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);
91 get_huge_page_tail(page);
95 } while (addr += PAGE_SIZE, addr != end);
97 get_head_page_multiple(head, refs);
101 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
102 int write, struct page **pages, int *nr)
107 pmdp = pmd_offset(&pud, addr);
111 next = pmd_addr_end(addr, end);
113 * The pmd_trans_splitting() check below explains why
114 * pmdp_splitting_flush has to flush the tlb, to stop
115 * this gup-fast code from running while we set the
116 * splitting bit in the pmd. Returning zero will take
117 * the slow path that will call wait_split_huge_page()
118 * if the pmd is still in splitting state. gup-fast
119 * can't because it has irq disabled and
120 * wait_split_huge_page() would never return as the
121 * tlb flush IPI wouldn't run.
123 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
125 if (unlikely(pmd_huge(pmd))) {
126 if (!gup_huge_pmd(pmd, addr, next, write, pages,nr))
129 if (!gup_pte_range(pmd, addr, next, write, pages,nr))
132 } while (pmdp++, addr = next, addr != end);
137 static int gup_huge_pud(pud_t pud, unsigned long addr, unsigned long end,
138 int write, struct page **pages, int *nr)
140 pte_t pte = *(pte_t *)&pud;
141 struct page *head, *page;
144 if (write && !pte_write(pte))
146 /* hugepages are never "special" */
147 VM_BUG_ON(pte_special(pte));
148 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
151 head = pte_page(pte);
152 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
154 VM_BUG_ON(compound_head(page) != head);
157 get_huge_page_tail(page);
161 } while (addr += PAGE_SIZE, addr != end);
163 get_head_page_multiple(head, refs);
167 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
168 int write, struct page **pages, int *nr)
173 pudp = pud_offset(&pgd, addr);
177 next = pud_addr_end(addr, end);
180 if (unlikely(pud_huge(pud))) {
181 if (!gup_huge_pud(pud, addr, next, write, pages,nr))
184 if (!gup_pmd_range(pud, addr, next, write, pages,nr))
187 } while (pudp++, addr = next, addr != end);
193 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
194 * back to the regular GUP.
196 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
199 struct mm_struct *mm = current->mm;
200 unsigned long addr, len, end;
208 len = (unsigned long) nr_pages << PAGE_SHIFT;
210 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
211 (void __user *)start, len)))
215 * XXX: batch / limit 'nr', to avoid large irq off latency
216 * needs some instrumenting to determine the common sizes used by
217 * important workloads (eg. DB2), and whether limiting the batch
218 * size will decrease performance.
220 * It seems like we're in the clear for the moment. Direct-IO is
221 * the main guy that batches up lots of get_user_pages, and even
222 * they are limited to 64-at-a-time which is not so many.
225 * This doesn't prevent pagetable teardown, but does prevent
226 * the pagetables and pages from being freed.
228 * So long as we atomically load page table pointers versus teardown,
229 * we can follow the address down to the page and take a ref on it.
231 local_irq_save(flags);
232 pgdp = pgd_offset(mm, addr);
236 next = pgd_addr_end(addr, end);
239 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
241 } while (pgdp++, addr = next, addr != end);
242 local_irq_restore(flags);
248 * get_user_pages_fast() - pin user pages in memory
249 * @start: starting user address
250 * @nr_pages: number of pages from start to pin
251 * @write: whether pages will be written to
252 * @pages: array that receives pointers to the pages pinned.
253 * Should be at least nr_pages long.
255 * Attempt to pin user pages in memory without taking mm->mmap_sem.
256 * If not successful, it will fall back to taking the lock and
257 * calling get_user_pages().
259 * Returns number of pages pinned. This may be fewer than the number
260 * requested. If nr_pages is 0 or negative, returns 0. If no pages
261 * were pinned, returns -errno.
263 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
266 struct mm_struct *mm = current->mm;
267 unsigned long addr, len, end;
274 len = (unsigned long) nr_pages << PAGE_SHIFT;
277 if (end < start || cpu_has_dc_aliases)
280 /* XXX: batch / limit 'nr' */
282 pgdp = pgd_offset(mm, addr);
286 next = pgd_addr_end(addr, end);
289 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
291 } while (pgdp++, addr = next, addr != end);
294 VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
300 /* Try to get the remaining pages with get_user_pages */
301 start += nr << PAGE_SHIFT;
304 ret = get_user_pages_unlocked(current, mm, start,
305 (end - start) >> PAGE_SHIFT,
308 /* Have to be a bit careful with return values */