2 * Lockless get_user_pages_fast for sparc, cribbed from powerpc
4 * Copyright (C) 2008 Nick Piggin
5 * Copyright (C) 2008 Novell Inc.
8 #include <linux/sched.h>
10 #include <linux/vmstat.h>
11 #include <linux/pagemap.h>
12 #include <linux/rwsem.h>
13 #include <asm/pgtable.h>
15 static inline void get_huge_page_tail(struct page
*page
)
18 * __split_huge_page_refcount() cannot run
21 VM_BUG_ON(page_mapcount(page
) < 0);
22 VM_BUG_ON(atomic_read(&page
->_count
) != 0);
23 atomic_inc(&page
->_mapcount
);
27 * The performance critical leaf functions are made noinline otherwise gcc
28 * inlines everything into a single function which results in too much
31 static noinline
int gup_pte_range(pmd_t pmd
, unsigned long addr
,
32 unsigned long end
, int write
, struct page
**pages
, int *nr
)
34 unsigned long mask
, result
;
37 if (tlb_type
== hypervisor
) {
38 result
= _PAGE_PRESENT_4V
|_PAGE_P_4V
;
40 result
|= _PAGE_WRITE_4V
;
42 result
= _PAGE_PRESENT_4U
|_PAGE_P_4U
;
44 result
|= _PAGE_WRITE_4U
;
46 mask
= result
| _PAGE_SPECIAL
;
48 ptep
= pte_offset_kernel(&pmd
, addr
);
50 struct page
*page
, *head
;
53 if ((pte_val(pte
) & mask
) != result
)
55 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
57 /* The hugepage case is simplified on sparc64 because
58 * we encode the sub-page pfn offsets into the
59 * hugepage PTEs. We could optimize this in the future
60 * use page_cache_add_speculative() for the hugepage case.
63 head
= compound_head(page
);
64 if (!page_cache_get_speculative(head
))
66 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
71 get_huge_page_tail(page
);
75 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
80 static int gup_pmd_range(pud_t pud
, unsigned long addr
, unsigned long end
,
81 int write
, struct page
**pages
, int *nr
)
86 pmdp
= pmd_offset(&pud
, addr
);
90 next
= pmd_addr_end(addr
, end
);
93 if (!gup_pte_range(pmd
, addr
, next
, write
, pages
, nr
))
95 } while (pmdp
++, addr
= next
, addr
!= end
);
100 static int gup_pud_range(pgd_t pgd
, unsigned long addr
, unsigned long end
,
101 int write
, struct page
**pages
, int *nr
)
106 pudp
= pud_offset(&pgd
, addr
);
110 next
= pud_addr_end(addr
, end
);
113 if (!gup_pmd_range(pud
, addr
, next
, write
, pages
, nr
))
115 } while (pudp
++, addr
= next
, addr
!= end
);
120 int get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
123 struct mm_struct
*mm
= current
->mm
;
124 unsigned long addr
, len
, end
;
131 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
135 * XXX: batch / limit 'nr', to avoid large irq off latency
136 * needs some instrumenting to determine the common sizes used by
137 * important workloads (eg. DB2), and whether limiting the batch size
138 * will decrease performance.
140 * It seems like we're in the clear for the moment. Direct-IO is
141 * the main guy that batches up lots of get_user_pages, and even
142 * they are limited to 64-at-a-time which is not so many.
145 * This doesn't prevent pagetable teardown, but does prevent
146 * the pagetables from being freed on sparc.
148 * So long as we atomically load page table pointers versus teardown,
149 * we can follow the address down to the the page and take a ref on it.
153 pgdp
= pgd_offset(mm
, addr
);
157 next
= pgd_addr_end(addr
, end
);
160 if (!gup_pud_range(pgd
, addr
, next
, write
, pages
, &nr
))
162 } while (pgdp
++, addr
= next
, addr
!= end
);
166 VM_BUG_ON(nr
!= (end
- start
) >> PAGE_SHIFT
);
175 /* Try to get the remaining pages with get_user_pages */
176 start
+= nr
<< PAGE_SHIFT
;
179 down_read(&mm
->mmap_sem
);
180 ret
= get_user_pages(current
, mm
, start
,
181 (end
- start
) >> PAGE_SHIFT
, write
, 0, pages
, NULL
);
182 up_read(&mm
->mmap_sem
);
184 /* Have to be a bit careful with return values */