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/delay.h>
53 #include <linux/init.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>
64 #ifdef CONFIG_CMA_PINPAGE_MIGRATION
65 #include <linux/mm_inline.h>
69 #include <asm/pgalloc.h>
70 #include <asm/uaccess.h>
72 #include <asm/tlbflush.h>
73 #include <asm/pgtable.h>
77 #ifdef LAST_NID_NOT_IN_PAGE_FLAGS
78 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_nid.
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
);
90 unsigned long num_physpages
;
92 * A number of key systems in x86 including ioremap() rely on the assumption
93 * that high_memory defines the upper bound on direct map memory, then end
94 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
95 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
100 EXPORT_SYMBOL(num_physpages
);
101 EXPORT_SYMBOL(high_memory
);
104 * Randomize the address space (stacks, mmaps, brk, etc.).
106 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
107 * as ancient (libc5 based) binaries can segfault. )
109 int randomize_va_space __read_mostly
=
110 #ifdef CONFIG_COMPAT_BRK
116 static int __init
disable_randmaps(char *s
)
118 randomize_va_space
= 0;
121 __setup("norandmaps", disable_randmaps
);
123 unsigned long zero_pfn __read_mostly
;
124 unsigned long highest_memmap_pfn __read_mostly
;
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 int 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;
229 tlb
->local
.next
= NULL
;
231 tlb
->local
.max
= ARRAY_SIZE(tlb
->__pages
);
232 tlb
->active
= &tlb
->local
;
233 tlb
->batch_count
= 0;
235 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
240 void tlb_flush_mmu(struct mmu_gather
*tlb
)
242 struct mmu_gather_batch
*batch
;
244 if (!tlb
->need_flush
)
248 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
249 tlb_table_flush(tlb
);
252 for (batch
= &tlb
->local
; batch
; batch
= batch
->next
) {
253 free_pages_and_swap_cache(batch
->pages
, batch
->nr
);
256 tlb
->active
= &tlb
->local
;
260 * Called at the end of the shootdown operation to free up any resources
261 * that were required.
263 void tlb_finish_mmu(struct mmu_gather
*tlb
, unsigned long start
, unsigned long end
)
265 struct mmu_gather_batch
*batch
, *next
;
269 /* keep the page table cache within bounds */
272 for (batch
= tlb
->local
.next
; batch
; batch
= next
) {
274 free_pages((unsigned long)batch
, 0);
276 tlb
->local
.next
= NULL
;
280 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
281 * handling the additional races in SMP caused by other CPUs caching valid
282 * mappings in their TLBs. Returns the number of free page slots left.
283 * When out of page slots we must call tlb_flush_mmu().
285 int __tlb_remove_page(struct mmu_gather
*tlb
, struct page
*page
)
287 struct mmu_gather_batch
*batch
;
289 VM_BUG_ON(!tlb
->need_flush
);
292 batch
->pages
[batch
->nr
++] = page
;
293 if (batch
->nr
== batch
->max
) {
294 if (!tlb_next_batch(tlb
))
298 VM_BUG_ON(batch
->nr
> batch
->max
);
300 return batch
->max
- batch
->nr
;
303 #endif /* HAVE_GENERIC_MMU_GATHER */
305 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
308 * See the comment near struct mmu_table_batch.
311 static void tlb_remove_table_smp_sync(void *arg
)
313 /* Simply deliver the interrupt */
316 static void tlb_remove_table_one(void *table
)
319 * This isn't an RCU grace period and hence the page-tables cannot be
320 * assumed to be actually RCU-freed.
322 * It is however sufficient for software page-table walkers that rely on
323 * IRQ disabling. See the comment near struct mmu_table_batch.
325 smp_call_function(tlb_remove_table_smp_sync
, NULL
, 1);
326 __tlb_remove_table(table
);
329 static void tlb_remove_table_rcu(struct rcu_head
*head
)
331 struct mmu_table_batch
*batch
;
334 batch
= container_of(head
, struct mmu_table_batch
, rcu
);
336 for (i
= 0; i
< batch
->nr
; i
++)
337 __tlb_remove_table(batch
->tables
[i
]);
339 free_page((unsigned long)batch
);
342 void tlb_table_flush(struct mmu_gather
*tlb
)
344 struct mmu_table_batch
**batch
= &tlb
->batch
;
347 call_rcu_sched(&(*batch
)->rcu
, tlb_remove_table_rcu
);
352 void tlb_remove_table(struct mmu_gather
*tlb
, void *table
)
354 struct mmu_table_batch
**batch
= &tlb
->batch
;
359 * When there's less then two users of this mm there cannot be a
360 * concurrent page-table walk.
362 if (atomic_read(&tlb
->mm
->mm_users
) < 2) {
363 __tlb_remove_table(table
);
367 if (*batch
== NULL
) {
368 *batch
= (struct mmu_table_batch
*)__get_free_page(GFP_NOWAIT
| __GFP_NOWARN
);
369 if (*batch
== NULL
) {
370 tlb_remove_table_one(table
);
375 (*batch
)->tables
[(*batch
)->nr
++] = table
;
376 if ((*batch
)->nr
== MAX_TABLE_BATCH
)
377 tlb_table_flush(tlb
);
380 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
383 * If a p?d_bad entry is found while walking page tables, report
384 * the error, before resetting entry to p?d_none. Usually (but
385 * very seldom) called out from the p?d_none_or_clear_bad macros.
388 void pgd_clear_bad(pgd_t
*pgd
)
394 void pud_clear_bad(pud_t
*pud
)
400 void pmd_clear_bad(pmd_t
*pmd
)
407 * Note: this doesn't free the actual pages themselves. That
408 * has been handled earlier when unmapping all the memory regions.
410 static void free_pte_range(struct mmu_gather
*tlb
, pmd_t
*pmd
,
413 pgtable_t token
= pmd_pgtable(*pmd
);
415 pte_free_tlb(tlb
, token
, addr
);
419 static inline void free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
420 unsigned long addr
, unsigned long end
,
421 unsigned long floor
, unsigned long ceiling
)
428 pmd
= pmd_offset(pud
, addr
);
430 next
= pmd_addr_end(addr
, end
);
431 if (pmd_none_or_clear_bad(pmd
))
433 free_pte_range(tlb
, pmd
, addr
);
434 } while (pmd
++, addr
= next
, addr
!= end
);
444 if (end
- 1 > ceiling
- 1)
447 pmd
= pmd_offset(pud
, start
);
449 pmd_free_tlb(tlb
, pmd
, start
);
452 static inline void free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
453 unsigned long addr
, unsigned long end
,
454 unsigned long floor
, unsigned long ceiling
)
461 pud
= pud_offset(pgd
, addr
);
463 next
= pud_addr_end(addr
, end
);
464 if (pud_none_or_clear_bad(pud
))
466 free_pmd_range(tlb
, pud
, addr
, next
, floor
, ceiling
);
467 } while (pud
++, addr
= next
, addr
!= end
);
473 ceiling
&= PGDIR_MASK
;
477 if (end
- 1 > ceiling
- 1)
480 pud
= pud_offset(pgd
, start
);
482 pud_free_tlb(tlb
, pud
, start
);
486 * This function frees user-level page tables of a process.
488 * Must be called with pagetable lock held.
490 void free_pgd_range(struct mmu_gather
*tlb
,
491 unsigned long addr
, unsigned long end
,
492 unsigned long floor
, unsigned long ceiling
)
498 * The next few lines have given us lots of grief...
500 * Why are we testing PMD* at this top level? Because often
501 * there will be no work to do at all, and we'd prefer not to
502 * go all the way down to the bottom just to discover that.
504 * Why all these "- 1"s? Because 0 represents both the bottom
505 * of the address space and the top of it (using -1 for the
506 * top wouldn't help much: the masks would do the wrong thing).
507 * The rule is that addr 0 and floor 0 refer to the bottom of
508 * the address space, but end 0 and ceiling 0 refer to the top
509 * Comparisons need to use "end - 1" and "ceiling - 1" (though
510 * that end 0 case should be mythical).
512 * Wherever addr is brought up or ceiling brought down, we must
513 * be careful to reject "the opposite 0" before it confuses the
514 * subsequent tests. But what about where end is brought down
515 * by PMD_SIZE below? no, end can't go down to 0 there.
517 * Whereas we round start (addr) and ceiling down, by different
518 * masks at different levels, in order to test whether a table
519 * now has no other vmas using it, so can be freed, we don't
520 * bother to round floor or end up - the tests don't need that.
534 if (end
- 1 > ceiling
- 1)
539 pgd
= pgd_offset(tlb
->mm
, addr
);
541 next
= pgd_addr_end(addr
, end
);
542 if (pgd_none_or_clear_bad(pgd
))
544 free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
545 } while (pgd
++, addr
= next
, addr
!= end
);
548 void free_pgtables(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
549 unsigned long floor
, unsigned long ceiling
)
552 struct vm_area_struct
*next
= vma
->vm_next
;
553 unsigned long addr
= vma
->vm_start
;
556 * Hide vma from rmap and truncate_pagecache before freeing
559 unlink_anon_vmas(vma
);
560 unlink_file_vma(vma
);
562 if (is_vm_hugetlb_page(vma
)) {
563 hugetlb_free_pgd_range(tlb
, addr
, vma
->vm_end
,
564 floor
, next
? next
->vm_start
: ceiling
);
567 * Optimization: gather nearby vmas into one call down
569 while (next
&& next
->vm_start
<= vma
->vm_end
+ PMD_SIZE
570 && !is_vm_hugetlb_page(next
)) {
573 unlink_anon_vmas(vma
);
574 unlink_file_vma(vma
);
576 free_pgd_range(tlb
, addr
, vma
->vm_end
,
577 floor
, next
? next
->vm_start
: ceiling
);
583 int __pte_alloc(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
584 pmd_t
*pmd
, unsigned long address
)
586 pgtable_t
new = pte_alloc_one(mm
, address
);
587 int wait_split_huge_page
;
592 * Ensure all pte setup (eg. pte page lock and page clearing) are
593 * visible before the pte is made visible to other CPUs by being
594 * put into page tables.
596 * The other side of the story is the pointer chasing in the page
597 * table walking code (when walking the page table without locking;
598 * ie. most of the time). Fortunately, these data accesses consist
599 * of a chain of data-dependent loads, meaning most CPUs (alpha
600 * being the notable exception) will already guarantee loads are
601 * seen in-order. See the alpha page table accessors for the
602 * smp_read_barrier_depends() barriers in page table walking code.
604 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
606 spin_lock(&mm
->page_table_lock
);
607 wait_split_huge_page
= 0;
608 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
610 pmd_populate(mm
, pmd
, new);
612 } else if (unlikely(pmd_trans_splitting(*pmd
)))
613 wait_split_huge_page
= 1;
614 spin_unlock(&mm
->page_table_lock
);
617 if (wait_split_huge_page
)
618 wait_split_huge_page(vma
->anon_vma
, pmd
);
622 int __pte_alloc_kernel(pmd_t
*pmd
, unsigned long address
)
624 pte_t
*new = pte_alloc_one_kernel(&init_mm
, address
);
628 smp_wmb(); /* See comment in __pte_alloc */
630 spin_lock(&init_mm
.page_table_lock
);
631 if (likely(pmd_none(*pmd
))) { /* Has another populated it ? */
632 pmd_populate_kernel(&init_mm
, pmd
, new);
635 VM_BUG_ON(pmd_trans_splitting(*pmd
));
636 spin_unlock(&init_mm
.page_table_lock
);
638 pte_free_kernel(&init_mm
, new);
642 static inline void init_rss_vec(int *rss
)
644 memset(rss
, 0, sizeof(int) * NR_MM_COUNTERS
);
647 static inline void add_mm_rss_vec(struct mm_struct
*mm
, int *rss
)
651 if (current
->mm
== mm
)
653 for (i
= 0; i
< NR_MM_COUNTERS
; i
++)
655 add_mm_counter(mm
, i
, rss
[i
]);
659 * This function is called to print an error when a bad pte
660 * is found. For example, we might have a PFN-mapped pte in
661 * a region that doesn't allow it.
663 * The calling function must still handle the error.
665 static void print_bad_pte(struct vm_area_struct
*vma
, unsigned long addr
,
666 pte_t pte
, struct page
*page
)
668 pgd_t
*pgd
= pgd_offset(vma
->vm_mm
, addr
);
669 pud_t
*pud
= pud_offset(pgd
, addr
);
670 pmd_t
*pmd
= pmd_offset(pud
, addr
);
671 struct address_space
*mapping
;
673 static unsigned long resume
;
674 static unsigned long nr_shown
;
675 static unsigned long nr_unshown
;
678 * Allow a burst of 60 reports, then keep quiet for that minute;
679 * or allow a steady drip of one report per second.
681 if (nr_shown
== 60) {
682 if (time_before(jiffies
, resume
)) {
688 "BUG: Bad page map: %lu messages suppressed\n",
695 resume
= jiffies
+ 60 * HZ
;
697 mapping
= vma
->vm_file
? vma
->vm_file
->f_mapping
: NULL
;
698 index
= linear_page_index(vma
, addr
);
701 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
703 (long long)pte_val(pte
), (long long)pmd_val(*pmd
));
707 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
708 (void *)addr
, vma
->vm_flags
, vma
->anon_vma
, mapping
, index
);
710 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
713 printk(KERN_ALERT
"vma->vm_ops->fault: %pSR\n",
715 if (vma
->vm_file
&& vma
->vm_file
->f_op
)
716 printk(KERN_ALERT
"vma->vm_file->f_op->mmap: %pSR\n",
717 vma
->vm_file
->f_op
->mmap
);
719 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
722 static inline bool is_cow_mapping(vm_flags_t flags
)
724 return (flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
728 * vm_normal_page -- This function gets the "struct page" associated with a pte.
730 * "Special" mappings do not wish to be associated with a "struct page" (either
731 * it doesn't exist, or it exists but they don't want to touch it). In this
732 * case, NULL is returned here. "Normal" mappings do have a struct page.
734 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
735 * pte bit, in which case this function is trivial. Secondly, an architecture
736 * may not have a spare pte bit, which requires a more complicated scheme,
739 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
740 * special mapping (even if there are underlying and valid "struct pages").
741 * COWed pages of a VM_PFNMAP are always normal.
743 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
744 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
745 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
746 * mapping will always honor the rule
748 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
750 * And for normal mappings this is false.
752 * This restricts such mappings to be a linear translation from virtual address
753 * to pfn. To get around this restriction, we allow arbitrary mappings so long
754 * as the vma is not a COW mapping; in that case, we know that all ptes are
755 * special (because none can have been COWed).
758 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
760 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
761 * page" backing, however the difference is that _all_ pages with a struct
762 * page (that is, those where pfn_valid is true) are refcounted and considered
763 * normal pages by the VM. The disadvantage is that pages are refcounted
764 * (which can be slower and simply not an option for some PFNMAP users). The
765 * advantage is that we don't have to follow the strict linearity rule of
766 * PFNMAP mappings in order to support COWable mappings.
769 #ifdef __HAVE_ARCH_PTE_SPECIAL
770 # define HAVE_PTE_SPECIAL 1
772 # define HAVE_PTE_SPECIAL 0
774 struct page
*vm_normal_page(struct vm_area_struct
*vma
, unsigned long addr
,
777 unsigned long pfn
= pte_pfn(pte
);
779 if (HAVE_PTE_SPECIAL
) {
780 if (likely(!pte_special(pte
)))
782 if (vma
->vm_flags
& (VM_PFNMAP
| VM_MIXEDMAP
))
784 if (!is_zero_pfn(pfn
))
785 print_bad_pte(vma
, addr
, pte
, NULL
);
789 /* !HAVE_PTE_SPECIAL case follows: */
791 if (unlikely(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
))) {
792 if (vma
->vm_flags
& VM_MIXEDMAP
) {
798 off
= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
799 if (pfn
== vma
->vm_pgoff
+ off
)
801 if (!is_cow_mapping(vma
->vm_flags
))
806 if (is_zero_pfn(pfn
))
809 if (unlikely(pfn
> highest_memmap_pfn
)) {
810 print_bad_pte(vma
, addr
, pte
, NULL
);
815 * NOTE! We still have PageReserved() pages in the page tables.
816 * eg. VDSO mappings can cause them to exist.
819 return pfn_to_page(pfn
);
823 * copy one vm_area from one task to the other. Assumes the page tables
824 * already present in the new task to be cleared in the whole range
825 * covered by this vma.
828 static inline unsigned long
829 copy_one_pte(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
830 pte_t
*dst_pte
, pte_t
*src_pte
, struct vm_area_struct
*vma
,
831 unsigned long addr
, int *rss
)
833 unsigned long vm_flags
= vma
->vm_flags
;
834 pte_t pte
= *src_pte
;
837 /* pte contains position in swap or file, so copy. */
838 if (unlikely(!pte_present(pte
))) {
839 if (!pte_file(pte
)) {
840 swp_entry_t entry
= pte_to_swp_entry(pte
);
842 if (swap_duplicate(entry
) < 0)
845 /* make sure dst_mm is on swapoff's mmlist. */
846 if (unlikely(list_empty(&dst_mm
->mmlist
))) {
847 spin_lock(&mmlist_lock
);
848 if (list_empty(&dst_mm
->mmlist
))
849 list_add(&dst_mm
->mmlist
,
851 spin_unlock(&mmlist_lock
);
853 if (likely(!non_swap_entry(entry
)))
855 else if (is_migration_entry(entry
)) {
856 page
= migration_entry_to_page(entry
);
863 if (is_write_migration_entry(entry
) &&
864 is_cow_mapping(vm_flags
)) {
866 * COW mappings require pages in both
867 * parent and child to be set to read.
869 make_migration_entry_read(&entry
);
870 pte
= swp_entry_to_pte(entry
);
871 set_pte_at(src_mm
, addr
, src_pte
, pte
);
879 * If it's a COW mapping, write protect it both
880 * in the parent and the child
882 if (is_cow_mapping(vm_flags
)) {
883 ptep_set_wrprotect(src_mm
, addr
, src_pte
);
884 pte
= pte_wrprotect(pte
);
888 * If it's a shared mapping, mark it clean in
891 if (vm_flags
& VM_SHARED
)
892 pte
= pte_mkclean(pte
);
893 pte
= pte_mkold(pte
);
895 page
= vm_normal_page(vma
, addr
, pte
);
906 set_pte_at(dst_mm
, addr
, dst_pte
, pte
);
910 int copy_pte_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
911 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, struct vm_area_struct
*vma
,
912 unsigned long addr
, unsigned long end
)
914 pte_t
*orig_src_pte
, *orig_dst_pte
;
915 pte_t
*src_pte
, *dst_pte
;
916 spinlock_t
*src_ptl
, *dst_ptl
;
918 int rss
[NR_MM_COUNTERS
];
919 swp_entry_t entry
= (swp_entry_t
){0};
924 dst_pte
= pte_alloc_map_lock(dst_mm
, dst_pmd
, addr
, &dst_ptl
);
927 src_pte
= pte_offset_map(src_pmd
, addr
);
928 src_ptl
= pte_lockptr(src_mm
, src_pmd
);
929 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
930 orig_src_pte
= src_pte
;
931 orig_dst_pte
= dst_pte
;
932 arch_enter_lazy_mmu_mode();
936 * We are holding two locks at this point - either of them
937 * could generate latencies in another task on another CPU.
939 if (progress
>= 32) {
941 if (need_resched() ||
942 spin_needbreak(src_ptl
) || spin_needbreak(dst_ptl
))
945 if (pte_none(*src_pte
)) {
949 entry
.val
= copy_one_pte(dst_mm
, src_mm
, dst_pte
, src_pte
,
954 } while (dst_pte
++, src_pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
956 arch_leave_lazy_mmu_mode();
957 spin_unlock(src_ptl
);
958 pte_unmap(orig_src_pte
);
959 add_mm_rss_vec(dst_mm
, rss
);
960 pte_unmap_unlock(orig_dst_pte
, dst_ptl
);
964 if (add_swap_count_continuation(entry
, GFP_KERNEL
) < 0)
973 static inline int copy_pmd_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
974 pud_t
*dst_pud
, pud_t
*src_pud
, struct vm_area_struct
*vma
,
975 unsigned long addr
, unsigned long end
)
977 pmd_t
*src_pmd
, *dst_pmd
;
980 dst_pmd
= pmd_alloc(dst_mm
, dst_pud
, addr
);
983 src_pmd
= pmd_offset(src_pud
, addr
);
985 next
= pmd_addr_end(addr
, end
);
986 if (pmd_trans_huge(*src_pmd
)) {
988 VM_BUG_ON(next
-addr
!= HPAGE_PMD_SIZE
);
989 err
= copy_huge_pmd(dst_mm
, src_mm
,
990 dst_pmd
, src_pmd
, addr
, vma
);
997 if (pmd_none_or_clear_bad(src_pmd
))
999 if (copy_pte_range(dst_mm
, src_mm
, dst_pmd
, src_pmd
,
1002 } while (dst_pmd
++, src_pmd
++, addr
= next
, addr
!= end
);
1006 static inline int copy_pud_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1007 pgd_t
*dst_pgd
, pgd_t
*src_pgd
, struct vm_area_struct
*vma
,
1008 unsigned long addr
, unsigned long end
)
1010 pud_t
*src_pud
, *dst_pud
;
1013 dst_pud
= pud_alloc(dst_mm
, dst_pgd
, addr
);
1016 src_pud
= pud_offset(src_pgd
, addr
);
1018 next
= pud_addr_end(addr
, end
);
1019 if (pud_none_or_clear_bad(src_pud
))
1021 if (copy_pmd_range(dst_mm
, src_mm
, dst_pud
, src_pud
,
1024 } while (dst_pud
++, src_pud
++, addr
= next
, addr
!= end
);
1028 int copy_page_range(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1029 struct vm_area_struct
*vma
)
1031 pgd_t
*src_pgd
, *dst_pgd
;
1033 unsigned long addr
= vma
->vm_start
;
1034 unsigned long end
= vma
->vm_end
;
1035 unsigned long mmun_start
; /* For mmu_notifiers */
1036 unsigned long mmun_end
; /* For mmu_notifiers */
1041 * Don't copy ptes where a page fault will fill them correctly.
1042 * Fork becomes much lighter when there are big shared or private
1043 * readonly mappings. The tradeoff is that copy_page_range is more
1044 * efficient than faulting.
1046 if (!(vma
->vm_flags
& (VM_HUGETLB
| VM_NONLINEAR
|
1047 VM_PFNMAP
| VM_MIXEDMAP
))) {
1052 if (is_vm_hugetlb_page(vma
))
1053 return copy_hugetlb_page_range(dst_mm
, src_mm
, vma
);
1055 if (unlikely(vma
->vm_flags
& VM_PFNMAP
)) {
1057 * We do not free on error cases below as remove_vma
1058 * gets called on error from higher level routine
1060 ret
= track_pfn_copy(vma
);
1066 * We need to invalidate the secondary MMU mappings only when
1067 * there could be a permission downgrade on the ptes of the
1068 * parent mm. And a permission downgrade will only happen if
1069 * is_cow_mapping() returns true.
1071 is_cow
= is_cow_mapping(vma
->vm_flags
);
1075 mmu_notifier_invalidate_range_start(src_mm
, mmun_start
,
1079 dst_pgd
= pgd_offset(dst_mm
, addr
);
1080 src_pgd
= pgd_offset(src_mm
, addr
);
1082 next
= pgd_addr_end(addr
, end
);
1083 if (pgd_none_or_clear_bad(src_pgd
))
1085 if (unlikely(copy_pud_range(dst_mm
, src_mm
, dst_pgd
, src_pgd
,
1086 vma
, addr
, next
))) {
1090 } while (dst_pgd
++, src_pgd
++, addr
= next
, addr
!= end
);
1093 mmu_notifier_invalidate_range_end(src_mm
, mmun_start
, mmun_end
);
1097 static unsigned long zap_pte_range(struct mmu_gather
*tlb
,
1098 struct vm_area_struct
*vma
, pmd_t
*pmd
,
1099 unsigned long addr
, unsigned long end
,
1100 struct zap_details
*details
)
1102 struct mm_struct
*mm
= tlb
->mm
;
1103 int force_flush
= 0;
1104 int rss
[NR_MM_COUNTERS
];
1111 start_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1113 arch_enter_lazy_mmu_mode();
1116 if (pte_none(ptent
)) {
1120 if (pte_present(ptent
)) {
1123 page
= vm_normal_page(vma
, addr
, ptent
);
1124 if (unlikely(details
) && page
) {
1126 * unmap_shared_mapping_pages() wants to
1127 * invalidate cache without truncating:
1128 * unmap shared but keep private pages.
1130 if (details
->check_mapping
&&
1131 details
->check_mapping
!= page
->mapping
)
1134 * Each page->index must be checked when
1135 * invalidating or truncating nonlinear.
1137 if (details
->nonlinear_vma
&&
1138 (page
->index
< details
->first_index
||
1139 page
->index
> details
->last_index
))
1142 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
1144 tlb_remove_tlb_entry(tlb
, pte
, addr
);
1145 if (unlikely(!page
))
1147 if (unlikely(details
) && details
->nonlinear_vma
1148 && linear_page_index(details
->nonlinear_vma
,
1149 addr
) != page
->index
)
1150 set_pte_at(mm
, addr
, pte
,
1151 pgoff_to_pte(page
->index
));
1153 rss
[MM_ANONPAGES
]--;
1155 if (pte_dirty(ptent
))
1156 set_page_dirty(page
);
1157 if (pte_young(ptent
) &&
1158 likely(!VM_SequentialReadHint(vma
)))
1159 mark_page_accessed(page
);
1160 rss
[MM_FILEPAGES
]--;
1162 page_remove_rmap(page
);
1163 if (unlikely(page_mapcount(page
) < 0))
1164 print_bad_pte(vma
, addr
, ptent
, page
);
1165 force_flush
= !__tlb_remove_page(tlb
, page
);
1171 * If details->check_mapping, we leave swap entries;
1172 * if details->nonlinear_vma, we leave file entries.
1174 if (unlikely(details
))
1176 if (pte_file(ptent
)) {
1177 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
)))
1178 print_bad_pte(vma
, addr
, ptent
, NULL
);
1180 swp_entry_t entry
= pte_to_swp_entry(ptent
);
1182 if (!non_swap_entry(entry
))
1184 else if (is_migration_entry(entry
)) {
1187 page
= migration_entry_to_page(entry
);
1190 rss
[MM_ANONPAGES
]--;
1192 rss
[MM_FILEPAGES
]--;
1194 if (unlikely(!free_swap_and_cache(entry
)))
1195 print_bad_pte(vma
, addr
, ptent
, NULL
);
1197 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
1198 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1200 add_mm_rss_vec(mm
, rss
);
1201 arch_leave_lazy_mmu_mode();
1202 pte_unmap_unlock(start_pte
, ptl
);
1205 * mmu_gather ran out of room to batch pages, we break out of
1206 * the PTE lock to avoid doing the potential expensive TLB invalidate
1207 * and page-free while holding it.
1210 unsigned long old_end
;
1215 * Flush the TLB just for the previous segment,
1216 * then update the range to be the remaining
1234 static inline unsigned long zap_pmd_range(struct mmu_gather
*tlb
,
1235 struct vm_area_struct
*vma
, pud_t
*pud
,
1236 unsigned long addr
, unsigned long end
,
1237 struct zap_details
*details
)
1242 pmd
= pmd_offset(pud
, addr
);
1244 next
= pmd_addr_end(addr
, end
);
1245 if (pmd_trans_huge(*pmd
)) {
1246 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1247 #ifdef CONFIG_DEBUG_VM
1248 if (!rwsem_is_locked(&tlb
->mm
->mmap_sem
)) {
1249 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1250 __func__
, addr
, end
,
1256 split_huge_page_pmd(vma
, addr
, pmd
);
1257 } else if (zap_huge_pmd(tlb
, vma
, pmd
, addr
))
1262 * Here there can be other concurrent MADV_DONTNEED or
1263 * trans huge page faults running, and if the pmd is
1264 * none or trans huge it can change under us. This is
1265 * because MADV_DONTNEED holds the mmap_sem in read
1268 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
1270 next
= zap_pte_range(tlb
, vma
, pmd
, addr
, next
, details
);
1273 } while (pmd
++, addr
= next
, addr
!= end
);
1278 static inline unsigned long zap_pud_range(struct mmu_gather
*tlb
,
1279 struct vm_area_struct
*vma
, pgd_t
*pgd
,
1280 unsigned long addr
, unsigned long end
,
1281 struct zap_details
*details
)
1286 pud
= pud_offset(pgd
, addr
);
1288 next
= pud_addr_end(addr
, end
);
1289 if (pud_none_or_clear_bad(pud
))
1291 next
= zap_pmd_range(tlb
, vma
, pud
, addr
, next
, details
);
1292 } while (pud
++, addr
= next
, addr
!= end
);
1297 static void unmap_page_range(struct mmu_gather
*tlb
,
1298 struct vm_area_struct
*vma
,
1299 unsigned long addr
, unsigned long end
,
1300 struct zap_details
*details
)
1305 if (details
&& !details
->check_mapping
&& !details
->nonlinear_vma
)
1308 BUG_ON(addr
>= end
);
1309 mem_cgroup_uncharge_start();
1310 tlb_start_vma(tlb
, vma
);
1311 pgd
= pgd_offset(vma
->vm_mm
, addr
);
1313 next
= pgd_addr_end(addr
, end
);
1314 if (pgd_none_or_clear_bad(pgd
))
1316 next
= zap_pud_range(tlb
, vma
, pgd
, addr
, next
, details
);
1317 } while (pgd
++, addr
= next
, addr
!= end
);
1318 tlb_end_vma(tlb
, vma
);
1319 mem_cgroup_uncharge_end();
1323 static void unmap_single_vma(struct mmu_gather
*tlb
,
1324 struct vm_area_struct
*vma
, unsigned long start_addr
,
1325 unsigned long end_addr
,
1326 struct zap_details
*details
)
1328 unsigned long start
= max(vma
->vm_start
, start_addr
);
1331 if (start
>= vma
->vm_end
)
1333 end
= min(vma
->vm_end
, end_addr
);
1334 if (end
<= vma
->vm_start
)
1338 uprobe_munmap(vma
, start
, end
);
1340 if (unlikely(vma
->vm_flags
& VM_PFNMAP
))
1341 untrack_pfn(vma
, 0, 0);
1344 if (unlikely(is_vm_hugetlb_page(vma
))) {
1346 * It is undesirable to test vma->vm_file as it
1347 * should be non-null for valid hugetlb area.
1348 * However, vm_file will be NULL in the error
1349 * cleanup path of do_mmap_pgoff. When
1350 * hugetlbfs ->mmap method fails,
1351 * do_mmap_pgoff() nullifies vma->vm_file
1352 * before calling this function to clean up.
1353 * Since no pte has actually been setup, it is
1354 * safe to do nothing in this case.
1357 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
1358 __unmap_hugepage_range_final(tlb
, vma
, start
, end
, NULL
);
1359 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
1362 unmap_page_range(tlb
, vma
, start
, end
, details
);
1367 * unmap_vmas - unmap a range of memory covered by a list of vma's
1368 * @tlb: address of the caller's struct mmu_gather
1369 * @vma: the starting vma
1370 * @start_addr: virtual address at which to start unmapping
1371 * @end_addr: virtual address at which to end unmapping
1373 * Unmap all pages in the vma list.
1375 * Only addresses between `start' and `end' will be unmapped.
1377 * The VMA list must be sorted in ascending virtual address order.
1379 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1380 * range after unmap_vmas() returns. So the only responsibility here is to
1381 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1382 * drops the lock and schedules.
1384 void unmap_vmas(struct mmu_gather
*tlb
,
1385 struct vm_area_struct
*vma
, unsigned long start_addr
,
1386 unsigned long end_addr
)
1388 struct mm_struct
*mm
= vma
->vm_mm
;
1390 mmu_notifier_invalidate_range_start(mm
, start_addr
, end_addr
);
1391 for ( ; vma
&& vma
->vm_start
< end_addr
; vma
= vma
->vm_next
)
1392 unmap_single_vma(tlb
, vma
, start_addr
, end_addr
, NULL
);
1393 mmu_notifier_invalidate_range_end(mm
, start_addr
, end_addr
);
1397 * zap_page_range - remove user pages in a given range
1398 * @vma: vm_area_struct holding the applicable pages
1399 * @start: starting address of pages to zap
1400 * @size: number of bytes to zap
1401 * @details: details of nonlinear truncation or shared cache invalidation
1403 * Caller must protect the VMA list
1405 void zap_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1406 unsigned long size
, struct zap_details
*details
)
1408 struct mm_struct
*mm
= vma
->vm_mm
;
1409 struct mmu_gather tlb
;
1410 unsigned long end
= start
+ size
;
1413 tlb_gather_mmu(&tlb
, mm
, start
, end
);
1414 update_hiwater_rss(mm
);
1415 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1416 for ( ; vma
&& vma
->vm_start
< end
; vma
= vma
->vm_next
)
1417 unmap_single_vma(&tlb
, vma
, start
, end
, details
);
1418 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1419 tlb_finish_mmu(&tlb
, start
, end
);
1423 * zap_page_range_single - remove user pages in a given range
1424 * @vma: vm_area_struct holding the applicable pages
1425 * @address: starting address of pages to zap
1426 * @size: number of bytes to zap
1427 * @details: details of nonlinear truncation or shared cache invalidation
1429 * The range must fit into one VMA.
1431 static void zap_page_range_single(struct vm_area_struct
*vma
, unsigned long address
,
1432 unsigned long size
, struct zap_details
*details
)
1434 struct mm_struct
*mm
= vma
->vm_mm
;
1435 struct mmu_gather tlb
;
1436 unsigned long end
= address
+ size
;
1439 tlb_gather_mmu(&tlb
, mm
, address
, end
);
1440 update_hiwater_rss(mm
);
1441 mmu_notifier_invalidate_range_start(mm
, address
, end
);
1442 unmap_single_vma(&tlb
, vma
, address
, end
, details
);
1443 mmu_notifier_invalidate_range_end(mm
, address
, end
);
1444 tlb_finish_mmu(&tlb
, address
, end
);
1448 * FOLL_FORCE can write to even unwritable pte's, but only
1449 * after we've gone through a COW cycle and they are dirty.
1451 static inline bool can_follow_write_pte(pte_t pte
, unsigned int flags
)
1453 return pte_write(pte
) ||
1454 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pte_dirty(pte
));
1458 * zap_vma_ptes - remove ptes mapping the vma
1459 * @vma: vm_area_struct holding ptes to be zapped
1460 * @address: starting address of pages to zap
1461 * @size: number of bytes to zap
1463 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1465 * The entire address range must be fully contained within the vma.
1467 * Returns 0 if successful.
1469 int zap_vma_ptes(struct vm_area_struct
*vma
, unsigned long address
,
1472 if (address
< vma
->vm_start
|| address
+ size
> vma
->vm_end
||
1473 !(vma
->vm_flags
& VM_PFNMAP
))
1475 zap_page_range_single(vma
, address
, size
, NULL
);
1478 EXPORT_SYMBOL_GPL(zap_vma_ptes
);
1480 #ifdef CONFIG_CMA_PINPAGE_MIGRATION
1481 static struct page
*__alloc_nonmovable_userpage(struct page
*page
,
1482 unsigned long private, int **result
)
1484 return alloc_page(GFP_HIGHUSER
);
1487 static bool __need_migrate_cma_page(struct page
*page
,
1488 struct vm_area_struct
*vma
,
1489 unsigned long start
, unsigned int flags
)
1491 if (!(flags
& FOLL_GET
))
1494 if (!is_cma_pageblock(page
))
1497 if ((vma
->vm_flags
& VM_STACK_INCOMPLETE_SETUP
) ==
1498 VM_STACK_INCOMPLETE_SETUP
)
1501 if (!(flags
& FOLL_CMA
))
1504 migrate_prep_local();
1512 static int __migrate_cma_pinpage(struct page
*page
, struct vm_area_struct
*vma
)
1514 struct zone
*zone
= page_zone(page
);
1515 struct list_head migratepages
;
1516 struct lruvec
*lruvec
;
1520 INIT_LIST_HEAD(&migratepages
);
1522 if (__isolate_lru_page(page
, 0) != 0) {
1523 pr_warn("%s: failed to isolate lru page\n", __func__
);
1527 spin_lock_irq(&zone
->lru_lock
);
1528 lruvec
= mem_cgroup_page_lruvec(page
, page_zone(page
));
1529 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
1530 spin_unlock_irq(&zone
->lru_lock
);
1533 list_add(&page
->lru
, &migratepages
);
1534 inc_zone_page_state(page
, NR_ISOLATED_ANON
+ page_is_file_cache(page
));
1536 while (!list_empty(&migratepages
) && tries
++ < 5) {
1537 ret
= migrate_pages(&migratepages
,
1538 __alloc_nonmovable_userpage
, 0, MIGRATE_SYNC
, MR_CMA
);
1542 putback_movable_pages(&migratepages
);
1543 pr_err("%s: migration failed %p[%#lx]\n", __func__
,
1544 page
, page_to_pfn(page
));
1553 * follow_page_mask - look up a page descriptor from a user-virtual address
1554 * @vma: vm_area_struct mapping @address
1555 * @address: virtual address to look up
1556 * @flags: flags modifying lookup behaviour
1557 * @page_mask: on output, *page_mask is set according to the size of the page
1559 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1561 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1562 * an error pointer if there is a mapping to something not represented
1563 * by a page descriptor (see also vm_normal_page()).
1565 struct page
*follow_page_mask(struct vm_area_struct
*vma
,
1566 unsigned long address
, unsigned int flags
,
1567 unsigned int *page_mask
)
1575 struct mm_struct
*mm
= vma
->vm_mm
;
1579 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
1580 if (!IS_ERR(page
)) {
1581 BUG_ON(flags
& FOLL_GET
);
1586 pgd
= pgd_offset(mm
, address
);
1587 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
1590 pud
= pud_offset(pgd
, address
);
1593 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
1594 BUG_ON(flags
& FOLL_GET
);
1595 page
= follow_huge_pud(mm
, address
, pud
, flags
& FOLL_WRITE
);
1598 if (unlikely(pud_bad(*pud
)))
1601 pmd
= pmd_offset(pud
, address
);
1604 if (pmd_huge(*pmd
) && vma
->vm_flags
& VM_HUGETLB
) {
1605 BUG_ON(flags
& FOLL_GET
);
1606 page
= follow_huge_pmd(mm
, address
, pmd
, flags
& FOLL_WRITE
);
1609 if ((flags
& FOLL_NUMA
) && pmd_numa(*pmd
))
1611 if (pmd_trans_huge(*pmd
)) {
1612 if (flags
& FOLL_SPLIT
) {
1613 split_huge_page_pmd(vma
, address
, pmd
);
1614 goto split_fallthrough
;
1616 spin_lock(&mm
->page_table_lock
);
1617 if (likely(pmd_trans_huge(*pmd
))) {
1618 if (unlikely(pmd_trans_splitting(*pmd
))) {
1619 spin_unlock(&mm
->page_table_lock
);
1620 wait_split_huge_page(vma
->anon_vma
, pmd
);
1622 page
= follow_trans_huge_pmd(vma
, address
,
1624 spin_unlock(&mm
->page_table_lock
);
1625 *page_mask
= HPAGE_PMD_NR
- 1;
1629 spin_unlock(&mm
->page_table_lock
);
1633 if (unlikely(pmd_bad(*pmd
)))
1636 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1639 if (!pte_present(pte
)) {
1642 * KSM's break_ksm() relies upon recognizing a ksm page
1643 * even while it is being migrated, so for that case we
1644 * need migration_entry_wait().
1646 if (likely(!(flags
& FOLL_MIGRATION
)))
1648 if (pte_none(pte
) || pte_file(pte
))
1650 entry
= pte_to_swp_entry(pte
);
1651 if (!is_migration_entry(entry
))
1653 pte_unmap_unlock(ptep
, ptl
);
1654 migration_entry_wait(mm
, pmd
, address
);
1655 goto split_fallthrough
;
1657 if ((flags
& FOLL_NUMA
) && pte_numa(pte
))
1659 if ((flags
& FOLL_WRITE
) && !can_follow_write_pte(pte
, flags
))
1662 page
= vm_normal_page(vma
, address
, pte
);
1663 if (unlikely(!page
)) {
1664 if ((flags
& FOLL_DUMP
) ||
1665 !is_zero_pfn(pte_pfn(pte
)))
1667 page
= pte_page(pte
);
1670 #ifdef CONFIG_CMA_PINPAGE_MIGRATION
1671 if (__need_migrate_cma_page(page
, vma
, address
, flags
)) {
1672 pte_unmap_unlock(ptep
, ptl
);
1673 if (__migrate_cma_pinpage(page
, vma
)) {
1674 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1676 struct page
*old_page
= page
;
1678 migration_entry_wait(mm
, pmd
, address
);
1679 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1680 update_mmu_cache(vma
, address
, ptep
);
1682 set_pte_at_notify(mm
, address
, ptep
, pte
);
1683 page
= vm_normal_page(vma
, address
, pte
);
1686 pr_debug("cma: cma page %p[%#lx] migrated to new "
1687 "page %p[%#lx]\n", old_page
,
1688 page_to_pfn(old_page
),
1689 page
, page_to_pfn(page
));
1693 if (flags
& FOLL_GET
)
1694 get_page_foll(page
);
1695 if (flags
& FOLL_TOUCH
) {
1696 if ((flags
& FOLL_WRITE
) &&
1697 !pte_dirty(pte
) && !PageDirty(page
))
1698 set_page_dirty(page
);
1700 * pte_mkyoung() would be more correct here, but atomic care
1701 * is needed to avoid losing the dirty bit: it is easier to use
1702 * mark_page_accessed().
1704 mark_page_accessed(page
);
1706 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1708 * The preliminary mapping check is mainly to avoid the
1709 * pointless overhead of lock_page on the ZERO_PAGE
1710 * which might bounce very badly if there is contention.
1712 * If the page is already locked, we don't need to
1713 * handle it now - vmscan will handle it later if and
1714 * when it attempts to reclaim the page.
1716 if (page
->mapping
&& trylock_page(page
)) {
1717 lru_add_drain(); /* push cached pages to LRU */
1719 * Because we lock page here, and migration is
1720 * blocked by the pte's page reference, and we
1721 * know the page is still mapped, we don't even
1722 * need to check for file-cache page truncation.
1724 mlock_vma_page(page
);
1729 pte_unmap_unlock(ptep
, ptl
);
1734 pte_unmap_unlock(ptep
, ptl
);
1735 return ERR_PTR(-EFAULT
);
1738 pte_unmap_unlock(ptep
, ptl
);
1744 * When core dumping an enormous anonymous area that nobody
1745 * has touched so far, we don't want to allocate unnecessary pages or
1746 * page tables. Return error instead of NULL to skip handle_mm_fault,
1747 * then get_dump_page() will return NULL to leave a hole in the dump.
1748 * But we can only make this optimization where a hole would surely
1749 * be zero-filled if handle_mm_fault() actually did handle it.
1751 if ((flags
& FOLL_DUMP
) &&
1752 (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
1753 return ERR_PTR(-EFAULT
);
1758 * __get_user_pages() - pin user pages in memory
1759 * @tsk: task_struct of target task
1760 * @mm: mm_struct of target mm
1761 * @start: starting user address
1762 * @nr_pages: number of pages from start to pin
1763 * @gup_flags: flags modifying pin behaviour
1764 * @pages: array that receives pointers to the pages pinned.
1765 * Should be at least nr_pages long. Or NULL, if caller
1766 * only intends to ensure the pages are faulted in.
1767 * @vmas: array of pointers to vmas corresponding to each page.
1768 * Or NULL if the caller does not require them.
1769 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1771 * Returns number of pages pinned. This may be fewer than the number
1772 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1773 * were pinned, returns -errno. Each page returned must be released
1774 * with a put_page() call when it is finished with. vmas will only
1775 * remain valid while mmap_sem is held.
1777 * Must be called with mmap_sem held for read or write.
1779 * __get_user_pages walks a process's page tables and takes a reference to
1780 * each struct page that each user address corresponds to at a given
1781 * instant. That is, it takes the page that would be accessed if a user
1782 * thread accesses the given user virtual address at that instant.
1784 * This does not guarantee that the page exists in the user mappings when
1785 * __get_user_pages returns, and there may even be a completely different
1786 * page there in some cases (eg. if mmapped pagecache has been invalidated
1787 * and subsequently re faulted). However it does guarantee that the page
1788 * won't be freed completely. And mostly callers simply care that the page
1789 * contains data that was valid *at some point in time*. Typically, an IO
1790 * or similar operation cannot guarantee anything stronger anyway because
1791 * locks can't be held over the syscall boundary.
1793 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1794 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1795 * appropriate) must be called after the page is finished with, and
1796 * before put_page is called.
1798 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1799 * or mmap_sem contention, and if waiting is needed to pin all pages,
1800 * *@nonblocking will be set to 0.
1802 * In most cases, get_user_pages or get_user_pages_fast should be used
1803 * instead of __get_user_pages. __get_user_pages should be used only if
1804 * you need some special @gup_flags.
1806 long __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1807 unsigned long start
, unsigned long nr_pages
,
1808 unsigned int gup_flags
, struct page
**pages
,
1809 struct vm_area_struct
**vmas
, int *nonblocking
)
1812 unsigned long vm_flags
;
1813 unsigned int page_mask
;
1818 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
1821 * Require read or write permissions.
1822 * If FOLL_FORCE is set, we only require the "MAY" flags.
1824 vm_flags
= (gup_flags
& FOLL_WRITE
) ?
1825 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1826 vm_flags
&= (gup_flags
& FOLL_FORCE
) ?
1827 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1830 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1831 * would be called on PROT_NONE ranges. We must never invoke
1832 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1833 * page faults would unprotect the PROT_NONE ranges if
1834 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1835 * bitflag. So to avoid that, don't set FOLL_NUMA if
1836 * FOLL_FORCE is set.
1838 if (!(gup_flags
& FOLL_FORCE
))
1839 gup_flags
|= FOLL_NUMA
;
1844 struct vm_area_struct
*vma
;
1846 vma
= find_extend_vma(mm
, start
);
1847 if (!vma
&& in_gate_area(mm
, start
)) {
1848 unsigned long pg
= start
& PAGE_MASK
;
1854 /* user gate pages are read-only */
1855 if (gup_flags
& FOLL_WRITE
)
1856 return i
? : -EFAULT
;
1858 pgd
= pgd_offset_k(pg
);
1860 pgd
= pgd_offset_gate(mm
, pg
);
1861 BUG_ON(pgd_none(*pgd
));
1862 pud
= pud_offset(pgd
, pg
);
1863 BUG_ON(pud_none(*pud
));
1864 pmd
= pmd_offset(pud
, pg
);
1866 return i
? : -EFAULT
;
1867 VM_BUG_ON(pmd_trans_huge(*pmd
));
1868 pte
= pte_offset_map(pmd
, pg
);
1869 if (pte_none(*pte
)) {
1871 return i
? : -EFAULT
;
1873 vma
= get_gate_vma(mm
);
1877 page
= vm_normal_page(vma
, start
, *pte
);
1879 if (!(gup_flags
& FOLL_DUMP
) &&
1880 is_zero_pfn(pte_pfn(*pte
)))
1881 page
= pte_page(*pte
);
1884 return i
? : -EFAULT
;
1896 (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1897 !(vm_flags
& vma
->vm_flags
))
1898 return i
? : -EFAULT
;
1900 if (is_vm_hugetlb_page(vma
)) {
1901 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1902 &start
, &nr_pages
, i
, gup_flags
);
1908 unsigned int foll_flags
= gup_flags
;
1909 unsigned int page_increm
;
1912 * If we have a pending SIGKILL, don't keep faulting
1913 * pages and potentially allocating memory.
1915 if (unlikely(fatal_signal_pending(current
)))
1916 return i
? i
: -ERESTARTSYS
;
1919 while (!(page
= follow_page_mask(vma
, start
,
1920 foll_flags
, &page_mask
))) {
1922 unsigned int fault_flags
= 0;
1924 if (foll_flags
& FOLL_WRITE
)
1925 fault_flags
|= FAULT_FLAG_WRITE
;
1927 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
1928 if (foll_flags
& FOLL_NOWAIT
)
1929 fault_flags
|= (FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
);
1931 ret
= handle_mm_fault(mm
, vma
, start
,
1934 if (ret
& VM_FAULT_ERROR
) {
1935 if (ret
& VM_FAULT_OOM
)
1936 return i
? i
: -ENOMEM
;
1937 if (ret
& (VM_FAULT_HWPOISON
|
1938 VM_FAULT_HWPOISON_LARGE
)) {
1941 else if (gup_flags
& FOLL_HWPOISON
)
1946 if (ret
& (VM_FAULT_SIGBUS
|
1948 return i
? i
: -EFAULT
;
1953 if (ret
& VM_FAULT_MAJOR
)
1959 if (ret
& VM_FAULT_RETRY
) {
1966 * The VM_FAULT_WRITE bit tells us that
1967 * do_wp_page has broken COW when necessary,
1968 * even if maybe_mkwrite decided not to set
1969 * pte_write. We can thus safely do subsequent
1970 * page lookups as if they were reads. But only
1971 * do so when looping for pte_write is futile:
1972 * in some cases userspace may also be wanting
1973 * to write to the gotten user page, which a
1974 * read fault here might prevent (a readonly
1975 * page might get reCOWed by userspace write).
1977 if ((ret
& VM_FAULT_WRITE
) &&
1978 !(vma
->vm_flags
& VM_WRITE
))
1979 foll_flags
|= FOLL_COW
;
1984 return i
? i
: PTR_ERR(page
);
1988 flush_anon_page(vma
, page
, start
);
1989 flush_dcache_page(page
);
1997 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & page_mask
);
1998 if (page_increm
> nr_pages
)
1999 page_increm
= nr_pages
;
2001 start
+= page_increm
* PAGE_SIZE
;
2002 nr_pages
-= page_increm
;
2003 } while (nr_pages
&& start
< vma
->vm_end
);
2007 EXPORT_SYMBOL(__get_user_pages
);
2010 * fixup_user_fault() - manually resolve a user page fault
2011 * @tsk: the task_struct to use for page fault accounting, or
2012 * NULL if faults are not to be recorded.
2013 * @mm: mm_struct of target mm
2014 * @address: user address
2015 * @fault_flags:flags to pass down to handle_mm_fault()
2017 * This is meant to be called in the specific scenario where for locking reasons
2018 * we try to access user memory in atomic context (within a pagefault_disable()
2019 * section), this returns -EFAULT, and we want to resolve the user fault before
2022 * Typically this is meant to be used by the futex code.
2024 * The main difference with get_user_pages() is that this function will
2025 * unconditionally call handle_mm_fault() which will in turn perform all the
2026 * necessary SW fixup of the dirty and young bits in the PTE, while
2027 * handle_mm_fault() only guarantees to update these in the struct page.
2029 * This is important for some architectures where those bits also gate the
2030 * access permission to the page because they are maintained in software. On
2031 * such architectures, gup() will not be enough to make a subsequent access
2034 * This should be called with the mm_sem held for read.
2036 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
2037 unsigned long address
, unsigned int fault_flags
)
2039 struct vm_area_struct
*vma
;
2040 vm_flags_t vm_flags
;
2043 vma
= find_extend_vma(mm
, address
);
2044 if (!vma
|| address
< vma
->vm_start
)
2047 vm_flags
= (fault_flags
& FAULT_FLAG_WRITE
) ? VM_WRITE
: VM_READ
;
2048 if (!(vm_flags
& vma
->vm_flags
))
2051 ret
= handle_mm_fault(mm
, vma
, address
, fault_flags
);
2052 if (ret
& VM_FAULT_ERROR
) {
2053 if (ret
& VM_FAULT_OOM
)
2055 if (ret
& (VM_FAULT_HWPOISON
| VM_FAULT_HWPOISON_LARGE
))
2057 if (ret
& (VM_FAULT_SIGBUS
| VM_FAULT_SIGSEGV
))
2062 if (ret
& VM_FAULT_MAJOR
)
2071 * get_user_pages() - pin user pages in memory
2072 * @tsk: the task_struct to use for page fault accounting, or
2073 * NULL if faults are not to be recorded.
2074 * @mm: mm_struct of target mm
2075 * @start: starting user address
2076 * @nr_pages: number of pages from start to pin
2077 * @write: whether pages will be written to by the caller
2078 * @force: whether to force write access even if user mapping is
2079 * readonly. This will result in the page being COWed even
2080 * in MAP_SHARED mappings. You do not want this.
2081 * @pages: array that receives pointers to the pages pinned.
2082 * Should be at least nr_pages long. Or NULL, if caller
2083 * only intends to ensure the pages are faulted in.
2084 * @vmas: array of pointers to vmas corresponding to each page.
2085 * Or NULL if the caller does not require them.
2087 * Returns number of pages pinned. This may be fewer than the number
2088 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2089 * were pinned, returns -errno. Each page returned must be released
2090 * with a put_page() call when it is finished with. vmas will only
2091 * remain valid while mmap_sem is held.
2093 * Must be called with mmap_sem held for read or write.
2095 * get_user_pages walks a process's page tables and takes a reference to
2096 * each struct page that each user address corresponds to at a given
2097 * instant. That is, it takes the page that would be accessed if a user
2098 * thread accesses the given user virtual address at that instant.
2100 * This does not guarantee that the page exists in the user mappings when
2101 * get_user_pages returns, and there may even be a completely different
2102 * page there in some cases (eg. if mmapped pagecache has been invalidated
2103 * and subsequently re faulted). However it does guarantee that the page
2104 * won't be freed completely. And mostly callers simply care that the page
2105 * contains data that was valid *at some point in time*. Typically, an IO
2106 * or similar operation cannot guarantee anything stronger anyway because
2107 * locks can't be held over the syscall boundary.
2109 * If write=0, the page must not be written to. If the page is written to,
2110 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
2111 * after the page is finished with, and before put_page is called.
2113 * get_user_pages is typically used for fewer-copy IO operations, to get a
2114 * handle on the memory by some means other than accesses via the user virtual
2115 * addresses. The pages may be submitted for DMA to devices or accessed via
2116 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2117 * use the correct cache flushing APIs.
2119 * See also get_user_pages_fast, for performance critical applications.
2121 long get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
2122 unsigned long start
, unsigned long nr_pages
, int write
,
2123 int force
, struct page
**pages
, struct vm_area_struct
**vmas
)
2125 int flags
= FOLL_TOUCH
;
2130 flags
|= FOLL_WRITE
;
2132 flags
|= FOLL_FORCE
;
2134 return __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
, vmas
,
2137 EXPORT_SYMBOL(get_user_pages
);
2140 * get_dump_page() - pin user page in memory while writing it to core dump
2141 * @addr: user address
2143 * Returns struct page pointer of user page pinned for dump,
2144 * to be freed afterwards by page_cache_release() or put_page().
2146 * Returns NULL on any kind of failure - a hole must then be inserted into
2147 * the corefile, to preserve alignment with its headers; and also returns
2148 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2149 * allowing a hole to be left in the corefile to save diskspace.
2151 * Called without mmap_sem, but after all other threads have been killed.
2153 #ifdef CONFIG_ELF_CORE
2154 struct page
*get_dump_page(unsigned long addr
)
2156 struct vm_area_struct
*vma
;
2159 if (__get_user_pages(current
, current
->mm
, addr
, 1,
2160 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
2163 flush_cache_page(vma
, addr
, page_to_pfn(page
));
2166 #endif /* CONFIG_ELF_CORE */
2168 pte_t
*__get_locked_pte(struct mm_struct
*mm
, unsigned long addr
,
2171 pgd_t
* pgd
= pgd_offset(mm
, addr
);
2172 pud_t
* pud
= pud_alloc(mm
, pgd
, addr
);
2174 pmd_t
* pmd
= pmd_alloc(mm
, pud
, addr
);
2176 VM_BUG_ON(pmd_trans_huge(*pmd
));
2177 return pte_alloc_map_lock(mm
, pmd
, addr
, ptl
);
2184 * This is the old fallback for page remapping.
2186 * For historical reasons, it only allows reserved pages. Only
2187 * old drivers should use this, and they needed to mark their
2188 * pages reserved for the old functions anyway.
2190 static int insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2191 struct page
*page
, pgprot_t prot
)
2193 struct mm_struct
*mm
= vma
->vm_mm
;
2202 flush_dcache_page(page
);
2203 pte
= get_locked_pte(mm
, addr
, &ptl
);
2207 if (!pte_none(*pte
))
2210 /* Ok, finally just insert the thing.. */
2212 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
2213 page_add_file_rmap(page
);
2214 set_pte_at(mm
, addr
, pte
, mk_pte(page
, prot
));
2217 pte_unmap_unlock(pte
, ptl
);
2220 pte_unmap_unlock(pte
, ptl
);
2226 * vm_insert_page - insert single page into user vma
2227 * @vma: user vma to map to
2228 * @addr: target user address of this page
2229 * @page: source kernel page
2231 * This allows drivers to insert individual pages they've allocated
2234 * The page has to be a nice clean _individual_ kernel allocation.
2235 * If you allocate a compound page, you need to have marked it as
2236 * such (__GFP_COMP), or manually just split the page up yourself
2237 * (see split_page()).
2239 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2240 * took an arbitrary page protection parameter. This doesn't allow
2241 * that. Your vma protection will have to be set up correctly, which
2242 * means that if you want a shared writable mapping, you'd better
2243 * ask for a shared writable mapping!
2245 * The page does not need to be reserved.
2247 * Usually this function is called from f_op->mmap() handler
2248 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2249 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2250 * function from other places, for example from page-fault handler.
2252 int vm_insert_page(struct vm_area_struct
*vma
, unsigned long addr
,
2255 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2257 if (!page_count(page
))
2259 if (!(vma
->vm_flags
& VM_MIXEDMAP
)) {
2260 BUG_ON(down_read_trylock(&vma
->vm_mm
->mmap_sem
));
2261 BUG_ON(vma
->vm_flags
& VM_PFNMAP
);
2262 vma
->vm_flags
|= VM_MIXEDMAP
;
2264 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2266 EXPORT_SYMBOL(vm_insert_page
);
2268 static int insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2269 unsigned long pfn
, pgprot_t prot
)
2271 struct mm_struct
*mm
= vma
->vm_mm
;
2277 pte
= get_locked_pte(mm
, addr
, &ptl
);
2281 if (!pte_none(*pte
))
2284 /* Ok, finally just insert the thing.. */
2285 entry
= pte_mkspecial(pfn_pte(pfn
, prot
));
2286 set_pte_at(mm
, addr
, pte
, entry
);
2287 update_mmu_cache(vma
, addr
, pte
); /* XXX: why not for insert_page? */
2291 pte_unmap_unlock(pte
, ptl
);
2297 * vm_insert_pfn - insert single pfn into user vma
2298 * @vma: user vma to map to
2299 * @addr: target user address of this page
2300 * @pfn: source kernel pfn
2302 * Similar to vm_insert_page, this allows drivers to insert individual pages
2303 * they've allocated into a user vma. Same comments apply.
2305 * This function should only be called from a vm_ops->fault handler, and
2306 * in that case the handler should return NULL.
2308 * vma cannot be a COW mapping.
2310 * As this is called only for pages that do not currently exist, we
2311 * do not need to flush old virtual caches or the TLB.
2313 int vm_insert_pfn(struct vm_area_struct
*vma
, unsigned long addr
,
2317 pgprot_t pgprot
= vma
->vm_page_prot
;
2319 * Technically, architectures with pte_special can avoid all these
2320 * restrictions (same for remap_pfn_range). However we would like
2321 * consistency in testing and feature parity among all, so we should
2322 * try to keep these invariants in place for everybody.
2324 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
2325 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
2326 (VM_PFNMAP
|VM_MIXEDMAP
));
2327 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
2328 BUG_ON((vma
->vm_flags
& VM_MIXEDMAP
) && pfn_valid(pfn
));
2330 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2332 if (track_pfn_insert(vma
, &pgprot
, pfn
))
2335 ret
= insert_pfn(vma
, addr
, pfn
, pgprot
);
2339 EXPORT_SYMBOL(vm_insert_pfn
);
2341 int vm_insert_mixed(struct vm_area_struct
*vma
, unsigned long addr
,
2344 BUG_ON(!(vma
->vm_flags
& VM_MIXEDMAP
));
2346 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2350 * If we don't have pte special, then we have to use the pfn_valid()
2351 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2352 * refcount the page if pfn_valid is true (hence insert_page rather
2353 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2354 * without pte special, it would there be refcounted as a normal page.
2356 if (!HAVE_PTE_SPECIAL
&& pfn_valid(pfn
)) {
2359 page
= pfn_to_page(pfn
);
2360 return insert_page(vma
, addr
, page
, vma
->vm_page_prot
);
2362 return insert_pfn(vma
, addr
, pfn
, vma
->vm_page_prot
);
2364 EXPORT_SYMBOL(vm_insert_mixed
);
2367 * maps a range of physical memory into the requested pages. the old
2368 * mappings are removed. any references to nonexistent pages results
2369 * in null mappings (currently treated as "copy-on-access")
2371 static int remap_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2372 unsigned long addr
, unsigned long end
,
2373 unsigned long pfn
, pgprot_t prot
)
2378 pte
= pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2381 arch_enter_lazy_mmu_mode();
2383 BUG_ON(!pte_none(*pte
));
2384 set_pte_at(mm
, addr
, pte
, pte_mkspecial(pfn_pte(pfn
, prot
)));
2386 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
2387 arch_leave_lazy_mmu_mode();
2388 pte_unmap_unlock(pte
- 1, ptl
);
2392 static inline int remap_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2393 unsigned long addr
, unsigned long end
,
2394 unsigned long pfn
, pgprot_t prot
)
2399 pfn
-= addr
>> PAGE_SHIFT
;
2400 pmd
= pmd_alloc(mm
, pud
, addr
);
2403 VM_BUG_ON(pmd_trans_huge(*pmd
));
2405 next
= pmd_addr_end(addr
, end
);
2406 if (remap_pte_range(mm
, pmd
, addr
, next
,
2407 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
2409 } while (pmd
++, addr
= next
, addr
!= end
);
2413 static inline int remap_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2414 unsigned long addr
, unsigned long end
,
2415 unsigned long pfn
, pgprot_t prot
)
2420 pfn
-= addr
>> PAGE_SHIFT
;
2421 pud
= pud_alloc(mm
, pgd
, addr
);
2425 next
= pud_addr_end(addr
, end
);
2426 if (remap_pmd_range(mm
, pud
, addr
, next
,
2427 pfn
+ (addr
>> PAGE_SHIFT
), prot
))
2429 } while (pud
++, addr
= next
, addr
!= end
);
2434 * remap_pfn_range - remap kernel memory to userspace
2435 * @vma: user vma to map to
2436 * @addr: target user address to start at
2437 * @pfn: physical address of kernel memory
2438 * @size: size of map area
2439 * @prot: page protection flags for this mapping
2441 * Note: this is only safe if the mm semaphore is held when called.
2443 int remap_pfn_range(struct vm_area_struct
*vma
, unsigned long addr
,
2444 unsigned long pfn
, unsigned long size
, pgprot_t prot
)
2448 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2449 struct mm_struct
*mm
= vma
->vm_mm
;
2453 * Physically remapped pages are special. Tell the
2454 * rest of the world about it:
2455 * VM_IO tells people not to look at these pages
2456 * (accesses can have side effects).
2457 * VM_PFNMAP tells the core MM that the base pages are just
2458 * raw PFN mappings, and do not have a "struct page" associated
2461 * Disable vma merging and expanding with mremap().
2463 * Omit vma from core dump, even when VM_IO turned off.
2465 * There's a horrible special case to handle copy-on-write
2466 * behaviour that some programs depend on. We mark the "original"
2467 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2468 * See vm_normal_page() for details.
2470 if (is_cow_mapping(vma
->vm_flags
)) {
2471 if (addr
!= vma
->vm_start
|| end
!= vma
->vm_end
)
2473 vma
->vm_pgoff
= pfn
;
2476 err
= track_pfn_remap(vma
, &prot
, pfn
, addr
, PAGE_ALIGN(size
));
2480 vma
->vm_flags
|= VM_IO
| VM_PFNMAP
| VM_DONTEXPAND
| VM_DONTDUMP
;
2482 BUG_ON(addr
>= end
);
2483 pfn
-= addr
>> PAGE_SHIFT
;
2484 pgd
= pgd_offset(mm
, addr
);
2485 flush_cache_range(vma
, addr
, end
);
2487 next
= pgd_addr_end(addr
, end
);
2488 err
= remap_pud_range(mm
, pgd
, addr
, next
,
2489 pfn
+ (addr
>> PAGE_SHIFT
), prot
);
2492 } while (pgd
++, addr
= next
, addr
!= end
);
2495 untrack_pfn(vma
, pfn
, PAGE_ALIGN(size
));
2499 EXPORT_SYMBOL(remap_pfn_range
);
2502 * vm_iomap_memory - remap memory to userspace
2503 * @vma: user vma to map to
2504 * @start: start of area
2505 * @len: size of area
2507 * This is a simplified io_remap_pfn_range() for common driver use. The
2508 * driver just needs to give us the physical memory range to be mapped,
2509 * we'll figure out the rest from the vma information.
2511 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2512 * whatever write-combining details or similar.
2514 int vm_iomap_memory(struct vm_area_struct
*vma
, phys_addr_t start
, unsigned long len
)
2516 unsigned long vm_len
, pfn
, pages
;
2518 /* Check that the physical memory area passed in looks valid */
2519 if (start
+ len
< start
)
2522 * You *really* shouldn't map things that aren't page-aligned,
2523 * but we've historically allowed it because IO memory might
2524 * just have smaller alignment.
2526 len
+= start
& ~PAGE_MASK
;
2527 pfn
= start
>> PAGE_SHIFT
;
2528 pages
= (len
+ ~PAGE_MASK
) >> PAGE_SHIFT
;
2529 if (pfn
+ pages
< pfn
)
2532 /* We start the mapping 'vm_pgoff' pages into the area */
2533 if (vma
->vm_pgoff
> pages
)
2535 pfn
+= vma
->vm_pgoff
;
2536 pages
-= vma
->vm_pgoff
;
2538 /* Can we fit all of the mapping? */
2539 vm_len
= vma
->vm_end
- vma
->vm_start
;
2540 if (vm_len
>> PAGE_SHIFT
> pages
)
2543 /* Ok, let it rip */
2544 return io_remap_pfn_range(vma
, vma
->vm_start
, pfn
, vm_len
, vma
->vm_page_prot
);
2546 EXPORT_SYMBOL(vm_iomap_memory
);
2548 static int apply_to_pte_range(struct mm_struct
*mm
, pmd_t
*pmd
,
2549 unsigned long addr
, unsigned long end
,
2550 pte_fn_t fn
, void *data
)
2555 spinlock_t
*uninitialized_var(ptl
);
2557 pte
= (mm
== &init_mm
) ?
2558 pte_alloc_kernel(pmd
, addr
) :
2559 pte_alloc_map_lock(mm
, pmd
, addr
, &ptl
);
2563 BUG_ON(pmd_huge(*pmd
));
2565 arch_enter_lazy_mmu_mode();
2567 token
= pmd_pgtable(*pmd
);
2570 err
= fn(pte
++, token
, addr
, data
);
2573 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2575 arch_leave_lazy_mmu_mode();
2578 pte_unmap_unlock(pte
-1, ptl
);
2582 static int apply_to_pmd_range(struct mm_struct
*mm
, pud_t
*pud
,
2583 unsigned long addr
, unsigned long end
,
2584 pte_fn_t fn
, void *data
)
2590 BUG_ON(pud_huge(*pud
));
2592 pmd
= pmd_alloc(mm
, pud
, addr
);
2596 next
= pmd_addr_end(addr
, end
);
2597 err
= apply_to_pte_range(mm
, pmd
, addr
, next
, fn
, data
);
2600 } while (pmd
++, addr
= next
, addr
!= end
);
2604 static int apply_to_pud_range(struct mm_struct
*mm
, pgd_t
*pgd
,
2605 unsigned long addr
, unsigned long end
,
2606 pte_fn_t fn
, void *data
)
2612 pud
= pud_alloc(mm
, pgd
, addr
);
2616 next
= pud_addr_end(addr
, end
);
2617 err
= apply_to_pmd_range(mm
, pud
, addr
, next
, fn
, data
);
2620 } while (pud
++, addr
= next
, addr
!= end
);
2625 * Scan a region of virtual memory, filling in page tables as necessary
2626 * and calling a provided function on each leaf page table.
2628 int apply_to_page_range(struct mm_struct
*mm
, unsigned long addr
,
2629 unsigned long size
, pte_fn_t fn
, void *data
)
2633 unsigned long end
= addr
+ size
;
2636 BUG_ON(addr
>= end
);
2637 pgd
= pgd_offset(mm
, addr
);
2639 next
= pgd_addr_end(addr
, end
);
2640 err
= apply_to_pud_range(mm
, pgd
, addr
, next
, fn
, data
);
2643 } while (pgd
++, addr
= next
, addr
!= end
);
2647 EXPORT_SYMBOL_GPL(apply_to_page_range
);
2650 * handle_pte_fault chooses page fault handler according to an entry
2651 * which was read non-atomically. Before making any commitment, on
2652 * those architectures or configurations (e.g. i386 with PAE) which
2653 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2654 * must check under lock before unmapping the pte and proceeding
2655 * (but do_wp_page is only called after already making such a check;
2656 * and do_anonymous_page can safely check later on).
2658 static inline int pte_unmap_same(struct mm_struct
*mm
, pmd_t
*pmd
,
2659 pte_t
*page_table
, pte_t orig_pte
)
2662 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2663 if (sizeof(pte_t
) > sizeof(unsigned long)) {
2664 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
2666 same
= pte_same(*page_table
, orig_pte
);
2670 pte_unmap(page_table
);
2674 static inline void cow_user_page(struct page
*dst
, struct page
*src
, unsigned long va
, struct vm_area_struct
*vma
)
2677 * If the source page was a PFN mapping, we don't have
2678 * a "struct page" for it. We do a best-effort copy by
2679 * just copying from the original user address. If that
2680 * fails, we just zero-fill it. Live with it.
2682 if (unlikely(!src
)) {
2683 void *kaddr
= kmap_atomic(dst
);
2684 void __user
*uaddr
= (void __user
*)(va
& PAGE_MASK
);
2687 * This really shouldn't fail, because the page is there
2688 * in the page tables. But it might just be unreadable,
2689 * in which case we just give up and fill the result with
2692 if (__copy_from_user_inatomic(kaddr
, uaddr
, PAGE_SIZE
))
2694 kunmap_atomic(kaddr
);
2695 flush_dcache_page(dst
);
2697 copy_user_highpage(dst
, src
, va
, vma
);
2701 * This routine handles present pages, when users try to write
2702 * to a shared page. It is done by copying the page to a new address
2703 * and decrementing the shared-page counter for the old page.
2705 * Note that this routine assumes that the protection checks have been
2706 * done by the caller (the low-level page fault routine in most cases).
2707 * Thus we can safely just mark it writable once we've done any necessary
2710 * We also mark the page dirty at this point even though the page will
2711 * change only once the write actually happens. This avoids a few races,
2712 * and potentially makes it more efficient.
2714 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2715 * but allow concurrent faults), with pte both mapped and locked.
2716 * We return with mmap_sem still held, but pte unmapped and unlocked.
2718 static int do_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2719 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
2720 spinlock_t
*ptl
, pte_t orig_pte
)
2723 struct page
*old_page
, *new_page
= NULL
;
2726 int page_mkwrite
= 0;
2727 struct page
*dirty_page
= NULL
;
2728 unsigned long mmun_start
= 0; /* For mmu_notifiers */
2729 unsigned long mmun_end
= 0; /* For mmu_notifiers */
2731 old_page
= vm_normal_page(vma
, address
, orig_pte
);
2734 * VM_MIXEDMAP !pfn_valid() case
2736 * We should not cow pages in a shared writeable mapping.
2737 * Just mark the pages writable as we can't do any dirty
2738 * accounting on raw pfn maps.
2740 if ((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2741 (VM_WRITE
|VM_SHARED
))
2747 * Take out anonymous pages first, anonymous shared vmas are
2748 * not dirty accountable.
2750 if (PageAnon(old_page
) && !PageKsm(old_page
)) {
2751 if (!trylock_page(old_page
)) {
2752 page_cache_get(old_page
);
2753 pte_unmap_unlock(page_table
, ptl
);
2754 lock_page(old_page
);
2755 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2757 if (!pte_same(*page_table
, orig_pte
)) {
2758 unlock_page(old_page
);
2761 page_cache_release(old_page
);
2763 if (reuse_swap_page(old_page
)) {
2765 * The page is all ours. Move it to our anon_vma so
2766 * the rmap code will not search our parent or siblings.
2767 * Protected against the rmap code by the page lock.
2769 page_move_anon_rmap(old_page
, vma
, address
);
2770 unlock_page(old_page
);
2773 unlock_page(old_page
);
2774 } else if (unlikely((vma
->vm_flags
& (VM_WRITE
|VM_SHARED
)) ==
2775 (VM_WRITE
|VM_SHARED
))) {
2777 * Only catch write-faults on shared writable pages,
2778 * read-only shared pages can get COWed by
2779 * get_user_pages(.write=1, .force=1).
2781 if (vma
->vm_ops
&& vma
->vm_ops
->page_mkwrite
) {
2782 struct vm_fault vmf
;
2785 vmf
.virtual_address
= (void __user
*)(address
&
2787 vmf
.pgoff
= old_page
->index
;
2788 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
2789 vmf
.page
= old_page
;
2792 * Notify the address space that the page is about to
2793 * become writable so that it can prohibit this or wait
2794 * for the page to get into an appropriate state.
2796 * We do this without the lock held, so that it can
2797 * sleep if it needs to.
2799 page_cache_get(old_page
);
2800 pte_unmap_unlock(page_table
, ptl
);
2802 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
2804 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
2806 goto unwritable_page
;
2808 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
2809 lock_page(old_page
);
2810 if (!old_page
->mapping
) {
2811 ret
= 0; /* retry the fault */
2812 unlock_page(old_page
);
2813 goto unwritable_page
;
2816 VM_BUG_ON(!PageLocked(old_page
));
2819 * Since we dropped the lock we need to revalidate
2820 * the PTE as someone else may have changed it. If
2821 * they did, we just return, as we can count on the
2822 * MMU to tell us if they didn't also make it writable.
2824 page_table
= pte_offset_map_lock(mm
, pmd
, address
,
2826 if (!pte_same(*page_table
, orig_pte
)) {
2827 unlock_page(old_page
);
2833 dirty_page
= old_page
;
2834 get_page(dirty_page
);
2837 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2838 entry
= pte_mkyoung(orig_pte
);
2839 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2840 if (ptep_set_access_flags(vma
, address
, page_table
, entry
,1))
2841 update_mmu_cache(vma
, address
, page_table
);
2842 pte_unmap_unlock(page_table
, ptl
);
2843 ret
|= VM_FAULT_WRITE
;
2849 * Yes, Virginia, this is actually required to prevent a race
2850 * with clear_page_dirty_for_io() from clearing the page dirty
2851 * bit after it clear all dirty ptes, but before a racing
2852 * do_wp_page installs a dirty pte.
2854 * __do_fault is protected similarly.
2856 if (!page_mkwrite
) {
2857 wait_on_page_locked(dirty_page
);
2858 set_page_dirty_balance(dirty_page
, page_mkwrite
);
2859 /* file_update_time outside page_lock */
2861 file_update_time(vma
->vm_file
);
2863 put_page(dirty_page
);
2865 struct address_space
*mapping
= dirty_page
->mapping
;
2867 set_page_dirty(dirty_page
);
2868 unlock_page(dirty_page
);
2869 page_cache_release(dirty_page
);
2872 * Some device drivers do not set page.mapping
2873 * but still dirty their pages
2875 balance_dirty_pages_ratelimited(mapping
);
2883 * Ok, we need to copy. Oh, well..
2885 page_cache_get(old_page
);
2887 pte_unmap_unlock(page_table
, ptl
);
2889 if (unlikely(anon_vma_prepare(vma
)))
2892 if (is_zero_pfn(pte_pfn(orig_pte
))) {
2893 new_page
= alloc_zeroed_user_highpage_movable(vma
, address
);
2897 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2900 cow_user_page(new_page
, old_page
, address
, vma
);
2902 __SetPageUptodate(new_page
);
2904 if (mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
))
2907 mmun_start
= address
& PAGE_MASK
;
2908 mmun_end
= mmun_start
+ PAGE_SIZE
;
2909 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2912 * Re-check the pte - we dropped the lock
2914 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2915 if (likely(pte_same(*page_table
, orig_pte
))) {
2917 if (!PageAnon(old_page
)) {
2918 dec_mm_counter_fast(mm
, MM_FILEPAGES
);
2919 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2922 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
2923 flush_cache_page(vma
, address
, pte_pfn(orig_pte
));
2924 entry
= mk_pte(new_page
, vma
->vm_page_prot
);
2925 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
2927 * Clear the pte entry and flush it first, before updating the
2928 * pte with the new entry. This will avoid a race condition
2929 * seen in the presence of one thread doing SMC and another
2932 ptep_clear_flush(vma
, address
, page_table
);
2933 page_add_new_anon_rmap(new_page
, vma
, address
);
2935 * We call the notify macro here because, when using secondary
2936 * mmu page tables (such as kvm shadow page tables), we want the
2937 * new page to be mapped directly into the secondary page table.
2939 set_pte_at_notify(mm
, address
, page_table
, entry
);
2940 update_mmu_cache(vma
, address
, page_table
);
2943 * Only after switching the pte to the new page may
2944 * we remove the mapcount here. Otherwise another
2945 * process may come and find the rmap count decremented
2946 * before the pte is switched to the new page, and
2947 * "reuse" the old page writing into it while our pte
2948 * here still points into it and can be read by other
2951 * The critical issue is to order this
2952 * page_remove_rmap with the ptp_clear_flush above.
2953 * Those stores are ordered by (if nothing else,)
2954 * the barrier present in the atomic_add_negative
2955 * in page_remove_rmap.
2957 * Then the TLB flush in ptep_clear_flush ensures that
2958 * no process can access the old page before the
2959 * decremented mapcount is visible. And the old page
2960 * cannot be reused until after the decremented
2961 * mapcount is visible. So transitively, TLBs to
2962 * old page will be flushed before it can be reused.
2964 page_remove_rmap(old_page
);
2967 /* Free the old page.. */
2968 new_page
= old_page
;
2969 ret
|= VM_FAULT_WRITE
;
2971 mem_cgroup_uncharge_page(new_page
);
2974 page_cache_release(new_page
);
2976 pte_unmap_unlock(page_table
, ptl
);
2977 if (mmun_end
> mmun_start
)
2978 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2981 * Don't let another task, with possibly unlocked vma,
2982 * keep the mlocked page.
2984 if ((ret
& VM_FAULT_WRITE
) && (vma
->vm_flags
& VM_LOCKED
)) {
2985 lock_page(old_page
); /* LRU manipulation */
2986 munlock_vma_page(old_page
);
2987 unlock_page(old_page
);
2989 page_cache_release(old_page
);
2993 page_cache_release(new_page
);
2996 page_cache_release(old_page
);
2997 return VM_FAULT_OOM
;
3000 page_cache_release(old_page
);
3004 static void unmap_mapping_range_vma(struct vm_area_struct
*vma
,
3005 unsigned long start_addr
, unsigned long end_addr
,
3006 struct zap_details
*details
)
3008 zap_page_range_single(vma
, start_addr
, end_addr
- start_addr
, details
);
3011 static inline void unmap_mapping_range_tree(struct rb_root
*root
,
3012 struct zap_details
*details
)
3014 struct vm_area_struct
*vma
;
3015 pgoff_t vba
, vea
, zba
, zea
;
3017 vma_interval_tree_foreach(vma
, root
,
3018 details
->first_index
, details
->last_index
) {
3020 vba
= vma
->vm_pgoff
;
3021 vea
= vba
+ ((vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
) - 1;
3022 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
3023 zba
= details
->first_index
;
3026 zea
= details
->last_index
;
3030 unmap_mapping_range_vma(vma
,
3031 ((zba
- vba
) << PAGE_SHIFT
) + vma
->vm_start
,
3032 ((zea
- vba
+ 1) << PAGE_SHIFT
) + vma
->vm_start
,
3037 static inline void unmap_mapping_range_list(struct list_head
*head
,
3038 struct zap_details
*details
)
3040 struct vm_area_struct
*vma
;
3043 * In nonlinear VMAs there is no correspondence between virtual address
3044 * offset and file offset. So we must perform an exhaustive search
3045 * across *all* the pages in each nonlinear VMA, not just the pages
3046 * whose virtual address lies outside the file truncation point.
3048 list_for_each_entry(vma
, head
, shared
.nonlinear
) {
3049 details
->nonlinear_vma
= vma
;
3050 unmap_mapping_range_vma(vma
, vma
->vm_start
, vma
->vm_end
, details
);
3055 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
3056 * @mapping: the address space containing mmaps to be unmapped.
3057 * @holebegin: byte in first page to unmap, relative to the start of
3058 * the underlying file. This will be rounded down to a PAGE_SIZE
3059 * boundary. Note that this is different from truncate_pagecache(), which
3060 * must keep the partial page. In contrast, we must get rid of
3062 * @holelen: size of prospective hole in bytes. This will be rounded
3063 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3065 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3066 * but 0 when invalidating pagecache, don't throw away private data.
3068 void unmap_mapping_range(struct address_space
*mapping
,
3069 loff_t
const holebegin
, loff_t
const holelen
, int even_cows
)
3071 struct zap_details details
;
3072 pgoff_t hba
= holebegin
>> PAGE_SHIFT
;
3073 pgoff_t hlen
= (holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3075 /* Check for overflow. */
3076 if (sizeof(holelen
) > sizeof(hlen
)) {
3078 (holebegin
+ holelen
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
3079 if (holeend
& ~(long long)ULONG_MAX
)
3080 hlen
= ULONG_MAX
- hba
+ 1;
3083 details
.check_mapping
= even_cows
? NULL
: mapping
;
3084 details
.nonlinear_vma
= NULL
;
3085 details
.first_index
= hba
;
3086 details
.last_index
= hba
+ hlen
- 1;
3087 if (details
.last_index
< details
.first_index
)
3088 details
.last_index
= ULONG_MAX
;
3091 mutex_lock(&mapping
->i_mmap_mutex
);
3092 if (unlikely(!RB_EMPTY_ROOT(&mapping
->i_mmap
)))
3093 unmap_mapping_range_tree(&mapping
->i_mmap
, &details
);
3094 if (unlikely(!list_empty(&mapping
->i_mmap_nonlinear
)))
3095 unmap_mapping_range_list(&mapping
->i_mmap_nonlinear
, &details
);
3096 mutex_unlock(&mapping
->i_mmap_mutex
);
3098 EXPORT_SYMBOL(unmap_mapping_range
);
3101 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3102 * but allow concurrent faults), and pte mapped but not yet locked.
3103 * We return with mmap_sem still held, but pte unmapped and unlocked.
3105 static int do_swap_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3106 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3107 unsigned int flags
, pte_t orig_pte
)
3110 struct page
*page
, *swapcache
;
3114 struct mem_cgroup
*ptr
;
3118 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
3121 entry
= pte_to_swp_entry(orig_pte
);
3122 if (unlikely(non_swap_entry(entry
))) {
3123 if (is_migration_entry(entry
)) {
3124 #ifdef CONFIG_DMA_CMA
3126 * FIXME: mszyprow: cruel, brute-force method for
3127 * letting cma/migration to finish it's job without
3128 * stealing the lock migration_entry_wait() and creating
3129 * a live-lock on the faulted page
3130 * (page->_count == 2 migration failure issue)
3134 migration_entry_wait(mm
, pmd
, address
);
3135 } else if (is_hwpoison_entry(entry
)) {
3136 ret
= VM_FAULT_HWPOISON
;
3138 print_bad_pte(vma
, address
, orig_pte
, NULL
);
3139 ret
= VM_FAULT_SIGBUS
;
3143 delayacct_set_flag(DELAYACCT_PF_SWAPIN
);
3144 page
= lookup_swap_cache(entry
);
3146 page
= swapin_readahead(entry
,
3147 GFP_HIGHUSER_MOVABLE
, vma
, address
);
3150 * Back out if somebody else faulted in this pte
3151 * while we released the pte lock.
3153 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3154 if (likely(pte_same(*page_table
, orig_pte
)))
3156 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3160 /* Had to read the page from swap area: Major fault */
3161 ret
= VM_FAULT_MAJOR
;
3162 count_vm_event(PGMAJFAULT
);
3163 mem_cgroup_count_vm_event(mm
, PGMAJFAULT
);
3164 } else if (PageHWPoison(page
)) {
3166 * hwpoisoned dirty swapcache pages are kept for killing
3167 * owner processes (which may be unknown at hwpoison time)
3169 ret
= VM_FAULT_HWPOISON
;
3170 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3176 locked
= lock_page_or_retry(page
, mm
, flags
);
3178 delayacct_clear_flag(DELAYACCT_PF_SWAPIN
);
3180 ret
|= VM_FAULT_RETRY
;
3185 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3186 * release the swapcache from under us. The page pin, and pte_same
3187 * test below, are not enough to exclude that. Even if it is still
3188 * swapcache, we need to check that the page's swap has not changed.
3190 if (unlikely(!PageSwapCache(page
) || page_private(page
) != entry
.val
))
3193 page
= ksm_might_need_to_copy(page
, vma
, address
);
3194 if (unlikely(!page
)) {
3200 if (mem_cgroup_try_charge_swapin(mm
, page
, GFP_KERNEL
, &ptr
)) {
3206 * Back out if somebody else already faulted in this pte.
3208 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3209 if (unlikely(!pte_same(*page_table
, orig_pte
)))
3212 if (unlikely(!PageUptodate(page
))) {
3213 ret
= VM_FAULT_SIGBUS
;
3218 * The page isn't present yet, go ahead with the fault.
3220 * Be careful about the sequence of operations here.
3221 * To get its accounting right, reuse_swap_page() must be called
3222 * while the page is counted on swap but not yet in mapcount i.e.
3223 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3224 * must be called after the swap_free(), or it will never succeed.
3225 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3226 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3227 * in page->private. In this case, a record in swap_cgroup is silently
3228 * discarded at swap_free().
3231 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3232 dec_mm_counter_fast(mm
, MM_SWAPENTS
);
3233 pte
= mk_pte(page
, vma
->vm_page_prot
);
3234 if ((flags
& FAULT_FLAG_WRITE
) && reuse_swap_page(page
)) {
3235 pte
= maybe_mkwrite(pte_mkdirty(pte
), vma
);
3236 flags
&= ~FAULT_FLAG_WRITE
;
3237 ret
|= VM_FAULT_WRITE
;
3240 flush_icache_page(vma
, page
);
3241 set_pte_at(mm
, address
, page_table
, pte
);
3242 if (page
== swapcache
)
3243 do_page_add_anon_rmap(page
, vma
, address
, exclusive
);
3244 else /* ksm created a completely new copy */
3245 page_add_new_anon_rmap(page
, vma
, address
);
3246 /* It's better to call commit-charge after rmap is established */
3247 mem_cgroup_commit_charge_swapin(page
, ptr
);
3250 if (vm_swap_full() || (vma
->vm_flags
& VM_LOCKED
) || PageMlocked(page
))
3251 try_to_free_swap(page
);
3253 if (page
!= swapcache
) {
3255 * Hold the lock to avoid the swap entry to be reused
3256 * until we take the PT lock for the pte_same() check
3257 * (to avoid false positives from pte_same). For
3258 * further safety release the lock after the swap_free
3259 * so that the swap count won't change under a
3260 * parallel locked swapcache.
3262 unlock_page(swapcache
);
3263 page_cache_release(swapcache
);
3266 if (flags
& FAULT_FLAG_WRITE
) {
3267 ret
|= do_wp_page(mm
, vma
, address
, page_table
, pmd
, ptl
, pte
);
3268 if (ret
& VM_FAULT_ERROR
)
3269 ret
&= VM_FAULT_ERROR
;
3273 /* No need to invalidate - it was non-present before */
3274 update_mmu_cache(vma
, address
, page_table
);
3276 pte_unmap_unlock(page_table
, ptl
);
3280 mem_cgroup_cancel_charge_swapin(ptr
);
3281 pte_unmap_unlock(page_table
, ptl
);
3285 page_cache_release(page
);
3286 if (page
!= swapcache
) {
3287 unlock_page(swapcache
);
3288 page_cache_release(swapcache
);
3294 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3295 * but allow concurrent faults), and pte mapped but not yet locked.
3296 * We return with mmap_sem still held, but pte unmapped and unlocked.
3298 static int do_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3299 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3306 pte_unmap(page_table
);
3308 /* File mapping without ->vm_ops ? */
3309 if (vma
->vm_flags
& VM_SHARED
)
3310 return VM_FAULT_SIGBUS
;
3312 /* Use the zero-page for reads */
3313 if (!(flags
& FAULT_FLAG_WRITE
)) {
3314 entry
= pte_mkspecial(pfn_pte(my_zero_pfn(address
),
3315 vma
->vm_page_prot
));
3316 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3317 if (!pte_none(*page_table
))
3322 /* Allocate our own private page. */
3323 if (unlikely(anon_vma_prepare(vma
)))
3325 page
= alloc_zeroed_user_highpage_movable(vma
, address
);
3329 * The memory barrier inside __SetPageUptodate makes sure that
3330 * preceeding stores to the page contents become visible before
3331 * the set_pte_at() write.
3333 __SetPageUptodate(page
);
3335 if (mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
))
3338 entry
= mk_pte(page
, vma
->vm_page_prot
);
3339 if (vma
->vm_flags
& VM_WRITE
)
3340 entry
= pte_mkwrite(pte_mkdirty(entry
));
3342 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3343 if (!pte_none(*page_table
))
3346 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3347 page_add_new_anon_rmap(page
, vma
, address
);
3349 set_pte_at(mm
, address
, page_table
, entry
);
3351 /* No need to invalidate - it was non-present before */
3352 update_mmu_cache(vma
, address
, page_table
);
3354 pte_unmap_unlock(page_table
, ptl
);
3357 mem_cgroup_uncharge_page(page
);
3358 page_cache_release(page
);
3361 page_cache_release(page
);
3363 return VM_FAULT_OOM
;
3367 * __do_fault() tries to create a new page mapping. It aggressively
3368 * tries to share with existing pages, but makes a separate copy if
3369 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3370 * the next page fault.
3372 * As this is called only for pages that do not currently exist, we
3373 * do not need to flush old virtual caches or the TLB.
3375 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3376 * but allow concurrent faults), and pte neither mapped nor locked.
3377 * We return with mmap_sem still held, but pte unmapped and unlocked.
3379 static int __do_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3380 unsigned long address
, pmd_t
*pmd
,
3381 pgoff_t pgoff
, unsigned int flags
, pte_t orig_pte
)
3386 struct page
*cow_page
;
3389 struct page
*dirty_page
= NULL
;
3390 struct vm_fault vmf
;
3392 int page_mkwrite
= 0;
3395 * If we do COW later, allocate page befor taking lock_page()
3396 * on the file cache page. This will reduce lock holding time.
3398 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3400 if (unlikely(anon_vma_prepare(vma
)))
3401 return VM_FAULT_OOM
;
3403 cow_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
3405 return VM_FAULT_OOM
;
3407 if (mem_cgroup_newpage_charge(cow_page
, mm
, GFP_KERNEL
)) {
3408 page_cache_release(cow_page
);
3409 return VM_FAULT_OOM
;
3414 vmf
.virtual_address
= (void __user
*)(address
& PAGE_MASK
);
3419 ret
= vma
->vm_ops
->fault(vma
, &vmf
);
3420 if (unlikely(ret
& (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
|
3424 if (unlikely(PageHWPoison(vmf
.page
))) {
3425 if (ret
& VM_FAULT_LOCKED
)
3426 unlock_page(vmf
.page
);
3427 ret
= VM_FAULT_HWPOISON
;
3432 * For consistency in subsequent calls, make the faulted page always
3435 if (unlikely(!(ret
& VM_FAULT_LOCKED
)))
3436 lock_page(vmf
.page
);
3438 VM_BUG_ON(!PageLocked(vmf
.page
));
3441 * Should we do an early C-O-W break?
3444 if (flags
& FAULT_FLAG_WRITE
) {
3445 if (!(vma
->vm_flags
& VM_SHARED
)) {
3448 copy_user_highpage(page
, vmf
.page
, address
, vma
);
3449 __SetPageUptodate(page
);
3452 * If the page will be shareable, see if the backing
3453 * address space wants to know that the page is about
3454 * to become writable
3456 if (vma
->vm_ops
->page_mkwrite
) {
3460 vmf
.flags
= FAULT_FLAG_WRITE
|FAULT_FLAG_MKWRITE
;
3461 tmp
= vma
->vm_ops
->page_mkwrite(vma
, &vmf
);
3463 (VM_FAULT_ERROR
| VM_FAULT_NOPAGE
))) {
3465 goto unwritable_page
;
3467 if (unlikely(!(tmp
& VM_FAULT_LOCKED
))) {
3469 if (!page
->mapping
) {
3470 ret
= 0; /* retry the fault */
3472 goto unwritable_page
;
3475 VM_BUG_ON(!PageLocked(page
));
3482 page_table
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
3485 * This silly early PAGE_DIRTY setting removes a race
3486 * due to the bad i386 page protection. But it's valid
3487 * for other architectures too.
3489 * Note that if FAULT_FLAG_WRITE is set, we either now have
3490 * an exclusive copy of the page, or this is a shared mapping,
3491 * so we can make it writable and dirty to avoid having to
3492 * handle that later.
3494 /* Only go through if we didn't race with anybody else... */
3495 if (likely(pte_same(*page_table
, orig_pte
))) {
3496 flush_icache_page(vma
, page
);
3497 entry
= mk_pte(page
, vma
->vm_page_prot
);
3498 if (flags
& FAULT_FLAG_WRITE
)
3499 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
3501 inc_mm_counter_fast(mm
, MM_ANONPAGES
);
3502 page_add_new_anon_rmap(page
, vma
, address
);
3504 inc_mm_counter_fast(mm
, MM_FILEPAGES
);
3505 page_add_file_rmap(page
);
3506 if (flags
& FAULT_FLAG_WRITE
) {
3508 get_page(dirty_page
);
3511 set_pte_at(mm
, address
, page_table
, entry
);
3513 /* no need to invalidate: a not-present page won't be cached */
3514 update_mmu_cache(vma
, address
, page_table
);
3517 mem_cgroup_uncharge_page(cow_page
);
3519 page_cache_release(page
);
3521 anon
= 1; /* no anon but release faulted_page */
3524 pte_unmap_unlock(page_table
, ptl
);
3527 struct address_space
*mapping
= page
->mapping
;
3530 if (set_page_dirty(dirty_page
))
3532 unlock_page(dirty_page
);
3533 put_page(dirty_page
);
3534 if ((dirtied
|| page_mkwrite
) && mapping
) {
3536 * Some device drivers do not set page.mapping but still
3539 balance_dirty_pages_ratelimited(mapping
);
3542 /* file_update_time outside page_lock */
3543 if (vma
->vm_file
&& !page_mkwrite
)
3544 file_update_time(vma
->vm_file
);
3546 unlock_page(vmf
.page
);
3548 page_cache_release(vmf
.page
);
3554 page_cache_release(page
);
3557 /* fs's fault handler get error */
3559 mem_cgroup_uncharge_page(cow_page
);
3560 page_cache_release(cow_page
);
3565 static int do_linear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3566 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3567 unsigned int flags
, pte_t orig_pte
)
3569 pgoff_t pgoff
= (((address
& PAGE_MASK
)
3570 - vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
3572 pte_unmap(page_table
);
3573 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3574 if (!vma
->vm_ops
->fault
)
3575 return VM_FAULT_SIGBUS
;
3576 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3580 * Fault of a previously existing named mapping. Repopulate the pte
3581 * from the encoded file_pte if possible. This enables swappable
3584 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3585 * but allow concurrent faults), and pte mapped but not yet locked.
3586 * We return with mmap_sem still held, but pte unmapped and unlocked.
3588 static int do_nonlinear_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3589 unsigned long address
, pte_t
*page_table
, pmd_t
*pmd
,
3590 unsigned int flags
, pte_t orig_pte
)
3594 flags
|= FAULT_FLAG_NONLINEAR
;
3596 if (!pte_unmap_same(mm
, pmd
, page_table
, orig_pte
))
3599 if (unlikely(!(vma
->vm_flags
& VM_NONLINEAR
))) {
3601 * Page table corrupted: show pte and kill process.
3603 print_bad_pte(vma
, address
, orig_pte
, NULL
);
3604 return VM_FAULT_SIGBUS
;
3607 pgoff
= pte_to_pgoff(orig_pte
);
3608 return __do_fault(mm
, vma
, address
, pmd
, pgoff
, flags
, orig_pte
);
3611 int numa_migrate_prep(struct page
*page
, struct vm_area_struct
*vma
,
3612 unsigned long addr
, int page_nid
)
3616 count_vm_numa_event(NUMA_HINT_FAULTS
);
3617 if (page_nid
== numa_node_id())
3618 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
3620 return mpol_misplaced(page
, vma
, addr
);
3623 int do_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3624 unsigned long addr
, pte_t pte
, pte_t
*ptep
, pmd_t
*pmd
)
3626 struct page
*page
= NULL
;
3630 bool migrated
= false;
3633 * The "pte" at this point cannot be used safely without
3634 * validation through pte_unmap_same(). It's of NUMA type but
3635 * the pfn may be screwed if the read is non atomic.
3637 * ptep_modify_prot_start is not called as this is clearing
3638 * the _PAGE_NUMA bit and it is not really expected that there
3639 * would be concurrent hardware modifications to the PTE.
3641 ptl
= pte_lockptr(mm
, pmd
);
3643 if (unlikely(!pte_same(*ptep
, pte
))) {
3644 pte_unmap_unlock(ptep
, ptl
);
3648 pte
= pte_mknonnuma(pte
);
3649 set_pte_at(mm
, addr
, ptep
, pte
);
3650 update_mmu_cache(vma
, addr
, ptep
);
3652 page
= vm_normal_page(vma
, addr
, pte
);
3654 pte_unmap_unlock(ptep
, ptl
);
3658 page_nid
= page_to_nid(page
);
3659 target_nid
= numa_migrate_prep(page
, vma
, addr
, page_nid
);
3660 pte_unmap_unlock(ptep
, ptl
);
3661 if (target_nid
== -1) {
3666 /* Migrate to the requested node */
3667 migrated
= migrate_misplaced_page(page
, target_nid
);
3669 page_nid
= target_nid
;
3673 task_numa_fault(page_nid
, 1, migrated
);
3677 /* NUMA hinting page fault entry point for regular pmds */
3678 #ifdef CONFIG_NUMA_BALANCING
3679 static int do_pmd_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3680 unsigned long addr
, pmd_t
*pmdp
)
3683 pte_t
*pte
, *orig_pte
;
3684 unsigned long _addr
= addr
& PMD_MASK
;
3685 unsigned long offset
;
3689 spin_lock(&mm
->page_table_lock
);
3691 if (pmd_numa(pmd
)) {
3692 set_pmd_at(mm
, _addr
, pmdp
, pmd_mknonnuma(pmd
));
3695 spin_unlock(&mm
->page_table_lock
);
3700 /* we're in a page fault so some vma must be in the range */
3702 BUG_ON(vma
->vm_start
>= _addr
+ PMD_SIZE
);
3703 offset
= max(_addr
, vma
->vm_start
) & ~PMD_MASK
;
3704 VM_BUG_ON(offset
>= PMD_SIZE
);
3705 orig_pte
= pte
= pte_offset_map_lock(mm
, pmdp
, _addr
, &ptl
);
3706 pte
+= offset
>> PAGE_SHIFT
;
3707 for (addr
= _addr
+ offset
; addr
< _addr
+ PMD_SIZE
; pte
++, addr
+= PAGE_SIZE
) {
3708 pte_t pteval
= *pte
;
3712 bool migrated
= false;
3714 if (!pte_present(pteval
))
3716 if (!pte_numa(pteval
))
3718 if (addr
>= vma
->vm_end
) {
3719 vma
= find_vma(mm
, addr
);
3720 /* there's a pte present so there must be a vma */
3722 BUG_ON(addr
< vma
->vm_start
);
3724 if (pte_numa(pteval
)) {
3725 pteval
= pte_mknonnuma(pteval
);
3726 set_pte_at(mm
, addr
, pte
, pteval
);
3728 page
= vm_normal_page(vma
, addr
, pteval
);
3729 if (unlikely(!page
))
3731 /* only check non-shared pages */
3732 if (unlikely(page_mapcount(page
) != 1))
3735 page_nid
= page_to_nid(page
);
3736 target_nid
= numa_migrate_prep(page
, vma
, addr
, page_nid
);
3737 pte_unmap_unlock(pte
, ptl
);
3738 if (target_nid
!= -1) {
3739 migrated
= migrate_misplaced_page(page
, target_nid
);
3741 page_nid
= target_nid
;
3747 task_numa_fault(page_nid
, 1, migrated
);
3749 pte
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
3751 pte_unmap_unlock(orig_pte
, ptl
);
3756 static int do_pmd_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3757 unsigned long addr
, pmd_t
*pmdp
)
3762 #endif /* CONFIG_NUMA_BALANCING */
3765 * These routines also need to handle stuff like marking pages dirty
3766 * and/or accessed for architectures that don't do it in hardware (most
3767 * RISC architectures). The early dirtying is also good on the i386.
3769 * There is also a hook called "update_mmu_cache()" that architectures
3770 * with external mmu caches can use to update those (ie the Sparc or
3771 * PowerPC hashed page tables that act as extended TLBs).
3773 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3774 * but allow concurrent faults), and pte mapped but not yet locked.
3775 * We return with mmap_sem still held, but pte unmapped and unlocked.
3777 int handle_pte_fault(struct mm_struct
*mm
,
3778 struct vm_area_struct
*vma
, unsigned long address
,
3779 pte_t
*pte
, pmd_t
*pmd
, unsigned int flags
)
3785 if (!pte_present(entry
)) {
3786 if (pte_none(entry
)) {
3788 return do_linear_fault(mm
, vma
, address
,
3789 pte
, pmd
, flags
, entry
);
3790 return do_anonymous_page(mm
, vma
, address
,
3793 if (pte_file(entry
))
3794 return do_nonlinear_fault(mm
, vma
, address
,
3795 pte
, pmd
, flags
, entry
);
3796 return do_swap_page(mm
, vma
, address
,
3797 pte
, pmd
, flags
, entry
);
3800 if (pte_numa(entry
))
3801 return do_numa_page(mm
, vma
, address
, entry
, pte
, pmd
);
3803 ptl
= pte_lockptr(mm
, pmd
);
3805 if (unlikely(!pte_same(*pte
, entry
)))
3807 if (flags
& FAULT_FLAG_WRITE
) {
3808 if (!pte_write(entry
))
3809 return do_wp_page(mm
, vma
, address
,
3810 pte
, pmd
, ptl
, entry
);
3811 entry
= pte_mkdirty(entry
);
3813 entry
= pte_mkyoung(entry
);
3814 if (ptep_set_access_flags(vma
, address
, pte
, entry
, flags
& FAULT_FLAG_WRITE
)) {
3815 update_mmu_cache(vma
, address
, pte
);
3818 * This is needed only for protection faults but the arch code
3819 * is not yet telling us if this is a protection fault or not.
3820 * This still avoids useless tlb flushes for .text page faults
3823 if (flags
& FAULT_FLAG_WRITE
)
3824 flush_tlb_fix_spurious_fault(vma
, address
);
3827 pte_unmap_unlock(pte
, ptl
);
3832 * By the time we get here, we already hold the mm semaphore
3834 static int __handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3835 unsigned long address
, unsigned int flags
)
3842 if (unlikely(is_vm_hugetlb_page(vma
)))
3843 return hugetlb_fault(mm
, vma
, address
, flags
);
3846 pgd
= pgd_offset(mm
, address
);
3847 pud
= pud_alloc(mm
, pgd
, address
);
3849 return VM_FAULT_OOM
;
3850 pmd
= pmd_alloc(mm
, pud
, address
);
3852 return VM_FAULT_OOM
;
3853 if (pmd_none(*pmd
) && transparent_hugepage_enabled(vma
)) {
3855 return do_huge_pmd_anonymous_page(mm
, vma
, address
,
3858 pmd_t orig_pmd
= *pmd
;
3862 if (pmd_trans_huge(orig_pmd
)) {
3863 unsigned int dirty
= flags
& FAULT_FLAG_WRITE
;
3866 * If the pmd is splitting, return and retry the
3867 * the fault. Alternative: wait until the split
3868 * is done, and goto retry.
3870 if (pmd_trans_splitting(orig_pmd
))
3873 if (pmd_numa(orig_pmd
))
3874 return do_huge_pmd_numa_page(mm
, vma
, address
,
3877 if (dirty
&& !pmd_write(orig_pmd
)) {
3878 ret
= do_huge_pmd_wp_page(mm
, vma
, address
, pmd
,
3881 * If COW results in an oom, the huge pmd will
3882 * have been split, so retry the fault on the
3883 * pte for a smaller charge.
3885 if (unlikely(ret
& VM_FAULT_OOM
))
3889 huge_pmd_set_accessed(mm
, vma
, address
, pmd
,
3898 return do_pmd_numa_page(mm
, vma
, address
, pmd
);
3901 * Use __pte_alloc instead of pte_alloc_map, because we can't
3902 * run pte_offset_map on the pmd, if an huge pmd could
3903 * materialize from under us from a different thread.
3905 if (unlikely(pmd_none(*pmd
)) &&
3906 unlikely(__pte_alloc(mm
, vma
, pmd
, address
)))
3907 return VM_FAULT_OOM
;
3909 * If a huge pmd materialized under us just retry later. Use
3910 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
3911 * didn't become pmd_trans_huge under us and then back to pmd_none, as
3912 * a result of MADV_DONTNEED running immediately after a huge pmd fault
3913 * in a different thread of this mm, in turn leading to a misleading
3914 * pmd_trans_huge() retval. All we have to ensure is that it is a
3915 * regular pmd that we can walk with pte_offset_map() and we can do that
3916 * through an atomic read in C, which is what pmd_trans_unstable()
3919 if (unlikely(pmd_trans_unstable(pmd
)))
3922 * A regular pmd is established and it can't morph into a huge pmd
3923 * from under us anymore at this point because we hold the mmap_sem
3924 * read mode and khugepaged takes it in write mode. So now it's
3925 * safe to run pte_offset_map().
3927 pte
= pte_offset_map(pmd
, address
);
3929 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
3932 int handle_mm_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3933 unsigned long address
, unsigned int flags
)
3937 __set_current_state(TASK_RUNNING
);
3939 count_vm_event(PGFAULT
);
3940 mem_cgroup_count_vm_event(mm
, PGFAULT
);
3942 /* do counter updates before entering really critical section. */
3943 check_sync_rss_stat(current
);
3946 * Enable the memcg OOM handling for faults triggered in user
3947 * space. Kernel faults are handled more gracefully.
3949 if (flags
& FAULT_FLAG_USER
)
3950 mem_cgroup_oom_enable();
3952 ret
= __handle_mm_fault(mm
, vma
, address
, flags
);
3954 if (flags
& FAULT_FLAG_USER
) {
3955 mem_cgroup_oom_disable();
3957 * The task may have entered a memcg OOM situation but
3958 * if the allocation error was handled gracefully (no
3959 * VM_FAULT_OOM), there is no need to kill anything.
3960 * Just clean up the OOM state peacefully.
3962 if (task_in_memcg_oom(current
) && !(ret
& VM_FAULT_OOM
))
3963 mem_cgroup_oom_synchronize(false);
3969 #ifndef __PAGETABLE_PUD_FOLDED
3971 * Allocate page upper directory.
3972 * We've already handled the fast-path in-line.
3974 int __pud_alloc(struct mm_struct
*mm
, pgd_t
*pgd
, unsigned long address
)
3976 pud_t
*new = pud_alloc_one(mm
, address
);
3980 smp_wmb(); /* See comment in __pte_alloc */
3982 spin_lock(&mm
->page_table_lock
);
3983 if (pgd_present(*pgd
)) /* Another has populated it */
3986 pgd_populate(mm
, pgd
, new);
3987 spin_unlock(&mm
->page_table_lock
);
3990 #endif /* __PAGETABLE_PUD_FOLDED */
3992 #ifndef __PAGETABLE_PMD_FOLDED
3994 * Allocate page middle directory.
3995 * We've already handled the fast-path in-line.
3997 int __pmd_alloc(struct mm_struct
*mm
, pud_t
*pud
, unsigned long address
)
3999 pmd_t
*new = pmd_alloc_one(mm
, address
);
4003 smp_wmb(); /* See comment in __pte_alloc */
4005 spin_lock(&mm
->page_table_lock
);
4006 #ifndef __ARCH_HAS_4LEVEL_HACK
4007 if (pud_present(*pud
)) /* Another has populated it */
4010 pud_populate(mm
, pud
, new);
4012 if (pgd_present(*pud
)) /* Another has populated it */
4015 pgd_populate(mm
, pud
, new);
4016 #endif /* __ARCH_HAS_4LEVEL_HACK */
4017 spin_unlock(&mm
->page_table_lock
);
4020 #endif /* __PAGETABLE_PMD_FOLDED */
4022 #if !defined(__HAVE_ARCH_GATE_AREA)
4024 #if defined(AT_SYSINFO_EHDR)
4025 static struct vm_area_struct gate_vma
;
4027 static int __init
gate_vma_init(void)
4029 gate_vma
.vm_mm
= NULL
;
4030 gate_vma
.vm_start
= FIXADDR_USER_START
;
4031 gate_vma
.vm_end
= FIXADDR_USER_END
;
4032 gate_vma
.vm_flags
= VM_READ
| VM_MAYREAD
| VM_EXEC
| VM_MAYEXEC
;
4033 gate_vma
.vm_page_prot
= __P101
;
4037 __initcall(gate_vma_init
);
4040 struct vm_area_struct
*get_gate_vma(struct mm_struct
*mm
)
4042 #ifdef AT_SYSINFO_EHDR
4049 int in_gate_area_no_mm(unsigned long addr
)
4051 #ifdef AT_SYSINFO_EHDR
4052 if ((addr
>= FIXADDR_USER_START
) && (addr
< FIXADDR_USER_END
))
4058 #endif /* __HAVE_ARCH_GATE_AREA */
4060 static int __follow_pte(struct mm_struct
*mm
, unsigned long address
,
4061 pte_t
**ptepp
, spinlock_t
**ptlp
)
4068 pgd
= pgd_offset(mm
, address
);
4069 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
4072 pud
= pud_offset(pgd
, address
);
4073 if (pud_none(*pud
) || unlikely(pud_bad(*pud
)))
4076 pmd
= pmd_offset(pud
, address
);
4077 VM_BUG_ON(pmd_trans_huge(*pmd
));
4078 if (pmd_none(*pmd
) || unlikely(pmd_bad(*pmd
)))
4081 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
4085 ptep
= pte_offset_map_lock(mm
, pmd
, address
, ptlp
);
4088 if (!pte_present(*ptep
))
4093 pte_unmap_unlock(ptep
, *ptlp
);
4098 static inline int follow_pte(struct mm_struct
*mm
, unsigned long address
,
4099 pte_t
**ptepp
, spinlock_t
**ptlp
)
4103 /* (void) is needed to make gcc happy */
4104 (void) __cond_lock(*ptlp
,
4105 !(res
= __follow_pte(mm
, address
, ptepp
, ptlp
)));
4110 * follow_pfn - look up PFN at a user virtual address
4111 * @vma: memory mapping
4112 * @address: user virtual address
4113 * @pfn: location to store found PFN
4115 * Only IO mappings and raw PFN mappings are allowed.
4117 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4119 int follow_pfn(struct vm_area_struct
*vma
, unsigned long address
,
4126 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4129 ret
= follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
);
4132 *pfn
= pte_pfn(*ptep
);
4133 pte_unmap_unlock(ptep
, ptl
);
4136 EXPORT_SYMBOL(follow_pfn
);
4138 #ifdef CONFIG_HAVE_IOREMAP_PROT
4139 int follow_phys(struct vm_area_struct
*vma
,
4140 unsigned long address
, unsigned int flags
,
4141 unsigned long *prot
, resource_size_t
*phys
)
4147 if (!(vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)))
4150 if (follow_pte(vma
->vm_mm
, address
, &ptep
, &ptl
))
4154 if ((flags
& FOLL_WRITE
) && !pte_write(pte
))
4157 *prot
= pgprot_val(pte_pgprot(pte
));
4158 *phys
= (resource_size_t
)pte_pfn(pte
) << PAGE_SHIFT
;
4162 pte_unmap_unlock(ptep
, ptl
);
4167 int generic_access_phys(struct vm_area_struct
*vma
, unsigned long addr
,
4168 void *buf
, int len
, int write
)
4170 resource_size_t phys_addr
;
4171 unsigned long prot
= 0;
4172 void __iomem
*maddr
;
4173 int offset
= addr
& (PAGE_SIZE
-1);
4175 if (follow_phys(vma
, addr
, write
, &prot
, &phys_addr
))
4178 maddr
= ioremap_prot(phys_addr
, PAGE_ALIGN(len
+ offset
), prot
);
4180 memcpy_toio(maddr
+ offset
, buf
, len
);
4182 memcpy_fromio(buf
, maddr
+ offset
, len
);
4187 EXPORT_SYMBOL_GPL(generic_access_phys
);
4191 * Access another process' address space as given in mm. If non-NULL, use the
4192 * given task for page fault accounting.
4194 static int __access_remote_vm(struct task_struct
*tsk
, struct mm_struct
*mm
,
4195 unsigned long addr
, void *buf
, int len
, int write
)
4197 struct vm_area_struct
*vma
;
4198 void *old_buf
= buf
;
4200 down_read(&mm
->mmap_sem
);
4201 /* ignore errors, just check how much was successfully transferred */
4203 int bytes
, ret
, offset
;
4205 struct page
*page
= NULL
;
4207 ret
= get_user_pages(tsk
, mm
, addr
, 1,
4208 write
, 1, &page
, &vma
);
4211 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4212 * we can access using slightly different code.
4214 #ifdef CONFIG_HAVE_IOREMAP_PROT
4215 vma
= find_vma(mm
, addr
);
4216 if (!vma
|| vma
->vm_start
> addr
)
4218 if (vma
->vm_ops
&& vma
->vm_ops
->access
)
4219 ret
= vma
->vm_ops
->access(vma
, addr
, buf
,
4227 offset
= addr
& (PAGE_SIZE
-1);
4228 if (bytes
> PAGE_SIZE
-offset
)
4229 bytes
= PAGE_SIZE
-offset
;
4233 copy_to_user_page(vma
, page
, addr
,
4234 maddr
+ offset
, buf
, bytes
);
4235 set_page_dirty_lock(page
);
4237 copy_from_user_page(vma
, page
, addr
,
4238 buf
, maddr
+ offset
, bytes
);
4241 page_cache_release(page
);
4247 up_read(&mm
->mmap_sem
);
4249 return buf
- old_buf
;
4253 * access_remote_vm - access another process' address space
4254 * @mm: the mm_struct of the target address space
4255 * @addr: start address to access
4256 * @buf: source or destination buffer
4257 * @len: number of bytes to transfer
4258 * @write: whether the access is a write
4260 * The caller must hold a reference on @mm.
4262 int access_remote_vm(struct mm_struct
*mm
, unsigned long addr
,
4263 void *buf
, int len
, int write
)
4265 return __access_remote_vm(NULL
, mm
, addr
, buf
, len
, write
);
4269 * Access another process' address space.
4270 * Source/target buffer must be kernel space,
4271 * Do not walk the page table directly, use get_user_pages
4273 int access_process_vm(struct task_struct
*tsk
, unsigned long addr
,
4274 void *buf
, int len
, int write
)
4276 struct mm_struct
*mm
;
4279 mm
= get_task_mm(tsk
);
4283 ret
= __access_remote_vm(tsk
, mm
, addr
, buf
, len
, write
);
4290 * Print the name of a VMA.
4292 void print_vma_addr(char *prefix
, unsigned long ip
)
4294 struct mm_struct
*mm
= current
->mm
;
4295 struct vm_area_struct
*vma
;
4298 * Do not print if we are in atomic
4299 * contexts (in exception stacks, etc.):
4301 if (preempt_count())
4304 down_read(&mm
->mmap_sem
);
4305 vma
= find_vma(mm
, ip
);
4306 if (vma
&& vma
->vm_file
) {
4307 struct file
*f
= vma
->vm_file
;
4308 char *buf
= (char *)__get_free_page(GFP_KERNEL
);
4312 p
= d_path(&f
->f_path
, buf
, PAGE_SIZE
);
4315 printk("%s%s[%lx+%lx]", prefix
, kbasename(p
),
4317 vma
->vm_end
- vma
->vm_start
);
4318 free_page((unsigned long)buf
);
4321 up_read(&mm
->mmap_sem
);
4324 #ifdef CONFIG_PROVE_LOCKING
4325 void might_fault(void)
4328 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4329 * holding the mmap_sem, this is safe because kernel memory doesn't
4330 * get paged out, therefore we'll never actually fault, and the
4331 * below annotations will generate false positives.
4333 if (segment_eq(get_fs(), KERNEL_DS
))
4338 * it would be nicer only to annotate paths which are not under
4339 * pagefault_disable, however that requires a larger audit and
4340 * providing helpers like get_user_atomic.
4342 if (!in_atomic() && current
->mm
)
4343 might_lock_read(¤t
->mm
->mmap_sem
);
4345 EXPORT_SYMBOL(might_fault
);
4348 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4349 static void clear_gigantic_page(struct page
*page
,
4351 unsigned int pages_per_huge_page
)
4354 struct page
*p
= page
;
4357 for (i
= 0; i
< pages_per_huge_page
;
4358 i
++, p
= mem_map_next(p
, page
, i
)) {
4360 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
4363 void clear_huge_page(struct page
*page
,
4364 unsigned long addr
, unsigned int pages_per_huge_page
)
4368 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4369 clear_gigantic_page(page
, addr
, pages_per_huge_page
);
4374 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4376 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
4380 static void copy_user_gigantic_page(struct page
*dst
, struct page
*src
,
4382 struct vm_area_struct
*vma
,
4383 unsigned int pages_per_huge_page
)
4386 struct page
*dst_base
= dst
;
4387 struct page
*src_base
= src
;
4389 for (i
= 0; i
< pages_per_huge_page
; ) {
4391 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
4394 dst
= mem_map_next(dst
, dst_base
, i
);
4395 src
= mem_map_next(src
, src_base
, i
);
4399 void copy_user_huge_page(struct page
*dst
, struct page
*src
,
4400 unsigned long addr
, struct vm_area_struct
*vma
,
4401 unsigned int pages_per_huge_page
)
4405 if (unlikely(pages_per_huge_page
> MAX_ORDER_NR_PAGES
)) {
4406 copy_user_gigantic_page(dst
, src
, addr
, vma
,
4407 pages_per_huge_page
);
4412 for (i
= 0; i
< pages_per_huge_page
; i
++) {
4414 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
4417 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */