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
)
23 struct page
*head
, *page
;
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 head
= compound_head(page
);
41 if (!page_cache_get_speculative(head
))
43 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
47 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
51 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
56 static inline int gup_huge_pmd(pmd_t
*pmdp
, pmd_t pmd
, unsigned long addr
,
57 unsigned long end
, int write
, struct page
**pages
, int *nr
)
59 struct page
*head
, *page
;
63 mask
= (write
? _SEGMENT_ENTRY_PROTECT
: 0) | _SEGMENT_ENTRY_INVALID
;
64 if ((pmd_val(pmd
) & mask
) != 0)
66 VM_BUG_ON(!pfn_valid(pmd_val(pmd
) >> PAGE_SHIFT
));
70 page
= head
+ ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
72 VM_BUG_ON(compound_head(page
) != head
);
77 } while (addr
+= PAGE_SIZE
, addr
!= end
);
79 if (!page_cache_add_speculative(head
, refs
)) {
84 if (unlikely(pmd_val(pmd
) != pmd_val(*pmdp
))) {
95 static inline int gup_pmd_range(pud_t
*pudp
, pud_t pud
, unsigned long addr
,
96 unsigned long end
, int write
, struct page
**pages
, int *nr
)
101 pmdp
= (pmd_t
*) pudp
;
102 if ((pud_val(pud
) & _REGION_ENTRY_TYPE_MASK
) == _REGION_ENTRY_TYPE_R3
)
103 pmdp
= (pmd_t
*) pud_deref(pud
);
104 pmdp
+= pmd_index(addr
);
108 next
= pmd_addr_end(addr
, end
);
111 if (unlikely(pmd_large(pmd
))) {
113 * NUMA hinting faults need to be handled in the GUP
114 * slowpath for accounting purposes and so that they
115 * can be serialised against THP migration.
117 if (pmd_protnone(pmd
))
119 if (!gup_huge_pmd(pmdp
, pmd
, addr
, next
,
122 } else if (!gup_pte_range(pmdp
, pmd
, addr
, next
,
125 } while (pmdp
++, addr
= next
, addr
!= end
);
130 static int gup_huge_pud(pud_t
*pudp
, pud_t pud
, unsigned long addr
,
131 unsigned long end
, int write
, struct page
**pages
, int *nr
)
133 struct page
*head
, *page
;
137 mask
= (write
? _REGION_ENTRY_PROTECT
: 0) | _REGION_ENTRY_INVALID
;
138 if ((pud_val(pud
) & mask
) != 0)
140 VM_BUG_ON(!pfn_valid(pud_pfn(pud
)));
143 head
= pud_page(pud
);
144 page
= head
+ ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
146 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
151 } while (addr
+= PAGE_SIZE
, addr
!= end
);
153 if (!page_cache_add_speculative(head
, refs
)) {
158 if (unlikely(pud_val(pud
) != pud_val(*pudp
))) {
168 static inline int gup_pud_range(p4d_t
*p4dp
, p4d_t p4d
, unsigned long addr
,
169 unsigned long end
, int write
, struct page
**pages
, int *nr
)
174 pudp
= (pud_t
*) p4dp
;
175 if ((p4d_val(p4d
) & _REGION_ENTRY_TYPE_MASK
) == _REGION_ENTRY_TYPE_R2
)
176 pudp
= (pud_t
*) p4d_deref(p4d
);
177 pudp
+= pud_index(addr
);
181 next
= pud_addr_end(addr
, end
);
184 if (unlikely(pud_large(pud
))) {
185 if (!gup_huge_pud(pudp
, pud
, addr
, next
, write
, pages
,
188 } else if (!gup_pmd_range(pudp
, pud
, addr
, next
, write
, pages
,
191 } while (pudp
++, addr
= next
, addr
!= end
);
196 static inline int gup_p4d_range(pgd_t
*pgdp
, pgd_t pgd
, unsigned long addr
,
197 unsigned long end
, int write
, struct page
**pages
, int *nr
)
202 p4dp
= (p4d_t
*) pgdp
;
203 if ((pgd_val(pgd
) & _REGION_ENTRY_TYPE_MASK
) == _REGION_ENTRY_TYPE_R1
)
204 p4dp
= (p4d_t
*) pgd_deref(pgd
);
205 p4dp
+= p4d_index(addr
);
209 next
= p4d_addr_end(addr
, end
);
212 if (!gup_pud_range(p4dp
, p4d
, addr
, next
, write
, pages
, nr
))
214 } while (p4dp
++, addr
= next
, addr
!= end
);
220 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
221 * back to the regular GUP.
223 int __get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
226 struct mm_struct
*mm
= current
->mm
;
227 unsigned long addr
, len
, end
;
228 unsigned long next
, flags
;
234 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
236 if ((end
<= start
) || (end
> mm
->context
.asce_limit
))
239 * local_irq_save() doesn't prevent pagetable teardown, but does
240 * prevent the pagetables from being freed on s390.
242 * So long as we atomically load page table pointers versus teardown,
243 * we can follow the address down to the the page and take a ref on it.
245 local_irq_save(flags
);
246 pgdp
= pgd_offset(mm
, addr
);
250 next
= pgd_addr_end(addr
, end
);
253 if (!gup_p4d_range(pgdp
, pgd
, addr
, next
, write
, pages
, &nr
))
255 } while (pgdp
++, addr
= next
, addr
!= end
);
256 local_irq_restore(flags
);
262 * get_user_pages_fast() - pin user pages in memory
263 * @start: starting user address
264 * @nr_pages: number of pages from start to pin
265 * @write: whether pages will be written to
266 * @pages: array that receives pointers to the pages pinned.
267 * Should be at least nr_pages long.
269 * Attempt to pin user pages in memory without taking mm->mmap_sem.
270 * If not successful, it will fall back to taking the lock and
271 * calling get_user_pages().
273 * Returns number of pages pinned. This may be fewer than the number
274 * requested. If nr_pages is 0 or negative, returns 0. If no pages
275 * were pinned, returns -errno.
277 int get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
284 nr
= __get_user_pages_fast(start
, nr_pages
, write
, pages
);
288 /* Try to get the remaining pages with get_user_pages */
289 start
+= nr
<< PAGE_SHIFT
;
291 ret
= get_user_pages_unlocked(start
, nr_pages
- nr
, pages
,
292 write
? FOLL_WRITE
: 0);
293 /* Have to be a bit careful with return values */
295 ret
= (ret
< 0) ? nr
: ret
+ nr
;