2 * Lockless get_user_pages_fast for powerpc
4 * Copyright (C) 2008 Nick Piggin
5 * Copyright (C) 2008 Novell Inc.
9 #include <linux/sched.h>
11 #include <linux/hugetlb.h>
12 #include <linux/vmstat.h>
13 #include <linux/pagemap.h>
14 #include <linux/rwsem.h>
15 #include <asm/pgtable.h>
17 #ifdef __HAVE_ARCH_PTE_SPECIAL
19 static inline void get_huge_page_tail(struct page
*page
)
22 * __split_huge_page_refcount() cannot run
25 VM_BUG_ON(page_mapcount(page
) < 0);
26 VM_BUG_ON(atomic_read(&page
->_count
) != 0);
27 atomic_inc(&page
->_mapcount
);
31 * The performance critical leaf functions are made noinline otherwise gcc
32 * inlines everything into a single function which results in too much
35 static noinline
int gup_pte_range(pmd_t pmd
, unsigned long addr
,
36 unsigned long end
, int write
, struct page
**pages
, int *nr
)
38 unsigned long mask
, result
;
41 result
= _PAGE_PRESENT
|_PAGE_USER
;
44 mask
= result
| _PAGE_SPECIAL
;
46 ptep
= pte_offset_kernel(&pmd
, addr
);
51 if ((pte_val(pte
) & mask
) != result
)
53 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
55 if (!page_cache_get_speculative(page
))
57 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
62 get_huge_page_tail(page
);
66 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
71 static int gup_pmd_range(pud_t pud
, unsigned long addr
, unsigned long end
,
72 int write
, struct page
**pages
, int *nr
)
77 pmdp
= pmd_offset(&pud
, addr
);
81 next
= pmd_addr_end(addr
, end
);
84 if (is_hugepd(pmdp
)) {
85 if (!gup_hugepd((hugepd_t
*)pmdp
, PMD_SHIFT
,
86 addr
, next
, write
, pages
, nr
))
88 } else if (!gup_pte_range(pmd
, addr
, next
, write
, pages
, nr
))
90 } while (pmdp
++, addr
= next
, addr
!= end
);
95 static int gup_pud_range(pgd_t pgd
, unsigned long addr
, unsigned long end
,
96 int write
, struct page
**pages
, int *nr
)
101 pudp
= pud_offset(&pgd
, addr
);
105 next
= pud_addr_end(addr
, end
);
108 if (is_hugepd(pudp
)) {
109 if (!gup_hugepd((hugepd_t
*)pudp
, PUD_SHIFT
,
110 addr
, next
, write
, pages
, nr
))
112 } else if (!gup_pmd_range(pud
, addr
, next
, write
, pages
, nr
))
114 } while (pudp
++, addr
= next
, addr
!= end
);
119 int get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
122 struct mm_struct
*mm
= current
->mm
;
123 unsigned long addr
, len
, end
;
128 pr_devel("%s(%lx,%x,%s)\n", __func__
, start
, nr_pages
, write
? "write" : "read");
132 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
135 if (unlikely(!access_ok(write
? VERIFY_WRITE
: VERIFY_READ
,
139 pr_devel(" aligned: %lx .. %lx\n", start
, end
);
142 * XXX: batch / limit 'nr', to avoid large irq off latency
143 * needs some instrumenting to determine the common sizes used by
144 * important workloads (eg. DB2), and whether limiting the batch size
145 * will decrease performance.
147 * It seems like we're in the clear for the moment. Direct-IO is
148 * the main guy that batches up lots of get_user_pages, and even
149 * they are limited to 64-at-a-time which is not so many.
152 * This doesn't prevent pagetable teardown, but does prevent
153 * the pagetables from being freed on powerpc.
155 * So long as we atomically load page table pointers versus teardown,
156 * we can follow the address down to the the page and take a ref on it.
160 pgdp
= pgd_offset(mm
, addr
);
164 pr_devel(" %016lx: normal pgd %p\n", addr
,
165 (void *)pgd_val(pgd
));
166 next
= pgd_addr_end(addr
, end
);
169 if (is_hugepd(pgdp
)) {
170 if (!gup_hugepd((hugepd_t
*)pgdp
, PGDIR_SHIFT
,
171 addr
, next
, write
, pages
, &nr
))
173 } else if (!gup_pud_range(pgd
, addr
, next
, write
, pages
, &nr
))
175 } while (pgdp
++, addr
= next
, addr
!= end
);
179 VM_BUG_ON(nr
!= (end
- start
) >> PAGE_SHIFT
);
188 pr_devel(" slow path ! nr = %d\n", nr
);
190 /* Try to get the remaining pages with get_user_pages */
191 start
+= nr
<< PAGE_SHIFT
;
194 down_read(&mm
->mmap_sem
);
195 ret
= get_user_pages(current
, mm
, start
,
196 (end
- start
) >> PAGE_SHIFT
, write
, 0, pages
, NULL
);
197 up_read(&mm
->mmap_sem
);
199 /* Have to be a bit careful with return values */
211 #endif /* __HAVE_ARCH_PTE_SPECIAL */