b9e1c7ff5f6d33d8a111c10ab9b3551f6df6f335
[linux-2.6.git] / arch / powerpc / mm / gup.c
1 /*
2  * Lockless get_user_pages_fast for powerpc
3  *
4  * Copyright (C) 2008 Nick Piggin
5  * Copyright (C) 2008 Novell Inc.
6  */
7 #undef DEBUG
8
9 #include <linux/sched.h>
10 #include <linux/mm.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>
16
17 #ifdef __HAVE_ARCH_PTE_SPECIAL
18
19 static inline void get_huge_page_tail(struct page *page)
20 {
21         /*
22          * __split_huge_page_refcount() cannot run
23          * from under us.
24          */
25         VM_BUG_ON(page_mapcount(page) < 0);
26         VM_BUG_ON(atomic_read(&page->_count) != 0);
27         atomic_inc(&page->_mapcount);
28 }
29
30 /*
31  * The performance critical leaf functions are made noinline otherwise gcc
32  * inlines everything into a single function which results in too much
33  * register pressure.
34  */
35 static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
36                 unsigned long end, int write, struct page **pages, int *nr)
37 {
38         unsigned long mask, result;
39         pte_t *ptep;
40
41         result = _PAGE_PRESENT|_PAGE_USER;
42         if (write)
43                 result |= _PAGE_RW;
44         mask = result | _PAGE_SPECIAL;
45
46         ptep = pte_offset_kernel(&pmd, addr);
47         do {
48                 pte_t pte = *ptep;
49                 struct page *page;
50
51                 if ((pte_val(pte) & mask) != result)
52                         return 0;
53                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
54                 page = pte_page(pte);
55                 if (!page_cache_get_speculative(page))
56                         return 0;
57                 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
58                         put_page(page);
59                         return 0;
60                 }
61                 if (PageTail(page))
62                         get_huge_page_tail(page);
63                 pages[*nr] = page;
64                 (*nr)++;
65
66         } while (ptep++, addr += PAGE_SIZE, addr != end);
67
68         return 1;
69 }
70
71 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
72                 int write, struct page **pages, int *nr)
73 {
74         unsigned long next;
75         pmd_t *pmdp;
76
77         pmdp = pmd_offset(&pud, addr);
78         do {
79                 pmd_t pmd = *pmdp;
80
81                 next = pmd_addr_end(addr, end);
82                 if (pmd_none(pmd))
83                         return 0;
84                 if (is_hugepd(pmdp)) {
85                         if (!gup_hugepd((hugepd_t *)pmdp, PMD_SHIFT,
86                                         addr, next, write, pages, nr))
87                                 return 0;
88                 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
89                         return 0;
90         } while (pmdp++, addr = next, addr != end);
91
92         return 1;
93 }
94
95 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
96                 int write, struct page **pages, int *nr)
97 {
98         unsigned long next;
99         pud_t *pudp;
100
101         pudp = pud_offset(&pgd, addr);
102         do {
103                 pud_t pud = *pudp;
104
105                 next = pud_addr_end(addr, end);
106                 if (pud_none(pud))
107                         return 0;
108                 if (is_hugepd(pudp)) {
109                         if (!gup_hugepd((hugepd_t *)pudp, PUD_SHIFT,
110                                         addr, next, write, pages, nr))
111                                 return 0;
112                 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
113                         return 0;
114         } while (pudp++, addr = next, addr != end);
115
116         return 1;
117 }
118
119 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
120                         struct page **pages)
121 {
122         struct mm_struct *mm = current->mm;
123         unsigned long addr, len, end;
124         unsigned long next;
125         pgd_t *pgdp;
126         int nr = 0;
127
128         pr_devel("%s(%lx,%x,%s)\n", __func__, start, nr_pages, write ? "write" : "read");
129
130         start &= PAGE_MASK;
131         addr = start;
132         len = (unsigned long) nr_pages << PAGE_SHIFT;
133         end = start + len;
134
135         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
136                                         start, len)))
137                 goto slow_irqon;
138
139         pr_devel("  aligned: %lx .. %lx\n", start, end);
140
141         /*
142          * XXX: batch / limit 'nr', to avoid large irq off latency
143          * needs some instrumenting to determine the common sizes used by
144          * important workloads (eg. DB2), and whether limiting the batch size
145          * will decrease performance.
146          *
147          * It seems like we're in the clear for the moment. Direct-IO is
148          * the main guy that batches up lots of get_user_pages, and even
149          * they are limited to 64-at-a-time which is not so many.
150          */
151         /*
152          * This doesn't prevent pagetable teardown, but does prevent
153          * the pagetables from being freed on powerpc.
154          *
155          * So long as we atomically load page table pointers versus teardown,
156          * we can follow the address down to the the page and take a ref on it.
157          */
158         local_irq_disable();
159
160         pgdp = pgd_offset(mm, addr);
161         do {
162                 pgd_t pgd = *pgdp;
163
164                 pr_devel("  %016lx: normal pgd %p\n", addr,
165                          (void *)pgd_val(pgd));
166                 next = pgd_addr_end(addr, end);
167                 if (pgd_none(pgd))
168                         goto slow;
169                 if (is_hugepd(pgdp)) {
170                         if (!gup_hugepd((hugepd_t *)pgdp, PGDIR_SHIFT,
171                                         addr, next, write, pages, &nr))
172                                 goto slow;
173                 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
174                         goto slow;
175         } while (pgdp++, addr = next, addr != end);
176
177         local_irq_enable();
178
179         VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
180         return nr;
181
182         {
183                 int ret;
184
185 slow:
186                 local_irq_enable();
187 slow_irqon:
188                 pr_devel("  slow path ! nr = %d\n", nr);
189
190                 /* Try to get the remaining pages with get_user_pages */
191                 start += nr << PAGE_SHIFT;
192                 pages += nr;
193
194                 down_read(&mm->mmap_sem);
195                 ret = get_user_pages(current, mm, start,
196                         (end - start) >> PAGE_SHIFT, write, 0, pages, NULL);
197                 up_read(&mm->mmap_sem);
198
199                 /* Have to be a bit careful with return values */
200                 if (nr > 0) {
201                         if (ret < 0)
202                                 ret = nr;
203                         else
204                                 ret += nr;
205                 }
206
207                 return ret;
208         }
209 }
210
211 #endif /* __HAVE_ARCH_PTE_SPECIAL */