2 * Lockless get_user_pages_fast for x86
4 * Copyright (C) 2008 Nick Piggin
5 * Copyright (C) 2008 Novell Inc.
7 #include <linux/sched.h>
9 #include <linux/vmstat.h>
10 #include <linux/highmem.h>
11 #include <linux/swap.h>
13 #include <asm/pgtable.h>
15 static inline pte_t
gup_get_pte(pte_t
*ptep
)
17 #ifndef CONFIG_X86_PAE
18 return ACCESS_ONCE(*ptep
);
21 * With get_user_pages_fast, we walk down the pagetables without taking
22 * any locks. For this we would like to load the pointers atomically,
23 * but that is not possible (without expensive cmpxchg8b) on PAE. What
24 * we do have is the guarantee that a pte will only either go from not
25 * present to present, or present to not present or both -- it will not
26 * switch to a completely different present page without a TLB flush in
27 * between; something that we are blocking by holding interrupts off.
29 * Setting ptes from not present to present goes:
34 * And present to not present goes:
39 * We must ensure here that the load of pte_low sees l iff pte_high
40 * sees h. We load pte_high *after* loading pte_low, which ensures we
41 * don't see an older value of pte_high. *Then* we recheck pte_low,
42 * which ensures that we haven't picked up a changed pte high. We might
43 * have got rubbish values from pte_low and pte_high, but we are
44 * guaranteed that pte_low will not have the present bit set *unless*
45 * it is 'l'. And get_user_pages_fast only operates on present ptes, so
48 * gup_get_pte should not be used or copied outside gup.c without being
49 * very careful -- it does not atomically load the pte or anything that
50 * is likely to be useful for you.
55 pte
.pte_low
= ptep
->pte_low
;
57 pte
.pte_high
= ptep
->pte_high
;
59 if (unlikely(pte
.pte_low
!= ptep
->pte_low
))
67 * The performance critical leaf functions are made noinline otherwise gcc
68 * inlines everything into a single function which results in too much
71 static noinline
int gup_pte_range(pmd_t pmd
, unsigned long addr
,
72 unsigned long end
, int write
, struct page
**pages
, int *nr
)
77 mask
= _PAGE_PRESENT
|_PAGE_USER
;
81 ptep
= pte_offset_map(&pmd
, addr
);
83 pte_t pte
= gup_get_pte(ptep
);
86 if ((pte_flags(pte
) & (mask
| _PAGE_SPECIAL
)) != mask
) {
90 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
93 SetPageReferenced(page
);
97 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
103 static inline void get_head_page_multiple(struct page
*page
, int nr
)
105 VM_BUG_ON(page
!= compound_head(page
));
106 VM_BUG_ON(page_count(page
) == 0);
107 atomic_add(nr
, &page
->_count
);
108 SetPageReferenced(page
);
111 static inline void get_huge_page_tail(struct page
*page
)
114 * __split_huge_page_refcount() cannot run
117 VM_BUG_ON(page_mapcount(page
) < 0);
118 VM_BUG_ON(atomic_read(&page
->_count
) != 0);
119 atomic_inc(&page
->_mapcount
);
122 static noinline
int gup_huge_pmd(pmd_t pmd
, unsigned long addr
,
123 unsigned long end
, int write
, struct page
**pages
, int *nr
)
126 pte_t pte
= *(pte_t
*)&pmd
;
127 struct page
*head
, *page
;
130 mask
= _PAGE_PRESENT
|_PAGE_USER
;
133 if ((pte_flags(pte
) & mask
) != mask
)
135 /* hugepages are never "special" */
136 VM_BUG_ON(pte_flags(pte
) & _PAGE_SPECIAL
);
137 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
140 head
= pte_page(pte
);
141 page
= head
+ ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
143 VM_BUG_ON(compound_head(page
) != head
);
146 get_huge_page_tail(page
);
150 } while (addr
+= PAGE_SIZE
, addr
!= end
);
151 get_head_page_multiple(head
, refs
);
156 static int gup_pmd_range(pud_t pud
, unsigned long addr
, unsigned long end
,
157 int write
, struct page
**pages
, int *nr
)
162 pmdp
= pmd_offset(&pud
, addr
);
166 next
= pmd_addr_end(addr
, end
);
168 * The pmd_trans_splitting() check below explains why
169 * pmdp_splitting_flush has to flush the tlb, to stop
170 * this gup-fast code from running while we set the
171 * splitting bit in the pmd. Returning zero will take
172 * the slow path that will call wait_split_huge_page()
173 * if the pmd is still in splitting state. gup-fast
174 * can't because it has irq disabled and
175 * wait_split_huge_page() would never return as the
176 * tlb flush IPI wouldn't run.
178 if (pmd_none(pmd
) || pmd_trans_splitting(pmd
))
180 if (unlikely(pmd_large(pmd
))) {
181 if (!gup_huge_pmd(pmd
, addr
, next
, write
, pages
, nr
))
184 if (!gup_pte_range(pmd
, addr
, next
, write
, pages
, nr
))
187 } while (pmdp
++, addr
= next
, addr
!= end
);
192 static noinline
int gup_huge_pud(pud_t pud
, unsigned long addr
,
193 unsigned long end
, int write
, struct page
**pages
, int *nr
)
196 pte_t pte
= *(pte_t
*)&pud
;
197 struct page
*head
, *page
;
200 mask
= _PAGE_PRESENT
|_PAGE_USER
;
203 if ((pte_flags(pte
) & mask
) != mask
)
205 /* hugepages are never "special" */
206 VM_BUG_ON(pte_flags(pte
) & _PAGE_SPECIAL
);
207 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
210 head
= pte_page(pte
);
211 page
= head
+ ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
213 VM_BUG_ON(compound_head(page
) != head
);
218 } while (addr
+= PAGE_SIZE
, addr
!= end
);
219 get_head_page_multiple(head
, refs
);
224 static int gup_pud_range(pgd_t pgd
, unsigned long addr
, unsigned long end
,
225 int write
, struct page
**pages
, int *nr
)
230 pudp
= pud_offset(&pgd
, addr
);
234 next
= pud_addr_end(addr
, end
);
237 if (unlikely(pud_large(pud
))) {
238 if (!gup_huge_pud(pud
, addr
, next
, write
, pages
, nr
))
241 if (!gup_pmd_range(pud
, addr
, next
, write
, pages
, nr
))
244 } while (pudp
++, addr
= next
, addr
!= end
);
250 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
251 * back to the regular GUP.
253 int __get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
256 struct mm_struct
*mm
= current
->mm
;
257 unsigned long addr
, len
, end
;
265 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
267 if (unlikely(!access_ok(write
? VERIFY_WRITE
: VERIFY_READ
,
268 (void __user
*)start
, len
)))
272 * XXX: batch / limit 'nr', to avoid large irq off latency
273 * needs some instrumenting to determine the common sizes used by
274 * important workloads (eg. DB2), and whether limiting the batch size
275 * will decrease performance.
277 * It seems like we're in the clear for the moment. Direct-IO is
278 * the main guy that batches up lots of get_user_pages, and even
279 * they are limited to 64-at-a-time which is not so many.
282 * This doesn't prevent pagetable teardown, but does prevent
283 * the pagetables and pages from being freed on x86.
285 * So long as we atomically load page table pointers versus teardown
286 * (which we do on x86, with the above PAE exception), we can follow the
287 * address down to the the page and take a ref on it.
289 local_irq_save(flags
);
290 pgdp
= pgd_offset(mm
, addr
);
294 next
= pgd_addr_end(addr
, end
);
297 if (!gup_pud_range(pgd
, addr
, next
, write
, pages
, &nr
))
299 } while (pgdp
++, addr
= next
, addr
!= end
);
300 local_irq_restore(flags
);
306 * get_user_pages_fast() - pin user pages in memory
307 * @start: starting user address
308 * @nr_pages: number of pages from start to pin
309 * @write: whether pages will be written to
310 * @pages: array that receives pointers to the pages pinned.
311 * Should be at least nr_pages long.
313 * Attempt to pin user pages in memory without taking mm->mmap_sem.
314 * If not successful, it will fall back to taking the lock and
315 * calling get_user_pages().
317 * Returns number of pages pinned. This may be fewer than the number
318 * requested. If nr_pages is 0 or negative, returns 0. If no pages
319 * were pinned, returns -errno.
321 int get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
324 struct mm_struct
*mm
= current
->mm
;
325 unsigned long addr
, len
, end
;
332 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
339 if (end
>> __VIRTUAL_MASK_SHIFT
)
344 * XXX: batch / limit 'nr', to avoid large irq off latency
345 * needs some instrumenting to determine the common sizes used by
346 * important workloads (eg. DB2), and whether limiting the batch size
347 * will decrease performance.
349 * It seems like we're in the clear for the moment. Direct-IO is
350 * the main guy that batches up lots of get_user_pages, and even
351 * they are limited to 64-at-a-time which is not so many.
354 * This doesn't prevent pagetable teardown, but does prevent
355 * the pagetables and pages from being freed on x86.
357 * So long as we atomically load page table pointers versus teardown
358 * (which we do on x86, with the above PAE exception), we can follow the
359 * address down to the the page and take a ref on it.
362 pgdp
= pgd_offset(mm
, addr
);
366 next
= pgd_addr_end(addr
, end
);
369 if (!gup_pud_range(pgd
, addr
, next
, write
, pages
, &nr
))
371 } while (pgdp
++, addr
= next
, addr
!= end
);
374 VM_BUG_ON(nr
!= (end
- start
) >> PAGE_SHIFT
);
383 /* Try to get the remaining pages with get_user_pages */
384 start
+= nr
<< PAGE_SHIFT
;
387 down_read(&mm
->mmap_sem
);
388 ret
= get_user_pages(current
, mm
, start
,
389 (end
- start
) >> PAGE_SHIFT
, write
, 0, pages
, NULL
);
390 up_read(&mm
->mmap_sem
);
392 /* Have to be a bit careful with return values */