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);
37 pte_t pte = ACCESS_ONCE(*ptep);
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);
66 pmd_t pmd = ACCESS_ONCE(*pmdp);
68 next = pmd_addr_end(addr, end);
70 * If we find a splitting transparent hugepage we
71 * return zero. That will result in taking the slow
72 * path which will call wait_split_huge_page()
73 * if the pmd is still in splitting state
75 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
77 if (pmd_huge(pmd) || pmd_large(pmd)) {
78 if (!gup_hugepte((pte_t *)pmdp, PMD_SIZE, addr, next,
81 } else if (is_hugepd(pmdp)) {
82 if (!gup_hugepd((hugepd_t *)pmdp, PMD_SHIFT,
83 addr, next, write, pages, nr))
85 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
87 } while (pmdp++, addr = next, addr != end);
92 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
93 int write, struct page **pages, int *nr)
98 pudp = pud_offset(&pgd, addr);
100 pud_t pud = ACCESS_ONCE(*pudp);
102 next = pud_addr_end(addr, end);
106 if (!gup_hugepte((pte_t *)pudp, PUD_SIZE, addr, next,
109 } else if (is_hugepd(pudp)) {
110 if (!gup_hugepd((hugepd_t *)pudp, PUD_SHIFT,
111 addr, next, write, pages, nr))
113 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
115 } while (pudp++, addr = next, addr != end);
120 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
123 struct mm_struct *mm = current->mm;
124 unsigned long addr, len, end;
130 pr_devel("%s(%lx,%x,%s)\n", __func__, start, nr_pages, write ? "write" : "read");
134 len = (unsigned long) nr_pages << PAGE_SHIFT;
137 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
141 pr_devel(" aligned: %lx .. %lx\n", start, end);
144 * XXX: batch / limit 'nr', to avoid large irq off latency
145 * needs some instrumenting to determine the common sizes used by
146 * important workloads (eg. DB2), and whether limiting the batch size
147 * will decrease performance.
149 * It seems like we're in the clear for the moment. Direct-IO is
150 * the main guy that batches up lots of get_user_pages, and even
151 * they are limited to 64-at-a-time which is not so many.
154 * This doesn't prevent pagetable teardown, but does prevent
155 * the pagetables from being freed on powerpc.
157 * So long as we atomically load page table pointers versus teardown,
158 * we can follow the address down to the the page and take a ref on it.
160 local_irq_save(flags);
162 pgdp = pgd_offset(mm, addr);
164 pgd_t pgd = ACCESS_ONCE(*pgdp);
166 pr_devel(" %016lx: normal pgd %p\n", addr,
167 (void *)pgd_val(pgd));
168 next = pgd_addr_end(addr, end);
172 if (!gup_hugepte((pte_t *)pgdp, PGDIR_SIZE, addr, next,
175 } else if (is_hugepd(pgdp)) {
176 if (!gup_hugepd((hugepd_t *)pgdp, PGDIR_SHIFT,
177 addr, next, write, pages, &nr))
179 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
181 } while (pgdp++, addr = next, addr != end);
183 local_irq_restore(flags);
188 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
191 struct mm_struct *mm = current->mm;
195 nr = __get_user_pages_fast(start, nr_pages, write, pages);
199 pr_devel(" slow path ! nr = %d\n", nr);
201 /* Try to get the remaining pages with get_user_pages */
202 start += nr << PAGE_SHIFT;
205 down_read(&mm->mmap_sem);
206 ret = get_user_pages(current, mm, start,
207 nr_pages - nr, write, 0, pages, NULL);
208 up_read(&mm->mmap_sem);
210 /* Have to be a bit careful with return values */
222 #endif /* __HAVE_ARCH_PTE_SPECIAL */