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>
68 #include <asm/pgalloc.h>
69 #include <asm/uaccess.h>
71 #include <asm/tlbflush.h>
72 #include <asm/pgtable.h>
76 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
77 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
80 #ifndef CONFIG_NEED_MULTIPLE_NODES
81 /* use the per-pgdat data instead for discontigmem - mbligh */
82 unsigned long max_mapnr
;
85 EXPORT_SYMBOL(max_mapnr
);
86 EXPORT_SYMBOL(mem_map
);
90 * A number of key systems in x86 including ioremap() rely on the assumption
91 * that high_memory defines the upper bound on direct map memory, then end
92 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
93 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
98 EXPORT_SYMBOL(high_memory
);
101 * Randomize the address space (stacks, mmaps, brk, etc.).
103 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
104 * as ancient (libc5 based) binaries can segfault. )
106 int randomize_va_space __read_mostly
=
107 #ifdef CONFIG_COMPAT_BRK
113 static int __init
disable_randmaps(char *s
)
115 randomize_va_space
= 0;
118 __setup("norandmaps", disable_randmaps
);
120 unsigned long zero_pfn __read_mostly
;
121 unsigned long highest_memmap_pfn __read_mostly
;
123 EXPORT_SYMBOL(zero_pfn
);
126 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
128 static int __init
init_zero_pfn(void)
130 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
133 core_initcall(init_zero_pfn
);
136 #if defined(SPLIT_RSS_COUNTING)
138 void sync_mm_rss(struct mm_struct
*mm
)
142 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
143 if (current
->rss_stat
.count
[i
]) {
144 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
145 current
->rss_stat
.count
[i
] = 0;
148 current
->rss_stat
.events
= 0;
151 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
153 struct task_struct
*task
= current
;
155 if (likely(task
->mm
== mm
))
156 task
->rss_stat
.count
[member
] += val
;
158 add_mm_counter(mm
, member
, val
);
160 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
161 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
163 /* sync counter once per 64 page faults */
164 #define TASK_RSS_EVENTS_THRESH (64)
165 static void check_sync_rss_stat(struct task_struct
*task
)
167 if (unlikely(task
!= current
))
169 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
170 sync_mm_rss(task
->mm
);
172 #else /* SPLIT_RSS_COUNTING */
174 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
175 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
177 static void check_sync_rss_stat(struct task_struct
*task
)
181 #endif /* SPLIT_RSS_COUNTING */
183 #ifdef HAVE_GENERIC_MMU_GATHER
185 static bool tlb_next_batch(struct mmu_gather
*tlb
)
187 struct mmu_gather_batch
*batch
;
191 tlb
->active
= batch
->next
;
195 if (tlb
->batch_count
== MAX_GATHER_BATCH_COUNT
)
198 batch
= (void *)__get_free_pages(GFP_NOWAIT
| __GFP_NOWARN
, 0);
205 batch
->max
= MAX_GATHER_BATCH
;
207 tlb
->active
->next
= batch
;
214 * Called to initialize an (on-stack) mmu_gather structure for page-table
215 * tear-down from @mm. The @fullmm argument is used when @mm is without
216 * users and we're going to destroy the full address space (exit/execve).
218 void tlb_gather_mmu(struct mmu_gather
*tlb
, struct mm_struct
*mm
, unsigned long start
, unsigned long end
)
222 /* Is it from 0 to ~0? */
223 tlb
->fullmm
= !(start
| (end
+1));
224 tlb
->need_flush_all
= 0;
225 tlb
->local
.next
= NULL
;
227 tlb
->local
.max
= ARRAY_SIZE(tlb
->__pages
);
228 tlb
->active
= &tlb
->local
;
229 tlb
->batch_count
= 0;
231 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
235 __tlb_reset_range(tlb
);
238 static void tlb_flush_mmu_tlbonly(struct mmu_gather
*tlb
)
244 mmu_notifier_invalidate_range(tlb
->mm
, tlb
->start
, tlb
->end
);
245 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
246 tlb_table_flush(tlb
);
248 __tlb_reset_range(tlb
);
251 static void tlb_flush_mmu_free(struct mmu_gather
*tlb
)
253 struct mmu_gather_batch
*batch
;
255 for (batch
= &tlb
->local
; batch
&& batch
->nr
; batch
= batch
->next
) {
256 free_pages_and_swap_cache(batch
->pages
, batch
->nr
);
259 tlb
->active
= &tlb
->local
;
262 void tlb_flush_mmu(struct mmu_gather
*tlb
)
264 tlb_flush_mmu_tlbonly(tlb
);
265 tlb_flush_mmu_free(tlb
);
269 * Called at the end of the shootdown operation to free up any resources
270 * that were required.
272 void tlb_finish_mmu(struct mmu_gather
*tlb
, unsigned long start
, unsigned long end
)
274 struct mmu_gather_batch
*batch
, *next
;
278 /* keep the page table cache within bounds */
281 for (batch
= tlb
->local
.next
; batch
; batch
= next
) {
283 free_pages((unsigned long)batch
, 0);
285 tlb
->local
.next
= NULL
;
289 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
290 * handling the additional races in SMP caused by other CPUs caching valid
291 * mappings in their TLBs. Returns the number of free page slots left.
292 * When out of page slots we must call tlb_flush_mmu().
294 int __tlb_remove_page(struct mmu_gather
*tlb
, struct page
*page
)
296 struct mmu_gather_batch
*batch
;
298 VM_BUG_ON(!tlb
->end
);
301 batch
->pages
[batch
->nr
++] = page
;
302 if (batch
->nr
== batch
->max
) {
303 if (!tlb_next_batch(tlb
))
307 VM_BUG_ON_PAGE(batch
->nr
> batch
->max
, page
);
309 return batch
->max
- batch
->nr
;
312 #endif /* HAVE_GENERIC_MMU_GATHER */
314 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
317 * See the comment near struct mmu_table_batch.
320 static void tlb_remove_table_smp_sync(void *arg
)
322 /* Simply deliver the interrupt */
325 static void tlb_remove_table_one(void *table
)
328 * This isn't an RCU grace period and hence the page-tables cannot be
329 * assumed to be actually RCU-freed.
331 * It is however sufficient for software page-table walkers that rely on
332 * IRQ disabling. See the comment near struct mmu_table_batch.
334 smp_call_function(tlb_remove_table_smp_sync
, NULL
, 1);
335 __tlb_remove_table(table
);
338 static void tlb_remove_table_rcu(struct rcu_head
*head
)
340 struct mmu_table_batch
*batch
;
343 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
345 for (i
= 0; i
< batch
->nr
; i
++)
346 __tlb_remove_table(batch
->tables
[i
]);
348 free_page((unsigned long)batch
);
351 void tlb_table_flush(struct mmu_gather
*tlb
)
353 struct mmu_table_batch
**batch
= &tlb
->batch
;
356 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
361 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
363 struct mmu_table_batch
**batch
= &tlb
->batch
;
366 * When there's less then two users of this mm there cannot be a
367 * concurrent page-table walk.
369 if (atomic_read(&tlb
->mm
->mm_users
) < 2) {
370 __tlb_remove_table(table
);
374 if (*batch
== NULL
) {
375 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
376 if (*batch
== NULL
) {
377 tlb_remove_table_one(table
);
382 (*batch
)->tables
[(*batch
)->nr
++] = table
;
383 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
384 tlb_table_flush(tlb
);
387 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
390 * Note: this doesn't free the actual pages themselves. That
391 * has been handled earlier when unmapping all the memory regions.
393 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
396 pgtable_t token
= pmd_pgtable(*pmd
);
398 pte_free_tlb(tlb
, token
, addr
);
399 atomic_long_dec(&tlb
->mm
->nr_ptes
);
402 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
403 unsigned long addr
, unsigned long end
,
404 unsigned long floor
, unsigned long ceiling
)
411 pmd
= pmd_offset(pud
, addr
);
413 next
= pmd_addr_end(addr
, end
);
414 if (pmd_none_or_clear_bad(pmd
))
416 free_pte_range(tlb
, pmd
, addr
);
417 } while (pmd
++, addr
= next
, addr
!= end
);
427 if (end
- 1 > ceiling
- 1)
430 pmd
= pmd_offset(pud
, start
);
432 pmd_free_tlb(tlb
, pmd
, start
);
433 mm_dec_nr_pmds(tlb
->mm
);
436 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
437 unsigned long addr
, unsigned long end
,
438 unsigned long floor
, unsigned long ceiling
)
445 pud
= pud_offset(pgd
, addr
);
447 next
= pud_addr_end(addr
, end
);
448 if (pud_none_or_clear_bad(pud
))
450 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
451 } while (pud
++, addr
= next
, addr
!= end
);
457 ceiling
&= PGDIR_MASK
;
461 if (end
- 1 > ceiling
- 1)
464 pud
= pud_offset(pgd
, start
);
466 pud_free_tlb(tlb
, pud
, start
);
470 * This function frees user-level page tables of a process.
472 void free_pgd_range(struct mmu_gather
*tlb
,
473 unsigned long addr
, unsigned long end
,
474 unsigned long floor
, unsigned long ceiling
)
480 * The next few lines have given us lots of grief...
482 * Why are we testing PMD* at this top level? Because often
483 * there will be no work to do at all, and we'd prefer not to
484 * go all the way down to the bottom just to discover that.
486 * Why all these "- 1"s? Because 0 represents both the bottom
487 * of the address space and the top of it (using -1 for the
488 * top wouldn't help much: the masks would do the wrong thing).
489 * The rule is that addr 0 and floor 0 refer to the bottom of
490 * the address space, but end 0 and ceiling 0 refer to the top
491 * Comparisons need to use "end - 1" and "ceiling - 1" (though
492 * that end 0 case should be mythical).
494 * Wherever addr is brought up or ceiling brought down, we must
495 * be careful to reject "the opposite 0" before it confuses the
496 * subsequent tests. But what about where end is brought down
497 * by PMD_SIZE below? no, end can't go down to 0 there.
499 * Whereas we round start (addr) and ceiling down, by different
500 * masks at different levels, in order to test whether a table
501 * now has no other vmas using it, so can be freed, we don't
502 * bother to round floor or end up - the tests don't need that.
516 if (end
- 1 > ceiling
- 1)
521 pgd
= pgd_offset(tlb
->mm
, addr
);
523 next
= pgd_addr_end(addr
, end
);
524 if (pgd_none_or_clear_bad(pgd
))
526 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
527 } while (pgd
++, addr
= next
, addr
!= end
);
530 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
531 unsigned long floor
, unsigned long ceiling
)
534 struct vm_area_struct
*next
= vma
->vm_next
;
535 unsigned long addr
= vma
->vm_start
;
538 * Hide vma from rmap and truncate_pagecache before freeing
541 unlink_anon_vmas(vma
);
542 unlink_file_vma(vma
);
544 if (is_vm_hugetlb_page(vma
)) {
545 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
546 floor
, next
? next
->vm_start
: ceiling
);
549 * Optimization: gather nearby vmas into one call down
551 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
552 && !is_vm_hugetlb_page(next
)) {
555 unlink_anon_vmas(vma
);
556 unlink_file_vma(vma
);
558 free_pgd_range(tlb
, addr
, vma
->vm_end
,
559 floor
, next
? next
->vm_start
: ceiling
);
565 int __pte_alloc(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
566 pmd_t
*pmd
, unsigned long address
)
569 pgtable_t
new = pte_alloc_one(mm
, address
);
574 * Ensure all pte setup (eg. pte page lock and page clearing) are
575 * visible before the pte is made visible to other CPUs by being
576 * put into page tables.
578 * The other side of the story is the pointer chasing in the page
579 * table walking code (when walking the page table without locking;
580 * ie. most of the time). Fortunately, these data accesses consist
581 * of a chain of data-dependent loads, meaning most CPUs (alpha
582 * being the notable exception) will already guarantee loads are
583 * seen in-order. See the alpha page table accessors for the
584 * smp_read_barrier_depends() barriers in page table walking code.
586 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
588 ptl
= pmd_lock(mm
, pmd
);
589 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
590 atomic_long_inc(&mm
->nr_ptes
);
591 pmd_populate(mm
, pmd
, new);
600 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
602 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
606 smp_wmb(); /* See comment in __pte_alloc */
608 spin_lock(&init_mm
.page_table_lock
);
609 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
610 pmd_populate_kernel(&init_mm
, pmd
, new);
613 spin_unlock(&init_mm
.page_table_lock
);
615 pte_free_kernel(&init_mm
, new);
619 static inline void init_rss_vec(int *rss
)
621 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
624 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
628 if (current
->mm
== mm
)
630 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
632 add_mm_counter(mm
, i
, rss
[i
]);
636 * This function is called to print an error when a bad pte
637 * is found. For example, we might have a PFN-mapped pte in
638 * a region that doesn't allow it.
640 * The calling function must still handle the error.
642 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
643 pte_t pte
, struct page
*page
)
645 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
646 pud_t
*pud
= pud_offset(pgd
, addr
);
647 pmd_t
*pmd
= pmd_offset(pud
, addr
);
648 struct address_space
*mapping
;
650 static unsigned long resume
;
651 static unsigned long nr_shown
;
652 static unsigned long nr_unshown
;
655 * Allow a burst of 60 reports, then keep quiet for that minute;
656 * or allow a steady drip of one report per second.
658 if (nr_shown
== 60) {
659 if (time_before(jiffies
, resume
)) {
665 "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
);
678 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
680 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
682 dump_page(page
, "bad pte");
684 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
685 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
687 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
689 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
691 vma
->vm_ops
? vma
->vm_ops
->fault
: NULL
,
692 vma
->vm_file
? vma
->vm_file
->f_op
->mmap
: NULL
,
693 mapping
? mapping
->a_ops
->readpage
: NULL
);
695 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
699 * vm_normal_page -- This function gets the "struct page" associated with a pte.
701 * "Special" mappings do not wish to be associated with a "struct page" (either
702 * it doesn't exist, or it exists but they don't want to touch it). In this
703 * case, NULL is returned here. "Normal" mappings do have a struct page.
705 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
706 * pte bit, in which case this function is trivial. Secondly, an architecture
707 * may not have a spare pte bit, which requires a more complicated scheme,
710 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
711 * special mapping (even if there are underlying and valid "struct pages").
712 * COWed pages of a VM_PFNMAP are always normal.
714 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
715 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
716 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
717 * mapping will always honor the rule
719 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
721 * And for normal mappings this is false.
723 * This restricts such mappings to be a linear translation from virtual address
724 * to pfn. To get around this restriction, we allow arbitrary mappings so long
725 * as the vma is not a COW mapping; in that case, we know that all ptes are
726 * special (because none can have been COWed).
729 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
731 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
732 * page" backing, however the difference is that _all_ pages with a struct
733 * page (that is, those where pfn_valid is true) are refcounted and considered
734 * normal pages by the VM. The disadvantage is that pages are refcounted
735 * (which can be slower and simply not an option for some PFNMAP users). The
736 * advantage is that we don't have to follow the strict linearity rule of
737 * PFNMAP mappings in order to support COWable mappings.
740 #ifdef __HAVE_ARCH_PTE_SPECIAL
741 # define HAVE_PTE_SPECIAL 1
743 # define HAVE_PTE_SPECIAL 0
745 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
748 unsigned long pfn
= pte_pfn(pte
);
750 if (HAVE_PTE_SPECIAL
) {
751 if (likely(!pte_special(pte
)))
753 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
754 return vma
->vm_ops
->find_special_page(vma
, addr
);
755 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
757 if (!is_zero_pfn(pfn
))
758 print_bad_pte(vma
, addr
, pte
, NULL
);
762 /* !HAVE_PTE_SPECIAL case follows: */
764 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
765 if (vma
->vm_flags
& VM_MIXEDMAP
) {
771 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
772 if (pfn
== vma
->vm_pgoff
+ off
)
774 if (!is_cow_mapping(vma
->vm_flags
))
779 if (is_zero_pfn(pfn
))
782 if (unlikely(pfn
> highest_memmap_pfn
)) {
783 print_bad_pte(vma
, addr
, pte
, NULL
);
788 * NOTE! We still have PageReserved() pages in the page tables.
789 * eg. VDSO mappings can cause them to exist.
792 return pfn_to_page(pfn
);
796 * copy one vm_area from one task to the other. Assumes the page tables
797 * already present in the new task to be cleared in the whole range
798 * covered by this vma.
801 static inline unsigned long
802 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
803 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
804 unsigned long addr
, int *rss
)
806 unsigned long vm_flags
= vma
->vm_flags
;
807 pte_t pte
= *src_pte
;
810 /* pte contains position in swap or file, so copy. */
811 if (unlikely(!pte_present(pte
))) {
812 swp_entry_t entry
= pte_to_swp_entry(pte
);
814 if (likely(!non_swap_entry(entry
))) {
815 if (swap_duplicate(entry
) < 0)
818 /* make sure dst_mm is on swapoff's mmlist. */
819 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
820 spin_lock(&mmlist_lock
);
821 if (list_empty(&dst_mm
->mmlist
))
822 list_add(&dst_mm
->mmlist
,
824 spin_unlock(&mmlist_lock
);
827 } else if (is_migration_entry(entry
)) {
828 page
= migration_entry_to_page(entry
);
830 rss
[mm_counter(page
)]++;
832 if (is_write_migration_entry(entry
) &&
833 is_cow_mapping(vm_flags
)) {
835 * COW mappings require pages in both
836 * parent and child to be set to read.
838 make_migration_entry_read(&entry
);
839 pte
= swp_entry_to_pte(entry
);
840 if (pte_swp_soft_dirty(*src_pte
))
841 pte
= pte_swp_mksoft_dirty(pte
);
842 set_pte_at(src_mm
, addr
, src_pte
, pte
);
849 * If it's a COW mapping, write protect it both
850 * in the parent and the child
852 if (is_cow_mapping(vm_flags
)) {
853 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
854 pte
= pte_wrprotect(pte
);
858 * If it's a shared mapping, mark it clean in
861 if (vm_flags
& VM_SHARED
)
862 pte
= pte_mkclean(pte
);
863 pte
= pte_mkold(pte
);
865 page
= vm_normal_page(vma
, addr
, pte
);
868 page_dup_rmap(page
, false);
869 rss
[mm_counter(page
)]++;
873 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
877 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
878 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
879 unsigned long addr
, unsigned long end
)
881 pte_t
*orig_src_pte
, *orig_dst_pte
;
882 pte_t
*src_pte
, *dst_pte
;
883 spinlock_t
*src_ptl
, *dst_ptl
;
885 int rss
[NR_MM_COUNTERS
];
886 swp_entry_t entry
= (swp_entry_t
){0};
891 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
894 src_pte
= pte_offset_map(src_pmd
, addr
);
895 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
896 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
897 orig_src_pte
= src_pte
;
898 orig_dst_pte
= dst_pte
;
899 arch_enter_lazy_mmu_mode();
903 * We are holding two locks at this point - either of them
904 * could generate latencies in another task on another CPU.
906 if (progress
>= 32) {
908 if (need_resched() ||
909 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
912 if (pte_none(*src_pte
)) {
916 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
921 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
923 arch_leave_lazy_mmu_mode();
924 spin_unlock(src_ptl
);
925 pte_unmap(orig_src_pte
);
926 add_mm_rss_vec(dst_mm
, rss
);
927 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
931 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
940 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
941 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
942 unsigned long addr
, unsigned long end
)
944 pmd_t
*src_pmd
, *dst_pmd
;
947 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
950 src_pmd
= pmd_offset(src_pud
, addr
);
952 next
= pmd_addr_end(addr
, end
);
953 if (pmd_trans_huge(*src_pmd
) || pmd_devmap(*src_pmd
)) {
955 VM_BUG_ON(next
-addr
!= HPAGE_PMD_SIZE
);
956 err
= copy_huge_pmd(dst_mm
, src_mm
,
957 dst_pmd
, src_pmd
, addr
, vma
);
964 if (pmd_none_or_clear_bad(src_pmd
))
966 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
969 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
973 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
974 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
975 unsigned long addr
, unsigned long end
)
977 pud_t
*src_pud
, *dst_pud
;
980 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
983 src_pud
= pud_offset(src_pgd
, addr
);
985 next
= pud_addr_end(addr
, end
);
986 if (pud_none_or_clear_bad(src_pud
))
988 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
991 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
995 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
996 struct vm_area_struct
*vma
)
998 pgd_t
*src_pgd
, *dst_pgd
;
1000 unsigned long addr
= vma
->vm_start
;
1001 unsigned long end
= vma
->vm_end
;
1002 unsigned long mmun_start
; /* For mmu_notifiers */
1003 unsigned long mmun_end
; /* For mmu_notifiers */
1008 * Don't copy ptes where a page fault will fill them correctly.
1009 * Fork becomes much lighter when there are big shared or private
1010 * readonly mappings. The tradeoff is that copy_page_range is more
1011 * efficient than faulting.
1013 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_PFNMAP
| VM_MIXEDMAP
)) &&
1017 if (is_vm_hugetlb_page(vma
))
1018 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1020 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
1022 * We do not free on error cases below as remove_vma
1023 * gets called on error from higher level routine
1025 ret
= track_pfn_copy(vma
);
1031 * We need to invalidate the secondary MMU mappings only when
1032 * there could be a permission downgrade on the ptes of the
1033 * parent mm. And a permission downgrade will only happen if
1034 * is_cow_mapping() returns true.
1036 is_cow
= is_cow_mapping(vma
->vm_flags
);
1040 mmu_notifier_invalidate_range_start(src_mm
, mmun_start
,
1044 dst_pgd
= pgd_offset(dst_mm
, addr
);
1045 src_pgd
= pgd_offset(src_mm
, addr
);
1047 next
= pgd_addr_end(addr
, end
);
1048 if (pgd_none_or_clear_bad(src_pgd
))
1050 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1051 vma
, addr
, next
))) {
1055 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1058 mmu_notifier_invalidate_range_end(src_mm
, mmun_start
, mmun_end
);
1062 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1063 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1064 unsigned long addr
, unsigned long end
,
1065 struct zap_details
*details
)
1067 struct mm_struct
*mm
= tlb
->mm
;
1068 int force_flush
= 0;
1069 int rss
[NR_MM_COUNTERS
];
1077 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1079 arch_enter_lazy_mmu_mode();
1082 if (pte_none(ptent
)) {
1086 if (pte_present(ptent
)) {
1089 page
= vm_normal_page(vma
, addr
, ptent
);
1090 if (unlikely(details
) && page
) {
1092 * unmap_shared_mapping_pages() wants to
1093 * invalidate cache without truncating:
1094 * unmap shared but keep private pages.
1096 if (details
->check_mapping
&&
1097 details
->check_mapping
!= page
->mapping
)
1100 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1102 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1103 if (unlikely(!page
))
1106 if (!PageAnon(page
)) {
1107 if (pte_dirty(ptent
)) {
1109 set_page_dirty(page
);
1111 if (pte_young(ptent
) &&
1112 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1113 mark_page_accessed(page
);
1115 rss
[mm_counter(page
)]--;
1116 page_remove_rmap(page
, false);
1117 if (unlikely(page_mapcount(page
) < 0))
1118 print_bad_pte(vma
, addr
, ptent
, page
);
1119 if (unlikely(!__tlb_remove_page(tlb
, page
))) {
1126 /* If details->check_mapping, we leave swap entries. */
1127 if (unlikely(details
))
1130 entry
= pte_to_swp_entry(ptent
);
1131 if (!non_swap_entry(entry
))
1133 else if (is_migration_entry(entry
)) {
1136 page
= migration_entry_to_page(entry
);
1137 rss
[mm_counter(page
)]--;
1139 if (unlikely(!free_swap_and_cache(entry
)))
1140 print_bad_pte(vma
, addr
, ptent
, NULL
);
1141 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1142 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1144 add_mm_rss_vec(mm
, rss
);
1145 arch_leave_lazy_mmu_mode();
1147 /* Do the actual TLB flush before dropping ptl */
1149 tlb_flush_mmu_tlbonly(tlb
);
1150 pte_unmap_unlock(start_pte
, ptl
);
1153 * If we forced a TLB flush (either due to running out of
1154 * batch buffers or because we needed to flush dirty TLB
1155 * entries before releasing the ptl), free the batched
1156 * memory too. Restart if we didn't do everything.
1160 tlb_flush_mmu_free(tlb
);
1169 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1170 struct vm_area_struct
*vma
, pud_t
*pud
,
1171 unsigned long addr
, unsigned long end
,
1172 struct zap_details
*details
)
1177 pmd
= pmd_offset(pud
, addr
);
1179 next
= pmd_addr_end(addr
, end
);
1180 if (pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1181 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1182 #ifdef CONFIG_DEBUG_VM
1183 if (!rwsem_is_locked(&tlb
->mm
->mmap_sem
)) {
1184 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1185 __func__
, addr
, end
,
1191 split_huge_pmd(vma
, pmd
, addr
);
1192 } else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1197 * Here there can be other concurrent MADV_DONTNEED or
1198 * trans huge page faults running, and if the pmd is
1199 * none or trans huge it can change under us. This is
1200 * because MADV_DONTNEED holds the mmap_sem in read
1203 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1205 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1208 } while (pmd
++, addr
= next
, addr
!= end
);
1213 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1214 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1215 unsigned long addr
, unsigned long end
,
1216 struct zap_details
*details
)
1221 pud
= pud_offset(pgd
, addr
);
1223 next
= pud_addr_end(addr
, end
);
1224 if (pud_none_or_clear_bad(pud
))
1226 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1227 } while (pud
++, addr
= next
, addr
!= end
);
1232 static void unmap_page_range(struct mmu_gather
*tlb
,
1233 struct vm_area_struct
*vma
,
1234 unsigned long addr
, unsigned long end
,
1235 struct zap_details
*details
)
1240 if (details
&& !details
->check_mapping
)
1243 BUG_ON(addr
>= end
);
1244 tlb_start_vma(tlb
, vma
);
1245 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1247 next
= pgd_addr_end(addr
, end
);
1248 if (pgd_none_or_clear_bad(pgd
))
1250 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
1251 } while (pgd
++, addr
= next
, addr
!= end
);
1252 tlb_end_vma(tlb
, vma
);
1256 static void unmap_single_vma(struct mmu_gather
*tlb
,
1257 struct vm_area_struct
*vma
, unsigned long start_addr
,
1258 unsigned long end_addr
,
1259 struct zap_details
*details
)
1261 unsigned long start
= max(vma
->vm_start
, start_addr
);
1264 if (start
>= vma
->vm_end
)
1266 end
= min(vma
->vm_end
, end_addr
);
1267 if (end
<= vma
->vm_start
)
1271 uprobe_munmap(vma
, start
, end
);
1273 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1274 untrack_pfn(vma
, 0, 0);
1277 if (unlikely(is_vm_hugetlb_page(vma
))) {
1279 * It is undesirable to test vma->vm_file as it
1280 * should be non-null for valid hugetlb area.
1281 * However, vm_file will be NULL in the error
1282 * cleanup path of mmap_region. When
1283 * hugetlbfs ->mmap method fails,
1284 * mmap_region() nullifies vma->vm_file
1285 * before calling this function to clean up.
1286 * Since no pte has actually been setup, it is
1287 * safe to do nothing in this case.
1290 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
1291 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1292 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
1295 unmap_page_range(tlb
, vma
, start
, end
, details
);
1300 * unmap_vmas - unmap a range of memory covered by a list of vma's
1301 * @tlb: address of the caller's struct mmu_gather
1302 * @vma: the starting vma
1303 * @start_addr: virtual address at which to start unmapping
1304 * @end_addr: virtual address at which to end unmapping
1306 * Unmap all pages in the vma list.
1308 * Only addresses between `start' and `end' will be unmapped.
1310 * The VMA list must be sorted in ascending virtual address order.
1312 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1313 * range after unmap_vmas() returns. So the only responsibility here is to
1314 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1315 * drops the lock and schedules.
1317 void unmap_vmas(struct mmu_gather
*tlb
,
1318 struct vm_area_struct
*vma
, unsigned long start_addr
,
1319 unsigned long end_addr
)
1321 struct mm_struct
*mm
= vma
->vm_mm
;
1323 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1324 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1325 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1326 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1330 * zap_page_range - remove user pages in a given range
1331 * @vma: vm_area_struct holding the applicable pages
1332 * @start: starting address of pages to zap
1333 * @size: number of bytes to zap
1334 * @details: details of shared cache invalidation
1336 * Caller must protect the VMA list
1338 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1339 unsigned long size
, struct zap_details
*details
)
1341 struct mm_struct
*mm
= vma
->vm_mm
;
1342 struct mmu_gather tlb
;
1343 unsigned long end
= start
+ size
;
1346 tlb_gather_mmu(&tlb
, mm
, start
, end
);
1347 update_hiwater_rss(mm
);
1348 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1349 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
)
1350 unmap_single_vma(&tlb
, vma
, start
, end
, details
);
1351 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1352 tlb_finish_mmu(&tlb
, start
, end
);
1356 * zap_page_range_single - remove user pages in a given range
1357 * @vma: vm_area_struct holding the applicable pages
1358 * @address: starting address of pages to zap
1359 * @size: number of bytes to zap
1360 * @details: details of shared cache invalidation
1362 * The range must fit into one VMA.
1364 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1365 unsigned long size
, struct zap_details
*details
)
1367 struct mm_struct
*mm
= vma
->vm_mm
;
1368 struct mmu_gather tlb
;
1369 unsigned long end
= address
+ size
;
1372 tlb_gather_mmu(&tlb
, mm
, address
, end
);
1373 update_hiwater_rss(mm
);
1374 mmu_notifier_invalidate_range_start(mm
, address
, end
);
1375 unmap_single_vma(&tlb
, vma
, address
, end
, details
);
1376 mmu_notifier_invalidate_range_end(mm
, address
, end
);
1377 tlb_finish_mmu(&tlb
, address
, end
);
1381 * zap_vma_ptes - remove ptes mapping the vma
1382 * @vma: vm_area_struct holding ptes to be zapped
1383 * @address: starting address of pages to zap
1384 * @size: number of bytes to zap
1386 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1388 * The entire address range must be fully contained within the vma.
1390 * Returns 0 if successful.
1392 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1395 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1396 !(vma
->vm_flags
& VM_PFNMAP
))
1398 zap_page_range_single(vma
, address
, size
, NULL
);
1401 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1403 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1406 pgd_t
* pgd
= pgd_offset(mm
, addr
);
1407 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
1409 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
1411 VM_BUG_ON(pmd_trans_huge(*pmd
));
1412 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1419 * This is the old fallback for page remapping.
1421 * For historical reasons, it only allows reserved pages. Only
1422 * old drivers should use this, and they needed to mark their
1423 * pages reserved for the old functions anyway.
1425 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1426 struct page
*page
, pgprot_t prot
)
1428 struct mm_struct
*mm
= vma
->vm_mm
;
1437 flush_dcache_page(page
);
1438 pte
= get_locked_pte(mm
, addr
, &ptl
);
1442 if (!pte_none(*pte
))
1445 /* Ok, finally just insert the thing.. */
1447 inc_mm_counter_fast(mm
, mm_counter_file(page
));
1448 page_add_file_rmap(page
);
1449 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1452 pte_unmap_unlock(pte
, ptl
);
1455 pte_unmap_unlock(pte
, ptl
);
1461 * vm_insert_page - insert single page into user vma
1462 * @vma: user vma to map to
1463 * @addr: target user address of this page
1464 * @page: source kernel page
1466 * This allows drivers to insert individual pages they've allocated
1469 * The page has to be a nice clean _individual_ kernel allocation.
1470 * If you allocate a compound page, you need to have marked it as
1471 * such (__GFP_COMP), or manually just split the page up yourself
1472 * (see split_page()).
1474 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1475 * took an arbitrary page protection parameter. This doesn't allow
1476 * that. Your vma protection will have to be set up correctly, which
1477 * means that if you want a shared writable mapping, you'd better
1478 * ask for a shared writable mapping!
1480 * The page does not need to be reserved.
1482 * Usually this function is called from f_op->mmap() handler
1483 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1484 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1485 * function from other places, for example from page-fault handler.
1487 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1490 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1492 if (!page_count(page
))
1494 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1495 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
1496 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1497 vma
->vm_flags
|= VM_MIXEDMAP
;
1499 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1501 EXPORT_SYMBOL(vm_insert_page
);
1503 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1504 pfn_t pfn
, pgprot_t prot
)
1506 struct mm_struct
*mm
= vma
->vm_mm
;
1512 pte
= get_locked_pte(mm
, addr
, &ptl
);
1516 if (!pte_none(*pte
))
1519 /* Ok, finally just insert the thing.. */
1520 if (pfn_t_devmap(pfn
))
1521 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
1523 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
1524 set_pte_at(mm
, addr
, pte
, entry
);
1525 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1529 pte_unmap_unlock(pte
, ptl
);
1535 * vm_insert_pfn - insert single pfn into user vma
1536 * @vma: user vma to map to
1537 * @addr: target user address of this page
1538 * @pfn: source kernel pfn
1540 * Similar to vm_insert_page, this allows drivers to insert individual pages
1541 * they've allocated into a user vma. Same comments apply.
1543 * This function should only be called from a vm_ops->fault handler, and
1544 * in that case the handler should return NULL.
1546 * vma cannot be a COW mapping.
1548 * As this is called only for pages that do not currently exist, we
1549 * do not need to flush old virtual caches or the TLB.
1551 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1554 return vm_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
1556 EXPORT_SYMBOL(vm_insert_pfn
);
1559 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1560 * @vma: user vma to map to
1561 * @addr: target user address of this page
1562 * @pfn: source kernel pfn
1563 * @pgprot: pgprot flags for the inserted page
1565 * This is exactly like vm_insert_pfn, except that it allows drivers to
1566 * to override pgprot on a per-page basis.
1568 * This only makes sense for IO mappings, and it makes no sense for
1569 * cow mappings. In general, using multiple vmas is preferable;
1570 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1573 int vm_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
1574 unsigned long pfn
, pgprot_t pgprot
)
1578 * Technically, architectures with pte_special can avoid all these
1579 * restrictions (same for remap_pfn_range). However we would like
1580 * consistency in testing and feature parity among all, so we should
1581 * try to keep these invariants in place for everybody.
1583 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1584 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1585 (VM_PFNMAP
|VM_MIXEDMAP
));
1586 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1587 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1589 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1591 if (track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
)))
1594 ret
= insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
);
1598 EXPORT_SYMBOL(vm_insert_pfn_prot
);
1600 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1603 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
1605 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1609 * If we don't have pte special, then we have to use the pfn_valid()
1610 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1611 * refcount the page if pfn_valid is true (hence insert_page rather
1612 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1613 * without pte special, it would there be refcounted as a normal page.
1615 if (!HAVE_PTE_SPECIAL
&& !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
1619 * At this point we are committed to insert_page()
1620 * regardless of whether the caller specified flags that
1621 * result in pfn_t_has_page() == false.
1623 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
1624 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1626 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1628 EXPORT_SYMBOL(vm_insert_mixed
);
1631 * maps a range of physical memory into the requested pages. the old
1632 * mappings are removed. any references to nonexistent pages results
1633 * in null mappings (currently treated as "copy-on-access")
1635 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1636 unsigned long addr
, unsigned long end
,
1637 unsigned long pfn
, pgprot_t prot
)
1642 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1645 arch_enter_lazy_mmu_mode();
1647 BUG_ON(!pte_none(*pte
));
1648 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
1650 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1651 arch_leave_lazy_mmu_mode();
1652 pte_unmap_unlock(pte
- 1, ptl
);
1656 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1657 unsigned long addr
, unsigned long end
,
1658 unsigned long pfn
, pgprot_t prot
)
1663 pfn
-= addr
>> PAGE_SHIFT
;
1664 pmd
= pmd_alloc(mm
, pud
, addr
);
1667 VM_BUG_ON(pmd_trans_huge(*pmd
));
1669 next
= pmd_addr_end(addr
, end
);
1670 if (remap_pte_range(mm
, pmd
, addr
, next
,
1671 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1673 } while (pmd
++, addr
= next
, addr
!= end
);
1677 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1678 unsigned long addr
, unsigned long end
,
1679 unsigned long pfn
, pgprot_t prot
)
1684 pfn
-= addr
>> PAGE_SHIFT
;
1685 pud
= pud_alloc(mm
, pgd
, addr
);
1689 next
= pud_addr_end(addr
, end
);
1690 if (remap_pmd_range(mm
, pud
, addr
, next
,
1691 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1693 } while (pud
++, addr
= next
, addr
!= end
);
1698 * remap_pfn_range - remap kernel memory to userspace
1699 * @vma: user vma to map to
1700 * @addr: target user address to start at
1701 * @pfn: physical address of kernel memory
1702 * @size: size of map area
1703 * @prot: page protection flags for this mapping
1705 * Note: this is only safe if the mm semaphore is held when called.
1707 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1708 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1712 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1713 struct mm_struct
*mm
= vma
->vm_mm
;
1717 * Physically remapped pages are special. Tell the
1718 * rest of the world about it:
1719 * VM_IO tells people not to look at these pages
1720 * (accesses can have side effects).
1721 * VM_PFNMAP tells the core MM that the base pages are just
1722 * raw PFN mappings, and do not have a "struct page" associated
1725 * Disable vma merging and expanding with mremap().
1727 * Omit vma from core dump, even when VM_IO turned off.
1729 * There's a horrible special case to handle copy-on-write
1730 * behaviour that some programs depend on. We mark the "original"
1731 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1732 * See vm_normal_page() for details.
1734 if (is_cow_mapping(vma
->vm_flags
)) {
1735 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
1737 vma
->vm_pgoff
= pfn
;
1740 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
1744 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
1746 BUG_ON(addr
>= end
);
1747 pfn
-= addr
>> PAGE_SHIFT
;
1748 pgd
= pgd_offset(mm
, addr
);
1749 flush_cache_range(vma
, addr
, end
);
1751 next
= pgd_addr_end(addr
, end
);
1752 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1753 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1756 } while (pgd
++, addr
= next
, addr
!= end
);
1759 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
));
1763 EXPORT_SYMBOL(remap_pfn_range
);
1766 * vm_iomap_memory - remap memory to userspace
1767 * @vma: user vma to map to
1768 * @start: start of area
1769 * @len: size of area
1771 * This is a simplified io_remap_pfn_range() for common driver use. The
1772 * driver just needs to give us the physical memory range to be mapped,
1773 * we'll figure out the rest from the vma information.
1775 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1776 * whatever write-combining details or similar.
1778 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
1780 unsigned long vm_len
, pfn
, pages
;
1782 /* Check that the physical memory area passed in looks valid */
1783 if (start
+ len
< start
)
1786 * You *really* shouldn't map things that aren't page-aligned,
1787 * but we've historically allowed it because IO memory might
1788 * just have smaller alignment.
1790 len
+= start
& ~PAGE_MASK
;
1791 pfn
= start
>> PAGE_SHIFT
;
1792 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
1793 if (pfn
+ pages
< pfn
)
1796 /* We start the mapping 'vm_pgoff' pages into the area */
1797 if (vma
->vm_pgoff
> pages
)
1799 pfn
+= vma
->vm_pgoff
;
1800 pages
-= vma
->vm_pgoff
;
1802 /* Can we fit all of the mapping? */
1803 vm_len
= vma
->vm_end
- vma
->vm_start
;
1804 if (vm_len
>> PAGE_SHIFT
> pages
)
1807 /* Ok, let it rip */
1808 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
1810 EXPORT_SYMBOL(vm_iomap_memory
);
1812 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1813 unsigned long addr
, unsigned long end
,
1814 pte_fn_t fn
, void *data
)
1819 spinlock_t
*uninitialized_var(ptl
);
1821 pte
= (mm
== &init_mm
) ?
1822 pte_alloc_kernel(pmd
, addr
) :
1823 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1827 BUG_ON(pmd_huge(*pmd
));
1829 arch_enter_lazy_mmu_mode();
1831 token
= pmd_pgtable(*pmd
);
1834 err
= fn(pte
++, token
, addr
, data
);
1837 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1839 arch_leave_lazy_mmu_mode();
1842 pte_unmap_unlock(pte
-1, ptl
);
1846 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1847 unsigned long addr
, unsigned long end
,
1848 pte_fn_t fn
, void *data
)
1854 BUG_ON(pud_huge(*pud
));
1856 pmd
= pmd_alloc(mm
, pud
, addr
);
1860 next
= pmd_addr_end(addr
, end
);
1861 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
1864 } while (pmd
++, addr
= next
, addr
!= end
);
1868 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1869 unsigned long addr
, unsigned long end
,
1870 pte_fn_t fn
, void *data
)
1876 pud
= pud_alloc(mm
, pgd
, addr
);
1880 next
= pud_addr_end(addr
, end
);
1881 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
1884 } while (pud
++, addr
= next
, addr
!= end
);
1889 * Scan a region of virtual memory, filling in page tables as necessary
1890 * and calling a provided function on each leaf page table.
1892 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
1893 unsigned long size
, pte_fn_t fn
, void *data
)
1897 unsigned long end
= addr
+ size
;
1900 if (WARN_ON(addr
>= end
))
1903 pgd
= pgd_offset(mm
, addr
);
1905 next
= pgd_addr_end(addr
, end
);
1906 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
1909 } while (pgd
++, addr
= next
, addr
!= end
);
1913 EXPORT_SYMBOL_GPL(apply_to_page_range
);
1916 * handle_pte_fault chooses page fault handler according to an entry which was
1917 * read non-atomically. Before making any commitment, on those architectures
1918 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1919 * parts, do_swap_page must check under lock before unmapping the pte and
1920 * proceeding (but do_wp_page is only called after already making such a check;
1921 * and do_anonymous_page can safely check later on).
1923 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
1924 pte_t
*page_table
, pte_t orig_pte
)
1927 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1928 if (sizeof(pte_t
) > sizeof(unsigned long)) {
1929 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
1931 same
= pte_same(*page_table
, orig_pte
);
1935 pte_unmap(page_table
);
1939 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
1941 debug_dma_assert_idle(src
);
1944 * If the source page was a PFN mapping, we don't have
1945 * a "struct page" for it. We do a best-effort copy by
1946 * just copying from the original user address. If that
1947 * fails, we just zero-fill it. Live with it.
1949 if (unlikely(!src
)) {
1950 void *kaddr
= kmap_atomic(dst
);
1951 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
1954 * This really shouldn't fail, because the page is there
1955 * in the page tables. But it might just be unreadable,
1956 * in which case we just give up and fill the result with
1959 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
1961 kunmap_atomic(kaddr
);
1962 flush_dcache_page(dst
);
1964 copy_user_highpage(dst
, src
, va
, vma
);
1967 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
1969 struct file
*vm_file
= vma
->vm_file
;
1972 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
1975 * Special mappings (e.g. VDSO) do not have any file so fake
1976 * a default GFP_KERNEL for them.
1982 * Notify the address space that the page is about to become writable so that
1983 * it can prohibit this or wait for the page to get into an appropriate state.
1985 * We do this without the lock held, so that it can sleep if it needs to.
1987 static int do_page_mkwrite(struct vm_area_struct
*vma
, struct page
*page
,
1988 unsigned long address
)
1990 struct vm_fault vmf
;
1993 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
1994 vmf
.pgoff
= page
->index
;
1995 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
1996 vmf
.gfp_mask
= __get_fault_gfp_mask(vma
);
1998 vmf
.cow_page
= NULL
;
2000 ret
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
2001 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2003 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
2005 if (!page
->mapping
) {
2007 return 0; /* retry */
2009 ret
|= VM_FAULT_LOCKED
;
2011 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2016 * Handle write page faults for pages that can be reused in the current vma
2018 * This can happen either due to the mapping being with the VM_SHARED flag,
2019 * or due to us being the last reference standing to the page. In either
2020 * case, all we need to do here is to mark the page as writable and update
2021 * any related book-keeping.
2023 static inline int wp_page_reuse(struct mm_struct
*mm
,
2024 struct vm_area_struct
*vma
, unsigned long address
,
2025 pte_t
*page_table
, spinlock_t
*ptl
, pte_t orig_pte
,
2026 struct page
*page
, int page_mkwrite
,
2032 * Clear the pages cpupid information as the existing
2033 * information potentially belongs to a now completely
2034 * unrelated process.
2037 page_cpupid_xchg_last(page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2039 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2040 entry
= pte_mkyoung(orig_pte
);
2041 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2042 if (ptep_set_access_flags(vma
, address
, page_table
, entry
, 1))
2043 update_mmu_cache(vma
, address
, page_table
);
2044 pte_unmap_unlock(page_table
, ptl
);
2047 struct address_space
*mapping
;
2053 dirtied
= set_page_dirty(page
);
2054 VM_BUG_ON_PAGE(PageAnon(page
), page
);
2055 mapping
= page
->mapping
;
2057 page_cache_release(page
);
2059 if ((dirtied
|| page_mkwrite
) && mapping
) {
2061 * Some device drivers do not set page.mapping
2062 * but still dirty their pages
2064 balance_dirty_pages_ratelimited(mapping
);
2068 file_update_time(vma
->vm_file
);
2071 return VM_FAULT_WRITE
;
2075 * Handle the case of a page which we actually need to copy to a new page.
2077 * Called with mmap_sem locked and the old page referenced, but
2078 * without the ptl held.
2080 * High level logic flow:
2082 * - Allocate a page, copy the content of the old page to the new one.
2083 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2084 * - Take the PTL. If the pte changed, bail out and release the allocated page
2085 * - If the pte is still the way we remember it, update the page table and all
2086 * relevant references. This includes dropping the reference the page-table
2087 * held to the old page, as well as updating the rmap.
2088 * - In any case, unlock the PTL and drop the reference we took to the old page.
2090 static int wp_page_copy(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2091 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2092 pte_t orig_pte
, struct page
*old_page
)
2094 struct page
*new_page
= NULL
;
2095 spinlock_t
*ptl
= NULL
;
2097 int page_copied
= 0;
2098 const unsigned long mmun_start
= address
& PAGE_MASK
; /* For mmu_notifiers */
2099 const unsigned long mmun_end
= mmun_start
+ PAGE_SIZE
; /* For mmu_notifiers */
2100 struct mem_cgroup
*memcg
;
2102 if (unlikely(anon_vma_prepare(vma
)))
2105 if (is_zero_pfn(pte_pfn(orig_pte
))) {
2106 new_page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2110 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2113 cow_user_page(new_page
, old_page
, address
, vma
);
2116 if (mem_cgroup_try_charge(new_page
, mm
, GFP_KERNEL
, &memcg
, false))
2119 __SetPageUptodate(new_page
);
2121 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2124 * Re-check the pte - we dropped the lock
2126 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2127 if (likely(pte_same(*page_table
, orig_pte
))) {
2129 if (!PageAnon(old_page
)) {
2130 dec_mm_counter_fast(mm
,
2131 mm_counter_file(old_page
));
2132 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2135 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2137 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2138 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2139 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2141 * Clear the pte entry and flush it first, before updating the
2142 * pte with the new entry. This will avoid a race condition
2143 * seen in the presence of one thread doing SMC and another
2146 ptep_clear_flush_notify(vma
, address
, page_table
);
2147 page_add_new_anon_rmap(new_page
, vma
, address
, false);
2148 mem_cgroup_commit_charge(new_page
, memcg
, false, false);
2149 lru_cache_add_active_or_unevictable(new_page
, vma
);
2151 * We call the notify macro here because, when using secondary
2152 * mmu page tables (such as kvm shadow page tables), we want the
2153 * new page to be mapped directly into the secondary page table.
2155 set_pte_at_notify(mm
, address
, page_table
, entry
);
2156 update_mmu_cache(vma
, address
, page_table
);
2159 * Only after switching the pte to the new page may
2160 * we remove the mapcount here. Otherwise another
2161 * process may come and find the rmap count decremented
2162 * before the pte is switched to the new page, and
2163 * "reuse" the old page writing into it while our pte
2164 * here still points into it and can be read by other
2167 * The critical issue is to order this
2168 * page_remove_rmap with the ptp_clear_flush above.
2169 * Those stores are ordered by (if nothing else,)
2170 * the barrier present in the atomic_add_negative
2171 * in page_remove_rmap.
2173 * Then the TLB flush in ptep_clear_flush ensures that
2174 * no process can access the old page before the
2175 * decremented mapcount is visible. And the old page
2176 * cannot be reused until after the decremented
2177 * mapcount is visible. So transitively, TLBs to
2178 * old page will be flushed before it can be reused.
2180 page_remove_rmap(old_page
, false);
2183 /* Free the old page.. */
2184 new_page
= old_page
;
2187 mem_cgroup_cancel_charge(new_page
, memcg
, false);
2191 page_cache_release(new_page
);
2193 pte_unmap_unlock(page_table
, ptl
);
2194 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2197 * Don't let another task, with possibly unlocked vma,
2198 * keep the mlocked page.
2200 if (page_copied
&& (vma
->vm_flags
& VM_LOCKED
)) {
2201 lock_page(old_page
); /* LRU manipulation */
2202 if (PageMlocked(old_page
))
2203 munlock_vma_page(old_page
);
2204 unlock_page(old_page
);
2206 page_cache_release(old_page
);
2208 return page_copied
? VM_FAULT_WRITE
: 0;
2210 page_cache_release(new_page
);
2213 page_cache_release(old_page
);
2214 return VM_FAULT_OOM
;
2218 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2221 static int wp_pfn_shared(struct mm_struct
*mm
,
2222 struct vm_area_struct
*vma
, unsigned long address
,
2223 pte_t
*page_table
, spinlock_t
*ptl
, pte_t orig_pte
,
2226 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
2227 struct vm_fault vmf
= {
2229 .pgoff
= linear_page_index(vma
, address
),
2230 .virtual_address
= (void __user
*)(address
& PAGE_MASK
),
2231 .flags
= FAULT_FLAG_WRITE
| FAULT_FLAG_MKWRITE
,
2235 pte_unmap_unlock(page_table
, ptl
);
2236 ret
= vma
->vm_ops
->pfn_mkwrite(vma
, &vmf
);
2237 if (ret
& VM_FAULT_ERROR
)
2239 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2241 * We might have raced with another page fault while we
2242 * released the pte_offset_map_lock.
2244 if (!pte_same(*page_table
, orig_pte
)) {
2245 pte_unmap_unlock(page_table
, ptl
);
2249 return wp_page_reuse(mm
, vma
, address
, page_table
, ptl
, orig_pte
,
2253 static int wp_page_shared(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2254 unsigned long address
, pte_t
*page_table
,
2255 pmd_t
*pmd
, spinlock_t
*ptl
, pte_t orig_pte
,
2256 struct page
*old_page
)
2259 int page_mkwrite
= 0;
2261 page_cache_get(old_page
);
2263 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2266 pte_unmap_unlock(page_table
, ptl
);
2267 tmp
= do_page_mkwrite(vma
, old_page
, address
);
2268 if (unlikely(!tmp
|| (tmp
&
2269 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
2270 page_cache_release(old_page
);
2274 * Since we dropped the lock we need to revalidate
2275 * the PTE as someone else may have changed it. If
2276 * they did, we just return, as we can count on the
2277 * MMU to tell us if they didn't also make it writable.
2279 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2281 if (!pte_same(*page_table
, orig_pte
)) {
2282 unlock_page(old_page
);
2283 pte_unmap_unlock(page_table
, ptl
);
2284 page_cache_release(old_page
);
2290 return wp_page_reuse(mm
, vma
, address
, page_table
, ptl
,
2291 orig_pte
, old_page
, page_mkwrite
, 1);
2295 * This routine handles present pages, when users try to write
2296 * to a shared page. It is done by copying the page to a new address
2297 * and decrementing the shared-page counter for the old page.
2299 * Note that this routine assumes that the protection checks have been
2300 * done by the caller (the low-level page fault routine in most cases).
2301 * Thus we can safely just mark it writable once we've done any necessary
2304 * We also mark the page dirty at this point even though the page will
2305 * change only once the write actually happens. This avoids a few races,
2306 * and potentially makes it more efficient.
2308 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2309 * but allow concurrent faults), with pte both mapped and locked.
2310 * We return with mmap_sem still held, but pte unmapped and unlocked.
2312 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2313 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2314 spinlock_t
*ptl
, pte_t orig_pte
)
2317 struct page
*old_page
;
2319 old_page
= vm_normal_page(vma
, address
, orig_pte
);
2322 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2325 * We should not cow pages in a shared writeable mapping.
2326 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2328 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2329 (VM_WRITE
|VM_SHARED
))
2330 return wp_pfn_shared(mm
, vma
, address
, page_table
, ptl
,
2333 pte_unmap_unlock(page_table
, ptl
);
2334 return wp_page_copy(mm
, vma
, address
, page_table
, pmd
,
2335 orig_pte
, old_page
);
2339 * Take out anonymous pages first, anonymous shared vmas are
2340 * not dirty accountable.
2342 if (PageAnon(old_page
) && !PageKsm(old_page
)) {
2343 if (!trylock_page(old_page
)) {
2344 page_cache_get(old_page
);
2345 pte_unmap_unlock(page_table
, ptl
);
2346 lock_page(old_page
);
2347 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2349 if (!pte_same(*page_table
, orig_pte
)) {
2350 unlock_page(old_page
);
2351 pte_unmap_unlock(page_table
, ptl
);
2352 page_cache_release(old_page
);
2355 page_cache_release(old_page
);
2357 if (reuse_swap_page(old_page
)) {
2359 * The page is all ours. Move it to our anon_vma so
2360 * the rmap code will not search our parent or siblings.
2361 * Protected against the rmap code by the page lock.
2363 page_move_anon_rmap(old_page
, vma
, address
);
2364 unlock_page(old_page
);
2365 return wp_page_reuse(mm
, vma
, address
, page_table
, ptl
,
2366 orig_pte
, old_page
, 0, 0);
2368 unlock_page(old_page
);
2369 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2370 (VM_WRITE
|VM_SHARED
))) {
2371 return wp_page_shared(mm
, vma
, address
, page_table
, pmd
,
2372 ptl
, orig_pte
, old_page
);
2376 * Ok, we need to copy. Oh, well..
2378 page_cache_get(old_page
);
2380 pte_unmap_unlock(page_table
, ptl
);
2381 return wp_page_copy(mm
, vma
, address
, page_table
, pmd
,
2382 orig_pte
, old_page
);
2385 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2386 unsigned long start_addr
, unsigned long end_addr
,
2387 struct zap_details
*details
)
2389 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2392 static inline void unmap_mapping_range_tree(struct rb_root
*root
,
2393 struct zap_details
*details
)
2395 struct vm_area_struct
*vma
;
2396 pgoff_t vba
, vea
, zba
, zea
;
2398 vma_interval_tree_foreach(vma
, root
,
2399 details
->first_index
, details
->last_index
) {
2401 vba
= vma
->vm_pgoff
;
2402 vea
= vba
+ vma_pages(vma
) - 1;
2403 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2404 zba
= details
->first_index
;
2407 zea
= details
->last_index
;
2411 unmap_mapping_range_vma(vma
,
2412 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2413 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2419 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2420 * address_space corresponding to the specified page range in the underlying
2423 * @mapping: the address space containing mmaps to be unmapped.
2424 * @holebegin: byte in first page to unmap, relative to the start of
2425 * the underlying file. This will be rounded down to a PAGE_SIZE
2426 * boundary. Note that this is different from truncate_pagecache(), which
2427 * must keep the partial page. In contrast, we must get rid of
2429 * @holelen: size of prospective hole in bytes. This will be rounded
2430 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2432 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2433 * but 0 when invalidating pagecache, don't throw away private data.
2435 void unmap_mapping_range(struct address_space
*mapping
,
2436 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2438 struct zap_details details
;
2439 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2440 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2442 /* Check for overflow. */
2443 if (sizeof(holelen
) > sizeof(hlen
)) {
2445 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2446 if (holeend
& ~(long long)ULONG_MAX
)
2447 hlen
= ULONG_MAX
- hba
+ 1;
2450 details
.check_mapping
= even_cows
? NULL
: mapping
;
2451 details
.first_index
= hba
;
2452 details
.last_index
= hba
+ hlen
- 1;
2453 if (details
.last_index
< details
.first_index
)
2454 details
.last_index
= ULONG_MAX
;
2457 /* DAX uses i_mmap_lock to serialise file truncate vs page fault */
2458 i_mmap_lock_write(mapping
);
2459 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
)))
2460 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2461 i_mmap_unlock_write(mapping
);
2463 EXPORT_SYMBOL(unmap_mapping_range
);
2466 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2467 * but allow concurrent faults), and pte mapped but not yet locked.
2468 * We return with pte unmapped and unlocked.
2470 * We return with the mmap_sem locked or unlocked in the same cases
2471 * as does filemap_fault().
2473 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2474 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2475 unsigned int flags
, pte_t orig_pte
)
2478 struct page
*page
, *swapcache
;
2479 struct mem_cgroup
*memcg
;
2486 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2489 entry
= pte_to_swp_entry(orig_pte
);
2490 if (unlikely(non_swap_entry(entry
))) {
2491 if (is_migration_entry(entry
)) {
2492 migration_entry_wait(mm
, pmd
, address
);
2493 } else if (is_hwpoison_entry(entry
)) {
2494 ret
= VM_FAULT_HWPOISON
;
2496 print_bad_pte(vma
, address
, orig_pte
, NULL
);
2497 ret
= VM_FAULT_SIGBUS
;
2501 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2502 page
= lookup_swap_cache(entry
);
2504 page
= swapin_readahead(entry
,
2505 GFP_HIGHUSER_MOVABLE
, vma
, address
);
2508 * Back out if somebody else faulted in this pte
2509 * while we released the pte lock.
2511 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2512 if (likely(pte_same(*page_table
, orig_pte
)))
2514 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2518 /* Had to read the page from swap area: Major fault */
2519 ret
= VM_FAULT_MAJOR
;
2520 count_vm_event(PGMAJFAULT
);
2521 mem_cgroup_count_vm_event(mm
, PGMAJFAULT
);
2522 } else if (PageHWPoison(page
)) {
2524 * hwpoisoned dirty swapcache pages are kept for killing
2525 * owner processes (which may be unknown at hwpoison time)
2527 ret
= VM_FAULT_HWPOISON
;
2528 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2534 locked
= lock_page_or_retry(page
, mm
, flags
);
2536 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2538 ret
|= VM_FAULT_RETRY
;
2543 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2544 * release the swapcache from under us. The page pin, and pte_same
2545 * test below, are not enough to exclude that. Even if it is still
2546 * swapcache, we need to check that the page's swap has not changed.
2548 if (unlikely(!PageSwapCache(page
) || page_private(page
) != entry
.val
))
2551 page
= ksm_might_need_to_copy(page
, vma
, address
);
2552 if (unlikely(!page
)) {
2558 if (mem_cgroup_try_charge(page
, mm
, GFP_KERNEL
, &memcg
, false)) {
2564 * Back out if somebody else already faulted in this pte.
2566 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2567 if (unlikely(!pte_same(*page_table
, orig_pte
)))
2570 if (unlikely(!PageUptodate(page
))) {
2571 ret
= VM_FAULT_SIGBUS
;
2576 * The page isn't present yet, go ahead with the fault.
2578 * Be careful about the sequence of operations here.
2579 * To get its accounting right, reuse_swap_page() must be called
2580 * while the page is counted on swap but not yet in mapcount i.e.
2581 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2582 * must be called after the swap_free(), or it will never succeed.
2585 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2586 dec_mm_counter_fast(mm
, MM_SWAPENTS
);
2587 pte
= mk_pte(page
, vma
->vm_page_prot
);
2588 if ((flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
)) {
2589 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
2590 flags
&= ~FAULT_FLAG_WRITE
;
2591 ret
|= VM_FAULT_WRITE
;
2592 exclusive
= RMAP_EXCLUSIVE
;
2594 flush_icache_page(vma
, page
);
2595 if (pte_swp_soft_dirty(orig_pte
))
2596 pte
= pte_mksoft_dirty(pte
);
2597 set_pte_at(mm
, address
, page_table
, pte
);
2598 if (page
== swapcache
) {
2599 do_page_add_anon_rmap(page
, vma
, address
, exclusive
);
2600 mem_cgroup_commit_charge(page
, memcg
, true, false);
2601 } else { /* ksm created a completely new copy */
2602 page_add_new_anon_rmap(page
, vma
, address
, false);
2603 mem_cgroup_commit_charge(page
, memcg
, false, false);
2604 lru_cache_add_active_or_unevictable(page
, vma
);
2608 if (mem_cgroup_swap_full(page
) ||
2609 (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
2610 try_to_free_swap(page
);
2612 if (page
!= swapcache
) {
2614 * Hold the lock to avoid the swap entry to be reused
2615 * until we take the PT lock for the pte_same() check
2616 * (to avoid false positives from pte_same). For
2617 * further safety release the lock after the swap_free
2618 * so that the swap count won't change under a
2619 * parallel locked swapcache.
2621 unlock_page(swapcache
);
2622 page_cache_release(swapcache
);
2625 if (flags
& FAULT_FLAG_WRITE
) {
2626 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
2627 if (ret
& VM_FAULT_ERROR
)
2628 ret
&= VM_FAULT_ERROR
;
2632 /* No need to invalidate - it was non-present before */
2633 update_mmu_cache(vma
, address
, page_table
);
2635 pte_unmap_unlock(page_table
, ptl
);
2639 mem_cgroup_cancel_charge(page
, memcg
, false);
2640 pte_unmap_unlock(page_table
, ptl
);
2644 page_cache_release(page
);
2645 if (page
!= swapcache
) {
2646 unlock_page(swapcache
);
2647 page_cache_release(swapcache
);
2653 * This is like a special single-page "expand_{down|up}wards()",
2654 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2655 * doesn't hit another vma.
2657 static inline int check_stack_guard_page(struct vm_area_struct
*vma
, unsigned long address
)
2659 address
&= PAGE_MASK
;
2660 if ((vma
->vm_flags
& VM_GROWSDOWN
) && address
== vma
->vm_start
) {
2661 struct vm_area_struct
*prev
= vma
->vm_prev
;
2664 * Is there a mapping abutting this one below?
2666 * That's only ok if it's the same stack mapping
2667 * that has gotten split..
2669 if (prev
&& prev
->vm_end
== address
)
2670 return prev
->vm_flags
& VM_GROWSDOWN
? 0 : -ENOMEM
;
2672 return expand_downwards(vma
, address
- PAGE_SIZE
);
2674 if ((vma
->vm_flags
& VM_GROWSUP
) && address
+ PAGE_SIZE
== vma
->vm_end
) {
2675 struct vm_area_struct
*next
= vma
->vm_next
;
2677 /* As VM_GROWSDOWN but s/below/above/ */
2678 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
)
2679 return next
->vm_flags
& VM_GROWSUP
? 0 : -ENOMEM
;
2681 return expand_upwards(vma
, address
+ PAGE_SIZE
);
2687 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2688 * but allow concurrent faults), and pte mapped but not yet locked.
2689 * We return with mmap_sem still held, but pte unmapped and unlocked.
2691 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2692 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2695 struct mem_cgroup
*memcg
;
2700 pte_unmap(page_table
);
2702 /* File mapping without ->vm_ops ? */
2703 if (vma
->vm_flags
& VM_SHARED
)
2704 return VM_FAULT_SIGBUS
;
2706 /* Check if we need to add a guard page to the stack */
2707 if (check_stack_guard_page(vma
, address
) < 0)
2708 return VM_FAULT_SIGSEGV
;
2710 /* Use the zero-page for reads */
2711 if (!(flags
& FAULT_FLAG_WRITE
) && !mm_forbids_zeropage(mm
)) {
2712 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(address
),
2713 vma
->vm_page_prot
));
2714 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2715 if (!pte_none(*page_table
))
2717 /* Deliver the page fault to userland, check inside PT lock */
2718 if (userfaultfd_missing(vma
)) {
2719 pte_unmap_unlock(page_table
, ptl
);
2720 return handle_userfault(vma
, address
, flags
,
2726 /* Allocate our own private page. */
2727 if (unlikely(anon_vma_prepare(vma
)))
2729 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2733 if (mem_cgroup_try_charge(page
, mm
, GFP_KERNEL
, &memcg
, false))
2737 * The memory barrier inside __SetPageUptodate makes sure that
2738 * preceeding stores to the page contents become visible before
2739 * the set_pte_at() write.
2741 __SetPageUptodate(page
);
2743 entry
= mk_pte(page
, vma
->vm_page_prot
);
2744 if (vma
->vm_flags
& VM_WRITE
)
2745 entry
= pte_mkwrite(pte_mkdirty(entry
));
2747 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2748 if (!pte_none(*page_table
))
2751 /* Deliver the page fault to userland, check inside PT lock */
2752 if (userfaultfd_missing(vma
)) {
2753 pte_unmap_unlock(page_table
, ptl
);
2754 mem_cgroup_cancel_charge(page
, memcg
, false);
2755 page_cache_release(page
);
2756 return handle_userfault(vma
, address
, flags
,
2760 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2761 page_add_new_anon_rmap(page
, vma
, address
, false);
2762 mem_cgroup_commit_charge(page
, memcg
, false, false);
2763 lru_cache_add_active_or_unevictable(page
, vma
);
2765 set_pte_at(mm
, address
, page_table
, entry
);
2767 /* No need to invalidate - it was non-present before */
2768 update_mmu_cache(vma
, address
, page_table
);
2770 pte_unmap_unlock(page_table
, ptl
);
2773 mem_cgroup_cancel_charge(page
, memcg
, false);
2774 page_cache_release(page
);
2777 page_cache_release(page
);
2779 return VM_FAULT_OOM
;
2783 * The mmap_sem must have been held on entry, and may have been
2784 * released depending on flags and vma->vm_ops->fault() return value.
2785 * See filemap_fault() and __lock_page_retry().
2787 static int __do_fault(struct vm_area_struct
*vma
, unsigned long address
,
2788 pgoff_t pgoff
, unsigned int flags
,
2789 struct page
*cow_page
, struct page
**page
)
2791 struct vm_fault vmf
;
2794 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
2798 vmf
.gfp_mask
= __get_fault_gfp_mask(vma
);
2799 vmf
.cow_page
= cow_page
;
2801 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
2802 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
2807 if (unlikely(PageHWPoison(vmf
.page
))) {
2808 if (ret
& VM_FAULT_LOCKED
)
2809 unlock_page(vmf
.page
);
2810 page_cache_release(vmf
.page
);
2811 return VM_FAULT_HWPOISON
;
2814 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
2815 lock_page(vmf
.page
);
2817 VM_BUG_ON_PAGE(!PageLocked(vmf
.page
), vmf
.page
);
2825 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2827 * @vma: virtual memory area
2828 * @address: user virtual address
2829 * @page: page to map
2830 * @pte: pointer to target page table entry
2831 * @write: true, if new entry is writable
2832 * @anon: true, if it's anonymous page
2834 * Caller must hold page table lock relevant for @pte.
2836 * Target users are page handler itself and implementations of
2837 * vm_ops->map_pages.
2839 void do_set_pte(struct vm_area_struct
*vma
, unsigned long address
,
2840 struct page
*page
, pte_t
*pte
, bool write
, bool anon
)
2844 flush_icache_page(vma
, page
);
2845 entry
= mk_pte(page
, vma
->vm_page_prot
);
2847 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2849 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
2850 page_add_new_anon_rmap(page
, vma
, address
, false);
2852 inc_mm_counter_fast(vma
->vm_mm
, mm_counter_file(page
));
2853 page_add_file_rmap(page
);
2855 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
2857 /* no need to invalidate: a not-present page won't be cached */
2858 update_mmu_cache(vma
, address
, pte
);
2861 static unsigned long fault_around_bytes __read_mostly
=
2862 rounddown_pow_of_two(65536);
2864 #ifdef CONFIG_DEBUG_FS
2865 static int fault_around_bytes_get(void *data
, u64
*val
)
2867 *val
= fault_around_bytes
;
2872 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2873 * rounded down to nearest page order. It's what do_fault_around() expects to
2876 static int fault_around_bytes_set(void *data
, u64 val
)
2878 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
2880 if (val
> PAGE_SIZE
)
2881 fault_around_bytes
= rounddown_pow_of_two(val
);
2883 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
2886 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops
,
2887 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
2889 static int __init
fault_around_debugfs(void)
2893 ret
= debugfs_create_file("fault_around_bytes", 0644, NULL
, NULL
,
2894 &fault_around_bytes_fops
);
2896 pr_warn("Failed to create fault_around_bytes in debugfs");
2899 late_initcall(fault_around_debugfs
);
2903 * do_fault_around() tries to map few pages around the fault address. The hope
2904 * is that the pages will be needed soon and this will lower the number of
2907 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2908 * not ready to be mapped: not up-to-date, locked, etc.
2910 * This function is called with the page table lock taken. In the split ptlock
2911 * case the page table lock only protects only those entries which belong to
2912 * the page table corresponding to the fault address.
2914 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2917 * fault_around_pages() defines how many pages we'll try to map.
2918 * do_fault_around() expects it to return a power of two less than or equal to
2921 * The virtual address of the area that we map is naturally aligned to the
2922 * fault_around_pages() value (and therefore to page order). This way it's
2923 * easier to guarantee that we don't cross page table boundaries.
2925 static void do_fault_around(struct vm_area_struct
*vma
, unsigned long address
,
2926 pte_t
*pte
, pgoff_t pgoff
, unsigned int flags
)
2928 unsigned long start_addr
, nr_pages
, mask
;
2930 struct vm_fault vmf
;
2933 nr_pages
= READ_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
2934 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
2936 start_addr
= max(address
& mask
, vma
->vm_start
);
2937 off
= ((address
- start_addr
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
2942 * max_pgoff is either end of page table or end of vma
2943 * or fault_around_pages() from pgoff, depending what is nearest.
2945 max_pgoff
= pgoff
- ((start_addr
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
2947 max_pgoff
= min3(max_pgoff
, vma_pages(vma
) + vma
->vm_pgoff
- 1,
2948 pgoff
+ nr_pages
- 1);
2950 /* Check if it makes any sense to call ->map_pages */
2951 while (!pte_none(*pte
)) {
2952 if (++pgoff
> max_pgoff
)
2954 start_addr
+= PAGE_SIZE
;
2955 if (start_addr
>= vma
->vm_end
)
2960 vmf
.virtual_address
= (void __user
*) start_addr
;
2963 vmf
.max_pgoff
= max_pgoff
;
2965 vmf
.gfp_mask
= __get_fault_gfp_mask(vma
);
2966 vma
->vm_ops
->map_pages(vma
, &vmf
);
2969 static int do_read_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2970 unsigned long address
, pmd_t
*pmd
,
2971 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
2973 struct page
*fault_page
;
2979 * Let's call ->map_pages() first and use ->fault() as fallback
2980 * if page by the offset is not ready to be mapped (cold cache or
2983 if (vma
->vm_ops
->map_pages
&& fault_around_bytes
>> PAGE_SHIFT
> 1) {
2984 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2985 do_fault_around(vma
, address
, pte
, pgoff
, flags
);
2986 if (!pte_same(*pte
, orig_pte
))
2988 pte_unmap_unlock(pte
, ptl
);
2991 ret
= __do_fault(vma
, address
, pgoff
, flags
, NULL
, &fault_page
);
2992 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
2995 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2996 if (unlikely(!pte_same(*pte
, orig_pte
))) {
2997 pte_unmap_unlock(pte
, ptl
);
2998 unlock_page(fault_page
);
2999 page_cache_release(fault_page
);
3002 do_set_pte(vma
, address
, fault_page
, pte
, false, false);
3003 unlock_page(fault_page
);
3005 pte_unmap_unlock(pte
, ptl
);
3009 static int do_cow_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3010 unsigned long address
, pmd_t
*pmd
,
3011 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3013 struct page
*fault_page
, *new_page
;
3014 struct mem_cgroup
*memcg
;
3019 if (unlikely(anon_vma_prepare(vma
)))
3020 return VM_FAULT_OOM
;
3022 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
3024 return VM_FAULT_OOM
;
3026 if (mem_cgroup_try_charge(new_page
, mm
, GFP_KERNEL
, &memcg
, false)) {
3027 page_cache_release(new_page
);
3028 return VM_FAULT_OOM
;
3031 ret
= __do_fault(vma
, address
, pgoff
, flags
, new_page
, &fault_page
);
3032 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3036 copy_user_highpage(new_page
, fault_page
, address
, vma
);
3037 __SetPageUptodate(new_page
);
3039 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3040 if (unlikely(!pte_same(*pte
, orig_pte
))) {
3041 pte_unmap_unlock(pte
, ptl
);
3043 unlock_page(fault_page
);
3044 page_cache_release(fault_page
);
3047 * The fault handler has no page to lock, so it holds
3048 * i_mmap_lock for read to protect against truncate.
3050 i_mmap_unlock_read(vma
->vm_file
->f_mapping
);
3054 do_set_pte(vma
, address
, new_page
, pte
, true, true);
3055 mem_cgroup_commit_charge(new_page
, memcg
, false, false);
3056 lru_cache_add_active_or_unevictable(new_page
, vma
);
3057 pte_unmap_unlock(pte
, ptl
);
3059 unlock_page(fault_page
);
3060 page_cache_release(fault_page
);
3063 * The fault handler has no page to lock, so it holds
3064 * i_mmap_lock for read to protect against truncate.
3066 i_mmap_unlock_read(vma
->vm_file
->f_mapping
);
3070 mem_cgroup_cancel_charge(new_page
, memcg
, false);
3071 page_cache_release(new_page
);
3075 static int do_shared_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3076 unsigned long address
, pmd_t
*pmd
,
3077 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3079 struct page
*fault_page
;
3080 struct address_space
*mapping
;
3086 ret
= __do_fault(vma
, address
, pgoff
, flags
, NULL
, &fault_page
);
3087 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3091 * Check if the backing address space wants to know that the page is
3092 * about to become writable
3094 if (vma
->vm_ops
->page_mkwrite
) {
3095 unlock_page(fault_page
);
3096 tmp
= do_page_mkwrite(vma
, fault_page
, address
);
3097 if (unlikely(!tmp
||
3098 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3099 page_cache_release(fault_page
);
3104 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3105 if (unlikely(!pte_same(*pte
, orig_pte
))) {
3106 pte_unmap_unlock(pte
, ptl
);
3107 unlock_page(fault_page
);
3108 page_cache_release(fault_page
);
3111 do_set_pte(vma
, address
, fault_page
, pte
, true, false);
3112 pte_unmap_unlock(pte
, ptl
);
3114 if (set_page_dirty(fault_page
))
3117 * Take a local copy of the address_space - page.mapping may be zeroed
3118 * by truncate after unlock_page(). The address_space itself remains
3119 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3120 * release semantics to prevent the compiler from undoing this copying.
3122 mapping
= page_rmapping(fault_page
);
3123 unlock_page(fault_page
);
3124 if ((dirtied
|| vma
->vm_ops
->page_mkwrite
) && mapping
) {
3126 * Some device drivers do not set page.mapping but still
3129 balance_dirty_pages_ratelimited(mapping
);
3132 if (!vma
->vm_ops
->page_mkwrite
)
3133 file_update_time(vma
->vm_file
);
3139 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3140 * but allow concurrent faults).
3141 * The mmap_sem may have been released depending on flags and our
3142 * return value. See filemap_fault() and __lock_page_or_retry().
3144 static int do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3145 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3146 unsigned int flags
, pte_t orig_pte
)
3148 pgoff_t pgoff
= linear_page_index(vma
, address
);
3150 pte_unmap(page_table
);
3151 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3152 if (!vma
->vm_ops
->fault
)
3153 return VM_FAULT_SIGBUS
;
3154 if (!(flags
& FAULT_FLAG_WRITE
))
3155 return do_read_fault(mm
, vma
, address
, pmd
, pgoff
, flags
,
3157 if (!(vma
->vm_flags
& VM_SHARED
))
3158 return do_cow_fault(mm
, vma
, address
, pmd
, pgoff
, flags
,
3160 return do_shared_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3163 static int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3164 unsigned long addr
, int page_nid
,
3169 count_vm_numa_event(NUMA_HINT_FAULTS
);
3170 if (page_nid
== numa_node_id()) {
3171 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3172 *flags
|= TNF_FAULT_LOCAL
;
3175 return mpol_misplaced(page
, vma
, addr
);
3178 static int do_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3179 unsigned long addr
, pte_t pte
, pte_t
*ptep
, pmd_t
*pmd
)
3181 struct page
*page
= NULL
;
3186 bool migrated
= false;
3187 bool was_writable
= pte_write(pte
);
3190 /* A PROT_NONE fault should not end up here */
3191 BUG_ON(!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)));
3194 * The "pte" at this point cannot be used safely without
3195 * validation through pte_unmap_same(). It's of NUMA type but
3196 * the pfn may be screwed if the read is non atomic.
3198 * We can safely just do a "set_pte_at()", because the old
3199 * page table entry is not accessible, so there would be no
3200 * concurrent hardware modifications to the PTE.
3202 ptl
= pte_lockptr(mm
, pmd
);
3204 if (unlikely(!pte_same(*ptep
, pte
))) {
3205 pte_unmap_unlock(ptep
, ptl
);
3209 /* Make it present again */
3210 pte
= pte_modify(pte
, vma
->vm_page_prot
);
3211 pte
= pte_mkyoung(pte
);
3213 pte
= pte_mkwrite(pte
);
3214 set_pte_at(mm
, addr
, ptep
, pte
);
3215 update_mmu_cache(vma
, addr
, ptep
);
3217 page
= vm_normal_page(vma
, addr
, pte
);
3219 pte_unmap_unlock(ptep
, ptl
);
3223 /* TODO: handle PTE-mapped THP */
3224 if (PageCompound(page
)) {
3225 pte_unmap_unlock(ptep
, ptl
);
3230 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3231 * much anyway since they can be in shared cache state. This misses
3232 * the case where a mapping is writable but the process never writes
3233 * to it but pte_write gets cleared during protection updates and
3234 * pte_dirty has unpredictable behaviour between PTE scan updates,
3235 * background writeback, dirty balancing and application behaviour.
3237 if (!(vma
->vm_flags
& VM_WRITE
))
3238 flags
|= TNF_NO_GROUP
;
3241 * Flag if the page is shared between multiple address spaces. This
3242 * is later used when determining whether to group tasks together
3244 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
3245 flags
|= TNF_SHARED
;
3247 last_cpupid
= page_cpupid_last(page
);
3248 page_nid
= page_to_nid(page
);
3249 target_nid
= numa_migrate_prep(page
, vma
, addr
, page_nid
, &flags
);
3250 pte_unmap_unlock(ptep
, ptl
);
3251 if (target_nid
== -1) {
3256 /* Migrate to the requested node */
3257 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
3259 page_nid
= target_nid
;
3260 flags
|= TNF_MIGRATED
;
3262 flags
|= TNF_MIGRATE_FAIL
;
3266 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
3270 static int create_huge_pmd(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3271 unsigned long address
, pmd_t
*pmd
, unsigned int flags
)
3273 if (vma_is_anonymous(vma
))
3274 return do_huge_pmd_anonymous_page(mm
, vma
, address
, pmd
, flags
);
3275 if (vma
->vm_ops
->pmd_fault
)
3276 return vma
->vm_ops
->pmd_fault(vma
, address
, pmd
, flags
);
3277 return VM_FAULT_FALLBACK
;
3280 static int wp_huge_pmd(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3281 unsigned long address
, pmd_t
*pmd
, pmd_t orig_pmd
,
3284 if (vma_is_anonymous(vma
))
3285 return do_huge_pmd_wp_page(mm
, vma
, address
, pmd
, orig_pmd
);
3286 if (vma
->vm_ops
->pmd_fault
)
3287 return vma
->vm_ops
->pmd_fault(vma
, address
, pmd
, flags
);
3288 return VM_FAULT_FALLBACK
;
3292 * These routines also need to handle stuff like marking pages dirty
3293 * and/or accessed for architectures that don't do it in hardware (most
3294 * RISC architectures). The early dirtying is also good on the i386.
3296 * There is also a hook called "update_mmu_cache()" that architectures
3297 * with external mmu caches can use to update those (ie the Sparc or
3298 * PowerPC hashed page tables that act as extended TLBs).
3300 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3301 * but allow concurrent faults), and pte mapped but not yet locked.
3302 * We return with pte unmapped and unlocked.
3304 * The mmap_sem may have been released depending on flags and our
3305 * return value. See filemap_fault() and __lock_page_or_retry().
3307 static int handle_pte_fault(struct mm_struct
*mm
,
3308 struct vm_area_struct
*vma
, unsigned long address
,
3309 pte_t
*pte
, pmd_t
*pmd
, unsigned int flags
)
3315 * some architectures can have larger ptes than wordsize,
3316 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3317 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3318 * The code below just needs a consistent view for the ifs and
3319 * we later double check anyway with the ptl lock held. So here
3320 * a barrier will do.
3324 if (!pte_present(entry
)) {
3325 if (pte_none(entry
)) {
3326 if (vma_is_anonymous(vma
))
3327 return do_anonymous_page(mm
, vma
, address
,
3330 return do_fault(mm
, vma
, address
, pte
, pmd
,
3333 return do_swap_page(mm
, vma
, address
,
3334 pte
, pmd
, flags
, entry
);
3337 if (pte_protnone(entry
))
3338 return do_numa_page(mm
, vma
, address
, entry
, pte
, pmd
);
3340 ptl
= pte_lockptr(mm
, pmd
);
3342 if (unlikely(!pte_same(*pte
, entry
)))
3344 if (flags
& FAULT_FLAG_WRITE
) {
3345 if (!pte_write(entry
))
3346 return do_wp_page(mm
, vma
, address
,
3347 pte
, pmd
, ptl
, entry
);
3348 entry
= pte_mkdirty(entry
);
3350 entry
= pte_mkyoung(entry
);
3351 if (ptep_set_access_flags(vma
, address
, pte
, entry
, flags
& FAULT_FLAG_WRITE
)) {
3352 update_mmu_cache(vma
, address
, pte
);
3355 * This is needed only for protection faults but the arch code
3356 * is not yet telling us if this is a protection fault or not.
3357 * This still avoids useless tlb flushes for .text page faults
3360 if (flags
& FAULT_FLAG_WRITE
)
3361 flush_tlb_fix_spurious_fault(vma
, address
);
3364 pte_unmap_unlock(pte
, ptl
);
3369 * By the time we get here, we already hold the mm semaphore
3371 * The mmap_sem may have been released depending on flags and our
3372 * return value. See filemap_fault() and __lock_page_or_retry().
3374 static int __handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3375 unsigned long address
, unsigned int flags
)
3382 if (unlikely(is_vm_hugetlb_page(vma
)))
3383 return hugetlb_fault(mm
, vma
, address
, flags
);
3385 pgd
= pgd_offset(mm
, address
);
3386 pud
= pud_alloc(mm
, pgd
, address
);
3388 return VM_FAULT_OOM
;
3389 pmd
= pmd_alloc(mm
, pud
, address
);
3391 return VM_FAULT_OOM
;
3392 if (pmd_none(*pmd
) && transparent_hugepage_enabled(vma
)) {
3393 int ret
= create_huge_pmd(mm
, vma
, address
, pmd
, flags
);
3394 if (!(ret
& VM_FAULT_FALLBACK
))
3397 pmd_t orig_pmd
= *pmd
;
3401 if (pmd_trans_huge(orig_pmd
) || pmd_devmap(orig_pmd
)) {
3402 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
3404 if (pmd_protnone(orig_pmd
))
3405 return do_huge_pmd_numa_page(mm
, vma
, address
,
3408 if (dirty
&& !pmd_write(orig_pmd
)) {
3409 ret
= wp_huge_pmd(mm
, vma
, address
, pmd
,
3411 if (!(ret
& VM_FAULT_FALLBACK
))
3414 huge_pmd_set_accessed(mm
, vma
, address
, pmd
,
3422 * Use __pte_alloc instead of pte_alloc_map, because we can't
3423 * run pte_offset_map on the pmd, if an huge pmd could
3424 * materialize from under us from a different thread.
3426 if (unlikely(pmd_none(*pmd
)) &&
3427 unlikely(__pte_alloc(mm
, vma
, pmd
, address
)))
3428 return VM_FAULT_OOM
;
3430 * If a huge pmd materialized under us just retry later. Use
3431 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
3432 * didn't become pmd_trans_huge under us and then back to pmd_none, as
3433 * a result of MADV_DONTNEED running immediately after a huge pmd fault
3434 * in a different thread of this mm, in turn leading to a misleading
3435 * pmd_trans_huge() retval. All we have to ensure is that it is a
3436 * regular pmd that we can walk with pte_offset_map() and we can do that
3437 * through an atomic read in C, which is what pmd_trans_unstable()
3440 if (unlikely(pmd_trans_unstable(pmd
) || pmd_devmap(*pmd
)))
3443 * A regular pmd is established and it can't morph into a huge pmd
3444 * from under us anymore at this point because we hold the mmap_sem
3445 * read mode and khugepaged takes it in write mode. So now it's
3446 * safe to run pte_offset_map().
3448 pte
= pte_offset_map(pmd
, address
);
3450 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
3454 * By the time we get here, we already hold the mm semaphore
3456 * The mmap_sem may have been released depending on flags and our
3457 * return value. See filemap_fault() and __lock_page_or_retry().
3459 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3460 unsigned long address
, unsigned int flags
)
3464 __set_current_state(TASK_RUNNING
);
3466 count_vm_event(PGFAULT
);
3467 mem_cgroup_count_vm_event(mm
, PGFAULT
);
3469 /* do counter updates before entering really critical section. */
3470 check_sync_rss_stat(current
);
3473 * Enable the memcg OOM handling for faults triggered in user
3474 * space. Kernel faults are handled more gracefully.
3476 if (flags
& FAULT_FLAG_USER
)
3477 mem_cgroup_oom_enable();
3479 ret
= __handle_mm_fault(mm
, vma
, address
, flags
);
3481 if (flags
& FAULT_FLAG_USER
) {
3482 mem_cgroup_oom_disable();
3484 * The task may have entered a memcg OOM situation but
3485 * if the allocation error was handled gracefully (no
3486 * VM_FAULT_OOM), there is no need to kill anything.
3487 * Just clean up the OOM state peacefully.
3489 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
3490 mem_cgroup_oom_synchronize(false);
3495 EXPORT_SYMBOL_GPL(handle_mm_fault
);
3497 #ifndef __PAGETABLE_PUD_FOLDED
3499 * Allocate page upper directory.
3500 * We've already handled the fast-path in-line.
3502 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
3504 pud_t
*new = pud_alloc_one(mm
, address
);
3508 smp_wmb(); /* See comment in __pte_alloc */
3510 spin_lock(&mm
->page_table_lock
);
3511 if (pgd_present(*pgd
)) /* Another has populated it */
3514 pgd_populate(mm
, pgd
, new);
3515 spin_unlock(&mm
->page_table_lock
);
3518 #endif /* __PAGETABLE_PUD_FOLDED */
3520 #ifndef __PAGETABLE_PMD_FOLDED
3522 * Allocate page middle directory.
3523 * We've already handled the fast-path in-line.
3525 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
3527 pmd_t
*new = pmd_alloc_one(mm
, address
);
3531 smp_wmb(); /* See comment in __pte_alloc */
3533 spin_lock(&mm
->page_table_lock
);
3534 #ifndef __ARCH_HAS_4LEVEL_HACK
3535 if (!pud_present(*pud
)) {
3537 pud_populate(mm
, pud
, new);
3538 } else /* Another has populated it */
3541 if (!pgd_present(*pud
)) {
3543 pgd_populate(mm
, pud
, new);
3544 } else /* Another has populated it */
3546 #endif /* __ARCH_HAS_4LEVEL_HACK */
3547 spin_unlock(&mm
->page_table_lock
);
3550 #endif /* __PAGETABLE_PMD_FOLDED */
3552 static int __follow_pte(struct mm_struct
*mm
, unsigned long address
,
3553 pte_t
**ptepp
, spinlock_t
**ptlp
)
3560 pgd
= pgd_offset(mm
, address
);
3561 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
3564 pud
= pud_offset(pgd
, address
);
3565 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
3568 pmd
= pmd_offset(pud
, address
);
3569 VM_BUG_ON(pmd_trans_huge(*pmd
));
3570 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
3573 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3577 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
3580 if (!pte_present(*ptep
))
3585 pte_unmap_unlock(ptep
, *ptlp
);
3590 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
3591 pte_t
**ptepp
, spinlock_t
**ptlp
)
3595 /* (void) is needed to make gcc happy */
3596 (void) __cond_lock(*ptlp
,
3597 !(res
= __follow_pte(mm
, address
, ptepp
, ptlp
)));
3602 * follow_pfn - look up PFN at a user virtual address
3603 * @vma: memory mapping
3604 * @address: user virtual address
3605 * @pfn: location to store found PFN
3607 * Only IO mappings and raw PFN mappings are allowed.
3609 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3611 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
3618 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3621 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
3624 *pfn
= pte_pfn(*ptep
);
3625 pte_unmap_unlock(ptep
, ptl
);
3628 EXPORT_SYMBOL(follow_pfn
);
3630 #ifdef CONFIG_HAVE_IOREMAP_PROT
3631 int follow_phys(struct vm_area_struct
*vma
,
3632 unsigned long address
, unsigned int flags
,
3633 unsigned long *prot
, resource_size_t
*phys
)
3639 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3642 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
3646 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
3649 *prot
= pgprot_val(pte_pgprot(pte
));
3650 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
3654 pte_unmap_unlock(ptep
, ptl
);
3659 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
3660 void *buf
, int len
, int write
)
3662 resource_size_t phys_addr
;
3663 unsigned long prot
= 0;
3664 void __iomem
*maddr
;
3665 int offset
= addr
& (PAGE_SIZE
-1);
3667 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
3670 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
3672 memcpy_toio(maddr
+ offset
, buf
, len
);
3674 memcpy_fromio(buf
, maddr
+ offset
, len
);
3679 EXPORT_SYMBOL_GPL(generic_access_phys
);
3683 * Access another process' address space as given in mm. If non-NULL, use the
3684 * given task for page fault accounting.
3686 static int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
3687 unsigned long addr
, void *buf
, int len
, int write
)
3689 struct vm_area_struct
*vma
;
3690 void *old_buf
= buf
;
3692 down_read(&mm
->mmap_sem
);
3693 /* ignore errors, just check how much was successfully transferred */
3695 int bytes
, ret
, offset
;
3697 struct page
*page
= NULL
;
3699 ret
= get_user_pages(tsk
, mm
, addr
, 1,
3700 write
, 1, &page
, &vma
);
3702 #ifndef CONFIG_HAVE_IOREMAP_PROT
3706 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3707 * we can access using slightly different code.
3709 vma
= find_vma(mm
, addr
);
3710 if (!vma
|| vma
->vm_start
> addr
)
3712 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
3713 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
3721 offset
= addr
& (PAGE_SIZE
-1);
3722 if (bytes
> PAGE_SIZE
-offset
)
3723 bytes
= PAGE_SIZE
-offset
;
3727 copy_to_user_page(vma
, page
, addr
,
3728 maddr
+ offset
, buf
, bytes
);
3729 set_page_dirty_lock(page
);
3731 copy_from_user_page(vma
, page
, addr
,
3732 buf
, maddr
+ offset
, bytes
);
3735 page_cache_release(page
);
3741 up_read(&mm
->mmap_sem
);
3743 return buf
- old_buf
;
3747 * access_remote_vm - access another process' address space
3748 * @mm: the mm_struct of the target address space
3749 * @addr: start address to access
3750 * @buf: source or destination buffer
3751 * @len: number of bytes to transfer
3752 * @write: whether the access is a write
3754 * The caller must hold a reference on @mm.
3756 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
3757 void *buf
, int len
, int write
)
3759 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, write
);
3763 * Access another process' address space.
3764 * Source/target buffer must be kernel space,
3765 * Do not walk the page table directly, use get_user_pages
3767 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
3768 void *buf
, int len
, int write
)
3770 struct mm_struct
*mm
;
3773 mm
= get_task_mm(tsk
);
3777 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, write
);
3784 * Print the name of a VMA.
3786 void print_vma_addr(char *prefix
, unsigned long ip
)
3788 struct mm_struct
*mm
= current
->mm
;
3789 struct vm_area_struct
*vma
;
3792 * Do not print if we are in atomic
3793 * contexts (in exception stacks, etc.):
3795 if (preempt_count())
3798 down_read(&mm
->mmap_sem
);
3799 vma
= find_vma(mm
, ip
);
3800 if (vma
&& vma
->vm_file
) {
3801 struct file
*f
= vma
->vm_file
;
3802 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
3806 p
= file_path(f
, buf
, PAGE_SIZE
);
3809 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
3811 vma
->vm_end
- vma
->vm_start
);
3812 free_page((unsigned long)buf
);
3815 up_read(&mm
->mmap_sem
);
3818 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3819 void __might_fault(const char *file
, int line
)
3822 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3823 * holding the mmap_sem, this is safe because kernel memory doesn't
3824 * get paged out, therefore we'll never actually fault, and the
3825 * below annotations will generate false positives.
3827 if (segment_eq(get_fs(), KERNEL_DS
))
3829 if (pagefault_disabled())
3831 __might_sleep(file
, line
, 0);
3832 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3834 might_lock_read(¤t
->mm
->mmap_sem
);
3837 EXPORT_SYMBOL(__might_fault
);
3840 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3841 static void clear_gigantic_page(struct page
*page
,
3843 unsigned int pages_per_huge_page
)
3846 struct page
*p
= page
;
3849 for (i
= 0; i
< pages_per_huge_page
;
3850 i
++, p
= mem_map_next(p
, page
, i
)) {
3852 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
3855 void clear_huge_page(struct page
*page
,
3856 unsigned long addr
, unsigned int pages_per_huge_page
)
3860 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3861 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
3866 for (i
= 0; i
< pages_per_huge_page
; i
++) {
3868 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
3872 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
3874 struct vm_area_struct
*vma
,
3875 unsigned int pages_per_huge_page
)
3878 struct page
*dst_base
= dst
;
3879 struct page
*src_base
= src
;
3881 for (i
= 0; i
< pages_per_huge_page
; ) {
3883 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
3886 dst
= mem_map_next(dst
, dst_base
, i
);
3887 src
= mem_map_next(src
, src_base
, i
);
3891 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
3892 unsigned long addr
, struct vm_area_struct
*vma
,
3893 unsigned int pages_per_huge_page
)
3897 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3898 copy_user_gigantic_page(dst
, src
, addr
, vma
,
3899 pages_per_huge_page
);
3904 for (i
= 0; i
< pages_per_huge_page
; i
++) {
3906 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
3909 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3911 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3913 static struct kmem_cache
*page_ptl_cachep
;
3915 void __init
ptlock_cache_init(void)
3917 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
3921 bool ptlock_alloc(struct page
*page
)
3925 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
3932 void ptlock_free(struct page
*page
)
3934 kmem_cache_free(page_ptl_cachep
, page
->ptl
);