1/*
2 * Lockless get_user_pages_fast for sparc, cribbed from powerpc
3 *
4 * Copyright (C) 2008 Nick Piggin
5 * Copyright (C) 2008 Novell Inc.
6 */
7
8#include <linux/sched.h>
9#include <linux/mm.h>
10#include <linux/vmstat.h>
11#include <linux/pagemap.h>
12#include <linux/rwsem.h>
13#include <asm/pgtable.h>
14
15/*
16 * The performance critical leaf functions are made noinline otherwise gcc
17 * inlines everything into a single function which results in too much
18 * register pressure.
19 */
20static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
21		unsigned long end, int write, struct page **pages, int *nr)
22{
23	unsigned long mask, result;
24	pte_t *ptep;
25
26	if (tlb_type == hypervisor) {
27		result = _PAGE_PRESENT_4V|_PAGE_P_4V;
28		if (write)
29			result |= _PAGE_WRITE_4V;
30	} else {
31		result = _PAGE_PRESENT_4U|_PAGE_P_4U;
32		if (write)
33			result |= _PAGE_WRITE_4U;
34	}
35	mask = result | _PAGE_SPECIAL;
36
37	ptep = pte_offset_kernel(&pmd, addr);
38	do {
39		struct page *page, *head;
40		pte_t pte = *ptep;
41
42		if ((pte_val(pte) & mask) != result)
43			return 0;
44		VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
45
46		/* The hugepage case is simplified on sparc64 because
47		 * we encode the sub-page pfn offsets into the
48		 * hugepage PTEs.  We could optimize this in the future
49		 * use page_cache_add_speculative() for the hugepage case.
50		 */
51		page = pte_page(pte);
52		head = compound_head(page);
53		if (!page_cache_get_speculative(head))
54			return 0;
55		if (unlikely(pte_val(pte) != pte_val(*ptep))) {
56			put_page(head);
57			return 0;
58		}
59		if (head != page)
60			get_huge_page_tail(page);
61
62		pages[*nr] = page;
63		(*nr)++;
64	} while (ptep++, addr += PAGE_SIZE, addr != end);
65
66	return 1;
67}
68
69static int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
70			unsigned long end, int write, struct page **pages,
71			int *nr)
72{
73	struct page *head, *page, *tail;
74	int refs;
75
76	if (!(pmd_val(pmd) & _PAGE_VALID))
77		return 0;
78
79	if (write && !pmd_write(pmd))
80		return 0;
81
82	refs = 0;
83	head = pmd_page(pmd);
84	page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
85	tail = page;
86	do {
87		VM_BUG_ON(compound_head(page) != head);
88		pages[*nr] = page;
89		(*nr)++;
90		page++;
91		refs++;
92	} while (addr += PAGE_SIZE, addr != end);
93
94	if (!page_cache_add_speculative(head, refs)) {
95		*nr -= refs;
96		return 0;
97	}
98
99	if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) {
100		*nr -= refs;
101		while (refs--)
102			put_page(head);
103		return 0;
104	}
105
106	/* Any tail page need their mapcount reference taken before we
107	 * return.
108	 */
109	while (refs--) {
110		if (PageTail(tail))
111			get_huge_page_tail(tail);
112		tail++;
113	}
114
115	return 1;
116}
117
118static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
119		int write, struct page **pages, int *nr)
120{
121	unsigned long next;
122	pmd_t *pmdp;
123
124	pmdp = pmd_offset(&pud, addr);
125	do {
126		pmd_t pmd = *pmdp;
127
128		next = pmd_addr_end(addr, end);
129		if (pmd_none(pmd) || pmd_trans_splitting(pmd))
130			return 0;
131		if (unlikely(pmd_large(pmd))) {
132			if (!gup_huge_pmd(pmdp, pmd, addr, next,
133					  write, pages, nr))
134				return 0;
135		} else if (!gup_pte_range(pmd, addr, next, write,
136					  pages, nr))
137			return 0;
138	} while (pmdp++, addr = next, addr != end);
139
140	return 1;
141}
142
143static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
144		int write, struct page **pages, int *nr)
145{
146	unsigned long next;
147	pud_t *pudp;
148
149	pudp = pud_offset(&pgd, addr);
150	do {
151		pud_t pud = *pudp;
152
153		next = pud_addr_end(addr, end);
154		if (pud_none(pud))
155			return 0;
156		if (!gup_pmd_range(pud, addr, next, write, pages, nr))
157			return 0;
158	} while (pudp++, addr = next, addr != end);
159
160	return 1;
161}
162
163int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
164			  struct page **pages)
165{
166	struct mm_struct *mm = current->mm;
167	unsigned long addr, len, end;
168	unsigned long next, flags;
169	pgd_t *pgdp;
170	int nr = 0;
171
172	start &= PAGE_MASK;
173	addr = start;
174	len = (unsigned long) nr_pages << PAGE_SHIFT;
175	end = start + len;
176
177	local_irq_save(flags);
178	pgdp = pgd_offset(mm, addr);
179	do {
180		pgd_t pgd = *pgdp;
181
182		next = pgd_addr_end(addr, end);
183		if (pgd_none(pgd))
184			break;
185		if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
186			break;
187	} while (pgdp++, addr = next, addr != end);
188	local_irq_restore(flags);
189
190	return nr;
191}
192
193int get_user_pages_fast(unsigned long start, int nr_pages, int write,
194			struct page **pages)
195{
196	struct mm_struct *mm = current->mm;
197	unsigned long addr, len, end;
198	unsigned long next;
199	pgd_t *pgdp;
200	int nr = 0;
201
202	start &= PAGE_MASK;
203	addr = start;
204	len = (unsigned long) nr_pages << PAGE_SHIFT;
205	end = start + len;
206
207	/*
208	 * XXX: batch / limit 'nr', to avoid large irq off latency
209	 * needs some instrumenting to determine the common sizes used by
210	 * important workloads (eg. DB2), and whether limiting the batch size
211	 * will decrease performance.
212	 *
213	 * It seems like we're in the clear for the moment. Direct-IO is
214	 * the main guy that batches up lots of get_user_pages, and even
215	 * they are limited to 64-at-a-time which is not so many.
216	 */
217	/*
218	 * This doesn't prevent pagetable teardown, but does prevent
219	 * the pagetables from being freed on sparc.
220	 *
221	 * So long as we atomically load page table pointers versus teardown,
222	 * we can follow the address down to the the page and take a ref on it.
223	 */
224	local_irq_disable();
225
226	pgdp = pgd_offset(mm, addr);
227	do {
228		pgd_t pgd = *pgdp;
229
230		next = pgd_addr_end(addr, end);
231		if (pgd_none(pgd))
232			goto slow;
233		if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
234			goto slow;
235	} while (pgdp++, addr = next, addr != end);
236
237	local_irq_enable();
238
239	VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
240	return nr;
241
242	{
243		int ret;
244
245slow:
246		local_irq_enable();
247
248		/* Try to get the remaining pages with get_user_pages */
249		start += nr << PAGE_SHIFT;
250		pages += nr;
251
252		ret = get_user_pages_unlocked(current, mm, start,
253			(end - start) >> PAGE_SHIFT, write, 0, pages);
254
255		/* Have to be a bit careful with return values */
256		if (nr > 0) {
257			if (ret < 0)
258				ret = nr;
259			else
260				ret += nr;
261		}
262
263		return ret;
264	}
265}
266