2 * Lockless get_user_pages_fast for MIPS
4 * Copyright (C) 2008 Nick Piggin
5 * Copyright (C) 2008 Novell Inc.
6 * Copyright (C) 2011 Ralf Baechle
8 #include <linux/sched.h>
10 #include <linux/vmstat.h>
11 #include <linux/highmem.h>
12 #include <linux/swap.h>
13 #include <linux/hugetlb.h>
15 #include <asm/pgtable.h>
17 static inline pte_t gup_get_pte(pte_t *ptep)
19 #if defined(CONFIG_64BIT_PHYS_ADDR) && defined(CONFIG_CPU_MIPS32)
23 pte.pte_low = ptep->pte_low;
25 pte.pte_high = ptep->pte_high;
27 if (unlikely(pte.pte_low != ptep->pte_low))
32 return ACCESS_ONCE(*ptep);
36 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
37 int write, struct page **pages, int *nr)
39 pte_t *ptep = pte_offset_map(&pmd, addr);
41 pte_t pte = gup_get_pte(ptep);
44 if (!pte_present(pte) ||
45 pte_special(pte) || (write && !pte_write(pte))) {
49 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
52 SetPageReferenced(page);
56 } while (ptep++, addr += PAGE_SIZE, addr != end);
62 static inline void get_head_page_multiple(struct page *page, int nr)
64 VM_BUG_ON(page != compound_head(page));
65 VM_BUG_ON(page_count(page) == 0);
66 atomic_add(nr, &page->_count);
67 SetPageReferenced(page);
70 static int gup_huge_pmd(pmd_t pmd, unsigned long addr, unsigned long end,
71 int write, struct page **pages, int *nr)
73 pte_t pte = *(pte_t *)&pmd;
74 struct page *head, *page;
77 if (write && !pte_write(pte))
79 /* hugepages are never "special" */
80 VM_BUG_ON(pte_special(pte));
81 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
85 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
87 VM_BUG_ON(compound_head(page) != head);
90 get_huge_page_tail(page);
94 } while (addr += PAGE_SIZE, addr != end);
96 get_head_page_multiple(head, refs);
100 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
101 int write, struct page **pages, int *nr)
106 pmdp = pmd_offset(&pud, addr);
110 next = pmd_addr_end(addr, end);
112 * The pmd_trans_splitting() check below explains why
113 * pmdp_splitting_flush has to flush the tlb, to stop
114 * this gup-fast code from running while we set the
115 * splitting bit in the pmd. Returning zero will take
116 * the slow path that will call wait_split_huge_page()
117 * if the pmd is still in splitting state. gup-fast
118 * can't because it has irq disabled and
119 * wait_split_huge_page() would never return as the
120 * tlb flush IPI wouldn't run.
122 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
124 if (unlikely(pmd_huge(pmd))) {
125 if (!gup_huge_pmd(pmd, addr, next, write, pages,nr))
128 if (!gup_pte_range(pmd, addr, next, write, pages,nr))
131 } while (pmdp++, addr = next, addr != end);
136 static int gup_huge_pud(pud_t pud, unsigned long addr, unsigned long end,
137 int write, struct page **pages, int *nr)
139 pte_t pte = *(pte_t *)&pud;
140 struct page *head, *page;
143 if (write && !pte_write(pte))
145 /* hugepages are never "special" */
146 VM_BUG_ON(pte_special(pte));
147 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
150 head = pte_page(pte);
151 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
153 VM_BUG_ON(compound_head(page) != head);
158 } while (addr += PAGE_SIZE, addr != end);
160 get_head_page_multiple(head, refs);
164 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
165 int write, struct page **pages, int *nr)
170 pudp = pud_offset(&pgd, addr);
174 next = pud_addr_end(addr, end);
177 if (unlikely(pud_huge(pud))) {
178 if (!gup_huge_pud(pud, addr, next, write, pages,nr))
181 if (!gup_pmd_range(pud, addr, next, write, pages,nr))
184 } while (pudp++, addr = next, addr != end);
190 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
191 * back to the regular GUP.
193 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
196 struct mm_struct *mm = current->mm;
197 unsigned long addr, len, end;
205 len = (unsigned long) nr_pages << PAGE_SHIFT;
207 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
208 (void __user *)start, len)))
212 * XXX: batch / limit 'nr', to avoid large irq off latency
213 * needs some instrumenting to determine the common sizes used by
214 * important workloads (eg. DB2), and whether limiting the batch
215 * size will decrease performance.
217 * It seems like we're in the clear for the moment. Direct-IO is
218 * the main guy that batches up lots of get_user_pages, and even
219 * they are limited to 64-at-a-time which is not so many.
222 * This doesn't prevent pagetable teardown, but does prevent
223 * the pagetables and pages from being freed.
225 * So long as we atomically load page table pointers versus teardown,
226 * we can follow the address down to the page and take a ref on it.
228 local_irq_save(flags);
229 pgdp = pgd_offset(mm, addr);
233 next = pgd_addr_end(addr, end);
236 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
238 } while (pgdp++, addr = next, addr != end);
239 local_irq_restore(flags);
245 * get_user_pages_fast() - pin user pages in memory
246 * @start: starting user address
247 * @nr_pages: number of pages from start to pin
248 * @write: whether pages will be written to
249 * @pages: array that receives pointers to the pages pinned.
250 * Should be at least nr_pages long.
252 * Attempt to pin user pages in memory without taking mm->mmap_sem.
253 * If not successful, it will fall back to taking the lock and
254 * calling get_user_pages().
256 * Returns number of pages pinned. This may be fewer than the number
257 * requested. If nr_pages is 0 or negative, returns 0. If no pages
258 * were pinned, returns -errno.
260 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
263 struct mm_struct *mm = current->mm;
264 unsigned long addr, len, end;
271 len = (unsigned long) nr_pages << PAGE_SHIFT;
277 /* XXX: batch / limit 'nr' */
279 pgdp = pgd_offset(mm, addr);
283 next = pgd_addr_end(addr, end);
286 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
288 } while (pgdp++, addr = next, addr != end);
291 VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
297 /* Try to get the remaining pages with get_user_pages */
298 start += nr << PAGE_SHIFT;
301 down_read(&mm->mmap_sem);
302 ret = get_user_pages(current, mm, start,
303 (end - start) >> PAGE_SHIFT,
304 write, 0, pages, NULL);
305 up_read(&mm->mmap_sem);
307 /* Have to be a bit careful with return values */