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/mmu_context.h>
69 #include <asm/pgalloc.h>
70 #include <asm/uaccess.h>
72 #include <asm/tlbflush.h>
73 #include <asm/pgtable.h>
77 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
78 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
81 #ifndef CONFIG_NEED_MULTIPLE_NODES
82 /* use the per-pgdat data instead for discontigmem - mbligh */
83 unsigned long max_mapnr
;
86 EXPORT_SYMBOL(max_mapnr
);
87 EXPORT_SYMBOL(mem_map
);
91 * A number of key systems in x86 including ioremap() rely on the assumption
92 * that high_memory defines the upper bound on direct map memory, then end
93 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
94 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
99 EXPORT_SYMBOL(high_memory
);
102 * Randomize the address space (stacks, mmaps, brk, etc.).
104 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
105 * as ancient (libc5 based) binaries can segfault. )
107 int randomize_va_space __read_mostly
=
108 #ifdef CONFIG_COMPAT_BRK
114 static int __init
disable_randmaps(char *s
)
116 randomize_va_space
= 0;
119 __setup("norandmaps", disable_randmaps
);
121 unsigned long zero_pfn __read_mostly
;
122 unsigned long highest_memmap_pfn __read_mostly
;
124 EXPORT_SYMBOL(zero_pfn
);
127 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
129 static int __init
init_zero_pfn(void)
131 zero_pfn
= page_to_pfn(ZERO_PAGE(0));
134 core_initcall(init_zero_pfn
);
137 #if defined(SPLIT_RSS_COUNTING)
139 void sync_mm_rss(struct mm_struct
*mm
)
143 for (i
= 0; i
< NR_MM_COUNTERS
; i
++) {
144 if (current
->rss_stat
.count
[i
]) {
145 add_mm_counter(mm
, i
, current
->rss_stat
.count
[i
]);
146 current
->rss_stat
.count
[i
] = 0;
149 current
->rss_stat
.events
= 0;
152 static void add_mm_counter_fast(struct mm_struct
*mm
, int member
, int val
)
154 struct task_struct
*task
= current
;
156 if (likely(task
->mm
== mm
))
157 task
->rss_stat
.count
[member
] += val
;
159 add_mm_counter(mm
, member
, val
);
161 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
162 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
164 /* sync counter once per 64 page faults */
165 #define TASK_RSS_EVENTS_THRESH (64)
166 static void check_sync_rss_stat(struct task_struct
*task
)
168 if (unlikely(task
!= current
))
170 if (unlikely(task
->rss_stat
.events
++ > TASK_RSS_EVENTS_THRESH
))
171 sync_mm_rss(task
->mm
);
173 #else /* SPLIT_RSS_COUNTING */
175 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
176 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
178 static void check_sync_rss_stat(struct task_struct
*task
)
182 #endif /* SPLIT_RSS_COUNTING */
184 #ifdef HAVE_GENERIC_MMU_GATHER
186 static bool tlb_next_batch(struct mmu_gather
*tlb
)
188 struct mmu_gather_batch
*batch
;
192 tlb
->active
= batch
->next
;
196 if (tlb
->batch_count
== MAX_GATHER_BATCH_COUNT
)
199 batch
= (void *)__get_free_pages(GFP_NOWAIT
| __GFP_NOWARN
, 0);
206 batch
->max
= MAX_GATHER_BATCH
;
208 tlb
->active
->next
= batch
;
215 * Called to initialize an (on-stack) mmu_gather structure for page-table
216 * tear-down from @mm. The @fullmm argument is used when @mm is without
217 * users and we're going to destroy the full address space (exit/execve).
219 void tlb_gather_mmu(struct mmu_gather
*tlb
, struct mm_struct
*mm
, unsigned long start
, unsigned long end
)
223 /* Is it from 0 to ~0? */
224 tlb
->fullmm
= !(start
| (end
+1));
225 tlb
->need_flush_all
= 0;
226 tlb
->local
.next
= NULL
;
228 tlb
->local
.max
= ARRAY_SIZE(tlb
->__pages
);
229 tlb
->active
= &tlb
->local
;
230 tlb
->batch_count
= 0;
232 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
236 __tlb_reset_range(tlb
);
239 static void tlb_flush_mmu_tlbonly(struct mmu_gather
*tlb
)
245 mmu_notifier_invalidate_range(tlb
->mm
, tlb
->start
, tlb
->end
);
246 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
247 tlb_table_flush(tlb
);
249 __tlb_reset_range(tlb
);
252 static void tlb_flush_mmu_free(struct mmu_gather
*tlb
)
254 struct mmu_gather_batch
*batch
;
256 for (batch
= &tlb
->local
; batch
&& batch
->nr
; batch
= batch
->next
) {
257 free_pages_and_swap_cache(batch
->pages
, batch
->nr
);
260 tlb
->active
= &tlb
->local
;
263 void tlb_flush_mmu(struct mmu_gather
*tlb
)
265 tlb_flush_mmu_tlbonly(tlb
);
266 tlb_flush_mmu_free(tlb
);
270 * Called at the end of the shootdown operation to free up any resources
271 * that were required.
273 void tlb_finish_mmu(struct mmu_gather
*tlb
, unsigned long start
, unsigned long end
)
275 struct mmu_gather_batch
*batch
, *next
;
279 /* keep the page table cache within bounds */
282 for (batch
= tlb
->local
.next
; batch
; batch
= next
) {
284 free_pages((unsigned long)batch
, 0);
286 tlb
->local
.next
= NULL
;
290 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
291 * handling the additional races in SMP caused by other CPUs caching valid
292 * mappings in their TLBs. Returns the number of free page slots left.
293 * When out of page slots we must call tlb_flush_mmu().
295 int __tlb_remove_page(struct mmu_gather
*tlb
, struct page
*page
)
297 struct mmu_gather_batch
*batch
;
299 VM_BUG_ON(!tlb
->end
);
302 batch
->pages
[batch
->nr
++] = page
;
303 if (batch
->nr
== batch
->max
) {
304 if (!tlb_next_batch(tlb
))
308 VM_BUG_ON_PAGE(batch
->nr
> batch
->max
, page
);
310 return batch
->max
- batch
->nr
;
313 #endif /* HAVE_GENERIC_MMU_GATHER */
315 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
318 * See the comment near struct mmu_table_batch.
321 static void tlb_remove_table_smp_sync(void *arg
)
323 /* Simply deliver the interrupt */
326 static void tlb_remove_table_one(void *table
)
329 * This isn't an RCU grace period and hence the page-tables cannot be
330 * assumed to be actually RCU-freed.
332 * It is however sufficient for software page-table walkers that rely on
333 * IRQ disabling. See the comment near struct mmu_table_batch.
335 smp_call_function(tlb_remove_table_smp_sync
, NULL
, 1);
336 __tlb_remove_table(table
);
339 static void tlb_remove_table_rcu(struct rcu_head
*head
)
341 struct mmu_table_batch
*batch
;
344 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
346 for (i
= 0; i
< batch
->nr
; i
++)
347 __tlb_remove_table(batch
->tables
[i
]);
349 free_page((unsigned long)batch
);
352 void tlb_table_flush(struct mmu_gather
*tlb
)
354 struct mmu_table_batch
**batch
= &tlb
->batch
;
357 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
362 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
364 struct mmu_table_batch
**batch
= &tlb
->batch
;
367 * When there's less then two users of this mm there cannot be a
368 * concurrent page-table walk.
370 if (atomic_read(&tlb
->mm
->mm_users
) < 2) {
371 __tlb_remove_table(table
);
375 if (*batch
== NULL
) {
376 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
377 if (*batch
== NULL
) {
378 tlb_remove_table_one(table
);
383 (*batch
)->tables
[(*batch
)->nr
++] = table
;
384 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
385 tlb_table_flush(tlb
);
388 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
391 * Note: this doesn't free the actual pages themselves. That
392 * has been handled earlier when unmapping all the memory regions.
394 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
397 pgtable_t token
= pmd_pgtable(*pmd
);
399 pte_free_tlb(tlb
, token
, addr
);
400 atomic_long_dec(&tlb
->mm
->nr_ptes
);
403 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
404 unsigned long addr
, unsigned long end
,
405 unsigned long floor
, unsigned long ceiling
)
412 pmd
= pmd_offset(pud
, addr
);
414 next
= pmd_addr_end(addr
, end
);
415 if (pmd_none_or_clear_bad(pmd
))
417 free_pte_range(tlb
, pmd
, addr
);
418 } while (pmd
++, addr
= next
, addr
!= end
);
428 if (end
- 1 > ceiling
- 1)
431 pmd
= pmd_offset(pud
, start
);
433 pmd_free_tlb(tlb
, pmd
, start
);
434 mm_dec_nr_pmds(tlb
->mm
);
437 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
438 unsigned long addr
, unsigned long end
,
439 unsigned long floor
, unsigned long ceiling
)
446 pud
= pud_offset(pgd
, addr
);
448 next
= pud_addr_end(addr
, end
);
449 if (pud_none_or_clear_bad(pud
))
451 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
452 } while (pud
++, addr
= next
, addr
!= end
);
458 ceiling
&= PGDIR_MASK
;
462 if (end
- 1 > ceiling
- 1)
465 pud
= pud_offset(pgd
, start
);
467 pud_free_tlb(tlb
, pud
, start
);
471 * This function frees user-level page tables of a process.
473 void free_pgd_range(struct mmu_gather
*tlb
,
474 unsigned long addr
, unsigned long end
,
475 unsigned long floor
, unsigned long ceiling
)
481 * The next few lines have given us lots of grief...
483 * Why are we testing PMD* at this top level? Because often
484 * there will be no work to do at all, and we'd prefer not to
485 * go all the way down to the bottom just to discover that.
487 * Why all these "- 1"s? Because 0 represents both the bottom
488 * of the address space and the top of it (using -1 for the
489 * top wouldn't help much: the masks would do the wrong thing).
490 * The rule is that addr 0 and floor 0 refer to the bottom of
491 * the address space, but end 0 and ceiling 0 refer to the top
492 * Comparisons need to use "end - 1" and "ceiling - 1" (though
493 * that end 0 case should be mythical).
495 * Wherever addr is brought up or ceiling brought down, we must
496 * be careful to reject "the opposite 0" before it confuses the
497 * subsequent tests. But what about where end is brought down
498 * by PMD_SIZE below? no, end can't go down to 0 there.
500 * Whereas we round start (addr) and ceiling down, by different
501 * masks at different levels, in order to test whether a table
502 * now has no other vmas using it, so can be freed, we don't
503 * bother to round floor or end up - the tests don't need that.
517 if (end
- 1 > ceiling
- 1)
522 pgd
= pgd_offset(tlb
->mm
, addr
);
524 next
= pgd_addr_end(addr
, end
);
525 if (pgd_none_or_clear_bad(pgd
))
527 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
528 } while (pgd
++, addr
= next
, addr
!= end
);
531 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
532 unsigned long floor
, unsigned long ceiling
)
535 struct vm_area_struct
*next
= vma
->vm_next
;
536 unsigned long addr
= vma
->vm_start
;
539 * Hide vma from rmap and truncate_pagecache before freeing
542 unlink_anon_vmas(vma
);
543 unlink_file_vma(vma
);
545 if (is_vm_hugetlb_page(vma
)) {
546 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
547 floor
, next
? next
->vm_start
: ceiling
);
550 * Optimization: gather nearby vmas into one call down
552 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
553 && !is_vm_hugetlb_page(next
)) {
556 unlink_anon_vmas(vma
);
557 unlink_file_vma(vma
);
559 free_pgd_range(tlb
, addr
, vma
->vm_end
,
560 floor
, next
? next
->vm_start
: ceiling
);
566 int __pte_alloc(struct mm_struct
*mm
, 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
)) {
664 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
671 resume
= jiffies
+ 60 * HZ
;
673 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
674 index
= linear_page_index(vma
, addr
);
676 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
678 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
680 dump_page(page
, "bad pte");
681 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
682 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
684 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
686 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
688 vma
->vm_ops
? vma
->vm_ops
->fault
: NULL
,
689 vma
->vm_file
? vma
->vm_file
->f_op
->mmap
: NULL
,
690 mapping
? mapping
->a_ops
->readpage
: NULL
);
692 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
696 * vm_normal_page -- This function gets the "struct page" associated with a pte.
698 * "Special" mappings do not wish to be associated with a "struct page" (either
699 * it doesn't exist, or it exists but they don't want to touch it). In this
700 * case, NULL is returned here. "Normal" mappings do have a struct page.
702 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
703 * pte bit, in which case this function is trivial. Secondly, an architecture
704 * may not have a spare pte bit, which requires a more complicated scheme,
707 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
708 * special mapping (even if there are underlying and valid "struct pages").
709 * COWed pages of a VM_PFNMAP are always normal.
711 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
712 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
713 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
714 * mapping will always honor the rule
716 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
718 * And for normal mappings this is false.
720 * This restricts such mappings to be a linear translation from virtual address
721 * to pfn. To get around this restriction, we allow arbitrary mappings so long
722 * as the vma is not a COW mapping; in that case, we know that all ptes are
723 * special (because none can have been COWed).
726 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
728 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
729 * page" backing, however the difference is that _all_ pages with a struct
730 * page (that is, those where pfn_valid is true) are refcounted and considered
731 * normal pages by the VM. The disadvantage is that pages are refcounted
732 * (which can be slower and simply not an option for some PFNMAP users). The
733 * advantage is that we don't have to follow the strict linearity rule of
734 * PFNMAP mappings in order to support COWable mappings.
737 #ifdef __HAVE_ARCH_PTE_SPECIAL
738 # define HAVE_PTE_SPECIAL 1
740 # define HAVE_PTE_SPECIAL 0
742 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
745 unsigned long pfn
= pte_pfn(pte
);
747 if (HAVE_PTE_SPECIAL
) {
748 if (likely(!pte_special(pte
)))
750 if (vma
->vm_ops
&& vma
->vm_ops
->find_special_page
)
751 return vma
->vm_ops
->find_special_page(vma
, addr
);
752 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
754 if (!is_zero_pfn(pfn
))
755 print_bad_pte(vma
, addr
, pte
, NULL
);
759 /* !HAVE_PTE_SPECIAL case follows: */
761 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
762 if (vma
->vm_flags
& VM_MIXEDMAP
) {
768 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
769 if (pfn
== vma
->vm_pgoff
+ off
)
771 if (!is_cow_mapping(vma
->vm_flags
))
776 if (is_zero_pfn(pfn
))
779 if (unlikely(pfn
> highest_memmap_pfn
)) {
780 print_bad_pte(vma
, addr
, pte
, NULL
);
785 * NOTE! We still have PageReserved() pages in the page tables.
786 * eg. VDSO mappings can cause them to exist.
789 return pfn_to_page(pfn
);
792 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
793 struct page
*vm_normal_page_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
796 unsigned long pfn
= pmd_pfn(pmd
);
799 * There is no pmd_special() but there may be special pmds, e.g.
800 * in a direct-access (dax) mapping, so let's just replicate the
801 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
803 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
804 if (vma
->vm_flags
& VM_MIXEDMAP
) {
810 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
811 if (pfn
== vma
->vm_pgoff
+ off
)
813 if (!is_cow_mapping(vma
->vm_flags
))
818 if (is_zero_pfn(pfn
))
820 if (unlikely(pfn
> highest_memmap_pfn
))
824 * NOTE! We still have PageReserved() pages in the page tables.
825 * eg. VDSO mappings can cause them to exist.
828 return pfn_to_page(pfn
);
833 * copy one vm_area from one task to the other. Assumes the page tables
834 * already present in the new task to be cleared in the whole range
835 * covered by this vma.
838 static inline unsigned long
839 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
840 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
841 unsigned long addr
, int *rss
)
843 unsigned long vm_flags
= vma
->vm_flags
;
844 pte_t pte
= *src_pte
;
847 /* pte contains position in swap or file, so copy. */
848 if (unlikely(!pte_present(pte
))) {
849 swp_entry_t entry
= pte_to_swp_entry(pte
);
851 if (likely(!non_swap_entry(entry
))) {
852 if (swap_duplicate(entry
) < 0)
855 /* make sure dst_mm is on swapoff's mmlist. */
856 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
857 spin_lock(&mmlist_lock
);
858 if (list_empty(&dst_mm
->mmlist
))
859 list_add(&dst_mm
->mmlist
,
861 spin_unlock(&mmlist_lock
);
864 } else if (is_migration_entry(entry
)) {
865 page
= migration_entry_to_page(entry
);
867 rss
[mm_counter(page
)]++;
869 if (is_write_migration_entry(entry
) &&
870 is_cow_mapping(vm_flags
)) {
872 * COW mappings require pages in both
873 * parent and child to be set to read.
875 make_migration_entry_read(&entry
);
876 pte
= swp_entry_to_pte(entry
);
877 if (pte_swp_soft_dirty(*src_pte
))
878 pte
= pte_swp_mksoft_dirty(pte
);
879 set_pte_at(src_mm
, addr
, src_pte
, pte
);
886 * If it's a COW mapping, write protect it both
887 * in the parent and the child
889 if (is_cow_mapping(vm_flags
)) {
890 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
891 pte
= pte_wrprotect(pte
);
895 * If it's a shared mapping, mark it clean in
898 if (vm_flags
& VM_SHARED
)
899 pte
= pte_mkclean(pte
);
900 pte
= pte_mkold(pte
);
902 page
= vm_normal_page(vma
, addr
, pte
);
905 page_dup_rmap(page
, false);
906 rss
[mm_counter(page
)]++;
910 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
914 static int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
915 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
916 unsigned long addr
, unsigned long end
)
918 pte_t
*orig_src_pte
, *orig_dst_pte
;
919 pte_t
*src_pte
, *dst_pte
;
920 spinlock_t
*src_ptl
, *dst_ptl
;
922 int rss
[NR_MM_COUNTERS
];
923 swp_entry_t entry
= (swp_entry_t
){0};
928 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
931 src_pte
= pte_offset_map(src_pmd
, addr
);
932 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
933 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
934 orig_src_pte
= src_pte
;
935 orig_dst_pte
= dst_pte
;
936 arch_enter_lazy_mmu_mode();
940 * We are holding two locks at this point - either of them
941 * could generate latencies in another task on another CPU.
943 if (progress
>= 32) {
945 if (need_resched() ||
946 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
949 if (pte_none(*src_pte
)) {
953 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
958 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
960 arch_leave_lazy_mmu_mode();
961 spin_unlock(src_ptl
);
962 pte_unmap(orig_src_pte
);
963 add_mm_rss_vec(dst_mm
, rss
);
964 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
968 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
977 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
978 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
979 unsigned long addr
, unsigned long end
)
981 pmd_t
*src_pmd
, *dst_pmd
;
984 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
987 src_pmd
= pmd_offset(src_pud
, addr
);
989 next
= pmd_addr_end(addr
, end
);
990 if (pmd_trans_huge(*src_pmd
) || pmd_devmap(*src_pmd
)) {
992 VM_BUG_ON(next
-addr
!= HPAGE_PMD_SIZE
);
993 err
= copy_huge_pmd(dst_mm
, src_mm
,
994 dst_pmd
, src_pmd
, addr
, vma
);
1001 if (pmd_none_or_clear_bad(src_pmd
))
1003 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
1006 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1010 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1011 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
1012 unsigned long addr
, unsigned long end
)
1014 pud_t
*src_pud
, *dst_pud
;
1017 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
1020 src_pud
= pud_offset(src_pgd
, addr
);
1022 next
= pud_addr_end(addr
, end
);
1023 if (pud_none_or_clear_bad(src_pud
))
1025 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
1028 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1032 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1033 struct vm_area_struct
*vma
)
1035 pgd_t
*src_pgd
, *dst_pgd
;
1037 unsigned long addr
= vma
->vm_start
;
1038 unsigned long end
= vma
->vm_end
;
1039 unsigned long mmun_start
; /* For mmu_notifiers */
1040 unsigned long mmun_end
; /* For mmu_notifiers */
1045 * Don't copy ptes where a page fault will fill them correctly.
1046 * Fork becomes much lighter when there are big shared or private
1047 * readonly mappings. The tradeoff is that copy_page_range is more
1048 * efficient than faulting.
1050 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_PFNMAP
| VM_MIXEDMAP
)) &&
1054 if (is_vm_hugetlb_page(vma
))
1055 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1057 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
1059 * We do not free on error cases below as remove_vma
1060 * gets called on error from higher level routine
1062 ret
= track_pfn_copy(vma
);
1068 * We need to invalidate the secondary MMU mappings only when
1069 * there could be a permission downgrade on the ptes of the
1070 * parent mm. And a permission downgrade will only happen if
1071 * is_cow_mapping() returns true.
1073 is_cow
= is_cow_mapping(vma
->vm_flags
);
1077 mmu_notifier_invalidate_range_start(src_mm
, mmun_start
,
1081 dst_pgd
= pgd_offset(dst_mm
, addr
);
1082 src_pgd
= pgd_offset(src_mm
, addr
);
1084 next
= pgd_addr_end(addr
, end
);
1085 if (pgd_none_or_clear_bad(src_pgd
))
1087 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1088 vma
, addr
, next
))) {
1092 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1095 mmu_notifier_invalidate_range_end(src_mm
, mmun_start
, mmun_end
);
1099 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1100 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1101 unsigned long addr
, unsigned long end
,
1102 struct zap_details
*details
)
1104 struct mm_struct
*mm
= tlb
->mm
;
1105 int force_flush
= 0;
1106 int rss
[NR_MM_COUNTERS
];
1114 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1116 arch_enter_lazy_mmu_mode();
1119 if (pte_none(ptent
)) {
1123 if (pte_present(ptent
)) {
1126 page
= vm_normal_page(vma
, addr
, ptent
);
1127 if (unlikely(details
) && page
) {
1129 * unmap_shared_mapping_pages() wants to
1130 * invalidate cache without truncating:
1131 * unmap shared but keep private pages.
1133 if (details
->check_mapping
&&
1134 details
->check_mapping
!= page
->mapping
)
1137 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1139 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1140 if (unlikely(!page
))
1143 if (!PageAnon(page
)) {
1144 if (pte_dirty(ptent
)) {
1146 * oom_reaper cannot tear down dirty
1149 if (unlikely(details
&& details
->ignore_dirty
))
1152 set_page_dirty(page
);
1154 if (pte_young(ptent
) &&
1155 likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
1156 mark_page_accessed(page
);
1158 rss
[mm_counter(page
)]--;
1159 page_remove_rmap(page
, false);
1160 if (unlikely(page_mapcount(page
) < 0))
1161 print_bad_pte(vma
, addr
, ptent
, page
);
1162 if (unlikely(!__tlb_remove_page(tlb
, page
))) {
1169 /* only check swap_entries if explicitly asked for in details */
1170 if (unlikely(details
&& !details
->check_swap_entries
))
1173 entry
= pte_to_swp_entry(ptent
);
1174 if (!non_swap_entry(entry
))
1176 else if (is_migration_entry(entry
)) {
1179 page
= migration_entry_to_page(entry
);
1180 rss
[mm_counter(page
)]--;
1182 if (unlikely(!free_swap_and_cache(entry
)))
1183 print_bad_pte(vma
, addr
, ptent
, NULL
);
1184 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1185 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1187 add_mm_rss_vec(mm
, rss
);
1188 arch_leave_lazy_mmu_mode();
1190 /* Do the actual TLB flush before dropping ptl */
1192 tlb_flush_mmu_tlbonly(tlb
);
1193 pte_unmap_unlock(start_pte
, ptl
);
1196 * If we forced a TLB flush (either due to running out of
1197 * batch buffers or because we needed to flush dirty TLB
1198 * entries before releasing the ptl), free the batched
1199 * memory too. Restart if we didn't do everything.
1203 tlb_flush_mmu_free(tlb
);
1212 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1213 struct vm_area_struct
*vma
, pud_t
*pud
,
1214 unsigned long addr
, unsigned long end
,
1215 struct zap_details
*details
)
1220 pmd
= pmd_offset(pud
, addr
);
1222 next
= pmd_addr_end(addr
, end
);
1223 if (pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)) {
1224 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1225 VM_BUG_ON_VMA(vma_is_anonymous(vma
) &&
1226 !rwsem_is_locked(&tlb
->mm
->mmap_sem
), vma
);
1227 split_huge_pmd(vma
, pmd
, addr
);
1228 } else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1233 * Here there can be other concurrent MADV_DONTNEED or
1234 * trans huge page faults running, and if the pmd is
1235 * none or trans huge it can change under us. This is
1236 * because MADV_DONTNEED holds the mmap_sem in read
1239 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1241 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1244 } while (pmd
++, addr
= next
, addr
!= end
);
1249 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1250 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1251 unsigned long addr
, unsigned long end
,
1252 struct zap_details
*details
)
1257 pud
= pud_offset(pgd
, addr
);
1259 next
= pud_addr_end(addr
, end
);
1260 if (pud_none_or_clear_bad(pud
))
1262 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1263 } while (pud
++, addr
= next
, addr
!= end
);
1268 void unmap_page_range(struct mmu_gather
*tlb
,
1269 struct vm_area_struct
*vma
,
1270 unsigned long addr
, unsigned long end
,
1271 struct zap_details
*details
)
1276 BUG_ON(addr
>= end
);
1277 tlb_start_vma(tlb
, vma
);
1278 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1280 next
= pgd_addr_end(addr
, end
);
1281 if (pgd_none_or_clear_bad(pgd
))
1283 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
1284 } while (pgd
++, addr
= next
, addr
!= end
);
1285 tlb_end_vma(tlb
, vma
);
1289 static void unmap_single_vma(struct mmu_gather
*tlb
,
1290 struct vm_area_struct
*vma
, unsigned long start_addr
,
1291 unsigned long end_addr
,
1292 struct zap_details
*details
)
1294 unsigned long start
= max(vma
->vm_start
, start_addr
);
1297 if (start
>= vma
->vm_end
)
1299 end
= min(vma
->vm_end
, end_addr
);
1300 if (end
<= vma
->vm_start
)
1304 uprobe_munmap(vma
, start
, end
);
1306 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1307 untrack_pfn(vma
, 0, 0);
1310 if (unlikely(is_vm_hugetlb_page(vma
))) {
1312 * It is undesirable to test vma->vm_file as it
1313 * should be non-null for valid hugetlb area.
1314 * However, vm_file will be NULL in the error
1315 * cleanup path of mmap_region. When
1316 * hugetlbfs ->mmap method fails,
1317 * mmap_region() nullifies vma->vm_file
1318 * before calling this function to clean up.
1319 * Since no pte has actually been setup, it is
1320 * safe to do nothing in this case.
1323 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
1324 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1325 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
1328 unmap_page_range(tlb
, vma
, start
, end
, details
);
1333 * unmap_vmas - unmap a range of memory covered by a list of vma's
1334 * @tlb: address of the caller's struct mmu_gather
1335 * @vma: the starting vma
1336 * @start_addr: virtual address at which to start unmapping
1337 * @end_addr: virtual address at which to end unmapping
1339 * Unmap all pages in the vma list.
1341 * Only addresses between `start' and `end' will be unmapped.
1343 * The VMA list must be sorted in ascending virtual address order.
1345 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1346 * range after unmap_vmas() returns. So the only responsibility here is to
1347 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1348 * drops the lock and schedules.
1350 void unmap_vmas(struct mmu_gather
*tlb
,
1351 struct vm_area_struct
*vma
, unsigned long start_addr
,
1352 unsigned long end_addr
)
1354 struct mm_struct
*mm
= vma
->vm_mm
;
1356 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1357 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1358 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1359 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1363 * zap_page_range - remove user pages in a given range
1364 * @vma: vm_area_struct holding the applicable pages
1365 * @start: starting address of pages to zap
1366 * @size: number of bytes to zap
1367 * @details: details of shared cache invalidation
1369 * Caller must protect the VMA list
1371 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1372 unsigned long size
, struct zap_details
*details
)
1374 struct mm_struct
*mm
= vma
->vm_mm
;
1375 struct mmu_gather tlb
;
1376 unsigned long end
= start
+ size
;
1379 tlb_gather_mmu(&tlb
, mm
, start
, end
);
1380 update_hiwater_rss(mm
);
1381 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1382 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
)
1383 unmap_single_vma(&tlb
, vma
, start
, end
, details
);
1384 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1385 tlb_finish_mmu(&tlb
, start
, end
);
1389 * zap_page_range_single - remove user pages in a given range
1390 * @vma: vm_area_struct holding the applicable pages
1391 * @address: starting address of pages to zap
1392 * @size: number of bytes to zap
1393 * @details: details of shared cache invalidation
1395 * The range must fit into one VMA.
1397 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1398 unsigned long size
, struct zap_details
*details
)
1400 struct mm_struct
*mm
= vma
->vm_mm
;
1401 struct mmu_gather tlb
;
1402 unsigned long end
= address
+ size
;
1405 tlb_gather_mmu(&tlb
, mm
, address
, end
);
1406 update_hiwater_rss(mm
);
1407 mmu_notifier_invalidate_range_start(mm
, address
, end
);
1408 unmap_single_vma(&tlb
, vma
, address
, end
, details
);
1409 mmu_notifier_invalidate_range_end(mm
, address
, end
);
1410 tlb_finish_mmu(&tlb
, address
, end
);
1414 * zap_vma_ptes - remove ptes mapping the vma
1415 * @vma: vm_area_struct holding ptes to be zapped
1416 * @address: starting address of pages to zap
1417 * @size: number of bytes to zap
1419 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1421 * The entire address range must be fully contained within the vma.
1423 * Returns 0 if successful.
1425 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1428 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1429 !(vma
->vm_flags
& VM_PFNMAP
))
1431 zap_page_range_single(vma
, address
, size
, NULL
);
1434 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1436 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
1439 pgd_t
* pgd
= pgd_offset(mm
, addr
);
1440 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
1442 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
1444 VM_BUG_ON(pmd_trans_huge(*pmd
));
1445 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
1452 * This is the old fallback for page remapping.
1454 * For historical reasons, it only allows reserved pages. Only
1455 * old drivers should use this, and they needed to mark their
1456 * pages reserved for the old functions anyway.
1458 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1459 struct page
*page
, pgprot_t prot
)
1461 struct mm_struct
*mm
= vma
->vm_mm
;
1470 flush_dcache_page(page
);
1471 pte
= get_locked_pte(mm
, addr
, &ptl
);
1475 if (!pte_none(*pte
))
1478 /* Ok, finally just insert the thing.. */
1480 inc_mm_counter_fast(mm
, mm_counter_file(page
));
1481 page_add_file_rmap(page
);
1482 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
1485 pte_unmap_unlock(pte
, ptl
);
1488 pte_unmap_unlock(pte
, ptl
);
1494 * vm_insert_page - insert single page into user vma
1495 * @vma: user vma to map to
1496 * @addr: target user address of this page
1497 * @page: source kernel page
1499 * This allows drivers to insert individual pages they've allocated
1502 * The page has to be a nice clean _individual_ kernel allocation.
1503 * If you allocate a compound page, you need to have marked it as
1504 * such (__GFP_COMP), or manually just split the page up yourself
1505 * (see split_page()).
1507 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1508 * took an arbitrary page protection parameter. This doesn't allow
1509 * that. Your vma protection will have to be set up correctly, which
1510 * means that if you want a shared writable mapping, you'd better
1511 * ask for a shared writable mapping!
1513 * The page does not need to be reserved.
1515 * Usually this function is called from f_op->mmap() handler
1516 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1517 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1518 * function from other places, for example from page-fault handler.
1520 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
1523 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1525 if (!page_count(page
))
1527 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
1528 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
1529 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
1530 vma
->vm_flags
|= VM_MIXEDMAP
;
1532 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1534 EXPORT_SYMBOL(vm_insert_page
);
1536 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1537 pfn_t pfn
, pgprot_t prot
)
1539 struct mm_struct
*mm
= vma
->vm_mm
;
1545 pte
= get_locked_pte(mm
, addr
, &ptl
);
1549 if (!pte_none(*pte
))
1552 /* Ok, finally just insert the thing.. */
1553 if (pfn_t_devmap(pfn
))
1554 entry
= pte_mkdevmap(pfn_t_pte(pfn
, prot
));
1556 entry
= pte_mkspecial(pfn_t_pte(pfn
, prot
));
1557 set_pte_at(mm
, addr
, pte
, entry
);
1558 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
1562 pte_unmap_unlock(pte
, ptl
);
1568 * vm_insert_pfn - insert single pfn into user vma
1569 * @vma: user vma to map to
1570 * @addr: target user address of this page
1571 * @pfn: source kernel pfn
1573 * Similar to vm_insert_page, this allows drivers to insert individual pages
1574 * they've allocated into a user vma. Same comments apply.
1576 * This function should only be called from a vm_ops->fault handler, and
1577 * in that case the handler should return NULL.
1579 * vma cannot be a COW mapping.
1581 * As this is called only for pages that do not currently exist, we
1582 * do not need to flush old virtual caches or the TLB.
1584 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
1587 return vm_insert_pfn_prot(vma
, addr
, pfn
, vma
->vm_page_prot
);
1589 EXPORT_SYMBOL(vm_insert_pfn
);
1592 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1593 * @vma: user vma to map to
1594 * @addr: target user address of this page
1595 * @pfn: source kernel pfn
1596 * @pgprot: pgprot flags for the inserted page
1598 * This is exactly like vm_insert_pfn, except that it allows drivers to
1599 * to override pgprot on a per-page basis.
1601 * This only makes sense for IO mappings, and it makes no sense for
1602 * cow mappings. In general, using multiple vmas is preferable;
1603 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1606 int vm_insert_pfn_prot(struct vm_area_struct
*vma
, unsigned long addr
,
1607 unsigned long pfn
, pgprot_t pgprot
)
1611 * Technically, architectures with pte_special can avoid all these
1612 * restrictions (same for remap_pfn_range). However we would like
1613 * consistency in testing and feature parity among all, so we should
1614 * try to keep these invariants in place for everybody.
1616 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1617 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1618 (VM_PFNMAP
|VM_MIXEDMAP
));
1619 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1620 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
1622 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1624 if (track_pfn_insert(vma
, &pgprot
, __pfn_to_pfn_t(pfn
, PFN_DEV
)))
1627 ret
= insert_pfn(vma
, addr
, __pfn_to_pfn_t(pfn
, PFN_DEV
), pgprot
);
1631 EXPORT_SYMBOL(vm_insert_pfn_prot
);
1633 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
1636 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
1638 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1642 * If we don't have pte special, then we have to use the pfn_valid()
1643 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1644 * refcount the page if pfn_valid is true (hence insert_page rather
1645 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1646 * without pte special, it would there be refcounted as a normal page.
1648 if (!HAVE_PTE_SPECIAL
&& !pfn_t_devmap(pfn
) && pfn_t_valid(pfn
)) {
1652 * At this point we are committed to insert_page()
1653 * regardless of whether the caller specified flags that
1654 * result in pfn_t_has_page() == false.
1656 page
= pfn_to_page(pfn_t_to_pfn(pfn
));
1657 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
1659 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
1661 EXPORT_SYMBOL(vm_insert_mixed
);
1664 * maps a range of physical memory into the requested pages. the old
1665 * mappings are removed. any references to nonexistent pages results
1666 * in null mappings (currently treated as "copy-on-access")
1668 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1669 unsigned long addr
, unsigned long end
,
1670 unsigned long pfn
, pgprot_t prot
)
1675 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1678 arch_enter_lazy_mmu_mode();
1680 BUG_ON(!pte_none(*pte
));
1681 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
1683 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1684 arch_leave_lazy_mmu_mode();
1685 pte_unmap_unlock(pte
- 1, ptl
);
1689 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1690 unsigned long addr
, unsigned long end
,
1691 unsigned long pfn
, pgprot_t prot
)
1696 pfn
-= addr
>> PAGE_SHIFT
;
1697 pmd
= pmd_alloc(mm
, pud
, addr
);
1700 VM_BUG_ON(pmd_trans_huge(*pmd
));
1702 next
= pmd_addr_end(addr
, end
);
1703 if (remap_pte_range(mm
, pmd
, addr
, next
,
1704 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1706 } while (pmd
++, addr
= next
, addr
!= end
);
1710 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1711 unsigned long addr
, unsigned long end
,
1712 unsigned long pfn
, pgprot_t prot
)
1717 pfn
-= addr
>> PAGE_SHIFT
;
1718 pud
= pud_alloc(mm
, pgd
, addr
);
1722 next
= pud_addr_end(addr
, end
);
1723 if (remap_pmd_range(mm
, pud
, addr
, next
,
1724 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
1726 } while (pud
++, addr
= next
, addr
!= end
);
1731 * remap_pfn_range - remap kernel memory to userspace
1732 * @vma: user vma to map to
1733 * @addr: target user address to start at
1734 * @pfn: physical address of kernel memory
1735 * @size: size of map area
1736 * @prot: page protection flags for this mapping
1738 * Note: this is only safe if the mm semaphore is held when called.
1740 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
1741 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
1745 unsigned long end
= addr
+ PAGE_ALIGN(size
);
1746 struct mm_struct
*mm
= vma
->vm_mm
;
1747 unsigned long remap_pfn
= pfn
;
1751 * Physically remapped pages are special. Tell the
1752 * rest of the world about it:
1753 * VM_IO tells people not to look at these pages
1754 * (accesses can have side effects).
1755 * VM_PFNMAP tells the core MM that the base pages are just
1756 * raw PFN mappings, and do not have a "struct page" associated
1759 * Disable vma merging and expanding with mremap().
1761 * Omit vma from core dump, even when VM_IO turned off.
1763 * There's a horrible special case to handle copy-on-write
1764 * behaviour that some programs depend on. We mark the "original"
1765 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1766 * See vm_normal_page() for details.
1768 if (is_cow_mapping(vma
->vm_flags
)) {
1769 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
1771 vma
->vm_pgoff
= pfn
;
1774 err
= track_pfn_remap(vma
, &prot
, remap_pfn
, addr
, PAGE_ALIGN(size
));
1778 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
1780 BUG_ON(addr
>= end
);
1781 pfn
-= addr
>> PAGE_SHIFT
;
1782 pgd
= pgd_offset(mm
, addr
);
1783 flush_cache_range(vma
, addr
, end
);
1785 next
= pgd_addr_end(addr
, end
);
1786 err
= remap_pud_range(mm
, pgd
, addr
, next
,
1787 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
1790 } while (pgd
++, addr
= next
, addr
!= end
);
1793 untrack_pfn(vma
, remap_pfn
, PAGE_ALIGN(size
));
1797 EXPORT_SYMBOL(remap_pfn_range
);
1800 * vm_iomap_memory - remap memory to userspace
1801 * @vma: user vma to map to
1802 * @start: start of area
1803 * @len: size of area
1805 * This is a simplified io_remap_pfn_range() for common driver use. The
1806 * driver just needs to give us the physical memory range to be mapped,
1807 * we'll figure out the rest from the vma information.
1809 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1810 * whatever write-combining details or similar.
1812 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
1814 unsigned long vm_len
, pfn
, pages
;
1816 /* Check that the physical memory area passed in looks valid */
1817 if (start
+ len
< start
)
1820 * You *really* shouldn't map things that aren't page-aligned,
1821 * but we've historically allowed it because IO memory might
1822 * just have smaller alignment.
1824 len
+= start
& ~PAGE_MASK
;
1825 pfn
= start
>> PAGE_SHIFT
;
1826 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
1827 if (pfn
+ pages
< pfn
)
1830 /* We start the mapping 'vm_pgoff' pages into the area */
1831 if (vma
->vm_pgoff
> pages
)
1833 pfn
+= vma
->vm_pgoff
;
1834 pages
-= vma
->vm_pgoff
;
1836 /* Can we fit all of the mapping? */
1837 vm_len
= vma
->vm_end
- vma
->vm_start
;
1838 if (vm_len
>> PAGE_SHIFT
> pages
)
1841 /* Ok, let it rip */
1842 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
1844 EXPORT_SYMBOL(vm_iomap_memory
);
1846 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
1847 unsigned long addr
, unsigned long end
,
1848 pte_fn_t fn
, void *data
)
1853 spinlock_t
*uninitialized_var(ptl
);
1855 pte
= (mm
== &init_mm
) ?
1856 pte_alloc_kernel(pmd
, addr
) :
1857 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
1861 BUG_ON(pmd_huge(*pmd
));
1863 arch_enter_lazy_mmu_mode();
1865 token
= pmd_pgtable(*pmd
);
1868 err
= fn(pte
++, token
, addr
, data
);
1871 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1873 arch_leave_lazy_mmu_mode();
1876 pte_unmap_unlock(pte
-1, ptl
);
1880 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
1881 unsigned long addr
, unsigned long end
,
1882 pte_fn_t fn
, void *data
)
1888 BUG_ON(pud_huge(*pud
));
1890 pmd
= pmd_alloc(mm
, pud
, addr
);
1894 next
= pmd_addr_end(addr
, end
);
1895 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
1898 } while (pmd
++, addr
= next
, addr
!= end
);
1902 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
1903 unsigned long addr
, unsigned long end
,
1904 pte_fn_t fn
, void *data
)
1910 pud
= pud_alloc(mm
, pgd
, addr
);
1914 next
= pud_addr_end(addr
, end
);
1915 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
1918 } while (pud
++, addr
= next
, addr
!= end
);
1923 * Scan a region of virtual memory, filling in page tables as necessary
1924 * and calling a provided function on each leaf page table.
1926 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
1927 unsigned long size
, pte_fn_t fn
, void *data
)
1931 unsigned long end
= addr
+ size
;
1934 if (WARN_ON(addr
>= end
))
1937 pgd
= pgd_offset(mm
, addr
);
1939 next
= pgd_addr_end(addr
, end
);
1940 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
1943 } while (pgd
++, addr
= next
, addr
!= end
);
1947 EXPORT_SYMBOL_GPL(apply_to_page_range
);
1950 * handle_pte_fault chooses page fault handler according to an entry which was
1951 * read non-atomically. Before making any commitment, on those architectures
1952 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1953 * parts, do_swap_page must check under lock before unmapping the pte and
1954 * proceeding (but do_wp_page is only called after already making such a check;
1955 * and do_anonymous_page can safely check later on).
1957 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
1958 pte_t
*page_table
, pte_t orig_pte
)
1961 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1962 if (sizeof(pte_t
) > sizeof(unsigned long)) {
1963 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
1965 same
= pte_same(*page_table
, orig_pte
);
1969 pte_unmap(page_table
);
1973 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
1975 debug_dma_assert_idle(src
);
1978 * If the source page was a PFN mapping, we don't have
1979 * a "struct page" for it. We do a best-effort copy by
1980 * just copying from the original user address. If that
1981 * fails, we just zero-fill it. Live with it.
1983 if (unlikely(!src
)) {
1984 void *kaddr
= kmap_atomic(dst
);
1985 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
1988 * This really shouldn't fail, because the page is there
1989 * in the page tables. But it might just be unreadable,
1990 * in which case we just give up and fill the result with
1993 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
1995 kunmap_atomic(kaddr
);
1996 flush_dcache_page(dst
);
1998 copy_user_highpage(dst
, src
, va
, vma
);
2001 static gfp_t
__get_fault_gfp_mask(struct vm_area_struct
*vma
)
2003 struct file
*vm_file
= vma
->vm_file
;
2006 return mapping_gfp_mask(vm_file
->f_mapping
) | __GFP_FS
| __GFP_IO
;
2009 * Special mappings (e.g. VDSO) do not have any file so fake
2010 * a default GFP_KERNEL for them.
2016 * Notify the address space that the page is about to become writable so that
2017 * it can prohibit this or wait for the page to get into an appropriate state.
2019 * We do this without the lock held, so that it can sleep if it needs to.
2021 static int do_page_mkwrite(struct vm_area_struct
*vma
, struct page
*page
,
2022 unsigned long address
)
2024 struct vm_fault vmf
;
2027 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
2028 vmf
.pgoff
= page
->index
;
2029 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2030 vmf
.gfp_mask
= __get_fault_gfp_mask(vma
);
2032 vmf
.cow_page
= NULL
;
2034 ret
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
2035 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))
2037 if (unlikely(!(ret
& VM_FAULT_LOCKED
))) {
2039 if (!page
->mapping
) {
2041 return 0; /* retry */
2043 ret
|= VM_FAULT_LOCKED
;
2045 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2050 * Handle write page faults for pages that can be reused in the current vma
2052 * This can happen either due to the mapping being with the VM_SHARED flag,
2053 * or due to us being the last reference standing to the page. In either
2054 * case, all we need to do here is to mark the page as writable and update
2055 * any related book-keeping.
2057 static inline int wp_page_reuse(struct mm_struct
*mm
,
2058 struct vm_area_struct
*vma
, unsigned long address
,
2059 pte_t
*page_table
, spinlock_t
*ptl
, pte_t orig_pte
,
2060 struct page
*page
, int page_mkwrite
,
2066 * Clear the pages cpupid information as the existing
2067 * information potentially belongs to a now completely
2068 * unrelated process.
2071 page_cpupid_xchg_last(page
, (1 << LAST_CPUPID_SHIFT
) - 1);
2073 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2074 entry
= pte_mkyoung(orig_pte
);
2075 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2076 if (ptep_set_access_flags(vma
, address
, page_table
, entry
, 1))
2077 update_mmu_cache(vma
, address
, page_table
);
2078 pte_unmap_unlock(page_table
, ptl
);
2081 struct address_space
*mapping
;
2087 dirtied
= set_page_dirty(page
);
2088 VM_BUG_ON_PAGE(PageAnon(page
), page
);
2089 mapping
= page
->mapping
;
2093 if ((dirtied
|| page_mkwrite
) && mapping
) {
2095 * Some device drivers do not set page.mapping
2096 * but still dirty their pages
2098 balance_dirty_pages_ratelimited(mapping
);
2102 file_update_time(vma
->vm_file
);
2105 return VM_FAULT_WRITE
;
2109 * Handle the case of a page which we actually need to copy to a new page.
2111 * Called with mmap_sem locked and the old page referenced, but
2112 * without the ptl held.
2114 * High level logic flow:
2116 * - Allocate a page, copy the content of the old page to the new one.
2117 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2118 * - Take the PTL. If the pte changed, bail out and release the allocated page
2119 * - If the pte is still the way we remember it, update the page table and all
2120 * relevant references. This includes dropping the reference the page-table
2121 * held to the old page, as well as updating the rmap.
2122 * - In any case, unlock the PTL and drop the reference we took to the old page.
2124 static int wp_page_copy(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2125 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2126 pte_t orig_pte
, struct page
*old_page
)
2128 struct page
*new_page
= NULL
;
2129 spinlock_t
*ptl
= NULL
;
2131 int page_copied
= 0;
2132 const unsigned long mmun_start
= address
& PAGE_MASK
; /* For mmu_notifiers */
2133 const unsigned long mmun_end
= mmun_start
+ PAGE_SIZE
; /* For mmu_notifiers */
2134 struct mem_cgroup
*memcg
;
2136 if (unlikely(anon_vma_prepare(vma
)))
2139 if (is_zero_pfn(pte_pfn(orig_pte
))) {
2140 new_page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2144 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2147 cow_user_page(new_page
, old_page
, address
, vma
);
2150 if (mem_cgroup_try_charge(new_page
, mm
, GFP_KERNEL
, &memcg
, false))
2153 __SetPageUptodate(new_page
);
2155 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2158 * Re-check the pte - we dropped the lock
2160 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2161 if (likely(pte_same(*page_table
, orig_pte
))) {
2163 if (!PageAnon(old_page
)) {
2164 dec_mm_counter_fast(mm
,
2165 mm_counter_file(old_page
));
2166 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2169 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2171 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2172 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2173 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2175 * Clear the pte entry and flush it first, before updating the
2176 * pte with the new entry. This will avoid a race condition
2177 * seen in the presence of one thread doing SMC and another
2180 ptep_clear_flush_notify(vma
, address
, page_table
);
2181 page_add_new_anon_rmap(new_page
, vma
, address
, false);
2182 mem_cgroup_commit_charge(new_page
, memcg
, false, false);
2183 lru_cache_add_active_or_unevictable(new_page
, vma
);
2185 * We call the notify macro here because, when using secondary
2186 * mmu page tables (such as kvm shadow page tables), we want the
2187 * new page to be mapped directly into the secondary page table.
2189 set_pte_at_notify(mm
, address
, page_table
, entry
);
2190 update_mmu_cache(vma
, address
, page_table
);
2193 * Only after switching the pte to the new page may
2194 * we remove the mapcount here. Otherwise another
2195 * process may come and find the rmap count decremented
2196 * before the pte is switched to the new page, and
2197 * "reuse" the old page writing into it while our pte
2198 * here still points into it and can be read by other
2201 * The critical issue is to order this
2202 * page_remove_rmap with the ptp_clear_flush above.
2203 * Those stores are ordered by (if nothing else,)
2204 * the barrier present in the atomic_add_negative
2205 * in page_remove_rmap.
2207 * Then the TLB flush in ptep_clear_flush ensures that
2208 * no process can access the old page before the
2209 * decremented mapcount is visible. And the old page
2210 * cannot be reused until after the decremented
2211 * mapcount is visible. So transitively, TLBs to
2212 * old page will be flushed before it can be reused.
2214 page_remove_rmap(old_page
, false);
2217 /* Free the old page.. */
2218 new_page
= old_page
;
2221 mem_cgroup_cancel_charge(new_page
, memcg
, false);
2227 pte_unmap_unlock(page_table
, ptl
);
2228 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2231 * Don't let another task, with possibly unlocked vma,
2232 * keep the mlocked page.
2234 if (page_copied
&& (vma
->vm_flags
& VM_LOCKED
)) {
2235 lock_page(old_page
); /* LRU manipulation */
2236 if (PageMlocked(old_page
))
2237 munlock_vma_page(old_page
);
2238 unlock_page(old_page
);
2242 return page_copied
? VM_FAULT_WRITE
: 0;
2248 return VM_FAULT_OOM
;
2252 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2255 static int wp_pfn_shared(struct mm_struct
*mm
,
2256 struct vm_area_struct
*vma
, unsigned long address
,
2257 pte_t
*page_table
, spinlock_t
*ptl
, pte_t orig_pte
,
2260 if (vma
->vm_ops
&& vma
->vm_ops
->pfn_mkwrite
) {
2261 struct vm_fault vmf
= {
2263 .pgoff
= linear_page_index(vma
, address
),
2264 .virtual_address
= (void __user
*)(address
& PAGE_MASK
),
2265 .flags
= FAULT_FLAG_WRITE
| FAULT_FLAG_MKWRITE
,
2269 pte_unmap_unlock(page_table
, ptl
);
2270 ret
= vma
->vm_ops
->pfn_mkwrite(vma
, &vmf
);
2271 if (ret
& VM_FAULT_ERROR
)
2273 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2275 * We might have raced with another page fault while we
2276 * released the pte_offset_map_lock.
2278 if (!pte_same(*page_table
, orig_pte
)) {
2279 pte_unmap_unlock(page_table
, ptl
);
2283 return wp_page_reuse(mm
, vma
, address
, page_table
, ptl
, orig_pte
,
2287 static int wp_page_shared(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2288 unsigned long address
, pte_t
*page_table
,
2289 pmd_t
*pmd
, spinlock_t
*ptl
, pte_t orig_pte
,
2290 struct page
*old_page
)
2293 int page_mkwrite
= 0;
2297 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2300 pte_unmap_unlock(page_table
, ptl
);
2301 tmp
= do_page_mkwrite(vma
, old_page
, address
);
2302 if (unlikely(!tmp
|| (tmp
&
2303 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
2308 * Since we dropped the lock we need to revalidate
2309 * the PTE as someone else may have changed it. If
2310 * they did, we just return, as we can count on the
2311 * MMU to tell us if they didn't also make it writable.
2313 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2315 if (!pte_same(*page_table
, orig_pte
)) {
2316 unlock_page(old_page
);
2317 pte_unmap_unlock(page_table
, ptl
);
2324 return wp_page_reuse(mm
, vma
, address
, page_table
, ptl
,
2325 orig_pte
, old_page
, page_mkwrite
, 1);
2329 * This routine handles present pages, when users try to write
2330 * to a shared page. It is done by copying the page to a new address
2331 * and decrementing the shared-page counter for the old page.
2333 * Note that this routine assumes that the protection checks have been
2334 * done by the caller (the low-level page fault routine in most cases).
2335 * Thus we can safely just mark it writable once we've done any necessary
2338 * We also mark the page dirty at this point even though the page will
2339 * change only once the write actually happens. This avoids a few races,
2340 * and potentially makes it more efficient.
2342 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2343 * but allow concurrent faults), with pte both mapped and locked.
2344 * We return with mmap_sem still held, but pte unmapped and unlocked.
2346 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2347 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2348 spinlock_t
*ptl
, pte_t orig_pte
)
2351 struct page
*old_page
;
2353 old_page
= vm_normal_page(vma
, address
, orig_pte
);
2356 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2359 * We should not cow pages in a shared writeable mapping.
2360 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2362 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2363 (VM_WRITE
|VM_SHARED
))
2364 return wp_pfn_shared(mm
, vma
, address
, page_table
, ptl
,
2367 pte_unmap_unlock(page_table
, ptl
);
2368 return wp_page_copy(mm
, vma
, address
, page_table
, pmd
,
2369 orig_pte
, old_page
);
2373 * Take out anonymous pages first, anonymous shared vmas are
2374 * not dirty accountable.
2376 if (PageAnon(old_page
) && !PageKsm(old_page
)) {
2378 if (!trylock_page(old_page
)) {
2380 pte_unmap_unlock(page_table
, ptl
);
2381 lock_page(old_page
);
2382 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2384 if (!pte_same(*page_table
, orig_pte
)) {
2385 unlock_page(old_page
);
2386 pte_unmap_unlock(page_table
, ptl
);
2392 if (reuse_swap_page(old_page
, &total_mapcount
)) {
2393 if (total_mapcount
== 1) {
2395 * The page is all ours. Move it to
2396 * our anon_vma so the rmap code will
2397 * not search our parent or siblings.
2398 * Protected against the rmap code by
2401 page_move_anon_rmap(compound_head(old_page
),
2404 unlock_page(old_page
);
2405 return wp_page_reuse(mm
, vma
, address
, page_table
, ptl
,
2406 orig_pte
, old_page
, 0, 0);
2408 unlock_page(old_page
);
2409 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2410 (VM_WRITE
|VM_SHARED
))) {
2411 return wp_page_shared(mm
, vma
, address
, page_table
, pmd
,
2412 ptl
, orig_pte
, old_page
);
2416 * Ok, we need to copy. Oh, well..
2420 pte_unmap_unlock(page_table
, ptl
);
2421 return wp_page_copy(mm
, vma
, address
, page_table
, pmd
,
2422 orig_pte
, old_page
);
2425 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
2426 unsigned long start_addr
, unsigned long end_addr
,
2427 struct zap_details
*details
)
2429 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
2432 static inline void unmap_mapping_range_tree(struct rb_root
*root
,
2433 struct zap_details
*details
)
2435 struct vm_area_struct
*vma
;
2436 pgoff_t vba
, vea
, zba
, zea
;
2438 vma_interval_tree_foreach(vma
, root
,
2439 details
->first_index
, details
->last_index
) {
2441 vba
= vma
->vm_pgoff
;
2442 vea
= vba
+ vma_pages(vma
) - 1;
2443 zba
= details
->first_index
;
2446 zea
= details
->last_index
;
2450 unmap_mapping_range_vma(vma
,
2451 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
2452 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
2458 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2459 * address_space corresponding to the specified page range in the underlying
2462 * @mapping: the address space containing mmaps to be unmapped.
2463 * @holebegin: byte in first page to unmap, relative to the start of
2464 * the underlying file. This will be rounded down to a PAGE_SIZE
2465 * boundary. Note that this is different from truncate_pagecache(), which
2466 * must keep the partial page. In contrast, we must get rid of
2468 * @holelen: size of prospective hole in bytes. This will be rounded
2469 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2471 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2472 * but 0 when invalidating pagecache, don't throw away private data.
2474 void unmap_mapping_range(struct address_space
*mapping
,
2475 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
2477 struct zap_details details
= { };
2478 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
2479 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2481 /* Check for overflow. */
2482 if (sizeof(holelen
) > sizeof(hlen
)) {
2484 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2485 if (holeend
& ~(long long)ULONG_MAX
)
2486 hlen
= ULONG_MAX
- hba
+ 1;
2489 details
.check_mapping
= even_cows
? NULL
: mapping
;
2490 details
.first_index
= hba
;
2491 details
.last_index
= hba
+ hlen
- 1;
2492 if (details
.last_index
< details
.first_index
)
2493 details
.last_index
= ULONG_MAX
;
2496 /* DAX uses i_mmap_lock to serialise file truncate vs page fault */
2497 i_mmap_lock_write(mapping
);
2498 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
)))
2499 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
2500 i_mmap_unlock_write(mapping
);
2502 EXPORT_SYMBOL(unmap_mapping_range
);
2505 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2506 * but allow concurrent faults), and pte mapped but not yet locked.
2507 * We return with pte unmapped and unlocked.
2509 * We return with the mmap_sem locked or unlocked in the same cases
2510 * as does filemap_fault().
2512 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2513 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2514 unsigned int flags
, pte_t orig_pte
)
2517 struct page
*page
, *swapcache
;
2518 struct mem_cgroup
*memcg
;
2525 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
2528 entry
= pte_to_swp_entry(orig_pte
);
2529 if (unlikely(non_swap_entry(entry
))) {
2530 if (is_migration_entry(entry
)) {
2531 migration_entry_wait(mm
, pmd
, address
);
2532 } else if (is_hwpoison_entry(entry
)) {
2533 ret
= VM_FAULT_HWPOISON
;
2535 print_bad_pte(vma
, address
, orig_pte
, NULL
);
2536 ret
= VM_FAULT_SIGBUS
;
2540 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
2541 page
= lookup_swap_cache(entry
);
2543 page
= swapin_readahead(entry
,
2544 GFP_HIGHUSER_MOVABLE
, vma
, address
);
2547 * Back out if somebody else faulted in this pte
2548 * while we released the pte lock.
2550 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2551 if (likely(pte_same(*page_table
, orig_pte
)))
2553 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2557 /* Had to read the page from swap area: Major fault */
2558 ret
= VM_FAULT_MAJOR
;
2559 count_vm_event(PGMAJFAULT
);
2560 mem_cgroup_count_vm_event(mm
, PGMAJFAULT
);
2561 } else if (PageHWPoison(page
)) {
2563 * hwpoisoned dirty swapcache pages are kept for killing
2564 * owner processes (which may be unknown at hwpoison time)
2566 ret
= VM_FAULT_HWPOISON
;
2567 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2573 locked
= lock_page_or_retry(page
, mm
, flags
);
2575 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
2577 ret
|= VM_FAULT_RETRY
;
2582 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2583 * release the swapcache from under us. The page pin, and pte_same
2584 * test below, are not enough to exclude that. Even if it is still
2585 * swapcache, we need to check that the page's swap has not changed.
2587 if (unlikely(!PageSwapCache(page
) || page_private(page
) != entry
.val
))
2590 page
= ksm_might_need_to_copy(page
, vma
, address
);
2591 if (unlikely(!page
)) {
2597 if (mem_cgroup_try_charge(page
, mm
, GFP_KERNEL
, &memcg
, false)) {
2603 * Back out if somebody else already faulted in this pte.
2605 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2606 if (unlikely(!pte_same(*page_table
, orig_pte
)))
2609 if (unlikely(!PageUptodate(page
))) {
2610 ret
= VM_FAULT_SIGBUS
;
2615 * The page isn't present yet, go ahead with the fault.
2617 * Be careful about the sequence of operations here.
2618 * To get its accounting right, reuse_swap_page() must be called
2619 * while the page is counted on swap but not yet in mapcount i.e.
2620 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2621 * must be called after the swap_free(), or it will never succeed.
2624 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2625 dec_mm_counter_fast(mm
, MM_SWAPENTS
);
2626 pte
= mk_pte(page
, vma
->vm_page_prot
);
2627 if ((flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
, NULL
)) {
2628 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
2629 flags
&= ~FAULT_FLAG_WRITE
;
2630 ret
|= VM_FAULT_WRITE
;
2631 exclusive
= RMAP_EXCLUSIVE
;
2633 flush_icache_page(vma
, page
);
2634 if (pte_swp_soft_dirty(orig_pte
))
2635 pte
= pte_mksoft_dirty(pte
);
2636 set_pte_at(mm
, address
, page_table
, pte
);
2637 if (page
== swapcache
) {
2638 do_page_add_anon_rmap(page
, vma
, address
, exclusive
);
2639 mem_cgroup_commit_charge(page
, memcg
, true, false);
2640 } else { /* ksm created a completely new copy */
2641 page_add_new_anon_rmap(page
, vma
, address
, false);
2642 mem_cgroup_commit_charge(page
, memcg
, false, false);
2643 lru_cache_add_active_or_unevictable(page
, vma
);
2647 if (mem_cgroup_swap_full(page
) ||
2648 (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
2649 try_to_free_swap(page
);
2651 if (page
!= swapcache
) {
2653 * Hold the lock to avoid the swap entry to be reused
2654 * until we take the PT lock for the pte_same() check
2655 * (to avoid false positives from pte_same). For
2656 * further safety release the lock after the swap_free
2657 * so that the swap count won't change under a
2658 * parallel locked swapcache.
2660 unlock_page(swapcache
);
2661 put_page(swapcache
);
2664 if (flags
& FAULT_FLAG_WRITE
) {
2665 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
2666 if (ret
& VM_FAULT_ERROR
)
2667 ret
&= VM_FAULT_ERROR
;
2671 /* No need to invalidate - it was non-present before */
2672 update_mmu_cache(vma
, address
, page_table
);
2674 pte_unmap_unlock(page_table
, ptl
);
2678 mem_cgroup_cancel_charge(page
, memcg
, false);
2679 pte_unmap_unlock(page_table
, ptl
);
2684 if (page
!= swapcache
) {
2685 unlock_page(swapcache
);
2686 put_page(swapcache
);
2692 * This is like a special single-page "expand_{down|up}wards()",
2693 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2694 * doesn't hit another vma.
2696 static inline int check_stack_guard_page(struct vm_area_struct
*vma
, unsigned long address
)
2698 address
&= PAGE_MASK
;
2699 if ((vma
->vm_flags
& VM_GROWSDOWN
) && address
== vma
->vm_start
) {
2700 struct vm_area_struct
*prev
= vma
->vm_prev
;
2703 * Is there a mapping abutting this one below?
2705 * That's only ok if it's the same stack mapping
2706 * that has gotten split..
2708 if (prev
&& prev
->vm_end
== address
)
2709 return prev
->vm_flags
& VM_GROWSDOWN
? 0 : -ENOMEM
;
2711 return expand_downwards(vma
, address
- PAGE_SIZE
);
2713 if ((vma
->vm_flags
& VM_GROWSUP
) && address
+ PAGE_SIZE
== vma
->vm_end
) {
2714 struct vm_area_struct
*next
= vma
->vm_next
;
2716 /* As VM_GROWSDOWN but s/below/above/ */
2717 if (next
&& next
->vm_start
== address
+ PAGE_SIZE
)
2718 return next
->vm_flags
& VM_GROWSUP
? 0 : -ENOMEM
;
2720 return expand_upwards(vma
, address
+ PAGE_SIZE
);
2726 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2727 * but allow concurrent faults), and pte mapped but not yet locked.
2728 * We return with mmap_sem still held, but pte unmapped and unlocked.
2730 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2731 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2734 struct mem_cgroup
*memcg
;
2739 pte_unmap(page_table
);
2741 /* File mapping without ->vm_ops ? */
2742 if (vma
->vm_flags
& VM_SHARED
)
2743 return VM_FAULT_SIGBUS
;
2745 /* Check if we need to add a guard page to the stack */
2746 if (check_stack_guard_page(vma
, address
) < 0)
2747 return VM_FAULT_SIGSEGV
;
2749 /* Use the zero-page for reads */
2750 if (!(flags
& FAULT_FLAG_WRITE
) && !mm_forbids_zeropage(mm
)) {
2751 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(address
),
2752 vma
->vm_page_prot
));
2753 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2754 if (!pte_none(*page_table
))
2756 /* Deliver the page fault to userland, check inside PT lock */
2757 if (userfaultfd_missing(vma
)) {
2758 pte_unmap_unlock(page_table
, ptl
);
2759 return handle_userfault(vma
, address
, flags
,
2765 /* Allocate our own private page. */
2766 if (unlikely(anon_vma_prepare(vma
)))
2768 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2772 if (mem_cgroup_try_charge(page
, mm
, GFP_KERNEL
, &memcg
, false))
2776 * The memory barrier inside __SetPageUptodate makes sure that
2777 * preceeding stores to the page contents become visible before
2778 * the set_pte_at() write.
2780 __SetPageUptodate(page
);
2782 entry
= mk_pte(page
, vma
->vm_page_prot
);
2783 if (vma
->vm_flags
& VM_WRITE
)
2784 entry
= pte_mkwrite(pte_mkdirty(entry
));
2786 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2787 if (!pte_none(*page_table
))
2790 /* Deliver the page fault to userland, check inside PT lock */
2791 if (userfaultfd_missing(vma
)) {
2792 pte_unmap_unlock(page_table
, ptl
);
2793 mem_cgroup_cancel_charge(page
, memcg
, false);
2795 return handle_userfault(vma
, address
, flags
,
2799 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2800 page_add_new_anon_rmap(page
, vma
, address
, false);
2801 mem_cgroup_commit_charge(page
, memcg
, false, false);
2802 lru_cache_add_active_or_unevictable(page
, vma
);
2804 set_pte_at(mm
, address
, page_table
, entry
);
2806 /* No need to invalidate - it was non-present before */
2807 update_mmu_cache(vma
, address
, page_table
);
2809 pte_unmap_unlock(page_table
, ptl
);
2812 mem_cgroup_cancel_charge(page
, memcg
, false);
2818 return VM_FAULT_OOM
;
2822 * The mmap_sem must have been held on entry, and may have been
2823 * released depending on flags and vma->vm_ops->fault() return value.
2824 * See filemap_fault() and __lock_page_retry().
2826 static int __do_fault(struct vm_area_struct
*vma
, unsigned long address
,
2827 pgoff_t pgoff
, unsigned int flags
,
2828 struct page
*cow_page
, struct page
**page
)
2830 struct vm_fault vmf
;
2833 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
2837 vmf
.gfp_mask
= __get_fault_gfp_mask(vma
);
2838 vmf
.cow_page
= cow_page
;
2840 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
2841 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
2846 if (unlikely(PageHWPoison(vmf
.page
))) {
2847 if (ret
& VM_FAULT_LOCKED
)
2848 unlock_page(vmf
.page
);
2850 return VM_FAULT_HWPOISON
;
2853 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
2854 lock_page(vmf
.page
);
2856 VM_BUG_ON_PAGE(!PageLocked(vmf
.page
), vmf
.page
);
2864 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2866 * @vma: virtual memory area
2867 * @address: user virtual address
2868 * @page: page to map
2869 * @pte: pointer to target page table entry
2870 * @write: true, if new entry is writable
2871 * @anon: true, if it's anonymous page
2873 * Caller must hold page table lock relevant for @pte.
2875 * Target users are page handler itself and implementations of
2876 * vm_ops->map_pages.
2878 void do_set_pte(struct vm_area_struct
*vma
, unsigned long address
,
2879 struct page
*page
, pte_t
*pte
, bool write
, bool anon
, bool old
)
2883 flush_icache_page(vma
, page
);
2884 entry
= mk_pte(page
, vma
->vm_page_prot
);
2886 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2888 entry
= pte_mkold(entry
);
2890 inc_mm_counter_fast(vma
->vm_mm
, MM_ANONPAGES
);
2891 page_add_new_anon_rmap(page
, vma
, address
, false);
2893 inc_mm_counter_fast(vma
->vm_mm
, mm_counter_file(page
));
2894 page_add_file_rmap(page
);
2896 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
2898 /* no need to invalidate: a not-present page won't be cached */
2899 update_mmu_cache(vma
, address
, pte
);
2903 * If architecture emulates "accessed" or "young" bit without HW support,
2904 * there is no much gain with fault_around.
2906 static unsigned long fault_around_bytes __read_mostly
=
2907 #ifndef __HAVE_ARCH_PTEP_SET_ACCESS_FLAGS
2910 rounddown_pow_of_two(65536);
2913 #ifdef CONFIG_DEBUG_FS
2914 static int fault_around_bytes_get(void *data
, u64
*val
)
2916 *val
= fault_around_bytes
;
2921 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2922 * rounded down to nearest page order. It's what do_fault_around() expects to
2925 static int fault_around_bytes_set(void *data
, u64 val
)
2927 if (val
/ PAGE_SIZE
> PTRS_PER_PTE
)
2929 if (val
> PAGE_SIZE
)
2930 fault_around_bytes
= rounddown_pow_of_two(val
);
2932 fault_around_bytes
= PAGE_SIZE
; /* rounddown_pow_of_two(0) is undefined */
2935 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops
,
2936 fault_around_bytes_get
, fault_around_bytes_set
, "%llu\n");
2938 static int __init
fault_around_debugfs(void)
2942 ret
= debugfs_create_file("fault_around_bytes", 0644, NULL
, NULL
,
2943 &fault_around_bytes_fops
);
2945 pr_warn("Failed to create fault_around_bytes in debugfs");
2948 late_initcall(fault_around_debugfs
);
2952 * do_fault_around() tries to map few pages around the fault address. The hope
2953 * is that the pages will be needed soon and this will lower the number of
2956 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2957 * not ready to be mapped: not up-to-date, locked, etc.
2959 * This function is called with the page table lock taken. In the split ptlock
2960 * case the page table lock only protects only those entries which belong to
2961 * the page table corresponding to the fault address.
2963 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2966 * fault_around_pages() defines how many pages we'll try to map.
2967 * do_fault_around() expects it to return a power of two less than or equal to
2970 * The virtual address of the area that we map is naturally aligned to the
2971 * fault_around_pages() value (and therefore to page order). This way it's
2972 * easier to guarantee that we don't cross page table boundaries.
2974 static void do_fault_around(struct vm_area_struct
*vma
, unsigned long address
,
2975 pte_t
*pte
, pgoff_t pgoff
, unsigned int flags
)
2977 unsigned long start_addr
, nr_pages
, mask
;
2979 struct vm_fault vmf
;
2982 nr_pages
= READ_ONCE(fault_around_bytes
) >> PAGE_SHIFT
;
2983 mask
= ~(nr_pages
* PAGE_SIZE
- 1) & PAGE_MASK
;
2985 start_addr
= max(address
& mask
, vma
->vm_start
);
2986 off
= ((address
- start_addr
) >> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1);
2991 * max_pgoff is either end of page table or end of vma
2992 * or fault_around_pages() from pgoff, depending what is nearest.
2994 max_pgoff
= pgoff
- ((start_addr
>> PAGE_SHIFT
) & (PTRS_PER_PTE
- 1)) +
2996 max_pgoff
= min3(max_pgoff
, vma_pages(vma
) + vma
->vm_pgoff
- 1,
2997 pgoff
+ nr_pages
- 1);
2999 /* Check if it makes any sense to call ->map_pages */
3000 while (!pte_none(*pte
)) {
3001 if (++pgoff
> max_pgoff
)
3003 start_addr
+= PAGE_SIZE
;
3004 if (start_addr
>= vma
->vm_end
)
3009 vmf
.virtual_address
= (void __user
*) start_addr
;
3012 vmf
.max_pgoff
= max_pgoff
;
3014 vmf
.gfp_mask
= __get_fault_gfp_mask(vma
);
3015 vma
->vm_ops
->map_pages(vma
, &vmf
);
3018 static int do_read_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3019 unsigned long address
, pmd_t
*pmd
,
3020 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3022 struct page
*fault_page
;
3028 * Let's call ->map_pages() first and use ->fault() as fallback
3029 * if page by the offset is not ready to be mapped (cold cache or
3032 if (vma
->vm_ops
->map_pages
&& fault_around_bytes
>> PAGE_SHIFT
> 1) {
3033 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3034 if (!pte_same(*pte
, orig_pte
))
3036 do_fault_around(vma
, address
, pte
, pgoff
, flags
);
3037 /* Check if the fault is handled by faultaround */
3038 if (!pte_same(*pte
, orig_pte
)) {
3040 * Faultaround produce old pte, but the pte we've
3041 * handler fault for should be young.
3043 pte_t entry
= pte_mkyoung(*pte
);
3044 if (ptep_set_access_flags(vma
, address
, pte
, entry
, 0))
3045 update_mmu_cache(vma
, address
, pte
);
3048 pte_unmap_unlock(pte
, ptl
);
3051 ret
= __do_fault(vma
, address
, pgoff
, flags
, NULL
, &fault_page
);
3052 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3055 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3056 if (unlikely(!pte_same(*pte
, orig_pte
))) {
3057 pte_unmap_unlock(pte
, ptl
);
3058 unlock_page(fault_page
);
3059 put_page(fault_page
);
3062 do_set_pte(vma
, address
, fault_page
, pte
, false, false, false);
3063 unlock_page(fault_page
);
3065 pte_unmap_unlock(pte
, ptl
);
3069 static int do_cow_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3070 unsigned long address
, pmd_t
*pmd
,
3071 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3073 struct page
*fault_page
, *new_page
;
3074 struct mem_cgroup
*memcg
;
3079 if (unlikely(anon_vma_prepare(vma
)))
3080 return VM_FAULT_OOM
;
3082 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
3084 return VM_FAULT_OOM
;
3086 if (mem_cgroup_try_charge(new_page
, mm
, GFP_KERNEL
, &memcg
, false)) {
3088 return VM_FAULT_OOM
;
3091 ret
= __do_fault(vma
, address
, pgoff
, flags
, new_page
, &fault_page
);
3092 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3096 copy_user_highpage(new_page
, fault_page
, address
, vma
);
3097 __SetPageUptodate(new_page
);
3099 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3100 if (unlikely(!pte_same(*pte
, orig_pte
))) {
3101 pte_unmap_unlock(pte
, ptl
);
3103 unlock_page(fault_page
);
3104 put_page(fault_page
);
3107 * The fault handler has no page to lock, so it holds
3108 * i_mmap_lock for read to protect against truncate.
3110 i_mmap_unlock_read(vma
->vm_file
->f_mapping
);
3114 do_set_pte(vma
, address
, new_page
, pte
, true, true, false);
3115 mem_cgroup_commit_charge(new_page
, memcg
, false, false);
3116 lru_cache_add_active_or_unevictable(new_page
, vma
);
3117 pte_unmap_unlock(pte
, ptl
);
3119 unlock_page(fault_page
);
3120 put_page(fault_page
);
3123 * The fault handler has no page to lock, so it holds
3124 * i_mmap_lock for read to protect against truncate.
3126 i_mmap_unlock_read(vma
->vm_file
->f_mapping
);
3130 mem_cgroup_cancel_charge(new_page
, memcg
, false);
3135 static int do_shared_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3136 unsigned long address
, pmd_t
*pmd
,
3137 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3139 struct page
*fault_page
;
3140 struct address_space
*mapping
;
3146 ret
= __do_fault(vma
, address
, pgoff
, flags
, NULL
, &fault_page
);
3147 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
| VM_FAULT_RETRY
)))
3151 * Check if the backing address space wants to know that the page is
3152 * about to become writable
3154 if (vma
->vm_ops
->page_mkwrite
) {
3155 unlock_page(fault_page
);
3156 tmp
= do_page_mkwrite(vma
, fault_page
, address
);
3157 if (unlikely(!tmp
||
3158 (tmp
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
)))) {
3159 put_page(fault_page
);
3164 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3165 if (unlikely(!pte_same(*pte
, orig_pte
))) {
3166 pte_unmap_unlock(pte
, ptl
);
3167 unlock_page(fault_page
);
3168 put_page(fault_page
);
3171 do_set_pte(vma
, address
, fault_page
, pte
, true, false, false);
3172 pte_unmap_unlock(pte
, ptl
);
3174 if (set_page_dirty(fault_page
))
3177 * Take a local copy of the address_space - page.mapping may be zeroed
3178 * by truncate after unlock_page(). The address_space itself remains
3179 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3180 * release semantics to prevent the compiler from undoing this copying.
3182 mapping
= page_rmapping(fault_page
);
3183 unlock_page(fault_page
);
3184 if ((dirtied
|| vma
->vm_ops
->page_mkwrite
) && mapping
) {
3186 * Some device drivers do not set page.mapping but still
3189 balance_dirty_pages_ratelimited(mapping
);
3192 if (!vma
->vm_ops
->page_mkwrite
)
3193 file_update_time(vma
->vm_file
);
3199 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3200 * but allow concurrent faults).
3201 * The mmap_sem may have been released depending on flags and our
3202 * return value. See filemap_fault() and __lock_page_or_retry().
3204 static int do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3205 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3206 unsigned int flags
, pte_t orig_pte
)
3208 pgoff_t pgoff
= linear_page_index(vma
, address
);
3210 pte_unmap(page_table
);
3211 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3212 if (!vma
->vm_ops
->fault
)
3213 return VM_FAULT_SIGBUS
;
3214 if (!(flags
& FAULT_FLAG_WRITE
))
3215 return do_read_fault(mm
, vma
, address
, pmd
, pgoff
, flags
,
3217 if (!(vma
->vm_flags
& VM_SHARED
))
3218 return do_cow_fault(mm
, vma
, address
, pmd
, pgoff
, flags
,
3220 return do_shared_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3223 static int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3224 unsigned long addr
, int page_nid
,
3229 count_vm_numa_event(NUMA_HINT_FAULTS
);
3230 if (page_nid
== numa_node_id()) {
3231 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3232 *flags
|= TNF_FAULT_LOCAL
;
3235 return mpol_misplaced(page
, vma
, addr
);
3238 static int do_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3239 unsigned long addr
, pte_t pte
, pte_t
*ptep
, pmd_t
*pmd
)
3241 struct page
*page
= NULL
;
3246 bool migrated
= false;
3247 bool was_writable
= pte_write(pte
);
3250 /* A PROT_NONE fault should not end up here */
3251 BUG_ON(!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)));
3254 * The "pte" at this point cannot be used safely without
3255 * validation through pte_unmap_same(). It's of NUMA type but
3256 * the pfn may be screwed if the read is non atomic.
3258 * We can safely just do a "set_pte_at()", because the old
3259 * page table entry is not accessible, so there would be no
3260 * concurrent hardware modifications to the PTE.
3262 ptl
= pte_lockptr(mm
, pmd
);
3264 if (unlikely(!pte_same(*ptep
, pte
))) {
3265 pte_unmap_unlock(ptep
, ptl
);
3269 /* Make it present again */
3270 pte
= pte_modify(pte
, vma
->vm_page_prot
);
3271 pte
= pte_mkyoung(pte
);
3273 pte
= pte_mkwrite(pte
);
3274 set_pte_at(mm
, addr
, ptep
, pte
);
3275 update_mmu_cache(vma
, addr
, ptep
);
3277 page
= vm_normal_page(vma
, addr
, pte
);
3279 pte_unmap_unlock(ptep
, ptl
);
3283 /* TODO: handle PTE-mapped THP */
3284 if (PageCompound(page
)) {
3285 pte_unmap_unlock(ptep
, ptl
);
3290 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3291 * much anyway since they can be in shared cache state. This misses
3292 * the case where a mapping is writable but the process never writes
3293 * to it but pte_write gets cleared during protection updates and
3294 * pte_dirty has unpredictable behaviour between PTE scan updates,
3295 * background writeback, dirty balancing and application behaviour.
3297 if (!(vma
->vm_flags
& VM_WRITE
))
3298 flags
|= TNF_NO_GROUP
;
3301 * Flag if the page is shared between multiple address spaces. This
3302 * is later used when determining whether to group tasks together
3304 if (page_mapcount(page
) > 1 && (vma
->vm_flags
& VM_SHARED
))
3305 flags
|= TNF_SHARED
;
3307 last_cpupid
= page_cpupid_last(page
);
3308 page_nid
= page_to_nid(page
);
3309 target_nid
= numa_migrate_prep(page
, vma
, addr
, page_nid
, &flags
);
3310 pte_unmap_unlock(ptep
, ptl
);
3311 if (target_nid
== -1) {
3316 /* Migrate to the requested node */
3317 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
3319 page_nid
= target_nid
;
3320 flags
|= TNF_MIGRATED
;
3322 flags
|= TNF_MIGRATE_FAIL
;
3326 task_numa_fault(last_cpupid
, page_nid
, 1, flags
);
3330 static int create_huge_pmd(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3331 unsigned long address
, pmd_t
*pmd
, unsigned int flags
)
3333 if (vma_is_anonymous(vma
))
3334 return do_huge_pmd_anonymous_page(mm
, vma
, address
, pmd
, flags
);
3335 if (vma
->vm_ops
->pmd_fault
)
3336 return vma
->vm_ops
->pmd_fault(vma
, address
, pmd
, flags
);
3337 return VM_FAULT_FALLBACK
;
3340 static int wp_huge_pmd(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3341 unsigned long address
, pmd_t
*pmd
, pmd_t orig_pmd
,
3344 if (vma_is_anonymous(vma
))
3345 return do_huge_pmd_wp_page(mm
, vma
, address
, pmd
, orig_pmd
);
3346 if (vma
->vm_ops
->pmd_fault
)
3347 return vma
->vm_ops
->pmd_fault(vma
, address
, pmd
, flags
);
3348 return VM_FAULT_FALLBACK
;
3352 * These routines also need to handle stuff like marking pages dirty
3353 * and/or accessed for architectures that don't do it in hardware (most
3354 * RISC architectures). The early dirtying is also good on the i386.
3356 * There is also a hook called "update_mmu_cache()" that architectures
3357 * with external mmu caches can use to update those (ie the Sparc or
3358 * PowerPC hashed page tables that act as extended TLBs).
3360 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3361 * but allow concurrent faults), and pte mapped but not yet locked.
3362 * We return with pte unmapped and unlocked.
3364 * The mmap_sem may have been released depending on flags and our
3365 * return value. See filemap_fault() and __lock_page_or_retry().
3367 static int handle_pte_fault(struct mm_struct
*mm
,
3368 struct vm_area_struct
*vma
, unsigned long address
,
3369 pte_t
*pte
, pmd_t
*pmd
, unsigned int flags
)
3375 * some architectures can have larger ptes than wordsize,
3376 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3377 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3378 * The code below just needs a consistent view for the ifs and
3379 * we later double check anyway with the ptl lock held. So here
3380 * a barrier will do.
3384 if (!pte_present(entry
)) {
3385 if (pte_none(entry
)) {
3386 if (vma_is_anonymous(vma
))
3387 return do_anonymous_page(mm
, vma
, address
,
3390 return do_fault(mm
, vma
, address
, pte
, pmd
,
3393 return do_swap_page(mm
, vma
, address
,
3394 pte
, pmd
, flags
, entry
);
3397 if (pte_protnone(entry
))
3398 return do_numa_page(mm
, vma
, address
, entry
, pte
, pmd
);
3400 ptl
= pte_lockptr(mm
, pmd
);
3402 if (unlikely(!pte_same(*pte
, entry
)))
3404 if (flags
& FAULT_FLAG_WRITE
) {
3405 if (!pte_write(entry
))
3406 return do_wp_page(mm
, vma
, address
,
3407 pte
, pmd
, ptl
, entry
);
3408 entry
= pte_mkdirty(entry
);
3410 entry
= pte_mkyoung(entry
);
3411 if (ptep_set_access_flags(vma
, address
, pte
, entry
, flags
& FAULT_FLAG_WRITE
)) {
3412 update_mmu_cache(vma
, address
, pte
);
3415 * This is needed only for protection faults but the arch code
3416 * is not yet telling us if this is a protection fault or not.
3417 * This still avoids useless tlb flushes for .text page faults
3420 if (flags
& FAULT_FLAG_WRITE
)
3421 flush_tlb_fix_spurious_fault(vma
, address
);
3424 pte_unmap_unlock(pte
, ptl
);
3429 * By the time we get here, we already hold the mm semaphore
3431 * The mmap_sem may have been released depending on flags and our
3432 * return value. See filemap_fault() and __lock_page_or_retry().
3434 static int __handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3435 unsigned long address
, unsigned int flags
)
3442 if (!arch_vma_access_permitted(vma
, flags
& FAULT_FLAG_WRITE
,
3443 flags
& FAULT_FLAG_INSTRUCTION
,
3444 flags
& FAULT_FLAG_REMOTE
))
3445 return VM_FAULT_SIGSEGV
;
3447 if (unlikely(is_vm_hugetlb_page(vma
)))
3448 return hugetlb_fault(mm
, vma
, address
, flags
);
3450 pgd
= pgd_offset(mm
, address
);
3451 pud
= pud_alloc(mm
, pgd
, address
);
3453 return VM_FAULT_OOM
;
3454 pmd
= pmd_alloc(mm
, pud
, address
);
3456 return VM_FAULT_OOM
;
3457 if (pmd_none(*pmd
) && transparent_hugepage_enabled(vma
)) {
3458 int ret
= create_huge_pmd(mm
, vma
, address
, pmd
, flags
);
3459 if (!(ret
& VM_FAULT_FALLBACK
))
3462 pmd_t orig_pmd
= *pmd
;
3466 if (pmd_trans_huge(orig_pmd
) || pmd_devmap(orig_pmd
)) {
3467 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
3469 if (pmd_protnone(orig_pmd
))
3470 return do_huge_pmd_numa_page(mm
, vma
, address
,
3473 if (dirty
&& !pmd_write(orig_pmd
)) {
3474 ret
= wp_huge_pmd(mm
, vma
, address
, pmd
,
3476 if (!(ret
& VM_FAULT_FALLBACK
))
3479 huge_pmd_set_accessed(mm
, vma
, address
, pmd
,
3487 * Use pte_alloc() instead of pte_alloc_map, because we can't
3488 * run pte_offset_map on the pmd, if an huge pmd could
3489 * materialize from under us from a different thread.
3491 if (unlikely(pte_alloc(mm
, pmd
, address
)))
3492 return VM_FAULT_OOM
;
3494 * If a huge pmd materialized under us just retry later. Use
3495 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
3496 * didn't become pmd_trans_huge under us and then back to pmd_none, as
3497 * a result of MADV_DONTNEED running immediately after a huge pmd fault
3498 * in a different thread of this mm, in turn leading to a misleading
3499 * pmd_trans_huge() retval. All we have to ensure is that it is a
3500 * regular pmd that we can walk with pte_offset_map() and we can do that
3501 * through an atomic read in C, which is what pmd_trans_unstable()
3504 if (unlikely(pmd_trans_unstable(pmd
) || pmd_devmap(*pmd
)))
3507 * A regular pmd is established and it can't morph into a huge pmd
3508 * from under us anymore at this point because we hold the mmap_sem
3509 * read mode and khugepaged takes it in write mode. So now it's
3510 * safe to run pte_offset_map().
3512 pte
= pte_offset_map(pmd
, address
);
3514 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
3518 * By the time we get here, we already hold the mm semaphore
3520 * The mmap_sem may have been released depending on flags and our
3521 * return value. See filemap_fault() and __lock_page_or_retry().
3523 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3524 unsigned long address
, unsigned int flags
)
3528 __set_current_state(TASK_RUNNING
);
3530 count_vm_event(PGFAULT
);
3531 mem_cgroup_count_vm_event(mm
, PGFAULT
);
3533 /* do counter updates before entering really critical section. */
3534 check_sync_rss_stat(current
);
3537 * Enable the memcg OOM handling for faults triggered in user
3538 * space. Kernel faults are handled more gracefully.
3540 if (flags
& FAULT_FLAG_USER
)
3541 mem_cgroup_oom_enable();
3543 ret
= __handle_mm_fault(mm
, vma
, address
, flags
);
3545 if (flags
& FAULT_FLAG_USER
) {
3546 mem_cgroup_oom_disable();
3548 * The task may have entered a memcg OOM situation but
3549 * if the allocation error was handled gracefully (no
3550 * VM_FAULT_OOM), there is no need to kill anything.
3551 * Just clean up the OOM state peacefully.
3553 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
3554 mem_cgroup_oom_synchronize(false);
3559 EXPORT_SYMBOL_GPL(handle_mm_fault
);
3561 #ifndef __PAGETABLE_PUD_FOLDED
3563 * Allocate page upper directory.
3564 * We've already handled the fast-path in-line.
3566 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
3568 pud_t
*new = pud_alloc_one(mm
, address
);
3572 smp_wmb(); /* See comment in __pte_alloc */
3574 spin_lock(&mm
->page_table_lock
);
3575 if (pgd_present(*pgd
)) /* Another has populated it */
3578 pgd_populate(mm
, pgd
, new);
3579 spin_unlock(&mm
->page_table_lock
);
3582 #endif /* __PAGETABLE_PUD_FOLDED */
3584 #ifndef __PAGETABLE_PMD_FOLDED
3586 * Allocate page middle directory.
3587 * We've already handled the fast-path in-line.
3589 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
3591 pmd_t
*new = pmd_alloc_one(mm
, address
);
3595 smp_wmb(); /* See comment in __pte_alloc */
3597 spin_lock(&mm
->page_table_lock
);
3598 #ifndef __ARCH_HAS_4LEVEL_HACK
3599 if (!pud_present(*pud
)) {
3601 pud_populate(mm
, pud
, new);
3602 } else /* Another has populated it */
3605 if (!pgd_present(*pud
)) {
3607 pgd_populate(mm
, pud
, new);
3608 } else /* Another has populated it */
3610 #endif /* __ARCH_HAS_4LEVEL_HACK */
3611 spin_unlock(&mm
->page_table_lock
);
3614 #endif /* __PAGETABLE_PMD_FOLDED */
3616 static int __follow_pte(struct mm_struct
*mm
, unsigned long address
,
3617 pte_t
**ptepp
, spinlock_t
**ptlp
)
3624 pgd
= pgd_offset(mm
, address
);
3625 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
3628 pud
= pud_offset(pgd
, address
);
3629 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
3632 pmd
= pmd_offset(pud
, address
);
3633 VM_BUG_ON(pmd_trans_huge(*pmd
));
3634 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
3637 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3641 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
3644 if (!pte_present(*ptep
))
3649 pte_unmap_unlock(ptep
, *ptlp
);
3654 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
3655 pte_t
**ptepp
, spinlock_t
**ptlp
)
3659 /* (void) is needed to make gcc happy */
3660 (void) __cond_lock(*ptlp
,
3661 !(res
= __follow_pte(mm
, address
, ptepp
, ptlp
)));
3666 * follow_pfn - look up PFN at a user virtual address
3667 * @vma: memory mapping
3668 * @address: user virtual address
3669 * @pfn: location to store found PFN
3671 * Only IO mappings and raw PFN mappings are allowed.
3673 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3675 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
3682 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3685 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
3688 *pfn
= pte_pfn(*ptep
);
3689 pte_unmap_unlock(ptep
, ptl
);
3692 EXPORT_SYMBOL(follow_pfn
);
3694 #ifdef CONFIG_HAVE_IOREMAP_PROT
3695 int follow_phys(struct vm_area_struct
*vma
,
3696 unsigned long address
, unsigned int flags
,
3697 unsigned long *prot
, resource_size_t
*phys
)
3703 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
3706 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
3710 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
3713 *prot
= pgprot_val(pte_pgprot(pte
));
3714 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
3718 pte_unmap_unlock(ptep
, ptl
);
3723 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
3724 void *buf
, int len
, int write
)
3726 resource_size_t phys_addr
;
3727 unsigned long prot
= 0;
3728 void __iomem
*maddr
;
3729 int offset
= addr
& (PAGE_SIZE
-1);
3731 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
3734 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
3736 memcpy_toio(maddr
+ offset
, buf
, len
);
3738 memcpy_fromio(buf
, maddr
+ offset
, len
);
3743 EXPORT_SYMBOL_GPL(generic_access_phys
);
3747 * Access another process' address space as given in mm. If non-NULL, use the
3748 * given task for page fault accounting.
3750 static int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
3751 unsigned long addr
, void *buf
, int len
, int write
)
3753 struct vm_area_struct
*vma
;
3754 void *old_buf
= buf
;
3756 down_read(&mm
->mmap_sem
);
3757 /* ignore errors, just check how much was successfully transferred */
3759 int bytes
, ret
, offset
;
3761 struct page
*page
= NULL
;
3763 ret
= get_user_pages_remote(tsk
, mm
, addr
, 1,
3764 write
, 1, &page
, &vma
);
3766 #ifndef CONFIG_HAVE_IOREMAP_PROT
3770 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3771 * we can access using slightly different code.
3773 vma
= find_vma(mm
, addr
);
3774 if (!vma
|| vma
->vm_start
> addr
)
3776 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
3777 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
3785 offset
= addr
& (PAGE_SIZE
-1);
3786 if (bytes
> PAGE_SIZE
-offset
)
3787 bytes
= PAGE_SIZE
-offset
;
3791 copy_to_user_page(vma
, page
, addr
,
3792 maddr
+ offset
, buf
, bytes
);
3793 set_page_dirty_lock(page
);
3795 copy_from_user_page(vma
, page
, addr
,
3796 buf
, maddr
+ offset
, bytes
);
3805 up_read(&mm
->mmap_sem
);
3807 return buf
- old_buf
;
3811 * access_remote_vm - access another process' address space
3812 * @mm: the mm_struct of the target address space
3813 * @addr: start address to access
3814 * @buf: source or destination buffer
3815 * @len: number of bytes to transfer
3816 * @write: whether the access is a write
3818 * The caller must hold a reference on @mm.
3820 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
3821 void *buf
, int len
, int write
)
3823 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, write
);
3827 * Access another process' address space.
3828 * Source/target buffer must be kernel space,
3829 * Do not walk the page table directly, use get_user_pages
3831 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
3832 void *buf
, int len
, int write
)
3834 struct mm_struct
*mm
;
3837 mm
= get_task_mm(tsk
);
3841 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, write
);
3848 * Print the name of a VMA.
3850 void print_vma_addr(char *prefix
, unsigned long ip
)
3852 struct mm_struct
*mm
= current
->mm
;
3853 struct vm_area_struct
*vma
;
3856 * Do not print if we are in atomic
3857 * contexts (in exception stacks, etc.):
3859 if (preempt_count())
3862 down_read(&mm
->mmap_sem
);
3863 vma
= find_vma(mm
, ip
);
3864 if (vma
&& vma
->vm_file
) {
3865 struct file
*f
= vma
->vm_file
;
3866 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
3870 p
= file_path(f
, buf
, PAGE_SIZE
);
3873 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
3875 vma
->vm_end
- vma
->vm_start
);
3876 free_page((unsigned long)buf
);
3879 up_read(&mm
->mmap_sem
);
3882 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3883 void __might_fault(const char *file
, int line
)
3886 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3887 * holding the mmap_sem, this is safe because kernel memory doesn't
3888 * get paged out, therefore we'll never actually fault, and the
3889 * below annotations will generate false positives.
3891 if (segment_eq(get_fs(), KERNEL_DS
))
3893 if (pagefault_disabled())
3895 __might_sleep(file
, line
, 0);
3896 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3898 might_lock_read(¤t
->mm
->mmap_sem
);
3901 EXPORT_SYMBOL(__might_fault
);
3904 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3905 static void clear_gigantic_page(struct page
*page
,
3907 unsigned int pages_per_huge_page
)
3910 struct page
*p
= page
;
3913 for (i
= 0; i
< pages_per_huge_page
;
3914 i
++, p
= mem_map_next(p
, page
, i
)) {
3916 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
3919 void clear_huge_page(struct page
*page
,
3920 unsigned long addr
, unsigned int pages_per_huge_page
)
3924 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3925 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
3930 for (i
= 0; i
< pages_per_huge_page
; i
++) {
3932 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
3936 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
3938 struct vm_area_struct
*vma
,
3939 unsigned int pages_per_huge_page
)
3942 struct page
*dst_base
= dst
;
3943 struct page
*src_base
= src
;
3945 for (i
= 0; i
< pages_per_huge_page
; ) {
3947 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
3950 dst
= mem_map_next(dst
, dst_base
, i
);
3951 src
= mem_map_next(src
, src_base
, i
);
3955 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
3956 unsigned long addr
, struct vm_area_struct
*vma
,
3957 unsigned int pages_per_huge_page
)
3961 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
3962 copy_user_gigantic_page(dst
, src
, addr
, vma
,
3963 pages_per_huge_page
);
3968 for (i
= 0; i
< pages_per_huge_page
; i
++) {
3970 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
3973 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3975 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3977 static struct kmem_cache
*page_ptl_cachep
;
3979 void __init
ptlock_cache_init(void)
3981 page_ptl_cachep
= kmem_cache_create("page->ptl", sizeof(spinlock_t
), 0,
3985 bool ptlock_alloc(struct page
*page
)
3989 ptl
= kmem_cache_alloc(page_ptl_cachep
, GFP_KERNEL
);
3996 void ptlock_free(struct page
*page
)
3998 kmem_cache_free(page_ptl_cachep
, page
->ptl
);