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
20 * The performance critical leaf functions are made noinline otherwise gcc
21 * inlines everything into a single function which results in too much
24 static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
25 unsigned long end, int write, struct page **pages, int *nr)
27 unsigned long mask, result;
30 result = _PAGE_PRESENT|_PAGE_USER;
33 mask = result | _PAGE_SPECIAL;
35 ptep = pte_offset_kernel(&pmd, addr);
40 if ((pte_val(pte) & mask) != result)
42 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
44 if (!page_cache_get_speculative(page))
46 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
53 } while (ptep++, addr += PAGE_SIZE, addr != end);
58 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
59 int write, struct page **pages, int *nr)
64 pmdp = pmd_offset(&pud, addr);
68 next = pmd_addr_end(addr, end);
71 if (is_hugepd(pmdp)) {
72 if (!gup_hugepd((hugepd_t *)pmdp, PMD_SHIFT,
73 addr, next, write, pages, nr))
75 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
77 } while (pmdp++, addr = next, addr != end);
82 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
83 int write, struct page **pages, int *nr)
88 pudp = pud_offset(&pgd, addr);
92 next = pud_addr_end(addr, end);
95 if (is_hugepd(pudp)) {
96 if (!gup_hugepd((hugepd_t *)pudp, PUD_SHIFT,
97 addr, next, write, pages, nr))
99 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
101 } while (pudp++, addr = next, addr != end);
106 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
109 struct mm_struct *mm = current->mm;
110 unsigned long addr, len, end;
115 pr_devel("%s(%lx,%x,%s)\n", __func__, start, nr_pages, write ? "write" : "read");
119 len = (unsigned long) nr_pages << PAGE_SHIFT;
122 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
126 pr_devel(" aligned: %lx .. %lx\n", start, end);
129 * XXX: batch / limit 'nr', to avoid large irq off latency
130 * needs some instrumenting to determine the common sizes used by
131 * important workloads (eg. DB2), and whether limiting the batch size
132 * will decrease performance.
134 * It seems like we're in the clear for the moment. Direct-IO is
135 * the main guy that batches up lots of get_user_pages, and even
136 * they are limited to 64-at-a-time which is not so many.
139 * This doesn't prevent pagetable teardown, but does prevent
140 * the pagetables from being freed on powerpc.
142 * So long as we atomically load page table pointers versus teardown,
143 * we can follow the address down to the the page and take a ref on it.
147 pgdp = pgd_offset(mm, addr);
151 pr_devel(" %016lx: normal pgd %p\n", addr,
152 (void *)pgd_val(pgd));
153 next = pgd_addr_end(addr, end);
156 if (is_hugepd(pgdp)) {
157 if (!gup_hugepd((hugepd_t *)pgdp, PGDIR_SHIFT,
158 addr, next, write, pages, &nr))
160 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
162 } while (pgdp++, addr = next, addr != end);
166 VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
175 pr_devel(" slow path ! nr = %d\n", nr);
177 /* Try to get the remaining pages with get_user_pages */
178 start += nr << PAGE_SHIFT;
181 down_read(&mm->mmap_sem);
182 ret = get_user_pages(current, mm, start,
183 (end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
184 up_read(&mm->mmap_sem);
186 /* Have to be a bit careful with return values */
198 #endif /* __HAVE_ARCH_PTE_SPECIAL */