches can use to update those (ie the Sparc or
3354 * PowerPC hashed page tables that act as extended TLBs).
3356 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3357 * but allow concurrent faults), and pte mapped but not yet locked.
3358 * We return with pte unmapped and unlocked.
3360 * The mmap_sem may have been released depending on flags and our
3361 * return value. See filemap_fault() and __lock_page_or_retry().
3363 static int handle_pte_fault(struct mm_struct
*mm
,
3364 struct vm_area_struct
*vma
, unsigned long address
,
3365 pte_t
*pte
, pmd_t
*pmd
, unsigned int flags
)
3371 * some architectures can have larger ptes than wordsize,
3372 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3373 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3374 * The code below just needs a consistent view for the ifs and
3375 * we later double check anyway with the ptl lock held. So here
3376 * a barrier will do.
3380 if (!pte_present(entry
)) {
3381 if (pte_none(entry
)) {
3382 if (vma_is_anonymous(vma
))
3383 return do_anonymous_page(mm
, vma
, address
,
3386 return do_fault(mm
, vma
, address
, pte
, pmd
,
3389 return do_swap_page(mm
, vma
, address
,
3390 pte
, pmd
, flags
, entry
);
3393 if (pte_protnone(entry
))
3394 return do_numa_page(mm
, vma
, address
, entry
, pte
, pmd
);
3396 ptl
= pte_lockptr(mm
, pmd
);
3398 if (unlikely(!pte_same(*pte
, entry
)))
3400 if (flags
& FAULT_FLAG_WRITE
) {
3401 if (!pte_write(entry
))
3402 return do_wp_page(mm
, vma
, address
,
3403 pte
, pmd
, ptl
, entry
);
3404 entry
= pte_mkdirty(entry
);
3406 entry
= pte_mkyoung(entry
);
3407 if (ptep_set_access_flags(vma
, address
, pte
, entry
, flags
& FAULT_FLAG_WRITE
)) {
3408 update_mmu_cache(vma
, address
, pte
);
3411 * This is needed only for protection faults but the arch code
3412 * is not yet telling us if this is a protection fault or not.
3413 * This still avoids useless tlb flushes for .text page faults
3416 if (flags
& FAULT_FLAG_WRITE
)
3417 flush_tlb_fix_spurious_fault(vma
, address
);
3420 pte_unmap_unlock(pte
, ptl
);
3425 * By the time we get here, we already hold the mm semaphore
3427 * The mmap_sem may have been released depending on flags and our
3428 * return value. See filemap_fault() and __lock_page_or_retry().
3430 static int __handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3431 unsigned long address
, unsigned int flags
)
3438 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
3439 flags
& FAULT_FLAG_INSTRUCTION
,
3440 flags
& FAULT_FLAG_REMOTE
))
3441 return VM_FAULT_SIGSEGV
;
3443 if (unlikely(is_vm_hugetlb_page(vma
)))
3444 return hugetlb_fault(mm
, vma
, address
, flags
);
3446 pgd
= pgd_offset(mm
, address
);
3447 pud
= pud_alloc(mm
, pgd
, address
);
3449 return VM_FAULT_OOM
;
3450 pmd
= pmd_alloc(mm
, pud
, address
);
3452 return VM_FAULT_OOM
;
3453 if (pmd_none(*pmd
) && transparent_hugepage_enabled(vma
)) {
3454 int ret
= create_huge_pmd(mm
, vma
, address
, pmd
, flags
);
3455 if (!(ret
& VM_FAULT_FALLBACK
))
3458 pmd_t orig_pmd
= *pmd
;
3462 if (pmd_trans_huge(orig_pmd
) || pmd_devmap(orig_pmd
)) {
3463 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
3465 if (pmd_protnone(orig_pmd
))
3466 return do_huge_pmd_numa_page(mm
, vma
, address
,
3469 if (dirty
&& !pmd_write(orig_pmd
)) {
3470 ret
= wp_huge_pmd(mm
, vma
, address
, pmd
,
3472 if (!(ret
& VM_FAULT_FALLBACK
))
3475 huge_pmd_set_accessed(mm
, vma
, address
, pmd
,
3483 * Use pte_alloc() instead of pte_alloc_map, because we can't
3484 * run pte_offset_map on the pmd, if an huge pmd could
3485 * materialize from under us from a different thread.
3487 if (unlikely(pte_alloc(mm
, pmd
, address
)))
3488 return VM_FAULT_OOM
;
3490 * If a huge pmd materialized under us just retry later. Use
3491 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
3492 * didn't become pmd_trans_huge under us and then back to pmd_none, as
3493 * a result of MADV_DONTNEED running immediately after a huge pmd fault
3494 * in a different thread of this mm, in turn leading to a misleading
3495 * pmd_trans_huge() retval. All we have to ensure is that it is a
3496 * regular pmd that we can walk with pte_offset_map() and we can do that
3497 * through an atomic read in C, which is what pmd_trans_unstable()
3500 if (unlikely(pmd_trans_unstable(pmd
) || pmd_devmap(*pmd
)))
3503 * A regular pmd is established and it can't morph into a huge pmd
3504 * from under us anymore at this point because we hold the mmap_sem
3505 * read mode and khugepaged takes it in write mode. So now it's
3506 * safe to run pte_offset_map().
3508 pte
= pte_offset_map(pmd
, address
);
3510 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
3514 * By the time we get here, we already hold the mm semaphore
3516 * The mmap_sem may have been released depending on flags and our
3517 * return value. See filemap_fault() and __lock_page_or_retry().
3519 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3520 unsigned long address
, unsigned int flags
)
3524 __set_current_state(TASK_RUNNING
);
3526 count_vm_event(PGFAULT
);
3527 mem_cgroup_count_vm_event(mm
, PGFAULT
);
3529 /* do counter updates before entering really critical section. */
3530 check_sync_rss_stat(current
);
3533 * Enable the memcg OOM handling for faults triggered in user
3534 * space. Kernel faults are handled more gracefully.
3536 if (flags
& FAULT_FLAG_USER
)
3537 mem_cgroup_oom_enable();
3539 ret
= __handle_mm_fault(mm
, vma
, address
, flags
);
3541 if (flags
& FAULT_FLAG_USER
) {
3542 mem_cgroup_oom_disable();
3544 * The task may have entered a memcg OOM situation but
3545 * if the allocation error was handled gracefully (no
3546 * VM_FAULT_OOM), there is no need to kill anything.
3547 * Just clean up the OOM state peacefully.
3549 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
3550 mem_cgroup_oom_synchronize(false);
3555 EXPORT_SYMBOL_GPL(handle_mm_fault
);
3557 #ifndef __PAGETABLE_PUD_FOLDED
3559 * Allocate page upper directory.
3560 * We've already handled the fast-path in-line.
3562 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
3564 pud_t
*new = pud_alloc_one(mm
, address
);
3568 smp_wmb(); /* See comment in __pte_alloc */
3570 spin_lock(&mm
->page_table_lock
);
3571 if (pgd_present(*pgd
)) /* Another has populated it */
3574 pgd_populate(mm
, pgd
, new);
3575 spin_unlock(&mm
->page_table_lock
);
3578 #endif /* __PAGETABLE_PUD_FOLDED */
3580 #ifndef __PAGETABLE_PMD_FOLDED
3582 * Allocate page middle directory.
3583 * We've already handled the fast-path in-line.
3585 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
3587 pmd_t
*new = pmd_alloc_one(mm
, address
);
3591 smp_wmb(); /* See comment in __pte_alloc */
3593 spin_lock(&mm
->page_table_lock
);
3594 #ifndef __ARCH_HAS_4LEVEL_HACK
3595 if (!pud_present(*pud
)) {
3597 pud_populate(mm
, pud
, new);
3598 } else /* Another has populated it */
3601 if (!pgd_present(*pud
)) {
3603 pgd_populate(mm
, pud
, new);
3604 } else /* Another has populated it */
3606 #endif /* __ARCH_HAS_4LEVEL_HACK */
3607 spin_unlock(&mm
->page_table_lock
);
3610 #endif /* __PAGETABLE_PMD_FOLDED */
3612 static int __follow_pte(struct mm_struct
*mm
, unsigned long address
,
3613 pte_t
**ptepp
, spinlock_t
**ptlp
)
3620 pgd
= pgd_offset(mm
, address
);
3621 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
3624 pud
= pud_offset(pgd
, address
);
3625 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
3628 pmd
= pmd_offset(pud
, address
);
3629 VM_BUG_ON(pmd_trans_huge(*pmd
));
3630 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
3633 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3637 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
3640 if (!pte_present(*ptep
))
3645 pte_unmap_unlock(ptep
, *ptlp
);
3650 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
3651 pte_t
**ptepp
, spinlock_t
**ptlp
)
3655 /* (void) is needed to make gcc happy */
3656 (void) __cond_lock(*ptlp
,
3657 !(res
= __follow_pte(mm
, address
, ptepp
, ptlp
)));
3662 * follow_pfn - look up PFN at a user virtual address
3663 * @vma: memory mapping
3664 * @address: user virtual address
3665 * @pfn: location to store found PFN
3667 * Only IO mappings and raw PFN mappings are allowed.
3669 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3671 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
3678 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3681 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
3684 *pfn
= pte_pfn(*ptep
);
3685 pte_unmap_unlock(ptep
, ptl
);
3688 EXPORT_SYMBOL(follow_pfn
);
3690 #ifdef CONFIG_HAVE_IOREMAP_PROT
3691 int follow_phys(struct vm_area_struct
*vma
,
3692 unsigned long address
, unsigned int flags
,
3693 unsigned long *prot
, resource_size_t
*phys
)
3699 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3702 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
3706 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
3709 *prot
= pgprot_val(pte_pgprot(pte
));
3710 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
3714 pte_unmap_unlock(ptep
, ptl
);
3719 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
3720 void *buf
, int len
, int write
)
3722 resource_size_t phys_addr
;
3723 unsigned long prot
= 0;
3724 void __iomem
*maddr
;
3725 int offset
= addr
& (PAGE_SIZE
-1);
3727 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
3730 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
3732 memcpy_toio(maddr
+ offset
, buf
, len
);
3734 memcpy_fromio(buf
, maddr
+ offset
, len
);
3739 EXPORT_SYMBOL_GPL(generic_access_phys
);
3743 * Access another process' address space as given in mm. If non-NULL, use the
3744 * given task for page fault accounting.
3746 static int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
3747 unsigned long addr
, void *buf
, int len
, int write
)
3749 struct vm_area_struct
*vma
;
3750 void *old_buf
= buf
;
3752 down_read(&mm
->mmap_sem
);
3753 /* ignore errors, just check how much was successfully transferred */
3755 int bytes
, ret
, offset
;
3757 struct page
*page
= NULL
;
3759 ret
= get_user_pages_remote(tsk
, mm
, addr
, 1,
3760 write
, 1, &page
, &vma
);
3762 #ifndef CONFIG_HAVE_IOREMAP_PROT
3766 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3767 * we can access using slightly different code.
3769 vma
= find_vma(mm
, addr
);
3770 if (!vma
|| vma
->vm_start
> addr
)
3772 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
3773 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
3781 offset
= addr
& (PAGE_SIZE
-1);
3782 if (bytes
> PAGE_SIZE
-offset
)
3783 bytes
= PAGE_SIZE
-offset
;
3787 copy_to_user_page(vma
, page
, addr
,
3788 maddr
+ offset
, buf
, bytes
);
3789 set_page_dirty_lock(page
);
3791 copy_from_user_page(vma
, page
, addr
,
3792 buf
, maddr
+ offset
, bytes
);
3801 up_read(&mm
->mmap_sem
);
3803 return buf
- old_buf
;
3807 * access_remote_vm - access another process' address space
3808 * @mm: the mm_struct of the target address space
3809 * @addr: start address to access
3810 * @buf: source or destination buffer
3811 * @len: number of bytes to transfer
3812 * @write: whether the access is a write
3814 * The caller must hold a reference on @mm.
3816 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
3817 void *buf
, int len
, int write
)
3819 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, write
);
3823 * Access another process' address space.
3824 * Source/target buffer must be kernel space,
3825 * Do not walk the page table directly, use get_user_pages
3827 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
3828 void *buf
, int len
, int write
)
3830 struct mm_struct
*mm
;
3833 mm
= get_task_mm(tsk
);
3837 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, write
);
3844 * Print the name of a VMA.
3846 void print_vma_addr(char *prefix
, unsigned long ip
)
3848 struct mm_struct
*mm
= current
->mm
;
3849 struct vm_area_struct
*vma
;
3852 * Do not print if we are in atomic
3853 * contexts (in exception stacks, etc.):
3855 if (preempt_count())
3858 down_read(&mm
->mmap_sem
);
3859 vma
= find_vma(mm
, ip
);
3860 if (vma
&& vma
->vm_file
) {
3861 struct file
*f
= vma
->vm_file
;
3862 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
3866 p
= file_path(f
, buf
, PAGE_SIZE
);
3869 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
3871 vma
->vm_end
- vma
->vm_start
);
3872 free_page((unsigned long)buf
);
3875 up_read(&mm
->mmap_sem
);
3878 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3879 void __might_fault(const char *file
, int line
)
3882 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3883 * holding the mmap_sem, this is safe because kernel memory doesn't
3884 * get paged out, therefore we'll never actually fault, and the
3885 * below annotations will generate false positives.
3887 if (segment_eq(get_fs(), KERNEL_DS
))
3889 if (pagefault_disabled())
3891 __might_sleep(file
, line
, 0);
3892 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3894 might_lock_read(¤t
->mm
->mmap_sem
);
3897 EXPORT_SYMBOL(__might_fault
);
3900 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3901 static void clear_gigantic_page(struct page
*page
,
3903 unsigned int pages_per_huge_page
)
3906 struct page
*p
= page
;
3909 for (i
= 0; i
< pages_per_huge_page
;
3910 i
++, p
= mem_map_next(p
, page
, i
)) {
3912 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
3915 void clear_huge_page(struct page
*page
,
3916 unsigned long addr
, unsigned int pages_per_huge_page
)
3920 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3921 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
3926 for (i
= 0; i
< pages_per_huge_page
; i
++) {
3928 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
3932 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
3934 struct vm_area_struct
*vma
,
3935 unsigned int pages_per_huge_page
)
3938 struct page
*dst_base
= dst
;
3939 struct page
*src_base
= src
;
3941 for (i
= 0; i
< pages_per_huge_page
; ) {
3943 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
3946 dst
= mem_map_next(dst
, dst_base
, i
);
3947 src
= mem_map_next(src
, src_base
, i
);
3951 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
3952 unsigned long addr
, struct vm_area_struct
*vma
,
3953 unsigned int pages_per_huge_page
)
3957 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3958 copy_user_gigantic_page(dst
, src
, addr
, vma
,
3959 pages_per_huge_page
);
3964 for (i
= 0; i
< pages_per_huge_page
; i
++) {
3966 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
3969 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3971 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3973 static struct kmem_cache
*page_ptl_cachep
;
3975 void __init
ptlock_cache_init(void)
3977 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
3981 bool ptlock_alloc(struct page
*page
)
3985 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
3992 void ptlock_free(struct page
*page
)
3994 kmem_cache_free(page_ptl_cachep
, page
->ptl
);