tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
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
19static 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(atomic_read(&page->_count) < 0);
26 atomic_inc(&page->_count);
27}
28
29/*
30 * The performance critical leaf functions are made noinline otherwise gcc
31 * inlines everything into a single function which results in too much
32 * register pressure.
33 */
34static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
35 unsigned long end, int write, struct page **pages, int *nr)
36{
37 unsigned long mask, result;
38 pte_t *ptep;
39
40 result = _PAGE_PRESENT|_PAGE_USER;
41 if (write)
42 result |= _PAGE_RW;
43 mask = result | _PAGE_SPECIAL;
44
45 ptep = pte_offset_kernel(&pmd, addr);
46 do {
47 pte_t pte = *ptep;
48 struct page *page;
49
50 if ((pte_val(pte) & mask) != result)
51 return 0;
52 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
53 page = pte_page(pte);
54 if (!page_cache_get_speculative(page))
55 return 0;
56 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
57 put_page(page);
58 return 0;
59 }
60 if (PageTail(page))
61 get_huge_page_tail(page);
62 pages[*nr] = page;
63 (*nr)++;
64
65 } while (ptep++, addr += PAGE_SIZE, addr != end);
66
67 return 1;
68}
69
70static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
71 int write, struct page **pages, int *nr)
72{
73 unsigned long next;
74 pmd_t *pmdp;
75
76 pmdp = pmd_offset(&pud, addr);
77 do {
78 pmd_t pmd = *pmdp;
79
80 next = pmd_addr_end(addr, end);
81 if (pmd_none(pmd))
82 return 0;
83 if (is_hugepd(pmdp)) {
84 if (!gup_hugepd((hugepd_t *)pmdp, PMD_SHIFT,
85 addr, next, write, pages, nr))
86 return 0;
87 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
88 return 0;
89 } while (pmdp++, addr = next, addr != end);
90
91 return 1;
92}
93
94static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
95 int write, struct page **pages, int *nr)
96{
97 unsigned long next;
98 pud_t *pudp;
99
100 pudp = pud_offset(&pgd, addr);
101 do {
102 pud_t pud = *pudp;
103
104 next = pud_addr_end(addr, end);
105 if (pud_none(pud))
106 return 0;
107 if (is_hugepd(pudp)) {
108 if (!gup_hugepd((hugepd_t *)pudp, PUD_SHIFT,
109 addr, next, write, pages, nr))
110 return 0;
111 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
112 return 0;
113 } while (pudp++, addr = next, addr != end);
114
115 return 1;
116}
117
118int get_user_pages_fast(unsigned long start, int nr_pages, int write,
119 struct page **pages)
120{
121 struct mm_struct *mm = current->mm;
122 unsigned long addr, len, end;
123 unsigned long next;
124 pgd_t *pgdp;
125 int nr = 0;
126
127 pr_devel("%s(%lx,%x,%s)\n", __func__, start, nr_pages, write ? "write" : "read");
128
129 start &= PAGE_MASK;
130 addr = start;
131 len = (unsigned long) nr_pages << PAGE_SHIFT;
132 end = start + len;
133
134 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
135 start, len)))
136 goto slow_irqon;
137
138 pr_devel(" aligned: %lx .. %lx\n", start, end);
139
140 /*
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.
145 *
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.
149 */
150 /*
151 * This doesn't prevent pagetable teardown, but does prevent
152 * the pagetables from being freed on powerpc.
153 *
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.
156 */
157 local_irq_disable();
158
159 pgdp = pgd_offset(mm, addr);
160 do {
161 pgd_t pgd = *pgdp;
162
163 pr_devel(" %016lx: normal pgd %p\n", addr,
164 (void *)pgd_val(pgd));
165 next = pgd_addr_end(addr, end);
166 if (pgd_none(pgd))
167 goto slow;
168 if (is_hugepd(pgdp)) {
169 if (!gup_hugepd((hugepd_t *)pgdp, PGDIR_SHIFT,
170 addr, next, write, pages, &nr))
171 goto slow;
172 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
173 goto slow;
174 } while (pgdp++, addr = next, addr != end);
175
176 local_irq_enable();
177
178 VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
179 return nr;
180
181 {
182 int ret;
183
184slow:
185 local_irq_enable();
186slow_irqon:
187 pr_devel(" slow path ! nr = %d\n", nr);
188
189 /* Try to get the remaining pages with get_user_pages */
190 start += nr << PAGE_SHIFT;
191 pages += nr;
192
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);
197
198 /* Have to be a bit careful with return values */
199 if (nr > 0) {
200 if (ret < 0)
201 ret = nr;
202 else
203 ret += nr;
204 }
205
206 return ret;
207 }
208}
209
210#endif /* __HAVE_ARCH_PTE_SPECIAL */