1/*
2 *  Lockless get_user_pages_fast for s390
3 *
4 *  Copyright IBM Corp. 2010
5 *  Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
6 */
7#include <linux/sched.h>
8#include <linux/mm.h>
9#include <linux/hugetlb.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 inline int gup_pte_range(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
21		unsigned long end, int write, struct page **pages, int *nr)
22{
23	unsigned long mask;
24	pte_t *ptep, pte;
25	struct page *page;
26
27	mask = (write ? _PAGE_PROTECT : 0) | _PAGE_INVALID | _PAGE_SPECIAL;
28
29	ptep = ((pte_t *) pmd_deref(pmd)) + pte_index(addr);
30	do {
31		pte = *ptep;
32		barrier();
33		/* Similar to the PMD case, NUMA hinting must take slow path */
34		if (pte_protnone(pte))
35			return 0;
36		if ((pte_val(pte) & mask) != 0)
37			return 0;
38		VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
39		page = pte_page(pte);
40		if (!page_cache_get_speculative(page))
41			return 0;
42		if (unlikely(pte_val(pte) != pte_val(*ptep))) {
43			put_page(page);
44			return 0;
45		}
46		pages[*nr] = page;
47		(*nr)++;
48
49	} while (ptep++, addr += PAGE_SIZE, addr != end);
50
51	return 1;
52}
53
54static inline int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
55		unsigned long end, int write, struct page **pages, int *nr)
56{
57	unsigned long mask, result;
58	struct page *head, *page, *tail;
59	int refs;
60
61	result = write ? 0 : _SEGMENT_ENTRY_PROTECT;
62	mask = result | _SEGMENT_ENTRY_INVALID;
63	if ((pmd_val(pmd) & mask) != result)
64		return 0;
65	VM_BUG_ON(!pfn_valid(pmd_val(pmd) >> PAGE_SHIFT));
66
67	refs = 0;
68	head = pmd_page(pmd);
69	page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
70	tail = page;
71	do {
72		VM_BUG_ON(compound_head(page) != head);
73		pages[*nr] = page;
74		(*nr)++;
75		page++;
76		refs++;
77	} while (addr += PAGE_SIZE, addr != end);
78
79	if (!page_cache_add_speculative(head, refs)) {
80		*nr -= refs;
81		return 0;
82	}
83
84	if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) {
85		*nr -= refs;
86		while (refs--)
87			put_page(head);
88		return 0;
89	}
90
91	/*
92	 * Any tail page need their mapcount reference taken before we
93	 * return.
94	 */
95	while (refs--) {
96		if (PageTail(tail))
97			get_huge_page_tail(tail);
98		tail++;
99	}
100
101	return 1;
102}
103
104
105static inline int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr,
106		unsigned long end, int write, struct page **pages, int *nr)
107{
108	unsigned long next;
109	pmd_t *pmdp, pmd;
110
111	pmdp = (pmd_t *) pudp;
112	if ((pud_val(pud) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R3)
113		pmdp = (pmd_t *) pud_deref(pud);
114	pmdp += pmd_index(addr);
115	do {
116		pmd = *pmdp;
117		barrier();
118		next = pmd_addr_end(addr, end);
119		/*
120		 * The pmd_trans_splitting() check below explains why
121		 * pmdp_splitting_flush() has to serialize with
122		 * smp_call_function() against our disabled IRQs, to stop
123		 * this gup-fast code from running while we set the
124		 * splitting bit in the pmd. Returning zero will take
125		 * the slow path that will call wait_split_huge_page()
126		 * if the pmd is still in splitting state.
127		 */
128		if (pmd_none(pmd) || pmd_trans_splitting(pmd))
129			return 0;
130		if (unlikely(pmd_large(pmd))) {
131			/*
132			 * NUMA hinting faults need to be handled in the GUP
133			 * slowpath for accounting purposes and so that they
134			 * can be serialised against THP migration.
135			 */
136			if (pmd_protnone(pmd))
137				return 0;
138			if (!gup_huge_pmd(pmdp, pmd, addr, next,
139					  write, pages, nr))
140				return 0;
141		} else if (!gup_pte_range(pmdp, pmd, addr, next,
142					  write, pages, nr))
143			return 0;
144	} while (pmdp++, addr = next, addr != end);
145
146	return 1;
147}
148
149static inline int gup_pud_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr,
150		unsigned long end, int write, struct page **pages, int *nr)
151{
152	unsigned long next;
153	pud_t *pudp, pud;
154
155	pudp = (pud_t *) pgdp;
156	if ((pgd_val(pgd) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R2)
157		pudp = (pud_t *) pgd_deref(pgd);
158	pudp += pud_index(addr);
159	do {
160		pud = *pudp;
161		barrier();
162		next = pud_addr_end(addr, end);
163		if (pud_none(pud))
164			return 0;
165		if (!gup_pmd_range(pudp, pud, addr, next, write, pages, nr))
166			return 0;
167	} while (pudp++, addr = next, addr != end);
168
169	return 1;
170}
171
172/*
173 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
174 * back to the regular GUP.
175 */
176int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
177			  struct page **pages)
178{
179	struct mm_struct *mm = current->mm;
180	unsigned long addr, len, end;
181	unsigned long next, flags;
182	pgd_t *pgdp, pgd;
183	int nr = 0;
184
185	start &= PAGE_MASK;
186	addr = start;
187	len = (unsigned long) nr_pages << PAGE_SHIFT;
188	end = start + len;
189	if ((end <= start) || (end > TASK_SIZE))
190		return 0;
191	/*
192	 * local_irq_save() doesn't prevent pagetable teardown, but does
193	 * prevent the pagetables from being freed on s390.
194	 *
195	 * So long as we atomically load page table pointers versus teardown,
196	 * we can follow the address down to the the page and take a ref on it.
197	 */
198	local_irq_save(flags);
199	pgdp = pgd_offset(mm, addr);
200	do {
201		pgd = *pgdp;
202		barrier();
203		next = pgd_addr_end(addr, end);
204		if (pgd_none(pgd))
205			break;
206		if (!gup_pud_range(pgdp, pgd, addr, next, write, pages, &nr))
207			break;
208	} while (pgdp++, addr = next, addr != end);
209	local_irq_restore(flags);
210
211	return nr;
212}
213
214/**
215 * get_user_pages_fast() - pin user pages in memory
216 * @start:	starting user address
217 * @nr_pages:	number of pages from start to pin
218 * @write:	whether pages will be written to
219 * @pages:	array that receives pointers to the pages pinned.
220 *		Should be at least nr_pages long.
221 *
222 * Attempt to pin user pages in memory without taking mm->mmap_sem.
223 * If not successful, it will fall back to taking the lock and
224 * calling get_user_pages().
225 *
226 * Returns number of pages pinned. This may be fewer than the number
227 * requested. If nr_pages is 0 or negative, returns 0. If no pages
228 * were pinned, returns -errno.
229 */
230int get_user_pages_fast(unsigned long start, int nr_pages, int write,
231			struct page **pages)
232{
233	struct mm_struct *mm = current->mm;
234	int nr, ret;
235
236	start &= PAGE_MASK;
237	nr = __get_user_pages_fast(start, nr_pages, write, pages);
238	if (nr == nr_pages)
239		return nr;
240
241	/* Try to get the remaining pages with get_user_pages */
242	start += nr << PAGE_SHIFT;
243	pages += nr;
244	ret = get_user_pages_unlocked(current, mm, start,
245			     nr_pages - nr, write, 0, pages);
246	/* Have to be a bit careful with return values */
247	if (nr > 0)
248		ret = (ret < 0) ? nr : ret + nr;
249	return ret;
250}
251