2 * Lockless get_user_pages_fast for s390
4 * Copyright IBM Corp. 2010
5 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
7 #include <linux/sched.h>
9 #include <linux/hugetlb.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 inline int gup_pte_range(pmd_t
*pmdp
, pmd_t pmd
, unsigned long addr
,
21 unsigned long end
, int write
, struct page
**pages
, int *nr
)
27 mask
= (write
? _PAGE_PROTECT
: 0) | _PAGE_INVALID
| _PAGE_SPECIAL
;
29 ptep
= ((pte_t
*) pmd_deref(pmd
)) + pte_index(addr
);
33 /* Similar to the PMD case, NUMA hinting must take slow path */
34 if (pte_protnone(pte
))
36 if ((pte_val(pte
) & mask
) != 0)
38 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
40 if (!page_cache_get_speculative(page
))
42 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
49 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
54 static inline int gup_huge_pmd(pmd_t
*pmdp
, pmd_t pmd
, unsigned long addr
,
55 unsigned long end
, int write
, struct page
**pages
, int *nr
)
57 unsigned long mask
, result
;
58 struct page
*head
, *page
;
61 result
= write
? 0 : _SEGMENT_ENTRY_PROTECT
;
62 mask
= result
| _SEGMENT_ENTRY_INVALID
;
63 if ((pmd_val(pmd
) & mask
) != result
)
65 VM_BUG_ON(!pfn_valid(pmd_val(pmd
) >> PAGE_SHIFT
));
69 page
= head
+ ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
71 VM_BUG_ON(compound_head(page
) != head
);
76 } while (addr
+= PAGE_SIZE
, addr
!= end
);
78 if (!page_cache_add_speculative(head
, refs
)) {
83 if (unlikely(pmd_val(pmd
) != pmd_val(*pmdp
))) {
94 static inline int gup_pmd_range(pud_t
*pudp
, pud_t pud
, unsigned long addr
,
95 unsigned long end
, int write
, struct page
**pages
, int *nr
)
100 pmdp
= (pmd_t
*) pudp
;
101 if ((pud_val(pud
) & _REGION_ENTRY_TYPE_MASK
) == _REGION_ENTRY_TYPE_R3
)
102 pmdp
= (pmd_t
*) pud_deref(pud
);
103 pmdp
+= pmd_index(addr
);
107 next
= pmd_addr_end(addr
, end
);
110 if (unlikely(pmd_large(pmd
))) {
112 * NUMA hinting faults need to be handled in the GUP
113 * slowpath for accounting purposes and so that they
114 * can be serialised against THP migration.
116 if (pmd_protnone(pmd
))
118 if (!gup_huge_pmd(pmdp
, pmd
, addr
, next
,
121 } else if (!gup_pte_range(pmdp
, pmd
, addr
, next
,
124 } while (pmdp
++, addr
= next
, addr
!= end
);
129 static inline int gup_pud_range(pgd_t
*pgdp
, pgd_t pgd
, unsigned long addr
,
130 unsigned long end
, int write
, struct page
**pages
, int *nr
)
135 pudp
= (pud_t
*) pgdp
;
136 if ((pgd_val(pgd
) & _REGION_ENTRY_TYPE_MASK
) == _REGION_ENTRY_TYPE_R2
)
137 pudp
= (pud_t
*) pgd_deref(pgd
);
138 pudp
+= pud_index(addr
);
142 next
= pud_addr_end(addr
, end
);
145 if (!gup_pmd_range(pudp
, pud
, addr
, next
, write
, pages
, nr
))
147 } while (pudp
++, addr
= next
, addr
!= end
);
153 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
154 * back to the regular GUP.
156 int __get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
159 struct mm_struct
*mm
= current
->mm
;
160 unsigned long addr
, len
, end
;
161 unsigned long next
, flags
;
167 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
169 if ((end
<= start
) || (end
> TASK_SIZE
))
172 * local_irq_save() doesn't prevent pagetable teardown, but does
173 * prevent the pagetables from being freed on s390.
175 * So long as we atomically load page table pointers versus teardown,
176 * we can follow the address down to the the page and take a ref on it.
178 local_irq_save(flags
);
179 pgdp
= pgd_offset(mm
, addr
);
183 next
= pgd_addr_end(addr
, end
);
186 if (!gup_pud_range(pgdp
, pgd
, addr
, next
, write
, pages
, &nr
))
188 } while (pgdp
++, addr
= next
, addr
!= end
);
189 local_irq_restore(flags
);
195 * get_user_pages_fast() - pin user pages in memory
196 * @start: starting user address
197 * @nr_pages: number of pages from start to pin
198 * @write: whether pages will be written to
199 * @pages: array that receives pointers to the pages pinned.
200 * Should be at least nr_pages long.
202 * Attempt to pin user pages in memory without taking mm->mmap_sem.
203 * If not successful, it will fall back to taking the lock and
204 * calling get_user_pages().
206 * Returns number of pages pinned. This may be fewer than the number
207 * requested. If nr_pages is 0 or negative, returns 0. If no pages
208 * were pinned, returns -errno.
210 int get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
217 nr
= __get_user_pages_fast(start
, nr_pages
, write
, pages
);
221 /* Try to get the remaining pages with get_user_pages */
222 start
+= nr
<< PAGE_SHIFT
;
224 ret
= get_user_pages_unlocked(start
, nr_pages
- nr
, write
, 0, pages
);
225 /* Have to be a bit careful with return values */
227 ret
= (ret
< 0) ? nr
: ret
+ nr
;
This page took 0.034377 seconds and 5 git commands to generate.