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/cpu-features.h>
16 #include <asm/pgtable.h>
18 static inline pte_t gup_get_pte(pte_t *ptep)
20 #if defined(CONFIG_PHYS_ADDR_T_64BIT) && defined(CONFIG_CPU_MIPS32)
24 pte.pte_low = ptep->pte_low;
26 pte.pte_high = ptep->pte_high;
28 if (unlikely(pte.pte_low != ptep->pte_low))
33 return READ_ONCE(*ptep);
37 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
38 int write, struct page **pages, int *nr)
40 pte_t *ptep = pte_offset_map(&pmd, addr);
42 pte_t pte = gup_get_pte(ptep);
45 if (!pte_present(pte) ||
46 pte_special(pte) || (write && !pte_write(pte))) {
50 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
53 SetPageReferenced(page);
57 } while (ptep++, addr += PAGE_SIZE, addr != end);
63 static inline void get_head_page_multiple(struct page *page, int nr)
65 VM_BUG_ON(page != compound_head(page));
66 VM_BUG_ON(page_count(page) == 0);
67 atomic_add(nr, &page->_count);
68 SetPageReferenced(page);
71 static int gup_huge_pmd(pmd_t pmd, unsigned long addr, unsigned long end,
72 int write, struct page **pages, int *nr)
74 pte_t pte = *(pte_t *)&pmd;
75 struct page *head, *page;
78 if (write && !pte_write(pte))
80 /* hugepages are never "special" */
81 VM_BUG_ON(pte_special(pte));
82 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
86 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
88 VM_BUG_ON(compound_head(page) != head);
93 } while (addr += PAGE_SIZE, addr != end);
95 get_head_page_multiple(head, refs);
99 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
100 int write, struct page **pages, int *nr)
105 pmdp = pmd_offset(&pud, addr);
109 next = pmd_addr_end(addr, end);
112 if (unlikely(pmd_huge(pmd))) {
113 if (!gup_huge_pmd(pmd, addr, next, write, pages,nr))
116 if (!gup_pte_range(pmd, addr, next, write, pages,nr))
119 } while (pmdp++, addr = next, addr != end);
124 static int gup_huge_pud(pud_t pud, unsigned long addr, unsigned long end,
125 int write, struct page **pages, int *nr)
127 pte_t pte = *(pte_t *)&pud;
128 struct page *head, *page;
131 if (write && !pte_write(pte))
133 /* hugepages are never "special" */
134 VM_BUG_ON(pte_special(pte));
135 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
138 head = pte_page(pte);
139 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
141 VM_BUG_ON(compound_head(page) != head);
146 } while (addr += PAGE_SIZE, addr != end);
148 get_head_page_multiple(head, refs);
152 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
153 int write, struct page **pages, int *nr)
158 pudp = pud_offset(&pgd, addr);
162 next = pud_addr_end(addr, end);
165 if (unlikely(pud_huge(pud))) {
166 if (!gup_huge_pud(pud, addr, next, write, pages,nr))
169 if (!gup_pmd_range(pud, addr, next, write, pages,nr))
172 } while (pudp++, addr = next, addr != end);
178 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
179 * back to the regular GUP.
181 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
184 struct mm_struct *mm = current->mm;
185 unsigned long addr, len, end;
193 len = (unsigned long) nr_pages << PAGE_SHIFT;
195 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
196 (void __user *)start, len)))
200 * XXX: batch / limit 'nr', to avoid large irq off latency
201 * needs some instrumenting to determine the common sizes used by
202 * important workloads (eg. DB2), and whether limiting the batch
203 * size will decrease performance.
205 * It seems like we're in the clear for the moment. Direct-IO is
206 * the main guy that batches up lots of get_user_pages, and even
207 * they are limited to 64-at-a-time which is not so many.
210 * This doesn't prevent pagetable teardown, but does prevent
211 * the pagetables and pages from being freed.
213 * So long as we atomically load page table pointers versus teardown,
214 * we can follow the address down to the page and take a ref on it.
216 local_irq_save(flags);
217 pgdp = pgd_offset(mm, addr);
221 next = pgd_addr_end(addr, end);
224 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
226 } while (pgdp++, addr = next, addr != end);
227 local_irq_restore(flags);
233 * get_user_pages_fast() - pin user pages in memory
234 * @start: starting user address
235 * @nr_pages: number of pages from start to pin
236 * @write: whether pages will be written to
237 * @pages: array that receives pointers to the pages pinned.
238 * Should be at least nr_pages long.
240 * Attempt to pin user pages in memory without taking mm->mmap_sem.
241 * If not successful, it will fall back to taking the lock and
242 * calling get_user_pages().
244 * Returns number of pages pinned. This may be fewer than the number
245 * requested. If nr_pages is 0 or negative, returns 0. If no pages
246 * were pinned, returns -errno.
248 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
251 struct mm_struct *mm = current->mm;
252 unsigned long addr, len, end;
259 len = (unsigned long) nr_pages << PAGE_SHIFT;
262 if (end < start || cpu_has_dc_aliases)
265 /* XXX: batch / limit 'nr' */
267 pgdp = pgd_offset(mm, addr);
271 next = pgd_addr_end(addr, end);
274 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
276 } while (pgdp++, addr = next, addr != end);
279 VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
285 /* Try to get the remaining pages with get_user_pages */
286 start += nr << PAGE_SHIFT;
289 ret = get_user_pages_unlocked(current, mm, start,
290 (end - start) >> PAGE_SHIFT,
293 /* Have to be a bit careful with return values */