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>
16 * The performance critical leaf functions are made noinline otherwise gcc
17 * inlines everything into a single function which results in too much
20 static noinline
int gup_pte_range(pmd_t pmd
, unsigned long addr
,
21 unsigned long end
, int write
, struct page
**pages
, int *nr
)
23 unsigned long mask
, result
;
26 if (tlb_type
== hypervisor
) {
27 result
= _PAGE_PRESENT_4V
|_PAGE_P_4V
;
29 result
|= _PAGE_WRITE_4V
;
31 result
= _PAGE_PRESENT_4U
|_PAGE_P_4U
;
33 result
|= _PAGE_WRITE_4U
;
35 mask
= result
| _PAGE_SPECIAL
;
37 ptep
= pte_offset_kernel(&pmd
, addr
);
39 struct page
*page
, *head
;
42 if ((pte_val(pte
) & mask
) != result
)
44 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
46 /* The hugepage case is simplified on sparc64 because
47 * we encode the sub-page pfn offsets into the
48 * hugepage PTEs. We could optimize this in the future
49 * use page_cache_add_speculative() for the hugepage case.
52 head
= compound_head(page
);
53 if (!page_cache_get_speculative(head
))
55 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
60 get_huge_page_tail(page
);
64 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
69 static int gup_huge_pmd(pmd_t
*pmdp
, pmd_t pmd
, unsigned long addr
,
70 unsigned long end
, int write
, struct page
**pages
,
73 struct page
*head
, *page
, *tail
;
79 if (write
&& !pmd_write(pmd
))
84 page
= head
+ ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
87 VM_BUG_ON(compound_head(page
) != head
);
92 } while (addr
+= PAGE_SIZE
, addr
!= end
);
94 if (!page_cache_add_speculative(head
, refs
)) {
99 if (unlikely(pmd_val(pmd
) != pmd_val(*pmdp
))) {
106 /* Any tail page need their mapcount reference taken before we
111 get_huge_page_tail(tail
);
118 static int gup_pmd_range(pud_t pud
, unsigned long addr
, unsigned long end
,
119 int write
, struct page
**pages
, int *nr
)
124 pmdp
= pmd_offset(&pud
, addr
);
128 next
= pmd_addr_end(addr
, end
);
129 if (pmd_none(pmd
) || pmd_trans_splitting(pmd
))
131 if (unlikely(pmd_large(pmd
))) {
132 if (!gup_huge_pmd(pmdp
, pmd
, addr
, next
,
135 } else if (!gup_pte_range(pmd
, addr
, next
, write
,
138 } while (pmdp
++, addr
= next
, addr
!= end
);
143 static int gup_pud_range(pgd_t pgd
, unsigned long addr
, unsigned long end
,
144 int write
, struct page
**pages
, int *nr
)
149 pudp
= pud_offset(&pgd
, addr
);
153 next
= pud_addr_end(addr
, end
);
156 if (!gup_pmd_range(pud
, addr
, next
, write
, pages
, nr
))
158 } while (pudp
++, addr
= next
, addr
!= end
);
163 int get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
166 struct mm_struct
*mm
= current
->mm
;
167 unsigned long addr
, len
, end
;
174 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
178 * XXX: batch / limit 'nr', to avoid large irq off latency
179 * needs some instrumenting to determine the common sizes used by
180 * important workloads (eg. DB2), and whether limiting the batch size
181 * will decrease performance.
183 * It seems like we're in the clear for the moment. Direct-IO is
184 * the main guy that batches up lots of get_user_pages, and even
185 * they are limited to 64-at-a-time which is not so many.
188 * This doesn't prevent pagetable teardown, but does prevent
189 * the pagetables from being freed on sparc.
191 * So long as we atomically load page table pointers versus teardown,
192 * we can follow the address down to the the page and take a ref on it.
196 pgdp
= pgd_offset(mm
, addr
);
200 next
= pgd_addr_end(addr
, end
);
203 if (!gup_pud_range(pgd
, addr
, next
, write
, pages
, &nr
))
205 } while (pgdp
++, addr
= next
, addr
!= end
);
209 VM_BUG_ON(nr
!= (end
- start
) >> PAGE_SHIFT
);
218 /* Try to get the remaining pages with get_user_pages */
219 start
+= nr
<< PAGE_SHIFT
;
222 down_read(&mm
->mmap_sem
);
223 ret
= get_user_pages(current
, mm
, start
,
224 (end
- start
) >> PAGE_SHIFT
, write
, 0, pages
, NULL
);
225 up_read(&mm
->mmap_sem
);
227 /* Have to be a bit careful with return values */