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/*
20 * The performance critical leaf functions are made noinline otherwise gcc
21 * inlines everything into a single function which results in too much
22 * register pressure.
23 */
24static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
25		unsigned long end, int write, struct page **pages, int *nr)
26{
27	unsigned long mask, result;
28	pte_t *ptep;
29
30	result = _PAGE_PRESENT|_PAGE_USER;
31	if (write)
32		result |= _PAGE_RW;
33	mask = result | _PAGE_SPECIAL;
34
35	ptep = pte_offset_kernel(&pmd, addr);
36	do {
37		pte_t pte = *ptep;
38		struct page *page;
39
40		if ((pte_val(pte) & mask) != result)
41			return 0;
42		VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
43		page = pte_page(pte);
44		if (!page_cache_get_speculative(page))
45			return 0;
46		if (unlikely(pte_val(pte) != pte_val(*ptep))) {
47			put_page(page);
48			return 0;
49		}
50		pages[*nr] = page;
51		(*nr)++;
52
53	} while (ptep++, addr += PAGE_SIZE, addr != end);
54
55	return 1;
56}
57
58static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
59		int write, struct page **pages, int *nr)
60{
61	unsigned long next;
62	pmd_t *pmdp;
63
64	pmdp = pmd_offset(&pud, addr);
65	do {
66		pmd_t pmd = *pmdp;
67
68		next = pmd_addr_end(addr, end);
69		if (pmd_none(pmd))
70			return 0;
71		if (is_hugepd(pmdp)) {
72			if (!gup_hugepd((hugepd_t *)pmdp, PMD_SHIFT,
73					addr, next, write, pages, nr))
74				return 0;
75		} else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
76			return 0;
77	} while (pmdp++, addr = next, addr != end);
78
79	return 1;
80}
81
82static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
83		int write, struct page **pages, int *nr)
84{
85	unsigned long next;
86	pud_t *pudp;
87
88	pudp = pud_offset(&pgd, addr);
89	do {
90		pud_t pud = *pudp;
91
92		next = pud_addr_end(addr, end);
93		if (pud_none(pud))
94			return 0;
95		if (is_hugepd(pudp)) {
96			if (!gup_hugepd((hugepd_t *)pudp, PUD_SHIFT,
97					addr, next, write, pages, nr))
98				return 0;
99		} else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
100			return 0;
101	} while (pudp++, addr = next, addr != end);
102
103	return 1;
104}
105
106int get_user_pages_fast(unsigned long start, int nr_pages, int write,
107			struct page **pages)
108{
109	struct mm_struct *mm = current->mm;
110	unsigned long addr, len, end;
111	unsigned long next;
112	pgd_t *pgdp;
113	int nr = 0;
114
115	pr_devel("%s(%lx,%x,%s)\n", __func__, start, nr_pages, write ? "write" : "read");
116
117	start &= PAGE_MASK;
118	addr = start;
119	len = (unsigned long) nr_pages << PAGE_SHIFT;
120	end = start + len;
121
122	if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
123					start, len)))
124		goto slow_irqon;
125
126	pr_devel("  aligned: %lx .. %lx\n", start, end);
127
128	/*
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.
133	 *
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.
137	 */
138	/*
139	 * This doesn't prevent pagetable teardown, but does prevent
140	 * the pagetables from being freed on powerpc.
141	 *
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.
144	 */
145	local_irq_disable();
146
147	pgdp = pgd_offset(mm, addr);
148	do {
149		pgd_t pgd = *pgdp;
150
151		pr_devel("  %016lx: normal pgd %p\n", addr,
152			 (void *)pgd_val(pgd));
153		next = pgd_addr_end(addr, end);
154		if (pgd_none(pgd))
155			goto slow;
156		if (is_hugepd(pgdp)) {
157			if (!gup_hugepd((hugepd_t *)pgdp, PGDIR_SHIFT,
158					addr, next, write, pages, &nr))
159				goto slow;
160		} else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
161			goto slow;
162	} while (pgdp++, addr = next, addr != end);
163
164	local_irq_enable();
165
166	VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
167	return nr;
168
169	{
170		int ret;
171
172slow:
173		local_irq_enable();
174slow_irqon:
175		pr_devel("  slow path ! nr = %d\n", nr);
176
177		/* Try to get the remaining pages with get_user_pages */
178		start += nr << PAGE_SHIFT;
179		pages += nr;
180
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);
185
186		/* Have to be a bit careful with return values */
187		if (nr > 0) {
188			if (ret < 0)
189				ret = nr;
190			else
191				ret += nr;
192		}
193
194		return ret;
195	}
196}
197
198#endif /* __HAVE_ARCH_PTE_SPECIAL */
199