Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Lockless get_user_pages_fast for sparc, cribbed from powerpc
4 *
5 * Copyright (C) 2008 Nick Piggin
6 * Copyright (C) 2008 Novell Inc.
7 */
8
9#include <linux/sched.h>
10#include <linux/mm.h>
11#include <linux/vmstat.h>
12#include <linux/pagemap.h>
13#include <linux/rwsem.h>
14#include <asm/pgtable.h>
15#include <asm/adi.h>
16
17/*
18 * The performance critical leaf functions are made noinline otherwise gcc
19 * inlines everything into a single function which results in too much
20 * register pressure.
21 */
22static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
23 unsigned long end, int write, struct page **pages, int *nr)
24{
25 unsigned long mask, result;
26 pte_t *ptep;
27
28 if (tlb_type == hypervisor) {
29 result = _PAGE_PRESENT_4V|_PAGE_P_4V;
30 if (write)
31 result |= _PAGE_WRITE_4V;
32 } else {
33 result = _PAGE_PRESENT_4U|_PAGE_P_4U;
34 if (write)
35 result |= _PAGE_WRITE_4U;
36 }
37 mask = result | _PAGE_SPECIAL;
38
39 ptep = pte_offset_kernel(&pmd, addr);
40 do {
41 struct page *page, *head;
42 pte_t pte = *ptep;
43
44 if ((pte_val(pte) & mask) != result)
45 return 0;
46 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
47
48 /* The hugepage case is simplified on sparc64 because
49 * we encode the sub-page pfn offsets into the
50 * hugepage PTEs. We could optimize this in the future
51 * use page_cache_add_speculative() for the hugepage case.
52 */
53 page = pte_page(pte);
54 head = compound_head(page);
55 if (!page_cache_get_speculative(head))
56 return 0;
57 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
58 put_page(head);
59 return 0;
60 }
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;
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 page = pmd_page(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
84 head = compound_head(page);
85 do {
86 VM_BUG_ON(compound_head(page) != head);
87 pages[*nr] = page;
88 (*nr)++;
89 page++;
90 refs++;
91 } while (addr += PAGE_SIZE, addr != end);
92
93 if (!page_cache_add_speculative(head, refs)) {
94 *nr -= refs;
95 return 0;
96 }
97
98 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) {
99 *nr -= refs;
100 while (refs--)
101 put_page(head);
102 return 0;
103 }
104
105 return 1;
106}
107
108static int gup_huge_pud(pud_t *pudp, pud_t pud, unsigned long addr,
109 unsigned long end, int write, struct page **pages,
110 int *nr)
111{
112 struct page *head, *page;
113 int refs;
114
115 if (!(pud_val(pud) & _PAGE_VALID))
116 return 0;
117
118 if (write && !pud_write(pud))
119 return 0;
120
121 refs = 0;
122 page = pud_page(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
123 head = compound_head(page);
124 do {
125 VM_BUG_ON(compound_head(page) != head);
126 pages[*nr] = page;
127 (*nr)++;
128 page++;
129 refs++;
130 } while (addr += PAGE_SIZE, addr != end);
131
132 if (!page_cache_add_speculative(head, refs)) {
133 *nr -= refs;
134 return 0;
135 }
136
137 if (unlikely(pud_val(pud) != pud_val(*pudp))) {
138 *nr -= refs;
139 while (refs--)
140 put_page(head);
141 return 0;
142 }
143
144 return 1;
145}
146
147static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
148 int write, struct page **pages, int *nr)
149{
150 unsigned long next;
151 pmd_t *pmdp;
152
153 pmdp = pmd_offset(&pud, addr);
154 do {
155 pmd_t pmd = *pmdp;
156
157 next = pmd_addr_end(addr, end);
158 if (pmd_none(pmd))
159 return 0;
160 if (unlikely(pmd_large(pmd))) {
161 if (!gup_huge_pmd(pmdp, pmd, addr, next,
162 write, pages, nr))
163 return 0;
164 } else if (!gup_pte_range(pmd, addr, next, write,
165 pages, nr))
166 return 0;
167 } while (pmdp++, addr = next, addr != end);
168
169 return 1;
170}
171
172static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
173 int write, struct page **pages, int *nr)
174{
175 unsigned long next;
176 pud_t *pudp;
177
178 pudp = pud_offset(&pgd, addr);
179 do {
180 pud_t pud = *pudp;
181
182 next = pud_addr_end(addr, end);
183 if (pud_none(pud))
184 return 0;
185 if (unlikely(pud_large(pud))) {
186 if (!gup_huge_pud(pudp, pud, addr, next,
187 write, pages, nr))
188 return 0;
189 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
190 return 0;
191 } while (pudp++, addr = next, addr != end);
192
193 return 1;
194}
195
196/*
197 * Note a difference with get_user_pages_fast: this always returns the
198 * number of pages pinned, 0 if no pages were pinned.
199 */
200int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
201 struct page **pages)
202{
203 struct mm_struct *mm = current->mm;
204 unsigned long addr, len, end;
205 unsigned long next, flags;
206 pgd_t *pgdp;
207 int nr = 0;
208
209#ifdef CONFIG_SPARC64
210 if (adi_capable()) {
211 long addr = start;
212
213 /* If userspace has passed a versioned address, kernel
214 * will not find it in the VMAs since it does not store
215 * the version tags in the list of VMAs. Storing version
216 * tags in list of VMAs is impractical since they can be
217 * changed any time from userspace without dropping into
218 * kernel. Any address search in VMAs will be done with
219 * non-versioned addresses. Ensure the ADI version bits
220 * are dropped here by sign extending the last bit before
221 * ADI bits. IOMMU does not implement version tags.
222 */
223 addr = (addr << (long)adi_nbits()) >> (long)adi_nbits();
224 start = addr;
225 }
226#endif
227 start &= PAGE_MASK;
228 addr = start;
229 len = (unsigned long) nr_pages << PAGE_SHIFT;
230 end = start + len;
231
232 local_irq_save(flags);
233 pgdp = pgd_offset(mm, addr);
234 do {
235 pgd_t pgd = *pgdp;
236
237 next = pgd_addr_end(addr, end);
238 if (pgd_none(pgd))
239 break;
240 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
241 break;
242 } while (pgdp++, addr = next, addr != end);
243 local_irq_restore(flags);
244
245 return nr;
246}
247
248int get_user_pages_fast(unsigned long start, int nr_pages, int write,
249 struct page **pages)
250{
251 struct mm_struct *mm = current->mm;
252 unsigned long addr, len, end;
253 unsigned long next;
254 pgd_t *pgdp;
255 int nr = 0;
256
257#ifdef CONFIG_SPARC64
258 if (adi_capable()) {
259 long addr = start;
260
261 /* If userspace has passed a versioned address, kernel
262 * will not find it in the VMAs since it does not store
263 * the version tags in the list of VMAs. Storing version
264 * tags in list of VMAs is impractical since they can be
265 * changed any time from userspace without dropping into
266 * kernel. Any address search in VMAs will be done with
267 * non-versioned addresses. Ensure the ADI version bits
268 * are dropped here by sign extending the last bit before
269 * ADI bits. IOMMU does not implements version tags,
270 */
271 addr = (addr << (long)adi_nbits()) >> (long)adi_nbits();
272 start = addr;
273 }
274#endif
275 start &= PAGE_MASK;
276 addr = start;
277 len = (unsigned long) nr_pages << PAGE_SHIFT;
278 end = start + len;
279
280 /*
281 * XXX: batch / limit 'nr', to avoid large irq off latency
282 * needs some instrumenting to determine the common sizes used by
283 * important workloads (eg. DB2), and whether limiting the batch size
284 * will decrease performance.
285 *
286 * It seems like we're in the clear for the moment. Direct-IO is
287 * the main guy that batches up lots of get_user_pages, and even
288 * they are limited to 64-at-a-time which is not so many.
289 */
290 /*
291 * This doesn't prevent pagetable teardown, but does prevent
292 * the pagetables from being freed on sparc.
293 *
294 * So long as we atomically load page table pointers versus teardown,
295 * we can follow the address down to the the page and take a ref on it.
296 */
297 local_irq_disable();
298
299 pgdp = pgd_offset(mm, addr);
300 do {
301 pgd_t pgd = *pgdp;
302
303 next = pgd_addr_end(addr, end);
304 if (pgd_none(pgd))
305 goto slow;
306 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
307 goto slow;
308 } while (pgdp++, addr = next, addr != end);
309
310 local_irq_enable();
311
312 VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
313 return nr;
314
315 {
316 int ret;
317
318slow:
319 local_irq_enable();
320
321 /* Try to get the remaining pages with get_user_pages */
322 start += nr << PAGE_SHIFT;
323 pages += nr;
324
325 ret = get_user_pages_unlocked(start,
326 (end - start) >> PAGE_SHIFT, pages,
327 write ? FOLL_WRITE : 0);
328
329 /* Have to be a bit careful with return values */
330 if (nr > 0) {
331 if (ret < 0)
332 ret = nr;
333 else
334 ret += nr;
335 }
336
337 return ret;
338 }
339}
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
60 pages[*nr] = page;
61 (*nr)++;
62 } while (ptep++, addr += PAGE_SIZE, addr != end);
63
64 return 1;
65}
66
67static int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
68 unsigned long end, int write, struct page **pages,
69 int *nr)
70{
71 struct page *head, *page;
72 int refs;
73
74 if (!(pmd_val(pmd) & _PAGE_VALID))
75 return 0;
76
77 if (write && !pmd_write(pmd))
78 return 0;
79
80 refs = 0;
81 head = pmd_page(pmd);
82 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
83 do {
84 VM_BUG_ON(compound_head(page) != head);
85 pages[*nr] = page;
86 (*nr)++;
87 page++;
88 refs++;
89 } while (addr += PAGE_SIZE, addr != end);
90
91 if (!page_cache_add_speculative(head, refs)) {
92 *nr -= refs;
93 return 0;
94 }
95
96 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) {
97 *nr -= refs;
98 while (refs--)
99 put_page(head);
100 return 0;
101 }
102
103 return 1;
104}
105
106static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
107 int write, struct page **pages, int *nr)
108{
109 unsigned long next;
110 pmd_t *pmdp;
111
112 pmdp = pmd_offset(&pud, addr);
113 do {
114 pmd_t pmd = *pmdp;
115
116 next = pmd_addr_end(addr, end);
117 if (pmd_none(pmd))
118 return 0;
119 if (unlikely(pmd_large(pmd))) {
120 if (!gup_huge_pmd(pmdp, pmd, addr, next,
121 write, pages, nr))
122 return 0;
123 } else if (!gup_pte_range(pmd, addr, next, write,
124 pages, nr))
125 return 0;
126 } while (pmdp++, addr = next, addr != end);
127
128 return 1;
129}
130
131static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
132 int write, struct page **pages, int *nr)
133{
134 unsigned long next;
135 pud_t *pudp;
136
137 pudp = pud_offset(&pgd, addr);
138 do {
139 pud_t pud = *pudp;
140
141 next = pud_addr_end(addr, end);
142 if (pud_none(pud))
143 return 0;
144 if (!gup_pmd_range(pud, addr, next, write, pages, nr))
145 return 0;
146 } while (pudp++, addr = next, addr != end);
147
148 return 1;
149}
150
151int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
152 struct page **pages)
153{
154 struct mm_struct *mm = current->mm;
155 unsigned long addr, len, end;
156 unsigned long next, flags;
157 pgd_t *pgdp;
158 int nr = 0;
159
160 start &= PAGE_MASK;
161 addr = start;
162 len = (unsigned long) nr_pages << PAGE_SHIFT;
163 end = start + len;
164
165 local_irq_save(flags);
166 pgdp = pgd_offset(mm, addr);
167 do {
168 pgd_t pgd = *pgdp;
169
170 next = pgd_addr_end(addr, end);
171 if (pgd_none(pgd))
172 break;
173 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
174 break;
175 } while (pgdp++, addr = next, addr != end);
176 local_irq_restore(flags);
177
178 return nr;
179}
180
181int get_user_pages_fast(unsigned long start, int nr_pages, int write,
182 struct page **pages)
183{
184 struct mm_struct *mm = current->mm;
185 unsigned long addr, len, end;
186 unsigned long next;
187 pgd_t *pgdp;
188 int nr = 0;
189
190 start &= PAGE_MASK;
191 addr = start;
192 len = (unsigned long) nr_pages << PAGE_SHIFT;
193 end = start + len;
194
195 /*
196 * XXX: batch / limit 'nr', to avoid large irq off latency
197 * needs some instrumenting to determine the common sizes used by
198 * important workloads (eg. DB2), and whether limiting the batch size
199 * will decrease performance.
200 *
201 * It seems like we're in the clear for the moment. Direct-IO is
202 * the main guy that batches up lots of get_user_pages, and even
203 * they are limited to 64-at-a-time which is not so many.
204 */
205 /*
206 * This doesn't prevent pagetable teardown, but does prevent
207 * the pagetables from being freed on sparc.
208 *
209 * So long as we atomically load page table pointers versus teardown,
210 * we can follow the address down to the the page and take a ref on it.
211 */
212 local_irq_disable();
213
214 pgdp = pgd_offset(mm, addr);
215 do {
216 pgd_t pgd = *pgdp;
217
218 next = pgd_addr_end(addr, end);
219 if (pgd_none(pgd))
220 goto slow;
221 if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
222 goto slow;
223 } while (pgdp++, addr = next, addr != end);
224
225 local_irq_enable();
226
227 VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
228 return nr;
229
230 {
231 int ret;
232
233slow:
234 local_irq_enable();
235
236 /* Try to get the remaining pages with get_user_pages */
237 start += nr << PAGE_SHIFT;
238 pages += nr;
239
240 ret = get_user_pages_unlocked(start,
241 (end - start) >> PAGE_SHIFT, write, 0, pages);
242
243 /* Have to be a bit careful with return values */
244 if (nr > 0) {
245 if (ret < 0)
246 ret = nr;
247 else
248 ret += nr;
249 }
250
251 return ret;
252 }
253}