2 * Lockless get_user_pages_fast for powerpc
4 * Copyright (C) 2008 Nick Piggin
5 * Copyright (C) 2008 Novell Inc.
9 #include <linux/sched.h>
11 #include <linux/hugetlb.h>
12 #include <linux/vmstat.h>
13 #include <linux/pagemap.h>
14 #include <linux/rwsem.h>
15 #include <asm/pgtable.h>
17 #ifdef __HAVE_ARCH_PTE_SPECIAL
19 static inline void get_huge_page_tail(struct page *page)
22 * __split_huge_page_refcount() cannot run
25 VM_BUG_ON(atomic_read(&page->_count) < 0);
26 atomic_inc(&page->_count);
30 * The performance critical leaf functions are made noinline otherwise gcc
31 * inlines everything into a single function which results in too much
34 static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
35 unsigned long end, int write, struct page **pages, int *nr)
37 unsigned long mask, result;
40 result = _PAGE_PRESENT|_PAGE_USER;
43 mask = result | _PAGE_SPECIAL;
45 ptep = pte_offset_kernel(&pmd, addr);
50 if ((pte_val(pte) & mask) != result)
52 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
54 if (!page_cache_get_speculative(page))
56 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
61 get_huge_page_tail(page);
65 } while (ptep++, addr += PAGE_SIZE, addr != end);
70 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
71 int write, struct page **pages, int *nr)
76 pmdp = pmd_offset(&pud, addr);
80 next = pmd_addr_end(addr, end);
83 if (is_hugepd(pmdp)) {
84 if (!gup_hugepd((hugepd_t *)pmdp, PMD_SHIFT,
85 addr, next, write, pages, nr))
87 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
89 } while (pmdp++, addr = next, addr != end);
94 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
95 int write, struct page **pages, int *nr)
100 pudp = pud_offset(&pgd, addr);
104 next = pud_addr_end(addr, end);
107 if (is_hugepd(pudp)) {
108 if (!gup_hugepd((hugepd_t *)pudp, PUD_SHIFT,
109 addr, next, write, pages, nr))
111 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
113 } while (pudp++, addr = next, addr != end);
118 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
121 struct mm_struct *mm = current->mm;
122 unsigned long addr, len, end;
127 pr_devel("%s(%lx,%x,%s)\n", __func__, start, nr_pages, write ? "write" : "read");
131 len = (unsigned long) nr_pages << PAGE_SHIFT;
134 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
138 pr_devel(" aligned: %lx .. %lx\n", start, end);
141 * XXX: batch / limit 'nr', to avoid large irq off latency
142 * needs some instrumenting to determine the common sizes used by
143 * important workloads (eg. DB2), and whether limiting the batch size
144 * will decrease performance.
146 * It seems like we're in the clear for the moment. Direct-IO is
147 * the main guy that batches up lots of get_user_pages, and even
148 * they are limited to 64-at-a-time which is not so many.
151 * This doesn't prevent pagetable teardown, but does prevent
152 * the pagetables from being freed on powerpc.
154 * So long as we atomically load page table pointers versus teardown,
155 * we can follow the address down to the the page and take a ref on it.
159 pgdp = pgd_offset(mm, addr);
163 pr_devel(" %016lx: normal pgd %p\n", addr,
164 (void *)pgd_val(pgd));
165 next = pgd_addr_end(addr, end);
168 if (is_hugepd(pgdp)) {
169 if (!gup_hugepd((hugepd_t *)pgdp, PGDIR_SHIFT,
170 addr, next, write, pages, &nr))
172 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
174 } while (pgdp++, addr = next, addr != end);
178 VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
187 pr_devel(" slow path ! nr = %d\n", nr);
189 /* Try to get the remaining pages with get_user_pages */
190 start += nr << PAGE_SHIFT;
193 down_read(&mm->mmap_sem);
194 ret = get_user_pages(current, mm, start,
195 (end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
196 up_read(&mm->mmap_sem);
198 /* Have to be a bit careful with return values */
210 #endif /* __HAVE_ARCH_PTE_SPECIAL */